Document the v0.2.0 upgrade path

LICENSE

@@ -208,7 +208,7 @@ If you develop a new program, and you want it to be of the greatest possible use
 
 To do so, attach the following notices to the program. It is safest to attach them to the start of each source file to most effectively state the exclusion of warranty; and each file should have at least the “copyright” line and a pointer to where the full notice is found.
 
-     go-application-template
+     Promptkit
      Copyright (C) 2026  eric
 
      This program is free software: you can redistribute it and/or modify it under the terms of the GNU General Public License as published by the Free Software Foundation, either version 3 of the License, or (at your option) any later version.
@@ -221,7 +221,7 @@ Also add information on how to contact you by electronic and paper mail.
 
 If the program does terminal interaction, make it output a short notice like this when it starts in an interactive mode:
 
-     go-application-template  Copyright (C) 2026  eric
+     Promptkit  Copyright (C) 2026  eric
      This program comes with ABSOLUTELY NO WARRANTY; for details type `show w'.
      This is free software, and you are welcome to redistribute it under certain conditions; type `show c' for details.
 

README.md

@@ -1,3 +1,44 @@
-# PromptKit
+# Promptkit
 
-TODO
+Promptkit is a reusable Go library for preparing and executing prompt-defined
+LLM workflows. Its module path is:
+
+```text
+gitea.maximumdirect.net/eric/promptkit
+```
+
+The root `promptkit` package provides the supported public engine. Consumers
+can configure filesystem or in-memory prompt, profile, and schema sources,
+prepare requests without generation, run requests with the built-in
+OpenAI-compatible client, or inject their own model client and artifact reader.
+
+## Quickstart
+
+Run the maintained offline preparation example from the repository root:
+
+```sh
+go run ./examples/go-library/prepare
+```
+
+It loads a repository-local prompt, supplies an in-memory profile and inline
+input, and prints deterministic preparation metadata without contacting a
+provider or requiring credentials. Read the
+[example source](examples/go-library/prepare/main.go), the
+[Go package consumer guide](docs/consumers/pkg-promptkit.md), and the
+[framework format reference](docs/formats.md) to build a consumer workflow.
+
+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).

architecture_test.go

@@ -0,0 +1,89 @@
+package promptkit_test
+
+import (
+	"go/ast"
+	"go/parser"
+	"go/token"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strconv"
+	"strings"
+	"testing"
+)
+
+const formerModulePath = "gitea.maximumdirect.net/eric/" + "scrip" + "torium"
+
+func TestRepositoryDoesNotImportFormerModule(t *testing.T) {
+	violations, err := findFormerModuleImports(".")
+	if err != nil {
+		t.Fatalf("inspect repository imports: %v", err)
+	}
+	if len(violations) > 0 {
+		t.Fatalf("repository imports the former module:\n%s", strings.Join(violations, "\n"))
+	}
+}
+
+func TestFormerModuleGuardFindsNestedImport(t *testing.T) {
+	root := t.TempDir()
+	nested := filepath.Join(root, "nested", "package")
+	if err := os.MkdirAll(nested, 0o755); err != nil {
+		t.Fatalf("create nested package: %v", err)
+	}
+
+	sourcePath := filepath.Join(nested, "violation.go")
+	source := "package nested\n\nimport _ " + strconv.Quote(formerModulePath+"/internal/domain") + "\n"
+	if err := os.WriteFile(sourcePath, []byte(source), 0o600); err != nil {
+		t.Fatalf("write nested source: %v", err)
+	}
+
+	violations, err := findFormerModuleImports(root)
+	if err != nil {
+		t.Fatalf("inspect nested imports: %v", err)
+	}
+	if len(violations) != 1 {
+		t.Fatalf("violations = %v, want one nested import", violations)
+	}
+	if !strings.Contains(violations[0], "violation.go") ||
+		!strings.Contains(violations[0], formerModulePath+"/internal/domain") {
+		t.Fatalf("violation = %q, want file and import path", violations[0])
+	}
+}
+
+func findFormerModuleImports(root string) ([]string, error) {
+	var violations []string
+	err := filepath.WalkDir(root, func(path string, entry fs.DirEntry, walkErr error) error {
+		if walkErr != nil {
+			return walkErr
+		}
+		if entry.IsDir() {
+			switch entry.Name() {
+			case ".git", "generated", "vendor":
+				return filepath.SkipDir
+			}
+			return nil
+		}
+		if filepath.Ext(path) != ".go" {
+			return nil
+		}
+
+		file, err := parser.ParseFile(token.NewFileSet(), path, nil, parser.ImportsOnly|parser.ParseComments)
+		if err != nil {
+			return err
+		}
+		if ast.IsGenerated(file) {
+			return nil
+		}
+		for _, spec := range file.Imports {
+			importPath, err := strconv.Unquote(spec.Path.Value)
+			if err != nil {
+				return err
+			}
+			if importPath == formerModulePath || strings.HasPrefix(importPath, formerModulePath+"/") {
+				violations = append(violations, path+": "+importPath)
+			}
+		}
+		return nil
+	})
+	return violations, err
+}

artifact_reader.go

@@ -0,0 +1,39 @@
+package promptkit
+
+import (
+	"context"
+	"errors"
+
+	artifactadapter "gitea.maximumdirect.net/eric/promptkit/internal/artifact"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+var errNilArtifactReaderResponse = errors.New("artifact reader returned nil artifact without error")
+
+type publicArtifactReaderAdapter struct {
+	reader ArtifactReader
+}
+
+var _ artifactadapter.Reader = publicArtifactReaderAdapter{}
+
+func (a publicArtifactReaderAdapter) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	artifact, err := a.reader.Read(ctx, ArtifactRef{
+		Type: ArtifactRefType(ref.Type),
+		URI:  ref.URI,
+		Body: ref.Body,
+	})
+	if err != nil {
+		return nil, err
+	}
+	if artifact == nil {
+		return nil, errNilArtifactReaderResponse
+	}
+	return &domain.Artifact{
+		Name:        artifact.Name,
+		ContentType: artifact.ContentType,
+		Body:        copyBytes(artifact.Body),
+		URI:         artifact.URI,
+		Size:        artifact.Size,
+		Hash:        artifact.Hash,
+	}, nil
+}

artifact_reader_internal_test.go

@@ -0,0 +1,36 @@
+package promptkit
+
+import (
+	"context"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestPublicArtifactReaderAdapterCopiesBody(t *testing.T) {
+	reader := internalArtifactReaderFake{
+		artifact: &Artifact{Body: []byte("original")},
+	}
+	adapter := publicArtifactReaderAdapter{reader: &reader}
+
+	artifact, err := adapter.Read(context.Background(), domain.ArtifactRef{
+		Type: domain.ArtifactRefInline,
+		URI:  "memory://input",
+		Body: "input",
+	})
+	if err != nil {
+		t.Fatalf("read artifact: %v", err)
+	}
+	artifact.Body[0] = 'X'
+	if got := string(reader.artifact.Body); got != "original" {
+		t.Fatalf("reader artifact body was mutated: %q", got)
+	}
+}
+
+type internalArtifactReaderFake struct {
+	artifact *Artifact
+}
+
+func (r *internalArtifactReaderFake) Read(context.Context, ArtifactRef) (*Artifact, error) {
+	return r.artifact, nil
+}

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

@@ -0,0 +1,419 @@
+package promptkit
+
+import (
+	"reflect"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
+)
+
+func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
+	execution, err := toDomainExecutionTargetOverride(req.Execution)
+	if err != nil {
+		return domain.RunRequest{}, err
+	}
+	return domain.RunRequest{
+		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),
+	}, nil
+}
+
+func fromDomainPreparedRun(prepared *domain.PreparedRun) *PreparedRun {
+	if prepared == nil {
+		return nil
+	}
+	return &PreparedRun{
+		PromptID:             prepared.PromptID,
+		PromptVersion:        prepared.PromptVersion,
+		PromptHash:           prepared.PromptHash,
+		SelectedProfileID:    prepared.SelectedProfileID,
+		SelectedBackendID:    prepared.SelectedBackendID,
+		EffectiveModelParams: fromDomainExecutionTarget(prepared.EffectiveModelParams),
+		OutputContract:       fromDomainOutputContract(prepared.OutputContract),
+		StructuredOutput:     fromDomainStructuredOutputSpec(prepared.StructuredOutput),
+		InputHashes:          copyStringMap(prepared.InputHashes),
+		SessionID:            prepared.SessionID,
+		RenderedPromptHash:   prepared.RenderedPromptHash,
+		Messages:             fromDomainRenderedMessages(prepared.Messages),
+		StartTime:            prepared.StartTime,
+		EndTime:              prepared.EndTime,
+		DurationMS:           prepared.DurationMS,
+	}
+}
+
+func fromDomainRunResult(result *domain.RunResult) *RunResult {
+	if result == nil {
+		return nil
+	}
+	return &RunResult{
+		RunID:                result.RunID,
+		Artifact:             fromDomainArtifact(result.Artifact),
+		RawOutput:            result.RawOutput,
+		Validation:           fromDomainValidationResult(result.Validation),
+		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),
+		InputHashes:          copyStringMap(result.InputHashes),
+		Usage:                fromDomainTokenUsage(result.Usage),
+		StartTime:            result.StartTime,
+		EndTime:              result.EndTime,
+		Duration:             result.Duration,
+	}
+}
+
+func fromDomainGenerateRequest(req domain.GenerateRequest) GenerateRequest {
+	return GenerateRequest{
+		Prompt:           fromDomainRenderedPrompt(req.Prompt),
+		Target:           fromDomainExecutionTarget(req.Target),
+		TargetPresence:   fromDomainExecutionTargetPresence(req.TargetPresence),
+		StructuredOutput: fromDomainStructuredOutputSpec(req.StructuredOutput),
+		APIKey:           req.Target.APIKey,
+	}
+}
+
+func toDomainGenerateResponse(resp *GenerateResponse) *domain.GenerateResponse {
+	if resp == nil {
+		return nil
+	}
+	return &domain.GenerateResponse{
+		Content: resp.Content,
+		Usage:   toDomainTokenUsage(resp.Usage),
+	}
+}
+
+func fromDomainRenderedPrompt(prompt domain.RenderedPrompt) RenderedPrompt {
+	return RenderedPrompt{
+		SessionID: prompt.SessionID,
+		Messages:  fromDomainRenderedMessages(prompt.Messages),
+	}
+}
+
+func toDomainArtifactRefMap(src map[string]ArtifactRef) map[string]domain.ArtifactRef {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]domain.ArtifactRef, len(src))
+	for k, v := range src {
+		out[k] = toDomainArtifactRef(v)
+	}
+	return out
+}
+
+func toDomainArtifactRef(ref ArtifactRef) domain.ArtifactRef {
+	return domain.ArtifactRef{
+		Type: domain.ArtifactRefType(ref.Type),
+		URI:  ref.URI,
+		Body: ref.Body,
+	}
+}
+
+func fromDomainArtifact(artifact domain.Artifact) Artifact {
+	return Artifact{
+		Name:        artifact.Name,
+		ContentType: artifact.ContentType,
+		Body:        copyBytes(artifact.Body),
+		URI:         artifact.URI,
+		Size:        artifact.Size,
+		Hash:        artifact.Hash,
+	}
+}
+
+func toDomainExecutionTargetOverride(override *ExecutionTargetOverride) (*domain.ExecutionTargetOverride, error) {
+	if override == nil {
+		return nil, nil
+	}
+	extraParams, err := jsonvalue.CopyMap(override.ExtraParams)
+	if err != nil {
+		return nil, err
+	}
+	return &domain.ExecutionTargetOverride{
+		Endpoint:        override.Endpoint,
+		Model:           override.Model,
+		Temperature:     copyFloat64Ptr(override.Temperature),
+		MaxTokens:       copyIntPtr(override.MaxTokens),
+		TopP:            copyFloat64Ptr(override.TopP),
+		TimeoutSeconds:  copyIntPtr(override.TimeoutSeconds),
+		ServiceTier:     override.ServiceTier,
+		ReasoningEffort: copyStringPtr(override.ReasoningEffort),
+		APIKeyEnv:       override.APIKeyEnv,
+		ExtraParams:     extraParams,
+	}, nil
+}
+
+func fromDomainExecutionTarget(target domain.ExecutionTarget) ExecutionTarget {
+	return ExecutionTarget{
+		BackendID:       target.BackendID,
+		Endpoint:        target.Endpoint,
+		Model:           target.Model,
+		Temperature:     target.Temperature,
+		MaxTokens:       target.MaxTokens,
+		TopP:            target.TopP,
+		TimeoutSeconds:  target.TimeoutSeconds,
+		ServiceTier:     target.ServiceTier,
+		ReasoningEffort: target.ReasoningEffort,
+		APIKeyEnv:       target.APIKeyEnv,
+		ExtraParams:     copyAnyMap(target.ExtraParams),
+	}
+}
+
+func fromDomainExecutionTargetPresence(presence domain.ExecutionTargetPresence) ExecutionTargetPresence {
+	return ExecutionTargetPresence{
+		Temperature:    presence.Temperature,
+		MaxTokens:      presence.MaxTokens,
+		TopP:           presence.TopP,
+		TimeoutSeconds: presence.TimeoutSeconds,
+	}
+}
+
+func toDomainOutputContractPtr(contract *OutputContract) *domain.OutputContract {
+	if contract == nil {
+		return nil
+	}
+	out := toDomainOutputContract(*contract)
+	return &out
+}
+
+func toDomainOutputContract(contract OutputContract) domain.OutputContract {
+	return domain.OutputContract{
+		Format:         domain.OutputFormat(contract.Format),
+		ValidationMode: domain.ValidationMode(contract.ValidationMode),
+		SchemaPath:     contract.SchemaPath,
+		RepairAttempts: contract.RepairAttempts,
+	}
+}
+
+func fromDomainOutputContract(contract domain.OutputContract) OutputContract {
+	return OutputContract{
+		Format:         OutputFormat(contract.Format),
+		ValidationMode: ValidationMode(contract.ValidationMode),
+		SchemaPath:     contract.SchemaPath,
+		RepairAttempts: contract.RepairAttempts,
+	}
+}
+
+func fromDomainValidationResult(result domain.ValidationResult) ValidationResult {
+	return ValidationResult{
+		Status:         ValidationStatus(result.Status),
+		Mode:           ValidationMode(result.Mode),
+		Errors:         copyStringSlice(result.Errors),
+		SchemaPath:     result.SchemaPath,
+		RepairAttempts: result.RepairAttempts,
+		IsValid:        result.IsValid,
+	}
+}
+
+func fromDomainTokenUsage(usage domain.TokenUsage) TokenUsage {
+	return TokenUsage{
+		PromptTokens:     usage.PromptTokens,
+		CompletionTokens: usage.CompletionTokens,
+		TotalTokens:      usage.TotalTokens,
+		CachedTokens:     usage.CachedTokens,
+		CacheWriteTokens: usage.CacheWriteTokens,
+	}
+}
+
+func toDomainTokenUsage(usage TokenUsage) domain.TokenUsage {
+	return domain.TokenUsage{
+		PromptTokens:     usage.PromptTokens,
+		CompletionTokens: usage.CompletionTokens,
+		TotalTokens:      usage.TotalTokens,
+		CachedTokens:     usage.CachedTokens,
+		CacheWriteTokens: usage.CacheWriteTokens,
+	}
+}
+
+func fromDomainRenderedMessages(messages []domain.RenderedMessage) []RenderedMessage {
+	if messages == nil {
+		return nil
+	}
+	out := make([]RenderedMessage, len(messages))
+	for i, msg := range messages {
+		out[i] = RenderedMessage{
+			Role:         msg.Role,
+			Content:      msg.Content,
+			CacheControl: fromDomainCacheControl(msg.CacheControl),
+		}
+	}
+	return out
+}
+
+func fromDomainCacheControl(cacheControl *domain.CacheControl) *CacheControl {
+	if cacheControl == nil {
+		return nil
+	}
+	return &CacheControl{
+		Type: CacheControlType(cacheControl.Type),
+		TTL:  cacheControl.TTL,
+	}
+}
+
+func fromDomainStructuredOutputSpec(spec *domain.StructuredOutputSpec) *StructuredOutputSpec {
+	if spec == nil {
+		return nil
+	}
+	out := &StructuredOutputSpec{
+		Type: StructuredOutputType(spec.Type),
+	}
+	if spec.JSONSchema != nil {
+		out.JSONSchema = &StructuredOutputJSONSpec{
+			Name:   spec.JSONSchema.Name,
+			Strict: spec.JSONSchema.Strict,
+			Schema: copyAny(spec.JSONSchema.Schema),
+		}
+	}
+	return out
+}
+
+func copyStringMap(src map[string]string) map[string]string {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]string, len(src))
+	for k, v := range src {
+		out[k] = v
+	}
+	return out
+}
+
+func copyAnyMap(src map[string]any) map[string]any {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]any, len(src))
+	for k, v := range src {
+		out[k] = copyAny(v)
+	}
+	return out
+}
+
+func copyAny(value any) any {
+	if value == nil {
+		return nil
+	}
+	switch v := value.(type) {
+	case map[string]any:
+		return copyAnyMap(v)
+	case []any:
+		out := make([]any, len(v))
+		for i, item := range v {
+			out[i] = copyAny(item)
+		}
+		return out
+	case []string:
+		return copyStringSlice(v)
+	case []byte:
+		return copyBytes(v)
+	default:
+		return copyReflectValue(reflect.ValueOf(value)).Interface()
+	}
+}
+
+func copyReflectValue(value reflect.Value) reflect.Value {
+	if !value.IsValid() {
+		return value
+	}
+
+	switch value.Kind() {
+	case reflect.Interface:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		copied := copyReflectValue(value.Elem())
+		if copied.IsValid() && copied.Type().AssignableTo(value.Type()) {
+			return copied
+		}
+		out := reflect.New(value.Type()).Elem()
+		out.Set(copied)
+		return out
+	case reflect.Pointer:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.New(value.Type().Elem())
+		out.Elem().Set(copyReflectValue(value.Elem()))
+		return out
+	case reflect.Map:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.MakeMapWithSize(value.Type(), value.Len())
+		iter := value.MapRange()
+		for iter.Next() {
+			out.SetMapIndex(copyReflectValue(iter.Key()), copyReflectValue(iter.Value()))
+		}
+		return out
+	case reflect.Slice:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.MakeSlice(value.Type(), value.Len(), value.Cap())
+		for i := 0; i < value.Len(); i++ {
+			out.Index(i).Set(copyReflectValue(value.Index(i)))
+		}
+		return out
+	case reflect.Array:
+		out := reflect.New(value.Type()).Elem()
+		for i := 0; i < value.Len(); i++ {
+			out.Index(i).Set(copyReflectValue(value.Index(i)))
+		}
+		return out
+	default:
+		return value
+	}
+}
+
+func copyStringSlice(src []string) []string {
+	if src == nil {
+		return nil
+	}
+	out := make([]string, len(src))
+	copy(out, src)
+	return out
+}
+
+func copyBytes(src []byte) []byte {
+	if src == nil {
+		return nil
+	}
+	out := make([]byte, len(src))
+	copy(out, src)
+	return out
+}
+
+func copyFloat64Ptr(src *float64) *float64 {
+	if src == nil {
+		return nil
+	}
+	v := *src
+	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
+	}
+	v := *src
+	return &v
+}

doc.go

@@ -0,0 +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 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

@@ -0,0 +1,254 @@
+# Package `promptkit`
+
+## 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"
+```
+
+The following Go fragments are illustrative and omit surrounding package,
+import, and error-handling code. Use the maintained examples for complete
+programs.
+
+## Construct An Engine
+
+Create an engine with
+[`NewEngine`](../../engine.go). A directory-backed setup supplies a prompt
+directory and may supply profile and schema directories:
+
+```go
+engine, err := promptkit.NewEngine(promptkit.Config{
+	PromptDir:  "prompts",
+	ProfileDir: "profiles",
+	SchemaDir:  "schemas",
+})
+```
+
+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).
+
+## Prepare Without Model Execution
+
+[`Engine.Prepare`](../../engine.go) resolves the selected prompt and profile,
+loads inputs and any structured-output schema, and renders messages without
+calling a model client:
+
+```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)
+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).
+
+## Execute And Validate
+
+[`Engine.Run`](../../engine.go) performs the same preparation, invokes the
+configured model client, classifies the generated artifact, and validates the
+content. A completed content check may return `ValidationFailed` in the result;
+an operational inability to validate returns an error.
+
+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.
+
+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.
+
+## Inputs, Profiles, And Overrides
+
+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.
+
+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.
+
+For programmatic profiles,
+[`OpenAICompatibleProfile`](../../profiles.go) converts ordinary
+OpenAI-compatible settings into a value accepted by `WithProfiles`.
+
+### Set A Per-Run Session And Reasoning
+
+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
+
+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.
+
+## Handle Errors
+
+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`.
+
+When a limited backend has admitted all active and waiting calls, handle
+`ErrCapacityExceeded` separately from request errors and provider failures:
+
+```go
+result, err := engine.Run(ctx, request)
+if errors.Is(err, promptkit.ErrCapacityExceeded) {
+	// Apply application policy: shed work, report overload, or retry later.
+}
+```
+
+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.
+
+## Application Boundary
+
+Promptkit is an importable library. It does not own a command, inbound HTTP
+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/development.md

@@ -1,54 +1,141 @@
 # Development
 
-This is the contributor entry point for thsi application. Use the task-specific
-reading guide below before making changes. Canonical architecture, contracts,
-component behavior, and policies remain in their owning documents.
+This is the contributor entry point for Promptkit, a reusable Go library. All
+contributors must read the
+[architecture policy](policy/architecture.md) before making changes.
 
 ## Initial Orientation
 
 Before starting work:
 
 1. inspect the working tree and preserve unrelated changes;
-2. read the architecture policy for code or design work;
-3. read the policy, contract, and internal documents listed for the task;
-4. inspect the relevant implementation and tests before deciding how to change
-   them.
+2. read the policy, contract, and internal documents listed for the task;
+3. inspect the relevant implementation and tests before deciding how to change
+   them; and
+4. keep documentation limited to implemented behavior unless an accepted
+   decision or temporary roadmap explicitly owns future work.
 
 Start with:
 
-- [Architecture policy](policy/architecture.md) for system boundaries,
-  invariants, and non-goals;
-- [Internal component overview](internal/overview.md) for the current package
-  and component map;
-- [Documentation policy](policy/documentation.md) before changing
+- the [architecture policy](policy/architecture.md) for library boundaries,
+  dependency direction, invariants, and non-goals;
+- the [internal component overview](internal/overview.md) for the current
+  package and component inventory;
+- the [documentation policy](policy/documentation.md) before changing
   documentation;
-- [Testing policy](policy/testing.md) before adding, rewriting, or deleting
-  tests.
+- the [testing policy](policy/testing.md) before adding, rewriting, or deleting
+  tests; and
+- the [release procedure](release.md) for version and publication work.
 
 ## Task-Specific Reading Guide
 
 | Task | Read before changing |
 | --- | --- |
-| Repository orientation or component responsibility | [Internal component overview](internal/overview.md) and [architecture policy](policy/architecture.md) |
-| Examples or copyable assets | The owning contract for the demonstrated behavior and the related files under `examples/` |
-| Architecture decisions or future work | The [documentation policy](policy/documentation.md) and relevant accepted ADRs |
+| Root public API | The [architecture policy](policy/architecture.md), [consumer guide](consumers/pkg-promptkit.md), [testing policy](policy/testing.md), and existing GoDoc. |
+| Prompt, profile, or schema formats | The [framework format reference](formats.md), owning parser or validator package, and [documentation policy](policy/documentation.md). |
+| Source loading or validation | The [framework format reference](formats.md), [internal source document](internal/sources.md), and owning package tests. |
+| Model-client behavior | The [OpenAI-compatible integration contract](integrations/openai-compatible-chat.md), [internal model-client document](internal/llm.md), and owning package tests. |
+| Internal package implementation | The [architecture policy](policy/architecture.md), [internal component overview](internal/overview.md), and focused internal document listed for that package. |
+| Tests or test fixtures | The [testing policy](policy/testing.md), owning package, and focused internal document listed by the component overview. |
+| Maintained example | The [example](../examples/go-library/prepare/main.go), [consumer guide](consumers/pkg-promptkit.md), [framework format reference](formats.md), and [documentation policy](policy/documentation.md). |
+| Documentation | The [documentation policy](policy/documentation.md) and canonical owner of every affected contract. |
+| Release preparation or publication | The [release procedure](release.md). |
 
-For cross-cutting changes, follow every applicable row. Internal component
-documents own detailed subsystem change recipes.
+For cross-cutting changes, follow every applicable row. Do not create
+placeholder documents for packages, APIs, or integrations that do not yet
+exist.
 
-## Baseline Validation
+## Maintainer-Run Validation
 
-Use focused checks while iterating, then run validation proportionate to the
-change and the risks described by the testing policy.
+Promptkit does not currently use hosted CI. Maintainers are responsible for
+running the documented checks before accepting changes. Run the default Go
+validation from the Promptkit repository root:
 
-The repository-level baseline for code changes is:
-
-```bash
+```sh
 go test ./...
+go test -race ./...
 go vet ./...
-go build ./cmd/scriptorium
+go build ./...
+go run ./examples/go-library/prepare
 ```
 
-Documentation-only work does not require the full Go suite unless it changes
-commands, examples, generated output, or another behavior that the suite
-validates. Always check changed links, paths, examples, and canonical ownership.
+Check formatting across every tracked Go file:
+
+```sh
+gofmt -l $(git ls-files '*.go')
+```
+
+The formatting command must produce no paths. Follow every added or changed
+Markdown link and confirm its target exists. Finally, check whitespace:
+
+```sh
+git diff --check
+```
+
+Documentation-only work does not require unrelated new tests, but it still
+requires link validation and `git diff --check`. Run the Go validation whenever
+documentation changes commands, examples, generated output, or another
+behavior checked by the module.
+
+## Focused Validation
+
+Use focused checks while iterating, then run the complete validation sequence
+before accepting the change. The root package supports:
+
+```sh
+go test .
+go vet .
+go build .
+```
+
+Filter tests by name without assuming a fixed internal package layout:
+
+```sh
+go test ./... -run 'TestName'
+```
+
+Replace `TestName` with a useful regular expression. Target only paths that
+exist, and consult the internal component overview for their owning
+documentation. A filtered or package-specific run does not replace the
+complete repository validation.
+
+## Coordinated Work With Scriptorium
+
+Promptkit and Scriptorium must remain independently valid. For temporary local
+integration, use either a Go workspace outside both repositories or an
+uncommitted replacement in the consuming module.
+
+If the repositories are sibling directories, run the workspace commands from
+their parent directory:
+
+```sh
+go work init ./promptkit ./scriptorium
+go work sync
+```
+
+Use the workspace only for coordinated local checks. From the same parent
+directory, remove it when finished:
+
+```sh
+rm -f go.work go.work.sum
+```
+
+Alternatively, from the Scriptorium repository root, temporarily point its
+Promptkit dependency at the sibling checkout:
+
+```sh
+go mod edit -replace gitea.maximumdirect.net/eric/promptkit=../promptkit
+```
+
+After coordinated checks, remove the replacement and reconcile module
+metadata:
+
+```sh
+go mod edit -dropreplace gitea.maximumdirect.net/eric/promptkit
+go mod tidy
+```
+
+Never commit `go.work`, `go.work.sum`, or a local filesystem `replace`
+directive. Before committing in either repository, inspect its module files and
+working tree independently. Published consumer versions must depend on a tagged
+Promptkit version, not a workspace, local replacement, or unpublished commit.

docs/formats.md

@@ -0,0 +1,300 @@
+# Framework Format Reference
+
+## Purpose
+
+This document is the canonical contract for Promptkit prompt-definition,
+profile, and schema files. The [Go package consumer guide](consumers/pkg-promptkit.md)
+explains how to select these sources and invoke the engine. The
+[OpenAI-compatible integration contract](integrations/openai-compatible-chat.md)
+owns the resulting outbound wire behavior.
+
+Prompt and profile sources recursively discover files ending in `.yaml` or
+`.yml`. YAML decoding is strict: unknown fields are errors for the selected
+definition. Definitions are selected by their YAML `id`, not their file name
+or directory.
+
+## Prompt Definitions
+
+A prompt definition describes inputs, Go-template messages, an optional
+default profile, and an output contract.
+
+```yaml
+id: meeting.summary
+version: "1.0.0"
+default_profile: local-summary
+description: Summarize a synthetic meeting note.
+session_id: '{{.session}}'
+inputs:
+  - name: note
+    required: true
+    content_type: text/plain
+    description: Meeting note to summarize.
+messages:
+  - role: system
+    content: Return a concise summary.
+    cache_control:
+      type: ephemeral
+      ttl: 1h
+  - role: user
+    content_file: ./summary.user.md
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+```
+
+| Field | Required | Meaning |
+| --- | --- | --- |
+| `id` | yes | Non-empty prompt identifier used by `RunRequest.PromptID`. |
+| `version` | yes | Non-empty version selected by an optional `RunRequest.PromptVersion`. |
+| `default_profile` | no | Non-empty profile ID used when the request omits `ProfileID`. |
+| `description` | no | Human-readable description. |
+| `session_id` | no | Go template rendered from request variables and input helpers. |
+| `inputs` | no | Declared input metadata. |
+| `messages` | yes | One or more chat-message templates. |
+| `output` | yes | Output format and validation settings. |
+
+When a request omits a version, the selected prompt ID must identify exactly
+one definition. When it supplies a version, the ID and version pair must be
+unique.
+
+### Inputs
+
+Each `inputs` item has these fields:
+
+| Field | Required | Meaning |
+| --- | --- | --- |
+| `name` | yes | Non-empty name used by the request input map and `input` template helper. Names must be unique. |
+| `required` | no | When true, preparation fails if the request omits the input. The default is false. |
+| `content_type` | no | Expected media-type metadata. |
+| `description` | no | Human-readable input description. |
+
+Requests supply inputs as inline or file-backed `ArtifactRef` values. Declared
+required inputs must be present. A template reference also requires the named
+input to exist, whether or not it was declared. Extra request inputs are
+allowed.
+
+### Messages And Templates
+
+Each message has a non-empty `role` and exactly one of:
+
+- `content`, containing an inline Go template; or
+- `content_file`, naming a file whose contents are the Go template.
+
+For directory and `fs.FS` prompt sources, `content_file` resolves relative to
+the prompt file and remains within the source root. `WithPromptFile` also
+resolves it relative to that file.
+
+Request variables are the template data, so a variable named `audience` is
+referenced as `{{.audience}}`. The `{{input "note"}}` helper renders the body
+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. A nonblank direct request session ID bypasses this template completely;
+a blank direct value leaves the template behavior unchanged.
+
+### Cache Control
+
+`cache_control` is optional and has these fields:
+
+| Field | Required | Values |
+| --- | --- | --- |
+| `type` | yes | `ephemeral` |
+| `ttl` | no | Empty or `1h` |
+
+Promptkit preserves cache-control metadata on the rendered message. The
+outbound integration determines its wire representation.
+
+### Output Contract
+
+| Field | Required | Values or behavior |
+| --- | --- | --- |
+| `format` | yes | `text`, `markdown`, or `json`. |
+| `validation_mode` | yes | `none`, `basic`, `json`, or `json_schema`. |
+| `schema_path` | for `json_schema` | Path to a schema in the configured schema source. |
+| `repair_attempts` | no | Integer zero or greater; omitted means zero. |
+
+The validation modes behave as follows:
+
+- `none` skips content validation;
+- `basic` requires non-empty generated content;
+- `json` requires valid JSON; and
+- `json_schema` requires valid JSON that satisfies the selected schema.
+
+`format` controls output artifact metadata. JSON Schema mode also supplies the
+schema to compatible model clients as structured-output metadata. The public
+engine does not install an output repairer, so its validation is single-pass
+even when a positive `repair_attempts` value is present.
+
+A request-level `OutputContract` replaces the complete prompt output contract.
+It does not merge individual fields. If its format is empty, Promptkit uses
+`text`.
+
+## Profile Definitions
+
+A profile supplies model execution settings:
+
+```yaml
+id: local-summary
+backend: openrouter
+model: example-model
+temperature: 0.2
+max_tokens: 500
+top_p: 0.95
+timeout_seconds: 90
+service_tier: flex
+reasoning_effort: medium
+extra_params:
+  provider_option: enabled
+```
+
+| Field | Required | Meaning |
+| --- | --- | --- |
+| `id` | yes | Non-empty profile identifier. IDs must be unique within one source. |
+| `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. |
+| `top_p` | no | Number from 0 through 1. |
+| `timeout_seconds` | no | Per-generation deadline in whole seconds; integer zero or greater. |
+| `service_tier` | no | Provider-specific request tier. |
+| `reasoning_effort` | no | Provider-specific reasoning setting. |
+| `api_key_env` | no | Name of an environment variable containing the API key. |
+| `extra_params` | no | JSON-compatible provider-specific outbound fields. |
+
+Raw `api_key` is prohibited in profile YAML. Store only an environment
+variable name in `api_key_env`.
+
+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
+[outbound request contract](integrations/openai-compatible-chat.md#request-body).
+
+### Defaults And Overrides
+
+Execution settings resolve in this order:
+
+1. framework defaults;
+2. the selected backend, when the profile names one;
+3. the selected profile; and
+4. request `ExecutionTargetOverride` values.
+
+The framework defaults are:
+
+| Setting | Default |
+| --- | --- |
+| `temperature` | `0` |
+| `max_tokens` | `0` |
+| `top_p` | `1` |
+| `timeout_seconds` | `600` |
+
+Numeric zero in a file or in-memory profile means that the profile does not
+replace the framework default. Numeric request overrides use pointers, so an
+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 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.
+
+### Source And Profile Precedence
+
+An explicit request profile ID takes precedence over the prompt's
+`default_profile`. If neither is present, preparation fails.
+
+Profile sources resolve matching IDs in this order:
+
+1. in-memory profiles supplied with `WithProfiles`;
+2. a profile file, `fs.FS`, or configured profile directory; and
+3. embedded built-in profiles.
+
+A higher-precedence source falls back only when the profile is absent. An
+invalid matching profile is an error and does not fall back. In-memory
+`Profile` values follow the same ranges as YAML profiles. They use
+`APIKeyRequired` for request-scoped credentials instead of `api_key_env`.
+
+## Built-In Profile Catalog
+
+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 |
+| --- | --- | --- |
+| aion-labs | `aion-2` | `aion-labs/aion-2.0` |
+| anthropic | `claude-fable-latest` | `~anthropic/claude-fable-latest` |
+| anthropic | `claude-haiku-latest` | `~anthropic/claude-haiku-latest` |
+| anthropic | `claude-opus-latest` | `~anthropic/claude-opus-latest` |
+| anthropic | `claude-sonnet-latest` | `~anthropic/claude-sonnet-latest` |
+| deepseek | `deepseek-3-2` | `deepseek/deepseek-v3.2` |
+| deepseek | `deepseek-4-flash` | `deepseek/deepseek-v4-flash` |
+| deepseek | `deepseek-4-pro` | `deepseek/deepseek-v4-pro` |
+| google | `gemini-2-flash` | `google/gemini-2.5-flash` |
+| google | `gemini-2-flash-lite` | `google/gemini-2.5-flash-lite` |
+| google | `gemini-2-pro` | `google/gemini-2.5-pro` |
+| google | `gemini-3-flash-lite` | `google/gemini-3.1-flash-lite` |
+| google | `gemini-flash-latest` | `~google/gemini-flash-latest` |
+| google | `gemini-pro-latest` | `~google/gemini-pro-latest` |
+| google | `gemma-4-31b` | `google/gemma-4-31b-it:exacto` |
+| minimax | `minimax-m2` | `minimax/minimax-m2.5` |
+| minimax | `minimax-m3` | `minimax/minimax-m3` |
+| mistral | `mistral-large-2512` | `mistralai/mistral-large-2512` |
+| mistral | `mistral-medium-3-5` | `mistralai/mistral-medium-3-5` |
+| mistral | `mistral-small-3` | `mistralai/mistral-small-3.2-24b-instruct` |
+| mistral | `mistral-small-4` | `mistralai/mistral-small-2603` |
+| nvidia | `nemotron-3-ultra` | `nvidia/nemotron-3-ultra-550b-a55b` |
+| openai | `gpt-5-mini` | `openai/gpt-5.4-mini` |
+| openai | `gpt-5-nano` | `openai/gpt-5.4-nano` |
+
+## Schemas
+
+Schemas are JSON documents selected by a prompt or request
+`schema_path`. For a directory or `fs.FS` source, paths resolve within the
+configured source root. Referenced nested schemas resolve relative to the
+owning schema document. `WithSchemaFile` exposes one schema, addressed by its
+base name.
+
+An unreadable, invalid, or unresolvable schema produces an operational
+validation error. Generated content that is valid JSON but does not satisfy the
+schema produces a failed validation result.
+
+## Credentials
+
+Credential values belong at the request or environment boundary, never in
+prompt, profile, schema, or example files:
+
+- a file profile names an environment variable with `api_key_env`;
+- an in-memory profile may set `APIKeyRequired`;
+- 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
+their values do not.

docs/integrations/openai-compatible-chat.md

@@ -0,0 +1,106 @@
+# OpenAI-Compatible Chat Integration
+
+## Purpose
+
+This document defines the outbound HTTP behavior implemented by Promptkit's
+internal OpenAI-compatible model client. The
+[internal model-client document](../internal/llm.md) owns implementation flow,
+errors, and test ownership. The root Promptkit engine uses this client by
+default unless a consumer injects another implementation. The
+[framework format reference](../formats.md) owns the profile and prompt values
+that produce these outbound settings.
+
+## Endpoint And Method
+
+Generation sends an HTTP `POST` with `Content-Type: application/json`.
+Before the client is called, the engine resolves framework, backend, profile,
+and request values into one execution target. A non-empty endpoint from that
+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
+
+A non-empty API key supplied directly on the execution target takes
+precedence. Otherwise, when an API-key environment-variable name is supplied,
+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
+target's model takes precedence over the client's configured model, and one
+must be available.
+
+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.
+
+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 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.
+
+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`
+- `messages`
+- `temperature`
+- `max_tokens`
+- `top_p`
+- `service_tier`
+- `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
+returns the first choice's non-empty message content and maps prompt,
+completion, total, cached, and cache-write token counts.
+
+Invalid JSON, absent choices, and empty first-choice content are malformed
+responses. For a non-2xx status, the error includes the status code but never
+the provider response body.
+
+## Timeout And Cancellation
+
+Timeouts are layered:
+
+- the caller context remains the outer cancellation boundary;
+- a positive generation timeout adds a request context deadline;
+- zero adds no generation-specific deadline;
+- a negative generation timeout is invalid; and
+- the cloned `http.Client` supplies the whole-request transport cap, retaining
+  a positive supplied-client timeout or applying the configured/default
+  timeout when the supplied value is not positive.
+
+The earliest applicable caller, generation, or transport deadline controls the
+request. Constructing the internal client does not mutate a supplied
+`http.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

@@ -0,0 +1,66 @@
+# Internal Model Client
+
+## Purpose
+
+This document describes Promptkit's internal model-client implementation. The
+[architecture policy](../policy/architecture.md) owns the library boundary,
+and the
+[OpenAI-compatible chat integration](../integrations/openai-compatible-chat.md)
+owns the observable outbound HTTP contract. The
+[framework format reference](../formats.md) owns the profile and prompt
+settings consumed by the client.
+
+The concrete client remains under `internal/llm`. The root engine assembles it
+as the default implementation behind Promptkit's public client boundary.
+
+## Components And Flow
+
+`Client` is the provider-neutral generation boundary consumed by later
+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:
+
+1. validates request-level timeout and endpoint requirements;
+2. maps the internal request into the OpenAI-compatible chat payload;
+3. validates and merges extra parameters;
+4. resolves authentication;
+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.
+
+## Failure Categories
+
+The package preserves distinct error identities for invalid client
+configuration, invalid generation requests, request execution failures,
+non-success provider statuses, and malformed successful responses. Provider
+response bodies are not included in non-success errors.
+
+Caller cancellation and deadline failures during the outbound request are
+reported as request execution failures. The runner classifies these identities
+without depending on HTTP status mapping.
+
+## Test Ownership
+
+The
+[OpenAI-compatible client tests](../../internal/llm/openai_compatible_client_test.go)
+own configuration, client cloning, deterministic deadline precedence,
+authentication, request and response mapping, malformed data, error identity,
+cancellation, and response-body suppression. The root transport contract test
+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

@@ -2,12 +2,39 @@
 
 ## Purpose
 
-This is the inventory of this application's implemented components for contributors.
-The [architecture policy](../policy/architecture.md) owns normative boundaries
-and invariants; public behavior belongs in the linked contracts.
+This document inventories Promptkit's implemented components for contributors.
+The [architecture policy](../policy/architecture.md) owns durable boundary and
+dependency rules. See the [development guide](../development.md) for
+contributor workflow and validation.
 
-TODO: Add tables below, using the following format:
+## Implemented Components
 
 | Component | Implemented responsibility | References |
 | --- | --- | --- |
-| | | |
+| Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request and result values, profile construction, extension interfaces, value conversion, redacted formatting, public error mapping, and engine-local assembly. | [Package GoDoc](../../doc.go), [backend API](../../backends.go), [engine assembly](../../engine.go) |
+| `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, 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 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.
+
+## Maintenance
+
+Update this inventory as implementation adds packages or changes component
+responsibilities. List only implemented components; proposed package
+boundaries belong in temporary planning documents until their implementation
+lands.

docs/internal/runner.md

@@ -0,0 +1,159 @@
+# Internal Runner
+
+## Purpose
+
+This document describes Promptkit's implemented internal orchestration. The
+[architecture policy](../policy/architecture.md) owns dependency and consumer
+boundaries. The [source and validation document](sources.md) owns repository,
+artifact, rendering, and validation behavior, while the
+[model-client document](llm.md) owns generation behavior and failure
+categories.
+
+The runner remains under `internal/usecase` and is assembled by the root
+Promptkit engine. Its concrete type is not part of the public API.
+The [framework format reference](../formats.md) owns prompt, profile, schema,
+and override semantics consumed by the runner.
+
+## Collaborators
+
+`Runner` coordinates narrow internal interfaces for prompt definitions,
+profiles, backend resolution, artifacts, rendering, model generation, and
+validation. The root engine supplies one immutable registry containing the
+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.
+
+Each invocation carries its state in request, prepared-run, and result values.
+The runner has no durable run or session store.
+
+## Shared Preparation Pipeline
+
+`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 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 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. 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` 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 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, 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
+
+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.
+
+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, 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 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

@@ -0,0 +1,73 @@
+# Internal Sources And Validation
+
+## Purpose
+
+This document describes Promptkit's implemented internal source, artifact,
+rendering, and output-validation behavior. The
+[architecture policy](../policy/architecture.md) owns the library boundary and
+dependency rules. None of these internal packages is a supported consumer API,
+and the root engine assembles them behind its public source options and values.
+The [framework format reference](../formats.md) owns the exact file fields,
+validation modes, built-in catalog, and source precedence.
+
+## Prompt Definitions
+
+`internal/promptdef` discovers YAML deterministically, decodes and validates
+definitions, selects an ID and optional version, and resolves file-backed
+message content within the selected operating-system or `fs.FS` source.
+
+Its package tests own prompt selection, strict decoding, definition validation,
+duplicate detection, and source containment:
+[prompt-definition repository tests](../../internal/promptdef/repository_test.go).
+
+## Profiles And Built-Ins
+
+`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. 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. 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, 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
+
+`internal/artifact` resolves inline references and unrestricted,
+caller-selected file paths. It copies content into an artifact, records
+metadata and a content hash, applies a content-type fallback, and honors
+context cancellation.
+
+This ordinary reader does not implement an inbound HTTP security boundary. In
+particular, it does not constrain files to an application root or impose an
+HTTP request-size policy. Scriptorium's restricted HTTP reader remains an
+application concern outside Promptkit. The
+[artifact reader tests](../../internal/artifact/reader_test.go) own the
+implemented reader behavior and failures.
+
+## Rendering
+
+`internal/prompt` renders definition messages as Go templates using named
+artifacts and variables. It carries message roles, session IDs, and cache
+control into the rendered prompt. The
+[renderer tests](../../internal/prompt/renderer_test.go) own rendering behavior.
+
+## Schemas And Output Validation
+
+`internal/validate` provides validators backed by an operating-system
+filesystem or an `fs.FS`. Invalid generated content is returned as a validation
+result; inability to load, register, or compile a schema is an operational
+error.
+
+The [validator tests](../../internal/validate/standard_validator_test.go) own
+basic, JSON, JSON Schema, source resolution, schema loading, compilation, and
+content-failure behavior.

docs/policy/architecture.md

@@ -1,13 +1,172 @@
-# Architecture
+# Architecture Policy
 
-This document defines the intended high-level architecture of this application and the
-invariants that changes must preserve. Implemented component details belong in
-[Internal Overview](../internal/overview.md) and its linked documents. The
-reasoning behind significant architectural choices belongs in
-[ADRs](../adr/).
+## Purpose
+
+This document defines Promptkit's current high-level architecture and the
+durable boundaries that implementation changes must preserve. The
+[internal component overview](../internal/overview.md) inventories concrete
+implemented packages without redefining these rules.
 
 ## System Shape
 
-This application is a small, dependency-light Go application for ...
+Promptkit is an importable Go library. It does not ship a command, an HTTP
+service, or another application process. Repository examples demonstrate
+library use but are not Promptkit applications or release artifacts.
 
-TODO: Complete this document.
+The module root contains package `promptkit`, which is the public facade. It
+provides the supported engine, configuration and source options, requests,
+results, public values, extension interfaces, profiles, and error sentinels.
+
+The implemented internal components consist of:
+
+- `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
+  from filesystem and `fs.FS` sources;
+- `internal/profile/builtin`, which embeds the built-in execution profile
+  catalog;
+- `internal/prompt`, which renders prompt messages from Go templates;
+- `internal/artifact`, which resolves ordinary inline and unrestricted
+  caller-selected file references;
+- `internal/validate`, which validates basic, JSON, and JSON Schema output
+  using filesystem and `fs.FS` schema sources;
+- `internal/llm`, which defines the provider-neutral generation boundary and
+  implements outbound OpenAI-compatible chat requests; and
+- `internal/usecase`, which coordinates preparation and execution across the
+  internal framework components.
+
+The `examples/go-library/prepare` and `examples/go-library/run` packages are
+maintained downstream consumers of the root facade. They do not expose library
+packages or participate in internal assembly.
+
+The root facade assembles one immutable backend registry, one capacity manager,
+the internal repositories, renderer, validator, outbound client, and use-case
+runner while translating public values and errors at the library boundary. The
+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, including backend lookup and run admission.
+
+The current implementation follows this dependency direction:
+
+```text
+downstream consumers, including Scriptorium
+                    |
+                    v
+       root promptkit public facade
+                    |
+                    v
+       internal framework components
+                    |
+                    v
+        narrow injected abstractions
+```
+
+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
+
+Scriptorium is a downstream application that consumes Promptkit through
+the supported public facade. It is not a Promptkit package and must not become
+an internal dependency.
+
+Promptkit owns reusable, application-neutral library behavior. It does not own:
+
+- binaries or executable packaging;
+- CLI commands, parsing, streams, or exit codes;
+- HTTP routes, servers, request DTOs, status mapping, or deployment policy;
+- application configuration discovery or precedence;
+- process lifecycle, operational state, or application logging; or
+- consumer-specific filesystem or security policy.
+
+Those concerns remain with Scriptorium or another consuming application.
+
+## Package Ownership
+
+The module root is the supported public facade. Framework implementation
+packages belong under Go's `internal/` boundary unless a demonstrated, stable
+consumer contract requires a public package.
+
+Each package must have one cohesive responsibility and a clear dependency
+direction. Internal packages must not expose their types merely to simplify
+wiring, and the public facade must not leak internal representations through
+exported signatures. New public packages require a durable consumer need that
+cannot be served cleanly by the root facade.
+
+The [internal component overview](../internal/overview.md) must be updated as
+packages are implemented or their responsibilities change.
+
+## Exported API Discipline
+
+Export the smallest contract required by real consumers. Exported declarations
+must have accurate GoDoc, stable semantics, and tests proportionate to their
+compatibility risk. Avoid speculative extension points, aliases for internal
+types, and public constructors that expose assembly details.
+
+Once an exported API exists, its Go declaration and GoDoc own its exact public
+contract. Architecture documentation owns boundary rules, not a duplicate API
+reference.
+
+## Error Boundaries
+
+Internal failures must cross the public facade as errors meaningful to a Go
+consumer without exposing private package types or transport-specific policy.
+Wrapping should add useful context while preserving any public error identity
+needed with `errors.Is` or `errors.As`.
+
+Promptkit must not assign CLI exit codes or HTTP status codes. Consumers map
+public library outcomes into their own transport behavior.
+
+## Dependency Injection
+
+External effects and consumer-selected policy must enter through narrow
+interfaces or functions at the boundary that uses them. Dependencies should be
+explicitly supplied during construction or invocation rather than read from
+consumer configuration or hidden process-global state.
+
+Interfaces should be owned by the code that consumes the behavior and should
+contain only the operations that code requires. Provide defaults only for
+application-neutral behavior; consumer-specific restrictions and adapters
+remain injected from the consuming project.
+
+## Repository Independence
+
+Promptkit must build, test, and validate independently of Scriptorium. Do not
+commit `go.work`, `go.work.sum`, or a local filesystem `replace` directive.
+Temporary workspace or replacement configuration may support coordinated local
+development, but it is not part of either repository's architecture or release
+state.
+
+## Current-State Maintenance
+
+Do not list planned packages as implemented components. When implementation
+introduces a package, update the internal inventory and the owning contract or
+subsystem document in the same change.

docs/policy/documentation.md

@@ -2,112 +2,137 @@
 
 ## Purpose
 
-This policy assigns each documentation topic to one canonical owner. Its goal is
-to keep this application's documentation accurate, concise, discoverable, and resistant
-to drift for users, operators, developers, integrators, and LLM coding agents.
+This policy assigns each Promptkit documentation topic to one canonical owner.
+Its goal is to keep documentation for this reusable Go library accurate,
+concise, discoverable, and resistant to drift for consumers, contributors,
+maintainers, integrators, and coding agents.
 
 ## Core Rules
 
 ### One Canonical Owner
 
-Each authoritative fact belongs in one document. A non-owning document may give
-a short, stable summary for orientation, but it must link to the canonical owner
-instead of repeating volatile details.
+Each authoritative fact belongs in one document or source form. A non-owning
+document may give a short, stable summary for orientation, but it must link to
+the canonical owner instead of repeating exact contracts.
 
-Volatile details include commands, flags, configuration fields and defaults,
-module keys, schemas, file names, paths, status codes, retry behavior, and
-runtime guarantees. If readers could reasonably treat a statement as a
-contract, maintain it only in the owning document.
+Volatile details include exported declarations, accepted inputs, defaults,
+schemas, file names, paths, error identities, retry behavior, and runtime
+guarantees. If readers could reasonably treat a statement as a contract,
+maintain its exact definition only in the owning source.
 
-### Current And Future Behavior
+### Current State, Decisions, And Future Work
 
 Outside `docs/roadmap/`, documentation describes implemented behavior only.
 Partial features may be described only to their implemented boundary.
 
-ADRs are the narrow exception: an ADR may record an accepted architectural
-decision before implementation, but acceptance must not be presented as proof
-that the behavior exists. The roadmap owns implementation status and sequencing
-until the decision is implemented. Current architecture, user, operator,
-integration, and internal documentation are updated when the behavior lands.
+An accepted architecture decision may describe an approved direction before it
+is implemented, but acceptance is not evidence that the behavior exists.
+Current-state documents change when the implementation lands. Temporary
+roadmaps own future work, sequencing, and implementation status; they do not
+replace durable policies or current contracts.
 
 ### Audience And Detail
 
 Write for the document's stated audience and include only the detail needed for
-its owned topic. User and operator docs should not expose implementation detail.
-Developer docs should link to user-facing and external contracts rather than
-restate them.
+its owned topic. Consumer guidance should not expose incidental implementation
+detail. Contributor documentation should link to public contracts and durable
+policies instead of restating them.
 
-### Examples
+### Links
 
-Complete copyable files belong in `examples/`. Documentation may use the
-smallest illustrative snippet needed to explain its owned topic, but should link
-to maintained examples instead of embedding a second complete copy.
+Use descriptive link text and repository-relative links for repository
+documents. Link to the canonical owner rather than to a duplicate summary.
+Check every added or changed link and repair or remove links when their target
+moves or is retired.
 
-Examples must be valid, secret-free, and tested where practical. Commands and
-configuration used in documentation should match the application.
+### Examples And Code Fences
+
+Complete copyable files belong in `examples/` when maintained examples exist.
+Documentation may use the smallest illustrative snippet needed for its owned
+topic, but should link to a maintained example instead of embedding a second
+complete copy.
+
+Examples must be valid, secret-free, and tested where practical. Commands,
+imports, and Go snippets must match the implemented library. Use a language tag
+on fenced code blocks and make clear when a fragment is illustrative rather
+than directly runnable.
 
 ### Security And Privacy
 
 Documentation and examples must not contain real credentials, private keys,
-private environment dumps, sensitive source material, or private infrastructure
-details unless intentionally public. Document secret-handling mechanisms, not
-secret values.
+private environment dumps, sensitive source material, or private
+infrastructure details unless intentionally public. Document secret-handling
+mechanisms, not secret values.
 
 ## Canonical Ownership
 
 | Topic | Canonical owner | Owned content | Content owned elsewhere |
 | --- | --- | --- | --- |
-| Product orientation and minimal end-to-end quickstart | `README.md` | What this application is, why it is useful, one shortest successful invocation, and links onward. | Complete command reference, configuration reference, operational procedures, implementation detail. |
-| Contributor entry point | `docs/development.md` | Task-oriented reading guide, minimal contributor orientation, baseline validation commands, and links to canonical docs. | Package inventory, architecture rules, subsystem behavior, detailed change recipes. |
-| Current application architecture | `docs/policy/architecture.md` | System shape, normative ownership, dependency direction, architectural boundaries, invariants, safety properties, and non-goals. | Concrete package inventory, implementation mechanics, contributor procedures, decision history, future work. |
-| Documentation organization | `docs/policy/documentation.md` | Documentation ownership, audience boundaries, maintenance rules, and ADR/document lifecycle. | Application architecture or product behavior. |
-| Testing policy | `docs/policy/testing.md` | Test philosophy, risk-based sufficiency, test boundaries, doubles, coverage guidance, regression-test policy, and criteria for adding, rewriting, or deleting tests. | Subsystem behavior, application contracts, subsystem-specific test inventories, and implementation plans. |
-| CLI contract | `docs/cli.md` | Commands, arguments, flags, invocation semantics, and exit codes. | End-to-end operating procedures, configuration field definitions, runtime filesystem layout, module implementation details. |
-| Configuration contract | `docs/config.md` | Discovery and precedence, file schema, fields, defaults, environment overrides, validation rules, and user-selectable module or validator keys. | Complete example files, CLI syntax, runtime state lifecycle, module implementation details. |
-| Operations | `docs/operations.md` | Runtime workflows, physical filesystem and state layout, output, cache, and debug handling, resume, cleanup, permissions, recovery, and operational limits. | CLI flag syntax, configuration field definitions, logical output schemas, implementation mechanics. |
-| Public HTTP contract, if introduced | `docs/api.md` | Routes, authentication, media types, request and response schemas, status codes, pagination, caching, idempotency, rate limits, and HTTP retry semantics. | Client walkthroughs, upstream or downstream integration internals, implementation detail. |
-| Consumer guidance, if a public package or API is introduced | `docs/consumers/` | Task-oriented use of the public interface, minimal client examples, and consumer responsibilities. | HTTP wire semantics, external protocol contracts, internal implementation detail. |
-| External and durable integration contracts | `docs/integrations/` | External file formats and protocols, upstream and downstream contracts, logical output bundle paths and schemas, media types, and compatibility behavior. | Physical runtime placement and lifecycle, internal transformations, CLI syntax, configuration defaults. |
-| Implemented component inventory | `docs/internal/overview.md` | Current packages and components, their implemented responsibilities, and links to focused internal docs. | Normative architecture, contributor reading policy, external contracts. |
-| Internal component behavior | Other files under `docs/internal/` | Implementation flow, internal collaborators and state transitions, package-local guarantees and failures, and relevant tests. | Global architecture invariants, configuration definitions and defaults, external schemas, operator procedures. |
-| Architectural decision history | `docs/adr/` | Significant decisions, context, alternatives, rationale, consequences, and supersession history. | Current behavior reference, implementation status, task sequencing. |
-| Future work and implementation status | `docs/roadmap/` | Proposed, accepted, deferred, or rejected work; implementation status; sequencing; and task breakdowns. | Implemented behavior reference and architectural decision rationale. |
-| Complete copyable artifacts | `examples/` | Maintained configuration, inputs, and other files intended to be copied or run. | Field-by-field reference, command reference, prose explanation. |
+| Project orientation | `README.md` | What Promptkit is, its current usability, module identity, license summary, and links onward. | Exact API contracts, contributor procedures, architecture detail, and release steps. |
+| Contributor workflow | `docs/development.md` | Task-oriented reading guide, local workflow, validation commands, and repository hygiene. | Architecture rules, API semantics, subsystem behavior, and release procedure. |
+| Current architecture | `docs/policy/architecture.md` | System shape, normative ownership, dependency direction, package boundaries, invariants, and non-goals. | Concrete component inventory, implementation mechanics, contributor procedures, decision history, and future work. |
+| Documentation organization | `docs/policy/documentation.md` | Documentation ownership, audience boundaries, maintenance rules, and document lifecycle. | Library architecture or runtime behavior. |
+| Testing policy | `docs/policy/testing.md` | Test philosophy, risk-based sufficiency, test boundaries, doubles, coverage guidance, regression policy, and test maintenance. | Subsystem behavior, exact public contracts, subsystem-specific test inventories, and implementation plans. |
+| Release procedure | `docs/release.md`, when present | Required release validation, version and tag procedure, release ordering, and post-publication checks. | General contributor workflow, public API semantics, and decision history. |
+| Exact exported Go API | Go declarations and GoDoc, as APIs are implemented | Exported names, signatures, types, values, errors, and exact behavioral contracts. | Task-oriented consumer walkthroughs, implementation details, and future API proposals. |
+| Framework file formats | `docs/formats.md` | Prompt-definition and profile YAML fields, schema references, defaults, validation modes, built-in profiles, credentials, and file-to-request precedence. | Exported Go declarations, outbound wire behavior, internal parsing mechanics, and application configuration. |
+| Consumer guidance | `docs/consumers/`, when consumer workflows require dedicated guidance | Task-oriented use of implemented public APIs, minimal examples, and consumer responsibilities. | Exact exported declarations and internal mechanics. |
+| Durable integration contracts | `docs/integrations/`, when integrations exist | External formats and protocols, compatibility behavior, and upstream or downstream responsibilities. | Internal transformations and public Go declarations. |
+| Supplemental release guidance | None. `docs/releases/` may be used when a release benefits from a changelog or migration guide. | No canonical content. These files may briefly summarize release-specific changes, compatibility, and consumer migration paths, and may be corrected, consolidated, archived, or removed when no longer useful. | Public API and behavior contracts, formats, integrations, architecture, release procedure, and the authoritative annotated-tag release record. |
+| 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. |
+| Temporary feature roadmaps | `docs/roadmap/`, while planned work needs coordination | Proposed or accepted scope, sequencing, gates, and implementation status. | Implemented behavior reference and durable decision rationale. |
+| 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. |
 
-Documents that do not exist are required only when the corresponding interface
-or responsibility exists. Do not create placeholder API, consumer, integration,
-or operations documents for behavior the application does not have.
+Conditional owners do not require placeholder files or directories. Create a
+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
 
 ### Orientation
 
-The README owns product orientation. The developer guide routes contributors.
-Architecture owns normative structure. Internal overview owns the current
-concrete component map. These documents may link to one another but should not
-maintain parallel package or behavior descriptions.
+The README owns project orientation. The development guide routes
+contributors. Architecture owns normative structure. The internal overview
+owns the current concrete component map. These documents may link to one
+another but must not maintain parallel package or behavior descriptions.
 
-### Commands, Configuration, And Operations
+### Public Contracts And Implementation
 
-CLI documentation answers how to invoke the application. Configuration
-documentation answers what settings mean. Operations answers what happens to
-runtime state and how to operate or recover the application. When a workflow
-crosses these topics, choose the document that owns the task and link to the
-other contracts.
+Go declarations and GoDoc own exact exported API contracts once those APIs
+exist. Consumer and integration documents explain how to use those contracts
+for a task. Internal documents explain how Promptkit implements them. Internal
+documentation may identify a public type or external format as a dependency,
+but must link to its canonical definition rather than restate it.
 
-### Contracts And Implementation
+The [framework format reference](../formats.md) owns exact prompt, profile, and
+schema-file contracts. Integration documents own external wire formats.
 
-Integration and API documents define externally observable shapes and
-semantics. Internal documents explain how thos application implements or consumes those
-contracts. Internal docs may name a field, file, or protocol to identify a
-dependency, but must link to its canonical contract for the definition.
+### 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
-application security invariants. Configuration owns credential-supply
-mechanisms. Operations owns permissions and handling of sensitive runtime
-artifacts. Internal docs own implementation mechanisms only.
+library security boundaries and invariants. Public declarations and integration
+documents own consumer-visible security contracts. Internal documents own
+implementation mechanisms only.
 
 ## Architecture Decision Records
 
@@ -122,23 +147,48 @@ Use sequentially numbered ADR filenames such as
 6. alternatives considered;
 7. consequences.
 
-Treat the decision content of an accepted ADR as immutable. When a decision
-changes, create a new ADR and update the earlier ADR's status to superseded.
-Rejected architectural alternatives belong in the ADR; rejected product ideas
-belong in the roadmap.
+Use one of these statuses:
 
-## Maintenance
+- **Proposed:** the decision is under consideration and may change;
+- **Accepted:** the decision is approved, whether or not implementation is
+  complete;
+- **Rejected:** the proposed decision was considered and not adopted;
+- **Superseded:** a later accepted ADR replaces the accepted decision.
 
-When behavior changes, update its canonical owner in the same change. If
-ownership moves, remove the old definition and replace it with a link where
-navigation remains useful.
+A proposed ADR transitions to Accepted or Rejected. An Accepted ADR transitions
+to Superseded only when a later Accepted ADR replaces it. An ADR may be created
+as Accepted when the decision has already been made.
+
+Treat the decision content of an Accepted ADR as immutable. A changed decision
+requires a later ADR rather than a rewrite of the accepted record. A Superseded
+ADR must link to its replacement, and the replacement must link back. Rejected
+architectural alternatives belong in the ADR; rejected feature ideas belong in
+a roadmap when they need to be retained.
+
+## Document Lifecycle
+
+Create durable current-state documentation with the implementation it
+describes. Update its canonical owner in the same change when behavior changes.
+If ownership moves, remove the old definition and leave a link where navigation
+remains useful.
+
+Roadmaps are temporary coordination documents. When their work is complete,
+record completion, move any still-useful decisions or contracts to their
+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;
-- check commands, flags, fields, defaults, schemas, and paths against their
-  implementation;
-- keep unimplemented behavior in the roadmap, subject to the ADR exception;
-- remove stale references and validate links;
-- confirm that non-owning documents summarize and link rather than redefine;
+- check commands, imports, declarations, defaults, schemas, and paths against
+  their implementation;
+- keep unimplemented behavior in a roadmap, subject to the ADR exception;
+- validate links and fenced examples;
+- confirm non-owning documents summarize and link rather than redefine;
+- remove stale or unsupported claims; and
 - confirm that no secrets or sensitive private data were added.

docs/policy/testing.md

@@ -14,7 +14,7 @@ A test must be:
 
 - written and reviewed;
 - understood by future maintainers and coding agents;
-- executed in local and CI workflows;
+- executed in maintainer-run validation;
 - diagnosed when it fails;
 - updated when legitimate behavior changes;
 - maintained as fixtures, APIs, and dependencies evolve; and
@@ -49,9 +49,53 @@ Examples of appropriate seams include clocks, randomness, subprocesses, remote A
 
 ## Test execution requirements
 
-Tests in the default suite must be deterministic, offline, and independent of real credentials. They must not invoke paid APIs or depend on mutable external services. Tests that require live infrastructure must be explicitly opt-in and clearly separated from the default suite.
+Promptkit currently uses maintainer-run validation rather than hosted CI.
+Maintainers run the repository-documented test, vet, build, formatting,
+documentation-link, and repository-hygiene checks before accepting changes.
+Introducing hosted CI later would supplement, not silently redefine, this
+documented validation model.
 
-Control clocks, randomness, environment variables, and other process-global or machine-specific state when they affect behavior. Tests should be safe to run repeatedly and alongside other tests without depending on execution order or state left by an earlier test.
+The complete test sequence includes ordinary and race-enabled package tests.
+The maintained offline consumer workflow is also run from the repository root:
+
+```sh
+go test ./...
+go test -race ./...
+go run ./examples/go-library/prepare
+```
+
+Tests in the default suite must be deterministic, offline, and independent of
+real credentials. They must not invoke paid APIs, use live network
+dependencies, or depend on mutable external services. Tests that require live
+infrastructure must be explicitly opt-in and clearly separated from the
+default suite.
+
+Control clocks, randomness, environment variables, and other process-global or
+machine-specific state when they affect behavior. Tests must be parallel-safe:
+they should run repeatedly and alongside other tests without depending on
+execution order, shared mutable state, fixed ports, or state left by an earlier
+test.
+
+## Test types and assets
+
+Use each test type where it protects a distinct risk:
+
+- Unit and package tests protect focused behavior and invariants through the
+  narrowest stable boundary.
+- Contract tests protect exported behavior, compatibility, and error identity
+  relied upon by consumers.
+- Integration tests use real collaborators when correctness depends on their
+  interaction, while replacing live or nondeterministic external boundaries.
+- External-package root tests exercise the public facade as a Go consumer,
+  while internal package tests own focused implementation behavior.
+- The maintained offline preparation example protects one representative
+  assembled consumer workflow without contacting a model provider.
+- Fixtures should be minimal, synthetic, versioned with the behavior they
+  exercise, and free of credentials or private data.
+- Golden files are appropriate only when the complete output is intentionally
+  stable and semantic review of updates is practical.
+- Failure-path tests should cover consequential malformed input, dependency
+  failure, cancellation, partial results, and recovery behavior.
 
 ## What deserves tests
 
@@ -63,7 +107,7 @@ Prioritize tests for:
 4. Failure handling, cancellation, retries, recovery, and partial success.
 5. Serialization, schemas, compatibility, and round trips.
 6. Previously observed or plausible regressions.
-7. Representative integration and end-to-end workflows.
+7. Representative integration and consumer workflows.
 
 A package-level contract is behavior relied upon by another package or major collaborator, not every observable detail of a package implementation.
 
@@ -81,7 +125,9 @@ This is often the package API, but it may instead be:
 - a package-level operation when several internal collaborators jointly produce the behavior; or
 - a larger integration boundary when correctness emerges from interaction with a real dependency.
 
-Do not force all behavior through oversized end-to-end tests. Do not test every private helper merely because it exists. Choose the boundary that gives durable confidence with the least incidental coupling.
+Do not force all behavior through oversized consumer-workflow tests. Do not
+test every private helper merely because it exists. Choose the boundary that
+gives durable confidence with the least incidental coupling.
 
 ## Test behavior, not implementation
 
@@ -121,7 +167,8 @@ A test failing is not the same as a test needing to be edited. Many tests may co
 
 Configurable thresholds and defaults must not be duplicated throughout the test suite.
 
-For example, do not encode an internal concurrency limit indirectly:
+The following fragments are illustrative rather than standalone Go programs.
+Do not encode an internal concurrency limit indirectly:
 
 ```go
 // Production policy:
@@ -165,10 +212,10 @@ Each behavior should have a clear test owner.
 - Parser tests own parsing cases.
 - Validator tests own validation rules.
 - Domain tests own transformations and invariants.
-- Adapter tests own external integration behavior.
+- Boundary tests own external integration behavior.
 - Orchestrator tests own coordination and failure propagation.
-- CLI tests own argument and configuration mapping.
-- End-to-end tests prove that representative assembled workflows work.
+- Consumer-workflow tests prove that representative assembled library use
+  works.
 
 Higher-level tests should not repeat every lower-level case. A single intentional policy change should not require unrelated edits across many test files.
 
@@ -205,13 +252,17 @@ Use:
 - fuzz tests for parsers, normalization, path handling, and broad input spaces;
 - golden files only when the complete output is intentionally stable;
 - integration tests where correctness depends on component interaction; and
-- a small number of representative end-to-end tests.
+- a small number of representative consumer-workflow tests.
 
 Avoid exact error-string assertions unless the wording is itself contractual. Prefer `errors.Is`, `errors.As`, typed errors, or structured error fields.
 
-At CLI boundaries, prefer exit classifications, structured output, and the smallest stable semantic fragment needed to identify the error. Do not snapshot complete diagnostic wording unless it is contractual.
+At public API boundaries, prefer stable error identity, structured values, and
+the smallest semantic fragment needed to identify the failure. Do not snapshot
+complete diagnostic wording unless it is contractual.
 
-Golden-file updates must require an explicit local flag. CI must not update golden files automatically, and reviewers must inspect the semantic diff before accepting an update.
+Golden-file updates must require an explicit local flag. Ordinary validation
+runs must never update golden files automatically, and maintainers must inspect
+the semantic diff before accepting an update.
 
 Keep tests readable and direct. Test helpers and fixture frameworks must earn their own maintenance cost; do not build elaborate test infrastructure for small or isolated needs.
 
@@ -221,7 +272,8 @@ Coverage is a diagnostic, not a target.
 
 Use it to find untested critical branches and unexpectedly weak packages. Do not write low-value tests solely to increase a percentage, and do not infer test quality from coverage alone.
 
-Pure domain logic will often warrant higher coverage than CLI wiring or external adapters. Uneven coverage is acceptable when it reflects risk.
+Pure domain logic will often warrant higher coverage than facade wiring or
+external adapters. Uneven coverage is acceptable when it reflects risk.
 
 Increasing coverage is valuable only when the newly covered behavior protects a meaningful risk at an acceptable cost.
 
@@ -289,7 +341,9 @@ A test suite is sufficient when:
 - legitimate internal changes usually do not require test edits; and
 - additional tests would mostly repeat existing protection or preserve inconsequential implementation details.
 
-Sufficiency is a risk judgment, not a coverage percentage or test count. Reassess it as the application, its users, and the consequences of failure evolve.
+Sufficiency is a risk judgment, not a coverage percentage or test count.
+Reassess it as the library, its consumers, and the consequences of failure
+evolve.
 
 The governing rule is:
 

docs/release.md

@@ -0,0 +1,278 @@
+# Release Procedure
+
+## Release Model
+
+Promptkit publishes a Go library through source commits and semantic Go module
+tags. It does not publish runnable binaries or binary packages and does not
+currently use hosted CI. The release maintainer performs and records the
+required validation.
+
+`v0.1.0` is the initial published release. Later releases use semantic
+`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.
+
+Promptkit releases are source-only. The annotated tag message is the release
+note; there is no separate hosted release or binary packaging step.
+
+## Establish The Candidate
+
+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
+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: $RELEASE_VERSION" >&2
+        return 1
+    fi
+}
+
+check_release_candidate
+```
+
+Do not continue unless the guard completes successfully. In particular, push
+the intended commit through the normal `main` branch workflow before release;
+the tag procedure is not a substitute for publishing the source commit.
+
+## Validate The Candidate
+
+Confirm the module and root package metadata:
+
+```sh
+go list -m -f '{{.Path}} {{.GoVersion}}'
+go list -f '{{.Name}} {{.ImportPath}}' .
+```
+
+The output must be:
+
+```text
+gitea.maximumdirect.net/eric/promptkit 1.25.5
+promptkit gitea.maximumdirect.net/eric/promptkit
+```
+
+Run the complete maintainer validation required by the
+[development guide](development.md):
+
+```sh
+go test ./...
+go test -race ./...
+go vet ./...
+go build ./...
+go run ./examples/go-library/prepare
+```
+
+Check every tracked Go file. This command must produce no output:
+
+```sh
+unformatted=$(
+    git ls-files '*.go' |
+        while IFS= read -r go_file
+        do
+            gofmt -l "$go_file"
+        done
+)
+test -z "$unformatted"
+```
+
+Follow every maintained Markdown link and confirm that its local or published
+target exists. Review the repository for generated binaries, test or coverage
+output, credentials, template residue, downloaded assets, and other files that
+do not belong in source control.
+
+Recheck module and repository hygiene, whitespace, and the clean checkout:
+
+```sh
+test -z "$(git ls-files go.work go.work.sum)"
+test ! -e vendor
+if grep -Eq '^[[:space:]]*replace([[:space:]]|\()' go.mod
+then
+    printf '%s\n' 'go.mod contains a replacement' >&2
+    exit 1
+fi
+git diff --check
+test -z "$(git status --porcelain)"
+```
+
+## Write The Release Note
+
+Prepare a plain-text annotated-tag message outside the repository and export
+its path as `RELEASE_NOTES_FILE`. Use this form, replacing each summary with
+release-specific text; write `None.` when there are no public API changes or
+consumer actions:
+
+```text
+Promptkit vMAJOR.MINOR.PATCH
+
+Validated commit: full commit ID
+Compatibility: compatibility summary
+Public API changes: changes or None.
+Consumer action: required action or None.
+```
+
+After writing it, require all release-note fields, the selected version, and
+the validated commit to be present:
+
+```sh
+: "${RELEASE_NOTES_FILE:?export the path to the release-note file}"
+test -f "$RELEASE_NOTES_FILE"
+test -s "$RELEASE_NOTES_FILE"
+grep -F "Promptkit $RELEASE_VERSION" "$RELEASE_NOTES_FILE"
+grep -F "Validated commit: $RELEASE_COMMIT" "$RELEASE_NOTES_FILE"
+grep -F 'Compatibility:' "$RELEASE_NOTES_FILE"
+grep -F 'Public API changes:' "$RELEASE_NOTES_FILE"
+grep -F 'Consumer action:' "$RELEASE_NOTES_FILE"
+```
+
+Inspect the complete message and confirm that it accurately records the
+compatibility impact, public API changes, and required consumer action.
+
+## Create And Inspect The Tag
+
+Run the candidate guard again immediately before tag creation. This ensures
+that validation or release-note preparation did not change the checkout and
+that the commit is still published and untagged:
+
+```sh
+check_release_candidate
+```
+
+Create the annotated tag from the prepared release note and bind it explicitly
+to the validated commit:
+
+```sh
+git tag --annotate "$RELEASE_VERSION" \
+    --file "$RELEASE_NOTES_FILE" \
+    "$RELEASE_COMMIT"
+```
+
+Inspect both the tag message and its source commit before publication:
+
+```sh
+test "$(git cat-file -t "refs/tags/$RELEASE_VERSION")" = tag
+git show --no-patch --decorate "refs/tags/$RELEASE_VERSION"
+test "$(
+    git rev-parse --verify "refs/tags/$RELEASE_VERSION^{commit}"
+)" = "$RELEASE_COMMIT"
+```
+
+If inspection finds an error, delete the unpublished local tag, correct the
+release note or candidate, and repeat the guards. Never move or recreate a tag
+that has been published.
+
+## Publish The Selected Tag
+
+Push only the selected tag ref. Do not use `git push --tags`:
+
+```sh
+git push origin \
+    "refs/tags/$RELEASE_VERSION:refs/tags/$RELEASE_VERSION"
+```
+
+## Verify Publication
+
+Compare the remote annotated-tag object with the local object, then compare the
+remote peeled source commit with the validated commit:
+
+```sh
+remote_tag=$(
+    git ls-remote --tags origin "refs/tags/$RELEASE_VERSION" |
+        awk 'NR == 1 { print $1 }'
+)
+remote_commit=$(
+    git ls-remote --tags origin "refs/tags/$RELEASE_VERSION^{}" |
+        awk 'NR == 1 { print $1 }'
+)
+test -n "$remote_tag"
+test "$remote_tag" = \
+    "$(git rev-parse --verify "refs/tags/$RELEASE_VERSION")"
+test "$remote_commit" = "$RELEASE_COMMIT"
+```
+
+Finally, resolve the version as an ordinary Go module in a temporary module
+outside this repository and without a workspace or replacement:
+
+```sh
+resolution_dir=$(mktemp -d)
+(
+    trap 'rm -rf "$resolution_dir"' 0 1 2 15
+    cd "$resolution_dir"
+    GOWORK=off go mod init example.com/promptkit-release-check
+    GOWORK=off go mod download \
+        "gitea.maximumdirect.net/eric/promptkit@$RELEASE_VERSION"
+    resolved_version=$(
+        GOWORK=off go list -m -f '{{.Version}}' \
+            "gitea.maximumdirect.net/eric/promptkit@$RELEASE_VERSION"
+    )
+    test "$resolved_version" = "$RELEASE_VERSION"
+)
+```
+
+Promptkit must publish and verify the required version before Scriptorium or
+another consumer publishes a release that depends on it. This ordering does
+not replace the consumer project's own release procedure. Released consumers
+must select the published Promptkit tag through ordinary module resolution,
+without a workspace, replacement, vendored Promptkit source, or unpublished
+revision.
+
+## Policy Changes
+
+Document and approve a durable policy change before introducing hosted
+automation, binary artifacts, or different release governance. Update this
+procedure in the same change so maintainers do not rely on hidden release
+requirements.

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

@@ -0,0 +1,493 @@
+package promptkit
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io/fs"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	artifactadapter "gitea.maximumdirect.net/eric/promptkit/internal/artifact"
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/capacity"
+	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/llm"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile/builtin"
+	"gitea.maximumdirect.net/eric/promptkit/internal/prompt"
+	"gitea.maximumdirect.net/eric/promptkit/internal/promptdef"
+	"gitea.maximumdirect.net/eric/promptkit/internal/usecase"
+	"gitea.maximumdirect.net/eric/promptkit/internal/validate"
+)
+
+// ErrInvalidConfig identifies invalid engine construction, including missing
+// required configuration, invalid options or backend registrations, and a nil
+// Engine receiver.
+var ErrInvalidConfig = errors.New("invalid engine configuration")
+
+var (
+	// 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 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 selects the directory-backed sources and built-in model-client
+// transport used by [NewEngine]. Config has no stable JSON representation.
+type Config struct {
+	// 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 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. 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 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
+}
+
+type optionFunc func(*engineOptions) error
+
+func (f optionFunc) apply(options *engineOptions) error {
+	return f(options)
+}
+
+type engineOptions struct {
+	llmClient       llm.Client
+	artifactReader  artifactadapter.Reader
+	promptDefs      promptdef.Repository
+	profiles        profile.Repository
+	memoryProfiles  profile.Repository
+	backends        []domain.Backend
+	validator       validate.Validator
+	promptSource    bool
+	profileSource   bool
+	memorySource    bool
+	validatorSource bool
+	artifactSource  bool
+}
+
+// WithLLMClient replaces the built-in model client used by [Engine.Run].
+//
+// 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 {
+			return ErrInvalidConfig
+		}
+		options.llmClient = publicLLMClientAdapter{client: client}
+		return nil
+	})
+}
+
+// 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 {
+			return ErrInvalidConfig
+		}
+		options.artifactReader = publicArtifactReaderAdapter{reader: reader}
+		options.artifactSource = true
+		return nil
+	})
+}
+
+// WithPromptFS loads prompt definitions from fsys under root.
+//
+// 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 {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.promptDefs = promptdef.NewFSRepository(fsys, root)
+		options.promptSource = true
+		return nil
+	})
+}
+
+// WithPromptFile loads prompt definitions from the single prompt file at path.
+//
+// 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)
+		if err != nil {
+			return err
+		}
+		options.promptDefs = promptdef.NewFSRepository(fsys, root)
+		options.promptSource = true
+		return nil
+	})
+}
+
+// WithProfileFS loads execution profiles from fsys under root.
+//
+// Profiles from this source overlay built-in profiles. Profile YAML must use
+// 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 {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.profiles = profile.NewFSRepository(fsys, root)
+		options.profileSource = true
+		return nil
+	})
+}
+
+// WithProfileFile loads execution profiles from the single profile file at path.
+//
+// The profile overlays built-in profiles. Profile YAML must use api_key_env for
+// 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)
+		if err != nil {
+			return err
+		}
+		options.profiles = profile.NewFSRepository(fsys, root)
+		options.profileSource = true
+		return nil
+	})
+}
+
+// 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)
+		if err != nil {
+			return err
+		}
+		options.memoryProfiles = repo
+		options.memorySource = true
+		return nil
+	})
+}
+
+// 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. 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 {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.validator = validate.NewFSValidator(fsys, root)
+		options.validatorSource = true
+		return nil
+	})
+}
+
+// WithSchemaFile loads JSON Schema documents from the single schema file at path.
+//
+// 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)
+		if err != nil {
+			return err
+		}
+		options.validator = validate.NewFSValidator(fsys, root)
+		options.validatorSource = true
+		return nil
+	})
+}
+
+// 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 {
+		if opt == nil {
+			continue
+		}
+		if err := opt.apply(&options); err != nil {
+			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
+		}
+	}
+
+	promptDefs := options.promptDefs
+	if !options.promptSource {
+		if strings.TrimSpace(cfg.PromptDir) == "" {
+			return nil, fmt.Errorf("%w: prompt directory is required", ErrInvalidConfig)
+		}
+		promptDefs = promptdef.NewFilesystemRepository(cfg.PromptDir)
+	}
+
+	profiles := builtin.NewRepositoryWithDirectory(cfg.ProfileDir)
+	if options.profileSource {
+		profiles = builtin.NewRepositoryWithPrimary(options.profiles)
+	}
+	if options.memorySource {
+		profiles = profile.NewOverlayRepository(options.memoryProfiles, profiles)
+	}
+
+	backendRegistry, err := backend.NewRegistry(options.backends)
+	if err != nil {
+		return nil, fmt.Errorf("%w: failed to construct backend registry: %v", ErrInvalidConfig, err)
+	}
+
+	capacityManager, err := capacity.NewManager(backendRegistry.CapacityPolicies())
+	if err != nil {
+		return nil, fmt.Errorf("%w: failed to construct backend capacity manager: %v", ErrInvalidConfig, err)
+	}
+
+	validator := options.validator
+	if !options.validatorSource {
+		schemaDir := cfg.SchemaDir
+		if strings.TrimSpace(schemaDir) == "" {
+			schemaDir = defaults.SchemaDirDefault
+		}
+		validator = validate.NewStandardValidator(schemaDir)
+	}
+
+	llmClient := options.llmClient
+	if llmClient == nil {
+		var err error
+		llmClient, err = llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
+			Timeout:    cfg.Timeout,
+			HTTPClient: cfg.HTTPClient,
+		})
+		if err != nil {
+			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
+		}
+	}
+	llmClient = capacity.NewClient(capacityManager, llmClient)
+
+	artifacts := options.artifactReader
+	if !options.artifactSource {
+		artifacts = artifactadapter.NewCompositeReader()
+	}
+
+	return &Engine{
+		runner: usecase.NewRunner(
+			promptDefs,
+			profiles,
+			backendRegistry,
+			artifacts,
+			prompt.NewGoRenderer(),
+			llmClient,
+			validator,
+			capacityManager,
+		),
+	}, nil
+}
+
+func fileSource(name string) (fs.FS, string, error) {
+	cleanName := strings.TrimSpace(name)
+	if cleanName == "" {
+		return nil, "", ErrInvalidConfig
+	}
+	dir := filepath.Dir(cleanName)
+	base := filepath.Base(cleanName)
+	if base == "." || base == string(filepath.Separator) || strings.TrimSpace(base) == "" {
+		return nil, "", ErrInvalidConfig
+	}
+	info, err := os.Stat(cleanName)
+	if err != nil {
+		return nil, "", fmt.Errorf("%w: failed to access source file %q: %v", ErrInvalidConfig, cleanName, err)
+	}
+	if info.IsDir() {
+		return nil, "", fmt.Errorf("%w: source path %q must be a file", ErrInvalidConfig, cleanName)
+	}
+	return os.DirFS(dir), filepath.ToSlash(base), nil
+}
+
+// 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)
+	}
+
+	domainReq, err := toDomainRunRequest(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	prepared, err := e.runner.Prepare(ctx, domainReq)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return fromDomainPreparedRun(prepared), nil
+}
+
+// Run prepares a request, invokes the configured LLMClient, and validates the
+// 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)
+	}
+
+	domainReq, err := toDomainRunRequest(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	result, err := e.runner.Run(ctx, domainReq)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return fromDomainRunResult(result), nil
+}

errors.go

@@ -0,0 +1,63 @@
+package promptkit
+
+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"
+)
+
+func mapPublicError(err error) error {
+	if err == nil {
+		return nil
+	}
+	publicErr := publicErrorFor(err)
+	if publicErr == nil {
+		return err
+	}
+	return fmt.Errorf("%w: %w", publicErr, err)
+}
+
+func publicErrorFor(err error) error {
+	switch {
+	case errors.Is(err, promptdef.ErrPromptDefinitionNotFound):
+		return ErrPromptNotFound
+	case errors.Is(err, profile.ErrProfileNotFound):
+		return ErrProfileNotFound
+	case errors.Is(err, usecase.ErrProfileRequired):
+		return errors.Join(ErrInvalidRequest, ErrProfileRequired)
+	case errors.Is(err, usecase.ErrPromptLoad):
+		return ErrPromptLoad
+	case errors.Is(err, usecase.ErrProfileLoad):
+		return ErrProfileLoad
+	case errors.Is(err, promptdef.ErrInvalidYAML), errors.Is(err, promptdef.ErrInvalidPromptDefinition):
+		return ErrPromptLoad
+	case isProfileLoadCause(err):
+		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):
+		return ErrPromptRender
+	case errors.Is(err, usecase.ErrLLMGenerate):
+		return ErrLLMGenerate
+	case errors.Is(err, usecase.ErrValidation):
+		return ErrValidation
+	case errors.Is(err, usecase.ErrInvalidRequest):
+		return ErrInvalidRequest
+	default:
+		return nil
+	}
+}
+
+func isProfileLoadCause(err error) bool {
+	return errors.Is(err, profile.ErrInvalidYAML) ||
+		errors.Is(err, profile.ErrInvalidProfile) ||
+		errors.Is(err, profile.ErrRawAPIKeyNotAllowed)
+}

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/prepare/main.go

@@ -0,0 +1,60 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"os"
+
+	"gitea.maximumdirect.net/eric/promptkit"
+)
+
+type summary struct {
+	PromptID        string `json:"prompt_id"`
+	PromptVersion   string `json:"prompt_version"`
+	SelectedProfile string `json:"selected_profile"`
+	Model           string `json:"model"`
+	MessageCount    int    `json:"message_count"`
+}
+
+func main() {
+	engine, err := promptkit.NewEngine(
+		promptkit.Config{},
+		promptkit.WithPromptFile("examples/go-library/prepare/prompt.yaml"),
+		promptkit.WithProfiles(promptkit.Profile{
+			ID:       "offline-example",
+			Endpoint: "https://example.invalid/v1",
+			Model:    "offline-model",
+		}),
+	)
+	if err != nil {
+		exit(err)
+	}
+
+	prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
+		PromptID: "example.prepare",
+		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{
+		PromptID:        prepared.PromptID,
+		PromptVersion:   prepared.PromptVersion,
+		SelectedProfile: prepared.SelectedProfileID,
+		Model:           prepared.EffectiveModelParams.Model,
+		MessageCount:    len(prepared.Messages),
+	}); err != nil {
+		exit(err)
+	}
+}
+
+func exit(err error) {
+	fmt.Fprintln(os.Stderr, err)
+	os.Exit(1)
+}

examples/go-library/prepare/prompt.yaml

@@ -0,0 +1,16 @@
+id: example.prepare
+version: "1.0.0"
+default_profile: offline-example
+description: Prepare a prompt without contacting a model provider.
+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

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

@@ -0,0 +1,56 @@
+package promptkit
+
+import "fmt"
+
+// 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 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}",
+		r.PromptID,
+		r.PromptVersion,
+		r.ProfileID,
+		r.APIKey != "",
+		len(r.Inputs),
+		len(r.Vars),
+		r.Execution != nil,
+		r.Validation != nil,
+	)
+}
+
+// String returns a concise request summary without exposing direct API keys or
+// 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. Reflection-based formatting does not carry this
+// guarantee.
+func (r GenerateRequest) GoString() string {
+	return r.redactedString()
+}
+
+func (r GenerateRequest) redactedString() string {
+	return fmt.Sprintf(
+		"promptkit.GenerateRequest{Messages:%d Model:%q APIKeySet:%t StructuredOutputSet:%t ExtraParams:%d}",
+		len(r.Prompt.Messages),
+		r.Target.Model,
+		r.APIKey != "",
+		r.StructuredOutput != nil,
+		len(r.Target.ExtraParams),
+	)
+}

go.mod

@@ -0,0 +1,10 @@
+module gitea.maximumdirect.net/eric/promptkit
+
+go 1.25.5
+
+require (
+	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
+	gopkg.in/yaml.v3 v3.0.1
+)
+
+require golang.org/x/text v0.14.0 // indirect

go.sum

@@ -0,0 +1,10 @@
+github.com/dlclark/regexp2 v1.11.0 h1:G/nrcoOa7ZXlpoa/91N3X7mM3r8eIlMBBJZvsz/mxKI=
+github.com/dlclark/regexp2 v1.11.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=
+github.com/santhosh-tekuri/jsonschema/v6 v6.0.2/go.mod h1:JXeL+ps8p7/KNMjDQk3TCwPpBy0wYklyWTfbkIzdIFU=
+golang.org/x/text v0.14.0 h1:ScX5w1eTa3QqT8oi6+ziP7dTV1S2+ALU0bI+0zXKWiQ=
+golang.org/x/text v0.14.0/go.mod h1:18ZOQIKpY8NJVqYksKHtTdi31H5itFRjB5/qKTNYzSU=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
+gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
+gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=
+gopkg.in/yaml.v3 v3.0.1/go.mod h1:K4uyk7z7BCEPqu6E+C64Yfv1cQ7kz7rIZviUmN+EgEM=

internal/artifact/reader.go

@@ -0,0 +1,122 @@
+package artifact
+
+import (
+	"context"
+	"crypto/sha256"
+	"errors"
+	"fmt"
+	"io"
+	"mime"
+	"os"
+	"path/filepath"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+var (
+	ErrUnsupportedRefType = errors.New("unsupported artifact reference type")
+	ErrMissingInlineBody  = errors.New("missing body for inline artifact")
+	ErrMissingFilePath    = errors.New("missing file path for file artifact")
+)
+
+// Reader resolves artifact references into actual artifacts.
+type Reader interface {
+	Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error)
+}
+
+// CompositeReader routes artifact resolution based on the reference type.
+type CompositeReader struct {
+	inlineReader *inlineReader
+	fileReader   Reader
+}
+
+func NewCompositeReader() Reader {
+	return &CompositeReader{
+		inlineReader: &inlineReader{},
+		fileReader:   &fileReader{},
+	}
+}
+
+func (c *CompositeReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
+	}
+
+	switch ref.Type {
+	case domain.ArtifactRefInline:
+		return c.inlineReader.Read(ctx, ref)
+	case domain.ArtifactRefFile:
+		return c.fileReader.Read(ctx, ref)
+	default:
+		return nil, fmt.Errorf("%w: %s", ErrUnsupportedRefType, ref.Type)
+	}
+}
+
+type inlineReader struct{}
+
+func (r *inlineReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
+	}
+
+	if ref.Body == "" {
+		return nil, ErrMissingInlineBody
+	}
+
+	body := []byte(ref.Body)
+	return &domain.Artifact{
+		ContentType: defaults.ContentTypeTextPlain,
+		Body:        body,
+		Size:        int64(len(body)),
+		Hash:        fmt.Sprintf("%x", sha256.Sum256(body)),
+		URI:         ref.URI,
+	}, nil
+}
+
+type fileReader struct{}
+
+func (r *fileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
+	}
+
+	if ref.URI == "" {
+		return nil, ErrMissingFilePath
+	}
+
+	return readFileArtifact(ref.URI)
+}
+
+func readFileArtifact(path string) (*domain.Artifact, error) {
+	file, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+	}
+	defer file.Close()
+
+	data, err := io.ReadAll(file)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+	}
+
+	contentType := mime.TypeByExtension(filepath.Ext(path))
+	if contentType == "" {
+		contentType = defaults.ContentTypeTextPlain
+	}
+
+	return &domain.Artifact{
+		Name:        filepath.Base(path),
+		ContentType: contentType,
+		Body:        data,
+		URI:         path,
+		Size:        int64(len(data)),
+		Hash:        fmt.Sprintf("%x", sha256.Sum256(data)),
+	}, nil
+}

internal/artifact/reader_test.go

@@ -0,0 +1,180 @@
+package artifact
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestCompositeReader_Read(t *testing.T) {
+	reader := NewCompositeReader()
+	ctx := context.Background()
+
+	t.Run("inline artifact", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefInline,
+			Body: "hello world",
+		}
+		art, err := reader.Read(ctx, ref)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if string(art.Body) != "hello world" {
+			t.Errorf("expected 'hello world', got %s", string(art.Body))
+		}
+		if art.ContentType != "text/plain" {
+			t.Errorf("expected text/plain content type, got %q", art.ContentType)
+		}
+		if art.Hash != "b94d27b9934d3e08a52e52d7da7dabfac484efe37a5380ee9088f7ace2efcde9" {
+			t.Errorf("unexpected hash: %s", art.Hash)
+		}
+		if art.Size != int64(len(ref.Body)) {
+			t.Errorf("expected size %d, got %d", len(ref.Body), art.Size)
+		}
+	})
+
+	t.Run("inline artifact missing body", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefInline,
+			Body: "",
+		}
+		_, err := reader.Read(ctx, ref)
+		if !errors.Is(err, ErrMissingInlineBody) {
+			t.Errorf("expected ErrMissingInlineBody, got %v", err)
+		}
+	})
+
+	t.Run("unsupported ref type", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefType("unsupported"),
+			URI:  "unsupported://bucket/key",
+		}
+		_, err := reader.Read(ctx, ref)
+		if !errors.Is(err, ErrUnsupportedRefType) {
+			t.Error("expected error for unsupported type")
+		}
+	})
+}
+
+func TestCompositeReaderCopiesInlineData(t *testing.T) {
+	reader := NewCompositeReader()
+	ref := domain.ArtifactRef{
+		Type: domain.ArtifactRefInline,
+		Body: "hello",
+		URI:  "inline:greeting",
+	}
+
+	first, err := reader.Read(context.Background(), ref)
+	if err != nil {
+		t.Fatalf("read first artifact: %v", err)
+	}
+	first.Body[0] = 'j'
+
+	second, err := reader.Read(context.Background(), ref)
+	if err != nil {
+		t.Fatalf("read second artifact: %v", err)
+	}
+	if got := string(second.Body); got != ref.Body {
+		t.Fatalf("expected an independent body %q, got %q", ref.Body, got)
+	}
+	if second.URI != ref.URI {
+		t.Fatalf("expected URI %q, got %q", ref.URI, second.URI)
+	}
+}
+
+func TestCompositeReaderHonorsCancellation(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	_, err := NewCompositeReader().Read(ctx, domain.ArtifactRef{
+		Type: domain.ArtifactRefInline,
+		Body: "ignored",
+	})
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected context cancellation, got %v", err)
+	}
+}
+
+func TestFileReader_Read(t *testing.T) {
+	content := []byte("test file content")
+	filePath := filepath.Join(t.TempDir(), "artifact.txt")
+	if err := os.WriteFile(filePath, content, 0o600); err != nil {
+		t.Fatal(err)
+	}
+
+	reader := NewCompositeReader()
+	ctx := context.Background()
+
+	t.Run("file artifact loading", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  filePath,
+		}
+		art, err := reader.Read(ctx, ref)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if string(art.Body) != string(content) {
+			t.Errorf("expected %s, got %s", string(content), string(art.Body))
+		}
+		if art.Name != filepath.Base(filePath) {
+			t.Errorf("expected name %q, got %q", filepath.Base(filePath), art.Name)
+		}
+		if !strings.HasPrefix(art.ContentType, "text/plain") {
+			t.Errorf("expected text content type, got %q", art.ContentType)
+		}
+		if art.URI != filePath {
+			t.Errorf("expected URI %q, got %q", filePath, art.URI)
+		}
+		if art.Size != int64(len(content)) {
+			t.Errorf("expected size %d, got %d", len(content), art.Size)
+		}
+		if art.Hash != "60f5237ed4049f0382661ef009d2bc42e48c3ceb3edb6600f7024e7ab3b838f3" {
+			t.Errorf("unexpected hash: %s", art.Hash)
+		}
+	})
+
+	t.Run("missing file path", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  "",
+		}
+		_, err := reader.Read(ctx, ref)
+		if !errors.Is(err, ErrMissingFilePath) {
+			t.Errorf("expected ErrMissingFilePath, got %v", err)
+		}
+	})
+
+	t.Run("missing file", func(t *testing.T) {
+		ref := domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  filepath.Join(t.TempDir(), "missing.txt"),
+		}
+		if _, err := reader.Read(ctx, ref); err == nil {
+			t.Fatal("expected missing file error")
+		}
+	})
+
+	t.Run("unknown extension uses text fallback", func(t *testing.T) {
+		path := filepath.Join(t.TempDir(), "artifact.unknownextension")
+		if err := os.WriteFile(path, content, 0o600); err != nil {
+			t.Fatal(err)
+		}
+
+		art, err := reader.Read(ctx, domain.ArtifactRef{
+			Type: domain.ArtifactRefFile,
+			URI:  path,
+		})
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if art.ContentType != "text/plain" {
+			t.Errorf("expected text/plain fallback, got %q", art.ContentType)
+		}
+	})
+}

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/defaults/defaults.go

@@ -0,0 +1,34 @@
+package defaults
+
+import (
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+const (
+	SchemaDirDefault           = "."
+	OutputArtifactName         = "output"
+	ContentTypeTextPlain       = "text/plain"
+	ContentTypeTextMarkdown    = "text/markdown"
+	ContentTypeApplicationJSON = "application/json"
+	OpenAIChatCompletionsPath  = "/chat/completions"
+
+	ExecutionDefaultTemperature    = 0.0
+	ExecutionDefaultMaxTokens      = 0
+	ExecutionDefaultTopP           = 1.0
+	ExecutionDefaultTimeoutSeconds = 600
+)
+
+var (
+	LLMRequestTimeoutDefault = 10 * time.Minute
+)
+
+func ExecutionTargetDefault() domain.ExecutionTarget {
+	return domain.ExecutionTarget{
+		Temperature:    ExecutionDefaultTemperature,
+		MaxTokens:      ExecutionDefaultMaxTokens,
+		TopP:           ExecutionDefaultTopP,
+		TimeoutSeconds: ExecutionDefaultTimeoutSeconds,
+	}
+}

internal/domain/domain.go

@@ -0,0 +1,311 @@
+package domain
+
+import (
+	"time"
+)
+
+// ArtifactRefType defines how an artifact is referenced.
+type ArtifactRefType string
+
+const (
+	ArtifactRefInline ArtifactRefType = "inline"
+	ArtifactRefFile   ArtifactRefType = "file"
+)
+
+// OutputFormat defines the desired format of the generated artifact.
+type OutputFormat string
+
+const (
+	FormatText     OutputFormat = "text"
+	FormatMarkdown OutputFormat = "markdown"
+	FormatJSON     OutputFormat = "json"
+)
+
+// ValidationMode defines how the output should be validated.
+type ValidationMode string
+
+const (
+	ValidationNone       ValidationMode = "none"
+	ValidationBasic      ValidationMode = "basic"
+	ValidationJSON       ValidationMode = "json"
+	ValidationJSONSchema ValidationMode = "json_schema"
+)
+
+// ValidationStatus defines the result of a validation check.
+type ValidationStatus string
+
+const (
+	ValidationPassed  ValidationStatus = "passed"
+	ValidationFailed  ValidationStatus = "failed"
+	ValidationSkipped ValidationStatus = "skipped"
+)
+
+// CacheControlType defines provider cache behavior for prompt content.
+type CacheControlType string
+
+const (
+	CacheControlEphemeral CacheControlType = "ephemeral"
+)
+
+const (
+	// SessionIDMaxLength is OpenRouter's documented maximum session_id length.
+	SessionIDMaxLength = 256
+)
+
+// CacheControl describes provider cache metadata attached to prompt content.
+type CacheControl struct {
+	Type CacheControlType `yaml:"type" json:"type"`
+	TTL  string           `yaml:"ttl,omitempty" json:"ttl,omitempty"`
+}
+
+// RunRequest represents a request to generate a single artifact.
+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
+}
+
+// RunResult represents the complete result of a prompt execution run.
+type RunResult struct {
+	RunID                string
+	Artifact             Artifact
+	RawOutput            string
+	Validation           ValidationResult
+	PromptID             string
+	PromptVersion        string
+	PromptHash           string
+	SessionID            string
+	RenderedPromptHash   string
+	SelectedProfileID    string
+	SelectedBackendID    string
+	ModelName            string
+	Endpoint             string
+	EffectiveModelParams ExecutionTarget
+	InputHashes          map[string]string
+	Usage                TokenUsage
+	StartTime            time.Time
+	EndTime              time.Time
+	Duration             time.Duration
+}
+
+// PreparedRun contains pre-LLM execution state from the prepare/render phase.
+// It must never include resolved API key values, model output, or validation data.
+type PreparedRun struct {
+	PromptID             string                  `json:"prompt_id"`
+	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"`
+	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"`
+}
+
+// ArtifactRef represents a reference to an input artifact.
+type ArtifactRef struct {
+	Type ArtifactRefType
+	URI  string
+	Body string // Used for inline
+}
+
+// Artifact represents the actual loaded content of a reference.
+type Artifact struct {
+	Name        string
+	ContentType string
+	Body        []byte
+	URI         string
+	Size        int64
+	Hash        string
+}
+
+// PromptDefinition represents a configured prompt execution definition.
+type PromptDefinition struct {
+	ID             string                  `yaml:"id"`
+	Version        string                  `yaml:"version"`
+	DefaultProfile string                  `yaml:"default_profile"`
+	Description    string                  `yaml:"description"`
+	SessionID      string                  `yaml:"session_id" json:"session_id,omitempty"`
+	Inputs         []PromptInput           `yaml:"inputs"`
+	Templates      []PromptMessageTemplate `yaml:"templates"`
+	OutputFormat   OutputFormat            `yaml:"output_format"`
+	Validation     OutputContract          `yaml:"validation"`
+}
+
+// PromptInput describes one named input expected by a prompt definition.
+type PromptInput struct {
+	Name        string `yaml:"name"`
+	Required    bool   `yaml:"required"`
+	ContentType string `yaml:"content_type"`
+	Description string `yaml:"description"`
+}
+
+// PromptMessageTemplate defines a template for a chat message.
+type PromptMessageTemplate struct {
+	Role         string        `yaml:"role"`
+	Content      string        `yaml:"content"`
+	ContentFile  string        `yaml:"content_file"`
+	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"`
+	MaxTokens       int            `yaml:"max_tokens"`
+	TopP            float64        `yaml:"top_p"`
+	TimeoutSeconds  int            `yaml:"timeout_seconds"`
+	ServiceTier     string         `yaml:"service_tier"`
+	ReasoningEffort string         `yaml:"reasoning_effort"`
+	APIKeyEnv       string         `yaml:"api_key_env"`
+	APIKeyRequired  bool           `yaml:"-" json:"-"`
+	ExtraParams     map[string]any `yaml:"extra_params"`
+}
+
+// ExecutionTargetOverride represents per-request runtime setting overrides.
+type ExecutionTargetOverride struct {
+	Endpoint        string         `json:"endpoint,omitempty"`
+	Model           string         `json:"model,omitempty"`
+	Temperature     *float64       `json:"temperature,omitempty"`
+	MaxTokens       *int           `json:"max_tokens,omitempty"`
+	TopP            *float64       `json:"top_p,omitempty"`
+	TimeoutSeconds  *int           `json:"timeout_seconds,omitempty"`
+	ServiceTier     string         `json:"service_tier,omitempty"`
+	ReasoningEffort *string        `json:"reasoning_effort,omitempty"`
+	APIKeyEnv       string         `json:"api_key_env,omitempty"`
+	ExtraParams     map[string]any `json:"extra_params,omitempty"`
+}
+
+// ExecutionTargetPresence tracks which effective runtime fields came from an
+// explicit request override even when the resolved value is a zero value.
+type ExecutionTargetPresence struct {
+	Temperature    bool
+	MaxTokens      bool
+	TopP           bool
+	TimeoutSeconds bool
+}
+
+// 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"`
+	MaxTokens       int            `yaml:"max_tokens" json:"max_tokens"`
+	TopP            float64        `yaml:"top_p" json:"top_p"`
+	TimeoutSeconds  int            `yaml:"timeout_seconds" json:"timeout_seconds"`
+	ServiceTier     string         `yaml:"service_tier" json:"service_tier"`
+	ReasoningEffort string         `yaml:"reasoning_effort" json:"reasoning_effort"`
+	APIKeyEnv       string         `yaml:"api_key_env" json:"api_key_env"`
+	APIKey          string         `yaml:"-" json:"-"`
+	APIKeyRequired  bool           `yaml:"-" json:"-"`
+	ExtraParams     map[string]any `yaml:"extra_params" json:"extra_params"`
+}
+
+// OutputContract defines the requirements for the output artifact.
+type OutputContract struct {
+	Format         OutputFormat   `yaml:"format"`
+	ValidationMode ValidationMode `yaml:"validation_mode"`
+	SchemaPath     string         `yaml:"schema_path"`
+	RepairAttempts int            `yaml:"repair_attempts"`
+}
+
+// RenderedPrompt represents the prompt after template application.
+type RenderedPrompt struct {
+	SessionID string            `json:"session_id,omitempty"`
+	Messages  []RenderedMessage `json:"messages"`
+}
+
+// RenderedMessage is a single message in a rendered prompt.
+type RenderedMessage struct {
+	Role         string        `json:"role"`
+	Content      string        `json:"content"`
+	CacheControl *CacheControl `json:"cache_control,omitempty"`
+}
+
+// GenerateRequest is the internal request passed to the LLM client.
+type GenerateRequest struct {
+	Prompt           RenderedPrompt
+	Target           ExecutionTarget
+	TargetPresence   ExecutionTargetPresence
+	StructuredOutput *StructuredOutputSpec
+}
+
+// StructuredOutputType indicates which provider-level output mode is requested.
+type StructuredOutputType string
+
+const (
+	StructuredOutputJSONSchema StructuredOutputType = "json_schema"
+)
+
+// StructuredOutputSpec describes provider-level structured output requirements.
+type StructuredOutputSpec struct {
+	Type       StructuredOutputType      `json:"type"`
+	JSONSchema *StructuredOutputJSONSpec `json:"json_schema,omitempty"`
+}
+
+// StructuredOutputJSONSpec contains json_schema output constraints.
+type StructuredOutputJSONSpec struct {
+	Name   string `json:"name"`
+	Strict bool   `json:"strict"`
+	Schema any    `json:"schema"`
+}
+
+// GenerateResponse is the response received from the LLM client.
+type GenerateResponse struct {
+	Content string
+	Usage   TokenUsage
+}
+
+// TokenUsage tracks token consumption.
+type TokenUsage struct {
+	PromptTokens     int
+	CompletionTokens int
+	TotalTokens      int
+	CachedTokens     int
+	CacheWriteTokens int
+}
+
+// ValidationResult represents the outcome of an output validation.
+type ValidationResult struct {
+	Status         ValidationStatus
+	Mode           ValidationMode
+	Errors         []string
+	SchemaPath     string
+	RepairAttempts int
+	IsValid        bool
+}

internal/domain/prepared_run_test.go

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

internal/domain/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/filecatalog/catalog.go

@@ -0,0 +1,142 @@
+package filecatalog
+
+import (
+	"context"
+	"fmt"
+	"io/fs"
+	"os"
+	"path"
+	"path/filepath"
+	"sort"
+	"strings"
+)
+
+// FindYAMLFiles returns sorted full paths for .yaml and .yml files under root.
+func FindYAMLFiles(ctx context.Context, root string) ([]string, error) {
+	var files []string
+	err := filepath.WalkDir(root, func(path string, d os.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		default:
+		}
+		if d.IsDir() {
+			return nil
+		}
+		if !IsYAMLFile(d.Name()) {
+			return nil
+		}
+		files = append(files, path)
+		return nil
+	})
+	sort.Strings(files)
+	return files, err
+}
+
+// FindFSYAMLFiles returns sorted paths for .yaml and .yml files under root in fsys.
+func FindFSYAMLFiles(ctx context.Context, fsys fs.FS, root string) ([]string, error) {
+	cleanRoot := CleanFSRoot(root)
+	var files []string
+	err := fs.WalkDir(fsys, cleanRoot, func(name string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		default:
+		}
+		if d.IsDir() {
+			return nil
+		}
+		if !IsYAMLFile(d.Name()) {
+			return nil
+		}
+		files = append(files, name)
+		return nil
+	})
+	sort.Strings(files)
+	return files, err
+}
+
+// RelativePath computes a clean relative path from root to path.
+func RelativePath(root string, filePath string) string {
+	rel, err := filepath.Rel(root, filePath)
+	if err != nil {
+		return filepath.Clean(filePath)
+	}
+	return filepath.Clean(rel)
+}
+
+// CleanFSRoot normalizes a root path for use with fs.FS.
+func CleanFSRoot(root string) string {
+	root = strings.TrimSpace(root)
+	if root == "" || root == "." {
+		return "."
+	}
+	return path.Clean(root)
+}
+
+// DisplayPath returns name relative to root for messages about fs.FS paths.
+func DisplayPath(root string, name string) string {
+	cleanRoot := CleanFSRoot(root)
+	cleanName := path.Clean(name)
+	if cleanRoot == "." {
+		return cleanName
+	}
+	prefix := strings.TrimSuffix(cleanRoot, "/") + "/"
+	if strings.HasPrefix(cleanName, prefix) {
+		return strings.TrimPrefix(cleanName, prefix)
+	}
+	return cleanName
+}
+
+// ResolveFSPath resolves userPath from baseDir and keeps it inside root.
+func ResolveFSPath(root string, baseDir string, userPath string) (string, string, error) {
+	cleanRoot := CleanFSRoot(root)
+	cleanBase := path.Clean(strings.TrimSpace(baseDir))
+	if cleanBase == "" {
+		cleanBase = cleanRoot
+	}
+	if !containsFSPath(cleanRoot, cleanBase) {
+		return "", "", fmt.Errorf("base path %q is outside source root %q", cleanBase, cleanRoot)
+	}
+
+	cleanUserPath := strings.TrimSpace(userPath)
+	if cleanUserPath == "" {
+		return "", "", fmt.Errorf("path is required")
+	}
+	cleanUserPath = path.Clean(cleanUserPath)
+	if path.IsAbs(cleanUserPath) {
+		return "", "", fmt.Errorf("path %q must be relative", userPath)
+	}
+
+	resolved := path.Clean(path.Join(cleanBase, cleanUserPath))
+	if !containsFSPath(cleanRoot, resolved) {
+		return "", "", fmt.Errorf("path %q escapes source root %q", userPath, cleanRoot)
+	}
+	return resolved, DisplayPath(cleanRoot, resolved), nil
+}
+
+func containsFSPath(root string, name string) bool {
+	root = CleanFSRoot(root)
+	name = path.Clean(name)
+	if root == "." {
+		return name == "." || (name != ".." && !strings.HasPrefix(name, "../"))
+	}
+	return name == root || strings.HasPrefix(name, strings.TrimSuffix(root, "/")+"/")
+}
+
+// Stem strips .yaml or .yml from a file name.
+func Stem(name string) string {
+	name = strings.TrimSuffix(name, ".yaml")
+	name = strings.TrimSuffix(name, ".yml")
+	return name
+}
+
+func IsYAMLFile(name string) bool {
+	return strings.HasSuffix(name, ".yaml") || strings.HasSuffix(name, ".yml")
+}

internal/filecatalog/catalog_test.go

@@ -0,0 +1,270 @@
+package filecatalog
+
+import (
+	"context"
+	"errors"
+	"os"
+	"path/filepath"
+	"reflect"
+	"strings"
+	"testing"
+	"testing/fstest"
+)
+
+func TestFindYAMLFilesNestedSortedAndFiltered(t *testing.T) {
+	root := t.TempDir()
+	mustWriteFile(t, filepath.Join(root, "z", "prompt.yml"), "id: z")
+	mustWriteFile(t, filepath.Join(root, "a", "profile.yaml"), "id: a")
+	mustWriteFile(t, filepath.Join(root, "a", "ignore.txt"), "not yaml")
+	mustWriteFile(t, filepath.Join(root, "b", "ignore.yaml.bak"), "not yaml")
+
+	got, err := FindYAMLFiles(context.Background(), root)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+
+	want := []string{
+		filepath.Join(root, "a", "profile.yaml"),
+		filepath.Join(root, "z", "prompt.yml"),
+	}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("expected sorted YAML files %v, got %v", want, got)
+	}
+}
+
+func TestFindYAMLFilesHonorsContextCancellation(t *testing.T) {
+	root := t.TempDir()
+	mustWriteFile(t, filepath.Join(root, "one.yaml"), "id: one")
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	_, err := FindYAMLFiles(ctx, root)
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected context.Canceled, got %v", err)
+	}
+}
+
+func TestFindFSYAMLFilesNestedSortedAndFiltered(t *testing.T) {
+	fsys := fstest.MapFS{
+		"prompts/z/prompt.yml":      &fstest.MapFile{Data: []byte("id: z")},
+		"prompts/a/profile.yaml":    &fstest.MapFile{Data: []byte("id: a")},
+		"prompts/a/ignore.txt":      &fstest.MapFile{Data: []byte("not yaml")},
+		"prompts/b/ignore.yaml.bak": &fstest.MapFile{Data: []byte("not yaml")},
+		"other/ignored.yaml":        &fstest.MapFile{Data: []byte("id: ignored")},
+	}
+
+	got, err := FindFSYAMLFiles(context.Background(), fsys, " prompts ")
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+
+	want := []string{
+		"prompts/a/profile.yaml",
+		"prompts/z/prompt.yml",
+	}
+	if !reflect.DeepEqual(got, want) {
+		t.Fatalf("expected sorted YAML files %v, got %v", want, got)
+	}
+}
+
+func TestFindFSYAMLFilesHonorsContextCancellation(t *testing.T) {
+	fsys := fstest.MapFS{
+		"one.yaml": &fstest.MapFile{Data: []byte("id: one")},
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	_, err := FindFSYAMLFiles(ctx, fsys, ".")
+	if !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected context.Canceled, got %v", err)
+	}
+}
+
+func TestRelativePathNested(t *testing.T) {
+	root := t.TempDir()
+	path := filepath.Join(root, "nested", "profiles", "local.yaml")
+	got := RelativePath(root, path)
+	want := filepath.Join("nested", "profiles", "local.yaml")
+	if got != want {
+		t.Fatalf("expected relative path %q, got %q", want, got)
+	}
+}
+
+func TestCleanFSRoot(t *testing.T) {
+	tests := []struct {
+		name string
+		root string
+		want string
+	}{
+		{name: "empty", root: "", want: "."},
+		{name: "dot", root: ".", want: "."},
+		{name: "trimmed", root: " prompts/../profiles ", want: "profiles"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := CleanFSRoot(tc.root); got != tc.want {
+				t.Fatalf("expected %q, got %q", tc.want, got)
+			}
+		})
+	}
+}
+
+func TestDisplayPath(t *testing.T) {
+	tests := []struct {
+		name string
+		root string
+		path string
+		want string
+	}{
+		{name: "root dot", root: ".", path: "profiles/local.yaml", want: "profiles/local.yaml"},
+		{name: "nested root", root: "profiles", path: "profiles/local.yaml", want: "local.yaml"},
+		{name: "outside root", root: "profiles", path: "other/local.yaml", want: "other/local.yaml"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := DisplayPath(tc.root, tc.path); got != tc.want {
+				t.Fatalf("expected %q, got %q", tc.want, got)
+			}
+		})
+	}
+}
+
+func TestResolveFSPath(t *testing.T) {
+	tests := []struct {
+		name        string
+		root        string
+		baseDir     string
+		userPath    string
+		wantPath    string
+		wantDisplay string
+		wantErr     string
+	}{
+		{
+			name:        "sibling inside root",
+			root:        "prompts",
+			baseDir:     "prompts/nested",
+			userPath:    "./messages/user.tmpl",
+			wantPath:    "prompts/nested/messages/user.tmpl",
+			wantDisplay: "nested/messages/user.tmpl",
+		},
+		{
+			name:        "parent inside root",
+			root:        "prompts",
+			baseDir:     "prompts/nested",
+			userPath:    "../shared/user.tmpl",
+			wantPath:    "prompts/shared/user.tmpl",
+			wantDisplay: "shared/user.tmpl",
+		},
+		{
+			name:     "escape rejected",
+			root:     "prompts",
+			baseDir:  "prompts/nested",
+			userPath: "../../outside.tmpl",
+			wantErr:  "escapes source root",
+		},
+		{
+			name:     "absolute path rejected",
+			root:     "prompts",
+			baseDir:  "prompts/nested",
+			userPath: "/outside.tmpl",
+			wantErr:  "must be relative",
+		},
+		{
+			name:     "empty path rejected",
+			root:     "prompts",
+			baseDir:  "prompts/nested",
+			userPath: " ",
+			wantErr:  "path is required",
+		},
+		{
+			name:        "dot root allows normal relative path",
+			root:        ".",
+			baseDir:     ".",
+			userPath:    "schemas/events.schema.json",
+			wantPath:    "schemas/events.schema.json",
+			wantDisplay: "schemas/events.schema.json",
+		},
+		{
+			name:     "dot root rejects parent escape",
+			root:     ".",
+			baseDir:  ".",
+			userPath: "../outside.tmpl",
+			wantErr:  "escapes source root",
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			gotPath, gotDisplay, err := ResolveFSPath(tc.root, tc.baseDir, tc.userPath)
+			if tc.wantErr != "" {
+				if err == nil {
+					t.Fatalf("expected error containing %q", tc.wantErr)
+				}
+				if !strings.Contains(err.Error(), tc.wantErr) {
+					t.Fatalf("expected error to contain %q, got %v", tc.wantErr, err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("expected no error, got %v", err)
+			}
+			if gotPath != tc.wantPath || gotDisplay != tc.wantDisplay {
+				t.Fatalf("expected path/display %q/%q, got %q/%q", tc.wantPath, tc.wantDisplay, gotPath, gotDisplay)
+			}
+		})
+	}
+}
+
+func TestStemStripsYAMLExtensions(t *testing.T) {
+	tests := []struct {
+		name string
+		in   string
+		want string
+	}{
+		{name: "yaml", in: "prompt.yaml", want: "prompt"},
+		{name: "yml", in: "profile.yml", want: "profile"},
+		{name: "other", in: "file.txt", want: "file.txt"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := Stem(tc.in); got != tc.want {
+				t.Fatalf("expected %q, got %q", tc.want, got)
+			}
+		})
+	}
+}
+
+func TestIsYAMLFile(t *testing.T) {
+	tests := []struct {
+		name string
+		in   string
+		want bool
+	}{
+		{name: "yaml", in: "prompt.yaml", want: true},
+		{name: "yml", in: "profile.yml", want: true},
+		{name: "backup", in: "profile.yaml.bak", want: false},
+		{name: "uppercase", in: "profile.YAML", want: false},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			if got := IsYAMLFile(tc.in); got != tc.want {
+				t.Fatalf("expected %v, got %v", tc.want, got)
+			}
+		})
+	}
+}
+
+func mustWriteFile(t *testing.T, path string, content string) {
+	t.Helper()
+	if err := os.MkdirAll(filepath.Dir(path), 0o755); err != nil {
+		t.Fatalf("failed to create directory: %v", err)
+	}
+	if err := os.WriteFile(path, []byte(content), 0o644); err != nil {
+		t.Fatalf("failed to write file %q: %v", path, err)
+	}
+}

internal/jsonvalue/jsonvalue.go

@@ -0,0 +1,232 @@
+// 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 visit struct {
+	typ reflect.Type
+	ptr uintptr
+}
+
+// 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 := copyValue(reflect.ValueOf(src), "extra_params", make(map[visit]struct{}))
+	if err != nil {
+		return nil, err
+	}
+	out, ok := copied.(map[string]any)
+	if !ok {
+		return nil, fmt.Errorf("extra_params: expected object")
+	}
+	return out, nil
+}
+
+func copyValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+	if !value.IsValid() {
+		return nil, nil
+	}
+	if value.Kind() == reflect.Interface {
+		if value.IsNil() {
+			return nil, nil
+		}
+		return copyValue(value.Elem(), path, seen)
+	}
+	if !value.CanInterface() {
+		return nil, fmt.Errorf("%s: value cannot be copied", path)
+	}
+	if number, ok := value.Interface().(json.Number); ok {
+		if _, err := json.Marshal(number); err != nil {
+			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)
+		}
+		return number, nil
+	}
+
+	switch value.Kind() {
+	case reflect.Bool, reflect.String:
+		return value.Interface(), nil
+	case reflect.Int, reflect.Int8, reflect.Int16, reflect.Int32, reflect.Int64:
+		if value.Int() < -maxSafeJSONInteger || value.Int() > maxSafeJSONInteger {
+			return nil, fmt.Errorf("%s: integer is outside the JSON-safe range", path)
+		}
+		return value.Interface(), nil
+	case reflect.Uint, reflect.Uint8, reflect.Uint16, reflect.Uint32, reflect.Uint64, reflect.Uintptr:
+		if value.Uint() > maxSafeJSONInteger {
+			return nil, fmt.Errorf("%s: integer is outside the JSON-safe range", path)
+		}
+		return value.Interface(), nil
+	case reflect.Float32, reflect.Float64:
+		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
+	case reflect.Pointer:
+		if value.IsNil() {
+			return nil, nil
+		}
+		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[current] = struct{}{}
+		defer delete(seen, current)
+		return copyValue(value.Elem(), path, seen)
+	case reflect.Map:
+		return copyMapValue(value, path, seen)
+	case reflect.Slice:
+		if value.IsNil() {
+			return nil, nil
+		}
+		return copySequenceValue(value, path, seen)
+	case reflect.Array:
+		return copySequenceValue(value, path, seen)
+	default:
+		return nil, fmt.Errorf("%s: unsupported JSON value type %s", path, value.Type())
+	}
+}
+
+func copyMapValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+	if value.IsNil() {
+		return nil, nil
+	}
+	if value.Type().Key().Kind() != reflect.String {
+		return nil, fmt.Errorf("%s: map key type %s is not supported", path, value.Type().Key())
+	}
+
+	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[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, len(keys))
+	preserveType := true
+	elementType := value.Type().Elem()
+	for _, key := range keys {
+		name := key.String()
+		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(elementType) {
+				preserveType = false
+			}
+			continue
+		}
+		if !reflect.TypeOf(copied).AssignableTo(elementType) {
+			preserveType = false
+		}
+	}
+
+	if preserveType {
+		out := reflect.MakeMapWithSize(value.Type(), len(entries))
+		for _, entry := range entries {
+			if entry.value == nil {
+				out.SetMapIndex(entry.key, reflect.Zero(elementType))
+				continue
+			}
+			out.SetMapIndex(entry.key, reflect.ValueOf(entry.value))
+		}
+		return out.Interface(), nil
+	}
+
+	out := make(map[string]any, len(entries))
+	for _, entry := range entries {
+		out[entry.name] = entry.value
+	}
+	return out, nil
+}
+
+func copySequenceValue(value reflect.Value, path string, seen map[visit]struct{}) (any, error) {
+	var current visit
+	if value.Kind() == reflect.Slice {
+		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[current] = struct{}{}
+		defer delete(seen, current)
+	}
+
+	values := make([]any, value.Len())
+	preserveType := true
+	elementType := value.Type().Elem()
+	for i := 0; i < value.Len(); i++ {
+		copied, err := copyValue(value.Index(i), fmt.Sprintf("%s[%d]", path, i), seen)
+		if err != nil {
+			return nil, err
+		}
+		values[i] = copied
+		if copied == nil {
+			if !canAssignNil(elementType) {
+				preserveType = false
+			}
+			continue
+		}
+		if !reflect.TypeOf(copied).AssignableTo(elementType) {
+			preserveType = false
+		}
+	}
+
+	if preserveType {
+		out := reflect.New(value.Type()).Elem()
+		if value.Kind() == reflect.Slice {
+			out = reflect.MakeSlice(value.Type(), value.Len(), value.Len())
+		}
+		for i, copied := range values {
+			if copied == nil {
+				out.Index(i).Set(reflect.Zero(elementType))
+				continue
+			}
+			out.Index(i).Set(reflect.ValueOf(copied))
+		}
+		return out.Interface(), nil
+	}
+
+	out := make([]any, len(values))
+	copy(out, values)
+	return out, nil
+}
+
+func canAssignNil(typ reflect.Type) bool {
+	switch typ.Kind() {
+	case reflect.Chan, reflect.Func, reflect.Interface, reflect.Map, reflect.Pointer, reflect.Slice:
+		return true
+	default:
+		return false
+	}
+}

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/client.go

@@ -0,0 +1,11 @@
+package llm
+
+import (
+	"context"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// Client executes a rendered prompt against an LLM endpoint.
+type Client interface {
+	Generate(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error)
+}

internal/llm/openai_compatible_client.go

@@ -0,0 +1,390 @@
+package llm
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"net/http"
+	"net/url"
+	"os"
+	"strings"
+	"time"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/defaults"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+var (
+	ErrInvalidConfig     = errors.New("invalid llm client configuration")
+	ErrInvalidRequest    = errors.New("invalid generate request")
+	ErrRequestFailed     = errors.New("llm request failed")
+	ErrUnexpectedStatus  = errors.New("llm returned non-success status")
+	ErrMalformedResponse = errors.New("malformed llm response")
+)
+
+type OpenAICompatibleConfig struct {
+	BaseURL    string
+	Model      string
+	Timeout    time.Duration
+	HTTPClient *http.Client
+}
+
+type OpenAICompatibleClient struct {
+	baseURL      string
+	defaultModel string
+	httpClient   *http.Client
+}
+
+func NewOpenAICompatibleClient(cfg OpenAICompatibleConfig) (*OpenAICompatibleClient, error) {
+	baseURL := strings.TrimSpace(cfg.BaseURL)
+	if baseURL != "" {
+		if _, err := url.ParseRequestURI(baseURL); err != nil {
+			return nil, fmt.Errorf("%w: invalid base URL: %v", ErrInvalidConfig, err)
+		}
+	}
+
+	timeout := cfg.Timeout
+	if timeout <= 0 {
+		timeout = defaults.LLMRequestTimeoutDefault
+	}
+
+	var client *http.Client
+	if cfg.HTTPClient != nil {
+		cloned := *cfg.HTTPClient
+		if cloned.Timeout <= 0 {
+			cloned.Timeout = timeout
+		}
+		client = &cloned
+	} else {
+		client = &http.Client{Timeout: timeout}
+	}
+
+	return &OpenAICompatibleClient{
+		baseURL:      strings.TrimRight(baseURL, "/"),
+		defaultModel: cfg.Model,
+		httpClient:   client,
+	}, nil
+}
+
+func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	if req.Target.TimeoutSeconds < 0 {
+		return nil, fmt.Errorf("%w: timeout_seconds must be greater than or equal to 0", ErrInvalidRequest)
+	}
+
+	endpoint := strings.TrimSpace(req.Target.Endpoint)
+	if endpoint == "" {
+		endpoint = c.baseURL
+	}
+	if endpoint == "" {
+		return nil, fmt.Errorf("%w: endpoint is required", ErrInvalidRequest)
+	}
+	endpoint = strings.TrimRight(endpoint, "/") + defaults.OpenAIChatCompletionsPath
+
+	wireReq, err := openAIChatRequestFromGenerateRequest(req, c.defaultModel)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	wirePayload, err := openAIChatRequestPayload(wireReq)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	payload, err := json.Marshal(wirePayload)
+	if err != nil {
+		return nil, fmt.Errorf("%w: failed to encode request: %v", ErrRequestFailed, err)
+	}
+
+	requestContext := ctx
+	if req.Target.TimeoutSeconds > 0 {
+		var cancel context.CancelFunc
+		requestContext, cancel = context.WithTimeout(
+			ctx,
+			time.Duration(req.Target.TimeoutSeconds)*time.Second,
+		)
+		defer cancel()
+	}
+
+	httpReq, err := http.NewRequestWithContext(requestContext, http.MethodPost, endpoint, bytes.NewReader(payload))
+	if err != nil {
+		return nil, fmt.Errorf("%w: failed to create request: %v", ErrRequestFailed, err)
+	}
+	httpReq.Header.Set("Content-Type", "application/json")
+	if apiKey := strings.TrimSpace(req.Target.APIKey); apiKey != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
+	} else if envName := strings.TrimSpace(req.Target.APIKeyEnv); envName != "" {
+		apiKey := strings.TrimSpace(os.Getenv(envName))
+		if apiKey == "" {
+			return nil, fmt.Errorf("%w: api key environment variable %q is not set", ErrInvalidRequest, envName)
+		}
+		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
+	}
+
+	httpClient := c.httpClient
+	if httpClient == nil {
+		httpClient = &http.Client{Timeout: defaults.LLMRequestTimeoutDefault}
+	}
+
+	httpResp, err := httpClient.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrRequestFailed, err)
+	}
+	defer httpResp.Body.Close()
+
+	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
+		_, _ = io.Copy(io.Discard, io.LimitReader(httpResp.Body, 4096))
+		return nil, fmt.Errorf("%w: status=%d", ErrUnexpectedStatus, httpResp.StatusCode)
+	}
+
+	var wireResp openAIChatResponse
+	if err := json.NewDecoder(httpResp.Body).Decode(&wireResp); err != nil {
+		return nil, fmt.Errorf("%w: failed to decode response: %v", ErrMalformedResponse, err)
+	}
+
+	if len(wireResp.Choices) == 0 {
+		return nil, fmt.Errorf("%w: no choices returned", ErrMalformedResponse)
+	}
+	content := wireResp.Choices[0].Message.Content
+	if content == "" {
+		return nil, fmt.Errorf("%w: first choice has empty message content", ErrMalformedResponse)
+	}
+
+	return &domain.GenerateResponse{
+		Content: content,
+		Usage: domain.TokenUsage{
+			PromptTokens:     wireResp.Usage.PromptTokens,
+			CompletionTokens: wireResp.Usage.CompletionTokens,
+			TotalTokens:      wireResp.Usage.TotalTokens,
+			CachedTokens:     wireResp.Usage.PromptTokensDetails.CachedTokens,
+			CacheWriteTokens: wireResp.Usage.CacheWriteTokens,
+		},
+	}, nil
+}
+
+func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultModel string) (openAIChatRequest, error) {
+	model := strings.TrimSpace(req.Target.Model)
+	if model == "" {
+		model = strings.TrimSpace(defaultModel)
+	}
+	if model == "" {
+		return openAIChatRequest{}, errors.New("model is required")
+	}
+
+	wireReq := openAIChatRequest{
+		Model: model,
+	}
+	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 {
+		wireReq.Messages = append(wireReq.Messages, openAIChatRequestMessageFromRenderedMessage(msg))
+	}
+
+	if req.Target.Temperature != 0 || req.TargetPresence.Temperature {
+		wireReq.Temperature = &req.Target.Temperature
+	}
+	if req.Target.MaxTokens != 0 || req.TargetPresence.MaxTokens {
+		wireReq.MaxTokens = &req.Target.MaxTokens
+	}
+	if req.Target.TopP != 0 || req.TargetPresence.TopP {
+		wireReq.TopP = &req.Target.TopP
+	}
+	if strings.TrimSpace(req.Target.ServiceTier) != "" {
+		wireReq.ServiceTier = req.Target.ServiceTier
+	}
+	if strings.TrimSpace(req.Target.ReasoningEffort) != "" {
+		wireReq.ReasoningEffort = req.Target.ReasoningEffort
+	}
+	if len(req.Target.ExtraParams) > 0 {
+		wireReq.ExtraParams = req.Target.ExtraParams
+	}
+	if req.StructuredOutput != nil {
+		responseFormat, err := toOpenAIResponseFormat(req.StructuredOutput)
+		if err != nil {
+			return openAIChatRequest{}, err
+		}
+		wireReq.ResponseFormat = responseFormat
+	}
+
+	return wireReq, nil
+}
+
+type openAIChatRequest struct {
+	Model           string                     `json:"model"`
+	SessionID       string                     `json:"session_id,omitempty"`
+	Messages        []openAIChatRequestMessage `json:"messages"`
+	Temperature     *float64                   `json:"temperature,omitempty"`
+	MaxTokens       *int                       `json:"max_tokens,omitempty"`
+	TopP            *float64                   `json:"top_p,omitempty"`
+	ServiceTier     string                     `json:"service_tier,omitempty"`
+	ReasoningEffort string                     `json:"reasoning_effort,omitempty"`
+	ResponseFormat  *openAIResponseFormat      `json:"response_format,omitempty"`
+	ExtraParams     map[string]any             `json:"-"`
+}
+
+func openAIChatRequestPayload(req openAIChatRequest) (map[string]any, error) {
+	out := map[string]any{
+		"model":    req.Model,
+		"messages": req.Messages,
+	}
+	if req.SessionID != "" {
+		out["session_id"] = req.SessionID
+	}
+	if req.Temperature != nil {
+		out["temperature"] = *req.Temperature
+	}
+	if req.MaxTokens != nil {
+		out["max_tokens"] = *req.MaxTokens
+	}
+	if req.TopP != nil {
+		out["top_p"] = *req.TopP
+	}
+	if req.ServiceTier != "" {
+		out["service_tier"] = req.ServiceTier
+	}
+	if req.ReasoningEffort != "" {
+		out["reasoning_effort"] = req.ReasoningEffort
+	}
+	if req.ResponseFormat != nil {
+		out["response_format"] = req.ResponseFormat
+	}
+
+	for key, value := range req.ExtraParams {
+		if key == "" {
+			return nil, errors.New("extra_params key must not be empty")
+		}
+		if IsReservedOpenAIChatRequestField(key) {
+			return nil, fmt.Errorf("extra_params key %q collides with reserved request field", key)
+		}
+		if _, err := json.Marshal(value); err != nil {
+			return nil, fmt.Errorf("extra_params.%s must be JSON-serializable: %w", key, err)
+		}
+		out[key] = value
+	}
+
+	return out, nil
+}
+
+// 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 {
+	Role    string `json:"role"`
+	Content any    `json:"content"`
+}
+
+type openAIChatTextContentBlock struct {
+	Type         string              `json:"type"`
+	Text         string              `json:"text"`
+	CacheControl *openAICacheControl `json:"cache_control,omitempty"`
+}
+
+type openAICacheControl struct {
+	Type string `json:"type"`
+	TTL  string `json:"ttl,omitempty"`
+}
+
+type openAIChatResponseMessage struct {
+	Role    string `json:"role"`
+	Content string `json:"content"`
+}
+
+type openAIChatResponse struct {
+	Choices []struct {
+		Message openAIChatResponseMessage `json:"message"`
+	} `json:"choices"`
+	Usage struct {
+		PromptTokens        int `json:"prompt_tokens"`
+		CompletionTokens    int `json:"completion_tokens"`
+		TotalTokens         int `json:"total_tokens"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+		CacheWriteTokens int `json:"cache_write_tokens"`
+	} `json:"usage"`
+}
+
+type openAIResponseFormat struct {
+	Type       string                    `json:"type"`
+	JSONSchema *openAIJSONSchemaEnvelope `json:"json_schema,omitempty"`
+}
+
+type openAIJSONSchemaEnvelope struct {
+	Name   string `json:"name"`
+	Strict bool   `json:"strict"`
+	Schema any    `json:"schema"`
+}
+
+func openAIChatRequestMessageFromRenderedMessage(msg domain.RenderedMessage) openAIChatRequestMessage {
+	wireMsg := openAIChatRequestMessage{
+		Role:    msg.Role,
+		Content: msg.Content,
+	}
+	if msg.CacheControl == nil {
+		return wireMsg
+	}
+
+	wireMsg.Content = []openAIChatTextContentBlock{
+		{
+			Type: "text",
+			Text: msg.Content,
+			CacheControl: &openAICacheControl{
+				Type: string(msg.CacheControl.Type),
+				TTL:  msg.CacheControl.TTL,
+			},
+		},
+	}
+	return wireMsg
+}
+
+func toOpenAIResponseFormat(spec *domain.StructuredOutputSpec) (*openAIResponseFormat, error) {
+	if spec == nil {
+		return nil, nil
+	}
+
+	switch spec.Type {
+	case domain.StructuredOutputJSONSchema:
+		if spec.JSONSchema == nil {
+			return nil, errors.New("json_schema structured output requires schema payload")
+		}
+		if strings.TrimSpace(spec.JSONSchema.Name) == "" {
+			return nil, errors.New("json_schema structured output requires non-empty schema name")
+		}
+		if spec.JSONSchema.Schema == nil {
+			return nil, errors.New("json_schema structured output requires schema document")
+		}
+		return &openAIResponseFormat{
+			Type: "json_schema",
+			JSONSchema: &openAIJSONSchemaEnvelope{
+				Name:   spec.JSONSchema.Name,
+				Strict: spec.JSONSchema.Strict,
+				Schema: spec.JSONSchema.Schema,
+			},
+		}, nil
+	default:
+		return nil, fmt.Errorf("unsupported structured output type %q", spec.Type)
+	}
+}

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

@@ -0,0 +1,8 @@
+id: aion-2
+backend: openrouter
+model: aion-labs/aion-2.0
+temperature: 0.72
+reasoning_effort: high
+top_p: 0.95
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: claude-fable-latest
+backend: openrouter
+model: "~anthropic/claude-fable-latest"
+reasoning_effort: high
+timeout_seconds: 600
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: claude-haiku-latest
+backend: openrouter
+model: "~anthropic/claude-haiku-latest"
+reasoning_effort: medium
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: claude-opus-latest
+backend: openrouter
+model: "~anthropic/claude-opus-latest"
+reasoning_effort: high
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: claude-sonnet-latest
+backend: openrouter
+model: "~anthropic/claude-sonnet-latest"
+reasoning_effort: high
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: deepseek-3-2
+backend: openrouter
+model: deepseek/deepseek-v3.2
+reasoning_effort: high
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: deepseek-4-flash
+backend: openrouter
+model: deepseek/deepseek-v4-flash
+#reasoning_effort: medium
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: deepseek-4-pro
+backend: openrouter
+model: deepseek/deepseek-v4-pro
+reasoning_effort: high
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-2-flash-lite
+backend: openrouter
+model: "google/gemini-2.5-flash-lite"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-2-flash
+backend: openrouter
+model: "google/gemini-2.5-flash"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-2-pro
+backend: openrouter
+model: "google/gemini-2.5-pro"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-3-flash-lite
+backend: openrouter
+model: "google/gemini-3.1-flash-lite"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-flash-latest
+backend: openrouter
+model: "~google/gemini-flash-latest"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemini-pro-latest
+backend: openrouter
+model: "~google/gemini-pro-latest"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: gemma-4-31b
+backend: openrouter
+model: google/gemma-4-31b-it:exacto
+temperature: 0.15
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: minimax-m2
+backend: openrouter
+model: minimax/minimax-m2.5
+temperature: 0.5
+reasoning_effort: high
+top_p: 0.95
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,8 @@
+id: minimax-m3
+backend: openrouter
+model: minimax/minimax-m3
+#temperature: 0.5
+reasoning_effort: high
+#top_p: 0.95
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: mistral-large-2512
+backend: openrouter
+model: mistralai/mistral-large-2512
+temperature: 0.15
+top_p: 0.98
+timeout_seconds: 180

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

@@ -0,0 +1,7 @@
+id: mistral-medium-3-5
+backend: openrouter
+model: mistralai/mistral-medium-3-5
+temperature: 0.15
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 180

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

@@ -0,0 +1,6 @@
+id: mistral-small-3
+backend: openrouter
+model: mistralai/mistral-small-3.2-24b-instruct
+temperature: 0.05
+top_p: 1.0
+timeout_seconds: 180

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

@@ -0,0 +1,7 @@
+id: mistral-small-4
+backend: openrouter
+model: mistralai/mistral-small-2603
+temperature: 0.1
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 180

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

@@ -0,0 +1,6 @@
+id: nemotron-3-ultra
+backend: openrouter
+model: nvidia/nemotron-3-ultra-550b-a55b
+reasoning_effort: high
+timeout_seconds: 180
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: gpt-5-mini
+backend: openrouter
+model: "openai/gpt-5.4-mini"
+reasoning_effort: high
+timeout_seconds: 240
+service_tier: flex

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

@@ -0,0 +1,6 @@
+id: gpt-5-nano
+backend: openrouter
+model: "openai/gpt-5.4-nano"
+reasoning_effort: high
+timeout_seconds: 240
+service_tier: flex

internal/profile/builtin/repository.go

@@ -0,0 +1,31 @@
+package builtin
+
+import (
+	"embed"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+)
+
+const assetRoot = "assets"
+
+//go:embed assets/**/*.yml
+var assets embed.FS
+
+func NewRepository() profile.Repository {
+	return profile.NewFSRepository(assets, assetRoot)
+}
+
+func NewRepositoryWithPrimary(primary profile.Repository) profile.Repository {
+	if primary == nil {
+		return NewRepository()
+	}
+	return profile.NewOverlayRepository(primary, NewRepository())
+}
+
+func NewRepositoryWithDirectory(dir string) profile.Repository {
+	if strings.TrimSpace(dir) == "" {
+		return NewRepository()
+	}
+	return NewRepositoryWithPrimary(profile.NewFilesystemRepository(dir))
+}

internal/profile/builtin/repository_test.go

@@ -0,0 +1,143 @@
+package builtin
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"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"
+)
+
+func TestBuiltInProfilesValidateThroughRepository(t *testing.T) {
+	repo := NewRepository()
+	ids := loadBuiltInProfileIDs(t)
+	if len(ids) == 0 {
+		t.Fatal("expected built-in profiles")
+	}
+
+	for id := range ids {
+		t.Run(id, func(t *testing.T) {
+			p, err := repo.GetProfile(context.Background(), id)
+			if err != nil {
+				t.Fatalf("expected built-in profile %q to load, got %v", id, err)
+			}
+			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)
+			}
+		})
+	}
+}
+
+func TestBuiltInProfilesDoNotContainDuplicateIDsOrRawAPIKeys(t *testing.T) {
+	loadBuiltInProfileIDs(t)
+}
+
+func loadBuiltInProfileIDs(t *testing.T) map[string]string {
+	t.Helper()
+
+	ids := map[string]string{}
+	err := fs.WalkDir(assets, assetRoot, func(name string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.IsDir() || !strings.HasSuffix(name, ".yml") {
+			return nil
+		}
+
+		data, err := assets.ReadFile(name)
+		if err != nil {
+			t.Fatalf("failed to read built-in profile %s: %v", name, err)
+		}
+
+		var raw map[string]any
+		if err := yaml.Unmarshal(data, &raw); err != nil {
+			t.Fatalf("failed to decode built-in profile %s: %v", name, err)
+		}
+		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)
+		}
+		if previous, ok := ids[id]; ok {
+			t.Fatalf("duplicate built-in profile id %q in %s and %s", id, previous, name)
+		}
+		ids[id] = name
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("failed to walk built-in profiles: %v", err)
+	}
+	return ids
+}
+
+func TestRepositoryWithPrimaryUsesPrimaryBeforeBuiltIns(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{
+		profiles: map[string]string{"mistral-small-3": "custom-model"},
+	})
+
+	p, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if err != nil {
+		t.Fatalf("expected profile to load, got %v", err)
+	}
+	if p.Model != "custom-model" {
+		t.Fatalf("expected primary profile to override built-in, got %+v", p)
+	}
+}
+
+func TestRepositoryWithPrimaryFallsBackToBuiltIns(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{})
+
+	p, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if err != nil {
+		t.Fatalf("expected built-in profile to load, got %v", err)
+	}
+	if p.ID != "mistral-small-3" {
+		t.Fatalf("unexpected profile: %+v", p)
+	}
+}
+
+func TestRepositoryWithPrimaryDoesNotFallBackAfterPrimaryError(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{err: profile.ErrInvalidProfile})
+
+	_, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if !errors.Is(err, profile.ErrInvalidProfile) {
+		t.Fatalf("expected primary error, got %v", err)
+	}
+}
+
+type staticProfileRepo struct {
+	profiles map[string]string
+	err      error
+}
+
+func (r staticProfileRepo) GetProfile(_ context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.err != nil {
+		return nil, r.err
+	}
+	if model, ok := r.profiles[id]; ok {
+		return &domain.ExecutionProfile{ID: id, Endpoint: "http://primary/v1", Model: model}, nil
+	}
+	return nil, profile.ErrProfileNotFound
+}

internal/profile/filesystem_repository.go

@@ -0,0 +1,214 @@
+package profile
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"path"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/filecatalog"
+	"gopkg.in/yaml.v3"
+)
+
+var (
+	ErrProfileNotFound     = errors.New("execution profile not found")
+	ErrInvalidYAML         = errors.New("invalid YAML format")
+	ErrInvalidProfile      = errors.New("invalid execution profile configuration")
+	ErrRawAPIKeyNotAllowed = errors.New("raw api_key is not allowed; use api_key_env")
+)
+
+type filesystemRepository struct {
+	dir string
+}
+
+func NewFilesystemRepository(dir string) Repository {
+	return &filesystemRepository{dir: dir}
+}
+
+func (r *filesystemRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	return loadProfile(ctx, os.DirFS(r.dir), ".", id)
+}
+
+type fsRepository struct {
+	fsys fs.FS
+	root string
+}
+
+func NewFSRepository(fsys fs.FS, root string) Repository {
+	return &fsRepository{fsys: fsys, root: root}
+}
+
+func (r *fsRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	return loadProfile(ctx, r.fsys, r.root, id)
+}
+
+type overlayRepository struct {
+	primary  Repository
+	fallback Repository
+}
+
+func NewOverlayRepository(primary, fallback Repository) Repository {
+	return &overlayRepository{primary: primary, fallback: fallback}
+}
+
+func (r *overlayRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.primary != nil {
+		prof, err := r.primary.GetProfile(ctx, id)
+		if err == nil {
+			return prof, nil
+		}
+		if !errors.Is(err, ErrProfileNotFound) {
+			return nil, err
+		}
+	}
+	if r.fallback == nil {
+		return nil, ErrProfileNotFound
+	}
+	return r.fallback.GetProfile(ctx, id)
+}
+
+func loadProfile(ctx context.Context, fsys fs.FS, root string, id string) (*domain.ExecutionProfile, error) {
+	if strings.TrimSpace(id) == "" {
+		return nil, fmt.Errorf("%w: profile id is required", ErrInvalidProfile)
+	}
+	if fsys == nil {
+		return nil, fmt.Errorf("failed to read profile directory: filesystem is nil")
+	}
+
+	files, err := filecatalog.FindFSYAMLFiles(ctx, fsys, root)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read profile directory: %w", err)
+	}
+
+	var matches []profileMatch
+	for _, fullPath := range files {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		default:
+		}
+
+		relPath := filecatalog.DisplayPath(root, fullPath)
+		fileMatch := filecatalog.Stem(path.Base(fullPath)) == id
+		data, err := fs.ReadFile(fsys, fullPath)
+		if err != nil {
+			return nil, fmt.Errorf("failed to read profile file %s: %w", relPath, err)
+		}
+		metadata := readProfileFileMetadata(data)
+		idMatch := fileMatch || metadata.id == id
+		if metadata.hasRawAPIKey {
+			if idMatch {
+				return nil, fmt.Errorf("%w: %s", ErrRawAPIKeyNotAllowed, relPath)
+			}
+			continue
+		}
+
+		var prof domain.ExecutionProfile
+		decoder := yaml.NewDecoder(bytes.NewReader(data))
+		decoder.KnownFields(true)
+		if err := decoder.Decode(&prof); err != nil {
+			if idMatch {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidYAML, relPath, err)
+			}
+			continue
+		}
+
+		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)
+			}
+			return nil, fmt.Errorf("%w: %s: %v", ErrInvalidProfile, relPath, err)
+		}
+		matches = append(matches, profileMatch{
+			profile: &prof,
+			path:    relPath,
+		})
+	}
+
+	if len(matches) > 1 {
+		paths := make([]string, 0, len(matches))
+		for _, match := range matches {
+			paths = append(paths, match.path)
+		}
+		return nil, fmt.Errorf("%w: duplicate execution profile id %q found in: %s", ErrInvalidProfile, id, strings.Join(paths, ", "))
+	}
+
+	if len(matches) == 1 {
+		return matches[0].profile, nil
+	}
+
+	return nil, ErrProfileNotFound
+}
+
+type profileMatch struct {
+	profile *domain.ExecutionProfile
+	path    string
+}
+
+type profileFileMetadata struct {
+	id           string
+	hasRawAPIKey bool
+}
+
+func readProfileFileMetadata(data []byte) profileFileMetadata {
+	var node yaml.Node
+	if err := yaml.NewDecoder(bytes.NewReader(data)).Decode(&node); err != nil {
+		return profileFileMetadata{}
+	}
+	if node.Kind != yaml.DocumentNode || len(node.Content) == 0 {
+		return profileFileMetadata{}
+	}
+	mapping := node.Content[0]
+	if mapping.Kind != yaml.MappingNode {
+		return profileFileMetadata{}
+	}
+
+	var metadata profileFileMetadata
+	for i := 0; i+1 < len(mapping.Content); i += 2 {
+		key := mapping.Content[i]
+		value := mapping.Content[i+1]
+		switch key.Value {
+		case "id":
+			metadata.id = strings.TrimSpace(value.Value)
+		case "api_key":
+			metadata.hasRawAPIKey = true
+		}
+	}
+	return metadata
+}
+
+func validateProfile(p *domain.ExecutionProfile) error {
+	if strings.TrimSpace(p.ID) == "" {
+		return errors.New("id 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")
+	}
+
+	if p.Temperature < 0 || p.Temperature > 2 {
+		return errors.New("temperature must be between 0 and 2")
+	}
+	if p.MaxTokens < 0 {
+		return errors.New("max_tokens must be greater than or equal to 0")
+	}
+	if p.TopP < 0 || p.TopP > 1 {
+		return errors.New("top_p must be between 0 and 1")
+	}
+	if p.TimeoutSeconds < 0 {
+		return errors.New("timeout_seconds must be greater than or equal to 0")
+	}
+
+	return nil
+}

internal/profile/repository.go

@@ -0,0 +1,12 @@
+package profile
+
+import (
+	"context"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// Repository loads execution profiles.
+type Repository interface {
+	GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error)
+}

internal/profile/repository_test.go

@@ -0,0 +1,516 @@
+package profile
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"testing/fstest"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestFilesystemRepository_GetProfile(t *testing.T) {
+	tmpDir, err := os.MkdirTemp("", "execution_profile_test")
+	if err != nil {
+		t.Fatal(err)
+	}
+	defer os.RemoveAll(tmpDir)
+
+	files, err := os.ReadDir("testdata")
+	if err != nil {
+		t.Fatalf("failed to read testdata: %v", err)
+	}
+	for _, f := range files {
+		src := filepath.Join("testdata", f.Name())
+		dst := filepath.Join(tmpDir, f.Name())
+		data, err := os.ReadFile(src)
+		if err != nil {
+			t.Fatal(err)
+		}
+		if err := os.WriteFile(dst, data, 0644); err != nil {
+			t.Fatal(err)
+		}
+	}
+
+	repo := NewFilesystemRepository(tmpDir)
+	ctx := context.Background()
+
+	t.Run("valid local profile", func(t *testing.T) {
+		p, err := repo.GetProfile(ctx, "local-default")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "local-default" {
+			t.Fatalf("unexpected id: %q", p.ID)
+		}
+		if p.Endpoint == "" || p.Model == "" {
+			t.Fatalf("expected endpoint/model to be set: %+v", p)
+		}
+	})
+
+	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 {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.APIKeyEnv != "PROMPTKIT_API_KEY" {
+			t.Fatalf("unexpected api_key_env: %q", p.APIKeyEnv)
+		}
+		if p.ReasoningEffort != "medium" {
+			t.Fatalf("unexpected reasoning_effort: %q", p.ReasoningEffort)
+		}
+		if p.ServiceTier != "priority" {
+			t.Fatalf("unexpected service_tier: %q", p.ServiceTier)
+		}
+	})
+
+	t.Run("valid nested profile", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "local")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "nested-local.yaml"), `
+id: nested-local
+endpoint: http://localhost:8000/v1
+model: nested-model
+temperature: 0.1
+`)
+
+		p, err := repo.GetProfile(ctx, "nested-local")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "nested-model" {
+			t.Fatalf("unexpected model: %q", p.Model)
+		}
+	})
+
+	t.Run("valid profile with JSON-compatible extra params", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "json-extra-params.yaml"), `
+id: json-extra-params
+endpoint: http://localhost:8000/v1
+model: nested-model
+extra_params:
+  string_value: enabled
+  number_value: 42
+  boolean_value: true
+  object_value:
+    nested: value
+    count: 2
+  array_value:
+    - first
+    - 3
+    - false
+`)
+
+		p, err := repo.GetProfile(ctx, "json-extra-params")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+
+		var got map[string]any
+		encoded, err := json.Marshal(p.ExtraParams)
+		if err != nil {
+			t.Fatalf("expected extra_params to marshal as JSON, got %v", err)
+		}
+		if err := json.Unmarshal(encoded, &got); err != nil {
+			t.Fatalf("expected extra_params JSON to decode, got %v", err)
+		}
+
+		if got["string_value"] != "enabled" {
+			t.Fatalf("unexpected string extra param: %#v", got["string_value"])
+		}
+		if got["number_value"] != float64(42) {
+			t.Fatalf("unexpected number extra param: %#v", got["number_value"])
+		}
+		if got["boolean_value"] != true {
+			t.Fatalf("unexpected boolean extra param: %#v", got["boolean_value"])
+		}
+		objectValue, ok := got["object_value"].(map[string]any)
+		if !ok {
+			t.Fatalf("expected object extra param, got %#v", got["object_value"])
+		}
+		if objectValue["nested"] != "value" || objectValue["count"] != float64(2) {
+			t.Fatalf("unexpected object extra param: %#v", objectValue)
+		}
+		arrayValue, ok := got["array_value"].([]any)
+		if !ok {
+			t.Fatalf("expected array extra param, got %#v", got["array_value"])
+		}
+		if len(arrayValue) != 3 || arrayValue[0] != "first" || arrayValue[1] != float64(3) || arrayValue[2] != false {
+			t.Fatalf("unexpected array extra param: %#v", arrayValue)
+		}
+	})
+
+	t.Run("duplicate profile IDs fail as ambiguous", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "duplicate-profile-a.yaml"), `
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: first-model
+`)
+		nestedDir := filepath.Join(tmpDir, "duplicates")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "duplicate-profile-b.yaml"), `
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: second-model
+`)
+
+		_, err := repo.GetProfile(ctx, "duplicate-profile")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected duplicate profile to return ErrInvalidProfile, got %v", err)
+		}
+		for _, want := range []string{"duplicate execution profile id", "duplicate-profile-a.yaml", filepath.Join("duplicates", "duplicate-profile-b.yaml")} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+
+	t.Run("nested raw api_key rejected for likely target file", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "secure")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "not_named_like_id.yaml"), `
+id: nested_raw_api_key
+endpoint: http://localhost:8000/v1
+model: m
+api_key: secret
+`)
+
+		_, err := repo.GetProfile(ctx, "nested_raw_api_key")
+		if !errors.Is(err, ErrRawAPIKeyNotAllowed) {
+			t.Fatalf("expected ErrRawAPIKeyNotAllowed, got %v", err)
+		}
+		if !strings.Contains(err.Error(), filepath.Join("secure", "not_named_like_id.yaml")) {
+			t.Fatalf("expected nested path in error, got %v", err)
+		}
+	})
+
+	t.Run("raw api_key in non-target profile is ignored", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "raw-api-key-non-target.yaml"), `
+id: raw-api-key-non-target
+endpoint: http://localhost:8000/v1
+model: m
+api_key: secret
+`)
+
+		_, err := repo.GetProfile(ctx, "does-not-exist-with-raw-key-nearby")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound for non-target raw api_key file, got %v", err)
+		}
+	})
+
+	t.Run("invalid yaml", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "invalid_yaml")
+		if !errors.Is(err, ErrInvalidYAML) {
+			t.Fatalf("expected ErrInvalidYAML, got %v", err)
+		}
+	})
+
+	t.Run("missing id", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "missing_id")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+
+	t.Run("missing endpoint", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "missing-endpoint")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected ErrInvalidProfile, got %v", err)
+		}
+	})
+
+	t.Run("missing model", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "missing-model")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected ErrInvalidProfile, got %v", err)
+		}
+	})
+
+	t.Run("unknown field", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "unknown_field")
+		if !errors.Is(err, ErrInvalidYAML) {
+			t.Fatalf("expected ErrInvalidYAML for strict decode unknown field, got %v", err)
+		}
+	})
+
+	t.Run("raw api_key rejected", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "raw_api_key")
+		if !errors.Is(err, ErrRawAPIKeyNotAllowed) {
+			t.Fatalf("expected ErrRawAPIKeyNotAllowed, got %v", err)
+		}
+	})
+
+	t.Run("profile not found", func(t *testing.T) {
+		_, err := repo.GetProfile(ctx, "does-not-exist")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+}
+
+func writeProfileTestFile(t *testing.T, path string, content string) {
+	t.Helper()
+	if err := os.WriteFile(path, []byte(strings.TrimLeft(content, "\n")), 0o644); err != nil {
+		t.Fatalf("failed to write profile test file %q: %v", path, err)
+	}
+}
+
+func TestFSRepository(t *testing.T) {
+	ctx := context.Background()
+
+	t.Run("loads valid profiles from nested directories", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/provider/nested.yaml": profileMapFile(`
+id: nested-profile
+endpoint: http://localhost:8000/v1
+model: nested-model
+temperature: 0.1
+`),
+		}, "profiles")
+
+		p, err := repo.GetProfile(ctx, "nested-profile")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "nested-profile" || p.Model != "nested-model" {
+			t.Fatalf("unexpected profile: %+v", p)
+		}
+	})
+
+	t.Run("rejects unknown YAML fields", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/unknown.yaml": profileMapFile(`
+id: unknown-profile
+endpoint: http://localhost:8000/v1
+model: model
+unknown: value
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "unknown-profile")
+		if !errors.Is(err, ErrInvalidYAML) {
+			t.Fatalf("expected ErrInvalidYAML, got %v", err)
+		}
+	})
+
+	t.Run("rejects raw api_key in selected profile", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/raw.yaml": profileMapFile(`
+id: raw-profile
+endpoint: http://localhost:8000/v1
+model: model
+api_key: secret
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "raw-profile")
+		if !errors.Is(err, ErrRawAPIKeyNotAllowed) {
+			t.Fatalf("expected ErrRawAPIKeyNotAllowed, got %v", err)
+		}
+	})
+
+	t.Run("ignores raw api_key in non-selected profiles", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/raw.yaml": profileMapFile(`
+id: raw-profile
+endpoint: http://localhost:8000/v1
+model: model
+api_key: secret
+`),
+			"profiles/valid.yaml": profileMapFile(`
+id: valid-profile
+endpoint: http://localhost:8000/v1
+model: model
+`),
+		}, "profiles")
+
+		p, err := repo.GetProfile(ctx, "valid-profile")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "valid-profile" {
+			t.Fatalf("unexpected profile: %+v", p)
+		}
+	})
+
+	t.Run("rejects duplicate IDs within one source", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/a.yaml": profileMapFile(`
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: first
+`),
+			"profiles/nested/b.yaml": profileMapFile(`
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: second
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "duplicate-profile")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected ErrInvalidProfile, got %v", err)
+		}
+		for _, want := range []string{"duplicate execution profile id", "a.yaml", "nested/b.yaml"} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+}
+
+func TestOverlayRepository(t *testing.T) {
+	ctx := context.Background()
+	primaryProfile := &domain.ExecutionProfile{ID: "shared", Endpoint: "http://primary", Model: "primary"}
+	fallbackProfile := &domain.ExecutionProfile{ID: "shared", Endpoint: "http://fallback", Model: "fallback"}
+
+	t.Run("returns primary matches before fallback matches", func(t *testing.T) {
+		repo := NewOverlayRepository(
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": primaryProfile}},
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+		)
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "primary" {
+			t.Fatalf("expected primary profile, got %+v", p)
+		}
+	})
+
+	t.Run("falls back on primary not found", func(t *testing.T) {
+		repo := NewOverlayRepository(
+			staticProfileRepo{},
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+		)
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "fallback" {
+			t.Fatalf("expected fallback profile, got %+v", p)
+		}
+	})
+
+	t.Run("does not fall back after primary load errors", func(t *testing.T) {
+		for _, tc := range []struct {
+			name string
+			err  error
+		}{
+			{name: "invalid yaml", err: ErrInvalidYAML},
+			{name: "invalid profile", err: ErrInvalidProfile},
+			{name: "raw api key", err: ErrRawAPIKeyNotAllowed},
+		} {
+			t.Run(tc.name, func(t *testing.T) {
+				repo := NewOverlayRepository(
+					staticProfileRepo{err: tc.err},
+					staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+				)
+
+				_, err := repo.GetProfile(ctx, "shared")
+				if !errors.Is(err, tc.err) {
+					t.Fatalf("expected %v, got %v", tc.err, err)
+				}
+			})
+		}
+	})
+
+	t.Run("returns not found when both sources miss", func(t *testing.T) {
+		repo := NewOverlayRepository(staticProfileRepo{}, staticProfileRepo{})
+
+		_, err := repo.GetProfile(ctx, "missing")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+
+	t.Run("nil primary uses fallback", func(t *testing.T) {
+		repo := NewOverlayRepository(nil, staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}})
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "fallback" {
+			t.Fatalf("expected fallback profile, got %+v", p)
+		}
+	})
+
+	t.Run("nil fallback returns not found after primary miss", func(t *testing.T) {
+		repo := NewOverlayRepository(staticProfileRepo{}, nil)
+
+		_, err := repo.GetProfile(ctx, "missing")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+}
+
+func profileMapFile(content string) *fstest.MapFile {
+	return &fstest.MapFile{Data: []byte(strings.TrimLeft(content, "\n"))}
+}
+
+type staticProfileRepo struct {
+	profiles map[string]*domain.ExecutionProfile
+	err      error
+}
+
+func (r staticProfileRepo) GetProfile(_ context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.err != nil {
+		return nil, r.err
+	}
+	if p, ok := r.profiles[id]; ok {
+		cp := *p
+		return &cp, nil
+	}
+	return nil, ErrProfileNotFound
+}

internal/profile/testdata/invalid_yaml.yaml

@@ -0,0 +1,3 @@
+id: invalid_yaml
+endpoint: http://localhost:8000/v1
+model: [broken

internal/profile/testdata/missing_endpoint.yaml

@@ -0,0 +1,2 @@
+id: missing-endpoint
+model: gpt-4o-mini

internal/profile/testdata/missing_id.yaml

@@ -0,0 +1,2 @@
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini

internal/profile/testdata/missing_model.yaml

@@ -0,0 +1,2 @@
+id: missing-model
+endpoint: http://localhost:8000/v1

internal/profile/testdata/raw_api_key.yaml

@@ -0,0 +1,4 @@
+id: raw-api-key
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini
+api_key: super-secret-should-not-be-here

internal/profile/testdata/unknown_field.yaml

@@ -0,0 +1,4 @@
+id: unknown-field
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini
+foo: bar

internal/profile/testdata/valid_local_profile.yaml

@@ -0,0 +1,7 @@
+id: local-default
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini
+temperature: 0.2
+max_tokens: 700
+top_p: 1.0
+timeout_seconds: 120

internal/profile/testdata/valid_with_api_key_env.yaml

@@ -0,0 +1,8 @@
+id: local-secure
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini
+api_key_env: PROMPTKIT_API_KEY
+service_tier: priority
+reasoning_effort: medium
+extra_params:
+  provider: local

internal/prompt/go_renderer.go

@@ -0,0 +1,120 @@
+package prompt
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"text/template"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+var (
+	ErrMissingRequiredInput = errors.New("missing required input artifact")
+	ErrUnknownInput         = errors.New("referenced unknown input artifact")
+	ErrInvalidTemplate      = errors.New("invalid prompt template")
+	ErrRenderFailure        = errors.New("prompt render failure")
+	ErrInvalidMessageRole   = errors.New("invalid or empty message role")
+)
+
+type goRenderer struct{}
+
+func NewGoRenderer() Renderer {
+	return &goRenderer{}
+}
+
+func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefinition, inputs map[string]*domain.Artifact, vars map[string]string) (*domain.RenderedPrompt, error) {
+	if definition == nil {
+		return nil, fmt.Errorf("%w: nil prompt definition", ErrRenderFailure)
+	}
+
+	// 1. Verify required inputs
+	for _, in := range definition.Inputs {
+		if !in.Required {
+			continue
+		}
+		art, ok := inputs[in.Name]
+		if !ok || art == nil {
+			return nil, fmt.Errorf("%w: %s", ErrMissingRequiredInput, in.Name)
+		}
+	}
+
+	// 2. Setup template functions
+	funcs := template.FuncMap{
+		"input": func(name string) (string, error) {
+			art, ok := inputs[name]
+			if !ok || art == nil {
+				return "", fmt.Errorf("%w: %s", ErrUnknownInput, name)
+			}
+			return string(art.Body), nil
+		},
+	}
+
+	sessionID, err := renderSessionID(definition.SessionID, funcs, vars)
+	if err != nil {
+		return nil, err
+	}
+
+	var renderedMessages []domain.RenderedMessage
+
+	for i, tmplMsg := range definition.Templates {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		default:
+		}
+
+		if tmplMsg.Role == "" {
+			return nil, fmt.Errorf("%w: message %d", ErrInvalidMessageRole, i)
+		}
+
+		// Parse and execute template
+		tmpl, err := template.New(fmt.Sprintf("msg_%d", i)).Funcs(funcs).Option("missingkey=error").Parse(tmplMsg.Content)
+		if err != nil {
+			return nil, fmt.Errorf("%w: message %d: %v", ErrInvalidTemplate, i, err)
+		}
+
+		var buf bytes.Buffer
+		if err := tmpl.Execute(&buf, vars); err != nil {
+			return nil, fmt.Errorf("%w: message %d: %w", ErrRenderFailure, i, err)
+		}
+
+		renderedMessages = append(renderedMessages, domain.RenderedMessage{
+			Role:         tmplMsg.Role,
+			Content:      buf.String(),
+			CacheControl: cloneCacheControl(tmplMsg.CacheControl),
+		})
+	}
+
+	return &domain.RenderedPrompt{
+		SessionID: sessionID,
+		Messages:  renderedMessages,
+	}, nil
+}
+
+func renderSessionID(raw string, funcs template.FuncMap, vars map[string]string) (string, error) {
+	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)
+	}
+
+	var buf bytes.Buffer
+	if err := tmpl.Execute(&buf, vars); err != nil {
+		return "", fmt.Errorf("%w: session_id: %w", ErrRenderFailure, err)
+	}
+
+	sessionID, err := domain.NormalizeSessionID(buf.String())
+	if err != nil {
+		return "", fmt.Errorf("%w: session_id: %v", ErrRenderFailure, err)
+	}
+	return sessionID, nil
+}
+
+func cloneCacheControl(in *domain.CacheControl) *domain.CacheControl {
+	if in == nil {
+		return nil
+	}
+	out := *in
+	return &out
+}

internal/prompt/renderer.go

@@ -0,0 +1,11 @@
+package prompt
+
+import (
+	"context"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// Renderer renders prompt templates using named artifacts and variables.
+type Renderer interface {
+	Render(ctx context.Context, definition *domain.PromptDefinition, inputs map[string]*domain.Artifact, vars map[string]string) (*domain.RenderedPrompt, error)
+}

internal/prompt/renderer_test.go

@@ -0,0 +1,345 @@
+package prompt
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestGoRenderer_Render(t *testing.T) {
+	renderer := NewGoRenderer()
+	ctx := context.Background()
+
+	inputs := map[string]*domain.Artifact{
+		"transcript": {Body: []byte("The quick brown fox.")},
+	}
+	vars := map[string]string{
+		"role": "helpful assistant",
+		"tone": "concise",
+	}
+
+	t.Run("rendering inline message content", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Analyze this: {{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if len(res.Messages) != 1 {
+			t.Fatalf("expected 1 message, got %d", len(res.Messages))
+		}
+		if res.Messages[0].Content != "Analyze this: The quick brown fox." {
+			t.Fatalf("unexpected rendered content: %q", res.Messages[0].Content)
+		}
+	})
+
+	t.Run("rendering file-backed message content loaded into prompt definition", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "From file: {{input \"transcript\"}}", ContentFile: "/tmp/user.tmpl"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if got := res.Messages[0].Content; got != "From file: The quick brown fox." {
+			t.Fatalf("unexpected file-backed render result: %q", got)
+		}
+	})
+
+	t.Run("rendering system and user messages", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "You are a {{.role}}."},
+				{Role: "user", Content: "Analyze this: {{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if len(res.Messages) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(res.Messages))
+		}
+		if res.Messages[0].Role != "system" || res.Messages[1].Role != "user" {
+			t.Fatalf("unexpected roles: %#v", res.Messages)
+		}
+	})
+
+	t.Run("copying cache control to rendered messages", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{
+					Role:    "system",
+					Content: "You are concise.",
+					CacheControl: &domain.CacheControl{
+						Type: domain.CacheControlEphemeral,
+						TTL:  "1h",
+					},
+				},
+				{Role: "user", Content: "Analyze this: {{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if len(res.Messages) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(res.Messages))
+		}
+		if res.Messages[0].CacheControl == nil {
+			t.Fatal("expected rendered cache control")
+		}
+		if res.Messages[0].CacheControl.Type != domain.CacheControlEphemeral {
+			t.Fatalf("unexpected cache control type: %q", res.Messages[0].CacheControl.Type)
+		}
+		if res.Messages[0].CacheControl.TTL != "1h" {
+			t.Fatalf("unexpected cache control ttl: %q", res.Messages[0].CacheControl.TTL)
+		}
+		if res.Messages[1].CacheControl != nil {
+			t.Fatalf("expected no cache control on second message, got %#v", res.Messages[1].CacheControl)
+		}
+	})
+
+	t.Run("rendered cache control does not alias source template", func(t *testing.T) {
+		source := &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "You are concise.", CacheControl: source},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.Messages[0].CacheControl == source {
+			t.Fatal("expected rendered cache control to be cloned")
+		}
+
+		res.Messages[0].CacheControl.TTL = ""
+		if source.TTL != "1h" {
+			t.Fatalf("source cache control was mutated, ttl=%q", source.TTL)
+		}
+	})
+
+	t.Run("accessing vars", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.Messages[0].Content != "Speak in a concise tone." {
+			t.Fatalf("unexpected vars rendering: %q", res.Messages[0].Content)
+		}
+	})
+
+	t.Run("rendering session id from vars", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "  {{ .session_id }}  ",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, map[string]string{
+			"tone":       "concise",
+			"session_id": "agent-session-123",
+		})
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.SessionID != "agent-session-123" {
+			t.Fatalf("unexpected session id: %q", res.SessionID)
+		}
+	})
+
+	t.Run("empty rendered session id is omitted", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "   ",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.SessionID != "" {
+			t.Fatalf("expected empty session id, got %q", res.SessionID)
+		}
+	})
+
+	t.Run("missing session id var fails rendering", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "{{ .session_id }}",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+	})
+
+	t.Run("too long rendered session id fails rendering", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "{{ .session_id }}",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, map[string]string{
+			"tone":       "concise",
+			"session_id": strings.Repeat("x", domain.SessionIDMaxLength+1),
+		})
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+	})
+
+	t.Run("inserting required input artifact", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "{{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.Messages[0].Content != "The quick brown fox." {
+			t.Fatalf("unexpected required input rendering: %q", res.Messages[0].Content)
+		}
+	})
+
+	t.Run("optional input absent and not referenced", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{
+				{Name: "transcript", Required: true},
+				{Name: "glossary", Required: false},
+			},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Transcript: {{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if len(res.Messages) != 1 {
+			t.Fatalf("expected one rendered message, got %d", len(res.Messages))
+		}
+	})
+
+	t.Run("optional input absent but referenced, expecting failure", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{
+				{Name: "transcript", Required: true},
+				{Name: "glossary", Required: false},
+			},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Glossary: {{input \"glossary\"}}"},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+		if !errors.Is(err, ErrUnknownInput) {
+			t.Fatalf("expected ErrUnknownInput, got %v", err)
+		}
+	})
+
+	t.Run("required input missing, expecting failure", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Analyze this: {{input \"transcript\"}}"},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, map[string]*domain.Artifact{}, vars)
+		if !errors.Is(err, ErrMissingRequiredInput) {
+			t.Fatalf("expected ErrMissingRequiredInput, got %v", err)
+		}
+	})
+
+	t.Run("invalid template syntax", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Hello {{.unclosed"},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrInvalidTemplate) {
+			t.Fatalf("expected ErrInvalidTemplate, got %v", err)
+		}
+	})
+
+	t.Run("unknown input reference", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "user", Content: "Hello {{input \"ghost\"}}"},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+		if !errors.Is(err, ErrUnknownInput) {
+			t.Fatalf("expected ErrUnknownInput, got %v", err)
+		}
+	})
+
+	t.Run("empty message role", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "", Content: "Hello"},
+			},
+		}
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrInvalidMessageRole) {
+			t.Fatalf("expected ErrInvalidMessageRole, got %v", err)
+		}
+	})
+}

internal/promptdef/filesystem_repository.go

@@ -0,0 +1,484 @@
+package promptdef
+
+import (
+	"bytes"
+	"context"
+	"errors"
+	"fmt"
+	"io/fs"
+	"os"
+	"path"
+	"path/filepath"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/filecatalog"
+	"gopkg.in/yaml.v3"
+)
+
+var (
+	ErrPromptDefinitionNotFound = errors.New("prompt definition not found")
+	ErrInvalidYAML              = errors.New("invalid YAML format")
+	ErrInvalidPromptDefinition  = errors.New("invalid prompt definition configuration")
+)
+
+type filesystemRepository struct {
+	dir string
+}
+
+type fsRepository struct {
+	fsys fs.FS
+	root string
+}
+
+type promptDefinitionFile struct {
+	ID             string                   `yaml:"id"`
+	Version        string                   `yaml:"version"`
+	DefaultProfile *string                  `yaml:"default_profile"`
+	Description    string                   `yaml:"description"`
+	SessionID      string                   `yaml:"session_id"`
+	Inputs         []promptInputFile        `yaml:"inputs"`
+	Messages       []promptMessageFile      `yaml:"messages"`
+	Output         promptOutputContractFile `yaml:"output"`
+}
+
+type promptInputFile struct {
+	Name        string `yaml:"name"`
+	Required    bool   `yaml:"required"`
+	ContentType string `yaml:"content_type"`
+	Description string `yaml:"description"`
+}
+
+type promptMessageFile struct {
+	Role         string            `yaml:"role"`
+	Content      string            `yaml:"content"`
+	ContentFile  string            `yaml:"content_file"`
+	CacheControl *cacheControlFile `yaml:"cache_control"`
+}
+
+type cacheControlFile struct {
+	Type string `yaml:"type"`
+	TTL  string `yaml:"ttl"`
+}
+
+type promptOutputContractFile struct {
+	Format         domain.OutputFormat   `yaml:"format"`
+	ValidationMode domain.ValidationMode `yaml:"validation_mode"`
+	SchemaPath     string                `yaml:"schema_path"`
+	RepairAttempts int                   `yaml:"repair_attempts"`
+}
+
+func NewFilesystemRepository(dir string) Repository {
+	return &filesystemRepository{dir: dir}
+}
+
+func NewFSRepository(fsys fs.FS, root string) Repository {
+	return &fsRepository{fsys: fsys, root: root}
+}
+
+func (r *filesystemRepository) GetPromptDefinition(ctx context.Context, id string, version string) (*domain.PromptDefinition, error) {
+	if strings.TrimSpace(id) == "" {
+		return nil, fmt.Errorf("%w: prompt id is required", ErrInvalidPromptDefinition)
+	}
+
+	files, err := filecatalog.FindYAMLFiles(ctx, r.dir)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read prompt definition directory: %w", err)
+	}
+
+	var matches []promptDefinitionMatch
+	for _, fullPath := range files {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		default:
+		}
+
+		relPath := filecatalog.RelativePath(r.dir, fullPath)
+		fileMatch := filecatalog.Stem(filepath.Base(fullPath)) == id
+
+		raw, err := loadPromptDefinitionFile(fullPath)
+		if err != nil {
+			if fileMatch || promptDefinitionFileHasID(fullPath, id) {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidYAML, relPath, err)
+			}
+			continue
+		}
+
+		def, err := normalizePromptDefinition(raw, fullPath)
+		if err != nil {
+			if fileMatch || strings.TrimSpace(raw.ID) == id {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidPromptDefinition, relPath, err)
+			}
+			continue
+		}
+
+		if def.ID != id {
+			continue
+		}
+		if version != "" && def.Version != version {
+			continue
+		}
+		matches = append(matches, promptDefinitionMatch{
+			def:  def,
+			path: relPath,
+		})
+	}
+
+	if len(matches) > 1 {
+		paths := make([]string, 0, len(matches))
+		for _, match := range matches {
+			paths = append(paths, match.path)
+		}
+		if version != "" {
+			return nil, fmt.Errorf("%w: duplicate prompt definition id %q version %q found in: %s", ErrInvalidPromptDefinition, id, version, strings.Join(paths, ", "))
+		}
+		return nil, fmt.Errorf("%w: duplicate prompt definition id %q found in: %s", ErrInvalidPromptDefinition, id, strings.Join(paths, ", "))
+	}
+
+	if len(matches) == 1 {
+		return matches[0].def, nil
+	}
+
+	return nil, ErrPromptDefinitionNotFound
+}
+
+func (r *fsRepository) GetPromptDefinition(ctx context.Context, id string, version string) (*domain.PromptDefinition, error) {
+	return loadPromptDefinition(ctx, r.fsys, r.root, id, version)
+}
+
+type promptDefinitionMatch struct {
+	def  *domain.PromptDefinition
+	path string
+}
+
+func loadPromptDefinitionFile(path string) (*promptDefinitionFile, error) {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read prompt definition file: %w", err)
+	}
+
+	var raw promptDefinitionFile
+	decoder := yaml.NewDecoder(bytes.NewReader(data))
+	decoder.KnownFields(true)
+	if err := decoder.Decode(&raw); err != nil {
+		return nil, err
+	}
+	return &raw, nil
+}
+
+func promptDefinitionFileHasID(path string, id string) bool {
+	data, err := os.ReadFile(path)
+	if err != nil {
+		return false
+	}
+	var raw struct {
+		ID string `yaml:"id"`
+	}
+	if err := yaml.NewDecoder(bytes.NewReader(data)).Decode(&raw); err != nil {
+		return false
+	}
+	return strings.TrimSpace(raw.ID) == id
+}
+
+func loadPromptDefinition(ctx context.Context, fsys fs.FS, root string, id string, version string) (*domain.PromptDefinition, error) {
+	if strings.TrimSpace(id) == "" {
+		return nil, fmt.Errorf("%w: prompt id is required", ErrInvalidPromptDefinition)
+	}
+	if fsys == nil {
+		return nil, fmt.Errorf("failed to read prompt definition directory: filesystem is nil")
+	}
+
+	files, err := filecatalog.FindFSYAMLFiles(ctx, fsys, root)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read prompt definition directory: %w", err)
+	}
+	cleanRoot := filecatalog.CleanFSRoot(root)
+	rootInfo, err := fs.Stat(fsys, cleanRoot)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read prompt definition directory: %w", err)
+	}
+
+	var matches []promptDefinitionMatch
+	for _, fullPath := range files {
+		select {
+		case <-ctx.Done():
+			return nil, ctx.Err()
+		default:
+		}
+
+		relPath := filecatalog.DisplayPath(root, fullPath)
+		fileMatch := filecatalog.Stem(path.Base(fullPath)) == id
+		data, err := fs.ReadFile(fsys, fullPath)
+		if err != nil {
+			if fileMatch {
+				return nil, fmt.Errorf("%w: %s: failed to read prompt definition file: %v", ErrInvalidYAML, relPath, err)
+			}
+			continue
+		}
+
+		raw, err := decodePromptDefinition(data)
+		if err != nil {
+			if fileMatch || promptDefinitionDataHasID(data, id) {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidYAML, relPath, err)
+			}
+			continue
+		}
+
+		def, err := normalizePromptDefinitionFromFS(raw, fsys, root, fullPath, rootInfo.IsDir())
+		if err != nil {
+			if fileMatch || strings.TrimSpace(raw.ID) == id {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidPromptDefinition, relPath, err)
+			}
+			continue
+		}
+
+		if def.ID != id {
+			continue
+		}
+		if version != "" && def.Version != version {
+			continue
+		}
+		matches = append(matches, promptDefinitionMatch{
+			def:  def,
+			path: relPath,
+		})
+	}
+
+	if len(matches) > 1 {
+		paths := make([]string, 0, len(matches))
+		for _, match := range matches {
+			paths = append(paths, match.path)
+		}
+		if version != "" {
+			return nil, fmt.Errorf("%w: duplicate prompt definition id %q version %q found in: %s", ErrInvalidPromptDefinition, id, version, strings.Join(paths, ", "))
+		}
+		return nil, fmt.Errorf("%w: duplicate prompt definition id %q found in: %s", ErrInvalidPromptDefinition, id, strings.Join(paths, ", "))
+	}
+
+	if len(matches) == 1 {
+		return matches[0].def, nil
+	}
+
+	return nil, ErrPromptDefinitionNotFound
+}
+
+func decodePromptDefinition(data []byte) (*promptDefinitionFile, error) {
+	var raw promptDefinitionFile
+	decoder := yaml.NewDecoder(bytes.NewReader(data))
+	decoder.KnownFields(true)
+	if err := decoder.Decode(&raw); err != nil {
+		return nil, err
+	}
+	return &raw, nil
+}
+
+func promptDefinitionDataHasID(data []byte, id string) bool {
+	var raw struct {
+		ID string `yaml:"id"`
+	}
+	if err := yaml.NewDecoder(bytes.NewReader(data)).Decode(&raw); err != nil {
+		return false
+	}
+	return strings.TrimSpace(raw.ID) == id
+}
+
+func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*domain.PromptDefinition, error) {
+	promptDir := filepath.Dir(sourcePath)
+	return normalizePromptDefinitionWithContent(raw, func(contentFile string) (string, string, error) {
+		resolvedPath := strings.TrimSpace(contentFile)
+		if !filepath.IsAbs(resolvedPath) {
+			resolvedPath = filepath.Join(promptDir, resolvedPath)
+		}
+		resolvedPath = filepath.Clean(resolvedPath)
+
+		body, err := os.ReadFile(resolvedPath)
+		if err != nil {
+			return "", "", err
+		}
+		return string(body), resolvedPath, nil
+	})
+}
+
+func normalizePromptDefinitionFromFS(raw *promptDefinitionFile, fsys fs.FS, root string, sourcePath string, rootIsDir bool) (*domain.PromptDefinition, error) {
+	promptDir := path.Dir(sourcePath)
+	return normalizePromptDefinitionWithContent(raw, func(contentFile string) (string, string, error) {
+		var resolvedPath string
+		if rootIsDir {
+			var err error
+			resolvedPath, _, err = filecatalog.ResolveFSPath(root, promptDir, contentFile)
+			if err != nil {
+				return "", "", err
+			}
+		} else {
+			resolvedPath = strings.TrimSpace(contentFile)
+			if !path.IsAbs(resolvedPath) {
+				resolvedPath = path.Join(promptDir, resolvedPath)
+			}
+			resolvedPath = strings.TrimPrefix(path.Clean(resolvedPath), "/")
+		}
+
+		body, err := fs.ReadFile(fsys, resolvedPath)
+		if err != nil {
+			return "", "", err
+		}
+		return string(body), resolvedPath, nil
+	})
+}
+
+func normalizePromptDefinitionWithContent(raw *promptDefinitionFile, readContentFile func(string) (string, string, error)) (*domain.PromptDefinition, error) {
+	if raw == nil {
+		return nil, errors.New("prompt definition is nil")
+	}
+
+	id := strings.TrimSpace(raw.ID)
+	if id == "" {
+		return nil, errors.New("id is required")
+	}
+
+	version := strings.TrimSpace(raw.Version)
+	if version == "" {
+		return nil, errors.New("version is required")
+	}
+
+	if len(raw.Messages) == 0 {
+		return nil, errors.New("at least one message is required")
+	}
+
+	inputs := make([]domain.PromptInput, 0, len(raw.Inputs))
+	seenInputNames := make(map[string]struct{}, len(raw.Inputs))
+	for i, in := range raw.Inputs {
+		name := strings.TrimSpace(in.Name)
+		if name == "" {
+			return nil, fmt.Errorf("input %d has empty name", i)
+		}
+		if _, exists := seenInputNames[name]; exists {
+			return nil, fmt.Errorf("duplicate input name %q", name)
+		}
+		seenInputNames[name] = struct{}{}
+
+		inputs = append(inputs, domain.PromptInput{
+			Name:        name,
+			Required:    in.Required,
+			ContentType: strings.TrimSpace(in.ContentType),
+			Description: strings.TrimSpace(in.Description),
+		})
+	}
+
+	templates := make([]domain.PromptMessageTemplate, 0, len(raw.Messages))
+	for i, msg := range raw.Messages {
+		role := strings.TrimSpace(msg.Role)
+		if role == "" {
+			return nil, fmt.Errorf("message %d role is required", i)
+		}
+
+		hasContent := strings.TrimSpace(msg.Content) != ""
+		hasContentFile := strings.TrimSpace(msg.ContentFile) != ""
+		if hasContent == hasContentFile {
+			return nil, fmt.Errorf("message %d (%s) must set exactly one of content or content_file", i, role)
+		}
+
+		cacheControl, err := normalizeCacheControl(msg.CacheControl)
+		if err != nil {
+			return nil, fmt.Errorf("message %d (%s) cache_control: %w", i, role, err)
+		}
+
+		templateContent := msg.Content
+		resolvedContentFile := ""
+		if hasContentFile {
+			body, resolvedPath, err := readContentFile(msg.ContentFile)
+			if err != nil {
+				return nil, fmt.Errorf("prompt %q message %d (%s): failed to read content_file %q: %w", id, i, role, msg.ContentFile, err)
+			}
+			templateContent = body
+			resolvedContentFile = resolvedPath
+		}
+
+		templates = append(templates, domain.PromptMessageTemplate{
+			Role:         role,
+			Content:      templateContent,
+			ContentFile:  resolvedContentFile,
+			CacheControl: cacheControl,
+		})
+	}
+
+	if !isValidOutputFormat(raw.Output.Format) {
+		return nil, fmt.Errorf("invalid output format: %q", raw.Output.Format)
+	}
+	if !isValidValidationMode(raw.Output.ValidationMode) {
+		return nil, fmt.Errorf("invalid validation mode: %q", raw.Output.ValidationMode)
+	}
+	if raw.Output.ValidationMode == domain.ValidationJSONSchema && strings.TrimSpace(raw.Output.SchemaPath) == "" {
+		return nil, errors.New("output.schema_path is required when output.validation_mode is json_schema")
+	}
+	if raw.Output.RepairAttempts < 0 {
+		return nil, errors.New("output.repair_attempts must be greater than or equal to 0")
+	}
+
+	defaultProfile := ""
+	if raw.DefaultProfile != nil {
+		defaultProfile = strings.TrimSpace(*raw.DefaultProfile)
+		if defaultProfile == "" {
+			return nil, errors.New("default_profile must be a non-empty string when set")
+		}
+	}
+
+	return &domain.PromptDefinition{
+		ID:             id,
+		Version:        version,
+		DefaultProfile: defaultProfile,
+		Description:    strings.TrimSpace(raw.Description),
+		SessionID:      strings.TrimSpace(raw.SessionID),
+		Inputs:         inputs,
+		Templates:      templates,
+		OutputFormat:   raw.Output.Format,
+		Validation: domain.OutputContract{
+			Format:         raw.Output.Format,
+			ValidationMode: raw.Output.ValidationMode,
+			SchemaPath:     strings.TrimSpace(raw.Output.SchemaPath),
+			RepairAttempts: raw.Output.RepairAttempts,
+		},
+	}, nil
+}
+
+func normalizeCacheControl(raw *cacheControlFile) (*domain.CacheControl, error) {
+	if raw == nil {
+		return nil, nil
+	}
+
+	cacheType := strings.TrimSpace(raw.Type)
+	if cacheType == "" {
+		return nil, errors.New("type is required")
+	}
+	if domain.CacheControlType(cacheType) != domain.CacheControlEphemeral {
+		return nil, fmt.Errorf("unsupported type %q", cacheType)
+	}
+
+	ttl := strings.TrimSpace(raw.TTL)
+	if ttl != "" && ttl != "1h" {
+		return nil, fmt.Errorf("unsupported ttl %q", ttl)
+	}
+
+	return &domain.CacheControl{
+		Type: domain.CacheControlType(cacheType),
+		TTL:  ttl,
+	}, nil
+}
+
+func isValidOutputFormat(f domain.OutputFormat) bool {
+	switch f {
+	case domain.FormatText, domain.FormatMarkdown, domain.FormatJSON:
+		return true
+	default:
+		return false
+	}
+}
+
+func isValidValidationMode(m domain.ValidationMode) bool {
+	switch m {
+	case domain.ValidationNone, domain.ValidationBasic, domain.ValidationJSON, domain.ValidationJSONSchema:
+		return true
+	default:
+		return false
+	}
+}

internal/promptdef/repository.go

@@ -0,0 +1,12 @@
+package promptdef
+
+import (
+	"context"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+// Repository loads prompt definitions.
+type Repository interface {
+	GetPromptDefinition(ctx context.Context, id string, version string) (*domain.PromptDefinition, error)
+}

internal/promptdef/repository_test.go

@@ -0,0 +1,526 @@
+package promptdef
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"strings"
+	"testing"
+	"testing/fstest"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+)
+
+func TestFilesystemRepository_GetPromptDefinition(t *testing.T) {
+	tmpDir := t.TempDir()
+	if err := copyTree("testdata", tmpDir); err != nil {
+		t.Fatalf("failed to copy testdata: %v", err)
+	}
+
+	repo := NewFilesystemRepository(tmpDir)
+	ctx := context.Background()
+
+	t.Run("valid inline prompt", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-inline", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "valid-inline" {
+			t.Fatalf("unexpected id: %q", p.ID)
+		}
+		if p.Version != "1.0.0" {
+			t.Fatalf("unexpected version: %q", p.Version)
+		}
+		if p.OutputFormat != domain.FormatMarkdown {
+			t.Fatalf("unexpected output format: %q", p.OutputFormat)
+		}
+		if p.Validation.ValidationMode != domain.ValidationBasic {
+			t.Fatalf("unexpected validation mode: %q", p.Validation.ValidationMode)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		if len(p.Inputs) != 1 {
+			t.Fatalf("expected 1 input, got %d", len(p.Inputs))
+		}
+		if p.Inputs[0].ContentType != "text/markdown" {
+			t.Fatalf("expected input content_type to be preserved, got %q", p.Inputs[0].ContentType)
+		}
+	})
+
+	t.Run("valid file-backed prompt", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-file-backed", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		if !strings.Contains(p.Templates[1].Content, "{{input \"transcript\"}}") {
+			t.Fatalf("expected content_file template body to be loaded, got %q", p.Templates[1].Content)
+		}
+		if p.Templates[1].ContentFile == "" {
+			t.Fatal("expected ContentFile source metadata to be preserved")
+		}
+		if !filepath.IsAbs(p.Templates[1].ContentFile) {
+			t.Fatalf("expected resolved content_file path to be absolute, got %q", p.Templates[1].ContentFile)
+		}
+	})
+
+	t.Run("valid cache control with ttl", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-cache-control-ttl", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		assertCacheControl(t, p.Templates[0].CacheControl, domain.CacheControlEphemeral, "1h")
+		if p.Templates[1].CacheControl != nil {
+			t.Fatalf("expected second message cache control to be nil, got %#v", p.Templates[1].CacheControl)
+		}
+	})
+
+	t.Run("valid cache control without ttl", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-cache-control-without-ttl", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		assertCacheControl(t, p.Templates[0].CacheControl, domain.CacheControlEphemeral, "")
+		if p.Templates[1].CacheControl != nil {
+			t.Fatalf("expected second message cache control to be nil, got %#v", p.Templates[1].CacheControl)
+		}
+	})
+
+	t.Run("valid session id template", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-session-id", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.SessionID != "{{ .session_id }}" {
+			t.Fatalf("expected trimmed session_id template, got %q", p.SessionID)
+		}
+	})
+
+	t.Run("valid nested file-backed prompt resolves content file relative to nested YAML", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "dnd", "recap")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePromptTestFile(t, filepath.Join(nestedDir, "nested_recap.yaml"), `
+id: nested-recap
+version: "1.0.0"
+messages:
+  - role: user
+    content_file: ./nested_recap.user.tmpl
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+		writePromptTestFile(t, filepath.Join(nestedDir, "nested_recap.user.tmpl"), `Nested recap: {{input "transcript"}}`)
+
+		p, err := repo.GetPromptDefinition(ctx, "nested-recap", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 1 {
+			t.Fatalf("expected one template, got %d", len(p.Templates))
+		}
+		if !strings.Contains(p.Templates[0].Content, "Nested recap") {
+			t.Fatalf("expected nested content file body, got %q", p.Templates[0].Content)
+		}
+		if !strings.Contains(p.Templates[0].ContentFile, filepath.Join("dnd", "recap", "nested_recap.user.tmpl")) {
+			t.Fatalf("expected nested content file path, got %q", p.Templates[0].ContentFile)
+		}
+	})
+
+	t.Run("prompt with default_profile", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "with-default-profile", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.DefaultProfile != "local-default" {
+			t.Fatalf("unexpected default profile: %q", p.DefaultProfile)
+		}
+		if len(p.Inputs) != 1 {
+			t.Fatalf("expected one input, got %d", len(p.Inputs))
+		}
+		if p.Inputs[0].ContentType != "" {
+			t.Fatalf("expected missing content_type to remain empty, got %q", p.Inputs[0].ContentType)
+		}
+	})
+
+	t.Run("duplicate prompt IDs fail as ambiguous", func(t *testing.T) {
+		writePromptTestFile(t, filepath.Join(tmpDir, "duplicate_a.yaml"), `
+id: duplicate-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content: First duplicate.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+		nestedDir := filepath.Join(tmpDir, "nested")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePromptTestFile(t, filepath.Join(nestedDir, "duplicate_b.yaml"), `
+id: duplicate-prompt
+version: "2.0.0"
+messages:
+  - role: user
+    content: Second duplicate.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+
+		_, err := repo.GetPromptDefinition(ctx, "duplicate-prompt", "")
+		if !errors.Is(err, ErrInvalidPromptDefinition) {
+			t.Fatalf("expected duplicate prompt to return ErrInvalidPromptDefinition, got %v", err)
+		}
+		for _, want := range []string{"duplicate prompt definition id", "duplicate_a.yaml", filepath.Join("nested", "duplicate_b.yaml")} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+
+	t.Run("duplicate prompt ID and requested version fails as ambiguous", func(t *testing.T) {
+		writePromptTestFile(t, filepath.Join(tmpDir, "version_duplicate_a.yaml"), `
+id: duplicate-version-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content: First duplicate version.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+		nestedDir := filepath.Join(tmpDir, "versioned")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePromptTestFile(t, filepath.Join(nestedDir, "version_duplicate_b.yaml"), `
+id: duplicate-version-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content: Second duplicate version.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+
+		_, err := repo.GetPromptDefinition(ctx, "duplicate-version-prompt", "1.0.0")
+		if !errors.Is(err, ErrInvalidPromptDefinition) {
+			t.Fatalf("expected duplicate prompt version to return ErrInvalidPromptDefinition, got %v", err)
+		}
+		for _, want := range []string{"duplicate prompt definition id", "version \"1.0.0\"", "version_duplicate_a.yaml", filepath.Join("versioned", "version_duplicate_b.yaml")} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+
+	t.Run("non-matching malformed nested prompt is ignored for not found lookup", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "broken")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePromptTestFile(t, filepath.Join(nestedDir, "unrelated.yaml"), "id: [")
+
+		_, err := repo.GetPromptDefinition(ctx, "does-not-exist-even-with-broken-nested-file", "")
+		if !errors.Is(err, ErrPromptDefinitionNotFound) {
+			t.Fatalf("expected ErrPromptDefinitionNotFound, got %v", err)
+		}
+	})
+
+	t.Run("strict decode failure in nested prompt matches by YAML ID", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "strict")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writePromptTestFile(t, filepath.Join(nestedDir, "not_named_like_id.yaml"), `
+id: nested-strict-error
+version: "1.0.0"
+unknown_field: true
+messages:
+  - role: user
+    content: Invalid because of unknown field.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)
+
+		_, err := repo.GetPromptDefinition(ctx, "nested-strict-error", "")
+		if !errors.Is(err, ErrInvalidYAML) {
+			t.Fatalf("expected ErrInvalidYAML, got %v", err)
+		}
+		if !strings.Contains(err.Error(), filepath.Join("strict", "not_named_like_id.yaml")) {
+			t.Fatalf("expected nested path in error, got %v", err)
+		}
+	})
+
+	t.Run("version lookup", func(t *testing.T) {
+		_, err := repo.GetPromptDefinition(ctx, "valid-inline", "9.9.9")
+		if !errors.Is(err, ErrPromptDefinitionNotFound) {
+			t.Fatalf("expected ErrPromptDefinitionNotFound, got %v", err)
+		}
+	})
+
+	cases := []struct {
+		name       string
+		id         string
+		targetErr  error
+		errSubstrs []string
+	}{
+		{name: "invalid YAML", id: "invalid_yaml", targetErr: ErrInvalidYAML},
+		{name: "missing id", id: "missing_id", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"id is required"}},
+		{name: "no messages", id: "no_messages", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"at least one message is required"}},
+		{name: "both content and content_file", id: "both_content_and_content_file", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"exactly one"}},
+		{name: "neither content nor content_file", id: "neither_content_nor_content_file", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"exactly one"}},
+		{name: "missing content_file", id: "missing_content_file", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"failed to read content_file"}},
+		{name: "duplicate input names", id: "duplicate_input_names", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"duplicate input name"}},
+		{name: "invalid validation mode", id: "invalid_validation_mode", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"invalid validation mode"}},
+		{name: "json_schema without schema_path", id: "json_schema_without_schema_path", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"schema_path"}},
+		{name: "unknown input field", id: "unknown_input_field", targetErr: ErrInvalidYAML, errSubstrs: []string{"field unknown_input_setting not found"}},
+		{name: "empty cache control type", id: "empty_cache_control_type", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "type is required"}},
+		{name: "unsupported cache control type", id: "unsupported_cache_control_type", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "unsupported type"}},
+		{name: "unsupported cache control ttl", id: "unsupported_cache_control_ttl", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "unsupported ttl"}},
+		{name: "unknown cache control field", id: "unknown_cache_control_field", targetErr: ErrInvalidYAML, errSubstrs: []string{"field unexpected not found"}},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := repo.GetPromptDefinition(ctx, tc.id, "")
+			if !errors.Is(err, tc.targetErr) {
+				t.Fatalf("expected %v, got %v", tc.targetErr, err)
+			}
+			for _, sub := range tc.errSubstrs {
+				if !strings.Contains(err.Error(), sub) {
+					t.Fatalf("expected error to contain %q, got %v", sub, err)
+				}
+			}
+		})
+	}
+
+	t.Run("prompt definition not found", func(t *testing.T) {
+		_, err := repo.GetPromptDefinition(ctx, "does-not-exist", "")
+		if !errors.Is(err, ErrPromptDefinitionNotFound) {
+			t.Fatalf("expected ErrPromptDefinitionNotFound, got %v", err)
+		}
+	})
+}
+
+func TestFSRepositoryGetPromptDefinition(t *testing.T) {
+	repo := NewFSRepository(fstest.MapFS{
+		"prompts/nested/prompt.yaml": &fstest.MapFile{Data: []byte(`
+id: fs-prompt
+version: "1.0.0"
+inputs:
+  - name: transcript
+    required: true
+messages:
+  - role: user
+    content_file: ./messages/user.tmpl
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)},
+		"prompts/nested/messages/user.tmpl": &fstest.MapFile{Data: []byte(`Summarize {{input "transcript"}}.`)},
+	}, "prompts")
+
+	got, err := repo.GetPromptDefinition(context.Background(), "fs-prompt", "")
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if got.ID != "fs-prompt" {
+		t.Fatalf("unexpected prompt id: %q", got.ID)
+	}
+	if len(got.Templates) != 1 || !strings.Contains(got.Templates[0].Content, `{{input "transcript"}}`) {
+		t.Fatalf("expected content_file body to be loaded, got %+v", got.Templates)
+	}
+	if got.Templates[0].ContentFile != "prompts/nested/messages/user.tmpl" {
+		t.Fatalf("unexpected content file path: %q", got.Templates[0].ContentFile)
+	}
+}
+
+func TestFSRepositoryContentFileContainment(t *testing.T) {
+	t.Run("nested prompt can reference file inside root", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"prompts/nested/prompt.yaml": &fstest.MapFile{Data: []byte(`
+id: fs-contained-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content_file: ../shared/user.tmpl
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)},
+			"prompts/shared/user.tmpl": &fstest.MapFile{Data: []byte(`Inside root.`)},
+		}, "prompts")
+
+		got, err := repo.GetPromptDefinition(context.Background(), "fs-contained-prompt", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(got.Templates) != 1 || got.Templates[0].Content != "Inside root." {
+			t.Fatalf("expected contained content file, got %+v", got.Templates)
+		}
+	})
+
+	tests := []struct {
+		name        string
+		contentFile string
+		wantErr     string
+	}{
+		{name: "parent escape rejected", contentFile: "../outside.tmpl", wantErr: "escapes source root"},
+		{name: "absolute path rejected", contentFile: "/outside.tmpl", wantErr: "must be relative"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			repo := NewFSRepository(fstest.MapFS{
+				"prompts/prompt.yaml": &fstest.MapFile{Data: []byte(`
+id: fs-escaped-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content_file: ` + tc.contentFile + `
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0
+`)},
+				"outside.tmpl": &fstest.MapFile{Data: []byte(`Outside root.`)},
+			}, "prompts")
+
+			_, err := repo.GetPromptDefinition(context.Background(), "fs-escaped-prompt", "")
+			if !errors.Is(err, ErrInvalidPromptDefinition) {
+				t.Fatalf("expected ErrInvalidPromptDefinition, got %v", err)
+			}
+			if !strings.Contains(err.Error(), tc.wantErr) {
+				t.Fatalf("expected error to contain %q, got %v", tc.wantErr, err)
+			}
+		})
+	}
+}
+
+func TestFSRepositoryRejectsDuplicatePromptIDs(t *testing.T) {
+	repo := NewFSRepository(fstest.MapFS{
+		"one.yaml": &fstest.MapFile{Data: []byte(`
+id: duplicate-fs-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content: First.
+output:
+  format: text
+  validation_mode: none
+  repair_attempts: 0
+`)},
+		"nested/two.yaml": &fstest.MapFile{Data: []byte(`
+id: duplicate-fs-prompt
+version: "1.0.0"
+messages:
+  - role: user
+    content: Second.
+output:
+  format: text
+  validation_mode: none
+  repair_attempts: 0
+`)},
+	}, ".")
+
+	_, err := repo.GetPromptDefinition(context.Background(), "duplicate-fs-prompt", "")
+	if !errors.Is(err, ErrInvalidPromptDefinition) {
+		t.Fatalf("expected ErrInvalidPromptDefinition, got %v", err)
+	}
+	if !strings.Contains(err.Error(), "one.yaml") || !strings.Contains(err.Error(), "nested/two.yaml") {
+		t.Fatalf("expected duplicate paths in error, got %v", err)
+	}
+}
+
+func TestFSRepositoryRejectsUnknownYAMLFields(t *testing.T) {
+	repo := NewFSRepository(fstest.MapFS{
+		"not_named_like_id.yaml": &fstest.MapFile{Data: []byte(`
+id: strict-fs-prompt
+version: "1.0.0"
+unknown: true
+messages:
+  - role: user
+    content: Invalid.
+output:
+  format: text
+  validation_mode: none
+  repair_attempts: 0
+`)},
+	}, ".")
+
+	_, err := repo.GetPromptDefinition(context.Background(), "strict-fs-prompt", "")
+	if !errors.Is(err, ErrInvalidYAML) {
+		t.Fatalf("expected ErrInvalidYAML, got %v", err)
+	}
+}
+
+func assertCacheControl(t *testing.T, got *domain.CacheControl, wantType domain.CacheControlType, wantTTL string) {
+	t.Helper()
+	if got == nil {
+		t.Fatal("expected cache control, got nil")
+	}
+	if got.Type != wantType {
+		t.Fatalf("unexpected cache control type: got %q want %q", got.Type, wantType)
+	}
+	if got.TTL != wantTTL {
+		t.Fatalf("unexpected cache control ttl: got %q want %q", got.TTL, wantTTL)
+	}
+}
+
+func writePromptTestFile(t *testing.T, path string, content string) {
+	t.Helper()
+	if err := os.WriteFile(path, []byte(strings.TrimLeft(content, "\n")), 0o644); err != nil {
+		t.Fatalf("failed to write prompt test file %q: %v", path, err)
+	}
+}
+
+func copyTree(src, dst string) error {
+	return filepath.WalkDir(src, func(path string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		rel, err := filepath.Rel(src, path)
+		if err != nil {
+			return err
+		}
+		if rel == "." {
+			return nil
+		}
+
+		target := filepath.Join(dst, rel)
+		if d.IsDir() {
+			return os.MkdirAll(target, 0o755)
+		}
+
+		data, err := os.ReadFile(path)
+		if err != nil {
+			return err
+		}
+		return os.WriteFile(target, data, 0o644)
+	})
+}