Document the v0.2.0 upgrade path

README.md

@@ -31,4 +31,14 @@ Contributors should start with the [development guide](docs/development.md).
 The [architecture policy](docs/policy/architecture.md) defines the library
 boundary and constraints that framework work must preserve.
 
+## Release Guidance
+
+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).
+
+## Related Project
+
+[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).

backends.go

@@ -0,0 +1,74 @@
+package promptkit
+
+import (
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// BackendOpenRouter is the reserved ID of Promptkit's built-in OpenRouter
+// backend.
+const BackendOpenRouter = backend.OpenRouterID
+
+// 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.
+	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
+	// key. NewEngine trims it and requires the portable form
+	// [A-Za-z_][A-Za-z0-9_]*. 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.
+	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
+	// beyond ConcurrencyLimit. Nil uses 1024 when ConcurrencyLimit is positive;
+	// a pointer uses its exact value, including zero. The pointed-to value must
+	// be non-negative, and QueueCapacity must be nil when ConcurrencyLimit is
+	// zero. Their sum must fit in an int. WithBackend copies the value and does
+	// not retain the pointer.
+	QueueCapacity *int
+}
+
+// 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,
+// BackendOpenRouter cannot be replaced. The immutable registration is scoped
+// to the resulting Engine and cannot be enumerated, replaced, removed, or
+// mutated after construction. WithBackend does not install package-global
+// state.
+func WithBackend(backend Backend) Option {
+	queueCapacity := 0
+	queueCapacitySet := backend.QueueCapacity != nil
+	if queueCapacitySet {
+		queueCapacity = *backend.QueueCapacity
+	}
+	return optionFunc(func(options *engineOptions) error {
+		options.backends = append(options.backends, domain.Backend{
+			ID:               backend.ID,
+			Endpoint:         backend.Endpoint,
+			APIKeyEnv:        backend.APIKeyEnv,
+			ExtraParams:      backend.ExtraParams,
+			ConcurrencyLimit: backend.ConcurrencyLimit,
+			QueueCapacity:    queueCapacity,
+			QueueCapacitySet: queueCapacitySet,
+		})
+		return nil
+	})
+}

capacity_contract_test.go

@@ -0,0 +1,369 @@
+package promptkit_test
+
+import (
+	"context"
+	"errors"
+	"sync"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+func TestEngineLimitsInjectedClientConcurrency(t *testing.T) {
+	release := make(chan struct{})
+	client := newCapacityGateClient(release, 8)
+	engine := newBackendCapacityEngine(t, client, 2, capacityInt(4), nil)
+
+	results := make(chan capacityRunResult, 6)
+	for i := 0; i < 6; i++ {
+		go runCapacityRequest(engine, context.Background(), promptkit.RunRequest{
+			PromptID: "prompt",
+			Execution: &promptkit.ExecutionTargetOverride{
+				Endpoint: "http://request.example/v1",
+			},
+		}, results)
+	}
+
+	first := awaitCapacityRequest(t, client.started)
+	second := awaitCapacityRequest(t, client.started)
+	if first.Target.BackendID != "limited" || second.Target.BackendID != "limited" {
+		t.Fatalf("endpoint override changed backend pool: first=%q second=%q",
+			first.Target.BackendID, second.Target.BackendID)
+	}
+	if active, peak, _ := client.snapshot(); active != 2 || peak != 2 {
+		t.Fatalf("client concurrency before release=(active=%d peak=%d), want 2", active, peak)
+	}
+
+	close(release)
+	for i := 0; i < 6; i++ {
+		outcome := awaitCapacityRun(t, results)
+		if outcome.err != nil || outcome.result == nil {
+			t.Fatalf("run outcome=(%+v, %v), want success", outcome.result, outcome.err)
+		}
+	}
+	if _, peak, calls := client.snapshot(); peak > 2 || calls != 6 {
+		t.Fatalf("client observations=(peak=%d calls=%d), want peak <= 2 and 6 calls", peak, calls)
+	}
+}
+
+func TestEngineRejectsRunBeforeCompletionWhenAdmissionIsFull(t *testing.T) {
+	artifactRelease := make(chan struct{})
+	reader := &capacityArtifactReader{
+		entered: make(chan struct{}, 2),
+		release: artifactRelease,
+	}
+	client := newCapacityGateClient(closedCapacityChannel(), 2)
+	engine := newBackendCapacityEngine(t, client, 1, capacityInt(0), reader)
+	firstResult := make(chan capacityRunResult, 1)
+	go runCapacityRequest(engine, context.Background(), capacityInputRequest("http://first.example/v1"), firstResult)
+
+	awaitCapacitySignal(t, reader.entered, "first artifact read")
+
+	result, err := engine.Run(context.Background(), capacityInputRequest("http://second.example/v1"))
+	if result != nil {
+		t.Fatalf("capacity rejection returned partial result: %+v", result)
+	}
+	if !errors.Is(err, promptkit.ErrCapacityExceeded) {
+		t.Fatalf("capacity rejection=%v, want ErrCapacityExceeded", err)
+	}
+	if errors.Is(err, promptkit.ErrInvalidRequest) || errors.Is(err, promptkit.ErrLLMGenerate) {
+		t.Fatalf("capacity rejection had an unrelated category: %v", err)
+	}
+	if calls := reader.callCount(); calls != 1 {
+		t.Fatalf("artifact calls=%d, want only the admitted run", calls)
+	}
+	if _, _, calls := client.snapshot(); calls != 0 {
+		t.Fatalf("client calls=%d before admitted run was released, want 0", calls)
+	}
+
+	close(artifactRelease)
+	outcome := awaitCapacityRun(t, firstResult)
+	if outcome.err != nil || outcome.result == nil {
+		t.Fatalf("first run outcome=(%+v, %v), want success", outcome.result, outcome.err)
+	}
+}
+
+func TestBackendCapacityIsIndependentBetweenEngines(t *testing.T) {
+	firstRelease := make(chan struct{})
+	firstClient := newCapacityGateClient(firstRelease, 1)
+	firstEngine := newBackendCapacityEngine(t, firstClient, 1, capacityInt(0), nil)
+	secondClient := newCapacityGateClient(closedCapacityChannel(), 1)
+	secondEngine := newBackendCapacityEngine(t, secondClient, 1, capacityInt(0), nil)
+
+	firstResult := make(chan capacityRunResult, 1)
+	go runCapacityRequest(firstEngine, context.Background(), promptkit.RunRequest{PromptID: "prompt"}, firstResult)
+	awaitCapacityRequest(t, firstClient.started)
+
+	result, err := secondEngine.Run(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil || result == nil {
+		t.Fatalf("second engine run=(%+v, %v), want independent success", result, err)
+	}
+	if _, _, calls := secondClient.snapshot(); calls != 1 {
+		t.Fatalf("second engine client calls=%d, want 1", calls)
+	}
+
+	close(firstRelease)
+	outcome := awaitCapacityRun(t, firstResult)
+	if outcome.err != nil || outcome.result == nil {
+		t.Fatalf("first engine run=(%+v, %v), want success", outcome.result, outcome.err)
+	}
+}
+
+func TestUnlimitedBackendsRetainInjectedClientConcurrency(t *testing.T) {
+	tests := []struct {
+		name      string
+		configure func(*testing.T, promptkit.LLMClient) *promptkit.Engine
+	}{
+		{
+			name: "custom backend",
+			configure: func(t *testing.T, client promptkit.LLMClient) *promptkit.Engine {
+				return newBackendCapacityEngine(t, client, 0, nil, nil)
+			},
+		},
+		{
+			name: "endpoint-only profile",
+			configure: func(t *testing.T, client promptkit.LLMClient) *promptkit.Engine {
+				engine, err := promptkit.NewEngine(promptkit.Config{},
+					promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+					promptkit.WithProfiles(promptkit.Profile{
+						ID: "profile", Endpoint: "http://endpoint.example/v1", Model: "model",
+					}),
+					promptkit.WithLLMClient(client),
+				)
+				if err != nil {
+					t.Fatalf("construct endpoint-only engine: %v", err)
+				}
+				return engine
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			release := make(chan struct{})
+			client := newCapacityGateClient(release, 2)
+			engine := tc.configure(t, client)
+			results := make(chan capacityRunResult, 2)
+			for i := 0; i < 2; i++ {
+				go runCapacityRequest(
+					engine,
+					context.Background(),
+					promptkit.RunRequest{PromptID: "prompt"},
+					results,
+				)
+			}
+
+			awaitCapacityRequest(t, client.started)
+			awaitCapacityRequest(t, client.started)
+			if active, peak, _ := client.snapshot(); active != 2 || peak != 2 {
+				t.Fatalf("unlimited concurrency=(active=%d peak=%d), want 2", active, peak)
+			}
+			close(release)
+			for i := 0; i < 2; i++ {
+				outcome := awaitCapacityRun(t, results)
+				if outcome.err != nil || outcome.result == nil {
+					t.Fatalf("run outcome=(%+v, %v), want success", outcome.result, outcome.err)
+				}
+			}
+		})
+	}
+}
+
+func TestCapacityExceededSentinelContract(t *testing.T) {
+	if promptkit.ErrCapacityExceeded == nil {
+		t.Fatal("ErrCapacityExceeded is nil")
+	}
+	for _, unrelated := range []error{
+		promptkit.ErrInvalidConfig,
+		promptkit.ErrInvalidRequest,
+		promptkit.ErrLLMGenerate,
+		promptkit.ErrValidation,
+	} {
+		if errors.Is(promptkit.ErrCapacityExceeded, unrelated) ||
+			errors.Is(unrelated, promptkit.ErrCapacityExceeded) {
+			t.Fatalf("ErrCapacityExceeded aliases unrelated sentinel %v", unrelated)
+		}
+	}
+}
+
+type capacityRunResult struct {
+	result *promptkit.RunResult
+	err    error
+}
+
+func runCapacityRequest(
+	engine *promptkit.Engine,
+	ctx context.Context,
+	request promptkit.RunRequest,
+	results chan<- capacityRunResult,
+) {
+	result, err := engine.Run(ctx, request)
+	results <- capacityRunResult{result: result, err: err}
+}
+
+func newBackendCapacityEngine(
+	t *testing.T,
+	client promptkit.LLMClient,
+	limit int,
+	queueCapacity *int,
+	reader promptkit.ArtifactReader,
+) *promptkit.Engine {
+	t.Helper()
+	promptFS := contractPromptFS("prompt", "profile", "message")
+	if reader != nil {
+		promptFS = contractInputPromptFS()
+	}
+	options := []promptkit.Option{
+		promptkit.WithPromptFS(promptFS, "."),
+		promptkit.WithBackend(promptkit.Backend{
+			ID:               "limited",
+			Endpoint:         "http://backend.example/v1",
+			ConcurrencyLimit: limit,
+			QueueCapacity:    queueCapacity,
+		}),
+		promptkit.WithProfiles(promptkit.Profile{
+			ID: "profile", BackendID: "limited", Model: "model",
+		}),
+		promptkit.WithLLMClient(client),
+	}
+	if reader != nil {
+		options = append(options, promptkit.WithArtifactReader(reader))
+	}
+	engine, err := promptkit.NewEngine(promptkit.Config{}, options...)
+	if err != nil {
+		t.Fatalf("construct capacity engine: %v", err)
+	}
+	return engine
+}
+
+func capacityInputRequest(endpoint string) promptkit.RunRequest {
+	return promptkit.RunRequest{
+		PromptID: "input-prompt",
+		Inputs: map[string]promptkit.ArtifactRef{
+			"input": promptkit.Inline("input"),
+		},
+		Execution: &promptkit.ExecutionTargetOverride{Endpoint: endpoint},
+	}
+}
+
+type capacityGateClient struct {
+	mu      sync.Mutex
+	active  int
+	peak    int
+	calls   int
+	started chan promptkit.GenerateRequest
+	release <-chan struct{}
+}
+
+func newCapacityGateClient(release <-chan struct{}, buffer int) *capacityGateClient {
+	return &capacityGateClient{
+		started: make(chan promptkit.GenerateRequest, buffer),
+		release: release,
+	}
+}
+
+func (c *capacityGateClient) Generate(
+	ctx context.Context,
+	request promptkit.GenerateRequest,
+) (*promptkit.GenerateResponse, error) {
+	c.mu.Lock()
+	c.calls++
+	c.active++
+	if c.active > c.peak {
+		c.peak = c.active
+	}
+	c.mu.Unlock()
+	defer func() {
+		c.mu.Lock()
+		c.active--
+		c.mu.Unlock()
+	}()
+
+	c.started <- request
+	select {
+	case <-c.release:
+		return &promptkit.GenerateResponse{Content: "ok"}, nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+}
+
+func (c *capacityGateClient) snapshot() (active, peak, calls int) {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.active, c.peak, c.calls
+}
+
+type capacityArtifactReader struct {
+	mu      sync.Mutex
+	calls   int
+	entered chan struct{}
+	release <-chan struct{}
+}
+
+func (r *capacityArtifactReader) Read(
+	ctx context.Context,
+	_ promptkit.ArtifactRef,
+) (*promptkit.Artifact, error) {
+	r.mu.Lock()
+	r.calls++
+	r.mu.Unlock()
+	r.entered <- struct{}{}
+	select {
+	case <-r.release:
+		return &promptkit.Artifact{Body: []byte("input")}, nil
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	}
+}
+
+func (r *capacityArtifactReader) callCount() int {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.calls
+}
+
+func awaitCapacityRequest(
+	t *testing.T,
+	requests <-chan promptkit.GenerateRequest,
+) promptkit.GenerateRequest {
+	t.Helper()
+	select {
+	case request := <-requests:
+		return request
+	case <-time.After(5 * time.Second):
+		t.Fatal("timed out waiting for client invocation")
+		return promptkit.GenerateRequest{}
+	}
+}
+
+func awaitCapacityRun(t *testing.T, results <-chan capacityRunResult) capacityRunResult {
+	t.Helper()
+	select {
+	case result := <-results:
+		return result
+	case <-time.After(5 * time.Second):
+		t.Fatal("timed out waiting for Run")
+		return capacityRunResult{}
+	}
+}
+
+func awaitCapacitySignal(t *testing.T, signal <-chan struct{}, name string) {
+	t.Helper()
+	select {
+	case <-signal:
+	case <-time.After(5 * time.Second):
+		t.Fatalf("timed out waiting for %s", name)
+	}
+}
+
+func capacityInt(value int) *int {
+	return &value
+}
+
+func closedCapacityChannel() <-chan struct{} {
+	channel := make(chan struct{})
+	close(channel)
+	return channel
+}

convert.go

@@ -4,6 +4,7 @@ import (
 	"reflect"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
 )
 
 func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
@@ -15,12 +16,12 @@ func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
 		PromptID:      req.PromptID,
 		PromptVersion: req.PromptVersion,
 		ProfileID:     req.ProfileID,
+		SessionID:     req.SessionID,
 		APIKey:        req.APIKey,
 		Inputs:        toDomainArtifactRefMap(req.Inputs),
 		Vars:          copyStringMap(req.Vars),
 		Execution:     execution,
 		Validation:    toDomainOutputContractPtr(req.Validation),
-		Metadata:      copyStringMap(req.Metadata),
 	}, nil
 }
 
@@ -33,6 +34,7 @@ func fromDomainPreparedRun(prepared *domain.PreparedRun) *PreparedRun {
 		PromptVersion:        prepared.PromptVersion,
 		PromptHash:           prepared.PromptHash,
 		SelectedProfileID:    prepared.SelectedProfileID,
+		SelectedBackendID:    prepared.SelectedBackendID,
 		EffectiveModelParams: fromDomainExecutionTarget(prepared.EffectiveModelParams),
 		OutputContract:       fromDomainOutputContract(prepared.OutputContract),
 		StructuredOutput:     fromDomainStructuredOutputSpec(prepared.StructuredOutput),
@@ -58,8 +60,10 @@ func fromDomainRunResult(result *domain.RunResult) *RunResult {
 		PromptID:             result.PromptID,
 		PromptVersion:        result.PromptVersion,
 		PromptHash:           result.PromptHash,
+		SessionID:            result.SessionID,
 		RenderedPromptHash:   result.RenderedPromptHash,
 		SelectedProfileID:    result.SelectedProfileID,
+		SelectedBackendID:    result.SelectedBackendID,
 		ModelName:            result.ModelName,
 		Endpoint:             result.Endpoint,
 		EffectiveModelParams: fromDomainExecutionTarget(result.EffectiveModelParams),
@@ -132,7 +136,7 @@ func toDomainExecutionTargetOverride(override *ExecutionTargetOverride) (*domain
 	if override == nil {
 		return nil, nil
 	}
-	extraParams, err := copyPublicJSONMap(override.ExtraParams)
+	extraParams, err := jsonvalue.CopyMap(override.ExtraParams)
 	if err != nil {
 		return nil, err
 	}
@@ -144,7 +148,7 @@ func toDomainExecutionTargetOverride(override *ExecutionTargetOverride) (*domain
 		TopP:            copyFloat64Ptr(override.TopP),
 		TimeoutSeconds:  copyIntPtr(override.TimeoutSeconds),
 		ServiceTier:     override.ServiceTier,
-		ReasoningEffort: override.ReasoningEffort,
+		ReasoningEffort: copyStringPtr(override.ReasoningEffort),
 		APIKeyEnv:       override.APIKeyEnv,
 		ExtraParams:     extraParams,
 	}, nil
@@ -152,6 +156,7 @@ func toDomainExecutionTargetOverride(override *ExecutionTargetOverride) (*domain
 
 func fromDomainExecutionTarget(target domain.ExecutionTarget) ExecutionTarget {
 	return ExecutionTarget{
+		BackendID:       target.BackendID,
 		Endpoint:        target.Endpoint,
 		Model:           target.Model,
 		Temperature:     target.Temperature,
@@ -397,6 +402,14 @@ func copyFloat64Ptr(src *float64) *float64 {
 	return &v
 }
 
+func copyStringPtr(src *string) *string {
+	if src == nil {
+		return nil
+	}
+	v := *src
+	return &v
+}
+
 func copyIntPtr(src *int) *int {
 	if src == nil {
 		return nil

doc.go

@@ -1,8 +1,57 @@
 // Package promptkit provides an embeddable engine for preparing and executing
 // prompt-defined LLM workflows.
 //
-// Applications construct an Engine with NewEngine, select filesystem or
-// in-memory definition sources with options, and use Prepare or Run to execute
-// requests. Concrete repositories, validators, and outbound clients remain
-// internal implementation details.
+// 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,
+// 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
+// policies bound admitted Run calls and model generations where configured,
+// while different backend pools and unlimited backends continue 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
+// copy request maps, slices, pointer values, and JSON-compatible extra
+// parameters before using them. Returned values and values 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.
+//
+// # Security and sensitive data
+//
+// The default artifact reader treats [File] paths as caller-selected operating
+// system paths. It does not restrict them to an application root or impose an
+// inbound request-size policy. Promptkit is not an inbound request or
+// untrusted-input security boundary. Applications must validate and restrict
+// untrusted input before constructing a request, or install an [ArtifactReader]
+// that enforces their filesystem, authorization, and size policies.
+//
+// Rendered messages, input and output [Artifact] bodies, [RunResult.RawOutput],
+// and [ValidationResult.Errors] may contain sensitive data. Credential
+// exclusion and redaction do not sanitize those values. Applications and
+// injected collaborators are responsible for access control, retention,
+// logging, and secret handling appropriate to their data.
+//
+// # JSON
+//
+// Stable JSON representations are provided for [PreparedRun], [RunResult],
+// [Artifact], [ExecutionTarget], [OutputContract], [ValidationResult],
+// [TokenUsage], [RenderedPrompt], [RenderedMessage], [CacheControl],
+// [StructuredOutputSpec], [StructuredOutputJSONSpec], [GenerateRequest],
+// [GenerateResponse], [ExecutionTargetPresence], and the string value types
+// used by those values.
+//
+// Construction values, including [Config], [Backend], [RunRequest],
+// [ArtifactRef], [ExecutionTargetOverride], [Profile], and
+// [OpenAICompatibleProfileConfig], do not have stable JSON representations.
+// Direct API keys are nevertheless excluded from JSON for every public value.
+//
+// JSON timestamps use time.Time's RFC 3339 encoding and are omitted when zero.
+// PreparedRun and RunResult durations are encoded as integer milliseconds in
+// duration_ms and omitted when zero. Run IDs and all exposed hashes are opaque:
+// their spelling, length, character set, and algorithm are not API contracts.
 package promptkit

docs/consumers/pkg-promptkit.md

@@ -1,151 +1,254 @@
 # Package `promptkit`
 
-Import path:
+## Purpose
+
+This guide helps Go consumers assemble Promptkit and choose the main
+preparation or execution workflow. The declarations and GoDoc in the
+[root package](../../doc.go) own exact field, option, serialization,
+concurrency, ownership, failure, and cancellation semantics. The
+[framework format reference](../formats.md) owns prompt, profile, and schema
+file contracts.
+
+Import the package as:
 
 ```go
 import "gitea.maximumdirect.net/eric/promptkit"
 ```
 
-Package `promptkit` is the supported Go contract for in-process prompt
-preparation and execution. The declarations and their GoDoc in the
-[root package](../../doc.go) own the exact API; this guide explains how the
-pieces are used together. The [framework format reference](../formats.md) owns
-prompt, profile, and schema file contracts.
+The following Go fragments are illustrative and omit surrounding package,
+import, and error-handling code. Use the maintained examples for complete
+programs.
 
-## Engine Construction And Sources
+## Construct An Engine
 
-Construct an engine with [`NewEngine`, `Config`, and
-`Option`](../../engine.go). `PromptDir` is required unless a prompt source
-option is supplied. `ProfileDir` optionally overlays built-in profiles, and an
-empty `SchemaDir` uses the current directory. `Timeout` is the transport-wide
-safety cap for the built-in OpenAI-compatible client. An optional `HTTPClient`
-is cloned; its positive timeout takes precedence.
+Create an engine with
+[`NewEngine`](../../engine.go). A directory-backed setup supplies a prompt
+directory and may supply profile and schema directories:
 
-Nil options are ignored. Invalid construction, including a nil injected client
-or artifact reader, returns an error matching `ErrInvalidConfig`.
+```go
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir:  "prompts",
+	ProfileDir: "profiles",
+	SchemaDir:  "schemas",
+})
+```
 
-The [source options](../../engine.go) replace their matching directory source:
-
-- `WithPromptFS` and `WithPromptFile` select prompt definitions;
-- `WithProfileFS` and `WithProfileFile` overlay built-in profiles;
-- `WithProfiles` adds in-memory profiles ahead of file and built-in profiles;
-- `WithSchemaFS` and `WithSchemaFile` select JSON Schema documents;
-- `WithLLMClient` replaces the built-in model client; and
-- `WithArtifactReader` replaces the default reader for every input.
-
-Source selection, path resolution, strict decoding, profile overlays, and
-file-to-request precedence are defined in the
+Options support single-file or `fs.FS` sources, in-memory profiles,
+engine-scoped backends, and injected artifact or model clients. Consult the
+[constructor and option GoDoc](../../engine.go) for composition, precedence,
+validation, and default transport behavior. Source discovery, format
+validation, and profile precedence are defined by the
 [framework format reference](../formats.md).
 
-Per-generation timeout values from profiles or requests are independent of
-the transport cap and caller context. An explicit request value of zero
-disables only the per-generation deadline. The
-[outbound integration contract](../integrations/openai-compatible-chat.md#timeout-and-cancellation)
-defines the complete timeout layering.
+## Prepare Without Model Execution
 
-## Preparation And 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:
 
-[`Engine.Prepare` and `Engine.Run`](../../engine.go) accept the public
-[`RunRequest`](../../types.go). `Prepare` resolves the prompt, profile, input
-artifacts, validation contract, and rendered messages without calling an LLM.
-`Run` performs the same preparation, calls the configured client, and validates
-the generated content. The maintained
+```go
+prepared, err := engine.Prepare(ctx, promptkit.RunRequest{
+	PromptID: "meeting.summary",
+	Inputs: map[string]promptkit.ArtifactRef{
+		"note": promptkit.Inline("Synthetic meeting notes"),
+	},
+})
+```
+
+The maintained
 [offline preparation example](../../examples/go-library/prepare/main.go)
-provides a complete runnable workflow using a prompt file, in-memory profile,
-and inline input.
+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).
 
-[`PreparedRun` and `RunResult`](../../types.go) expose copied public values.
-Preparation returns effective settings, hashes, rendered messages, selected
-profile, structured-output information, and timing without resolved secrets or
-model output. Execution adds the generated artifact and raw output, validation
-state, model metadata, usage, run ID, and duration.
+## Execute And Validate
 
-A generated-content validation failure returns a result with
-`Validation.Status == ValidationFailed`. An inability to perform validation
-returns an error matching `ErrValidation`.
+[`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;
+an operational inability to validate returns an error.
 
-## Requests, Inputs, And Overrides
+The maintained
+[offline execution example](../../examples/go-library/run/main.go) injects a
+deterministic model client and exercises `Run` without credentials, network
+access, or paid calls. It is intentionally separate from the preparation
+example so each workflow and its small prompt fixture can be copied and run on
+its own.
 
-The [request and value declarations](../../types.go) own the available fields,
-serialized constants, and result shapes. Use `File`, `Inline`, or
-`InlineWithURI` to construct artifact references. The
-[framework format reference](../formats.md) defines declared inputs, template
-references, output contracts, and the relationship between file values and
-request overrides.
+Use the [`RunResult` and `ValidationResult` GoDoc](../../types.go) for the
+returned data and the `Engine.Run` GoDoc for failure and cancellation
+semantics. The
+[OpenAI-compatible integration contract](../integrations/openai-compatible-chat.md)
+owns the built-in client's outbound HTTP behavior.
 
-`ExecutionTargetOverride` uses pointers for numeric settings so an explicit
-zero remains distinct from no override. `ExtraParams` accepts JSON-compatible
-strings, booleans, finite numbers, string-keyed objects, arrays or slices, and
-nil. Unsupported values, non-string map keys, non-finite numbers, and cycles
-match `ErrInvalidConfig` in profiles or `ErrInvalidRequest` in request
-overrides.
+## Inputs, Profiles, And Overrides
 
-Returned requests, profiles, prepared values, results, artifacts, maps, and
-slices are isolated from internal engine state. Consumers and injected
-extensions should not retain or mutate values owned by another caller.
+Use `File`, `Inline`, or `InlineWithURI` to construct request inputs. A request
+can select a profile explicitly or use the prompt's default profile, and can
+replace execution settings or the complete output contract.
 
-## Profiles And Credentials
+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,
+and framework defaults.
 
-[`OpenAICompatibleProfile`](../../profiles.go) constructs an ordinary
-in-memory profile for an OpenAI-compatible chat-completions endpoint.
-`WithProfiles` rejects duplicate IDs in one call and gives in-memory profiles
-precedence over explicit file sources and built-ins.
+For programmatic profiles,
+[`OpenAICompatibleProfile`](../../profiles.go) converts ordinary
+OpenAI-compatible settings into a value accepted by `WithProfiles`.
 
-Raw API keys do not belong in profiles. File-backed profiles may name an
-environment variable, while an in-memory profile can require a request key.
-A direct `RunRequest.APIKey` is request-scoped and takes precedence over an
-environment lookup for the built-in client. Profile fields, ranges, built-ins,
-precedence, and credential rules are owned by the
-[framework format reference](../formats.md).
+### Set A Per-Run Session And Reasoning
 
-API keys are excluded from JSON, prepared values, and results. The public
-`String` and `GoString` methods report only whether a direct key is present.
-Avoid reflection-based dumps of request structs, which can bypass that
-redaction.
+Supply a direct session ID when one prompt should be correlated with a
+consumer-managed conversation or workflow without changing prompt variables:
+
+```go
+reasoning := "high"
+result, err := engine.Run(ctx, promptkit.RunRequest{
+	PromptID: "meeting.summary",
+	SessionID: "conversation-42",
+	Inputs: map[string]promptkit.ArtifactRef{
+		"note": promptkit.Inline("Synthetic meeting notes"),
+	},
+	Execution: &promptkit.ExecutionTargetOverride{
+		ReasoningEffort: &reasoning,
+	},
+})
+```
+
+A nil reasoning pointer inherits the selected profile, a pointer to a
+nonblank string replaces it, and a pointer to a blank string disables
+reasoning for that run. Session IDs are correlation metadata, not credentials;
+use stable, non-secret values that are safe to expose to collaborators and
+providers. The
+[`RunRequest` and `ExecutionTargetOverride` GoDoc](../../types.go) owns the
+exact normalization, precedence, error, copying, and exposure contract.
+
+### Register A Custom Backend
+
+Register a reusable OpenAI-compatible connection once, then select it from a
+profile. This local backend limits model generation to two simultaneous calls;
+because `QueueCapacity` is omitted, the engine admits up to 1024 additional
+calls waiting behind them:
+
+```go
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir: "prompts",
+},
+	promptkit.WithBackend(promptkit.Backend{
+		ID:        "local",
+		Endpoint:  "http://localhost:8000/v1",
+		APIKeyEnv: "LOCAL_LLM_API_KEY",
+		ConcurrencyLimit: 2,
+	}),
+	promptkit.WithProfiles(promptkit.Profile{
+		ID:        "local-summary",
+		BackendID: "local",
+		Model:     "example-model",
+	}),
+)
+```
+
+Registrations belong to one engine and custom IDs cannot replace built-ins.
+The [`Backend` and `WithBackend` GoDoc](../../backends.go) defines validation,
+copying, uniqueness, exact concurrency-field semantics, and request-default
+behavior.
+
+Both file-backed and in-memory profiles select a registration through
+`backend` or `Profile.BackendID`. Profile and request endpoint overrides retain
+that routing and capacity identity. `PreparedRun.SelectedBackendID`,
+`RunResult.SelectedBackendID`, and the effective `ExecutionTarget.BackendID`
+expose it to consumers and injected model clients. Endpoint-only profiles
+remain supported and expose an empty backend ID.
+
+### Limit Backend Concurrency
+
+Set `Backend.ConcurrencyLimit` when a backend needs protection from too many
+simultaneous model calls. Leaving `QueueCapacity` nil, as in the local-backend
+example above, selects the default waiting capacity of 1024.
+
+To accept no waiting backlog beyond the active calls, provide an explicit
+zero:
+
+```go
+noWaiting := 0
+backend := promptkit.Backend{
+	ID:               "local",
+	Endpoint:         "http://localhost:8000/v1",
+	ConcurrencyLimit: 2,
+	QueueCapacity:    &noWaiting,
+}
+```
+
+The pointer distinguishes an explicit zero from omission. Keep using keyed
+`Backend` literals so additive configuration fields remain source-compatible.
+Capacity belongs to one engine and the selected backend ID; endpoint-only
+profiles and custom backends without a configured limit remain unrestricted.
+Exact validation, defaulting, ownership, and concurrency semantics belong to
+the [`Backend` GoDoc](../../backends.go).
+
+## Credentials
+
+File-backed profiles name an environment variable; in-memory profiles can
+require a direct request key. Direct keys are request-scoped and are excluded
+from supported JSON values and the package's `String` and `GoString`
+summaries. The exact precedence and redaction guarantees belong to
+[`RunRequest`, `GenerateRequest`, and the profile GoDoc](../../types.go).
+
+## Protect Files And Generated Data
+
+The default artifact reader opens a `File` reference as a caller-selected
+operating-system path. It does not constrain paths to an application root,
+impose an inbound request-size policy, or establish an untrusted-input security
+boundary. Applications must validate and restrict untrusted paths and payloads
+before constructing a request, or inject an artifact reader that enforces
+their filesystem, authorization, and size policies.
+
+Rendered messages, input and output artifact bodies, raw model output, and
+validation diagnostics can contain sensitive data. API-key redaction does not
+sanitize those values. Treat prepared values, results, collaborator requests,
+errors, and logs according to the application's data-access, retention, and
+secret-handling policies.
 
 ## Extension Interfaces
 
-The [`LLMClient`, `GenerateRequest`, and
-`GenerateResponse`](../../types.go) boundary lets a consumer replace model
-generation. Injected clients receive copied rendered messages, effective
-settings, explicit numeric-setting presence, structured-output constraints,
-and the request-scoped key. They return generated content and token usage.
+Inject an [`LLMClient` or `ArtifactReader`](../../types.go) when the built-in
+behavior does not fit the application. Their GoDoc defines concurrent use,
+context handling, ownership of copied values, nil responses, and preservation
+of collaborator errors. Implementations must honor cancellation, safely manage
+copies they retain, avoid unsafe logging of content or credentials, and enforce
+the application policy that motivated the injection.
 
-The [`ArtifactReader`](../../types.go) boundary replaces the default inline and
-file reader for every input. Readers provide artifact content and metadata; the
-engine fills an empty artifact name from the input-map key. A reader error
-matches `ErrArtifactLoad` while preserving the original identity for
-`errors.Is`. A nil artifact with a nil error is also an artifact-load failure.
+## Handle Errors
 
-Extensions should honor context cancellation and avoid logging raw prompts,
-artifacts, or credentials.
+Use `errors.Is` with the
+[public error sentinels and operation GoDoc](../../engine.go). The declarations
+distinguish invalid construction, invalid requests, absent sources,
+source-loading failures, collaborator failures, and operational validation
+failures. Specific request conditions may also match the broader
+`ErrInvalidRequest`, and injected collaborator identities are preserved where
+documented. Invalid or duplicate backend registrations match
+`ErrInvalidConfig`; selecting an unknown backend matches `ErrProfileLoad`.
 
-## Errors
+When a limited backend has admitted all active and waiting calls, handle
+`ErrCapacityExceeded` separately from request errors and provider failures:
 
-The [public error declarations](../../engine.go) and
-[mapping](../../errors.go) preserve these sentinel checks through `errors.Is`:
+```go
+result, err := engine.Run(ctx, request)
+if errors.Is(err, promptkit.ErrCapacityExceeded) {
+	// Apply application policy: shed work, report overload, or retry later.
+}
+```
 
-- `ErrInvalidConfig`
-- `ErrInvalidRequest`
-- `ErrPromptNotFound`
-- `ErrProfileNotFound`
-- `ErrProfileRequired`
-- `ErrPromptLoad`
-- `ErrProfileLoad`
-- `ErrAPIKeyEnvMissing`
-- `ErrArtifactLoad`
-- `ErrPromptRender`
-- `ErrLLMGenerate`
-- `ErrValidation`
+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
+cancellation identities.
 
-`ErrProfileRequired` and `ErrAPIKeyEnvMissing` also match
-`ErrInvalidRequest`, allowing either broad request handling or a specific
-condition. Wrapped collaborator errors retain their identity where the public
-contract promises it.
-
-## Consumer Boundary
+## Application Boundary
 
 Promptkit is an importable library. It does not own a command, inbound HTTP
-API, process configuration, or deployment policy. Scriptorium is one
-downstream application that maps this root package contract into those
-application concerns.
+API, process configuration, or deployment policy. Applications map the root
+package's results and errors into those concerns, including inbound size and
+trust policy.

docs/formats.md

@@ -91,7 +91,8 @@ of a named input. Missing variables and input references are errors.
 
 The optional `session_id` uses the same template data and input helper. Its
 rendered value is trimmed, omitted when empty, and limited to 256 Unicode code
-points.
+points. A nonblank direct request session ID bypasses this template completely;
+a blank direct value leaves the template behavior unchanged.
 
 ### Cache Control
 
@@ -136,7 +137,7 @@ A profile supplies model execution settings:
 
 ```yaml
 id: local-summary
-endpoint: http://localhost:8000/v1
+backend: openrouter
 model: example-model
 temperature: 0.2
 max_tokens: 500
@@ -144,7 +145,6 @@ top_p: 0.95
 timeout_seconds: 90
 service_tier: flex
 reasoning_effort: medium
-api_key_env: EXAMPLE_API_KEY
 extra_params:
   provider_option: enabled
 ```
@@ -152,7 +152,8 @@ extra_params:
 | Field | Required | Meaning |
 | --- | --- | --- |
 | `id` | yes | Non-empty profile identifier. IDs must be unique within one source. |
-| `endpoint` | yes | Non-empty OpenAI-compatible base URL, including an API version path when required. |
+| `backend` | unless `endpoint` is present | Backend registry ID. It is trimmed and registry membership is checked when the profile is prepared. |
+| `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. |
 | `model` | yes | Non-empty provider model name. |
 | `temperature` | no | Number from 0 through 2. |
 | `max_tokens` | no | Integer zero or greater. |
@@ -166,6 +167,13 @@ extra_params:
 Raw `api_key` is prohibited in profile YAML. Store only an environment
 variable name in `api_key_env`.
 
+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
+engine-scoped IDs with
+[`WithBackend`](../backends.go); exact registration validation belongs to its
+GoDoc.
+
 `extra_params` accepts null, booleans, finite numbers, strings, arrays, and
 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
@@ -176,8 +184,9 @@ they also cannot collide with the standard fields listed in the
 Execution settings resolve in this order:
 
 1. framework defaults;
-2. the selected profile; and
-3. request `ExecutionTargetOverride` values.
+2. the selected backend, when the profile names one;
+3. the selected profile; and
+4. request `ExecutionTargetOverride` values.
 
 The framework defaults are:
 
@@ -194,8 +203,16 @@ explicit zero is preserved. 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 request strings replace profile strings. A non-empty request
-`ExtraParams` map replaces the profile map rather than merging keys.
+Non-empty profile strings replace backend defaults, and non-empty request
+strings replace both. Request reasoning is the exception: a nil
+`ReasoningEffort` pointer inherits the profile, a pointer to a nonblank string
+trims and replaces it, and a pointer to a blank string clears it. Backend
+identity is retained when either layer overrides the endpoint, so the override
+also retains any engine-local capacity policy configured for that backend.
+Capacity configuration belongs to the Go
+[`Backend` API](../backends.go), not prompt or profile YAML. A non-empty
+`extra_params` map at each layer replaces the entire lower-precedence map
+rather than merging keys.
 The [outbound integration contract](integrations/openai-compatible-chat.md)
 defines how the effective settings are serialized.
 
@@ -217,8 +234,11 @@ invalid matching profile is an error and does not fall back. In-memory
 
 ## Built-In Profile Catalog
 
-Built-ins use the OpenRouter-compatible endpoint and
-`OPENROUTER_API_KEY`. A custom or in-memory profile with the same ID takes
+Every built-in selects the `openrouter` backend. The engine's built-in backend
+registry supplies `https://openrouter.ai/api/v1` and the environment-variable
+name `OPENROUTER_API_KEY`, so individual profiles contain only model and
+generation settings. Built-in profile files do not repeat those connection
+values. A custom or in-memory profile with the same profile ID takes
 precedence.
 
 | Provider | ID | Model |
@@ -270,6 +290,10 @@ prompt, profile, schema, or example files:
 - a request can provide a direct `APIKey` or override `APIKeyEnv`; 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.
 Direct keys are excluded from JSON results and redacted by public string
 formatters. Environment-variable names may appear in prepared metadata, but

docs/integrations/openai-compatible-chat.md

@@ -13,10 +13,15 @@ that produce these outbound settings.
 ## Endpoint And Method
 
 Generation sends an HTTP `POST` with `Content-Type: application/json`.
-A non-empty endpoint from the execution target overrides the client's
-configured base URL. After trailing slashes are removed,
-`/chat/completions` is appended. Generation fails before sending when neither
-source supplies an endpoint.
+Before the client is called, the engine resolves framework, backend, profile,
+and request values into one execution target. A non-empty endpoint from that
+target overrides the client's configured base URL. After trailing slashes are
+removed, `/chat/completions` is appended. Generation fails before sending when
+neither source supplies an 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
+does not serialize it in the provider request.
 
 ## Authentication
 
@@ -26,6 +31,12 @@ the client reads that variable and requires a non-empty value. The selected
 key is sent as `Authorization: Bearer <key>`. No authorization header is sent
 when neither mechanism is configured.
 
+The target contains the already resolved environment-variable name: an
+explicit request override takes precedence over profile metadata, which takes
+precedence over the backend default. Only the name reaches prepared metadata;
+the environment value is read just before the provider call and is never added
+to the JSON body.
+
 ## Request Body
 
 The request body always contains `model` and `messages`. The execution
@@ -36,20 +47,24 @@ 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.
 
-A non-empty session ID is trimmed, checked against the internal domain limit,
-and sent as top-level `session_id`. It is not sent as a session header.
+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
   present;
-- non-empty `service_tier` and `reasoning_effort`; and
+- 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.
 
-Extra parameters are merged directly into the top-level body after JSON
-serialization is verified. Empty keys and collisions with these reserved
-fields are rejected before any provider call:
+The engine resolves backend, profile, and request extra-parameter maps by
+whole-map replacement rather than key merging. The resulting effective map is
+then merged directly into the top-level body after JSON serialization is
+verified. Empty keys and collisions with these reserved fields are rejected
+before any provider call:
 
 - `model`
 - `session_id`
@@ -61,6 +76,9 @@ fields are rejected before any provider call:
 - `reasoning_effort`
 - `response_format`
 
+`backend_id`, `api_key_env`, and resolved credential values are not provider
+request fields.
+
 ## Response Handling
 
 Any 2xx response is decoded as an OpenAI-compatible chat response. The client

docs/internal/capacity.md

@@ -0,0 +1,100 @@
+# Internal Capacity Management
+
+## Purpose
+
+This document describes the implemented engine-local capacity coordination in
+`internal/capacity`. The [architecture policy](../policy/architecture.md) owns
+component boundaries, the [backend GoDoc](../../backends.go) owns exact public
+configuration semantics, and the
+[internal runner document](runner.md) owns orchestration around admission.
+
+Capacity scheduling is outside the provider wire contract. It does not add
+fields to execution targets, generated requests, prompt or profile YAML, or
+stable JSON values.
+
+## Construction And Pool Lifecycle
+
+Each root `NewEngine` call 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
+the same registrations still have independent capacity.
+
+Unlimited registered backends and endpoint-only profiles have no pool. Their
+admission and generation calls take the unrestricted fast path. An endpoint
+override does not change the selected backend ID and therefore does not change
+the pool.
+
+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
+
+The runner asks the manager to admit a run after resolving the prompt, profile,
+selected backend, effective execution target, credentials, and output contract,
+but before schema loading, artifact loading, or rendering. Admission is
+immediate: a limited pool either reserves a slot or returns the internal
+`ErrCapacityExceeded` identity. The root facade maps that identity to the
+public error 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
+preparation, initial generation, validation, every repair attempt, and all
+failure or cancellation 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
+client adaptation or built-in client construction. Initial generation and the
+default repairer receive the same wrapper.
+
+For each `Generate` call, the wrapper selects a pool from the request's
+effective backend ID. An unlimited call passes directly to the next client. A
+limited call acquires an active permit, invokes the next client, and defers
+permit release so ordinary returns and panic unwinding both restore capacity.
+Preparation and validation never hold an active permit.
+
+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
+through unchanged. It owns scheduling only; the concrete model client remains
+responsible for provider transport behavior.
+
+## 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
+client is not invoked. If a concurrent grant wins first, the call owns the
+permit and invokes the client with the original context, allowing the client
+to observe cancellation normally.
+
+This grant-or-cancel decision prevents lost and double-released permits.
+Admission leases and active permits are released after success, collaborator
+errors, validation failures, cancellation, and panic unwinding. Canceled
+waiters are unlinked so their contexts and requests are not retained by the
+pool.
+
+## Test Ownership
+
+The [manager tests](../../internal/capacity/manager_test.go) own policy
+validation, bounded admission, idempotent release, context handling, and
+unlimited admission. The
+[client tests](../../internal/capacity/client_test.go) own peak enforcement,
+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,
+lease lifetime, failure release, and shared initial/repair scheduling. 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
+[root error-boundary tests](../../errors_internal_test.go) own preservation of
+the public generation category and context identity when generation is
+canceled.

docs/internal/llm.md

@@ -20,6 +20,11 @@ orchestration. `OpenAICompatibleClient` is the built-in implementation. It
 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,
+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
 `http.Client` so Promptkit can apply its timeout default without mutating the
 caller's client. Generation then:
@@ -31,6 +36,10 @@ caller's client. Generation then:
 5. performs the outbound request under the applicable deadlines; and
 6. decodes the first response choice and token usage.
 
+`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,
 inbound HTTP behavior, or durable session store.
 
@@ -51,5 +60,7 @@ 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. They use local test servers and
-test transports; the default suite makes no live or paid provider requests.
+cancellation, and response-body suppression. The root transport contract test
+also verifies that resolved backend settings reach this client without
+serializing backend identity. All use local test servers or test transports;
+the default suite makes no live or paid provider requests.

docs/internal/overview.md

@@ -11,19 +11,23 @@ contributor workflow and validation.
 
 | Component | Implemented responsibility | References |
 | --- | --- | --- |
-| Root `promptkit` package | Provides the supported engine facade, source and injection options, public request and result values, built-in profile construction, extension interfaces, value conversion, redacted formatting, and public error mapping. | [Package GoDoc](../../doc.go), [engine assembly](../../engine.go) |
+| 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) |
 | `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) |
+| `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/domain` | Defines internal framework values for requests, artifacts, prompt definitions, profiles, execution targets, rendering, generation, and validation. | [Domain declarations](../../internal/domain/domain.go) |
 | `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/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 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/builtin` | Embeds the built-in execution profile catalog and combines it with an optional primary repository. | [Built-in catalog](../formats.md#built-in-profile-catalog), [repository](../../internal/profile/builtin/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/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/llm` | Defines the internal generation boundary and implements outbound OpenAI-compatible chat requests, response decoding, authentication, and deadline handling. | [Internal model client](llm.md) |
-| `internal/usecase` | Coordinates preparation and execution across internal sources, rendering, artifact loading, generation, validation, and optional repair. | [Internal runner](runner.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/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) |
 
 The root package assembles these internal components without exposing their
 representations. Consumers depend only on the root facade.

docs/internal/runner.md

@@ -17,7 +17,10 @@ and override semantics consumed by the runner.
 ## Collaborators
 
 `Runner` coordinates narrow internal interfaces for prompt definitions,
-profiles, artifacts, rendering, model generation, and validation. Schema
+profiles, backend resolution, artifacts, rendering, model generation, and
+validation. The root 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
 documents are loaded through the validator's optional schema-loader interface.
 An output repairer can be injected internally, but the ordinary runner
 constructor does not enable one.
@@ -25,43 +28,93 @@ constructor does not enable one.
 Each invocation carries its state in request, prepared-run, and result values.
 The runner has no durable run or session store.
 
-## Preparation Flow
+## Shared Preparation Pipeline
 
-`Prepare` performs the reusable pre-generation workflow:
+`Prepare` and `Run` share one private preparation pipeline split at the point
+where a run can be assigned to its selected backend pool. The resolution phase
+performs only the work needed to validate routing and admission:
 
-1. validate the prompt selection and load the prompt definition;
-2. hash the loaded definition;
+1. validate the required prompt selection and normalize any direct session ID;
+2. load the prompt definition and hash the original definition;
 3. select the request profile or the prompt's default profile;
-4. resolve application-neutral defaults, profile values, and explicit request
-   overrides in that order;
-5. validate endpoint, model, numeric overrides, and credential requirements;
-6. resolve the output contract and load a structured-output schema when
-   required;
-7. load and hash input artifacts;
-8. render and hash the prompt; and
-9. return the effective settings, source identities, messages, hashes, and
-   preparation timing.
+4. resolve the profile's backend ID, when present;
+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
+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;
+2. load and hash input artifacts;
+3. render messages and the prompt-defined session;
+4. apply any direct session ID;
+5. hash the effective rendered prompt; and
+6. construct the prepared value and preparation timing.
+
+`Prepare` runs both phases consecutively and never performs capacity admission.
+`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.
 
 Pointer-based numeric overrides preserve an explicit zero. Invalid negative or
-out-of-range values fail as invalid requests. A direct API key takes
-precedence over environment lookup for execution; secret values remain
-excluded from serialized metadata.
+out-of-range values fail as invalid requests. Endpoint overrides do not change
+the selected backend identity. Non-empty extra-parameter maps replace whole
+lower-precedence maps. A direct API key takes precedence over environment
+lookup; otherwise request, profile, and backend environment-variable names
+apply in that order. A profile requiring a direct key clears an inherited
+backend environment name unless the request supplies its own name. Secret
+values remain excluded from serialized metadata.
+
+Reasoning overrides are tri-state: nil inherits the profile, a pointer to a
+nonblank string trims and replaces it, and a pointer to a blank string clears
+it. A nonblank direct session is normalized before source loading, bypasses
+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.
+
+The registry is read-only after engine construction. Concurrent `Prepare` and
+`Run` calls resolve independent defensive backend values and keep all
+invocation state local.
 
 ## Run Flow
 
-`Run` calls `Prepare` rather than maintaining a second preparation path. It
-performs one initial generation call, builds the named output artifact, and
-validates that artifact. Invalid generated content remains a validation result;
-an inability to perform validation is an operational error.
+`Run` records its start time, performs the shared resolution phase, and asks
+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
+result; an inability to perform validation is an operational error.
+
+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
+each actual generation call.
 
 When an internal repairer is present, a JSON or JSON Schema content failure can
 trigger bounded repair attempts. Repair receives the effective execution
-target, validation errors, prior output, and structured-output specification.
-This capability remains internal and is not a public option.
+target and session ID, validation errors, prior output, and structured-output
+specification. The default repairer uses the same wrapped client 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
+option.
 
 A successful result includes the output artifact and raw output, validation
-state, prompt and rendered-prompt hashes, selected profile, effective settings,
-input hashes, token usage, a generated run identifier, and UTC timing.
+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
+repair attempt.
 
 ## Failure Categories
 
@@ -69,17 +122,38 @@ Package errors distinguish invalid requests, required profile selection,
 credential failures, and prompt, profile, artifact, rendering, generation, and
 validation failures. Wrapping preserves the package identities mapped by the
 public facade and retains collaborator identities where they are part of the
-internal contract. Context cancellation propagates through the invoked
-collaborator and is classified by the owning operation.
+internal contract.
+
+Admission capacity exhaustion retains the internal capacity identity and adds
+the selected backend ID as context. It is not recategorized as an invalid
+request or generation failure, and no partial result is returned. A context
+already done at admission retains its context identity directly. Cancellation
+while waiting for an active generation permit prevents client invocation when
+it wins the grant race; the model-client 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.
+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,
+is classified as a profile-load failure.
 
 ## Test Ownership And Changes
 
 The [runner tests](../../internal/usecase/runner_test.go) own preparation order,
-selection and override precedence, schema-before-generation behavior, hashing,
-generation and validation outcomes, bounded repair, credentials and redaction,
-error categories, artifact metadata, usage, and timing.
+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,
+error categories, artifact metadata, usage, and timing. The
+[capacity subsystem document](capacity.md) identifies the focused pool,
+waiter, and wrapped-client tests.
 
 Changes to orchestration should continue to use the existing package
-interfaces, keep request state local to an invocation, and preserve `Run`'s use
-of `Prepare`. Source, renderer, validator, or model-client contract changes
-belong first in their owning package and document.
+interfaces, keep request state local to an invocation, and preserve the shared
+resolution and completion pipeline. Source, renderer, validator, or
+model-client contract changes belong first in their owning package and
+document.

docs/internal/sources.md

@@ -24,13 +24,20 @@ duplicate detection, and source containment:
 
 `internal/profile` loads and validates execution profiles from an
 operating-system filesystem or an `fs.FS`. It supports a primary repository
-with fallback only when the primary reports that a profile is absent.
+with fallback only when the primary reports that a profile is absent. Strict
+YAML decoding recognizes the optional `backend` field, trims its value, and
+requires a model plus at least one non-blank backend or endpoint. Loading does
+not check registry membership because the available registry belongs to the
+assembled engine; the runner checks membership during preparation.
 
 `internal/profile/builtin` embeds the maintained built-in profile catalog and
-can place a caller-selected repository ahead of that catalog. Profile behavior
-is owned by the
+can place a caller-selected repository ahead of that catalog. Every embedded
+profile selects `openrouter` and inherits its endpoint and credential
+environment-variable name from the built-in backend registry rather than
+repeating those values. Profile behavior is owned by the
 [profile repository tests](../../internal/profile/repository_test.go), while
-catalog completeness, duplicate IDs, and overlay behavior are owned by the
+catalog completeness, the backend-selection invariant, duplicate IDs, and
+overlay behavior are owned by the
 [built-in repository tests](../../internal/profile/builtin/repository_test.go).
 
 ## Ordinary Artifacts

docs/policy/architecture.md

@@ -21,10 +21,16 @@ The implemented internal components consist of:
 
 - `internal/domain`, which owns framework data values shared by later internal
   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
@@ -41,21 +47,25 @@ The implemented internal components consist of:
 - `internal/usecase`, which coordinates preparation and execution across the
   internal framework components.
 
-The `examples/go-library/prepare` package is a maintained downstream consumer
-of the root facade. It does not expose a library package or participate in
-internal assembly.
+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 the internal repositories, renderer, validator,
-outbound client, and use-case runner while translating public values and
-errors at the library boundary. 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 an
-embedded `fs.FS` to the profile package. Artifact reading uses the domain model
-and application-neutral defaults. Validation uses the domain model, file
+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
+registry contains built-ins plus validated engine-scoped consumer additions.
+The facade 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
+an embedded `fs.FS` to the profile package. Artifact reading uses the domain
+model and application-neutral defaults. Validation uses the domain model, file
 catalog, and JSON Schema implementation. The model client uses the domain
 model, 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.
+internal component, including backend lookup and run admission.
 
 The current implementation follows this dependency direction:
 
@@ -72,9 +82,15 @@ downstream consumers, including Scriptorium
         narrow injected abstractions
 ```
 
-The facade coordinates internal components and adapts the supported public
-extension interfaces to narrow internal abstractions. Internal components must
-not depend on consumers or on Scriptorium.
+The backend registry depends on the domain model and shared JSON-value
+validation, has no mutation API after construction, and consumes the
+OpenAI-compatible reserved request-field rule owned by the model client. The
+capacity component depends on the domain model and the narrow internal
+model-client boundary, not on provider transport implementation. The model
+client does not depend on registry or capacity configuration. The facade
+coordinates internal components and adapts the supported public extension
+interfaces to narrow internal abstractions. Internal components must not depend
+on consumers or on Scriptorium.
 
 ## Repository And Consumer Boundary
 

docs/policy/documentation.md

@@ -78,6 +78,7 @@ mechanisms, not secret values.
 | 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. |
 | 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. |
@@ -85,9 +86,9 @@ mechanisms, not secret values.
 | Complete copyable artifacts | `examples/` | Valid inputs, Go programs, and other files intended to be copied or run. | Field-by-field reference, exact API declarations, and prose explanation. |
 
 Conditional owners do not require placeholder files or directories. Create a
-consumer, integration, subsystem, ADR, roadmap, or example document only when
-the corresponding implemented interface, decision, planned effort, or
-maintained artifact exists.
+consumer, integration, release, subsystem, ADR, roadmap, or example document
+only when the corresponding implemented interface, release, decision, planned
+effort, or maintained artifact exists.
 
 ## Boundary Rules
 
@@ -109,6 +110,23 @@ 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
+
+Files under `docs/releases/` may provide changelog-style summaries and
+migration guidance for a particular release. They are navigation and
+orientation aids, not canonical owners of public APIs, behavior, formats,
+integrations, architecture, release procedure, or other durable facts. When a
+reader needs detail beyond a short release-specific note, the release document
+must link to the applicable canonical documentation rather than reproduce its
+contract.
+
+The annotated tag message required by the
+[release procedure](../release.md#write-the-release-note) remains the
+authoritative release record. Supplemental release documents may be corrected,
+consolidated, archived, or removed at any time when they are no longer useful,
+provided maintained documentation does not depend on them and the annotated
+tag record remains intact.
+
 ### Security Topics
 
 This policy owns what documentation and examples may contain. Architecture owns
@@ -160,6 +178,10 @@ 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
+replacement. Before removal, update maintained incoming links so current
+documentation does not depend on an optional historical guide.
+
 Before completing documentation work:
 
 - verify affected behavior and examples;

docs/release.md

@@ -7,27 +7,91 @@ 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.
 
-The first planned release is `v0.1.0`. Do not create that tag until the
-framework has been extracted and the resulting public library has passed this
-procedure. Later tags use the `vMAJOR.MINOR.PATCH` form. While Promptkit remains
-pre-`v1`, release notes must identify intentional public API changes and any
-consumer migration required by them.
+`v0.1.0` is the initial published release. Later releases use semantic
+`vMAJOR.MINOR.PATCH` tags. Before `v1`, minor releases may change the public
+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.
 
-## Prepare The Release
+Promptkit releases are source-only. The annotated tag message is the release
+note; there is no separate hosted release or binary packaging step.
 
-Work from a clean checkout of the intended release commit, outside any Go
-workspace and without a local module replacement. Confirm the source commit is
-already published through the normal branch workflow.
+## Establish The Candidate
 
-From the Promptkit repository root, verify the checkout:
+Choose a version that has not been published and export it as
+`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
+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
-gowork=$(go env GOWORK)
-test -z "$gowork" || test "$gowork" = off
-test -z "$(git status --short)"
-git fetch --tags origin
+set -eu
+
+: "${RELEASE_VERSION:?export an unpublished vMAJOR.MINOR.PATCH version}"
+if ! printf '%s\n' "$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 intended commit through the normal `main` branch workflow before release;
+the tag procedure is not a substitute for publishing the source commit.
+
+## Validate The Candidate
+
 Confirm the module and root package metadata:
 
 ```sh
@@ -42,7 +106,7 @@ gitea.maximumdirect.net/eric/promptkit 1.25.5
 promptkit gitea.maximumdirect.net/eric/promptkit
 ```
 
-Run the same default Go validation required by the
+Run the complete maintainer validation required by the
 [development guide](development.md):
 
 ```sh
@@ -53,84 +117,158 @@ go build ./...
 go run ./examples/go-library/prepare
 ```
 
-Check every tracked Go file and repository whitespace:
+Check every tracked Go file. This command must produce no output:
 
 ```sh
-gofmt -l $(git ls-files '*.go')
+unformatted=$(
+    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)"
 ```
 
-The formatting command must produce no paths. Follow every maintained Markdown
-link and confirm its target exists. Review the repository for generated
-binaries, test or coverage output, credentials, template residue, and other
-files that do not belong in source control.
+## Write The Release Note
 
-Confirm that no workspace override is tracked and that `go.mod` contains no
-`replace` directive:
+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:
 
-```sh
-git ls-files go.work go.work.sum
-rg -n '^replace\b' go.mod
+```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.
 ```
 
-Both commands must produce no output. Re-run `git status --short` and require a
-clean result after every validation and review check.
-
-## Create And Publish The Tag
-
-Choose the semantic version from the intended compatibility change. Record the
-release commit before tagging:
+After writing it, require all release-note fields, the selected version, and
+the validated commit to be present:
 
 ```sh
-release_version=v0.1.0
-release_commit=$(git rev-parse HEAD)
+: "${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"
 ```
 
-Replace the example version for later releases and keep both values in the same
-shell for the remaining commands. Confirm the tag does not already exist
-locally or remotely:
+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
-test -z "$(git tag --list "$release_version")"
-test -z "$(git ls-remote --tags origin "refs/tags/$release_version")"
+check_release_candidate
 ```
 
-Create an annotated tag whose message identifies the release and records that
-the documented validation passed for the tagged commit:
+Create the annotated tag from the prepared release note and bind it explicitly
+to the validated commit:
 
 ```sh
-git tag --annotate "$release_version" \
-  --message "Promptkit $release_version; documented validation passed for $release_commit"
+git tag --annotate "$RELEASE_VERSION" \
+    --file "$RELEASE_NOTES_FILE" \
+    "$RELEASE_COMMIT"
 ```
 
-Inspect the tag before publication:
+Inspect both the tag message and its source commit before publication:
 
 ```sh
-git show --no-patch --decorate "$release_version"
-test "$(git rev-list -n 1 "$release_version")" = "$release_commit"
+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"
 ```
 
-Publish the tag without relying on a hosting-provider-specific release
-interface:
+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/$release_version"
+git push origin \
+    "refs/tags/$RELEASE_VERSION:refs/tags/$RELEASE_VERSION"
 ```
 
 ## Verify Publication
 
-Confirm that the remote tag object matches the local annotated tag and still
-resolves to the intended source commit:
+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 '{print $1}')
-test "$remote_tag" = "$(git rev-parse "refs/tags/$release_version")"
-test "$(git rev-list -n 1 "refs/tags/$release_version")" = "$release_commit"
+remote_tag=$(
+    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"
 ```
 
-Promptkit must publish the required tag before Scriptorium or another consumer
-publishes a release that depends on that version. Released consumer modules
-must not use a local replacement or unpublished Promptkit revision.
+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
 

docs/releases/v0.2.0.md

@@ -0,0 +1,243 @@
+# Promptkit v0.2.0
+
+This supplemental changelog and migration guide summarizes the consumer-facing
+changes from `v0.1.0` to `v0.2.0`. The annotated `v0.2.0` tag is the
+authoritative release record. Exact current contracts belong to the linked
+GoDoc and durable documentation.
+
+## Summary
+
+`v0.2.0` adds three major capabilities:
+
+- an engine-scoped registry for reusable OpenAI-compatible backend
+  definitions;
+- bounded, backend-specific run admission and model-generation concurrency;
+  and
+- direct per-run session IDs and tri-state reasoning-effort overrides.
+
+Existing endpoint-only profiles remain supported. Consumers can adopt backend
+registration and runtime overrides incrementally rather than rewriting all
+profiles during the upgrade.
+
+## Compatibility At A Glance
+
+Promptkit remains pre-`v1`, and this minor release includes source-level and
+behavioral changes that deserve review.
+
+| Area | `v0.1.0` consumer impact |
+| --- | --- |
+| Endpoint-only profiles | Continue to work without migration. |
+| Built-in profiles | Continue to use OpenRouter and `OPENROUTER_API_KEY`; they now select the built-in `openrouter` backend. |
+| Custom backends | Registration is optional. Existing profiles may keep their endpoint and credential configuration. |
+| Reasoning overrides | String assignments must migrate to the new pointer field. |
+| `RunRequest.Metadata` | Removed; delete assignments to this field. |
+| OpenRouter concurrency | Now limited to 16 active generations with waiting capacity of 1024 per engine. |
+| Public JSON | `v0.2.0` formalizes supported JSON representations; consumers relying on `v0.1.0` encodings should review the notes below. |
+| Unkeyed public struct literals | May require updates because fields were added. Keyed literals are recommended. |
+
+## Upgrade
+
+After the `v0.2.0` tag is published, update the module dependency with:
+
+```sh
+go get gitea.maximumdirect.net/eric/promptkit@v0.2.0
+go mod tidy
+```
+
+Run the consuming project's ordinary tests and race-enabled tests after the
+upgrade, especially if it calls one engine concurrently or persists Promptkit
+JSON values.
+
+## Backend Registry
+
+Consumers may now register reusable OpenAI-compatible backend definitions with
+`WithBackend`, then select them by ID from file-backed or in-memory profiles.
+A backend can supply its endpoint, API-key environment-variable name,
+request-wide extra parameters, and optional capacity policy.
+
+Registrations are immutable and belong to one engine. Consumer registrations
+can add new IDs but cannot replace Promptkit's reserved `openrouter` backend.
+Profiles that select a backend may still override its endpoint without losing
+the backend's routing or capacity identity.
+
+An existing endpoint-only in-memory profile remains valid:
+
+```go
+promptkit.Profile{
+	ID:       "local",
+	Endpoint: "http://localhost:8000/v1",
+	Model:    "example-model",
+}
+```
+
+Adopting the registry is optional and can be done when several profiles should
+share connection or capacity settings:
+
+```go
+engine, err := promptkit.NewEngine(
+	promptkit.Config{PromptDir: "prompts"},
+	promptkit.WithBackend(promptkit.Backend{
+		ID:        "local",
+		Endpoint:  "http://localhost:8000/v1",
+		APIKeyEnv: "LOCAL_LLM_API_KEY",
+	}),
+	promptkit.WithProfiles(promptkit.Profile{
+		ID:        "local-summary",
+		BackendID: "local",
+		Model:     "example-model",
+	}),
+)
+```
+
+See the
+[custom-backend consumer guide](../consumers/pkg-promptkit.md#register-a-custom-backend)
+for task-oriented usage. The
+[`Backend` and `WithBackend` GoDoc](../../backends.go) owns exact registration,
+validation, copying, defaulting, and uniqueness semantics. The
+[framework format reference](../formats.md) owns the profile `backend` field
+and execution precedence.
+
+## Backend-Specific Concurrency
+
+Each registered backend may now define:
+
+- an active model-generation limit; and
+- a bounded number of additional admitted `Run` calls.
+
+Promptkit owns scheduling for both its built-in model client and an injected
+`LLMClient`. `Run` remains synchronous: an admitted caller waits for its
+ordinary result, while a call beyond the bounded admission capacity returns
+`ErrCapacityExceeded`. Capacity is engine-local and keyed by backend ID.
+Endpoint-only profiles and custom backends without a configured limit remain
+unlimited.
+
+The built-in OpenRouter backend now permits 16 active generations and 1024
+additional admitted calls per engine. Applications that can exceed this bound
+should handle capacity exhaustion separately from provider and request
+failures:
+
+```go
+result, err := engine.Run(ctx, request)
+if errors.Is(err, promptkit.ErrCapacityExceeded) {
+	// Apply application-specific overload or retry policy.
+}
+```
+
+Promptkit does not prescribe retries or map this error to an HTTP status. See
+the
+[concurrency consumer guidance](../consumers/pkg-promptkit.md#limit-backend-concurrency)
+and the [`Backend` GoDoc](../../backends.go) for the canonical configuration
+contract. Runtime behavior and public error identities belong to the
+[`Engine.Run` GoDoc](../../engine.go).
+
+## Per-Run Session IDs
+
+`RunRequest.SessionID` can now supply a consumer-managed correlation ID for one
+`Prepare` or `Run` invocation. A nonblank direct value overrides the prompt's
+session template and is exposed in prepared values, results, injected-client
+requests, and provider observability. Session IDs should therefore be stable,
+non-secret values.
+
+```go
+result, err := engine.Run(ctx, promptkit.RunRequest{
+	PromptID: "meeting.summary",
+	SessionID: "conversation-42",
+})
+```
+
+The built-in OpenAI-compatible client sends a nonempty effective session as the
+top-level `session_id` request-body field, not as an `x-session-id` header. See
+the
+[session and reasoning consumer guide](../consumers/pkg-promptkit.md#set-a-per-run-session-and-reasoning),
+the [`RunRequest` GoDoc](../../types.go), and the
+[OpenAI-compatible request contract](../integrations/openai-compatible-chat.md#request-body)
+for exact normalization, length, exposure, and wire behavior.
+
+## Per-Run Reasoning Effort
+
+`ExecutionTargetOverride.ReasoningEffort` changed from `string` to `*string` so
+one request can distinguish inheritance, replacement, and explicit clearing.
+
+Update a `v0.1.0` override like this:
+
+```go
+// v0.1.0
+Execution: &promptkit.ExecutionTargetOverride{
+	ReasoningEffort: "high",
+}
+```
+
+to:
+
+```go
+// v0.2.0
+reasoning := "high"
+Execution: &promptkit.ExecutionTargetOverride{
+	ReasoningEffort: &reasoning,
+}
+```
+
+The three states are:
+
+- `nil` inherits the selected profile's value;
+- a pointer to a nonblank string replaces it for that invocation; and
+- a pointer to an empty or whitespace-only string clears it for that
+  invocation.
+
+This allows consumers to consolidate profiles that differed only by reasoning
+effort. The [`ExecutionTargetOverride` GoDoc](../../types.go) owns the exact
+override contract.
+
+## Other Migration Notes
+
+### Remove `RunRequest.Metadata`
+
+`RunRequest.Metadata` is no longer part of the public request. Remove any
+assignment to that field. Use application-owned state keyed by `RunResult.RunID`
+or a direct `SessionID` when correlation is needed; these identifiers have
+different purposes, so choose according to the application's lifecycle.
+
+### Review Persisted JSON
+
+`v0.2.0` defines stable JSON representations for the public result, artifact,
+execution, validation, and model-client values listed in the
+[package documentation](../../doc.go). Consumers that treated `v0.1.0`
+reflection-derived encodings as stable should update fixtures and stored-data
+adapters.
+
+In particular:
+
+- `RunResult` encodes elapsed time as integer milliseconds in `duration_ms`
+  instead of encoding `time.Duration` under `duration`;
+- result JSON can include the new `session_id` and `selected_backend_id`
+  fields;
+- execution-target JSON can include `backend_id`; and
+- artifact and target-presence fields now use their documented lower-case
+  names.
+
+The `v0.2.0` `RunResult` decoder reads `duration_ms`; it does not translate a
+persisted `v0.1.0` `duration` field. Transform old payloads before decoding
+when preserving their elapsed duration matters.
+
+### Prefer Keyed Struct Literals
+
+New fields were added to several public structs. Replace positional composite
+literals with keyed literals so future additive fields do not cause another
+source migration.
+
+## Migration Checklist
+
+- Update the module dependency and run the consumer's tests.
+- Change reasoning overrides from strings to pointers.
+- Remove uses of `RunRequest.Metadata`.
+- Review unkeyed Promptkit struct literals.
+- Decide whether shared endpoints should move into registered backends.
+- If using built-in OpenRouter profiles at high concurrency, handle
+  `ErrCapacityExceeded` and review the new engine-local bound.
+- Review stored JSON, fixtures, and downstream decoders.
+- Optionally replace profile-specific session or reasoning variants with
+  per-run overrides.
+
+For complete consumer workflows, use the
+[package consumer guide](../consumers/pkg-promptkit.md) and maintained
+[offline execution example](../../examples/go-library/run/main.go).

docs/roadmap/concurrency.md

@@ -0,0 +1,236 @@
+# 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/future.md

@@ -0,0 +1,53 @@
+# Future Feature Ideas
+
+## Purpose
+
+This document catalogs reasonably specific ideas that may be useful in future
+Promptkit development. It is an idea pool, not a commitment, schedule, or
+description of current behavior.
+
+Ideas belong here while they are worth retaining but have not been selected
+for active development. Keep each entry at the level of intended capability,
+consumer value, and important scope boundaries. Defer API design,
+implementation details, sequencing, and acceptance criteria until an idea is
+selected.
+
+## Using This Catalog
+
+- Add an idea when its purpose and likely value can be stated clearly.
+- Keep entries independent enough that maintainers can evaluate and select
+  them individually.
+- Note significant dependencies or boundary concerns, but do not turn entries
+  into implementation plans.
+- Treat inclusion as an invitation to evaluate, not as approval or priority.
+- 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.
+- Remove ideas that are no longer relevant. Retain a rejected idea only when
+  its rationale is likely to prevent repeated reconsideration.
+
+Future capabilities must continue to respect the
+[architecture policy](../policy/architecture.md), particularly Promptkit's
+role as an application-neutral library and its boundary with downstream
+consumers.
+
+## Ideas
+
+No ideas are currently cataloged. Backend-specific concurrency management has
+been selected for active planning in the
+[focused concurrency roadmap](concurrency.md).
+
+## Entry Format
+
+Use a short heading followed by a concise summary. Add focused bullets when
+they help preserve important scope boundaries without becoming an
+implementation plan:
+
+```markdown
+### Idea name
+
+Describe the intended capability, who benefits, and the most important scope
+boundary or dependency.
+
+- Optionally record an important behavior or boundary.
+```

docs/roadmap/implementation.md

@@ -0,0 +1,841 @@
+# 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

@@ -12,7 +12,10 @@ import (
 	"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"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile/builtin"
@@ -22,42 +25,102 @@ import (
 	"gitea.maximumdirect.net/eric/promptkit/internal/validate"
 )
 
-// ErrInvalidConfig indicates invalid public engine configuration.
+// ErrInvalidConfig identifies invalid engine construction, including missing
+// required configuration, invalid options or backend registrations, and a nil
+// Engine receiver.
 var ErrInvalidConfig = errors.New("invalid engine configuration")
 
 var (
-	ErrInvalidRequest   = errors.New("invalid run request")
-	ErrPromptNotFound   = errors.New("prompt not found")
-	ErrProfileNotFound  = errors.New("profile not found")
-	ErrProfileRequired  = errors.New("profile selection is required")
-	ErrPromptLoad       = errors.New("failed to load prompt definition")
-	ErrProfileLoad      = errors.New("failed to load execution profile")
+	// ErrInvalidRequest identifies a request whose required values, overrides,
+	// credentials, or effective settings are invalid.
+	ErrInvalidRequest = errors.New("invalid run request")
+	// ErrPromptNotFound identifies a requested prompt ID or version that is not
+	// present in the selected prompt source. It does not also match
+	// ErrPromptLoad.
+	ErrPromptNotFound = errors.New("prompt not found")
+	// ErrProfileNotFound identifies a selected profile ID that is absent from
+	// every configured profile source. It does not also match ErrProfileLoad.
+	ErrProfileNotFound = errors.New("profile not found")
+	// ErrProfileRequired identifies a request for which neither RunRequest.ProfileID
+	// nor the selected prompt's default profile is present. Such an error also
+	// matches ErrInvalidRequest.
+	ErrProfileRequired = errors.New("profile selection is required")
+	// ErrPromptLoad identifies a failure to read, decode, validate, select, or
+	// hash a prompt definition, except for the not-found case represented by
+	// ErrPromptNotFound.
+	ErrPromptLoad = errors.New("failed to load prompt definition")
+	// ErrProfileLoad identifies a failure to read, decode, validate, or select
+	// 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
+	// unset or empty when no direct RunRequest.APIKey takes precedence. Such an
+	// error also matches ErrInvalidRequest.
 	ErrAPIKeyEnvMissing = errors.New("api_key_env points to an unset environment variable")
-	ErrArtifactLoad     = errors.New("failed to load artifact")
-	ErrPromptRender     = errors.New("failed to render prompt")
-	ErrLLMGenerate      = errors.New("failed to generate output")
-	ErrValidation       = errors.New("failed to validate output")
+	// ErrArtifactLoad identifies a failure to resolve an input artifact. Errors
+	// returned by an injected ArtifactReader remain available through errors.Is.
+	ErrArtifactLoad = errors.New("failed to load artifact")
+	// 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
+	// already admitted ConcurrencyLimit + QueueCapacity calls. 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
+	// through errors.Is.
+	ErrLLMGenerate = errors.New("failed to generate output")
+	// 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.
+//
+// An Engine is safe for concurrent calls to [Engine.Prepare] and [Engine.Run].
+// Each Engine owns independent backend-capacity pools that coordinate Run
+// 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
 }
 
-// Config configures a public Promptkit engine.
+// Config selects the directory-backed sources and built-in model-client
+// transport used by [NewEngine]. Config has no stable JSON representation.
 type Config struct {
-	PromptDir  string
+	// PromptDir is the directory searched recursively for prompt definitions.
+	// 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
+	// embedded built-in profiles. An empty value selects only built-ins unless
+	// profile options are also supplied.
 	ProfileDir string
-	SchemaDir  string
+	// SchemaDir is the root for JSON Schema files. An empty value uses the
+	// current directory. WithSchemaFS or WithSchemaFile replaces this source.
+	SchemaDir string
 	// Timeout is the transport-wide safety cap for the built-in LLM client
-	// when HTTPClient is absent or has a non-positive timeout.
+	// when HTTPClient is absent or has a non-positive timeout. A zero or negative
+	// value selects the 10-minute default.
 	Timeout time.Duration
 	// HTTPClient is cloned for the built-in LLM client. Its positive Timeout
-	// takes precedence over Config.Timeout as the transport-wide safety cap.
+	// takes precedence over Timeout. A zero or negative client Timeout inherits
+	// Timeout or the 10-minute default. The supplied client is not mutated. This
+	// field is ignored when WithLLMClient is used.
 	HTTPClient *http.Client
 }
 
 // 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,
+// model-client, and artifact-reader category, the last non-nil valid option
+// replaces earlier options in that category. WithBackend is the additive
+// exception: unique registrations accumulate, and a repeated backend ID is an
+// error rather than a replacement. An invalid option fails construction even
+// if a later option would replace it.
 type Option interface {
 	apply(*engineOptions) error
 }
@@ -74,6 +137,7 @@ type engineOptions struct {
 	promptDefs      promptdef.Repository
 	profiles        profile.Repository
 	memoryProfiles  profile.Repository
+	backends        []domain.Backend
 	validator       validate.Validator
 	promptSource    bool
 	profileSource   bool
@@ -82,7 +146,12 @@ type engineOptions struct {
 	artifactSource  bool
 }
 
-// WithLLMClient injects a custom LLM client for execution.
+// WithLLMClient replaces the built-in model client used by [Engine.Run].
+//
+// 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].
 func WithLLMClient(client LLMClient) Option {
 	return optionFunc(func(options *engineOptions) error {
 		if client == nil {
@@ -93,7 +162,11 @@ func WithLLMClient(client LLMClient) Option {
 	})
 }
 
-// WithArtifactReader injects a reader for every input artifact reference.
+// WithArtifactReader replaces the default reader for every input artifact
+// reference, regardless of its ArtifactRef.Type.
+//
+// A nil reader makes NewEngine fail with ErrInvalidConfig. The reader may be
+// called concurrently.
 func WithArtifactReader(reader ArtifactReader) Option {
 	return optionFunc(func(options *engineOptions) error {
 		if reader == nil {
@@ -109,6 +182,9 @@ func WithArtifactReader(reader ArtifactReader) Option {
 //
 // The source uses the same strict prompt YAML rules as configured prompt
 // directories, and prompt content_file paths resolve within this source.
+// fsys must be non-nil and root must be non-empty; otherwise NewEngine fails
+// with ErrInvalidConfig. This option replaces Config.PromptDir and earlier
+// prompt-source options.
 func WithPromptFS(fsys fs.FS, root string) Option {
 	return optionFunc(func(options *engineOptions) error {
 		if fsys == nil {
@@ -125,7 +201,9 @@ func WithPromptFS(fsys fs.FS, root string) Option {
 
 // WithPromptFile loads prompt definitions from the single prompt file at path.
 //
-// Relative prompt content_file paths resolve from the file's directory.
+// Relative prompt content_file paths resolve from the file's directory. path
+// must name an existing non-directory file when NewEngine applies the 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)
@@ -142,6 +220,10 @@ func WithPromptFile(path string) Option {
 //
 // Profiles from this source overlay built-in profiles. Profile YAML must use
 // api_key_env for environment-based credentials; raw API keys are rejected.
+// fsys must be non-nil and root must be non-empty; otherwise NewEngine fails
+// with ErrInvalidConfig. This option replaces Config.ProfileDir and earlier
+// file or 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 {
@@ -159,7 +241,10 @@ func WithProfileFS(fsys fs.FS, root string) Option {
 // 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
-// environment-based credentials; raw API keys are rejected.
+// environment-based credentials; raw API keys are rejected. path must name an
+// existing non-directory file when NewEngine applies the option. This option
+// replaces Config.ProfileDir and 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)
@@ -174,6 +259,11 @@ func WithProfileFile(path string) Option {
 
 // WithProfiles configures in-memory profiles that take precedence over
 // configured profile files and built-in profiles.
+//
+// NewEngine validates and copies every profile. IDs must be unique within one
+// call. An invalid profile, duplicate ID, or unsupported ExtraParams value
+// makes construction fail with ErrInvalidConfig. 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)
@@ -189,7 +279,9 @@ func WithProfiles(profiles ...Profile) Option {
 // WithSchemaFS loads JSON Schema documents from fsys under root.
 //
 // Prompt schema_path values resolve within this source when schema validation
-// or structured output is requested.
+// or structured output is requested. fsys must be non-nil and root must be
+// non-empty; otherwise NewEngine fails with ErrInvalidConfig. This 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 {
@@ -206,7 +298,9 @@ func WithSchemaFS(fsys fs.FS, root string) Option {
 
 // WithSchemaFile loads JSON Schema documents from the single schema file at path.
 //
-// Prompt schema_path values refer to the file's base name.
+// Prompt schema_path values refer to the file's base name. path must name an
+// existing non-directory file when NewEngine applies the option. This 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)
@@ -220,6 +314,17 @@ func WithSchemaFile(path string) Option {
 }
 
 // NewEngine constructs an Engine from configuration and options.
+//
+// Options are applied in order according to [Option]. PromptDir is required
+// unless a prompt-source option is present. Construction validates option
+// arguments, in-memory profiles, and backend registrations but defers reading
+// and validating prompt, file-backed profile, and schema contents until Prepare
+// or Run needs them.
+//
+// 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.
 func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 	var options engineOptions
 	for _, opt := range opts {
@@ -247,6 +352,16 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 		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)
+	}
+
+	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
@@ -267,6 +382,7 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
 		}
 	}
+	llmClient = capacity.NewClient(capacityManager, llmClient)
 
 	artifacts := options.artifactReader
 	if !options.artifactSource {
@@ -277,10 +393,12 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 		runner: usecase.NewRunner(
 			promptDefs,
 			profiles,
+			backendRegistry,
 			artifacts,
 			prompt.NewGoRenderer(),
 			llmClient,
 			validator,
+			capacityManager,
 		),
 	}, nil
 }
@@ -305,7 +423,23 @@ func fileSource(name string) (fs.FS, string, error) {
 	return os.DirFS(dir), filepath.ToSlash(base), nil
 }
 
-// Prepare resolves a prompt request without calling an LLM.
+// Prepare resolves and renders a prompt request without calling an LLM.
+//
+// Prepare selects the prompt and profile, resolves any selected backend and
+// effective execution settings, resolves the output contract, loads and hashes
+// inputs, loads structured-output schema metadata when required, and renders
+// the session ID and messages. The returned PreparedRun is owned by the caller
+// and never contains a resolved API-key value, model output, or validation
+// result.
+//
+// A nil Engine returns an error matching ErrInvalidConfig. Request and
+// preparation failures may match ErrInvalidRequest, ErrPromptNotFound,
+// ErrPromptLoad, ErrProfileNotFound, ErrProfileLoad, ErrProfileRequired,
+// ErrAPIKeyEnvMissing, ErrArtifactLoad, ErrPromptRender, or ErrValidation as
+// 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.
 func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, error) {
 	if e == nil || e.runner == nil {
 		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
@@ -323,7 +457,24 @@ func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, err
 	return fromDomainPreparedRun(prepared), nil
 }
 
-// Run executes a prompt request and returns the generated artifact and metadata.
+// Run prepares a request, invokes the configured LLMClient, and validates the
+// generated output.
+//
+// A content-validation failure is a successful run whose
+// RunResult.Validation has Status ValidationFailed. An inability to perform
+// validation returns an error matching ErrValidation and no partial result.
+// The public Engine does not perform output repair, so validation is
+// single-pass even when OutputContract.RepairAttempts is positive.
+//
+// Run can return every error category documented by [Engine.Prepare], plus
+// ErrCapacityExceeded and ErrLLMGenerate. ErrCapacityExceeded identifies
+// rejection before artifacts, schemas, rendering, or model generation because
+// the selected backend's admission capacity is full; it does not match
+// ErrInvalidRequest or ErrLLMGenerate. Errors from injected clients 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) {
 	if e == nil || e.runner == nil {
 		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)

engine_test.go

@@ -281,6 +281,9 @@ func TestEngineExecutionSettingPrecedence(t *testing.T) {
 	intPointer := func(value int) *int {
 		return &value
 	}
+	stringPointer := func(value string) *string {
+		return &value
+	}
 
 	defaultsProfile := executionProfileFixture{
 		id:               "settings-defaults",
@@ -388,7 +391,7 @@ func TestEngineExecutionSettingPrecedence(t *testing.T) {
 				TopP:            floatPointer(requestTarget.TopP),
 				TimeoutSeconds:  intPointer(requestTarget.TimeoutSeconds),
 				ServiceTier:     requestTarget.ServiceTier,
-				ReasoningEffort: requestTarget.ReasoningEffort,
+				ReasoningEffort: stringPointer(requestTarget.ReasoningEffort),
 				APIKeyEnv:       requestTarget.APIKeyEnv,
 				ExtraParams:     requestTarget.ExtraParams,
 			},
@@ -452,6 +455,43 @@ func TestEngineExecutionSettingPrecedence(t *testing.T) {
 			}
 		})
 	}
+
+	t.Run("blank request reasoning clears profile setting", func(t *testing.T) {
+		profile := executionProfileFixture{
+			id:              "settings-reasoning-clear",
+			endpoint:        "http://profile-reasoning.test/v1",
+			model:           "profile-reasoning-model",
+			reasoningEffort: "medium",
+		}
+		profileDir := t.TempDir()
+		writeExecutionProfileFixture(t, profileDir, profile)
+		engine, err := promptkit.NewEngine(promptkit.Config{
+			PromptDir:  frameworkPromptDir,
+			ProfileDir: profileDir,
+			SchemaDir:  frameworkSchemaDir,
+		})
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
+			PromptID:  frameworkMarkdownSummaryPromptID,
+			ProfileID: profile.id,
+			Inputs: map[string]promptkit.ArtifactRef{
+				"transcript": promptkit.Inline("Nia labels the archive."),
+				"glossary":   promptkit.Inline("archive: A catalogued collection."),
+			},
+			Execution: &promptkit.ExecutionTargetOverride{
+				ReasoningEffort: stringPointer(" \t "),
+			},
+		})
+		if err != nil {
+			t.Fatalf("prepare engine: %v", err)
+		}
+		if prepared.EffectiveModelParams.ReasoningEffort != "" {
+			t.Fatalf("expected blank request reasoning to clear profile value, got %q", prepared.EffectiveModelParams.ReasoningEffort)
+		}
+	})
 }
 
 func TestRunSucceedsWithInjectedLLMClient(t *testing.T) {
@@ -571,19 +611,26 @@ func TestEngineRunWithDirectorySourcesAndFileInputs(t *testing.T) {
 
 func TestRunPassesPreparedRequestToInjectedLLMClient(t *testing.T) {
 	const directKey = "direct-injected-key"
+	const directSession = "assembled-session"
 	fake := &fakeLLMClient{
 		response: &promptkit.GenerateResponse{Content: `{"events":[{"title":"Archive labelled"}]}`},
 	}
 	engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(fake))
 
-	_, err := engine.Run(context.Background(), promptkit.RunRequest{
-		PromptID: frameworkStructuredEventsPromptID,
-		APIKey:   directKey,
+	runRequest := promptkit.RunRequest{
+		PromptID:  frameworkStructuredEventsPromptID,
+		SessionID: "  " + directSession + "  ",
+		APIKey:    directKey,
 		Inputs: map[string]promptkit.ArtifactRef{
 			"transcript": promptkit.Inline("Rin opens the gate."),
 			"glossary":   promptkit.Inline("gate: A guarded passage."),
 		},
-	})
+	}
+	prepared, err := engine.Prepare(context.Background(), runRequest)
+	if err != nil {
+		t.Fatalf("expected prepare to succeed, got %v", err)
+	}
+	result, err := engine.Run(context.Background(), runRequest)
 	if err != nil {
 		t.Fatalf("expected run to succeed, got %v", err)
 	}
@@ -594,6 +641,16 @@ func TestRunPassesPreparedRequestToInjectedLLMClient(t *testing.T) {
 	if len(req.Prompt.Messages) != 2 || !strings.Contains(req.Prompt.Messages[1].Content, "Rin opens the gate.") {
 		t.Fatalf("expected rendered prompt in generate request, got %+v", req.Prompt)
 	}
+	if prepared.SessionID != directSession ||
+		req.Prompt.SessionID != directSession ||
+		result.SessionID != directSession {
+		t.Fatalf(
+			"direct session did not propagate consistently: prepared=%q generated=%q result=%q",
+			prepared.SessionID,
+			req.Prompt.SessionID,
+			result.SessionID,
+		)
+	}
 	if req.StructuredOutput == nil || req.StructuredOutput.Type != promptkit.StructuredOutputJSONSchema || req.StructuredOutput.JSONSchema == nil {
 		t.Fatalf("expected structured output handoff, got %+v", req.StructuredOutput)
 	}
@@ -755,6 +812,92 @@ func TestRunUsesDirectAPIKeyWithDefaultLLMClient(t *testing.T) {
 	}
 }
 
+func TestRunUsesResolvedBackendWithBuiltInLLMClient(t *testing.T) {
+	const (
+		backendID = "local-test"
+		envName   = "PROMPTKIT_BACKEND_TRANSPORT_KEY"
+		apiKey    = "synthetic-backend-key"
+	)
+	t.Setenv(envName, apiKey)
+
+	var (
+		gotAuth string
+		gotBody map[string]any
+	)
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotAuth = r.Header.Get("Authorization")
+		if r.URL.Path != "/v1/chat/completions" {
+			t.Errorf("unexpected path: %s", r.URL.Path)
+		}
+		if err := json.NewDecoder(r.Body).Decode(&gotBody); err != nil {
+			t.Errorf("decode request body: %v", err)
+		}
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = w.Write([]byte(`{
+  "choices": [{"message": {"role": "assistant", "content": "# Summary\n\nDone."}}],
+  "usage": {"prompt_tokens": 3, "completion_tokens": 4, "total_tokens": 7}
+}`))
+	}))
+	defer server.Close()
+
+	engine, err := promptkit.NewEngine(promptkit.Config{
+		PromptDir: frameworkPromptDir,
+		SchemaDir: frameworkSchemaDir,
+	},
+		promptkit.WithBackend(promptkit.Backend{
+			ID:        backendID,
+			Endpoint:  server.URL + "/v1",
+			APIKeyEnv: envName,
+			ExtraParams: map[string]any{
+				"provider": "synthetic",
+			},
+		}),
+		promptkit.WithProfiles(promptkit.Profile{
+			ID:        "backend-transport",
+			BackendID: backendID,
+			Model:     "test-model",
+		}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID:  frameworkMarkdownSummaryPromptID,
+		ProfileID: "backend-transport",
+		Inputs: map[string]promptkit.ArtifactRef{
+			"transcript": promptkit.Inline("Rin opens the gate."),
+			"glossary":   promptkit.Inline("gate: A guarded passage."),
+		},
+	})
+	if err != nil {
+		t.Fatalf("run with resolved backend: %v", err)
+	}
+	if gotAuth != "Bearer "+apiKey {
+		t.Fatalf("unexpected Authorization header: %q", gotAuth)
+	}
+	if gotBody["model"] != "test-model" || gotBody["provider"] != "synthetic" {
+		t.Fatalf("backend defaults did not reach provider payload: %#v", gotBody)
+	}
+	for _, field := range []string{"backend_id", "api_key_env"} {
+		if _, ok := gotBody[field]; ok {
+			t.Fatalf("internal metadata field %q was serialized to provider payload: %#v", field, gotBody)
+		}
+	}
+	bodyJSON, err := json.Marshal(gotBody)
+	if err != nil {
+		t.Fatalf("marshal captured provider payload: %v", err)
+	}
+	if strings.Contains(string(bodyJSON), apiKey) {
+		t.Fatalf("credential value was serialized to provider payload: %s", bodyJSON)
+	}
+	if result.SelectedBackendID != backendID ||
+		result.EffectiveModelParams.Endpoint != server.URL+"/v1" ||
+		result.EffectiveModelParams.APIKeyEnv != envName {
+		t.Fatalf("unexpected resolved backend metadata: %+v", result)
+	}
+}
+
 func TestPrepareDirectAPIKeyBypassesMissingEnvWithoutLeakingOrHashing(t *testing.T) {
 	const missingEnv = "PROMPTKIT_PUBLIC_PREPARE_MISSING"
 	const firstKey = "first-direct-key"
@@ -1020,6 +1163,7 @@ func TestPublicErrorsSupportErrorsIs(t *testing.T) {
 		client    promptkit.LLMClient
 		schemaDir string
 		want      error
+		notWant   error
 	}{
 		{
 			name:   "invalid request",
@@ -1028,10 +1172,11 @@ func TestPublicErrorsSupportErrorsIs(t *testing.T) {
 			want:   promptkit.ErrInvalidRequest,
 		},
 		{
-			name:   "prompt not found",
-			req:    promptkit.RunRequest{PromptID: "missing.prompt"},
-			client: &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}},
-			want:   promptkit.ErrPromptNotFound,
+			name:    "prompt not found",
+			req:     promptkit.RunRequest{PromptID: "missing.prompt"},
+			client:  &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}},
+			want:    promptkit.ErrPromptNotFound,
+			notWant: promptkit.ErrPromptLoad,
 		},
 		{
 			name: "profile not found",
@@ -1042,8 +1187,9 @@ func TestPublicErrorsSupportErrorsIs(t *testing.T) {
 					"transcript": promptkit.Inline("Rin opens the gate."),
 				},
 			},
-			client: &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}},
-			want:   promptkit.ErrProfileNotFound,
+			client:  &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}},
+			want:    promptkit.ErrProfileNotFound,
+			notWant: promptkit.ErrProfileLoad,
 		},
 		{
 			name: "artifact load",
@@ -1114,6 +1260,9 @@ func TestPublicErrorsSupportErrorsIs(t *testing.T) {
 			if !errors.Is(err, tc.want) {
 				t.Fatalf("expected errors.Is(%v), got %v", tc.want, err)
 			}
+			if tc.notWant != nil && errors.Is(err, tc.notWant) {
+				t.Fatalf("did not expect errors.Is(%v), got %v", tc.notWant, err)
+			}
 		})
 	}
 }
@@ -1265,6 +1414,18 @@ func TestPrepareUsesBuiltInProfileWithoutProfileDir(t *testing.T) {
 	if prepared.SelectedProfileID != "mistral-small-3" {
 		t.Fatalf("unexpected selected profile: %q", prepared.SelectedProfileID)
 	}
+	if prepared.SelectedBackendID != promptkit.BackendOpenRouter {
+		t.Fatalf("unexpected selected backend: %q", prepared.SelectedBackendID)
+	}
+	if prepared.EffectiveModelParams.BackendID != promptkit.BackendOpenRouter {
+		t.Fatalf("unexpected effective backend: %q", prepared.EffectiveModelParams.BackendID)
+	}
+	if prepared.EffectiveModelParams.Endpoint != "https://openrouter.ai/api/v1" {
+		t.Fatalf("unexpected built-in endpoint: %q", prepared.EffectiveModelParams.Endpoint)
+	}
+	if prepared.EffectiveModelParams.APIKeyEnv != "OPENROUTER_API_KEY" {
+		t.Fatalf("unexpected built-in api key environment name: %q", prepared.EffectiveModelParams.APIKeyEnv)
+	}
 	if prepared.EffectiveModelParams.Model != "mistralai/mistral-small-3.2-24b-instruct" {
 		t.Fatalf("unexpected built-in model: %q", prepared.EffectiveModelParams.Model)
 	}
@@ -1635,6 +1796,7 @@ func TestOpenAICompatibleProfileRunsThroughNormalProfilePath(t *testing.T) {
 	fake := &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}}
 	prof := promptkit.OpenAICompatibleProfile(promptkit.OpenAICompatibleProfileConfig{
 		ID:             "template-profile",
+		BackendID:      "  openrouter  ",
 		Endpoint:       "http://template/v1",
 		Model:          "template-model",
 		APIKeyRequired: true,
@@ -1666,7 +1828,9 @@ func TestOpenAICompatibleProfileRunsThroughNormalProfilePath(t *testing.T) {
 	if len(fake.requests) != 1 {
 		t.Fatalf("expected one request, got %d", len(fake.requests))
 	}
-	if fake.requests[0].Target.Model != "template-model" || fake.requests[0].APIKey != "template-key" {
+	if fake.requests[0].Target.BackendID != promptkit.BackendOpenRouter ||
+		fake.requests[0].Target.Model != "template-model" ||
+		fake.requests[0].APIKey != "template-key" {
 		t.Fatalf("unexpected generated request: %+v", fake.requests[0])
 	}
 	if !reflect.DeepEqual(fake.requests[0].Target.ExtraParams, map[string]any{"provider": "template"}) {
@@ -1699,6 +1863,16 @@ func TestEngineRunLayersTransportAndGenerationTimeouts(t *testing.T) {
 			configTimeout:          5 * time.Second,
 			wantRemainingAtRequest: 5 * time.Second,
 		},
+		{
+			name:                   "zero configuration uses ten minute transport default",
+			wantRemainingAtRequest: 10 * time.Minute,
+		},
+		{
+			name:                   "negative configuration uses ten minute transport default",
+			configTimeout:          -2 * time.Second,
+			suppliedClientTimeout:  -3 * time.Second,
+			wantRemainingAtRequest: 10 * time.Minute,
+		},
 		{
 			name:                   "profile deadline is shorter than transport cap",
 			suppliedClientTimeout:  6 * time.Second,
@@ -2240,6 +2414,11 @@ func TestExtraParamsTypedNestedValuesAreCopiedAcrossPublicBoundary(t *testing.T)
 }
 
 func TestRunRejectsInvalidExtraParams(t *testing.T) {
+	cyclicMap := map[string]any{}
+	cyclicMap["self"] = cyclicMap
+	cyclicSlice := []any{nil}
+	cyclicSlice[0] = cyclicSlice
+
 	tests := []struct {
 		name        string
 		extraParams map[string]any
@@ -2251,42 +2430,10 @@ func TestRunRejectsInvalidExtraParams(t *testing.T) {
 		{name: "nan", extraParams: map[string]any{"bad": math.NaN()}},
 		{name: "positive infinity", extraParams: map[string]any{"bad": math.Inf(1)}},
 		{name: "negative infinity", extraParams: map[string]any{"bad": math.Inf(-1)}},
-	}
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			fake := &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}}
-			engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(fake))
-
-			_, err := engine.Run(context.Background(), promptkit.RunRequest{
-				PromptID: frameworkMarkdownSummaryPromptID,
-				Inputs: map[string]promptkit.ArtifactRef{
-					"transcript": promptkit.Inline("Rin opens the gate."),
-					"glossary":   promptkit.Inline("gate: A guarded passage."),
-				},
-				Execution: &promptkit.ExecutionTargetOverride{ExtraParams: tc.extraParams},
-			})
-			if !errors.Is(err, promptkit.ErrInvalidRequest) {
-				t.Fatalf("expected ErrInvalidRequest, got %v", err)
-			}
-			if len(fake.requests) != 0 {
-				t.Fatalf("expected invalid request to fail before LLM call, got %d requests", len(fake.requests))
-			}
-		})
-	}
-}
-
-func TestRunRejectsCyclicExtraParams(t *testing.T) {
-	cyclicMap := map[string]any{}
-	cyclicMap["self"] = cyclicMap
-	cyclicSlice := []any{nil}
-	cyclicSlice[0] = cyclicSlice
-
-	tests := []struct {
-		name        string
-		extraParams map[string]any
-	}{
-		{name: "map", extraParams: cyclicMap},
-		{name: "slice", extraParams: map[string]any{"cycle": cyclicSlice}},
+		{name: "cyclic map", extraParams: cyclicMap},
+		{name: "cyclic slice", extraParams: map[string]any{"cycle": cyclicSlice}},
+		{name: "malformed JSON number", extraParams: map[string]any{"value": json.Number("+1")}},
+		{name: "empty nested key", extraParams: map[string]any{"nested": map[string]any{"": true}}},
 	}
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {

errors.go

@@ -4,6 +4,7 @@ import (
 	"errors"
 	"fmt"
 
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
 	"gitea.maximumdirect.net/eric/promptkit/internal/promptdef"
 	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
@@ -38,6 +39,8 @@ func publicErrorFor(err error) error {
 		return ErrProfileLoad
 	case errors.Is(err, usecase.ErrAPIKeyEnvMissing):
 		return errors.Join(ErrInvalidRequest, ErrAPIKeyEnvMissing)
+	case errors.Is(err, capacity.ErrCapacityExceeded):
+		return ErrCapacityExceeded
 	case errors.Is(err, usecase.ErrArtifactLoad):
 		return ErrArtifactLoad
 	case errors.Is(err, usecase.ErrPromptRender):

errors_internal_test.go

@@ -0,0 +1,22 @@
+package promptkit
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
+)
+
+func TestMapPublicErrorPreservesGenerationCancellation(t *testing.T) {
+	internalErr := fmt.Errorf("%w: %w", usecase.ErrLLMGenerate, context.Canceled)
+
+	err := mapPublicError(internalErr)
+	if !errors.Is(err, ErrLLMGenerate) {
+		t.Fatalf("mapped error=%v, want ErrLLMGenerate", err)
+	}
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("mapped error=%v, want context.Canceled", err)
+	}
+}

examples/go-library/run/main.go

@@ -0,0 +1,81 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+type deterministicClient struct{}
+
+func (deterministicClient) Generate(
+	ctx context.Context,
+	_ promptkit.GenerateRequest,
+) (*promptkit.GenerateResponse, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+
+	return &promptkit.GenerateResponse{
+		Content: "Ada finished the migration review.",
+		Usage: promptkit.TokenUsage{
+			PromptTokens:     12,
+			CompletionTokens: 6,
+			TotalTokens:      18,
+		},
+	}, nil
+}
+
+type summary struct {
+	Output           string                     `json:"output"`
+	ValidationStatus promptkit.ValidationStatus `json:"validation_status"`
+	IsValid          bool                       `json:"is_valid"`
+	Model            string                     `json:"model"`
+	TotalTokens      int                        `json:"total_tokens"`
+}
+
+func main() {
+	engine, err := promptkit.NewEngine(
+		promptkit.Config{},
+		promptkit.WithPromptFile("examples/go-library/run/prompt.yaml"),
+		promptkit.WithProfiles(promptkit.Profile{
+			ID:       "offline-example",
+			Endpoint: "https://example.invalid/v1",
+			Model:    "offline-model",
+		}),
+		promptkit.WithLLMClient(deterministicClient{}),
+	)
+	if err != nil {
+		exit(err)
+	}
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID: "example.run",
+		Inputs: map[string]promptkit.ArtifactRef{
+			"note": promptkit.Inline("Ada finished the migration review."),
+		},
+	})
+	if err != nil {
+		exit(err)
+	}
+
+	encoder := json.NewEncoder(os.Stdout)
+	encoder.SetIndent("", "  ")
+	if err := encoder.Encode(summary{
+		Output:           result.RawOutput,
+		ValidationStatus: result.Validation.Status,
+		IsValid:          result.Validation.IsValid,
+		Model:            result.ModelName,
+		TotalTokens:      result.Usage.TotalTokens,
+	}); err != nil {
+		exit(err)
+	}
+}
+
+func exit(err error) {
+	fmt.Fprintln(os.Stderr, err)
+	os.Exit(1)
+}

examples/go-library/run/prompt.yaml

@@ -0,0 +1,16 @@
+id: example.run
+version: "1.0.0"
+default_profile: offline-example
+description: Run a prompt with a deterministic injected model client.
+inputs:
+  - name: note
+    required: true
+    content_type: text/plain
+messages:
+  - role: system
+    content: Summarize the note in one sentence.
+  - role: user
+    content: '{{input "note"}}'
+output:
+  format: text
+  validation_mode: basic

formatting.go

@@ -2,19 +2,23 @@ package promptkit
 
 import "fmt"
 
-// String returns a concise request summary without exposing direct API keys.
+// String returns a concise request summary without exposing the direct API key
+// or input and variable contents. Reflection-based formatting does not carry
+// this guarantee.
 func (r RunRequest) String() string {
 	return r.redactedString()
 }
 
-// GoString returns a concise request summary without exposing direct API keys.
+// GoString returns a concise request summary without exposing the direct API
+// key or input and variable contents. Reflection-based formatting does not
+// carry this guarantee.
 func (r RunRequest) GoString() string {
 	return r.redactedString()
 }
 
 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 Metadata:%d}",
+		"promptkit.RunRequest{PromptID:%q PromptVersion:%q ProfileID:%q APIKeySet:%t Inputs:%d Vars:%d ExecutionSet:%t ValidationSet:%t}",
 		r.PromptID,
 		r.PromptVersion,
 		r.ProfileID,
@@ -23,18 +27,19 @@ func (r RunRequest) redactedString() string {
 		len(r.Vars),
 		r.Execution != nil,
 		r.Validation != nil,
-		len(r.Metadata),
 	)
 }
 
 // String returns a concise request summary without exposing direct API keys or
-// rendered prompt content.
+// rendered prompt content. Reflection-based formatting does not carry this
+// guarantee.
 func (r GenerateRequest) String() string {
 	return r.redactedString()
 }
 
 // GoString returns a concise request summary without exposing direct API keys or
-// rendered prompt content.
+// rendered prompt content. Reflection-based formatting does not carry this
+// guarantee.
 func (r GenerateRequest) GoString() string {
 	return r.redactedString()
 }

internal/backend/registry.go

@@ -0,0 +1,212 @@
+// Package backend owns validated, immutable OpenAI-compatible backend
+// definitions.
+package backend
+
+import (
+	"errors"
+	"fmt"
+	"net/url"
+	"regexp"
+	"sort"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+)
+
+const (
+	// OpenRouterID is the reserved ID of Promptkit's built-in OpenRouter
+	// backend.
+	OpenRouterID = "openrouter"
+
+	openRouterEndpoint  = "https://openrouter.ai/api/v1"
+	openRouterAPIKeyEnv = "OPENROUTER_API_KEY"
+
+	openRouterConcurrencyLimit = 16
+	defaultQueueCapacity       = 1024
+)
+
+// ErrBackendNotFound identifies a registry lookup for an unknown backend ID.
+var ErrBackendNotFound = errors.New("backend not found")
+
+var environmentVariableName = regexp.MustCompile(`^[A-Za-z_][A-Za-z0-9_]*$`)
+
+// Registry is an immutable collection of validated backend definitions.
+type Registry struct {
+	backends map[string]domain.Backend
+}
+
+// NewRegistry constructs a registry containing the built-in OpenRouter
+// definition followed by the supplied additions. Every ID must be unique.
+func NewRegistry(additions []domain.Backend) (*Registry, error) {
+	registry := &Registry{
+		backends: make(map[string]domain.Backend, len(additions)+1),
+	}
+
+	definitions := make([]domain.Backend, 0, len(additions)+1)
+	definitions = append(definitions, domain.Backend{
+		ID:               OpenRouterID,
+		Endpoint:         openRouterEndpoint,
+		APIKeyEnv:        openRouterAPIKeyEnv,
+		ConcurrencyLimit: openRouterConcurrencyLimit,
+	})
+	definitions = append(definitions, additions...)
+
+	for _, definition := range definitions {
+		definition.ID = strings.TrimSpace(definition.ID)
+		if definition.ID == "" {
+			return nil, errors.New("backend ID must not be blank")
+		}
+		if _, exists := registry.backends[definition.ID]; exists {
+			return nil, fmt.Errorf("backend ID %q is already registered", definition.ID)
+		}
+
+		normalized, err := normalizeBackend(definition)
+		if err != nil {
+			return nil, err
+		}
+		registry.backends[normalized.ID] = normalized
+	}
+
+	return registry, nil
+}
+
+// GetBackend returns a defensive copy of the backend registered with id.
+func (r *Registry) GetBackend(id string) (domain.Backend, error) {
+	if r == nil {
+		return domain.Backend{}, fmt.Errorf("%w: %q", ErrBackendNotFound, id)
+	}
+	definition, ok := r.backends[id]
+	if !ok {
+		return domain.Backend{}, fmt.Errorf("%w: %q", ErrBackendNotFound, id)
+	}
+	extraParams, err := jsonvalue.CopyMap(definition.ExtraParams)
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("copy backend %q: %w", id, err)
+	}
+	definition.ExtraParams = extraParams
+	return definition, nil
+}
+
+// CapacityPolicies returns a copy of the normalized policies for limited
+// backends.
+func (r *Registry) CapacityPolicies() map[string]domain.BackendCapacityPolicy {
+	policies := make(map[string]domain.BackendCapacityPolicy)
+	if r == nil {
+		return policies
+	}
+	for id, definition := range r.backends {
+		if definition.ConcurrencyLimit == 0 {
+			continue
+		}
+		policies[id] = domain.BackendCapacityPolicy{
+			ConcurrencyLimit: definition.ConcurrencyLimit,
+			QueueCapacity:    definition.QueueCapacity,
+		}
+	}
+	return policies
+}
+
+func normalizeBackend(definition domain.Backend) (domain.Backend, error) {
+	definition.Endpoint = strings.TrimSpace(definition.Endpoint)
+	if err := validateEndpoint(definition.Endpoint); err != nil {
+		return domain.Backend{}, fmt.Errorf("backend %q endpoint: %w", definition.ID, err)
+	}
+
+	definition.APIKeyEnv = strings.TrimSpace(definition.APIKeyEnv)
+	if definition.APIKeyEnv != "" && !environmentVariableName.MatchString(definition.APIKeyEnv) {
+		return domain.Backend{}, fmt.Errorf(
+			"backend %q api key environment variable %q is invalid",
+			definition.ID,
+			definition.APIKeyEnv,
+		)
+	}
+
+	if definition.ConcurrencyLimit < 0 {
+		return domain.Backend{}, fmt.Errorf(
+			"backend %q concurrency limit must not be negative",
+			definition.ID,
+		)
+	}
+	if definition.QueueCapacity < 0 {
+		return domain.Backend{}, fmt.Errorf(
+			"backend %q queue capacity must not be negative",
+			definition.ID,
+		)
+	}
+	if definition.ConcurrencyLimit == 0 {
+		if definition.QueueCapacitySet {
+			return domain.Backend{}, fmt.Errorf(
+				"backend %q queue capacity requires a positive concurrency limit",
+				definition.ID,
+			)
+		}
+		definition.QueueCapacity = 0
+	} else {
+		if !definition.QueueCapacitySet {
+			definition.QueueCapacity = defaultQueueCapacity
+			definition.QueueCapacitySet = true
+		}
+		maxInt := int(^uint(0) >> 1)
+		if definition.QueueCapacity > maxInt-definition.ConcurrencyLimit {
+			return domain.Backend{}, fmt.Errorf(
+				"backend %q total capacity overflows int",
+				definition.ID,
+			)
+		}
+	}
+
+	keys := make([]string, 0, len(definition.ExtraParams))
+	for key := range definition.ExtraParams {
+		keys = append(keys, key)
+	}
+	sort.Strings(keys)
+	for _, key := range keys {
+		if key == "" {
+			return domain.Backend{}, fmt.Errorf("backend %q extra parameter key must not be empty", definition.ID)
+		}
+		if llm.IsReservedOpenAIChatRequestField(key) {
+			return domain.Backend{}, fmt.Errorf(
+				"backend %q extra parameter %q collides with a reserved request field",
+				definition.ID,
+				key,
+			)
+		}
+	}
+
+	extraParams, err := jsonvalue.CopyMap(definition.ExtraParams)
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("backend %q extra parameters: %w", definition.ID, err)
+	}
+	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

@@ -0,0 +1,368 @@
+package backend_test
+
+import (
+	"errors"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+const validEndpoint = "https://backend.example/v1"
+
+func TestRegistryIncludesExactOpenRouterDefinition(t *testing.T) {
+	registry, err := backend.NewRegistry(nil)
+	if err != nil {
+		t.Fatalf("construct registry: %v", err)
+	}
+
+	definition, err := registry.GetBackend(backend.OpenRouterID)
+	if err != nil {
+		t.Fatalf("look up OpenRouter: %v", err)
+	}
+	if definition.ID != "openrouter" ||
+		definition.Endpoint != "https://openrouter.ai/api/v1" ||
+		definition.APIKeyEnv != "OPENROUTER_API_KEY" ||
+		definition.ConcurrencyLimit != 16 ||
+		definition.QueueCapacity != 1024 ||
+		!definition.QueueCapacitySet ||
+		definition.ExtraParams != nil {
+		t.Fatalf("unexpected OpenRouter definition: %#v", definition)
+	}
+	policies := registry.CapacityPolicies()
+	if len(policies) != 1 ||
+		policies["openrouter"] != (domain.BackendCapacityPolicy{
+			ConcurrencyLimit: 16,
+			QueueCapacity:    1024,
+		}) {
+		t.Fatalf("unexpected OpenRouter capacity policies: %#v", policies)
+	}
+}
+
+func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
+	nested := map[string]int{"limit": 2}
+	extraParams := map[string]any{
+		"count":  int64(7),
+		"nested": nested,
+	}
+	registry, err := backend.NewRegistry([]domain.Backend{
+		{
+			ID:               " custom ",
+			Endpoint:         " https://custom.example/openai/v1 ",
+			APIKeyEnv:        " CUSTOM_API_KEY ",
+			ExtraParams:      extraParams,
+			ConcurrencyLimit: 3,
+			QueueCapacity:    2,
+			QueueCapacitySet: true,
+		},
+		{
+			ID:       "Custom",
+			Endpoint: validEndpoint,
+		},
+	})
+	if err != nil {
+		t.Fatalf("construct registry: %v", err)
+	}
+
+	nested["limit"] = 99
+	extraParams["added"] = true
+
+	got, err := registry.GetBackend("custom")
+	if err != nil {
+		t.Fatalf("look up custom backend: %v", err)
+	}
+	if got.ID != "custom" ||
+		got.Endpoint != "https://custom.example/openai/v1" ||
+		got.APIKeyEnv != "CUSTOM_API_KEY" ||
+		got.ConcurrencyLimit != 3 ||
+		got.QueueCapacity != 2 ||
+		!got.QueueCapacitySet {
+		t.Fatalf("unexpected normalized definition: %#v", got)
+	}
+	if count, ok := got.ExtraParams["count"].(int64); !ok || count != 7 {
+		t.Fatalf("integer type or value changed: %#v", got.ExtraParams["count"])
+	}
+	gotNested, ok := got.ExtraParams["nested"].(map[string]int)
+	if !ok || gotNested["limit"] != 2 {
+		t.Fatalf("container type or value changed: %#v", got.ExtraParams["nested"])
+	}
+	if _, exists := got.ExtraParams["added"]; exists {
+		t.Fatalf("registry retained caller map: %#v", got.ExtraParams)
+	}
+
+	gotNested["limit"] = 100
+	got.ExtraParams["added"] = true
+	again, err := registry.GetBackend("custom")
+	if err != nil {
+		t.Fatalf("look up custom backend again: %v", err)
+	}
+	if again.ExtraParams["nested"].(map[string]int)["limit"] != 2 {
+		t.Fatalf("lookup exposed registry nested map: %#v", again.ExtraParams)
+	}
+	if _, exists := again.ExtraParams["added"]; exists {
+		t.Fatalf("lookup exposed registry map: %#v", again.ExtraParams)
+	}
+
+	if _, err := registry.GetBackend("Custom"); err != nil {
+		t.Fatalf("backend IDs should be case-sensitive: %v", err)
+	}
+
+	policies := registry.CapacityPolicies()
+	if len(policies) != 2 {
+		t.Fatalf("unexpected capacity policy count: %#v", policies)
+	}
+	policies["custom"] = domain.BackendCapacityPolicy{}
+	delete(policies, backend.OpenRouterID)
+	againPolicies := registry.CapacityPolicies()
+	if againPolicies["custom"] != (domain.BackendCapacityPolicy{
+		ConcurrencyLimit: 3,
+		QueueCapacity:    2,
+	}) {
+		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) {
+	maxInt := int(^uint(0) >> 1)
+	tests := []struct {
+		name       string
+		definition domain.Backend
+		want       domain.BackendCapacityPolicy
+		wantSet    bool
+		wantError  bool
+	}{
+		{
+			name:       "unlimited when omitted",
+			definition: domain.Backend{},
+		},
+		{
+			name: "default queue",
+			definition: domain.Backend{
+				ConcurrencyLimit: 2,
+			},
+			want: domain.BackendCapacityPolicy{
+				ConcurrencyLimit: 2,
+				QueueCapacity:    1024,
+			},
+			wantSet: true,
+		},
+		{
+			name: "explicit zero queue",
+			definition: domain.Backend{
+				ConcurrencyLimit: 2,
+				QueueCapacitySet: true,
+			},
+			want: domain.BackendCapacityPolicy{
+				ConcurrencyLimit: 2,
+			},
+			wantSet: true,
+		},
+		{
+			name: "negative concurrency limit",
+			definition: domain.Backend{
+				ConcurrencyLimit: -1,
+			},
+			wantError: true,
+		},
+		{
+			name: "negative queue capacity",
+			definition: domain.Backend{
+				ConcurrencyLimit: 1,
+				QueueCapacity:    -1,
+				QueueCapacitySet: true,
+			},
+			wantError: true,
+		},
+		{
+			name: "queue without limit",
+			definition: domain.Backend{
+				QueueCapacitySet: true,
+			},
+			wantError: true,
+		},
+		{
+			name: "total overflow",
+			definition: domain.Backend{
+				ConcurrencyLimit: maxInt,
+				QueueCapacity:    1,
+				QueueCapacitySet: true,
+			},
+			wantError: true,
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			tc.definition.ID = "custom"
+			tc.definition.Endpoint = validEndpoint
+			registry, err := backend.NewRegistry([]domain.Backend{tc.definition})
+			if tc.wantError {
+				if err == nil {
+					t.Fatal("expected invalid capacity policy error")
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("construct registry: %v", err)
+			}
+
+			definition, err := registry.GetBackend("custom")
+			if err != nil {
+				t.Fatalf("look up custom backend: %v", err)
+			}
+			if definition.ConcurrencyLimit != tc.want.ConcurrencyLimit ||
+				definition.QueueCapacity != tc.want.QueueCapacity ||
+				definition.QueueCapacitySet != tc.wantSet {
+				t.Fatalf("normalized capacity=(%d, %d, %t), want (%d, %d, %t)",
+					definition.ConcurrencyLimit,
+					definition.QueueCapacity,
+					definition.QueueCapacitySet,
+					tc.want.ConcurrencyLimit,
+					tc.want.QueueCapacity,
+					tc.wantSet,
+				)
+			}
+			policies := registry.CapacityPolicies()
+			got, ok := policies["custom"]
+			if ok != tc.wantSet || got != tc.want {
+				t.Fatalf("capacity policy=(%#v, %t), want (%#v, %t)", got, ok, tc.want, tc.wantSet)
+			}
+		})
+	}
+}
+
+func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
+	tests := []struct {
+		name      string
+		additions []domain.Backend
+		wantID    string
+	}{
+		{
+			name: "built-in collision after normalization",
+			additions: []domain.Backend{{
+				ID: " openrouter ",
+			}},
+			wantID: "openrouter",
+		},
+		{
+			name: "consumer collision after normalization",
+			additions: []domain.Backend{
+				{ID: "custom", Endpoint: validEndpoint},
+				{ID: " custom ", Endpoint: "https://other.example/v1"},
+			},
+			wantID: "custom",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := backend.NewRegistry(tc.additions)
+			if err == nil {
+				t.Fatal("expected duplicate ID error")
+			}
+			if !strings.Contains(err.Error(), tc.wantID) {
+				t.Fatalf("expected error to identify %q, got %v", tc.wantID, err)
+			}
+		})
+	}
+}
+
+func TestNewRegistryValidatesIDs(t *testing.T) {
+	for _, id := range []string{"", " \t\n "} {
+		t.Run(id, func(t *testing.T) {
+			_, err := backend.NewRegistry([]domain.Backend{{
+				ID:       id,
+				Endpoint: validEndpoint,
+			}})
+			if err == nil {
+				t.Fatal("expected blank ID error")
+			}
+		})
+	}
+}
+
+func TestNewRegistryValidatesEndpoints(t *testing.T) {
+	tests := []struct {
+		name     string
+		endpoint string
+	}{
+		{name: "blank", endpoint: ""},
+		{name: "relative", endpoint: "/v1"},
+		{name: "missing host", endpoint: "https:///v1"},
+		{name: "unsupported scheme", endpoint: "ftp://backend.example/v1"},
+		{name: "user information", endpoint: "https://user@backend.example/v1"},
+		{name: "query", endpoint: "https://backend.example/v1?mode=chat"},
+		{name: "empty query", endpoint: "https://backend.example/v1?"},
+		{name: "fragment", endpoint: "https://backend.example/v1#chat"},
+		{name: "empty fragment", endpoint: "https://backend.example/v1#"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := backend.NewRegistry([]domain.Backend{{
+				ID:       "custom",
+				Endpoint: tc.endpoint,
+			}})
+			if err == nil {
+				t.Fatal("expected invalid endpoint error")
+			}
+		})
+	}
+}
+
+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{{
+				ID:        "custom",
+				Endpoint:  validEndpoint,
+				APIKeyEnv: name,
+			}})
+			if err == nil {
+				t.Fatal("expected invalid environment-variable name error")
+			}
+		})
+	}
+}
+
+func TestNewRegistryRejectsInvalidAndReservedExtraParameters(t *testing.T) {
+	tests := []struct {
+		name        string
+		extraParams map[string]any
+	}{
+		{name: "unsupported value", extraParams: map[string]any{"value": make(chan int)}},
+		{name: "reserved key", extraParams: map[string]any{"model": "override"}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := backend.NewRegistry([]domain.Backend{{
+				ID:          "custom",
+				Endpoint:    validEndpoint,
+				ExtraParams: tc.extraParams,
+			}})
+			if err == nil {
+				t.Fatal("expected invalid extra parameters error")
+			}
+		})
+	}
+}
+
+func TestRegistryLookupReportsNotFound(t *testing.T) {
+	registry, err := backend.NewRegistry(nil)
+	if err != nil {
+		t.Fatalf("construct registry: %v", err)
+	}
+
+	_, err = registry.GetBackend("missing")
+	if !errors.Is(err, backend.ErrBackendNotFound) {
+		t.Fatalf("expected ErrBackendNotFound, got %v", err)
+	}
+	if !strings.Contains(err.Error(), "missing") {
+		t.Fatalf("expected error to identify backend, got %v", err)
+	}
+}

internal/capacity/client.go

@@ -0,0 +1,40 @@
+package capacity
+
+import (
+	"context"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+)
+
+type client struct {
+	manager *Manager
+	next    llm.Client
+}
+
+// NewClient wraps next with configured active-generation limits. A nil manager
+// leaves next unchanged.
+func NewClient(manager *Manager, next llm.Client) llm.Client {
+	if manager == nil {
+		return next
+	}
+	return &client{
+		manager: manager,
+		next:    next,
+	}
+}
+
+func (c *client) Generate(
+	ctx context.Context,
+	req domain.GenerateRequest,
+) (*domain.GenerateResponse, error) {
+	pool := c.manager.getPool(req.Target.BackendID)
+	if pool == nil {
+		return c.next.Generate(ctx, req)
+	}
+	if err := pool.acquire(ctx); err != nil {
+		return nil, err
+	}
+	defer pool.releaseActive()
+	return c.next.Generate(ctx, req)
+}

internal/capacity/client_test.go

@@ -0,0 +1,517 @@
+package capacity
+
+import (
+	"context"
+	"errors"
+	"reflect"
+	"runtime"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+)
+
+type generateResult struct {
+	response *domain.GenerateResponse
+	err      error
+}
+
+type clientFunc func(
+	context.Context,
+	domain.GenerateRequest,
+) (*domain.GenerateResponse, error)
+
+func (f clientFunc) Generate(
+	ctx context.Context,
+	req domain.GenerateRequest,
+) (*domain.GenerateResponse, error) {
+	return f(ctx, req)
+}
+
+type blockingClient struct {
+	mu       sync.Mutex
+	active   int
+	peak     int
+	calls    map[string]int
+	started  chan string
+	releases map[string]chan struct{}
+}
+
+func newBlockingClient(releases map[string]chan struct{}) *blockingClient {
+	return &blockingClient{
+		calls:    make(map[string]int),
+		started:  make(chan string, 64),
+		releases: releases,
+	}
+}
+
+func (c *blockingClient) Generate(
+	ctx context.Context,
+	req domain.GenerateRequest,
+) (*domain.GenerateResponse, error) {
+	id := req.Prompt.SessionID
+	c.mu.Lock()
+	c.active++
+	if c.active > c.peak {
+		c.peak = c.active
+	}
+	c.calls[id]++
+	c.mu.Unlock()
+	defer func() {
+		c.mu.Lock()
+		c.active--
+		c.mu.Unlock()
+	}()
+
+	c.started <- id
+	if release := c.releases[id]; release != nil {
+		select {
+		case <-release:
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		}
+	}
+	return &domain.GenerateResponse{Content: id}, nil
+}
+
+func (c *blockingClient) callCount(id string) int {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.calls[id]
+}
+
+func (c *blockingClient) peakConcurrency() int {
+	c.mu.Lock()
+	defer c.mu.Unlock()
+	return c.peak
+}
+
+func generateAsync(
+	client llm.Client,
+	ctx context.Context,
+	backendID string,
+	id string,
+) <-chan generateResult {
+	result := make(chan generateResult, 1)
+	go func() {
+		response, err := client.Generate(ctx, domain.GenerateRequest{
+			Prompt: domain.RenderedPrompt{SessionID: id},
+			Target: domain.ExecutionTarget{BackendID: backendID},
+		})
+		result <- generateResult{response: response, err: err}
+	}()
+	return result
+}
+
+func waitForWaiterCount(t *testing.T, manager *Manager, backendID string, want int) {
+	t.Helper()
+	pool := manager.pools[backendID]
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		pool.mu.Lock()
+		got := pool.waiters.Len()
+		pool.mu.Unlock()
+		if got == want {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatalf("waiter count=%d, want %d", got, want)
+		}
+		runtime.Gosched()
+	}
+}
+
+func receiveStarted(t *testing.T, started <-chan string) string {
+	t.Helper()
+	select {
+	case id := <-started:
+		return id
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for wrapped client invocation")
+		return ""
+	}
+}
+
+func receiveResult(t *testing.T, result <-chan generateResult) generateResult {
+	t.Helper()
+	select {
+	case got := <-result:
+		return got
+	case <-time.After(2 * time.Second):
+		t.Fatal("timed out waiting for generation result")
+		return generateResult{}
+	}
+}
+
+func newTestManager(t *testing.T, policies map[string]domain.BackendCapacityPolicy) *Manager {
+	t.Helper()
+	manager, err := NewManager(policies)
+	if err != nil {
+		t.Fatalf("construct manager: %v", err)
+	}
+	return manager
+}
+
+func TestClientLimitsPeakConcurrencyAndServesWaitersFIFO(t *testing.T) {
+	manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1},
+	})
+	firstRelease := make(chan struct{})
+	secondRelease := make(chan struct{})
+	thirdRelease := make(chan struct{})
+	next := newBlockingClient(map[string]chan struct{}{
+		"first":  firstRelease,
+		"second": secondRelease,
+		"third":  thirdRelease,
+	})
+	client := NewClient(manager, next)
+
+	first := generateAsync(client, context.Background(), "limited", "first")
+	if got := receiveStarted(t, next.started); got != "first" {
+		t.Fatalf("first invocation=%q, want first", got)
+	}
+	second := generateAsync(client, context.Background(), "limited", "second")
+	waitForWaiterCount(t, manager, "limited", 1)
+	third := generateAsync(client, context.Background(), "limited", "third")
+	waitForWaiterCount(t, manager, "limited", 2)
+
+	close(firstRelease)
+	if got := receiveResult(t, first); got.err != nil {
+		t.Fatalf("first generation: %v", got.err)
+	}
+	if got := receiveStarted(t, next.started); got != "second" {
+		t.Fatalf("second invocation=%q, want second", got)
+	}
+	close(secondRelease)
+	if got := receiveResult(t, second); got.err != nil {
+		t.Fatalf("second generation: %v", got.err)
+	}
+	if got := receiveStarted(t, next.started); got != "third" {
+		t.Fatalf("third invocation=%q, want third", got)
+	}
+	close(thirdRelease)
+	if got := receiveResult(t, third); got.err != nil {
+		t.Fatalf("third generation: %v", got.err)
+	}
+	if peak := next.peakConcurrency(); peak != 1 {
+		t.Fatalf("peak concurrency=%d, want 1", peak)
+	}
+}
+
+func TestClientPeakConcurrencyDoesNotExceedConfiguredLimit(t *testing.T) {
+	const limit = 2
+	manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: limit},
+	})
+	gate := make(chan struct{})
+	releases := make(map[string]chan struct{})
+	for i := range 5 {
+		releases[string(rune('a'+i))] = gate
+	}
+	next := newBlockingClient(releases)
+	client := NewClient(manager, next)
+
+	results := make([]<-chan generateResult, 0, len(releases))
+	for id := range releases {
+		results = append(results, generateAsync(client, context.Background(), "limited", id))
+	}
+	for range limit {
+		receiveStarted(t, next.started)
+	}
+	waitForWaiterCount(t, manager, "limited", len(releases)-limit)
+
+	close(gate)
+	for _, result := range results {
+		if got := receiveResult(t, result); got.err != nil {
+			t.Fatalf("generation: %v", got.err)
+		}
+	}
+	if peak := next.peakConcurrency(); peak != limit {
+		t.Fatalf("peak concurrency=%d, want %d", peak, limit)
+	}
+}
+
+func TestClientRemovesCanceledWaiters(t *testing.T) {
+	tests := []struct {
+		name      string
+		cancelID  string
+		wantOrder []string
+	}{
+		{name: "first waiter", cancelID: "one", wantOrder: []string{"two", "three"}},
+		{name: "middle waiter", cancelID: "two", wantOrder: []string{"one", "three"}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+				"limited": {ConcurrencyLimit: 1},
+			})
+			holderRelease := make(chan struct{})
+			releases := map[string]chan struct{}{
+				"holder": holderRelease,
+				"one":    make(chan struct{}),
+				"two":    make(chan struct{}),
+				"three":  make(chan struct{}),
+			}
+			next := newBlockingClient(releases)
+			client := NewClient(manager, next)
+
+			holder := generateAsync(client, context.Background(), "limited", "holder")
+			if got := receiveStarted(t, next.started); got != "holder" {
+				t.Fatalf("initial invocation=%q, want holder", got)
+			}
+
+			contexts := make(map[string]context.Context)
+			cancels := make(map[string]context.CancelFunc)
+			results := make(map[string]<-chan generateResult)
+			for _, id := range []string{"one", "two", "three"} {
+				contexts[id], cancels[id] = context.WithCancel(context.Background())
+				results[id] = generateAsync(client, contexts[id], "limited", id)
+				waitForWaiterCount(t, manager, "limited", len(results))
+			}
+
+			cancels[tc.cancelID]()
+			if got := receiveResult(t, results[tc.cancelID]); !errors.Is(got.err, context.Canceled) {
+				t.Fatalf("canceled waiter error=%v, want context.Canceled", got.err)
+			}
+			waitForWaiterCount(t, manager, "limited", 2)
+
+			close(holderRelease)
+			if got := receiveResult(t, holder); got.err != nil {
+				t.Fatalf("holder generation: %v", got.err)
+			}
+			for _, id := range tc.wantOrder {
+				if got := receiveStarted(t, next.started); got != id {
+					t.Fatalf("next invocation=%q, want %q", got, id)
+				}
+				close(releases[id])
+				if got := receiveResult(t, results[id]); got.err != nil {
+					t.Fatalf("%s generation: %v", id, got.err)
+				}
+			}
+			if calls := next.callCount(tc.cancelID); calls != 0 {
+				t.Fatalf("canceled waiter invoked wrapped client %d times", calls)
+			}
+			for _, cancel := range cancels {
+				cancel()
+			}
+		})
+	}
+}
+
+func TestClientGrantCancellationRaceDoesNotLeakPermit(t *testing.T) {
+	const iterations = 200
+	for i := range iterations {
+		manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+			"limited": {ConcurrencyLimit: 1},
+		})
+		holderRelease := make(chan struct{})
+		var waiterCalls atomic.Int64
+		next := clientFunc(func(
+			_ context.Context,
+			req domain.GenerateRequest,
+		) (*domain.GenerateResponse, error) {
+			if req.Prompt.SessionID == "holder" {
+				<-holderRelease
+			} else if req.Prompt.SessionID == "waiter" {
+				waiterCalls.Add(1)
+			}
+			return &domain.GenerateResponse{Content: req.Prompt.SessionID}, nil
+		})
+		client := NewClient(manager, next)
+
+		holder := generateAsync(client, context.Background(), "limited", "holder")
+		waitForActiveCount(t, manager, "limited", 1)
+		ctx, cancel := context.WithCancel(context.Background())
+		waiterResult := generateAsync(client, ctx, "limited", "waiter")
+		waitForWaiterCount(t, manager, "limited", 1)
+
+		start := make(chan struct{})
+		var race sync.WaitGroup
+		race.Add(2)
+		go func() {
+			defer race.Done()
+			<-start
+			cancel()
+		}()
+		go func() {
+			defer race.Done()
+			<-start
+			close(holderRelease)
+		}()
+		close(start)
+		race.Wait()
+
+		if got := receiveResult(t, holder); got.err != nil {
+			t.Fatalf("iteration %d holder generation: %v", i, got.err)
+		}
+		got := receiveResult(t, waiterResult)
+		switch calls := waiterCalls.Load(); {
+		case calls == 0 && errors.Is(got.err, context.Canceled):
+		case calls == 1 && got.err == nil:
+		default:
+			t.Fatalf("iteration %d waiter calls=%d error=%v", i, calls, got.err)
+		}
+
+		probe := generateAsync(client, context.Background(), "limited", "probe")
+		if got := receiveResult(t, probe); got.err != nil {
+			t.Fatalf("iteration %d probe generation: %v", i, got.err)
+		}
+		waitForActiveCount(t, manager, "limited", 0)
+		waitForWaiterCount(t, manager, "limited", 0)
+	}
+}
+
+func waitForActiveCount(t *testing.T, manager *Manager, backendID string, want int) {
+	t.Helper()
+	pool := manager.pools[backendID]
+	deadline := time.Now().Add(2 * time.Second)
+	for {
+		pool.mu.Lock()
+		got := pool.active
+		pool.mu.Unlock()
+		if got == want {
+			return
+		}
+		if time.Now().After(deadline) {
+			t.Fatalf("active count=%d, want %d", got, want)
+		}
+		runtime.Gosched()
+	}
+}
+
+func TestClientUsesIndependentPoolsAndUnlimitedFastPaths(t *testing.T) {
+	manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+		"alpha": {ConcurrencyLimit: 1},
+		"beta":  {ConcurrencyLimit: 1},
+	})
+	alphaRelease := make(chan struct{})
+	betaRelease := make(chan struct{})
+	next := newBlockingClient(map[string]chan struct{}{
+		"alpha": alphaRelease,
+		"beta":  betaRelease,
+	})
+	client := NewClient(manager, next)
+
+	alpha := generateAsync(client, context.Background(), "alpha", "alpha")
+	beta := generateAsync(client, context.Background(), "beta", "beta")
+	started := map[string]bool{
+		receiveStarted(t, next.started): true,
+		receiveStarted(t, next.started): true,
+	}
+	if !started["alpha"] || !started["beta"] {
+		t.Fatalf("independent pools did not both start: %#v", started)
+	}
+	close(alphaRelease)
+	close(betaRelease)
+	if got := receiveResult(t, alpha); got.err != nil {
+		t.Fatalf("alpha generation: %v", got.err)
+	}
+	if got := receiveResult(t, beta); got.err != nil {
+		t.Fatalf("beta generation: %v", got.err)
+	}
+
+	for _, backendID := range []string{"", "unknown"} {
+		response, err := client.Generate(context.Background(), domain.GenerateRequest{
+			Prompt: domain.RenderedPrompt{SessionID: backendID},
+			Target: domain.ExecutionTarget{BackendID: backendID},
+		})
+		if err != nil || response == nil {
+			t.Fatalf("unlimited backend %q response=(%#v, %v)", backendID, response, err)
+		}
+	}
+	if got := NewClient(nil, next); got != next {
+		t.Fatal("nil manager did not return the wrapped client unchanged")
+	}
+}
+
+func TestClientPreservesRequestsResponsesAndErrors(t *testing.T) {
+	manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1},
+	})
+	request := domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{
+			SessionID: "session",
+			Messages: []domain.RenderedMessage{
+				{Role: "user", Content: "content"},
+			},
+		},
+		Target: domain.ExecutionTarget{
+			BackendID:   "limited",
+			Model:       "model",
+			ExtraParams: map[string]any{"key": "value"},
+		},
+	}
+	response := &domain.GenerateResponse{
+		Content: "output",
+		Usage:   domain.TokenUsage{TotalTokens: 7},
+	}
+	collaboratorErr := errors.New("collaborator failure")
+	tests := []struct {
+		name     string
+		response *domain.GenerateResponse
+		err      error
+	}{
+		{name: "successful response", response: response},
+		{name: "nil response"},
+		{name: "collaborator error", response: response, err: collaboratorErr},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			var captured domain.GenerateRequest
+			next := clientFunc(func(
+				_ context.Context,
+				req domain.GenerateRequest,
+			) (*domain.GenerateResponse, error) {
+				captured = req
+				return tc.response, tc.err
+			})
+			gotResponse, gotErr := NewClient(manager, next).Generate(context.Background(), request)
+			if !reflect.DeepEqual(captured, request) {
+				t.Fatalf("request changed: %#v", captured)
+			}
+			if gotResponse != tc.response || gotErr != tc.err {
+				t.Fatalf("response=(%p, %v), want (%p, %v)",
+					gotResponse, gotErr, tc.response, tc.err)
+			}
+		})
+	}
+}
+
+func TestClientReleasesPermitDuringPanicUnwinding(t *testing.T) {
+	manager := newTestManager(t, map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1},
+	})
+	var calls atomic.Int64
+	next := clientFunc(func(
+		_ context.Context,
+		_ domain.GenerateRequest,
+	) (*domain.GenerateResponse, error) {
+		if calls.Add(1) == 1 {
+			panic("test panic")
+		}
+		return &domain.GenerateResponse{Content: "recovered"}, nil
+	})
+	client := NewClient(manager, next)
+	request := domain.GenerateRequest{
+		Target: domain.ExecutionTarget{BackendID: "limited"},
+	}
+
+	func() {
+		defer func() {
+			if recover() == nil {
+				t.Fatal("expected wrapped client panic")
+			}
+		}()
+		_, _ = client.Generate(context.Background(), request)
+	}()
+
+	response, err := client.Generate(context.Background(), request)
+	if err != nil || response == nil || response.Content != "recovered" {
+		t.Fatalf("generation after panic=(%#v, %v)", response, err)
+	}
+}

internal/capacity/manager.go

@@ -0,0 +1,160 @@
+// Package capacity coordinates engine-local run admission and model-generation
+// concurrency for configured backends.
+package capacity
+
+import (
+	"container/list"
+	"context"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// ErrCapacityExceeded identifies an admission rejected because a backend's
+// configured run capacity is full.
+var ErrCapacityExceeded = errors.New("backend capacity exceeded")
+
+// Manager owns independent backend capacity pools with immutable limits.
+type Manager struct {
+	pools map[string]*pool
+}
+
+type pool struct {
+	mu               sync.Mutex
+	concurrencyLimit int
+	totalCapacity    int
+	admitted         int
+	active           int
+	waiters          list.List
+}
+
+type waiter struct {
+	ready   chan struct{}
+	element *list.Element
+	granted bool
+}
+
+// NewManager constructs independent pools from normalized backend policies.
+func NewManager(policies map[string]domain.BackendCapacityPolicy) (*Manager, error) {
+	manager := &Manager{
+		pools: make(map[string]*pool, len(policies)),
+	}
+	maxInt := int(^uint(0) >> 1)
+	for id, policy := range policies {
+		if strings.TrimSpace(id) == "" {
+			return nil, errors.New("backend capacity policy ID must not be blank")
+		}
+		if policy.ConcurrencyLimit <= 0 {
+			return nil, fmt.Errorf(
+				"backend %q concurrency limit must be positive",
+				id,
+			)
+		}
+		if policy.QueueCapacity < 0 {
+			return nil, fmt.Errorf(
+				"backend %q queue capacity must not be negative",
+				id,
+			)
+		}
+		if policy.QueueCapacity > maxInt-policy.ConcurrencyLimit {
+			return nil, fmt.Errorf("backend %q total capacity overflows int", id)
+		}
+		manager.pools[id] = &pool{
+			concurrencyLimit: policy.ConcurrencyLimit,
+			totalCapacity:    policy.ConcurrencyLimit + policy.QueueCapacity,
+		}
+	}
+	return manager, nil
+}
+
+// Admit immediately reserves one configured backend run slot. Backends without
+// a configured pool are unlimited.
+func (m *Manager) Admit(ctx context.Context, backendID string) (func(), error) {
+	pool := m.getPool(backendID)
+	if pool == nil {
+		return releaseNothing, nil
+	}
+
+	pool.mu.Lock()
+	defer pool.mu.Unlock()
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	if pool.admitted >= pool.totalCapacity {
+		return nil, ErrCapacityExceeded
+	}
+	pool.admitted++
+
+	var once sync.Once
+	return func() {
+		once.Do(func() {
+			pool.mu.Lock()
+			pool.admitted--
+			pool.mu.Unlock()
+		})
+	}, nil
+}
+
+func releaseNothing() {}
+
+func (m *Manager) getPool(backendID string) *pool {
+	if m == nil || backendID == "" {
+		return nil
+	}
+	return m.pools[backendID]
+}
+
+func (p *pool) acquire(ctx context.Context) error {
+	p.mu.Lock()
+	if err := ctx.Err(); err != nil {
+		p.mu.Unlock()
+		return err
+	}
+	if p.active < p.concurrencyLimit && p.waiters.Len() == 0 {
+		p.active++
+		p.mu.Unlock()
+		return nil
+	}
+
+	waiter := &waiter{ready: make(chan struct{})}
+	waiter.element = p.waiters.PushBack(waiter)
+	p.mu.Unlock()
+
+	select {
+	case <-waiter.ready:
+		return nil
+	case <-ctx.Done():
+		p.mu.Lock()
+		if !waiter.granted {
+			p.waiters.Remove(waiter.element)
+			waiter.element = nil
+			p.mu.Unlock()
+			return ctx.Err()
+		}
+		p.mu.Unlock()
+		return nil
+	}
+}
+
+func (p *pool) releaseActive() {
+	var ready chan struct{}
+
+	p.mu.Lock()
+	if element := p.waiters.Front(); element != nil {
+		waiter := element.Value.(*waiter)
+		p.waiters.Remove(element)
+		waiter.element = nil
+		waiter.granted = true
+		ready = waiter.ready
+	} else {
+		p.active--
+	}
+	p.mu.Unlock()
+
+	if ready != nil {
+		close(ready)
+	}
+}

internal/capacity/manager_test.go

@@ -0,0 +1,158 @@
+package capacity
+
+import (
+	"context"
+	"errors"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestNewManagerRejectsInvalidPolicies(t *testing.T) {
+	maxInt := int(^uint(0) >> 1)
+	tests := []struct {
+		name   string
+		id     string
+		policy domain.BackendCapacityPolicy
+	}{
+		{
+			name:   "blank ID",
+			id:     " \t ",
+			policy: domain.BackendCapacityPolicy{ConcurrencyLimit: 1},
+		},
+		{
+			name:   "zero concurrency",
+			id:     "backend",
+			policy: domain.BackendCapacityPolicy{},
+		},
+		{
+			name:   "negative concurrency",
+			id:     "backend",
+			policy: domain.BackendCapacityPolicy{ConcurrencyLimit: -1},
+		},
+		{
+			name: "negative queue",
+			id:   "backend",
+			policy: domain.BackendCapacityPolicy{
+				ConcurrencyLimit: 1,
+				QueueCapacity:    -1,
+			},
+		},
+		{
+			name: "total overflow",
+			id:   "backend",
+			policy: domain.BackendCapacityPolicy{
+				ConcurrencyLimit: maxInt,
+				QueueCapacity:    1,
+			},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := NewManager(map[string]domain.BackendCapacityPolicy{
+				tc.id: tc.policy,
+			})
+			if err == nil {
+				t.Fatal("expected invalid policy error")
+			}
+		})
+	}
+}
+
+func TestManagerAdmissionIsBoundedAndReleaseIsIdempotent(t *testing.T) {
+	policies := map[string]domain.BackendCapacityPolicy{
+		"limited": {
+			ConcurrencyLimit: 2,
+			QueueCapacity:    1,
+		},
+		"independent": {
+			ConcurrencyLimit: 1,
+		},
+	}
+	manager, err := NewManager(policies)
+	if err != nil {
+		t.Fatalf("construct manager: %v", err)
+	}
+	policies["limited"] = domain.BackendCapacityPolicy{
+		ConcurrencyLimit: 100,
+		QueueCapacity:    100,
+	}
+
+	releases := make([]func(), 0, 3)
+	for range 3 {
+		release, err := manager.Admit(context.Background(), "limited")
+		if err != nil {
+			t.Fatalf("admit within configured capacity: %v", err)
+		}
+		releases = append(releases, release)
+	}
+	if release, err := manager.Admit(context.Background(), "limited"); release != nil ||
+		!errors.Is(err, ErrCapacityExceeded) {
+		t.Fatalf("admission beyond capacity=(release=%t, err=%v), want ErrCapacityExceeded",
+			release != nil, err)
+	}
+	independentRelease, err := manager.Admit(context.Background(), "independent")
+	if err != nil {
+		t.Fatalf("admit independent backend while first is full: %v", err)
+	}
+	independentRelease()
+
+	releases[0]()
+	releases[0]()
+	replacement, err := manager.Admit(context.Background(), "limited")
+	if err != nil {
+		t.Fatalf("admit after release: %v", err)
+	}
+	replacement()
+	releases[1]()
+	releases[2]()
+
+	pool := manager.pools["limited"]
+	pool.mu.Lock()
+	admitted := pool.admitted
+	pool.mu.Unlock()
+	if admitted != 0 {
+		t.Fatalf("admitted runs after releases=%d, want 0", admitted)
+	}
+}
+
+func TestManagerAdmissionHonorsContextAndUnlimitedBackends(t *testing.T) {
+	manager, err := NewManager(map[string]domain.BackendCapacityPolicy{
+		"limited": {ConcurrencyLimit: 1},
+	})
+	if err != nil {
+		t.Fatalf("construct manager: %v", err)
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if release, err := manager.Admit(ctx, "limited"); release != nil ||
+		!errors.Is(err, context.Canceled) {
+		t.Fatalf("canceled limited admission=(release=%t, err=%v), want context cancellation",
+			release != nil, err)
+	}
+
+	var nilManager *Manager
+	for _, tc := range []struct {
+		name      string
+		manager   *Manager
+		backendID string
+	}{
+		{name: "nil manager", manager: nilManager, backendID: "limited"},
+		{name: "blank ID", manager: manager},
+		{name: "unknown ID", manager: manager, backendID: "unknown"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			release, err := tc.manager.Admit(ctx, tc.backendID)
+			if err != nil {
+				t.Fatalf("unlimited admission: %v", err)
+			}
+			if release == nil {
+				t.Fatal("unlimited admission returned nil release")
+			}
+			release()
+			release()
+		})
+	}
+}

internal/domain/domain.go

@@ -63,12 +63,12 @@ type RunRequest struct {
 	PromptID      string
 	PromptVersion string
 	ProfileID     string
+	SessionID     string
 	APIKey        string `json:"-" yaml:"-"`
 	Inputs        map[string]ArtifactRef
 	Vars          map[string]string
 	Execution     *ExecutionTargetOverride
 	Validation    *OutputContract
-	Metadata      map[string]string
 }
 
 // RunResult represents the complete result of a prompt execution run.
@@ -80,8 +80,10 @@ type RunResult struct {
 	PromptID             string
 	PromptVersion        string
 	PromptHash           string
+	SessionID            string
 	RenderedPromptHash   string
 	SelectedProfileID    string
+	SelectedBackendID    string
 	ModelName            string
 	Endpoint             string
 	EffectiveModelParams ExecutionTarget
@@ -99,6 +101,7 @@ type PreparedRun struct {
 	PromptVersion        string                  `json:"prompt_version,omitempty"`
 	PromptHash           string                  `json:"prompt_hash,omitempty"`
 	SelectedProfileID    string                  `json:"selected_profile_id"`
+	SelectedBackendID    string                  `json:"selected_backend_id,omitempty"`
 	EffectiveModelParams ExecutionTarget         `json:"effective_model_params"`
 	TargetPresence       ExecutionTargetPresence `json:"-"`
 	OutputContract       OutputContract          `json:"output_contract"`
@@ -158,9 +161,28 @@ type PromptMessageTemplate struct {
 	CacheControl *CacheControl `yaml:"cache_control,omitempty" json:"cache_control,omitempty"`
 }
 
+// Backend describes reusable OpenAI-compatible connection defaults.
+type Backend struct {
+	ID               string
+	Endpoint         string
+	APIKeyEnv        string
+	ExtraParams      map[string]any
+	ConcurrencyLimit int
+	QueueCapacity    int
+	QueueCapacitySet bool
+}
+
+// BackendCapacityPolicy describes normalized run and generation capacity for
+// one limited backend.
+type BackendCapacityPolicy struct {
+	ConcurrencyLimit int
+	QueueCapacity    int
+}
+
 // ExecutionProfile describes how and where to execute a model.
 type ExecutionProfile struct {
 	ID              string         `yaml:"id"`
+	BackendID       string         `yaml:"backend"`
 	Endpoint        string         `yaml:"endpoint"`
 	Model           string         `yaml:"model"`
 	Temperature     float64        `yaml:"temperature"`
@@ -183,7 +205,7 @@ type ExecutionTargetOverride struct {
 	TopP            *float64       `json:"top_p,omitempty"`
 	TimeoutSeconds  *int           `json:"timeout_seconds,omitempty"`
 	ServiceTier     string         `json:"service_tier,omitempty"`
-	ReasoningEffort string         `json:"reasoning_effort,omitempty"`
+	ReasoningEffort *string        `json:"reasoning_effort,omitempty"`
 	APIKeyEnv       string         `json:"api_key_env,omitempty"`
 	ExtraParams     map[string]any `json:"extra_params,omitempty"`
 }
@@ -199,6 +221,7 @@ type ExecutionTargetPresence struct {
 
 // ExecutionTarget represents effective model runtime settings for a run.
 type ExecutionTarget struct {
+	BackendID       string         `yaml:"backend" json:"backend_id,omitempty"`
 	Endpoint        string         `yaml:"endpoint" json:"endpoint"`
 	Model           string         `yaml:"model" json:"model"`
 	Temperature     float64        `yaml:"temperature" json:"temperature"`

internal/domain/session.go

@@ -0,0 +1,19 @@
+package domain
+
+import (
+	"fmt"
+	"strings"
+	"unicode/utf8"
+)
+
+// NormalizeSessionID applies the shared session identifier rule.
+func NormalizeSessionID(raw string) (string, error) {
+	normalized := strings.TrimSpace(raw)
+	if normalized == "" {
+		return "", nil
+	}
+	if length := utf8.RuneCountInString(normalized); length > SessionIDMaxLength {
+		return "", fmt.Errorf("session_id length %d exceeds maximum %d", length, SessionIDMaxLength)
+	}
+	return normalized, nil
+}

internal/domain/session_test.go

@@ -0,0 +1,57 @@
+package domain
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestNormalizeSessionID(t *testing.T) {
+	tests := []struct {
+		name    string
+		raw     string
+		want    string
+		wantErr bool
+	}{
+		{
+			name: "trims surrounding Unicode whitespace",
+			raw:  "\u2003 session-123 \u2003",
+			want: "session-123",
+		},
+		{
+			name: "blank input is omitted",
+			raw:  " \t\u2003 ",
+			want: "",
+		},
+		{
+			name: "maximum Unicode length is accepted",
+			raw:  strings.Repeat("界", SessionIDMaxLength),
+			want: strings.Repeat("界", SessionIDMaxLength),
+		},
+		{
+			name:    "one Unicode code point over maximum is rejected",
+			raw:     strings.Repeat("界", SessionIDMaxLength+1),
+			wantErr: true,
+		},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			got, err := NormalizeSessionID(tt.raw)
+			if tt.wantErr {
+				if err == nil {
+					t.Fatal("expected normalization error")
+				}
+				if !strings.Contains(err.Error(), "exceeds maximum") {
+					t.Fatalf("expected useful length diagnostic, got %v", err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("normalize session id: %v", err)
+			}
+			if got != tt.want {
+				t.Fatalf("normalized session id = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}

internal/jsonvalue/jsonvalue.go

@@ -1,25 +1,30 @@
-package promptkit
+// Package jsonvalue validates and defensively copies JSON-compatible value
+// trees used by public configuration and request boundaries.
+package jsonvalue
 
 import (
 	"encoding/json"
 	"fmt"
 	"math"
 	"reflect"
+	"sort"
 	"strconv"
 )
 
 const maxSafeJSONInteger = 1<<53 - 1
 
-type jsonVisit struct {
+type visit struct {
 	typ reflect.Type
 	ptr uintptr
 }
 
-func copyPublicJSONMap(src map[string]any) (map[string]any, error) {
+// CopyMap validates and deeply copies an extra-parameter map while preserving
+// compatible concrete map, slice, array, scalar, and number types.
+func CopyMap(src map[string]any) (map[string]any, error) {
 	if src == nil {
 		return nil, nil
 	}
-	copied, err := copyPublicJSONValue(reflect.ValueOf(src), "extra_params", make(map[jsonVisit]struct{}))
+	copied, err := copyValue(reflect.ValueOf(src), "extra_params", make(map[visit]struct{}))
 	if err != nil {
 		return nil, err
 	}
@@ -30,7 +35,7 @@ func copyPublicJSONMap(src map[string]any) (map[string]any, error) {
 	return out, nil
 }
 
-func copyPublicJSONValue(value reflect.Value, path string, seen map[jsonVisit]struct{}) (any, error) {
+func copyValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
 	if !value.IsValid() {
 		return nil, nil
 	}
@@ -38,12 +43,15 @@ func copyPublicJSONValue(value reflect.Value, path string, seen map[jsonVisit]st
 		if value.IsNil() {
 			return nil, nil
 		}
-		return copyPublicJSONValue(value.Elem(), path, seen)
+		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 {
+			return nil, fmt.Errorf("%s: invalid JSON number", path)
+		}
 		f, err := strconv.ParseFloat(number.String(), 64)
 		if err != nil || math.IsNaN(f) || math.IsInf(f, 0) {
 			return nil, fmt.Errorf("%s: invalid JSON number", path)
@@ -65,8 +73,8 @@ func copyPublicJSONValue(value reflect.Value, path string, seen map[jsonVisit]st
 		}
 		return value.Interface(), nil
 	case reflect.Float32, reflect.Float64:
-		f := value.Convert(reflect.TypeOf(float64(0))).Float()
-		if math.IsNaN(f) || math.IsInf(f, 0) {
+		number := value.Convert(reflect.TypeOf(float64(0))).Float()
+		if math.IsNaN(number) || math.IsInf(number, 0) {
 			return nil, fmt.Errorf("%s: floating-point value must be finite", path)
 		}
 		return value.Interface(), nil
@@ -74,28 +82,28 @@ func copyPublicJSONValue(value reflect.Value, path string, seen map[jsonVisit]st
 		if value.IsNil() {
 			return nil, nil
 		}
-		visit := jsonVisit{typ: value.Type(), ptr: value.Pointer()}
-		if _, ok := seen[visit]; ok {
+		current := visit{typ: value.Type(), ptr: value.Pointer()}
+		if _, ok := seen[current]; ok {
 			return nil, fmt.Errorf("%s: cyclic value is not supported", path)
 		}
-		seen[visit] = struct{}{}
-		defer delete(seen, visit)
-		return copyPublicJSONValue(value.Elem(), path, seen)
+		seen[current] = struct{}{}
+		defer delete(seen, current)
+		return copyValue(value.Elem(), path, seen)
 	case reflect.Map:
-		return copyPublicJSONMapValue(value, path, seen)
+		return copyMapValue(value, path, seen)
 	case reflect.Slice:
 		if value.IsNil() {
 			return nil, nil
 		}
-		return copyPublicJSONSequenceValue(value, path, seen)
+		return copySequenceValue(value, path, seen)
 	case reflect.Array:
-		return copyPublicJSONSequenceValue(value, path, seen)
+		return copySequenceValue(value, path, seen)
 	default:
 		return nil, fmt.Errorf("%s: unsupported JSON value type %s", path, value.Type())
 	}
 }
 
-func copyPublicJSONMapValue(value reflect.Value, path string, seen map[jsonVisit]struct{}) (any, error) {
+func copyMapValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
 	if value.IsNil() {
 		return nil, nil
 	}
@@ -103,37 +111,43 @@ func copyPublicJSONMapValue(value reflect.Value, path string, seen map[jsonVisit
 		return nil, fmt.Errorf("%s: map key type %s is not supported", path, value.Type().Key())
 	}
 
-	visit := jsonVisit{typ: value.Type(), ptr: value.Pointer()}
-	if _, ok := seen[visit]; ok {
+	current := visit{typ: value.Type(), ptr: value.Pointer()}
+	if _, ok := seen[current]; ok {
 		return nil, fmt.Errorf("%s: cyclic value is not supported", path)
 	}
-	seen[visit] = struct{}{}
-	defer delete(seen, visit)
+	seen[current] = struct{}{}
+	defer delete(seen, current)
+
+	keys := value.MapKeys()
+	sort.Slice(keys, func(i, j int) bool {
+		return keys[i].String() < keys[j].String()
+	})
 
 	type entry struct {
 		key   reflect.Value
 		name  string
 		value any
 	}
-	entries := make([]entry, 0, value.Len())
+	entries := make([]entry, 0, len(keys))
 	preserveType := true
-	elemType := value.Type().Elem()
-	iter := value.MapRange()
-	for iter.Next() {
-		key := iter.Key()
+	elementType := value.Type().Elem()
+	for _, key := range keys {
 		name := key.String()
-		copied, err := copyPublicJSONValue(iter.Value(), path+"."+name, seen)
+		if name == "" {
+			return nil, fmt.Errorf("%s: map key must not be empty", path)
+		}
+		copied, err := copyValue(value.MapIndex(key), path+"."+name, seen)
 		if err != nil {
 			return nil, err
 		}
 		entries = append(entries, entry{key: key, name: name, value: copied})
 		if copied == nil {
-			if !canAssignNil(elemType) {
+			if !canAssignNil(elementType) {
 				preserveType = false
 			}
 			continue
 		}
-		if !reflect.TypeOf(copied).AssignableTo(elemType) {
+		if !reflect.TypeOf(copied).AssignableTo(elementType) {
 			preserveType = false
 		}
 	}
@@ -142,7 +156,7 @@ func copyPublicJSONMapValue(value reflect.Value, path string, seen map[jsonVisit
 		out := reflect.MakeMapWithSize(value.Type(), len(entries))
 		for _, entry := range entries {
 			if entry.value == nil {
-				out.SetMapIndex(entry.key, reflect.Zero(elemType))
+				out.SetMapIndex(entry.key, reflect.Zero(elementType))
 				continue
 			}
 			out.SetMapIndex(entry.key, reflect.ValueOf(entry.value))
@@ -157,33 +171,33 @@ func copyPublicJSONMapValue(value reflect.Value, path string, seen map[jsonVisit
 	return out, nil
 }
 
-func copyPublicJSONSequenceValue(value reflect.Value, path string, seen map[jsonVisit]struct{}) (any, error) {
-	var visit jsonVisit
+func copySequenceValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+	var current visit
 	if value.Kind() == reflect.Slice {
-		visit = jsonVisit{typ: value.Type(), ptr: value.Pointer()}
-		if _, ok := seen[visit]; ok {
+		current = visit{typ: value.Type(), ptr: value.Pointer()}
+		if _, ok := seen[current]; ok {
 			return nil, fmt.Errorf("%s: cyclic value is not supported", path)
 		}
-		seen[visit] = struct{}{}
-		defer delete(seen, visit)
+		seen[current] = struct{}{}
+		defer delete(seen, current)
 	}
 
 	values := make([]any, value.Len())
 	preserveType := true
-	elemType := value.Type().Elem()
+	elementType := value.Type().Elem()
 	for i := 0; i < value.Len(); i++ {
-		copied, err := copyPublicJSONValue(value.Index(i), fmt.Sprintf("%s[%d]", path, i), seen)
+		copied, err := copyValue(value.Index(i), fmt.Sprintf("%s[%d]", path, i), seen)
 		if err != nil {
 			return nil, err
 		}
 		values[i] = copied
 		if copied == nil {
-			if !canAssignNil(elemType) {
+			if !canAssignNil(elementType) {
 				preserveType = false
 			}
 			continue
 		}
-		if !reflect.TypeOf(copied).AssignableTo(elemType) {
+		if !reflect.TypeOf(copied).AssignableTo(elementType) {
 			preserveType = false
 		}
 	}
@@ -195,7 +209,7 @@ func copyPublicJSONSequenceValue(value reflect.Value, path string, seen map[json
 		}
 		for i, copied := range values {
 			if copied == nil {
-				out.Index(i).Set(reflect.Zero(elemType))
+				out.Index(i).Set(reflect.Zero(elementType))
 				continue
 			}
 			out.Index(i).Set(reflect.ValueOf(copied))

internal/jsonvalue/jsonvalue_test.go

@@ -0,0 +1,97 @@
+package jsonvalue_test
+
+import (
+	"encoding/json"
+	"math"
+	"reflect"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
+)
+
+func TestCopyMapPreservesTypesAndIsolatesMutations(t *testing.T) {
+	nested := map[string]int{"limit": 2}
+	sequence := []string{"one", "two"}
+	input := map[string]any{
+		"count":    int64(7),
+		"number":   json.Number("-1.25e+2"),
+		"nested":   nested,
+		"sequence": sequence,
+	}
+
+	copied, err := jsonvalue.CopyMap(input)
+	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
+
+	tests := []struct {
+		name  string
+		value any
+	}{
+		{name: "empty nested key", value: map[string]int{"": 1}},
+		{name: "non-string map key", value: map[int]string{1: "one"}},
+		{name: "unsupported value", value: make(chan int)},
+		{name: "cyclic map", value: cyclicMap},
+		{name: "cyclic slice", value: cyclicSlice},
+		{name: "NaN", value: math.NaN()},
+		{name: "positive infinity", value: math.Inf(1)},
+		{name: "unsafe signed integer", value: int64(1 << 53)},
+		{name: "unsafe unsigned integer", value: uint64(1 << 53)},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if _, err := jsonvalue.CopyMap(map[string]any{"value": tc.value}); err == nil {
+				t.Fatal("expected validation error")
+			}
+		})
+	}
+}
+
+func TestCopyMapValidatesJSONNumberSyntaxAndRange(t *testing.T) {
+	for _, number := range []json.Number{"0", "-1", "1.25", "-1.25e+2"} {
+		t.Run("valid "+number.String(), func(t *testing.T) {
+			got, err := jsonvalue.CopyMap(map[string]any{"value": number})
+			if err != nil {
+				t.Fatalf("copy valid JSON number: %v", err)
+			}
+			if !reflect.DeepEqual(got["value"], number) {
+				t.Fatalf("JSON number changed: got %#v want %#v", got["value"], number)
+			}
+		})
+	}
+
+	for _, number := range []json.Number{"", "01", "+1", "1.", ".1", "1e9999", "not-a-number"} {
+		t.Run("invalid "+number.String(), func(t *testing.T) {
+			if _, err := jsonvalue.CopyMap(map[string]any{"value": number}); err == nil {
+				t.Fatal("expected invalid JSON number error")
+			}
+		})
+	}
+}

internal/llm/openai_compatible_client.go

@@ -12,7 +12,6 @@ import (
 	"os"
 	"strings"
 	"time"
-	"unicode/utf8"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
@@ -177,12 +176,11 @@ func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultMod
 	wireReq := openAIChatRequest{
 		Model: model,
 	}
-	if sessionID := strings.TrimSpace(req.Prompt.SessionID); sessionID != "" {
-		if n := utf8.RuneCountInString(sessionID); n > domain.SessionIDMaxLength {
-			return openAIChatRequest{}, fmt.Errorf("session_id length %d exceeds maximum %d", n, domain.SessionIDMaxLength)
-		}
-		wireReq.SessionID = sessionID
+	sessionID, err := domain.NormalizeSessionID(req.Prompt.SessionID)
+	if err != nil {
+		return openAIChatRequest{}, err
 	}
+	wireReq.SessionID = sessionID
 
 	wireReq.Messages = make([]openAIChatRequestMessage, 0, len(req.Prompt.Messages))
 	for _, msg := range req.Prompt.Messages {
@@ -262,7 +260,7 @@ func openAIChatRequestPayload(req openAIChatRequest) (map[string]any, error) {
 		if key == "" {
 			return nil, errors.New("extra_params key must not be empty")
 		}
-		if _, reserved := reservedOpenAIChatRequestFields[key]; reserved {
+		if IsReservedOpenAIChatRequestField(key) {
 			return nil, fmt.Errorf("extra_params key %q collides with reserved request field", key)
 		}
 		if _, err := json.Marshal(value); err != nil {
@@ -274,16 +272,23 @@ func openAIChatRequestPayload(req openAIChatRequest) (map[string]any, error) {
 	return out, nil
 }
 
-var reservedOpenAIChatRequestFields = map[string]struct{}{
-	"model":            {},
-	"session_id":       {},
-	"messages":         {},
-	"temperature":      {},
-	"max_tokens":       {},
-	"top_p":            {},
-	"service_tier":     {},
-	"reasoning_effort": {},
-	"response_format":  {},
+// IsReservedOpenAIChatRequestField reports whether name is owned by the
+// standard OpenAI-compatible chat request rather than extra parameters.
+func IsReservedOpenAIChatRequestField(name string) bool {
+	switch name {
+	case "model",
+		"session_id",
+		"messages",
+		"temperature",
+		"max_tokens",
+		"top_p",
+		"service_tier",
+		"reasoning_effort",
+		"response_format":
+		return true
+	default:
+		return false
+	}
 }
 
 type openAIChatRequestMessage struct {

internal/profile/builtin/assets/aion-labs/aion-2.yml

@@ -1,9 +1,8 @@
 id: aion-2
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: aion-labs/aion-2.0
 temperature: 0.72
 reasoning_effort: high
 top_p: 0.95
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/anthropic/claude-fable-latest.yml

@@ -1,7 +1,6 @@
 id: claude-fable-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~anthropic/claude-fable-latest"
 reasoning_effort: high
 timeout_seconds: 600
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/anthropic/claude-haiku-latest.yml

@@ -1,7 +1,6 @@
 id: claude-haiku-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~anthropic/claude-haiku-latest"
 reasoning_effort: medium
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/anthropic/claude-opus-latest.yml

@@ -1,7 +1,6 @@
 id: claude-opus-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~anthropic/claude-opus-latest"
 reasoning_effort: high
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/anthropic/claude-sonnet-latest.yml

@@ -1,7 +1,6 @@
 id: claude-sonnet-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~anthropic/claude-sonnet-latest"
 reasoning_effort: high
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/deepseek/deepseek-3-2.yml

@@ -1,7 +1,6 @@
 id: deepseek-3-2
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: deepseek/deepseek-v3.2
 reasoning_effort: high
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/deepseek/deepseek-4-flash.yml

@@ -1,7 +1,6 @@
 id: deepseek-4-flash
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: deepseek/deepseek-v4-flash
 #reasoning_effort: medium
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/deepseek/deepseek-4-pro.yml

@@ -1,7 +1,6 @@
 id: deepseek-4-pro
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: deepseek/deepseek-v4-pro
 reasoning_effort: high
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-2-flash-lite.yml

@@ -1,9 +1,8 @@
 id: gemini-2-flash-lite
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "google/gemini-2.5-flash-lite"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-2-flash.yml

@@ -1,9 +1,8 @@
 id: gemini-2-flash
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "google/gemini-2.5-flash"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-2-pro.yml

@@ -1,9 +1,8 @@
 id: gemini-2-pro
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "google/gemini-2.5-pro"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-3-flash-lite.yml

@@ -1,9 +1,8 @@
 id: gemini-3-flash-lite
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "google/gemini-3.1-flash-lite"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-flash-latest.yml

@@ -1,9 +1,8 @@
 id: gemini-flash-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~google/gemini-flash-latest"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemini-pro-latest.yml

@@ -1,9 +1,8 @@
 id: gemini-pro-latest
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "~google/gemini-pro-latest"
 #temperature: 0.15
 reasoning_effort: high
 #top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/google/gemma-4-31b.yml

@@ -1,9 +1,8 @@
 id: gemma-4-31b
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: google/gemma-4-31b-it:exacto
 temperature: 0.15
 reasoning_effort: high
 top_p: 0.98
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/minimax/minimax-m2.yml

@@ -1,9 +1,8 @@
 id: minimax-m2
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: minimax/minimax-m2.5
 temperature: 0.5
 reasoning_effort: high
 top_p: 0.95
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/minimax/minimax-m3.yml

@@ -1,9 +1,8 @@
 id: minimax-m3
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: minimax/minimax-m3
 #temperature: 0.5
 reasoning_effort: high
 #top_p: 0.95
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/mistral/mistral-large-2512.yml

@@ -1,7 +1,6 @@
 id: mistral-large-2512
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: mistralai/mistral-large-2512
 temperature: 0.15
 top_p: 0.98
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY

internal/profile/builtin/assets/mistral/mistral-medium-3-5.yml

@@ -1,8 +1,7 @@
 id: mistral-medium-3-5
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: mistralai/mistral-medium-3-5
 temperature: 0.15
 reasoning_effort: high
 top_p: 0.98
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY

internal/profile/builtin/assets/mistral/mistral-small-3.yml

@@ -1,7 +1,6 @@
 id: mistral-small-3
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: mistralai/mistral-small-3.2-24b-instruct
 temperature: 0.05
 top_p: 1.0
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY

internal/profile/builtin/assets/mistral/mistral-small-4.yml

@@ -1,8 +1,7 @@
 id: mistral-small-4
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: mistralai/mistral-small-2603
 temperature: 0.1
 reasoning_effort: high
 top_p: 0.98
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY

internal/profile/builtin/assets/nvidia/nemotron-3-ultra.yml

@@ -1,7 +1,6 @@
 id: nemotron-3-ultra
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: nvidia/nemotron-3-ultra-550b-a55b
 reasoning_effort: high
 timeout_seconds: 180
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/openai/gpt-5-mini.yml

@@ -1,7 +1,6 @@
 id: gpt-5-mini
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "openai/gpt-5.4-mini"
 reasoning_effort: high
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/assets/openai/gpt-5-nano.yml

@@ -1,7 +1,6 @@
 id: gpt-5-nano
-endpoint: https://openrouter.ai/api/v1
+backend: openrouter
 model: "openai/gpt-5.4-nano"
 reasoning_effort: high
 timeout_seconds: 240
-api_key_env: OPENROUTER_API_KEY
 service_tier: flex

internal/profile/builtin/repository_test.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 	"testing"
 
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
 	"gopkg.in/yaml.v3"
@@ -28,6 +29,12 @@ func TestBuiltInProfilesValidateThroughRepository(t *testing.T) {
 			if p.ID != id {
 				t.Fatalf("expected profile id %q, got %q", id, p.ID)
 			}
+			if p.BackendID != backend.OpenRouterID {
+				t.Fatalf("expected profile %q to select %q, got %q", id, backend.OpenRouterID, p.BackendID)
+			}
+			if p.Endpoint != "" || p.APIKeyEnv != "" {
+				t.Fatalf("expected profile %q to inherit backend connection settings, got endpoint=%q api_key_env=%q", id, p.Endpoint, p.APIKeyEnv)
+			}
 		})
 	}
 }
@@ -60,6 +67,15 @@ func loadBuiltInProfileIDs(t *testing.T) map[string]string {
 		if _, ok := raw["api_key"]; ok {
 			t.Fatalf("built-in profile %s contains raw api_key", name)
 		}
+		if raw["backend"] != backend.OpenRouterID {
+			t.Fatalf("built-in profile %s does not select %q", name, backend.OpenRouterID)
+		}
+		if _, ok := raw["endpoint"]; ok {
+			t.Fatalf("built-in profile %s repeats endpoint", name)
+		}
+		if _, ok := raw["api_key_env"]; ok {
+			t.Fatalf("built-in profile %s repeats api_key_env", name)
+		}
 		id, ok := raw["id"].(string)
 		if !ok || strings.TrimSpace(id) == "" {
 			t.Fatalf("built-in profile %s has missing id", name)

internal/profile/filesystem_repository.go

@@ -121,6 +121,7 @@ func loadProfile(ctx context.Context, fsys fs.FS, root string, id string) (*doma
 		if prof.ID != id {
 			continue
 		}
+		prof.BackendID = strings.TrimSpace(prof.BackendID)
 		if err := validateProfile(&prof); err != nil {
 			if errors.Is(err, ErrRawAPIKeyNotAllowed) {
 				return nil, fmt.Errorf("%w: %s", err, relPath)
@@ -189,8 +190,8 @@ func validateProfile(p *domain.ExecutionProfile) error {
 	if strings.TrimSpace(p.ID) == "" {
 		return errors.New("id is required")
 	}
-	if strings.TrimSpace(p.Endpoint) == "" {
-		return errors.New("endpoint is required")
+	if strings.TrimSpace(p.BackendID) == "" && strings.TrimSpace(p.Endpoint) == "" {
+		return errors.New("backend or endpoint is required")
 	}
 	if strings.TrimSpace(p.Model) == "" {
 		return errors.New("model is required")

internal/profile/repository_test.go

@@ -52,6 +52,43 @@ func TestFilesystemRepository_GetProfile(t *testing.T) {
 		}
 	})
 
+	t.Run("backend and endpoint connection matrix", func(t *testing.T) {
+		tests := []struct {
+			name         string
+			connection   string
+			wantBackend  string
+			wantEndpoint string
+			wantErr      bool
+		}{
+			{name: "backend only", connection: "backend: '  openrouter  '", wantBackend: "openrouter"},
+			{name: "endpoint only", connection: "endpoint: http://localhost:8000/v1", wantEndpoint: "http://localhost:8000/v1"},
+			{name: "both", connection: "backend: openrouter\nendpoint: http://localhost:8000/v1", wantBackend: "openrouter", wantEndpoint: "http://localhost:8000/v1"},
+			{name: "neither", wantErr: true},
+			{name: "blank backend", connection: "backend: '   '", wantErr: true},
+		}
+
+		for _, tt := range tests {
+			t.Run(tt.name, func(t *testing.T) {
+				id := "connection-" + strings.ReplaceAll(tt.name, " ", "-")
+				writeProfileTestFile(t, filepath.Join(tmpDir, id+".yaml"), "id: "+id+"\nmodel: model\n"+tt.connection+"\n")
+
+				p, err := repo.GetProfile(ctx, id)
+				if tt.wantErr {
+					if !errors.Is(err, ErrInvalidProfile) {
+						t.Fatalf("expected ErrInvalidProfile, got %v", err)
+					}
+					return
+				}
+				if err != nil {
+					t.Fatalf("expected profile to load, got %v", err)
+				}
+				if p.BackendID != tt.wantBackend || p.Endpoint != tt.wantEndpoint {
+					t.Fatalf("unexpected connection values: backend=%q endpoint=%q", p.BackendID, p.Endpoint)
+				}
+			})
+		}
+	})
+
 	t.Run("valid profile with api_key_env", func(t *testing.T) {
 		p, err := repo.GetProfile(ctx, "local-secure")
 		if err != nil {

internal/prompt/go_renderer.go

@@ -5,10 +5,9 @@ import (
 	"context"
 	"errors"
 	"fmt"
-	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
-	"strings"
 	"text/template"
-	"unicode/utf8"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 )
 
 var (
@@ -95,10 +94,6 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 }
 
 func renderSessionID(raw string, funcs template.FuncMap, vars map[string]string) (string, error) {
-	if strings.TrimSpace(raw) == "" {
-		return "", nil
-	}
-
 	tmpl, err := template.New("session_id").Funcs(funcs).Option("missingkey=error").Parse(raw)
 	if err != nil {
 		return "", fmt.Errorf("%w: session_id: %v", ErrInvalidTemplate, err)
@@ -109,9 +104,9 @@ func renderSessionID(raw string, funcs template.FuncMap, vars map[string]string)
 		return "", fmt.Errorf("%w: session_id: %w", ErrRenderFailure, err)
 	}
 
-	sessionID := strings.TrimSpace(buf.String())
-	if n := utf8.RuneCountInString(sessionID); n > domain.SessionIDMaxLength {
-		return "", fmt.Errorf("%w: session_id length %d exceeds maximum %d", ErrRenderFailure, n, domain.SessionIDMaxLength)
+	sessionID, err := domain.NormalizeSessionID(buf.String())
+	if err != nil {
+		return "", fmt.Errorf("%w: session_id: %v", ErrRenderFailure, err)
 	}
 	return sessionID, nil
 }

internal/usecase/repairer.go

@@ -17,6 +17,7 @@ type OutputRepairer interface {
 type RepairRequest struct {
 	PreviousOutput   string
 	ValidationErrors []string
+	SessionID        string
 	Target           domain.ExecutionTarget
 	StructuredOutput *domain.StructuredOutputSpec
 	Attempt          int
@@ -42,23 +43,26 @@ func (r *defaultOutputRepairer) Repair(ctx context.Context, req RepairRequest) (
 		errs = strings.Join(req.ValidationErrors, "\n")
 	}
 
-	prompt := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
-		{
-			Role:    "system",
-			Content: "You repair invalid JSON output. Return only corrected JSON. Do not include explanations or markdown code fences.",
+	prompt := domain.RenderedPrompt{
+		SessionID: req.SessionID,
+		Messages: []domain.RenderedMessage{
+			{
+				Role:    "system",
+				Content: "You repair invalid JSON output. Return only corrected JSON. Do not include explanations or markdown code fences.",
+			},
+			{
+				Role: "user",
+				Content: fmt.Sprintf(
+					"Repair attempt %d of %d for validation mode %s.\n\nValidation errors:\n%s\n\nPrevious output:\n%s\n\nReturn only corrected JSON.",
+					req.Attempt,
+					req.MaxAttempts,
+					req.Mode,
+					errs,
+					req.PreviousOutput,
+				),
+			},
 		},
-		{
-			Role: "user",
-			Content: fmt.Sprintf(
-				"Repair attempt %d of %d for validation mode %s.\n\nValidation errors:\n%s\n\nPrevious output:\n%s\n\nReturn only corrected JSON.",
-				req.Attempt,
-				req.MaxAttempts,
-				req.Mode,
-				errs,
-				req.PreviousOutput,
-			),
-		},
-	}}
+	}
 
 	resp, err := r.llm.Generate(ctx, domain.GenerateRequest{
 		Prompt:           prompt,

internal/usecase/runner.go

@@ -14,6 +14,7 @@ import (
 	"unicode"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/artifact"
+	"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"
@@ -40,41 +41,80 @@ var (
 type Runner struct {
 	promptDefs promptdef.Repository
 	profiles   profile.Repository
+	backends   BackendResolver
 	artifacts  artifact.Reader
 	renderer   prompt.Renderer
 	llm        llm.Client
 	validator  validate.Validator
 	repairer   OutputRepairer
+	admitter   RunAdmitter
+}
+
+// BackendResolver resolves one normalized backend ID.
+type BackendResolver interface {
+	GetBackend(string) (domain.Backend, error)
+}
+
+// RunAdmitter reserves capacity for one resolved backend run.
+type RunAdmitter interface {
+	Admit(context.Context, string) (func(), error)
+}
+
+type preparationState struct {
+	definition           *domain.PromptDefinition
+	directSessionID      string
+	promptDefinitionHash string
+	selectedProfileID    string
+	effectiveModel       domain.ExecutionTarget
+	targetPresence       domain.ExecutionTargetPresence
+	effectiveContract    domain.OutputContract
+	start                time.Time
 }
 
 func NewRunner(
 	promptDefs promptdef.Repository,
 	profiles profile.Repository,
+	backends BackendResolver,
 	artifacts artifact.Reader,
 	renderer prompt.Renderer,
 	llmClient llm.Client,
 	validator validate.Validator,
+	admitter RunAdmitter,
 ) *Runner {
-	return NewRunnerWithRepairer(promptDefs, profiles, artifacts, renderer, llmClient, validator, nil)
+	return NewRunnerWithRepairer(
+		promptDefs,
+		profiles,
+		backends,
+		artifacts,
+		renderer,
+		llmClient,
+		validator,
+		nil,
+		admitter,
+	)
 }
 
 func NewRunnerWithRepairer(
 	promptDefs promptdef.Repository,
 	profiles profile.Repository,
+	backends BackendResolver,
 	artifacts artifact.Reader,
 	renderer prompt.Renderer,
 	llmClient llm.Client,
 	validator validate.Validator,
 	repairer OutputRepairer,
+	admitter RunAdmitter,
 ) *Runner {
 	return &Runner{
 		promptDefs: promptDefs,
 		profiles:   profiles,
+		backends:   backends,
 		artifacts:  artifacts,
 		renderer:   renderer,
 		llm:        llmClient,
 		validator:  validator,
 		repairer:   repairer,
+		admitter:   admitter,
 	}
 }
 
@@ -86,7 +126,27 @@ func (r *Runner) Run(ctx context.Context, req domain.RunRequest) (*domain.RunRes
 
 	start := time.Now().UTC()
 
-	prepared, err := r.Prepare(ctx, req)
+	state, err := r.resolvePreparation(ctx, req, time.Now().UTC())
+	if err != nil {
+		return nil, err
+	}
+
+	if r.admitter != nil {
+		release, admitErr := r.admitter.Admit(ctx, state.effectiveModel.BackendID)
+		if admitErr != nil {
+			if errors.Is(admitErr, capacity.ErrCapacityExceeded) {
+				return nil, fmt.Errorf(
+					"backend %q admission: %w",
+					state.effectiveModel.BackendID,
+					admitErr,
+				)
+			}
+			return nil, admitErr
+		}
+		defer release()
+	}
+
+	prepared, err := r.completePreparation(ctx, req, state)
 	if err != nil {
 		return nil, err
 	}
@@ -118,6 +178,7 @@ func (r *Runner) Run(ctx context.Context, req domain.RunRequest) (*domain.RunRes
 			repairResp, repairErr := r.repairer.Repair(ctx, RepairRequest{
 				PreviousOutput:   genResp.Content,
 				ValidationErrors: validationResult.Errors,
+				SessionID:        prepared.SessionID,
 				Target:           prepared.EffectiveModelParams,
 				StructuredOutput: prepared.StructuredOutput,
 				Attempt:          attemptsUsed,
@@ -151,8 +212,10 @@ func (r *Runner) Run(ctx context.Context, req domain.RunRequest) (*domain.RunRes
 		PromptID:             prepared.PromptID,
 		PromptVersion:        prepared.PromptVersion,
 		PromptHash:           prepared.PromptHash,
+		SessionID:            prepared.SessionID,
 		RenderedPromptHash:   prepared.RenderedPromptHash,
 		SelectedProfileID:    prepared.SelectedProfileID,
+		SelectedBackendID:    prepared.SelectedBackendID,
 		ModelName:            prepared.EffectiveModelParams.Model,
 		Endpoint:             prepared.EffectiveModelParams.Endpoint,
 		EffectiveModelParams: prepared.EffectiveModelParams,
@@ -165,11 +228,25 @@ func (r *Runner) Run(ctx context.Context, req domain.RunRequest) (*domain.RunRes
 }
 
 func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.PreparedRun, error) {
+	state, err := r.resolvePreparation(ctx, req, time.Now().UTC())
+	if err != nil {
+		return nil, err
+	}
+	return r.completePreparation(ctx, req, state)
+}
+
+func (r *Runner) resolvePreparation(
+	ctx context.Context,
+	req domain.RunRequest,
+	start time.Time,
+) (*preparationState, error) {
 	if strings.TrimSpace(req.PromptID) == "" {
 		return nil, fmt.Errorf("%w: prompt id is required", ErrInvalidRequest)
 	}
-
-	start := time.Now().UTC()
+	directSessionID, err := domain.NormalizeSessionID(req.SessionID)
+	if err != nil {
+		return nil, fmt.Errorf("%w: session_id: %v", ErrInvalidRequest, err)
+	}
 
 	def, err := r.promptDefs.GetPromptDefinition(ctx, req.PromptID, req.PromptVersion)
 	if err != nil {
@@ -193,7 +270,20 @@ func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.Pr
 		return nil, fmt.Errorf("%w: %w", ErrProfileLoad, err)
 	}
 
-	effectiveModel, targetPresence, err := resolveExecutionTarget(execProfile, req.Execution)
+	var selectedBackend *domain.Backend
+	if backendID := strings.TrimSpace(execProfile.BackendID); backendID != "" {
+		execProfile.BackendID = backendID
+		if r.backends == nil {
+			return nil, fmt.Errorf("%w: backend %q cannot be resolved", ErrProfileLoad, backendID)
+		}
+		resolvedBackend, resolveErr := r.backends.GetBackend(backendID)
+		if resolveErr != nil {
+			return nil, fmt.Errorf("%w: backend %q: %w", ErrProfileLoad, backendID, resolveErr)
+		}
+		selectedBackend = &resolvedBackend
+	}
+
+	effectiveModel, targetPresence, err := resolveExecutionTarget(selectedBackend, execProfile, req.Execution)
 	if err != nil {
 		return nil, fmt.Errorf("%w: %w", ErrInvalidRequest, err)
 	}
@@ -209,7 +299,28 @@ func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.Pr
 	}
 
 	effectiveContract := resolveOutputContract(def, req.Validation)
-	structuredOutput, err := r.resolveStructuredOutput(ctx, def, effectiveContract)
+	return &preparationState{
+		definition:           def,
+		directSessionID:      directSessionID,
+		promptDefinitionHash: promptDefinitionHash,
+		selectedProfileID:    selectedProfileID,
+		effectiveModel:       effectiveModel,
+		targetPresence:       targetPresence,
+		effectiveContract:    effectiveContract,
+		start:                start,
+	}, nil
+}
+
+func (r *Runner) completePreparation(
+	ctx context.Context,
+	req domain.RunRequest,
+	state *preparationState,
+) (*domain.PreparedRun, error) {
+	structuredOutput, err := r.resolveStructuredOutput(
+		ctx,
+		state.definition,
+		state.effectiveContract,
+	)
 	if err != nil {
 		return nil, err
 	}
@@ -228,28 +339,38 @@ func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.Pr
 		inputHashes[name] = art.Hash
 	}
 
-	renderedPrompt, err := r.renderer.Render(ctx, def, resolvedInputs, req.Vars)
+	definitionToRender := state.definition
+	if state.directSessionID != "" {
+		definitionCopy := *state.definition
+		definitionCopy.SessionID = ""
+		definitionToRender = &definitionCopy
+	}
+	renderedPrompt, err := r.renderer.Render(ctx, definitionToRender, resolvedInputs, req.Vars)
 	if err != nil {
 		return nil, fmt.Errorf("%w: %w", ErrPromptRender, err)
 	}
+	if state.directSessionID != "" {
+		renderedPrompt.SessionID = state.directSessionID
+	}
 
 	end := time.Now().UTC()
 	return &domain.PreparedRun{
-		PromptID:             def.ID,
-		PromptVersion:        def.Version,
-		PromptHash:           promptDefinitionHash,
-		SelectedProfileID:    selectedProfileID,
-		EffectiveModelParams: effectiveModel,
-		TargetPresence:       targetPresence,
-		OutputContract:       effectiveContract,
+		PromptID:             state.definition.ID,
+		PromptVersion:        state.definition.Version,
+		PromptHash:           state.promptDefinitionHash,
+		SelectedProfileID:    state.selectedProfileID,
+		SelectedBackendID:    state.effectiveModel.BackendID,
+		EffectiveModelParams: state.effectiveModel,
+		TargetPresence:       state.targetPresence,
+		OutputContract:       state.effectiveContract,
 		StructuredOutput:     structuredOutput,
 		InputHashes:          inputHashes,
 		SessionID:            renderedPrompt.SessionID,
 		RenderedPromptHash:   hashRenderedPrompt(*renderedPrompt),
 		Messages:             renderedPrompt.Messages,
-		StartTime:            start,
+		StartTime:            state.start,
 		EndTime:              end,
-		DurationMS:           end.Sub(start).Milliseconds(),
+		DurationMS:           end.Sub(state.start).Milliseconds(),
 	}, nil
 }
 
@@ -338,7 +459,10 @@ func (r *Runner) shouldAttemptRepair(contract domain.OutputContract, validationR
 
 func mergeExecutionTarget(base domain.ExecutionTarget, override domain.ExecutionTarget) domain.ExecutionTarget {
 	out := base
-	if override.Endpoint != "" {
+	if strings.TrimSpace(override.BackendID) != "" {
+		out.BackendID = override.BackendID
+	}
+	if strings.TrimSpace(override.Endpoint) != "" {
 		out.Endpoint = override.Endpoint
 	}
 	if override.Model != "" {
@@ -367,6 +491,7 @@ func mergeExecutionTarget(base domain.ExecutionTarget, override domain.Execution
 	}
 	if override.APIKeyRequired {
 		out.APIKeyRequired = true
+		out.APIKeyEnv = ""
 	}
 	if len(override.ExtraParams) > 0 {
 		out.ExtraParams = copyExtraParams(override.ExtraParams)
@@ -414,8 +539,8 @@ func mergeExecutionTargetOverride(base domain.ExecutionTarget, override domain.E
 	if strings.TrimSpace(override.ServiceTier) != "" {
 		out.ServiceTier = override.ServiceTier
 	}
-	if strings.TrimSpace(override.ReasoningEffort) != "" {
-		out.ReasoningEffort = override.ReasoningEffort
+	if override.ReasoningEffort != nil {
+		out.ReasoningEffort = strings.TrimSpace(*override.ReasoningEffort)
 	}
 	if strings.TrimSpace(override.APIKeyEnv) != "" {
 		out.APIKeyEnv = override.APIKeyEnv
@@ -426,8 +551,9 @@ func mergeExecutionTargetOverride(base domain.ExecutionTarget, override domain.E
 	return out, presence, nil
 }
 
-func resolveExecutionTarget(profileValue *domain.ExecutionProfile, override *domain.ExecutionTargetOverride) (domain.ExecutionTarget, domain.ExecutionTargetPresence, error) {
+func resolveExecutionTarget(backendValue *domain.Backend, profileValue *domain.ExecutionProfile, override *domain.ExecutionTargetOverride) (domain.ExecutionTarget, domain.ExecutionTargetPresence, error) {
 	out := defaults.ExecutionTargetDefault()
+	out = mergeExecutionTarget(out, backendToTarget(backendValue))
 	out = mergeExecutionTarget(out, executionProfileToTarget(profileValue))
 	var presence domain.ExecutionTargetPresence
 	if override != nil {
@@ -461,8 +587,13 @@ func executionProfileToTarget(p *domain.ExecutionProfile) domain.ExecutionTarget
 	if p == nil {
 		return domain.ExecutionTarget{}
 	}
+	endpoint := p.Endpoint
+	if strings.TrimSpace(endpoint) == "" {
+		endpoint = ""
+	}
 	return domain.ExecutionTarget{
-		Endpoint:        p.Endpoint,
+		BackendID:       p.BackendID,
+		Endpoint:        endpoint,
 		Model:           p.Model,
 		Temperature:     p.Temperature,
 		MaxTokens:       p.MaxTokens,
@@ -476,6 +607,18 @@ func executionProfileToTarget(p *domain.ExecutionProfile) domain.ExecutionTarget
 	}
 }
 
+func backendToTarget(value *domain.Backend) domain.ExecutionTarget {
+	if value == nil {
+		return domain.ExecutionTarget{}
+	}
+	return domain.ExecutionTarget{
+		BackendID:   value.ID,
+		Endpoint:    value.Endpoint,
+		APIKeyEnv:   value.APIKeyEnv,
+		ExtraParams: copyExtraParams(value.ExtraParams),
+	}
+}
+
 func copyExtraParams(src map[string]any) map[string]any {
 	if len(src) == 0 {
 		return nil

internal/validate/standard_validator.go

@@ -6,6 +6,7 @@ import (
 	"errors"
 	"fmt"
 	"io/fs"
+	"net/url"
 	"os"
 	"path"
 	"path/filepath"
@@ -16,6 +17,8 @@ import (
 	"github.com/santhosh-tekuri/jsonschema/v6"
 )
 
+const jsonSchemaDraft2020 = "https://json-schema.org/draft/2020-12/schema"
+
 // StandardValidator provides basic, JSON, and JSON Schema output validation.
 type StandardValidator struct {
 	schemaBaseDir string
@@ -121,7 +124,11 @@ func (v *StandardValidator) validateJSONSchema(instance any, schemaPath string)
 		return nil, err
 	}
 
-	compiler := jsonschema.NewCompiler()
+	schemaRoot, err := v.schemaRoot()
+	if err != nil {
+		return nil, err
+	}
+	compiler := newSchemaCompiler(standardSchemaLoader{root: schemaRoot})
 	schema, err := compiler.Compile(resolvedSchemaPath)
 	if err != nil {
 		return nil, fmt.Errorf("failed to compile JSON schema %q: %w", resolvedSchemaPath, err)
@@ -140,7 +147,10 @@ func (v *FSValidator) validateJSONSchema(instance any, schemaPath string) ([]str
 	}
 
 	resourceURL := fsSchemaResourceURL(schemaName)
-	compiler := jsonschema.NewCompiler()
+	if err := validateSchemaDialect(schemaDoc); err != nil {
+		return nil, fmt.Errorf("failed to compile JSON schema %q: %w", schemaName, err)
+	}
+	compiler := newSchemaCompiler(fsSchemaLoader{fsys: v.fsys, root: filecatalog.CleanFSRoot(v.root)})
 	if err := compiler.AddResource(resourceURL, schemaDoc); err != nil {
 		return nil, fmt.Errorf("failed to register JSON schema %q: %w", schemaName, err)
 	}
@@ -184,6 +194,9 @@ func (v *StandardValidator) LoadSchemaDocument(ctx context.Context, schemaPath s
 	if err := json.Unmarshal(raw, &doc); err != nil {
 		return nil, fmt.Errorf("failed to decode JSON schema %q: %w", resolved, err)
 	}
+	if err := validateSchemaDialect(doc); err != nil {
+		return nil, fmt.Errorf("failed to decode JSON schema %q: %w", resolved, err)
+	}
 	return doc, nil
 }
 
@@ -206,12 +219,14 @@ func (v *StandardValidator) resolveSchemaPath(schemaPath string) (string, error)
 		return "", errors.New("schema path is required for json_schema validation")
 	}
 
-	resolved := schemaPath
-	if !filepath.IsAbs(schemaPath) {
-		resolved = filepath.Join(v.schemaBaseDir, schemaPath)
+	root, err := v.schemaRoot()
+	if err != nil {
+		return "", err
+	}
+	resolved, err := containedFilesystemPath(root, schemaPath)
+	if err != nil {
+		return "", err
 	}
-
-	resolved = filepath.Clean(resolved)
 	if _, err := os.Stat(resolved); err != nil {
 		return "", fmt.Errorf("failed to access schema file %q: %w", resolved, err)
 	}
@@ -219,6 +234,22 @@ func (v *StandardValidator) resolveSchemaPath(schemaPath string) (string, error)
 	return resolved, nil
 }
 
+func (v *StandardValidator) schemaRoot() (string, error) {
+	root := v.schemaBaseDir
+	if strings.TrimSpace(root) == "" {
+		root = "."
+	}
+	absolute, err := filepath.Abs(root)
+	if err != nil {
+		return "", fmt.Errorf("failed to resolve schema source %q: %w", root, err)
+	}
+	resolved, err := filepath.EvalSymlinks(absolute)
+	if err != nil {
+		return "", fmt.Errorf("failed to access schema source %q: %w", root, err)
+	}
+	return resolved, nil
+}
+
 func (v *FSValidator) loadSchemaDocument(schemaPath string) (string, any, error) {
 	resolved, err := v.resolveSchemaPath(schemaPath)
 	if err != nil {
@@ -234,6 +265,9 @@ func (v *FSValidator) loadSchemaDocument(schemaPath string) (string, any, error)
 	if err := json.Unmarshal(raw, &doc); err != nil {
 		return "", nil, fmt.Errorf("failed to decode JSON schema %q: %w", resolved, err)
 	}
+	if err := validateSchemaDialect(doc); err != nil {
+		return "", nil, fmt.Errorf("failed to decode JSON schema %q: %w", resolved, err)
+	}
 	return resolved, doc, nil
 }
 
@@ -290,3 +324,131 @@ func cleanSchemaFSPath(schemaPath string) (string, error) {
 func fsSchemaResourceURL(schemaName string) string {
 	return "promptkit-schema:///" + strings.TrimPrefix(path.Clean(schemaName), "/")
 }
+
+func newSchemaCompiler(loader jsonschema.URLLoader) *jsonschema.Compiler {
+	compiler := jsonschema.NewCompiler()
+	compiler.DefaultDraft(jsonschema.Draft2020)
+	compiler.UseLoader(loader)
+	return compiler
+}
+
+func validateSchemaDialect(doc any) error {
+	object, ok := doc.(map[string]any)
+	if !ok {
+		return nil
+	}
+	value, ok := object["$schema"]
+	if !ok {
+		return nil
+	}
+	dialect, ok := value.(string)
+	if !ok {
+		return errors.New("$schema must be a string")
+	}
+	if dialect != jsonSchemaDraft2020 && dialect != jsonSchemaDraft2020+"#" {
+		return fmt.Errorf("unsupported JSON Schema dialect %q; expected %q", dialect, jsonSchemaDraft2020)
+	}
+	return nil
+}
+
+type standardSchemaLoader struct {
+	root string
+}
+
+func (l standardSchemaLoader) Load(resourceURL string) (any, error) {
+	fileName, err := (jsonschema.FileLoader{}).ToFile(resourceURL)
+	if err != nil {
+		return nil, fmt.Errorf("schema reference %q is not a contained file reference: %w", resourceURL, err)
+	}
+	resolved, err := containedFilesystemPath(l.root, fileName)
+	if err != nil {
+		return nil, err
+	}
+	return loadJSONSchemaFile(resolved)
+}
+
+func containedFilesystemPath(root, name string) (string, error) {
+	candidate := name
+	if !filepath.IsAbs(candidate) {
+		candidate = filepath.Join(root, candidate)
+	}
+	candidate, err := filepath.Abs(candidate)
+	if err != nil {
+		return "", fmt.Errorf("failed to resolve schema path %q: %w", name, err)
+	}
+	candidate, err = filepath.EvalSymlinks(candidate)
+	if err != nil {
+		return "", fmt.Errorf("failed to access schema file %q: %w", candidate, err)
+	}
+	relative, err := filepath.Rel(root, candidate)
+	if err != nil {
+		return "", fmt.Errorf("failed to compare schema path %q with source root: %w", candidate, err)
+	}
+	if relative == ".." || strings.HasPrefix(relative, ".."+string(filepath.Separator)) {
+		return "", fmt.Errorf("schema path %q escapes source root", name)
+	}
+	return candidate, nil
+}
+
+func loadJSONSchemaFile(name string) (any, error) {
+	raw, err := os.ReadFile(name)
+	if err != nil {
+		return nil, err
+	}
+	var doc any
+	if err := json.Unmarshal(raw, &doc); err != nil {
+		return nil, err
+	}
+	if err := validateSchemaDialect(doc); err != nil {
+		return nil, err
+	}
+	return doc, nil
+}
+
+type fsSchemaLoader struct {
+	fsys fs.FS
+	root string
+}
+
+func (l fsSchemaLoader) Load(resourceURL string) (any, error) {
+	parsed, err := url.Parse(resourceURL)
+	if err != nil {
+		return nil, fmt.Errorf("invalid schema reference %q: %w", resourceURL, err)
+	}
+	if parsed.Scheme != "promptkit-schema" || parsed.Host != "" {
+		return nil, fmt.Errorf("schema reference %q is not allowed", resourceURL)
+	}
+	name, err := url.PathUnescape(strings.TrimPrefix(parsed.Path, "/"))
+	if err != nil {
+		return nil, fmt.Errorf("invalid schema reference %q: %w", resourceURL, err)
+	}
+	name = path.Clean(name)
+	if l.root == "." {
+		if strings.HasPrefix(name, "../") || name == ".." {
+			return nil, fmt.Errorf("schema reference %q escapes source root", resourceURL)
+		}
+	} else if name != l.root && !strings.HasPrefix(name, l.root+"/") {
+		return nil, fmt.Errorf("schema reference %q escapes source root", resourceURL)
+	}
+
+	rootInfo, err := fs.Stat(l.fsys, l.root)
+	if err != nil {
+		return nil, err
+	}
+	if !rootInfo.IsDir() && name != l.root {
+		return nil, fmt.Errorf("schema reference %q is outside the configured schema file", resourceURL)
+	}
+
+	raw, err := fs.ReadFile(l.fsys, name)
+	if err != nil {
+		return nil, err
+	}
+	var doc any
+	if err := json.Unmarshal(raw, &doc); err != nil {
+		return nil, err
+	}
+	if err := validateSchemaDialect(doc); err != nil {
+		return nil, err
+	}
+	return doc, nil
+}

internal/validate/standard_validator_test.go

@@ -2,8 +2,10 @@ package validate
 
 import (
 	"context"
+	"encoding/json"
 	"os"
 	"path/filepath"
+	"strconv"
 	"strings"
 	"testing"
 	"testing/fstest"
@@ -411,3 +413,138 @@ func TestFSValidatorLoadSchemaDocument(t *testing.T) {
 		t.Fatalf("unexpected schema document: %#v", doc)
 	}
 }
+
+func TestStandardValidatorJSONSchemaReferenceBoundaries(t *testing.T) {
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "child.json"), []byte(`{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "type": "string"
+}`), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	tests := []struct {
+		name       string
+		reference  string
+		wantError  string
+		writeOuter bool
+	}{
+		{name: "contained relative reference", reference: "child.json"},
+		{name: "remote reference", reference: "https://example.test/schema.json", wantError: "not a contained file reference"},
+		{name: "escaping reference", reference: "../outside.json", wantError: "escapes source root", writeOuter: true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if tc.writeOuter {
+				if err := os.WriteFile(filepath.Join(filepath.Dir(root), "outside.json"), []byte(`{"type":"string"}`), 0o644); err != nil {
+					t.Fatal(err)
+				}
+			}
+			schema := `{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$ref": ` + strconv.Quote(tc.reference) + `
+}`
+			if err := os.WriteFile(filepath.Join(root, "root.json"), []byte(schema), 0o644); err != nil {
+				t.Fatal(err)
+			}
+
+			v := NewStandardValidator(root)
+			result, err := v.Validate(context.Background(), &domain.Artifact{Body: []byte(`"value"`)}, domain.OutputContract{
+				ValidationMode: domain.ValidationJSONSchema,
+				SchemaPath:     "root.json",
+			})
+			if tc.wantError == "" {
+				if err != nil || !result.IsValid {
+					t.Fatalf("expected contained reference to validate, got result=%#v error=%v", result, err)
+				}
+				return
+			}
+			if err == nil || !strings.Contains(err.Error(), tc.wantError) {
+				t.Fatalf("expected error containing %q, got %v", tc.wantError, err)
+			}
+		})
+	}
+}
+
+func TestFSValidatorJSONSchemaReferenceBoundaries(t *testing.T) {
+	tests := []struct {
+		name      string
+		reference string
+		wantError string
+	}{
+		{name: "same document fragment", reference: "#/$defs/value"},
+		{name: "contained relative reference", reference: "child.json"},
+		{name: "remote reference", reference: "https://example.test/schema.json", wantError: "is not allowed"},
+		{name: "escaping reference", reference: "../outside.json", wantError: "escapes source root"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			rootSchema := `{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "$defs": {"value": {"type": "string"}},
+  "$ref": ` + strconv.Quote(tc.reference) + `
+}`
+			v := NewFSValidator(fstest.MapFS{
+				"schemas/root.json":  &fstest.MapFile{Data: []byte(rootSchema)},
+				"schemas/child.json": &fstest.MapFile{Data: []byte(`{"type":"string"}`)},
+				"outside.json":       &fstest.MapFile{Data: []byte(`{"type":"string"}`)},
+			}, "schemas")
+			result, err := v.Validate(context.Background(), &domain.Artifact{Body: []byte(`"value"`)}, domain.OutputContract{
+				ValidationMode: domain.ValidationJSONSchema,
+				SchemaPath:     "root.json",
+			})
+			if tc.wantError == "" {
+				if err != nil || !result.IsValid {
+					t.Fatalf("expected supported reference to validate, got result=%#v error=%v", result, err)
+				}
+				return
+			}
+			if err == nil || !strings.Contains(err.Error(), tc.wantError) {
+				t.Fatalf("expected error containing %q, got %v", tc.wantError, err)
+			}
+		})
+	}
+}
+
+func TestJSONSchemaDialectIsDraft2020(t *testing.T) {
+	tests := []struct {
+		name      string
+		dialect   string
+		wantError bool
+	}{
+		{name: "omitted uses supported default"},
+		{name: "draft 2020-12", dialect: "https://json-schema.org/draft/2020-12/schema"},
+		{name: "draft 7 rejected", dialect: "http://json-schema.org/draft-07/schema#", wantError: true},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			schema := map[string]any{"type": "object"}
+			if tc.dialect != "" {
+				schema["$schema"] = tc.dialect
+			}
+			data, err := json.Marshal(schema)
+			if err != nil {
+				t.Fatal(err)
+			}
+			v := NewFSValidator(fstest.MapFS{
+				"schema.json": &fstest.MapFile{Data: data},
+			}, ".")
+			_, err = v.Validate(context.Background(), &domain.Artifact{Body: []byte(`{}`)}, domain.OutputContract{
+				ValidationMode: domain.ValidationJSONSchema,
+				SchemaPath:     "schema.json",
+			})
+			if tc.wantError {
+				if err == nil || !strings.Contains(err.Error(), "unsupported JSON Schema dialect") {
+					t.Fatalf("expected unsupported-dialect error, got %v", err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("expected supported dialect, got %v", err)
+			}
+		})
+	}
+}

json.go

@@ -0,0 +1,162 @@
+package promptkit
+
+import (
+	"encoding/json"
+	"time"
+)
+
+// MarshalJSON implements json.Marshaler for PreparedRun. It uses RFC 3339
+// timestamps, integer duration_ms, and omits zero timing values.
+func (r PreparedRun) MarshalJSON() ([]byte, error) {
+	var startTime, endTime *time.Time
+	if !r.StartTime.IsZero() {
+		startTime = &r.StartTime
+	}
+	if !r.EndTime.IsZero() {
+		endTime = &r.EndTime
+	}
+
+	var durationMS *int64
+	if r.DurationMS != 0 {
+		durationMS = &r.DurationMS
+	}
+
+	return json.Marshal(struct {
+		PromptID             string                `json:"prompt_id"`
+		PromptVersion        string                `json:"prompt_version,omitempty"`
+		PromptHash           string                `json:"prompt_hash,omitempty"`
+		SelectedProfileID    string                `json:"selected_profile_id"`
+		SelectedBackendID    string                `json:"selected_backend_id,omitempty"`
+		EffectiveModelParams ExecutionTarget       `json:"effective_model_params"`
+		OutputContract       OutputContract        `json:"output_contract"`
+		StructuredOutput     *StructuredOutputSpec `json:"structured_output,omitempty"`
+		InputHashes          map[string]string     `json:"input_hashes,omitempty"`
+		SessionID            string                `json:"session_id,omitempty"`
+		RenderedPromptHash   string                `json:"rendered_prompt_hash"`
+		Messages             []RenderedMessage     `json:"messages"`
+		StartTime            *time.Time            `json:"start_time,omitempty"`
+		EndTime              *time.Time            `json:"end_time,omitempty"`
+		DurationMS           *int64                `json:"duration_ms,omitempty"`
+	}{
+		PromptID:             r.PromptID,
+		PromptVersion:        r.PromptVersion,
+		PromptHash:           r.PromptHash,
+		SelectedProfileID:    r.SelectedProfileID,
+		SelectedBackendID:    r.SelectedBackendID,
+		EffectiveModelParams: r.EffectiveModelParams,
+		OutputContract:       r.OutputContract,
+		StructuredOutput:     r.StructuredOutput,
+		InputHashes:          r.InputHashes,
+		SessionID:            r.SessionID,
+		RenderedPromptHash:   r.RenderedPromptHash,
+		Messages:             r.Messages,
+		StartTime:            startTime,
+		EndTime:              endTime,
+		DurationMS:           durationMS,
+	})
+}
+
+// MarshalJSON implements json.Marshaler for RunResult. It encodes Duration as
+// integer milliseconds in duration_ms and omits zero timing values.
+func (r RunResult) MarshalJSON() ([]byte, error) {
+	var startTime, endTime *time.Time
+	if !r.StartTime.IsZero() {
+		startTime = &r.StartTime
+	}
+	if !r.EndTime.IsZero() {
+		endTime = &r.EndTime
+	}
+
+	var durationMS *int64
+	if r.Duration != 0 {
+		value := r.Duration.Milliseconds()
+		durationMS = &value
+	}
+
+	return json.Marshal(runResultJSON{
+		RunID:                r.RunID,
+		Artifact:             r.Artifact,
+		RawOutput:            r.RawOutput,
+		Validation:           r.Validation,
+		PromptID:             r.PromptID,
+		PromptVersion:        r.PromptVersion,
+		PromptHash:           r.PromptHash,
+		SessionID:            r.SessionID,
+		RenderedPromptHash:   r.RenderedPromptHash,
+		SelectedProfileID:    r.SelectedProfileID,
+		SelectedBackendID:    r.SelectedBackendID,
+		ModelName:            r.ModelName,
+		Endpoint:             r.Endpoint,
+		EffectiveModelParams: r.EffectiveModelParams,
+		InputHashes:          r.InputHashes,
+		Usage:                r.Usage,
+		StartTime:            startTime,
+		EndTime:              endTime,
+		DurationMS:           durationMS,
+	})
+}
+
+// UnmarshalJSON implements json.Unmarshaler for RunResult. It decodes
+// duration_ms into Duration with millisecond precision.
+func (r *RunResult) UnmarshalJSON(data []byte) error {
+	var wire runResultJSON
+	if err := json.Unmarshal(data, &wire); err != nil {
+		return err
+	}
+
+	*r = RunResult{
+		RunID:                wire.RunID,
+		Artifact:             wire.Artifact,
+		RawOutput:            wire.RawOutput,
+		Validation:           wire.Validation,
+		PromptID:             wire.PromptID,
+		PromptVersion:        wire.PromptVersion,
+		PromptHash:           wire.PromptHash,
+		SessionID:            wire.SessionID,
+		RenderedPromptHash:   wire.RenderedPromptHash,
+		SelectedProfileID:    wire.SelectedProfileID,
+		SelectedBackendID:    wire.SelectedBackendID,
+		ModelName:            wire.ModelName,
+		Endpoint:             wire.Endpoint,
+		EffectiveModelParams: wire.EffectiveModelParams,
+		InputHashes:          wire.InputHashes,
+		Usage:                wire.Usage,
+		Duration:             time.Duration(valueOrZero(wire.DurationMS)) * time.Millisecond,
+	}
+	if wire.StartTime != nil {
+		r.StartTime = *wire.StartTime
+	}
+	if wire.EndTime != nil {
+		r.EndTime = *wire.EndTime
+	}
+	return nil
+}
+
+type runResultJSON struct {
+	RunID                string            `json:"run_id"`
+	Artifact             Artifact          `json:"artifact"`
+	RawOutput            string            `json:"raw_output"`
+	Validation           ValidationResult  `json:"validation"`
+	PromptID             string            `json:"prompt_id"`
+	PromptVersion        string            `json:"prompt_version,omitempty"`
+	PromptHash           string            `json:"prompt_hash,omitempty"`
+	SessionID            string            `json:"session_id,omitempty"`
+	RenderedPromptHash   string            `json:"rendered_prompt_hash"`
+	SelectedProfileID    string            `json:"selected_profile_id"`
+	SelectedBackendID    string            `json:"selected_backend_id,omitempty"`
+	ModelName            string            `json:"model_name"`
+	Endpoint             string            `json:"endpoint"`
+	EffectiveModelParams ExecutionTarget   `json:"effective_model_params"`
+	InputHashes          map[string]string `json:"input_hashes,omitempty"`
+	Usage                TokenUsage        `json:"usage"`
+	StartTime            *time.Time        `json:"start_time,omitempty"`
+	EndTime              *time.Time        `json:"end_time,omitempty"`
+	DurationMS           *int64            `json:"duration_ms,omitempty"`
+}
+
+func valueOrZero(value *int64) int64 {
+	if value == nil {
+		return 0
+	}
+	return *value
+}

profiles.go

@@ -7,6 +7,7 @@ import (
 	"strings"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
 	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
 )
 
@@ -15,10 +16,16 @@ import (
 //
 // It does not register global state, maintain a model catalog, or resolve
 // credentials. If APIKeyRequired is true, callers satisfy it with
-// RunRequest.APIKey. Raw API keys do not belong in profiles.
+// RunRequest.APIKey or an explicit request ExecutionTargetOverride.APIKeyEnv.
+// Raw API keys do not belong in profiles.
+//
+// The function copies the ExtraParams map itself but does not recursively copy
+// nested values. Validation and a deep copy occur when NewEngine applies a
+// WithProfiles option containing the returned Profile.
 func OpenAICompatibleProfile(cfg OpenAICompatibleProfileConfig) Profile {
 	return Profile{
 		ID:              cfg.ID,
+		BackendID:       cfg.BackendID,
 		Endpoint:        cfg.Endpoint,
 		Model:           cfg.Model,
 		Temperature:     cfg.Temperature,
@@ -75,12 +82,13 @@ func (r *memoryProfileRepository) GetProfile(_ context.Context, id string) (*dom
 }
 
 func toDomainProfile(publicProfile Profile) (domain.ExecutionProfile, error) {
-	extraParams, err := copyPublicJSONMap(publicProfile.ExtraParams)
+	extraParams, err := jsonvalue.CopyMap(publicProfile.ExtraParams)
 	if err != nil {
 		return domain.ExecutionProfile{}, err
 	}
 	prof := domain.ExecutionProfile{
 		ID:              strings.TrimSpace(publicProfile.ID),
+		BackendID:       strings.TrimSpace(publicProfile.BackendID),
 		Endpoint:        publicProfile.Endpoint,
 		Model:           publicProfile.Model,
 		Temperature:     publicProfile.Temperature,
@@ -102,8 +110,8 @@ func validatePublicProfile(prof domain.ExecutionProfile) error {
 	if strings.TrimSpace(prof.ID) == "" {
 		return errors.New("id is required")
 	}
-	if strings.TrimSpace(prof.Endpoint) == "" {
-		return errors.New("endpoint is required")
+	if strings.TrimSpace(prof.BackendID) == "" && strings.TrimSpace(prof.Endpoint) == "" {
+		return errors.New("backend or endpoint is required")
 	}
 	if strings.TrimSpace(prof.Model) == "" {
 		return errors.New("model is required")

public_contract_test.go

@@ -0,0 +1,768 @@
+package promptkit_test
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"testing/fstest"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+func TestPreparedRunJSONOmitsZeroTimingValues(t *testing.T) {
+	payload, err := json.Marshal(promptkit.PreparedRun{})
+	if err != nil {
+		t.Fatalf("marshal prepared run: %v", err)
+	}
+	for _, field := range []string{"start_time", "end_time", "duration_ms"} {
+		if strings.Contains(string(payload), `"`+field+`"`) {
+			t.Fatalf("expected zero %s to be omitted, got %s", field, payload)
+		}
+	}
+}
+
+func TestBackendIdentityJSONNamesAndOmission(t *testing.T) {
+	t.Run("execution target round trip", func(t *testing.T) {
+		value := promptkit.ExecutionTarget{BackendID: promptkit.BackendOpenRouter}
+		payload, err := json.Marshal(value)
+		if err != nil {
+			t.Fatalf("marshal execution target: %v", err)
+		}
+		var decoded promptkit.ExecutionTarget
+		if err := json.Unmarshal(payload, &decoded); err != nil {
+			t.Fatalf("unmarshal execution target: %v", err)
+		}
+		if decoded.BackendID != value.BackendID {
+			t.Fatalf("backend identity did not round trip: got %q want %q", decoded.BackendID, value.BackendID)
+		}
+	})
+
+	t.Run("prepared run round trip", func(t *testing.T) {
+		value := promptkit.PreparedRun{SelectedBackendID: promptkit.BackendOpenRouter}
+		payload, err := json.Marshal(value)
+		if err != nil {
+			t.Fatalf("marshal prepared run: %v", err)
+		}
+		var decoded promptkit.PreparedRun
+		if err := json.Unmarshal(payload, &decoded); err != nil {
+			t.Fatalf("unmarshal prepared run: %v", err)
+		}
+		if decoded.SelectedBackendID != value.SelectedBackendID {
+			t.Fatalf("backend identity did not round trip: got %q want %q", decoded.SelectedBackendID, value.SelectedBackendID)
+		}
+	})
+
+	t.Run("run result round trip", func(t *testing.T) {
+		value := promptkit.RunResult{SelectedBackendID: promptkit.BackendOpenRouter}
+		payload, err := json.Marshal(value)
+		if err != nil {
+			t.Fatalf("marshal run result: %v", err)
+		}
+		var decoded promptkit.RunResult
+		if err := json.Unmarshal(payload, &decoded); err != nil {
+			t.Fatalf("unmarshal run result: %v", err)
+		}
+		if decoded.SelectedBackendID != value.SelectedBackendID {
+			t.Fatalf("backend identity did not round trip: got %q want %q", decoded.SelectedBackendID, value.SelectedBackendID)
+		}
+	})
+
+	payload, err := json.Marshal(promptkit.ExecutionTarget{})
+	if err != nil {
+		t.Fatalf("marshal empty execution target: %v", err)
+	}
+	if strings.Contains(string(payload), `"backend_id"`) {
+		t.Fatalf("empty backend identity was not omitted: %s", payload)
+	}
+}
+
+func TestEndpointOnlyProfileOmitsBackendIdentityFromStableJSON(t *testing.T) {
+	client := &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}}
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+		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)
+	}
+
+	prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("prepare endpoint-only profile: %v", err)
+	}
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("run endpoint-only profile: %v", err)
+	}
+	if prepared.SelectedBackendID != "" ||
+		prepared.EffectiveModelParams.BackendID != "" ||
+		result.SelectedBackendID != "" ||
+		result.EffectiveModelParams.BackendID != "" {
+		t.Fatalf("endpoint-only profile acquired backend identity: prepared=%+v result=%+v", prepared, result)
+	}
+	for _, value := range []any{prepared, result} {
+		payload, err := json.Marshal(value)
+		if err != nil {
+			t.Fatalf("marshal endpoint-only value: %v", err)
+		}
+		if strings.Contains(string(payload), `"backend_id"`) || strings.Contains(string(payload), `"selected_backend_id"`) {
+			t.Fatalf("endpoint-only backend identity was not omitted: %s", payload)
+		}
+	}
+}
+
+func TestUnknownProfileBackendHasProfileLoadIdentity(t *testing.T) {
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+		promptkit.WithProfiles(promptkit.Profile{ID: "profile", BackendID: "unknown", Model: "model"}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+
+	_, err = engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if !errors.Is(err, promptkit.ErrProfileLoad) {
+		t.Fatalf("expected ErrProfileLoad, got %v", err)
+	}
+	if errors.Is(err, promptkit.ErrInvalidRequest) {
+		t.Fatalf("unknown backend should not have invalid-request identity: %v", err)
+	}
+}
+
+func TestCustomBackendFlowsThroughProfilesOverridesAndInjectedClient(t *testing.T) {
+	t.Setenv("CUSTOM_LLM_KEY", "test-key")
+	client := &fakeLLMClient{response: &promptkit.GenerateResponse{Content: "ok"}}
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "backend-profile", "message"), "."),
+		promptkit.WithBackend(promptkit.Backend{
+			ID:        "  custom  ",
+			Endpoint:  "  http://backend.example/v1  ",
+			APIKeyEnv: "  CUSTOM_LLM_KEY  ",
+			ExtraParams: map[string]any{
+				"provider": "custom",
+			},
+		}),
+		promptkit.WithProfiles(
+			promptkit.Profile{ID: "backend-profile", BackendID: "custom", Model: "backend-model"},
+			promptkit.Profile{ID: "profile-endpoint", BackendID: "custom", Endpoint: "http://profile.example/v1", Model: "profile-model"},
+			promptkit.Profile{ID: "blank-profile-endpoint", BackendID: "custom", Endpoint: " \t ", Model: "profile-model"},
+		),
+		promptkit.WithLLMClient(client),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("run with custom backend: %v", err)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("expected one injected-client request, got %d", len(client.requests))
+	}
+	target := client.requests[0].Target
+	if target.BackendID != "custom" ||
+		target.Endpoint != "http://backend.example/v1" ||
+		target.APIKeyEnv != "CUSTOM_LLM_KEY" ||
+		target.Model != "backend-model" ||
+		target.ExtraParams["provider"] != "custom" ||
+		result.SelectedBackendID != "custom" {
+		t.Fatalf("unexpected custom backend settings: target=%+v result_backend=%q", target, result.SelectedBackendID)
+	}
+
+	prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
+		PromptID: "prompt", ProfileID: "profile-endpoint",
+	})
+	if err != nil {
+		t.Fatalf("prepare profile endpoint override: %v", err)
+	}
+	if prepared.SelectedBackendID != "custom" || prepared.EffectiveModelParams.Endpoint != "http://profile.example/v1" {
+		t.Fatalf("profile endpoint override changed backend identity: %+v", prepared)
+	}
+
+	prepared, err = engine.Prepare(context.Background(), promptkit.RunRequest{
+		PromptID: "prompt", ProfileID: "blank-profile-endpoint",
+	})
+	if err != nil {
+		t.Fatalf("prepare blank profile endpoint: %v", err)
+	}
+	if prepared.SelectedBackendID != "custom" || prepared.EffectiveModelParams.Endpoint != "http://backend.example/v1" {
+		t.Fatalf("blank profile endpoint did not inherit backend endpoint: %+v", prepared)
+	}
+
+	prepared, err = engine.Prepare(context.Background(), promptkit.RunRequest{
+		PromptID: "prompt",
+		Execution: &promptkit.ExecutionTargetOverride{
+			Endpoint: "http://request.example/v1",
+		},
+	})
+	if err != nil {
+		t.Fatalf("prepare request endpoint override: %v", err)
+	}
+	if prepared.SelectedBackendID != "custom" || prepared.EffectiveModelParams.Endpoint != "http://request.example/v1" {
+		t.Fatalf("request endpoint override changed backend identity: %+v", prepared)
+	}
+}
+
+func TestCustomBackendSupportsFileProfileAndBothSelectionPaths(t *testing.T) {
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "file-profile", "message"), "."),
+		promptkit.WithProfileFS(fstest.MapFS{
+			"profile.yaml": &fstest.MapFile{Data: []byte(`id: file-profile
+backend: file-backend
+endpoint: "  "
+model: file-model
+`)},
+		}, "."),
+		promptkit.WithBackend(promptkit.Backend{
+			ID:       "file-backend",
+			Endpoint: "http://file-backend.example/v1",
+		}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+
+	for _, request := range []promptkit.RunRequest{
+		{PromptID: "prompt"},
+		{PromptID: "prompt", ProfileID: "file-profile"},
+	} {
+		prepared, err := engine.Prepare(context.Background(), request)
+		if err != nil {
+			t.Fatalf("prepare file profile: %v", err)
+		}
+		if prepared.SelectedBackendID != "file-backend" ||
+			prepared.EffectiveModelParams.Endpoint != "http://file-backend.example/v1" {
+			t.Fatalf("unexpected file-profile backend resolution: %+v", prepared)
+		}
+	}
+}
+
+func TestBackendOptionsAccumulateAndRegistrationsAreEngineLocal(t *testing.T) {
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "first-profile", "message"), "."),
+		promptkit.WithBackend(promptkit.Backend{ID: "first", Endpoint: "http://first.example/v1"}),
+		promptkit.WithBackend(promptkit.Backend{ID: "second", Endpoint: "http://second.example/v1"}),
+		promptkit.WithProfiles(
+			promptkit.Profile{ID: "first-profile", BackendID: "first", Model: "model"},
+			promptkit.Profile{ID: "second-profile", BackendID: "second", Model: "model"},
+		),
+	)
+	if err != nil {
+		t.Fatalf("construct engine with accumulated registrations: %v", err)
+	}
+	for profileID, wantEndpoint := range map[string]string{
+		"first-profile":  "http://first.example/v1",
+		"second-profile": "http://second.example/v1",
+	} {
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
+			PromptID: "prompt", ProfileID: profileID,
+		})
+		if err != nil {
+			t.Fatalf("prepare %s: %v", profileID, err)
+		}
+		if prepared.EffectiveModelParams.Endpoint != wantEndpoint {
+			t.Fatalf("profile %s endpoint=%q, want %q", profileID, prepared.EffectiveModelParams.Endpoint, wantEndpoint)
+		}
+	}
+
+	newEngine := func(endpoint string) *promptkit.Engine {
+		t.Helper()
+		value, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+			promptkit.WithBackend(promptkit.Backend{ID: "same-id", Endpoint: endpoint}),
+			promptkit.WithProfiles(promptkit.Profile{ID: "profile", BackendID: "same-id", Model: "model"}),
+		)
+		if err != nil {
+			t.Fatalf("construct isolated engine: %v", err)
+		}
+		return value
+	}
+	firstEngine := newEngine("http://one.example/v1")
+	secondEngine := newEngine("http://two.example/v1")
+	for engine, wantEndpoint := range map[*promptkit.Engine]string{
+		firstEngine:  "http://one.example/v1",
+		secondEngine: "http://two.example/v1",
+	} {
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+		if err != nil {
+			t.Fatalf("prepare isolated engine: %v", err)
+		}
+		if prepared.EffectiveModelParams.Endpoint != wantEndpoint {
+			t.Fatalf("isolated engine endpoint=%q, want %q", prepared.EffectiveModelParams.Endpoint, wantEndpoint)
+		}
+	}
+}
+
+func TestWithBackendCopiesQueueCapacity(t *testing.T) {
+	queueCapacity := 4
+	option := promptkit.WithBackend(promptkit.Backend{
+		ID:               "custom",
+		Endpoint:         "http://custom.example/v1",
+		ConcurrencyLimit: 1,
+		QueueCapacity:    &queueCapacity,
+	})
+	queueCapacity = -1
+
+	_, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+		option,
+		promptkit.WithProfiles(promptkit.Profile{
+			ID: "profile", BackendID: "custom", Model: "model",
+		}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine after mutating queue pointer: %v", err)
+	}
+}
+
+func TestBackendRegistrationRejectsInvalidAndDuplicateDefinitions(t *testing.T) {
+	cycle := map[string]any{}
+	cycle["self"] = cycle
+	tests := []struct {
+		name     string
+		backends []promptkit.Backend
+	}{
+		{name: "blank id", backends: []promptkit.Backend{{Endpoint: "http://example.test/v1"}}},
+		{name: "invalid endpoint", backends: []promptkit.Backend{{ID: "custom", Endpoint: "ftp://example.test/v1"}}},
+		{name: "invalid environment", backends: []promptkit.Backend{{ID: "custom", Endpoint: "http://example.test/v1", APIKeyEnv: "BAD-NAME"}}},
+		{name: "reserved extra parameter", backends: []promptkit.Backend{{ID: "custom", Endpoint: "http://example.test/v1", ExtraParams: map[string]any{"model": "override"}}}},
+		{name: "cyclic extra parameter", backends: []promptkit.Backend{{ID: "custom", Endpoint: "http://example.test/v1", ExtraParams: cycle}}},
+		{name: "malformed JSON number", backends: []promptkit.Backend{{ID: "custom", Endpoint: "http://example.test/v1", ExtraParams: map[string]any{"value": json.Number("01")}}}},
+		{name: "duplicate consumer id", backends: []promptkit.Backend{
+			{ID: " custom ", Endpoint: "http://one.example/v1"},
+			{ID: "custom", Endpoint: "http://two.example/v1"},
+		}},
+		{name: "reserved built-in id", backends: []promptkit.Backend{{
+			ID: promptkit.BackendOpenRouter, Endpoint: "http://replacement.example/v1",
+		}}},
+	}
+
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			options := []promptkit.Option{
+				promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+			}
+			for _, backend := range tt.backends {
+				options = append(options, promptkit.WithBackend(backend))
+			}
+			_, err := promptkit.NewEngine(promptkit.Config{}, options...)
+			if !errors.Is(err, promptkit.ErrInvalidConfig) {
+				t.Fatalf("expected ErrInvalidConfig, got %v", err)
+			}
+		})
+	}
+}
+
+func TestBackendExtraParamsAreDeeplyCopiedAtConstructionAndLookup(t *testing.T) {
+	nested := map[string]any{"value": "original"}
+	extraParams := map[string]any{"nested": nested}
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+		promptkit.WithBackend(promptkit.Backend{
+			ID: "custom", Endpoint: "http://example.test/v1", ExtraParams: extraParams,
+		}),
+		promptkit.WithProfiles(promptkit.Profile{ID: "profile", BackendID: "custom", Model: "model"}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+	nested["value"] = "mutated input"
+	extraParams["later"] = true
+
+	prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("first prepare: %v", err)
+	}
+	gotNested := prepared.EffectiveModelParams.ExtraParams["nested"].(map[string]any)
+	if gotNested["value"] != "original" || prepared.EffectiveModelParams.ExtraParams["later"] != nil {
+		t.Fatalf("backend retained caller mutations: %#v", prepared.EffectiveModelParams.ExtraParams)
+	}
+	gotNested["value"] = "mutated lookup"
+
+	prepared, err = engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+	if err != nil {
+		t.Fatalf("second prepare: %v", err)
+	}
+	gotNested = prepared.EffectiveModelParams.ExtraParams["nested"].(map[string]any)
+	if gotNested["value"] != "original" {
+		t.Fatalf("backend retained lookup mutation: %#v", prepared.EffectiveModelParams.ExtraParams)
+	}
+}
+
+func TestPreparedRunJSONTimingRoundTrips(t *testing.T) {
+	start := time.Date(2026, time.July, 29, 12, 0, 0, 0, time.UTC)
+	prepared := promptkit.PreparedRun{
+		PromptID:   "prompt",
+		StartTime:  start,
+		EndTime:    start.Add(1250 * time.Millisecond),
+		DurationMS: 1250,
+	}
+
+	payload, err := json.Marshal(prepared)
+	if err != nil {
+		t.Fatalf("marshal prepared run: %v", err)
+	}
+	var decoded promptkit.PreparedRun
+	if err := json.Unmarshal(payload, &decoded); err != nil {
+		t.Fatalf("unmarshal prepared run: %v", err)
+	}
+	if decoded.DurationMS != prepared.DurationMS ||
+		!decoded.StartTime.Equal(prepared.StartTime) ||
+		!decoded.EndTime.Equal(prepared.EndTime) {
+		t.Fatalf("timing values did not round trip: got %#v, want %#v", decoded, prepared)
+	}
+}
+
+func TestRunResultJSONUsesMillisecondsAndRoundTrips(t *testing.T) {
+	start := time.Date(2026, time.July, 29, 12, 0, 0, 0, time.UTC)
+	result := promptkit.RunResult{
+		RunID:     "opaque-run-id",
+		Artifact:  promptkit.Artifact{Name: "output", ContentType: "text/plain", Body: []byte("ok")},
+		SessionID: "session-123",
+		StartTime: start,
+		EndTime:   start.Add(1500 * time.Millisecond),
+		Duration:  1500 * time.Millisecond,
+	}
+
+	payload, err := json.Marshal(result)
+	if err != nil {
+		t.Fatalf("marshal run result: %v", err)
+	}
+	var object map[string]any
+	if err := json.Unmarshal(payload, &object); err != nil {
+		t.Fatalf("decode run result JSON: %v", err)
+	}
+	if got := object["duration_ms"]; got != float64(1500) {
+		t.Fatalf("expected duration_ms=1500, got %#v in %s", got, payload)
+	}
+	if _, exists := object["duration"]; exists {
+		t.Fatalf("unexpected nanosecond duration field in %s", payload)
+	}
+	if got := object["session_id"]; got != result.SessionID {
+		t.Fatalf("expected session_id=%q, got %#v in %s", result.SessionID, got, payload)
+	}
+	artifact, ok := object["artifact"].(map[string]any)
+	if !ok || artifact["content_type"] != "text/plain" {
+		t.Fatalf("expected stable artifact JSON fields, got %#v", object["artifact"])
+	}
+
+	var decoded promptkit.RunResult
+	if err := json.Unmarshal(payload, &decoded); err != nil {
+		t.Fatalf("unmarshal run result: %v", err)
+	}
+	if decoded.SessionID != result.SessionID ||
+		decoded.Duration != result.Duration ||
+		!decoded.StartTime.Equal(result.StartTime) ||
+		!decoded.EndTime.Equal(result.EndTime) {
+		t.Fatalf("timing values did not round trip: got %#v, want %#v", decoded, result)
+	}
+
+	payload, err = json.Marshal(promptkit.RunResult{})
+	if err != nil {
+		t.Fatalf("marshal zero run result: %v", err)
+	}
+	for _, field := range []string{"session_id", "start_time", "end_time", "duration_ms"} {
+		if strings.Contains(string(payload), `"`+field+`"`) {
+			t.Fatalf("expected zero %s to be omitted, got %s", field, payload)
+		}
+	}
+	var decodedEmpty promptkit.RunResult
+	if err := json.Unmarshal(payload, &decodedEmpty); err != nil {
+		t.Fatalf("unmarshal run result without session_id: %v", err)
+	}
+	if decodedEmpty.SessionID != "" {
+		t.Fatalf("expected absent session_id to decode empty, got %q", decodedEmpty.SessionID)
+	}
+}
+
+func TestEngineValidationIsSinglePass(t *testing.T) {
+	client := &fakeLLMClient{
+		response: &promptkit.GenerateResponse{Content: "not-json"},
+	}
+	engine := newContractEngineWithOptions(t, frameworkSchemaDir, promptkit.WithLLMClient(client))
+
+	result, err := engine.Run(context.Background(), promptkit.RunRequest{
+		PromptID: frameworkMarkdownSummaryPromptID,
+		Inputs: map[string]promptkit.ArtifactRef{
+			"transcript": promptkit.Inline("Rin opens the gate."),
+			"glossary":   promptkit.Inline("gate: A guarded passage."),
+		},
+		Validation: &promptkit.OutputContract{
+			Format:         promptkit.FormatJSON,
+			ValidationMode: promptkit.ValidationJSON,
+			RepairAttempts: 3,
+		},
+	})
+	if err != nil {
+		t.Fatalf("run with failed content validation: %v", err)
+	}
+	if result.Validation.Status != promptkit.ValidationFailed ||
+		result.Validation.RepairAttempts != 0 {
+		t.Fatalf("expected failed single-pass validation, got %#v", result.Validation)
+	}
+	if len(client.requests) != 1 {
+		t.Fatalf("expected one model generation, got %d", len(client.requests))
+	}
+}
+
+func TestRepeatedOptionsUseLastValueInEachCategory(t *testing.T) {
+	profile := promptkit.Profile{ID: "profile", Endpoint: "http://example.test/v1", Model: "model"}
+
+	t.Run("prompt source", func(t *testing.T) {
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractPromptFS("first", "profile", "first"), "."),
+			promptkit.WithPromptFS(contractPromptFS("second", "profile", "second"), "."),
+			promptkit.WithProfiles(profile),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "second"})
+		if err != nil {
+			t.Fatalf("prepare from last prompt source: %v", err)
+		}
+		if prepared.Messages[0].Content != "second" {
+			t.Fatalf("expected last prompt source, got %#v", prepared.Messages)
+		}
+	})
+
+	t.Run("profile source", func(t *testing.T) {
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+			promptkit.WithProfileFS(contractProfileFS("profile", "first-model"), "."),
+			promptkit.WithProfileFS(contractProfileFS("profile", "second-model"), "."),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+		if err != nil {
+			t.Fatalf("prepare from last profile source: %v", err)
+		}
+		if prepared.EffectiveModelParams.Model != "second-model" {
+			t.Fatalf("expected last profile source, got %q", prepared.EffectiveModelParams.Model)
+		}
+	})
+
+	t.Run("in-memory profiles", func(t *testing.T) {
+		first := profile
+		first.Model = "first-model"
+		second := profile
+		second.Model = "second-model"
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+			promptkit.WithProfiles(first),
+			promptkit.WithProfiles(second),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "prompt"})
+		if err != nil {
+			t.Fatalf("prepare from last in-memory profile option: %v", err)
+		}
+		if prepared.EffectiveModelParams.Model != "second-model" {
+			t.Fatalf("expected last in-memory profiles, got %q", prepared.EffectiveModelParams.Model)
+		}
+	})
+
+	t.Run("schema source", func(t *testing.T) {
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractSchemaPromptFS(), "."),
+			promptkit.WithProfiles(profile),
+			promptkit.WithSchemaFS(contractSchemaFS("first"), "."),
+			promptkit.WithSchemaFS(contractSchemaFS("second"), "."),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{PromptID: "schema-prompt"})
+		if err != nil {
+			t.Fatalf("prepare from last schema source: %v", err)
+		}
+		schema := prepared.StructuredOutput.JSONSchema.Schema.(map[string]any)
+		if schema["title"] != "second" {
+			t.Fatalf("expected last schema source, got %#v", schema)
+		}
+	})
+
+	t.Run("model client", func(t *testing.T) {
+		var firstCalls, secondCalls atomic.Int64
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+			promptkit.WithProfiles(profile),
+			promptkit.WithLLMClient(countingLLMClient{calls: &firstCalls}),
+			promptkit.WithLLMClient(countingLLMClient{calls: &secondCalls}),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		if _, err := engine.Run(context.Background(), promptkit.RunRequest{PromptID: "prompt"}); err != nil {
+			t.Fatalf("run with last model client: %v", err)
+		}
+		if firstCalls.Load() != 0 || secondCalls.Load() != 1 {
+			t.Fatalf("expected only last client call, got first=%d second=%d", firstCalls.Load(), secondCalls.Load())
+		}
+	})
+
+	t.Run("artifact reader", func(t *testing.T) {
+		engine, err := promptkit.NewEngine(promptkit.Config{},
+			promptkit.WithPromptFS(contractInputPromptFS(), "."),
+			promptkit.WithProfiles(profile),
+			promptkit.WithArtifactReader(fixedArtifactReader("first")),
+			promptkit.WithArtifactReader(fixedArtifactReader("second")),
+		)
+		if err != nil {
+			t.Fatalf("construct engine: %v", err)
+		}
+		prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
+			PromptID: "input-prompt",
+			Inputs:   map[string]promptkit.ArtifactRef{"input": promptkit.Inline("ignored")},
+		})
+		if err != nil {
+			t.Fatalf("prepare with last artifact reader: %v", err)
+		}
+		if prepared.Messages[0].Content != "second" {
+			t.Fatalf("expected last artifact reader, got %#v", prepared.Messages)
+		}
+	})
+}
+
+func TestEngineSupportsConcurrentPrepareAndRun(t *testing.T) {
+	engine, err := promptkit.NewEngine(promptkit.Config{},
+		promptkit.WithPromptFS(contractPromptFS("prompt", "profile", "message"), "."),
+		promptkit.WithBackend(promptkit.Backend{
+			ID:       "concurrent",
+			Endpoint: "http://example.test/v1",
+		}),
+		promptkit.WithProfiles(promptkit.Profile{
+			ID:        "profile",
+			BackendID: "concurrent",
+			Model:     "model",
+		}),
+		promptkit.WithLLMClient(countingLLMClient{}),
+	)
+	if err != nil {
+		t.Fatalf("construct engine: %v", err)
+	}
+
+	const calls = 40
+	errs := make(chan error, calls)
+	var wg sync.WaitGroup
+	for i := 0; i < calls; i++ {
+		wg.Add(1)
+		go func(run bool) {
+			defer wg.Done()
+			request := promptkit.RunRequest{PromptID: "prompt"}
+			if run {
+				_, err := engine.Run(context.Background(), request)
+				errs <- err
+				return
+			}
+			_, err := engine.Prepare(context.Background(), request)
+			errs <- err
+		}(i%2 == 0)
+	}
+	wg.Wait()
+	close(errs)
+	for err := range errs {
+		if err != nil {
+			t.Fatalf("concurrent call failed: %v", err)
+		}
+	}
+}
+
+type countingLLMClient struct {
+	calls *atomic.Int64
+}
+
+func (c countingLLMClient) Generate(context.Context, promptkit.GenerateRequest) (*promptkit.GenerateResponse, error) {
+	if c.calls != nil {
+		c.calls.Add(1)
+	}
+	return &promptkit.GenerateResponse{Content: "ok"}, nil
+}
+
+type fixedArtifactReader string
+
+func (r fixedArtifactReader) Read(context.Context, promptkit.ArtifactRef) (*promptkit.Artifact, error) {
+	return &promptkit.Artifact{Body: []byte(r)}, nil
+}
+
+func contractPromptFS(id, profileID, message string) fstest.MapFS {
+	return fstest.MapFS{
+		"prompt.yaml": &fstest.MapFile{Data: []byte(fmt.Sprintf(`id: %s
+version: "1"
+default_profile: %s
+messages:
+  - role: user
+    content: %q
+output:
+  format: text
+  validation_mode: none
+`, id, profileID, message))},
+	}
+}
+
+func contractInputPromptFS() fstest.MapFS {
+	return fstest.MapFS{
+		"prompt.yaml": &fstest.MapFile{Data: []byte(`id: input-prompt
+version: "1"
+default_profile: profile
+inputs:
+  - name: input
+    required: true
+messages:
+  - role: user
+    content: '{{input "input"}}'
+output:
+  format: text
+  validation_mode: none
+`)},
+	}
+}
+
+func contractProfileFS(id, model string) fstest.MapFS {
+	return fstest.MapFS{
+		"profile.yaml": &fstest.MapFile{Data: []byte(fmt.Sprintf(`id: %s
+endpoint: http://example.test/v1
+model: %s
+`, id, model))},
+	}
+}
+
+func contractSchemaPromptFS() fstest.MapFS {
+	return fstest.MapFS{
+		"prompt.yaml": &fstest.MapFile{Data: []byte(`id: schema-prompt
+version: "1"
+default_profile: profile
+messages:
+  - role: user
+    content: message
+output:
+  format: json
+  validation_mode: json_schema
+  schema_path: schema.json
+`)},
+	}
+}
+
+func contractSchemaFS(title string) fstest.MapFS {
+	return fstest.MapFS{
+		"schema.json": &fstest.MapFile{Data: []byte(fmt.Sprintf(`{
+  "$schema": "https://json-schema.org/draft/2020-12/schema",
+  "title": %q,
+  "type": "object"
+}`, title))},
+	}
+}

types.go

@@ -5,302 +5,630 @@ import (
 	"time"
 )
 
-// ArtifactRefType defines how an artifact is referenced.
+// ArtifactRefType identifies how an [ArtifactRef] supplies content.
 type ArtifactRefType string
 
 const (
+	// ArtifactRefInline selects ArtifactRef.Body as the content.
 	ArtifactRefInline ArtifactRefType = "inline"
-	ArtifactRefFile   ArtifactRefType = "file"
+	// ArtifactRefFile selects the filesystem path in ArtifactRef.URI.
+	ArtifactRefFile ArtifactRefType = "file"
 )
 
-// OutputFormat defines the desired output format.
+// OutputFormat identifies the media format of generated output.
+// OutputFormat has a stable JSON string representation.
 type OutputFormat string
 
 const (
-	FormatText     OutputFormat = "text"
+	// FormatText identifies plain-text output.
+	FormatText OutputFormat = "text"
+	// FormatMarkdown identifies Markdown output.
 	FormatMarkdown OutputFormat = "markdown"
-	FormatJSON     OutputFormat = "json"
+	// FormatJSON identifies JSON output.
+	FormatJSON OutputFormat = "json"
 )
 
-// ValidationMode defines the output validation strategy.
+// ValidationMode identifies how generated output is checked.
+// ValidationMode has a stable JSON string representation.
 type ValidationMode string
 
 const (
-	ValidationNone       ValidationMode = "none"
-	ValidationBasic      ValidationMode = "basic"
-	ValidationJSON       ValidationMode = "json"
+	// ValidationNone skips content validation.
+	ValidationNone ValidationMode = "none"
+	// ValidationBasic requires non-empty output.
+	ValidationBasic ValidationMode = "basic"
+	// ValidationJSON requires syntactically valid JSON.
+	ValidationJSON ValidationMode = "json"
+	// ValidationJSONSchema requires JSON that satisfies OutputContract.SchemaPath.
 	ValidationJSONSchema ValidationMode = "json_schema"
 )
 
-// ValidationStatus defines the result of a validation check.
+// ValidationStatus identifies the completed state of an output check.
+// ValidationStatus has a stable JSON string representation.
 type ValidationStatus string
 
 const (
-	ValidationPassed  ValidationStatus = "passed"
-	ValidationFailed  ValidationStatus = "failed"
+	// ValidationPassed means the generated output satisfied its contract.
+	ValidationPassed ValidationStatus = "passed"
+	// ValidationFailed means validation completed and rejected the generated
+	// output. Engine.Run returns this status in a result, not as an error.
+	ValidationFailed ValidationStatus = "failed"
+	// ValidationSkipped means ValidationNone selected no content check.
 	ValidationSkipped ValidationStatus = "skipped"
 )
 
-// CacheControlType defines provider cache behavior for prompt content.
+// CacheControlType identifies provider cache behavior for prompt content.
+// CacheControlType has a stable JSON string representation.
 type CacheControlType string
 
 const (
+	// CacheControlEphemeral requests provider-defined ephemeral caching.
 	CacheControlEphemeral CacheControlType = "ephemeral"
 )
 
 // StructuredOutputType identifies provider-level structured output modes.
+// StructuredOutputType has a stable JSON string representation.
 type StructuredOutputType string
 
 const (
+	// StructuredOutputJSONSchema supplies JSON Schema response constraints.
 	StructuredOutputJSONSchema StructuredOutputType = "json_schema"
 )
 
-// RunRequest represents a request to prepare or run a single prompt.
+// RunRequest selects one prompt execution. It has no stable JSON
+// representation.
+//
+// Prepare and Run copy the request's maps, pointers, and nested
+// JSON-compatible values before using them. The caller may mutate the request
+// after either method returns.
 type RunRequest struct {
-	PromptID      string
+	// PromptID is the required non-empty prompt identifier.
+	PromptID string
+	// PromptVersion optionally selects one version of PromptID. When empty, the
+	// prompt source must contain exactly one matching version.
 	PromptVersion string
-	ProfileID     string
-	APIKey        string `json:"-"`
-	Inputs        map[string]ArtifactRef
-	Vars          map[string]string
-	Execution     *ExecutionTargetOverride
-	Validation    *OutputContract
-	Metadata      map[string]string
+	// ProfileID selects an execution profile. When empty, the prompt's default
+	// profile is used; if both are empty, the error matches ErrProfileRequired
+	// and ErrInvalidRequest.
+	ProfileID string
+	// SessionID optionally supplies a direct per-run session identifier. A
+	// nonblank value is trimmed and overrides the prompt definition's
+	// session_id template. A blank value supplies no direct override. The
+	// maximum is 256 Unicode code points after trimming. A direct value is
+	// opaque consumer metadata, not a credential, and may be exposed in
+	// prepared values, results, collaborator requests, provider requests, and
+	// provider observability. Callers should use stable, non-sensitive
+	// identifiers. An overlong direct value makes Prepare or Run return an
+	// error matching ErrInvalidRequest.
+	SessionID string
+	// APIKey is a request-scoped direct credential. It takes precedence over
+	// APIKeyEnv, is passed to the selected LLMClient, and is never included in
+	// prepared values, results, hashes, JSON, String, or GoString output.
+	APIKey string `json:"-"`
+	// Inputs maps prompt input names to references. A nil or empty map is valid
+	// only when the selected prompt and its templates require no inputs.
+	Inputs map[string]ArtifactRef
+	// Vars supplies Go-template data for messages and the session ID. Nil and
+	// empty maps are equivalent.
+	Vars map[string]string
+	// Execution optionally overrides individual execution settings. Nil uses
+	// the selected profile over its backend, when any, and framework defaults.
+	Execution *ExecutionTargetOverride
+	// Validation optionally replaces the prompt's complete output contract. It
+	// does not merge individual fields. Nil uses the prompt contract.
+	Validation *OutputContract
 }
 
 // PreparedRun contains prepared prompt execution state. It does not include
 // resolved API key values, model output, validation results, or internal target
-// presence metadata.
+// presence metadata. PreparedRun has a stable JSON representation.
+//
+// All maps, slices, pointers, and schema values are caller-owned copies. JSON
+// timestamps use RFC 3339 and zero timing values are omitted. Hash formats are
+// opaque.
 type PreparedRun struct {
-	PromptID             string                `json:"prompt_id"`
-	PromptVersion        string                `json:"prompt_version,omitempty"`
-	PromptHash           string                `json:"prompt_hash,omitempty"`
-	SelectedProfileID    string                `json:"selected_profile_id"`
-	EffectiveModelParams ExecutionTarget       `json:"effective_model_params"`
-	OutputContract       OutputContract        `json:"output_contract"`
-	StructuredOutput     *StructuredOutputSpec `json:"structured_output,omitempty"`
-	InputHashes          map[string]string     `json:"input_hashes,omitempty"`
-	SessionID            string                `json:"session_id,omitempty"`
-	RenderedPromptHash   string                `json:"rendered_prompt_hash"`
-	Messages             []RenderedMessage     `json:"messages"`
-	StartTime            time.Time             `json:"start_time,omitempty"`
-	EndTime              time.Time             `json:"end_time,omitempty"`
-	DurationMS           int64                 `json:"duration_ms,omitempty"`
+	// PromptID is the selected prompt identifier.
+	PromptID string `json:"prompt_id"`
+	// PromptVersion is the selected prompt version.
+	PromptVersion string `json:"prompt_version,omitempty"`
+	// PromptHash is an opaque equality value for the selected definition.
+	PromptHash string `json:"prompt_hash,omitempty"`
+	// SelectedProfileID is the explicit request profile or prompt default that
+	// supplied execution settings.
+	SelectedProfileID string `json:"selected_profile_id"`
+	// SelectedBackendID equals EffectiveModelParams.BackendID. It is empty for
+	// an endpoint-only profile.
+	SelectedBackendID string `json:"selected_backend_id,omitempty"`
+	// EffectiveModelParams contains framework defaults overlaid by the selected
+	// backend, profile, and then request overrides. It excludes resolved API-key
+	// values.
+	EffectiveModelParams ExecutionTarget `json:"effective_model_params"`
+	// OutputContract is the complete effective output contract.
+	OutputContract OutputContract `json:"output_contract"`
+	// StructuredOutput is non-nil for JSON Schema validation and contains the
+	// provider-facing response constraint passed to an LLM client.
+	StructuredOutput *StructuredOutputSpec `json:"structured_output,omitempty"`
+	// InputHashes maps every supplied input name to its opaque artifact hash.
+	InputHashes map[string]string `json:"input_hashes,omitempty"`
+	// SessionID is the effective direct or rendered session identifier, if any.
+	SessionID string `json:"session_id,omitempty"`
+	// RenderedPromptHash is an opaque equality value for SessionID and Messages.
+	RenderedPromptHash string `json:"rendered_prompt_hash"`
+	// Messages are the rendered messages that Run passes to the LLM client.
+	Messages []RenderedMessage `json:"messages"`
+	// StartTime is the UTC time at which preparation began.
+	StartTime time.Time `json:"start_time,omitempty"`
+	// EndTime is the UTC time at which preparation completed.
+	EndTime time.Time `json:"end_time,omitempty"`
+	// DurationMS is preparation elapsed time in integer milliseconds. JSON uses
+	// duration_ms and omits a zero value.
+	DurationMS int64 `json:"duration_ms,omitempty"`
 }
 
 // RunResult contains generated output, validation state, and run metadata.
+// RunResult has a stable JSON representation and round-trips its Duration
+// through the duration_ms JSON field.
+//
+// All maps, slices, and nested values are caller-owned copies. JSON timestamps
+// use RFC 3339 and zero timing values are omitted. Run IDs and hash formats are
+// opaque.
 type RunResult struct {
-	RunID                string            `json:"run_id"`
-	Artifact             Artifact          `json:"artifact"`
-	RawOutput            string            `json:"raw_output"`
-	Validation           ValidationResult  `json:"validation"`
-	PromptID             string            `json:"prompt_id"`
-	PromptVersion        string            `json:"prompt_version,omitempty"`
-	PromptHash           string            `json:"prompt_hash,omitempty"`
-	RenderedPromptHash   string            `json:"rendered_prompt_hash"`
-	SelectedProfileID    string            `json:"selected_profile_id"`
-	ModelName            string            `json:"model_name"`
-	Endpoint             string            `json:"endpoint"`
-	EffectiveModelParams ExecutionTarget   `json:"effective_model_params"`
-	InputHashes          map[string]string `json:"input_hashes,omitempty"`
-	Usage                TokenUsage        `json:"usage"`
-	StartTime            time.Time         `json:"start_time,omitempty"`
-	EndTime              time.Time         `json:"end_time,omitempty"`
-	Duration             time.Duration     `json:"duration,omitempty"`
+	// RunID is an opaque identifier for this invocation.
+	RunID string `json:"run_id"`
+	// Artifact contains the generated output and derived metadata.
+	Artifact Artifact `json:"artifact"`
+	// RawOutput is the exact generated content before artifact classification
+	// and validation.
+	RawOutput string `json:"raw_output"`
+	// Validation records the completed content check.
+	Validation ValidationResult `json:"validation"`
+	// PromptID is the selected prompt identifier.
+	PromptID string `json:"prompt_id"`
+	// PromptVersion is the selected prompt version.
+	PromptVersion string `json:"prompt_version,omitempty"`
+	// PromptHash is the same opaque definition equality value exposed by
+	// PreparedRun.
+	PromptHash string `json:"prompt_hash,omitempty"`
+	// SessionID is the effective direct or rendered session identifier, if any.
+	// JSON omits an empty value.
+	SessionID string `json:"session_id,omitempty"`
+	// RenderedPromptHash is the same opaque rendered-prompt equality value
+	// computed during preparation.
+	RenderedPromptHash string `json:"rendered_prompt_hash"`
+	// SelectedProfileID identifies the profile used for execution.
+	SelectedProfileID string `json:"selected_profile_id"`
+	// SelectedBackendID equals EffectiveModelParams.BackendID. It is empty for
+	// an endpoint-only profile.
+	SelectedBackendID string `json:"selected_backend_id,omitempty"`
+	// ModelName is the effective model name and equals
+	// EffectiveModelParams.Model.
+	ModelName string `json:"model_name"`
+	// Endpoint is the effective base endpoint and equals
+	// EffectiveModelParams.Endpoint.
+	Endpoint string `json:"endpoint"`
+	// EffectiveModelParams contains the settings supplied to the LLM client,
+	// excluding resolved API-key values.
+	EffectiveModelParams ExecutionTarget `json:"effective_model_params"`
+	// InputHashes are the opaque input equality values computed during
+	// preparation.
+	InputHashes map[string]string `json:"input_hashes,omitempty"`
+	// Usage is the token accounting reported by the LLM client.
+	Usage TokenUsage `json:"usage"`
+	// StartTime is the UTC time immediately before preparation begins.
+	StartTime time.Time `json:"start_time,omitempty"`
+	// EndTime is the UTC time after generation and validation complete.
+	EndTime time.Time `json:"end_time,omitempty"`
+	// Duration covers preparation, generation, and validation. JSON represents
+	// it as integer milliseconds in duration_ms and omits a zero value.
+	Duration time.Duration `json:"-"`
 }
 
-// ArtifactRef represents a reference to prompt input content.
+// ArtifactRef identifies prompt input content. It has no stable JSON
+// representation. Prefer [File], [Inline], or [InlineWithURI] to construct one.
 type ArtifactRef struct {
+	// Type must be ArtifactRefInline or ArtifactRefFile.
 	Type ArtifactRefType
-	URI  string
+	// URI is the file path for ArtifactRefFile and optional provenance metadata
+	// for ArtifactRefInline.
+	URI string
+	// Body is the content for ArtifactRefInline and is ignored for
+	// ArtifactRefFile.
 	Body string
 }
 
-// Artifact represents loaded artifact content.
+// Artifact represents loaded or generated content and has a stable JSON
+// representation. Body uses encoding/json's base64 representation for []byte.
 type Artifact struct {
-	Name        string
-	ContentType string
-	Body        []byte
-	URI         string
-	Size        int64
-	Hash        string
+	// Name is artifact metadata. During input preparation the engine fills an
+	// empty reader-supplied name with the request input-map key.
+	Name string `json:"name"`
+	// ContentType is the media type reported by the reader or derived for
+	// generated output.
+	ContentType string `json:"content_type"`
+	// Body is the artifact content. Engine boundaries copy this slice.
+	Body []byte `json:"body"`
+	// URI is optional source or result provenance metadata.
+	URI string `json:"uri"`
+	// Size is content-size metadata in bytes.
+	Size int64 `json:"size"`
+	// Hash is an opaque content equality value when the producing reader
+	// supplies one. Its format and algorithm are not API contracts.
+	Hash string `json:"hash"`
 }
 
 // ArtifactReader resolves a prompt input reference into its content.
 //
-// Readers are responsible for supplying artifact metadata. The engine assigns
-// an input-map name only when the returned artifact name is empty.
+// Read may be called concurrently. It must honor ctx cancellation to make
+// Prepare and Run responsive to cancellation. The engine passes a copied ref
+// and immediately copies the returned Artifact.Body; it does not retain either
+// value. Readers supply artifact metadata, and the engine assigns an input-map
+// name only when the returned artifact name is empty.
+//
+// An injected reader owns any application-specific path containment,
+// authorization, content-size, and content-type policy. It must protect
+// sensitive references and bodies in its logging and in any copies it retains.
+// It may reuse or mutate the returned artifact and body after Read returns.
+//
+// Returning a non-nil error makes the engine return an error matching
+// ErrArtifactLoad while preserving the reader error through errors.Is.
+// Returning a nil artifact with a nil error also produces ErrArtifactLoad.
 type ArtifactReader interface {
 	Read(context.Context, ArtifactRef) (*Artifact, error)
 }
 
-// ExecutionTarget represents effective model runtime settings.
+// ExecutionTarget represents effective model runtime settings and has a stable
+// JSON representation. It never exposes a resolved API-key value.
 type ExecutionTarget struct {
-	Endpoint        string         `json:"endpoint"`
-	Model           string         `json:"model"`
-	Temperature     float64        `json:"temperature"`
-	MaxTokens       int            `json:"max_tokens"`
-	TopP            float64        `json:"top_p"`
-	TimeoutSeconds  int            `json:"timeout_seconds"`
-	ServiceTier     string         `json:"service_tier"`
-	ReasoningEffort string         `json:"reasoning_effort"`
-	APIKeyEnv       string         `json:"api_key_env"`
-	ExtraParams     map[string]any `json:"extra_params"`
+	// BackendID is the effective routing identity selected by the profile. It
+	// remains unchanged when a profile or request overrides Endpoint and is
+	// empty for endpoint-only profiles. It is supplied to injected LLMClient
+	// implementations as part of the effective target.
+	BackendID string `json:"backend_id,omitempty"`
+	// Endpoint is the model-provider base URL.
+	Endpoint string `json:"endpoint"`
+	// Model is the provider model identifier.
+	Model string `json:"model"`
+	// Temperature is the effective sampling temperature from 0 through 2.
+	Temperature float64 `json:"temperature"`
+	// MaxTokens is the non-negative effective output-token limit. Zero leaves
+	// the limit unspecified to compatible providers unless it was an explicit
+	// request override.
+	MaxTokens int `json:"max_tokens"`
+	// TopP is the effective nucleus-sampling value from 0 through 1.
+	TopP float64 `json:"top_p"`
+	// TimeoutSeconds is the non-negative per-generation deadline. Zero disables
+	// this deadline without disabling caller cancellation or the transport cap.
+	TimeoutSeconds int `json:"timeout_seconds"`
+	// ServiceTier is an optional provider-specific request tier.
+	ServiceTier string `json:"service_tier"`
+	// ReasoningEffort is the effective opaque provider-specific reasoning
+	// setting. An empty value instructs model clients to omit reasoning.
+	ReasoningEffort string `json:"reasoning_effort"`
+	// APIKeyEnv is an environment-variable name, not its credential value.
+	APIKeyEnv string `json:"api_key_env"`
+	// ExtraParams contains copied JSON-compatible provider parameters.
+	ExtraParams map[string]any `json:"extra_params"`
 }
 
-// ExecutionTargetOverride represents per-request runtime setting overrides.
+// ExecutionTargetOverride represents per-request runtime setting overrides and
+// has no stable JSON representation.
+//
+// Non-empty string fields replace profile and backend values. Non-nil pointer
+// fields replace profile values and preserve explicit zero or empty values. A
+// non-empty ExtraParams map replaces the complete profile or backend map
+// rather than merging keys. Empty string fields, nil pointers, and a nil or
+// empty ExtraParams map inherit the selected profile over its backend, when
+// any, and framework defaults.
 type ExecutionTargetOverride struct {
-	Endpoint        string
-	Model           string
-	Temperature     *float64
-	MaxTokens       *int
-	TopP            *float64
-	TimeoutSeconds  *int
-	ServiceTier     string
-	ReasoningEffort string
-	APIKeyEnv       string
-	ExtraParams     map[string]any
+	// Endpoint replaces the profile or backend endpoint when non-empty without
+	// changing the effective BackendID.
+	Endpoint string
+	// Model replaces the profile model when non-empty.
+	Model string
+	// Temperature, when non-nil, must point to a value from 0 through 2.
+	Temperature *float64
+	// MaxTokens, when non-nil, must point to a non-negative value.
+	MaxTokens *int
+	// TopP, when non-nil, must point to a value from 0 through 1.
+	TopP *float64
+	// TimeoutSeconds, when non-nil, must point to a non-negative value. A
+	// pointed-to zero disables the per-generation deadline.
+	TimeoutSeconds *int
+	// ServiceTier replaces the profile value when non-blank.
+	ServiceTier string
+	// ReasoningEffort controls the per-run reasoning setting. Nil inherits the
+	// profile value. A pointer to a non-blank string trims and replaces the
+	// profile value. A pointer to an empty or whitespace-only string clears the
+	// inherited value and disables reasoning for this run. Non-blank values
+	// are opaque and are not validated against a fixed vocabulary.
+	ReasoningEffort *string
+	// APIKeyEnv replaces the profile or backend environment-variable name when
+	// non-blank. A direct RunRequest.APIKey still takes precedence over
+	// environment lookup.
+	APIKeyEnv string
+	// ExtraParams, when non-empty, replaces the complete profile or backend map.
+	// Values must be JSON-compatible: nil, booleans, finite numbers, strings,
+	// arrays or slices, and maps with non-empty string keys. Cycles are invalid.
+	ExtraParams map[string]any
 }
 
 // Profile is an in-memory execution profile for library consumers.
 //
 // It is equivalent to a loaded profile file after validation. Raw API keys do
-// not belong in profiles; use APIKeyRequired to require callers to provide
-// RunRequest.APIKey for each request, or use profile YAML api_key_env with file
-// and FS profile sources.
+// not belong in profiles; use APIKeyRequired to require callers to provide a
+// RunRequest.APIKey or explicit request ExecutionTargetOverride.APIKeyEnv, or
+// use profile YAML api_key_env with file and FS profile sources. Profile has no
+// stable JSON representation.
+//
+// WithProfiles validates and copies Profile values during NewEngine. Numeric
+// zero, blank strings, and an empty ExtraParams map inherit framework defaults;
+// use ExecutionTargetOverride pointer fields to request explicit numeric zero.
 type Profile struct {
-	ID              string
-	Endpoint        string
-	Model           string
-	Temperature     float64
-	MaxTokens       int
-	TopP            float64
-	TimeoutSeconds  int
-	ServiceTier     string
+	// ID is the required non-blank profile identifier. WithProfiles trims it.
+	ID string
+	// BackendID optionally selects an engine backend. WithProfiles trims it.
+	// Backend membership is checked when a request selects the profile; an
+	// unknown ID makes preparation fail with ErrProfileLoad.
+	BackendID string
+	// Endpoint is the model-provider base URL. It is required only when
+	// BackendID is blank and otherwise overrides the backend endpoint when
+	// non-blank.
+	Endpoint string
+	// Model is the required non-blank provider model identifier.
+	Model string
+	// Temperature is from 0 through 2. Zero inherits the framework default.
+	Temperature float64
+	// MaxTokens is non-negative. Zero inherits the framework default.
+	MaxTokens int
+	// TopP is from 0 through 1. Zero inherits the framework default rather than
+	// selecting an explicit zero.
+	TopP float64
+	// TimeoutSeconds is non-negative. Zero inherits the framework default.
+	TimeoutSeconds int
+	// ServiceTier is optional; a blank value inherits the framework default.
+	ServiceTier string
+	// ReasoningEffort is optional; a blank value inherits the framework
+	// default.
 	ReasoningEffort string
-	APIKeyRequired  bool
-	ExtraParams     map[string]any
+	// APIKeyRequired clears a backend's inherited API-key environment name and
+	// requires a non-blank RunRequest.APIKey unless the request explicitly
+	// supplies ExecutionTargetOverride.APIKeyEnv. It does not store a credential.
+	APIKeyRequired bool
+	// ExtraParams contains provider-specific JSON-compatible values. An empty
+	// map inherits backend request defaults, when any. WithProfiles validates
+	// and deeply copies it during NewEngine.
+	ExtraParams map[string]any
 }
 
 // OpenAICompatibleProfileConfig configures an OpenAI-compatible in-memory
 // profile.
 //
 // It contains ordinary profile fields for OpenAI-compatible chat-completions
-// endpoints. APIKeyRequired is satisfied by RunRequest.APIKey. Raw API keys do
-// not belong in this config.
+// endpoints. APIKeyRequired follows Profile.APIKeyRequired. Raw API keys do not
+// belong in this config. OpenAICompatibleProfileConfig has no stable JSON
+// representation and is not validated until its resulting Profile is supplied
+// through WithProfiles to NewEngine.
 type OpenAICompatibleProfileConfig struct {
-	ID              string
-	Endpoint        string
-	Model           string
-	APIKeyRequired  bool
-	Temperature     float64
-	MaxTokens       int
-	TopP            float64
-	TimeoutSeconds  int
-	ServiceTier     string
+	// ID becomes Profile.ID.
+	ID string
+	// BackendID becomes Profile.BackendID.
+	BackendID string
+	// Endpoint becomes Profile.Endpoint.
+	Endpoint string
+	// Model becomes Profile.Model.
+	Model string
+	// APIKeyRequired becomes Profile.APIKeyRequired.
+	APIKeyRequired bool
+	// Temperature becomes Profile.Temperature.
+	Temperature float64
+	// MaxTokens becomes Profile.MaxTokens.
+	MaxTokens int
+	// TopP becomes Profile.TopP.
+	TopP float64
+	// TimeoutSeconds becomes Profile.TimeoutSeconds.
+	TimeoutSeconds int
+	// ServiceTier becomes Profile.ServiceTier.
+	ServiceTier string
+	// ReasoningEffort becomes Profile.ReasoningEffort.
 	ReasoningEffort string
-	ExtraParams     map[string]any
+	// ExtraParams becomes a shallow-copied Profile.ExtraParams map. NewEngine
+	// performs validation and a deep copy when WithProfiles applies the result.
+	ExtraParams map[string]any
 }
 
 // ExecutionTargetPresence tracks which numeric runtime settings were explicit
-// request overrides.
+// request overrides, including explicit zero values. It has a stable JSON
+// representation and is supplied to injected LLM clients so they can preserve
+// omission semantics.
 type ExecutionTargetPresence struct {
-	Temperature    bool
-	MaxTokens      bool
-	TopP           bool
-	TimeoutSeconds bool
+	// Temperature reports a non-nil ExecutionTargetOverride.Temperature.
+	Temperature bool `json:"temperature"`
+	// MaxTokens reports a non-nil ExecutionTargetOverride.MaxTokens.
+	MaxTokens bool `json:"max_tokens"`
+	// TopP reports a non-nil ExecutionTargetOverride.TopP.
+	TopP bool `json:"top_p"`
+	// TimeoutSeconds reports a non-nil ExecutionTargetOverride.TimeoutSeconds.
+	TimeoutSeconds bool `json:"timeout_seconds"`
 }
 
-// OutputContract defines output and validation requirements.
+// OutputContract defines output and validation requirements and has a stable
+// JSON representation.
+//
+// A non-nil RunRequest.Validation replaces the complete prompt contract. It
+// does not merge fields. The public Engine validates generated output once and
+// does not install an output repairer.
 type OutputContract struct {
-	Format         OutputFormat   `json:"format"`
+	// Format selects generated artifact metadata. An empty effective value
+	// defaults to FormatText.
+	Format OutputFormat `json:"format"`
+	// ValidationMode selects the content check. Use one of the declared
+	// ValidationMode constants.
 	ValidationMode ValidationMode `json:"validation_mode"`
-	SchemaPath     string         `json:"schema_path"`
-	RepairAttempts int            `json:"repair_attempts"`
+	// SchemaPath is required when ValidationMode is ValidationJSONSchema and is
+	// ignored by other modes.
+	SchemaPath string `json:"schema_path"`
+	// RepairAttempts is a requested repair limit. A non-positive value requests
+	// no repairs. The public Engine performs no repairs even when this value is
+	// positive, so its runs report zero attempts used.
+	RepairAttempts int `json:"repair_attempts"`
 }
 
-// ValidationResult represents output validation state.
+// ValidationResult represents a completed output check and has a stable JSON
+// representation. An operational inability to perform validation is returned
+// as ErrValidation instead of a ValidationResult.
 type ValidationResult struct {
-	Status         ValidationStatus `json:"status"`
-	Mode           ValidationMode   `json:"mode"`
-	Errors         []string         `json:"errors,omitempty"`
-	SchemaPath     string           `json:"schema_path,omitempty"`
-	RepairAttempts int              `json:"repair_attempts"`
-	IsValid        bool             `json:"is_valid"`
+	// Status is Passed, Failed, or Skipped.
+	Status ValidationStatus `json:"status"`
+	// Mode is the effective validation mode.
+	Mode ValidationMode `json:"mode"`
+	// Errors contains validation diagnostics when Status is ValidationFailed.
+	Errors []string `json:"errors,omitempty"`
+	// SchemaPath is the effective schema path for JSON Schema validation.
+	SchemaPath string `json:"schema_path,omitempty"`
+	// RepairAttempts is the number of repairs actually attempted. It is always
+	// zero for the public Engine.
+	RepairAttempts int `json:"repair_attempts"`
+	// IsValid is true for ValidationPassed and ValidationSkipped and false for
+	// ValidationFailed.
+	IsValid bool `json:"is_valid"`
 }
 
-// TokenUsage tracks token consumption.
+// TokenUsage contains model-client token accounting and has a stable JSON
+// representation. Promptkit preserves values reported by the client and does
+// not derive or reconcile them.
 type TokenUsage struct {
-	PromptTokens     int `json:"prompt_tokens"`
+	// PromptTokens is the reported input-token count.
+	PromptTokens int `json:"prompt_tokens"`
+	// CompletionTokens is the reported generated-token count.
 	CompletionTokens int `json:"completion_tokens"`
-	TotalTokens      int `json:"total_tokens"`
-	CachedTokens     int `json:"cached_tokens"`
+	// TotalTokens is the reported total-token count.
+	TotalTokens int `json:"total_tokens"`
+	// CachedTokens is the reported cached-input-token count.
+	CachedTokens int `json:"cached_tokens"`
+	// CacheWriteTokens is the reported cache-write-token count.
 	CacheWriteTokens int `json:"cache_write_tokens"`
 }
 
-// RenderedPrompt is the fully rendered prompt passed to an LLM client.
+// RenderedPrompt is the fully rendered prompt passed to an LLM client and has
+// a stable JSON representation.
 type RenderedPrompt struct {
-	SessionID string            `json:"session_id,omitempty"`
-	Messages  []RenderedMessage `json:"messages"`
+	// SessionID is the optional effective direct or rendered session
+	// identifier supplied to the model client.
+	SessionID string `json:"session_id,omitempty"`
+	// Messages contains rendered messages in definition order.
+	Messages []RenderedMessage `json:"messages"`
 }
 
-// RenderedMessage is a rendered chat message.
+// RenderedMessage is a rendered chat message and has a stable JSON
+// representation.
 type RenderedMessage struct {
-	Role         string        `json:"role"`
-	Content      string        `json:"content"`
+	// Role is the definition-supplied chat role.
+	Role string `json:"role"`
+	// Content is the rendered message text.
+	Content string `json:"content"`
+	// CacheControl is optional provider cache metadata.
 	CacheControl *CacheControl `json:"cache_control,omitempty"`
 }
 
-// CacheControl describes provider cache metadata attached to prompt content.
+// CacheControl describes provider cache metadata attached to prompt content
+// and has a stable JSON representation.
 type CacheControl struct {
+	// Type identifies the cache behavior.
 	Type CacheControlType `json:"type"`
-	TTL  string           `json:"ttl,omitempty"`
+	// TTL is an optional provider cache lifetime.
+	TTL string `json:"ttl,omitempty"`
 }
 
-// StructuredOutputSpec describes provider-level structured output.
+// StructuredOutputSpec describes provider-level structured output and has a
+// stable JSON representation.
 type StructuredOutputSpec struct {
-	Type       StructuredOutputType      `json:"type"`
+	// Type identifies the structured-output mechanism.
+	Type StructuredOutputType `json:"type"`
+	// JSONSchema contains constraints when Type is StructuredOutputJSONSchema.
 	JSONSchema *StructuredOutputJSONSpec `json:"json_schema,omitempty"`
 }
 
-// StructuredOutputJSONSpec contains JSON Schema output constraints.
+// StructuredOutputJSONSpec contains provider-facing JSON Schema output
+// constraints and has a stable JSON representation.
 type StructuredOutputJSONSpec struct {
-	Name   string `json:"name"`
-	Strict bool   `json:"strict"`
-	Schema any    `json:"schema"`
+	// Name is the provider-facing schema name.
+	Name string `json:"name"`
+	// Strict requests strict provider enforcement of Schema.
+	Strict bool `json:"strict"`
+	// Schema is a caller-owned copy of the loaded JSON Schema document.
+	Schema any `json:"schema"`
 }
 
-// LLMClient executes rendered prompts for Engine.Run.
+// LLMClient executes rendered prompts for [Engine.Run].
+//
+// Generate is scheduled according to the resolved backend's capacity policy.
+// It may still be called concurrently for different backend pools or unlimited
+// backends. Cancellation while waiting for capacity can prevent Generate from
+// being called. Once invoked, it must honor context cancellation to make Run
+// responsive to cancellation. The request and all nested maps, slices, and
+// pointers are client-owned copies and may be mutated or retained without
+// affecting engine state.
+//
+// Generate receives rendered messages and may receive a direct API key. A
+// client must protect those values and any raw output in its logging, storage,
+// and retained copies. It is responsible for the cancellation behavior of any
+// work it starts and for synchronizing access to retained or shared data.
+//
+// A returned error makes Run return ErrLLMGenerate while preserving the client
+// error through errors.Is. A nil response with a nil error also produces
+// ErrLLMGenerate. Promptkit copies the non-nil response before returning from
+// Run.
 type LLMClient interface {
 	Generate(context.Context, GenerateRequest) (*GenerateResponse, error)
 }
 
-// GenerateRequest is passed to an injected LLM client.
+// GenerateRequest is passed to an injected LLM client and has a stable JSON
+// representation. Its String and GoString methods omit rendered content and
+// direct credentials.
 type GenerateRequest struct {
-	Prompt           RenderedPrompt          `json:"prompt"`
-	Target           ExecutionTarget         `json:"target"`
-	TargetPresence   ExecutionTargetPresence `json:"target_presence"`
-	StructuredOutput *StructuredOutputSpec   `json:"structured_output,omitempty"`
-	APIKey           string                  `json:"-"`
+	// Prompt contains the rendered session ID and messages.
+	Prompt RenderedPrompt `json:"prompt"`
+	// Target contains effective model settings without the direct API key.
+	Target ExecutionTarget `json:"target"`
+	// TargetPresence distinguishes inherited numeric zeros from explicit
+	// request overrides.
+	TargetPresence ExecutionTargetPresence `json:"target_presence"`
+	// StructuredOutput contains provider response constraints when requested.
+	StructuredOutput *StructuredOutputSpec `json:"structured_output,omitempty"`
+	// APIKey is the direct request-scoped credential, if any. It is excluded
+	// from JSON, String, and GoString output.
+	APIKey string `json:"-"`
 }
 
-// GenerateResponse is returned by an injected LLM client.
+// GenerateResponse is returned by an injected LLM client and has a stable JSON
+// representation.
 type GenerateResponse struct {
-	Content string     `json:"content"`
-	Usage   TokenUsage `json:"usage"`
+	// Content is the generated output. It must be non-empty when using the
+	// built-in client; injected clients may return empty content for Promptkit
+	// validation to classify.
+	Content string `json:"content"`
+	// Usage is the client's token accounting.
+	Usage TokenUsage `json:"usage"`
 }
 
-// File returns a file-backed artifact reference.
+// File returns a file-backed artifact reference whose URI is path.
+//
+// The default artifact reader opens path as a caller-selected operating-system
+// path without restricting it to an application root or imposing a size limit.
+// Applications accepting untrusted paths must validate them before calling
+// Promptkit or use [WithArtifactReader] to enforce application policy.
 func File(path string) ArtifactRef {
 	return ArtifactRef{Type: ArtifactRefFile, URI: path}
 }
 
-// Inline returns an inline artifact reference.
+// Inline returns an inline artifact reference whose Body is body and whose URI
+// is empty.
 func Inline(body string) ArtifactRef {
 	return ArtifactRef{Type: ArtifactRefInline, Body: body}
 }
 
-// InlineWithURI returns an inline artifact reference with URI metadata.
+// InlineWithURI returns an inline artifact reference with body content and uri
+// provenance metadata.
 func InlineWithURI(uri string, body string) ArtifactRef {
 	return ArtifactRef{Type: ArtifactRefInline, URI: uri, Body: body}
 }