Align the COPYRIGHT and LICENSE files to the house standard

AGENTS.md

@@ -1,2 +1,5 @@
 Please review `docs/development.md` for initial orientation in this repository
 and follow its task-specific reading guide.
+
+Review and follow `docs/policy/agents.md` before recommending, planning, or
+implementing any changes to this repository.

COPYRIGHT

@@ -0,0 +1,19 @@
+PromptKit
+
+Copyright (C) 2026 Eric Rakestraw
+
+SPDX-License-Identifier: LGPL-3.0-or-later
+
+This library is free software: you can redistribute it and/or modify it
+under the terms of the GNU Lesser General Public License as published by
+the Free Software Foundation, either version 3 of the License, or
+(at your option) any later version.
+
+The GNU Lesser General Public License version 3 is provided in LICENSE.
+
+This library is distributed in the hope that it will be useful,
+but WITHOUT ANY WARRANTY; without even the implied warranty of
+MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the
+GNU Lesser General Public License for more details.
+
+Third-party code and assets are subject to their respective license terms.

README.md

@@ -33,6 +33,34 @@ boundary and constraints that framework work must preserve.
 
 ## Release Guidance
 
+Consumers upgrading to `v0.11.0` should read the
+[v0.11.0 changelog and migration guide](docs/releases/v0.11.0.md), particularly
+the default retry behavior and license change.
+
+Consumers upgrading to the `v0.10.0` profile-policy transition
+should read the [v0.10.0 transition guide](docs/releases/v0.10.0.md).
+
+Consumers upgrading from `v0.8.0` to `v0.9.0` should read the
+[v0.9.0 changelog and migration guide](docs/releases/v0.9.0.md).
+
+Consumers upgrading from `v0.7.0` to `v0.8.0` can consult the
+[v0.8.0 changelog and migration guide](docs/releases/v0.8.0.md).
+
+Earlier adopters can consult the
+[v0.7.0 changelog and migration guide](docs/releases/v0.7.0.md).
+
+Consumers upgrading from `v0.5.0` to `v0.6.0` can consult the
+[v0.6.0 changelog and migration guide](docs/releases/v0.6.0.md).
+
+Consumers upgrading from `v0.4.0` to `v0.5.0` can consult the
+[v0.5.0 changelog and migration guide](docs/releases/v0.5.0.md).
+
+Consumers upgrading from `v0.3.0` to `v0.4.0` should read the
+[v0.4.0 changelog and adoption guide](docs/releases/v0.4.0.md).
+
+Consumers upgrading from `v0.2.0` to `v0.3.0` should read the
+[v0.3.0 changelog](docs/releases/v0.3.0.md).
+
 Consumers moving from `v0.1.0` to `v0.2.0` should read the
 [v0.2.0 changelog and migration guide](docs/releases/v0.2.0.md).
 
@@ -41,4 +69,18 @@ Consumers moving from `v0.1.0` to `v0.2.0` should read the
 [Scriptorium](https://gitea.maximumdirect.net/eric/scriptorium) is the CLI and
 HTTP application built on Promptkit.
 
-Promptkit is licensed under the [GNU General Public License version 3](LICENSE).
+## License
+
+Promptkit is licensed under the [GNU Lesser General Public License version 3
+or later](LICENSE) (`LGPL-3.0-or-later`). The accompanying [GPLv3 text](COPYING)
+is incorporated by LGPLv3. This maintainer-authorized change applies from this
+revision forward; existing release tags and their license notices are unchanged.
+Dependencies, including the backend catalogs, retain their own licenses.
+
+Applications may use their own licenses subject to the LGPL's combined-work
+conditions. Distributors of statically linked Go binaries must account for the
+applicable notices, license copies, library source, and materials and permissions
+needed to recombine or relink with a modified library; installation information
+may also be required. Providing library source alone is not necessarily enough.
+See section 4 of the [LGPL text](LICENSE) for the governing requirements, rather
+than treating this summary as a complete distribution checklist.

appended_messages_contract_test.go

@@ -0,0 +1,145 @@
+package promptkit_test
+
+import (
+	"context"
+	"errors"
+	"reflect"
+	"sync/atomic"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+func TestAppendedMessageRoleConstantsAreStrings(t *testing.T) {
+	var (
+		developer string = promptkit.RoleDeveloper
+		system    string = promptkit.RoleSystem
+		user      string = promptkit.RoleUser
+		assistant string = promptkit.RoleAssistant
+	)
+	if developer != "developer" || system != "system" || user != "user" || assistant != "assistant" {
+		t.Fatalf("unexpected role constants: %q %q %q %q", developer, system, user, assistant)
+	}
+}
+
+func TestAppendedMessagesComposeFrozenEffectivePrompt(t *testing.T) {
+	cacheControl := &promptkit.CacheControl{Type: promptkit.CacheControlEphemeral, TTL: "1h"}
+	appended := []promptkit.RenderedMessage{
+		{Role: " \tAsSiStAnT\n", Content: "previous response"},
+		{Role: promptkit.RoleUser, Content: "corrective request", CacheControl: cacheControl},
+	}
+	client := &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "output"}}
+	engine, err := promptkit.NewEngine(
+		promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "configured prompt"), "."),
+		promptkit.WithProfiles(promptkit.Profile{ID: "profile", Endpoint: "http://example.test/v1", Model: "model"}),
+		promptkit.WithLLMClient(client),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+	request := promptkit.RunRequest{PromptID: "prompt", AppendedMessages: appended}
+
+	plain, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("prepare plain request: %v", err)
+	}
+	prepared, err := engine.Prepare(context.Background(), request)
+	if err != nil {
+		t.Fatalf("prepare appended request: %v", err)
+	}
+	expected := []promptkit.RenderedMessage{
+		{Role: promptkit.RoleUser, Content: "configured prompt"},
+		{Role: promptkit.RoleAssistant, Content: "previous response"},
+		{Role: promptkit.RoleUser, Content: "corrective request", CacheControl: &promptkit.CacheControl{Type: promptkit.CacheControlEphemeral, TTL: "1h"}},
+	}
+	if !reflect.DeepEqual(prepared.Messages, expected) || prepared.PromptHash != plain.PromptHash || prepared.RenderedPromptHash == plain.RenderedPromptHash {
+		t.Fatalf("prepared effective prompt = %#v, plain = %#v", prepared, plain)
+	}
+
+	equivalent, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt", AppendedMessages: []promptkit.RenderedMessage{
+		{Role: promptkit.RoleAssistant, Content: "previous response"},
+		{Role: promptkit.RoleUser, Content: "corrective request", CacheControl: &promptkit.CacheControl{Type: promptkit.CacheControlEphemeral, TTL: "1h"}},
+	}})
+	if err != nil || equivalent.RenderedPromptHash != prepared.RenderedPromptHash {
+		t.Fatalf("normalized-equivalent request = (%#v, %v), want matching rendered hash %q", equivalent, err, prepared.RenderedPromptHash)
+	}
+
+	for _, variant := range []promptkit.RunRequest{
+		{PromptID: "prompt", AppendedMessages: []promptkit.RenderedMessage{{Role: promptkit.RoleAssistant, Content: "changed"}, expected[2]}},
+		{PromptID: "prompt", AppendedMessages: []promptkit.RenderedMessage{expected[2], expected[1]}},
+		{PromptID: "prompt", AppendedMessages: []promptkit.RenderedMessage{{Role: promptkit.RoleAssistant, Content: "previous response"}, {Role: promptkit.RoleUser, Content: "corrective request"}}},
+	} {
+		variantPrepared, prepareErr := engine.Prepare(context.Background(), variant)
+		if prepareErr != nil || variantPrepared.RenderedPromptHash == prepared.RenderedPromptHash || variantPrepared.PromptHash != prepared.PromptHash {
+			t.Fatalf("variant preparation = (%#v, %v)", variantPrepared, prepareErr)
+		}
+	}
+
+	result, err := engine.Run(context.Background(), request)
+	if err != nil || result.RenderedPromptHash != prepared.RenderedPromptHash || !reflect.DeepEqual(client.requests[0].Prompt.Messages, expected) {
+		t.Fatalf("run result = (%#v, %v), request = %#v", result, err, client.requests)
+	}
+
+	execution, err := engine.PrepareExecution(context.Background(), request)
+	if err != nil {
+		t.Fatalf("prepare execution: %v", err)
+	}
+	appended[0].Content = "changed caller content"
+	cacheControl.TTL = ""
+	firstDetails := execution.Details()
+	firstDetails.Messages[1].Content = "changed details content"
+	firstDetails.Messages[2].CacheControl.Type = "changed details cache"
+	secondDetails := execution.Details()
+	if !reflect.DeepEqual(secondDetails.Messages, expected) || secondDetails.RenderedPromptHash != prepared.RenderedPromptHash {
+		t.Fatalf("prepared execution details = %#v, want frozen %#v", secondDetails, expected)
+	}
+
+	result, err = engine.RunPrepared(context.Background(), execution)
+	if err != nil || result.RenderedPromptHash != prepared.RenderedPromptHash || !reflect.DeepEqual(client.requests[1].Prompt.Messages, expected) {
+		t.Fatalf("prepared result = (%#v, %v), requests = %#v", result, err, client.requests)
+	}
+}
+
+func TestInvalidAppendedMessagesFailBeforeSourceOrModelWork(t *testing.T) {
+	promptSource := &inspectionCountingFS{}
+	var modelCalls atomic.Int64
+	engine, err := promptkit.NewEngine(
+		promptkit.Config{},
+		promptkit.WithPromptFS(promptSource, "."),
+		promptkit.WithLLMClient(countingLLMClient{calls: &modelCalls}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+	request := promptkit.RunRequest{
+		PromptID: "unreached",
+		AppendedMessages: []promptkit.RenderedMessage{{
+			Role:    "unsupported-role",
+			Content: "sensitive appended content",
+		}},
+	}
+
+	operations := []struct {
+		name string
+		run  func() error
+	}{
+		{name: "Prepare", run: func() error { _, err := engine.Prepare(context.Background(), request); return err }},
+		{name: "PrepareExecution", run: func() error { _, err := engine.PrepareExecution(context.Background(), request); return err }},
+		{name: "Run", run: func() error { _, err := engine.Run(context.Background(), request); return err }},
+	}
+	for _, operation := range operations {
+		t.Run(operation.name, func(t *testing.T) {
+			err := operation.run()
+			if !errors.Is(err, promptkit.ErrInvalidRequest) {
+				t.Fatalf("error = %v, want ErrInvalidRequest", err)
+			}
+		})
+	}
+	if promptSource.opens.Load() != 0 {
+		t.Fatalf("invalid appended message opened prompt sources %d times", promptSource.opens.Load())
+	}
+	if modelCalls.Load() != 0 {
+		t.Fatalf("invalid appended message invoked the model %d times", modelCalls.Load())
+	}
+}

backends.go

@@ -9,50 +9,83 @@ import (
 // backend.
 const BackendOpenRouter = backend.OpenRouterID
 
+// BackendRakestrawHome is the reserved ID of Promptkit's built-in
+// Rakestrawhome backend.
+const BackendRakestrawHome = backend.RakestrawHomeID
+
+// BackendLocal is the case-sensitive conventional ID used by [LocalBackend].
+// It is not a built-in or reserved backend and must be registered with
+// [WithBackend].
+const BackendLocal = "local"
+
 // Backend configures one engine-scoped OpenAI-compatible backend.
 //
 // Backend has no stable JSON representation. Use keyed literals so additions
 // to this configuration value do not break source compatibility.
 type Backend struct {
 	// ID is the stable, case-sensitive registry key. NewEngine trims it and
-	// requires a non-blank value. BackendOpenRouter is reserved.
+	// requires a non-blank value. Built-in backend IDs are reserved.
 	ID string
 	// Endpoint is the OpenAI-compatible base endpoint. NewEngine trims it and
 	// requires an absolute HTTP or HTTPS URL with a host and without user
 	// information, a query string, or a fragment. Paths are allowed.
 	Endpoint string
-	// APIKeyEnv optionally names the environment variable containing the API
+	// APIKeyEnv optionally names an environment lookup source for an API key.
-	// key. NewEngine trims it and requires the portable form
+	// NewEngine trims it and requires the portable form [A-Za-z_][A-Za-z0-9_]*.
-	// [A-Za-z_][A-Za-z0-9_]*. Store only the name, never a credential value.
+	// A direct RunRequest.APIKey takes precedence. When no usable credential is
+	// available, the built-in client omits Authorization; injected clients own
+	// their own credential-resolution behavior. Store only the name, never a
+	// credential value.
 	APIKeyEnv string
 	// ExtraParams contains backend-wide request defaults. Values must be
 	// JSON-compatible, finite, acyclic, and keyed by non-empty strings. Keys
 	// must not be model, session_id, messages, temperature, max_tokens, top_p,
 	// service_tier, reasoning_effort, or response_format. An empty map supplies
-	// no defaults. NewEngine deeply copies the map.
+	// no defaults. NewEngine deeply copies the map and rejects excessively deep
+	// or large values for safety.
 	ExtraParams map[string]any
 	// ConcurrencyLimit is the maximum number of simultaneous model-generation
 	// calls allowed for this backend within one Engine. Zero leaves the backend
 	// unlimited. A negative value makes NewEngine fail with ErrInvalidConfig.
 	ConcurrencyLimit int
-	// QueueCapacity controls how many additional Run calls may be admitted
+	// QueueCapacity controls how many additional Run or RunPrepared calls may
-	// beyond ConcurrencyLimit. Nil uses 1024 when ConcurrencyLimit is positive;
+	// be admitted beyond ConcurrencyLimit. Nil uses 1024 when ConcurrencyLimit
-	// a pointer uses its exact value, including zero. The pointed-to value must
+	// is positive; a pointer uses its exact value, including zero. The pointed-to
-	// be non-negative, and QueueCapacity must be nil when ConcurrencyLimit is
+	// value must be non-negative, and QueueCapacity must be nil when
-	// zero. Their sum must fit in an int. WithBackend copies the value and does
+	// ConcurrencyLimit is zero. Their sum must fit in an int. WithBackend copies
-	// not retain the pointer.
+	// the value and does not retain the pointer.
 	QueueCapacity *int
 }
 
+// LocalBackend returns a caller-owned Backend for a conventional local
+// OpenAI-compatible endpoint. It sets ID to BackendLocal and copies endpoint
+// and concurrencyLimit into Endpoint and ConcurrencyLimit without
+// normalization or validation. APIKeyEnv, ExtraParams, and QueueCapacity keep
+// their zero values.
+//
+// LocalBackend does not read environment variables, register the value, or
+// mutate engine or package state. Supply the returned value through
+// [WithBackend]; [NewEngine] then applies the ordinary backend validation and
+// concurrency semantics, including default queue capacity for a positive
+// limit, unlimited behavior for zero, and ErrInvalidConfig for a negative
+// limit.
+func LocalBackend(endpoint string, concurrencyLimit int) Backend {
+	return Backend{
+		ID:               BackendLocal,
+		Endpoint:         endpoint,
+		ConcurrencyLimit: concurrencyLimit,
+	}
+}
+
 // WithBackend adds one Backend registration to the constructed Engine.
 //
 // Registrations accumulate in option order. Every normalized ID must be unique
 // across consumer registrations and built-ins; a duplicate or invalid
-// definition makes NewEngine fail with ErrInvalidConfig. In particular,
+// definition makes NewEngine fail with ErrInvalidConfig. Built-in IDs,
-// BackendOpenRouter cannot be replaced. The immutable registration is scoped
+// including [BackendOpenRouter] and [BackendRakestrawHome], cannot be
-// to the resulting Engine and cannot be enumerated, replaced, removed, or
+// replaced. The immutable registration is scoped to the resulting Engine and
-// mutated after construction. WithBackend does not install package-global
+// cannot be enumerated, replaced, removed, or mutated after construction.
-// state.
+// WithBackend does not install package-global state.
 func WithBackend(backend Backend) Option {
 	queueCapacity := 0
 	queueCapacitySet := backend.QueueCapacity != nil
@@ -60,7 +93,7 @@ func WithBackend(backend Backend) Option {
 		queueCapacity = *backend.QueueCapacity
 	}
 	return optionFunc(func(options *engineOptions) error {
-		options.backends = append(options.backends, domain.Backend{
+		options.internal.AddBackend(domain.Backend{
 			ID:               backend.ID,
 			Endpoint:         backend.Endpoint,
 			APIKeyEnv:        backend.APIKeyEnv,

capacity_contract_test.go

@@ -60,7 +60,18 @@ func TestEngineRejectsRunBeforeCompletionWhenAdmissionIsFull(t *testing.T) {
 
 	awaitCapacitySignal(t, reader.entered, "first artifact read")
 
-	result, err := engine.Run(context.Background(), capacityInputRequest("http://second.example/v1"))
+	canceledContext, cancel := context.WithCancel(context.Background())
+	cancel()
+	result, err := engine.Run(canceledContext, capacityInputRequest("http://canceled.example/v1"))
+	if result != nil || !errors.Is(err, context.Canceled) {
+		t.Fatalf("canceled capacity admission=(%+v, %v), want context cancellation", result, err)
+	}
+	var canceledCapacityErr *promptkit.CapacityError
+	if errors.Is(err, promptkit.ErrCapacityExceeded) || errors.As(err, &canceledCapacityErr) {
+		t.Fatalf("canceled admission exposed capacity rejection: %v", err)
+	}
+
+	result, err = engine.Run(context.Background(), capacityInputRequest("http://second.example/v1"))
 	if result != nil {
 		t.Fatalf("capacity rejection returned partial result: %+v", result)
 	}
@@ -70,6 +81,21 @@ func TestEngineRejectsRunBeforeCompletionWhenAdmissionIsFull(t *testing.T) {
 	if errors.Is(err, promptkit.ErrInvalidRequest) || errors.Is(err, promptkit.ErrLLMGenerate) {
 		t.Fatalf("capacity rejection had an unrelated category: %v", err)
 	}
+	var capacityErr *promptkit.CapacityError
+	if !errors.As(err, &capacityErr) || capacityErr == nil {
+		t.Fatalf("capacity rejection=%v, want CapacityError", err)
+	}
+	if capacityErr.BackendID != "limited" {
+		t.Fatalf("capacity backend ID=%q, want limited", capacityErr.BackendID)
+	}
+	capacityErr.BackendID = "changed"
+
+	result, err = engine.Run(context.Background(), capacityInputRequest("http://third.example/v1"))
+	var subsequentCapacityErr *promptkit.CapacityError
+	if result != nil || !errors.As(err, &subsequentCapacityErr) ||
+		subsequentCapacityErr == nil || subsequentCapacityErr.BackendID != "limited" {
+		t.Fatalf("subsequent capacity rejection=(%+v, %v), want independent limited CapacityError", result, err)
+	}
 	if calls := reader.callCount(); calls != 1 {
 		t.Fatalf("artifact calls=%d, want only the admitted run", calls)
 	}
@@ -174,6 +200,19 @@ func TestCapacityExceededSentinelContract(t *testing.T) {
 	if promptkit.ErrCapacityExceeded == nil {
 		t.Fatal("ErrCapacityExceeded is nil")
 	}
+	var nilCapacityErr *promptkit.CapacityError
+	zeroCapacityErr := &promptkit.CapacityError{}
+	populatedCapacityErr := &promptkit.CapacityError{BackendID: "limited"}
+	for _, capacityErr := range []error{nilCapacityErr, zeroCapacityErr} {
+		if !errors.Is(capacityErr, promptkit.ErrCapacityExceeded) {
+			t.Fatalf("capacity error=%v, want ErrCapacityExceeded", capacityErr)
+		}
+	}
+	var discoveredCapacityErr *promptkit.CapacityError
+	if !errors.As(populatedCapacityErr, &discoveredCapacityErr) || discoveredCapacityErr != populatedCapacityErr {
+		t.Fatalf("populated capacity error is not discoverable: %v", populatedCapacityErr)
+	}
+
 	for _, unrelated := range []error{
 		promptkit.ErrInvalidConfig,
 		promptkit.ErrInvalidRequest,
@@ -181,7 +220,8 @@ func TestCapacityExceededSentinelContract(t *testing.T) {
 		promptkit.ErrValidation,
 	} {
 		if errors.Is(promptkit.ErrCapacityExceeded, unrelated) ||
-			errors.Is(unrelated, promptkit.ErrCapacityExceeded) {
+			errors.Is(unrelated, promptkit.ErrCapacityExceeded) ||
+			errors.Is(populatedCapacityErr, unrelated) {
 			t.Fatalf("ErrCapacityExceeded aliases unrelated sentinel %v", unrelated)
 		}
 	}

capacity_error.go

@@ -0,0 +1,40 @@
+package promptkit
+
+import (
+	"fmt"
+	"strings"
+)
+
+// CapacityError reports bounded admission rejected for a selected backend.
+//
+// Engine-produced values identify only rejection at Promptkit's bounded
+// [Engine.Run] or [Engine.RunPrepared] admission boundary. BackendID is the
+// normalized registered backend ID used for routing and capacity; endpoint
+// overrides do not change it. Every engine-produced value is nonnil and has a
+// nonblank BackendID. Provider errors, active-generation waiting, and caller
+// cancellation are not represented by this type.
+//
+// Callers own returned values and may mutate BackendID without affecting engine
+// state or another error. CapacityError and its default Go encoding have no
+// stable JSON contract. Consumer-constructed values do not establish that an
+// engine rejected work.
+type CapacityError struct {
+	// BackendID is the normalized registered backend ID whose admission was
+	// rejected.
+	BackendID string
+}
+
+// Error returns diagnostic wording that is not a parsing contract. It is safe
+// to call on a nil receiver or a value with a blank BackendID.
+func (e *CapacityError) Error() string {
+	if e == nil || strings.TrimSpace(e.BackendID) == "" {
+		return ErrCapacityExceeded.Error()
+	}
+	return fmt.Sprintf("backend %q admission: %v", e.BackendID, ErrCapacityExceeded)
+}
+
+// Unwrap returns ErrCapacityExceeded so errors.Is and errors.As can be used
+// together. It is safe to call on a nil receiver or a zero value.
+func (e *CapacityError) Unwrap() error {
+	return ErrCapacityExceeded
+}

convert.go

@@ -1,7 +1,9 @@
 package promptkit
 
 import (
+	"fmt"
 	"reflect"
+	"unicode/utf8"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
@@ -12,19 +14,59 @@ func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
 	if err != nil {
 		return domain.RunRequest{}, err
 	}
+	appendedMessages, err := toDomainAppendedMessages(req.AppendedMessages)
+	if err != nil {
+		return domain.RunRequest{}, err
+	}
+	var observer domain.ExecutionObserver
+	if req.Observer != nil {
+		callback := req.Observer
+		observer = func(event domain.ExecutionEvent) { callback(fromDomainExecutionEvent(event)) }
+	}
 	return domain.RunRequest{
-		PromptID:      req.PromptID,
+		PromptID:         req.PromptID,
-		PromptVersion: req.PromptVersion,
+		PromptVersion:    req.PromptVersion,
-		ProfileID:     req.ProfileID,
+		ProfileID:        req.ProfileID,
-		SessionID:     req.SessionID,
+		SessionID:        req.SessionID,
-		APIKey:        req.APIKey,
+		APIKey:           req.APIKey,
-		Inputs:        toDomainArtifactRefMap(req.Inputs),
+		Inputs:           toDomainArtifactRefMap(req.Inputs),
-		Vars:          copyStringMap(req.Vars),
+		Vars:             copyStringMap(req.Vars),
-		Execution:     execution,
+		Execution:        execution,
-		Validation:    toDomainOutputContractPtr(req.Validation),
+		Validation:       toDomainOutputContractPtr(req.Validation),
+		AppendedMessages: appendedMessages,
+		TransientRetries: copyIntPtr(req.TransientRetries),
+		Observer:         observer,
 	}, nil
 }
 
+func toDomainAppendedMessages(messages []RenderedMessage) ([]domain.RenderedMessage, error) {
+	if messages == nil {
+		return nil, nil
+	}
+
+	converted := make([]domain.RenderedMessage, len(messages))
+	for index, message := range messages {
+		if !utf8.ValidString(message.Content) {
+			return nil, fmt.Errorf("appended message %d content must be valid UTF-8", index)
+		}
+		role, err := domain.NormalizeMessageRole(message.Role)
+		if err != nil {
+			return nil, fmt.Errorf("appended message %d role: %w", index, err)
+		}
+
+		cacheControl, err := domain.NormalizeCacheControl(toDomainCacheControl(message.CacheControl))
+		if err != nil {
+			return nil, fmt.Errorf("appended message %d cache_control: %w", index, err)
+		}
+		converted[index] = domain.RenderedMessage{
+			Role:         role,
+			Content:      message.Content,
+			CacheControl: cacheControl,
+		}
+	}
+	return converted, nil
+}
+
 func fromDomainPreparedRun(prepared *domain.PreparedRun) *PreparedRun {
 	if prepared == nil {
 		return nil
@@ -90,8 +132,22 @@ func toDomainGenerateResponse(resp *GenerateResponse) *domain.GenerateResponse {
 		return nil
 	}
 	return &domain.GenerateResponse{
-		Content: resp.Content,
+		Content:       resp.Content,
-		Usage:   toDomainTokenUsage(resp.Usage),
+		Usage:         toDomainTokenUsage(resp.Usage),
+		UsagePresence: toDomainTokenUsagePresence(resp.UsagePresence),
+	}
+}
+
+func toDomainTokenUsagePresence(presence *TokenUsagePresence) *domain.TokenUsagePresence {
+	if presence == nil {
+		return nil
+	}
+	return &domain.TokenUsagePresence{
+		PromptTokens:     presence.PromptTokens,
+		CompletionTokens: presence.CompletionTokens,
+		TotalTokens:      presence.TotalTokens,
+		CachedTokens:     presence.CachedTokens,
+		CacheWriteTokens: presence.CacheWriteTokens,
 	}
 }
 
@@ -170,6 +226,40 @@ func fromDomainExecutionTarget(target domain.ExecutionTarget) ExecutionTarget {
 	}
 }
 
+func fromDomainProfileInspection(inspection *domain.ProfileInspection) *ProfileInspection {
+	if inspection == nil {
+		return nil
+	}
+	return &ProfileInspection{
+		ProfileID:            inspection.ProfileID,
+		EffectiveModelParams: fromDomainExecutionTarget(inspection.EffectiveModelParams),
+		APIKeyRequired:       inspection.APIKeyRequired,
+	}
+}
+
+func fromDomainPromptInspection(inspection *domain.PromptInspection) *PromptInspection {
+	if inspection == nil {
+		return nil
+	}
+	inputs := make([]PromptInputDefinition, len(inspection.Inputs))
+	for i, input := range inspection.Inputs {
+		inputs[i] = PromptInputDefinition{
+			Name:        input.Name,
+			Required:    input.Required,
+			ContentType: input.ContentType,
+			Description: input.Description,
+		}
+	}
+	return &PromptInspection{
+		PromptID:         inspection.PromptID,
+		PromptVersion:    inspection.PromptVersion,
+		PromptHash:       inspection.PromptHash,
+		DefaultProfileID: inspection.DefaultProfileID,
+		Inputs:           inputs,
+		OutputContract:   fromDomainOutputContract(inspection.OutputContract),
+	}
+}
+
 func fromDomainExecutionTargetPresence(presence domain.ExecutionTargetPresence) ExecutionTargetPresence {
 	return ExecutionTargetPresence{
 		Temperature:    presence.Temperature,
@@ -251,6 +341,16 @@ func fromDomainRenderedMessages(messages []domain.RenderedMessage) []RenderedMes
 	return out
 }
 
+func toDomainCacheControl(cacheControl *CacheControl) *domain.CacheControl {
+	if cacheControl == nil {
+		return nil
+	}
+	return &domain.CacheControl{
+		Type: domain.CacheControlType(cacheControl.Type),
+		TTL:  cacheControl.TTL,
+	}
+}
+
 func fromDomainCacheControl(cacheControl *domain.CacheControl) *CacheControl {
 	if cacheControl == nil {
 		return nil

convert_appended_messages_test.go

@@ -0,0 +1,108 @@
+package promptkit
+
+import (
+	"reflect"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestToDomainAppendedMessages(t *testing.T) {
+	tests := []struct {
+		name        string
+		messages    []RenderedMessage
+		want        []domain.RenderedMessage
+		wantErr     string
+		privateRole string
+		privateText string
+	}{
+		{
+			name: "normalizes supported roles",
+			messages: []RenderedMessage{
+				{Role: " DeVeLoPeR ", Content: "developer"},
+				{Role: "SyStEm", Content: "system"},
+				{Role: "\tUsEr\n", Content: "user"},
+				{Role: "assistant", Content: "assistant"},
+			},
+			want: []domain.RenderedMessage{
+				{Role: domain.RoleDeveloper, Content: "developer"},
+				{Role: domain.RoleSystem, Content: "system"},
+				{Role: domain.RoleUser, Content: "user"},
+				{Role: domain.RoleAssistant, Content: "assistant"},
+			},
+		},
+		{
+			name:     "invalid role UTF-8",
+			messages: []RenderedMessage{{Role: string([]byte{0xff}), Content: "private-content"}},
+			wantErr:  "appended message 0 role",
+		},
+		{
+			name:     "invalid content UTF-8",
+			messages: []RenderedMessage{{Role: "private-role", Content: string([]byte{0xff})}},
+			wantErr:  "appended message 0 content",
+		},
+		{
+			name:        "unsupported role",
+			messages:    []RenderedMessage{{Role: "private-role", Content: "private-content"}},
+			wantErr:     "appended message 0 role",
+			privateRole: "private-role",
+			privateText: "private-content",
+		},
+		{
+			name:        "invalid cache control",
+			messages:    []RenderedMessage{{Role: RoleUser, Content: "private-content", CacheControl: &CacheControl{Type: "private-cache"}}},
+			wantErr:     "appended message 0 cache_control",
+			privateText: "private-content",
+		},
+		{
+			name: "empty and whitespace content",
+			messages: []RenderedMessage{
+				{Role: RoleUser, Content: ""},
+				{Role: RoleAssistant, Content: " \t\n "},
+			},
+			want: []domain.RenderedMessage{
+				{Role: domain.RoleUser, Content: ""},
+				{Role: domain.RoleAssistant, Content: " \t\n "},
+			},
+		},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := toDomainAppendedMessages(test.messages)
+			if test.wantErr != "" {
+				if err == nil || !strings.Contains(err.Error(), test.wantErr) {
+					t.Fatalf("error = %v, want %q", err, test.wantErr)
+				}
+				for _, privateValue := range []string{test.privateRole, test.privateText} {
+					if privateValue != "" && strings.Contains(err.Error(), privateValue) {
+						t.Fatalf("error exposed appended message data %q: %v", privateValue, err)
+					}
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("toDomainAppendedMessages() error = %v", err)
+			}
+			if !reflect.DeepEqual(got, test.want) {
+				t.Fatalf("toDomainAppendedMessages() = %#v, want %#v", got, test.want)
+			}
+		})
+	}
+}
+
+func TestToDomainAppendedMessagesCopiesCacheControl(t *testing.T) {
+	cacheControl := &CacheControl{Type: CacheControlEphemeral, TTL: "1h"}
+	messages := []RenderedMessage{{Role: RoleUser, Content: "content", CacheControl: cacheControl}}
+	request, err := toDomainRunRequest(RunRequest{AppendedMessages: messages})
+	if err != nil {
+		t.Fatalf("toDomainRunRequest() error = %v", err)
+	}
+
+	messages[0].Content = "changed"
+	cacheControl.TTL = ""
+	if request.AppendedMessages[0].Content != "content" || request.AppendedMessages[0].CacheControl.TTL != "1h" {
+		t.Fatalf("domain request did not retain an independent appended-message copy: %#v", request.AppendedMessages)
+	}
+}

doc.go

@@ -3,23 +3,29 @@
 //
 // Applications construct an [Engine] with [NewEngine], select filesystem or
 // in-memory sources and optional engine-scoped [Backend] registrations, and
-// call [Engine.Prepare] or [Engine.Run]. Concrete registries, repositories,
+// call [Engine.InspectPrompt], [Engine.InspectProfile], [Engine.Prepare],
-// validators, and the built-in OpenAI-compatible client remain internal
+// [Engine.PrepareExecution], [Engine.Run], or [Engine.RunPrepared]. Concrete
-// implementation details.
+// registries, repositories, validators, and the built-in OpenAI-compatible
+// client remain internal implementation details.
 //
 // # Concurrency and ownership
 //
-// An Engine supports concurrent Prepare and Run calls. Engine-local backend
+// An Engine supports concurrent InspectPrompt, InspectProfile, Prepare,
-// policies bound admitted Run calls and model generations where configured,
+// PrepareExecution, Run, and RunPrepared calls. Engine-local backend policies
-// while different backend pools and unlimited backends continue independently.
+// bound admitted Run and RunPrepared calls and model generations where
-// An injected [LLMClient] or [ArtifactReader] can therefore still receive
+// configured, while different backend pools and unlimited backends continue
-// concurrent calls and must be safe for that use.
+// independently. An injected [LLMClient] or [ArtifactReader] can therefore
+// still receive concurrent calls and must be safe for that use.
 //
-// NewEngine copies in-memory profiles and backend definitions. Prepare and Run
+// NewEngine copies in-memory profiles and backend definitions. Prepare,
-// copy request maps, slices, pointer values, and JSON-compatible extra
+// PrepareExecution, and Run copy request maps, slices, pointer values, and
-// parameters before using them. Returned values and values passed to extension
+// JSON-compatible extra parameters before using them. InspectPrompt and
-// interfaces are likewise isolated from engine state. Callers own those copies
+// InspectProfile return copied inspection values. Returned values and values
-// and may mutate them after the call that supplied or returned them.
+// passed to extension interfaces are likewise isolated from engine state.
+// Callers own those copies and may mutate them after the call that supplied or
+// returned them. [CapacityError] values are caller-owned and may be mutated
+// without affecting engine state or another error. Immutable [GenerationError]
+// values are also caller-owned and do not retain shared engine state.
 //
 // # Security and sensitive data
 //
@@ -45,10 +51,18 @@
 // [GenerateResponse], [ExecutionTargetPresence], and the string value types
 // used by those values.
 //
-// Construction values, including [Config], [Backend], [RunRequest],
+// Construction, inspection, handle, and error values, including [Config],
-// [ArtifactRef], [ExecutionTargetOverride], [Profile], and
+// [Backend], [RunRequest], [ArtifactRef], [ExecutionTargetOverride], [Profile],
-// [OpenAICompatibleProfileConfig], do not have stable JSON representations.
+// [OpenAICompatibleProfileConfig], [ProfileInspection],
-// Direct API keys are nevertheless excluded from JSON for every public value.
+// [PromptInputDefinition], [PromptInspection], [PreparedExecution],
+// [CapacityError], [GenerationError], [RetryExhaustedError], and
+// [ExecutionEvent], do not have stable JSON
+// representations. Direct API keys are nevertheless excluded from JSON for
+// every public value.
+
+// Provider-derived [GenerationError] accessor values are untrusted and can
+// contain sensitive request or schema fragments. Applications must apply their
+// own disclosure policy before logging, displaying, or returning them.
 //
 // JSON timestamps use time.Time's RFC 3339 encoding and are omitted when zero.
 // PreparedRun and RunResult durations are encoded as integer milliseconds in

docs/consumers/pkg-promptkit.md

@@ -40,11 +40,67 @@ validation, and default transport behavior. Source discovery, format
 validation, and profile precedence are defined by the
 [framework format reference](../formats.md).
 
+## Supply Embedded Application Defaults
+
+Use `WithFallbackProfileFS` when an application packages profile definitions
+that should apply unless an operator provides an ordinary configured profile
+with the same ID. For example, an application can embed its defaults while
+continuing to use `ProfileDir` for operator overrides:
+
+```go
+//go:embed profiles/*.yaml
+var applicationProfiles embed.FS
+
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir:  "prompts",
+	ProfileDir: operatorProfileDir,
+},
+	promptkit.WithFallbackProfileFS(applicationProfiles, "profiles"),
+)
+```
+
+Keep application-owned profile IDs and definitions in the embedded source.
+Use the ordinary configured profile source for operator overrides. Leave
+`operatorProfileDir` empty when the operator did not configure an override
+directory; a non-empty path names an authoritative higher-precedence source,
+so an unavailable or unreadable directory is an error rather than a reason to
+fall back. The
+[framework format reference](../formats.md#source-and-profile-precedence)
+owns the exact profile format and lookup order; the
+[`WithFallbackProfileFS` GoDoc](../../engine.go) owns its option contract and
+validation rules.
+
+## Inspect A Prompt Before Preparation
+
+Use [`Engine.InspectPrompt`](../../engine.go) to check one configured prompt's
+declared inputs and output workflow without creating placeholder inputs or
+resolving a profile:
+
+```go
+inspection, err := engine.InspectPrompt(ctx, "meeting.summary", "")
+if err != nil {
+	return err
+}
+
+for _, input := range inspection.Inputs {
+	// Compare the declared input with application configuration.
+}
+```
+
+Use this configuration-time boundary when the application needs only the
+declared prompt interface. Use `InspectProfile` separately when it must also
+check a configured profile. Use `Prepare` when it needs inputs, schemas, or
+rendered messages, and use prepared execution when that work must remain tied
+to later execution. The method's [GoDoc](../../engine.go) owns exact fields,
+hash, ownership, and error semantics.
+
 ## Prepare Without Model Execution
 
 [`Engine.Prepare`](../../engine.go) resolves the selected prompt and profile,
 loads inputs and any structured-output schema, and renders messages without
-calling a model client:
+calling a model client. Choose it when the prepared value is the final
+inspection or persistence result and no later execution must be tied to that
+exact snapshot:
 
 ```go
 prepared, err := engine.Prepare(ctx, promptkit.RunRequest{
@@ -61,12 +117,56 @@ shows a complete runnable setup with a prompt file, in-memory profile, and
 inline input. Exact request requirements and prepared-result fields belong to
 the [`RunRequest` and `PreparedRun` GoDoc](../../types.go).
 
+## Prepare Now And Execute The Same Snapshot Later
+
+Use [`Engine.PrepareExecution`](../../engine.go) when an application must
+inspect or persist preflight details before deciding whether to start model
+work, while ensuring that later execution uses those exact rendered messages,
+inputs, target settings, and validation resources:
+
+```go
+preparedExecution, err := engine.PrepareExecution(ctx, promptkit.RunRequest{
+	PromptID: "meeting.summary",
+	Inputs: map[string]promptkit.ArtifactRef{
+		"note": promptkit.Inline("Synthetic meeting notes"),
+	},
+})
+if err != nil {
+	return err
+}
+defer preparedExecution.Discard()
+
+details := preparedExecution.Details()
+// Inspect or persist an application-selected safe subset of details.
+
+result, err := engine.RunPrepared(ctx, preparedExecution)
+```
+
+Preparation does not call the model or reserve backend capacity.
+`RunPrepared` executes from the retained snapshot rather than reloading
+consumer sources. The handle is opaque in-process state, while `Details`
+contains rendered content and remains subject to the application's data
+handling policy. The
+[`PreparedExecution` and method GoDoc](../../prepared_execution.go) and
+[engine operation GoDoc](../../engine.go) own exact lifecycle, engine-binding,
+credential, cancellation, timing, and error semantics.
+
+If the request supplied `Observer`, successful preparation retains that
+function for `RunPrepared`; preparation itself emits nothing. Observation
+starts only when the owning engine successfully claims the handle. Reused,
+discarded, foreign, and invalid handles emit nothing. `Discard` clears the
+retained callback without invoking it. The function's closure is retained by
+identity, so synchronize any shared state it uses. Prepared start events can
+include identities already resolved in the frozen snapshot.
+
 ## Execute And Validate
 
 [`Engine.Run`](../../engine.go) performs the same preparation, invokes the
 configured model client, classifies the generated artifact, and validates the
-content. A completed content check may return `ValidationFailed` in the result;
+content in one call. Choose it when the application does not need a preflight
-an operational inability to validate returns an error.
+boundary tied to the eventual execution. A completed content check may return
+`ValidationFailed` in the result; an operational inability to validate returns
+an error.
 
 The maintained
 [offline execution example](../../examples/go-library/run/main.go) injects a
@@ -81,6 +181,98 @@ semantics. The
 [OpenAI-compatible integration contract](../integrations/openai-compatible-chat.md)
 owns the built-in client's outbound HTTP behavior.
 
+### Observe An Ordinary Run
+
+Set `RunRequest.Observer` to receive synchronous metadata events from `Run`:
+
+```go
+request.Observer = func(event promptkit.ExecutionEvent) {
+	// Translate selected metadata into an application-owned record.
+}
+result, err := engine.Run(ctx, request)
+```
+
+An ordinary run emits `started` before request conversion, followed by any
+attempt and retry-delay completions, then one `finished` event on a normal
+return. `Prepare` and `PrepareExecution` emit no events. A start event has no
+resolved prompt, profile, or model; later events gain identities only as
+selection succeeds. The terminal event has dispatched-attempt, transient-retry,
+repair-call, duration, and presence-aware usage totals. Its usage is an
+aggregate, not an additional billable attempt. A known zero differs from an
+unknown count, and incomplete totals must not be treated as complete cost
+accounting. A failed operation still returns no partial `RunResult`.
+An injected client that leaves `UsagePresence` nil can report known nonzero
+counts, but its zero counts remain unknown to observations; supply explicit
+presence when measured zeros matter. Existing `RunResult.Usage` sums returned
+values independently of observation presence.
+
+Callbacks run synchronously, in order for one operation, and can overlap for
+different operations. Synchronize shared callback state and avoid blocking on
+work that needs the same run's admission slot. Callback panics propagate;
+there is no durable event history or promise of a terminal event after panic.
+Events omit content and error objects, but selected identifiers may still be
+sensitive. See the [observation GoDoc](../../observations.go) for exact fields.
+The [offline proposal and assessment example](../../examples/go-library/observe/main.go)
+shows application-owned event collection and a separate semantic acceptance
+decision using the built-in HTTP path without network access.
+
+### Repair A Structured Result
+
+Set a small additional-call budget when a structurally invalid result can be
+corrected automatically:
+
+```go
+request.Validation = &promptkit.OutputContract{
+	Format:         promptkit.FormatJSON,
+	ValidationMode: promptkit.ValidationJSONSchema,
+	SchemaPath:     "events.schema.json",
+	RepairAttempts: 1,
+}
+```
+
+Each repair attempt is another model generation, so it can increase latency and
+usage; `RunResult.Usage` is cumulative and `Validation.RepairAttempts` reports
+repair generations rather than transient HTTP attempts. Exhaustion still
+returns the final failed validation result. `basic` validation can also repair
+an empty candidate, but structural
+validity is not evidence of factual or domain correctness. See the
+[output-contract format reference](../formats.md#output-contract) and
+[`OutputContract` GoDoc](../../types.go) for the exact budget and eligibility
+rules.
+
+The built-in client also retries eligible temporary HTTP statuses. By default,
+each generation can make two additional attempts; a repair generation receives
+its own budget. Set `Config.TransientRetries` to a pointer to zero for the
+previous single-attempt behavior, or set `RunRequest.TransientRetries` to
+override one execution. A custom HTTP transport still uses this policy;
+injected `LLMClient` implementations retain their own retry policy. Application
+semantic retries are separate runs and can multiply provider traffic further.
+The caller context covers capacity and retry waiting as well as attempts;
+without a deadline a valid provider delay can be long. Promptkit does not
+infer exactly-once or complete cost accounting from a retry outcome.
+See the [provider integration contract](../integrations/openai-compatible-chat.md#transient-status-retries)
+for eligible statuses, waits, timeouts, and provider timing.
+
+### Append Already-Rendered Messages
+
+An application can include an earlier assistant response and its own corrective
+instruction in a fresh request without changing the configured prompt:
+
+```go
+request.AppendedMessages = []promptkit.RenderedMessage{
+	{Role: promptkit.RoleAssistant, Content: previousResponse},
+	{Role: promptkit.RoleUser, Content: correction},
+}
+result, err := engine.Run(ctx, request)
+```
+
+These messages are already rendered: Promptkit does not template or resolve
+files in them, and they can contain sensitive model output or application
+feedback. Promptkit remains stateless; every `Run` call re-resolves its current
+sources and the application owns any semantic retry budget. When a
+pre-execution equality check is required, use `PrepareExecution` and compare
+its opaque rendered-prompt hash before invoking `RunPrepared`.
+
 ## Inputs, Profiles, And Overrides
 
 Use `File`, `Inline`, or `InlineWithURI` to construct request inputs. A request
@@ -90,13 +282,115 @@ replace execution settings or the complete output contract.
 The [public value GoDoc](../../types.go) defines nil, empty, zero, replacement,
 copy, and credential behavior. The
 [framework format reference](../formats.md) defines how those request values
-interact with prompt definitions, file-backed profiles, built-ins, schemas,
+interact with prompt definitions, file-backed and application fallback
-and framework defaults.
+profiles, built-ins, schemas, and framework defaults.
 
 For programmatic profiles,
 [`OpenAICompatibleProfile`](../../profiles.go) converts ordinary
 OpenAI-compatible settings into a value accepted by `WithProfiles`.
 
+### Choose A Profile Ownership Model
+
+Choose the profile source that matches the application's continuity needs:
+
+- Select `promptkit-openrouter-standard` or `promptkit-openrouter-auto` when
+  the maintained literal name is the desired consumer contract. Their current
+  selections are in the [format reference](../formats.md#built-in-profile-catalog).
+- Select an external catalog profile directly when the application accepts that
+  profile's catalog volatility. The selected external inventories are linked
+  from the [format reference](../formats.md#external-profile-inventories).
+- Define a complete application-owned profile with an explicit backend or
+  endpoint, model, and settings when the application owns those choices.
+
+An application profile that inherits an external base follows later changes to
+that base and fails if the base disappears; it is not a frozen copy. Promptkit
+tests the catalog versions selected in its module graph. Independently advancing
+a catalog dependency can remove an inherited base and produce an untested
+combination. Pinning the resolved module graph preserves the embedded data for
+a build, but cannot preserve provider availability. Running binaries neither
+download catalog updates nor automatically replace withdrawn models. See the
+[release policy](../policy/release.md#catalog-profile-compatibility) for the exact
+cross-version guarantees.
+
+### Alias A Built-In Profile
+
+Give an application-owned profile ID a built-in base when prompts should select
+the application ID while inheriting the built-in target. The child can override
+only the setting it owns:
+
+```go
+promptkit.WithProfiles(promptkit.Profile{
+	ID:              "weather-light",
+	BaseProfileID:   "promptkit-openrouter-standard",
+	ReasoningEffort: "high",
+})
+```
+
+Select `weather-light` in a prompt or `RunRequest.ProfileID`; it remains the
+reported selected profile. See the [profile inheritance format
+reference](../formats.md#profile-inheritance) and the
+[`Profile` GoDoc](../../types.go) for exact lookup, merging, and validation
+behavior. If the base is an external profile, apply the volatility guidance
+above when deciding whether to use an application-owned complete definition
+instead.
+
+### Use The Rakestrawhome Built-In Profile
+
+Set `RAKESTRAWHOME_INFERENCE_API_KEY` in the application environment, then
+select `rakestrawhome-gemma-4-31b` as an ordinary profile ID. For example, a
+prepared result identifies the selected built-in through
+`BackendRakestrawHome`:
+
+```go
+prepared, err := engine.Prepare(ctx, promptkit.RunRequest{
+	PromptID:  "meeting.summary",
+	ProfileID: "rakestrawhome-gemma-4-31b",
+	Inputs:    inputs,
+})
+if err != nil {
+	return err
+}
+if prepared.SelectedBackendID != promptkit.BackendRakestrawHome {
+	return fmt.Errorf("unexpected backend %q", prepared.SelectedBackendID)
+}
+```
+
+Do not register `rakestrawhome` manually. When adopting this built-in, remove
+an existing `WithBackend` registration with that exact ID; retaining it causes
+the intentional duplicate-ID configuration error. Direct request credentials
+and runtime endpoint overrides remain supported under their ordinary GoDoc and
+format contracts.
+
+### Inspect A Profile Before Prompt Work
+
+Use [`Engine.InspectProfile`](../../engine.go) to validate one configured
+profile without constructing a synthetic prompt or placeholder inputs. It
+resolves the profile's effective target but does not prepare or execute a
+prompt:
+
+```go
+inspection, err := engine.InspectProfile(ctx, profileID)
+if err != nil {
+	return err
+}
+
+target := inspection.EffectiveModelParams
+if target.APIKeyEnv != "" {
+	// This is a configured optional environment lookup source.
+} else if inspection.APIKeyRequired {
+	// Arrange a direct credential before later execution.
+}
+```
+
+Use this configuration-time boundary when only the profile and its target need
+checking. Use `Prepare` when the application also needs prompt, input, schema,
+or rendering work; use prepared execution when that work must remain tied to a
+later execution. A reported `APIKeyEnv` is a configured optional source, while
+`APIKeyRequired` is the explicit local requirement. The
+[credential format reference](../formats.md#credentials) and the method's
+[GoDoc](../../engine.go) own the exact precedence, timing, result, and error
+contracts.
+
 ### Set A Per-Run Session And Reasoning
 
 Supply a direct session ID when one prompt should be correlated with a
@@ -124,35 +418,86 @@ providers. The
 [`RunRequest` and `ExecutionTargetOverride` GoDoc](../../types.go) owns the
 exact normalization, precedence, error, copying, and exposure contract.
 
-### Register A Custom Backend
+### Configure A Local OpenAI-Compatible Endpoint
 
-Register a reusable OpenAI-compatible connection once, then select it from a
+Choose the smallest configuration that fits how the endpoint will be reused.
-profile. This local backend limits model generation to two simultaneous calls;
+
-because `QueueCapacity` is omitted, the engine admits up to 1024 additional
+#### Use An Endpoint-Only Profile
-calls waiting behind them:
+
+Put the endpoint directly on an in-memory profile when only that profile needs
+it and shared backend identity or capacity policy is unnecessary:
 
 ```go
 engine, err := promptkit.NewEngine(promptkit.Config{
 	PromptDir: "prompts",
 },
-	promptkit.WithBackend(promptkit.Backend{
+	promptkit.WithProfiles(promptkit.Profile{
-		ID:        "local",
+		ID:       "local-summary",
-		Endpoint:  "http://localhost:8000/v1",
+		Endpoint: "http://localhost:8000/v1",
-		APIKeyEnv: "LOCAL_LLM_API_KEY",
+		Model:    "example-model",
-		ConcurrencyLimit: 2,
 	}),
+)
+```
+
+Endpoint-only profiles have an empty backend ID and remain unrestricted by
+backend capacity policy.
+
+#### Use The Conventional Local Backend
+
+Use `LocalBackend` when profiles should share the conventional `local`
+identity, endpoint, and concurrency limit:
+
+```go
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir: "prompts",
+},
+	promptkit.WithBackend(
+		promptkit.LocalBackend("http://localhost:8000/v1", 2),
+	),
 	promptkit.WithProfiles(promptkit.Profile{
 		ID:        "local-summary",
-		BackendID: "local",
+		BackendID: promptkit.BackendLocal,
 		Model:     "example-model",
 	}),
 )
 ```
 
+The helper is explicit: it does not pre-register a backend or read environment
+variables. Supplying a positive limit leaves queue capacity omitted, so normal
+backend registration selects the existing default waiting capacity of 1024.
+The returned value still enters the engine through `WithBackend`.
+
+#### Configure A Complete Backend
+
+Use a keyed `Backend` value for authentication, extra request parameters, an
+explicit queue capacity, a custom ID, or multiple local endpoints:
+
+```go
+noWaiting := 0
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir: "prompts",
+},
+	promptkit.WithBackend(promptkit.Backend{
+		ID:               "local-gpu",
+		Endpoint:         "http://gpu-host:8000/v1",
+		APIKeyEnv:        "LOCAL_GPU_API_KEY",
+		ExtraParams:      map[string]any{"provider_option": "enabled"},
+		ConcurrencyLimit: 2,
+		QueueCapacity:    &noWaiting,
+	}),
+	promptkit.WithProfiles(promptkit.Profile{
+		ID:        "gpu-summary",
+		BackendID: "local-gpu",
+		Model:     "example-model",
+	}),
+)
+```
+
+Use distinct custom IDs when registering multiple local endpoints.
 Registrations belong to one engine and custom IDs cannot replace built-ins.
-The [`Backend` and `WithBackend` GoDoc](../../backends.go) defines validation,
+The [`Backend`, `LocalBackend`, and `WithBackend` GoDoc](../../backends.go)
-copying, uniqueness, exact concurrency-field semantics, and request-default
+defines exact construction, validation, copying, uniqueness, concurrency, and
-behavior.
+request-default behavior.
 
 Both file-backed and in-memory profiles select a registration through
 `backend` or `Profile.BackendID`. Profile and request endpoint overrides retain
@@ -173,8 +518,8 @@ zero:
 ```go
 noWaiting := 0
 backend := promptkit.Backend{
-	ID:               "local",
+	ID:               "local-gpu",
-	Endpoint:         "http://localhost:8000/v1",
+	Endpoint:         "http://gpu-host:8000/v1",
 	ConcurrencyLimit: 2,
 	QueueCapacity:    &noWaiting,
 }
@@ -236,6 +581,11 @@ When a limited backend has admitted all active and waiting calls, handle
 ```go
 result, err := engine.Run(ctx, request)
 if errors.Is(err, promptkit.ErrCapacityExceeded) {
+	var capacityErr *promptkit.CapacityError
+	if errors.As(err, &capacityErr) {
+		// Record capacityErr.BackendID using application-owned diagnostics.
+	}
+
 	// Apply application policy: shed work, report overload, or retry later.
 }
 ```
@@ -243,8 +593,41 @@ if errors.Is(err, promptkit.ErrCapacityExceeded) {
 A rejected call returns no partial result and does not invoke the model
 client. Promptkit does not prescribe retries or map this error to an HTTP
 status; those choices remain with the consuming application. The
-[`Engine.Run` and error GoDoc](../../engine.go) owns exact error and
+[`CapacityError` GoDoc](../../capacity_error.go) owns the exact typed-error
-cancellation identities.
+contract, while the [`Engine.Run` and error GoDoc](../../engine.go) owns broad
+error and cancellation identities.
+
+For a non-2xx response from the built-in OpenAI-compatible client, inspect the
+status and deliberately selected provider diagnostic when useful:
+
+```go
+var generationErr *promptkit.GenerationError
+if errors.As(err, &generationErr) {
+	status := generationErr.StatusCode()
+	message := generationErr.ProviderMessage()
+	_, _ = status, message // Apply application presentation policy.
+}
+```
+
+All provider fields are untrusted and can contain sensitive request or schema
+fragments. Do not log, display, or return them without an application-specific
+disclosure policy. Promptkit does not assign presentation behavior.
+The [`GenerationError` GoDoc](../../generation_error.go) owns the exact typed
+error contract.
+
+Generation failures also match `ErrTransport` when the built-in HTTP request
+fails, or `ErrMalformedResponse` when a successful provider response cannot be
+decoded or an injected client returns a nil successful response. Both retain
+`ErrLLMGenerate`; transport failures retain their original causes for
+`errors.Is` and `errors.As`. Caller cancellation and deadlines retain their
+standard context identities. Arbitrary injected-client errors retain their own
+identities without Promptkit guessing a failure category from their text.
+An eligible built-in status that consumes a positive retry budget additionally
+matches `ErrRetriesExhausted`; use `errors.As` to read
+`RetryExhaustedError.Attempts()` and the final `GenerationError` status. With
+explicit zero retries, the first eligible status is an ordinary generation
+error, not exhaustion. Completed structural rejection remains a normal result
+with failed validation, while an operational error returns no partial result.
 
 ## Application Boundary
 

docs/development.md

@@ -2,7 +2,8 @@
 
 This is the contributor entry point for Promptkit, a reusable Go library. All
 contributors must read the
-[architecture policy](policy/architecture.md) before making changes.
+[architecture policy](policy/architecture.md) before making changes. Agents must
+also follow the [agent policy](policy/agents.md) before recommendations or work.
 
 ## Initial Orientation
 
@@ -25,7 +26,8 @@ Start with:
   documentation;
 - the [testing policy](policy/testing.md) before adding, rewriting, or deleting
   tests; and
-- the [release procedure](release.md) for version and publication work.
+- the [release policy](policy/release.md) and [release procedure](release.md)
+  for version and publication work.
 
 ## Task-Specific Reading Guide
 
@@ -34,108 +36,54 @@ Start with:
 | Root public API | The [architecture policy](policy/architecture.md), [consumer guide](consumers/pkg-promptkit.md), [testing policy](policy/testing.md), and existing GoDoc. |
 | Prompt, profile, or schema formats | The [framework format reference](formats.md), owning parser or validator package, and [documentation policy](policy/documentation.md). |
 | Source loading or validation | The [framework format reference](formats.md), [internal source document](internal/sources.md), and owning package tests. |
+| Maintained external catalogs or built-ins | The [internal source document](internal/sources.md), [framework format reference](formats.md), [testing policy](policy/testing.md), both catalog module repositories, each catalog's release procedure, and the selected module versions in `go.mod`. |
 | Model-client behavior | The [OpenAI-compatible integration contract](integrations/openai-compatible-chat.md), [internal model-client document](internal/llm.md), and owning package tests. |
 | Internal package implementation | The [architecture policy](policy/architecture.md), [internal component overview](internal/overview.md), and focused internal document listed for that package. |
 | Tests or test fixtures | The [testing policy](policy/testing.md), owning package, and focused internal document listed by the component overview. |
-| Maintained example | The [example](../examples/go-library/prepare/main.go), [consumer guide](consumers/pkg-promptkit.md), [framework format reference](formats.md), and [documentation policy](policy/documentation.md). |
+| Maintained example | The [preparation](../examples/go-library/prepare/main.go), [execution](../examples/go-library/run/main.go), and [observation](../examples/go-library/observe/main.go) examples, [consumer guide](consumers/pkg-promptkit.md), [framework format reference](formats.md), and [documentation policy](policy/documentation.md). |
 | Documentation | The [documentation policy](policy/documentation.md) and canonical owner of every affected contract. |
-| Release preparation or publication | The [release procedure](release.md). |
+| Release preparation or publication | The [release policy](policy/release.md) and [release procedure](release.md). |
 
 For cross-cutting changes, follow every applicable row. Do not create
 placeholder documents for packages, APIs, or integrations that do not yet
 exist.
 
-## Maintainer-Run Validation
+## Maintainer Validation
 
-Promptkit does not currently use hosted CI. Maintainers are responsible for
+For ordinary work, select validation proportionate to the change. Do not repeat
-running the documented checks before accepting changes. Run the default Go
+successful checks for unchanged content merely to create a commit. The full
-validation from the Promptkit repository root:
+release-candidate checks have one executable entry point from the repository root:
 
 ```sh
-go test ./...
+./scripts/release-check.sh
-go test -race ./...
-go vet ./...
-go build ./...
-go run ./examples/go-library/prepare
 ```
 
-Check formatting across every tracked Go file:
+The [script](../scripts/release-check.sh) owns the executable checks:
+race-enabled tests, vet, the three offline consumer examples with semantic output
+assertions, tracked Go formatting, local Markdown links, module/workspace
+hygiene, whitespace, and the credential-pattern scan. It does not publish,
+change files, validate release-note structure, or retire roadmaps. It requires Go,
+Git, a POSIX shell, and Python 3. Tests and examples require no provider
+credentials or live inference calls.
 
-```sh
+The script can run on intended uncommitted changes during development.
-gofmt -l $(git ls-files '*.go')
+Formatting and documentation checks use tracked files; stage intended new
-```
+files before final validation so those checks include them. The release
+procedure separately enforces a clean, committed candidate.
 
-The formatting command must produce no paths. Follow every added or changed
+### Change Review
-Markdown link and confirm its target exists. Finally, check whitespace:
 
-```sh
+Inspect `git status --short --untracked-files=all` and the complete diff before
-git diff --check
+accepting a change. Reject unintended files, credentials, private data,
-```
+generated binaries, test output, downloaded assets, and template residue.
+The automated credential scan does not replace review of the actual change.
 
-Documentation-only work does not require unrelated new tests, but it still
+For releases, follow the [release procedure](release.md). Its publication
-requires link validation and `git diff --check`. Run the Go validation whenever
+command runs the full local checks once on the final committed candidate and
-documentation changes commands, examples, generated output, or another
+automates release-note structure and Git checks. Do not repeat ordinary change
-behavior checked by the module.
+review as a release sign-off. Roadmap retirement remains preparation housekeeping.
 
-## Focused Validation
+Follow the [agent commit policy](policy/agents.md#git-commit-policy) for local
-
+commits. Ordinary commits may leave unrelated work in the workspace; preserve
-Use focused checks while iterating, then run the complete validation sequence
+and report it. A clean worktree and index are required for release candidates,
-before accepting the change. The root package supports:
+as enforced by the publication command, not for every ordinary commit.
-
-```sh
-go test .
-go vet .
-go build .
-```
-
-Filter tests by name without assuming a fixed internal package layout:
-
-```sh
-go test ./... -run 'TestName'
-```
-
-Replace `TestName` with a useful regular expression. Target only paths that
-exist, and consult the internal component overview for their owning
-documentation. A filtered or package-specific run does not replace the
-complete repository validation.
-
-## Coordinated Work With Scriptorium
-
-Promptkit and Scriptorium must remain independently valid. For temporary local
-integration, use either a Go workspace outside both repositories or an
-uncommitted replacement in the consuming module.
-
-If the repositories are sibling directories, run the workspace commands from
-their parent directory:
-
-```sh
-go work init ./promptkit ./scriptorium
-go work sync
-```
-
-Use the workspace only for coordinated local checks. From the same parent
-directory, remove it when finished:
-
-```sh
-rm -f go.work go.work.sum
-```
-
-Alternatively, from the Scriptorium repository root, temporarily point its
-Promptkit dependency at the sibling checkout:
-
-```sh
-go mod edit -replace gitea.maximumdirect.net/eric/promptkit=../promptkit
-```
-
-After coordinated checks, remove the replacement and reconcile module
-metadata:
-
-```sh
-go mod edit -dropreplace gitea.maximumdirect.net/eric/promptkit
-go mod tidy
-```
-
-Never commit `go.work`, `go.work.sum`, or a local filesystem `replace`
-directive. Before committing in either repository, inspect its module files and
-working tree independently. Published consumer versions must depend on a tagged
-Promptkit version, not a workspace, local replacement, or unpublished commit.

docs/formats.md

@@ -9,9 +9,10 @@ explains how to select these sources and invoke the engine. The
 owns the resulting outbound wire behavior.
 
 Prompt and profile sources recursively discover files ending in `.yaml` or
-`.yml`. YAML decoding is strict: unknown fields are errors for the selected
+`.yml`. Each prompt-definition and profile file contains exactly one YAML
-definition. Definitions are selected by their YAML `id`, not their file name
+document; comments and trailing whitespace are allowed. YAML decoding is
-or directory.
+strict: unknown fields are errors for the selected definition. Definitions are
+selected by their YAML `id`, not their file name or directory.
 
 ## Prompt Definitions
 
@@ -58,6 +59,11 @@ When a request omits a version, the selected prompt ID must identify exactly
 one definition. When it supplies a version, the ID and version pair must be
 unique.
 
+Exact prompt inspection uses this same configured source, strict decoding,
+referenced content-file resolution, and ID/version selection. It reports the
+selected definition's declared metadata without changing the prompt format or
+executing the definition.
+
 ### Inputs
 
 Each `inputs` item has these fields:
@@ -76,14 +82,27 @@ allowed.
 
 ### Messages And Templates
 
-Each message has a non-empty `role` and exactly one of:
+Each message has a `role` that Promptkit trims and lowercases. It must then be
+exactly one of `developer`, `system`, `user`, or `assistant`; blank, custom,
+`tool`, and `function` roles are invalid. This intentionally tightens the
+previous nonblank-string rule. Consumers migrating to the next minor release
+must update any nonstandard prompt-definition roles before upgrading.
+
+Each message also has exactly one of:
 
 - `content`, containing an inline Go template; or
 - `content_file`, naming a file whose contents are the Go template.
 
-For directory and `fs.FS` prompt sources, `content_file` resolves relative to
+`content_file` must be a relative path. It resolves from the directory that
-the prompt file and remains within the source root. `WithPromptFile` also
+contains the prompt file and must remain within the configured prompt source
-resolves it relative to that file.
+root; parent components are allowed only when the resolved target remains
+inside that root. Absolute paths and paths that escape the root are rejected.
+Operating-system directory and single-file sources also reject symlink targets
+outside the root, while injected `fs.FS` sources apply containment in that
+filesystem's relative path namespace. For `WithPromptFile`, the source root is
+the directory containing the selected prompt file. Promptkit uses the parsed
+path text exactly after checking separately that it is not blank, so leading
+and trailing whitespace can name real filesystem entries.
 
 Request variables are the template data, so a variable named `audience` is
 referenced as `{{.audience}}`. The `{{input "note"}}` helper renders the body
@@ -113,7 +132,7 @@ outbound integration determines its wire representation.
 | `format` | yes | `text`, `markdown`, or `json`. |
 | `validation_mode` | yes | `none`, `basic`, `json`, or `json_schema`. |
 | `schema_path` | for `json_schema` | Path to a schema in the configured schema source. |
-| `repair_attempts` | no | Integer zero or greater; omitted means zero. |
+| `repair_attempts` | no | Integer from zero through three; omitted means zero. A positive value requires `basic`, `json`, or `json_schema` validation. |
 
 The validation modes behave as follows:
 
@@ -123,14 +142,31 @@ The validation modes behave as follows:
 - `json_schema` requires valid JSON that satisfies the selected schema.
 
 `format` controls output artifact metadata. JSON Schema mode also supplies the
-schema to compatible model clients as structured-output metadata. The public
+schema to compatible model clients as structured-output metadata. Plain `json`
-engine does not install an output repairer, so its validation is single-pass
+validation accepts every valid JSON value and does not request a provider-native
-even when a positive `repair_attempts` value is present.
+JSON-object constraint.
+
+`repair_attempts` counts additional generation calls after a failed validation.
+Zero is single-pass. With a positive eligible budget, Promptkit stops at the
+first valid candidate. If the budget is exhausted, it returns the final
+candidate and its complete failed validation result; generation and operational
+validation failures remain errors. `none` never permits repair.
 
 A request-level `OutputContract` replaces the complete prompt output contract.
 It does not merge individual fields. If its format is empty, Promptkit uses
 `text`.
 
+## Built-In Backends
+
+Every engine provides these reserved OpenAI-compatible backend IDs. Consumers
+must not register either ID with `WithBackend`; exact registration and
+reservation behavior belongs to the [`Backend` GoDoc](../backends.go).
+
+| ID | Base endpoint | API-key environment variable | Active generation limit | Default queue capacity |
+| --- | --- | --- | ---: | ---: |
+| `openrouter` | `https://openrouter.ai/api/v1` | `OPENROUTER_API_KEY` | 16 | 1024 |
+| `rakestrawhome` | `https://inference.ai.rakestrawhome.com/v1` | `RAKESTRAWHOME_INFERENCE_API_KEY` | 4 | 1024 |
+
 ## Profile Definitions
 
 A profile supplies model execution settings:
@@ -149,11 +185,21 @@ extra_params:
   provider_option: enabled
 ```
 
+A derived profile can use a named base and override only the settings it owns:
+
+```yaml
+id: local-summary-fast
+base_profile: local-summary
+timeout_seconds: 30
+reasoning_effort: low
+```
+
 | Field | Required | Meaning |
 | --- | --- | --- |
-| `id` | yes | Non-empty profile identifier. IDs must be unique within one source. |
+| `id` | yes | Profile identifier, trimmed before selection and publication. It must be non-empty after trimming and unique within one source after normalization. |
-| `backend` | unless `endpoint` is present | Backend registry ID. It is trimmed and registry membership is checked when the profile is prepared. |
+| `base_profile` | no | One optional parent profile ID. A derived profile may inherit target fields from it. |
-| `endpoint` | unless `backend` is present | Non-empty OpenAI-compatible base URL, including an API version path when required. When both connection fields are present, this overrides the backend endpoint without changing backend identity. |
+| `backend` | unless `endpoint` is present | Backend registry ID. It is trimmed and registry membership is checked when the profile is prepared or inspected. |
+| `endpoint` | unless `backend` is present | OpenAI-compatible base URL, including an API version path when required. A nonempty value is trimmed and must be absolute HTTP or HTTPS with a host and without user information, a query, or a fragment. When both connection fields are present, this overrides the backend endpoint without changing backend identity. |
 | `model` | yes | Non-empty provider model name. |
 | `temperature` | no | Number from 0 through 2. |
 | `max_tokens` | no | Integer zero or greater. |
@@ -161,16 +207,21 @@ extra_params:
 | `timeout_seconds` | no | Per-generation deadline in whole seconds; integer zero or greater. |
 | `service_tier` | no | Provider-specific request tier. |
 | `reasoning_effort` | no | Provider-specific reasoning setting. |
-| `api_key_env` | no | Name of an environment variable containing the API key. |
+| `api_key_env` | no | Optional environment-variable lookup source for an API key. |
 | `extra_params` | no | JSON-compatible provider-specific outbound fields. |
 
 Raw `api_key` is prohibited in profile YAML. Store only an environment
 variable name in `api_key_env`.
 
+A standalone profile must provide a model and at least one of `backend` or
+`endpoint`. A derived profile may omit those target fields because its selected
+base chain can provide them. Local parsing still validates a derived profile's
+own ID, supplied endpoint, execution-setting bounds, and `extra_params`.
+
 Promptkit does not infer a backend from a model or endpoint. Endpoint-only
 profiles remain supported and have no effective backend ID.
-The engine always provides the built-in `openrouter` ID. Consumers can add
+The engine always provides the built-in `openrouter` and `rakestrawhome` IDs.
-engine-scoped IDs with
+Consumers can add engine-scoped IDs with
 [`WithBackend`](../backends.go); exact registration validation belongs to its
 GoDoc.
 
@@ -178,30 +229,33 @@ GoDoc.
 objects with string keys. Keys must be non-empty. With the built-in client,
 they also cannot collide with the standard fields listed in the
 [outbound request contract](integrations/openai-compatible-chat.md#request-body).
+Excessively deep or large JSON-shaped values are rejected for safety.
 
 ### Defaults And Overrides
 
 Execution settings resolve in this order:
 
-1. framework defaults;
+1. the framework timeout baseline;
 2. the selected backend, when the profile names one;
 3. the selected profile; and
 4. request `ExecutionTargetOverride` values.
 
-The framework defaults are:
+The framework baseline is:
 
 | Setting | Default |
 | --- | --- |
-| `temperature` | `0` |
+| `temperature` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
-| `max_tokens` | `0` |
+| `max_tokens` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
-| `top_p` | `1` |
+| `top_p` | Unspecified and omitted from compatible provider requests unless a profile or runtime override selects it. |
 | `timeout_seconds` | `600` |
 
-Numeric zero in a file or in-memory profile means that the profile does not
+Numeric zero in a file or in-memory profile does not select a numeric value.
-replace the framework default. Numeric request overrides use pointers, so an
+For `temperature`, `max_tokens`, and `top_p`, it leaves the provider control
-explicit zero is preserved. In particular, an explicit request
+unspecified. For `timeout_seconds`, it retains the framework deadline. Numeric
-`timeout_seconds` of zero disables the per-generation deadline while leaving
+request overrides use pointers, so an explicit zero is retained and sent to
-the caller context and transport timeout intact.
+compatible providers. In particular, an explicit request `timeout_seconds` of
+zero disables the per-generation deadline while leaving the caller context and
+transport timeout intact.
 
 Non-empty profile strings replace backend defaults, and non-empty request
 strings replace both. Request reasoning is the exception: a nil
@@ -219,54 +273,86 @@ defines how the effective settings are serialized.
 ### Source And Profile Precedence
 
 An explicit request profile ID takes precedence over the prompt's
-`default_profile`. If neither is present, preparation fails.
+`default_profile`. If neither is present, preparation fails. Exact profile
+inspection instead takes one explicit profile ID and does not use a prompt
+default.
 
 Profile sources resolve matching IDs in this order:
 
 1. in-memory profiles supplied with `WithProfiles`;
-2. a profile file, `fs.FS`, or configured profile directory; and
+2. the ordinary configured source selected by a profile file, `fs.FS`, or
-3. embedded built-in profiles.
+   configured profile directory;
+3. application fallback profiles supplied with `WithFallbackProfileFS`; and
+4. Promptkit-owned profiles; and
+5. maintained external catalog profiles.
 
-A higher-precedence source falls back only when the profile is absent. An
+A profile source supplies a complete definition; definitions and their fields
-invalid matching profile is an error and does not fall back. In-memory
+are not merged across sources. A higher-precedence source falls back only when
-`Profile` values follow the same ranges as YAML profiles. They use
+the requested profile ID is absent. An invalid matching profile is an error and
-`APIKeyRequired` for request-scoped credentials instead of `api_key_env`.
+does not fall back. In-memory `Profile` values follow the same ranges as YAML
+profiles. They use `APIKeyRequired` for request-scoped credentials instead of
+`api_key_env`. Preparation and exact profile inspection use this same source
+precedence.
+
+When a selected definition names `base_profile`, every profile ID in that
+chain is looked up through this same precedence order. A higher-precedence
+definition therefore shadows a lower-precedence definition of the same base
+ID, including a built-in. References are not source-qualified.
+
+### Profile Inheritance
+
+Promptkit resolves one linear base chain of at most 32 profiles, including the
+selected profile. It merges settings from the root base to the selected leaf.
+The leaf's `id` remains the selected profile identity. Nonblank string fields
+(`backend`, `endpoint`, `model`, `service_tier`, `reasoning_effort`, and
+`api_key_env`) and nonzero numeric fields replace inherited values. A nonempty
+`extra_params` map replaces the complete inherited map rather than merging
+keys, and `APIKeyRequired: true` remains true through the chain. Backend and
+endpoint are independent: replacing one does not clear the other.
+
+There is no profile-level clearing syntax. Blank strings, zero numbers, false,
+and empty maps remain unspecified and inherit from a base. Use existing
+presence-aware request overrides where an execution needs an explicit zero or
+empty reasoning setting.
+
+An absent directly selected profile reports the ordinary not-found error. Once
+the selected profile exists, a missing base, cycle, overlong chain, or
+incomplete resolved target is a profile-load failure. Ordinary operations
+resolve chains afresh; prepared execution retains the fully resolved target.
 
 ## Built-In Profile Catalog
 
-Every built-in selects the `openrouter` backend. The engine's built-in backend
+Every built-in profile selects one maintained built-in backend and inherits
-registry supplies `https://openrouter.ai/api/v1` and the environment-variable
+that backend's connection and credential metadata. Profile files do not repeat
-name `OPENROUTER_API_KEY`, so individual profiles contain only model and
+those values. A configured, application fallback, or in-memory profile with
-generation settings. Built-in profile files do not repeat those connection
+the same profile ID takes precedence.
-values. A custom or in-memory profile with the same profile ID takes
-precedence.
 
-| Provider | ID | Model |
+Promptkit owns two opt-in profiles. The following definitions are current
-| --- | --- | --- |
+selections, not automatic defaults:
-| aion-labs | `aion-2` | `aion-labs/aion-2.0` |
+
-| anthropic | `claude-fable-latest` | `~anthropic/claude-fable-latest` |
+| ID | Current definition |
-| anthropic | `claude-haiku-latest` | `~anthropic/claude-haiku-latest` |
+| --- | --- |
-| anthropic | `claude-opus-latest` | `~anthropic/claude-opus-latest` |
+| `promptkit-openrouter-standard` | Inherits `gemini-flash-latest`. |
-| anthropic | `claude-sonnet-latest` | `~anthropic/claude-sonnet-latest` |
+| `promptkit-openrouter-auto` | Selects backend `openrouter` and model `openrouter/auto`. |
-| deepseek | `deepseek-3-2` | `deepseek/deepseek-v3.2` |
+
-| deepseek | `deepseek-4-flash` | `deepseek/deepseek-v4-flash` |
+Select either name through an existing request `profile_id` or prompt
-| deepseek | `deepseek-4-pro` | `deepseek/deepseek-v4-pro` |
+`default_profile`. Consumers can replace either definition or its inherited
-| google | `gemini-2-flash` | `google/gemini-2.5-flash` |
+base through the ordinary profile-source precedence above. The
-| google | `gemini-2-flash-lite` | `google/gemini-2.5-flash-lite` |
+[release policy](policy/release.md#catalog-profile-compatibility) defines the
-| google | `gemini-2-pro` | `google/gemini-2.5-pro` |
+cross-version guarantee for these names.
-| google | `gemini-3-flash-lite` | `google/gemini-3.1-flash-lite` |
+
-| google | `gemini-flash-latest` | `~google/gemini-flash-latest` |
+## External Profile Inventories
-| google | `gemini-pro-latest` | `~google/gemini-pro-latest` |
+
-| google | `gemma-4-31b` | `google/gemma-4-31b-it:exacto` |
+The external profile inventory is defined by the selected catalog modules, not
-| minimax | `minimax-m2` | `minimax/minimax-m2.5` |
+duplicated here. Browse the exact data selected by this Promptkit version:
-| minimax | `minimax-m3` | `minimax/minimax-m3` |
+
-| mistral | `mistral-large-2512` | `mistralai/mistral-large-2512` |
+- [OpenRouter catalog v2.20260922.1 profiles](https://gitea.maximumdirect.net/eric/promptkit-backend-openrouter/src/tag/v2.20260922.1/catalog/profiles)
-| mistral | `mistral-medium-3-5` | `mistralai/mistral-medium-3-5` |
+- [Rakestrawhome catalog v2.20260922.0 profiles](https://gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/src/tag/v2.20260922.0/catalog/profiles)
-| mistral | `mistral-small-3` | `mistralai/mistral-small-3.2-24b-instruct` |
+
-| mistral | `mistral-small-4` | `mistralai/mistral-small-2603` |
+External profile entries are volatile across catalog releases. The
-| nvidia | `nemotron-3-ultra` | `nvidia/nemotron-3-ultra-550b-a55b` |
+[release policy](policy/release.md#catalog-profile-compatibility) owns the precise
-| openai | `gpt-5-mini` | `openai/gpt-5.4-mini` |
+compatibility boundary; a selected module version identifies tested local data,
-| openai | `gpt-5-nano` | `openai/gpt-5.4-nano` |
+not live provider availability.
 
 ## Schemas
 
@@ -285,16 +371,25 @@ schema produces a failed validation result.
 Credential values belong at the request or environment boundary, never in
 prompt, profile, schema, or example files:
 
-- a file profile names an environment variable with `api_key_env`;
+- a backend or file profile can name an optional environment lookup source
-- an in-memory profile may set `APIKeyRequired`;
+  with `APIKeyEnv` or `api_key_env`;
-- a request can provide a direct `APIKey` or override `APIKeyEnv`; and
+- an in-memory profile may set `APIKeyRequired` as an explicit local
+  requirement;
+- a request can provide a direct `APIKey` or override the optional `APIKeyEnv`
+  source; and
 - a direct request key takes precedence over environment lookup.
 
 After a direct request key, the credential-source precedence is request
 `APIKeyEnv`, profile `api_key_env`, then the backend default. An in-memory
 profile with `APIKeyRequired` clears an inherited backend environment name and
 requires a direct key unless the request explicitly supplies `APIKeyEnv`.
-Promptkit validates required credential availability during preparation.
+Named environment sources are optional: when the selected source is absent,
+empty, or whitespace-only, the built-in client omits the `Authorization`
+header and handles the provider response normally. `APIKeyRequired` is the
+only explicit local availability requirement. Promptkit validates required
+credential availability during preparation and rechecks it when a prepared
+execution runs. Injected clients receive resolved source metadata but define
+their own credential-resolution behavior.
 Direct keys are excluded from JSON results and redacted by public string
 formatters. Environment-variable names may appear in prepared metadata, but
 their values do not.

docs/integrations/openai-compatible-chat.md

@@ -14,10 +14,16 @@ that produce these outbound settings.
 
 Generation sends an HTTP `POST` with `Content-Type: application/json`.
 Before the client is called, the engine resolves framework, backend, profile,
-and request values into one execution target. A non-empty endpoint from that
+and request values into one execution target. Endpoint configuration is trimmed
-target overrides the client's configured base URL. After trailing slashes are
+and must be an absolute HTTP or HTTPS URL with a host and without user
-removed, `/chat/completions` is appended. Generation fails before sending when
+information, a query, or a fragment. A non-empty endpoint from the target
-neither source supplies an endpoint.
+overrides the client's configured base URL. The final selected endpoint is
+validated again before transport.
+
+The completion URL is composed through parsed URL path operations. Nested base
+paths are retained, repeated trailing slashes are normalized, and the result
+has exactly one appended `/chat/completions` suffix. Generation fails before
+sending when neither source supplies a valid endpoint.
 
 The target's backend ID is routing metadata for prepared values, results, and
 injected clients. The built-in client does not derive the URL from that ID and
@@ -25,11 +31,13 @@ does not serialize it in the provider request.
 
 ## Authentication
 
-A non-empty API key supplied directly on the execution target takes
+A usable API key supplied directly on the execution target takes precedence.
-precedence. Otherwise, when an API-key environment-variable name is supplied,
+Otherwise, when an API-key environment-variable name is supplied, the client
-the client reads that variable and requires a non-empty value. The selected
+reads and trims that variable. A bearer header is sent only when the resolved
-key is sent as `Authorization: Bearer <key>`. No authorization header is sent
+direct or environment credential is non-empty. When neither source is usable,
-when neither mechanism is configured.
+the client omits `Authorization` and handles the provider response normally.
+An explicitly required target with no usable source is rejected before
+transport.
 
 The target contains the already resolved environment-variable name: an
 explicit request override takes precedence over profile metadata, which takes
@@ -47,18 +55,26 @@ Each ordinary message contains its `role` and string `content`. A
 cache-controlled message instead uses a text content block containing `type`,
 `text`, and `cache_control`; an empty cache-control TTL is omitted.
 
+Promptkit sends only `developer`, `system`, `user`, and `assistant` roles and
+does so without provider-specific translation. Tool and deprecated function
+payloads are outside this text-message contract. A backend or model that
+rejects an otherwise supported role or context returns its ordinary provider
+error, which follows the normal generation-error path.
+
 The effective direct or prompt-rendered session ID is trimmed, limited to 256
 Unicode code points, and sent when nonempty as top-level `session_id`. It is
 never also sent as a session header.
 
 The client conditionally includes:
 
-- `temperature`, `max_tokens`, and `top_p` when non-zero or explicitly
+- `temperature`, `max_tokens`, and `top_p` only when selected by a profile or
-  present;
+  runtime override, including an explicit runtime zero; they are absent when
+  unspecified;
 - non-empty `service_tier` and effective `reasoning_effort`; an explicitly
   disabled reasoning setting is empty and therefore omitted; and
 - `response_format` for JSON Schema structured output, including its name,
-  strict flag, and schema document.
+  strict flag, and schema document. Plain JSON validation does not add an
+  object-only response constraint.
 
 The engine resolves backend, profile, and request extra-parameter maps by
 whole-map replacement rather than key merging. The resulting effective map is
@@ -81,13 +97,82 @@ request fields.
 
 ## Response Handling
 
-Any 2xx response is decoded as an OpenAI-compatible chat response. The client
+Any 2xx response body is limited to 16 MiB (16,777,216 bytes). A larger
-returns the first choice's non-empty message content and maps prompt,
+declared `Content-Length` is rejected before the body is read, and streamed,
-completion, total, cached, and cache-write token counts.
+chunked, or underreported bodies are read through the same bound with at most
+one additional byte used to detect overflow. A body exactly at the limit is
+allowed. The body is closed on every outcome and an oversized stream is not
+drained.
 
-Invalid JSON, absent choices, and empty first-choice content are malformed
+The bounded body must contain exactly one OpenAI-compatible JSON response
-responses. For a non-2xx status, the error includes the status code but never
+object followed only by JSON whitespace and EOF. The client returns the first
-the provider response body.
+choice's explicitly present string message content, including an empty or
+whitespace-only string, and maps prompt, completion, total, cached, and
+cache-write token counts. Invalid or truncated JSON, trailing non-whitespace
+data, a second JSON value, absent choices, missing content, `null` content,
+non-string content, and size overflow are malformed responses and return no
+partial result.
+
+Each of the five token fields retains whether the provider reported a
+non-null integer, including zero. Omitted and `null` fields are unknown for
+observations. A strictly decoded successful envelope can yield observed usage
+even when a later missing choice or content makes the attempt malformed.
+Invalid framing and non-2xx error bodies supply no observed usage; provider
+work may still have occurred. Reported counts are not inferred, reconciled,
+or proof of complete cost accounting.
+
+For a non-2xx status, Promptkit recognizes one JSON document with a top-level
+object-valued `error` member. Its optional `message` and `type` fields must be
+strings, and `code` may be a string or JSON number. Valid supported fields are
+handled independently, numeric codes retain their JSON number text, and
+unknown fields are ignored. Missing, invalid, malformed, or multiply framed
+envelopes contribute no provider detail.
+
+Non-success bodies have a 65,536-byte limit. A larger declared
+`Content-Length` is not read; otherwise the client reads at most one additional
+byte to detect streamed or underreported overflow. Empty, unreadable,
+oversized, malformed, and unrecognized bodies retain only the received status.
+The body is always closed and no oversized stream is drained beyond that probe.
+
+Extracted strings are made valid UTF-8, trimmed, and converted to one line by
+collapsing Unicode whitespace, control, and format-character runs. Blank
+values are omitted. Codes and types longer than 256 Unicode code points are
+omitted; messages longer than 4,096 code points are truncated at a code-point
+boundary with an ellipsis inside the limit. Promptkit never exposes raw bodies,
+headers, endpoints, credentials, request data, schemas, generated content, or
+unsupported provider metadata through this handling.
+
+An outbound `http.Client.Do` failure retains both Promptkit's request-failure
+identity and the exact transport error for `errors.Is` and `errors.As` checks.
+The rendered error does not include the selected endpoint, request headers,
+request content, credentials, or provider body.
+
+## Transient Status Retries
+
+The built-in path retries only HTTP 429, 500, 502, 503, and 504. It makes two
+additional attempts by default; engine configuration or an individual run can
+choose a different nonnegative count, including zero for single-attempt
+behavior. The same budget applies separately to initial generation and each
+structural repair generation. An injected `LLMClient` is dispatched once by
+Promptkit regardless of this setting. A custom `http.Client` on the built-in
+path remains subject to these retries.
+
+Each retry uses equal jitter between half and all of an exponential local
+delay, starting at 500 milliseconds and capped at 8 seconds. The first
+`Retry-After` header value sets a minimum not-before time measured from
+response receipt if it is valid unsigned decimal seconds or an HTTP date.
+Invalid values use local delay; past dates impose no extra delay. The provider
+minimum is not capped at 8 seconds. Retry waiting is cancellable and holds no
+active backend permit. A caller without a deadline can wait a long time for
+provider timing.
+
+The request body, selected URL, and resolved authorization are fixed across
+attempts of one generation. Each attempt creates a fresh HTTP request and
+receives its own configured generation and HTTP-client timeout after capacity
+admission. The caller context covers every attempt and wait without a reset.
+Transport failures, malformed responses, noneligible statuses, and invalid
+generated content are not transient retries. A positive budget ending in
+eligible statuses reports retry exhaustion with the final provider cause.
 
 ## Timeout And Cancellation
 
@@ -102,5 +187,7 @@ Timeouts are layered:
   timeout when the supplied value is not positive.
 
 The earliest applicable caller, generation, or transport deadline controls the
-request. Constructing the internal client does not mutate a supplied
+request. Caller cancellation retains `context.Canceled`; caller, generation,
-`http.Client`.
+and whole-request timeout failures retain `context.DeadlineExceeded`, together
+with the request-failure identity. Constructing the internal client does not
+mutate a supplied `http.Client`.

docs/internal/capacity.md

@@ -14,7 +14,7 @@ stable JSON values.
 
 ## Construction And Pool Lifecycle
 
-Each root `NewEngine` call obtains a normalized capacity-policy snapshot from
+Each internal engine construction obtains a normalized capacity-policy snapshot from
 its immutable backend registry and constructs a new `Manager`. The manager
 creates one pool for each limited backend ID. It has no package-global mutable
 state, background workers, shutdown protocol, or persistence, so engines with
@@ -29,51 +29,73 @@ One pool owns immutable active and total limits plus mutex-protected admission
 count, active count, and ordered waiter list. Pool state exists only for the
 lifetime of its engine.
 
-## Bounded Run Admission
+## Bounded Execution Admission
 
-The runner asks the manager to admit a run after resolving the prompt, profile,
+For ordinary `Run`, the runner asks the manager to admit after resolving the
-selected backend, effective execution target, credentials, and output contract,
+prompt, profile, selected backend, effective execution target, credentials, and
-but before schema loading, artifact loading, or rendering. Admission is
+output contract, but before schema loading, artifact loading, or rendering.
-immediate: a limited pool either reserves a slot or returns the internal
+`PrepareExecution` performs no admission. `RunPrepared` claims its handle,
-`ErrCapacityExceeded` identity. The root facade maps that identity to the
+rechecks credential availability, and then asks the manager to admit the
-public error without treating it as an invalid request or generation failure.
+frozen backend before generation.
+
+Admission is immediate: a limited pool either reserves a slot or returns only
+the internal `ErrCapacityExceeded` identity. The runner attaches the selected
+backend identity at its use-case boundary, and the root facade translates that
+typed value without treating it as an invalid request or generation failure.
 
 The total admitted bound is the active-generation limit plus its configured
 waiting capacity. The returned release function is idempotent. The runner
-defers it as soon as admission succeeds and holds the lease across remaining
+defers it as soon as admission succeeds. An ordinary run holds the lease across
-preparation, initial generation, validation, every repair attempt, and all
+remaining preparation, initial generation, validation, every repair attempt,
-failure or cancellation exits. A repair is part of its original admission and
+and all failure or cancellation exits. Prepared execution holds the normal
-does not reserve another bounded slot.
+lease across generation, validation, every internal repair attempt, and all
+execution exits. A repair is part of its original admission and does not
+reserve another bounded slot.
 
 ## FIFO Generation Permits
 
-`NewClient` wraps the engine's selected internal model client after public
+The engine supplies one [generation coordinator](generation.md) to initial
-client adaptation or built-in client construction. Initial generation and the
+generation and the default repairer. Their requests retain the same effective
-default repairer receive the same wrapper.
+backend, credential, numeric-presence metadata, structured-output settings,
+and retry budget. Scheduling does not change provider omission semantics
+between calls.
 
-For each `Generate` call, the wrapper selects a pool from the request's
+For each model attempt, the coordinator selects a pool from the request's
-effective backend ID. An unlimited call passes directly to the next client. A
+effective backend ID. An unlimited call uses the unrestricted path. A limited
-limited call acquires an active permit, invokes the next client, and defers
+call acquires an active permit and defers release around that attempt, so
-permit release so ordinary returns and panic unwinding both restore capacity.
+ordinary returns and panic unwinding both restore capacity.
 Preparation and validation never hold an active permit.
 
+The manager's `Acquire` operation owns that active-permit acquisition and
+returns an idempotent release function. The coordinator uses it around
+individual attempts. The earlier one-shot wrapper still delegates to the same
+operation for focused internal callers and tests, but engine assembly no
+longer uses it. There is one queue and one active limit per limited backend pool.
+
 When all active permits are occupied, calls join a mutex-protected FIFO waiter
 list. Releasing a permit transfers it directly to the oldest remaining waiter
 before making it generally available. Pools do not order work relative to
 other backend IDs.
 
-The wrapper passes generation requests, responses, and collaborator errors
+The coordinator does not retry injected clients; it normalizes a nil successful
-through unchanged. It owns scheduling only; the concrete model client remains
+response as malformed generation. The built-in HTTP client remains responsible
-responsible for provider transport behavior.
+for provider transport behavior. The runner, rather than the
+capacity layer, sums all five usage fields from the initial response and every
+completed repair response into the successful run result.
+
+An invocation-local observer receives attempt events after active permit
+release, including canceled acquisition without dispatch. A retry wait holds
+the logical admission lease but no active permit. Observer callbacks must not
+block on work that requires the same run's admission slot.
 
 ## Cancellation And Release
 
 Admission checks the caller context before reserving a slot. A call canceled
 while waiting for an active permit removes its waiter under the same pool lock
-used to grant permits. If cancellation removes the waiter first, the wrapped
+used to grant permits. If cancellation removes the waiter first, no model
-client is not invoked. If a concurrent grant wins first, the call owns the
+attempt is dispatched. If a concurrent grant wins first, the call owns the
-permit and invokes the client with the original context, allowing the client
+permit; the coordinator checks the context before dispatch and releases the
-to observe cancellation normally.
+permit if cancellation has already occurred.
 
 This grant-or-cancel decision prevents lost and double-released permits.
 Admission leases and active permits are released after success, collaborator
@@ -90,11 +112,19 @@ unlimited admission. The
 FIFO transfer, canceled-waiter removal, grant/cancel races, independent pools,
 unlimited calls, passthrough behavior, and panic release.
 
-The [runner tests](../../internal/usecase/runner_test.go) own early admission,
+The [runner tests](../../internal/usecase/runner_test.go) own ordinary early
-lease lifetime, failure release, and shared initial/repair scheduling. The
+admission, lease lifetime, failure release, and shared initial/repair
+scheduling. The
+[prepared-execution use-case tests](../../internal/usecase/prepared_execution_test.go)
+own deferred admission, credential ordering, and prepared-execution lease
+release. The
 [external package capacity tests](../../capacity_contract_test.go) own the
 assembled public-engine behavior for configured limits, capacity errors,
 endpoint identity, engine independence, and injected clients. The
+[reliable execution contract tests](../../reliable_execution_contract_test.go)
+cover shared and independent backend pools during retry and repair. The
+[prepared-execution contract tests](../../prepared_execution_contract_test.go)
+own the public prepared-capacity boundary. The
 [root error-boundary tests](../../errors_internal_test.go) own preservation of
 the public generation category and context identity when generation is
 canceled.

docs/internal/generation.md

@@ -0,0 +1,50 @@
+# Internal Generation Coordination
+
+`internal/generation` coordinates one model-generation request through the
+existing engine-local [capacity manager](capacity.md) and the
+[built-in HTTP client](llm.md). The engine supplies the same coordinator to
+initial generation and the default output repairer.
+
+The coordinator has explicit built-in and injected-client constructors. The
+built-in path prepares the provider request once, then acquires and releases an
+active backend permit around each actual HTTP attempt. Only HTTP 429, 500, 502,
+503, and 504 can lead to another attempt. The injected-client path acquires a
+permit and invokes its client once, regardless of the internal retry budget.
+The runner continues to own the logical execution's admission lease. A request
+can override the engine's retry budget; the runner freezes that choice for
+prepared execution and passes it to each repair generation.
+
+Retry waits hold no active permit. The local delay uses equal jitter between
+half and all of a capped exponential base, beginning at 500 milliseconds and
+capping at 8 seconds. A valid provider not-before time can extend that delay;
+the caller context interrupts it. The next attempt reacquires the ordinary
+FIFO permit. Preparation failures dispatch nothing. When a positive retry
+budget ends with eligible failures, the internal exhaustion error retains the
+final cause and actual dispatched-attempt count.
+
+When an invocation-local [observation recorder](observation.md) is present in
+the context, each acquisition emits an attempt event, including canceled
+capacity waits without dispatch. Each actual retry wait emits a delay event,
+including interrupted waits. Active permits are released before callbacks;
+generation duration ends before callback delivery. Built-in successful-envelope
+usage retains reported-field presence, while injected clients use explicit
+presence when supplied and otherwise treat zero fields as unknown.
+Malformed successful envelopes can retain strictly decoded usage; error
+response bodies are never mined for it. Arbitrary injected failures carry
+unknown usage. Transport, timeout, malformed, and noneligible status failures
+stop the sequence without retry exhaustion. Context is checked before each
+wait, acquisition, and dispatch.
+
+The public client adapter translates explicitly supplied public transport and
+malformed-response identities into internal categories while retaining the
+original error chain. This keeps injected attempt classifications consistent
+with terminal observations without parsing diagnostic text or retrying injected
+clients.
+
+Focused coordinator tests use controlled transport and wait functions. They
+cover eligibility, exhaustion, cancellation, backoff, request freezing, permit
+release, and the injected-client path. The capacity manager tests own permit
+acquisition and idempotent release. Public
+[integration tests](../../reliable_execution_contract_test.go) cover retry,
+repair, observation, frozen prepared execution, and shared-backend behavior
+through the assembled facade.

docs/internal/llm.md

@@ -10,7 +10,7 @@ owns the observable outbound HTTP contract. The
 [framework format reference](../formats.md) owns the profile and prompt
 settings consumed by the client.
 
-The concrete client remains under `internal/llm`. The root engine assembles it
+The concrete client remains under `internal/llm`. The internal engine assembles it
 as the default implementation behind Promptkit's public client boundary.
 
 ## Components And Flow
@@ -21,38 +21,109 @@ uses internal domain values for rendered prompts, execution targets,
 structured output, responses, and token usage.
 
 The runner supplies a fully resolved target after applying backend, profile,
-and request precedence. The client uses its endpoint, credential metadata,
+and request precedence, plus canonical provider-bound text messages. The
-generation fields, and extra parameters. `BackendID` remains routing metadata
+client uses its endpoint, credential metadata, generation fields, and extra
+parameters. `BackendID` remains routing metadata
 for the generation boundary and is not mapped into the provider payload.
 
-Construction validates the configured base URL and clones any supplied
+Construction trims and validates a nonempty configured base URL and clones any
-`http.Client` so Promptkit can apply its timeout default without mutating the
+supplied `http.Client` so Promptkit can apply its timeout default without
-caller's client. Generation then:
+mutating the caller's client. An empty configured base remains valid because a
+resolved request target may supply the endpoint. Generation preparation then:
 
-1. validates request-level timeout and endpoint requirements;
+1. validates shared execution-setting invariants and the final selected base
+   endpoint;
 2. maps the internal request into the OpenAI-compatible chat payload;
 3. validates and merges extra parameters;
-4. resolves authentication;
+4. composes `/chat/completions` through parsed URL path operations;
-5. performs the outbound request under the applicable deadlines; and
+5. resolves authentication and freezes the URL, body, and optional credential.
-6. decodes the first response choice and token usage.
+
+Each execution of that prepared request creates a fresh HTTP request, body
+reader, and attempt timeout context. It closes the response body and returns
+the HTTP status, decoded response or error, reported token values and presence,
+and any valid provider `Retry-After` not-before time. A strictly decoded
+successful envelope can retain usage even if its first choice or content is
+missing. Invalid framing and non-success error bodies supply no usage. The
+ordinary `Generate` method prepares and executes once and returns no partial
+response on failure.
 
 `internal/llm` owns the set of reserved OpenAI-compatible request fields used
 when validating extra parameters. Backend registration consumes the same rule
 without making the model client depend on registry configuration.
 
-The implementation has no retry loop, tool-call support, provider catalog,
+This client does not schedule retries; the [generation coordinator](generation.md)
-inbound HTTP behavior, or durable session store.
+does so around its prepared attempts. The client has no tool-call support,
+provider catalog, inbound HTTP behavior, or durable session store.
+
+## Prepared Generation
+
+For [`RunPrepared`](../../engine.go), the runner supplies the model client with
+the target, rendered messages, and structured-output constraint retained by
+executable preparation. Execution does not reopen or rerender consumer
+sources.
+
+Before backend admission, the runner rechecks a frozen credential
+environment-variable name only when the target explicitly requires a
+credential. The handle does not retain the environment value; the model client
+resolves the value visible when generation begins. For optional sources with no
+usable value, the built-in client omits `Authorization` and continues to the
+provider. A direct request key remains in private execution state only until
+the claimed execution finishes or an unclaimed handle is discarded. Exact
+public ownership and redaction semantics belong to the
+[`PreparedExecution` GoDoc](../../prepared_execution.go).
 
 ## Failure Categories
 
 The package preserves distinct error identities for invalid client
 configuration, invalid generation requests, request execution failures,
 non-success provider statuses, and malformed successful responses. Provider
-response bodies are not included in non-success errors.
+response bodies are never exposed in raw form through non-success errors.
 
-Caller cancellation and deadline failures during the outbound request are
+Invalid nonempty configured endpoints are configuration failures. A missing or
-reported as request execution failures. The runner classifies these identities
+invalid final selected endpoint is an invalid generation request and is
-without depending on HTTP status mapping.
+rejected before transport.
+
+Authentication resolves a trimmed direct key before a trimmed configured
+environment value. Optional missing, empty, or whitespace-only sources do not
+block transport and produce no `Authorization` header. An explicitly required
+target with no usable source is rejected before transport with the existing
+invalid-request diagnostics.
+
+Successful response bodies have a fixed 16 MiB limit enforced by declared
+length and by reading at most one byte beyond the boundary. The decoder accepts
+exactly one JSON object plus trailing whitespace and EOF. Size overflow,
+truncation, malformed JSON, trailing data, and a second value are malformed
+responses with no partial result or provider content in the error. Every body
+is closed, and an unbounded oversized stream is not drained.
+
+After framing succeeds, the first choice must contain an explicitly present
+string `message.content`. The string is returned exactly, including empty or
+whitespace-only content. Missing choices, missing or `null` content, and
+non-string content are malformed responses. Output validation and correction
+eligibility remain outside this package.
+
+Successful-envelope usage is decoded with independent presence bits for each
+token field: omission and `null` are unknown, while a reported integer zero is
+known. Strictly decoded usage can still be observed if later choice/content
+checks fail; invalid framing or a non-success error body contributes none.
+Existing successful-result accounting remains value-based.
+
+For a non-success response, `ProviderHTTPError` retains the HTTP status and
+only normalized detail from the bounded recognized envelope. It retains
+`ErrUnexpectedStatus` through unwrapping. The client owns response closure;
+its bounded reader and parser never close or drain a body themselves. The root
+facade converts this concrete internal error into the public
+[`GenerationError`](../../generation_error.go), while arbitrary injected-client
+errors continue through the ordinary generation-error mapping unchanged.
+
+An `http.Client.Do` failure is represented by a redacting multi-cause error:
+the package request-failure sentinel and the exact returned transport error are
+both available through `errors.Is` and `errors.As`, while the rendered text
+does not expose the endpoint, headers, request content, credential, transport
+detail, or provider body. Caller cancellation retains `context.Canceled`;
+caller deadlines, generation deadlines, and whole-request client timeouts
+retain `context.DeadlineExceeded`. The runner adds its generation category
+without discarding those identities or depending on HTTP status mapping.
 
 ## Test Ownership
 
@@ -60,7 +131,15 @@ The
 [OpenAI-compatible client tests](../../internal/llm/openai_compatible_client_test.go)
 own configuration, client cloning, deterministic deadline precedence,
 authentication, request and response mapping, malformed data, error identity,
-cancellation, and response-body suppression. The root transport contract test
+cancellation, endpoint selection and composition, pre-transport rejection, and
-also verifies that resolved backend settings reach this client without
+bounded single-document successful-response framing, closure, and
-serializing backend identity. All use local test servers or test transports;
+response-body suppression. The focused
-the default suite makes no live or paid provider requests.
+[provider HTTP error tests](../../internal/llm/provider_http_error_test.go)
+own envelope parsing, normalization, and bounded-reader cases; their
+[transport tests](../../internal/llm/provider_http_error_transport_test.go)
+own non-success response closure and integration. Root transport contract tests
+own public `GenerationError` conversion, while also verifying that resolved
+backend settings reach this client without serializing backend identity and
+that ordinary-run cancellation retains its public generation and context
+identities. All use local test servers or controlled test transports; the
+default suite makes no live or paid provider requests.

docs/internal/observation.md

@@ -0,0 +1,48 @@
+# Internal Execution Observation
+
+`internal/observation` owns one invocation's run ID, presence-aware token
+aggregation, measured phase totals, and synchronous metadata event delivery.
+Its event values and callback type live in `internal/domain` so internal
+requests can carry them without depending on the recorder package. The
+recorder supplies run IDs to the runner and receives attempt and
+retry-delay measurements from the generation coordinator. Ordinary `Run`
+lifecycle instrumentation spans the facade and runner boundaries.
+Prepared execution starts only after its private payload is atomically claimed;
+the facade receives that invocation's recorder privately and emits the terminal
+event after public error mapping. Failed claims and preparation emit nothing.
+
+The facade masks inherited recorder and prepared-claim state when entering
+preparation or execution. Delegated calls from an injected collaborator retain
+the caller's cancellation, deadline, and unrelated context values, but do not
+reuse the enclosing invocation's accounting. Only the new request's observer
+(or the claimed handle's retained observer) enables its observation lifecycle.
+
+A recorder keeps only counters, resolved identity fields, and the selected
+callback. It has no global history, queue, goroutine, or mutex held across a
+callback. The caller supplies normalized failure metadata; the recorder never
+parses an error string or receives provider bodies, credentials, headers, or
+other content. Events are values with no mutable containers. Callback panics
+propagate to the caller.
+Per-operation callbacks are ordered and synchronous; different operations may
+invoke the same callback concurrently. The recorder marks each measured phase's
+end before delivery, so callback work does not inflate provider duration.
+Terminal phase durations sum measured phases, while wall duration may also
+include preparation, validation, and callbacks. Panics do not promise a
+terminal event.
+
+Attempt usage is known when its explicit presence bit is true. For legacy
+injected clients without presence metadata, nonzero counts are known and
+zeros are unknown. Aggregate values sum known contributions; `Known` means
+at least one reported value, while `Complete` means every dispatched attempt
+reported the field. Zero dispatched attempts produce known, complete zero
+totals. Non-dispatched acquisitions contribute capacity wait but no usage or
+generation duration. Terminal totals count dispatched attempts, transient
+attempts above the first in a generation, and repair generations with at least
+one dispatch. A terminal event is an aggregate and must not be added to its
+attempt events.
+
+The [recorder tests](../../internal/observation/recorder_test.go) own usage
+presence, phase sums, event ordering, empty-run accounting, identity snapshots,
+and callback panic behavior. The [generation coordinator](generation.md) owns
+attempt and retry-delay measurements and emits their completion events after
+releasing active permits.

docs/internal/overview.md

@@ -11,25 +11,29 @@ contributor workflow and validation.
 
 | Component | Implemented responsibility | References |
 | --- | --- | --- |
-| Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request and result values, profile construction, extension interfaces, value conversion, redacted formatting, public error mapping, and engine-local assembly. | [Package GoDoc](../../doc.go), [backend API](../../backends.go), [engine assembly](../../engine.go) |
+| Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request, result, prompt-inspection, profile-inspection, and ordinary-run observation values, opaque prepared-execution handles, profile construction, extension interfaces, value conversion, redacted formatting, and typed capacity and generation error mapping. | [Package GoDoc](../../doc.go), [observations](../../observations.go), [prepared execution](../../prepared_execution.go), [backend API](../../backends.go), [engine facade](../../engine.go) |
 | `examples/go-library/prepare` | Demonstrates an offline downstream consumer using a prompt file, in-memory profile, inline input, and `Prepare`. It is not a public library package. | [Example program](../../examples/go-library/prepare/main.go) |
 | `examples/go-library/run` | Demonstrates an offline downstream consumer using a prompt file, in-memory profile, inline input, an injected deterministic model client, and `Run`. It is not a public library package. | [Example program](../../examples/go-library/run/main.go) |
-| `internal/backend` | Constructs each engine's immutable registry from the built-in OpenRouter definition and consumer additions, validates and defensively copies definitions through the shared JSON-value package, and consumes the LLM-owned OpenAI-compatible reserved request-field rule. | [Backend registry](../../internal/backend/registry.go) |
+| `examples/go-library/observe` | Demonstrates offline built-in HTTP retry and synchronous metadata observation across structured proposal and assessment runs using embedded sources, two profiles on one finite backend, and an application decision. It is not a public library package. | [Example program](../../examples/go-library/observe/main.go) |
-| `internal/capacity` | Owns engine-local bounded run admission and FIFO model-generation permits for limited backend IDs, including cancellation-safe waiter removal and client wrapping. | [Internal capacity management](capacity.md) |
+| `internal/backend` | Constructs each engine's immutable registry from maintained definitions and consumer additions, validates and defensively copies definitions through the shared JSON-value package, and consumes the LLM-owned OpenAI-compatible reserved request-field rule. | [Backend registry](../../internal/backend/registry.go) |
-| `internal/domain` | Defines internal framework values for requests, artifacts, prompt definitions, profiles, execution targets, rendering, generation, and validation. | [Domain declarations](../../internal/domain/domain.go) |
+| `internal/catalog` | Strictly validates imported immutable maintained backend and profile catalog assets before engine assembly uses them. | [Catalog adapter](../../internal/catalog/catalog.go), [internal sources](sources.md#profiles-and-built-ins) |
+| `internal/capacity` | Owns engine-local bounded execution admission and FIFO model-generation permits for limited backend IDs, including cancellation-safe waiter removal and client wrapping. | [Internal capacity management](capacity.md) |
+| `internal/domain` | Defines internal framework values for requests, artifacts, prompt definitions, profiles, execution targets, rendering, generation, and validation, and owns source-neutral invariants for shared execution settings, OpenAI-compatible base endpoints, session identifiers, and output contracts. Source parsing, required fields, other source-specific normalization, defaulting, and boundary-specific error classification remain with their callers. | [Domain declarations](../../internal/domain/domain.go), [endpoint invariant](../../internal/domain/endpoint.go) |
+| `internal/engine` | Assembles external catalogs, Promptkit-owned profiles, repositories, registry, capacity, collaborators, repairer, and the one use-case runner for each engine instance. | [Engine assembly](../../internal/engine/engine.go), [internal sources](sources.md#profiles-and-built-ins), [internal runner](runner.md) |
+| `internal/generation` | Coordinates frozen built-in HTTP attempts and one-shot injected calls through active backend permits, including bounded status retries, cancellable backoff, and attempt observations. | [Internal generation coordination](generation.md) |
 | `internal/defaults` | Defines application-neutral framework constants and constructs the default execution target. It contains no CLI, server, or inbound HTTP limits. | [Framework defaults](../../internal/defaults/defaults.go) |
 | `internal/filecatalog` | Provides deterministic YAML discovery and path helpers for operating-system filesystems and `fs.FS` sources. | [File catalog](../../internal/filecatalog/catalog.go) |
-| `internal/jsonvalue` | Validates and deeply copies JSON-compatible extra-parameter trees while preserving supported concrete value types. | [JSON values](../../internal/jsonvalue/jsonvalue.go) |
+| `internal/jsonvalue` | Validates and deeply copies bounded JSON-compatible extra-parameter and prepared-schema trees while preserving supported concrete value types and rejecting cycles or excessive depth and work. | [JSON values](../../internal/jsonvalue/jsonvalue.go) |
 | `internal/promptdef` | Loads strictly decoded, validated prompt definitions from filesystem and `fs.FS` sources, including version selection and contained file-backed message content. | [Framework formats](../formats.md), [prompt-definition repository](../../internal/promptdef/filesystem_repository.go) |
-| `internal/profile` | Loads strictly decoded, validated execution profiles, including backend selection, from filesystem and `fs.FS` sources and composes repositories with error-preserving fallback. | [Framework formats](../formats.md), [profile repositories](../../internal/profile/filesystem_repository.go) |
+| `internal/profile` | Owns immutable in-memory and filesystem profile repositories, local profile validation, raw-source overlays with error-preserving fallback, and inherited-profile resolution. | [Framework formats](../formats.md), [profile repositories](../../internal/profile/filesystem_repository.go), [internal sources](sources.md#profiles-and-built-ins) |
-| `internal/profile/builtin` | Embeds the built-in profile catalog, whose entries select OpenRouter, and combines it with an optional primary repository. | [Built-in catalog](../formats.md#built-in-profile-catalog), [repository](../../internal/profile/builtin/repository.go) |
 | `internal/prompt` | Renders prompt messages from Go templates with artifact, variable, session, and cache-control data. | [Go-template renderer](../../internal/prompt/go_renderer.go) |
 | `internal/artifact` | Resolves ordinary inline and unrestricted caller-selected file references into copied artifacts with metadata and hashes. | [Internal sources and validation](sources.md) |
-| `internal/validate` | Validates basic, JSON, and JSON Schema output using operating-system filesystem or `fs.FS` schema sources. | [Framework formats](../formats.md#schemas), [internal sources and validation](sources.md) |
+| `internal/validate` | Validates basic, JSON, and JSON Schema output using operating-system filesystem or `fs.FS` schema sources and creates operation-local validation plans with canonical contained schema resources. | [Framework formats](../formats.md#schemas), [internal sources and validation](sources.md) |
-| `internal/llm` | Defines the internal generation boundary and implements outbound OpenAI-compatible chat requests from resolved execution targets, including response decoding, authentication, deadline handling, and ownership of the OpenAI-compatible reserved request-field policy. | [Internal model client](llm.md) |
+| `internal/llm` | Defines the internal generation boundary and implements outbound OpenAI-compatible chat requests from resolved execution targets, including bounded structured non-success response decoding, successful-response decoding, authentication, deadline handling, and ownership of the OpenAI-compatible reserved request-field policy. | [Internal model client](llm.md) |
-| `internal/usecase` | Resolves backend, profile, and request settings and coordinates preparation and execution across internal sources, rendering, artifact loading, generation, validation, and optional repair. | [Internal runner](runner.md) |
+| `internal/observation` | Owns invocation-local run IDs, presence-aware usage and phase accounting, and synchronous metadata event delivery for ordinary and claimed prepared runs and generation attempts. | [Internal execution observation](observation.md) |
+| `internal/usecase` | Resolves prompt definitions and hashes, profiles, backends, and targets for exact inspection and request settings for preparation, and coordinates ordinary and single-claim prepared execution across internal sources, rendering, artifact loading, operation-local validation plans, generation, capacity, and bounded repair. | [Internal runner](runner.md), [prepared-execution implementation](../../internal/usecase/prepared_execution.go) |
 
-The root package assembles these internal components without exposing their
+The internal engine assembles these internal components without exposing their
 representations. Consumers depend only on the root facade.
 
 ## Maintenance

docs/internal/runner.md

@@ -9,8 +9,8 @@ artifact, rendering, and validation behavior, while the
 [model-client document](llm.md) owns generation behavior and failure
 categories.
 
-The runner remains under `internal/usecase` and is assembled by the root
+The runner remains under `internal/usecase` and is assembled by
-Promptkit engine. Its concrete type is not part of the public API.
+`internal/engine`. Its concrete type is not part of the public API.
 The [framework format reference](../formats.md) owns prompt, profile, schema,
 and override semantics consumed by the runner.
 
@@ -18,16 +18,47 @@ and override semantics consumed by the runner.
 
 `Runner` coordinates narrow internal interfaces for prompt definitions,
 profiles, backend resolution, artifacts, rendering, model generation, and
-validation. The root engine supplies one immutable registry containing the
+validation. The internal engine supplies one immutable registry containing the
 built-in backend and validated consumer additions, one engine-local run
-admitter, and a model client wrapped by the same capacity manager. Schema
+admitter, and a generation coordinator using the same capacity manager. Validation
-documents are loaded through the validator's optional schema-loader interface.
+plans and provider-facing schema metadata come from the validator's preparation
-An output repairer can be injected internally, but the ordinary runner
+interface.
-constructor does not enable one.
+The internal engine supplies one default output repairer through the explicit
+runner constructor, using the same coordinator as initial
+generation. The no-repair runner constructor remains available for focused
+internal callers and tests.
 
 Each invocation carries its state in request, prepared-run, and result values.
 The runner has no durable run or session store.
 
+## Shared Prompt Selection
+
+The runner uses one prompt-selection and hashing boundary for ordinary
+preparation and exact prompt inspection. Preparation retains its early
+request-ID check before direct-session normalization; both operations then use
+the configured prompt repository to select one definition, load referenced
+message content, and calculate the same prompt hash.
+
+Inspection stops after that structural lookup. It does not parse templates or
+touch profile, artifact, schema, renderer, validator, admission, or model
+collaborators. The root [`Engine.InspectPrompt`](../../engine.go) GoDoc owns
+the public operation's exact contract.
+
+## Shared Profile Selection
+
+The runner uses one profile-selection and target-resolution boundary for
+ordinary preparation and exact profile inspection. Preparation first selects a
+request profile or a prompt default; inspection begins with its required
+explicit profile ID. Both then apply the ordinary source precedence, resolve a
+named backend, and construct the effective target from framework, backend, and
+profile values.
+
+Inspection stops after the resulting endpoint and model are structurally
+validated. It does not check credential availability or perform prompt,
+artifact, schema, rendering, admission, or model-client work. The root
+[`Engine.InspectProfile`](../../engine.go) GoDoc owns the public operation's
+exact contract.
+
 ## Shared Preparation Pipeline
 
 `Prepare` and `Run` share one private preparation pipeline split at the point
@@ -41,17 +72,20 @@ performs only the work needed to validate routing and admission:
 5. resolve application-neutral defaults, backend defaults, profile values,
    and explicit request overrides in that order;
 6. validate endpoint, model, numeric overrides, and credential requirements;
-7. resolve the effective output contract without loading its schema; and
+7. resolve and validate the effective output contract without loading its
+   schema; and
 8. retain the definition, source identities, effective settings, output
    contract, and preparation start time in invocation-local state.
 
 The completion phase consumes that state without reloading the prompt,
 profile, or backend:
 
-1. load structured-output schema metadata when required;
+1. create one operation-local validation plan and derive structured-output
+   schema metadata from it when required;
 2. load and hash input artifacts;
 3. render messages and the prompt-defined session;
-4. apply any direct session ID;
+4. apply any direct session ID, then append already-normalized request messages
+   after the rendered definition messages;
 5. hash the effective rendered prompt; and
 6. construct the prepared value and preparation timing.
 
@@ -59,7 +93,10 @@ profile, or backend:
 `Run` performs backend admission between the phases. This structure preserves
 one execution-precedence and error-ordering implementation while allowing a
 full backend pool to reject work before expensive schema, artifact, and
-rendering operations.
+rendering operations. `Prepare` discards the plan after returning its public
+metadata. `Run` retains the plan through initial and repaired-output validation
+and discards it when the operation ends. Prepared execution stores the same
+kind of plan only in its private payload.
 
 Pointer-based numeric overrides preserve an explicit zero. Invalid negative or
 out-of-range values fail as invalid requests. Endpoint overrides do not change
@@ -77,7 +114,10 @@ the prompt session template, and is applied after ordinary message rendering.
 A blank direct value retains prompt-template behavior. The runner clears the
 template only on a value copy of the definition, so the definition hash always
 describes the original source while the rendered-prompt hash includes the
-effective direct or rendered session.
+effective direct or rendered session and the complete effective message
+sequence. The rendered-prompt hash uses a versioned, length-framed SHA-256
+encoding of that session, every role and content value, and cache-control
+presence and values; its hexadecimal value is opaque.
 
 The registry is read-only after engine construction. Concurrent `Prepare` and
 `Run` calls resolve independent defensive backend values and keep all
@@ -90,31 +130,67 @@ its `RunAdmitter` to reserve capacity for the effective backend ID. A nil
 admitter is an internal unlimited fallback. After successful admission, `Run`
 immediately defers the returned release function, performs the completion
 phase, makes one initial generation call, builds the named output artifact,
-and validates that artifact. Invalid generated content remains a validation
+and validates that artifact with the plan compiled during completion. Invalid
-result; an inability to perform validation is an operational error.
+generated content remains a validation result; an inability to generate or
+perform validation is an operational error.
+
+When the public facade installs an invocation-local observation recorder,
+ordinary `Run` reuses its run ID and carries it to the generation coordinator
+through context. Prompt and profile identities are enriched only after their
+respective resolutions succeed, before admission. The facade starts observation
+before request conversion and emits its terminal event after public error
+mapping and after the runner's deferred admission release. Preparation and
+prepared execution follows the same terminal mapping after a successful claim;
+invalid or discarded handles do not acquire an observer.
+
+Validation preparation and execution honor cancellation at every
+Promptkit-controlled boundary and do not publish a partial plan or result.
+Schema reads are bounded and context-checked between chunks; JSON decoding,
+schema compilation, and schema execution are checked immediately before and
+after their synchronous calls. Promptkit does not move arbitrary filesystem or
+JSON Schema work to background goroutines, so an already-blocked dependency
+method must return before cancellation can take precedence over its outcome.
 
 The admission lease covers completion-phase preparation, initial generation,
 validation, every repair, and every exit. It bounds accepted work without
 serializing preparation or validation behind the active-generation limit.
-The wrapped model client separately acquires a FIFO active permit only around
+The coordinator separately acquires a FIFO active permit only around each
-each actual generation call.
+actual model attempt. It releases the permit before a transient retry wait.
 
-When an internal repairer is present, a JSON or JSON Schema content failure can
+After a failed `basic`, JSON, or JSON Schema validation with a positive frozen
-trigger bounded repair attempts. Repair receives the effective execution
+budget, the installed repairer can make a bounded corrective call. Each request
-target and session ID, validation errors, prior output, and structured-output
+starts with a fresh copy of the complete effective message sequence (the
-specification. The default repairer uses the same wrapped client as initial
+configured prefix followed by the request suffix), includes only the latest
+nonempty candidate as an assistant message, and appends one corrective user
+message. Empty candidates omit that assistant message. The
+correction carries validation diagnostics as JSON data bounded to 64 KiB; the
+full diagnostics remain in the validation result.
+
+Repair receives the effective execution target, explicit numeric-presence bits,
+credential, backend identity, session ID, structured-output specification, and
+frozen transient retry budget.
+The same request constructor supplies those common fields to initial and repair
+generation. The default repairer uses the same coordinator as initial
 generation, so each repair reacquires the selected backend's active permit
 while remaining inside its original admission lease. Repair never performs a
-second bounded admission. This capability remains internal and is not a public
+second bounded admission, and repaired outputs use the operation's existing
-option.
+validation plan. The runner stops at the first valid candidate, sums completed
+generation usage, reports calls actually started, and returns the final failed
+validation result on exhaustion. A repair generation failure follows the
+ordinary generation-error category rather than becoming a validation error.
 
 A successful result includes the output artifact and raw output, validation
 state, effective session ID, prompt and rendered-prompt hashes, selected
 profile and backend, effective settings, input hashes, token usage, a generated
 run identifier, and UTC timing. The same effective session reaches initial
 generation and any repair attempt through the rendered prompt. The same
-effective target, including backend identity, reaches generation and any
+effective target and presence metadata, including backend identity and direct
-repair attempt.
+credential during execution, reaches generation and every repair attempt.
+Result usage is the field-wise sum of all five usage values from the initial
+response and every completed repair response. Final raw output, artifact, and
+validation state still come from the last candidate. A repair error returns no
+partial run result or partial result usage, although the invocation-local
+observer retains completed-attempt usage and reports it in the terminal event.
 
 ## Failure Categories
 
@@ -124,18 +200,21 @@ validation failures. Wrapping preserves the package identities mapped by the
 public facade and retains collaborator identities where they are part of the
 internal contract.
 
-Admission capacity exhaustion retains the internal capacity identity and adds
+Admission capacity exhaustion retains the internal capacity identity. At the
-the selected backend ID as context. It is not recategorized as an invalid
+use-case boundary, the runner attaches the selected backend ID in an internal
-request or generation failure, and no partial result is returned. A context
+typed error, and the root facade copies that value into the public
-already done at admission retains its context identity directly. Cancellation
+[`CapacityError`](../../capacity_error.go) without parsing diagnostic text. It
-while waiting for an active generation permit prevents client invocation when
+is not recategorized as an invalid request or generation failure, and no
-it wins the grant race; the model-client boundary then preserves the context
+partial result is returned. A context already done at admission retains its
-error through the generation-failure category. Deferred release restores the
+context identity directly. Cancellation while waiting for an active generation
-admission lease on preparation, generation, validation, repair, and
+permit prevents client invocation when it wins the grant race; the model-client
-cancellation failures.
+boundary then preserves the context error through the generation-failure
+category. Deferred release restores the admission lease on preparation,
+generation, validation, repair, and cancellation failures.
 
 Other context cancellation propagates through the invoked collaborator and is
-classified by the owning operation.
+classified by the owning operation. In particular, cancellation observed by
+validation retains the context identity through the validation error category.
 An overlong direct session is an invalid request before source loading, while
 an invalid or overlong prompt session template remains a prompt-render failure.
 An unknown selected backend, or a selected backend with no configured resolver,
@@ -147,10 +226,12 @@ The [runner tests](../../internal/usecase/runner_test.go) own preparation order,
 selection and override precedence, the two-phase boundary, early admission,
 lease lifetime and release, direct-session resolution, schema-before-generation
 behavior, hashing, generation and validation outcomes, backend propagation,
-bounded repair, shared initial/repair capacity, credentials and redaction,
+bounded repair progression, initial/repair request parity, cumulative usage,
-error categories, artifact metadata, usage, and timing. The
+shared initial/repair capacity, credentials and redaction, error categories,
-[capacity subsystem document](capacity.md) identifies the focused pool,
+artifact metadata, and timing. The
-waiter, and wrapped-client tests.
+[capacity subsystem document](capacity.md) identifies the focused pool and
+waiter tests. The [generation document](generation.md) owns attempt and retry
+coordination tests.
 
 Changes to orchestration should continue to use the existing package
 interfaces, keep request state local to an invocation, and preserve the shared

docs/internal/sources.md

@@ -6,15 +6,38 @@ This document describes Promptkit's implemented internal source, artifact,
 rendering, and output-validation behavior. The
 [architecture policy](../policy/architecture.md) owns the library boundary and
 dependency rules. None of these internal packages is a supported consumer API,
-and the root engine assembles them behind its public source options and values.
+and the internal engine assembles them behind the public facade's source options and values.
 The [framework format reference](../formats.md) owns the exact file fields,
 validation modes, built-in catalog, and source precedence.
 
 ## Prompt Definitions
 
-`internal/promptdef` discovers YAML deterministically, decodes and validates
+`internal/promptdef` uses one source-neutral flow for prompt selection and
-definitions, selects an ID and optional version, and resolves file-backed
+normalization. That flow scans normalized YAML ID and version metadata,
-message content within the selected operating-system or `fs.FS` source.
+requires one strictly decoded document per file, classifies errors for the
+selected definition, detects duplicates, and normalizes the exact match.
+Small operating-system and `fs.FS` adapters own discovery, byte reads, display
+paths, content opening, and root containment. Each lookup remains a
+point-in-time scan: definitions and catalogs are not cached, and file-backed
+message content is opened only for the exact selected candidate.
+
+Message roles are normalized through the shared domain owner by trimming
+Unicode whitespace and lowercasing. Only `developer`, `system`, `user`, and
+`assistant` are published; invalid roles remain selected prompt-definition
+failures rather than becoming request errors. Cache-control metadata uses the
+same shared domain normalization and defensive-copy rule.
+
+Operating-system sources enforce containment against canonical roots and
+targets so symlinks cannot escape. Injected `fs.FS` sources enforce containment
+in their clean relative path namespace. A single-file source uses the selected
+prompt file's containing directory as its root. Every content path must be
+relative and is opened from its exact parsed text after a separate blank check;
+contained parent components and whitespace-bearing names remain valid.
+
+Exact prompt inspection performs one point-in-time lookup through that same
+repository and validates referenced message content before returning declared
+metadata. It does not parse templates or read profile, input, or schema
+sources, and it does not retain the definition for a later execution.
 
 Its package tests own prompt selection, strict decoding, definition validation,
 duplicate detection, and source containment:
@@ -22,30 +45,88 @@ duplicate detection, and source containment:
 
 ## Profiles And Built-Ins
 
-`internal/profile` loads and validates execution profiles from an
+`internal/profile` owns immutable in-memory and filesystem repositories, local
-operating-system filesystem or an `fs.FS`. It supports a primary repository
+validation, overlays, and resolution for execution profiles. In-memory
-with fallback only when the primary reports that a profile is absent. Strict
+definitions are normalized and validated once at that repository boundary and
-YAML decoding recognizes the optional `backend` field, trims its value, and
+are defensively copied on storage and lookup. A file contains
-requires a model plus at least one non-blank backend or endpoint. Loading does
+exactly one YAML document and its trimmed YAML `id` is its only selection
-not check registry membership because the available registry belongs to the
+identity; filenames do not confer authority. Each point lookup reads discovered
-assembled engine; the runner checks membership during preparation.
+files once for their metadata and reuses the selected file's bytes for strict
+decoding; unrelated profiles are not fully decoded. Strict selected decoding
+recognizes `base_profile` and the optional `backend` field, trims their values,
+and permits inherited target fields only when a base is named. File-backed
+`extra_params` values are validated and defensively copied through the shared
+bounded JSON-value owner before a profile is published. OpenAI-compatible
+reserved-field policy remains with the model-client and backend-registry owners.
 
-`internal/profile/builtin` embeds the maintained built-in profile catalog and
+`LoadFSRepository` is the eager immutable loading boundary for internal
-can place a caller-selected repository ahead of that catalog. Every embedded
+catalog consumers. It discovers and strictly validates every raw profile once,
-profile selects `openrouter` and inherits its endpoint and credential
+preserves safe source metadata including explicitly present YAML fields, and
-environment-variable name from the built-in backend registry rather than
+publishes independently copied values from memory. It does not resolve profile
-repeating those values. Profile behavior is owned by the
+inheritance. Configured consumer sources continue to use the lazy point lookup
-[profile repository tests](../../internal/profile/repository_test.go), while
+repositories described above.
-catalog completeness, the backend-selection invariant, duplicate IDs, and
+
-overlay behavior are owned by the
+`internal/catalog` validates the imported immutable OpenRouter and
-[built-in repository tests](../../internal/profile/builtin/repository_test.go).
+Rakestrawhome asset modules as one private adapter boundary. It enforces their
+manifest, layout, profile ownership, inheritance, and secret-safety rules
+before returning raw catalog sources. Internal engine assembly uses those validated
+catalogs as the lowest-precedence profile source.
+
+Normalized external catalog profile IDs with the exact `promptkit-` prefix are
+rejected at this adapter boundary. That reservation applies only to imported
+catalogs; configured consumer sources retain the ordinary profile-ID rules.
+
+The overlay repository consults the next repository only when the
+higher-precedence repository reports that a profile is absent. A reliably
+selected malformed profile stops fallback, while an unrelated malformed file
+does not become authoritative through its filename. Loading does not check
+backend registry membership because the available registry belongs to the
+assembled engine; the runner checks membership during preparation and exact
+profile inspection.
+
+The internal engine assembles one raw composite catalog in precedence order:
+in-memory profiles, one ordinary configured source, an application fallback
+source, Promptkit-owned profiles, then the maintained external catalog. It
+constructs the owned definitions through the in-memory repository and validates
+their maintained-only resolution and backend membership before consumer sources
+are overlaid. An explicit file or `fs.FS` profile source replaces
+`Config.ProfileDir` within the ordinary configured-source category. One outer
+resolving repository wraps that complete raw catalog, so each base lookup
+observes the same precedence and shadowing rules.
+
+The resolving repository traverses every selected chain afresh, retains no
+cache, detects cycles, limits a chain to 32 profiles, merges root-to-leaf into a
+new caller-owned value, and validates the final target before publishing it. It
+does not check backend registry membership. Exact `base_profile` syntax, merge
+rules, and consumer-visible failure behavior belong to the [framework format
+reference](../formats.md#profile-inheritance).
+
+Exact profile inspection performs one point-in-time resolved lookup through
+those profile sources and checks the final target without reading prompt, input,
+or schema sources. It does not retain that lookup for a later execution.
+Prepared execution instead freezes the fully resolved target; a later ordinary
+operation performs a fresh traversal.
+
+The imported external catalog provides external profiles and matching backend
+definitions. The internal engine owns the Promptkit profiles documented in the
+[framework format reference](../formats.md#built-in-profile-catalog). Profile
+loading and overlay behavior are owned by the [profile repository tests](../../internal/profile/repository_test.go), while
+catalog completeness, backend selection, and duplicate IDs are owned by the
+[catalog adapter tests](../../internal/catalog/catalog_test.go).
 
 ## Ordinary Artifacts
 
-`internal/artifact` resolves inline references and unrestricted,
+`internal/artifact` accepts explicitly typed inline references even when their
-caller-selected file paths. It copies content into an artifact, records
+body is empty. It also resolves unrestricted, caller-selected paths only when
-metadata and a content hash, applies a content-type fallback, and honors
+they identify regular operating-system files, checking that condition before
-context cancellation.
+and after opening the file. It copies content into an artifact, records
+metadata and an opaque content-equality value, and applies a content-type
+fallback.
+
+Regular files are read synchronously in bounded chunks. Cancellation is
+checked before opening, before and after every read, and before publishing the
+artifact, so a canceled read never publishes partial content. The ordinary
+reader does not detach file reads into background goroutines.
 
 This ordinary reader does not implement an inbound HTTP security boundary. In
 particular, it does not constrain files to an application root or impose an
@@ -57,7 +138,17 @@ implemented reader behavior and failures.
 ## Rendering
 
 `internal/prompt` renders definition messages as Go templates using named
-artifacts and variables. It carries message roles, session IDs, and cache
+artifacts and variables. Within one render, each referenced artifact body is
+converted to text lazily and cached by input name for reuse across the session
+and every message; the cache is not shared across renders. Conversion uses
+bounded chunks and preserves the artifact bytes exactly.
+
+Session and message parsing and execution remain synchronous. The renderer
+checks cancellation before and after each parse and execution boundary,
+between artifact conversion chunks, around each message, and before publishing
+the complete prompt. It cannot interrupt template work already in progress and
+never publishes a partial prompt after observing cancellation. It validates and
+canonicalizes message roles before carrying roles, session IDs, and cache
 control into the rendered prompt. The
 [renderer tests](../../internal/prompt/renderer_test.go) own rendering behavior.
 
@@ -68,6 +159,35 @@ filesystem or an `fs.FS`. Invalid generated content is returned as a validation
 result; inability to load, register, or compile a schema is an operational
 error.
 
+Every preparation operation creates one operation-local validation plan. None,
+basic, and JSON modes retain the effective output contract without source
+access. JSON Schema mode loads the root document once, resolves and compiles
+each transitive reference, and retains the compiled validator. Schema compiler
+resources use canonical escaped file or private-scheme URLs; loaders decode
+their paths once and enforce the configured source boundary. The
+provider-facing structured-output metadata uses the root document captured by
+the same plan.
+
+Schema preparation and execution remain synchronous. Promptkit checks
+cancellation before and after source resolution, JSON decoding, compilation,
+and validation, and between bounded schema-read chunks. Once cancellation is
+observed it returns the context error without publishing a partial plan or
+validation result, even when a compiler or validator has just returned a
+different error or a successful result. An `fs.FS` method or JSON Schema
+dependency call already in progress cannot be preempted; Promptkit waits for
+that call to return and then gives cancellation precedence. Validation does
+not detach dependency work into background goroutines.
+
+`Prepare` discards its validation plan after returning metadata. `Run` retains
+its plan for initial and repaired-output validation, then discards it with the
+operation. `PrepareExecution` retains the plan in its private frozen payload;
+`RunPrepared` uses that plan without reopening prompt, profile, input, or
+schema sources or rerendering the request. A later ordinary `Run` always
+performs fresh source resolution and preparation.
+
 The [validator tests](../../internal/validate/standard_validator_test.go) own
-basic, JSON, JSON Schema, source resolution, schema loading, compilation, and
+basic, JSON, JSON Schema, source resolution, schema loading, compilation,
-content-failure behavior.
+frozen-reference behavior, content-failure behavior, and the synchronous
+cancellation boundary. Prepared execution
+orchestration is owned by the
+[use-case tests](../../internal/usecase/prepared_execution_test.go).

docs/policy/agents.md

@@ -0,0 +1,88 @@
+# Agent Policy
+
+**Base Policy Version:** 1.1
+
+**Last Updated:** 2026-09-23
+
+## Purpose
+
+Agents must follow these house rules before recommending, planning, or
+implementing repository changes.
+
+## Implementation Discipline
+
+Write the simplest idiomatic code that satisfies the user's request.
+Simplify the mechanism, not the user's choices. Do not implicitly strengthen
+requirements, restrictions, or guarantees.
+
+**Keep implementation machinery proportionate.** Use straightforward,
+idiomatic code and introduce only the machinery needed to support the requested
+behavior and flexibility. Optimize for conceptual simplicity and
+maintainability, not for the fewest lines, fewest options, or narrowest product
+surface.
+
+**Do not invent restrictions.** Do not reject, prohibit, constrain, cap,
+harden, or otherwise narrow behavior unless required by the user, an explicit
+repository contract, or an unavoidable correctness constraint. Preserve
+operator and caller choice where policy has not been established. If an
+additional restriction appears advisable, present it as a decision rather than
+silently implementing it.
+
+A correctness constraint must follow from the requested behavior or an
+established invariant, not from the convenience of the chosen implementation.
+Explain the specific failure it prevents and why a less restrictive
+implementation cannot satisfy the contract.
+
+**An implementation detail is not repository policy.** Tests and documentation
+must not elevate incidental behavior into a durable guarantee or restriction
+unless that behavior was expressly selected as part of the contract. Existing
+explicit contracts remain authoritative; changing or weakening one requires a
+deliberate decision rather than being inferred from this principle.
+
+Agent-authored plans, tests, documentation, and prior implementations are not
+evidence that an additional restriction was deliberately approved. Identify the
+authorizing requirement or decision; when that authority is unclear, surface
+the ambiguity rather than treating repetition as approval.
+
+**Do not invent guarantees.** Do not strengthen public contracts, guarantees,
+or invariants beyond those expressly requested by the user.
+
+**Use the standard library.** Prefer simple, idiomatic standard-library
+implementations when they satisfy the requested behavior. Do not add custom
+power-loss protection or defenses against privileged concurrent filesystem
+actors unless expressly required. Preserve existing repository contracts.
+
+## Git Commit Policy
+
+By default, create a commit after completing each coherent unit of work, using a
+concise, plain-English message describing what changed. This policy is standing
+authorization to create those local commits; a separate user request is not
+required. An explicit user instruction to leave work uncommitted takes precedence.
+
+A unit of work is a focused, reviewable outcome, including its relevant tests and
+documentation. Do not commit after every intermediate edit or create empty commits.
+Complete the appropriate validation before committing. If work is incomplete or
+validation is blocked, report that state rather than presenting it as completed.
+
+Before committing:
+
+1. Run `git status --short` and review the relevant diff, including any staged
+   changes. Identify changes that predate the current unit of work.
+2. Exclude secrets, local environment files, generated junk, and unrelated changes.
+3. Stage only the files or hunks belonging to the completed unit. Use `git add .`
+   only when every included change belongs to that unit. Do not disturb unrelated
+   staged or unstaged work.
+4. Review `git diff --staged` and run appropriate checks for the change. Do not
+   repeat checks already passed for unchanged content merely to create a commit.
+5. Review `git log --oneline -n 10` and use
+   `git commit -m "<concise plain-English commit message>"`.
+6. Report the commit identifier and any remaining uncommitted work.
+
+If the unit cannot be safely separated from pre-existing work, leave it uncommitted
+and explain the overlap rather than silently including that work. If a commit
+fails, preserve the changes and report the failure; do not bypass hooks.
+
+Default commit authorization does not authorize pushing, publishing, amending,
+resetting, reverting, or rewriting history. Those actions require a separate user
+instruction. Do not use `git commit -a` or `git commit --no-verify` to bypass the
+staging and validation discipline above.

docs/policy/architecture.md

@@ -1,5 +1,9 @@
 # Architecture Policy
 
+**Base Policy Version:** 1.0
+
+**Last Updated:** 2026-09-23
+
 ## Purpose
 
 This document defines Promptkit's current high-level architecture and the
@@ -17,53 +21,25 @@ The module root contains package `promptkit`, which is the public facade. It
 provides the supported engine, configuration and source options, requests,
 results, public values, extension interfaces, profiles, and error sentinels.
 
-The implemented internal components consist of:
+The [internal component overview](../internal/overview.md) owns the inventory
+of implemented packages and their concrete responsibilities. This policy owns
+the boundaries and dependency rules those components must preserve.
 
-- `internal/domain`, which owns framework data values shared by later internal
+The internal engine assembles one immutable backend registry, one capacity manager,
-  components;
-- `internal/backend`, which owns validated immutable OpenAI-compatible backend
-  definitions and the built-in OpenRouter definition;
-- `internal/capacity`, which owns engine-local bounded run admission and
-  model-generation scheduling for limited backends;
-- `internal/defaults`, which owns application-neutral framework defaults and
-  constructs the default execution target;
-- `internal/filecatalog`, which discovers YAML files and provides source-path
-  helpers for filesystem and `fs.FS` consumers;
-- `internal/jsonvalue`, which validates and defensively copies JSON-compatible
-  extra-parameter trees;
-- `internal/promptdef`, which loads and validates prompt definitions from
-  filesystem and `fs.FS` sources;
-- `internal/profile`, which loads, validates, and overlays execution profiles
-  from filesystem and `fs.FS` sources;
-- `internal/profile/builtin`, which embeds the built-in execution profile
-  catalog;
-- `internal/prompt`, which renders prompt messages from Go templates;
-- `internal/artifact`, which resolves ordinary inline and unrestricted
-  caller-selected file references;
-- `internal/validate`, which validates basic, JSON, and JSON Schema output
-  using filesystem and `fs.FS` schema sources;
-- `internal/llm`, which defines the provider-neutral generation boundary and
-  implements outbound OpenAI-compatible chat requests; and
-- `internal/usecase`, which coordinates preparation and execution across the
-  internal framework components.
-
-The `examples/go-library/prepare` and `examples/go-library/run` packages are
-maintained downstream consumers of the root facade. They do not expose library
-packages or participate in internal assembly.
-
-The root facade assembles one immutable backend registry, one capacity manager,
 the internal repositories, renderer, validator, outbound client, and use-case
-runner while translating public values and errors at the library boundary. The
+runner. The registry contains validated maintained definitions plus engine-scoped
-registry contains built-ins plus validated engine-scoped consumer additions.
+consumer additions. The root facade translates public values and errors at the
-The facade constructs the capacity manager from the registry's immutable
+library boundary.
+The internal engine constructs the capacity manager from the registry's immutable
 policy snapshot, wraps the selected built-in or injected model client, and
 supplies bounded admission to the runner. The defaults and renderer depend on
 the domain model. Prompt-definition and profile repositories use the domain
-model, file catalog, and YAML decoder. The built-in profile repository supplies
+model, file catalog, and YAML decoder. The catalog adapter validates imported
-an embedded `fs.FS` to the profile package. Artifact reading uses the domain
+immutable maintained data, including its reserved namespace, before internal
-model and application-neutral defaults. Validation uses the domain model, file
+engine assembly supplies it to the profile package. Artifact reading uses the
-catalog, and JSON Schema implementation. The model client uses the domain
+domain model and application-neutral defaults. Validation uses the domain model,
-model, application-neutral defaults, and an injected or standard-library HTTP
+file catalog, and JSON Schema implementation. The model client uses the domain
+model and application-neutral defaults, and an injected or standard-library HTTP
 client. The use-case runner depends on the narrow interfaces owned by each
 internal component, including backend lookup and run admission.
 
@@ -92,6 +68,13 @@ coordinates internal components and adapts the supported public extension
 interfaces to narrow internal abstractions. Internal components must not depend
 on consumers or on Scriptorium.
 
+`internal/domain` owns source-neutral invariants for values shared across
+multiple input and execution boundaries, including execution-setting bounds,
+OpenAI-compatible base endpoints, session identifiers, and output-contract
+legality. Callers retain source parsing, required-field rules, other
+source-specific normalization, defaulting, error classification, and policy
+specific to their own boundary.
+
 ## Repository And Consumer Boundary
 
 Scriptorium is a downstream application that consumes Promptkit through
@@ -118,8 +101,10 @@ consumer contract requires a public package.
 Each package must have one cohesive responsibility and a clear dependency
 direction. Internal packages must not expose their types merely to simplify
 wiring, and the public facade must not leak internal representations through
-exported signatures. New public packages require a durable consumer need that
+exported signatures. Production packages under `internal/` must not import the
-cannot be served cleanly by the root facade.
+module root; the root facade may depend on internal implementation packages.
+New public packages require a durable consumer need that cannot be served
+cleanly by the root facade.
 
 The [internal component overview](../internal/overview.md) must be updated as
 packages are implemented or their responsibilities change.

docs/policy/documentation.md

@@ -1,5 +1,9 @@
 # Documentation Policy
 
+**Base Policy Version:** 1.1
+
+**Last Updated:** 2026-09-23
+
 ## Purpose
 
 This policy assigns each Promptkit documentation topic to one canonical owner.
@@ -68,17 +72,19 @@ mechanisms, not secret values.
 
 | Topic | Canonical owner | Owned content | Content owned elsewhere |
 | --- | --- | --- | --- |
-| Project orientation | `README.md` | What Promptkit is, its current usability, module identity, license summary, and links onward. | Exact API contracts, contributor procedures, architecture detail, and release steps. |
+| Project orientation | `README.md` | What Promptkit is, its current usability, module identity, license summary, a minimal successful quickstart, and links onward. | Exact API contracts, contributor procedures, architecture detail, and release steps. |
 | Contributor workflow | `docs/development.md` | Task-oriented reading guide, local workflow, validation commands, and repository hygiene. | Architecture rules, API semantics, subsystem behavior, and release procedure. |
 | Current architecture | `docs/policy/architecture.md` | System shape, normative ownership, dependency direction, package boundaries, invariants, and non-goals. | Concrete component inventory, implementation mechanics, contributor procedures, decision history, and future work. |
+| Agent policy | `docs/policy/agents.md` | Agent implementation discipline, proportionality, idiomatic simplicity, local commit authorization and discipline, and rules against inventing restrictions, guarantees, or repository policy. | Public contracts, architecture, testing, and contributor procedures. |
 | Documentation organization | `docs/policy/documentation.md` | Documentation ownership, audience boundaries, maintenance rules, and document lifecycle. | Library architecture or runtime behavior. |
 | Testing policy | `docs/policy/testing.md` | Test philosophy, risk-based sufficiency, test boundaries, doubles, coverage guidance, regression policy, and test maintenance. | Subsystem behavior, exact public contracts, subsystem-specific test inventories, and implementation plans. |
-| Release procedure | `docs/release.md`, when present | Required release validation, version and tag procedure, release ordering, and post-publication checks. | General contributor workflow, public API semantics, and decision history. |
+| Release policy | `docs/policy/release.md` | Versioning, compatibility, release-note requirements, validation obligations, immutability, recovery, and adoption scope. | Executable commands, publication mechanics, and historical release summaries. |
+| Release procedure | `docs/release.md` | Maintainer prerequisites, candidate preparation, concrete publication guards and commands, verification, and recovery steps. | Release-policy requirements, public API semantics, and historical summaries. |
 | Exact exported Go API | Go declarations and GoDoc, as APIs are implemented | Exported names, signatures, types, values, errors, and exact behavioral contracts. | Task-oriented consumer walkthroughs, implementation details, and future API proposals. |
 | Framework file formats | `docs/formats.md` | Prompt-definition and profile YAML fields, schema references, defaults, validation modes, built-in profiles, credentials, and file-to-request precedence. | Exported Go declarations, outbound wire behavior, internal parsing mechanics, and application configuration. |
 | Consumer guidance | `docs/consumers/`, when consumer workflows require dedicated guidance | Task-oriented use of implemented public APIs, minimal examples, and consumer responsibilities. | Exact exported declarations and internal mechanics. |
 | Durable integration contracts | `docs/integrations/`, when integrations exist | External formats and protocols, compatibility behavior, and upstream or downstream responsibilities. | Internal transformations and public Go declarations. |
-| Supplemental release guidance | None. `docs/releases/` may be used when a release benefits from a changelog or migration guide. | No canonical content. These files may briefly summarize release-specific changes, compatibility, and consumer migration paths, and may be corrected, consolidated, archived, or removed when no longer useful. | Public API and behavior contracts, formats, integrations, architecture, release procedure, and the authoritative annotated-tag release record. |
+| Release-note history | `docs/releases/` | Mandatory version-matched historical summaries for releases made under the adopted release policy; the note at the immutable tag records that release. | Current APIs, behavior, formats, compatibility contracts, and release procedures. |
 | Implemented component inventory | `docs/internal/overview.md` | Current packages and components, their implemented responsibilities, and links to focused internal documents. | Normative architecture, contributor workflow, external contracts, and proposed components. |
 | Internal subsystem behavior | Other files under `docs/internal/`, when a subsystem needs durable detail | Implementation flow, internal collaborators and state transitions, package-local guarantees and failures, and relevant tests. | Global architecture invariants, public API definitions, and future package plans. |
 | Architectural decision history | `docs/adr/`, when repository-local decisions require records | Significant decisions, context, alternatives, rationale, consequences, and supersession history. | Current behavior reference, implementation status, and task sequencing. |
@@ -110,22 +116,22 @@ but must link to its canonical definition rather than restate it.
 The [framework format reference](../formats.md) owns exact prompt, profile, and
 schema-file contracts. Integration documents own external wire formats.
 
-### Supplemental Release Guidance
+### Release Documentation
 
-Files under `docs/releases/` may provide changelog-style summaries and
+The [release policy](release.md) owns release requirements; the
-migration guidance for a particular release. They are navigation and
+[repository release procedure](../release.md) owns their concrete execution.
-orientation aids, not canonical owners of public APIs, behavior, formats,
+Version-matched checked-in notes are required for future releases under that
-integrations, architecture, release procedure, or other durable facts. When a
+policy. The source tag is the release identity, and the note in its immutable
-reader needs detail beyond a short release-specific note, the release document
+tree is the historical summary; annotated tag messages complement that record.
-must link to the applicable canonical documentation rather than reproduce its
-contract.
 
-The annotated tag message required by the
+Release notes do not own current API or behavior contracts. Link to the
-[release procedure](../release.md#write-the-release-note) remains the
+canonical owners for details instead of reproducing those contracts. Published
-authoritative release record. Supplemental release documents may be corrected,
+notes are not disposable guidance and must not be replaced under the same
-consolidated, archived, or removed at any time when they are no longer useful,
+version. Corrections belong in a subsequent release.
-provided maintained documentation does not depend on them and the annotated
+
-tag record remains intact.
+Adoption is prospective. Previously published tags and historical notes retain
+their original records and are not reformatted or rewritten to satisfy the new
+requirements.
 
 ### Security Topics
 
@@ -178,9 +184,8 @@ durable owners, update incoming links, and archive or remove the roadmap
 according to repository practice. Do not preserve completed roadmaps as a
 second current-state reference.
 
-Supplemental release documents may likewise be removed without preserving a
+Retain release notes as historical records under the release policy. Do not
-replacement. Before removal, update maintained incoming links so current
+apply temporary-roadmap retirement rules to published release notes.
-documentation does not depend on an optional historical guide.
 
 Before completing documentation work:
 

docs/policy/release.md

@@ -0,0 +1,146 @@
+# Release Policy
+
+**Base Policy Version:** 1.0
+
+**Last Updated:** 2026-09-23
+
+## Purpose And Adoption Scope
+
+This policy owns Promptkit's release requirements and compatibility rules.
+The [release procedure](../release.md) owns the commands used to satisfy them.
+This scoped adoption applies to releases prepared after its adoption on
+2026-09-23. Published tags and historical notes are not rewritten to retrofit
+the new note format.
+
+## Release Model And Compatibility
+
+Promptkit publishes source-only Go module releases from committed, pushed
+`main` candidates through annotated `vMAJOR.MINOR.PATCH` tags. It publishes
+no binaries, hosted releases, or other artifacts and uses maintainer-run
+validation rather than hosted CI. Versions must be unused locally and on
+`origin`; the existing procedure accepts stable version tags only.
+
+Before v1, minor releases may change the public API and patches preserve
+compatibility within their minor line, subject to the catalog-data provision
+below. Release notes must identify compatibility impact, public API changes,
+and required consumer action. The source tag identifies the exact release;
+the checked-in note at that tag records its historical summary.
+
+## Catalog Profile Compatibility
+
+External catalog profile data may be added, removed, renamed, or modified
+without deprecation or advance notice. This exception is limited to profile
+data imported from the maintained catalog modules. It does not weaken ordinary
+compatibility for Promptkit's public Go API, configuration and format semantics,
+runtime behavior, catalog-module APIs, manifest schema, backend definitions, or
+consumer-owned definitions.
+
+Within the catalog versions selected and tested by a Promptkit release, and
+unless a consumer overrides the name or its base, Promptkit keeps the literal
+profile IDs `promptkit-openrouter-standard` and
+`promptkit-openrouter-auto` available. The names do not promise a permanent
+model, backend, routing behavior, settings, cost, output, or upstream
+availability. The [framework format reference](../formats.md#built-in-profile-catalog)
+owns their current definitions.
+
+`v0.10.0` is the initial minor release for this compatibility transition.
+Afterward, catalog-only updates and adjustments to the current selections may
+ship in a patch release when all other compatibility obligations remain intact.
+External profile removals can require consumer action; they are not automatic
+migrations.
+
+Catalog adoption and release validation remain local and offline. They do not
+probe provider endpoints, request model lists, or make inference calls.
+
+## Release Documentation
+
+Every new release must include `docs/releases/<version>.md` in the tagged
+source, using the complete tag as the filename. Its title contains Promptkit
+and that exact version. Use this structure:
+
+```markdown
+# Promptkit vMAJOR.MINOR.PATCH
+
+This release ...
+
+## Summary
+
+## Compatibility
+
+## Upgrade
+
+## Changes
+```
+
+State “No action required” or “None” where appropriate rather than omitting a
+section. Keep exact current contracts in their canonical documents and link
+to them. Update affected canonical documentation in the same candidate.
+
+These notes are historical records, not optional disposable guidance. Published
+release-note contents must not be replaced under the same version. Corrections
+belong in a subsequent release. Annotated-tag messages complement the checked-in
+note: generate the annotation from that note and the validated commit rather
+than maintaining a second summary.
+
+## Validation And Publication
+
+The [publication script](../../scripts/release.sh) runs the
+[local checks](../../scripts/release-check.sh) once for the final candidate.
+Retain clean-candidate, synchronized-main, unused-version, module-hygiene,
+release-note structure, and tag-target checks. Publish only the selected tag
+ref. Changing the candidate requires fresh validation. No provider endpoint
+probes or inference calls are release checks.
+
+Release-note content and retirement of completed roadmaps are preparation
+responsibilities, not separate human sign-off gates. Preserve planned and
+deferred work and repair links. Ordinary change review covers unintended files,
+secrets, and private data; do not repeat it as a release-specific approval.
+
+A successful tag push to origin completes the release. Never wait for CI to
+finish or succeed. CI results, remote read-back, and Go module availability
+are informational follow-ups, not completion gates. An ambiguous push result
+requires confirming publication on origin, not waiting for CI.
+
+## Immutability And Failure Recovery
+
+Never move, replace, recreate, or delete a published tag or replace published
+release contents. If validation fails before publication, correct the candidate,
+rerun the complete checks, and repeat the publication guards. Remove an
+unpublished local tag only after confirming it was not published.
+
+If publication or subsequent investigation reveals a defect, leave the
+published release intact. Correct the defect on main, select a new patch or
+other semantically appropriate version, write its matching release note, and
+repeat the complete procedure. If publication status is uncertain, inspect
+the authoritative remote before deciding whether a tag is unpublished; do not
+overwrite it to retry.
+
+Transient network, authentication, CI, or module-availability failures do not
+by themselves warrant a new version. Investigate separately and publish a
+corrective release only when published contents actually need changing.
+
+## Policy Exceptions And Deliberate Scope
+
+This adoption follows the maintainer-approved lean workflow rather than every
+base-policy guard:
+
+- Collision checks concern origin, the authoritative publication destination,
+  rather than every configured remote. Local and origin tag checks, synchronized
+  main, and an explicit selected-ref push protect the actual release channel.
+- Automated local validation and publication-state checks replace manual
+  mechanical inspections and separate human sign-offs. Ordinary change review,
+  editorial responsibility for release notes, and roadmap housekeeping remain;
+  repeating them as approvals would add friction without distinct protection.
+- Successful tag publication completes this source-only release. CI can be slow
+  or unreliable, and module availability depends on external services. Neither
+  is a completion gate. Local candidate validation and explicit tag-target
+  checks protect publication; ambiguous push results are resolved against origin.
+  Actual defects are corrected in a new release, never by replacing a tag.
+
+Ignored local files are not prohibited merely because they exist: they are not
+included in the tagged commit. The clean-candidate guard and ordinary review of
+tracked changes protect the published contents.
+
+The catalog profile provision preserves an already approved compatibility
+boundary; it does not extend volatility to backend definitions or runtime
+contracts. Further release-governance changes require deliberate approval.

docs/policy/testing.md

@@ -1,5 +1,9 @@
 # Testing Policy
 
+**Base Policy Version:** 1.0
+
+**Last Updated:** 2026-09-23
+
 ## Purpose
 
 Our tests exist to make **incorrect changes expensive and correct changes cheap**.
@@ -50,19 +54,25 @@ Examples of appropriate seams include clocks, randomness, subprocesses, remote A
 ## Test execution requirements
 
 Promptkit currently uses maintainer-run validation rather than hosted CI.
-Maintainers run the repository-documented test, vet, build, formatting,
+Ordinary work requires validation proportionate to its scope and risk. Do not
-documentation-link, and repository-hygiene checks before accepting changes.
+repeat successful checks for unchanged content merely to create a commit.
+Release candidates require the complete local workflow in the
+[development guide](../development.md#maintainer-validation). That guide routes
+contributors to the executable checks for formatting, documentation links,
+and repository hygiene.
 Introducing hosted CI later would supplement, not silently redefine, this
 documented validation model.
 
-The complete test sequence includes ordinary and race-enabled package tests.
+Full release-candidate validation includes race-enabled package tests, explicit
-The maintained offline consumer workflow is also run from the repository root:
+static analysis, and execution with automated semantic checks of the three maintained
-
+offline consumer examples. The test run builds the packages; the examples
-```sh
+also exercise ordinary compilation. Separate ordinary tests and a standalone
-go test ./...
+build are not additional release gates. Ordinary tests remain useful during
-go test -race ./...
+development. The preparation example protects assembled preparation and
-go run ./examples/go-library/prepare
+inspection behavior. The execution example separately protects assembled
-```
+`Run`, injected-client, validation, usage, and result behavior. The observation
+example protects assembled built-in retries, structured proposal and assessment
+runs, and application-owned interpretation of metadata.
 
 Tests in the default suite must be deterministic, offline, and independent of
 real credentials. They must not invoke paid APIs, use live network
@@ -88,8 +98,9 @@ Use each test type where it protects a distinct risk:
   interaction, while replacing live or nondeterministic external boundaries.
 - External-package root tests exercise the public facade as a Go consumer,
   while internal package tests own focused implementation behavior.
-- The maintained offline preparation example protects one representative
+- The maintained offline preparation, execution, and observation examples protect distinct
-  assembled consumer workflow without contacting a model provider.
+  representative assembled consumer workflows without contacting a model
+  provider.
 - Fixtures should be minimal, synthetic, versioned with the behavior they
   exercise, and free of credentials or private data.
 - Golden files are appropriate only when the complete output is intentionally

docs/release.md

@@ -1,278 +1,82 @@
 # Release Procedure
 
-## Release Model
+This runbook implements the [release policy](policy/release.md). A successful
+push of the selected annotated tag to origin completes the release. Never wait
+for CI to finish or succeed.
 
-Promptkit publishes a Go library through source commits and semantic Go module
+## Prepare The Candidate
-tags. It does not publish runnable binaries or binary packages and does not
-currently use hosted CI. The release maintainer performs and records the
-required validation.
 
-`v0.1.0` is the initial published release. Later releases use semantic
+Required tools are Git, Go with race-detector support, POSIX shell, and Python
-`vMAJOR.MINOR.PATCH` tags. Before `v1`, minor releases may change the public
+3. Publication requires push access to origin. Work from the repository root.
-API and patch releases preserve compatibility within their minor line. Every
-pre-`v1` release note must summarize compatibility, identify public API
-changes, and state any action required of consumers.
 
-Promptkit releases are source-only. The annotated tag message is the release
+Select an unused version and create `docs/releases/<version>.md` using the
-note; there is no separate hosted release or binary packaging step.
+[required structure](policy/release.md#release-documentation). Describe
+compatibility impact and consumer actions and update affected canonical docs.
+For catalog updates, record selected catalog versions and any consumer action,
+linking to the catalog policy rather than repeating its contracts.
 
-## Establish The Candidate
+Retire completed roadmaps, preserving planned or deferred work and repairing
+links. This is housekeeping, not a separate approval gate. Review changes
+through the ordinary development workflow, commit them, and push main to
+origin. No repeated release-specific diff review is required.
 
-Choose a version that has not been published and export it as
+## Validate And Publish
-`RELEASE_VERSION`. Run every command in this procedure from the Promptkit
-repository root in the same POSIX shell. Do not reuse `v0.1.0` or another
-existing version.
 
-The following guard derives the release commit from `HEAD` and stops on a
+Run the following with the chosen unused version (replace the placeholder):
-missing or malformed version, a checkout other than synchronized `main`,
-uncommitted changes, an active Go workspace, a module replacement, a vendor
-tree, or an existing local or remote tag:
 
 ```sh
-set -eu
+./scripts/release.sh vMAJOR.MINOR.PATCH
-
-: "${RELEASE_VERSION:?export an unpublished vMAJOR.MINOR.PATCH version}"
-if ! printf '%s\n' "$RELEASE_VERSION" |
-    grep -Eq '^v(0|[1-9][0-9]*)\.(0|[1-9][0-9]*)\.(0|[1-9][0-9]*)$'
-then
-    printf '%s\n' "invalid release version: $RELEASE_VERSION" >&2
-    exit 1
-fi
-
-RELEASE_COMMIT=$(git rev-parse --verify 'HEAD^{commit}')
-export RELEASE_COMMIT
-
-check_release_candidate() {
-    test "$(git branch --show-current)" = main
-    test -z "$(git status --porcelain)"
-
-    gowork_value=$(go env GOWORK)
-    case "$gowork_value" in
-        ''|off) ;;
-        *)
-            printf '%s\n' "active Go workspace: $gowork_value" >&2
-            return 1
-            ;;
-    esac
-
-    test -z "$(git ls-files go.work go.work.sum)"
-    test ! -e vendor
-    if grep -Eq '^[[:space:]]*replace([[:space:]]|\()' go.mod
-    then
-        printf '%s\n' 'go.mod contains a replacement' >&2
-        return 1
-    fi
-
-    git fetch origin main --tags
-    test "$RELEASE_COMMIT" = \
-        "$(git rev-parse --verify 'refs/remotes/origin/main^{commit}')"
-
-    if git show-ref --verify --quiet "refs/tags/$RELEASE_VERSION"
-    then
-        printf '%s\n' "local tag already exists: $RELEASE_VERSION" >&2
-        return 1
-    fi
-    if test -n "$(
-        git ls-remote --tags origin \
-            "refs/tags/$RELEASE_VERSION" \
-            "refs/tags/$RELEASE_VERSION^{}"
-    )"
-    then
-        printf '%s\n' "remote tag already exists: $RELEASE_VERSION" >&2
-        return 1
-    fi
-}
-
-check_release_candidate
 ```
 
-Do not continue unless the guard completes successfully. In particular, push
+The [publication script](../scripts/release.sh) checks a clean, synchronized
-the intended commit through the normal `main` branch workflow before release;
+main candidate, the version and unused tag, and the release note's title and
-the tag procedure is not a substitute for publishing the source commit.
+required headings. It runs the [local checks](../scripts/release-check.sh)
+once, rechecks candidate identity and publication state, generates the
+annotation from the committed note plus the validated commit, and pushes only
+the selected annotated tag.
 
-## Validate The Candidate
+There is no separate tag-message authoring, example-output inspection,
+metadata inspection, or human sign-off. Compatibility and release-note content
+remain ordinary editorial responsibilities. Development validation is useful
+but does not require a second full run before invoking the publication script.
 
-Confirm the module and root package metadata:
+The script stops successfully after the tag push. CI completion, CI success,
+remote read-back, and Go module availability are not release completion gates.
+Do not wait or poll for them. Provider probes and inference calls are never
+part of release validation.
+
+## Catalog Profile Compatibility
+
+The [release policy](policy/release.md#catalog-profile-compatibility) owns the
+catalog compatibility boundary. This heading preserves older incoming links.
+
+## Failure Recovery
+
+If local checks fail, fix the candidate, commit and push changes, then rerun
+the publication command. It never automatically removes a tag.
+
+If a push fails or its result is ambiguous, set `RELEASE_VERSION` to the
+attempted version and inspect origin before retrying:
 
 ```sh
-go list -m -f '{{.Path}} {{.GoVersion}}'
+git ls-remote --tags origin "refs/tags/$RELEASE_VERSION" "refs/tags/$RELEASE_VERSION^{}"
-go list -f '{{.Name}} {{.ImportPath}}' .
+git rev-parse "refs/tags/$RELEASE_VERSION" "refs/tags/$RELEASE_VERSION^{}"
 ```
 
-The output must be:
+Matching remote tag and peeled commit values confirm publication, even if the
+push command reported an error. If the remote query fails, publication status
+is still unknown. Do not delete or recreate the local tag until absence on
+origin is confirmed. Once confirmed unpublished, an unsuccessful attempt's
+local tag can be removed and the publication command rerun. Never overwrite
+a published tag.
 
-```text
+CI or module-resolution problems may be investigated separately; temporary
-gitea.maximumdirect.net/eric/promptkit 1.25.5
+service, network, or authentication failures do not require another release.
-promptkit gitea.maximumdirect.net/eric/promptkit
+Only an actual defect requiring changes to published contents warrants a new
-```
+appropriate version. See the [recovery policy](policy/release.md#immutability-and-failure-recovery).
-
-Run the complete maintainer validation required by the
-[development guide](development.md):
-
-```sh
-go test ./...
-go test -race ./...
-go vet ./...
-go build ./...
-go run ./examples/go-library/prepare
-```
-
-Check every tracked Go file. This command must produce no output:
-
-```sh
-unformatted=$(
-    git ls-files '*.go' |
-        while IFS= read -r go_file
-        do
-            gofmt -l "$go_file"
-        done
-)
-test -z "$unformatted"
-```
-
-Follow every maintained Markdown link and confirm that its local or published
-target exists. Review the repository for generated binaries, test or coverage
-output, credentials, template residue, downloaded assets, and other files that
-do not belong in source control.
-
-Recheck module and repository hygiene, whitespace, and the clean checkout:
-
-```sh
-test -z "$(git ls-files go.work go.work.sum)"
-test ! -e vendor
-if grep -Eq '^[[:space:]]*replace([[:space:]]|\()' go.mod
-then
-    printf '%s\n' 'go.mod contains a replacement' >&2
-    exit 1
-fi
-git diff --check
-test -z "$(git status --porcelain)"
-```
-
-## Write The Release Note
-
-Prepare a plain-text annotated-tag message outside the repository and export
-its path as `RELEASE_NOTES_FILE`. Use this form, replacing each summary with
-release-specific text; write `None.` when there are no public API changes or
-consumer actions:
-
-```text
-Promptkit vMAJOR.MINOR.PATCH
-
-Validated commit: full commit ID
-Compatibility: compatibility summary
-Public API changes: changes or None.
-Consumer action: required action or None.
-```
-
-After writing it, require all release-note fields, the selected version, and
-the validated commit to be present:
-
-```sh
-: "${RELEASE_NOTES_FILE:?export the path to the release-note file}"
-test -f "$RELEASE_NOTES_FILE"
-test -s "$RELEASE_NOTES_FILE"
-grep -F "Promptkit $RELEASE_VERSION" "$RELEASE_NOTES_FILE"
-grep -F "Validated commit: $RELEASE_COMMIT" "$RELEASE_NOTES_FILE"
-grep -F 'Compatibility:' "$RELEASE_NOTES_FILE"
-grep -F 'Public API changes:' "$RELEASE_NOTES_FILE"
-grep -F 'Consumer action:' "$RELEASE_NOTES_FILE"
-```
-
-Inspect the complete message and confirm that it accurately records the
-compatibility impact, public API changes, and required consumer action.
-
-## Create And Inspect The Tag
-
-Run the candidate guard again immediately before tag creation. This ensures
-that validation or release-note preparation did not change the checkout and
-that the commit is still published and untagged:
-
-```sh
-check_release_candidate
-```
-
-Create the annotated tag from the prepared release note and bind it explicitly
-to the validated commit:
-
-```sh
-git tag --annotate "$RELEASE_VERSION" \
-    --file "$RELEASE_NOTES_FILE" \
-    "$RELEASE_COMMIT"
-```
-
-Inspect both the tag message and its source commit before publication:
-
-```sh
-test "$(git cat-file -t "refs/tags/$RELEASE_VERSION")" = tag
-git show --no-patch --decorate "refs/tags/$RELEASE_VERSION"
-test "$(
-    git rev-parse --verify "refs/tags/$RELEASE_VERSION^{commit}"
-)" = "$RELEASE_COMMIT"
-```
-
-If inspection finds an error, delete the unpublished local tag, correct the
-release note or candidate, and repeat the guards. Never move or recreate a tag
-that has been published.
-
-## Publish The Selected Tag
-
-Push only the selected tag ref. Do not use `git push --tags`:
-
-```sh
-git push origin \
-    "refs/tags/$RELEASE_VERSION:refs/tags/$RELEASE_VERSION"
-```
-
-## Verify Publication
-
-Compare the remote annotated-tag object with the local object, then compare the
-remote peeled source commit with the validated commit:
-
-```sh
-remote_tag=$(
-    git ls-remote --tags origin "refs/tags/$RELEASE_VERSION" |
-        awk 'NR == 1 { print $1 }'
-)
-remote_commit=$(
-    git ls-remote --tags origin "refs/tags/$RELEASE_VERSION^{}" |
-        awk 'NR == 1 { print $1 }'
-)
-test -n "$remote_tag"
-test "$remote_tag" = \
-    "$(git rev-parse --verify "refs/tags/$RELEASE_VERSION")"
-test "$remote_commit" = "$RELEASE_COMMIT"
-```
-
-Finally, resolve the version as an ordinary Go module in a temporary module
-outside this repository and without a workspace or replacement:
-
-```sh
-resolution_dir=$(mktemp -d)
-(
-    trap 'rm -rf "$resolution_dir"' 0 1 2 15
-    cd "$resolution_dir"
-    GOWORK=off go mod init example.com/promptkit-release-check
-    GOWORK=off go mod download \
-        "gitea.maximumdirect.net/eric/promptkit@$RELEASE_VERSION"
-    resolved_version=$(
-        GOWORK=off go list -m -f '{{.Version}}' \
-            "gitea.maximumdirect.net/eric/promptkit@$RELEASE_VERSION"
-    )
-    test "$resolved_version" = "$RELEASE_VERSION"
-)
-```
-
-Promptkit must publish and verify the required version before Scriptorium or
-another consumer publishes a release that depends on it. This ordering does
-not replace the consumer project's own release procedure. Released consumers
-must select the published Promptkit tag through ordinary module resolution,
-without a workspace, replacement, vendored Promptkit source, or unpublished
-revision.
 
 ## Policy Changes
 
-Document and approve a durable policy change before introducing hosted
+Update the policy and this procedure together when changing release governance.
-automation, binary artifacts, or different release governance. Update this
+Consumers own their dependency-availability and release checks; their readiness
-procedure in the same change so maintainers do not rely on hidden release
+does not delay completion of a Promptkit release.
-requirements.

docs/releases/v0.10.0.md

@@ -0,0 +1,44 @@
+# Promptkit v0.10.0
+
+This supplemental guide summarizes the `v0.10.0` changes and migration needs.
+The annotated tag remains the authoritative release record.
+
+## Transition Summary
+
+`v0.10.0` adopts independently versioned OpenRouter and Rakestrawhome catalog
+modules whose external profile data can change between catalog releases. It
+also introduces the maintained literal names
+`promptkit-openrouter-standard` and `promptkit-openrouter-auto`.
+
+The public package now delegates assembly to an internal engine while retaining
+the existing Go API. This refactor requires no consumer code changes.
+
+This release selects OpenRouter `v2.20260922.1` and Rakestrawhome
+`v2.20260922.0`; their exact inventories are linked from the
+[format reference](../formats.md#external-profile-inventories).
+
+The maintained names provide continuity only within the catalog versions
+selected and tested by a Promptkit release and when consumers do not override
+the name or its base. Their current selections may change. External catalog
+profile IDs can be removed, renamed, or modified, so an application that
+selects one directly may need to update its configuration when upgrading.
+
+## Consumer Migration
+
+Before upgrading, review every profile ID used by prompts, requests, and
+application-owned inheritance. Choose a maintained name, accept an external
+catalog profile's volatility, or define a complete application-owned profile
+when the application must control its target and settings. An alias that
+inherits from an external base is not a frozen copy.
+
+In particular, the selected OpenRouter catalog removes `deepseek-3-2`,
+`deepseek-4-flash`, and `deepseek-4-pro`. Consumers using those IDs must choose
+an available profile or supply their own complete definition; the new
+`deepseek-flash-latest` and `deepseek-pro-latest` profiles are not automatic
+aliases and need not preserve the previous settings.
+
+See the [release policy](../release.md#catalog-profile-compatibility) for the
+compatibility boundary, the [format reference](../formats.md#built-in-profile-catalog)
+for current Promptkit-owned definitions and selected external inventories, and
+the [consumer guide](../consumers/pkg-promptkit.md#choose-a-profile-ownership-model)
+for configuration choices.

docs/releases/v0.11.0.md

@@ -0,0 +1,76 @@
+# Promptkit v0.11.0
+
+This release adds reliable, observable model execution and changes Promptkit's
+license to LGPL-3.0-or-later.
+
+## Summary
+
+- Built-in model calls can recover from selected transient HTTP failures.
+- Optional execution observations report attempts, timing, outcomes, and known
+  token usage, including work completed before a run fails.
+- Public errors distinguish transport failures, malformed responses, and retry
+  exhaustion without parsing diagnostic text.
+
+## Compatibility
+
+The built-in client now defaults to two additional transient attempts. This can
+increase latency, request counts, and cost; retries do not guarantee exactly-once
+generation. Injected `LLMClient` implementations retain their own retry policy
+and are not automatically replayed. Existing generation timeouts apply per
+attempt, while the caller context governs the complete operation. See the
+[provider integration contract](../integrations/openai-compatible-chat.md).
+
+Public retry and observation controls are additive. Existing successful-result
+usage semantics remain unchanged, and operational failures still return no
+partial `RunResult`. Observations distinguish unknown usage from reported zero;
+they are not a complete billing ledger. See the
+[consumer guide](../consumers/pkg-promptkit.md) and
+[observation GoDoc](../../observations.go).
+
+Promptkit is now licensed under LGPL-3.0-or-later. Review the
+[license summary](../../README.md#license), [LGPL text](../../LICENSE), and
+[accompanying GPL text](../../COPYING) before adoption or distribution.
+Previously published releases and dependency licenses are unchanged.
+
+## Upgrade
+
+Update the dependency:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.11.0
+```
+
+To retain single-attempt behavior for built-in generation, set
+`Config.TransientRetries` to a pointer to zero. `RunRequest.TransientRetries`
+can override the engine choice per request. Review any existing application or
+HTTP-transport retry loops to avoid unintentionally multiplying attempts.
+Transient retries and structural repairs have separate budgets; provide a
+caller deadline when a total execution time limit is needed.
+
+Observation is opt-in through `RunRequest.Observer`, including for prepared
+execution. Callbacks are synchronous; shared observers must support concurrent
+runs. Injected clients can supply `GenerateResponse.UsagePresence` to explicitly
+report known zero counts. No observer or usage-presence changes are required
+for existing consumers. See the [consumer guide](../consumers/pkg-promptkit.md)
+for adoption and the [offline example](../../examples/go-library/observe/main.go)
+for a structured proposal-and-assessment workflow.
+
+## Changes
+
+- Centralized built-in retry coordination with cancellable backoff and provider
+  retry timing. Retries and structural repair share existing backend capacity;
+  backoff releases active permits without releasing logical run admission.
+- Added `ErrTransport`, `ErrMalformedResponse`, `ErrRetriesExhausted`, and
+  `RetryExhaustedError`, retaining applicable underlying error identities.
+- Added request-scoped lifecycle observations and presence-aware token accounting
+  for ordinary and prepared execution, without recording prompt/response bodies.
+- Isolated delegated operations' observation state and made injected-client
+  attempt classifications consistent with terminal classifications.
+- Added an offline observation example and regression coverage for retry,
+  repair, capacity, error, and observation interactions.
+- Adopted LGPL-3.0-or-later and streamlined the documented local release process.
+
+Backend catalog selections are unchanged from v0.10.0: OpenRouter
+`v2.20260922.1` and Rakestrawhome `v2.20260922.0`. See the
+[catalog compatibility policy](../policy/release.md#catalog-profile-compatibility)
+and [profile inventories](../formats.md#external-profile-inventories).

docs/releases/v0.2.0.md

@@ -90,7 +90,7 @@ engine, err := promptkit.NewEngine(
 ```
 
 See the
-[custom-backend consumer guide](../consumers/pkg-promptkit.md#register-a-custom-backend)
+[local-endpoint consumer guide](../consumers/pkg-promptkit.md#configure-a-local-openai-compatible-endpoint)
 for task-oriented usage. The
 [`Backend` and `WithBackend` GoDoc](../../backends.go) owns exact registration,
 validation, copying, defaulting, and uniqueness semantics. The

docs/releases/v0.3.0.md

@@ -0,0 +1,74 @@
+# Promptkit v0.3.0
+
+This supplemental changelog summarizes the consumer-facing changes from
+`v0.2.0` to `v0.3.0`. The annotated `v0.3.0` tag is the authoritative release
+record. Exact current contracts belong to the linked GoDoc and durable
+documentation.
+
+## Summary
+
+`v0.3.0` adds a concise way to register the common local OpenAI-compatible
+backend configuration:
+
+- `BackendLocal` provides the conventional, non-reserved backend ID `"local"`;
+  and
+- `LocalBackend` constructs an ordinary `Backend` from an endpoint and
+  concurrency limit.
+
+The helper is explicit and additive. It does not pre-register a backend, read
+environment variables, select a model, or replace the complete `Backend`
+configuration interface.
+
+## Compatibility
+
+Existing `v0.2.0` consumers require no migration. Endpoint-only profiles,
+complete custom `Backend` values, the built-in OpenRouter backend, and existing
+registrations using the literal ID `"local"` continue to work unchanged.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.3.0
+go mod tidy
+```
+
+Run the consuming project's ordinary tests and race-enabled tests after the
+upgrade.
+
+## Configure A Local Backend
+
+Register the convenience value through the existing `WithBackend` option and
+select it from one or more profiles:
+
+```go
+engine, err := promptkit.NewEngine(
+	promptkit.Config{PromptDir: "prompts"},
+	promptkit.WithBackend(
+		promptkit.LocalBackend("http://localhost:8000/v1", 2),
+	),
+	promptkit.WithProfiles(promptkit.Profile{
+		ID:        "local-summary",
+		BackendID: promptkit.BackendLocal,
+		Model:     "example-model",
+	}),
+)
+```
+
+Use an endpoint-only profile when shared backend identity and capacity policy
+are unnecessary. Continue to use a complete keyed `Backend` value for custom
+IDs, authentication, extra request parameters, explicit queue capacity, or
+multiple local endpoints.
+
+See the
+[local-endpoint consumer guide](../consumers/pkg-promptkit.md#configure-a-local-openai-compatible-endpoint)
+for task-oriented configuration choices. The
+[`BackendLocal`, `LocalBackend`, and `WithBackend` GoDoc](../../backends.go)
+owns their exact construction, registration, validation, and concurrency
+semantics.
+
+## Consumer Action
+
+None. Adopt the convenience constructor when it simplifies local endpoint
+configuration.

docs/releases/v0.4.0.md

@@ -0,0 +1,189 @@
+# Promptkit v0.4.0
+
+This supplemental changelog and adoption guide summarizes the consumer-facing
+changes from `v0.3.0` to `v0.4.0`. The annotated `v0.4.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.4.0` adds four complementary capabilities:
+
+- opaque prepared-execution handles for preparing once, inspecting safe
+  details, and executing the same frozen snapshot;
+- exact prompt-definition inspection without profile resolution or execution;
+- exact profile inspection without selecting a prompt or checking credential
+  availability; and
+- structured backend identity on engine admission-capacity rejection.
+
+These APIs let consumers perform more precise preflight work and retain useful
+operational context without reproducing Promptkit's internal resolution logic.
+
+## Compatibility
+
+The release is additive for `v0.3.0` consumers. Existing uses of `Prepare`,
+`Run`, backend registration, endpoint-only profiles, local-backend helpers,
+runtime overrides, public JSON values, and error sentinels continue to work
+without migration.
+
+Capacity rejection now returns a structured error while continuing to match
+`ErrCapacityExceeded` through `errors.Is`. Error-string wording and direct
+sentinel equality were not public contracts.
+
+The new inspection values, capacity error, and prepared-execution handle do not
+have stable JSON representations. `PreparedExecution.Details` returns the
+existing stable `PreparedRun` value.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.4.0
+go mod tidy
+```
+
+Run the consuming project's ordinary and race-enabled tests after upgrading.
+No source migration is required.
+
+## Prepare Once And Execute The Same Snapshot
+
+Consumers that need to persist preparation details before generation can now
+prepare an opaque, engine-bound execution:
+
+```go
+prepared, err := engine.PrepareExecution(ctx, request)
+if err != nil {
+	// Handle preparation failure.
+}
+defer prepared.Discard()
+
+details := prepared.Details()
+// Persist a consumer-selected, appropriately protected preparation record.
+
+result, err := engine.RunPrepared(ctx, prepared)
+```
+
+Preparation freezes the selected sources, rendered messages, effective
+settings, input content, structured-output metadata, and validation resources
+needed by execution. `Details` returns a fresh, caller-owned,
+credential-redacted `PreparedRun`.
+
+A handle belongs to its creating engine and permits one execution attempt.
+`RunPrepared` consumes that attempt on success and on operational failure.
+`Discard` is idempotent and releases an unclaimed handle's execution-only
+state. Consumers should discard handles they will not execute, particularly
+when a direct request API key may be retained privately until claim or
+discard.
+
+Prepared execution does not reserve backend admission during preparation.
+Credential availability and backend admission are checked when execution
+begins. The execution context is independent of the preparation context.
+
+See the
+[prepared-execution consumer guide](../consumers/pkg-promptkit.md#prepare-now-and-execute-the-same-snapshot-later),
+the [`PreparedExecution` GoDoc](../../prepared_execution.go), and the
+[`Engine.PrepareExecution` and `Engine.RunPrepared` GoDoc](../../engine.go)
+for the exact lifecycle, ownership, cancellation, capacity, timing, and
+failure contracts.
+
+## Inspect A Prompt
+
+`Engine.InspectPrompt` resolves one prompt ID and optional version through the
+engine's configured prompt source:
+
+```go
+inspection, err := engine.InspectPrompt(ctx, "report.summary", "")
+```
+
+The result includes prompt identity, the opaque prompt hash, declared default
+profile ID, declared input metadata, and normalized output contract. It
+structurally loads the selected definition and referenced message content but
+does not resolve a profile, load schemas or artifacts, render templates,
+reserve capacity, or contact a model.
+
+Use inspection for exact configuration checks and metadata discovery. Use
+`PrepareExecution` rather than relying on a prior inspection when later
+execution must freeze one exact source state, because filesystem-backed
+inspection is only a point-in-time lookup.
+
+See the
+[prompt-inspection consumer guide](../consumers/pkg-promptkit.md#inspect-a-prompt-before-preparation)
+and [`Engine.InspectPrompt` GoDoc](../../engine.go) for exact selection,
+ownership, and error behavior.
+
+## Inspect A Profile
+
+`Engine.InspectProfile` resolves one explicit profile independently of a
+prompt:
+
+```go
+inspection, err := engine.InspectProfile(ctx, "report-production")
+```
+
+The result includes the resolved effective execution target and whether a
+later request must provide a direct credential. Environment-variable names may
+be reported, but inspection does not read credential values or require the
+named variable to be populated.
+
+Inspection applies the engine's profile source precedence and resolves any
+selected backend. It does not load a prompt, render content, reserve capacity,
+or contact a model.
+
+See the
+[profile-inspection consumer guide](../consumers/pkg-promptkit.md#inspect-a-profile-before-prompt-work)
+and [`Engine.InspectProfile` GoDoc](../../engine.go) for the exact resolution,
+credential, ownership, and error contracts.
+
+## Identify Capacity-Rejected Backends
+
+Calls rejected at Promptkit's bounded engine admission boundary continue to
+match `ErrCapacityExceeded`. Consumers can additionally obtain the selected
+registered backend ID without parsing diagnostic text:
+
+```go
+result, err := engine.Run(ctx, request)
+if errors.Is(err, promptkit.ErrCapacityExceeded) {
+	var capacityErr *promptkit.CapacityError
+	if errors.As(err, &capacityErr) {
+		// Record capacityErr.BackendID using application-owned diagnostics.
+	}
+
+	// Apply application-owned overload or retry policy.
+}
+```
+
+The structured error applies to `Run` and `RunPrepared` admission rejection.
+It does not represent provider throttling, quota exhaustion, cancellation
+while waiting for generation capacity, or another model-client failure.
+Promptkit does not prescribe retry timing or transport status mapping.
+
+See the
+[error-handling consumer guide](../consumers/pkg-promptkit.md#handle-errors),
+the [`CapacityError` GoDoc](../../capacity_error.go), and the
+[`ErrCapacityExceeded` GoDoc](../../engine.go) for the canonical contracts.
+
+## Public API Additions
+
+The release adds:
+
+- `Engine.PrepareExecution`;
+- `Engine.RunPrepared`;
+- `PreparedExecution`, including `Details`, `Discard`, `String`, and
+  `GoString`;
+- `Engine.InspectPrompt`;
+- `PromptInspection`;
+- `PromptInputDefinition`;
+- `Engine.InspectProfile`;
+- `ProfileInspection`; and
+- `CapacityError`.
+
+No public API was removed.
+
+## Consumer Action
+
+None. Existing `v0.3.0` workflows may upgrade without adopting the new APIs.
+
+Consumers that adopt prepared execution should discard unused handles.
+Consumers that need backend-specific capacity diagnostics may add an
+`errors.As` check while retaining their existing `errors.Is` classification.

docs/releases/v0.5.0.md

@@ -0,0 +1,125 @@
+# Promptkit v0.5.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.4.0` to `v0.5.0`. The annotated `v0.5.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.5.0` makes provider requests less prescriptive and adds an application
+fallback layer for profile definitions:
+
+- unset optional provider controls are omitted from OpenAI-compatible request
+  bodies instead of being populated with framework values; and
+- `WithFallbackProfileFS` lets an application package profile defaults that
+  operators can override through the existing ordinary profile sources.
+
+These changes let compatible providers apply their own model defaults while
+giving applications stable embedded profile IDs without weakening operator
+configuration precedence.
+
+## Compatibility
+
+The release adds one public function and removes no public declaration.
+Existing source code should continue to compile.
+
+There is one intentional behavior change: when no profile or runtime override
+selects `top_p`, Promptkit no longer sends the former framework value of `1`.
+It omits `top_p` and lets the provider choose its behavior. Unset
+`temperature` and `max_tokens` are likewise omitted. Explicit nonzero profile
+values and runtime values—including explicit runtime zero values—retain their
+precedence and wire effect.
+
+Consumers that relied on Promptkit always sending `top_p: 1` should add that
+value to the relevant profile or runtime override before upgrading. Consumers
+that did not rely on the implicit sampling value require no migration.
+
+Application fallback profiles are opt-in. Engines that do not call
+`WithFallbackProfileFS` retain the previous profile-source behavior.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.5.0
+go mod tidy
+```
+
+Run the consuming project's ordinary and race-enabled tests after upgrading.
+If request payloads or model behavior are asserted in fixtures, review them for
+the optional-parameter omission described below.
+
+## Omitted Optional Provider Controls
+
+The built-in OpenAI-compatible client now includes `temperature`,
+`max_tokens`, and `top_p` only when a profile or runtime override selects the
+value. An explicit runtime zero remains present because runtime override
+pointers distinguish zero from an unspecified value.
+
+Promptkit's positive generation deadline remains a framework concern and is
+not a provider request-body default. Required request fields, session IDs,
+structured output, reasoning selection, credentials, and explicit extra
+parameters retain their existing behavior.
+
+See the [framework default and precedence reference](../formats.md#defaults-and-overrides),
+the [`ExecutionTargetOverride` GoDoc](../../types.go), and the
+[OpenAI-compatible request-body contract](../integrations/openai-compatible-chat.md#request-body)
+for current details.
+
+## Embedded Application Fallback Profiles
+
+Applications can package ordinary profile YAML in an `fs.FS` and register it
+as a fallback source:
+
+```go
+//go:embed profiles/*.yaml
+var applicationProfiles embed.FS
+
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir:  "prompts",
+	ProfileDir: operatorProfileDir,
+},
+	promptkit.WithFallbackProfileFS(applicationProfiles, "profiles"),
+)
+```
+
+Leave `operatorProfileDir` empty when no operator source is configured. A
+configured ordinary source is authoritative: a matching definition overrides
+the application fallback, while a read or validation failure remains an error
+instead of silently reaching a lower layer.
+
+Profile definitions resolve in this order:
+
+1. in-memory profiles supplied with `WithProfiles`;
+2. the ordinary configured source selected by `WithProfileFile`,
+   `WithProfileFS`, or `Config.ProfileDir`;
+3. the application source supplied with `WithFallbackProfileFS`; and
+4. Promptkit's embedded built-in profiles.
+
+Only an absent profile ID falls through. Sources provide complete profiles and
+do not merge fields. Loading remains lazy, and the new source uses the existing
+strict profile YAML and credential rules.
+
+See the
+[embedded-default consumer guidance](../consumers/pkg-promptkit.md#supply-embedded-application-defaults),
+the [`WithFallbackProfileFS` GoDoc](../../engine.go), and the
+[profile source reference](../formats.md#source-and-profile-precedence) for
+current details.
+
+## Public API Changes
+
+The release adds:
+
+- `WithFallbackProfileFS`.
+
+No public declaration was removed or changed.
+
+## Consumer Action
+
+- Review any workflow that depended on Promptkit's implicit `top_p: 1` and
+  configure the value explicitly when required.
+- Optionally adopt `WithFallbackProfileFS` when an application should package
+  overridable profile defaults.
+- Run consumer tests after updating the module dependency.

docs/releases/v0.6.0.md

@@ -0,0 +1,131 @@
+# Promptkit v0.6.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.5.0` to `v0.6.0`. The annotated `v0.6.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.6.0` is a broad correctness, safety, efficiency, and maintainability
+release. It does not add or remove public declarations. The release:
+
+- centralizes shared execution-setting, output-contract, endpoint, and
+  JSON-compatible-value rules;
+- unifies prompt repository behavior and avoids unnecessary prompt and profile
+  decoding;
+- bounds consumer-controlled JSON trees and successful provider responses;
+- hardens prompt content paths, artifact files, provider URLs, JSON framing,
+  and error propagation;
+- reuses compiled schema plans and rendered artifact text within an operation;
+  and
+- improves cancellation behavior, prepared-value ownership, deterministic
+  transport testing, and maintainer validation.
+
+## Compatibility
+
+No public declaration was added, removed, or changed. Ordinary valid `v0.5.0`
+configurations and requests should continue to compile and behave as before.
+
+The release intentionally rejects or reports several inputs that were
+previously accepted, altered, or misclassified:
+
+- execution settings must be finite, within their documented ranges, and safe
+  to convert to Go durations;
+- output formats, validation modes, repair counts, and JSON Schema dependencies
+  are validated consistently;
+- file-backed prompt and profile identity comes from normalized YAML metadata,
+  not filenames;
+- prompt `content_file` values must be exact relative paths contained by their
+  configured source root;
+- built-in file artifacts must resolve to regular files;
+- selected provider endpoints must be absolute HTTP or HTTPS URLs without user
+  information, query strings, or fragments;
+- JSON documents and successful provider responses must contain exactly one
+  value, and successful provider bodies are limited to 16 MiB; and
+- excessively deep or expansive JSON-compatible values fail with ordinary
+  validation errors.
+
+These are compatibility corrections and safety boundaries rather than new
+consumer configuration requirements. Consumers relying on an invalid or
+ambiguous input should correct that input before upgrading.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.6.0
+go mod tidy
+```
+
+Run the consuming project's ordinary and race-enabled tests after upgrading.
+Applications with custom prompt/profile sources, local provider endpoints,
+unusual artifact paths, or assertions over provider error identities should
+pay particular attention to the compatibility notes below.
+
+## Source Loading And Identity
+
+Prompt definitions now share one source-neutral selection and normalization
+flow across operating-system and `fs.FS` sources. YAML `id` and `version`
+metadata are authoritative; filenames do not create a second identity system.
+Only selected content bodies are loaded, malformed unrelated definitions do
+not shadow valid exact matches, and per-file read failures are reported as
+prompt-load failures rather than false absence.
+
+File-backed profiles likewise use normalized YAML IDs, reuse their metadata
+read for selected strict decoding, and avoid fully decoding unrelated files.
+Selected malformed definitions remain authoritative and do not silently fall
+through to a lower-precedence source.
+
+Prompt `content_file` paths are opened exactly as declared after a separate
+blank check. They must remain relative to and contained by the configured
+prompt source root, including across operating-system symlinks.
+
+See the [framework source and identity reference](../formats.md) and
+[internal source overview](../internal/sources.md) for the current contracts.
+
+## Validation, Cancellation, And Efficiency
+
+JSON Schema documents preserve exact JSON-number representations. Schema
+resource URLs safely escape legal filesystem names, and each operation loads
+and compiles its schema graph once. `Run` and prepared execution reuse that
+operation-local plan; Promptkit does not introduce a cross-operation cache.
+
+Artifact reading, rendering, schema loading, compilation, and validation now
+check cancellation at the synchronous boundaries Promptkit controls. Rendering
+memoizes each artifact's text within one render operation, while plain JSON
+validation avoids materializing an unnecessary generic tree.
+
+The shared JSON-compatible-value owner now limits nesting and produced work so
+unsafe consumer-controlled structures return errors instead of risking
+unbounded recursion or allocation. See the
+[architecture policy](../policy/architecture.md) for invariant ownership and
+the [format reference](../formats.md) for validation behavior.
+
+## Provider Transport Hardening
+
+OpenAI-compatible endpoints are parsed and composed structurally, including
+nested base paths. Underlying transport cancellation and deadline errors remain
+discoverable with `errors.Is` through Promptkit's generation error category.
+
+Successful provider bodies are read with a fixed 16 MiB bound and must contain
+exactly one JSON response object followed only by whitespace. Oversized,
+truncated, malformed, or multiply framed responses fail without returning a
+partial result. See the
+[OpenAI-compatible integration contract](../integrations/openai-compatible-chat.md)
+for the canonical request, endpoint, error, and response behavior.
+
+## Public API Changes
+
+None.
+
+## Consumer Action
+
+- Correct any configuration or request that depends on the formerly permissive
+  cases described under Compatibility.
+- Confirm custom local provider endpoints are absolute HTTP or HTTPS base URLs
+  without credentials, queries, or fragments.
+- Confirm prompt content paths remain within their configured source root and
+  file artifacts resolve to regular files.
+- Run ordinary and race-enabled consumer tests after updating the dependency.

docs/releases/v0.7.0.md

@@ -0,0 +1,155 @@
+# Promptkit v0.7.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.6.0` to `v0.7.0`. The annotated `v0.7.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.7.0` expands provider integration and profile composition while making
+credential and generation-failure handling more flexible:
+
+- Promptkit now includes the `rakestrawhome` backend and its Gemma profile;
+- built-in generation failures expose bounded structured provider details;
+- an unavailable optional API-key environment source no longer prevents a
+  request from reaching an upstream that permits unauthenticated access; and
+- profiles can inherit from and selectively refine another profile.
+
+## Compatibility
+
+This release adds public declarations and fields but removes none. Existing
+keyed configuration literals and ordinary `errors.Is` handling continue to
+work.
+
+Adding `BaseProfileID` to `Profile` and `OpenAICompatibleProfileConfig` changes
+their struct shape. Consumers using positional composite literals for either
+type must convert them to keyed literals. Existing keyed literals require no
+change.
+
+The `rakestrawhome` backend ID is now built in and reserved. A consumer that
+previously registered that exact ID with `WithBackend` must remove its manual
+registration before upgrading. Other custom backend registrations are
+unchanged.
+
+When an optional backend, profile, or request `APIKeyEnv` is unset, empty, or
+whitespace-only, the built-in client now omits `Authorization` and sends the
+request. Previously this condition could fail before transport. Set
+`Profile.APIKeyRequired` when missing credentials must remain a local
+preflight error.
+
+Provider non-success responses continue to match `ErrLLMGenerate`. Their
+rendered wording is not a compatibility contract; consumers can now use
+`errors.As` with `*GenerationError` when structured status information is
+needed.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.7.0
+go mod tidy
+```
+
+Remove any manual `rakestrawhome` backend registration, convert positional
+profile literals to keyed literals, and run the consuming project's ordinary
+and race-enabled tests.
+
+## Rakestrawhome Built-In Backend And Profile
+
+Every engine now includes the reserved `rakestrawhome` backend, identified by
+`BackendRakestrawHome`. The built-in `rakestrawhome-gemma-4-31b` profile
+selects that backend. Consumers can use the maintained endpoint, credential,
+capacity, and model defaults without registering either definition themselves.
+
+See the [built-in backend and profile catalogs](../formats.md#built-in-backends)
+and the [consumer adoption example](../consumers/pkg-promptkit.md#use-the-rakestrawhome-built-in-profile)
+for the current contracts.
+
+## Structured Generation Errors
+
+Non-2xx responses from the built-in OpenAI-compatible client now return an
+immutable `*GenerationError`. Consumers can inspect the HTTP status and any
+safely extracted provider code, type, or message while retaining the ordinary
+generation-error category:
+
+```go
+var generationErr *promptkit.GenerationError
+if errors.As(err, &generationErr) {
+	status := generationErr.StatusCode()
+	_ = status
+}
+```
+
+Provider fields are bounded and normalized but remain untrusted and may
+contain sensitive request or schema details. Default and Go-syntax formatting
+omit those fields. Applications must apply their own disclosure policy before
+logging or presenting accessor values.
+
+See the [`GenerationError` GoDoc](../../generation_error.go), the
+[consumer error-handling guide](../consumers/pkg-promptkit.md#handle-errors),
+and the [OpenAI-compatible response contract](../integrations/openai-compatible-chat.md#response-handling).
+
+## Optional Credential Sources
+
+`APIKeyEnv` names an optional environment lookup source unless the selected
+profile explicitly sets `APIKeyRequired`. When neither a direct request key nor
+a usable environment value exists, the built-in client omits the bearer header
+and handles the upstream response normally. This supports local and other
+OpenAI-compatible providers that permit unauthenticated requests without
+hiding an authentication error returned by a provider that requires one.
+
+The [credential format reference](../formats.md#credentials), the
+[`Backend` GoDoc](../../backends.go), the
+[`ExecutionTargetOverride` GoDoc](../../types.go), and the
+[authentication integration contract](../integrations/openai-compatible-chat.md#authentication)
+define the current precedence and availability rules.
+
+## Profile Inheritance
+
+YAML profiles can name one parent with `base_profile`; in-memory profiles use
+`Profile.BaseProfileID`, and `OpenAICompatibleProfileConfig` forwards the same
+field. A profile can act as an application-owned alias of a built-in or refine
+selected inherited settings:
+
+```go
+promptkit.WithProfiles(promptkit.Profile{
+	ID:              "weather-light",
+	BaseProfileID:   "deepseek-4-flash",
+	ReasoningEffort: "high",
+})
+```
+
+Base lookup observes the existing source precedence. Chains are linear,
+cycle-safe, and resolved afresh for ordinary operations. Prepared execution
+freezes the fully resolved target. The selected leaf ID remains public while
+effective execution settings reflect the resolved chain.
+
+See the [profile inheritance format reference](../formats.md#profile-inheritance),
+the [consumer alias example](../consumers/pkg-promptkit.md#alias-a-built-in-profile),
+and the [`Profile` GoDoc](../../types.go) for exact merge and validation
+behavior.
+
+## Public API Changes
+
+The release adds:
+
+- `BackendRakestrawHome`;
+- `GenerationError`, including `StatusCode`, `ProviderCode`, `ProviderType`,
+  `ProviderMessage`, `Error`, `GoString`, and `Unwrap`;
+- `Profile.BaseProfileID`; and
+- `OpenAICompatibleProfileConfig.BaseProfileID`.
+
+No public declaration was removed.
+
+## Consumer Action
+
+- Remove a manual backend registration whose ID is exactly `rakestrawhome`.
+- Convert positional `Profile` or `OpenAICompatibleProfileConfig` literals to
+  keyed literals.
+- Set `Profile.APIKeyRequired` where a missing credential must fail locally
+  instead of reaching the provider unauthenticated.
+- Treat `GenerationError` provider fields as untrusted and potentially
+  sensitive when adopting the new accessors.
+- Run consumer ordinary and race-enabled tests after updating the module.

docs/releases/v0.8.0.md

@@ -0,0 +1,109 @@
+# Promptkit v0.8.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.7.0` to `v0.8.0`. The annotated `v0.8.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.8.0` activates Promptkit's bounded output-repair workflow:
+
+- failed nonempty-text, JSON, and JSON Schema validation can make a limited
+  number of corrective model calls;
+- corrective calls preserve the original rendered conversation, effective
+  target, session, structured-output contract, and backend capacity policy;
+- results report cumulative usage and the number of corrective calls actually
+  made; and
+- explicitly empty OpenAI-compatible response content now reaches output
+  validation instead of being classified as a malformed provider envelope.
+
+## Compatibility
+
+This release adds no public declarations or fields and removes none. Existing
+source code remains source-compatible.
+
+The behavior of the existing `OutputContract.RepairAttempts` field and prompt
+YAML `repair_attempts` field has changed. A positive value now authorizes real
+additional model calls after eligible validation failures; earlier releases
+accepted the field but the public engine remained single-pass. Consumers that
+set a positive value should expect additional latency, token usage, and
+provider cost when repair is needed.
+
+Repair budgets must now be between zero and three. A positive budget requires
+`basic`, `json`, or `json_schema` validation. Values above three and a positive
+budget paired with `none` are invalid contracts rather than ignored settings.
+
+An explicitly present empty or whitespace-only string returned by the built-in
+OpenAI-compatible client is now a completed generation candidate. `none`
+validation permits it, while `basic`, `json`, and `json_schema` classify it
+under their ordinary validation rules and may repair it when configured.
+Missing, `null`, or non-string content remains a malformed provider response.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.8.0
+go mod tidy
+```
+
+Review every prompt definition and request override that sets a positive repair
+budget. Use zero or omit the field to retain single-pass execution. Ensure each
+positive budget is no greater than three and uses an eligible validation mode,
+then run the consuming project's ordinary and race-enabled tests.
+
+## Bounded Output Repair
+
+`repair_attempts` counts corrective calls in addition to the initial model
+call. Promptkit validates each completed candidate, stops at the first valid
+one, and never exceeds the configured bound. If every candidate remains
+invalid, the run completes successfully with the final candidate and its
+failed validation result rather than returning an operational error.
+
+Each correction starts from the original rendered messages and includes only
+the latest invalid candidate and latest validation diagnostics. JSON Schema
+mode retains the provider-native structured-output request as its first line of
+defense. Promptkit performs only deterministic structural validation; a valid
+response is not necessarily factual or correct for an application's domain.
+
+Usage in the final result is cumulative across the initial response and every
+completed corrective response. `ValidationResult.RepairAttempts` reports the
+number of corrective calls actually made. Corrective generation failures use
+the same public generation-error categories and structured provider details as
+an initial generation failure.
+
+See the [output-contract format reference](../formats.md#output-contract), the
+[consumer repair example](../consumers/pkg-promptkit.md#repair-a-structured-result),
+and the [`OutputContract` and `ValidationResult` GoDoc](../../types.go) for the
+current contracts.
+
+## Explicit Empty Content
+
+The built-in OpenAI-compatible client now distinguishes an explicitly present
+empty string from a missing or malformed `content` field. This aligns built-in
+and injected clients by letting the selected output contract decide whether an
+empty candidate is acceptable, invalid, or eligible for repair.
+
+See the
+[OpenAI-compatible response contract](../integrations/openai-compatible-chat.md#response-handling)
+for the exact envelope behavior.
+
+## Public API Changes
+
+None. This release activates and tightens the documented behavior of existing
+fields.
+
+## Consumer Action
+
+- Remove or set `repair_attempts` to zero where execution must remain
+  single-pass.
+- Keep every positive repair budget at three or fewer and pair it with
+  `basic`, `json`, or `json_schema` validation.
+- Account for additional latency, usage, and provider cost when enabling
+  repair.
+- Continue checking the returned validation status because bounded repair can
+  exhaust without producing a valid candidate.
+- Review workflows that previously treated explicit empty provider content as
+  a generation error.

docs/releases/v0.9.0.md

@@ -0,0 +1,125 @@
+# Promptkit v0.9.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.8.0` to `v0.9.0`. The annotated `v0.9.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.9.0` adds stateless request-message composition and separates maintained
+provider data from Promptkit's core implementation:
+
+- callers can append already-rendered messages to a configured prompt for
+  application-owned conversations and semantic correction workflows;
+- the public package now publishes constants for the four supported text-chat
+  roles, and prompt definitions use the same normalized role vocabulary; and
+- the OpenRouter and Rakestrawhome backend/profile catalogs now come from two
+  independently versioned Go module dependencies.
+
+## Compatibility
+
+This release adds one field and four constants to the public API and removes no
+public declarations. Existing keyed `RunRequest` literals that omit
+`AppendedMessages` retain their behavior. Adding the field changes the struct
+shape, so consumers using positional `RunRequest` literals must convert them
+to keyed literals.
+
+Message roles in prompt definitions are now trimmed, lowercased, and required
+to be `developer`, `system`, `user`, or `assistant`. Definitions using another
+role that earlier releases accepted as an arbitrary nonblank string now fail
+prompt loading. In particular, the text-only message contract does not support
+`tool` or the deprecated `function` role. Otherwise valid roles with different
+case or surrounding whitespace are normalized rather than rejected.
+
+The catalog extraction preserves Promptkit's public API, built-in backend and
+profile IDs, configuration, precedence, credential handling, and capacity
+behavior. Consumers do not import or register either catalog themselves, and
+no configuration migration is required. Promptkit now selects two
+independently versioned data dependencies and guarantees only the catalog
+versions selected and tested by this Promptkit release.
+
+## Upgrade
+
+Update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.9.0
+go mod tidy
+```
+
+Convert any positional `RunRequest` literals to keyed literals. Review prompt
+definitions for unsupported roles, then run the consuming project's ordinary
+and race-enabled tests.
+
+## Appended Request Messages
+
+`RunRequest.AppendedMessages` accepts caller-owned `RenderedMessage` values
+that Promptkit validates, copies, and appends after every rendered
+prompt-definition message in caller order. Promptkit does not template this
+content, retain conversation state between calls, impose a retry policy, or
+apply a message-count, byte-size, token, or context-window limit. Upstream
+rejections continue through the ordinary generation-error boundary.
+
+This primitive supports application-owned conversation continuations and
+domain-aware correction loops while preserving Promptkit's existing
+preparation, hashing, prepared-execution, structural repair, backend-capacity,
+credential, and cancellation behavior. Appended content can include sensitive
+model output or application feedback; prepared values expose the complete
+effective messages by design, while default request formatting reports only
+the appended-message count.
+
+See the
+[consumer appended-message example](../consumers/pkg-promptkit.md#append-already-rendered-messages),
+the [`RunRequest`, `RenderedMessage`, and `CacheControl` GoDoc](../../types.go),
+and the [OpenAI-compatible request contract](../integrations/openai-compatible-chat.md#request-body)
+for current behavior.
+
+## Supported Message Roles
+
+The new `RoleDeveloper`, `RoleSystem`, `RoleUser`, and `RoleAssistant`
+constants identify the complete role vocabulary accepted by Promptkit's
+text-chat message model. The same validation and normalization now apply to
+prompt-definition messages and request-supplied appended messages. Promptkit
+does not translate between roles; provider- or model-specific rejection of an
+otherwise supported role remains an upstream generation error.
+
+See the [message format reference](../formats.md#messages-and-templates) for
+the canonical prompt-definition contract.
+
+## Independently Versioned Backend Catalogs
+
+Promptkit imports immutable catalog data for its maintained OpenRouter and
+Rakestrawhome backends and profiles. Engine construction validates and
+assembles both catalogs behind the existing built-in registry and profile
+precedence rules. Promptkit no longer keeps duplicate embedded profile assets
+or hard-coded definitions for those maintained backends.
+
+The module versions in Promptkit's `go.mod` identify the catalog releases
+tested with this release. The [built-in backend and profile format
+reference](../formats.md#built-in-backends) remains the canonical consumer
+contract, while the [internal source documentation](../internal/sources.md#profiles-and-built-ins)
+describes the dependency boundary.
+
+## Public API Changes
+
+The release adds:
+
+- `RunRequest.AppendedMessages`;
+- `RoleDeveloper`;
+- `RoleSystem`;
+- `RoleUser`; and
+- `RoleAssistant`.
+
+No public declaration was removed.
+
+## Consumer Action
+
+- Convert positional `RunRequest` literals to keyed literals.
+- Replace unsupported prompt-definition roles with an appropriate supported
+  role, or keep richer tool-call protocols in an application-owned client.
+- Treat appended messages and prepared effective messages according to the
+  application's sensitive-data policy.
+- Do not add direct catalog imports or registration calls; existing Promptkit
+  construction and configuration remain correct.
+- Run consumer ordinary and race-enabled tests after updating the module.

docs/roadmap/concurrency.md

@@ -1,236 +0,0 @@
-# Backend-Specific Concurrency Management
-
-**Status:** Complete.
-
-## Purpose
-
-This roadmap defines the scope and target end state for engine-local,
-backend-specific concurrency management. It records the intended capability,
-consumer value, and important policy choices.
-
-This document is planning material, not a description of current behavior.
-Current exported contracts remain owned by Go declarations and GoDoc, backend
-registration guidance by the
-[consumer guide](../consumers/pkg-promptkit.md#register-a-custom-backend), and
-implemented orchestration by the
-[internal runner document](../internal/runner.md).
-
-## Motivation
-
-Different model backends can sustain very different request loads. A local
-network endpoint may need a small concurrency limit, while OpenRouter can
-usually accept substantially more simultaneous work. Requiring every consumer
-to build its own semaphores and queues would duplicate routing knowledge,
-create inconsistent cancellation behavior, and make it easy for one caller to
-bypass the intended backend limit.
-
-Promptkit should own this coordination because it already resolves each run to
-an engine-scoped backend identity and owns every model-generation call made by
-the runner. Consumers should continue submitting ready-to-run requests through
-the synchronous API, including concurrently from multiple goroutines, without
-implementing their own backend scheduler.
-
-The buffered queue is a safety boundary, not an ordinary throughput
-restriction. Its primary purpose is to prevent a bug or unintended submission
-loop from creating an unbounded in-memory backlog.
-
-## Scope
-
-The feature will add optional concurrency policy to registered backends and
-coordinate `Run` calls against independent per-backend capacity pools.
-
-Each policy has two distinct controls:
-
-- an active-generation limit, which protects the backend from too many
-  simultaneous model requests; and
-- a bounded waiting capacity, which protects the process from admitting an
-  unbounded backlog.
-
-Concurrency policy belongs to a backend registration. It is not a profile
-model parameter and cannot be overridden per run. Profiles select the policy
-through their backend ID, while a profile or request endpoint override remains
-in the selected backend's pool.
-
-`Prepare` does not call a model and will remain outside concurrency admission.
-
-## Defaults And Configuration
-
-The built-in OpenRouter backend will use:
-
-- an active-generation limit of 16; and
-- a waiting capacity of 1024.
-
-The waiting default is intentionally generous. Reaching it should indicate
-abnormal submission pressure rather than normal application behavior.
-
-Consumer-registered backends will remain unlimited unless the consumer
-configures an active-generation limit. When a consumer enables a limit and
-does not specify waiting capacity, the waiting capacity will default to 1024.
-Consumers may configure a different bounded capacity, including zero when
-they want no admitted backlog beyond the active-limit-sized run set.
-
-The public representation must distinguish an omitted waiting capacity from
-an explicit zero.
-
-Endpoint-only profiles have no backend registration from which to obtain
-policy and will remain unlimited. A future engine-wide or endpoint-keyed
-policy can be considered separately if consumers demonstrate that need.
-
-Invalid limits or capacities will fail engine construction as invalid
-configuration. Policy values will be copied into engine-owned immutable state
-along with the rest of the backend registration.
-
-## Admission And Execution Behavior
-
-`Run` remains a synchronous, wait-for-result operation. Concurrent callers may
-block inside `Run` while waiting for their selected backend, then receive the
-ordinary result or error from that invocation.
-
-For a configured pool, the active-generation limit plus the waiting capacity
-defines the maximum number of concurrent `Run` invocations that Promptkit will
-accept for that backend. A waiting capacity of zero therefore accepts no more
-runs than the active limit. Admission is immediate: a call either reserves one
-of those bounded slots or receives the capacity error. An accepted run may
-then wait internally for active-generation capacity.
-
-For a limited backend, Promptkit will bound the number of accepted runs before
-expensive artifact loading, prompt rendering, and large defensive copies where
-practical. Lightweight prompt, profile, and backend resolution may occur first
-when it is required to identify the correct capacity pool. This pre-admission
-resolution must not become a second execution-precedence path with behavior
-that can drift from `Prepare`.
-
-An accepted run retains its admission until it completes or fails. Every
-actual model-generation call for that run must separately observe the
-backend's active-generation limit. This includes:
-
-- the initial generation;
-- every output-repair generation; and
-- calls made through either the built-in or an injected model client.
-
-Preparation and output validation should not hold an active-generation permit.
-A repair remains part of its already-admitted run, but reacquires active
-generation capacity so repairs cannot exceed the backend limit. It must not be
-rejected merely because new runs filled the waiting queue after its initial
-generation.
-
-Within one backend pool, waiting generation calls should be served in FIFO
-order, subject to canceled calls being removed. Different backend pools make
-progress independently; a saturated local backend must not consume
-OpenRouter's active or waiting capacity.
-
-The feature will not promise ordering across backend pools or completion order
-among admitted runs.
-
-## Capacity Failure And Cancellation
-
-When a backend's bounded waiting capacity is full, a new `Run` call will fail
-promptly rather than waiting outside the bounded admission system. The public
-API will expose a recognizable capacity-exhaustion error identity distinct
-from invalid configuration, invalid requests, and model-client failures.
-Rejected calls return no partial result and do not invoke the model client.
-
-Waiting within the admitted backlog or for active-generation capacity must
-honor the caller's context. Cancellation or deadline expiry will:
-
-- stop waiting promptly;
-- release any admission or generation capacity held by that invocation;
-- preserve the applicable context error identity; and
-- avoid invoking the model client if cancellation wins before generation
-  starts.
-
-Capacity must also be released after preparation, generation, validation,
-repair, or collaborator failure. One failed or canceled run must not reduce
-the backend's future usable capacity.
-
-Elapsed `Run` timing will include time spent waiting after the call is
-accepted. `PreparedRun` timing will continue to describe preparation rather
-than queue waiting.
-
-## Engine And Client Boundaries
-
-All pools and queued state belong to one `Engine`. Separate engines do not
-share capacity, even when they register the same backend ID or endpoint. The
-feature introduces no process-global scheduler.
-
-The engine will apply policy consistently to the built-in model client and an
-injected `LLMClient`. Consumers calling their own client outside Promptkit are
-outside this boundary. Injected clients remain responsible for their internal
-thread safety and cancellation behavior.
-
-Backend policy is keyed by the resolved backend ID rather than endpoint text.
-This preserves stable routing when a selected backend's endpoint is overridden
-and avoids accidentally combining unrelated registrations that happen to use
-the same URL.
-
-## Queue Lifetime And Observability
-
-Admission state is buffered, ephemeral, and in-process. It is not persisted
-and has no survival guarantee across engine disposal or process termination.
-Promptkit will not introduce background job ownership or require consumers to
-start or stop workers.
-
-The initial feature does not require public queue-depth metrics, callbacks, or
-inspection APIs. Capacity errors and ordinary call timing provide the
-consumer-visible behavior. Operational observability can be added later
-without coupling the scheduling mechanism to an application logging or
-metrics system.
-
-## Compatibility
-
-Consumer-registered backends and endpoint-only profiles remain unlimited
-unless concurrency is explicitly configured, preserving their existing
-behavior.
-
-The built-in OpenRouter backend will change from unlimited concurrency to a
-limit of 16 with a bounded waiting capacity of 1024. Ordinary synchronous
-calls remain unchanged, while unusually high concurrent use may now wait or
-return the capacity error. This behavioral change must be identified in the
-release notes for the version that publishes it.
-
-Adding backend policy fields and a public capacity error is otherwise
-additive. The change will use a pre-`v1` minor release under Promptkit's
-[release policy](../release.md#release-model).
-
-## Non-Goals
-
-This scope does not include:
-
-- asynchronous job handles, polling, or detached result delivery;
-- durable or cross-process queues;
-- persistence or recovery across engine or process shutdown;
-- priorities, scheduling weights, or consumer-defined fairness classes;
-- automatic retries, backoff, rate-limit interpretation, or provider quota
-  discovery;
-- token-per-minute or request-per-minute rate limiting;
-- dynamic reconfiguration after engine construction;
-- per-profile or per-run concurrency overrides;
-- endpoint-keyed pooling for profiles without a backend ID;
-- process-global coordination across engines;
-- application worker lifecycle, logging, tracing, or metrics policy; or
-- changes to prompt, profile, schema, or model-provider wire formats.
-
-## Target End State
-
-This roadmap reaches its target end state when:
-
-- each engine independently coordinates configured backend capacity;
-- the built-in OpenRouter backend allows 16 active generations and up to 1024
-  waiting runs;
-- consumer backends can opt into their own active and waiting limits while
-  remaining unlimited by default;
-- endpoint overrides retain the selected backend's capacity pool and
-  endpoint-only profiles remain unlimited;
-- synchronous `Run` callers wait for and receive their ordinary result;
-- admission is bounded before expensive preparation work where practical;
-- every initial and repair generation observes the backend's active limit
-  without serializing preparation or validation;
-- a full waiting queue returns a recognizable capacity error without invoking
-  the model client;
-- cancellation and all failure paths promptly release capacity and preserve
-  context error identity;
-- built-in and injected model clients receive the same scheduling behavior;
-- pools remain ephemeral, engine-scoped, and independent across backend IDs;
-  and
-- current-state GoDoc, consumer, internal, and release documentation describe
-  the implemented behavior once it lands.

docs/roadmap/deferred.md

@@ -0,0 +1,61 @@
+# Deferred Feature Ideas
+
+## Purpose
+
+This document catalogs feature ideas that remain potentially useful but have
+been deliberately postponed. These ideas are not awaiting ordinary selection
+from the [future feature catalog](future.md); each has a stated reason to wait
+and should be reconsidered only when its trigger becomes relevant.
+
+Deferred entries are not commitments, schedules, active implementation plans,
+or descriptions of current behavior. When an entry is reactivated, move it to
+`future.md` for evaluation or directly into a focused roadmap after its open
+design dependencies have been resolved.
+
+## Deferred Ideas
+
+### Semantic Execution-Target Fingerprints
+
+**Reason for deferral:** A stable digest requires a deliberate semantic-
+equality and versioning design. Notarius can safely use conservative source
+hashes and a Promptkit release marker today, while Weatherreporter does not
+currently reuse LLM-dependent checkpoints.
+
+Promptkit could expose an opaque equality value for a resolved profile and its
+effective generation target. This would let checkpointing consumers detect
+generation-affecting configuration changes without hashing YAML presentation
+or depending on Promptkit's built-in catalog layout.
+
+The digest should change with semantically relevant state such as the resolved
+model, endpoint, backend routing identity, request defaults, extra parameters,
+profile generation settings, and selected built-in profile semantics. It
+should exclude credential values, concurrency and queue policy, source paths,
+comments, formatting, and other representation-only changes. Whether a
+credential environment-variable name affects equality must be decided
+explicitly. The encoding should remain opaque and internally versioned so
+Promptkit can deliberately invalidate earlier digests when its resolution
+semantics change.
+
+Reconsider this idea when a downstream consumer needs Promptkit-owned
+checkpoint equality or when a broader semantic identity design is selected.
+
+### Eager Source Validation
+
+**Reason for deferral:** Exact prompt and profile inspection may already
+provide a sufficiently small validation surface. Experience from downstream
+adoption should establish whether an engine-wide operation would add enough
+value to justify its broader contract.
+
+Promptkit could provide an explicit offline operation that discovers and
+structurally validates configured prompt, profile, and schema sources without
+model generation. The normal `NewEngine` path would remain lazy.
+
+An eager operation would need coherent handling for duplicate prompt IDs and
+versions, strict YAML decoding, referenced content files, profile/backend
+membership, schema syntax and transitive references, context cancellation,
+and source-specific public errors. Credential declarations must remain
+separate from credential values; checking current environment availability,
+if supported at all, should be an explicit option and must not expose secrets.
+
+Reconsider this idea after downstream use of `InspectPrompt`,
+`InspectProfile`, and fixture-based preparation demonstrates a concrete gap.

docs/roadmap/future.md

@@ -12,6 +12,9 @@ consumer value, and important scope boundaries. Defer API design,
 implementation details, sequencing, and acceptance criteria until an idea is
 selected.
 
+Ideas that have been deliberately postponed rather than left available for
+ordinary selection belong in the [deferred catalog](deferred.md).
+
 ## Using This Catalog
 
 - Add an idea when its purpose and likely value can be stated clearly.
@@ -23,6 +26,8 @@ selected.
 - When an idea is selected, move its active planning to a focused roadmap or,
   when it requires a durable architectural decision, an ADR. Update
   current-state documentation only when implementation lands.
+- Move an idea to `deferred.md` when maintainers decide to retain it but wait
+  for a stated design dependency, demand signal, or reconsideration trigger.
 - Remove ideas that are no longer relevant. Retain a rejected idea only when
   its rationale is likely to prevent repeated reconsideration.
 
@@ -33,9 +38,8 @@ consumers.
 
 ## Ideas
 
-No ideas are currently cataloged. Backend-specific concurrency management has
+No ideas are currently awaiting selection. Active feature work belongs in its
-been selected for active planning in the
+focused roadmap rather than this catalog.
-[focused concurrency roadmap](concurrency.md).
 
 ## Entry Format
 

docs/roadmap/implementation.md

@@ -1,841 +0,0 @@
-# Backend-Specific Concurrency Management Implementation Plan
-
-**Status:** Complete.
-
-## Purpose
-
-This document is the decision-complete implementation plan for
-[backend-specific concurrency management](concurrency.md). It is written for a
-coding agent that will implement each stage in order.
-
-The feature roadmap owns the intended capability, consumer value, policy
-choices, compatibility decision, and target end state. This document owns the
-concrete API, internal representation, scheduling architecture, implementation
-sequence, test ownership, documentation updates, and completion gates.
-
-## Implementation Rules
-
-- Complete the stages in order. Keep the repository compiling and the focused
-  tests passing at every stage boundary.
-- Preserve unrelated working-tree changes. In particular, `concurrency.md` and
-  the removal of its source idea from `future.md` may already be uncommitted
-  when implementation begins; retain both.
-- Follow every policy under `docs/policy/`, the task-specific reading guide in
-  `docs/development.md`, and the target behavior in `concurrency.md`.
-- Keep the public API in the root `promptkit` package and implementation
-  details under `internal/`. Do not expose scheduler types or create another
-  public package.
-- Use only the Go standard library for scheduling. Do not add a queue,
-  semaphore, worker-pool, or metrics dependency.
-- Preserve synchronous, wait-for-result `Run`, unrestricted `Prepare`,
-  engine-local state, endpoint-only profiles, backend-selected profiles,
-  backend identity through endpoint overrides, and injected `LLMClient`
-  behavior.
-- Do not broaden the work into asynchronous jobs, durable queues, retries,
-  rate limiting, dynamic configuration, priorities, worker lifecycle,
-  endpoint-keyed pools, or public queue observability.
-- Keep all tests deterministic, bounded, offline, and race-safe. Coordinate
-  concurrent tests with channels and barriers rather than timing assumptions
-  or live providers.
-- Update exact GoDoc with each exported declaration change. Update durable
-  current-state documents only after the corresponding behavior is
-  implemented.
-- Test configurable mechanisms with small test-owned limits. Assert the exact
-  OpenRouter `16` and default queue `1024` values only at the registry contract
-  that owns those operational defaults.
-- Do not create a release, change a module version, or tag a commit. The final
-  implementation handoff must identify the built-in OpenRouter behavior change
-  for the next pre-`v1` minor release.
-
-## Fixed Design
-
-### Public Backend Configuration
-
-Append these fields to the existing root `Backend` type in `backends.go`:
-
-```go
-type Backend struct {
-	// Existing fields remain unchanged and in their current order.
-
-	ConcurrencyLimit int
-	QueueCapacity    *int
-}
-```
-
-Use these exact semantics:
-
-| Public values | Meaning |
-| --- | --- |
-| `ConcurrencyLimit == 0`, `QueueCapacity == nil` | Unlimited backend; preserve current behavior. |
-| `ConcurrencyLimit > 0`, `QueueCapacity == nil` | Limit active generations and use the default waiting capacity of 1024. |
-| `ConcurrencyLimit > 0`, `QueueCapacity != nil` | Limit active generations and use the pointed-to capacity exactly, including zero. |
-| `ConcurrencyLimit < 0` | Invalid engine configuration. |
-| `QueueCapacity != nil` and `*QueueCapacity < 0` | Invalid engine configuration. |
-| `ConcurrencyLimit == 0` and `QueueCapacity != nil` | Invalid engine configuration because a queue without an active limit has no defined consumer value. |
-
-`ConcurrencyLimit` counts simultaneous calls to the engine-owned internal
-model-client boundary for this backend. `QueueCapacity` controls additional
-accepted `Run` invocations beyond that limit. The maximum admitted runs for a
-limited backend is therefore:
-
-```text
-ConcurrencyLimit + effective QueueCapacity
-```
-
-Guard that addition against integer overflow during backend validation.
-Do not impose an arbitrary upper bound beyond non-negativity and overflow
-safety.
-
-The `QueueCapacity` pointer exists only to distinguish omission from explicit
-zero. `WithBackend` and `NewEngine` must not retain the caller's pointer.
-`Backend` continues to have no stable JSON representation, and consumers
-remain directed to keyed literals.
-
-Do not add concurrency fields to `Profile`, `ExecutionTarget`,
-`ExecutionTargetOverride`, `RunRequest`, prompt or profile files, or stable
-prepared/result JSON.
-
-### Built-In And Custom Defaults
-
-The backend registry owns these exact operational defaults:
-
-```go
-const (
-	openRouterConcurrencyLimit = 16
-	defaultQueueCapacity       = 1024
-)
-```
-
-The built-in `openrouter` definition has a normalized concurrency limit of 16
-and queue capacity of 1024.
-
-Consumer registrations remain unlimited when concurrency is omitted. For a
-consumer backend with a positive limit and omitted queue capacity, normalize
-the queue capacity to 1024. Preserve an explicitly configured zero.
-
-Consumers still cannot replace the reserved `openrouter` registration.
-Endpoint-only profiles have no backend policy and remain unlimited. A selected
-backend retains its pool when a profile or request overrides only its endpoint.
-
-### Internal Backend Representation
-
-Extend `internal/domain.Backend` with scalar policy values and explicit
-presence rather than retaining a pointer:
-
-```go
-type Backend struct {
-	// Existing fields...
-	ConcurrencyLimit int
-	QueueCapacity    int
-	QueueCapacitySet bool
-}
-
-type BackendCapacityPolicy struct {
-	ConcurrencyLimit int
-	QueueCapacity    int
-}
-```
-
-`WithBackend` converts the public pointer into `QueueCapacity` plus
-`QueueCapacitySet`. Registry normalization validates the combinations above,
-fills the default, and leaves every limited stored backend with
-`QueueCapacitySet == true`. Unlimited stored backends retain zero values and
-`QueueCapacitySet == false`.
-
-Add this internal registry method:
-
-```go
-func (r *Registry) CapacityPolicies() map[string]domain.BackendCapacityPolicy
-```
-
-It returns a newly allocated map containing only limited backends. Values are
-scalars, so callers cannot mutate registry state. The built-in OpenRouter
-policy is included. `GetBackend` continues returning a defensive backend copy,
-now including normalized scalar capacity metadata.
-
-Capacity policy is operational registry metadata. Do not merge it into an
-execution target or expose it to injected model clients.
-
-### Public Capacity Error
-
-Add this root sentinel beside the other run errors in `engine.go`:
-
-```go
-var ErrCapacityExceeded = errors.New("backend capacity exceeded")
-```
-
-Its GoDoc must state that it identifies a `Run` rejected because the selected
-backend has already admitted `ConcurrencyLimit + QueueCapacity` runs. It is
-not an invalid request, an LLM/provider rate-limit response, or an
-`ErrLLMGenerate` failure.
-
-The internal capacity component owns a corresponding internal
-`ErrCapacityExceeded`. Add its mapping in `publicErrorFor` before the broader
-generation and invalid-request cases. The public error must preserve the
-internal error through wrapping while matching `ErrCapacityExceeded` with
-`errors.Is`.
-
-A capacity rejection returns no partial result and must not invoke the
-artifact reader, renderer, schema loader, validator, or model client. Prompt,
-profile, and backend loading needed to select the pool may already have
-occurred.
-
-### Internal Capacity Component
-
-Add `internal/capacity` as the single owner of engine-local run admission and
-active-generation permits.
-
-Use these package-level boundaries:
-
-```go
-var ErrCapacityExceeded error
-
-type Manager struct {
-	// Private immutable pool map.
-}
-
-func NewManager(
-	policies map[string]domain.BackendCapacityPolicy,
-) (*Manager, error)
-
-func (m *Manager) Admit(
-	ctx context.Context,
-	backendID string,
-) (release func(), err error)
-
-func NewClient(m *Manager, next llm.Client) llm.Client
-```
-
-`NewManager` copies the supplied map and creates one independent pool per
-limited backend. Defensively reject blank IDs, non-positive concurrency
-limits, negative queue capacities, or total-capacity overflow even though the
-registry normally supplies normalized values. Construction creates no worker
-goroutines.
-
-An absent manager, blank backend ID, or ID absent from the policy map is
-unlimited:
-
-- `Admit` succeeds with a non-nil no-op release function; and
-- the client wrapper calls the next client directly.
-
-For a limited pool, `Admit` is immediate and context-aware:
-
-1. return `ctx.Err()` if the context is already done;
-2. under the pool lock, compare admitted runs with
-   `ConcurrencyLimit + QueueCapacity`;
-3. return an error matching internal `ErrCapacityExceeded` when full; or
-4. increment admitted runs and return an idempotent release function.
-
-The release function decrements admission exactly once, even if accidentally
-called more than once. It does not release an active-generation permit; those
-permits have their own lifetime.
-
-### FIFO Generation Permits
-
-`NewClient` returns an internal `llm.Client` wrapper around either the built-in
-client or the public-client adapter. It must preserve requests, successful
-responses, nil responses, and collaborator error identities exactly.
-`next` must be non-nil; `NewEngine` and internal runner construction maintain
-that invariant. A nil manager returns `next` unchanged.
-
-For a configured backend ID, the wrapper:
-
-1. acquires one active-generation permit from the matching pool;
-2. waits in FIFO order when the active count equals `ConcurrencyLimit`;
-3. removes a canceled waiter and returns `ctx.Err()` when cancellation wins
-   before the permit is granted;
-4. invokes the next client only after a permit is granted; and
-5. releases the permit with `defer` after every success, nil response,
-   collaborator error, panic unwinding, or context outcome.
-
-Implement FIFO and cancellation explicitly with a mutex and an ordered waiter
-list. A channel used only as a counting semaphore is insufficient because it
-does not define FIFO ordering or safe removal of canceled waiters.
-
-Permit grant and cancellation must have one lock-protected linearization
-point. If cancellation removes the waiter first, do not invoke the next
-client. If grant wins first, invoke the next client with the caller's context;
-the next client may then observe cancellation normally. Never lose or
-double-release a permit in this race.
-
-Releasing a permit transfers it to the oldest non-canceled waiter before
-making it generally available. Different backend pools never share admission
-or active counts.
-
-The active wrapper enforces its limit even if an internal caller invokes it
-without a run admission lease. Bounded backlog is guaranteed for ordinary
-engine `Run` calls by the runner admission path; no public API exposes the
-wrapped internal client directly.
-
-### Engine Assembly
-
-In `NewEngine`, after constructing the validated backend registry:
-
-1. obtain `backendRegistry.CapacityPolicies()`;
-2. construct one `capacity.Manager`;
-3. construct the selected base internal LLM client exactly as today;
-4. wrap that base client with `capacity.NewClient`; and
-5. pass both the wrapped client and manager-as-admitter to the runner.
-
-Every `NewEngine` call constructs a distinct manager. Do not cache managers,
-pools, or policies in package globals. The wrapper must be applied after a
-public injected client is adapted to `internal/llm.Client`, so built-in and
-injected clients receive identical scheduling behavior.
-
-If `NewManager` reports a defensive configuration error, make `NewEngine`
-return an error matching `ErrInvalidConfig`.
-
-`Prepare` does not use the manager. An injected client remains required to be
-safe for concurrent calls because different backend pools and unlimited
-backends may still invoke it concurrently.
-
-### Shared Two-Phase Preparation
-
-Refactor `internal/usecase.Runner` so `Prepare` and `Run` share one preparation
-pipeline with two private phases. Do not duplicate prompt/profile/backend
-selection or execution precedence.
-
-The first phase resolves only the state required before admission:
-
-1. validate `PromptID`;
-2. normalize the direct session ID;
-3. load the prompt definition;
-4. hash the original prompt definition at its existing error-order position;
-5. select and load the execution profile;
-6. resolve the selected backend;
-7. resolve and validate the effective execution target and credentials; and
-8. resolve the effective output contract without loading its schema.
-
-Return a private state value containing the loaded definition, normalized
-direct session, prompt-definition hash, selected profile ID, effective target,
-numeric-presence metadata, effective output contract, and preparation start
-time. Keep this value private to `internal/usecase`.
-
-The second phase consumes that state and performs:
-
-1. structured-output schema loading;
-2. artifact loading and input hashing;
-3. message and prompt-session rendering;
-4. direct-session application;
-5. rendered-prompt hashing; and
-6. `PreparedRun` construction and timing.
-
-Preserve every existing precedence rule, error identity, direct-session
-template bypass, hash input, selected identity, copy guarantee, and timing
-field. Do not reload the prompt, profile, or backend between phases.
-
-`Runner.Prepare` records its start time, runs both phases consecutively, and
-never calls admission. Its behavior and error ordering remain unchanged.
-
-`Runner.Run` records its existing run start time, runs the first preparation
-phase, and then calls:
-
-```go
-release, err := r.admitter.Admit(ctx, effectiveBackendID)
-```
-
-Use a narrow use-case-owned interface with the same signature:
-
-```go
-type RunAdmitter interface {
-	Admit(context.Context, string) (func(), error)
-}
-```
-
-A nil admitter means unlimited behavior for internal constructors and tests.
-On successful admission, immediately `defer release()` around the remainder of
-the run. Then run the second preparation phase, initial generation,
-validation, and all repair attempts.
-
-If admission returns internal `capacity.ErrCapacityExceeded`, add useful
-backend context without changing its identity. If it returns `ctx.Err()`,
-preserve that identity directly rather than recategorizing it as invalid
-request or generation failure.
-
-This refactor intentionally replaces the current literal `Run`-calls-`Prepare`
-implementation with shared private phases. Update current-state documentation
-to describe one shared pipeline rather than retaining an inaccurate call-graph
-claim.
-
-### Generation And Repair Lifetime
-
-The admission lease covers the entire accepted run:
-
-- second-phase preparation;
-- initial generation;
-- validation;
-- every repair; and
-- all failure and cancellation exits.
-
-Preparation and validation do not hold an active-generation permit. The
-wrapped client acquires a permit only around each actual `Generate` call.
-
-The runner's initial generation already carries the effective backend ID in
-`GenerateRequest.Target`. Preserve that value. `RepairRequest.Target` and the
-default repairer's generated request must continue carrying the same backend
-ID, allowing each repair to reacquire the same pool's active permit.
-
-When testing or constructing `NewRunnerWithRepairer`, pass the same wrapped
-client to both the runner and `NewDefaultOutputRepairer`. Do not add capacity
-state to `RepairRequest`, `ExecutionTarget`, or public generation values.
-
-A repair remains within its existing admission lease. It waits for a FIFO
-active permit but never performs a second bounded admission and therefore
-cannot fail merely because later runs filled the admission capacity.
-
-### Error And Cancellation Semantics
-
-The required public outcomes are:
-
-| Situation | Required error identity |
-| --- | --- |
-| Admission capacity is full | `ErrCapacityExceeded` only; not `ErrInvalidRequest` or `ErrLLMGenerate`. |
-| Context is done before admission succeeds | Preserve `ctx.Err()`; do not return capacity exhaustion. |
-| Context cancels while waiting for an active permit | Preserve `ctx.Err()` through the existing `ErrLLMGenerate` generation category. |
-| Wrapped client fails after permit acquisition | Preserve existing `ErrLLMGenerate` and collaborator identities. |
-| Preparation or validation fails after admission | Preserve its existing category and release admission. |
-
-Maintain the existing rule that `Run` returns no partial result on any
-operational error. Do not add queue status to errors or results.
-
-`RunResult.Duration` continues to start at runner entry and therefore includes
-pre-admission resolution, accepted preparation, and active-permit waiting.
-`PreparedRun.DurationMS` continues to cover only its shared preparation phases;
-it does not include later generation waiting. Capacity-rejected calls have no
-result or timing value.
-
-### Ownership And Concurrency Safety
-
-The registry, capacity policy map, pool map, and per-pool limits are immutable
-after engine construction. Only admission counts, active counts, and waiter
-lists are mutable and must be protected by the owning pool mutex.
-
-Do not retain public queue pointers, caller request values, contexts, or
-generation requests after their call completes. A canceled waiter must be
-unlinked so its context and request cannot remain reachable from the pool.
-
-Do not hold a pool mutex while:
-
-- loading or rendering prompts;
-- reading artifacts or schemas;
-- invoking a model client;
-- validating output;
-- closing a waiter notification channel if the implementation could re-enter
-  pool code; or
-- calling consumer code.
-
-No scheduler operation may spawn a goroutine whose lifetime outlasts the
-calling `Run`. The zero steady-state goroutine count is part of the
-in-process/no-worker-lifecycle design.
-
-## Test Ownership
-
-Use this ownership split and avoid repeating the full policy matrix at every
-layer:
-
-- `internal/backend/registry_test.go` owns normalization, validation, the exact
-  OpenRouter policy, the custom default queue, explicit zero, unlimited
-  omission, and policy-map copying.
-- `internal/capacity/manager_test.go` owns admission bounds, idempotent release,
-  FIFO active permits, cancellation races, capacity recovery, independent
-  pools, unlimited IDs, and observed peak concurrency.
-- `internal/capacity/client_test.go` owns wrapper request/response/error
-  transparency and the rule that cancellation before grant does not invoke the
-  next client. Combine these with manager tests if one coherent package test
-  expresses the behavior more clearly.
-- `internal/usecase/runner_test.go` owns two-phase preparation parity, pool
-  selection, admission before expensive work, admission release across run
-  exits, `Prepare` bypass, and repair reuse of the admitted backend.
-- Root external-package tests own public configuration conversion, assembled
-  engine-local behavior, endpoint-override routing, injected-client limiting,
-  and public capacity/context error identities.
-- Existing model-client HTTP tests remain unchanged because scheduling does
-  not alter the OpenAI-compatible wire contract.
-
-Concurrency tests must use test-owned limits such as one or two and
-channel-controlled blocking clients. Record observed active and peak counts
-under a mutex or atomics. Do not use `time.Sleep` to infer queue state.
-Package-internal tests may inspect a waiter list under its mutex through a
-small test helper when necessary to establish deterministic FIFO ordering; do
-not add production metrics or hooks solely for tests.
-
-Do not add separate tests for trivial scalar copies when registry or assembled
-behavior already protects them.
-
-## Stage 1 — Backend Policy And Public Configuration
-
-**Status:** Complete.
-
-### Goal
-
-Add the public and internal backend policy representation, normalize all
-configured states, and expose immutable normalized policies without changing
-runtime scheduling yet.
-
-### Work
-
-1. Add `ConcurrencyLimit` and `QueueCapacity` to `Backend` in `backends.go`
-   with exact GoDoc for unlimited, defaulted, explicit-zero, invalid, and
-   engine-scoped behavior.
-2. Convert the public queue pointer into scalar value plus presence in
-   `WithBackend`; do not retain the pointer.
-3. Add the internal backend policy fields and
-   `BackendCapacityPolicy` to `internal/domain/domain.go`.
-4. Add the two registry-owned constants and configure the built-in OpenRouter
-   definition with 16 and 1024.
-5. Extend `normalizeBackend` with the fixed validation, defaulting, explicit
-   zero, and overflow rules.
-6. Add `Registry.CapacityPolicies`, returning only limited policies in a fresh
-   map.
-7. Update existing backend composite literals and assertions only where the
-   new fields are relevant. Continue using keyed literals.
-
-### Tests
-
-1. Extend the exact built-in registry test with the OpenRouter limit and queue.
-2. Add one coherent table covering unlimited omission, default queue,
-   explicit-zero queue, negative values, queue-without-limit, and total
-   overflow.
-3. Extend the registry copy/normalization test to prove returned policy maps
-   cannot mutate registry state.
-4. Add root coverage only if needed to prove the public pointer/presence
-   conversion; do not reproduce registry validation cases at the facade.
-
-### Focused Validation
-
-Run:
-
-```sh
-gofmt -w backends.go internal/domain/domain.go \
-  internal/backend/registry.go internal/backend/registry_test.go
-go test . ./internal/backend
-go vet . ./internal/backend
-git diff --check
-```
-
-Include another touched Go test file in `gofmt` only if it actually changed.
-
-### Completion Gate
-
-This stage is complete when every public configuration state has one normalized
-internal meaning, OpenRouter exposes exactly 16/1024, custom backends remain
-unlimited by omission, and no runtime call is scheduled yet.
-
-## Stage 2 — Engine-Local Capacity Manager
-
-**Status:** Complete.
-
-### Goal
-
-Implement and prove the bounded admission mechanism and FIFO active-generation
-client wrapper independently of runner orchestration.
-
-### Work
-
-1. Add `internal/capacity/manager.go` with the manager, immutable policy copy,
-   per-backend pools, internal error, immediate admission, idempotent release,
-   and FIFO context-aware active permits.
-2. Add `internal/capacity/client.go` with the transparent `llm.Client` wrapper.
-3. Use mutex-protected waiter state and an ordered list; explicitly resolve
-   grant-versus-cancel races.
-4. Ensure unlimited and independent-pool fast paths avoid queue allocation.
-5. Do not start workers, timers, cleanup goroutines, or process-global state.
-
-### Tests
-
-1. Add a compact constructor-validation table for blank IDs, non-positive
-   limits, negative queues, and total-capacity overflow.
-2. With a small configured policy, prove that exactly
-   `limit + queueCapacity` admissions succeed, the next matches
-   `ErrCapacityExceeded`, and a release permits another admission.
-3. Prove release is idempotent.
-4. Drive more blocked client calls than the active limit and assert observed
-   peak concurrency never exceeds that limit.
-5. Prove FIFO order with deterministic queue-entry synchronization.
-6. Cancel the first and a middle waiter and prove they are removed, never call
-   the wrapped client, and do not block later waiters.
-7. Exercise the grant/cancel race repeatedly under `go test -race`, asserting
-   no permit leak or double invocation.
-8. Prove different backend IDs proceed independently and blank, unknown, or
-   nil-manager paths remain unlimited.
-9. Prove request values, successful and nil responses, and collaborator errors
-   pass through unchanged after permit acquisition.
-
-### Focused Validation
-
-Run:
-
-```sh
-gofmt -w internal/capacity/manager.go \
-  internal/capacity/manager_test.go \
-  internal/capacity/client.go \
-  internal/capacity/client_test.go
-go test ./internal/capacity
-go test -race ./internal/capacity
-go vet ./internal/capacity
-git diff --check
-```
-
-If tests are combined into one file, omit the nonexistent file from `gofmt`.
-
-### Completion Gate
-
-This stage is complete when the standalone component enforces relational
-admission and active limits, FIFO cancellation is race-safe, separate pools
-are independent, and the wrapper is transparent apart from waiting.
-
-## Stage 3 — Shared Preparation And Early Run Admission
-
-**Status:** Complete.
-
-### Goal
-
-Refactor runner preparation into one shared two-phase pipeline and place
-bounded admission after backend resolution but before schema, artifact, and
-rendering work.
-
-### Work
-
-1. Add the private pre-admission preparation state and split the existing
-   `Prepare` logic according to the fixed design.
-2. Make `Runner.Prepare` call both phases without an admitter.
-3. Add the `RunAdmitter` interface and runner field.
-4. Update `NewRunner` and `NewRunnerWithRepairer` to accept the optional
-   admitter; update internal call sites with nil until root assembly is wired.
-5. Change `Runner.Run` to use the first phase, admit by effective backend ID,
-   defer the returned release, and then use the second phase.
-6. Preserve all existing error precedence, target resolution, hashes,
-   metadata, session behavior, and timing.
-7. Return capacity and context errors with the fixed identities. Do not invoke
-   later collaborators after rejection.
-
-### Tests
-
-1. Keep the existing `Run`/`Prepare` parity coverage passing to prove the
-   shared phases do not drift.
-2. Add a fake admitter that records backend IDs and release calls.
-3. Prove a backend-selected run admits with the selected ID even when the
-   endpoint is overridden.
-4. Prove an endpoint-only run uses the unlimited/blank identity and that
-   `Prepare` never calls admission.
-5. Reject admission and assert schema, artifact, renderer, validator, repairer,
-   and LLM collaborators are not invoked.
-6. Prove admission is released after one successful run and representative
-   second-phase, generation, and validation errors. Prefer a small table around
-   the single `defer` invariant rather than duplicating every error test.
-7. Retain direct-session, backend precedence, credential, hashing, and repair
-   tests unchanged except for constructor arguments.
-
-### Focused Validation
-
-Run:
-
-```sh
-gofmt -w internal/usecase/runner.go \
-  internal/usecase/runner_test.go
-go test ./internal/usecase
-go test -race ./internal/usecase
-go vet ./internal/usecase
-git diff --check
-```
-
-### Completion Gate
-
-This stage is complete when `Prepare` remains unrestricted, `Run` admits after
-one canonical routing phase and before expensive completion work, every exit
-releases admission, and existing preparation semantics remain unchanged.
-
-## Stage 4 — Engine Assembly And Public Runtime Contract
-
-**Status:** Complete.
-
-### Goal
-
-Wire one manager into each engine, schedule built-in and injected clients,
-expose the capacity error, and prove assembled runtime behavior.
-
-### Work
-
-1. Add public `ErrCapacityExceeded` and its exact GoDoc in `engine.go`.
-2. Map internal capacity exhaustion in `errors.go`.
-3. Construct the manager from the registry policy snapshot in `NewEngine`.
-4. Wrap the selected internal client after built-in or injected-client
-   selection and pass the manager and wrapped client to the runner.
-5. Update `Engine`, `NewEngine`, `Run`, `WithLLMClient`, and `LLMClient` GoDoc
-   only where concurrency, capacity, or cancellation statements change.
-6. Ensure manager-construction errors match `ErrInvalidConfig`.
-7. For internal repair coverage, construct the default repairer with the same
-   wrapped client used by its runner and confirm repair target backend identity
-   remains intact.
-
-### Tests
-
-1. Add an external-package assembled test with a small custom limit and a
-   blocking injected client; assert peak generation equals or remains below
-   the configured limit.
-2. With queue capacity zero, block one accepted run before generation and
-   assert the next matching-backend run returns `ErrCapacityExceeded`, does not
-   match `ErrInvalidRequest` or `ErrLLMGenerate`, returns no result, and never
-   reaches expensive collaborators or the client.
-3. In the same or another focused workflow, prove an endpoint override remains
-   in the selected backend's pool.
-4. Prove two engines with the same backend ID have independent capacity.
-5. Prove an unlimited custom backend and an endpoint-only profile preserve
-   concurrent behavior.
-6. Cancel a call waiting for an active permit; assert it matches both
-   `context.Canceled` and `ErrLLMGenerate`, never invokes the injected client,
-   and leaves capacity reusable.
-7. Add one internal repair workflow with concurrent runs or controlled permits
-   showing initial and repair generations never exceed the same backend limit
-   and repairs do not perform a second admission.
-8. Extend the public error sentinel contract test with
-   `ErrCapacityExceeded`.
-
-Avoid a second HTTP-level concurrency suite: the capacity client tests and one
-assembled injected-client workflow already protect the shared wrapper used by
-the built-in client.
-
-### Focused Validation
-
-Run:
-
-```sh
-gofmt -w engine.go errors.go backends.go \
-  internal/usecase/runner.go internal/usecase/runner_test.go \
-  engine_test.go public_contract_test.go
-go test . ./internal/backend ./internal/capacity ./internal/usecase
-go test -race . ./internal/capacity ./internal/usecase
-go vet . ./internal/backend ./internal/capacity ./internal/usecase
-git diff --check
-```
-
-Add any newly created capacity files to `gofmt` when they changed in this
-stage.
-
-### Completion Gate
-
-This stage is complete when every engine has independent pools, limited runs
-are bounded and FIFO at generation, endpoint routing is correct, capacity and
-context errors are stable, repairs reuse admission, and both client kinds pass
-through the same wrapper.
-
-## Stage 5 — Durable Documentation And Final Validation
-
-**Status:** Complete.
-
-### Goal
-
-Move implemented contracts into their durable owners, record compatibility
-impact, and validate the complete repository.
-
-### Work
-
-1. Review every changed exported declaration. Ensure GoDoc is the canonical
-   owner of exact field types, nil/zero semantics, defaulting, error identity,
-   engine scope, concurrency safety, cancellation, and source compatibility.
-2. Update `doc.go` so its concurrency summary acknowledges backend scheduling
-   while continuing to require injected collaborators to be concurrency-safe.
-3. Update `docs/consumers/pkg-promptkit.md` with task-oriented examples for:
-   - a limited local backend;
-   - omitted queue capacity selecting 1024;
-   - explicit zero queue capacity; and
-   - handling `ErrCapacityExceeded`.
-   Keep exact field semantics in GoDoc rather than duplicating a full table.
-4. Add `docs/internal/capacity.md` as the durable owner of pool lifecycle,
-   admission, FIFO active permits, cancellation, client wrapping, and test
-   ownership.
-5. Add `internal/capacity` to `docs/internal/overview.md`.
-6. Update `docs/policy/architecture.md` to include the implemented component
-   and root assembly dependency without turning policy into an API reference.
-7. Update `docs/internal/runner.md` to describe the shared two-phase
-   preparation pipeline, early bounded admission, lease lifetime, generation
-   permits, repairs, capacity failures, and cancellation.
-8. Review `docs/formats.md`; add only a concise link or clarification if needed
-   to explain that endpoint overrides preserve backend capacity identity.
-   Do not add concurrency fields to YAML.
-9. Do not change the OpenAI-compatible integration contract or
-   `docs/internal/llm.md` unless implementation changes their current
-   statements; scheduling is outside the provider wire contract and concrete
-   model-client implementation.
-10. Record in the implementation handoff that built-in OpenRouter now limits
-    active generations to 16 with queue capacity 1024 and that the release
-    must be a pre-`v1` minor release. Do not edit the release procedure or
-    create a tag.
-11. After every check passes, set `concurrency.md`, this implementation plan,
-    and each stage status to `Complete`. Do not remove the roadmaps in the
-    implementation change; lifecycle retirement follows review.
-
-### Full Validation
-
-Run the complete sequence from `docs/development.md`:
-
-```sh
-go test ./...
-go test -race ./...
-go vet ./...
-go build ./...
-go run ./examples/go-library/prepare
-gofmt -l $(git ls-files '*.go')
-git diff --check
-```
-
-The formatting command must produce no paths. Follow every added or changed
-Markdown link and confirm its target and heading exist.
-
-Also inspect:
-
-```sh
-git status --short
-git diff --stat
-git diff
-```
-
-Confirm that:
-
-- only intended backend, capacity, runner, facade, test, documentation, and
-  roadmap files changed;
-- no `go.work`, `go.work.sum`, local module replacement, credential,
-  generated binary, coverage output, or unrelated change was introduced;
-- the built-in OpenRouter policy is exactly 16/1024;
-- custom and endpoint-only backends remain unlimited by omission;
-- explicit queue zero is distinguishable from omission;
-- no capacity value enters execution targets, generated requests, stable JSON,
-  prompt/profile YAML, or provider payloads;
-- every engine owns distinct pools with no package-global mutable state;
-- every initial and repair generation uses the active permit wrapper;
-- capacity and waiter state is released on success, error, panic unwinding,
-  and cancellation;
-- concurrency tests use deterministic coordination rather than sleeps;
-- current-state documentation describes implemented behavior rather than
-  referring readers to the roadmaps; and
-- the feature and implementation roadmaps contain no unresolved work marked
-  complete.
-
-### Completion Gate
-
-The implementation is complete only when every target-end-state item in
-`concurrency.md` is implemented, race-enabled tests demonstrate the configured
-limits and cancellation safety, durable contracts no longer depend on roadmap
-prose, and the OpenRouter compatibility change is clearly reported for the
-next minor release.
-
-## Implementation Handoff
-
-Backend-specific capacity management is implemented and has passed the complete
-repository validation sequence. The built-in OpenRouter backend now permits 16
-active generations and a waiting capacity of 1024. Custom backends remain
-unlimited when their limit is omitted, and endpoint-only profiles remain
-unlimited.
-
-Publishing this behavior requires a pre-`v1` minor release. Its release notes
-must identify that unusually high concurrent OpenRouter use can now wait or
-return `ErrCapacityExceeded`. This implementation does not change a module
-version or create a tag.
-
-## Open Questions
-
-None. The feature roadmap and this plan fix the public representation,
-registry defaults, admission bound, FIFO generation behavior, early-routing
-refactor, cancellation races, error identities, engine and repair lifetimes,
-test ownership, compatibility treatment, and non-goals required for
-implementation.

engine.go

@@ -6,23 +6,12 @@ import (
 	"fmt"
 	"io/fs"
 	"net/http"
-	"os"
-	"path/filepath"
-	"strings"
 	"time"
 
-	artifactadapter "gitea.maximumdirect.net/eric/promptkit/internal/artifact"
-	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
-	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
-	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
-	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+	internalengine "gitea.maximumdirect.net/eric/promptkit/internal/engine"
-	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+	"gitea.maximumdirect.net/eric/promptkit/internal/observation"
-	"gitea.maximumdirect.net/eric/promptkit/internal/profile/builtin"
-	"gitea.maximumdirect.net/eric/promptkit/internal/prompt"
-	"gitea.maximumdirect.net/eric/promptkit/internal/promptdef"
 	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
-	"gitea.maximumdirect.net/eric/promptkit/internal/validate"
 )
 
 // ErrInvalidConfig identifies invalid engine construction, including missing
@@ -53,9 +42,9 @@ var (
 	// an execution profile or resolve its backend, except for the profile
 	// not-found case represented by ErrProfileNotFound.
 	ErrProfileLoad = errors.New("failed to load execution profile")
-	// ErrAPIKeyEnvMissing identifies an APIKeyEnv whose environment variable is
+	// ErrAPIKeyEnvMissing identifies an explicitly required APIKeyEnv whose
-	// unset or empty when no direct RunRequest.APIKey takes precedence. Such an
+	// environment variable is unset or empty after direct RunRequest.APIKey
-	// error also matches ErrInvalidRequest.
+	// precedence is applied. Such an error also matches ErrInvalidRequest.
 	ErrAPIKeyEnvMissing = errors.New("api_key_env points to an unset environment variable")
 	// ErrArtifactLoad identifies a failure to resolve an input artifact. Errors
 	// returned by an injected ArtifactReader remain available through errors.Is.
@@ -63,28 +52,46 @@ var (
 	// ErrPromptRender identifies a failure to render prompt messages or the
 	// session ID from the resolved inputs and variables.
 	ErrPromptRender = errors.New("failed to render prompt")
-	// ErrCapacityExceeded identifies a Run rejected because the selected backend
+	// ErrCapacityExceeded identifies a Run or RunPrepared rejected because the
-	// already admitted ConcurrencyLimit + QueueCapacity calls. It is not an
+	// selected backend already admitted ConcurrencyLimit + QueueCapacity calls.
-	// invalid request, an LLM or provider rate-limit response, or ErrLLMGenerate.
+	// A [CapacityError] reports the selected backend ID. It is not an invalid
+	// request, an LLM or provider rate-limit response, or ErrLLMGenerate.
 	ErrCapacityExceeded = errors.New("backend capacity exceeded")
 	// ErrLLMGenerate identifies a model-client failure or a nil successful
-	// response. Errors returned by an injected LLMClient remain available
+	// response. A built-in OpenAI-compatible non-2xx response is available as a
-	// through errors.Is.
+	// [GenerationError]. Errors returned by an injected LLMClient remain
+	// available through errors.Is.
 	ErrLLMGenerate = errors.New("failed to generate output")
+	// ErrTransport identifies a failed HTTP request by the built-in model
+	// client. Generation errors in this category also match ErrLLMGenerate and
+	// retain the transport cause for errors.Is and errors.As. It does not imply
+	// that the provider received no work.
+	ErrTransport = errors.New("model transport failed")
+	// ErrMalformedResponse identifies an invalid successful model response or
+	// a nil successful response from an injected client. Generation errors in
+	// this category also match ErrLLMGenerate.
+	ErrMalformedResponse = errors.New("malformed model response")
+	// ErrRetriesExhausted identifies a built-in generation sequence whose
+	// positive transient retry budget ended with another eligible HTTP status.
+	// RetryExhaustedError retains the final cause and actual attempt count.
+	ErrRetriesExhausted = errors.New("transient model retries exhausted")
 	// ErrValidation identifies an operational failure to load or compile a
 	// schema or validate output. A completed validation whose Status is
 	// ValidationFailed is returned in RunResult without this error.
 	ErrValidation = errors.New("failed to validate output")
 )
 
-// Engine prepares and runs Promptkit prompt requests.
+// Engine inspects prompts and profiles and prepares and runs Promptkit prompt
+// requests.
 //
-// An Engine is safe for concurrent calls to [Engine.Prepare] and [Engine.Run].
+// An Engine is safe for concurrent calls to [Engine.InspectPrompt],
-// Each Engine owns independent backend-capacity pools that coordinate Run
+// [Engine.InspectProfile], [Engine.Prepare], [Engine.PrepareExecution],
-// admission and model generation. Injected collaborators may still be invoked
+// [Engine.Run], and [Engine.RunPrepared]. Each Engine owns independent
-// concurrently across different backend pools or for unlimited backends.
+// backend-capacity pools that coordinate Run and RunPrepared admission and
+// model generation. Injected collaborators may still be invoked concurrently
+// across different backend pools or for unlimited backends.
 type Engine struct {
-	runner *usecase.Runner
+	internal *internalengine.Engine
 }
 
 // Config selects the directory-backed sources and built-in model-client
@@ -94,9 +101,10 @@ type Config struct {
 	// It is required unless a WithPromptFS or WithPromptFile option supplies the
 	// prompt source.
 	PromptDir string
-	// ProfileDir is an optional directory whose profiles take precedence over
+	// ProfileDir is an optional ordinary configured source whose profiles take
-	// embedded built-in profiles. An empty value selects only built-ins unless
+	// precedence over application fallback and maintained catalog profiles. An
-	// profile options are also supplied.
+	// empty value selects the lower-precedence sources unless a profile-source
+	// option supplies the ordinary source.
 	ProfileDir string
 	// SchemaDir is the root for JSON Schema files. An empty value uses the
 	// current directory. WithSchemaFS or WithSchemaFile replaces this source.
@@ -110,17 +118,22 @@ type Config struct {
 	// Timeout or the 10-minute default. The supplied client is not mutated. This
 	// field is ignored when WithLLMClient is used.
 	HTTPClient *http.Client
+	// TransientRetries selects additional built-in HTTP attempts after eligible
+	// provider statuses. Nil defaults to two; zero disables retries. Negative
+	// values are invalid. Injected LLMClient implementations are never retried
+	// automatically.
+	TransientRetries *int
 }
 
 // Option customizes engine construction.
 //
 // NewEngine applies options in argument order and ignores nil options. Within
-// each prompt-source, profile-source, in-memory-profile, schema-source,
+// each prompt-source, ordinary-profile-source, fallback-profile-source,
-// model-client, and artifact-reader category, the last non-nil valid option
+// in-memory-profile, schema-source, model-client, and artifact-reader
-// replaces earlier options in that category. WithBackend is the additive
+// category, the last non-nil valid option replaces earlier options in that
-// exception: unique registrations accumulate, and a repeated backend ID is an
+// category. WithBackend is the additive exception: unique registrations
-// error rather than a replacement. An invalid option fails construction even
+// accumulate, and a repeated backend ID is an error rather than a replacement.
-// if a later option would replace it.
+// An invalid option fails construction even if a later option would replace it.
 type Option interface {
 	apply(*engineOptions) error
 }
@@ -132,32 +145,23 @@ func (f optionFunc) apply(options *engineOptions) error {
 }
 
 type engineOptions struct {
-	llmClient       llm.Client
+	internal internalengine.Options
-	artifactReader  artifactadapter.Reader
-	promptDefs      promptdef.Repository
-	profiles        profile.Repository
-	memoryProfiles  profile.Repository
-	backends        []domain.Backend
-	validator       validate.Validator
-	promptSource    bool
-	profileSource   bool
-	memorySource    bool
-	validatorSource bool
-	artifactSource  bool
 }
 
-// WithLLMClient replaces the built-in model client used by [Engine.Run].
+// WithLLMClient replaces the built-in model client used by [Engine.Run] and
+// [Engine.RunPrepared].
 //
 // A nil client makes NewEngine fail with ErrInvalidConfig. The Engine schedules
 // Generate calls according to the selected backend's capacity policy, but the
 // client may still be called concurrently across different backend pools or for
-// unlimited backends. The client is not used by [Engine.Prepare].
+// unlimited backends. The client is not used by [Engine.Prepare] or
+// [Engine.PrepareExecution].
 func WithLLMClient(client LLMClient) Option {
 	return optionFunc(func(options *engineOptions) error {
 		if client == nil {
 			return ErrInvalidConfig
 		}
-		options.llmClient = publicLLMClientAdapter{client: client}
+		options.internal.SetLLMClient(publicLLMClientAdapter{client: client})
 		return nil
 	})
 }
@@ -172,8 +176,7 @@ func WithArtifactReader(reader ArtifactReader) Option {
 		if reader == nil {
 			return ErrInvalidConfig
 		}
-		options.artifactReader = publicArtifactReaderAdapter{reader: reader}
+		options.internal.SetArtifactReader(publicArtifactReaderAdapter{reader: reader})
-		options.artifactSource = true
 		return nil
 	})
 }
@@ -187,15 +190,7 @@ func WithArtifactReader(reader ArtifactReader) Option {
 // prompt-source options.
 func WithPromptFS(fsys fs.FS, root string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		if fsys == nil {
+		return options.internal.SetPromptFS(fsys, root)
-			return ErrInvalidConfig
-		}
-		if strings.TrimSpace(root) == "" {
-			return ErrInvalidConfig
-		}
-		options.promptDefs = promptdef.NewFSRepository(fsys, root)
-		options.promptSource = true
-		return nil
 	})
 }
 
@@ -206,73 +201,79 @@ func WithPromptFS(fsys fs.FS, root string) Option {
 // This option replaces Config.PromptDir and earlier prompt-source options.
 func WithPromptFile(path string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		fsys, root, err := fileSource(path)
+		return options.internal.SetPromptFile(path)
-		if err != nil {
-			return err
-		}
-		options.promptDefs = promptdef.NewFSRepository(fsys, root)
-		options.promptSource = true
-		return nil
 	})
 }
 
 // WithProfileFS loads execution profiles from fsys under root.
 //
-// Profiles from this source overlay built-in profiles. Profile YAML must use
+// Profiles from this ordinary configured source take precedence over
-// api_key_env for environment-based credentials; raw API keys are rejected.
+// application fallback and built-in profiles. Profile YAML must use api_key_env
-// fsys must be non-nil and root must be non-empty; otherwise NewEngine fails
+// for environment-based credentials; raw API keys are rejected. fsys must be
-// with ErrInvalidConfig. This option replaces Config.ProfileDir and earlier
+// non-nil and root must be non-empty; otherwise NewEngine fails with
-// file or FS profile-source options, but remains below WithProfiles in
+// ErrInvalidConfig. This option replaces Config.ProfileDir and earlier file or
-// precedence.
+// FS profile-source options, but remains below WithProfiles in precedence.
 func WithProfileFS(fsys fs.FS, root string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		if fsys == nil {
+		return options.internal.SetProfileFS(fsys, root)
-			return ErrInvalidConfig
-		}
-		if strings.TrimSpace(root) == "" {
-			return ErrInvalidConfig
-		}
-		options.profiles = profile.NewFSRepository(fsys, root)
-		options.profileSource = true
-		return nil
 	})
 }
 
 // WithProfileFile loads execution profiles from the single profile file at path.
 //
-// The profile overlays built-in profiles. Profile YAML must use api_key_env for
+// The profile takes precedence over application fallback and built-in profiles.
-// environment-based credentials; raw API keys are rejected. path must name an
+// Profile YAML must use api_key_env for environment-based credentials; raw API
-// existing non-directory file when NewEngine applies the option. This option
+// keys are rejected. path must name an existing non-directory file when
-// replaces Config.ProfileDir and earlier file or FS profile-source options,
+// NewEngine applies the option. This option replaces Config.ProfileDir and
-// but remains below WithProfiles in precedence.
+// earlier file or FS profile-source options, but remains below WithProfiles in
+// precedence.
 func WithProfileFile(path string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		fsys, root, err := fileSource(path)
+		return options.internal.SetProfileFile(path)
-		if err != nil {
+	})
-			return err
+}
-		}
+
-		options.profiles = profile.NewFSRepository(fsys, root)
+// WithFallbackProfileFS supplies application-owned fallback profile
-		options.profileSource = true
+// definitions from fsys under root.
-		return nil
+//
+// Profile lookup checks, in order, profiles supplied by WithProfiles; the
+// ordinary configured source selected by WithProfileFile, WithProfileFS, or
+// Config.ProfileDir; this fallback source; and Promptkit's maintained catalog
+// profiles. Each source supplies a complete profile definition; profile fields
+// are not merged between sources. Only an absent profile ID proceeds to the
+// next source. A matching read, parse, duplicate, validation, or credential
+// format failure stops resolution.
+//
+// Files use the ordinary strict profile YAML and api_key_env credential rules.
+// Loading and validation are lazy: NewEngine validates this option's arguments
+// but does not read profile files. fsys must be non-nil and root must be
+// nonblank; otherwise NewEngine returns an error matching ErrInvalidConfig.
+// Repeating this option replaces the earlier valid fallback source.
+//
+// This option controls profile-definition lookup, not provider or generation
+// failover.
+func WithFallbackProfileFS(fsys fs.FS, root string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		return options.internal.SetFallbackProfileFS(fsys, root)
 	})
 }
 
 // WithProfiles configures in-memory profiles that take precedence over
-// configured profile files and built-in profiles.
+// ordinary configured, application fallback, and built-in profiles.
 //
-// NewEngine validates and copies every profile. IDs must be unique within one
+// NewEngine locally validates and copies every profile. IDs must be unique
-// call. An invalid profile, duplicate ID, or unsupported ExtraParams value
+// within one call. An invalid local definition, duplicate ID, or unsupported
-// makes construction fail with ErrInvalidConfig. Repeating WithProfiles
+// ExtraParams value makes construction fail with ErrInvalidConfig. A derived
-// replaces the complete earlier in-memory set rather than merging it.
+// profile's base reference and resolved target completeness are checked when it
+// is selected or inspected. Repeating WithProfiles replaces the complete
+// earlier in-memory set rather than merging it.
 func WithProfiles(profiles ...Profile) Option {
 	return optionFunc(func(options *engineOptions) error {
-		repo, err := newMemoryProfileRepository(profiles)
+		definitions := make([]domain.ExecutionProfile, len(profiles))
-		if err != nil {
+		for index, publicProfile := range profiles {
-			return err
+			definitions[index] = toDomainProfile(publicProfile)
 		}
-		options.memoryProfiles = repo
+		return options.internal.SetMemoryProfiles(definitions)
-		options.memorySource = true
-		return nil
 	})
 }
 
@@ -284,15 +285,7 @@ func WithProfiles(profiles ...Profile) Option {
 // replaces Config.SchemaDir and earlier schema-source options.
 func WithSchemaFS(fsys fs.FS, root string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		if fsys == nil {
+		return options.internal.SetSchemaFS(fsys, root)
-			return ErrInvalidConfig
-		}
-		if strings.TrimSpace(root) == "" {
-			return ErrInvalidConfig
-		}
-		options.validator = validate.NewFSValidator(fsys, root)
-		options.validatorSource = true
-		return nil
 	})
 }
 
@@ -303,13 +296,7 @@ func WithSchemaFS(fsys fs.FS, root string) Option {
 // replaces Config.SchemaDir and earlier schema-source options.
 func WithSchemaFile(path string) Option {
 	return optionFunc(func(options *engineOptions) error {
-		fsys, root, err := fileSource(path)
+		return options.internal.SetSchemaFile(path)
-		if err != nil {
-			return err
-		}
-		options.validator = validate.NewFSValidator(fsys, root)
-		options.validatorSource = true
-		return nil
 	})
 }
 
@@ -324,7 +311,8 @@ func WithSchemaFile(path string) Option {
 // NewEngine returns an error matching ErrInvalidConfig for invalid
 // configuration, options, or backend-capacity policies. Each constructed
 // Engine has independent backend-capacity pools. Construction does not perform
-// model requests or require credentials.
+// model requests or require credentials. A negative Config.TransientRetries
+// is invalid.
 func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 	var options engineOptions
 	for _, opt := range opts {
@@ -336,94 +324,106 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 		}
 	}
 
-	promptDefs := options.promptDefs
+	assembled, err := internalengine.New(
-	if !options.promptSource {
+		internalengine.Config{PromptDir: cfg.PromptDir, ProfileDir: cfg.ProfileDir, SchemaDir: cfg.SchemaDir, Timeout: cfg.Timeout, HTTPClient: cfg.HTTPClient, TransientRetries: copyIntPtr(cfg.TransientRetries)},
-		if strings.TrimSpace(cfg.PromptDir) == "" {
+		options.internal,
-			return nil, fmt.Errorf("%w: prompt directory is required", ErrInvalidConfig)
+	)
-		}
-		promptDefs = promptdef.NewFilesystemRepository(cfg.PromptDir)
-	}
-
-	profiles := builtin.NewRepositoryWithDirectory(cfg.ProfileDir)
-	if options.profileSource {
-		profiles = builtin.NewRepositoryWithPrimary(options.profiles)
-	}
-	if options.memorySource {
-		profiles = profile.NewOverlayRepository(options.memoryProfiles, profiles)
-	}
-
-	backendRegistry, err := backend.NewRegistry(options.backends)
 	if err != nil {
-		return nil, fmt.Errorf("%w: failed to construct backend registry: %v", ErrInvalidConfig, err)
+		return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
 	}
-
+	return &Engine{internal: assembled}, nil
-	capacityManager, err := capacity.NewManager(backendRegistry.CapacityPolicies())
-	if err != nil {
-		return nil, fmt.Errorf("%w: failed to construct backend capacity manager: %v", ErrInvalidConfig, err)
-	}
-
-	validator := options.validator
-	if !options.validatorSource {
-		schemaDir := cfg.SchemaDir
-		if strings.TrimSpace(schemaDir) == "" {
-			schemaDir = defaults.SchemaDirDefault
-		}
-		validator = validate.NewStandardValidator(schemaDir)
-	}
-
-	llmClient := options.llmClient
-	if llmClient == nil {
-		var err error
-		llmClient, err = llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
-			Timeout:    cfg.Timeout,
-			HTTPClient: cfg.HTTPClient,
-		})
-		if err != nil {
-			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
-		}
-	}
-	llmClient = capacity.NewClient(capacityManager, llmClient)
-
-	artifacts := options.artifactReader
-	if !options.artifactSource {
-		artifacts = artifactadapter.NewCompositeReader()
-	}
-
-	return &Engine{
-		runner: usecase.NewRunner(
-			promptDefs,
-			profiles,
-			backendRegistry,
-			artifacts,
-			prompt.NewGoRenderer(),
-			llmClient,
-			validator,
-			capacityManager,
-		),
-	}, nil
 }
 
-func fileSource(name string) (fs.FS, string, error) {
+// InspectPrompt resolves one explicit prompt definition without selecting a
-	cleanName := strings.TrimSpace(name)
+// profile or starting execution work.
-	if cleanName == "" {
+//
-		return nil, "", ErrInvalidConfig
+// InspectPrompt requires a nonblank promptID. It passes nonblank promptID and
+// promptVersion values unchanged to the engine's ordinary, case-sensitive
+// prompt selection. An empty version succeeds only when that source has one
+// selected ID; a nonempty version selects one exact ID/version pair. The
+// configured prompt source is used without merging, fallback, or enumeration.
+//
+// A successful result proves that the selected definition and any referenced
+// message content files were structurally loaded. Inputs are returned in
+// definition order. DefaultProfileID is declared metadata only and is not
+// resolved. OutputContract is the normalized declared contract, with a JSON
+// Schema path when declared but without loading or compiling that schema.
+// PromptHash is the same opaque equality value as PreparedRun.PromptHash for
+// the selected definition and observed source state; its spelling, length,
+// encoding, algorithm, and security properties are not contracts.
+//
+// This method does not return prompt bodies, templates, source paths, schemas,
+// rendered messages, or execution settings. It does not resolve a profile or
+// credential, read artifacts or schemas, render, validate, admit capacity,
+// contact a provider, or generate model output. The returned PromptInspection
+// and its input slice are caller-owned. Filesystem-backed inspection is a
+// point-in-time lookup and does not freeze a definition for later execution.
+//
+// A nil Engine returns an error matching ErrInvalidConfig. A blank prompt ID
+// matches ErrInvalidRequest. An absent exact ID or version matches
+// ErrPromptNotFound and not ErrPromptLoad. Malformed, unreadable, duplicate,
+// ambiguous, referenced-content, or hashing failures match ErrPromptLoad.
+// Cancellation during lookup matches ErrPromptLoad while preserving the
+// context error. InspectPrompt returns no partial result on error.
+func (e *Engine) InspectPrompt(
+	ctx context.Context,
+	promptID string,
+	promptVersion string,
+) (*PromptInspection, error) {
+	if e == nil || e.internal == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
 	}
-	dir := filepath.Dir(cleanName)
+
-	base := filepath.Base(cleanName)
+	inspection, err := e.internal.InspectPrompt(ctx, promptID, promptVersion)
-	if base == "." || base == string(filepath.Separator) || strings.TrimSpace(base) == "" {
-		return nil, "", ErrInvalidConfig
-	}
-	info, err := os.Stat(cleanName)
 	if err != nil {
-		return nil, "", fmt.Errorf("%w: failed to access source file %q: %v", ErrInvalidConfig, cleanName, err)
+		return nil, mapPublicError(err)
 	}
-	if info.IsDir() {
+	return fromDomainPromptInspection(inspection), nil
-		return nil, "", fmt.Errorf("%w: source path %q must be a file", ErrInvalidConfig, cleanName)
+}
+
+// InspectProfile resolves one explicit profile without selecting a prompt or
+// starting execution work.
+//
+// InspectProfile trims surrounding whitespace from profileID and looks up the
+// resulting nonblank ID exactly and case-sensitively through the engine's
+// in-memory, ordinary configured-source, application fallback, and built-in
+// profile precedence. It applies the framework timeout baseline, selected
+// backend, and then selected profile to EffectiveModelParams without a request
+// override. BackendID is empty for an endpoint-only profile.
+//
+// APIKeyEnv in the returned target is an environment-variable name, never its
+// value. APIKeyRequired instead reports a direct credential requirement and is
+// mutually exclusive with a nonblank APIKeyEnv. InspectProfile neither derives
+// an ID from a prompt default_profile nor checks credential availability, so an
+// absent or blank named environment variable is not an error.
+//
+// The returned ProfileInspection and all nested mutable values are
+// caller-owned. Filesystem-backed inspection is a point-in-time lookup and
+// does not freeze the profile for a later execution. This method does not load
+// a prompt, render, read artifacts or schemas, admit backend capacity, contact
+// a provider, or generate model output.
+//
+// A nil Engine returns an error matching ErrInvalidConfig. A blank profile ID
+// matches ErrInvalidRequest. An absent exact ID matches ErrProfileNotFound and
+// not ErrProfileLoad. Malformed or unreadable profile data, an unknown backend,
+// or an invalid resolved target matches ErrProfileLoad. Cancellation during
+// profile loading matches ErrProfileLoad while preserving the context error.
+// InspectProfile returns no partial result on error.
+func (e *Engine) InspectProfile(ctx context.Context, profileID string) (*ProfileInspection, error) {
+	if e == nil || e.internal == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
 	}
-	return os.DirFS(dir), filepath.ToSlash(base), nil
+
+	inspection, err := e.internal.InspectProfile(ctx, profileID)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return fromDomainProfileInspection(inspection), nil
 }
 
 // Prepare resolves and renders a prompt request without calling an LLM.
+// It appends any validated RunRequest.AppendedMessages after rendered
+// definition messages in the returned caller-owned snapshot.
 //
 // Prepare selects the prompt and profile, resolves any selected backend and
 // effective execution settings, resolves the output contract, loads and hashes
@@ -439,9 +439,10 @@ func fileSource(name string) (fs.FS, string, error) {
 // applicable. Cancellation is passed to the active collaborator and is
 // reported in the applicable operation category; no general errors.Is
 // relationship to ctx.Err is promised. Prepare returns no partial result on
-// error.
+// error. It validates RunRequest.TransientRetries but makes no model attempts.
 func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, error) {
-	if e == nil || e.runner == nil {
+	ctx = observation.WithoutExecutionState(ctx)
+	if e == nil || e.internal == nil {
 		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
 	}
 
@@ -450,33 +451,34 @@ func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, err
 		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
 
-	prepared, err := e.runner.Prepare(ctx, domainReq)
+	prepared, err := e.internal.Prepare(ctx, domainReq)
 	if err != nil {
 		return nil, mapPublicError(err)
 	}
 	return fromDomainPreparedRun(prepared), nil
 }
 
-// Run prepares a request, invokes the configured LLMClient, and validates the
+// PrepareExecution completely prepares a prompt request without calling the
-// generated output.
+// configured LLMClient or reserving backend admission capacity. Validated
+// RunRequest.AppendedMessages are included in the frozen effective messages.
 //
-// A content-validation failure is a successful run whose
+// The returned opaque handle is bound to this Engine and permits one
-// RunResult.Validation has Status ValidationFailed. An inability to perform
+// [Engine.RunPrepared] invocation. Preparation freezes the selected sources,
-// validation returns an error matching ErrValidation and no partial result.
+// rendered messages, effective settings, inputs, provider structured-output
-// The public Engine does not perform output repair, so validation is
+// metadata, and validation resources needed by that invocation. The handle
-// single-pass even when OutputContract.RepairAttempts is positive.
+// retains a direct RunRequest.APIKey only in private execution state;
+// [PreparedExecution.Details] is credential-redacted.
 //
-// Run can return every error category documented by [Engine.Prepare], plus
+// The context governs preparation only. Cancellation after this method
-// ErrCapacityExceeded and ErrLLMGenerate. ErrCapacityExceeded identifies
+// returns does not invalidate the handle or propagate to RunPrepared.
-// rejection before artifacts, schemas, rendering, or model generation because
+// PrepareExecution returns the same error categories as [Engine.Prepare] and
-// the selected backend's admission capacity is full; it does not match
+// returns no handle on error. A nil Engine returns an error matching
-// ErrInvalidRequest or ErrLLMGenerate. Errors from injected clients remain
+// ErrInvalidConfig. The effective transient retry budget is frozen privately
-// available through errors.Is. Cancellation while waiting for model-generation
+// for RunPrepared and is absent from Details. A supplied Observer is retained
-// capacity matches both ErrLLMGenerate and the context error. Cancellation
+// privately for RunPrepared; preparation itself emits no events.
-// otherwise follows the active collaborator's documented behavior. A nil
+func (e *Engine) PrepareExecution(ctx context.Context, req RunRequest) (*PreparedExecution, error) {
-// Engine returns ErrInvalidConfig. Run returns no partial result on error.
+	ctx = observation.WithoutExecutionState(ctx)
-func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error) {
+	if e == nil || e.internal == nil {
-	if e == nil || e.runner == nil {
 		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
 	}
 
@@ -485,9 +487,149 @@ func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error) {
 		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
 
-	result, err := e.runner.Run(ctx, domainReq)
+	prepared, err := e.internal.PrepareExecution(ctx, domainReq)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return &PreparedExecution{internal: prepared}, nil
+}
+
+// Run prepares a request, invokes the configured LLMClient, and validates the
+// generated output. Each call resolves current sources and composes a fresh,
+// stateless effective prompt with any validated RunRequest.AppendedMessages.
+//
+// A content-validation failure is a successful run whose
+// RunResult.Validation has Status ValidationFailed. When its output contract
+// has a positive repair budget, a failed eligible validation can make bounded
+// additional model calls and stops at the first valid candidate. Exhaustion
+// returns the final failed validation result with cumulative usage and actual
+// repair attempts. An inability to generate or validate returns an error and
+// no partial result.
+//
+// The built-in client retries eligible provider statuses using the effective
+// Config or request budget. With R additional transient attempts and P repair
+// generations, one run can dispatch at most (R+1)×(P+1) Promptkit-controlled
+// model attempts. Each dispatched attempt receives its own generation and
+// HTTP-client timeout; the caller context governs the full operation.
+// Injected LLMClient implementations are not retried automatically.
+// With Observer, Run emits started before facade request conversion or a nil
+// engine check, and finished on every normal return after public error mapping
+// and resource release. The event RunID matches a successful result. Without
+// Observer, result timing retains its existing convention; observed wall
+// duration can be slightly longer because it spans the facade boundary.
+//
+// Run can return every error category documented by [Engine.Prepare], plus
+// ErrCapacityExceeded and ErrLLMGenerate. Built-in HTTP execution failures
+// additionally match ErrTransport; malformed successful responses and nil
+// successful injected responses additionally match ErrMalformedResponse.
+// An engine admission rejection is
+// discoverable as [CapacityError] and still matches ErrCapacityExceeded. It
+// occurs before artifacts, schemas, rendering, or model generation because the
+// selected backend's admission capacity is full; it does not match
+// ErrInvalidRequest or ErrLLMGenerate. A built-in OpenAI-compatible non-2xx
+// response is discoverable as [GenerationError]. Eligible-status exhaustion
+// with a positive budget is discoverable as [RetryExhaustedError]. Injected
+// errors remain available through errors.Is. Cancellation while waiting for
+// model-generation capacity matches both ErrLLMGenerate and the context error.
+// Cancellation otherwise follows the active collaborator's documented
+// behavior. A nil Engine returns ErrInvalidConfig. Run returns no partial
+// result on error.
+func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error) {
+	ctx = observation.WithoutExecutionState(ctx)
+	if req.Observer == nil {
+		return e.run(ctx, req)
+	}
+	recorder, idErr := observation.New(func(event domain.ExecutionEvent) { req.Observer(fromDomainExecutionEvent(event)) })
+	recorder.Start()
+	if idErr != nil {
+		err := fmt.Errorf("failed to create run id: %w", idErr)
+		terminal := observedFinish(nil, err)
+		recorder.Finish(observation.FinishData{Outcome: terminal.Outcome, Failure: terminal.Failure})
+		return nil, err
+	}
+	result, err := e.run(observation.WithRecorder(ctx, recorder), req)
+	terminal := observedFinish(result, err)
+	recorder.Finish(observation.FinishData{Outcome: terminal.Outcome, Failure: terminal.Failure,
+		StatusCode: terminal.StatusCode, RetriesExhausted: terminal.RetriesExhausted})
+	return result, err
+}
+
+func (e *Engine) run(ctx context.Context, req RunRequest) (*RunResult, error) {
+	if e == nil || e.internal == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
+	}
+
+	domainReq, err := toDomainRunRequest(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	result, err := e.internal.Run(ctx, domainReq)
 	if err != nil {
 		return nil, mapPublicError(err)
 	}
 	return fromDomainRunResult(result), nil
 }
+
+// RunPrepared atomically claims and executes a handle created by
+// [Engine.PrepareExecution].
+//
+// A valid owning-Engine invocation consumes the handle's one attempt before
+// credential revalidation, backend admission, generation, or validation.
+// Cancellation, capacity rejection, generation failure, operational
+// validation failure, and success all leave the handle unusable. A nil,
+// zero-value, foreign-Engine, discarded, claimed, or used handle returns an
+// error matching ErrInvalidRequest; a nil Engine returns ErrInvalidConfig and
+// does not claim the handle.
+//
+// The supplied context governs this execution attempt independently of the
+// preparation context. It covers credential revalidation, admission,
+// generation, validation, and any bounded output repair. Result timing begins
+// after the claim and excludes preparation and consumer-held delay.
+//
+// RunPrepared can return ErrInvalidRequest, ErrAPIKeyEnvMissing,
+// ErrCapacityExceeded, ErrLLMGenerate, or ErrValidation as applicable while
+// preserving documented collaborator and context identities. An engine
+// admission rejection is discoverable as [CapacityError] and still matches
+// ErrCapacityExceeded. A built-in OpenAI-compatible non-2xx response is
+// discoverable as [GenerationError]. Built-in HTTP execution failures match
+// ErrTransport; malformed successful responses and nil successful injected
+// responses match ErrMalformedResponse. Positive-budget eligible-status
+// exhaustion is discoverable as [RetryExhaustedError]. A completed
+// content-validation rejection, including repair exhaustion, is returned in
+// RunResult, not as an operational
+// error. RunPrepared uses the retry choice frozen during PrepareExecution. An
+// operational error returns no partial RunResult. A retained Observer begins
+// only after a successful atomic claim, with identities from the frozen
+// snapshot, and receives a terminal event on normal return after public error
+// mapping. Invalid, foreign, discarded, or reused handles emit no events.
+// Callbacks run outside the handle and capacity locks; panics propagate and
+// do not promise a terminal event.
+func (e *Engine) RunPrepared(ctx context.Context, prepared *PreparedExecution) (*RunResult, error) {
+	ctx = observation.WithoutExecutionState(ctx)
+	if e == nil || e.internal == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
+	}
+
+	var internal *usecase.PreparedExecution
+	if prepared != nil {
+		internal = prepared.internal
+	}
+	var recorder *observation.Recorder
+	result, err := e.internal.RunPrepared(observation.WithClaimSink(ctx, &recorder), internal)
+	if err != nil {
+		mapped := mapPublicError(err)
+		if recorder != nil {
+			terminal := observedFinish(nil, mapped)
+			recorder.Finish(observation.FinishData{Outcome: terminal.Outcome, Failure: terminal.Failure,
+				StatusCode: terminal.StatusCode, RetriesExhausted: terminal.RetriesExhausted})
+		}
+		return nil, mapped
+	}
+	publicResult := fromDomainRunResult(result)
+	if recorder != nil {
+		terminal := observedFinish(publicResult, nil)
+		recorder.Finish(observation.FinishData{Outcome: terminal.Outcome})
+	}
+	return publicResult, nil
+}

errors.go

@@ -3,8 +3,11 @@ package promptkit
 import (
 	"errors"
 	"fmt"
+	"strings"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/generation"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
 	"gitea.maximumdirect.net/eric/promptkit/internal/promptdef"
 	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
@@ -14,11 +17,47 @@ func mapPublicError(err error) error {
 	if err == nil {
 		return nil
 	}
+	var internalCapacityError *usecase.CapacityError
+	if errors.As(err, &internalCapacityError) && internalCapacityError != nil &&
+		strings.TrimSpace(internalCapacityError.BackendID) != "" {
+		return &CapacityError{BackendID: internalCapacityError.BackendID}
+	}
+	var internalExhausted *generation.ExhaustedError
+	if errors.As(err, &internalExhausted) && internalExhausted != nil {
+		mapped := newRetryExhaustedError(internalExhausted.Attempts(), mapPublicError(internalExhausted.Unwrap()))
+		if errors.Is(err, usecase.ErrValidation) && !errors.Is(mapped, ErrValidation) {
+			return fmt.Errorf("%w: %w", ErrValidation, mapped)
+		}
+		if publicErr := publicErrorFor(err); publicErr != nil && !errors.Is(mapped, publicErr) {
+			return fmt.Errorf("%w: %w", publicErr, mapped)
+		}
+		return mapped
+	}
 	publicErr := publicErrorFor(err)
+	var providerHTTPError *llm.ProviderHTTPError
+	if errors.As(err, &providerHTTPError) && providerHTTPError != nil {
+		generationErr := newGenerationError(
+			providerHTTPError.StatusCode(),
+			providerHTTPError.ProviderCode(),
+			providerHTTPError.ProviderType(),
+			providerHTTPError.ProviderMessage(),
+		)
+		if publicErr != nil && !errors.Is(publicErr, ErrLLMGenerate) {
+			return fmt.Errorf("%w: %w", publicErr, generationErr)
+		}
+		return generationErr
+	}
 	if publicErr == nil {
 		return err
 	}
-	return fmt.Errorf("%w: %w", publicErr, err)
+	switch {
+	case errors.Is(err, llm.ErrTransport):
+		return fmt.Errorf("%w: %w: %w", publicErr, ErrTransport, err)
+	case errors.Is(err, llm.ErrMalformedResponse):
+		return fmt.Errorf("%w: %w: %w", publicErr, ErrMalformedResponse, err)
+	default:
+		return fmt.Errorf("%w: %w", publicErr, err)
+	}
 }
 
 func publicErrorFor(err error) error {

errors_internal_test.go

@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"testing"
 
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
 	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
 )
 
@@ -20,3 +21,62 @@ func TestMapPublicErrorPreservesGenerationCancellation(t *testing.T) {
 		t.Fatalf("mapped error=%v, want context.Canceled", err)
 	}
 }
+
+func TestMapPublicErrorTranslatesCapacityError(t *testing.T) {
+	internalErr := &usecase.CapacityError{BackendID: "limited"}
+
+	err := mapPublicError(internalErr)
+	var publicErr *CapacityError
+	if !errors.As(err, &publicErr) || publicErr == nil {
+		t.Fatalf("mapped error=%v, want public CapacityError", err)
+	}
+	if publicErr.BackendID != "limited" {
+		t.Fatalf("mapped backend ID=%q, want limited", publicErr.BackendID)
+	}
+	if !errors.Is(err, ErrCapacityExceeded) {
+		t.Fatalf("mapped error=%v, want ErrCapacityExceeded", err)
+	}
+	if errors.Is(err, ErrInvalidRequest) || errors.Is(err, ErrLLMGenerate) {
+		t.Fatalf("mapped capacity error has an unrelated category: %v", err)
+	}
+	var leakedInternalErr *usecase.CapacityError
+	if errors.As(err, &leakedInternalErr) {
+		t.Fatalf("mapped error exposes internal CapacityError: %v", err)
+	}
+
+	internalErr.BackendID = "changed"
+	if publicErr.BackendID != "limited" {
+		t.Fatalf("mapped backend ID changed with source error: %q", publicErr.BackendID)
+	}
+}
+
+func TestMapPublicErrorPreservesValidationAroundGenerationError(t *testing.T) {
+	internalErr := fmt.Errorf(
+		"%w: %w",
+		usecase.ErrValidation,
+		&llm.ProviderHTTPError{},
+	)
+
+	err := mapPublicError(internalErr)
+	if !errors.Is(err, ErrValidation) {
+		t.Fatalf("mapped error=%v, want ErrValidation", err)
+	}
+	if !errors.Is(err, ErrLLMGenerate) {
+		t.Fatalf("mapped error=%v, want ErrLLMGenerate", err)
+	}
+	var generationErr *GenerationError
+	if !errors.As(err, &generationErr) || generationErr == nil {
+		t.Fatalf("mapped error=%v, want GenerationError", err)
+	}
+	var leakedInternalErr *llm.ProviderHTTPError
+	if errors.As(err, &leakedInternalErr) {
+		t.Fatalf("mapped error exposes internal ProviderHTTPError: %v", err)
+	}
+}
+
+func TestMapPublicErrorDoesNotTreatRequestEncodingAsTransport(t *testing.T) {
+	err := mapPublicError(fmt.Errorf("%w: %w", usecase.ErrLLMGenerate, llm.ErrRequestFailed))
+	if !errors.Is(err, ErrLLMGenerate) || errors.Is(err, ErrTransport) {
+		t.Fatalf("request encoding category = %v", err)
+	}
+}

examples/go-library/observe/main.go

@@ -0,0 +1,157 @@
+package main
+
+import (
+	"context"
+	"embed"
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"os"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+//go:embed prompts/*.yaml schemas/*.json
+var assets embed.FS
+
+type scriptedTransport struct {
+	proposalCalls int
+}
+
+func (t *scriptedTransport) RoundTrip(req *http.Request) (*http.Response, error) {
+	if err := req.Context().Err(); err != nil {
+		return nil, err
+	}
+	var request struct {
+		Model string `json:"model"`
+	}
+	if err := json.NewDecoder(req.Body).Decode(&request); err != nil {
+		return nil, err
+	}
+	switch request.Model {
+	case "proposal-model":
+		t.proposalCalls++
+		if t.proposalCalls == 1 {
+			return httpResponse(http.StatusServiceUnavailable, `{"error":{}}`), nil
+		}
+		return modelResponse(`{"title":"Queue cleanup","risk":"low"}`, 10, 8), nil
+	case "assessment-model":
+		return modelResponse(`{"recommendation":"accept","confidence":0.9}`, 7, 5), nil
+	default:
+		return nil, fmt.Errorf("unexpected model in offline transport")
+	}
+}
+
+func modelResponse(content string, promptTokens, completionTokens int) *http.Response {
+	body, _ := json.Marshal(map[string]any{
+		"choices": []any{map[string]any{"message": map[string]string{"content": content}}},
+		"usage":   map[string]int{"prompt_tokens": promptTokens, "completion_tokens": completionTokens},
+	})
+	return httpResponse(http.StatusOK, string(body))
+}
+
+func httpResponse(status int, body string) *http.Response {
+	return &http.Response{StatusCode: status, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(body))}
+}
+
+type eventLog struct{ events []promptkit.ExecutionEvent }
+
+func (l *eventLog) record(event promptkit.ExecutionEvent) { l.events = append(l.events, event) }
+
+func (l *eventLog) terminal() (promptkit.ExecutionEvent, error) {
+	if len(l.events) == 0 || l.events[len(l.events)-1].Kind != promptkit.ExecutionFinished {
+		return promptkit.ExecutionEvent{}, fmt.Errorf("execution did not finish")
+	}
+	return l.events[len(l.events)-1], nil
+}
+
+type summary struct {
+	ProposalAttempts           uint64 `json:"proposal_attempts"`
+	ProposalTransientRetries   uint64 `json:"proposal_transient_retries"`
+	ProposalUsageComplete      bool   `json:"proposal_usage_complete"`
+	AssessmentAttempts         uint64 `json:"assessment_attempts"`
+	AssessmentTransientRetries uint64 `json:"assessment_transient_retries"`
+	AssessmentOutcome          string `json:"assessment_outcome"`
+}
+
+func run() (summary, error) {
+	transport := &scriptedTransport{}
+	engine, err := promptkit.NewEngine(promptkit.Config{HTTPClient: &http.Client{Transport: transport}},
+		promptkit.WithPromptFS(assets, "prompts"),
+		promptkit.WithSchemaFS(assets, "schemas"),
+		promptkit.WithBackend(promptkit.Backend{ID: "offline-backend", Endpoint: "https://example.invalid/v1", ConcurrencyLimit: 1}),
+		promptkit.WithProfiles(
+			promptkit.Profile{ID: "proposal-profile", BackendID: "offline-backend", Model: "proposal-model"},
+			promptkit.Profile{ID: "assessment-profile", BackendID: "offline-backend", Model: "assessment-model"},
+		))
+	if err != nil {
+		return summary{}, err
+	}
+
+	var proposalLog eventLog
+	proposal, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID: "example.proposal", Inputs: map[string]promptkit.ArtifactRef{"idea": promptkit.Inline("Reduce queue cleanup toil.")}, Observer: proposalLog.record,
+	})
+	if err != nil {
+		return summary{}, err
+	}
+	if !proposal.Validation.IsValid {
+		return summary{}, fmt.Errorf("proposal failed structural validation")
+	}
+	proposalTerminal, err := proposalLog.terminal()
+	if err != nil {
+		return summary{}, err
+	}
+	var assessmentLog eventLog
+	assessment, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID: "example.assessment", Inputs: map[string]promptkit.ArtifactRef{"proposal": promptkit.Inline(proposal.RawOutput)}, Observer: assessmentLog.record,
+	})
+	if err != nil {
+		return summary{}, err
+	}
+	if !assessment.Validation.IsValid {
+		return summary{}, fmt.Errorf("assessment failed structural validation")
+	}
+	assessmentTerminal, err := assessmentLog.terminal()
+	if err != nil {
+		return summary{}, err
+	}
+	var proposalValue struct {
+		Risk string `json:"risk"`
+	}
+	var assessmentValue struct {
+		Recommendation string `json:"recommendation"`
+	}
+	if err := json.Unmarshal([]byte(proposal.RawOutput), &proposalValue); err != nil {
+		return summary{}, err
+	}
+	if err := json.Unmarshal([]byte(assessment.RawOutput), &assessmentValue); err != nil {
+		return summary{}, err
+	}
+	decision := "rejected"
+	if proposalValue.Risk == "low" && assessmentValue.Recommendation == "accept" {
+		decision = "accepted"
+	}
+	return summary{
+		ProposalAttempts:           proposalTerminal.Attempts,
+		ProposalTransientRetries:   proposalTerminal.TransientRetries,
+		ProposalUsageComplete:      proposalTerminal.Usage.PromptTokens.Complete,
+		AssessmentAttempts:         assessmentTerminal.Attempts,
+		AssessmentTransientRetries: assessmentTerminal.TransientRetries,
+		AssessmentOutcome:          decision,
+	}, nil
+}
+
+func main() {
+	value, err := run()
+	if err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+	if err := json.NewEncoder(os.Stdout).Encode(value); err != nil {
+		fmt.Fprintln(os.Stderr, err)
+		os.Exit(1)
+	}
+}

examples/go-library/observe/prompts/assessment.yaml

@@ -0,0 +1,17 @@
+id: example.assessment
+version: "1.0.0"
+default_profile: assessment-profile
+description: Assess a structured proposal independently of schema validity.
+inputs:
+  - name: proposal
+    required: true
+    content_type: application/json
+messages:
+  - role: system
+    content: Return a concise assessment as JSON.
+  - role: user
+    content: '{{input "proposal"}}'
+output:
+  format: json
+  validation_mode: json_schema
+  schema_path: assessment.schema.json

examples/go-library/observe/prompts/proposal.yaml

@@ -0,0 +1,17 @@
+id: example.proposal
+version: "1.0.0"
+default_profile: proposal-profile
+description: Propose a bounded improvement from a synthetic idea.
+inputs:
+  - name: idea
+    required: true
+    content_type: text/plain
+messages:
+  - role: system
+    content: Return a concise proposal as JSON.
+  - role: user
+    content: '{{input "idea"}}'
+output:
+  format: json
+  validation_mode: json_schema
+  schema_path: proposal.schema.json

examples/go-library/observe/schemas/assessment.schema.json

@@ -0,0 +1,10 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "type": "object",
+  "required": ["recommendation", "confidence"],
+  "properties": {
+    "recommendation": {"type": "string"},
+    "confidence": {"type": "number"}
+  },
+  "additionalProperties": false
+}

examples/go-library/observe/schemas/proposal.schema.json

@@ -0,0 +1,10 @@
+{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "type": "object",
+  "required": ["title", "risk"],
+  "properties": {
+    "title": {"type": "string"},
+    "risk": {"type": "string"}
+  },
+  "additionalProperties": false
+}

formatting.go

@@ -18,7 +18,7 @@ func (r RunRequest) GoString() string {
 
 func (r RunRequest) redactedString() string {
 	return fmt.Sprintf(
-		"promptkit.RunRequest{PromptID:%q PromptVersion:%q ProfileID:%q APIKeySet:%t Inputs:%d Vars:%d ExecutionSet:%t ValidationSet:%t}",
+		"promptkit.RunRequest{PromptID:%q PromptVersion:%q ProfileID:%q APIKeySet:%t Inputs:%d Vars:%d ExecutionSet:%t ValidationSet:%t AppendedMessages:%d}",
 		r.PromptID,
 		r.PromptVersion,
 		r.ProfileID,
@@ -27,6 +27,7 @@ func (r RunRequest) redactedString() string {
 		len(r.Vars),
 		r.Execution != nil,
 		r.Validation != nil,
+		len(r.AppendedMessages),
 	)
 }
 

generation_error.go

@@ -0,0 +1,87 @@
+package promptkit
+
+import "fmt"
+
+// GenerationError reports a non-2xx response from Promptkit's built-in
+// OpenAI-compatible client during [Engine.Run] or [Engine.RunPrepared].
+//
+// Engine-produced values are immutable, caller-owned values. Use errors.Is to
+// match [ErrLLMGenerate] and errors.As with a *GenerationError target to obtain
+// this type. The four provider accessors expose untrusted provider-controlled
+// values that can contain sensitive request or schema fragments. Applications
+// must apply their own disclosure policy before logging, displaying, or
+// returning them to another caller.
+//
+// Accessors, Error, GoString, and Unwrap are safe on a nil receiver and a zero
+// value. Default and Go-syntax formatting deliberately redact provider details.
+// GenerationError has no stable JSON representation.
+type GenerationError struct {
+	statusCode      int
+	providerCode    string
+	providerType    string
+	providerMessage string
+}
+
+func newGenerationError(statusCode int, providerCode, providerType, providerMessage string) *GenerationError {
+	return &GenerationError{
+		statusCode:      statusCode,
+		providerCode:    providerCode,
+		providerType:    providerType,
+		providerMessage: providerMessage,
+	}
+}
+
+// StatusCode returns the received provider HTTP status code, or zero for a nil
+// receiver or zero value.
+func (e *GenerationError) StatusCode() int {
+	if e == nil {
+		return 0
+	}
+	return e.statusCode
+}
+
+// ProviderCode returns the normalized provider error code, if present. Its
+// value is untrusted and may contain sensitive data.
+func (e *GenerationError) ProviderCode() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerCode
+}
+
+// ProviderType returns the normalized provider error type, if present. Its
+// value is untrusted and may contain sensitive data.
+func (e *GenerationError) ProviderType() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerType
+}
+
+// ProviderMessage returns the bounded normalized provider diagnostic, if
+// present. Its value is untrusted and may contain sensitive data.
+func (e *GenerationError) ProviderMessage() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerMessage
+}
+
+// Error returns a redacted diagnostic that is not a parsing contract.
+func (e *GenerationError) Error() string {
+	if e == nil || e.statusCode == 0 {
+		return ErrLLMGenerate.Error()
+	}
+	return fmt.Sprintf("%s: provider returned HTTP status %d", ErrLLMGenerate, e.statusCode)
+}
+
+// GoString returns the same redacted diagnostic as Error.
+func (e *GenerationError) GoString() string {
+	return e.Error()
+}
+
+// Unwrap returns ErrLLMGenerate. It is safe to call on a nil receiver or zero
+// value.
+func (e *GenerationError) Unwrap() error {
+	return ErrLLMGenerate
+}

generation_error_contract_test.go

@@ -0,0 +1,235 @@
+package promptkit_test
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+func TestBuiltInGenerationError(t *testing.T) {
+	const (
+		codeMarker    = "provider-code-marker"
+		typeMarker    = "provider-type-marker"
+		messageMarker = "provider-message-marker"
+	)
+	engine := newBuiltInGenerationErrorEngine(t, http.StatusUnprocessableEntity,
+		`{"error":{"code":"`+codeMarker+`","type":"`+typeMarker+`","message":"`+messageMarker+`"}}`)
+
+	result, err := engine.Run(context.Background(), generationErrorRunRequest())
+	if result != nil {
+		t.Fatalf("Run result = %#v, want nil", result)
+	}
+	assertGenerationError(t, err, http.StatusUnprocessableEntity, codeMarker, typeMarker, messageMarker)
+
+	preparedEngine := newBuiltInGenerationErrorEngine(t, http.StatusServiceUnavailable, `{"error":{}}`)
+	prepared, err := preparedEngine.PrepareExecution(context.Background(), generationErrorRunRequest())
+	if err != nil {
+		t.Fatalf("PrepareExecution: %v", err)
+	}
+	result, err = preparedEngine.RunPrepared(context.Background(), prepared)
+	if result != nil {
+		t.Fatalf("RunPrepared result = %#v, want nil", result)
+	}
+	assertGenerationError(t, err, http.StatusServiceUnavailable, "", "", "")
+}
+
+func TestBuiltInGenerationErrorWithAppendedMessages(t *testing.T) {
+	const messageMarker = "combined-request-provider-marker"
+	engine := newBuiltInGenerationErrorEngine(t, http.StatusUnprocessableEntity,
+		`{"error":{"message":"`+messageMarker+`"}}`)
+	request := generationErrorRunRequest()
+	request.AppendedMessages = []promptkit.RenderedMessage{
+		{Role: promptkit.RoleAssistant, Content: "previous response"},
+		{Role: promptkit.RoleUser, Content: "consumer correction"},
+	}
+
+	result, err := engine.Run(context.Background(), request)
+	if result != nil {
+		t.Fatalf("Run result = %#v, want nil", result)
+	}
+	assertGenerationError(t, err, http.StatusUnprocessableEntity, "", "", messageMarker)
+}
+
+func TestBuiltInRepairGenerationError(t *testing.T) {
+	const (
+		codeMarker    = "repair-code-marker"
+		typeMarker    = "repair-type-marker"
+		messageMarker = "repair-message-marker"
+	)
+	calls := 0
+	config := contractConfig(frameworkSchemaDir)
+	config.HTTPClient = &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+		calls++
+		if calls == 1 {
+			body := `{"choices":[{"message":{"content":"not-json"}}]}`
+			return &http.Response{StatusCode: http.StatusOK, ContentLength: int64(len(body)), Body: io.NopCloser(strings.NewReader(body))}, nil
+		}
+		body := `{"error":{"code":"` + codeMarker + `","type":"` + typeMarker + `","message":"` + messageMarker + `"}}`
+		return &http.Response{StatusCode: http.StatusUnprocessableEntity, ContentLength: int64(len(body)), Body: io.NopCloser(strings.NewReader(body))}, nil
+	})}
+	engine, err := promptkit.NewEngine(config)
+	if err != nil {
+		t.Fatalf("NewEngine: %v", err)
+	}
+	req := generationErrorRunRequest()
+	req.Validation = &promptkit.OutputContract{
+		Format:         promptkit.FormatJSON,
+		ValidationMode: promptkit.ValidationJSON,
+		RepairAttempts: 1,
+	}
+
+	result, err := engine.Run(context.Background(), req)
+	if result != nil {
+		t.Fatalf("Run result = %#v, want nil", result)
+	}
+	if calls != 2 {
+		t.Fatalf("provider calls = %d, want 2", calls)
+	}
+	assertGenerationError(t, err, http.StatusUnprocessableEntity, codeMarker, typeMarker, messageMarker)
+}
+
+func TestBuiltInGenerationFailureCategories(t *testing.T) {
+	transportCause := errors.New("private transport diagnostic")
+	config := contractConfig(frameworkSchemaDir)
+	config.HTTPClient = &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+		return nil, transportCause
+	})}
+	engine, err := promptkit.NewEngine(config)
+	if err != nil {
+		t.Fatal(err)
+	}
+	result, err := engine.Run(context.Background(), generationErrorRunRequest())
+	if result != nil || !errors.Is(err, promptkit.ErrLLMGenerate) || !errors.Is(err, promptkit.ErrTransport) || !errors.Is(err, transportCause) {
+		t.Fatalf("transport result=%#v, error=%v", result, err)
+	}
+	if errors.Is(err, promptkit.ErrMalformedResponse) || strings.Contains(fmt.Sprintf("%#v", err), transportCause.Error()) {
+		t.Fatalf("transport category or formatting = %v", err)
+	}
+
+	for _, tc := range []struct {
+		name string
+		body string
+	}{
+		{name: "invalid framing", body: `{"choices":[]}{}`},
+		{name: "missing choices", body: `{"choices":[]}`},
+		{name: "missing content", body: `{"choices":[{"message":{}}]}`},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			engine := newBuiltInGenerationErrorEngine(t, http.StatusOK, tc.body)
+			result, err := engine.Run(context.Background(), generationErrorRunRequest())
+			if result != nil || !errors.Is(err, promptkit.ErrLLMGenerate) || !errors.Is(err, promptkit.ErrMalformedResponse) || errors.Is(err, promptkit.ErrTransport) {
+				t.Fatalf("malformed result=%#v, error=%v", result, err)
+			}
+		})
+	}
+}
+
+func TestBuiltInGenerationPreservesContextIdentity(t *testing.T) {
+	for _, tc := range []struct {
+		name  string
+		cause error
+	}{
+		{name: "canceled", cause: context.Canceled},
+		{name: "deadline", cause: context.DeadlineExceeded},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			config := contractConfig(frameworkSchemaDir)
+			config.HTTPClient = &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+				return nil, tc.cause
+			})}
+			engine, err := promptkit.NewEngine(config)
+			if err != nil {
+				t.Fatal(err)
+			}
+			result, err := engine.Run(context.Background(), generationErrorRunRequest())
+			if result != nil || !errors.Is(err, tc.cause) || !errors.Is(err, promptkit.ErrLLMGenerate) || !errors.Is(err, promptkit.ErrTransport) {
+				t.Fatalf("context result=%#v, error=%v", result, err)
+			}
+		})
+	}
+}
+
+func TestInjectedGenerationFailureCategories(t *testing.T) {
+	for _, tc := range []struct {
+		name      string
+		client    *fakeLLMClient
+		malformed bool
+	}{
+		{name: "nil response", client: &fakeLLMClient{}, malformed: true},
+		{name: "arbitrary error", client: &fakeLLMClient{err: errors.New("injected failure")}},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(tc.client))
+			result, err := engine.Run(context.Background(), generationErrorRunRequest())
+			if result != nil || !errors.Is(err, promptkit.ErrLLMGenerate) || errors.Is(err, promptkit.ErrTransport) || errors.Is(err, promptkit.ErrMalformedResponse) != tc.malformed {
+				t.Fatalf("injected result=%#v, error=%v", result, err)
+			}
+			if tc.client.err != nil && !errors.Is(err, tc.client.err) {
+				t.Fatalf("injected cause lost: %v", err)
+			}
+		})
+	}
+}
+
+func assertGenerationError(t *testing.T, err error, statusCode int, code, providerType, message string) {
+	t.Helper()
+
+	if !errors.Is(err, promptkit.ErrLLMGenerate) {
+		t.Fatalf("errors.Is(%v, ErrLLMGenerate) = false", err)
+	}
+	var generationErr *promptkit.GenerationError
+	if !errors.As(err, &generationErr) || generationErr == nil {
+		t.Fatalf("error = %T, want *GenerationError", err)
+	}
+	if generationErr.StatusCode() != statusCode || generationErr.ProviderCode() != code || generationErr.ProviderType() != providerType || generationErr.ProviderMessage() != message {
+		t.Fatalf("GenerationError = %#v", generationErr)
+	}
+
+	wantFormatted := fmt.Sprintf("failed to generate output: provider returned HTTP status %d", statusCode)
+	for _, rendered := range []string{fmt.Sprintf("%v", generationErr), fmt.Sprintf("%+v", generationErr), fmt.Sprintf("%#v", generationErr)} {
+		if rendered != wantFormatted {
+			t.Fatalf("formatted error = %q, want %q", rendered, wantFormatted)
+		}
+		for _, marker := range []string{code, providerType, message} {
+			if marker != "" && strings.Contains(rendered, marker) {
+				t.Fatalf("formatted error exposed provider marker %q: %q", marker, rendered)
+			}
+		}
+	}
+}
+
+func newBuiltInGenerationErrorEngine(t *testing.T, statusCode int, body string) *promptkit.Engine {
+	t.Helper()
+
+	config := contractConfig(frameworkSchemaDir)
+	noRetries := 0
+	config.TransientRetries = &noRetries
+	config.HTTPClient = &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+		return &http.Response{
+			StatusCode:    statusCode,
+			ContentLength: int64(len(body)),
+			Body:          io.NopCloser(strings.NewReader(body)),
+		}, nil
+	})}
+	engine, err := promptkit.NewEngine(config)
+	if err != nil {
+		t.Fatalf("NewEngine: %v", err)
+	}
+	return engine
+}
+
+func generationErrorRunRequest() promptkit.RunRequest {
+	return promptkit.RunRequest{
+		PromptID: frameworkMarkdownSummaryPromptID,
+		Inputs: map[string]promptkit.ArtifactRef{
+			"transcript": promptkit.Inline("Rin opens the gate."),
+			"glossary":   promptkit.Inline("gate: A guarded passage."),
+		},
+	}
+}

generation_error_internal_test.go

@@ -0,0 +1,32 @@
+package promptkit
+
+import (
+	"errors"
+	"fmt"
+	"testing"
+)
+
+func TestGenerationErrorNilAndZeroValue(t *testing.T) {
+	var nilError *GenerationError
+	zeroError := &GenerationError{}
+
+	for name, err := range map[string]*GenerationError{
+		"nil":  nilError,
+		"zero": zeroError,
+	} {
+		t.Run(name, func(t *testing.T) {
+			if err.StatusCode() != 0 || err.ProviderCode() != "" || err.ProviderType() != "" || err.ProviderMessage() != "" {
+				t.Fatalf("accessors returned provider details: %#v", err)
+			}
+			if err.Error() != "failed to generate output" || err.GoString() != "failed to generate output" {
+				t.Fatalf("redacted formatting = (%q, %q)", err.Error(), err.GoString())
+			}
+			if fmt.Sprintf("%v", err) != "failed to generate output" || fmt.Sprintf("%#v", err) != "failed to generate output" {
+				t.Fatalf("formatted error = (%q, %q)", fmt.Sprintf("%v", err), fmt.Sprintf("%#v", err))
+			}
+			if !errors.Is(err, ErrLLMGenerate) {
+				t.Fatalf("errors.Is(%v, ErrLLMGenerate) = false", err)
+			}
+		})
+	}
+}

go.mod

@@ -3,6 +3,8 @@ module gitea.maximumdirect.net/eric/promptkit
 go 1.25.5
 
 require (
+	gitea.maximumdirect.net/eric/promptkit-backend-openrouter/v2 v2.20260922.1
+	gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/v2 v2.20260922.0
 	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
 	gopkg.in/yaml.v3 v3.0.1
 )

go.sum

@@ -1,3 +1,7 @@
+gitea.maximumdirect.net/eric/promptkit-backend-openrouter/v2 v2.20260922.1 h1:S2/Q/JDE+m8Ca3pDhAjVbfYrdlEhs9blj0STgNopHo0=
+gitea.maximumdirect.net/eric/promptkit-backend-openrouter/v2 v2.20260922.1/go.mod h1:a3VBbh9oaGXD6ruxhtwHj+AWW6DY/SSyby1QtMmYSUo=
+gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/v2 v2.20260922.0 h1:Y0EXc3eh12eKW2NrOviVqdk3oBo6UUjopt1YAa1B3Oc=
+gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/v2 v2.20260922.0/go.mod h1:mtgD1E0SGm0e/I8Ffhz+AYX7GN7RKUWJXehNFIKHs4Y=
 github.com/dlclark/regexp2 v1.11.0 h1:G/nrcoOa7ZXlpoa/91N3X7mM3r8eIlMBBJZvsz/mxKI=
 github.com/dlclark/regexp2 v1.11.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=

internal/artifact/reader.go

@@ -16,10 +16,12 @@ import (
 
 var (
 	ErrUnsupportedRefType = errors.New("unsupported artifact reference type")
-	ErrMissingInlineBody  = errors.New("missing body for inline artifact")
 	ErrMissingFilePath    = errors.New("missing file path for file artifact")
+	ErrUnsupportedFile    = errors.New("file artifact path is not a regular file")
 )
 
+const fileReadChunkSize = 64 * 1024
+
 // Reader resolves artifact references into actual artifacts.
 type Reader interface {
 	Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error)
@@ -34,7 +36,7 @@ type CompositeReader struct {
 func NewCompositeReader() Reader {
 	return &CompositeReader{
 		inlineReader: &inlineReader{},
-		fileReader:   &fileReader{},
+		fileReader:   &fileReader{open: openArtifactFile},
 	}
 }
 
@@ -64,10 +66,6 @@ func (r *inlineReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domai
 	default:
 	}
 
-	if ref.Body == "" {
-		return nil, ErrMissingInlineBody
-	}
-
 	body := []byte(ref.Body)
 	return &domain.Artifact{
 		ContentType: defaults.ContentTypeTextPlain,
@@ -78,7 +76,15 @@ func (r *inlineReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domai
 	}, nil
 }
 
-type fileReader struct{}
+type artifactFile interface {
+	Read([]byte) (int, error)
+	Stat() (os.FileInfo, error)
+	Close() error
+}
+
+type fileReader struct {
+	open func(string) (artifactFile, error)
+}
 
 func (r *fileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
 	select {
@@ -91,25 +97,71 @@ func (r *fileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.
 		return nil, ErrMissingFilePath
 	}
 
-	return readFileArtifact(ref.URI)
+	return readFileArtifact(ctx, ref.URI, r.open)
 }
 
-func readFileArtifact(path string) (*domain.Artifact, error) {
+func openArtifactFile(path string) (artifactFile, error) {
-	file, err := os.Open(path)
+	return os.Open(path)
+}
+
+func readFileArtifact(ctx context.Context, path string, open func(string) (artifactFile, error)) (*domain.Artifact, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	info, err := os.Stat(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+	}
+	if !info.Mode().IsRegular() {
+		return nil, fmt.Errorf("%w: %s", ErrUnsupportedFile, path)
+	}
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+
+	file, err := open(path)
 	if err != nil {
 		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
 	}
 	defer file.Close()
 
-	data, err := io.ReadAll(file)
+	openedInfo, err := file.Stat()
 	if err != nil {
-		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+		return nil, fmt.Errorf("failed to inspect opened file %s: %w", path, err)
+	}
+	if !openedInfo.Mode().IsRegular() {
+		return nil, fmt.Errorf("%w: %s", ErrUnsupportedFile, path)
+	}
+
+	data := make([]byte, 0)
+	chunk := make([]byte, fileReadChunkSize)
+	for {
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		n, readErr := file.Read(chunk)
+		if n > 0 {
+			data = append(data, chunk[:n]...)
+		}
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		if errors.Is(readErr, io.EOF) {
+			break
+		}
+		if readErr != nil {
+			return nil, fmt.Errorf("failed to read file %s: %w", path, readErr)
+		}
 	}
 
 	contentType := mime.TypeByExtension(filepath.Ext(path))
 	if contentType == "" {
 		contentType = defaults.ContentTypeTextPlain
 	}
+	hash := fmt.Sprintf("%x", sha256.Sum256(data))
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
 
 	return &domain.Artifact{
 		Name:        filepath.Base(path),
@@ -117,6 +169,6 @@ func readFileArtifact(path string) (*domain.Artifact, error) {
 		Body:        data,
 		URI:         path,
 		Size:        int64(len(data)),
-		Hash:        fmt.Sprintf("%x", sha256.Sum256(data)),
+		Hash:        hash,
 	}, nil
 }

internal/artifact/reader_fifo_linux_test.go

@@ -0,0 +1,43 @@
+//go:build linux
+
+package artifact
+
+import (
+	"context"
+	"errors"
+	"path/filepath"
+	"syscall"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestFileReaderRejectsFIFOBeforeOpen(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "artifact.fifo")
+	if err := syscall.Mkfifo(path, 0o600); err != nil {
+		t.Fatalf("create fifo: %v", err)
+	}
+
+	type result struct {
+		artifact *domain.Artifact
+		err      error
+	}
+	done := make(chan result, 1)
+	go func() {
+		artifact, err := NewCompositeReader().Read(context.Background(), domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  path,
+		})
+		done <- result{artifact: artifact, err: err}
+	}()
+
+	select {
+	case got := <-done:
+		if got.artifact != nil || !errors.Is(got.err, ErrUnsupportedFile) {
+			t.Fatalf("artifact=%#v err=%v, want nil/ErrUnsupportedFile", got.artifact, got.err)
+		}
+	case <-time.After(time.Second):
+		t.Fatal("FIFO read blocked instead of rejecting the non-regular file")
+	}
+}

internal/artifact/reader_test.go

@@ -1,6 +1,7 @@
 package artifact
 
 import (
+	"bytes"
 	"context"
 	"errors"
 	"os"
@@ -11,54 +12,97 @@ import (
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 )
 
-func TestCompositeReader_Read(t *testing.T) {
+func TestCompositeReaderRejectsUnsupportedReferences(t *testing.T) {
+	_, err := NewCompositeReader().Read(context.Background(), domain.ArtifactRef{
+		Type: domain.ArtifactRefType("unsupported"),
+		URI:  "unsupported://bucket/key",
+	})
+	if !errors.Is(err, ErrUnsupportedRefType) {
+		t.Fatalf("expected ErrUnsupportedRefType, got %v", err)
+	}
+}
+
+func TestCompositeReaderSourceParityAndOpaqueHashes(t *testing.T) {
 	reader := NewCompositeReader()
-	ctx := context.Background()
+	hashes := make(map[string]string)
+	tests := []struct {
+		name    string
+		content string
+	}{
+		{name: "empty", content: ""},
+		{name: "ordinary", content: "same content"},
+		{name: "changed", content: "changed content"},
+	}
 
-	t.Run("inline artifact", func(t *testing.T) {
+	for _, tc := range tests {
-		ref := domain.ArtifactRef{
+		t.Run(tc.name, func(t *testing.T) {
-			Type: domain.ArtifactRefInline,
+			filePath := filepath.Join(t.TempDir(), "artifact.txt")
-			Body: "hello world",
+			if err := os.WriteFile(filePath, []byte(tc.content), 0o600); err != nil {
-		}
+				t.Fatal(err)
-		art, err := reader.Read(ctx, ref)
+			}
-		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
-		}
-		if string(art.Body) != "hello world" {
-			t.Errorf("expected 'hello world', got %s", string(art.Body))
-		}
-		if art.ContentType != "text/plain" {
-			t.Errorf("expected text/plain content type, got %q", art.ContentType)
-		}
-		if art.Hash != "b94d27b9934d3e08a52e52d7da7dabfac484efe37a5380ee9088f7ace2efcde9" {
-			t.Errorf("unexpected hash: %s", art.Hash)
-		}
-		if art.Size != int64(len(ref.Body)) {
-			t.Errorf("expected size %d, got %d", len(ref.Body), art.Size)
-		}
-	})
 
-	t.Run("inline artifact missing body", func(t *testing.T) {
+			sources := []struct {
-		ref := domain.ArtifactRef{
+				name    string
-			Type: domain.ArtifactRefInline,
+				ref     domain.ArtifactRef
-			Body: "",
+				wantURI string
-		}
+			}{
-		_, err := reader.Read(ctx, ref)
+				{
-		if !errors.Is(err, ErrMissingInlineBody) {
+					name: "inline",
-			t.Errorf("expected ErrMissingInlineBody, got %v", err)
+					ref:  domain.ArtifactRef{Type: domain.ArtifactRefInline, Body: tc.content},
-		}
+				},
-	})
+				{
+					name:    "inline with uri",
+					ref:     domain.ArtifactRef{Type: domain.ArtifactRefInline, URI: "memory://input", Body: tc.content},
+					wantURI: "memory://input",
+				},
+				{
+					name:    "file",
+					ref:     domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filePath},
+					wantURI: filePath,
+				},
+			}
 
-	t.Run("unsupported ref type", func(t *testing.T) {
+			var sourceHash string
-		ref := domain.ArtifactRef{
+			for _, source := range sources {
-			Type: domain.ArtifactRefType("unsupported"),
+				t.Run(source.name, func(t *testing.T) {
-			URI:  "unsupported://bucket/key",
+					first, err := reader.Read(context.Background(), source.ref)
-		}
+					if err != nil {
-		_, err := reader.Read(ctx, ref)
+						t.Fatalf("first read: %v", err)
-		if !errors.Is(err, ErrUnsupportedRefType) {
+					}
-			t.Error("expected error for unsupported type")
+					second, err := reader.Read(context.Background(), source.ref)
-		}
+					if err != nil {
-	})
+						t.Fatalf("second read: %v", err)
+					}
+					if string(first.Body) != tc.content || first.Size != int64(len(tc.content)) {
+						t.Fatalf("body=%q size=%d, want %q/%d", first.Body, first.Size, tc.content, len(tc.content))
+					}
+					if first.URI != source.wantURI {
+						t.Fatalf("URI = %q, want %q", first.URI, source.wantURI)
+					}
+					if first.Hash == "" || first.Hash != second.Hash {
+						t.Fatalf("hashes are not non-empty and stable: %q/%q", first.Hash, second.Hash)
+					}
+					if sourceHash == "" {
+						sourceHash = first.Hash
+					} else if first.Hash != sourceHash {
+						t.Fatalf("equal content hashes differ: %q/%q", sourceHash, first.Hash)
+					}
+					if source.ref.Type == domain.ArtifactRefFile {
+						if first.Name != filepath.Base(filePath) || !strings.HasPrefix(first.ContentType, "text/plain") {
+							t.Fatalf("unexpected file metadata: %+v", first)
+						}
+					} else if first.ContentType != "text/plain" {
+						t.Fatalf("inline content type = %q", first.ContentType)
+					}
+				})
+			}
+			hashes[tc.name] = sourceHash
+		})
+	}
+
+	if hashes["empty"] == hashes["ordinary"] || hashes["ordinary"] == hashes["changed"] {
+		t.Fatalf("changed content did not change opaque hash: %#v", hashes)
+	}
 }
 
 func TestCompositeReaderCopiesInlineData(t *testing.T) {
@@ -87,94 +131,154 @@ func TestCompositeReaderCopiesInlineData(t *testing.T) {
 	}
 }
 
-func TestCompositeReaderHonorsCancellation(t *testing.T) {
+func TestCompositeReaderHonorsPreCancellation(t *testing.T) {
-	ctx, cancel := context.WithCancel(context.Background())
+	filePath := filepath.Join(t.TempDir(), "artifact.txt")
-	cancel()
+	if err := os.WriteFile(filePath, []byte("ignored"), 0o600); err != nil {
+		t.Fatal(err)
+	}
+	tests := []struct {
+		name string
+		ref  domain.ArtifactRef
+	}{
+		{name: "inline", ref: domain.ArtifactRef{Type: domain.ArtifactRefInline, Body: "ignored"}},
+		{name: "file", ref: domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filePath}},
+	}
 
-	_, err := NewCompositeReader().Read(ctx, domain.ArtifactRef{
+	for _, tc := range tests {
-		Type: domain.ArtifactRefInline,
+		t.Run(tc.name, func(t *testing.T) {
-		Body: "ignored",
+			ctx, cancel := context.WithCancel(context.Background())
-	})
+			cancel()
-	if !errors.Is(err, context.Canceled) {
+
-		t.Fatalf("expected context cancellation, got %v", err)
+			artifact, err := NewCompositeReader().Read(ctx, tc.ref)
+			if artifact != nil || !errors.Is(err, context.Canceled) {
+				t.Fatalf("artifact=%#v err=%v, want nil/context.Canceled", artifact, err)
+			}
+		})
 	}
 }
 
-func TestFileReader_Read(t *testing.T) {
+func TestFileReaderFailuresAndMetadata(t *testing.T) {
-	content := []byte("test file content")
-	filePath := filepath.Join(t.TempDir(), "artifact.txt")
-	if err := os.WriteFile(filePath, content, 0o600); err != nil {
-		t.Fatal(err)
-	}
-
 	reader := NewCompositeReader()
-	ctx := context.Background()
-
-	t.Run("file artifact loading", func(t *testing.T) {
-		ref := domain.ArtifactRef{
-			Type: domain.ArtifactRefFile,
-			URI:  filePath,
-		}
-		art, err := reader.Read(ctx, ref)
-		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
-		}
-		if string(art.Body) != string(content) {
-			t.Errorf("expected %s, got %s", string(content), string(art.Body))
-		}
-		if art.Name != filepath.Base(filePath) {
-			t.Errorf("expected name %q, got %q", filepath.Base(filePath), art.Name)
-		}
-		if !strings.HasPrefix(art.ContentType, "text/plain") {
-			t.Errorf("expected text content type, got %q", art.ContentType)
-		}
-		if art.URI != filePath {
-			t.Errorf("expected URI %q, got %q", filePath, art.URI)
-		}
-		if art.Size != int64(len(content)) {
-			t.Errorf("expected size %d, got %d", len(content), art.Size)
-		}
-		if art.Hash != "60f5237ed4049f0382661ef009d2bc42e48c3ceb3edb6600f7024e7ab3b838f3" {
-			t.Errorf("unexpected hash: %s", art.Hash)
-		}
-	})
 
 	t.Run("missing file path", func(t *testing.T) {
-		ref := domain.ArtifactRef{
+		_, err := reader.Read(context.Background(), domain.ArtifactRef{Type: domain.ArtifactRefFile})
-			Type: domain.ArtifactRefFile,
-			URI:  "",
-		}
-		_, err := reader.Read(ctx, ref)
 		if !errors.Is(err, ErrMissingFilePath) {
-			t.Errorf("expected ErrMissingFilePath, got %v", err)
+			t.Fatalf("expected ErrMissingFilePath, got %v", err)
 		}
 	})
 
 	t.Run("missing file", func(t *testing.T) {
-		ref := domain.ArtifactRef{
+		_, err := reader.Read(context.Background(), domain.ArtifactRef{
 			Type: domain.ArtifactRefFile,
 			URI:  filepath.Join(t.TempDir(), "missing.txt"),
-		}
+		})
-		if _, err := reader.Read(ctx, ref); err == nil {
+		if err == nil {
 			t.Fatal("expected missing file error")
 		}
 	})
 
+	t.Run("directory rejected before open", func(t *testing.T) {
+		artifact, err := reader.Read(context.Background(), domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  t.TempDir(),
+		})
+		if artifact != nil || !errors.Is(err, ErrUnsupportedFile) {
+			t.Fatalf("artifact=%#v err=%v, want nil/ErrUnsupportedFile", artifact, err)
+		}
+	})
+
+	t.Run("non-regular opened target rejected", func(t *testing.T) {
+		filePath := filepath.Join(t.TempDir(), "artifact.txt")
+		if err := os.WriteFile(filePath, []byte("content"), 0o600); err != nil {
+			t.Fatal(err)
+		}
+		directoryInfo, err := os.Stat(t.TempDir())
+		if err != nil {
+			t.Fatal(err)
+		}
+		fileReader := &fileReader{open: func(path string) (artifactFile, error) {
+			file, err := os.Open(path)
+			if err != nil {
+				return nil, err
+			}
+			return &reportedInfoFile{artifactFile: file, info: directoryInfo}, nil
+		}}
+
+		artifact, err := fileReader.Read(context.Background(), domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  filePath,
+		})
+		if artifact != nil || !errors.Is(err, ErrUnsupportedFile) {
+			t.Fatalf("artifact=%#v err=%v, want nil/ErrUnsupportedFile", artifact, err)
+		}
+	})
+
 	t.Run("unknown extension uses text fallback", func(t *testing.T) {
-		path := filepath.Join(t.TempDir(), "artifact.unknownextension")
+		filePath := filepath.Join(t.TempDir(), "artifact.unknownextension")
-		if err := os.WriteFile(path, content, 0o600); err != nil {
+		if err := os.WriteFile(filePath, []byte("content"), 0o600); err != nil {
 			t.Fatal(err)
 		}
 
-		art, err := reader.Read(ctx, domain.ArtifactRef{
+		artifact, err := reader.Read(context.Background(), domain.ArtifactRef{
 			Type: domain.ArtifactRefFile,
-			URI:  path,
+			URI:  filePath,
 		})
 		if err != nil {
-			t.Fatalf("unexpected error: %v", err)
+			t.Fatalf("read artifact: %v", err)
 		}
-		if art.ContentType != "text/plain" {
+		if artifact.ContentType != "text/plain" {
-			t.Errorf("expected text/plain fallback, got %q", art.ContentType)
+			t.Fatalf("content type = %q", artifact.ContentType)
 		}
 	})
 }
+
+func TestFileReaderCancelsAfterReadProgress(t *testing.T) {
+	filePath := filepath.Join(t.TempDir(), "artifact.bin")
+	content := bytes.Repeat([]byte("x"), fileReadChunkSize*2)
+	if err := os.WriteFile(filePath, content, 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	var opened *cancelAfterProgressFile
+	reader := &fileReader{open: func(path string) (artifactFile, error) {
+		file, err := os.Open(path)
+		if err != nil {
+			return nil, err
+		}
+		opened = &cancelAfterProgressFile{artifactFile: file, cancel: cancel}
+		return opened, nil
+	}}
+
+	artifact, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filePath})
+	if artifact != nil || !errors.Is(err, context.Canceled) {
+		t.Fatalf("artifact=%#v err=%v, want nil/context.Canceled", artifact, err)
+	}
+	if opened == nil || opened.reads != 1 {
+		t.Fatalf("read count = %v, want one progressing read", opened)
+	}
+}
+
+type reportedInfoFile struct {
+	artifactFile
+	info os.FileInfo
+}
+
+func (f *reportedInfoFile) Stat() (os.FileInfo, error) {
+	return f.info, nil
+}
+
+type cancelAfterProgressFile struct {
+	artifactFile
+	cancel context.CancelFunc
+	reads  int
+}
+
+func (f *cancelAfterProgressFile) Read(buffer []byte) (int, error) {
+	n, err := f.artifactFile.Read(buffer)
+	if n > 0 {
+		f.reads++
+		f.cancel()
+	}
+	return n, err
+}

internal/backend/registry.go

@@ -5,7 +5,6 @@ package backend
 import (
 	"errors"
 	"fmt"
-	"net/url"
 	"regexp"
 	"sort"
 	"strings"
@@ -19,12 +18,11 @@ const (
 	// OpenRouterID is the reserved ID of Promptkit's built-in OpenRouter
 	// backend.
 	OpenRouterID = "openrouter"
+	// RakestrawHomeID is the reserved ID of Promptkit's built-in Rakestrawhome
+	// backend.
+	RakestrawHomeID = "rakestrawhome"
 
-	openRouterEndpoint  = "https://openrouter.ai/api/v1"
+	defaultQueueCapacity = 1024
-	openRouterAPIKeyEnv = "OPENROUTER_API_KEY"
-
-	openRouterConcurrencyLimit = 16
-	defaultQueueCapacity       = 1024
 )
 
 // ErrBackendNotFound identifies a registry lookup for an unknown backend ID.
@@ -37,20 +35,15 @@ type Registry struct {
 	backends map[string]domain.Backend
 }
 
-// NewRegistry constructs a registry containing the built-in OpenRouter
+// NewRegistry constructs a registry containing maintained definitions followed
-// definition followed by the supplied additions. Every ID must be unique.
+// by consumer additions. Every ID must be unique across both groups.
-func NewRegistry(additions []domain.Backend) (*Registry, error) {
+func NewRegistry(maintained, additions []domain.Backend) (*Registry, error) {
 	registry := &Registry{
-		backends: make(map[string]domain.Backend, len(additions)+1),
+		backends: make(map[string]domain.Backend, len(maintained)+len(additions)),
 	}
 
-	definitions := make([]domain.Backend, 0, len(additions)+1)
+	definitions := make([]domain.Backend, 0, len(maintained)+len(additions))
-	definitions = append(definitions, domain.Backend{
+	definitions = append(definitions, maintained...)
-		ID:               OpenRouterID,
-		Endpoint:         openRouterEndpoint,
-		APIKeyEnv:        openRouterAPIKeyEnv,
-		ConcurrencyLimit: openRouterConcurrencyLimit,
-	})
 	definitions = append(definitions, additions...)
 
 	for _, definition := range definitions {
@@ -62,7 +55,7 @@ func NewRegistry(additions []domain.Backend) (*Registry, error) {
 			return nil, fmt.Errorf("backend ID %q is already registered", definition.ID)
 		}
 
-		normalized, err := normalizeBackend(definition)
+		normalized, err := NormalizeDefinition(definition)
 		if err != nil {
 			return nil, err
 		}
@@ -108,11 +101,13 @@ func (r *Registry) CapacityPolicies() map[string]domain.BackendCapacityPolicy {
 	return policies
 }
 
-func normalizeBackend(definition domain.Backend) (domain.Backend, error) {
+// NormalizeDefinition validates and defensively copies one backend definition.
-	definition.Endpoint = strings.TrimSpace(definition.Endpoint)
+func NormalizeDefinition(definition domain.Backend) (domain.Backend, error) {
-	if err := validateEndpoint(definition.Endpoint); err != nil {
+	endpoint, err := domain.NormalizeOpenAICompatibleBaseEndpoint(definition.Endpoint)
+	if err != nil {
 		return domain.Backend{}, fmt.Errorf("backend %q endpoint: %w", definition.ID, err)
 	}
+	definition.Endpoint = endpoint
 
 	definition.APIKeyEnv = strings.TrimSpace(definition.APIKeyEnv)
 	if definition.APIKeyEnv != "" && !environmentVariableName.MatchString(definition.APIKeyEnv) {
@@ -182,31 +177,3 @@ func normalizeBackend(definition domain.Backend) (domain.Backend, error) {
 	definition.ExtraParams = extraParams
 	return definition, nil
 }
-
-func validateEndpoint(endpoint string) error {
-	if endpoint == "" {
-		return errors.New("must not be blank")
-	}
-	if strings.Contains(endpoint, "#") {
-		return errors.New("must not contain a fragment")
-	}
-
-	parsed, err := url.Parse(endpoint)
-	if err != nil {
-		return fmt.Errorf("must be a valid URL: %w", err)
-	}
-	scheme := strings.ToLower(parsed.Scheme)
-	if scheme != "http" && scheme != "https" {
-		return errors.New("must use http or https")
-	}
-	if !parsed.IsAbs() || parsed.Hostname() == "" {
-		return errors.New("must be absolute and include a host")
-	}
-	if parsed.User != nil {
-		return errors.New("must not contain user information")
-	}
-	if parsed.RawQuery != "" || parsed.ForceQuery {
-		return errors.New("must not contain a query string")
-	}
-	return nil
-}

internal/backend/registry_test.go

@@ -2,6 +2,7 @@ package backend_test
 
 import (
 	"errors"
+	"reflect"
 	"strings"
 	"testing"
 
@@ -11,32 +12,39 @@ import (
 
 const validEndpoint = "https://backend.example/v1"
 
-func TestRegistryIncludesExactOpenRouterDefinition(t *testing.T) {
+func TestRegistryIncludesMaintainedDefinitions(t *testing.T) {
-	registry, err := backend.NewRegistry(nil)
+	maintained := []domain.Backend{
+		{ID: backend.OpenRouterID, Endpoint: validEndpoint, ConcurrencyLimit: 2, QueueCapacity: 3, QueueCapacitySet: true},
+		{ID: backend.RakestrawHomeID, Endpoint: "https://second.example/v1", ConcurrencyLimit: 4, QueueCapacity: 5, QueueCapacitySet: true},
+	}
+	registry, err := backend.NewRegistry(maintained, nil)
 	if err != nil {
 		t.Fatalf("construct registry: %v", err)
 	}
 
-	definition, err := registry.GetBackend(backend.OpenRouterID)
+	for _, expected := range maintained {
-	if err != nil {
+		t.Run(expected.ID, func(t *testing.T) {
-		t.Fatalf("look up OpenRouter: %v", err)
+			definition, err := registry.GetBackend(expected.ID)
-	}
+			if err != nil {
-	if definition.ID != "openrouter" ||
+				t.Fatalf("look up maintained definition: %v", err)
-		definition.Endpoint != "https://openrouter.ai/api/v1" ||
+			}
-		definition.APIKeyEnv != "OPENROUTER_API_KEY" ||
+			if !reflect.DeepEqual(definition, expected) {
-		definition.ConcurrencyLimit != 16 ||
+				t.Fatalf("unexpected maintained definition: %#v", definition)
-		definition.QueueCapacity != 1024 ||
+			}
-		!definition.QueueCapacitySet ||
+		})
-		definition.ExtraParams != nil {
-		t.Fatalf("unexpected OpenRouter definition: %#v", definition)
 	}
+
 	policies := registry.CapacityPolicies()
-	if len(policies) != 1 ||
+	if len(policies) != 2 ||
-		policies["openrouter"] != (domain.BackendCapacityPolicy{
+		policies[backend.OpenRouterID] != (domain.BackendCapacityPolicy{
-			ConcurrencyLimit: 16,
+			ConcurrencyLimit: 2,
-			QueueCapacity:    1024,
+			QueueCapacity:    3,
+		}) ||
+		policies[backend.RakestrawHomeID] != (domain.BackendCapacityPolicy{
+			ConcurrencyLimit: 4,
+			QueueCapacity:    5,
 		}) {
-		t.Fatalf("unexpected OpenRouter capacity policies: %#v", policies)
+		t.Fatalf("unexpected built-in capacity policies: %#v", policies)
 	}
 }
 
@@ -46,7 +54,7 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 		"count":  int64(7),
 		"nested": nested,
 	}
-	registry, err := backend.NewRegistry([]domain.Backend{
+	registry, err := backend.NewRegistry(nil, []domain.Backend{
 		{
 			ID:               " custom ",
 			Endpoint:         " https://custom.example/openai/v1 ",
@@ -109,11 +117,11 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 	}
 
 	policies := registry.CapacityPolicies()
-	if len(policies) != 2 {
+	if len(policies) != 1 {
 		t.Fatalf("unexpected capacity policy count: %#v", policies)
 	}
 	policies["custom"] = domain.BackendCapacityPolicy{}
-	delete(policies, backend.OpenRouterID)
+	delete(policies, "custom")
 	againPolicies := registry.CapacityPolicies()
 	if againPolicies["custom"] != (domain.BackendCapacityPolicy{
 		ConcurrencyLimit: 3,
@@ -121,9 +129,6 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 	}) {
 		t.Fatalf("capacity policy map mutated registry state: %#v", againPolicies)
 	}
-	if _, ok := againPolicies[backend.OpenRouterID]; !ok {
-		t.Fatalf("capacity policy deletion mutated registry state: %#v", againPolicies)
-	}
 }
 
 func TestNewRegistryNormalizesCapacityPolicy(t *testing.T) {
@@ -199,7 +204,7 @@ func TestNewRegistryNormalizesCapacityPolicy(t *testing.T) {
 		t.Run(tc.name, func(t *testing.T) {
 			tc.definition.ID = "custom"
 			tc.definition.Endpoint = validEndpoint
-			registry, err := backend.NewRegistry([]domain.Backend{tc.definition})
+			registry, err := backend.NewRegistry(nil, []domain.Backend{tc.definition})
 			if tc.wantError {
 				if err == nil {
 					t.Fatal("expected invalid capacity policy error")
@@ -242,11 +247,14 @@ func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
 		wantID    string
 	}{
 		{
-			name: "built-in collision after normalization",
+			name:      "OpenRouter collision after normalization",
-			additions: []domain.Backend{{
+			additions: []domain.Backend{{ID: " openrouter "}},
-				ID: " openrouter ",
+			wantID:    backend.OpenRouterID,
-			}},
+		},
-			wantID: "openrouter",
+		{
+			name:      "Rakestrawhome collision after normalization",
+			additions: []domain.Backend{{ID: " rakestrawhome "}},
+			wantID:    backend.RakestrawHomeID,
 		},
 		{
 			name: "consumer collision after normalization",
@@ -260,7 +268,7 @@ func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry(tc.additions)
+			_, err := backend.NewRegistry(nil, tc.additions)
 			if err == nil {
 				t.Fatal("expected duplicate ID error")
 			}
@@ -274,7 +282,7 @@ func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
 func TestNewRegistryValidatesIDs(t *testing.T) {
 	for _, id := range []string{"", " \t\n "} {
 		t.Run(id, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:       id,
 				Endpoint: validEndpoint,
 			}})
@@ -303,7 +311,7 @@ func TestNewRegistryValidatesEndpoints(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:       "custom",
 				Endpoint: tc.endpoint,
 			}})
@@ -317,7 +325,7 @@ func TestNewRegistryValidatesEndpoints(t *testing.T) {
 func TestNewRegistryValidatesEnvironmentVariableNames(t *testing.T) {
 	for _, name := range []string{"1API_KEY", "API-KEY", "API KEY", "ÅPI_KEY"} {
 		t.Run(name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:        "custom",
 				Endpoint:  validEndpoint,
 				APIKeyEnv: name,
@@ -340,7 +348,7 @@ func TestNewRegistryRejectsInvalidAndReservedExtraParameters(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:          "custom",
 				Endpoint:    validEndpoint,
 				ExtraParams: tc.extraParams,
@@ -353,7 +361,7 @@ func TestNewRegistryRejectsInvalidAndReservedExtraParameters(t *testing.T) {
 }
 
 func TestRegistryLookupReportsNotFound(t *testing.T) {
-	registry, err := backend.NewRegistry(nil)
+	registry, err := backend.NewRegistry(nil, nil)
 	if err != nil {
 		t.Fatalf("construct registry: %v", err)
 	}

internal/capacity/client.go

@@ -28,13 +28,13 @@ func (c *client) Generate(
 	ctx context.Context,
 	req domain.GenerateRequest,
 ) (*domain.GenerateResponse, error) {
-	pool := c.manager.getPool(req.Target.BackendID)
+	if c.manager.getPool(req.Target.BackendID) == nil {
-	if pool == nil {
 		return c.next.Generate(ctx, req)
 	}
-	if err := pool.acquire(ctx); err != nil {
+	release, err := c.manager.Acquire(ctx, req.Target.BackendID)
+	if err != nil {
 		return nil, err
 	}
-	defer pool.releaseActive()
+	defer release()
 	return c.next.Generate(ctx, req)
 }

internal/capacity/manager.go

@@ -107,6 +107,26 @@ func (m *Manager) getPool(backendID string) *pool {
 	return m.pools[backendID]
 }
 
+// Limited reports whether a backend uses an active-generation pool.
+func (m *Manager) Limited(backendID string) bool { return m.getPool(backendID) != nil }
+
+// Acquire obtains one active-generation permit for a backend. Its release
+// function is idempotent. Backends without a configured pool are unlimited.
+func (m *Manager) Acquire(ctx context.Context, backendID string) (func(), error) {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	pool := m.getPool(backendID)
+	if pool == nil {
+		return releaseNothing, nil
+	}
+	if err := pool.acquire(ctx); err != nil {
+		return nil, err
+	}
+	var once sync.Once
+	return func() { once.Do(pool.releaseActive) }, nil
+}
+
 func (p *pool) acquire(ctx context.Context) error {
 	p.mu.Lock()
 	if err := ctx.Err(); err != nil {

internal/capacity/manager_test.go

@@ -60,6 +60,26 @@ func TestNewManagerRejectsInvalidPolicies(t *testing.T) {
 	}
 }
 
+func TestManagerAcquireReleaseIsIdempotent(t *testing.T) {
+	manager, err := NewManager(map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	release, err := manager.Acquire(context.Background(), "limited")
+	if err != nil {
+		t.Fatal(err)
+	}
+	release()
+	release()
+	other, err := manager.Acquire(context.Background(), "limited")
+	if err != nil {
+		t.Fatal(err)
+	}
+	other()
+}
+
 func TestManagerAdmissionIsBoundedAndReleaseIsIdempotent(t *testing.T) {
 	policies := map[string]domain.BackendCapacityPolicy{
 		"limited": {
@@ -124,6 +144,11 @@ func TestManagerAdmissionHonorsContextAndUnlimitedBackends(t *testing.T) {
 	if err != nil {
 		t.Fatalf("construct manager: %v", err)
 	}
+	release, err := manager.Admit(context.Background(), "limited")
+	if err != nil {
+		t.Fatalf("fill limited pool: %v", err)
+	}
+	defer release()
 
 	ctx, cancel := context.WithCancel(context.Background())
 	cancel()

internal/catalog/catalog.go

@@ -0,0 +1,249 @@
+// Package catalog validates immutable maintained backend catalog assets.
+package catalog
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"path"
+	"sort"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+)
+
+// Source identifies one immutable backend catalog asset tree.
+type Source struct {
+	Name              string
+	ExpectedBackendID string
+	FS                fs.FS
+	Root              string
+}
+
+// Set is the validated maintained backend and raw profile catalog.
+type Set struct {
+	Backends   []domain.Backend
+	Profiles   profile.Repository
+	profileIDs []string
+}
+
+// Load validates and combines immutable catalog sources in source order.
+func Load(sources ...Source) (Set, error) {
+	if len(sources) == 0 {
+		return Set{}, errors.New("at least one catalog source is required")
+	}
+	loaded := Set{Backends: make([]domain.Backend, 0, len(sources))}
+	names := map[string]bool{}
+	backendIDs := map[string]bool{}
+	profileIDs := map[string]bool{}
+	for _, source := range sources {
+		if err := validateSource(source, names); err != nil {
+			return Set{}, err
+		}
+		names[source.Name] = true
+		definition, err := loadBackend(source)
+		if err != nil {
+			return Set{}, err
+		}
+		if backendIDs[definition.ID] {
+			return Set{}, fmt.Errorf("catalog %s: backend ID duplicates an earlier catalog", source.Name)
+		}
+		backendIDs[definition.ID] = true
+		repository, metadata, err := profile.LoadFSRepository(context.Background(), source.FS, path.Join(source.Root, "profiles"))
+		if err != nil {
+			return Set{}, fmt.Errorf("catalog %s: %w", source.Name, err)
+		}
+		if len(metadata) == 0 {
+			return Set{}, fmt.Errorf("catalog %s: profiles must not be empty", source.Name)
+		}
+		resolvingRepository := profile.NewResolvingRepository(repository)
+		for _, entry := range metadata {
+			if strings.HasPrefix(entry.ID, "promptkit-") {
+				return Set{}, fmt.Errorf("catalog %s: %s uses the reserved promptkit- profile ID prefix", source.Name, entry.Path)
+			}
+			if profileIDs[entry.ID] {
+				return Set{}, fmt.Errorf("catalog %s: %s duplicates an earlier profile ID", source.Name, entry.Path)
+			}
+			if containsField(entry.ExplicitFields, "endpoint") || containsField(entry.ExplicitFields, "api_key_env") {
+				return Set{}, fmt.Errorf("catalog %s: %s contains connection metadata", source.Name, entry.Path)
+			}
+			value, err := repository.GetProfile(context.Background(), entry.ID)
+			if err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s: %w", source.Name, entry.Path, err)
+			}
+			if err := rejectSecretKeys(value.ExtraParams); err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s: prohibited extra parameter key", source.Name, entry.Path)
+			}
+			resolved, err := resolvingRepository.GetProfile(context.Background(), entry.ID)
+			if err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s has invalid profile inheritance", source.Name, entry.Path)
+			}
+			if resolved.BackendID != definition.ID {
+				return Set{}, fmt.Errorf("catalog %s: %s selects a different backend", source.Name, entry.Path)
+			}
+			profileIDs[entry.ID] = true
+			loaded.profileIDs = append(loaded.profileIDs, entry.ID)
+		}
+		if loaded.Profiles == nil {
+			loaded.Profiles = repository
+		} else {
+			loaded.Profiles = profile.NewOverlayRepository(loaded.Profiles, repository)
+		}
+		loaded.Backends = append(loaded.Backends, definition)
+	}
+	sort.Strings(loaded.profileIDs)
+	return loaded, nil
+}
+
+func validateSource(source Source, names map[string]bool) error {
+	if strings.TrimSpace(source.Name) == "" || source.Name != strings.TrimSpace(source.Name) {
+		return errors.New("catalog source name must not be blank")
+	}
+	if names[source.Name] {
+		return fmt.Errorf("duplicate catalog source name %q", source.Name)
+	}
+	if source.FS == nil {
+		return fmt.Errorf("catalog %s: filesystem is nil", source.Name)
+	}
+	if source.Root == "." || source.Root == "" || !fs.ValidPath(source.Root) {
+		return fmt.Errorf("catalog %s: asset root is invalid", source.Name)
+	}
+	if strings.TrimSpace(source.ExpectedBackendID) == "" {
+		return fmt.Errorf("catalog %s: expected backend ID is blank", source.Name)
+	}
+	return validateLayout(source)
+}
+
+func validateLayout(source Source) error {
+	manifestPath := path.Join(source.Root, "backend.json")
+	profilesRoot := path.Join(source.Root, "profiles")
+	return fs.WalkDir(source.FS, source.Root, func(assetPath string, entry fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if assetPath == source.Root {
+			if !entry.IsDir() {
+				return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+			}
+			return nil
+		}
+		if entry.IsDir() {
+			if assetPath == profilesRoot || strings.HasPrefix(assetPath, profilesRoot+"/") {
+				return nil
+			}
+			return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+		}
+		if assetPath == manifestPath && entry.Type().IsRegular() {
+			return nil
+		}
+		if strings.HasPrefix(assetPath, profilesRoot+"/") && entry.Type().IsRegular() && strings.HasSuffix(assetPath, ".yml") {
+			return nil
+		}
+		return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+	})
+}
+
+func loadBackend(source Source) (domain.Backend, error) {
+	data, err := fs.ReadFile(source.FS, path.Join(source.Root, "backend.json"))
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: %w", source.Name, err)
+	}
+	type manifest struct {
+		SchemaVersion    *int            `json:"schema_version"`
+		ID               *string         `json:"id"`
+		Endpoint         *string         `json:"endpoint"`
+		APIKeyEnv        *string         `json:"api_key_env"`
+		ConcurrencyLimit *int            `json:"concurrency_limit"`
+		QueueCapacity    *int            `json:"queue_capacity"`
+		ExtraParams      json.RawMessage `json:"extra_params"`
+	}
+	decoder := json.NewDecoder(bytes.NewReader(data))
+	decoder.DisallowUnknownFields()
+	var value manifest
+	if err := decoder.Decode(&value); err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid manifest", source.Name)
+	}
+	var trailing any
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid manifest", source.Name)
+	}
+	if value.SchemaVersion == nil || *value.SchemaVersion != 1 || value.ID == nil || value.Endpoint == nil || value.APIKeyEnv == nil || value.ConcurrencyLimit == nil || value.QueueCapacity == nil || value.ExtraParams == nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: required field is missing or unsupported", source.Name)
+	}
+	if *value.ID != source.ExpectedBackendID {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend ID does not match expected ID", source.Name)
+	}
+	if strings.TrimSpace(*value.APIKeyEnv) == "" {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: api key environment variable must not be blank", source.Name)
+	}
+	extraParams, err := decodeExtraParams(value.ExtraParams)
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid extra parameters", source.Name)
+	}
+	if err := rejectSecretKeys(extraParams); err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: prohibited extra parameter key", source.Name)
+	}
+	normalized, err := backend.NormalizeDefinition(domain.Backend{ID: *value.ID, Endpoint: *value.Endpoint, APIKeyEnv: *value.APIKeyEnv, ExtraParams: extraParams, ConcurrencyLimit: *value.ConcurrencyLimit, QueueCapacity: *value.QueueCapacity, QueueCapacitySet: true})
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid backend definition", source.Name)
+	}
+	return normalized, nil
+}
+
+func decodeExtraParams(data []byte) (map[string]any, error) {
+	if bytes.Equal(bytes.TrimSpace(data), []byte("null")) {
+		return nil, nil
+	}
+	decoder := json.NewDecoder(bytes.NewReader(data))
+	decoder.UseNumber()
+	var value map[string]any
+	if err := decoder.Decode(&value); err != nil || value == nil {
+		return nil, errors.New("extra parameters must be an object or null")
+	}
+	var trailing any
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
+		return nil, errors.New("extra parameters must contain exactly one JSON value")
+	}
+	return value, nil
+}
+
+func containsField(fields []string, target string) bool {
+	for _, field := range fields {
+		if field == target {
+			return true
+		}
+	}
+	return false
+}
+
+func rejectSecretKeys(value map[string]any) error {
+	return rejectSecretValue(value)
+}
+
+func rejectSecretValue(value any) error {
+	switch value := value.(type) {
+	case map[string]any:
+		for key, child := range value {
+			switch strings.ToLower(key) {
+			case "api_key", "apikey", "authorization", "credential", "credentials", "password", "secret", "token", "access_token":
+				return errors.New("prohibited key")
+			}
+			if err := rejectSecretValue(child); err != nil {
+				return err
+			}
+		}
+	case []any:
+		for _, child := range value {
+			if err := rejectSecretValue(child); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}

internal/catalog/catalog_test.go

@@ -0,0 +1,356 @@
+package catalog
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"io/fs"
+	"strings"
+	"testing"
+	"testing/fstest"
+
+	openrouter "gitea.maximumdirect.net/eric/promptkit-backend-openrouter/v2"
+	rakestrawhome "gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/v2"
+)
+
+func TestLoadPublishedCatalogs(t *testing.T) {
+	loaded, err := Load(
+		Source{Name: "OpenRouter", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root},
+		Source{Name: "Rakestrawhome", ExpectedBackendID: "rakestrawhome", FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+	)
+	if err != nil {
+		t.Fatalf("load published catalogs: %v", err)
+	}
+	if loaded.Profiles == nil || len(loaded.profileIDs) == 0 {
+		t.Fatalf("published catalogs have no usable profiles: %#v", loaded)
+	}
+	if _, err := loaded.Profiles.GetProfile(context.Background(), "gemini-flash-latest"); err != nil {
+		t.Fatalf("load maintained standard-profile base: %v", err)
+	}
+
+	expectedBackends := map[string]struct {
+		endpoint         string
+		apiKeyEnv        string
+		concurrencyLimit int
+		queueCapacity    int
+	}{
+		"openrouter": {
+			endpoint:         "https://openrouter.ai/api/v1",
+			apiKeyEnv:        "OPENROUTER_API_KEY",
+			concurrencyLimit: 16,
+			queueCapacity:    1024,
+		},
+		"rakestrawhome": {
+			endpoint:         "https://inference.ai.rakestrawhome.com/v1",
+			apiKeyEnv:        "RAKESTRAWHOME_INFERENCE_API_KEY",
+			concurrencyLimit: 4,
+			queueCapacity:    1024,
+		},
+	}
+	if len(loaded.Backends) != len(expectedBackends) {
+		t.Fatalf("published backend count = %d, want %d", len(loaded.Backends), len(expectedBackends))
+	}
+	for _, definition := range loaded.Backends {
+		expected, found := expectedBackends[definition.ID]
+		if !found {
+			t.Fatalf("unexpected published backend %q", definition.ID)
+		}
+		if definition.Endpoint != expected.endpoint ||
+			definition.APIKeyEnv != expected.apiKeyEnv ||
+			definition.ConcurrencyLimit != expected.concurrencyLimit ||
+			definition.QueueCapacity != expected.queueCapacity ||
+			!definition.QueueCapacitySet ||
+			definition.ExtraParams != nil {
+			t.Fatalf("unexpected published backend definition: %#v", definition)
+		}
+	}
+}
+
+func TestLoadRejectsInvalidSources(t *testing.T) {
+	for name, sources := range map[string][]Source{
+		"none":       nil,
+		"blank name": {{Name: " ", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root}},
+		"duplicate name": {
+			{Name: "same", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root},
+			{Name: "same", ExpectedBackendID: "rakestrawhome", FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+		},
+		"nil filesystem":         {{Name: "missing", ExpectedBackendID: "openrouter", Root: openrouter.Root}},
+		"invalid root":           {{Name: "invalid-root", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: "."}},
+		"blank expected backend": {{Name: "blank-backend", ExpectedBackendID: " ", FS: openrouter.FS(), Root: openrouter.Root}},
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := Load(sources...); err == nil {
+				t.Fatal("expected error")
+			}
+		})
+	}
+}
+
+func TestLoadRejectsInvalidLayouts(t *testing.T) {
+	tests := map[string]func(fstest.MapFS){
+		"unexpected file": func(fsys fstest.MapFS) {
+			fsys["catalog/notes.txt"] = &fstest.MapFile{Data: []byte("unexpected")}
+		},
+		"unexpected directory": func(fsys fstest.MapFS) {
+			fsys["catalog/unexpected"] = &fstest.MapFile{Mode: fs.ModeDir}
+		},
+		"nonregular manifest": func(fsys fstest.MapFS) {
+			fsys["catalog/backend.json"].Mode = fs.ModeSymlink
+		},
+		"wrong profile extension": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/extra.yaml"] = &fstest.MapFile{Data: []byte(validProfile("extra", "one"))}
+		},
+		"nonregular profile": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Mode = fs.ModeSymlink
+		},
+	}
+	for name, mutate := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			mutate(fsys)
+			if _, err := Load(testSource("one", fsys)); err == nil {
+				t.Fatal("expected invalid layout error")
+			}
+		})
+	}
+}
+
+func TestLoadRejectsInvalidManifests(t *testing.T) {
+	tests := map[string]string{
+		"malformed":             `{`,
+		"trailing value":        validManifest("one", "TEST_API_KEY", "null") + `{}`,
+		"missing fields":        `{"schema_version":1,"id":"one"}`,
+		"unsupported version":   strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"schema_version":1`, `"schema_version":2`, 1),
+		"unknown field":         strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"extra_params":null`, `"extra_params":null,"unknown":true`, 1),
+		"blank API key env":     validManifest("one", "   ", "null"),
+		"invalid API key env":   validManifest("one", "LEAK-MARKER", "null"),
+		"invalid endpoint":      strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `https://one.example/v1`, `ftp://leak-marker.invalid/v1`, 1),
+		"zero concurrency":      strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"concurrency_limit":2`, `"concurrency_limit":0`, 1),
+		"non-object parameters": validManifest("one", "TEST_API_KEY", `[]`),
+		"secret parameter":      validManifest("one", "TEST_API_KEY", `{"nested":{"token":"leak-marker"}}`),
+	}
+	for name, manifest := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			fsys["catalog/backend.json"].Data = []byte(manifest)
+			_, err := Load(testSource("one", fsys))
+			if err == nil {
+				t.Fatal("expected invalid manifest error")
+			}
+			if strings.Contains(strings.ToLower(err.Error()), "leak-marker") {
+				t.Fatalf("catalog error exposed manifest content: %v", err)
+			}
+		})
+	}
+}
+
+func TestLoadPreservesManifestJSONNumbers(t *testing.T) {
+	fsys := validCatalogFS("one")
+	fsys["catalog/backend.json"].Data = []byte(validManifest(
+		"one",
+		"TEST_API_KEY",
+		`{"large":9007199254740993,"nested":[1.25]}`,
+	))
+	loaded, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("load catalog: %v", err)
+	}
+	if got := loaded.Backends[0].ExtraParams["large"]; got != json.Number("9007199254740993") {
+		t.Fatalf("large JSON integer = %#v, want preserved json.Number", got)
+	}
+	nested := loaded.Backends[0].ExtraParams["nested"].([]any)
+	if nested[0] != json.Number("1.25") {
+		t.Fatalf("nested JSON number = %#v, want preserved json.Number", nested[0])
+	}
+}
+
+func TestLoadRejectsInvalidCatalogProfiles(t *testing.T) {
+	tests := map[string]func(fstest.MapFS){
+		"empty": func(fsys fstest.MapFS) {
+			delete(fsys, "catalog/profiles/one-profile.yml")
+			fsys["catalog/profiles"] = &fstest.MapFile{Mode: fs.ModeDir}
+		},
+		"malformed": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: [")
+		},
+		"raw API key": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "api_key: leak-marker\n")
+		},
+		"endpoint field": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "endpoint: ''\n")
+		},
+		"API key environment field": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "api_key_env: ''\n")
+		},
+		"owner mismatch": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "other"))
+		},
+		"missing base": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: one-profile\nbase_profile: leak-marker\n")
+		},
+		"cyclic base": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: one-profile\nbase_profile: second\n")
+			fsys["catalog/profiles/second.yml"] = &fstest.MapFile{Data: []byte("id: second\nbase_profile: one-profile\n")}
+		},
+		"secret profile parameter": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "extra_params:\n  nested:\n    password: leak-marker\n")
+		},
+		"unknown field is redacted": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "leak_marker: leak-marker\n")
+		},
+	}
+	for name, mutate := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			mutate(fsys)
+			_, err := Load(testSource("one", fsys))
+			if err == nil {
+				t.Fatal("expected invalid profile error")
+			}
+			if strings.Contains(strings.ToLower(err.Error()), "leak-marker") {
+				t.Fatalf("catalog error exposed profile content: %v", err)
+			}
+		})
+	}
+}
+
+func TestLoadReservesPromptkitProfilePrefixForCatalogs(t *testing.T) {
+	tests := map[string]struct {
+		profileID string
+		wantError bool
+	}{
+		"maintained name": {
+			profileID: "promptkit-openrouter-standard",
+			wantError: true,
+		},
+		"other reserved name": {
+			profileID: "promptkit-example",
+			wantError: true,
+		},
+		"ordinary name": {
+			profileID: "ordinary-profile",
+		},
+	}
+	for name, test := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile(test.profileID, "one"))
+
+			_, err := Load(testSource("one", fsys))
+			if (err != nil) != test.wantError {
+				t.Fatalf("Load() error = %v, want error %t", err, test.wantError)
+			}
+			if err != nil && !strings.Contains(err.Error(), "catalog one: one-profile.yml") {
+				t.Fatalf("Load() error = %v, want catalog source and path context", err)
+			}
+		})
+	}
+}
+
+func TestLoadRejectsCrossCatalogConflicts(t *testing.T) {
+	t.Run("duplicate backend", func(t *testing.T) {
+		second := catalogFS("one", map[string]string{
+			"catalog/profiles/second.yml": validProfile("second", "one"),
+		}, "null")
+		_, err := Load(
+			Source{Name: "first", ExpectedBackendID: "one", FS: validCatalogFS("one"), Root: "catalog"},
+			Source{Name: "second", ExpectedBackendID: "one", FS: second, Root: "catalog"},
+		)
+		if err == nil {
+			t.Fatal("expected duplicate backend error")
+		}
+	})
+
+	t.Run("duplicate profile", func(t *testing.T) {
+		first := catalogFS("one", map[string]string{
+			"catalog/profiles/shared.yml": validProfile("shared", "one"),
+		}, "null")
+		second := catalogFS("two", map[string]string{
+			"catalog/profiles/shared.yml": validProfile("shared", "two"),
+		}, "null")
+		_, err := Load(testSource("one", first), testSource("two", second))
+		if err == nil {
+			t.Fatal("expected duplicate profile error")
+		}
+	})
+
+	t.Run("cross-catalog base", func(t *testing.T) {
+		first := catalogFS("one", map[string]string{
+			"catalog/profiles/base.yml": validProfile("base", "one"),
+		}, "null")
+		second := catalogFS("two", map[string]string{
+			"catalog/profiles/child.yml": "id: child\nbase_profile: base\n",
+		}, "null")
+		_, err := Load(testSource("one", first), testSource("two", second))
+		if err == nil {
+			t.Fatal("expected cross-catalog base error")
+		}
+	})
+}
+
+func TestLoadReturnsDefensiveCatalogValues(t *testing.T) {
+	fsys := catalogFS("one", map[string]string{
+		"catalog/profiles/one-profile.yml": validProfile("one-profile", "one") + "extra_params:\n  nested:\n    value: profile\n",
+	}, `{"nested":{"value":"backend"}}`)
+	loaded, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("load catalog: %v", err)
+	}
+	loaded.Backends[0].ExtraParams["nested"].(map[string]any)["value"] = "changed"
+	profileValue, err := loaded.Profiles.GetProfile(context.Background(), "one-profile")
+	if err != nil {
+		t.Fatalf("load profile: %v", err)
+	}
+	profileValue.ExtraParams["nested"].(map[string]any)["value"] = "changed"
+
+	again, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("reload catalog: %v", err)
+	}
+	if got := again.Backends[0].ExtraParams["nested"].(map[string]any)["value"]; got != "backend" {
+		t.Fatalf("backend mutation escaped returned set: %#v", got)
+	}
+	againProfile, err := loaded.Profiles.GetProfile(context.Background(), "one-profile")
+	if err != nil {
+		t.Fatalf("reload profile: %v", err)
+	}
+	if got := againProfile.ExtraParams["nested"].(map[string]any)["value"]; got != "profile" {
+		t.Fatalf("profile mutation escaped returned value: %#v", got)
+	}
+}
+
+func testSource(id string, fsys fs.FS) Source {
+	return Source{Name: id, ExpectedBackendID: id, FS: fsys, Root: "catalog"}
+}
+
+func validCatalogFS(id string) fstest.MapFS {
+	return catalogFS(id, map[string]string{
+		"catalog/profiles/" + id + "-profile.yml": validProfile(id+"-profile", id),
+	}, "null")
+}
+
+func catalogFS(id string, profiles map[string]string, extraParams string) fstest.MapFS {
+	fsys := fstest.MapFS{
+		"catalog/backend.json": &fstest.MapFile{Data: []byte(validManifest(id, "TEST_API_KEY", extraParams))},
+	}
+	for name, content := range profiles {
+		fsys[name] = &fstest.MapFile{Data: []byte(content)}
+	}
+	return fsys
+}
+
+func validManifest(id, apiKeyEnv, extraParams string) string {
+	return fmt.Sprintf(
+		`{"schema_version":1,"id":%q,"endpoint":%q,"api_key_env":%q,"concurrency_limit":2,"queue_capacity":3,"extra_params":%s}`,
+		id,
+		"https://"+id+".example/v1",
+		apiKeyEnv,
+		extraParams,
+	)
+}
+
+func validProfile(id, backendID string) string {
+	return fmt.Sprintf("id: %s\nbackend: %s\nmodel: test-model\n", id, backendID)
+}
+
+var _ fs.FS = openrouter.FS()

internal/defaults/defaults.go

@@ -14,21 +14,12 @@ const (
 	ContentTypeApplicationJSON = "application/json"
 	OpenAIChatCompletionsPath  = "/chat/completions"
 
-	ExecutionDefaultTemperature    = 0.0
-	ExecutionDefaultMaxTokens      = 0
-	ExecutionDefaultTopP           = 1.0
 	ExecutionDefaultTimeoutSeconds = 600
-)
+	LLMRequestTimeoutDefault       = 10 * time.Minute
-
-var (
-	LLMRequestTimeoutDefault = 10 * time.Minute
 )
 
 func ExecutionTargetDefault() domain.ExecutionTarget {
 	return domain.ExecutionTarget{
-		Temperature:    ExecutionDefaultTemperature,
-		MaxTokens:      ExecutionDefaultMaxTokens,
-		TopP:           ExecutionDefaultTopP,
 		TimeoutSeconds: ExecutionDefaultTimeoutSeconds,
 	}
 }

internal/domain/domain.go

@@ -60,15 +60,18 @@ type CacheControl struct {
 
 // RunRequest represents a request to generate a single artifact.
 type RunRequest struct {
-	PromptID      string
+	PromptID         string
-	PromptVersion string
+	PromptVersion    string
-	ProfileID     string
+	ProfileID        string
-	SessionID     string
+	SessionID        string
-	APIKey        string `json:"-" yaml:"-"`
+	APIKey           string `json:"-" yaml:"-"`
-	Inputs        map[string]ArtifactRef
+	Inputs           map[string]ArtifactRef
-	Vars          map[string]string
+	Vars             map[string]string
-	Execution     *ExecutionTargetOverride
+	Execution        *ExecutionTargetOverride
-	Validation    *OutputContract
+	Validation       *OutputContract
+	AppendedMessages []RenderedMessage
+	TransientRetries *int
+	Observer         ExecutionObserver `json:"-"`
 }
 
 // RunResult represents the complete result of a prompt execution run.
@@ -97,22 +100,22 @@ type RunResult struct {
 // PreparedRun contains pre-LLM execution state from the prepare/render phase.
 // It must never include resolved API key values, model output, or validation data.
 type PreparedRun struct {
-	PromptID             string                  `json:"prompt_id"`
+	PromptID             string
-	PromptVersion        string                  `json:"prompt_version,omitempty"`
+	PromptVersion        string
-	PromptHash           string                  `json:"prompt_hash,omitempty"`
+	PromptHash           string
-	SelectedProfileID    string                  `json:"selected_profile_id"`
+	SelectedProfileID    string
-	SelectedBackendID    string                  `json:"selected_backend_id,omitempty"`
+	SelectedBackendID    string
-	EffectiveModelParams ExecutionTarget         `json:"effective_model_params"`
+	EffectiveModelParams ExecutionTarget
-	TargetPresence       ExecutionTargetPresence `json:"-"`
+	TargetPresence       ExecutionTargetPresence
-	OutputContract       OutputContract          `json:"output_contract"`
+	OutputContract       OutputContract
-	StructuredOutput     *StructuredOutputSpec   `json:"structured_output,omitempty"`
+	StructuredOutput     *StructuredOutputSpec
-	InputHashes          map[string]string       `json:"input_hashes,omitempty"`
+	InputHashes          map[string]string
-	SessionID            string                  `json:"session_id,omitempty"`
+	SessionID            string
-	RenderedPromptHash   string                  `json:"rendered_prompt_hash"`
+	RenderedPromptHash   string
-	Messages             []RenderedMessage       `json:"messages"`
+	Messages             []RenderedMessage
-	StartTime            time.Time               `json:"start_time,omitempty"`
+	StartTime            time.Time
-	EndTime              time.Time               `json:"end_time,omitempty"`
+	EndTime              time.Time
-	DurationMS           int64                   `json:"duration_ms,omitempty"`
+	DurationMS           int64
 }
 
 // ArtifactRef represents a reference to an input artifact.
@@ -145,6 +148,16 @@ type PromptDefinition struct {
 	Validation     OutputContract          `yaml:"validation"`
 }
 
+// PromptInspection is the resolved result of exact prompt inspection.
+type PromptInspection struct {
+	PromptID         string
+	PromptVersion    string
+	PromptHash       string
+	DefaultProfileID string
+	Inputs           []PromptInput
+	OutputContract   OutputContract
+}
+
 // PromptInput describes one named input expected by a prompt definition.
 type PromptInput struct {
 	Name        string `yaml:"name"`
@@ -182,6 +195,7 @@ type BackendCapacityPolicy struct {
 // ExecutionProfile describes how and where to execute a model.
 type ExecutionProfile struct {
 	ID              string         `yaml:"id"`
+	BaseProfileID   string         `yaml:"base_profile"`
 	BackendID       string         `yaml:"backend"`
 	Endpoint        string         `yaml:"endpoint"`
 	Model           string         `yaml:"model"`
@@ -236,6 +250,13 @@ type ExecutionTarget struct {
 	ExtraParams     map[string]any `yaml:"extra_params" json:"extra_params"`
 }
 
+// ProfileInspection is the resolved result of exact profile inspection.
+type ProfileInspection struct {
+	ProfileID            string
+	EffectiveModelParams ExecutionTarget
+	APIKeyRequired       bool
+}
+
 // OutputContract defines the requirements for the output artifact.
 type OutputContract struct {
 	Format         OutputFormat   `yaml:"format"`
@@ -263,6 +284,8 @@ type GenerateRequest struct {
 	Target           ExecutionTarget
 	TargetPresence   ExecutionTargetPresence
 	StructuredOutput *StructuredOutputSpec
+	TransientRetries uint64
+	Generation       uint64
 }
 
 // StructuredOutputType indicates which provider-level output mode is requested.
@@ -287,8 +310,18 @@ type StructuredOutputJSONSpec struct {
 
 // GenerateResponse is the response received from the LLM client.
 type GenerateResponse struct {
-	Content string
+	Content       string
-	Usage   TokenUsage
+	Usage         TokenUsage
+	UsagePresence *TokenUsagePresence
+}
+
+// TokenUsagePresence records which token fields a provider reported.
+type TokenUsagePresence struct {
+	PromptTokens     bool
+	CompletionTokens bool
+	TotalTokens      bool
+	CachedTokens     bool
+	CacheWriteTokens bool
 }
 
 // TokenUsage tracks token consumption.

internal/domain/endpoint.go

@@ -0,0 +1,39 @@
+package domain
+
+import (
+	"errors"
+	"net/url"
+	"strings"
+)
+
+// NormalizeOpenAICompatibleBaseEndpoint trims and validates a source-neutral
+// OpenAI-compatible provider base endpoint.
+func NormalizeOpenAICompatibleBaseEndpoint(endpoint string) (string, error) {
+	endpoint = strings.TrimSpace(endpoint)
+	if endpoint == "" {
+		return "", errors.New("endpoint must not be blank")
+	}
+	if strings.Contains(endpoint, "#") {
+		return "", errors.New("endpoint must not contain a fragment")
+	}
+
+	parsed, err := url.Parse(endpoint)
+	if err != nil {
+		return "", errors.New("endpoint must be a valid URL")
+	}
+	parsed.Scheme = strings.ToLower(parsed.Scheme)
+	if parsed.Scheme != "http" && parsed.Scheme != "https" {
+		return "", errors.New("endpoint must use http or https")
+	}
+	if !parsed.IsAbs() || parsed.Hostname() == "" {
+		return "", errors.New("endpoint must be absolute and include a host")
+	}
+	if parsed.User != nil {
+		return "", errors.New("endpoint must not contain user information")
+	}
+	if parsed.RawQuery != "" || parsed.ForceQuery {
+		return "", errors.New("endpoint must not contain a query string")
+	}
+
+	return parsed.String(), nil
+}

internal/domain/endpoint_test.go

@@ -0,0 +1,47 @@
+package domain
+
+import "testing"
+
+func TestNormalizeOpenAICompatibleBaseEndpoint(t *testing.T) {
+	tests := []struct {
+		name     string
+		endpoint string
+		want     string
+		wantErr  bool
+	}{
+		{name: "http host", endpoint: "http://provider.example", want: "http://provider.example"},
+		{name: "https nested path and whitespace", endpoint: "  HTTPS://provider.example/api/openai/v1  ", want: "https://provider.example/api/openai/v1"},
+		{name: "IPv4 host and port", endpoint: "http://127.0.0.1:8080/v1", want: "http://127.0.0.1:8080/v1"},
+		{name: "IPv6 host and port", endpoint: "https://[::1]:8443/v1", want: "https://[::1]:8443/v1"},
+		{name: "repeated trailing slashes", endpoint: "https://provider.example/v1///", want: "https://provider.example/v1///"},
+		{name: "blank", endpoint: " \t\n ", wantErr: true},
+		{name: "relative path", endpoint: "/api/v1", wantErr: true},
+		{name: "scheme relative", endpoint: "//provider.example/v1", wantErr: true},
+		{name: "missing host", endpoint: "https:///v1", wantErr: true},
+		{name: "unsupported scheme", endpoint: "ftp://provider.example/v1", wantErr: true},
+		{name: "user information", endpoint: "https://user:secret@provider.example/v1", wantErr: true},
+		{name: "query", endpoint: "https://provider.example/v1?mode=chat", wantErr: true},
+		{name: "empty query", endpoint: "https://provider.example/v1?", wantErr: true},
+		{name: "fragment", endpoint: "https://provider.example/v1#chat", wantErr: true},
+		{name: "empty fragment", endpoint: "https://provider.example/v1#", wantErr: true},
+		{name: "malformed URL", endpoint: "https://provider.example/%zz", wantErr: true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := NormalizeOpenAICompatibleBaseEndpoint(tc.endpoint)
+			if tc.wantErr {
+				if err == nil {
+					t.Fatalf("expected endpoint error, got %q", got)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("normalize endpoint: %v", err)
+			}
+			if got != tc.want {
+				t.Fatalf("normalized endpoint = %q, want %q", got, tc.want)
+			}
+		})
+	}
+}

internal/domain/execution_settings.go

@@ -0,0 +1,34 @@
+package domain
+
+import (
+	"errors"
+	"math"
+	"time"
+)
+
+const maxExecutionTimeoutSeconds int64 = math.MaxInt64 / int64(time.Second)
+
+// ValidateExecutionTargetSettings validates source-neutral execution-setting
+// invariants on a resolved target.
+func ValidateExecutionTargetSettings(target ExecutionTarget) error {
+	if !isFinite(target.Temperature) || target.Temperature < 0 || target.Temperature > 2 {
+		return errors.New("temperature must be finite and between 0 and 2")
+	}
+	if target.MaxTokens < 0 {
+		return errors.New("max_tokens must be greater than or equal to 0")
+	}
+	if !isFinite(target.TopP) || target.TopP < 0 || target.TopP > 1 {
+		return errors.New("top_p must be finite and between 0 and 1")
+	}
+	if target.TimeoutSeconds < 0 {
+		return errors.New("timeout_seconds must be greater than or equal to 0")
+	}
+	if int64(target.TimeoutSeconds) > maxExecutionTimeoutSeconds {
+		return errors.New("timeout_seconds exceeds the maximum supported duration")
+	}
+	return nil
+}
+
+func isFinite(value float64) bool {
+	return !math.IsNaN(value) && !math.IsInf(value, 0)
+}

internal/domain/execution_settings_test.go

@@ -0,0 +1,73 @@
+package domain
+
+import (
+	"math"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+func TestValidateExecutionTargetSettings(t *testing.T) {
+	valid := ExecutionTarget{
+		Temperature:    1,
+		MaxTokens:      1,
+		TopP:           0.5,
+		TimeoutSeconds: 1,
+	}
+
+	type testCase struct {
+		name    string
+		change  func(*ExecutionTarget)
+		wantErr string
+	}
+	tests := []testCase{
+		{name: "temperature lower boundary", change: func(v *ExecutionTarget) { v.Temperature = 0 }},
+		{name: "temperature finite lower neighbor", change: func(v *ExecutionTarget) { v.Temperature = math.Nextafter(0, 1) }},
+		{name: "temperature finite upper neighbor", change: func(v *ExecutionTarget) { v.Temperature = math.Nextafter(2, 0) }},
+		{name: "temperature upper boundary", change: func(v *ExecutionTarget) { v.Temperature = 2 }},
+		{name: "temperature below lower boundary", change: func(v *ExecutionTarget) { v.Temperature = math.Nextafter(0, math.Inf(-1)) }, wantErr: "temperature"},
+		{name: "temperature above upper boundary", change: func(v *ExecutionTarget) { v.Temperature = math.Nextafter(2, math.Inf(1)) }, wantErr: "temperature"},
+		{name: "temperature NaN", change: func(v *ExecutionTarget) { v.Temperature = math.NaN() }, wantErr: "temperature"},
+		{name: "temperature positive infinity", change: func(v *ExecutionTarget) { v.Temperature = math.Inf(1) }, wantErr: "temperature"},
+		{name: "temperature negative infinity", change: func(v *ExecutionTarget) { v.Temperature = math.Inf(-1) }, wantErr: "temperature"},
+		{name: "max tokens lower boundary", change: func(v *ExecutionTarget) { v.MaxTokens = 0 }},
+		{name: "max tokens finite neighbor", change: func(v *ExecutionTarget) { v.MaxTokens = 1 }},
+		{name: "max tokens below lower boundary", change: func(v *ExecutionTarget) { v.MaxTokens = -1 }, wantErr: "max_tokens"},
+		{name: "top p lower boundary", change: func(v *ExecutionTarget) { v.TopP = 0 }},
+		{name: "top p finite lower neighbor", change: func(v *ExecutionTarget) { v.TopP = math.Nextafter(0, 1) }},
+		{name: "top p finite upper neighbor", change: func(v *ExecutionTarget) { v.TopP = math.Nextafter(1, 0) }},
+		{name: "top p upper boundary", change: func(v *ExecutionTarget) { v.TopP = 1 }},
+		{name: "top p below lower boundary", change: func(v *ExecutionTarget) { v.TopP = math.Nextafter(0, math.Inf(-1)) }, wantErr: "top_p"},
+		{name: "top p above upper boundary", change: func(v *ExecutionTarget) { v.TopP = math.Nextafter(1, math.Inf(1)) }, wantErr: "top_p"},
+		{name: "top p NaN", change: func(v *ExecutionTarget) { v.TopP = math.NaN() }, wantErr: "top_p"},
+		{name: "top p positive infinity", change: func(v *ExecutionTarget) { v.TopP = math.Inf(1) }, wantErr: "top_p"},
+		{name: "top p negative infinity", change: func(v *ExecutionTarget) { v.TopP = math.Inf(-1) }, wantErr: "top_p"},
+		{name: "timeout lower boundary", change: func(v *ExecutionTarget) { v.TimeoutSeconds = 0 }},
+		{name: "timeout finite neighbor", change: func(v *ExecutionTarget) { v.TimeoutSeconds = 1 }},
+		{name: "timeout below lower boundary", change: func(v *ExecutionTarget) { v.TimeoutSeconds = -1 }, wantErr: "timeout_seconds"},
+	}
+	if strconv.IntSize == 64 {
+		durationLimit := maxExecutionTimeoutSeconds
+		tests = append(tests,
+			testCase{name: "timeout duration boundary", change: func(v *ExecutionTarget) { v.TimeoutSeconds = int(durationLimit) }},
+			testCase{name: "timeout above duration boundary", change: func(v *ExecutionTarget) { v.TimeoutSeconds = int(durationLimit) + 1 }, wantErr: "timeout_seconds"},
+		)
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			target := valid
+			tt.change(&target)
+			err := ValidateExecutionTargetSettings(target)
+			if tt.wantErr == "" {
+				if err != nil {
+					t.Fatalf("validate execution settings: %v", err)
+				}
+				return
+			}
+			if err == nil || !strings.Contains(err.Error(), tt.wantErr) {
+				t.Fatalf("error = %v, want diagnostic containing %q", err, tt.wantErr)
+			}
+		})
+	}
+}

internal/domain/message.go

@@ -0,0 +1,82 @@
+package domain
+
+import (
+	"errors"
+	"strings"
+	"unicode/utf8"
+)
+
+const (
+	RoleDeveloper = "developer"
+	RoleSystem    = "system"
+	RoleUser      = "user"
+	RoleAssistant = "assistant"
+)
+
+// NormalizeMessageRole validates and canonicalizes a provider-bound chat role.
+func NormalizeMessageRole(role string) (string, error) {
+	if !utf8.ValidString(role) {
+		return "", errors.New("message role must be valid UTF-8")
+	}
+
+	normalized := strings.ToLower(strings.TrimSpace(role))
+	switch normalized {
+	case RoleDeveloper, RoleSystem, RoleUser, RoleAssistant:
+		return normalized, nil
+	default:
+		return "", errors.New("message role must be developer, system, user, or assistant")
+	}
+}
+
+// NormalizeCacheControl validates, canonicalizes, and copies cache metadata.
+func NormalizeCacheControl(control *CacheControl) (*CacheControl, error) {
+	if control == nil {
+		return nil, nil
+	}
+	if !utf8.ValidString(string(control.Type)) {
+		return nil, errors.New("cache control type must be valid UTF-8")
+	}
+	if !utf8.ValidString(control.TTL) {
+		return nil, errors.New("cache control ttl must be valid UTF-8")
+	}
+
+	cacheType := strings.TrimSpace(string(control.Type))
+	if cacheType == "" {
+		return nil, errors.New("cache control type is required")
+	}
+	if CacheControlType(cacheType) != CacheControlEphemeral {
+		return nil, errors.New("unsupported type")
+	}
+
+	ttl := strings.TrimSpace(control.TTL)
+	if ttl != "" && ttl != "1h" {
+		return nil, errors.New("unsupported ttl")
+	}
+
+	return &CacheControl{Type: CacheControlType(cacheType), TTL: ttl}, nil
+}
+
+// CloneRenderedMessages returns a deep copy of rendered messages.
+func CloneRenderedMessages(messages []RenderedMessage) []RenderedMessage {
+	cloned := make([]RenderedMessage, len(messages))
+	copyRenderedMessages(cloned, messages)
+	return cloned
+}
+
+// ConcatRenderedMessages returns an independently owned concatenation of messages.
+func ConcatRenderedMessages(prefix, suffix []RenderedMessage) []RenderedMessage {
+	messages := make([]RenderedMessage, len(prefix)+len(suffix))
+	copyRenderedMessages(messages, prefix)
+	copyRenderedMessages(messages[len(prefix):], suffix)
+	return messages
+}
+
+func copyRenderedMessages(destination, source []RenderedMessage) {
+	for index, message := range source {
+		destination[index] = message
+		if message.CacheControl != nil {
+			cacheControl := *message.CacheControl
+			destination[index].CacheControl = &cacheControl
+		}
+	}
+}

internal/domain/message_test.go

@@ -0,0 +1,121 @@
+package domain
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestNormalizeMessageRole(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   string
+		want    string
+		wantErr bool
+	}{
+		{name: "developer", input: RoleDeveloper, want: RoleDeveloper},
+		{name: "system", input: RoleSystem, want: RoleSystem},
+		{name: "user", input: RoleUser, want: RoleUser},
+		{name: "assistant", input: RoleAssistant, want: RoleAssistant},
+		{name: "surrounding whitespace and mixed case", input: " \u2003UsEr\u2003 ", want: RoleUser},
+		{name: "blank", input: " \t\n ", wantErr: true},
+		{name: "tool", input: "tool", wantErr: true},
+		{name: "function", input: "function", wantErr: true},
+		{name: "custom", input: "custom-role", wantErr: true},
+		{name: "invalid UTF-8", input: string([]byte{0xff}), wantErr: true},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := NormalizeMessageRole(test.input)
+			if test.wantErr {
+				if err == nil {
+					t.Fatal("expected an error")
+				}
+				if strings.Contains(err.Error(), test.input) {
+					t.Fatalf("error exposed the input: %v", err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("NormalizeMessageRole() error = %v", err)
+			}
+			if got != test.want {
+				t.Fatalf("NormalizeMessageRole() = %q, want %q", got, test.want)
+			}
+		})
+	}
+}
+
+func TestNormalizeCacheControl(t *testing.T) {
+	tests := []struct {
+		name    string
+		input   *CacheControl
+		want    *CacheControl
+		wantErr bool
+	}{
+		{name: "nil", input: nil, want: nil},
+		{
+			name:  "canonical values",
+			input: &CacheControl{Type: CacheControlEphemeral, TTL: "1h"},
+			want:  &CacheControl{Type: CacheControlEphemeral, TTL: "1h"},
+		},
+		{
+			name:  "trims values",
+			input: &CacheControl{Type: " ephemeral ", TTL: " 1h\t"},
+			want:  &CacheControl{Type: CacheControlEphemeral, TTL: "1h"},
+		},
+		{name: "empty type", input: &CacheControl{}, wantErr: true},
+		{name: "unsupported type", input: &CacheControl{Type: "persistent"}, wantErr: true},
+		{name: "unsupported ttl", input: &CacheControl{Type: CacheControlEphemeral, TTL: "5m"}, wantErr: true},
+		{name: "invalid type UTF-8", input: &CacheControl{Type: CacheControlType(string([]byte{0xff}))}, wantErr: true},
+		{name: "invalid ttl UTF-8", input: &CacheControl{Type: CacheControlEphemeral, TTL: string([]byte{0xff})}, wantErr: true},
+	}
+
+	for _, test := range tests {
+		t.Run(test.name, func(t *testing.T) {
+			got, err := NormalizeCacheControl(test.input)
+			if test.wantErr {
+				if err == nil {
+					t.Fatal("expected an error")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("NormalizeCacheControl() error = %v", err)
+			}
+			if got == nil || test.want == nil {
+				if got != test.want {
+					t.Fatalf("NormalizeCacheControl() = %#v, want %#v", got, test.want)
+				}
+				return
+			}
+			if *got != *test.want {
+				t.Fatalf("NormalizeCacheControl() = %#v, want %#v", got, test.want)
+			}
+			if got == test.input {
+				t.Fatal("normalized cache control aliases its input")
+			}
+		})
+	}
+}
+
+func TestRenderedMessageCloning(t *testing.T) {
+	prefix := []RenderedMessage{{Role: RoleSystem, Content: "prefix", CacheControl: &CacheControl{Type: CacheControlEphemeral, TTL: "1h"}}}
+	suffix := []RenderedMessage{{Role: RoleUser, Content: " suffix "}, {Role: RoleAssistant, Content: "", CacheControl: &CacheControl{Type: CacheControlEphemeral}}}
+
+	cloned := CloneRenderedMessages(prefix)
+	combined := ConcatRenderedMessages(prefix, suffix)
+	if len(combined) != 3 || combined[0].Content != "prefix" || combined[1].Content != " suffix " || combined[2].Content != "" {
+		t.Fatalf("unexpected combined messages: %#v", combined)
+	}
+	if cloned[0].CacheControl == prefix[0].CacheControl || combined[0].CacheControl == prefix[0].CacheControl || combined[2].CacheControl == suffix[1].CacheControl {
+		t.Fatal("cloned cache controls alias their inputs")
+	}
+
+	prefix[0].Content = "changed"
+	prefix[0].CacheControl.TTL = ""
+	suffix[1].CacheControl.Type = "changed"
+	if cloned[0].Content != "prefix" || cloned[0].CacheControl.TTL != "1h" || combined[0].Content != "prefix" || combined[0].CacheControl.TTL != "1h" || combined[2].CacheControl.Type != CacheControlEphemeral {
+		t.Fatalf("cloned messages changed with their inputs: cloned=%#v combined=%#v", cloned, combined)
+	}
+}

internal/domain/observation.go

@@ -0,0 +1,90 @@
+package domain
+
+import "time"
+
+// ExecutionObserver receives one immutable metadata event synchronously.
+type ExecutionObserver func(ExecutionEvent)
+
+type ExecutionEventKind string
+
+const (
+	ExecutionStarted    ExecutionEventKind = "started"
+	ExecutionAttempt    ExecutionEventKind = "attempt"
+	ExecutionRetryDelay ExecutionEventKind = "retry_delay"
+	ExecutionFinished   ExecutionEventKind = "finished"
+)
+
+type ExecutionOutcome string
+
+const (
+	ExecutionOutcomeSuccess       ExecutionOutcome = "success"
+	ExecutionOutcomeInvalidOutput ExecutionOutcome = "invalid_output"
+	ExecutionOutcomeError         ExecutionOutcome = "error"
+)
+
+type ExecutionFailureKind string
+
+const (
+	ExecutionFailureNone              ExecutionFailureKind = ""
+	ExecutionFailureCanceled          ExecutionFailureKind = "canceled"
+	ExecutionFailureDeadline          ExecutionFailureKind = "deadline"
+	ExecutionFailureTransport         ExecutionFailureKind = "transport"
+	ExecutionFailureProviderStatus    ExecutionFailureKind = "provider_status"
+	ExecutionFailureMalformedResponse ExecutionFailureKind = "malformed_response"
+	ExecutionFailureCapacity          ExecutionFailureKind = "capacity"
+	ExecutionFailureInvalidConfig     ExecutionFailureKind = "invalid_config"
+	ExecutionFailureInvalidRequest    ExecutionFailureKind = "invalid_request"
+	ExecutionFailurePromptNotFound    ExecutionFailureKind = "prompt_not_found"
+	ExecutionFailureProfileNotFound   ExecutionFailureKind = "profile_not_found"
+	ExecutionFailurePromptLoad        ExecutionFailureKind = "prompt_load"
+	ExecutionFailureProfileLoad       ExecutionFailureKind = "profile_load"
+	ExecutionFailureArtifactLoad      ExecutionFailureKind = "artifact_load"
+	ExecutionFailurePromptRender      ExecutionFailureKind = "prompt_render"
+	ExecutionFailureValidation        ExecutionFailureKind = "validation"
+	ExecutionFailureGeneration        ExecutionFailureKind = "generation"
+	ExecutionFailureOther             ExecutionFailureKind = "other"
+)
+
+// ObservedTokenCount distinguishes omitted usage from reported zero.
+type ObservedTokenCount struct {
+	Value    int
+	Known    bool
+	Complete bool
+}
+
+// ObservedTokenUsage reports presence-aware token counts.
+type ObservedTokenUsage struct {
+	PromptTokens     ObservedTokenCount
+	CompletionTokens ObservedTokenCount
+	TotalTokens      ObservedTokenCount
+	CachedTokens     ObservedTokenCount
+	CacheWriteTokens ObservedTokenCount
+}
+
+// ExecutionEvent contains metadata only; it carries no mutable containers or
+// error values. Nonterminal events leave aggregate counters at zero.
+type ExecutionEvent struct {
+	Kind               ExecutionEventKind
+	RunID              string
+	Time               time.Time
+	PromptID           string
+	PromptVersion      string
+	ProfileID          string
+	BackendID          string
+	Model              string
+	Generation         uint64
+	Attempt            uint64
+	Dispatched         bool
+	CapacityWait       time.Duration
+	GenerationDuration time.Duration
+	RetryDelay         time.Duration
+	Duration           time.Duration
+	Outcome            ExecutionOutcome
+	Failure            ExecutionFailureKind
+	StatusCode         int
+	RetriesExhausted   bool
+	Usage              ObservedTokenUsage
+	Attempts           uint64
+	TransientRetries   uint64
+	RepairCalls        uint64
+}

internal/domain/output_contract.go

@@ -0,0 +1,38 @@
+package domain
+
+import (
+	"errors"
+	"fmt"
+	"strings"
+)
+
+const maxOutputRepairAttempts = 3
+
+// ValidateOutputContract validates source-neutral output-contract invariants.
+func ValidateOutputContract(contract OutputContract) error {
+	switch contract.Format {
+	case FormatText, FormatMarkdown, FormatJSON:
+	default:
+		return fmt.Errorf("invalid output format: %q", contract.Format)
+	}
+
+	switch contract.ValidationMode {
+	case ValidationNone, ValidationBasic, ValidationJSON, ValidationJSONSchema:
+	default:
+		return fmt.Errorf("invalid validation mode: %q", contract.ValidationMode)
+	}
+
+	if contract.ValidationMode == ValidationJSONSchema && strings.TrimSpace(contract.SchemaPath) == "" {
+		return errors.New("schema_path is required when validation_mode is json_schema")
+	}
+	if contract.RepairAttempts < 0 {
+		return errors.New("repair_attempts must be greater than or equal to 0")
+	}
+	if contract.RepairAttempts > maxOutputRepairAttempts {
+		return fmt.Errorf("repair_attempts must be less than or equal to %d", maxOutputRepairAttempts)
+	}
+	if contract.ValidationMode == ValidationNone && contract.RepairAttempts > 0 {
+		return errors.New("repair_attempts requires basic, json, or json_schema validation")
+	}
+	return nil
+}

internal/domain/output_contract_test.go

@@ -0,0 +1,87 @@
+package domain
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestValidateOutputContract(t *testing.T) {
+	valid := OutputContract{
+		Format:         FormatText,
+		ValidationMode: ValidationNone,
+	}
+
+	tests := []struct {
+		name    string
+		change  func(*OutputContract)
+		wantErr string
+	}{
+		{name: "text format", change: func(c *OutputContract) { c.Format = FormatText }},
+		{name: "markdown format", change: func(c *OutputContract) { c.Format = FormatMarkdown }},
+		{name: "json format", change: func(c *OutputContract) { c.Format = FormatJSON }},
+		{name: "empty format", change: func(c *OutputContract) { c.Format = "" }, wantErr: "format"},
+		{name: "unsupported format", change: func(c *OutputContract) { c.Format = OutputFormat("binary") }, wantErr: "format"},
+		{name: "none validation", change: func(c *OutputContract) { c.ValidationMode = ValidationNone }},
+		{name: "basic validation", change: func(c *OutputContract) { c.ValidationMode = ValidationBasic }},
+		{name: "json validation", change: func(c *OutputContract) { c.ValidationMode = ValidationJSON }},
+		{name: "json schema validation", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = "schema.json"
+		}},
+		{name: "empty validation mode", change: func(c *OutputContract) { c.ValidationMode = "" }, wantErr: "validation mode"},
+		{name: "unsupported validation mode", change: func(c *OutputContract) { c.ValidationMode = ValidationMode("unknown") }, wantErr: "validation mode"},
+		{name: "negative repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationBasic
+			c.RepairAttempts = -1
+		}, wantErr: "repair_attempts"},
+		{name: "zero repair attempts", change: func(c *OutputContract) { c.RepairAttempts = 0 }},
+		{name: "one repair attempt", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationBasic
+			c.RepairAttempts = 1
+		}},
+		{name: "maximum repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSON
+			c.RepairAttempts = 3
+		}},
+		{name: "too many repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = "schema.json"
+			c.RepairAttempts = 4
+		}, wantErr: "repair_attempts"},
+		{name: "none validation with repair attempts", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationNone
+			c.RepairAttempts = 1
+		}, wantErr: "repair_attempts"},
+		{name: "json schema empty path", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = ""
+		}, wantErr: "schema_path"},
+		{name: "json schema whitespace path", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = " \t "
+		}, wantErr: "schema_path"},
+		{name: "json schema nonblank path", change: func(c *OutputContract) {
+			c.ValidationMode = ValidationJSONSchema
+			c.SchemaPath = " schema.json "
+		}},
+		{name: "non-schema empty path", change: func(c *OutputContract) { c.SchemaPath = "" }},
+		{name: "non-schema populated path", change: func(c *OutputContract) { c.SchemaPath = "ignored.json" }},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			contract := valid
+			tt.change(&contract)
+			err := ValidateOutputContract(contract)
+			if tt.wantErr == "" {
+				if err != nil {
+					t.Fatalf("validate output contract: %v", err)
+				}
+				return
+			}
+			if err == nil || !strings.Contains(err.Error(), tt.wantErr) {
+				t.Fatalf("error = %v, want diagnostic containing %q", err, tt.wantErr)
+			}
+		})
+	}
+}

internal/domain/prepared_run_test.go

@@ -1,141 +0,0 @@
-package domain
-
-import (
-	"encoding/json"
-	"strings"
-	"testing"
-)
-
-func TestPreparedRunJSONDoesNotIncludeSecretValues(t *testing.T) {
-	const envName = "PROMPTKIT_TEST_API_KEY"
-	const secret = "super-secret-value"
-	t.Setenv(envName, secret)
-
-	prepared := PreparedRun{
-		PromptID:          "prompt.id",
-		PromptVersion:     "v1",
-		PromptHash:        "prompt-hash",
-		SelectedProfileID: "local-fast",
-		EffectiveModelParams: ExecutionTarget{
-			Endpoint:  "http://llm/v1",
-			Model:     "gpt-test",
-			APIKeyEnv: envName,
-			APIKey:    secret,
-		},
-		InputHashes:        map[string]string{"transcript": "hash-1"},
-		RenderedPromptHash: "rendered-hash",
-		Messages: []RenderedMessage{
-			{Role: "system", Content: "You are helpful."},
-			{Role: "user", Content: "Summarize this."},
-		},
-	}
-
-	b, err := json.Marshal(prepared)
-	if err != nil {
-		t.Fatalf("marshal failed: %v", err)
-	}
-
-	out := string(b)
-	if strings.Contains(out, secret) {
-		t.Fatalf("prepared run JSON unexpectedly contains secret value: %s", out)
-	}
-	if !strings.Contains(out, `"api_key_env":"`+envName+`"`) {
-		t.Fatalf("prepared run JSON should include api_key_env name: %s", out)
-	}
-
-	var top map[string]any
-	if err := json.Unmarshal(b, &top); err != nil {
-		t.Fatalf("unmarshal failed: %v", err)
-	}
-
-	for _, forbidden := range []string{"raw_output", "validation", "artifact"} {
-		if _, ok := top[forbidden]; ok {
-			t.Fatalf("prepared run JSON should not include %q", forbidden)
-		}
-	}
-}
-
-func TestPreparedRunJSONIncludesMessageCacheControlOnlyWhenPresent(t *testing.T) {
-	prepared := PreparedRun{
-		PromptID:          "prompt.id",
-		SelectedProfileID: "local-fast",
-		EffectiveModelParams: ExecutionTarget{
-			Endpoint: "http://llm/v1",
-			Model:    "gpt-test",
-		},
-		RenderedPromptHash: "rendered-hash",
-		Messages: []RenderedMessage{
-			{
-				Role:    "system",
-				Content: "You are helpful.",
-				CacheControl: &CacheControl{
-					Type: CacheControlEphemeral,
-					TTL:  "1h",
-				},
-			},
-			{Role: "user", Content: "Summarize this."},
-		},
-	}
-
-	b, err := json.Marshal(prepared)
-	if err != nil {
-		t.Fatalf("marshal failed: %v", err)
-	}
-
-	var decoded struct {
-		Messages []map[string]any `json:"messages"`
-	}
-	if err := json.Unmarshal(b, &decoded); err != nil {
-		t.Fatalf("unmarshal failed: %v", err)
-	}
-	if len(decoded.Messages) != 2 {
-		t.Fatalf("expected 2 messages, got %d", len(decoded.Messages))
-	}
-
-	cacheControl, ok := decoded.Messages[0]["cache_control"].(map[string]any)
-	if !ok {
-		t.Fatalf("expected cache_control on first message, got %#v", decoded.Messages[0])
-	}
-	if cacheControl["type"] != string(CacheControlEphemeral) || cacheControl["ttl"] != "1h" {
-		t.Fatalf("unexpected cache_control payload: %#v", cacheControl)
-	}
-	if _, ok := decoded.Messages[1]["cache_control"]; ok {
-		t.Fatalf("expected second message to omit cache_control, got %#v", decoded.Messages[1])
-	}
-}
-
-func TestPreparedRunJSONIncludesSessionIDOnlyWhenPresent(t *testing.T) {
-	prepared := PreparedRun{
-		PromptID:          "prompt.id",
-		SelectedProfileID: "local-fast",
-		EffectiveModelParams: ExecutionTarget{
-			Endpoint: "http://llm/v1",
-			Model:    "gpt-test",
-		},
-		SessionID:          "session-123",
-		RenderedPromptHash: "rendered-hash",
-		Messages:           []RenderedMessage{{Role: "user", Content: "Summarize this."}},
-	}
-
-	b, err := json.Marshal(prepared)
-	if err != nil {
-		t.Fatalf("marshal failed: %v", err)
-	}
-
-	var decoded map[string]any
-	if err := json.Unmarshal(b, &decoded); err != nil {
-		t.Fatalf("unmarshal failed: %v", err)
-	}
-	if decoded["session_id"] != "session-123" {
-		t.Fatalf("expected session_id in prepared run JSON, got %#v", decoded["session_id"])
-	}
-
-	prepared.SessionID = ""
-	b, err = json.Marshal(prepared)
-	if err != nil {
-		t.Fatalf("marshal failed: %v", err)
-	}
-	if strings.Contains(string(b), "session_id") {
-		t.Fatalf("expected empty session_id to be omitted, got %s", b)
-	}
-}

internal/domain/retries.go

@@ -0,0 +1,15 @@
+package domain
+
+import "fmt"
+
+// ResolveTransientRetries selects a nonnegative retry budget. A nil selection
+// inherits the supplied default.
+func ResolveTransientRetries(defaultBudget uint64, selection *int) (uint64, error) {
+	if selection == nil {
+		return defaultBudget, nil
+	}
+	if *selection < 0 {
+		return 0, fmt.Errorf("transient retries must not be negative")
+	}
+	return uint64(*selection), nil
+}

internal/domain/session.go

@@ -8,6 +8,9 @@ import (
 
 // NormalizeSessionID applies the shared session identifier rule.
 func NormalizeSessionID(raw string) (string, error) {
+	if !utf8.ValidString(raw) {
+		return "", fmt.Errorf("session_id must contain valid UTF-8")
+	}
 	normalized := strings.TrimSpace(raw)
 	if normalized == "" {
 		return "", nil

internal/domain/session_test.go

@@ -7,10 +7,10 @@ import (
 
 func TestNormalizeSessionID(t *testing.T) {
 	tests := []struct {
-		name    string
+		name            string
-		raw     string
+		raw             string
-		want    string
+		want            string
-		wantErr bool
+		wantErrContains string
 	}{
 		{
 			name: "trims surrounding Unicode whitespace",
@@ -28,21 +28,24 @@ func TestNormalizeSessionID(t *testing.T) {
 			want: strings.Repeat("界", SessionIDMaxLength),
 		},
 		{
-			name:    "one Unicode code point over maximum is rejected",
+			name:            "one Unicode code point over maximum is rejected",
-			raw:     strings.Repeat("界", SessionIDMaxLength+1),
+			raw:             strings.Repeat("界", SessionIDMaxLength+1),
-			wantErr: true,
+			wantErrContains: "exceeds maximum",
 		},
+		{name: "invalid UTF-8 before valid content", raw: string([]byte{0xff}) + "session", wantErrContains: "valid UTF-8"},
+		{name: "invalid UTF-8 within valid content", raw: "ses" + string([]byte{0xff}) + "sion", wantErrContains: "valid UTF-8"},
+		{name: "invalid UTF-8 after valid content", raw: "session" + string([]byte{0xff}), wantErrContains: "valid UTF-8"},
 	}
 
 	for _, tt := range tests {
 		t.Run(tt.name, func(t *testing.T) {
 			got, err := NormalizeSessionID(tt.raw)
-			if tt.wantErr {
+			if tt.wantErrContains != "" {
 				if err == nil {
 					t.Fatal("expected normalization error")
 				}
-				if !strings.Contains(err.Error(), "exceeds maximum") {
+				if !strings.Contains(err.Error(), tt.wantErrContains) {
-					t.Fatalf("expected useful length diagnostic, got %v", err)
+					t.Fatalf("expected diagnostic containing %q, got %v", tt.wantErrContains, err)
 				}
 				return
 			}

internal/engine/engine.go

@@ -0,0 +1,285 @@
+// Package engine assembles Promptkit's internal execution components.
+package engine
+
+import (
+	"context"
+	"fmt"
+	"io/fs"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	openrouter "gitea.maximumdirect.net/eric/promptkit-backend-openrouter/v2"
+	rakestrawhome "gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome/v2"
+	"gitea.maximumdirect.net/eric/promptkit/internal/artifact"
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/catalog"
+	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/generation"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+	"gitea.maximumdirect.net/eric/promptkit/internal/prompt"
+	"gitea.maximumdirect.net/eric/promptkit/internal/promptdef"
+	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
+	"gitea.maximumdirect.net/eric/promptkit/internal/validate"
+)
+
+// Config carries construction settings selected by the public facade.
+type Config struct {
+	PromptDir        string
+	ProfileDir       string
+	SchemaDir        string
+	Timeout          time.Duration
+	HTTPClient       *http.Client
+	TransientRetries *int
+}
+
+// Options collects construction choices from the public facade.
+type Options struct {
+	llmClient             llm.Client
+	artifactReader        artifact.Reader
+	promptDefinitions     promptdef.Repository
+	profiles              profile.Repository
+	fallbackProfiles      profile.Repository
+	memoryProfiles        profile.Repository
+	backends              []domain.Backend
+	validator             validate.Validator
+	promptSource          bool
+	profileSource         bool
+	fallbackProfileSource bool
+	memorySource          bool
+	validatorSource       bool
+	artifactSource        bool
+}
+
+// SetLLMClient selects the client collaborator for one engine.
+func (o *Options) SetLLMClient(client llm.Client) { o.llmClient = client }
+
+// SetArtifactReader selects the artifact collaborator for one engine.
+func (o *Options) SetArtifactReader(reader artifact.Reader) {
+	o.artifactReader = reader
+	o.artifactSource = true
+}
+
+// SetPromptFS selects a prompt source rooted in fsys.
+func (o *Options) SetPromptFS(fsys fs.FS, root string) error {
+	if err := validateSourceFS(fsys, root); err != nil {
+		return err
+	}
+	o.promptDefinitions = promptdef.NewFSRepository(fsys, root)
+	o.promptSource = true
+	return nil
+}
+
+// SetPromptFile selects a single prompt file.
+func (o *Options) SetPromptFile(path string) error {
+	fsys, root, directory, err := fileSource(path)
+	if err != nil {
+		return err
+	}
+	o.promptDefinitions = promptdef.NewFileRepository(fsys, root, directory)
+	o.promptSource = true
+	return nil
+}
+
+// SetProfileFS selects an ordinary profile source rooted in fsys.
+func (o *Options) SetProfileFS(fsys fs.FS, root string) error {
+	if err := validateSourceFS(fsys, root); err != nil {
+		return err
+	}
+	o.profiles = profile.NewFSRepository(fsys, root)
+	o.profileSource = true
+	return nil
+}
+
+// SetProfileFile selects a single ordinary profile file.
+func (o *Options) SetProfileFile(path string) error {
+	fsys, root, _, err := fileSource(path)
+	if err != nil {
+		return err
+	}
+	o.profiles = profile.NewFSRepository(fsys, root)
+	o.profileSource = true
+	return nil
+}
+
+// SetFallbackProfileFS selects a fallback profile source rooted in fsys.
+func (o *Options) SetFallbackProfileFS(fsys fs.FS, root string) error {
+	if err := validateSourceFS(fsys, root); err != nil {
+		return err
+	}
+	o.fallbackProfiles = profile.NewFSRepository(fsys, root)
+	o.fallbackProfileSource = true
+	return nil
+}
+
+// SetMemoryProfiles validates and stores application-owned profile definitions.
+func (o *Options) SetMemoryProfiles(definitions []domain.ExecutionProfile) error {
+	repository, err := profile.NewMemoryRepository(definitions)
+	if err != nil {
+		return err
+	}
+	o.memoryProfiles = repository
+	o.memorySource = true
+	return nil
+}
+
+// AddBackend appends one backend registration.
+func (o *Options) AddBackend(backend domain.Backend) { o.backends = append(o.backends, backend) }
+
+// SetSchemaFS selects a schema source rooted in fsys.
+func (o *Options) SetSchemaFS(fsys fs.FS, root string) error {
+	if err := validateSourceFS(fsys, root); err != nil {
+		return err
+	}
+	o.validator = validate.NewFSValidator(fsys, root)
+	o.validatorSource = true
+	return nil
+}
+
+// SetSchemaFile selects a single schema file.
+func (o *Options) SetSchemaFile(path string) error {
+	fsys, root, _, err := fileSource(path)
+	if err != nil {
+		return err
+	}
+	o.validator = validate.NewFSValidator(fsys, root)
+	o.validatorSource = true
+	return nil
+}
+
+// Engine forwards domain operations to the single assembled use-case runner.
+type Engine struct {
+	runner *usecase.Runner
+}
+
+// New assembles an engine from facade-selected configuration and collaborators.
+func New(cfg Config, options Options) (*Engine, error) {
+	defaultRetries, err := domain.ResolveTransientRetries(2, cfg.TransientRetries)
+	if err != nil {
+		return nil, err
+	}
+	promptDefinitions := options.promptDefinitions
+	if !options.promptSource {
+		if strings.TrimSpace(cfg.PromptDir) == "" {
+			return nil, fmt.Errorf("prompt directory is required")
+		}
+		promptDefinitions = promptdef.NewFilesystemRepository(cfg.PromptDir)
+	}
+
+	maintained, err := catalog.Load(
+		catalog.Source{Name: "OpenRouter", ExpectedBackendID: backend.OpenRouterID, FS: openrouter.FS(), Root: openrouter.Root},
+		catalog.Source{Name: "Rakestrawhome", ExpectedBackendID: backend.RakestrawHomeID, FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("failed to load maintained catalogs: %w", err)
+	}
+
+	maintainedProfiles, err := newMaintainedProfileRepository(maintained.Profiles, maintained.Backends)
+	if err != nil {
+		return nil, fmt.Errorf("failed to construct Promptkit-maintained profiles: %w", err)
+	}
+
+	profiles := newProfileRepository(cfg.ProfileDir, options, maintainedProfiles)
+	registry, err := backend.NewRegistry(maintained.Backends, options.backends)
+	if err != nil {
+		return nil, fmt.Errorf("failed to construct backend registry: %w", err)
+	}
+	manager, err := capacity.NewManager(registry.CapacityPolicies())
+	if err != nil {
+		return nil, fmt.Errorf("failed to construct backend capacity manager: %w", err)
+	}
+
+	validator := options.validator
+	if !options.validatorSource {
+		schemaDir := cfg.SchemaDir
+		if strings.TrimSpace(schemaDir) == "" {
+			schemaDir = defaults.SchemaDirDefault
+		}
+		validator = validate.NewStandardValidator(schemaDir)
+	}
+	var client llm.Client
+	if options.llmClient == nil {
+		builtIn, newErr := llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{Timeout: cfg.Timeout, HTTPClient: cfg.HTTPClient})
+		if newErr != nil {
+			return nil, newErr
+		}
+		client = generation.NewBuiltIn(manager, builtIn)
+	} else {
+		client = generation.NewInjected(manager, options.llmClient)
+	}
+	reader := options.artifactReader
+	if !options.artifactSource {
+		reader = artifact.NewCompositeReader()
+	}
+
+	return &Engine{runner: usecase.NewRunnerWithRepairer(promptDefinitions, profiles, registry, reader, prompt.NewGoRenderer(), client, validator, usecase.NewDefaultOutputRepairer(client), manager, defaultRetries)}, nil
+}
+
+func newProfileRepository(profileDir string, options Options, maintained profile.Repository) profile.Repository {
+	repository := maintained
+	if options.fallbackProfileSource {
+		repository = profile.NewOverlayRepository(options.fallbackProfiles, repository)
+	}
+	if options.profileSource {
+		repository = profile.NewOverlayRepository(options.profiles, repository)
+	} else if strings.TrimSpace(profileDir) != "" {
+		repository = profile.NewOverlayRepository(profile.NewFilesystemRepository(profileDir), repository)
+	}
+	if options.memorySource {
+		repository = profile.NewOverlayRepository(options.memoryProfiles, repository)
+	}
+	return profile.NewResolvingRepository(repository)
+}
+
+func validateSourceFS(fsys fs.FS, root string) error {
+	if fsys == nil {
+		return fmt.Errorf("source filesystem is nil")
+	}
+	if strings.TrimSpace(root) == "" {
+		return fmt.Errorf("source root is required")
+	}
+	return nil
+}
+
+func fileSource(name string) (fs.FS, string, string, error) {
+	if strings.TrimSpace(name) == "" {
+		return nil, "", "", fmt.Errorf("source path is required")
+	}
+	directory := filepath.Dir(name)
+	base := filepath.Base(name)
+	if base == "." || base == string(filepath.Separator) {
+		return nil, "", "", fmt.Errorf("source path must name a file")
+	}
+	info, err := os.Stat(name)
+	if err != nil {
+		return nil, "", "", fmt.Errorf("failed to access source file %q: %w", name, err)
+	}
+	if info.IsDir() {
+		return nil, "", "", fmt.Errorf("source path %q must be a file", name)
+	}
+	return os.DirFS(directory), filepath.ToSlash(base), directory, nil
+}
+
+func (e *Engine) InspectPrompt(ctx context.Context, id, version string) (*domain.PromptInspection, error) {
+	return e.runner.InspectPrompt(ctx, id, version)
+}
+func (e *Engine) InspectProfile(ctx context.Context, id string) (*domain.ProfileInspection, error) {
+	return e.runner.InspectProfile(ctx, id)
+}
+func (e *Engine) Prepare(ctx context.Context, request domain.RunRequest) (*domain.PreparedRun, error) {
+	return e.runner.Prepare(ctx, request)
+}
+func (e *Engine) PrepareExecution(ctx context.Context, request domain.RunRequest) (*usecase.PreparedExecution, error) {
+	return e.runner.PrepareExecution(ctx, request)
+}
+func (e *Engine) Run(ctx context.Context, request domain.RunRequest) (*domain.RunResult, error) {
+	return e.runner.Run(ctx, request)
+}
+func (e *Engine) RunPrepared(ctx context.Context, prepared *usecase.PreparedExecution) (*domain.RunResult, error) {
+	return e.runner.RunPrepared(ctx, prepared)
+}

internal/engine/maintained_profiles.go

@@ -0,0 +1,50 @@
+package engine
+
+import (
+	"context"
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+)
+
+func maintainedProfileDefinitions() []domain.ExecutionProfile {
+	return []domain.ExecutionProfile{
+		{
+			ID:            "promptkit-openrouter-standard",
+			BaseProfileID: "gemini-flash-latest",
+		},
+		{
+			ID:        "promptkit-openrouter-auto",
+			BackendID: backend.OpenRouterID,
+			Model:     "openrouter/auto",
+		},
+	}
+}
+
+func newMaintainedProfileRepository(catalogProfiles profile.Repository, maintainedBackends []domain.Backend) (profile.Repository, error) {
+	definitions := maintainedProfileDefinitions()
+	owned, err := profile.NewMemoryRepository(definitions)
+	if err != nil {
+		return nil, fmt.Errorf("create Promptkit-maintained profiles: %w", err)
+	}
+
+	combined := profile.NewOverlayRepository(owned, catalogProfiles)
+	resolved := profile.NewResolvingRepository(combined)
+	registry, err := backend.NewRegistry(maintainedBackends, nil)
+	if err != nil {
+		return nil, fmt.Errorf("construct maintained backend registry: %w", err)
+	}
+	for _, definition := range definitions {
+		effective, err := resolved.GetProfile(context.Background(), definition.ID)
+		if err != nil {
+			return nil, fmt.Errorf("resolve Promptkit-maintained profile %q: %w", definition.ID, err)
+		}
+		if _, err := registry.GetBackend(effective.BackendID); err != nil {
+			return nil, fmt.Errorf("Promptkit-maintained profile %q selects unavailable backend %q: %w", definition.ID, effective.BackendID, err)
+		}
+	}
+
+	return combined, nil
+}

internal/engine/maintained_profiles_test.go

@@ -0,0 +1,78 @@
+package engine
+
+import (
+	"context"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+)
+
+func TestNewMaintainedProfileRepository(t *testing.T) {
+	newCatalogProfiles := func(t *testing.T, definitions []domain.ExecutionProfile) profile.Repository {
+		t.Helper()
+		repository, err := profile.NewMemoryRepository(definitions)
+		if err != nil {
+			t.Fatalf("create catalog profiles: %v", err)
+		}
+		return repository
+	}
+	newBackend := func(id string) domain.Backend {
+		return domain.Backend{ID: id, Endpoint: "https://" + id + ".example/v1", APIKeyEnv: "TEST_API_KEY"}
+	}
+
+	t.Run("resolves inherited and direct definitions", func(t *testing.T) {
+		repository, err := newMaintainedProfileRepository(
+			newCatalogProfiles(t, []domain.ExecutionProfile{{
+				ID:        "gemini-flash-latest",
+				BackendID: backend.OpenRouterID,
+				Model:     "catalog-selected-model",
+			}}),
+			[]domain.Backend{newBackend(backend.OpenRouterID)},
+		)
+		if err != nil {
+			t.Fatalf("create maintained profiles: %v", err)
+		}
+
+		resolved := profile.NewResolvingRepository(repository)
+		standard, err := resolved.GetProfile(context.Background(), "promptkit-openrouter-standard")
+		if err != nil {
+			t.Fatalf("resolve inherited profile: %v", err)
+		}
+		if standard.BackendID != backend.OpenRouterID || standard.Model != "catalog-selected-model" {
+			t.Fatalf("unexpected inherited profile: %#v", standard)
+		}
+
+		auto, err := resolved.GetProfile(context.Background(), "promptkit-openrouter-auto")
+		if err != nil {
+			t.Fatalf("resolve direct profile: %v", err)
+		}
+		definitions := maintainedProfileDefinitions()
+		if auto.BackendID != definitions[1].BackendID || auto.Model != definitions[1].Model {
+			t.Fatalf("unexpected direct profile: %#v", auto)
+		}
+	})
+
+	t.Run("rejects a missing inherited base", func(t *testing.T) {
+		if _, err := newMaintainedProfileRepository(
+			newCatalogProfiles(t, nil),
+			[]domain.Backend{newBackend(backend.OpenRouterID)},
+		); err == nil {
+			t.Fatal("expected missing base error")
+		}
+	})
+
+	t.Run("rejects an unavailable effective backend", func(t *testing.T) {
+		if _, err := newMaintainedProfileRepository(
+			newCatalogProfiles(t, []domain.ExecutionProfile{{
+				ID:        "gemini-flash-latest",
+				BackendID: "other",
+				Model:     "catalog-selected-model",
+			}}),
+			[]domain.Backend{newBackend("other")},
+		); err == nil {
+			t.Fatal("expected unavailable backend error")
+		}
+	})
+}

internal/filecatalog/catalog.go

@@ -36,7 +36,8 @@ func FindYAMLFiles(ctx context.Context, root string) ([]string, error) {
 	return files, err
 }
 
-// FindFSYAMLFiles returns sorted paths for .yaml and .yml files under root in fsys.
+// FindFSYAMLFiles returns root itself when it names a file. For a directory
+// root, it returns sorted paths for .yaml and .yml files beneath that root.
 func FindFSYAMLFiles(ctx context.Context, fsys fs.FS, root string) ([]string, error) {
 	cleanRoot := CleanFSRoot(root)
 	var files []string
@@ -52,6 +53,10 @@ func FindFSYAMLFiles(ctx context.Context, fsys fs.FS, root string) ([]string, er
 		if d.IsDir() {
 			return nil
 		}
+		if name == cleanRoot {
+			files = append(files, name)
+			return nil
+		}
 		if !IsYAMLFile(d.Name()) {
 			return nil
 		}
@@ -71,10 +76,10 @@ func RelativePath(root string, filePath string) string {
 	return filepath.Clean(rel)
 }
 
-// CleanFSRoot normalizes a root path for use with fs.FS.
+// CleanFSRoot normalizes a root path for use with fs.FS while preserving
+// nonblank leading and trailing whitespace.
 func CleanFSRoot(root string) string {
-	root = strings.TrimSpace(root)
+	if strings.TrimSpace(root) == "" || root == "." {
-	if root == "" || root == "." {
 		return "."
 	}
 	return path.Clean(root)
@@ -97,22 +102,21 @@ func DisplayPath(root string, name string) string {
 // ResolveFSPath resolves userPath from baseDir and keeps it inside root.
 func ResolveFSPath(root string, baseDir string, userPath string) (string, string, error) {
 	cleanRoot := CleanFSRoot(root)
-	cleanBase := path.Clean(strings.TrimSpace(baseDir))
+	cleanBase := path.Clean(baseDir)
-	if cleanBase == "" {
+	if strings.TrimSpace(baseDir) == "" {
 		cleanBase = cleanRoot
 	}
 	if !containsFSPath(cleanRoot, cleanBase) {
 		return "", "", fmt.Errorf("base path %q is outside source root %q", cleanBase, cleanRoot)
 	}
 
-	cleanUserPath := strings.TrimSpace(userPath)
+	if strings.TrimSpace(userPath) == "" {
-	if cleanUserPath == "" {
 		return "", "", fmt.Errorf("path is required")
 	}
-	cleanUserPath = path.Clean(cleanUserPath)
+	if path.IsAbs(userPath) {
-	if path.IsAbs(cleanUserPath) {
 		return "", "", fmt.Errorf("path %q must be relative", userPath)
 	}
+	cleanUserPath := path.Clean(userPath)
 
 	resolved := path.Clean(path.Join(cleanBase, cleanUserPath))
 	if !containsFSPath(cleanRoot, resolved) {
@@ -130,13 +134,6 @@ func containsFSPath(root string, name string) bool {
 	return name == root || strings.HasPrefix(name, strings.TrimSuffix(root, "/")+"/")
 }
 
-// Stem strips .yaml or .yml from a file name.
-func Stem(name string) string {
-	name = strings.TrimSuffix(name, ".yaml")
-	name = strings.TrimSuffix(name, ".yml")
-	return name
-}
-
 func IsYAMLFile(name string) bool {
 	return strings.HasSuffix(name, ".yaml") || strings.HasSuffix(name, ".yml")
 }

internal/filecatalog/catalog_test.go

@@ -54,7 +54,7 @@ func TestFindFSYAMLFilesNestedSortedAndFiltered(t *testing.T) {
 		"other/ignored.yaml":        &fstest.MapFile{Data: []byte("id: ignored")},
 	}
 
-	got, err := FindFSYAMLFiles(context.Background(), fsys, " prompts ")
+	got, err := FindFSYAMLFiles(context.Background(), fsys, "prompts")
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
 	}
@@ -98,8 +98,11 @@ func TestCleanFSRoot(t *testing.T) {
 		want string
 	}{
 		{name: "empty", root: "", want: "."},
+		{name: "whitespace only", root: " \t ", want: "."},
 		{name: "dot", root: ".", want: "."},
-		{name: "trimmed", root: " prompts/../profiles ", want: "profiles"},
+		{name: "cleaned", root: "prompts/../profiles", want: "profiles"},
+		{name: "leading whitespace preserved", root: " profiles", want: " profiles"},
+		{name: "trailing whitespace preserved", root: "profiles ", want: "profiles "},
 	}
 
 	for _, tc := range tests {
@@ -158,6 +161,22 @@ func TestResolveFSPath(t *testing.T) {
 			wantPath:    "prompts/shared/user.tmpl",
 			wantDisplay: "shared/user.tmpl",
 		},
+		{
+			name:        "leading whitespace preserved",
+			root:        "prompts",
+			baseDir:     "prompts/nested",
+			userPath:    " user.tmpl",
+			wantPath:    "prompts/nested/ user.tmpl",
+			wantDisplay: "nested/ user.tmpl",
+		},
+		{
+			name:        "trailing whitespace preserved",
+			root:        "prompts",
+			baseDir:     "prompts/nested",
+			userPath:    "user.tmpl ",
+			wantPath:    "prompts/nested/user.tmpl ",
+			wantDisplay: "nested/user.tmpl ",
+		},
 		{
 			name:     "escape rejected",
 			root:     "prompts",
@@ -218,26 +237,6 @@ func TestResolveFSPath(t *testing.T) {
 	}
 }
 
-func TestStemStripsYAMLExtensions(t *testing.T) {
-	tests := []struct {
-		name string
-		in   string
-		want string
-	}{
-		{name: "yaml", in: "prompt.yaml", want: "prompt"},
-		{name: "yml", in: "profile.yml", want: "profile"},
-		{name: "other", in: "file.txt", want: "file.txt"},
-	}
-
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			if got := Stem(tc.in); got != tc.want {
-				t.Fatalf("expected %q, got %q", tc.want, got)
-			}
-		})
-	}
-}
-
 func TestIsYAMLFile(t *testing.T) {
 	tests := []struct {
 		name string

internal/generation/coordinator.go

@@ -0,0 +1,294 @@
+// Package generation coordinates model attempts, backend permits, and bounded
+// built-in HTTP retries for one generation request.
+package generation
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"math/rand/v2"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+	"gitea.maximumdirect.net/eric/promptkit/internal/observation"
+)
+
+const (
+	initialBackoff = 500 * time.Millisecond
+	maximumBackoff = 8 * time.Second
+)
+
+// Coordinator implements the internal model client using the existing backend
+// capacity manager. Exactly one of builtIn and injected is selected.
+type Coordinator struct {
+	manager  *capacity.Manager
+	builtIn  *llm.OpenAICompatibleClient
+	injected llm.Client
+	now      func() time.Time
+	jitter   func(time.Duration) time.Duration
+	wait     func(context.Context, time.Duration) error
+}
+
+// NewBuiltIn selects the built-in HTTP retry path.
+func NewBuiltIn(manager *capacity.Manager, client *llm.OpenAICompatibleClient) *Coordinator {
+	return &Coordinator{manager: manager, builtIn: client, now: time.Now, jitter: randomJitter, wait: waitContext}
+}
+
+// NewInjected selects one opaque client dispatch with no automatic retry.
+func NewInjected(manager *capacity.Manager, client llm.Client) *Coordinator {
+	return &Coordinator{manager: manager, injected: client, now: time.Now, jitter: randomJitter, wait: waitContext}
+}
+
+// ExhaustedError retains the final eligible status failure and the number of
+// attempts actually dispatched.
+type ExhaustedError struct {
+	attempts uint64
+	cause    error
+}
+
+func (e *ExhaustedError) Error() string {
+	if e == nil {
+		return "transient model retries exhausted"
+	}
+	return fmt.Sprintf("transient model retries exhausted after %d attempts", e.attempts)
+}
+
+func (e *ExhaustedError) Unwrap() error {
+	if e == nil {
+		return nil
+	}
+	return e.cause
+}
+
+func (e *ExhaustedError) Attempts() uint64 {
+	if e == nil {
+		return 0
+	}
+	return e.attempts
+}
+
+// Generate executes either one opaque client call or a frozen built-in HTTP
+// sequence with one permit per actual dispatch.
+func (c *Coordinator) Generate(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	if c.injected != nil {
+		return c.generateInjected(ctx, req)
+	}
+	prepared, err := c.builtIn.PrepareGeneration(req)
+	if err != nil {
+		return nil, err
+	}
+
+	var attempts, retriesUsed uint64
+	recorder := observation.FromContext(ctx)
+	generation := req.Generation
+	if generation == 0 {
+		generation = 1
+	}
+	for {
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		ordinal := attempts + 1
+		outcome, dispatched, capacityWait, generationDuration := c.execute(ctx, req.Target.BackendID, prepared)
+		if dispatched {
+			attempts++
+		}
+		if recorder != nil {
+			recorder.Attempt(observation.AttemptData{Generation: generation, Attempt: ordinal,
+				Dispatched: dispatched, CapacityWait: capacityWait, GenerationDuration: generationDuration,
+				Outcome: attemptOutcome(outcome.Err), Failure: attemptFailure(outcome.Err),
+				StatusCode: outcome.StatusCode, Usage: outcome.Usage, UsagePresence: outcome.UsagePresence})
+		}
+		if outcome.Err == nil {
+			if err := ctx.Err(); err != nil {
+				return nil, err
+			}
+			return outcome.Response, nil
+		}
+		if !dispatched || !eligibleStatus(outcome) {
+			return nil, outcome.Err
+		}
+		if retriesUsed == req.TransientRetries {
+			if retriesUsed == 0 {
+				return nil, outcome.Err
+			}
+			return nil, &ExhaustedError{attempts: attempts, cause: outcome.Err}
+		}
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		retryOrdinal := retriesUsed + 1
+		delay := retryDelay(c.backoff(retryOrdinal), outcome.RetryNotBefore, c.now())
+		waitStart := c.now()
+		waitErr := c.wait(ctx, delay)
+		if recorder != nil {
+			recorder.Delay(observation.DelayData{Generation: generation, Attempt: attempts,
+				Elapsed: elapsed(waitStart, c.now()), Failure: attemptFailure(waitErr)})
+		}
+		if waitErr != nil {
+			return nil, waitErr
+		}
+		if err := ctx.Err(); err != nil {
+			return nil, err
+		}
+		retriesUsed++
+	}
+}
+
+func (c *Coordinator) generateInjected(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	recorder := observation.FromContext(ctx)
+	generation := req.Generation
+	if generation == 0 {
+		generation = 1
+	}
+	start := c.now()
+	release, err := c.manager.Acquire(ctx, req.Target.BackendID)
+	capacityWait := c.capacityWait(req.Target.BackendID, start)
+	if err != nil {
+		if recorder != nil {
+			recorder.Attempt(observation.AttemptData{Generation: generation, Attempt: 1, CapacityWait: capacityWait, Outcome: domain.ExecutionOutcomeError, Failure: attemptFailure(err)})
+		}
+		return nil, err
+	}
+	defer release()
+	if err := ctx.Err(); err != nil {
+		release()
+		if recorder != nil {
+			recorder.Attempt(observation.AttemptData{Generation: generation, Attempt: 1, CapacityWait: capacityWait, Outcome: domain.ExecutionOutcomeError, Failure: attemptFailure(err)})
+		}
+		return nil, err
+	}
+	dispatchStart := c.now()
+	response, err := c.injected.Generate(ctx, req)
+	duration := elapsed(dispatchStart, c.now())
+	if err == nil && response == nil {
+		err = llm.ErrMalformedResponse
+	}
+	release()
+	if recorder != nil {
+		data := observation.AttemptData{Generation: generation, Attempt: 1, Dispatched: true, CapacityWait: capacityWait, GenerationDuration: duration,
+			Outcome: attemptOutcome(err), Failure: attemptFailure(err)}
+		if err == nil {
+			data.Usage = response.Usage
+			data.UsagePresence = response.UsagePresence
+		}
+		recorder.Attempt(data)
+	}
+	if err != nil {
+		return nil, err
+	}
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	return response, nil
+}
+
+func (c *Coordinator) execute(ctx context.Context, backendID string, prepared *llm.PreparedGeneration) (outcome llm.AttemptOutcome, dispatched bool, capacityWait, generationDuration time.Duration) {
+	start := c.now()
+	release, err := c.manager.Acquire(ctx, backendID)
+	capacityWait = c.capacityWait(backendID, start)
+	if err != nil {
+		return llm.AttemptOutcome{Err: err}, false, capacityWait, 0
+	}
+	defer release()
+	if err := ctx.Err(); err != nil {
+		return llm.AttemptOutcome{Err: err}, false, capacityWait, 0
+	}
+	dispatchStart := c.now()
+	outcome = prepared.Execute(ctx)
+	return outcome, true, capacityWait, elapsed(dispatchStart, c.now())
+}
+
+func (c *Coordinator) capacityWait(backendID string, start time.Time) time.Duration {
+	if !c.manager.Limited(backendID) {
+		return 0
+	}
+	return elapsed(start, c.now())
+}
+
+func elapsed(start, end time.Time) time.Duration {
+	if d := end.Sub(start); d > 0 {
+		return d
+	}
+	return 0
+}
+
+func attemptOutcome(err error) domain.ExecutionOutcome {
+	if err == nil {
+		return domain.ExecutionOutcomeSuccess
+	}
+	return domain.ExecutionOutcomeError
+}
+
+func attemptFailure(err error) domain.ExecutionFailureKind {
+	if err == nil {
+		return domain.ExecutionFailureNone
+	}
+	if errors.Is(err, context.Canceled) {
+		return domain.ExecutionFailureCanceled
+	}
+	if errors.Is(err, context.DeadlineExceeded) {
+		return domain.ExecutionFailureDeadline
+	}
+	if errors.Is(err, llm.ErrUnexpectedStatus) {
+		return domain.ExecutionFailureProviderStatus
+	}
+	if errors.Is(err, llm.ErrTransport) {
+		return domain.ExecutionFailureTransport
+	}
+	if errors.Is(err, llm.ErrMalformedResponse) {
+		return domain.ExecutionFailureMalformedResponse
+	}
+	if errors.Is(err, llm.ErrInvalidRequest) {
+		return domain.ExecutionFailureInvalidRequest
+	}
+	return domain.ExecutionFailureGeneration
+}
+
+func eligibleStatus(outcome llm.AttemptOutcome) bool {
+	if !errors.Is(outcome.Err, llm.ErrUnexpectedStatus) {
+		return false
+	}
+	switch outcome.StatusCode {
+	case 429, 500, 502, 503, 504:
+		return true
+	default:
+		return false
+	}
+}
+
+func (c *Coordinator) backoff(ordinal uint64) time.Duration {
+	base := initialBackoff
+	for step := uint64(1); step < ordinal && base < maximumBackoff; step++ {
+		base *= 2
+	}
+	half := base / 2
+	return half + c.jitter(half)
+}
+
+func retryDelay(local time.Duration, notBefore, now time.Time) time.Duration {
+	if providerDelay := notBefore.Sub(now); providerDelay > local {
+		return providerDelay
+	}
+	return local
+}
+
+func randomJitter(max time.Duration) time.Duration {
+	return time.Duration(rand.Int64N(int64(max) + 1))
+}
+
+func waitContext(ctx context.Context, delay time.Duration) error {
+	if err := ctx.Err(); err != nil {
+		return err
+	}
+	timer := time.NewTimer(delay)
+	defer timer.Stop()
+	select {
+	case <-ctx.Done():
+		return ctx.Err()
+	case <-timer.C:
+		return ctx.Err()
+	}
+}

internal/generation/coordinator_test.go

@@ -0,0 +1,278 @@
+package generation
+
+import (
+	"context"
+	"errors"
+	"io"
+	"math"
+	"net/http"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+)
+
+type transportFunc func(*http.Request) (*http.Response, error)
+
+func (f transportFunc) RoundTrip(req *http.Request) (*http.Response, error) { return f(req) }
+
+func response(status int, body string) *http.Response {
+	return &http.Response{StatusCode: status, Body: io.NopCloser(strings.NewReader(body)), Header: make(http.Header)}
+}
+
+func testRequest(retries uint64) domain.GenerateRequest {
+	return domain.GenerateRequest{
+		Prompt:           domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hello"}}},
+		Target:           domain.ExecutionTarget{BackendID: "limited", Endpoint: "https://example.invalid/v1", Model: "model"},
+		TransientRetries: retries,
+	}
+}
+
+func testCoordinator(t *testing.T, transport transportFunc) *Coordinator {
+	t.Helper()
+	manager, err := capacity.NewManager(map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1, QueueCapacity: 1},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	client, err := llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{HTTPClient: &http.Client{Transport: transport}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	coordinator := NewBuiltIn(manager, client)
+	coordinator.jitter = func(time.Duration) time.Duration { return 0 }
+	coordinator.wait = func(ctx context.Context, _ time.Duration) error { return ctx.Err() }
+	return coordinator
+}
+
+func TestBuiltInRetryEligibilityAndExhaustion(t *testing.T) {
+	for _, tc := range []struct {
+		status   int
+		eligible bool
+	}{
+		{status: 429, eligible: true}, {status: 500, eligible: true},
+		{status: 502, eligible: true}, {status: 503, eligible: true},
+		{status: 504, eligible: true}, {status: 400}, {status: 501},
+	} {
+		t.Run(http.StatusText(tc.status), func(t *testing.T) {
+			calls := 0
+			coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) {
+				calls++
+				if calls == 1 {
+					return response(tc.status, `{"error":{}}`), nil
+				}
+				return response(200, `{"choices":[{"message":{"content":"ok"}}]}`), nil
+			})
+			result, err := coordinator.Generate(context.Background(), testRequest(1))
+			if tc.eligible {
+				if err != nil || result == nil || result.Content != "ok" || calls != 2 {
+					t.Fatalf("result=%+v, error=%v, calls=%d", result, err, calls)
+				}
+			} else if result != nil || err == nil || calls != 1 {
+				t.Fatalf("result=%+v, error=%v, calls=%d", result, err, calls)
+			}
+		})
+	}
+
+	for _, budget := range []uint64{0, 2} {
+		calls := 0
+		coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) {
+			calls++
+			return response(503, `{"error":{}}`), nil
+		})
+		_, err := coordinator.Generate(context.Background(), testRequest(budget))
+		var exhausted *ExhaustedError
+		if !errors.Is(err, llm.ErrUnexpectedStatus) || calls != int(budget+1) || errors.As(err, &exhausted) != (budget > 0) {
+			t.Fatalf("budget=%d calls=%d error=%v", budget, calls, err)
+		}
+		if budget > 0 && exhausted.Attempts() != budget+1 {
+			t.Fatalf("attempts=%d", exhausted.Attempts())
+		}
+	}
+
+	calls := 0
+	coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) {
+		calls++
+		if calls == 1 {
+			return response(503, `{"error":{}}`), nil
+		}
+		return response(400, `{"error":{}}`), nil
+	})
+	_, err := coordinator.Generate(context.Background(), testRequest(3))
+	var exhausted *ExhaustedError
+	if calls != 2 || !errors.Is(err, llm.ErrUnexpectedStatus) || errors.As(err, &exhausted) {
+		t.Fatalf("later permanent failure: calls=%d, error=%v", calls, err)
+	}
+}
+
+func TestBuiltInRetryWaitHonorsContextAndProviderMinimum(t *testing.T) {
+	when := time.Date(2026, 9, 23, 12, 0, 0, 0, time.UTC)
+	coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) { return response(503, `{"error":{}}`), nil })
+	if got := coordinator.backoff(1); got != 250*time.Millisecond {
+		t.Fatalf("first backoff=%v", got)
+	}
+	if got := coordinator.backoff(math.MaxUint64); got != 4*time.Second {
+		t.Fatalf("capped backoff=%v", got)
+	}
+	coordinator.jitter = func(max time.Duration) time.Duration { return max }
+	if got := coordinator.backoff(math.MaxUint64); got != 8*time.Second {
+		t.Fatalf("maximum equal jitter=%v", got)
+	}
+	if got := retryDelay(500*time.Millisecond, when.Add(3*time.Second), when); got != 3*time.Second {
+		t.Fatalf("provider minimum=%v", got)
+	}
+	if got := retryDelay(500*time.Millisecond, when.Add(-time.Second), when); got != 500*time.Millisecond {
+		t.Fatalf("past provider minimum=%v", got)
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	calls := 0
+	coordinator = testCoordinator(t, func(*http.Request) (*http.Response, error) {
+		calls++
+		return response(503, `{"error":{}}`), nil
+	})
+	coordinator.wait = func(ctx context.Context, delay time.Duration) error {
+		if delay < 250*time.Millisecond {
+			t.Fatalf("delay=%v", delay)
+		}
+		cancel()
+		return ctx.Err()
+	}
+	_, err := coordinator.Generate(ctx, testRequest(2))
+	if !errors.Is(err, context.Canceled) || calls != 1 {
+		t.Fatalf("canceled wait: error=%v calls=%d", err, calls)
+	}
+}
+
+func TestBuiltInRetryUsesSeparateAttemptDeadlines(t *testing.T) {
+	var contexts []context.Context
+	coordinator := testCoordinator(t, func(req *http.Request) (*http.Response, error) {
+		contexts = append(contexts, req.Context())
+		if len(contexts) == 1 {
+			return response(503, `{"error":{}}`), nil
+		}
+		return response(200, `{"choices":[{"message":{"content":"ok"}}]}`), nil
+	})
+	req := testRequest(1)
+	req.Target.TimeoutSeconds = 1
+	if _, err := coordinator.Generate(context.Background(), req); err != nil {
+		t.Fatal(err)
+	}
+	if len(contexts) != 2 || contexts[0] == contexts[1] {
+		t.Fatalf("attempt contexts=%d", len(contexts))
+	}
+	for _, ctx := range contexts {
+		if _, ok := ctx.Deadline(); !ok || !errors.Is(ctx.Err(), context.Canceled) {
+			t.Fatalf("attempt deadline/error=%t/%v", ok, ctx.Err())
+		}
+	}
+}
+
+func TestCanceledCapacityWaitDispatchesNothing(t *testing.T) {
+	calls := 0
+	coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) {
+		calls++
+		return response(200, `{"choices":[{"message":{"content":"ok"}}]}`), nil
+	})
+	release, err := coordinator.manager.Acquire(context.Background(), "limited")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer release()
+	ctx, cancel := context.WithCancel(context.Background())
+	result := make(chan error, 1)
+	go func() { _, err := coordinator.Generate(ctx, testRequest(2)); result <- err }()
+	cancel()
+	if err := <-result; !errors.Is(err, context.Canceled) || calls != 0 {
+		t.Fatalf("capacity cancellation: error=%v calls=%d", err, calls)
+	}
+}
+
+func TestBuiltInRetryFreesPermitAndFreezesCredential(t *testing.T) {
+	const envName = "PROMPTKIT_GENERATION_RETRY_TEST_KEY"
+	t.Setenv(envName, "first-key")
+	var calls atomic.Int64
+	var firstAuth, retryAuth string
+	coordinator := testCoordinator(t, func(req *http.Request) (*http.Response, error) {
+		switch calls.Add(1) {
+		case 1:
+			firstAuth = req.Header.Get("Authorization")
+			return response(503, `{"error":{}}`), nil
+		case 2:
+			return response(200, `{"choices":[{"message":{"content":"other"}}]}`), nil
+		default:
+			retryAuth = req.Header.Get("Authorization")
+			return response(200, `{"choices":[{"message":{"content":"retry"}}]}`), nil
+		}
+	})
+	entered := make(chan struct{})
+	resume := make(chan struct{})
+	coordinator.wait = func(ctx context.Context, _ time.Duration) error {
+		close(entered)
+		select {
+		case <-resume:
+			return ctx.Err()
+		case <-ctx.Done():
+			return ctx.Err()
+		}
+	}
+	req := testRequest(1)
+	req.Target.APIKeyEnv = envName
+	first := make(chan error, 1)
+	go func() { _, err := coordinator.Generate(context.Background(), req); first <- err }()
+	<-entered
+	t.Setenv(envName, "later-key")
+	other, err := coordinator.Generate(context.Background(), testRequest(0))
+	if err != nil || other == nil || other.Content != "other" {
+		t.Fatalf("other call while backoff blocked = %+v, %v", other, err)
+	}
+	close(resume)
+	if err := <-first; err != nil {
+		t.Fatal(err)
+	}
+	if calls.Load() != 3 || firstAuth != "Bearer first-key" || retryAuth != firstAuth {
+		t.Fatalf("calls=%d, first auth=%q, retry auth=%q", calls.Load(), firstAuth, retryAuth)
+	}
+}
+
+func TestInjectedClientDispatchesOnceAndReleasesAfterPanic(t *testing.T) {
+	manager, err := capacity.NewManager(map[string]domain.BackendCapacityPolicy{"limited": {ConcurrencyLimit: 1}})
+	if err != nil {
+		t.Fatal(err)
+	}
+	client := &injectedClient{panicFirst: true}
+	coordinator := NewInjected(manager, client)
+	func() {
+		defer func() {
+			if recover() == nil {
+				t.Fatal("expected injected panic")
+			}
+		}()
+		_, _ = coordinator.Generate(context.Background(), testRequest(5))
+	}()
+	ctx, cancel := context.WithTimeout(context.Background(), 5*time.Second)
+	defer cancel()
+	result, err := coordinator.Generate(ctx, testRequest(5))
+	if err != nil || result == nil || client.calls != 2 {
+		t.Fatalf("after panic result=%+v error=%v calls=%d", result, err, client.calls)
+	}
+}
+
+type injectedClient struct {
+	calls      int
+	panicFirst bool
+}
+
+func (c *injectedClient) Generate(context.Context, domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	c.calls++
+	if c.panicFirst && c.calls == 1 {
+		panic("injected panic")
+	}
+	return &domain.GenerateResponse{Content: "ok"}, nil
+}

internal/generation/observation_test.go

@@ -0,0 +1,201 @@
+package generation
+
+import (
+	"context"
+	"errors"
+	"net/http"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/observation"
+)
+
+func recordEvents(t *testing.T, ctx context.Context, events *[]domain.ExecutionEvent) context.Context {
+	t.Helper()
+	recorder, err := observation.New(func(event domain.ExecutionEvent) { *events = append(*events, event) })
+	if err != nil {
+		t.Fatal(err)
+	}
+	return observation.WithRecorder(ctx, recorder)
+}
+
+func TestObservedRetryAndExhaustion(t *testing.T) {
+	for _, status := range []int{200, 503} {
+		calls := 0
+		coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) {
+			calls++
+			if calls == 1 || status == 503 {
+				return response(503, `{"error":{}}`), nil
+			}
+			return response(200, `{"choices":[{"message":{"content":"ok"}}],"usage":{"prompt_tokens":0,"completion_tokens":2}}`), nil
+		})
+		var events []domain.ExecutionEvent
+		ctx := recordEvents(t, context.Background(), &events)
+		_, err := coordinator.Generate(ctx, testRequest(1))
+		if status == 200 && err != nil {
+			t.Fatal(err)
+		}
+		if status == 503 {
+			var exhausted *ExhaustedError
+			if !errors.As(err, &exhausted) {
+				t.Fatalf("error=%v", err)
+			}
+		}
+		if len(events) != 3 || events[0].Kind != domain.ExecutionAttempt || events[1].Kind != domain.ExecutionRetryDelay || events[2].Kind != domain.ExecutionAttempt {
+			t.Fatalf("events=%+v", events)
+		}
+		if events[0].Attempt != 1 || events[1].Attempt != 1 || events[2].Attempt != 2 || !events[0].Dispatched || !events[2].Dispatched {
+			t.Fatalf("indexes=%+v", events)
+		}
+		if events[0].Failure != domain.ExecutionFailureProviderStatus || events[1].RetryDelay < 0 {
+			t.Fatalf("first=%+v delay=%+v", events[0], events[1])
+		}
+		if status == 200 && (!events[2].Usage.PromptTokens.Known || events[2].Usage.PromptTokens.Value != 0 || events[2].Usage.TotalTokens.Known) {
+			t.Fatalf("usage=%+v", events[2].Usage)
+		}
+	}
+}
+
+func TestObservedInterruptedWaitAndCanceledCapacity(t *testing.T) {
+	coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) { return response(503, `{"error":{}}`), nil })
+	ctx, cancel := context.WithCancel(context.Background())
+	var events []domain.ExecutionEvent
+	ctx = recordEvents(t, ctx, &events)
+	coordinator.wait = func(ctx context.Context, _ time.Duration) error { cancel(); return ctx.Err() }
+	_, err := coordinator.Generate(ctx, testRequest(1))
+	if !errors.Is(err, context.Canceled) || len(events) != 2 || events[1].Kind != domain.ExecutionRetryDelay || events[1].Failure != domain.ExecutionFailureCanceled {
+		t.Fatalf("error=%v events=%+v", err, events)
+	}
+
+	coordinator = testCoordinator(t, func(*http.Request) (*http.Response, error) { t.Fatal("unexpected dispatch"); return nil, nil })
+	release, err := coordinator.manager.Acquire(context.Background(), "limited")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer release()
+	ctx, cancel = context.WithCancel(context.Background())
+	events = nil
+	ctx = recordEvents(t, ctx, &events)
+	entered := make(chan struct{})
+	ctx = &notifyingContext{Context: ctx, entered: entered}
+	done := make(chan error, 1)
+	go func() { _, err := coordinator.Generate(ctx, testRequest(0)); done <- err }()
+	<-entered
+	cancel()
+	if err := <-done; !errors.Is(err, context.Canceled) {
+		t.Fatalf("error=%v", err)
+	}
+	if len(events) != 1 || events[0].Dispatched || events[0].Failure != domain.ExecutionFailureCanceled {
+		t.Fatalf("events=%+v", events)
+	}
+}
+
+type notifyingContext struct {
+	context.Context
+	entered chan struct{}
+}
+
+func (c *notifyingContext) Done() <-chan struct{} {
+	select {
+	case <-c.entered:
+	default:
+		close(c.entered)
+	}
+	return c.Context.Done()
+}
+
+func TestObservedCallbackReleasesPermitAndCanCancelRetry(t *testing.T) {
+	coordinator := testCoordinator(t, func(*http.Request) (*http.Response, error) { return response(503, `{"error":{}}`), nil })
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+	var events []domain.ExecutionEvent
+	recorder, err := observation.New(func(event domain.ExecutionEvent) {
+		events = append(events, event)
+		if event.Kind != domain.ExecutionAttempt {
+			return
+		}
+		permitCtx, stop := context.WithTimeout(context.Background(), time.Second)
+		defer stop()
+		release, err := coordinator.manager.Acquire(permitCtx, "limited")
+		if err != nil {
+			t.Errorf("callback could not acquire permit: %v", err)
+		} else {
+			release()
+		}
+		cancel()
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	_, err = coordinator.Generate(observation.WithRecorder(ctx, recorder), testRequest(2))
+	if !errors.Is(err, context.Canceled) || len(events) != 1 {
+		t.Fatalf("error=%v events=%+v", err, events)
+	}
+}
+
+type fixedClient struct {
+	response *domain.GenerateResponse
+	err      error
+}
+
+func (c fixedClient) Generate(context.Context, domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	return c.response, c.err
+}
+
+func TestObservedInjectedUsageAndRepairGeneration(t *testing.T) {
+	manager, err := capacity.NewManager(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var events []domain.ExecutionEvent
+	ctx := recordEvents(t, context.Background(), &events)
+	legacy := NewInjected(manager, fixedClient{response: &domain.GenerateResponse{Usage: domain.TokenUsage{PromptTokens: 3}}})
+	if _, err := legacy.Generate(ctx, testRequest(4)); err != nil {
+		t.Fatal(err)
+	}
+	presence := &domain.TokenUsagePresence{CompletionTokens: true}
+	repair := NewInjected(manager, fixedClient{response: &domain.GenerateResponse{UsagePresence: presence}})
+	req := testRequest(4)
+	req.Generation = 2
+	if _, err := repair.Generate(ctx, req); err != nil {
+		t.Fatal(err)
+	}
+	if len(events) != 2 || !events[0].Usage.PromptTokens.Known || events[0].Usage.CompletionTokens.Known || !events[1].Usage.CompletionTokens.Known || events[1].Usage.CompletionTokens.Value != 0 || events[1].Usage.PromptTokens.Known {
+		t.Fatalf("events=%+v", events)
+	}
+	if events[1].Generation != 2 || events[1].Attempt != 1 {
+		t.Fatalf("repair indexes=%+v", events[1])
+	}
+}
+
+func TestObservedSuccessfulGenerationBeforeRepairFailure(t *testing.T) {
+	manager, err := capacity.NewManager(nil)
+	if err != nil {
+		t.Fatal(err)
+	}
+	var events []domain.ExecutionEvent
+	recorder, err := observation.New(func(event domain.ExecutionEvent) { events = append(events, event) })
+	if err != nil {
+		t.Fatal(err)
+	}
+	ctx := observation.WithRecorder(context.Background(), recorder)
+	initial := NewInjected(manager, fixedClient{response: &domain.GenerateResponse{Usage: domain.TokenUsage{PromptTokens: 5}}})
+	if _, err := initial.Generate(ctx, testRequest(0)); err != nil {
+		t.Fatal(err)
+	}
+	repair := NewInjected(manager, fixedClient{err: errors.New("repair unavailable")})
+	req := testRequest(0)
+	req.Generation = 2
+	if _, err := repair.Generate(ctx, req); err == nil {
+		t.Fatal("expected repair failure")
+	}
+	recorder.Finish(observation.FinishData{Outcome: domain.ExecutionOutcomeError, Failure: domain.ExecutionFailureGeneration})
+	if len(events) != 3 || events[0].Outcome != domain.ExecutionOutcomeSuccess || events[1].Outcome != domain.ExecutionOutcomeError || events[2].Attempts != 2 || events[2].RepairCalls != 1 {
+		t.Fatalf("events=%+v", events)
+	}
+	if got := events[2].Usage.PromptTokens; got.Value != 5 || !got.Known || got.Complete {
+		t.Fatalf("prompt usage=%+v", got)
+	}
+}

internal/jsonvalue/jsonvalue.go

@@ -1,5 +1,5 @@
-// Package jsonvalue validates and defensively copies JSON-compatible value
+// Package jsonvalue validates and defensively copies bounded JSON-compatible
-// trees used by public configuration and request boundaries.
+// value trees used by configuration, request, and prepared-state boundaries.
 package jsonvalue
 
 import (
@@ -8,23 +8,39 @@ import (
 	"math"
 	"reflect"
 	"sort"
-	"strconv"
 )
 
-const maxSafeJSONInteger = 1<<53 - 1
+const (
+	maxContainerDepth = 100
+	maxProducedNodes  = 100_000
+)
 
 type visit struct {
 	typ reflect.Type
 	ptr uintptr
 }
 
+type traversalState struct {
+	active        map[visit]struct{}
+	producedNodes int
+}
+
+// Copy validates and deeply copies a JSON-compatible value while preserving
+// compatible concrete map, slice, array, scalar, and number types. It rejects
+// cycles and values that exceed the package's traversal limits.
+func Copy(src any) (any, error) {
+	return copyValue(reflect.ValueOf(src), "value", newTraversalState(), true, 0)
+}
+
 // CopyMap validates and deeply copies an extra-parameter map while preserving
-// compatible concrete map, slice, array, scalar, and number types.
+// compatible concrete map, slice, array, scalar, and number types. It rejects
+// empty object keys, cycles, and values that exceed the package's traversal
+// limits.
 func CopyMap(src map[string]any) (map[string]any, error) {
 	if src == nil {
 		return nil, nil
 	}
-	copied, err := copyValue(reflect.ValueOf(src), "extra_params", make(map[visit]struct{}))
+	copied, err := copyValue(reflect.ValueOf(src), "extra_params", newTraversalState(), false, 0)
 	if err != nil {
 		return nil, err
 	}
@@ -35,88 +51,121 @@ func CopyMap(src map[string]any) (map[string]any, error) {
 	return out, nil
 }
 
-func copyValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+func copyValue(
-	if !value.IsValid() {
+	value reflect.Value,
+	path string,
+	state *traversalState,
+	allowEmptyMapKeys bool,
+	containerDepth int,
+) (any, error) {
+	resolved, cleanup, isNull, err := state.resolveIndirection(value, path)
+	if err != nil {
+		return nil, err
+	}
+	defer cleanup()
+	if isNull {
+		if err := state.produceNode(path); err != nil {
+			return nil, err
+		}
 		return nil, nil
 	}
-	if value.Kind() == reflect.Interface {
+	value = resolved
-		if value.IsNil() {
-			return nil, nil
-		}
-		return copyValue(value.Elem(), path, seen)
-	}
 	if !value.CanInterface() {
 		return nil, fmt.Errorf("%s: value cannot be copied", path)
 	}
 	if number, ok := value.Interface().(json.Number); ok {
-		if _, err := json.Marshal(number); err != nil {
+		if !validJSONNumber(number) {
 			return nil, fmt.Errorf("%s: invalid JSON number", path)
 		}
-		f, err := strconv.ParseFloat(number.String(), 64)
+		if err := state.produceNode(path); err != nil {
-		if err != nil || math.IsNaN(f) || math.IsInf(f, 0) {
+			return nil, err
-			return nil, fmt.Errorf("%s: invalid JSON number", path)
 		}
 		return number, nil
 	}
 
 	switch value.Kind() {
 	case reflect.Bool, reflect.String:
+		if err := state.produceNode(path); err != nil {
+			return nil, err
+		}
 		return value.Interface(), nil
 	case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
-		if value.Int() < -maxSafeJSONInteger || value.Int() > maxSafeJSONInteger {
+		if err := state.produceNode(path); err != nil {
-			return nil, fmt.Errorf("%s: integer is outside the JSON-safe range", path)
+			return nil, err
 		}
 		return value.Interface(), nil
 	case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
-		if value.Uint() > maxSafeJSONInteger {
+		if err := state.produceNode(path); err != nil {
-			return nil, fmt.Errorf("%s: integer is outside the JSON-safe range", path)
+			return nil, err
 		}
 		return value.Interface(), nil
 	case reflect.Float32, reflect.Float64:
-		number := value.Convert(reflect.TypeOf(float64(0))).Float()
+		number := value.Float()
 		if math.IsNaN(number) || math.IsInf(number, 0) {
 			return nil, fmt.Errorf("%s: floating-point value must be finite", path)
 		}
+		if err := state.produceNode(path); err != nil {
+			return nil, err
+		}
 		return value.Interface(), nil
-	case reflect.Pointer:
+	case reflect.Map:
 		if value.IsNil() {
+			if err := state.produceNode(path); err != nil {
+				return nil, err
+			}
 			return nil, nil
 		}
-		current := visit{typ: value.Type(), ptr: value.Pointer()}
+		nextDepth, err := state.enterContainer(path, containerDepth)
-		if _, ok := seen[current]; ok {
+		if err != nil {
-			return nil, fmt.Errorf("%s: cyclic value is not supported", path)
+			return nil, err
 		}
-		seen[current] = struct{}{}
+		return copyMapValue(value, path, state, allowEmptyMapKeys, nextDepth)
-		defer delete(seen, current)
-		return copyValue(value.Elem(), path, seen)
-	case reflect.Map:
-		return copyMapValue(value, path, seen)
 	case reflect.Slice:
 		if value.IsNil() {
+			if err := state.produceNode(path); err != nil {
+				return nil, err
+			}
 			return nil, nil
 		}
-		return copySequenceValue(value, path, seen)
+		nextDepth, err := state.enterContainer(path, containerDepth)
+		if err != nil {
+			return nil, err
+		}
+		return copySequenceValue(value, path, state, allowEmptyMapKeys, nextDepth)
 	case reflect.Array:
-		return copySequenceValue(value, path, seen)
+		nextDepth, err := state.enterContainer(path, containerDepth)
+		if err != nil {
+			return nil, err
+		}
+		return copySequenceValue(value, path, state, allowEmptyMapKeys, nextDepth)
 	default:
 		return nil, fmt.Errorf("%s: unsupported JSON value type %s", path, value.Type())
 	}
 }
 
-func copyMapValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+func copyMapValue(
-	if value.IsNil() {
+	value reflect.Value,
-		return nil, nil
+	path string,
-	}
+	state *traversalState,
+	allowEmptyMapKeys bool,
+	containerDepth int,
+) (any, error) {
 	if value.Type().Key().Kind() != reflect.String {
 		return nil, fmt.Errorf("%s: map key type %s is not supported", path, value.Type().Key())
 	}
+	if err := state.produceNode(path); err != nil {
+		return nil, err
+	}
+	if err := state.ensureChildCapacity(path, value.Len()); err != nil {
+		return nil, err
+	}
 
 	current := visit{typ: value.Type(), ptr: value.Pointer()}
-	if _, ok := seen[current]; ok {
+	if _, ok := state.active[current]; ok {
 		return nil, fmt.Errorf("%s: cyclic value is not supported", path)
 	}
-	seen[current] = struct{}{}
+	state.active[current] = struct{}{}
-	defer delete(seen, current)
+	defer delete(state.active, current)
 
 	keys := value.MapKeys()
 	sort.Slice(keys, func(i, j int) bool {
@@ -133,10 +182,16 @@ func copyMapValue(value reflect.Value, path string, seen map[visit]struct{}) (an
 	elementType := value.Type().Elem()
 	for _, key := range keys {
 		name := key.String()
-		if name == "" {
+		if name == "" && !allowEmptyMapKeys {
 			return nil, fmt.Errorf("%s: map key must not be empty", path)
 		}
-		copied, err := copyValue(value.MapIndex(key), path+"."+name, seen)
+		copied, err := copyValue(
+			value.MapIndex(key),
+			path+"."+name,
+			state,
+			allowEmptyMapKeys,
+			containerDepth,
+		)
 		if err != nil {
 			return nil, err
 		}
@@ -171,22 +226,41 @@ func copyMapValue(value reflect.Value, path string, seen map[visit]struct{}) (an
 	return out, nil
 }
 
-func copySequenceValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+func copySequenceValue(
+	value reflect.Value,
+	path string,
+	state *traversalState,
+	allowEmptyMapKeys bool,
+	containerDepth int,
+) (any, error) {
+	if err := state.produceNode(path); err != nil {
+		return nil, err
+	}
+	if err := state.ensureChildCapacity(path, value.Len()); err != nil {
+		return nil, err
+	}
+
 	var current visit
 	if value.Kind() == reflect.Slice {
 		current = visit{typ: value.Type(), ptr: value.Pointer()}
-		if _, ok := seen[current]; ok {
+		if _, ok := state.active[current]; ok {
 			return nil, fmt.Errorf("%s: cyclic value is not supported", path)
 		}
-		seen[current] = struct{}{}
+		state.active[current] = struct{}{}
-		defer delete(seen, current)
+		defer delete(state.active, current)
 	}
 
 	values := make([]any, value.Len())
 	preserveType := true
 	elementType := value.Type().Elem()
 	for i := 0; i < value.Len(); i++ {
-		copied, err := copyValue(value.Index(i), fmt.Sprintf("%s[%d]", path, i), seen)
+		copied, err := copyValue(
+			value.Index(i),
+			fmt.Sprintf("%s[%d]", path, i),
+			state,
+			allowEmptyMapKeys,
+			containerDepth,
+		)
 		if err != nil {
 			return nil, err
 		}
@@ -222,6 +296,73 @@ func copySequenceValue(value reflect.Value, path string, seen map[visit]struct{}
 	return out, nil
 }
 
+func validJSONNumber(number json.Number) bool {
+	var parsed json.Number
+	if err := json.Unmarshal([]byte(number.String()), &parsed); err != nil {
+		return false
+	}
+	return parsed.String() == number.String()
+}
+
+func newTraversalState() *traversalState {
+	return &traversalState{active: make(map[visit]struct{})}
+}
+
+func (state *traversalState) produceNode(path string) error {
+	if state.producedNodes >= maxProducedNodes {
+		return fmt.Errorf("%s: JSON value work limit exceeded", path)
+	}
+	state.producedNodes++
+	return nil
+}
+
+func (state *traversalState) enterContainer(path string, depth int) (int, error) {
+	depth++
+	if depth > maxContainerDepth {
+		return 0, fmt.Errorf("%s: JSON container depth limit exceeded", path)
+	}
+	return depth, nil
+}
+
+func (state *traversalState) ensureChildCapacity(path string, count int) error {
+	if count > maxProducedNodes-state.producedNodes {
+		return fmt.Errorf("%s: JSON value work limit exceeded", path)
+	}
+	return nil
+}
+
+func (state *traversalState) resolveIndirection(
+	value reflect.Value,
+	path string,
+) (reflect.Value, func(), bool, error) {
+	var visits []visit
+	cleanup := func() {
+		for _, current := range visits {
+			delete(state.active, current)
+		}
+	}
+
+	for value.IsValid() && (value.Kind() == reflect.Interface || value.Kind() == reflect.Pointer) {
+		if value.IsNil() {
+			return reflect.Value{}, cleanup, true, nil
+		}
+		if value.Kind() == reflect.Pointer {
+			current := visit{typ: value.Type(), ptr: value.Pointer()}
+			if _, ok := state.active[current]; ok {
+				cleanup()
+				return reflect.Value{}, nil, false, fmt.Errorf("%s: cyclic value is not supported", path)
+			}
+			state.active[current] = struct{}{}
+			visits = append(visits, current)
+		}
+		value = value.Elem()
+	}
+	if !value.IsValid() {
+		return reflect.Value{}, cleanup, true, nil
+	}
+	return value, cleanup, false, nil
+}
+
 func canAssignNil(typ reflect.Type) bool {
 	switch typ.Kind() {
 	case reflect.Chan, reflect.Func, reflect.Interface, reflect.Map, reflect.Pointer, reflect.Slice:

internal/jsonvalue/jsonvalue_test.go

@@ -1,97 +1,332 @@
-package jsonvalue_test
+package jsonvalue
 
 import (
 	"encoding/json"
 	"math"
 	"reflect"
+	"strings"
 	"testing"
-
-	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
 )
 
-func TestCopyMapPreservesTypesAndIsolatesMutations(t *testing.T) {
+type (
-	nested := map[string]int{"limit": 2}
+	namedBool    bool
-	sequence := []string{"one", "two"}
+	namedString  string
-	input := map[string]any{
+	namedInt64   int64
-		"count":    int64(7),
+	namedUint64  uint64
-		"number":   json.Number("-1.25e+2"),
+	namedFloat32 float32
-		"nested":   nested,
+	namedFloat64 float64
-		"sequence": sequence,
+	namedKey     string
-	}
+	namedMap     map[namedKey]namedInt64
-
+	namedSlice   []namedString
-	copied, err := jsonvalue.CopyMap(input)
+	namedArray   [1]map[string]int
-	if err != nil {
+)
-		t.Fatalf("copy map: %v", err)
-	}
-	nested["limit"] = 99
-	sequence[0] = "changed"
-	input["added"] = true
-
-	if got, ok := copied["count"].(int64); !ok || got != 7 {
-		t.Fatalf("integer type or value changed: %#v", copied["count"])
-	}
-	if got, ok := copied["number"].(json.Number); !ok || got != "-1.25e+2" {
-		t.Fatalf("JSON number type or value changed: %#v", copied["number"])
-	}
-	if got := copied["nested"].(map[string]int)["limit"]; got != 2 {
-		t.Fatalf("nested map was not isolated: %d", got)
-	}
-	if got := copied["sequence"].([]string)[0]; got != "one" {
-		t.Fatalf("sequence was not isolated: %q", got)
-	}
-	if _, ok := copied["added"]; ok {
-		t.Fatalf("top-level map was not isolated: %#v", copied)
-	}
-}
-
-func TestCopyMapRejectsInvalidValues(t *testing.T) {
-	cyclicMap := map[string]any{}
-	cyclicMap["self"] = cyclicMap
-	cyclicSlice := []any{nil}
-	cyclicSlice[0] = cyclicSlice
 
+func TestCopyPreservesSupportedScalarAndNumberTypes(t *testing.T) {
+	maxInt := int(^uint(0) >> 1)
+	minInt := -maxInt - 1
 	tests := []struct {
 		name  string
 		value any
 	}{
-		{name: "empty nested key", value: map[string]int{"": 1}},
+		{name: "bool", value: true},
-		{name: "non-string map key", value: map[int]string{1: "one"}},
+		{name: "named bool", value: namedBool(true)},
-		{name: "unsupported value", value: make(chan int)},
+		{name: "string", value: "value"},
-		{name: "cyclic map", value: cyclicMap},
+		{name: "named string", value: namedString("value")},
-		{name: "cyclic slice", value: cyclicSlice},
+		{name: "int", value: minInt},
-		{name: "NaN", value: math.NaN()},
+		{name: "int8", value: int8(-1 << 7)},
-		{name: "positive infinity", value: math.Inf(1)},
+		{name: "int16", value: int16(-1 << 15)},
-		{name: "unsafe signed integer", value: int64(1 << 53)},
+		{name: "int32", value: int32(-1 << 31)},
-		{name: "unsafe unsigned integer", value: uint64(1 << 53)},
+		{name: "int64", value: int64(-1 << 63)},
+		{name: "named int64", value: namedInt64(1<<63 - 1)},
+		{name: "uint", value: ^uint(0)},
+		{name: "uint8", value: ^uint8(0)},
+		{name: "uint16", value: ^uint16(0)},
+		{name: "uint32", value: ^uint32(0)},
+		{name: "uint64", value: ^uint64(0)},
+		{name: "uintptr", value: ^uintptr(0)},
+		{name: "named uint64", value: namedUint64(^uint64(0))},
+		{name: "float32", value: float32(1.25)},
+		{name: "float64", value: float64(-2.5e100)},
+		{name: "named float32", value: namedFloat32(3.5)},
+		{name: "named float64", value: namedFloat64(-4.5e200)},
+		{name: "JSON number integer", value: json.Number("18446744073709551615")},
+		{name: "JSON number fraction", value: json.Number("-1.25e+2")},
+		{name: "JSON number beyond float64", value: json.Number("1e9999")},
 	}
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			if _, err := jsonvalue.CopyMap(map[string]any{"value": tc.value}); err == nil {
+			got, err := Copy(tc.value)
+			if err != nil {
+				t.Fatalf("copy value: %v", err)
+			}
+			if !reflect.DeepEqual(got, tc.value) {
+				t.Fatalf("value or concrete type changed: got %#v (%T), want %#v (%T)", got, got, tc.value, tc.value)
+			}
+		})
+	}
+}
+
+func TestCopyRejectsInvalidNumbers(t *testing.T) {
+	tests := []struct {
+		name  string
+		value any
+	}{
+		{name: "float32 NaN", value: float32(math.NaN())},
+		{name: "float64 NaN", value: math.NaN()},
+		{name: "named float NaN", value: namedFloat64(math.NaN())},
+		{name: "positive infinity", value: math.Inf(1)},
+		{name: "negative infinity", value: math.Inf(-1)},
+		{name: "empty JSON number", value: json.Number("")},
+		{name: "leading zero JSON number", value: json.Number("01")},
+		{name: "leading plus JSON number", value: json.Number("+1")},
+		{name: "trailing decimal JSON number", value: json.Number("1.")},
+		{name: "leading decimal JSON number", value: json.Number(".1")},
+		{name: "non-number JSON number", value: json.Number("NaN")},
+		{name: "spaced JSON number", value: json.Number(" 1")},
+		{name: "quoted JSON number", value: json.Number(`"1"`)},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if _, err := Copy(tc.value); err == nil {
 				t.Fatal("expected validation error")
 			}
 		})
 	}
 }
 
-func TestCopyMapValidatesJSONNumberSyntaxAndRange(t *testing.T) {
+func TestCopyPreservesCompatibleCollectionsAndNilEmptyDistinctions(t *testing.T) {
-	for _, number := range []json.Number{"0", "-1", "1.25", "-1.25e+2"} {
+	collections := []struct {
-		t.Run("valid "+number.String(), func(t *testing.T) {
+		name  string
-			got, err := jsonvalue.CopyMap(map[string]any{"value": number})
+		value any
+	}{
+		{name: "unnamed map", value: map[string]int{"limit": 2}},
+		{name: "named map", value: namedMap{"limit": 2}},
+		{name: "unnamed slice", value: []string{"one", "two"}},
+		{name: "named slice", value: namedSlice{"one", "two"}},
+		{name: "unnamed array", value: [2]int{1, 2}},
+		{name: "named array", value: namedArray{{"limit": 2}}},
+		{name: "empty map", value: map[string]int{}},
+		{name: "empty named map", value: namedMap{}},
+		{name: "empty slice", value: []string{}},
+		{name: "empty named slice", value: namedSlice{}},
+		{name: "empty array", value: [0]string{}},
+	}
+
+	for _, tc := range collections {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := Copy(tc.value)
 			if err != nil {
-				t.Fatalf("copy valid JSON number: %v", err)
+				t.Fatalf("copy collection: %v", err)
 			}
-			if !reflect.DeepEqual(got["value"], number) {
+			if !reflect.DeepEqual(got, tc.value) || reflect.TypeOf(got) != reflect.TypeOf(tc.value) {
-				t.Fatalf("JSON number changed: got %#v want %#v", got["value"], number)
+				t.Fatalf("collection changed: got %#v (%T), want %#v (%T)", got, got, tc.value, tc.value)
+			}
+			kind := reflect.ValueOf(got).Kind()
+			if (kind == reflect.Map || kind == reflect.Slice) && reflect.ValueOf(got).IsNil() {
+				t.Fatal("non-nil collection became nil")
 			}
 		})
 	}
 
-	for _, number := range []json.Number{"", "01", "+1", "1.", ".1", "1e9999", "not-a-number"} {
+	var nilMap map[string]int
-		t.Run("invalid "+number.String(), func(t *testing.T) {
+	var nilSlice []string
-			if _, err := jsonvalue.CopyMap(map[string]any{"value": number}); err == nil {
+	var nilPointer *namedInt64
-				t.Fatal("expected invalid JSON number error")
+	for _, value := range []any{nil, nilMap, nilSlice, nilPointer} {
+		got, err := Copy(value)
+		if err != nil {
+			t.Fatalf("copy null value: %v", err)
+		}
+		if got != nil {
+			t.Fatalf("null value became %#v (%T)", got, got)
+		}
+	}
+
+	gotNil, err := CopyMap(nil)
+	if err != nil || gotNil != nil {
+		t.Fatalf("nil CopyMap result = %#v, %v", gotNil, err)
+	}
+	gotEmpty, err := CopyMap(map[string]any{})
+	if err != nil || gotEmpty == nil || len(gotEmpty) != 0 {
+		t.Fatalf("empty CopyMap result = %#v, %v", gotEmpty, err)
+	}
+}
+
+func TestCopyHandlesIndirectionAndIsolatesNestedMutations(t *testing.T) {
+	integer := namedInt64(7)
+	nestedMap := namedMap{"limit": 2}
+	nestedSlice := namedSlice{"original"}
+	nestedArray := namedArray{{"limit": 3}}
+	shared := []any{map[string]int{"value": 4}}
+	input := map[string]any{
+		"integer": &integer,
+		"map":     nestedMap,
+		"slice":   nestedSlice,
+		"array":   nestedArray,
+		"first":   shared,
+		"second":  shared,
+	}
+
+	copiedValue, err := Copy(input)
+	if err != nil {
+		t.Fatalf("copy mixed tree: %v", err)
+	}
+	copied := copiedValue.(map[string]any)
+	nestedMap["limit"] = 20
+	nestedSlice[0] = "changed"
+	nestedArray[0]["limit"] = 30
+	shared[0].(map[string]int)["value"] = 40
+
+	if got, ok := copied["integer"].(namedInt64); !ok || got != 7 {
+		t.Fatalf("pointer target changed: %#v", copied["integer"])
+	}
+	if got := copied["map"].(namedMap)["limit"]; got != 2 {
+		t.Fatalf("nested map aliased input: %d", got)
+	}
+	if got := copied["slice"].(namedSlice)[0]; got != "original" {
+		t.Fatalf("nested slice aliased input: %q", got)
+	}
+	if got := copied["array"].(namedArray)[0]["limit"]; got != 3 {
+		t.Fatalf("nested array aliased input: %d", got)
+	}
+	first := copied["first"].([]any)
+	second := copied["second"].([]any)
+	if got := first[0].(map[string]int)["value"]; got != 4 {
+		t.Fatalf("shared child aliased input: %d", got)
+	}
+	first[0].(map[string]int)["value"] = 99
+	if got := second[0].(map[string]int)["value"]; got != 4 {
+		t.Fatalf("repeated acyclic value shared copied output: %d", got)
+	}
+}
+
+func TestCopyAndCopyMapApplyDistinctEmptyKeyRules(t *testing.T) {
+	nested := map[string]any{"": []any{"original"}}
+	copiedValue, err := Copy(nested)
+	if err != nil {
+		t.Fatalf("Copy rejected empty schema key: %v", err)
+	}
+	nested[""].([]any)[0] = "changed"
+	if got := copiedValue.(map[string]any)[""].([]any)[0]; got != "original" {
+		t.Fatalf("copied schema value was not isolated: %v", got)
+	}
+
+	_, err = CopyMap(map[string]any{"nested": map[string]any{"": true}})
+	if err == nil || !strings.Contains(err.Error(), "extra_params.nested") {
+		t.Fatalf("CopyMap empty-key error = %v", err)
+	}
+}
+
+func TestCopyRejectsUnsupportedValuesAndActiveCycles(t *testing.T) {
+	cyclicMap := map[string]any{}
+	cyclicMap["self"] = cyclicMap
+	cyclicSlice := []any{nil}
+	cyclicSlice[0] = cyclicSlice
+	var cyclicPointer any
+	cyclicPointer = &cyclicPointer
+
+	tests := []struct {
+		name     string
+		value    any
+		wantPath string
+	}{
+		{name: "non-string map key", value: map[int]string{1: "one"}, wantPath: "value"},
+		{name: "unsupported channel", value: make(chan int), wantPath: "value"},
+		{name: "deterministic map path", value: map[string]any{"z": make(chan int), "a": make(chan int)}, wantPath: "value.a"},
+		{name: "cyclic map", value: cyclicMap, wantPath: "value.self"},
+		{name: "cyclic slice", value: cyclicSlice, wantPath: "value[0]"},
+		{name: "cyclic pointer", value: cyclicPointer, wantPath: "value"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := Copy(tc.value)
+			if err == nil || !strings.Contains(err.Error(), tc.wantPath) {
+				t.Fatalf("error = %v, want structural path %q", err, tc.wantPath)
 			}
 		})
 	}
 }
+
+func TestCopyEnforcesContainerDepth(t *testing.T) {
+	tests := []struct {
+		name    string
+		depth   int
+		wantErr bool
+	}{
+		{name: "just below", depth: maxContainerDepth - 1},
+		{name: "at limit", depth: maxContainerDepth},
+		{name: "over limit", depth: maxContainerDepth + 1, wantErr: true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := Copy(alternatingContainers(tc.depth))
+			if tc.wantErr {
+				if err == nil || !strings.HasPrefix(err.Error(), "value") || !strings.Contains(err.Error(), "container depth limit") {
+					t.Fatalf("depth error = %v", err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("copy depth %d: %v", tc.depth, err)
+			}
+		})
+	}
+}
+
+func TestCopyEnforcesProducedNodeBudgetForRepeatedAcyclicValues(t *testing.T) {
+	shared := []any{true}
+	sharedOccurrences := (maxProducedNodes - 2) / 2
+	justBelow := repeatedValues(shared, sharedOccurrences, 0)
+	atLimit := repeatedValues(shared, sharedOccurrences, 1)
+	overLimit := repeatedValues(shared, sharedOccurrences, 2)
+
+	for name, value := range map[string]any{
+		"just below": justBelow,
+		"at limit":   atLimit,
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := Copy(value); err != nil {
+				t.Fatalf("copy value within work budget: %v", err)
+			}
+		})
+	}
+
+	_, err := Copy(overLimit)
+	if err == nil || !strings.HasPrefix(err.Error(), "value[") || !strings.Contains(err.Error(), "value work limit") {
+		t.Fatalf("work-budget error = %v", err)
+	}
+
+	_, err = Copy(make([]any, maxProducedNodes))
+	if err == nil || !strings.HasPrefix(err.Error(), "value:") || !strings.Contains(err.Error(), "value work limit") {
+		t.Fatalf("flat work-budget error = %v", err)
+	}
+}
+
+func alternatingContainers(depth int) any {
+	var value any = true
+	for level := 0; level < depth; level++ {
+		switch level % 3 {
+		case 0:
+			value = map[string]any{"child": value}
+		case 1:
+			value = []any{value}
+		default:
+			value = [1]any{value}
+		}
+	}
+	return value
+}
+
+func repeatedValues(shared []any, occurrences, leadingScalars int) []any {
+	values := make([]any, 0, leadingScalars+occurrences)
+	for i := 0; i < leadingScalars; i++ {
+		values = append(values, false)
+	}
+	for i := 0; i < occurrences; i++ {
+		values = append(values, shared)
+	}
+	return values
+}

internal/llm/openai_compatible_client.go

@@ -7,9 +7,11 @@ import (
 	"errors"
 	"fmt"
 	"io"
+	"math"
 	"net/http"
 	"net/url"
 	"os"
+	"strconv"
 	"strings"
 	"time"
 
@@ -21,10 +23,25 @@ var (
 	ErrInvalidConfig     = errors.New("invalid llm client configuration")
 	ErrInvalidRequest    = errors.New("invalid generate request")
 	ErrRequestFailed     = errors.New("llm request failed")
+	ErrTransport         = errors.New("llm transport failed")
 	ErrUnexpectedStatus  = errors.New("llm returned non-success status")
 	ErrMalformedResponse = errors.New("malformed llm response")
 )
 
+const maxOpenAIChatResponseBytes int64 = 16 << 20
+
+type requestFailedError struct {
+	cause error
+}
+
+func (e *requestFailedError) Error() string {
+	return ErrRequestFailed.Error()
+}
+
+func (e *requestFailedError) Unwrap() []error {
+	return []error{ErrRequestFailed, ErrTransport, e.cause}
+}
+
 type OpenAICompatibleConfig struct {
 	BaseURL    string
 	Model      string
@@ -39,9 +56,11 @@ type OpenAICompatibleClient struct {
 }
 
 func NewOpenAICompatibleClient(cfg OpenAICompatibleConfig) (*OpenAICompatibleClient, error) {
-	baseURL := strings.TrimSpace(cfg.BaseURL)
+	baseURL := ""
-	if baseURL != "" {
+	if strings.TrimSpace(cfg.BaseURL) != "" {
-		if _, err := url.ParseRequestURI(baseURL); err != nil {
+		var err error
+		baseURL, err = domain.NormalizeOpenAICompatibleBaseEndpoint(cfg.BaseURL)
+		if err != nil {
 			return nil, fmt.Errorf("%w: invalid base URL: %v", ErrInvalidConfig, err)
 		}
 	}
@@ -63,25 +82,63 @@ func NewOpenAICompatibleClient(cfg OpenAICompatibleConfig) (*OpenAICompatibleCli
 	}
 
 	return &OpenAICompatibleClient{
-		baseURL:      strings.TrimRight(baseURL, "/"),
+		baseURL:      baseURL,
 		defaultModel: cfg.Model,
 		httpClient:   client,
 	}, nil
 }
 
 func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
-	if req.Target.TimeoutSeconds < 0 {
+	prepared, err := c.PrepareGeneration(req)
-		return nil, fmt.Errorf("%w: timeout_seconds must be greater than or equal to 0", ErrInvalidRequest)
+	if err != nil {
+		return nil, err
+	}
+	outcome := prepared.Execute(ctx)
+	if outcome.Err != nil {
+		return nil, outcome.Err
+	}
+	return outcome.Response, nil
+}
+
+// PreparedGeneration holds one encoded request and resolved authentication for
+// repeated attempts. Each Execute call creates a fresh HTTP request and context.
+type PreparedGeneration struct {
+	endpoint       string
+	payload        []byte
+	apiKey         string
+	timeoutSeconds int
+	httpClient     *http.Client
+}
+
+// AttemptOutcome describes one HTTP execution without exposing response bodies
+// or headers to the orchestration layer.
+type AttemptOutcome struct {
+	Response       *domain.GenerateResponse
+	Err            error
+	StatusCode     int
+	Usage          domain.TokenUsage
+	UsagePresence  *domain.TokenUsagePresence
+	RetryNotBefore time.Time
+}
+
+// PrepareGeneration validates and freezes one built-in request before dispatch.
+func (c *OpenAICompatibleClient) PrepareGeneration(req domain.GenerateRequest) (*PreparedGeneration, error) {
+	if err := domain.ValidateExecutionTargetSettings(req.Target); err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
 
-	endpoint := strings.TrimSpace(req.Target.Endpoint)
+	selectedEndpoint := req.Target.Endpoint
-	if endpoint == "" {
+	if strings.TrimSpace(selectedEndpoint) == "" {
-		endpoint = c.baseURL
+		selectedEndpoint = c.baseURL
 	}
-	if endpoint == "" {
+	endpoint, err := domain.NormalizeOpenAICompatibleBaseEndpoint(selectedEndpoint)
-		return nil, fmt.Errorf("%w: endpoint is required", ErrInvalidRequest)
+	if err != nil {
+		return nil, fmt.Errorf("%w: invalid endpoint: %v", ErrInvalidRequest, err)
+	}
+	endpoint, err = url.JoinPath(endpoint, defaults.OpenAIChatCompletionsPath)
+	if err != nil {
+		return nil, fmt.Errorf("%w: invalid endpoint path: %v", ErrInvalidRequest, err)
 	}
-	endpoint = strings.TrimRight(endpoint, "/") + defaults.OpenAIChatCompletionsPath
 
 	wireReq, err := openAIChatRequestFromGenerateRequest(req, c.defaultModel)
 	if err != nil {
@@ -97,71 +154,197 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 	if err != nil {
 		return nil, fmt.Errorf("%w: failed to encode request: %v", ErrRequestFailed, err)
 	}
-
+	apiKey := strings.TrimSpace(req.Target.APIKey)
-	requestContext := ctx
+	envName := strings.TrimSpace(req.Target.APIKeyEnv)
-	if req.Target.TimeoutSeconds > 0 {
+	if apiKey == "" && envName != "" {
-		var cancel context.CancelFunc
+		apiKey = strings.TrimSpace(os.Getenv(envName))
-		requestContext, cancel = context.WithTimeout(
-			ctx,
-			time.Duration(req.Target.TimeoutSeconds)*time.Second,
-		)
-		defer cancel()
 	}
-
+	if apiKey == "" && req.Target.APIKeyRequired {
-	httpReq, err := http.NewRequestWithContext(requestContext, http.MethodPost, endpoint, bytes.NewReader(payload))
+		if envName != "" {
-	if err != nil {
-		return nil, fmt.Errorf("%w: failed to create request: %v", ErrRequestFailed, err)
-	}
-	httpReq.Header.Set("Content-Type", "application/json")
-	if apiKey := strings.TrimSpace(req.Target.APIKey); apiKey != "" {
-		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
-	} else if envName := strings.TrimSpace(req.Target.APIKeyEnv); envName != "" {
-		apiKey := strings.TrimSpace(os.Getenv(envName))
-		if apiKey == "" {
 			return nil, fmt.Errorf("%w: api key environment variable %q is not set", ErrInvalidRequest, envName)
 		}
-		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
+		return nil, fmt.Errorf("%w: api key is required", ErrInvalidRequest)
 	}
-
 	httpClient := c.httpClient
 	if httpClient == nil {
 		httpClient = &http.Client{Timeout: defaults.LLMRequestTimeoutDefault}
 	}
+	return &PreparedGeneration{
+		endpoint: endpoint, payload: payload, apiKey: apiKey,
+		timeoutSeconds: req.Target.TimeoutSeconds, httpClient: httpClient,
+	}, nil
+}
 
-	httpResp, err := httpClient.Do(httpReq)
+// Execute dispatches one attempt. The returned response is non-nil only when
-	if err != nil {
+// the successful envelope contains a usable first choice.
-		return nil, fmt.Errorf("%w: %v", ErrRequestFailed, err)
+func (p *PreparedGeneration) Execute(ctx context.Context) AttemptOutcome {
+	requestContext := ctx
+	if p.timeoutSeconds > 0 {
+		var cancel context.CancelFunc
+		requestContext, cancel = context.WithTimeout(ctx, time.Duration(p.timeoutSeconds)*time.Second)
+		defer cancel()
 	}
+	httpReq, err := http.NewRequestWithContext(requestContext, http.MethodPost, p.endpoint, bytes.NewReader(p.payload))
+	if err != nil {
+		return AttemptOutcome{Err: fmt.Errorf("%w: failed to create request: %v", ErrRequestFailed, err)}
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if p.apiKey != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+p.apiKey)
+	}
+
+	httpResp, err := p.httpClient.Do(httpReq)
+	if err != nil {
+		return AttemptOutcome{Err: &requestFailedError{cause: err}}
+	}
+	receivedAt := time.Now()
 	defer httpResp.Body.Close()
+	outcome := AttemptOutcome{StatusCode: httpResp.StatusCode}
 
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		_, _ = io.Copy(io.Discard, io.LimitReader(httpResp.Body, 4096))
+		outcome.RetryNotBefore = retryNotBefore(httpResp.Header.Values("Retry-After"), receivedAt)
-		return nil, fmt.Errorf("%w: status=%d", ErrUnexpectedStatus, httpResp.StatusCode)
+		outcome.Err = providerHTTPErrorFromBody(
+			httpResp.StatusCode,
+			httpResp.ContentLength,
+			httpResp.Body,
+		)
+		if err := observedContextError(requestContext, ctx, outcome.Err); err != nil {
+			outcome.Err = err
+		}
+		return outcome
+	}
+	if httpResp.ContentLength > maxOpenAIChatResponseBytes {
+		outcome.Err = openAIChatResponseTooLargeError()
+		return outcome
 	}
 
-	var wireResp openAIChatResponse
+	wireResp, err := decodeOpenAIChatResponse(httpResp.Body)
-	if err := json.NewDecoder(httpResp.Body).Decode(&wireResp); err != nil {
+	if err != nil {
-		return nil, fmt.Errorf("%w: failed to decode response: %v", ErrMalformedResponse, err)
+		outcome.Err = err
+		if contextErr := observedContextError(requestContext, ctx, err); contextErr != nil {
+			outcome.Err = contextErr
+		}
+		return outcome
+	}
+	outcome.Usage = domain.TokenUsage{
+		PromptTokens: tokenCount(wireResp.Usage.PromptTokens), CompletionTokens: tokenCount(wireResp.Usage.CompletionTokens),
+		TotalTokens: tokenCount(wireResp.Usage.TotalTokens), CachedTokens: tokenCount(wireResp.Usage.PromptTokensDetails.CachedTokens),
+		CacheWriteTokens: tokenCount(wireResp.Usage.CacheWriteTokens),
+	}
+	outcome.UsagePresence = &domain.TokenUsagePresence{
+		PromptTokens: wireResp.Usage.PromptTokens != nil, CompletionTokens: wireResp.Usage.CompletionTokens != nil,
+		TotalTokens: wireResp.Usage.TotalTokens != nil, CachedTokens: wireResp.Usage.PromptTokensDetails.CachedTokens != nil,
+		CacheWriteTokens: wireResp.Usage.CacheWriteTokens != nil,
 	}
 
 	if len(wireResp.Choices) == 0 {
-		return nil, fmt.Errorf("%w: no choices returned", ErrMalformedResponse)
+		outcome.Err = fmt.Errorf("%w: no choices returned", ErrMalformedResponse)
+		return outcome
 	}
 	content := wireResp.Choices[0].Message.Content
-	if content == "" {
+	if content == nil {
-		return nil, fmt.Errorf("%w: first choice has empty message content", ErrMalformedResponse)
+		outcome.Err = fmt.Errorf("%w: first choice has missing message content", ErrMalformedResponse)
+		return outcome
+	}
+	outcome.Response = &domain.GenerateResponse{Content: *content, Usage: outcome.Usage, UsagePresence: outcome.UsagePresence}
+	return outcome
+}
+
+func observedContextError(attempt, caller context.Context, cause error) error {
+	if err := caller.Err(); err != nil {
+		return &requestFailedError{cause: err}
+	}
+	if err := attempt.Err(); err != nil {
+		return &requestFailedError{cause: err}
+	}
+	if errors.Is(cause, context.Canceled) || errors.Is(cause, context.DeadlineExceeded) {
+		return &requestFailedError{cause: cause}
+	}
+	return nil
+}
+
+func retryNotBefore(values []string, receivedAt time.Time) time.Time {
+	if len(values) == 0 {
+		return time.Time{}
+	}
+	value := strings.TrimSpace(values[0])
+	if value == "" {
+		return time.Time{}
+	}
+	digits := true
+	for _, r := range value {
+		if r < '0' || r > '9' {
+			digits = false
+			break
+		}
+	}
+	if digits {
+		seconds, err := strconv.ParseUint(value, 10, 64)
+		if err != nil || seconds > uint64(math.MaxInt64/int64(time.Second)) {
+			return receivedAt.Add(time.Duration(math.MaxInt64))
+		}
+		return receivedAt.Add(time.Duration(seconds) * time.Second)
+	}
+	if when, err := http.ParseTime(value); err == nil {
+		if when.Before(receivedAt) {
+			return receivedAt
+		}
+		return when
+	}
+	return time.Time{}
+}
+
+func tokenCount(value *int) int {
+	if value == nil {
+		return 0
+	}
+	return *value
+}
+
+func decodeOpenAIChatResponse(body io.Reader) (openAIChatResponse, error) {
+	limited := &io.LimitedReader{
+		R: body,
+		N: maxOpenAIChatResponseBytes + 1,
+	}
+	decoder := json.NewDecoder(limited)
+
+	var response openAIChatResponse
+	if err := decoder.Decode(&response); err != nil {
+		if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
+			return openAIChatResponse{}, &requestFailedError{cause: err}
+		}
+		if limited.N == 0 {
+			return openAIChatResponse{}, openAIChatResponseTooLargeError()
+		}
+		return openAIChatResponse{}, fmt.Errorf("%w: failed to decode response", ErrMalformedResponse)
+	}
+	if limited.N == 0 {
+		return openAIChatResponse{}, openAIChatResponseTooLargeError()
 	}
 
-	return &domain.GenerateResponse{
+	var trailing any
-		Content: content,
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
-		Usage: domain.TokenUsage{
+		if errors.Is(err, context.Canceled) || errors.Is(err, context.DeadlineExceeded) {
-			PromptTokens:     wireResp.Usage.PromptTokens,
+			return openAIChatResponse{}, &requestFailedError{cause: err}
-			CompletionTokens: wireResp.Usage.CompletionTokens,
+		}
-			TotalTokens:      wireResp.Usage.TotalTokens,
+		if limited.N == 0 {
-			CachedTokens:     wireResp.Usage.PromptTokensDetails.CachedTokens,
+			return openAIChatResponse{}, openAIChatResponseTooLargeError()
-			CacheWriteTokens: wireResp.Usage.CacheWriteTokens,
+		}
-		},
+		return openAIChatResponse{}, fmt.Errorf("%w: response contains trailing data", ErrMalformedResponse)
-	}, nil
+	}
+	if limited.N == 0 {
+		return openAIChatResponse{}, openAIChatResponseTooLargeError()
+	}
+
+	return response, nil
+}
+
+func openAIChatResponseTooLargeError() error {
+	return fmt.Errorf(
+		"%w: response exceeds %d-byte limit",
+		ErrMalformedResponse,
+		maxOpenAIChatResponseBytes,
+	)
 }
 
 func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultModel string) (openAIChatRequest, error) {
@@ -308,8 +491,8 @@ type openAICacheControl struct {
 }
 
 type openAIChatResponseMessage struct {
-	Role    string `json:"role"`
+	Role    string  `json:"role"`
-	Content string `json:"content"`
+	Content *string `json:"content"`
 }
 
 type openAIChatResponse struct {
@@ -317,13 +500,13 @@ type openAIChatResponse struct {
 		Message openAIChatResponseMessage `json:"message"`
 	} `json:"choices"`
 	Usage struct {
-		PromptTokens        int `json:"prompt_tokens"`
+		PromptTokens        *int `json:"prompt_tokens"`
-		CompletionTokens    int `json:"completion_tokens"`
+		CompletionTokens    *int `json:"completion_tokens"`
-		TotalTokens         int `json:"total_tokens"`
+		TotalTokens         *int `json:"total_tokens"`
 		PromptTokensDetails struct {
-			CachedTokens int `json:"cached_tokens"`
+			CachedTokens *int `json:"cached_tokens"`
 		} `json:"prompt_tokens_details"`
-		CacheWriteTokens int `json:"cache_write_tokens"`
+		CacheWriteTokens *int `json:"cache_write_tokens"`
 	} `json:"usage"`
 }
 

internal/llm/prepared_generation_test.go

@@ -0,0 +1,225 @@
+package llm
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"reflect"
+	"strings"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestPreparedGenerationFreezesPayloadAndAuthentication(t *testing.T) {
+	for _, tc := range []struct {
+		name       string
+		initialKey string
+		wantAuth   string
+	}{
+		{name: "absent optional key"},
+		{name: "present key", initialKey: "original-key", wantAuth: "Bearer original-key"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			const envName = "PROMPTKIT_PREPARED_GENERATION_TEST_KEY"
+			t.Setenv(envName, tc.initialKey)
+			var bodies, auth, urls []string
+			client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+				BaseURL: "https://example.invalid/v1", Model: "default-model",
+				HTTPClient: &http.Client{Transport: roundTripFunc(func(req *http.Request) (*http.Response, error) {
+					body, err := io.ReadAll(req.Body)
+					if err != nil {
+						t.Fatal(err)
+					}
+					bodies = append(bodies, string(body))
+					auth = append(auth, req.Header.Get("Authorization"))
+					urls = append(urls, req.URL.String())
+					return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(`{"choices":[{"message":{"content":"ok"}}]}`))}, nil
+				})},
+			})
+			if err != nil {
+				t.Fatal(err)
+			}
+			req := domain.GenerateRequest{
+				Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "original"}}},
+				Target: domain.ExecutionTarget{APIKeyEnv: envName},
+			}
+			prepared, err := client.PrepareGeneration(req)
+			if err != nil {
+				t.Fatal(err)
+			}
+			req.Prompt.Messages[0].Content = "changed"
+			t.Setenv(envName, "later-key")
+			for range 2 {
+				outcome := prepared.Execute(context.Background())
+				if outcome.Err != nil || outcome.Response == nil || outcome.Response.Content != "ok" {
+					t.Fatalf("Execute = %+v", outcome)
+				}
+			}
+			if len(bodies) != 2 || bodies[0] != bodies[1] || !strings.Contains(bodies[0], "original") || strings.Contains(bodies[0], "changed") || !reflect.DeepEqual(auth, []string{tc.wantAuth, tc.wantAuth}) || !reflect.DeepEqual(urls, []string{"https://example.invalid/v1/chat/completions", "https://example.invalid/v1/chat/completions"}) {
+				t.Fatalf("bodies=%v auth=%v urls=%v", bodies, auth, urls)
+			}
+		})
+	}
+}
+
+func TestPreparedGenerationAttemptsHaveIndependentContextsAndCloseBodies(t *testing.T) {
+	var contexts []context.Context
+	var closed int
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "https://example.invalid/v1", Model: "model",
+		HTTPClient: &http.Client{Transport: roundTripFunc(func(req *http.Request) (*http.Response, error) {
+			if err := req.Context().Err(); err != nil {
+				return nil, err
+			}
+			contexts = append(contexts, req.Context())
+			return &http.Response{StatusCode: http.StatusOK, Body: &countingBody{
+				Reader: strings.NewReader(`{"choices":[{"message":{"content":"ok"}}]}`),
+				closed: &closed,
+			}}, nil
+		})},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := client.PrepareGeneration(domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{TimeoutSeconds: 1},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	for range 2 {
+		if outcome := prepared.Execute(context.Background()); outcome.Err != nil {
+			t.Fatal(outcome.Err)
+		}
+	}
+	if len(contexts) != 2 || contexts[0] == contexts[1] || closed != 2 {
+		t.Fatalf("contexts=%d, closed=%d", len(contexts), closed)
+	}
+	for _, ctx := range contexts {
+		if _, ok := ctx.Deadline(); !ok || !errors.Is(ctx.Err(), context.Canceled) {
+			t.Fatalf("attempt context deadline/error = %v/%v", ok, ctx.Err())
+		}
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if outcome := prepared.Execute(ctx); !errors.Is(outcome.Err, context.Canceled) {
+		t.Fatalf("canceled attempt = %+v", outcome)
+	}
+	if outcome := prepared.Execute(context.Background()); outcome.Err != nil || outcome.Response == nil {
+		t.Fatalf("later attempt = %+v", outcome)
+	}
+}
+
+func TestPreparedGenerationRetainsUsageFromMalformedChoice(t *testing.T) {
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "https://example.invalid/v1", Model: "model",
+		HTTPClient: &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(`{"usage":{"prompt_tokens":0},"choices":[]}`))}, nil
+		})},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := client.PrepareGeneration(domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	outcome := prepared.Execute(context.Background())
+	if !errors.Is(outcome.Err, ErrMalformedResponse) || outcome.Response != nil || outcome.UsagePresence == nil || !outcome.UsagePresence.PromptTokens || outcome.Usage.PromptTokens != 0 {
+		t.Fatalf("malformed choice = %+v", outcome)
+	}
+}
+
+func TestPreparedGenerationDoesNotUseInvalidlyFramedUsage(t *testing.T) {
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "https://example.invalid/v1", Model: "model",
+		HTTPClient: &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			return &http.Response{StatusCode: http.StatusOK, Body: io.NopCloser(strings.NewReader(`{"usage":{"prompt_tokens":9},"choices":[]}{}`))}, nil
+		})},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := client.PrepareGeneration(domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	outcome := prepared.Execute(context.Background())
+	if !errors.Is(outcome.Err, ErrMalformedResponse) || outcome.UsagePresence != nil || outcome.Usage != (domain.TokenUsage{}) {
+		t.Fatalf("invalid framing = %+v", outcome)
+	}
+}
+
+func TestPreparedGenerationRetainsReadCancellation(t *testing.T) {
+	closed := false
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "https://example.invalid/v1", Model: "model",
+		HTTPClient: &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			return &http.Response{StatusCode: http.StatusOK, Body: &readErrorBody{err: context.Canceled, closed: &closed}}, nil
+		})},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	prepared, err := client.PrepareGeneration(domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+	outcome := prepared.Execute(context.Background())
+	if !errors.Is(outcome.Err, context.Canceled) || !errors.Is(outcome.Err, ErrTransport) || errors.Is(outcome.Err, ErrMalformedResponse) || !closed {
+		t.Fatalf("read cancellation = %+v, closed=%t", outcome, closed)
+	}
+}
+
+func TestRetryNotBefore(t *testing.T) {
+	received := time.Date(2026, 9, 23, 12, 0, 0, 0, time.UTC)
+	for _, tc := range []struct {
+		name  string
+		value string
+		want  time.Time
+	}{
+		{name: "seconds", value: " 12 ", want: received.Add(12 * time.Second)},
+		{name: "date", value: "Wed, 23 Sep 2026 12:00:15 GMT", want: received.Add(15 * time.Second)},
+		{name: "past date", value: "Wed, 23 Sep 2026 11:59:59 GMT", want: received},
+		{name: "negative", value: "-2"},
+		{name: "fractional", value: "1.5"},
+		{name: "malformed", value: "tomorrow"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := retryNotBefore([]string{tc.value, "999"}, received); !got.Equal(tc.want) {
+				t.Fatalf("not before = %v, want %v", got, tc.want)
+			}
+		})
+	}
+	if got := retryNotBefore([]string{"999999999999999999999999999999999999"}, received); !got.After(received) {
+		t.Fatalf("large valid delay wrapped: %v", got)
+	}
+}
+
+type countingBody struct {
+	io.Reader
+	closed *int
+}
+
+func (b *countingBody) Close() error {
+	*b.closed++
+	return nil
+}
+
+type readErrorBody struct {
+	err    error
+	closed *bool
+}
+
+func (b *readErrorBody) Read([]byte) (int, error) { return 0, b.err }
+func (b *readErrorBody) Close() error             { *b.closed = true; return nil }

internal/llm/provider_http_error.go

@@ -0,0 +1,189 @@
+package llm
+
+import (
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"strings"
+	"unicode"
+)
+
+const (
+	maxProviderErrorResponseBytes   int64 = 64 << 10
+	maxProviderErrorIdentifierRunes       = 256
+	maxProviderErrorMessageRunes          = 4096
+)
+
+// ProviderHTTPError describes a non-success response from an LLM provider.
+type ProviderHTTPError struct {
+	statusCode      int
+	providerCode    string
+	providerType    string
+	providerMessage string
+}
+
+func (e *ProviderHTTPError) StatusCode() int {
+	if e == nil {
+		return 0
+	}
+	return e.statusCode
+}
+
+func (e *ProviderHTTPError) ProviderCode() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerCode
+}
+
+func (e *ProviderHTTPError) ProviderType() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerType
+}
+
+func (e *ProviderHTTPError) ProviderMessage() string {
+	if e == nil {
+		return ""
+	}
+	return e.providerMessage
+}
+
+func (e *ProviderHTTPError) Error() string {
+	if e == nil || e.statusCode == 0 {
+		return ErrUnexpectedStatus.Error()
+	}
+	return fmt.Sprintf("%s: status=%d", ErrUnexpectedStatus, e.statusCode)
+}
+
+func (e *ProviderHTTPError) GoString() string {
+	return e.Error()
+}
+
+func (e *ProviderHTTPError) Unwrap() error {
+	return ErrUnexpectedStatus
+}
+
+type providerErrorDetails struct {
+	providerCode    string
+	providerType    string
+	providerMessage string
+}
+
+func newProviderHTTPError(statusCode int, details providerErrorDetails) *ProviderHTTPError {
+	return &ProviderHTTPError{
+		statusCode:      statusCode,
+		providerCode:    details.providerCode,
+		providerType:    details.providerType,
+		providerMessage: details.providerMessage,
+	}
+}
+
+func providerHTTPErrorFromBody(statusCode int, contentLength int64, body io.Reader) *ProviderHTTPError {
+	if contentLength > maxProviderErrorResponseBytes {
+		return newProviderHTTPError(statusCode, providerErrorDetails{})
+	}
+
+	limited := &io.LimitedReader{
+		R: body,
+		N: maxProviderErrorResponseBytes + 1,
+	}
+	contents, err := io.ReadAll(limited)
+	if err != nil || limited.N == 0 {
+		return newProviderHTTPError(statusCode, providerErrorDetails{})
+	}
+	return newProviderHTTPError(statusCode, parseProviderErrorEnvelope(contents))
+}
+
+func parseProviderErrorEnvelope(body []byte) providerErrorDetails {
+	decoder := json.NewDecoder(strings.NewReader(string(body)))
+	decoder.UseNumber()
+
+	var envelope map[string]json.RawMessage
+	if err := decoder.Decode(&envelope); err != nil {
+		return providerErrorDetails{}
+	}
+	var trailing any
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
+		return providerErrorDetails{}
+	}
+
+	rawError, ok := envelope["error"]
+	if !ok {
+		return providerErrorDetails{}
+	}
+	var providerError map[string]json.RawMessage
+	if err := json.Unmarshal(rawError, &providerError); err != nil || providerError == nil {
+		return providerErrorDetails{}
+	}
+
+	var details providerErrorDetails
+	if raw, ok := providerError["message"]; ok {
+		var value string
+		if json.Unmarshal(raw, &value) == nil {
+			details.providerMessage = normalizeProviderErrorMessage(value)
+		}
+	}
+	if raw, ok := providerError["type"]; ok {
+		var value string
+		if json.Unmarshal(raw, &value) == nil {
+			details.providerType = normalizeProviderErrorIdentifier(value)
+		}
+	}
+	if raw, ok := providerError["code"]; ok {
+		var value any
+		fieldDecoder := json.NewDecoder(strings.NewReader(string(raw)))
+		fieldDecoder.UseNumber()
+		if fieldDecoder.Decode(&value) == nil {
+			switch value := value.(type) {
+			case string:
+				details.providerCode = normalizeProviderErrorIdentifier(value)
+			case json.Number:
+				details.providerCode = normalizeProviderErrorIdentifier(value.String())
+			}
+		}
+	}
+
+	return details
+}
+
+func normalizeProviderErrorIdentifier(value string) string {
+	normalized := normalizeProviderErrorText(value)
+	if len([]rune(normalized)) > maxProviderErrorIdentifierRunes {
+		return ""
+	}
+	return normalized
+}
+
+func normalizeProviderErrorMessage(value string) string {
+	normalized := normalizeProviderErrorText(value)
+	runes := []rune(normalized)
+	if len(runes) <= maxProviderErrorMessageRunes {
+		return normalized
+	}
+	return string(runes[:maxProviderErrorMessageRunes-1]) + "…"
+}
+
+func normalizeProviderErrorText(value string) string {
+	value = strings.ToValidUTF8(value, "<22>")
+
+	var result strings.Builder
+	result.Grow(len(value))
+	separatorPending := false
+	for _, r := range value {
+		if unicode.IsSpace(r) || unicode.IsControl(r) || unicode.In(r, unicode.Cf) {
+			if result.Len() > 0 {
+				separatorPending = true
+			}
+			continue
+		}
+		if separatorPending {
+			result.WriteByte(' ')
+			separatorPending = false
+		}
+		result.WriteRune(r)
+	}
+	return result.String()
+}

internal/llm/provider_http_error_test.go

@@ -0,0 +1,255 @@
+package llm
+
+import (
+	"errors"
+	"fmt"
+	"io"
+	"reflect"
+	"strings"
+	"testing"
+	"unicode/utf8"
+)
+
+type guardedReader struct {
+	reader    io.Reader
+	remaining int64
+	bytes     int64
+	violated  bool
+}
+
+func (r *guardedReader) Read(buffer []byte) (int, error) {
+	if int64(len(buffer)) > r.remaining {
+		r.violated = true
+		return 0, errors.New("reader was read past its allowed boundary")
+	}
+	n, err := r.reader.Read(buffer)
+	r.bytes += int64(n)
+	r.remaining -= int64(n)
+	return n, err
+}
+
+type failingReader struct {
+	err error
+}
+
+func (r failingReader) Read([]byte) (int, error) {
+	return 0, r.err
+}
+
+func TestProviderHTTPErrorEnvelopeParsing(t *testing.T) {
+	tests := []struct {
+		name string
+		body string
+		want providerErrorDetails
+	}{
+		{
+			name: "all supported string fields",
+			body: `{"error":{"message":"diagnostic","type":"invalid_request_error","code":"unsupported_parameter"}}`,
+			want: providerErrorDetails{providerMessage: "diagnostic", providerType: "invalid_request_error", providerCode: "unsupported_parameter"},
+		},
+		{
+			name: "integer code",
+			body: `{"error":{"code":17}}`,
+			want: providerErrorDetails{providerCode: "17"},
+		},
+		{
+			name: "fractional code",
+			body: `{"error":{"code":1.25}}`,
+			want: providerErrorDetails{providerCode: "1.25"},
+		},
+		{
+			name: "exponent code",
+			body: `{"error":{"code":6.02e+23}}`,
+			want: providerErrorDetails{providerCode: "6.02e+23"},
+		},
+		{
+			name: "invalid fields do not discard valid fields",
+			body: `{"error":{"message":null,"type":"invalid_request_error","code":false}}`,
+			want: providerErrorDetails{providerType: "invalid_request_error"},
+		},
+		{
+			name: "unknown fields are ignored",
+			body: `{"trace":"do not retain","error":{"param":"temperature","metadata":{"secret":"x"}}}`,
+			want: providerErrorDetails{},
+		},
+		{name: "missing error", body: `{}`, want: providerErrorDetails{}},
+		{name: "null error", body: `{"error":null}`, want: providerErrorDetails{}},
+		{name: "scalar error", body: `{"error":"nope"}`, want: providerErrorDetails{}},
+		{name: "empty error", body: `{"error":{}}`, want: providerErrorDetails{}},
+		{name: "malformed", body: `{"error":`, want: providerErrorDetails{}},
+		{name: "truncated", body: `{"error":{"message":"x"`, want: providerErrorDetails{}},
+		{name: "trailing garbage", body: `{"error":{"message":"x"}} garbage`, want: providerErrorDetails{}},
+		{name: "second document", body: `{"error":{"message":"x"}} {}`, want: providerErrorDetails{}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := parseProviderErrorEnvelope([]byte(tc.body)); !reflect.DeepEqual(got, tc.want) {
+				t.Fatalf("parseProviderErrorEnvelope() = %#v, want %#v", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestProviderErrorTextNormalizationAndLimits(t *testing.T) {
+	validIdentifier := strings.Repeat("界", maxProviderErrorIdentifierRunes)
+	validMessage := strings.Repeat("界", maxProviderErrorMessageRunes)
+	tests := []struct {
+		name string
+		got  string
+		want string
+	}{
+		{name: "multibyte text", got: "Grüße 世界", want: "Grüße 世界"},
+		{name: "invalid UTF-8", got: string([]byte{'a', 0xff, 'b'}), want: "a<>b"},
+		{name: "whitespace control and format runs", got: " \n\talpha\x00\u200b\u200bbeta \r ", want: "alpha beta"},
+		{name: "blank normalization", got: "\t\u200b\n", want: ""},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := normalizeProviderErrorText(tc.got); got != tc.want {
+				t.Fatalf("normalizeProviderErrorText() = %q, want %q", got, tc.want)
+			}
+		})
+	}
+
+	if got := normalizeProviderErrorIdentifier(validIdentifier); got != validIdentifier {
+		t.Fatalf("exact identifier boundary = %q, want retained value", got)
+	}
+	if got := normalizeProviderErrorIdentifier(validIdentifier + "界"); got != "" {
+		t.Fatalf("overlong identifier = %q, want empty", got)
+	}
+	if got := normalizeProviderErrorMessage(validMessage); got != validMessage {
+		t.Fatalf("exact message boundary = %q, want retained value", got)
+	}
+	wantTruncatedMessage := strings.Repeat("界", maxProviderErrorMessageRunes-1) + "…"
+	if got := normalizeProviderErrorMessage(validMessage + "界"); got != wantTruncatedMessage {
+		t.Fatalf("overlong message length = %d, want %d", utf8.RuneCountInString(got), maxProviderErrorMessageRunes)
+	}
+}
+
+func TestProviderHTTPErrorIdentityAndFormatting(t *testing.T) {
+	const marker = "provider-secret-marker"
+	err := newProviderHTTPError(429, providerErrorDetails{
+		providerCode:    marker + "-code",
+		providerType:    marker + "-type",
+		providerMessage: marker + "-message",
+	})
+
+	if err.StatusCode() != 429 || err.ProviderCode() != marker+"-code" || err.ProviderType() != marker+"-type" || err.ProviderMessage() != marker+"-message" {
+		t.Fatalf("accessors returned unexpected values: %#v", err)
+	}
+	if !errors.Is(err, ErrUnexpectedStatus) {
+		t.Fatalf("errors.Is(%v, ErrUnexpectedStatus) = false", err)
+	}
+	for _, rendered := range []string{fmt.Sprintf("%v", err), fmt.Sprintf("%+v", err), fmt.Sprintf("%#v", err)} {
+		if rendered != "llm returned non-success status: status=429" {
+			t.Fatalf("formatted error = %q", rendered)
+		}
+		if strings.Contains(rendered, marker) {
+			t.Fatalf("formatted error exposed provider marker: %q", rendered)
+		}
+	}
+
+	var nilError *ProviderHTTPError
+	if nilError.StatusCode() != 0 || nilError.ProviderCode() != "" || nilError.ProviderType() != "" || nilError.ProviderMessage() != "" {
+		t.Fatal("nil accessors returned provider values")
+	}
+	if nilError.Error() != "llm returned non-success status" || nilError.GoString() != "llm returned non-success status" || !errors.Is(nilError, ErrUnexpectedStatus) {
+		t.Fatalf("nil error behavior is not safe: %v", nilError)
+	}
+
+	zero := &ProviderHTTPError{}
+	if zero.Error() != "llm returned non-success status" || zero.GoString() != "llm returned non-success status" || !errors.Is(zero, ErrUnexpectedStatus) {
+		t.Fatalf("zero error behavior is not safe: %v", zero)
+	}
+}
+
+func TestProviderHTTPErrorBodyBounds(t *testing.T) {
+	const (
+		statusCode = 502
+		marker     = "provider-body-marker"
+	)
+
+	ordinaryBody := `{"error":{"message":"` + marker + `"}}`
+	exactLimitBody := ordinaryBody + strings.Repeat(" ", int(maxProviderErrorResponseBytes)-len(ordinaryBody))
+	overLimitBody := ordinaryBody + strings.Repeat(" ", int(maxProviderErrorResponseBytes)+1-len(ordinaryBody))
+	tests := []struct {
+		name          string
+		contentLength int64
+		reader        io.Reader
+		wantRead      int64
+		wantMessage   string
+	}{
+		{
+			name:          "recognized envelope",
+			contentLength: int64(len(ordinaryBody)),
+			reader:        strings.NewReader(ordinaryBody),
+			wantRead:      int64(len(ordinaryBody)),
+			wantMessage:   marker,
+		},
+		{
+			name:          "exact limit",
+			contentLength: maxProviderErrorResponseBytes,
+			reader:        strings.NewReader(exactLimitBody),
+			wantRead:      maxProviderErrorResponseBytes,
+			wantMessage:   marker,
+		},
+		{
+			name:          "declared oversize does not read",
+			contentLength: maxProviderErrorResponseBytes + 1,
+			reader:        strings.NewReader(ordinaryBody),
+			wantRead:      0,
+		},
+		{
+			name:          "unknown length oversize",
+			contentLength: -1,
+			reader:        strings.NewReader(overLimitBody),
+			wantRead:      maxProviderErrorResponseBytes + 1,
+		},
+		{
+			name:          "underreported oversize",
+			contentLength: maxProviderErrorResponseBytes,
+			reader:        strings.NewReader(overLimitBody),
+			wantRead:      maxProviderErrorResponseBytes + 1,
+		},
+		{
+			name:          "read failure",
+			contentLength: -1,
+			reader:        failingReader{err: errors.New("read failure")},
+			wantRead:      0,
+		},
+		{
+			name:          "empty body",
+			contentLength: 0,
+			reader:        strings.NewReader(""),
+			wantRead:      0,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			reader := &guardedReader{
+				reader:    tc.reader,
+				remaining: maxProviderErrorResponseBytes + 1,
+			}
+			err := providerHTTPErrorFromBody(statusCode, tc.contentLength, reader)
+			if err == nil || err.StatusCode() != statusCode {
+				t.Fatalf("error status = %v, want %d", err, statusCode)
+			}
+			if reader.bytes != tc.wantRead {
+				t.Fatalf("body bytes read = %d, want %d", reader.bytes, tc.wantRead)
+			}
+			if reader.violated {
+				t.Fatal("body reader was asked to read beyond the overflow probe")
+			}
+			if got := err.ProviderMessage(); got != tc.wantMessage {
+				t.Fatalf("provider message = %q, want %q", got, tc.wantMessage)
+			}
+			if tc.wantMessage == "" {
+				if err.ProviderCode() != "" || err.ProviderType() != "" || strings.Contains(err.Error(), marker) {
+					t.Fatalf("discarded details were retained: %#v", err)
+				}
+			}
+		})
+	}
+}

internal/llm/provider_http_error_transport_test.go

@@ -0,0 +1,145 @@
+package llm
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+func TestOpenAICompatibleClientStructuredNonSuccessResponse(t *testing.T) {
+	body := `{"error":{"message":"  provider\nmessage\u200b","type":"invalid\ttype","code":1.5e+4}}`
+	responseBody := &countingReadCloser{reader: strings.NewReader(body)}
+	client := newNonSuccessResponseClient(t, http.StatusBadRequest, int64(len(body)), responseBody)
+
+	response, err := client.Generate(context.Background(), ordinaryGenerateRequest())
+	if response != nil {
+		t.Fatalf("response = %#v, want nil", response)
+	}
+	if !errors.Is(err, ErrUnexpectedStatus) {
+		t.Fatalf("errors.Is(%v, ErrUnexpectedStatus) = false", err)
+	}
+	var providerHTTPError *ProviderHTTPError
+	if !errors.As(err, &providerHTTPError) {
+		t.Fatalf("error = %T, want *ProviderHTTPError", err)
+	}
+	if providerHTTPError.StatusCode() != http.StatusBadRequest || providerHTTPError.ProviderCode() != "1.5e+4" || providerHTTPError.ProviderType() != "invalid type" || providerHTTPError.ProviderMessage() != "provider message" {
+		t.Fatalf("provider error = %#v", providerHTTPError)
+	}
+	if !responseBody.closed {
+		t.Fatal("non-success response body was not closed")
+	}
+}
+
+func TestOpenAICompatibleClientNonSuccessBodyOwnership(t *testing.T) {
+	const marker = "provider-body-marker"
+	normalBody := `{"error":{"message":"` + marker + `"}}`
+	overLimitBody := normalBody + strings.Repeat(" ", int(maxProviderErrorResponseBytes)+1-len(normalBody))
+	tests := []struct {
+		name          string
+		contentLength int64
+		reader        io.Reader
+		wantRead      int64
+		wantMessage   string
+	}{
+		{
+			name:          "normal",
+			contentLength: int64(len(normalBody)),
+			reader:        strings.NewReader(normalBody),
+			wantRead:      int64(len(normalBody)),
+			wantMessage:   marker,
+		},
+		{
+			name:          "declared oversize",
+			contentLength: maxProviderErrorResponseBytes + 1,
+			reader:        strings.NewReader(normalBody),
+			wantRead:      0,
+		},
+		{
+			name:          "streamed oversize",
+			contentLength: -1,
+			reader: &guardedReader{
+				reader:    strings.NewReader(overLimitBody),
+				remaining: maxProviderErrorResponseBytes + 1,
+			},
+			wantRead: maxProviderErrorResponseBytes + 1,
+		},
+		{
+			name:          "underreported oversize",
+			contentLength: maxProviderErrorResponseBytes,
+			reader: &guardedReader{
+				reader:    strings.NewReader(overLimitBody),
+				remaining: maxProviderErrorResponseBytes + 1,
+			},
+			wantRead: maxProviderErrorResponseBytes + 1,
+		},
+		{
+			name:          "malformed",
+			contentLength: 1,
+			reader:        strings.NewReader("{"),
+			wantRead:      1,
+		},
+		{
+			name:          "read failure",
+			contentLength: -1,
+			reader:        failingReader{err: errors.New("response read failed")},
+			wantRead:      0,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			body := &countingReadCloser{reader: tc.reader}
+			client := newNonSuccessResponseClient(t, http.StatusBadGateway, tc.contentLength, body)
+
+			response, err := client.Generate(context.Background(), ordinaryGenerateRequest())
+			if response != nil {
+				t.Fatalf("response = %#v, want nil", response)
+			}
+			var providerHTTPError *ProviderHTTPError
+			if !errors.As(err, &providerHTTPError) {
+				t.Fatalf("error = %T, want *ProviderHTTPError", err)
+			}
+			if !body.closed {
+				t.Fatal("response body was not closed")
+			}
+			if body.bytesRead != tc.wantRead {
+				t.Fatalf("body bytes read = %d, want %d", body.bytesRead, tc.wantRead)
+			}
+			if body.bytesRead > maxProviderErrorResponseBytes+1 {
+				t.Fatalf("body bytes read = %d, exceeds overflow probe", body.bytesRead)
+			}
+			if guarded, ok := tc.reader.(*guardedReader); ok && guarded.violated {
+				t.Fatal("body reader was asked to read beyond the overflow probe")
+			}
+			if got := providerHTTPError.ProviderMessage(); got != tc.wantMessage {
+				t.Fatalf("provider message = %q, want %q", got, tc.wantMessage)
+			}
+			if tc.wantMessage == "" && (providerHTTPError.ProviderCode() != "" || providerHTTPError.ProviderType() != "" || strings.Contains(providerHTTPError.Error(), marker)) {
+				t.Fatalf("discarded details were retained: %#v", providerHTTPError)
+			}
+		})
+	}
+}
+
+func newNonSuccessResponseClient(t *testing.T, statusCode int, contentLength int64, body io.ReadCloser) *OpenAICompatibleClient {
+	t.Helper()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "https://provider.example/v1",
+		Model:   "m",
+		HTTPClient: &http.Client{Transport: roundTripFunc(func(*http.Request) (*http.Response, error) {
+			return &http.Response{
+				StatusCode:    statusCode,
+				ContentLength: contentLength,
+				Body:          body,
+			}, nil
+		})},
+	})
+	if err != nil {
+		t.Fatalf("construct client: %v", err)
+	}
+	return client
+}

internal/observation/recorder.go

@@ -0,0 +1,318 @@
+// Package observation owns invocation-local execution identities, metadata
+// accounting, and synchronous event delivery.
+package observation
+
+import (
+	"context"
+	"crypto/rand"
+	"fmt"
+	"math"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// Recorder accumulates one execution's measured attempts and usage. Its owner
+// calls methods in execution order; it has no shared history or background work.
+type Recorder struct {
+	runID                string
+	startedAt            time.Time
+	now                  func() time.Time
+	observer             domain.ExecutionObserver
+	selection            Selection
+	attempts             uint64
+	retries              uint64
+	repairCalls          uint64
+	lastRepairGeneration uint64
+	finished             bool
+	capacityWait         time.Duration
+	generationDuration   time.Duration
+	retryDelay           time.Duration
+	usage                usageTotals
+}
+
+// Selection contains resolved identities only. Unresolved fields stay empty.
+type Selection struct {
+	PromptID      string
+	PromptVersion string
+	ProfileID     string
+	BackendID     string
+	Model         string
+}
+
+// AttemptData describes one completed active-permit acquisition and optional
+// generation dispatch. Non-dispatched attempts contribute no token usage.
+type AttemptData struct {
+	Generation         uint64
+	Attempt            uint64
+	Dispatched         bool
+	CapacityWait       time.Duration
+	GenerationDuration time.Duration
+	Outcome            domain.ExecutionOutcome
+	Failure            domain.ExecutionFailureKind
+	StatusCode         int
+	Usage              domain.TokenUsage
+	UsagePresence      *domain.TokenUsagePresence
+}
+
+// DelayData describes one completed or interrupted retry wait.
+type DelayData struct {
+	Generation uint64
+	Attempt    uint64
+	Elapsed    time.Duration
+	Failure    domain.ExecutionFailureKind
+}
+
+// FinishData describes the terminal outcome after resource release.
+type FinishData struct {
+	Outcome          domain.ExecutionOutcome
+	Failure          domain.ExecutionFailureKind
+	StatusCode       int
+	RetriesExhausted bool
+}
+
+// New creates an invocation-local recorder and run ID. It returns the recorder
+// even if ID creation fails, allowing an error terminal event with an empty ID.
+func New(observer domain.ExecutionObserver) (*Recorder, error) {
+	runID, err := NewRunID()
+	return newRecorder(runID, time.Now, observer), err
+}
+
+func newRecorder(runID string, now func() time.Time, observer domain.ExecutionObserver) *Recorder {
+	return &Recorder{runID: runID, startedAt: now(), now: now, observer: observer,
+		usage: newUsageTotals()}
+}
+
+// NewRunID creates the identity shared by observations and successful results.
+func NewRunID() (string, error) {
+	var b [16]byte
+	if _, err := rand.Read(b[:]); err != nil {
+		return "", err
+	}
+	b[6] = (b[6] & 0x0f) | 0x40
+	b[8] = (b[8] & 0x3f) | 0x80
+	return fmt.Sprintf("%08x-%04x-%04x-%04x-%012x", b[0:4], b[4:6], b[6:8], b[8:10], b[10:16]), nil
+}
+
+// RunID returns this invocation's identity.
+func (r *Recorder) RunID() string { return r.runID }
+
+// Start delivers the opening event with only identities already resolved.
+func (r *Recorder) Start() {
+	event := r.event(domain.ExecutionStarted)
+	r.deliver(event)
+}
+
+// UpdateSelection replaces the known resolved identities for later events.
+func (r *Recorder) UpdateSelection(selection Selection) { r.selection = selection }
+
+// Attempt records one completed acquisition and optional dispatch.
+func (r *Recorder) Attempt(data AttemptData) {
+	r.capacityWait = addDuration(r.capacityWait, data.CapacityWait)
+	if data.Dispatched {
+		r.attempts++
+		if data.Attempt > 1 {
+			r.retries++
+		}
+		if data.Generation > 1 && data.Generation != r.lastRepairGeneration {
+			r.repairCalls++
+			r.lastRepairGeneration = data.Generation
+		}
+		r.generationDuration = addDuration(r.generationDuration, data.GenerationDuration)
+		r.usage.add(data.Usage, data.UsagePresence)
+	}
+	event := r.event(domain.ExecutionAttempt)
+	event.Generation = data.Generation
+	event.Attempt = data.Attempt
+	event.Dispatched = data.Dispatched
+	event.CapacityWait = data.CapacityWait
+	if data.Dispatched {
+		event.GenerationDuration = data.GenerationDuration
+	}
+	event.Outcome = data.Outcome
+	event.Failure = data.Failure
+	event.StatusCode = data.StatusCode
+	if data.Dispatched {
+		event.Usage = observedUsage(data.Usage, data.UsagePresence)
+	}
+	r.deliver(event)
+}
+
+// Delay records the measured time spent waiting after one failed attempt.
+func (r *Recorder) Delay(data DelayData) {
+	r.retryDelay = addDuration(r.retryDelay, data.Elapsed)
+	event := r.event(domain.ExecutionRetryDelay)
+	event.Generation = data.Generation
+	event.Attempt = data.Attempt
+	event.RetryDelay = data.Elapsed
+	event.Failure = data.Failure
+	r.deliver(event)
+}
+
+// Finish delivers terminal totals once on a normal return. Its caller owns the
+// operation's resource cleanup and failure classification.
+func (r *Recorder) Finish(data FinishData) {
+	if r.finished {
+		return
+	}
+	r.finished = true
+	event := r.event(domain.ExecutionFinished)
+	event.Outcome = data.Outcome
+	event.Failure = data.Failure
+	event.StatusCode = data.StatusCode
+	event.RetriesExhausted = data.RetriesExhausted
+	event.Attempts = r.attempts
+	event.TransientRetries = r.retries
+	event.RepairCalls = r.repairCalls
+	event.CapacityWait = r.capacityWait
+	event.GenerationDuration = r.generationDuration
+	event.RetryDelay = r.retryDelay
+	event.Duration = event.Time.Sub(r.startedAt)
+	if event.Duration < 0 {
+		event.Duration = 0
+	}
+	event.Usage = r.usage.snapshot(r.attempts)
+	r.deliver(event)
+}
+
+func (r *Recorder) event(kind domain.ExecutionEventKind) domain.ExecutionEvent {
+	return domain.ExecutionEvent{
+		Kind: kind, RunID: r.runID, Time: r.now(), PromptID: r.selection.PromptID,
+		PromptVersion: r.selection.PromptVersion, ProfileID: r.selection.ProfileID,
+		BackendID: r.selection.BackendID, Model: r.selection.Model,
+	}
+}
+
+func (r *Recorder) deliver(event domain.ExecutionEvent) {
+	if r.observer != nil {
+		r.observer(event)
+	}
+}
+
+type contextKey struct{}
+
+// WithoutExecutionState starts a public operation without inheriting another
+// invocation's recorder or prepared-claim sink. Other context values and all
+// cancellation/deadline behavior remain with the caller's context.
+func WithoutExecutionState(ctx context.Context) context.Context {
+	return executionBoundary{Context: ctx}
+}
+
+type executionBoundary struct{ context.Context }
+
+func (c executionBoundary) Value(key any) any {
+	switch key.(type) {
+	case contextKey, claimSinkKey:
+		return nil
+	default:
+		return c.Context.Value(key)
+	}
+}
+
+// WithRecorder carries one invocation's recorder through internal client calls.
+func WithRecorder(ctx context.Context, recorder *Recorder) context.Context {
+	return context.WithValue(ctx, contextKey{}, recorder)
+}
+
+// FromContext returns the invocation-local recorder, if one was installed.
+func FromContext(ctx context.Context) *Recorder {
+	recorder, _ := ctx.Value(contextKey{}).(*Recorder)
+	return recorder
+}
+
+type claimSinkKey struct{}
+
+// WithClaimSink lets the facade learn which invocation actually claimed a
+// prepared handle, so it can emit the terminal event after public mapping.
+func WithClaimSink(ctx context.Context, sink **Recorder) context.Context {
+	return context.WithValue(ctx, claimSinkKey{}, sink)
+}
+
+// RegisterClaim records an invocation after its prepared handle is claimed.
+func RegisterClaim(ctx context.Context, recorder *Recorder) {
+	if sink, ok := ctx.Value(claimSinkKey{}).(**Recorder); ok && sink != nil {
+		*sink = recorder
+	}
+}
+
+type usageField struct {
+	value           int
+	known, complete bool
+}
+type usageTotals struct {
+	prompt, completion, total, cached, cacheWrite usageField
+}
+
+func newUsageTotals() usageTotals {
+	return usageTotals{prompt: usageField{complete: true}, completion: usageField{complete: true},
+		total: usageField{complete: true}, cached: usageField{complete: true}, cacheWrite: usageField{complete: true}}
+}
+
+func (u *usageTotals) add(usage domain.TokenUsage, presence *domain.TokenUsagePresence) {
+	known := usagePresence(usage, presence)
+	u.prompt.add(usage.PromptTokens, known.PromptTokens)
+	u.completion.add(usage.CompletionTokens, known.CompletionTokens)
+	u.total.add(usage.TotalTokens, known.TotalTokens)
+	u.cached.add(usage.CachedTokens, known.CachedTokens)
+	u.cacheWrite.add(usage.CacheWriteTokens, known.CacheWriteTokens)
+}
+
+func (u usageTotals) snapshot(attempts uint64) domain.ObservedTokenUsage {
+	return domain.ObservedTokenUsage{
+		PromptTokens: u.prompt.snapshot(attempts), CompletionTokens: u.completion.snapshot(attempts),
+		TotalTokens: u.total.snapshot(attempts), CachedTokens: u.cached.snapshot(attempts),
+		CacheWriteTokens: u.cacheWrite.snapshot(attempts),
+	}
+}
+
+func (f *usageField) add(value int, known bool) {
+	if known {
+		f.value += value
+		f.known = true
+	} else {
+		f.complete = false
+	}
+}
+
+func (f usageField) snapshot(attempts uint64) domain.ObservedTokenCount {
+	if attempts == 0 {
+		return domain.ObservedTokenCount{Known: true, Complete: true}
+	}
+	return domain.ObservedTokenCount{Value: f.value, Known: f.known, Complete: f.complete}
+}
+
+func observedUsage(usage domain.TokenUsage, presence *domain.TokenUsagePresence) domain.ObservedTokenUsage {
+	known := usagePresence(usage, presence)
+	return domain.ObservedTokenUsage{
+		PromptTokens:     observedCount(usage.PromptTokens, known.PromptTokens),
+		CompletionTokens: observedCount(usage.CompletionTokens, known.CompletionTokens),
+		TotalTokens:      observedCount(usage.TotalTokens, known.TotalTokens),
+		CachedTokens:     observedCount(usage.CachedTokens, known.CachedTokens),
+		CacheWriteTokens: observedCount(usage.CacheWriteTokens, known.CacheWriteTokens),
+	}
+}
+
+func usagePresence(usage domain.TokenUsage, presence *domain.TokenUsagePresence) domain.TokenUsagePresence {
+	if presence != nil {
+		return *presence
+	}
+	return domain.TokenUsagePresence{
+		PromptTokens: usage.PromptTokens != 0, CompletionTokens: usage.CompletionTokens != 0,
+		TotalTokens: usage.TotalTokens != 0, CachedTokens: usage.CachedTokens != 0,
+		CacheWriteTokens: usage.CacheWriteTokens != 0,
+	}
+}
+
+func observedCount(value int, known bool) domain.ObservedTokenCount {
+	if !known {
+		return domain.ObservedTokenCount{}
+	}
+	return domain.ObservedTokenCount{Value: value, Known: true, Complete: true}
+}
+
+func addDuration(total, next time.Duration) time.Duration {
+	if next > 0 && total > time.Duration(math.MaxInt64)-next {
+		return time.Duration(math.MaxInt64)
+	}
+	return total + next
+}

internal/observation/recorder_test.go

@@ -0,0 +1,159 @@
+package observation
+
+import (
+	"context"
+	"errors"
+	"reflect"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestExecutionBoundaryPreservesCallerContext(t *testing.T) {
+	type callerKey struct{}
+	parent, stop := context.WithTimeout(context.WithValue(context.Background(), callerKey{}, "caller"), time.Minute)
+	defer stop()
+	parent, cancel := context.WithCancelCause(parent)
+	defer cancel(nil)
+	recorder := newRecorder("outer", time.Now, nil)
+	sink := recorder
+	parent = WithClaimSink(WithRecorder(parent, recorder), &sink)
+	child := WithoutExecutionState(parent)
+	if FromContext(child) != nil || FromContext(parent) != recorder {
+		t.Fatal("execution boundary did not isolate recorder state")
+	}
+	RegisterClaim(child, nil)
+	if sink != recorder {
+		t.Fatal("execution boundary retained the parent claim sink")
+	}
+	wantDeadline, _ := parent.Deadline()
+	gotDeadline, ok := child.Deadline()
+	if !ok || gotDeadline != wantDeadline || child.Value(callerKey{}) != "caller" || child.Done() != parent.Done() {
+		t.Fatal("execution boundary changed caller context")
+	}
+	cause := errors.New("caller canceled")
+	cancel(cause)
+	if !errors.Is(child.Err(), context.Canceled) || context.Cause(child) != cause {
+		t.Fatal("execution boundary lost cancellation or its cause")
+	}
+}
+
+func TestRecorderAccountsDispatchedAttemptsAndReportedUsage(t *testing.T) {
+	start := time.Date(2026, 9, 23, 12, 0, 0, 0, time.UTC)
+	tick := 0
+	now := func() time.Time {
+		value := start.Add(time.Duration(tick) * time.Second)
+		tick++
+		return value
+	}
+	var events []domain.ExecutionEvent
+	recorder := newRecorder("run-1", now, func(event domain.ExecutionEvent) { events = append(events, event) })
+	recorder.Start()
+	recorder.UpdateSelection(Selection{PromptID: "prompt", PromptVersion: "v1", ProfileID: "profile", BackendID: "backend", Model: "model"})
+	recorder.Attempt(AttemptData{
+		Generation: 1, Attempt: 1, Dispatched: true,
+		CapacityWait: 2 * time.Millisecond, GenerationDuration: 10 * time.Millisecond,
+		Outcome:       domain.ExecutionOutcomeSuccess,
+		UsagePresence: &domain.TokenUsagePresence{PromptTokens: true},
+	})
+	recorder.Delay(DelayData{Generation: 1, Attempt: 1, Elapsed: 7 * time.Millisecond})
+	recorder.Attempt(AttemptData{
+		Generation: 1, Attempt: 2, Dispatched: true,
+		CapacityWait: 3 * time.Millisecond, GenerationDuration: 20 * time.Millisecond,
+		Outcome:       domain.ExecutionOutcomeSuccess,
+		Usage:         domain.TokenUsage{PromptTokens: 2, TotalTokens: 5},
+		UsagePresence: &domain.TokenUsagePresence{PromptTokens: true, TotalTokens: true},
+	})
+	recorder.Attempt(AttemptData{
+		Generation: 2, Attempt: 1, Dispatched: false,
+		CapacityWait: 4 * time.Millisecond, GenerationDuration: 99 * time.Millisecond,
+		Outcome: domain.ExecutionOutcomeError, Failure: domain.ExecutionFailureCanceled,
+		Usage: domain.TokenUsage{PromptTokens: 99},
+	})
+	recorder.Attempt(AttemptData{
+		Generation: 2, Attempt: 1, Dispatched: true,
+		CapacityWait: 5 * time.Millisecond, GenerationDuration: 30 * time.Millisecond,
+		Outcome: domain.ExecutionOutcomeError, Failure: domain.ExecutionFailureProviderStatus,
+		StatusCode: 503, Usage: domain.TokenUsage{PromptTokens: 3},
+	})
+	recorder.Finish(FinishData{Outcome: domain.ExecutionOutcomeError, Failure: domain.ExecutionFailureProviderStatus, StatusCode: 503, RetriesExhausted: true})
+	recorder.Finish(FinishData{})
+
+	if len(events) != 7 {
+		t.Fatalf("event count=%d", len(events))
+	}
+	if events[0].PromptID != "" || events[1].PromptID != "prompt" || events[0].RunID != "run-1" || events[1].RunID != "run-1" {
+		t.Fatalf("identity snapshots=%+v, %+v", events[0], events[1])
+	}
+	if events[1].Usage.PromptTokens != (domain.ObservedTokenCount{Known: true, Complete: true}) ||
+		events[4].Dispatched || events[4].GenerationDuration != 0 || events[4].Usage.PromptTokens.Known {
+		t.Fatalf("attempt usage or undispatched event=%+v, %+v", events[1], events[4])
+	}
+	last := events[len(events)-1]
+	if last.Attempts != 3 || last.TransientRetries != 1 || last.RepairCalls != 1 ||
+		last.CapacityWait != 14*time.Millisecond || last.GenerationDuration != 60*time.Millisecond ||
+		last.RetryDelay != 7*time.Millisecond || last.Duration != 7*time.Second {
+		t.Fatalf("terminal counters=%+v", last)
+	}
+	if last.Usage.PromptTokens != (domain.ObservedTokenCount{Value: 5, Known: true, Complete: true}) ||
+		last.Usage.TotalTokens != (domain.ObservedTokenCount{Value: 5, Known: true, Complete: false}) ||
+		last.Usage.CachedTokens != (domain.ObservedTokenCount{}) {
+		t.Fatalf("aggregate usage=%+v", last.Usage)
+	}
+	if last.Failure != domain.ExecutionFailureProviderStatus || !last.RetriesExhausted || last.StatusCode != 503 {
+		t.Fatalf("terminal failure=%+v", last)
+	}
+	if events[1].Attempts != 0 || events[1].TransientRetries != 0 || events[1].RepairCalls != 0 {
+		t.Fatalf("attempt event carried terminal totals=%+v", events[1])
+	}
+}
+
+func TestRecorderZeroAttemptsAndExplicitUnknownUsage(t *testing.T) {
+	var events []domain.ExecutionEvent
+	recorder := newRecorder("", time.Now, func(event domain.ExecutionEvent) { events = append(events, event) })
+	recorder.Finish(FinishData{Outcome: domain.ExecutionOutcomeError, Failure: domain.ExecutionFailureOther})
+	if len(events) != 1 || events[0].RunID != "" || events[0].Attempts != 0 ||
+		events[0].Usage.PromptTokens != (domain.ObservedTokenCount{Known: true, Complete: true}) ||
+		events[0].Usage.CacheWriteTokens != (domain.ObservedTokenCount{Known: true, Complete: true}) {
+		t.Fatalf("zero-attempt terminal=%+v", events)
+	}
+
+	recorder = newRecorder("run-2", time.Now, func(event domain.ExecutionEvent) { events = append(events, event) })
+	recorder.Attempt(AttemptData{Dispatched: true, Generation: 1, Attempt: 1,
+		Usage: domain.TokenUsage{PromptTokens: 7}, UsagePresence: &domain.TokenUsagePresence{}})
+	recorder.Finish(FinishData{Outcome: domain.ExecutionOutcomeSuccess})
+	last := events[len(events)-1]
+	if last.Usage.PromptTokens != (domain.ObservedTokenCount{}) {
+		t.Fatalf("explicitly unknown nonzero field=%+v", last.Usage.PromptTokens)
+	}
+	_ = newRecorder("run-3", time.Now, nil) // A nil callback still permits recording.
+}
+
+func TestRecorderOrderContextAndCallbackPanic(t *testing.T) {
+	var kinds []domain.ExecutionEventKind
+	recorder := newRecorder("run", time.Now, func(event domain.ExecutionEvent) { kinds = append(kinds, event.Kind) })
+	ctx := WithRecorder(context.Background(), recorder)
+	if FromContext(ctx) != recorder || FromContext(context.Background()) != nil {
+		t.Fatal("recorder context propagation failed")
+	}
+	recorder.Start()
+	recorder.Attempt(AttemptData{Generation: 1, Attempt: 1})
+	recorder.Delay(DelayData{Generation: 1, Attempt: 1})
+	recorder.Finish(FinishData{Outcome: domain.ExecutionOutcomeSuccess})
+	if !reflect.DeepEqual(kinds, []domain.ExecutionEventKind{
+		domain.ExecutionStarted, domain.ExecutionAttempt, domain.ExecutionRetryDelay, domain.ExecutionFinished,
+	}) {
+		t.Fatalf("callback order=%v", kinds)
+	}
+
+	recorder = newRecorder("panic", time.Now, func(domain.ExecutionEvent) { panic("observer panic") })
+	func() {
+		defer func() {
+			if recover() == nil {
+				t.Fatal("observer panic was swallowed")
+			}
+		}()
+		recorder.Start()
+	}()
+}