Prepare the v0.9.0 release

README.md

@@ -33,7 +33,10 @@ boundary and constraints that framework work must preserve.
 
 ## Release Guidance
 
-Consumers upgrading from `v0.7.0` to `v0.8.0` should read the
+Consumers upgrading from `v0.8.0` to `v0.9.0` should read the
+[v0.9.0 changelog and migration guide](docs/releases/v0.9.0.md).
+
+Consumers upgrading from `v0.7.0` to `v0.8.0` can consult the
 [v0.8.0 changelog and migration guide](docs/releases/v0.8.0.md).
 
 Earlier adopters can consult the

appended_messages_contract_test.go

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

convert.go

@@ -1,7 +1,9 @@
 package promptkit
 
 import (
+	"fmt"
 	"reflect"
+	"unicode/utf8"
 
 	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
 	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
@@ -12,6 +14,10 @@ func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
 	if err != nil {
 		return domain.RunRequest{}, err
 	}
+	appendedMessages, err := toDomainAppendedMessages(req.AppendedMessages)
+	if err != nil {
+		return domain.RunRequest{}, err
+	}
 	return domain.RunRequest{
 		PromptID:         req.PromptID,
 		PromptVersion:    req.PromptVersion,
@@ -22,9 +28,38 @@ func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
 		Vars:             copyStringMap(req.Vars),
 		Execution:        execution,
 		Validation:       toDomainOutputContractPtr(req.Validation),
+		AppendedMessages: appendedMessages,
 	}, nil
 }
 
+func toDomainAppendedMessages(messages []RenderedMessage) ([]domain.RenderedMessage, error) {
+	if messages == nil {
+		return nil, nil
+	}
+
+	converted := make([]domain.RenderedMessage, len(messages))
+	for index, message := range messages {
+		if !utf8.ValidString(message.Content) {
+			return nil, fmt.Errorf("appended message %d content must be valid UTF-8", index)
+		}
+		role, err := domain.NormalizeMessageRole(message.Role)
+		if err != nil {
+			return nil, fmt.Errorf("appended message %d role: %w", index, err)
+		}
+
+		cacheControl, err := domain.NormalizeCacheControl(toDomainCacheControl(message.CacheControl))
+		if err != nil {
+			return nil, fmt.Errorf("appended message %d cache_control: %w", index, err)
+		}
+		converted[index] = domain.RenderedMessage{
+			Role:         role,
+			Content:      message.Content,
+			CacheControl: cacheControl,
+		}
+	}
+	return converted, nil
+}
+
 func fromDomainPreparedRun(prepared *domain.PreparedRun) *PreparedRun {
 	if prepared == nil {
 		return nil
@@ -285,6 +320,16 @@ func fromDomainRenderedMessages(messages []domain.RenderedMessage) []RenderedMes
 	return out
 }
 
+func toDomainCacheControl(cacheControl *CacheControl) *domain.CacheControl {
+	if cacheControl == nil {
+		return nil
+	}
+	return &domain.CacheControl{
+		Type: domain.CacheControlType(cacheControl.Type),
+		TTL:  cacheControl.TTL,
+	}
+}
+
 func fromDomainCacheControl(cacheControl *domain.CacheControl) *CacheControl {
 	if cacheControl == nil {
 		return nil

convert_appended_messages_test.go

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

docs/consumers/pkg-promptkit.md

@@ -196,6 +196,26 @@ validity is not evidence of factual or domain correctness. See the
 [`OutputContract` GoDoc](../../types.go) for the exact budget and eligibility
 rules.
 
+### Append Already-Rendered Messages
+
+An application can include an earlier assistant response and its own corrective
+instruction in a fresh request without changing the configured prompt:
+
+```go
+request.AppendedMessages = []promptkit.RenderedMessage{
+	{Role: promptkit.RoleAssistant, Content: previousResponse},
+	{Role: promptkit.RoleUser, Content: correction},
+}
+result, err := engine.Run(ctx, request)
+```
+
+These messages are already rendered: Promptkit does not template or resolve
+files in them, and they can contain sensitive model output or application
+feedback. Promptkit remains stateless; every `Run` call re-resolves its current
+sources and the application owns any semantic retry budget. When a
+pre-execution equality check is required, use `PrepareExecution` and compare
+its opaque rendered-prompt hash before invoking `RunPrepared`.
+
 ## Inputs, Profiles, And Overrides
 
 Use `File`, `Inline`, or `InlineWithURI` to construct request inputs. A request

docs/development.md

@@ -34,6 +34,7 @@ Start with:
 | Root public API | The [architecture policy](policy/architecture.md), [consumer guide](consumers/pkg-promptkit.md), [testing policy](policy/testing.md), and existing GoDoc. |
 | Prompt, profile, or schema formats | The [framework format reference](formats.md), owning parser or validator package, and [documentation policy](policy/documentation.md). |
 | Source loading or validation | The [framework format reference](formats.md), [internal source document](internal/sources.md), and owning package tests. |
+| Maintained external catalogs or built-ins | The [internal source document](internal/sources.md), [framework format reference](formats.md), [testing policy](policy/testing.md), both catalog module repositories, and each catalog's release procedure. |
 | 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. |

docs/formats.md

@@ -82,7 +82,13 @@ allowed.
 
 ### Messages And Templates
 
-Each message has a non-empty `role` and exactly one of:
+Each message has a `role` that Promptkit trims and lowercases. It must then be
+exactly one of `developer`, `system`, `user`, or `assistant`; blank, custom,
+`tool`, and `function` roles are invalid. This intentionally tightens the
+previous nonblank-string rule. Consumers migrating to the next minor release
+must update any nonstandard prompt-definition roles before upgrading.
+
+Each message also has exactly one of:
 
 - `content`, containing an inline Go template; or
 - `content_file`, naming a file whose contents are the Go template.
@@ -277,7 +283,7 @@ Profile sources resolve matching IDs in this order:
 2. the ordinary configured source selected by a profile file, `fs.FS`, or
    configured profile directory;
 3. application fallback profiles supplied with `WithFallbackProfileFS`; and
-4. embedded built-in profiles.
+4. maintained external catalog profiles.
 
 A profile source supplies a complete definition; definitions and their fields
 are not merged across sources. A higher-precedence source falls back only when

docs/integrations/openai-compatible-chat.md

@@ -55,6 +55,12 @@ Each ordinary message contains its `role` and string `content`. A
 cache-controlled message instead uses a text content block containing `type`,
 `text`, and `cache_control`; an empty cache-control TTL is omitted.
 
+Promptkit sends only `developer`, `system`, `user`, and `assistant` roles and
+does so without provider-specific translation. Tool and deprecated function
+payloads are outside this text-message contract. A backend or model that
+rejects an otherwise supported role or context returns its ordinary provider
+error, which follows the normal generation-error path.
+
 The effective direct or prompt-rendered session ID is trimmed, limited to 256
 Unicode code points, and sent when nonempty as top-level `session_id`. It is
 never also sent as a session header.

docs/internal/llm.md

@@ -21,8 +21,9 @@ uses internal domain values for rendered prompts, execution targets,
 structured output, responses, and token usage.
 
 The runner supplies a fully resolved target after applying backend, profile,
-and request precedence. The client uses its endpoint, credential metadata,
+and request precedence, plus canonical provider-bound text messages. The
-generation fields, and extra parameters. `BackendID` remains routing metadata
+client uses its endpoint, credential metadata, generation fields, and extra
+parameters. `BackendID` remains routing metadata
 for the generation boundary and is not mapped into the provider payload.
 
 Construction trims and validates a nonempty configured base URL and clones any

docs/internal/overview.md

@@ -14,7 +14,8 @@ contributor workflow and validation.
 | Root `promptkit` package | Provides the supported engine facade, source, backend-registration, and injection options, public request, result, prompt-inspection, and profile-inspection values, opaque prepared-execution handles, profile construction, extension interfaces, value conversion, redacted formatting, typed capacity and generation error mapping, engine-local profile-source assembly including application fallbacks, and bounded output-repair assembly. | [Package GoDoc](../../doc.go), [prepared execution](../../prepared_execution.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 maintained built-in definitions and consumer additions, validates and defensively copies definitions through the shared JSON-value package, and consumes the LLM-owned OpenAI-compatible reserved request-field rule. | [Backend registry](../../internal/backend/registry.go) |
+| `internal/backend` | Constructs each engine's immutable registry from maintained definitions and consumer additions, validates and defensively copies definitions through the shared JSON-value package, and consumes the LLM-owned OpenAI-compatible reserved request-field rule. | [Backend registry](../../internal/backend/registry.go) |
+| `internal/catalog` | Strictly validates imported immutable maintained backend and profile catalog assets before engine assembly uses them. | [Catalog adapter](../../internal/catalog/catalog.go), [internal sources](sources.md#profiles-and-built-ins) |
 | `internal/capacity` | Owns engine-local bounded execution admission and FIFO model-generation permits for limited backend IDs, including cancellation-safe waiter removal and client wrapping. | [Internal capacity management](capacity.md) |
 | `internal/domain` | Defines internal framework values for requests, artifacts, prompt definitions, profiles, execution targets, rendering, generation, and validation, and owns source-neutral invariants for shared execution settings, OpenAI-compatible base endpoints, session identifiers, and output contracts. Source parsing, required fields, other source-specific normalization, defaulting, and boundary-specific error classification remain with their callers. | [Domain declarations](../../internal/domain/domain.go), [endpoint invariant](../../internal/domain/endpoint.go) |
 | `internal/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) |
@@ -22,7 +23,6 @@ contributor workflow and validation.
 | `internal/jsonvalue` | Validates and deeply copies bounded JSON-compatible extra-parameter and prepared-schema trees while preserving supported concrete value types and rejecting cycles or excessive depth and work. | [JSON values](../../internal/jsonvalue/jsonvalue.go) |
 | `internal/promptdef` | Loads strictly decoded, validated prompt definitions from filesystem and `fs.FS` sources, including version selection and contained file-backed message content. | [Framework formats](../formats.md), [prompt-definition repository](../../internal/promptdef/filesystem_repository.go) |
 | `internal/profile` | Loads strictly decoded, locally validated execution profiles from filesystem and `fs.FS` sources, overlays raw sources with error-preserving fallback, and resolves inherited profiles. | [Framework formats](../formats.md), [profile repositories](../../internal/profile/filesystem_repository.go), [internal sources](sources.md#profiles-and-built-ins) |
-| `internal/profile/builtin` | Embeds the built-in profile catalog, whose entries select maintained built-in backends. | [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 and creates operation-local validation plans with canonical contained schema resources. | [Framework formats](../formats.md#schemas), [internal sources and validation](sources.md) |

docs/internal/runner.md

@@ -84,7 +84,8 @@ profile, or backend:
    schema metadata from it when required;
 2. load and hash input artifacts;
 3. render messages and the prompt-defined session;
-4. apply any direct session ID;
+4. apply any direct session ID, then append already-normalized request messages
+   after the rendered definition messages;
 5. hash the effective rendered prompt; and
 6. construct the prepared value and preparation timing.
 
@@ -113,7 +114,10 @@ the prompt session template, and is applied after ordinary message rendering.
 A blank direct value retains prompt-template behavior. The runner clears the
 template only on a value copy of the definition, so the definition hash always
 describes the original source while the rendered-prompt hash includes the
-effective direct or rendered session.
+effective direct or rendered session and the complete effective message
+sequence. The rendered-prompt hash uses a versioned, length-framed SHA-256
+encoding of that session, every role and content value, and cache-control
+presence and values; its hexadecimal value is opaque.
 
 The registry is read-only after engine construction. Concurrent `Prepare` and
 `Run` calls resolve independent defensive backend values and keep all
@@ -146,9 +150,10 @@ each actual generation call.
 
 After a failed `basic`, JSON, or JSON Schema validation with a positive frozen
 budget, the installed repairer can make a bounded corrective call. Each request
-starts with a fresh copy of the complete original rendered messages, includes
+starts with a fresh copy of the complete effective message sequence (the
-only the latest nonempty candidate as an assistant message, and appends one
+configured prefix followed by the request suffix), includes only the latest
-corrective user message. Empty candidates omit that assistant message. The
+nonempty candidate as an assistant message, and appends one corrective user
+message. Empty candidates omit that assistant message. The
 correction carries validation diagnostics as JSON data bounded to 64 KiB; the
 full diagnostics remain in the validation result.
 

docs/internal/sources.md

@@ -21,6 +21,12 @@ paths, content opening, and root containment. Each lookup remains a
 point-in-time scan: definitions and catalogs are not cached, and file-backed
 message content is opened only for the exact selected candidate.
 
+Message roles are normalized through the shared domain owner by trimming
+Unicode whitespace and lowercasing. Only `developer`, `system`, `user`, and
+`assistant` are published; invalid roles remain selected prompt-definition
+failures rather than becoming request errors. Cache-control metadata uses the
+same shared domain normalization and defensive-copy rule.
+
 Operating-system sources enforce containment against canonical roots and
 targets so symlinks cannot escape. Injected `fs.FS` sources enforce containment
 in their clean relative path namespace. A single-file source uses the selected
@@ -51,6 +57,19 @@ and permits inherited target fields only when a base is named. File-backed
 bounded JSON-value owner before a profile is published. OpenAI-compatible
 reserved-field policy remains with the model-client and backend-registry owners.
 
+`LoadFSRepository` is the eager immutable loading boundary for internal
+catalog consumers. It discovers and strictly validates every raw profile once,
+preserves safe source metadata including explicitly present YAML fields, and
+publishes independently copied values from memory. It does not resolve profile
+inheritance. Configured consumer sources continue to use the lazy point lookup
+repositories described above.
+
+`internal/catalog` validates the imported immutable OpenRouter and
+Rakestrawhome asset modules as one private adapter boundary. It enforces their
+manifest, layout, profile ownership, inheritance, and secret-safety rules
+before returning raw catalog sources. Root assembly uses those validated
+catalogs as the maintained lowest-precedence profile source.
+
 The overlay repository consults the next repository only when the
 higher-precedence repository reports that a profile is absent. A reliably
 selected malformed profile stops fallback, while an unrelated malformed file
@@ -61,7 +80,7 @@ profile inspection.
 
 The root engine assembles one raw composite catalog in precedence order:
 in-memory profiles, one ordinary configured source, an application fallback
-source, then the embedded built-in catalog. An explicit file or `fs.FS` profile
+source, then the maintained external catalog. An explicit file or `fs.FS` profile
 source replaces `Config.ProfileDir` within the ordinary configured-source
 category. One outer resolving repository wraps that complete raw catalog, so
 each base lookup observes the same precedence and shadowing rules.
@@ -79,15 +98,11 @@ or schema sources. It does not retain that lookup for a later execution.
 Prepared execution instead freezes the fully resolved target; a later ordinary
 operation performs a fresh traversal.
 
-`internal/profile/builtin` embeds the maintained built-in profile catalog.
+The maintained external catalog provides every built-in profile and its
-Every embedded profile selects a maintained built-in backend and inherits that
+matching backend definition. Profile loading and overlay behavior are owned by the
-backend's endpoint and credential environment-variable name from the built-in
-backend registry rather than repeating those values. Profile loading and
-overlay behavior are owned by the
 [profile repository tests](../../internal/profile/repository_test.go), while
-catalog completeness, the backend-selection invariant, and duplicate IDs are
+catalog completeness, backend selection, and duplicate IDs are owned by the
-owned by the
+[catalog adapter tests](../../internal/catalog/catalog_test.go).
-[built-in repository tests](../../internal/profile/builtin/repository_test.go).
 
 ## Ordinary Artifacts
 
@@ -122,8 +137,9 @@ Session and message parsing and execution remain synchronous. The renderer
 checks cancellation before and after each parse and execution boundary,
 between artifact conversion chunks, around each message, and before publishing
 the complete prompt. It cannot interrupt template work already in progress and
-never publishes a partial prompt after observing cancellation. It carries
+never publishes a partial prompt after observing cancellation. It validates and
-message roles, session IDs, and cache control into the rendered prompt. The
+canonicalizes message roles before carrying roles, session IDs, and cache
+control into the rendered prompt. The
 [renderer tests](../../internal/prompt/renderer_test.go) own rendering behavior.
 
 ## Schemas And Output Validation

docs/policy/architecture.md

@@ -22,21 +22,21 @@ The implemented internal components consist of:
 - `internal/domain`, which owns framework data values and source-neutral
   invariants shared by later internal components;
 - `internal/backend`, which owns validated immutable OpenAI-compatible backend
-  definitions and the maintained built-in definitions;
+  definitions;
 - `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/catalog`, which validates immutable external maintained-catalog
+  assets before they are eligible for engine assembly;
 - `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;
@@ -54,13 +54,15 @@ 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.
+registry contains validated maintained definitions plus 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
+model, file catalog, and YAML decoder. The catalog adapter validates imported
-an embedded `fs.FS` to the profile package. Artifact reading uses the domain
+immutable maintained data before root assembly supplies it 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

docs/release.md

@@ -146,6 +146,13 @@ 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.
 
+When a release introduces or updates the maintained external catalogs, state
+that it selects two independently versioned data dependencies, preserves the
+public API and configuration, requires no consumer migration, and guarantees
+only the catalog versions selected and tested by that Promptkit release. Link
+to the current architecture and format documentation rather than duplicating
+catalog contracts in the release note.
+
 ## Create And Inspect The Tag
 
 Run the candidate guard again immediately before tag creation. This ensures

docs/releases/v0.9.0.md

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

engine.go

@@ -11,14 +11,16 @@ import (
 	"strings"
 	"time"
 
+	openrouter "gitea.maximumdirect.net/eric/promptkit-backend-openrouter"
+	rakestrawhome "gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome"
 	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/catalog"
 	"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"
@@ -100,7 +102,7 @@ type Config struct {
 	// prompt source.
 	PromptDir string
 	// ProfileDir is an optional ordinary configured source whose profiles take
-	// precedence over application fallback and embedded built-in profiles. An
+	// precedence over application fallback and maintained catalog profiles. An
 	// empty value selects the lower-precedence sources unless a profile-source
 	// option supplies the ordinary source.
 	ProfileDir string
@@ -273,7 +275,7 @@ func WithProfileFile(path string) Option {
 //
 // Profile lookup checks, in order, profiles supplied by WithProfiles; the
 // ordinary configured source selected by WithProfileFile, WithProfileFS, or
-// Config.ProfileDir; this fallback source; and Promptkit's embedded built-in
+// Config.ProfileDir; this fallback source; and Promptkit's maintained catalog
 // profiles. Each source supplies a complete profile definition; profile fields
 // are not merged between sources. Only an absent profile ID proceeds to the
 // next source. A matching read, parse, duplicate, validation, or credential
@@ -390,9 +392,17 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 		promptDefs = promptdef.NewFilesystemRepository(cfg.PromptDir)
 	}
 
-	profiles := newProfileRepository(cfg.ProfileDir, options)
+	maintainedCatalogs, err := catalog.Load(
+		catalog.Source{Name: "OpenRouter", ExpectedBackendID: backend.OpenRouterID, FS: openrouter.FS(), Root: openrouter.Root},
+		catalog.Source{Name: "Rakestrawhome", ExpectedBackendID: backend.RakestrawHomeID, FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+	)
+	if err != nil {
+		return nil, fmt.Errorf("%w: failed to load maintained catalogs: %v", ErrInvalidConfig, err)
+	}
 
-	backendRegistry, err := backend.NewRegistry(options.backends)
+	profiles := newProfileRepository(cfg.ProfileDir, options, maintainedCatalogs.Profiles)
+
+	backendRegistry, err := backend.NewRegistry(maintainedCatalogs.Backends, options.backends)
 	if err != nil {
 		return nil, fmt.Errorf("%w: failed to construct backend registry: %v", ErrInvalidConfig, err)
 	}
@@ -444,8 +454,8 @@ func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
 	}, nil
 }
 
-func newProfileRepository(profileDir string, options engineOptions) profile.Repository {
+func newProfileRepository(profileDir string, options engineOptions, maintained profile.Repository) profile.Repository {
-	repository := builtin.NewRepository()
+	repository := maintained
 
 	if options.fallbackProfileSource {
 		repository = profile.NewOverlayRepository(options.fallbackProfiles, repository)
@@ -571,6 +581,8 @@ func (e *Engine) InspectProfile(ctx context.Context, profileID string) (*Profile
 }
 
 // Prepare resolves and renders a prompt request without calling an LLM.
+// It appends any validated RunRequest.AppendedMessages after rendered
+// definition messages in the returned caller-owned snapshot.
 //
 // Prepare selects the prompt and profile, resolves any selected backend and
 // effective execution settings, resolves the output contract, loads and hashes
@@ -605,7 +617,8 @@ func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, err
 }
 
 // PrepareExecution completely prepares a prompt request without calling the
-// configured LLMClient or reserving backend admission capacity.
+// configured LLMClient or reserving backend admission capacity. Validated
+// RunRequest.AppendedMessages are included in the frozen effective messages.
 //
 // The returned opaque handle is bound to this Engine and permits one
 // [Engine.RunPrepared] invocation. Preparation freezes the selected sources,
@@ -637,7 +650,8 @@ func (e *Engine) PrepareExecution(ctx context.Context, req RunRequest) (*Prepare
 }
 
 // Run prepares a request, invokes the configured LLMClient, and validates the
-// generated output.
+// generated output. Each call resolves current sources and composes a fresh,
+// stateless effective prompt with any validated RunRequest.AppendedMessages.
 //
 // A content-validation failure is a successful run whose
 // RunResult.Validation has Status ValidationFailed. When its output contract

engine_test.go

@@ -220,6 +220,14 @@ func TestRunRequestFormattingRedactsDirectAPIKey(t *testing.T) {
 			"transcript": promptkit.InlineWithURI(inputURI, inputBody),
 		},
 		Vars: map[string]string{"audience": variableValue},
+		AppendedMessages: []promptkit.RenderedMessage{{
+			Role:    "private-role-sentinel",
+			Content: "private-appended-content-sentinel",
+			CacheControl: &promptkit.CacheControl{
+				Type: "private-cache-sentinel",
+				TTL:  "private-ttl-sentinel",
+			},
+		}},
 	}
 
 	for _, formatted := range []string{
@@ -229,7 +237,7 @@ func TestRunRequestFormattingRedactsDirectAPIKey(t *testing.T) {
 		fmt.Sprintf("%+v", req),
 		fmt.Sprintf("%#v", req),
 	} {
-		for _, privateValue := range []string{secret, inputURI, inputBody, variableValue} {
+		for _, privateValue := range []string{secret, inputURI, inputBody, variableValue, "private-role-sentinel", "private-appended-content-sentinel", "private-cache-sentinel", "private-ttl-sentinel"} {
 			if strings.Contains(formatted, privateValue) {
 				t.Fatalf("formatted RunRequest leaked private value %q: %s", privateValue, formatted)
 			}
@@ -240,6 +248,7 @@ func TestRunRequestFormattingRedactsDirectAPIKey(t *testing.T) {
 			"APIKeySet:true",
 			"Inputs:1",
 			"Vars:1",
+			"AppendedMessages:1",
 		} {
 			if !strings.Contains(formatted, summary) {
 				t.Fatalf("formatted RunRequest omitted structural summary %q: %s", summary, formatted)
@@ -1536,31 +1545,21 @@ func TestPrepareUsesBuiltInProfileWithoutProfileDir(t *testing.T) {
 		name      string
 		profileID string
 		backendID string
-		endpoint  string
-		apiKeyEnv string
-		model     string
 	}{
 		{
 			name:      "OpenRouter",
 			profileID: "mistral-small-3",
 			backendID: promptkit.BackendOpenRouter,
-			endpoint:  "https://openrouter.ai/api/v1",
-			apiKeyEnv: "OPENROUTER_API_KEY",
-			model:     "mistralai/mistral-small-3.2-24b-instruct",
 		},
 		{
 			name:      "Rakestrawhome",
 			profileID: "rakestrawhome-gemma-4-31b",
 			backendID: promptkit.BackendRakestrawHome,
-			endpoint:  "https://inference.ai.rakestrawhome.com/v1",
-			apiKeyEnv: "RAKESTRAWHOME_INFERENCE_API_KEY",
-			model:     "google/gemma-4-31b-it",
 		},
 	}
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			t.Setenv(tc.apiKeyEnv, "test-key")
 			engine, err := promptkit.NewEngine(promptkit.Config{
 				PromptDir: frameworkPromptDir,
 				SchemaDir: frameworkSchemaDir,
@@ -1572,6 +1571,7 @@ func TestPrepareUsesBuiltInProfileWithoutProfileDir(t *testing.T) {
 			prepared, err := engine.Prepare(context.Background(), promptkit.RunRequest{
 				PromptID:  frameworkMarkdownSummaryPromptID,
 				ProfileID: tc.profileID,
+				APIKey:    "test-key",
 				Inputs: map[string]promptkit.ArtifactRef{
 					"transcript": promptkit.Inline("Rin opens the gate."),
 					"glossary":   promptkit.Inline("gate: A guarded passage."),
@@ -1582,10 +1582,7 @@ func TestPrepareUsesBuiltInProfileWithoutProfileDir(t *testing.T) {
 			}
 			if prepared.SelectedProfileID != tc.profileID ||
 				prepared.SelectedBackendID != tc.backendID ||
-				prepared.EffectiveModelParams.BackendID != tc.backendID ||
+				prepared.EffectiveModelParams.BackendID != tc.backendID {
-				prepared.EffectiveModelParams.Endpoint != tc.endpoint ||
-				prepared.EffectiveModelParams.APIKeyEnv != tc.apiKeyEnv ||
-				prepared.EffectiveModelParams.Model != tc.model {
 				t.Fatalf("unexpected built-in preparation: %#v", prepared)
 			}
 		})

formatting.go

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

generation_error_contract_test.go

@@ -39,6 +39,23 @@ func TestBuiltInGenerationError(t *testing.T) {
 	assertGenerationError(t, err, http.StatusServiceUnavailable, "", "", "")
 }
 
+func TestBuiltInGenerationErrorWithAppendedMessages(t *testing.T) {
+	const messageMarker = "combined-request-provider-marker"
+	engine := newBuiltInGenerationErrorEngine(t, http.StatusUnprocessableEntity,
+		`{"error":{"message":"`+messageMarker+`"}}`)
+	request := generationErrorRunRequest()
+	request.AppendedMessages = []promptkit.RenderedMessage{
+		{Role: promptkit.RoleAssistant, Content: "previous response"},
+		{Role: promptkit.RoleUser, Content: "consumer correction"},
+	}
+
+	result, err := engine.Run(context.Background(), request)
+	if result != nil {
+		t.Fatalf("Run result = %#v, want nil", result)
+	}
+	assertGenerationError(t, err, http.StatusUnprocessableEntity, "", "", messageMarker)
+}
+
 func TestBuiltInRepairGenerationError(t *testing.T) {
 	const (
 		codeMarker    = "repair-code-marker"

go.mod

@@ -3,6 +3,8 @@ module gitea.maximumdirect.net/eric/promptkit
 go 1.25.5
 
 require (
+	gitea.maximumdirect.net/eric/promptkit-backend-openrouter v1.0.0
+	gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome v1.0.0
 	github.com/santhosh-tekuri/jsonschema/v6 v6.0.2
 	gopkg.in/yaml.v3 v3.0.1
 )

go.sum

@@ -1,3 +1,7 @@
+gitea.maximumdirect.net/eric/promptkit-backend-openrouter v1.0.0 h1:lc062euk2qseO//D762i3JaFyulDNML3eQQX7DkYTho=
+gitea.maximumdirect.net/eric/promptkit-backend-openrouter v1.0.0/go.mod h1:AIa7kAu2mfrRQgcspe4L+DW51WqgnALQT60lqkEywJI=
+gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome v1.0.0 h1:j9YY7wsTVjzke2kHH4YAzpU0oUpM+x+nXwl1IeS+2eg=
+gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome v1.0.0/go.mod h1:4RNS+LILDg4JbS4Ts9Lwy1C92wauXJIbeQaalps4Koo=
 github.com/dlclark/regexp2 v1.11.0 h1:G/nrcoOa7ZXlpoa/91N3X7mM3r8eIlMBBJZvsz/mxKI=
 github.com/dlclark/regexp2 v1.11.0/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/santhosh-tekuri/jsonschema/v6 v6.0.2 h1:KRzFb2m7YtdldCEkzs6KqmJw4nqEVZGK7IN2kJkjTuQ=

internal/backend/registry.go

@@ -22,15 +22,6 @@ const (
 	// backend.
 	RakestrawHomeID = "rakestrawhome"
 
-	openRouterEndpoint  = "https://openrouter.ai/api/v1"
-	openRouterAPIKeyEnv = "OPENROUTER_API_KEY"
-
-	openRouterConcurrencyLimit = 16
-
-	rakestrawHomeEndpoint  = "https://inference.ai.rakestrawhome.com/v1"
-	rakestrawHomeAPIKeyEnv = "RAKESTRAWHOME_INFERENCE_API_KEY"
-
-	rakestrawHomeConcurrencyLimit = 4
 	defaultQueueCapacity = 1024
 )
 
@@ -44,16 +35,15 @@ type Registry struct {
 	backends map[string]domain.Backend
 }
 
-// NewRegistry constructs a registry containing the built-in definitions
+// NewRegistry constructs a registry containing maintained definitions followed
-// followed by the supplied additions. Every ID must be unique.
+// by consumer additions. Every ID must be unique across both groups.
-func NewRegistry(additions []domain.Backend) (*Registry, error) {
+func NewRegistry(maintained, additions []domain.Backend) (*Registry, error) {
-	builtIns := builtInBackends()
 	registry := &Registry{
-		backends: make(map[string]domain.Backend, len(builtIns)+len(additions)),
+		backends: make(map[string]domain.Backend, len(maintained)+len(additions)),
 	}
 
-	definitions := make([]domain.Backend, 0, len(builtIns)+len(additions))
+	definitions := make([]domain.Backend, 0, len(maintained)+len(additions))
-	definitions = append(definitions, builtIns...)
+	definitions = append(definitions, maintained...)
 	definitions = append(definitions, additions...)
 
 	for _, definition := range definitions {
@@ -65,7 +55,7 @@ func NewRegistry(additions []domain.Backend) (*Registry, error) {
 			return nil, fmt.Errorf("backend ID %q is already registered", definition.ID)
 		}
 
-		normalized, err := normalizeBackend(definition)
+		normalized, err := NormalizeDefinition(definition)
 		if err != nil {
 			return nil, err
 		}
@@ -75,23 +65,6 @@ func NewRegistry(additions []domain.Backend) (*Registry, error) {
 	return registry, nil
 }
 
-func builtInBackends() []domain.Backend {
-	return []domain.Backend{
-		{
-			ID:               OpenRouterID,
-			Endpoint:         openRouterEndpoint,
-			APIKeyEnv:        openRouterAPIKeyEnv,
-			ConcurrencyLimit: openRouterConcurrencyLimit,
-		},
-		{
-			ID:               RakestrawHomeID,
-			Endpoint:         rakestrawHomeEndpoint,
-			APIKeyEnv:        rakestrawHomeAPIKeyEnv,
-			ConcurrencyLimit: rakestrawHomeConcurrencyLimit,
-		},
-	}
-}
-
 // GetBackend returns a defensive copy of the backend registered with id.
 func (r *Registry) GetBackend(id string) (domain.Backend, error) {
 	if r == nil {
@@ -128,7 +101,8 @@ func (r *Registry) CapacityPolicies() map[string]domain.BackendCapacityPolicy {
 	return policies
 }
 
-func normalizeBackend(definition domain.Backend) (domain.Backend, error) {
+// NormalizeDefinition validates and defensively copies one backend definition.
+func NormalizeDefinition(definition domain.Backend) (domain.Backend, error) {
 	endpoint, err := domain.NormalizeOpenAICompatibleBaseEndpoint(definition.Endpoint)
 	if err != nil {
 		return domain.Backend{}, fmt.Errorf("backend %q endpoint: %w", definition.ID, err)

internal/backend/registry_test.go

@@ -2,6 +2,7 @@ package backend_test
 
 import (
 	"errors"
+	"reflect"
 	"strings"
 	"testing"
 
@@ -11,48 +12,24 @@ import (
 
 const validEndpoint = "https://backend.example/v1"
 
-func TestRegistryIncludesExactBuiltInDefinitions(t *testing.T) {
+func TestRegistryIncludesMaintainedDefinitions(t *testing.T) {
-	registry, err := backend.NewRegistry(nil)
+	maintained := []domain.Backend{
+		{ID: backend.OpenRouterID, Endpoint: validEndpoint, ConcurrencyLimit: 2, QueueCapacity: 3, QueueCapacitySet: true},
+		{ID: backend.RakestrawHomeID, Endpoint: "https://second.example/v1", ConcurrencyLimit: 4, QueueCapacity: 5, QueueCapacitySet: true},
+	}
+	registry, err := backend.NewRegistry(maintained, nil)
 	if err != nil {
 		t.Fatalf("construct registry: %v", err)
 	}
 
-	tests := []struct {
+	for _, expected := range maintained {
-		name       string
+		t.Run(expected.ID, func(t *testing.T) {
-		id         string
+			definition, err := registry.GetBackend(expected.ID)
-		endpoint   string
-		apiKeyEnv  string
-		concurrent int
-	}{
-		{
-			name:       "OpenRouter",
-			id:         backend.OpenRouterID,
-			endpoint:   "https://openrouter.ai/api/v1",
-			apiKeyEnv:  "OPENROUTER_API_KEY",
-			concurrent: 16,
-		},
-		{
-			name:       "Rakestrawhome",
-			id:         backend.RakestrawHomeID,
-			endpoint:   "https://inference.ai.rakestrawhome.com/v1",
-			apiKeyEnv:  "RAKESTRAWHOME_INFERENCE_API_KEY",
-			concurrent: 4,
-		},
-	}
-	for _, tc := range tests {
-		t.Run(tc.name, func(t *testing.T) {
-			definition, err := registry.GetBackend(tc.id)
 			if err != nil {
-				t.Fatalf("look up built-in: %v", err)
+				t.Fatalf("look up maintained definition: %v", err)
 			}
-			if definition.ID != tc.id ||
+			if !reflect.DeepEqual(definition, expected) {
-				definition.Endpoint != tc.endpoint ||
+				t.Fatalf("unexpected maintained definition: %#v", definition)
-				definition.APIKeyEnv != tc.apiKeyEnv ||
-				definition.ConcurrencyLimit != tc.concurrent ||
-				definition.QueueCapacity != 1024 ||
-				!definition.QueueCapacitySet ||
-				definition.ExtraParams != nil {
-				t.Fatalf("unexpected built-in definition: %#v", definition)
 			}
 		})
 	}
@@ -60,12 +37,12 @@ func TestRegistryIncludesExactBuiltInDefinitions(t *testing.T) {
 	policies := registry.CapacityPolicies()
 	if len(policies) != 2 ||
 		policies[backend.OpenRouterID] != (domain.BackendCapacityPolicy{
-			ConcurrencyLimit: 16,
+			ConcurrencyLimit: 2,
-			QueueCapacity:    1024,
+			QueueCapacity:    3,
 		}) ||
 		policies[backend.RakestrawHomeID] != (domain.BackendCapacityPolicy{
 			ConcurrencyLimit: 4,
-			QueueCapacity:    1024,
+			QueueCapacity:    5,
 		}) {
 		t.Fatalf("unexpected built-in capacity policies: %#v", policies)
 	}
@@ -77,7 +54,7 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 		"count":  int64(7),
 		"nested": nested,
 	}
-	registry, err := backend.NewRegistry([]domain.Backend{
+	registry, err := backend.NewRegistry(nil, []domain.Backend{
 		{
 			ID:               " custom ",
 			Endpoint:         " https://custom.example/openai/v1 ",
@@ -140,12 +117,11 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 	}
 
 	policies := registry.CapacityPolicies()
-	if len(policies) != 3 {
+	if len(policies) != 1 {
 		t.Fatalf("unexpected capacity policy count: %#v", policies)
 	}
 	policies["custom"] = domain.BackendCapacityPolicy{}
-	delete(policies, backend.OpenRouterID)
+	delete(policies, "custom")
-	delete(policies, backend.RakestrawHomeID)
 	againPolicies := registry.CapacityPolicies()
 	if againPolicies["custom"] != (domain.BackendCapacityPolicy{
 		ConcurrencyLimit: 3,
@@ -153,12 +129,6 @@ func TestRegistryNormalizesUniqueAdditionsAndIsolatesMutations(t *testing.T) {
 	}) {
 		t.Fatalf("capacity policy map mutated registry state: %#v", againPolicies)
 	}
-	if _, ok := againPolicies[backend.OpenRouterID]; !ok {
-		t.Fatalf("OpenRouter capacity policy deletion mutated registry state: %#v", againPolicies)
-	}
-	if _, ok := againPolicies[backend.RakestrawHomeID]; !ok {
-		t.Fatalf("Rakestrawhome capacity policy deletion mutated registry state: %#v", againPolicies)
-	}
 }
 
 func TestNewRegistryNormalizesCapacityPolicy(t *testing.T) {
@@ -234,7 +204,7 @@ func TestNewRegistryNormalizesCapacityPolicy(t *testing.T) {
 		t.Run(tc.name, func(t *testing.T) {
 			tc.definition.ID = "custom"
 			tc.definition.Endpoint = validEndpoint
-			registry, err := backend.NewRegistry([]domain.Backend{tc.definition})
+			registry, err := backend.NewRegistry(nil, []domain.Backend{tc.definition})
 			if tc.wantError {
 				if err == nil {
 					t.Fatal("expected invalid capacity policy error")
@@ -298,7 +268,7 @@ func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry(tc.additions)
+			_, err := backend.NewRegistry(nil, tc.additions)
 			if err == nil {
 				t.Fatal("expected duplicate ID error")
 			}
@@ -312,7 +282,7 @@ func TestNewRegistryRejectsDuplicateIDs(t *testing.T) {
 func TestNewRegistryValidatesIDs(t *testing.T) {
 	for _, id := range []string{"", " \t\n "} {
 		t.Run(id, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:       id,
 				Endpoint: validEndpoint,
 			}})
@@ -341,7 +311,7 @@ func TestNewRegistryValidatesEndpoints(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:       "custom",
 				Endpoint: tc.endpoint,
 			}})
@@ -355,7 +325,7 @@ func TestNewRegistryValidatesEndpoints(t *testing.T) {
 func TestNewRegistryValidatesEnvironmentVariableNames(t *testing.T) {
 	for _, name := range []string{"1API_KEY", "API-KEY", "API KEY", "ÅPI_KEY"} {
 		t.Run(name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:        "custom",
 				Endpoint:  validEndpoint,
 				APIKeyEnv: name,
@@ -378,7 +348,7 @@ func TestNewRegistryRejectsInvalidAndReservedExtraParameters(t *testing.T) {
 
 	for _, tc := range tests {
 		t.Run(tc.name, func(t *testing.T) {
-			_, err := backend.NewRegistry([]domain.Backend{{
+			_, err := backend.NewRegistry(nil, []domain.Backend{{
 				ID:          "custom",
 				Endpoint:    validEndpoint,
 				ExtraParams: tc.extraParams,
@@ -391,7 +361,7 @@ func TestNewRegistryRejectsInvalidAndReservedExtraParameters(t *testing.T) {
 }
 
 func TestRegistryLookupReportsNotFound(t *testing.T) {
-	registry, err := backend.NewRegistry(nil)
+	registry, err := backend.NewRegistry(nil, nil)
 	if err != nil {
 		t.Fatalf("construct registry: %v", err)
 	}

internal/catalog/catalog.go

@@ -0,0 +1,246 @@
+// Package catalog validates immutable maintained backend catalog assets.
+package catalog
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"io"
+	"io/fs"
+	"path"
+	"sort"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/profile"
+)
+
+// Source identifies one immutable backend catalog asset tree.
+type Source struct {
+	Name              string
+	ExpectedBackendID string
+	FS                fs.FS
+	Root              string
+}
+
+// Set is the validated maintained backend and raw profile catalog.
+type Set struct {
+	Backends   []domain.Backend
+	Profiles   profile.Repository
+	profileIDs []string
+}
+
+// Load validates and combines immutable catalog sources in source order.
+func Load(sources ...Source) (Set, error) {
+	if len(sources) == 0 {
+		return Set{}, errors.New("at least one catalog source is required")
+	}
+	loaded := Set{Backends: make([]domain.Backend, 0, len(sources))}
+	names := map[string]bool{}
+	backendIDs := map[string]bool{}
+	profileIDs := map[string]bool{}
+	for _, source := range sources {
+		if err := validateSource(source, names); err != nil {
+			return Set{}, err
+		}
+		names[source.Name] = true
+		definition, err := loadBackend(source)
+		if err != nil {
+			return Set{}, err
+		}
+		if backendIDs[definition.ID] {
+			return Set{}, fmt.Errorf("catalog %s: backend ID duplicates an earlier catalog", source.Name)
+		}
+		backendIDs[definition.ID] = true
+		repository, metadata, err := profile.LoadFSRepository(context.Background(), source.FS, path.Join(source.Root, "profiles"))
+		if err != nil {
+			return Set{}, fmt.Errorf("catalog %s: %w", source.Name, err)
+		}
+		if len(metadata) == 0 {
+			return Set{}, fmt.Errorf("catalog %s: profiles must not be empty", source.Name)
+		}
+		resolvingRepository := profile.NewResolvingRepository(repository)
+		for _, entry := range metadata {
+			if profileIDs[entry.ID] {
+				return Set{}, fmt.Errorf("catalog %s: %s duplicates an earlier profile ID", source.Name, entry.Path)
+			}
+			if containsField(entry.ExplicitFields, "endpoint") || containsField(entry.ExplicitFields, "api_key_env") {
+				return Set{}, fmt.Errorf("catalog %s: %s contains connection metadata", source.Name, entry.Path)
+			}
+			value, err := repository.GetProfile(context.Background(), entry.ID)
+			if err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s: %w", source.Name, entry.Path, err)
+			}
+			if err := rejectSecretKeys(value.ExtraParams); err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s: prohibited extra parameter key", source.Name, entry.Path)
+			}
+			resolved, err := resolvingRepository.GetProfile(context.Background(), entry.ID)
+			if err != nil {
+				return Set{}, fmt.Errorf("catalog %s: %s has invalid profile inheritance", source.Name, entry.Path)
+			}
+			if resolved.BackendID != definition.ID {
+				return Set{}, fmt.Errorf("catalog %s: %s selects a different backend", source.Name, entry.Path)
+			}
+			profileIDs[entry.ID] = true
+			loaded.profileIDs = append(loaded.profileIDs, entry.ID)
+		}
+		if loaded.Profiles == nil {
+			loaded.Profiles = repository
+		} else {
+			loaded.Profiles = profile.NewOverlayRepository(loaded.Profiles, repository)
+		}
+		loaded.Backends = append(loaded.Backends, definition)
+	}
+	sort.Strings(loaded.profileIDs)
+	return loaded, nil
+}
+
+func validateSource(source Source, names map[string]bool) error {
+	if strings.TrimSpace(source.Name) == "" || source.Name != strings.TrimSpace(source.Name) {
+		return errors.New("catalog source name must not be blank")
+	}
+	if names[source.Name] {
+		return fmt.Errorf("duplicate catalog source name %q", source.Name)
+	}
+	if source.FS == nil {
+		return fmt.Errorf("catalog %s: filesystem is nil", source.Name)
+	}
+	if source.Root == "." || source.Root == "" || !fs.ValidPath(source.Root) {
+		return fmt.Errorf("catalog %s: asset root is invalid", source.Name)
+	}
+	if strings.TrimSpace(source.ExpectedBackendID) == "" {
+		return fmt.Errorf("catalog %s: expected backend ID is blank", source.Name)
+	}
+	return validateLayout(source)
+}
+
+func validateLayout(source Source) error {
+	manifestPath := path.Join(source.Root, "backend.json")
+	profilesRoot := path.Join(source.Root, "profiles")
+	return fs.WalkDir(source.FS, source.Root, func(assetPath string, entry fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if assetPath == source.Root {
+			if !entry.IsDir() {
+				return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+			}
+			return nil
+		}
+		if entry.IsDir() {
+			if assetPath == profilesRoot || strings.HasPrefix(assetPath, profilesRoot+"/") {
+				return nil
+			}
+			return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+		}
+		if assetPath == manifestPath && entry.Type().IsRegular() {
+			return nil
+		}
+		if strings.HasPrefix(assetPath, profilesRoot+"/") && entry.Type().IsRegular() && strings.HasSuffix(assetPath, ".yml") {
+			return nil
+		}
+		return fmt.Errorf("catalog %s: invalid asset path %s", source.Name, assetPath)
+	})
+}
+
+func loadBackend(source Source) (domain.Backend, error) {
+	data, err := fs.ReadFile(source.FS, path.Join(source.Root, "backend.json"))
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: %w", source.Name, err)
+	}
+	type manifest struct {
+		SchemaVersion    *int            `json:"schema_version"`
+		ID               *string         `json:"id"`
+		Endpoint         *string         `json:"endpoint"`
+		APIKeyEnv        *string         `json:"api_key_env"`
+		ConcurrencyLimit *int            `json:"concurrency_limit"`
+		QueueCapacity    *int            `json:"queue_capacity"`
+		ExtraParams      json.RawMessage `json:"extra_params"`
+	}
+	decoder := json.NewDecoder(bytes.NewReader(data))
+	decoder.DisallowUnknownFields()
+	var value manifest
+	if err := decoder.Decode(&value); err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid manifest", source.Name)
+	}
+	var trailing any
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid manifest", source.Name)
+	}
+	if value.SchemaVersion == nil || *value.SchemaVersion != 1 || value.ID == nil || value.Endpoint == nil || value.APIKeyEnv == nil || value.ConcurrencyLimit == nil || value.QueueCapacity == nil || value.ExtraParams == nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: required field is missing or unsupported", source.Name)
+	}
+	if *value.ID != source.ExpectedBackendID {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend ID does not match expected ID", source.Name)
+	}
+	if strings.TrimSpace(*value.APIKeyEnv) == "" {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: api key environment variable must not be blank", source.Name)
+	}
+	extraParams, err := decodeExtraParams(value.ExtraParams)
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid extra parameters", source.Name)
+	}
+	if err := rejectSecretKeys(extraParams); err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: prohibited extra parameter key", source.Name)
+	}
+	normalized, err := backend.NormalizeDefinition(domain.Backend{ID: *value.ID, Endpoint: *value.Endpoint, APIKeyEnv: *value.APIKeyEnv, ExtraParams: extraParams, ConcurrencyLimit: *value.ConcurrencyLimit, QueueCapacity: *value.QueueCapacity, QueueCapacitySet: true})
+	if err != nil {
+		return domain.Backend{}, fmt.Errorf("catalog %s: backend.json: invalid backend definition", source.Name)
+	}
+	return normalized, nil
+}
+
+func decodeExtraParams(data []byte) (map[string]any, error) {
+	if bytes.Equal(bytes.TrimSpace(data), []byte("null")) {
+		return nil, nil
+	}
+	decoder := json.NewDecoder(bytes.NewReader(data))
+	decoder.UseNumber()
+	var value map[string]any
+	if err := decoder.Decode(&value); err != nil || value == nil {
+		return nil, errors.New("extra parameters must be an object or null")
+	}
+	var trailing any
+	if err := decoder.Decode(&trailing); !errors.Is(err, io.EOF) {
+		return nil, errors.New("extra parameters must contain exactly one JSON value")
+	}
+	return value, nil
+}
+
+func containsField(fields []string, target string) bool {
+	for _, field := range fields {
+		if field == target {
+			return true
+		}
+	}
+	return false
+}
+
+func rejectSecretKeys(value map[string]any) error {
+	return rejectSecretValue(value)
+}
+
+func rejectSecretValue(value any) error {
+	switch value := value.(type) {
+	case map[string]any:
+		for key, child := range value {
+			switch strings.ToLower(key) {
+			case "api_key", "apikey", "authorization", "credential", "credentials", "password", "secret", "token", "access_token":
+				return errors.New("prohibited key")
+			}
+			if err := rejectSecretValue(child); err != nil {
+				return err
+			}
+		}
+	case []any:
+		for _, child := range value {
+			if err := rejectSecretValue(child); err != nil {
+				return err
+			}
+		}
+	}
+	return nil
+}

internal/catalog/catalog_test.go

@@ -0,0 +1,364 @@
+package catalog
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"fmt"
+	"io/fs"
+	"os"
+	"path/filepath"
+	"reflect"
+	"sort"
+	"strings"
+	"testing"
+	"testing/fstest"
+
+	openrouter "gitea.maximumdirect.net/eric/promptkit-backend-openrouter"
+	rakestrawhome "gitea.maximumdirect.net/eric/promptkit-backend-rakestrawhome"
+)
+
+func TestLoadPublishedCatalogsMatchCompatibilityFixture(t *testing.T) {
+	loaded, err := Load(
+		Source{Name: "OpenRouter", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root},
+		Source{Name: "Rakestrawhome", ExpectedBackendID: "rakestrawhome", FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+	)
+	if err != nil {
+		t.Fatalf("load published catalogs: %v", err)
+	}
+	expected := loadCompatibilityFixture(t)
+	expectedIDs := fixtureProfileIDs(t, expected)
+	if !reflect.DeepEqual(loaded.profileIDs, expectedIDs) {
+		t.Fatalf("published profile IDs differ from compatibility fixture: got %q, want %q", loaded.profileIDs, expectedIDs)
+	}
+	actual := catalogValue(t, loaded)
+	actualJSON, err := json.Marshal(actual)
+	if err != nil {
+		t.Fatalf("encode loaded catalogs: %v", err)
+	}
+	expectedJSON, err := json.Marshal(expected)
+	if err != nil {
+		t.Fatalf("encode compatibility fixture: %v", err)
+	}
+	if !bytes.Equal(actualJSON, expectedJSON) {
+		t.Fatalf("published catalogs differ from compatibility fixture: got %#v, want %#v", actual, expected)
+	}
+}
+
+func TestLoadRejectsInvalidSources(t *testing.T) {
+	for name, sources := range map[string][]Source{
+		"none":       nil,
+		"blank name": {{Name: " ", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root}},
+		"duplicate name": {
+			{Name: "same", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: openrouter.Root},
+			{Name: "same", ExpectedBackendID: "rakestrawhome", FS: rakestrawhome.FS(), Root: rakestrawhome.Root},
+		},
+		"nil filesystem":         {{Name: "missing", ExpectedBackendID: "openrouter", Root: openrouter.Root}},
+		"invalid root":           {{Name: "invalid-root", ExpectedBackendID: "openrouter", FS: openrouter.FS(), Root: "."}},
+		"blank expected backend": {{Name: "blank-backend", ExpectedBackendID: " ", FS: openrouter.FS(), Root: openrouter.Root}},
+	} {
+		t.Run(name, func(t *testing.T) {
+			if _, err := Load(sources...); err == nil {
+				t.Fatal("expected error")
+			}
+		})
+	}
+}
+
+func TestLoadRejectsInvalidLayouts(t *testing.T) {
+	tests := map[string]func(fstest.MapFS){
+		"unexpected file": func(fsys fstest.MapFS) {
+			fsys["catalog/notes.txt"] = &fstest.MapFile{Data: []byte("unexpected")}
+		},
+		"unexpected directory": func(fsys fstest.MapFS) {
+			fsys["catalog/unexpected"] = &fstest.MapFile{Mode: fs.ModeDir}
+		},
+		"nonregular manifest": func(fsys fstest.MapFS) {
+			fsys["catalog/backend.json"].Mode = fs.ModeSymlink
+		},
+		"wrong profile extension": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/extra.yaml"] = &fstest.MapFile{Data: []byte(validProfile("extra", "one"))}
+		},
+		"nonregular profile": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Mode = fs.ModeSymlink
+		},
+	}
+	for name, mutate := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			mutate(fsys)
+			if _, err := Load(testSource("one", fsys)); err == nil {
+				t.Fatal("expected invalid layout error")
+			}
+		})
+	}
+}
+
+func TestLoadRejectsInvalidManifests(t *testing.T) {
+	tests := map[string]string{
+		"malformed":             `{`,
+		"trailing value":        validManifest("one", "TEST_API_KEY", "null") + `{}`,
+		"missing fields":        `{"schema_version":1,"id":"one"}`,
+		"unsupported version":   strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"schema_version":1`, `"schema_version":2`, 1),
+		"unknown field":         strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"extra_params":null`, `"extra_params":null,"unknown":true`, 1),
+		"blank API key env":     validManifest("one", "   ", "null"),
+		"invalid API key env":   validManifest("one", "LEAK-MARKER", "null"),
+		"invalid endpoint":      strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `https://one.example/v1`, `ftp://leak-marker.invalid/v1`, 1),
+		"zero concurrency":      strings.Replace(validManifest("one", "TEST_API_KEY", "null"), `"concurrency_limit":2`, `"concurrency_limit":0`, 1),
+		"non-object parameters": validManifest("one", "TEST_API_KEY", `[]`),
+		"secret parameter":      validManifest("one", "TEST_API_KEY", `{"nested":{"token":"leak-marker"}}`),
+	}
+	for name, manifest := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			fsys["catalog/backend.json"].Data = []byte(manifest)
+			_, err := Load(testSource("one", fsys))
+			if err == nil {
+				t.Fatal("expected invalid manifest error")
+			}
+			if strings.Contains(strings.ToLower(err.Error()), "leak-marker") {
+				t.Fatalf("catalog error exposed manifest content: %v", err)
+			}
+		})
+	}
+}
+
+func TestLoadPreservesManifestJSONNumbers(t *testing.T) {
+	fsys := validCatalogFS("one")
+	fsys["catalog/backend.json"].Data = []byte(validManifest(
+		"one",
+		"TEST_API_KEY",
+		`{"large":9007199254740993,"nested":[1.25]}`,
+	))
+	loaded, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("load catalog: %v", err)
+	}
+	if got := loaded.Backends[0].ExtraParams["large"]; got != json.Number("9007199254740993") {
+		t.Fatalf("large JSON integer = %#v, want preserved json.Number", got)
+	}
+	nested := loaded.Backends[0].ExtraParams["nested"].([]any)
+	if nested[0] != json.Number("1.25") {
+		t.Fatalf("nested JSON number = %#v, want preserved json.Number", nested[0])
+	}
+}
+
+func TestLoadRejectsInvalidCatalogProfiles(t *testing.T) {
+	tests := map[string]func(fstest.MapFS){
+		"empty": func(fsys fstest.MapFS) {
+			delete(fsys, "catalog/profiles/one-profile.yml")
+			fsys["catalog/profiles"] = &fstest.MapFile{Mode: fs.ModeDir}
+		},
+		"malformed": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: [")
+		},
+		"raw API key": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "api_key: leak-marker\n")
+		},
+		"endpoint field": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "endpoint: ''\n")
+		},
+		"API key environment field": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "api_key_env: ''\n")
+		},
+		"owner mismatch": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "other"))
+		},
+		"missing base": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: one-profile\nbase_profile: leak-marker\n")
+		},
+		"cyclic base": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte("id: one-profile\nbase_profile: second\n")
+			fsys["catalog/profiles/second.yml"] = &fstest.MapFile{Data: []byte("id: second\nbase_profile: one-profile\n")}
+		},
+		"secret profile parameter": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "extra_params:\n  nested:\n    password: leak-marker\n")
+		},
+		"unknown field is redacted": func(fsys fstest.MapFS) {
+			fsys["catalog/profiles/one-profile.yml"].Data = []byte(validProfile("one-profile", "one") + "leak_marker: leak-marker\n")
+		},
+	}
+	for name, mutate := range tests {
+		t.Run(name, func(t *testing.T) {
+			fsys := validCatalogFS("one")
+			mutate(fsys)
+			_, err := Load(testSource("one", fsys))
+			if err == nil {
+				t.Fatal("expected invalid profile error")
+			}
+			if strings.Contains(strings.ToLower(err.Error()), "leak-marker") {
+				t.Fatalf("catalog error exposed profile content: %v", err)
+			}
+		})
+	}
+}
+
+func TestLoadRejectsCrossCatalogConflicts(t *testing.T) {
+	t.Run("duplicate backend", func(t *testing.T) {
+		second := catalogFS("one", map[string]string{
+			"catalog/profiles/second.yml": validProfile("second", "one"),
+		}, "null")
+		_, err := Load(
+			Source{Name: "first", ExpectedBackendID: "one", FS: validCatalogFS("one"), Root: "catalog"},
+			Source{Name: "second", ExpectedBackendID: "one", FS: second, Root: "catalog"},
+		)
+		if err == nil {
+			t.Fatal("expected duplicate backend error")
+		}
+	})
+
+	t.Run("duplicate profile", func(t *testing.T) {
+		first := catalogFS("one", map[string]string{
+			"catalog/profiles/shared.yml": validProfile("shared", "one"),
+		}, "null")
+		second := catalogFS("two", map[string]string{
+			"catalog/profiles/shared.yml": validProfile("shared", "two"),
+		}, "null")
+		_, err := Load(testSource("one", first), testSource("two", second))
+		if err == nil {
+			t.Fatal("expected duplicate profile error")
+		}
+	})
+
+	t.Run("cross-catalog base", func(t *testing.T) {
+		first := catalogFS("one", map[string]string{
+			"catalog/profiles/base.yml": validProfile("base", "one"),
+		}, "null")
+		second := catalogFS("two", map[string]string{
+			"catalog/profiles/child.yml": "id: child\nbase_profile: base\n",
+		}, "null")
+		_, err := Load(testSource("one", first), testSource("two", second))
+		if err == nil {
+			t.Fatal("expected cross-catalog base error")
+		}
+	})
+}
+
+func TestLoadReturnsDefensiveCatalogValues(t *testing.T) {
+	fsys := catalogFS("one", map[string]string{
+		"catalog/profiles/one-profile.yml": validProfile("one-profile", "one") + "extra_params:\n  nested:\n    value: profile\n",
+	}, `{"nested":{"value":"backend"}}`)
+	loaded, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("load catalog: %v", err)
+	}
+	loaded.Backends[0].ExtraParams["nested"].(map[string]any)["value"] = "changed"
+	profileValue, err := loaded.Profiles.GetProfile(context.Background(), "one-profile")
+	if err != nil {
+		t.Fatalf("load profile: %v", err)
+	}
+	profileValue.ExtraParams["nested"].(map[string]any)["value"] = "changed"
+
+	again, err := Load(testSource("one", fsys))
+	if err != nil {
+		t.Fatalf("reload catalog: %v", err)
+	}
+	if got := again.Backends[0].ExtraParams["nested"].(map[string]any)["value"]; got != "backend" {
+		t.Fatalf("backend mutation escaped returned set: %#v", got)
+	}
+	againProfile, err := loaded.Profiles.GetProfile(context.Background(), "one-profile")
+	if err != nil {
+		t.Fatalf("reload profile: %v", err)
+	}
+	if got := againProfile.ExtraParams["nested"].(map[string]any)["value"]; got != "profile" {
+		t.Fatalf("profile mutation escaped returned value: %#v", got)
+	}
+}
+
+func loadCompatibilityFixture(t *testing.T) map[string]any {
+	t.Helper()
+	data, err := os.ReadFile(filepath.Join("..", "..", "testdata", "builtin-catalog-v1.json"))
+	if err != nil {
+		t.Fatalf("read fixture: %v", err)
+	}
+	var value map[string]any
+	if err := json.Unmarshal(data, &value); err != nil {
+		t.Fatalf("decode fixture: %v", err)
+	}
+	return value
+}
+
+func fixtureProfileIDs(t *testing.T, fixture map[string]any) []string {
+	t.Helper()
+	profiles, ok := fixture["profiles"].([]any)
+	if !ok {
+		t.Fatal("compatibility fixture profiles are malformed")
+	}
+	ids := make([]string, 0, len(profiles))
+	for _, entry := range profiles {
+		profileValue, ok := entry.(map[string]any)
+		if !ok {
+			t.Fatal("compatibility fixture profile is malformed")
+		}
+		id, ok := profileValue["id"].(string)
+		if !ok {
+			t.Fatal("compatibility fixture profile ID is malformed")
+		}
+		ids = append(ids, id)
+	}
+	sort.Strings(ids)
+	return ids
+}
+
+func catalogValue(t *testing.T, loaded Set) map[string]any {
+	t.Helper()
+	backends := make([]any, 0, len(loaded.Backends))
+	for _, backend := range loaded.Backends {
+		backends = append(backends, map[string]any{"id": backend.ID, "endpoint": backend.Endpoint, "api_key_env": backend.APIKeyEnv, "extra_params": interfaceValue(backend.ExtraParams), "concurrency_limit": backend.ConcurrencyLimit, "queue_capacity": backend.QueueCapacity, "queue_capacity_set": backend.QueueCapacitySet})
+	}
+	sort.Slice(backends, func(left, right int) bool {
+		return backends[left].(map[string]any)["id"].(string) < backends[right].(map[string]any)["id"].(string)
+	})
+	actualProfiles := make([]any, 0, len(loaded.profileIDs))
+	for _, id := range loaded.profileIDs {
+		profile, err := loaded.Profiles.GetProfile(context.Background(), id)
+		if err != nil {
+			t.Fatalf("load profile %q: %v", id, err)
+		}
+		actualProfiles = append(actualProfiles, map[string]any{"id": profile.ID, "base_profile": profile.BaseProfileID, "backend": profile.BackendID, "endpoint": profile.Endpoint, "model": profile.Model, "temperature": profile.Temperature, "max_tokens": profile.MaxTokens, "top_p": profile.TopP, "timeout_seconds": profile.TimeoutSeconds, "service_tier": profile.ServiceTier, "reasoning_effort": profile.ReasoningEffort, "api_key_env": profile.APIKeyEnv, "api_key_required": profile.APIKeyRequired, "extra_params": interfaceValue(profile.ExtraParams)})
+	}
+	return map[string]any{"backends": backends, "profiles": actualProfiles}
+}
+
+func interfaceValue(value map[string]any) any {
+	if value == nil {
+		return nil
+	}
+	return value
+}
+
+func testSource(id string, fsys fs.FS) Source {
+	return Source{Name: id, ExpectedBackendID: id, FS: fsys, Root: "catalog"}
+}
+
+func validCatalogFS(id string) fstest.MapFS {
+	return catalogFS(id, map[string]string{
+		"catalog/profiles/" + id + "-profile.yml": validProfile(id+"-profile", id),
+	}, "null")
+}
+
+func catalogFS(id string, profiles map[string]string, extraParams string) fstest.MapFS {
+	fsys := fstest.MapFS{
+		"catalog/backend.json": &fstest.MapFile{Data: []byte(validManifest(id, "TEST_API_KEY", extraParams))},
+	}
+	for name, content := range profiles {
+		fsys[name] = &fstest.MapFile{Data: []byte(content)}
+	}
+	return fsys
+}
+
+func validManifest(id, apiKeyEnv, extraParams string) string {
+	return fmt.Sprintf(
+		`{"schema_version":1,"id":%q,"endpoint":%q,"api_key_env":%q,"concurrency_limit":2,"queue_capacity":3,"extra_params":%s}`,
+		id,
+		"https://"+id+".example/v1",
+		apiKeyEnv,
+		extraParams,
+	)
+}
+
+func validProfile(id, backendID string) string {
+	return fmt.Sprintf("id: %s\nbackend: %s\nmodel: test-model\n", id, backendID)
+}
+
+var _ fs.FS = openrouter.FS()

internal/domain/domain.go

@@ -69,6 +69,7 @@ type RunRequest struct {
 	Vars             map[string]string
 	Execution        *ExecutionTargetOverride
 	Validation       *OutputContract
+	AppendedMessages []RenderedMessage
 }
 
 // RunResult represents the complete result of a prompt execution run.

internal/domain/message.go

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

internal/domain/message_test.go

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

internal/llm/openai_compatible_client_test.go

@@ -363,6 +363,7 @@ func TestOpenAICompatibleClientRequestMapping(t *testing.T) {
 		run  func(*testing.T)
 	}{
 		{name: "complete request and response mapping", run: checkCompleteRequestAndResponseMapping},
+		{name: "canonical role and exact content", run: checkCanonicalRoleAndExactContentMapping},
 		{name: "cache-controlled message", run: checkCacheControlledMessageMapping},
 		{name: "empty cache-control TTL", run: checkEmptyCacheControlTTLOmission},
 		{name: "session ID", run: checkSessionIDMapping},
@@ -377,6 +378,34 @@ func TestOpenAICompatibleClientRequestMapping(t *testing.T) {
 	}
 }
 
+func checkCanonicalRoleAndExactContentMapping(t *testing.T) {
+	provider := newRecordingProvider(t)
+	client := newProviderClient(t, provider, OpenAICompatibleConfig{})
+	const content = "  exact content\nwith whitespace  "
+	_, err := client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{
+			Role:    domain.RoleDeveloper,
+			Content: content,
+		}}},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("Generate() error = %v", err)
+	}
+
+	var messages []struct {
+		Role    string `json:"role"`
+		Content string `json:"content"`
+	}
+	request := provider.lastRequest(t)
+	if !request.decodeField(t, "messages", &messages) || len(messages) != 1 {
+		t.Fatalf("messages = %#v, want one message", messages)
+	}
+	if messages[0].Role != domain.RoleDeveloper || messages[0].Content != content {
+		t.Fatalf("message = %#v, want role %q and exact content %q", messages[0], domain.RoleDeveloper, content)
+	}
+}
+
 func checkCompleteRequestAndResponseMapping(t *testing.T) {
 	provider := newRecordingProvider(t)
 	provider.respond(http.StatusOK, `{

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

@@ -1,8 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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/google/rakestrawhome-gemma-4-31b.yml

@@ -1,3 +0,0 @@
-id: rakestrawhome-gemma-4-31b
-backend: rakestrawhome
-model: google/gemma-4-31b-it

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

@@ -1,8 +0,0 @@
-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

@@ -1,8 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,7 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,7 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,6 +0,0 @@
-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

@@ -1,16 +0,0 @@
-package builtin
-
-import (
-	"embed"
-
-	"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)
-}

internal/profile/builtin/repository_test.go

@@ -1,118 +0,0 @@
-package builtin
-
-import (
-	"context"
-	"io/fs"
-	"strings"
-	"testing"
-
-	"gitea.maximumdirect.net/eric/promptkit/internal/backend"
-	"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 !builtInBackendIDs[p.BackendID] {
-				t.Fatalf("expected profile %q to select a maintained built-in, got %q", id, 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 TestRakestrawhomeGemmaProfileUsesNativeDefaults(t *testing.T) {
-	p, err := NewRepository().GetProfile(context.Background(), "rakestrawhome-gemma-4-31b")
-	if err != nil {
-		t.Fatalf("load Rakestrawhome Gemma profile: %v", err)
-	}
-	if p.ID != "rakestrawhome-gemma-4-31b" ||
-		p.BackendID != backend.RakestrawHomeID ||
-		p.Model != "google/gemma-4-31b-it" ||
-		p.Endpoint != "" ||
-		p.Temperature != 0 ||
-		p.MaxTokens != 0 ||
-		p.TopP != 0 ||
-		p.TimeoutSeconds != 0 ||
-		p.ServiceTier != "" ||
-		p.ReasoningEffort != "" ||
-		p.APIKeyEnv != "" ||
-		p.APIKeyRequired ||
-		p.ExtraParams != nil {
-		t.Fatalf("unexpected Rakestrawhome Gemma profile: %#v", p)
-	}
-}
-
-var builtInBackendIDs = map[string]bool{
-	backend.OpenRouterID:    true,
-	backend.RakestrawHomeID: true,
-}
-
-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)
-		}
-		backendID, ok := raw["backend"].(string)
-		if !ok || !builtInBackendIDs[backendID] {
-			t.Fatalf("built-in profile %s does not select a maintained built-in: %#v", name, raw["backend"])
-		}
-		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
-}

internal/profile/eager_repository.go

@@ -0,0 +1,98 @@
+package profile
+
+import (
+	"context"
+	"fmt"
+	"io/fs"
+	"sort"
+	"strings"
+
+	"gitea.maximumdirect.net/eric/promptkit/internal/domain"
+	"gitea.maximumdirect.net/eric/promptkit/internal/filecatalog"
+	"gitea.maximumdirect.net/eric/promptkit/internal/jsonvalue"
+)
+
+// LoadedProfileMetadata identifies one profile accepted by LoadFSRepository.
+type LoadedProfileMetadata struct {
+	// ID is the normalized profile ID.
+	ID string
+	// Path is the safe root-relative source path.
+	Path string
+	// ExplicitFields lists the sorted top-level YAML fields present in source.
+	ExplicitFields []string
+}
+
+// LoadFSRepository eagerly validates every profile under root and returns an
+// immutable raw repository and independently owned source metadata.
+func LoadFSRepository(ctx context.Context, fsys fs.FS, root string) (Repository, []LoadedProfileMetadata, error) {
+	if fsys == nil {
+		return nil, nil, fmt.Errorf("failed to read profile directory: filesystem is nil")
+	}
+	paths, err := filecatalog.FindFSYAMLFiles(ctx, fsys, root)
+	if err != nil {
+		return nil, nil, fmt.Errorf("failed to read profile directory: %w", err)
+	}
+
+	repository := &loadedRepository{profiles: make(map[string]domain.ExecutionProfile, len(paths))}
+	metadata := make([]LoadedProfileMetadata, 0, len(paths))
+	for _, path := range paths {
+		if err := ctx.Err(); err != nil {
+			return nil, nil, err
+		}
+		data, err := fs.ReadFile(fsys, path)
+		if err != nil {
+			return nil, nil, fmt.Errorf("failed to read profile file %s: %w", filecatalog.DisplayPath(root, path), err)
+		}
+		fileMetadata, err := readProfileFileMetadata(data)
+		if err != nil {
+			return nil, nil, fmt.Errorf("%w: %s", ErrInvalidYAML, filecatalog.DisplayPath(root, path))
+		}
+		if fileMetadata.hasRawAPIKey {
+			return nil, nil, fmt.Errorf("%w: %s", ErrRawAPIKeyNotAllowed, filecatalog.DisplayPath(root, path))
+		}
+		definition, err := decodeProfile(data)
+		if err != nil {
+			return nil, nil, fmt.Errorf("%w: %s", ErrInvalidYAML, filecatalog.DisplayPath(root, path))
+		}
+		definition.ExtraParams, err = jsonvalue.CopyMap(definition.ExtraParams)
+		if err != nil {
+			return nil, nil, fmt.Errorf("%w: %s", ErrInvalidProfile, filecatalog.DisplayPath(root, path))
+		}
+		if err := NormalizeAndValidateDefinition(definition); err != nil {
+			return nil, nil, fmt.Errorf("%w: %s", ErrInvalidProfile, filecatalog.DisplayPath(root, path))
+		}
+		if _, exists := repository.profiles[definition.ID]; exists {
+			return nil, nil, fmt.Errorf("%w: %s: duplicate profile ID", ErrInvalidProfile, filecatalog.DisplayPath(root, path))
+		}
+		repository.profiles[definition.ID] = *definition
+		fields := append([]string(nil), fileMetadata.explicitFields...)
+		sort.Strings(fields)
+		metadata = append(metadata, LoadedProfileMetadata{
+			ID:             definition.ID,
+			Path:           filecatalog.DisplayPath(root, path),
+			ExplicitFields: fields,
+		})
+	}
+	sort.Slice(metadata, func(left, right int) bool { return metadata[left].ID < metadata[right].ID })
+	return repository, metadata, nil
+}
+
+type loadedRepository struct {
+	profiles map[string]domain.ExecutionProfile
+}
+
+func (r *loadedRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	if err := ctx.Err(); err != nil {
+		return nil, err
+	}
+	definition, found := r.profiles[strings.TrimSpace(id)]
+	if !found {
+		return nil, ErrProfileNotFound
+	}
+	extraParams, err := jsonvalue.CopyMap(definition.ExtraParams)
+	if err != nil {
+		return nil, fmt.Errorf("copy loaded profile %q: %w", definition.ID, err)
+	}
+	definition.ExtraParams = extraParams
+	return &definition, nil
+}

internal/profile/eager_repository_test.go

@@ -0,0 +1,110 @@
+package profile
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"reflect"
+	"testing"
+	"testing/fstest"
+)
+
+func TestLoadFSRepositoryLoadsSortedIndependentProfiles(t *testing.T) {
+	fsys := fstest.MapFS{
+		"profiles/z.yml": {Data: []byte("id: z\nbackend: local\nmodel: z-model\nextra_params:\n  nested:\n    value: one\n")},
+		"profiles/a.yml": {Data: []byte("id: a\nbackend: local\nmodel: a-model\nendpoint: ''\n")},
+	}
+
+	repository, metadata, err := LoadFSRepository(context.Background(), fsys, "profiles")
+	if err != nil {
+		t.Fatalf("load repository: %v", err)
+	}
+	if len(metadata) != 2 || metadata[0].ID != "a" || metadata[1].ID != "z" || metadata[0].Path != "a.yml" {
+		t.Fatalf("unexpected metadata: %#v", metadata)
+	}
+	if len(metadata[0].ExplicitFields) != 4 || metadata[0].ExplicitFields[0] != "backend" || metadata[0].ExplicitFields[1] != "endpoint" {
+		t.Fatalf("expected explicitly empty endpoint metadata, got %#v", metadata[0].ExplicitFields)
+	}
+	metadata[1].ExplicitFields[0] = "changed"
+	first, err := repository.GetProfile(context.Background(), "z")
+	if err != nil {
+		t.Fatalf("load profile: %v", err)
+	}
+	first.ExtraParams["nested"].(map[string]any)["value"] = "changed"
+	second, err := repository.GetProfile(context.Background(), "z")
+	if err != nil {
+		t.Fatalf("reload profile: %v", err)
+	}
+	if second.ExtraParams["nested"].(map[string]any)["value"] != "one" {
+		t.Fatalf("profile value was not defensively copied: %#v", second)
+	}
+}
+
+func TestLoadFSRepositoryRejectsInvalidProfiles(t *testing.T) {
+	tests := map[string]fstest.MapFS{
+		"duplicate IDs": {
+			"profiles/one.yml": {Data: []byte("id: duplicate\nbackend: local\nmodel: one\n")},
+			"profiles/two.yml": {Data: []byte("id: duplicate\nbackend: local\nmodel: two\n")},
+		},
+		"raw API key": {
+			"profiles/one.yml": {Data: []byte("id: one\nbackend: local\nmodel: one\napi_key: forbidden\n")},
+		},
+		"multiple documents": {
+			"profiles/one.yml": {Data: []byte("id: one\nbackend: local\nmodel: one\n---\nid: two\n")},
+		},
+	}
+	for name, fsys := range tests {
+		t.Run(name, func(t *testing.T) {
+			_, _, err := LoadFSRepository(context.Background(), fsys, "profiles")
+			if err == nil {
+				t.Fatal("expected load failure")
+			}
+		})
+	}
+}
+
+func TestLoadedRepositoryHonorsCancellationAndMissingProfiles(t *testing.T) {
+	repository, _, err := LoadFSRepository(context.Background(), fstest.MapFS{
+		"profiles/one.yml": {Data: []byte("id: one\nbackend: local\nmodel: one\n")},
+	}, "profiles")
+	if err != nil {
+		t.Fatalf("load repository: %v", err)
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+	if _, err := repository.GetProfile(ctx, "one"); !errors.Is(err, context.Canceled) {
+		t.Fatalf("expected cancellation, got %v", err)
+	}
+	if _, err := repository.GetProfile(context.Background(), "missing"); !errors.Is(err, ErrProfileNotFound) {
+		t.Fatalf("expected missing profile, got %v", err)
+	}
+}
+
+func TestLoadFSRepositoryMatchesPointLookupForRawProfiles(t *testing.T) {
+	fsys := fstest.MapFS{
+		"profiles/base.yml":    {Data: []byte("id: base\nbackend: local\nmodel: base-model\n")},
+		"profiles/derived.yml": {Data: []byte("id: derived\nbase_profile: base\nreasoning_effort: high\n")},
+	}
+	eager, _, err := LoadFSRepository(context.Background(), fsys, "profiles")
+	if err != nil {
+		t.Fatalf("load eager repository: %v", err)
+	}
+	pointLookup := NewFSRepository(fsys, "profiles")
+	for _, id := range []string{"base", "derived"} {
+		t.Run(id, func(t *testing.T) {
+			got, err := eager.GetProfile(context.Background(), id)
+			if err != nil {
+				t.Fatalf("load eager profile: %v", err)
+			}
+			want, err := pointLookup.GetProfile(context.Background(), id)
+			if err != nil {
+				t.Fatalf("load point-in-time profile: %v", err)
+			}
+			if !reflect.DeepEqual(got, want) {
+				t.Fatalf("raw profiles differ: got %#v, want %#v", got, want)
+			}
+		})
+	}
+}
+
+var _ fs.FS = fstest.MapFS{}

internal/profile/filesystem_repository.go

@@ -166,6 +166,7 @@ type profileMatch struct {
 type profileFileMetadata struct {
 	ids            []string
 	hasRawAPIKey   bool
+	explicitFields []string
 }
 
 func readProfileFileMetadata(data []byte) (profileFileMetadata, error) {
@@ -206,6 +207,7 @@ func profileMetadataFromNode(node *yaml.Node) profileFileMetadata {
 	for i := 0; i+1 < len(mapping.Content); i += 2 {
 		key := mapping.Content[i]
 		value := mapping.Content[i+1]
+		metadata.explicitFields = append(metadata.explicitFields, key.Value)
 		switch key.Value {
 		case "id":
 			metadata.ids = append(metadata.ids, strings.TrimSpace(value.Value))
@@ -228,6 +230,7 @@ func (m profileFileMetadata) matchesID(id string) bool {
 func (m *profileFileMetadata) merge(other profileFileMetadata) {
 	m.ids = append(m.ids, other.ids...)
 	m.hasRawAPIKey = m.hasRawAPIKey || other.hasRawAPIKey
+	m.explicitFields = append(m.explicitFields, other.explicitFields...)
 }
 
 func decodeProfile(data []byte) (*domain.ExecutionProfile, error) {

internal/prompt/go_renderer.go

@@ -68,8 +68,9 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 			return nil, err
 		}
 
-		if tmplMsg.Role == "" {
+		role, err := domain.NormalizeMessageRole(tmplMsg.Role)
-			return nil, fmt.Errorf("%w: message %d", ErrInvalidMessageRole, i)
+		if err != nil {
+			return nil, fmt.Errorf("%w: message %d: %v", ErrInvalidMessageRole, i, err)
 		}
 
 		if err := ctx.Err(); err != nil {
@@ -96,7 +97,7 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 		}
 
 		renderedMessages = append(renderedMessages, domain.RenderedMessage{
-			Role:         tmplMsg.Role,
+			Role:         role,
 			Content:      buf.String(),
 			CacheControl: cloneCacheControl(tmplMsg.CacheControl),
 		})

internal/prompt/renderer_test.go

@@ -361,6 +361,35 @@ func TestGoRenderer_Render(t *testing.T) {
 			t.Fatalf("expected ErrInvalidMessageRole, got %v", err)
 		}
 	})
+
+	t.Run("canonicalizes directly supplied message roles", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Templates: []domain.PromptMessageTemplate{
+				{Role: " \u2003DeVeLoPeR\u2003 ", Content: "Hello"},
+			},
+		}
+		res, err := renderer.Render(ctx, def, nil, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if got := res.Messages[0].Role; got != domain.RoleDeveloper {
+			t.Fatalf("rendered role = %q, want %q", got, domain.RoleDeveloper)
+		}
+	})
+
+	t.Run("rejects unsupported directly supplied message roles", func(t *testing.T) {
+		const unsupportedRole = "consumer-private-role"
+		def := &domain.PromptDefinition{
+			Templates: []domain.PromptMessageTemplate{{Role: unsupportedRole, Content: "Hello"}},
+		}
+		_, err := renderer.Render(ctx, def, nil, vars)
+		if !errors.Is(err, ErrInvalidMessageRole) {
+			t.Fatalf("expected ErrInvalidMessageRole, got %v", err)
+		}
+		if strings.Contains(err.Error(), unsupportedRole) {
+			t.Fatalf("renderer error exposed unsupported role: %v", err)
+		}
+	})
 }
 
 func TestGoRendererCancellation(t *testing.T) {

internal/promptdef/filesystem_repository.go

@@ -254,20 +254,27 @@ func normalizePromptDefinitionWithContent(raw *promptDefinitionFile, readContent
 
 	templates := make([]domain.PromptMessageTemplate, 0, len(raw.Messages))
 	for i, msg := range raw.Messages {
-		role := strings.TrimSpace(msg.Role)
+		role, err := domain.NormalizeMessageRole(msg.Role)
-		if role == "" {
+		if err != nil {
-			return nil, fmt.Errorf("message %d role is required", i)
+			return nil, fmt.Errorf("message %d role: %w", i, err)
 		}
 
 		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)
+			return nil, fmt.Errorf("message %d must set exactly one of content or content_file", i)
 		}
 
-		cacheControl, err := normalizeCacheControl(msg.CacheControl)
+		var rawCacheControl *domain.CacheControl
+		if msg.CacheControl != nil {
+			rawCacheControl = &domain.CacheControl{
+				Type: domain.CacheControlType(msg.CacheControl.Type),
+				TTL:  msg.CacheControl.TTL,
+			}
+		}
+		cacheControl, err := domain.NormalizeCacheControl(rawCacheControl)
 		if err != nil {
-			return nil, fmt.Errorf("message %d (%s) cache_control: %w", i, role, err)
+			return nil, fmt.Errorf("message %d cache_control: %w", i, err)
 		}
 
 		templateContent := msg.Content
@@ -275,7 +282,7 @@ func normalizePromptDefinitionWithContent(raw *promptDefinitionFile, readContent
 		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)
+				return nil, fmt.Errorf("prompt %q message %d: failed to read content_file %q: %w", id, i, msg.ContentFile, err)
 			}
 			templateContent = body
 			resolvedContentFile = resolvedPath
@@ -319,27 +326,3 @@ func normalizePromptDefinitionWithContent(raw *promptDefinitionFile, readContent
 		Validation:     outputContract,
 	}, 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
-}

internal/promptdef/repository_test.go

@@ -1087,6 +1087,48 @@ output:
 	}
 }
 
+func TestPromptDefinitionMessageRoleNormalization(t *testing.T) {
+	const invalidRole = "consumer-private-role"
+	repo := NewFSRepository(fstest.MapFS{
+		"canonical.yaml": {Data: []byte(`
+id: canonical
+version: "1"
+messages:
+  - role: " \u2003SyStEm\u2003 "
+    content: test
+output:
+  format: text
+  validation_mode: none
+`)},
+		"invalid.yaml": {Data: []byte(`
+id: invalid-role
+version: "1"
+messages:
+  - role: consumer-private-role
+    content: test
+output:
+  format: text
+  validation_mode: none
+`)},
+	}, ".")
+
+	definition, err := repo.GetPromptDefinition(context.Background(), "canonical", "")
+	if err != nil {
+		t.Fatalf("GetPromptDefinition() error = %v", err)
+	}
+	if got := definition.Templates[0].Role; got != domain.RoleSystem {
+		t.Fatalf("normalized role = %q, want %q", got, domain.RoleSystem)
+	}
+
+	_, err = repo.GetPromptDefinition(context.Background(), "invalid-role", "")
+	if !errors.Is(err, ErrInvalidPromptDefinition) {
+		t.Fatalf("expected ErrInvalidPromptDefinition, got %v", err)
+	}
+	if strings.Contains(err.Error(), invalidRole) {
+		t.Fatalf("invalid prompt error exposed the supplied role: %v", err)
+	}
+}
+
 func assertCacheControl(t *testing.T, got *domain.CacheControl, wantType domain.CacheControlType, wantTTL string) {
 	t.Helper()
 	if got == nil {

internal/usecase/prepared_execution.go

@@ -213,16 +213,7 @@ func clonePreparedRun(source *domain.PreparedRun) (*domain.PreparedRun, error) {
 			copied.InputHashes[name] = hash
 		}
 	}
-	if source.Messages != nil {
+	copied.Messages = domain.CloneRenderedMessages(source.Messages)
-		copied.Messages = make([]domain.RenderedMessage, len(source.Messages))
-		for i, message := range source.Messages {
-			copied.Messages[i] = message
-			if message.CacheControl != nil {
-				cacheControl := *message.CacheControl
-				copied.Messages[i].CacheControl = &cacheControl
-			}
-		}
-	}
 
 	if source.StructuredOutput != nil {
 		structuredOutput := *source.StructuredOutput

internal/usecase/repairer.go

@@ -55,21 +55,21 @@ func (r *defaultOutputRepairer) Repair(ctx context.Context, req RepairRequest) (
 		return nil, err
 	}
 
-	messages := make([]domain.RenderedMessage, len(req.OriginalMessages), len(req.OriginalMessages)+2)
+	hasPreviousOutput := strings.TrimSpace(req.PreviousOutput) != ""
-	copy(messages, req.OriginalMessages)
+	suffix := make([]domain.RenderedMessage, 0, 2)
-	if strings.TrimSpace(req.PreviousOutput) != "" {
+	if hasPreviousOutput {
-		messages = append(messages, domain.RenderedMessage{
+		suffix = append(suffix, domain.RenderedMessage{
-			Role:    "assistant",
+			Role:    domain.RoleAssistant,
 			Content: req.PreviousOutput,
 		})
 	}
 
 	previousResponse := "The previous response was empty."
-	if strings.TrimSpace(req.PreviousOutput) != "" {
+	if hasPreviousOutput {
 		previousResponse = "The previous response is included immediately before this instruction."
 	}
-	messages = append(messages, domain.RenderedMessage{
+	suffix = append(suffix, domain.RenderedMessage{
-		Role: "user",
+		Role: domain.RoleUser,
 		Content: fmt.Sprintf(
 			"Repair attempt %d of %d for validation mode %s.\n"+
 				"Preserve valid values and change only what is necessary.\n"+
@@ -83,6 +83,7 @@ func (r *defaultOutputRepairer) Repair(ctx context.Context, req RepairRequest) (
 			formatRepairDiagnostics(req.ValidationErrors),
 		),
 	})
+	messages := domain.ConcatRenderedMessages(req.OriginalMessages, suffix)
 
 	resp, err := r.llm.Generate(ctx, newGenerationRequest(
 		domain.RenderedPrompt{Messages: messages},

internal/usecase/repairer_test.go

@@ -30,6 +30,8 @@ func TestDefaultOutputRepairerBuildsFullContextRequest(t *testing.T) {
 	original := []domain.RenderedMessage{
 		{Role: "system", Content: "Follow the task.", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}},
 		{Role: "user", Content: "Summarize the report."},
+		{Role: "assistant", Content: "Consumer-supplied previous response."},
+		{Role: "user", Content: "Consumer-supplied correction."},
 	}
 	before := append([]domain.RenderedMessage(nil), original...)
 	previous := strings.Repeat("candidate ", 12_000)

internal/usecase/runner.go

@@ -4,10 +4,13 @@ import (
 	"context"
 	"crypto/rand"
 	"crypto/sha256"
+	"encoding/binary"
 	"encoding/hex"
 	"encoding/json"
 	"errors"
 	"fmt"
+	"hash"
+	"io"
 	"os"
 	"strings"
 	"time"
@@ -444,9 +447,11 @@ func (r *Runner) completePreparationWithStructuredOutput(
 	if err != nil {
 		return nil, fmt.Errorf("%w: %w", ErrPromptRender, err)
 	}
+	effectivePrompt := *renderedPrompt
 	if state.directSessionID != "" {
-		renderedPrompt.SessionID = state.directSessionID
+		effectivePrompt.SessionID = state.directSessionID
 	}
+	effectivePrompt.Messages = domain.ConcatRenderedMessages(renderedPrompt.Messages, req.AppendedMessages)
 
 	end := time.Now().UTC()
 	effectiveModel := state.effectiveModel
@@ -462,9 +467,9 @@ func (r *Runner) completePreparationWithStructuredOutput(
 		OutputContract:       state.effectiveContract,
 		StructuredOutput:     structuredOutput,
 		InputHashes:          inputHashes,
-		SessionID:            renderedPrompt.SessionID,
+		SessionID:            effectivePrompt.SessionID,
-		RenderedPromptHash:   hashRenderedPrompt(*renderedPrompt),
+		RenderedPromptHash:   hashRenderedPrompt(effectivePrompt),
-		Messages:             renderedPrompt.Messages,
+		Messages:             effectivePrompt.Messages,
 		StartTime:            state.start,
 		EndTime:              end,
 		DurationMS:           end.Sub(state.start).Milliseconds(),
@@ -714,29 +719,38 @@ func resolveOutputContract(def *domain.PromptDefinition, override *domain.Output
 	return contract, nil
 }
 
+const renderedPromptHashVersion = "promptkit/rendered-prompt/v2\x00"
+
 func hashRenderedPrompt(p domain.RenderedPrompt) string {
-	var b strings.Builder
+	hasher := sha256.New()
-	if p.SessionID != "" {
+	var lengthBuffer [8]byte
-		b.WriteString("session_id=")
+	_, _ = io.WriteString(hasher, renderedPromptHashVersion)
-		b.WriteString(p.SessionID)
+	writeRenderedPromptHashString(hasher, &lengthBuffer, p.SessionID)
-		b.WriteString("\n---\n")
+	writeRenderedPromptHashLength(hasher, &lengthBuffer, len(p.Messages))
+	for _, message := range p.Messages {
+		writeRenderedPromptHashString(hasher, &lengthBuffer, message.Role)
+		writeRenderedPromptHashString(hasher, &lengthBuffer, message.Content)
+		if message.CacheControl == nil {
+			lengthBuffer[0] = 0
+			_, _ = hasher.Write(lengthBuffer[:1])
+			continue
 		}
-	for _, msg := range p.Messages {
+		lengthBuffer[0] = 1
-		b.WriteString(msg.Role)
+		_, _ = hasher.Write(lengthBuffer[:1])
-		b.WriteByte('\n')
+		writeRenderedPromptHashString(hasher, &lengthBuffer, string(message.CacheControl.Type))
-		b.WriteString(msg.Content)
+		writeRenderedPromptHashString(hasher, &lengthBuffer, message.CacheControl.TTL)
-		if msg.CacheControl != nil {
-			b.WriteString("\ncache_control.type=")
-			b.WriteString(string(msg.CacheControl.Type))
-			if msg.CacheControl.TTL != "" {
-				b.WriteString("\ncache_control.ttl=")
-				b.WriteString(msg.CacheControl.TTL)
 	}
-		}
+	return hex.EncodeToString(hasher.Sum(nil))
-		b.WriteString("\n---\n")
+}
-	}
+
-	h := sha256.Sum256([]byte(b.String()))
+func writeRenderedPromptHashLength(hasher hash.Hash, buffer *[8]byte, length int) {
-	return hex.EncodeToString(h[:])
+	binary.BigEndian.PutUint64(buffer[:], uint64(length))
+	_, _ = hasher.Write(buffer[:])
+}
+
+func writeRenderedPromptHashString(hasher hash.Hash, buffer *[8]byte, value string) {
+	writeRenderedPromptHashLength(hasher, buffer, len(value))
+	_, _ = io.WriteString(hasher, value)
 }
 
 func buildOutputArtifact(content string, format domain.OutputFormat) domain.Artifact {

internal/usecase/runner_test.go

@@ -1110,14 +1110,18 @@ func TestDeriveStructuredSchemaName(t *testing.T) {
 	}
 }
 
-func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
+func TestHashRenderedPromptIncludesEveryContractedField(t *testing.T) {
-	uncached := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+	base := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
-		{Role: "system", Content: "sys"},
+		{Role: domain.RoleSystem, Content: "sys"},
-		{Role: "user", Content: "usr"},
+		{Role: domain.RoleUser, Content: "usr"},
 	}}
-	wantLegacyHash := hashString("system\nsys\n---\nuser\nusr\n---\n")
+	identical := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
-	if got := hashRenderedPrompt(uncached); got != wantLegacyHash {
+		{Role: domain.RoleSystem, Content: "sys"},
-		t.Fatalf("expected no-cache hash to preserve legacy input, got %q want %q", got, wantLegacyHash)
+		{Role: domain.RoleUser, Content: "usr"},
+	}}
+	baseHash := hashRenderedPrompt(base)
+	if baseHash != hashRenderedPrompt(identical) {
+		t.Fatal("identical rendered prompts must have the same hash")
 	}
 
 	withCache := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
@@ -1129,7 +1133,7 @@ func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
 				TTL:  "1h",
 			},
 		},
-		{Role: "user", Content: "usr"},
+		{Role: domain.RoleUser, Content: "usr"},
 	}}
 	alsoWithCache := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
 		{
@@ -1140,7 +1144,7 @@ func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
 				TTL:  "1h",
 			},
 		},
-		{Role: "user", Content: "usr"},
+		{Role: domain.RoleUser, Content: "usr"},
 	}}
 	withoutTTL := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
 		{
@@ -1150,11 +1154,11 @@ func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
 				Type: domain.CacheControlEphemeral,
 			},
 		},
-		{Role: "user", Content: "usr"},
+		{Role: domain.RoleUser, Content: "usr"},
 	}}
 
 	cachedHash := hashRenderedPrompt(withCache)
-	if cachedHash == hashRenderedPrompt(uncached) {
+	if cachedHash == baseHash {
 		t.Fatal("expected cache control to change rendered prompt hash")
 	}
 	if cachedHash != hashRenderedPrompt(alsoWithCache) {
@@ -1163,6 +1167,29 @@ func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
 	if cachedHash == hashRenderedPrompt(withoutTTL) {
 		t.Fatal("expected ttl changes to affect rendered prompt hash")
 	}
+
+	variants := []struct {
+		name   string
+		prompt domain.RenderedPrompt
+	}{
+		{name: "role", prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: domain.RoleDeveloper, Content: "sys"}, {Role: domain.RoleUser, Content: "usr"}}}},
+		{name: "content", prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: domain.RoleSystem, Content: "changed"}, {Role: domain.RoleUser, Content: "usr"}}}},
+		{name: "order", prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: domain.RoleUser, Content: "usr"}, {Role: domain.RoleSystem, Content: "sys"}}}},
+		{name: "session", prompt: domain.RenderedPrompt{SessionID: "session", Messages: base.Messages}},
+	}
+	for _, variant := range variants {
+		t.Run(variant.name, func(t *testing.T) {
+			if hashRenderedPrompt(variant.prompt) == baseHash {
+				t.Fatalf("%s did not change the rendered-prompt hash", variant.name)
+			}
+		})
+	}
+
+	oneMessage := domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: domain.RoleUser, Content: "x\n---\nassistant\ny"}}}
+	twoMessages := domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: domain.RoleUser, Content: "x"}, {Role: domain.RoleAssistant, Content: "y"}}}
+	if hashRenderedPrompt(oneMessage) == hashRenderedPrompt(twoMessages) {
+		t.Fatal("length-framed hashing did not distinguish legacy separator collision")
+	}
 }
 
 func TestHashRenderedPromptIncludesSessionIDWhenPresent(t *testing.T) {
@@ -1204,6 +1231,75 @@ func TestHashRenderedPromptIncludesSessionIDWhenPresent(t *testing.T) {
 	}
 }
 
+func TestRunnerComposesAppendedMessagesIntoEveryRenderedPrompt(t *testing.T) {
+	renderer := &fakeRenderer{rendered: &domain.RenderedPrompt{
+		SessionID: "rendered-session",
+		Messages: []domain.RenderedMessage{
+			{Role: domain.RoleSystem, Content: "ordinary prefix", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}},
+			{Role: domain.RoleUser, Content: "ordinary request"},
+		},
+	}}
+	client := &fakeLLM{resp: &domain.GenerateResponse{Content: "output"}}
+	runner := NewRunner(
+		&fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)},
+		&fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}},
+		nil,
+		defaultArtifactReader(),
+		renderer,
+		client,
+		nil,
+		nil,
+	)
+	appended := []domain.RenderedMessage{
+		{Role: domain.RoleAssistant, Content: "consumer response"},
+		{Role: domain.RoleUser, Content: "consumer correction", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral}},
+	}
+	request := domain.RunRequest{PromptID: "p", ProfileID: "exec", Inputs: singleInputRef(), AppendedMessages: appended}
+
+	plain, err := runner.Prepare(context.Background(), domain.RunRequest{PromptID: "p", ProfileID: "exec", Inputs: singleInputRef()})
+	if err != nil {
+		t.Fatalf("prepare without appended messages: %v", err)
+	}
+	empty, err := runner.Prepare(context.Background(), domain.RunRequest{PromptID: "p", ProfileID: "exec", Inputs: singleInputRef(), AppendedMessages: []domain.RenderedMessage{}})
+	if err != nil {
+		t.Fatalf("prepare with empty appended messages: %v", err)
+	}
+	if !reflect.DeepEqual(empty.Messages, plain.Messages) || empty.RenderedPromptHash != plain.RenderedPromptHash {
+		t.Fatalf("empty appended messages changed preparation: empty=%#v plain=%#v", empty, plain)
+	}
+	prepared, err := runner.Prepare(context.Background(), request)
+	if err != nil {
+		t.Fatalf("prepare with appended messages: %v", err)
+	}
+	if prepared.PromptHash != plain.PromptHash || prepared.RenderedPromptHash == plain.RenderedPromptHash {
+		t.Fatalf("hashes with appended messages = (%q, %q), plain = (%q, %q)", prepared.PromptHash, prepared.RenderedPromptHash, plain.PromptHash, plain.RenderedPromptHash)
+	}
+
+	expected := []domain.RenderedMessage{
+		{Role: domain.RoleSystem, Content: "ordinary prefix", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}},
+		{Role: domain.RoleUser, Content: "ordinary request"},
+		{Role: domain.RoleAssistant, Content: "consumer response"},
+		{Role: domain.RoleUser, Content: "consumer correction", CacheControl: &domain.CacheControl{Type: domain.CacheControlEphemeral}},
+	}
+	if !reflect.DeepEqual(prepared.Messages, expected) {
+		t.Fatalf("prepared messages = %#v, want %#v", prepared.Messages, expected)
+	}
+
+	result, err := runner.Run(context.Background(), request)
+	if err != nil {
+		t.Fatalf("run with appended messages: %v", err)
+	}
+	if result.RenderedPromptHash != prepared.RenderedPromptHash || !reflect.DeepEqual(client.lastReq.Prompt.Messages, expected) {
+		t.Fatalf("run did not use the prepared effective prompt: result=%#v request=%#v prepared=%#v", result, client.lastReq.Prompt.Messages, prepared)
+	}
+
+	renderer.rendered.Messages[0].Content = "changed source"
+	appended[1].CacheControl.Type = "changed caller value"
+	if !reflect.DeepEqual(prepared.Messages, expected) {
+		t.Fatalf("prepared messages changed after source or caller mutation: %#v", prepared.Messages)
+	}
+}
+
 func TestRunnerRunSuccessful(t *testing.T) {
 	promptRepo := &fakePromptRepo{def: promptDef(domain.FormatMarkdown, domain.ValidationBasic, 0)}
 	execRepo := &fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}}

message_roles.go

@@ -0,0 +1,14 @@
+package promptkit
+
+import "gitea.maximumdirect.net/eric/promptkit/internal/domain"
+
+const (
+	// RoleDeveloper identifies a provider-bound developer instruction message.
+	RoleDeveloper = domain.RoleDeveloper
+	// RoleSystem identifies a provider-bound system instruction message.
+	RoleSystem = domain.RoleSystem
+	// RoleUser identifies a provider-bound user message.
+	RoleUser = domain.RoleUser
+	// RoleAssistant identifies a provider-bound assistant message.
+	RoleAssistant = domain.RoleAssistant
+)

prepared_execution.go

@@ -17,7 +17,9 @@ type PreparedExecution struct {
 
 // Details returns a fresh caller-owned, credential-redacted copy of the
 // prepared request details. Mutating the result cannot affect execution or a
-// later Details call. Details remains available after execution or discard.
+// later Details call. Its complete effective message content, including any
+// appended request messages, remains subject to the caller's data-handling
+// policy. Details remains available after execution or discard.
 //
 // A nil receiver or zero-value PreparedExecution returns a zero [PreparedRun].
 func (p *PreparedExecution) Details() PreparedRun {

testdata/builtin-catalog-v1.json

@@ -0,0 +1,424 @@
+{
+    "backends": [
+        {
+            "id": "openrouter",
+            "endpoint": "https://openrouter.ai/api/v1",
+            "api_key_env": "OPENROUTER_API_KEY",
+            "extra_params": null,
+            "concurrency_limit": 16,
+            "queue_capacity": 1024,
+            "queue_capacity_set": true
+        },
+        {
+            "id": "rakestrawhome",
+            "endpoint": "https://inference.ai.rakestrawhome.com/v1",
+            "api_key_env": "RAKESTRAWHOME_INFERENCE_API_KEY",
+            "extra_params": null,
+            "concurrency_limit": 4,
+            "queue_capacity": 1024,
+            "queue_capacity_set": true
+        }
+    ],
+    "profiles": [
+        {
+            "id": "aion-2",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "aion-labs/aion-2.0",
+            "temperature": 0.72,
+            "max_tokens": 0,
+            "top_p": 0.95,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "claude-fable-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~anthropic/claude-fable-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 600,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "claude-haiku-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~anthropic/claude-haiku-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "medium",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "claude-opus-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~anthropic/claude-opus-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "claude-sonnet-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~anthropic/claude-sonnet-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "deepseek-3-2",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "deepseek/deepseek-v3.2",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "deepseek-4-flash",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "deepseek/deepseek-v4-flash",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "deepseek-4-pro",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "deepseek/deepseek-v4-pro",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-2-flash",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "google/gemini-2.5-flash",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-2-flash-lite",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "google/gemini-2.5-flash-lite",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-2-pro",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "google/gemini-2.5-pro",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-3-flash-lite",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "google/gemini-3.1-flash-lite",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-flash-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~google/gemini-flash-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemini-pro-latest",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "~google/gemini-pro-latest",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gemma-4-31b",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "google/gemma-4-31b-it:exacto",
+            "temperature": 0.15,
+            "max_tokens": 0,
+            "top_p": 0.98,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gpt-5-mini",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "openai/gpt-5.4-mini",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "gpt-5-nano",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "openai/gpt-5.4-nano",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 240,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "minimax-m2",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "minimax/minimax-m2.5",
+            "temperature": 0.5,
+            "max_tokens": 0,
+            "top_p": 0.95,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "minimax-m3",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "minimax/minimax-m3",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "mistral-large-2512",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "mistralai/mistral-large-2512",
+            "temperature": 0.15,
+            "max_tokens": 0,
+            "top_p": 0.98,
+            "timeout_seconds": 180,
+            "service_tier": "",
+            "reasoning_effort": "",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "mistral-medium-3-5",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "mistralai/mistral-medium-3-5",
+            "temperature": 0.15,
+            "max_tokens": 0,
+            "top_p": 0.98,
+            "timeout_seconds": 180,
+            "service_tier": "",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "mistral-small-3",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "mistralai/mistral-small-3.2-24b-instruct",
+            "temperature": 0.05,
+            "max_tokens": 0,
+            "top_p": 1,
+            "timeout_seconds": 180,
+            "service_tier": "",
+            "reasoning_effort": "",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "mistral-small-4",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "mistralai/mistral-small-2603",
+            "temperature": 0.1,
+            "max_tokens": 0,
+            "top_p": 0.98,
+            "timeout_seconds": 180,
+            "service_tier": "",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "nemotron-3-ultra",
+            "base_profile": "",
+            "backend": "openrouter",
+            "endpoint": "",
+            "model": "nvidia/nemotron-3-ultra-550b-a55b",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 180,
+            "service_tier": "flex",
+            "reasoning_effort": "high",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        },
+        {
+            "id": "rakestrawhome-gemma-4-31b",
+            "base_profile": "",
+            "backend": "rakestrawhome",
+            "endpoint": "",
+            "model": "google/gemma-4-31b-it",
+            "temperature": 0,
+            "max_tokens": 0,
+            "top_p": 0,
+            "timeout_seconds": 0,
+            "service_tier": "",
+            "reasoning_effort": "",
+            "api_key_env": "",
+            "api_key_required": false,
+            "extra_params": null
+        }
+    ]
+}

types.go

@@ -123,6 +123,21 @@ type RunRequest struct {
 	// Validation optionally replaces the prompt's complete output contract. It
 	// does not merge individual fields. Nil uses the prompt contract.
 	Validation *OutputContract
+	// AppendedMessages are already-rendered messages appended in caller order
+	// after every prompt definition message. Promptkit neither templates nor
+	// resolves files in them. Content must be valid UTF-8 and is preserved
+	// exactly, including empty or whitespace-only content. Roles are trimmed and
+	// lowercased, then must be [RoleDeveloper], [RoleSystem], [RoleUser], or
+	// [RoleAssistant]. Any CacheControl is normalized as documented on that
+	// type.
+	//
+	// Nil and empty slices are equivalent. Promptkit imposes no message-count,
+	// byte-size, token, or context-window limit and does not truncate content;
+	// an upstream rejection follows the ordinary generation-error contract.
+	// Prepare, PrepareExecution, and Run validate and copy the messages before
+	// source or model work. Malformed values return an error matching
+	// ErrInvalidRequest.
+	AppendedMessages []RenderedMessage
 }
 
 // PreparedRun contains prepared prompt execution state returned by
@@ -162,8 +177,9 @@ type PreparedRun struct {
 	SessionID string `json:"session_id,omitempty"`
 	// RenderedPromptHash is an opaque equality value for SessionID and Messages.
 	RenderedPromptHash string `json:"rendered_prompt_hash"`
-	// Messages are the rendered messages that Run or RunPrepared passes to the
+	// Messages are the complete effective messages that Run or RunPrepared passes
-	// LLM client.
+	// to the LLM client: definition messages in rendered order followed by any
+	// RunRequest.AppendedMessages in caller order.
 	Messages []RenderedMessage `json:"messages"`
 	// StartTime is the UTC time at which preparation began.
 	StartTime time.Time `json:"start_time,omitempty"`
@@ -625,23 +641,29 @@ type RenderedPrompt struct {
 	// SessionID is the optional effective direct or rendered session
 	// identifier supplied to the model client.
 	SessionID string `json:"session_id,omitempty"`
-	// Messages contains rendered messages in definition order.
+	// Messages contains the effective messages in provider order: definition
+	// messages in rendered order followed by any RunRequest.AppendedMessages.
 	Messages []RenderedMessage `json:"messages"`
 }
 
-// RenderedMessage is a rendered chat message and has a stable JSON
+// RenderedMessage is a prepared, provider-bound text chat message and has a
-// representation.
+// stable JSON representation. At a request boundary its role is trimmed and
+// lowercased, then must be one of [RoleDeveloper], [RoleSystem], [RoleUser],
+// or [RoleAssistant].
 type RenderedMessage struct {
-	// Role is the definition-supplied chat role.
+	// Role is the provider-bound chat role.
 	Role string `json:"role"`
-	// Content is the rendered message text.
+	// Content is the provider-bound message text. Request-supplied content must
+	// be valid UTF-8 and may be empty or whitespace-only.
 	Content string `json:"content"`
 	// CacheControl is optional provider cache metadata.
 	CacheControl *CacheControl `json:"cache_control,omitempty"`
 }
 
 // CacheControl describes provider cache metadata attached to prompt content
-// and has a stable JSON representation.
+// and has a stable JSON representation. At a request boundary Type and TTL
+// must be valid UTF-8 and are trimmed. Type must be [CacheControlEphemeral],
+// and TTL must be empty or "1h"; other values make the request invalid.
 type CacheControl struct {
 	// Type identifies the cache behavior.
 	Type CacheControlType `json:"type"`