Update documentation and testing policies and add a migration plan to cleanly separate the scriptorium CLI from the promptkit internals

.gitignore

@@ -1,6 +1,5 @@
 # ---> Codex
 .codex
-AGENTS.md
 
 # ---> Go
 # If you prefer the allow list template instead of the deny list, see community template:

.woodpecker/release.yml

@@ -28,10 +28,6 @@ steps:
 
         build_binary linux amd64 ""
         build_binary linux arm64 ""
-        build_binary darwin amd64 ""
-        build_binary darwin arm64 ""
-        build_binary windows amd64 ".exe"
-        build_binary windows arm64 ".exe"
 
   - name: publish-release
     image: woodpeckerci/plugin-release

AGENTS.md

@@ -0,0 +1 @@
+Please review `docs/development.md` for initial orientation in this repository and follow its task-specific reading guide.

LICENSE

@@ -1,4 +1,4 @@
-Copyright (c) 2026 eric. 
+Copyright (c) 2026 Eric Rakestraw. 
 
 Redistribution and use in source and binary forms, with or without modification, are permitted provided that the following conditions are met:
 

README.md

@@ -1,449 +1,45 @@
 # scriptorium
 
-Scriptorium is a generic prompt execution engine.
+Scriptorium is a narrow prompt-execution application for rendering prompt
+requests, running them against OpenAI-compatible chat-completions endpoints, and
+serving the same run workflow over HTTP.
 
-It takes:
-- a prompt definition
-- a selected or default execution profile
-- named input artifacts
-- template variables
-- optional runtime overrides
+It keeps prompt definitions, execution profiles, schemas, and input artifacts as
+separate files so prompts can be reviewed and reused without baking model
+runtime settings into application code.
 
-It returns:
-- for `run`: generated artifact, validation result, metadata
-- for `render`: prepared/rendered prompt data (no model output)
+## Quickstart
 
-## Prompt vs Profile
+From the repository root:
 
-Scriptorium separates **what** to do (Prompt) from **how** to do it (Profile).
-
-### Prompt Definition
-Defines the task logic and output contract.
-- Task description and version.
-- Message templates (system, user, etc.).
-- Required and optional input artifacts.
-- Output format and validation rules.
-- Repair settings for structured output.
-- Optional `default_profile` for convenience.
-
-### Execution Profile
-Defines the runtime environment and model settings.
-- LLM endpoint (URL).
-- Model name.
-- Generation parameters: `temperature`, `max_tokens`, `top_p`.
-- Runtime settings: `timeout`, `reasoning_effort`.
-- API key source via `api_key_env`.
-
-Callers can explicitly provide a `profile_id` to override the prompt's `default_profile`.
-
-## Precedence
-
-Scriptorium uses two precedence layers:
-
-### Application Configuration Precedence
-
-For application-level adapter settings (for example prompt/profile/schema directories, server address, and render output default), precedence is:
-
-1. **CLI Flags**
-2. **`config.yml`**
-3. **Built-in application defaults**
-
-Application config loading behavior:
-- Default config path: `/etc/scriptorium/config.yml`
-- Override path: `--config <PATH>` (supported by `run`, `render`, and `serve`)
-- If `--config` is provided, the file must exist and be valid.
-- If `--config` is omitted, missing `/etc/scriptorium/config.yml` is allowed.
-
-### Runtime Model Precedence
-
-When resolving runtime model settings, Scriptorium follows this precedence model (highest to lowest):
-
-1. **Runtime Overrides**: Provided via CLI flags or HTTP request `model` object.
-2. **Execution Profile**: Settings defined in the selected profile.
-3. **Application Defaults**: Built-in fallback values.
-
-### Profile Selection Logic
-The engine determines which profile to use in this order:
-1. Explicit `profile_id` (via `--profile` or HTTP request).
-2. The `default_profile` named in the Prompt Definition.
-3. Error: If neither is provided and no default exists.
-
-## API Key Policy
-
-To ensure security, Scriptorium does not support raw API keys in configuration files, CLI arguments, or HTTP requests.
-
-- **`api_key_env`**: Profiles and overrides specify the name of an environment variable (e.g., `SCRIPTORIUM_API_KEY`).
-- **Runtime Resolution**: The value of the environment variable is read directly from the process environment at runtime.
-- **Zero Leakage**: API key values are never included in metadata, logs, or response bodies.
-
-## CLI Usage
-
-Available CLI commands:
-- `scriptorium run`
-- `scriptorium render`
-- `scriptorium serve`
-
-All commands accept `--config <PATH>`.
-
-`prompt_dir` and `profile_dir` may be supplied by CLI flags or `config.yml`:
-- `--prompt-dir` or `config.yml` `prompt_dir`
-- `--profile-dir` or `config.yml` `profile_dir`
-
-`schema_dir` and `serve` `addr` may also be supplied by `config.yml` where applicable:
-- `--schema-dir` or `config.yml` `schema_dir`
-- `--addr` or `config.yml` `server.addr`
-
-### `scriptorium run`
-
-Runs a single prompt execution.
-
-**Required Flags:**
-- `--prompt`: The prompt ID to execute.
-- `--input`: Input mapping `name=path` (repeatable).
-
-**Required Effective Settings:**
-- Prompt directory: `--prompt-dir` or `config.yml` `prompt_dir`
-- Profile directory: `--profile-dir` or `config.yml` `profile_dir`
-
-**Optional Flags:**
-- `--config`: Application config file path. Default discovery path is `/etc/scriptorium/config.yml`.
-- `--prompt-dir`: Override prompt directory from config.
-- `--profile-dir`: Override profile directory from config.
-- `--profile`: Override the prompt's default profile.
-- `--var`: Template variable `name=value` (repeatable).
-- `--out`: Write output to a file instead of stdout.
-- `--llm-base-url`: Override endpoint.
-- `--model`: Override model name.
-- `--api-key-env`: Override API key environment variable name.
-- `--temperature`: Override temperature.
-- `--max-tokens`: Override max tokens.
-- `--top-p`: Override top_p.
-- `--timeout`: Override request timeout (e.g., `30s`, `1m`).
-- `--schema-dir`: Base directory for validation schemas.
-
-**Examples:**
-
-Using `config.yml` for prompt/profile directories:
 ```bash
-scriptorium run \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Overriding config directories explicitly:
-```bash
-scriptorium run \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Overriding the profile:
-```bash
-scriptorium run \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --profile local-quality \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Overriding model and runtime values:
-```bash
-scriptorium run \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --model gpt-4o \
-  --temperature 0.7 \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Using a local OpenAI-compatible vLLM endpoint:
-```bash
-scriptorium run \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --llm-base-url http://localhost:8000/v1 \
-  --model meta-llama-3-8b \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-### `scriptorium render`
-
-Prepares and renders a prompt without calling the LLM.
-
-`render` uses the same prompt/profile/input/variable/runtime override resolution as `run`:
-- Profile selection precedence: `--profile` -> prompt `default_profile` -> error.
-- Runtime precedence: CLI runtime overrides -> selected profile -> built-in defaults.
-
-`render` is useful for debugging:
-- prompt template rendering
-- input mappings
-- selected profile behavior
-- runtime override behavior
-
-`render` does not:
-- call the LLM
-- validate model output
-- perform repair
-- expose resolved API key values
-
-It may include `api_key_env` names where relevant.
-
-**Required Flags:**
-- `--prompt`: The prompt ID to render.
-- `--input`: Input mapping `name=path` (repeatable).
-
-**Required Effective Settings:**
-- Prompt directory: `--prompt-dir` or `config.yml` `prompt_dir`
-- Profile directory: `--profile-dir` or `config.yml` `profile_dir`
-
-**Optional Flags:**
-- `--config`: Application config file path. Default discovery path is `/etc/scriptorium/config.yml`.
-- `--prompt-dir`: Override prompt directory from config.
-- `--profile-dir`: Override profile directory from config.
-- `--profile`: Override the prompt's default profile.
-- `--var`: Template variable `name=value` (repeatable).
-- `--out`: Write output to a file instead of stdout.
-- `--format`: Render output format (`text` or `json`). Default: `text`.
-- `--llm-base-url`: Runtime override for endpoint.
-- `--model`: Runtime override for model name.
-- `--api-key-env`: Runtime override for API key environment variable name.
-- `--temperature`: Runtime override for temperature.
-- `--max-tokens`: Runtime override for max tokens.
-- `--top-p`: Runtime override for top_p.
-- `--timeout`: Runtime override for timeout (e.g., `30s`, `1m`).
-
-**Render Output Formats:**
-- `text`: Human-readable output (default).
-- `json`: Machine-readable structured output.
-
-Render formatting is modular; additional output formats can be added later without changing prepare/run core logic.
-
-**Examples:**
-
-Default text output using `config.yml` directories:
-```bash
-scriptorium render \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Explicit config path:
-```bash
-scriptorium render \
+go run ./cmd/scriptorium render \
   --config ./examples/config.yml \
   --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Explicit directory overrides:
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
-
-Explicit JSON output:
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
   --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
   --format json
 ```
 
-Using prompt `default_profile` (omit `--profile`):
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md
-```
+This command renders the prepared prompt and effective runtime settings without calling an LLM.
 
-Overriding profile selection:
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --profile local-quality \
-  --input transcript=./examples/fixtures/transcript.md
-```
+## Documentation
 
-Overriding runtime settings:
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md \
-  --llm-base-url http://localhost:8000/v1 \
-  --model gpt-4o-mini \
-  --temperature 0.2 \
-  --max-tokens 800 \
-  --top-p 1.0 \
-  --timeout 45s
-```
-
-Writing rendered output to a file:
-```bash
-scriptorium render \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --prompt generic.markdown_summary \
-  --input transcript=./examples/fixtures/transcript.md \
-  --format text \
-  --out ./rendered_prompt.txt
-```
-
-### `scriptorium serve`
-
-Starts the HTTP API.
-
-**Required Effective Settings:**
-- Prompt directory: `--prompt-dir` or `config.yml` `prompt_dir`
-- Profile directory: `--profile-dir` or `config.yml` `profile_dir`
-
-**Optional Flags:**
-- `--config`: Application config file path. Default discovery path is `/etc/scriptorium/config.yml`.
-- `--addr`: Listen address (default `:8080`).
-- `--schema-dir`: Base directory for validation schemas.
-
-**Examples:**
-
-Using `config.yml`:
-```bash
-scriptorium serve
-```
-
-Overriding config for local use:
-```bash
-scriptorium serve \
-  --prompt-dir ./prompts \
-  --profile-dir ./profiles \
-  --addr :9090
-```
-
-## HTTP API
-
-### `POST /v1/runs`
-
-Executes a prompt. No built-in authentication is provided; deploy behind a trusted gateway.
-
-**Request Body:**
-```json
-{
-  "prompt_id": "generic.structured_events",
-  "profile_id": "local-quality",
-  "include_raw_output": false,
-  "inputs": {
-    "transcript": {"type": "file", "uri": "./examples/fixtures/transcript.md"}
-  },
-  "vars": {
-    "session_date": "2026-05-04"
-  },
-  "model": {
-    "endpoint": "http://localhost:8000/v1",
-    "model": "gpt-4o-mini",
-    "temperature": 0.0
-  }
-}
-```
-
-`profile_id` is optional. If omitted, Scriptorium uses the prompt's `default_profile`. If neither is available, the run fails.
-
-**Response:**
-Returns a `200 OK` with the generated artifact, validation results, and metadata including the `prompt_id` and the `selected_profile_id`.
-
-**Validation Failures:**
-If the model output fails validation (e.g., invalid JSON), the API returns `200 OK` with `validation.status = "failed"`.
-
-**Raw Output Exposure:**
-- `raw_model_output` is omitted by default.
-- Set `include_raw_output: true` in the request to include it in the response.
-- Raw output is preserved internally in run results regardless of HTTP exposure.
-
-## Prompt Definition Authoring
-
-Prompts are defined in YAML.
-
-### Canonical Shape
-```yaml
-id: generic.structured_events
-version: "1.0.0"
-description: "Extracts structured events from a transcript"
-default_profile: local-quality
-
-inputs:
-  - name: transcript
-    required: true
-    content_type: text/markdown
-    description: "The raw session transcript"
-  - name: glossary
-    required: false
-    content_type: application/yaml
-    description: "Optional glossary terms"
-
-messages:
-  - role: system
-    content: "You are a helpful assistant."
-  - role: user
-    content_file: messages/extract_events.tmpl
-
-output:
-  format: json
-  validation_mode: json_schema
-  schema_path: structured_events.schema.json
-  repair_attempts: 2
-```
-
-**Key Features:**
-- **Inline vs File**: Use `content` for short prompts or `content_file` for larger templates. Exactly one must be set per message.
-- **Path Resolution**: `content_file` paths are resolved relative to the prompt YAML file.
-- **Inputs**: Mark inputs as `required` to ensure the runner fails early if they are missing.
-- **Input Metadata**: `content_type` is currently descriptive metadata and not enforced yet.
-- **Validation**: Support `none`, `basic`, `json`, and `json_schema`.
-- **Repair**: `repair_attempts` enables bounded retries to fix structured output.
-
-## Execution Profile Authoring
-
-Profiles are defined in YAML.
-
-### Canonical Shape
-```yaml
-id: local-quality
-endpoint: http://localhost:8000/v1
-model: gpt-4o
-temperature: 0.0
-max_tokens: 4096
-top_p: 1.0
-timeout_seconds: 300
-reasoning_effort: high
-api_key_env: SCRIPTORIUM_API_KEY
-```
-
-**Constraints:**
-- **No Raw Keys**: Do not include actual API keys. Only specify the environment variable name in `api_key_env`.
-- **Local Profiles**: For local endpoints that don't require auth, `api_key_env` can be omitted.
+- [CLI reference](docs/cli.md)
+- [Configuration reference](docs/config.md)
+- [HTTP API reference](docs/api.md)
+- [Operations guide](docs/operations.md)
+- [Troubleshooting](docs/troubleshooting.md)
+- [Consumer integration overview](docs/consumers/api.md)
+- [Go library package](docs/consumers/pkg-scriptorium.md)
+- [Subprocess integration](docs/integrations/subprocess.md)
+- [OpenAI-compatible chat integration](docs/integrations/openai-compatible-chat.md)
+- [Architecture policy](docs/policy/architecture.md)
 
 ## Examples
 
-- **Prompt Definitions**: `prompts/`
-- **Execution Profiles**: `profiles/`
-- **Schemas**: `schemas/`
-- **Fixtures**: `examples/fixtures/`
-
-## Build and Test
-
-```bash
-go build -o scriptorium ./cmd/scriptorium
-go test ./...
-```
+- `examples/config.yml`
+- `examples/config.full.yml`
+- `examples/render-markdown-summary.sh`
+- `examples/http-run.json`
+- `examples/go-library/prepare`

architecture.md

@@ -1,483 +0,0 @@
-# Scriptorium Architecture
-
-## 1. Purpose and Non-Goals
-
-Scriptorium is a prompt-definition execution engine.
-
-It accepts named input artifacts, renders prompt templates, calls an LLM, validates output, optionally performs bounded structured-output repair, and returns an artifact with metadata.
-
-Scriptorium also supports rendering/preparing a prompt without calling an LLM. This allows users to inspect the fully rendered prompt messages and effective runtime settings before executing a run.
-
-Scriptorium is not an orchestrator. It must not own transcription, transcript merge/polish steps, notifications, or cross-step workflow control.
-
-For the motivating D&D workflow:
-- Narratio orchestrates.
-- WhisperX transcribes.
-- Seriatim merges transcripts.
-- Audita polishes transcripts.
-- Scriptorium generates final artifacts from prepared inputs.
-
-Core Go code remains generic.
-
-## 2. Current Architecture
-
-Scriptorium uses a ports-and-adapters architecture to decouple the core execution logic from external dependencies.
-
-### Package Responsibilities
-
-- `cmd/scriptorium`: Binary entrypoint for CLI and HTTP server.
-- `internal/domain`: Core domain contracts, including `PromptDefinition`, `ExecutionProfile`, `PreparedRun`, `RunResult`, and related metadata.
-- `internal/usecase`: `Runner` use case logic, including prompt preparation, profile selection, runtime override resolution, full run execution, validation, and bounded repair.
-- `internal/config`: Application-level config model/loader for adapter settings (for example prompt/profile/schema directories, server address, and render format default).
-- `internal/promptdef`: Repository for loading and validating Prompt Definitions from the filesystem.
-- `internal/profile`: Repository for loading Execution Profiles from the filesystem.
-- `internal/artifact`: Input artifact resolution (`inline`, `file`).
-- `internal/prompt`: Template rendering via Go templates.
-- `internal/llm`: Provider-neutral client interface and OpenAI-compatible HTTP adapter.
-- `internal/validate`: Output validation implementation (`none/basic/json/json_schema`).
-- `internal/adapter/cli`: CLI flag parsing, command dispatch, and output handling.
-- `internal/adapter/http`: HTTP request/response mapping.
-- `internal/format` or equivalent: Formatting of prepared/rendered prompt output for CLI or other adapters, if formatting grows beyond simple CLI-local helpers.
-
-Exact package names may evolve, but the architectural boundaries should remain stable.
-
-### Application Configuration
-
-`config.yml` is adapter/application setup, not domain logic.
-
-Application config is intended for application-level settings such as:
-- `prompt_dir`
-- `profile_dir`
-- `schema_dir`
-- `server.addr`
-- `defaults.render_format`
-
-Application config precedence is:
-1. CLI flags
-2. `config.yml`
-3. Built-in application defaults
-
-Runtime model settings are intentionally separate:
-- Execution profiles and runtime overrides continue to own endpoint/model/runtime behavior.
-- `config.yml` does not replace execution profiles.
-
-The core use case (`Runner.Prepare`/`Runner.Run`) does not need to know whether adapter-level settings came from CLI flags or `config.yml`; it receives resolved dependencies and requests from adapters.
-
-## 3. Core Execution Model
-
-Scriptorium has two closely related execution paths:
-
-1. Prepare/render path.
-2. Full run path.
-
-The full run path should reuse the prepare path rather than duplicating its logic.
-
-### 3.1 Prepare / Render Data Flow
-
-The prepare path should be represented in the use case layer, preferably as `Runner.Prepare(ctx, RunRequest)` or an equivalent method.
-
-It executes all pre-LLM work:
-
-1. **Validate Request**: Ensure the request includes a prompt ID and all minimum required fields.
-2. **Load Prompt Definition**: Retrieve the `PromptDefinition` by ID from the prompt repository.
-3. **Select Profile**: Determine the `profile_id` using this precedence:
-   - Explicit `profile_id` in `RunRequest`.
-   - `default_profile` specified in the `PromptDefinition`.
-   - Error if neither is available.
-4. **Load Execution Profile**: Retrieve the `ExecutionProfile` from the profile repository.
-5. **Resolve Runtime Overrides**: Merge settings based on precedence, highest to lowest:
-   - Runtime overrides from CLI flags or HTTP request `model` object.
-   - Execution Profile settings.
-   - Built-in application defaults.
-6. **Resolve Artifacts**: Load all named input artifacts defined in the request.
-7. **Render Prompt**: Apply template variables and input artifacts to the prompt templates.
-8. **Compute Metadata**: Compute hashes, selected profile ID, prompt ID/version, effective runtime settings, input hashes, rendered prompt hash, and timing information as appropriate.
-9. **Return PreparedRun**: Return a `PreparedRun` containing the rendered messages, effective runtime settings, resolved metadata, and input/prompt hashes.
-
-The prepare path must not call the LLM.
-
-The prepare path must not validate model output, because there is no model output.
-
-The prepare path must not perform structured-output repair, because repair only applies after model output exists.
-
-The prepare path should not resolve or expose raw API key values. It may include the selected `api_key_env` name in effective runtime settings or metadata, but never the environment variable value.
-
-### 3.2 Full Run Data Flow
-
-The `Runner.Run(ctx, RunRequest)` flow should reuse the prepare path:
-
-1. **Prepare**: Call the shared prepare flow to load the prompt, select the profile, resolve artifacts, render prompt messages, and compute pre-run metadata.
-2. **Call LLM**: Execute the generation request using the effective runtime settings from the prepared run.
-3. **Build Output Artifact**: Convert the model response into the configured output artifact.
-4. **Validate Output**:
-   - Validate the model output against the prompt definition's output contract.
-   - Validation content failures remain successful run results with `validation.status=failed`.
-   - Validator runtime/config errors are run errors.
-5. **Repair If Configured**:
-   - If structured validation fails and `repair_attempts > 0`, perform bounded repair attempts.
-   - Re-validate after each repair attempt.
-   - Repair loops must remain strictly bounded.
-6. **Return RunResult**: Produce a `RunResult` containing the final artifact, validation status, raw model output, usage information, and metadata.
-
-`Runner.Run` should not duplicate profile selection, artifact resolution, or prompt rendering logic that already exists in `Runner.Prepare`.
-
-## 4. Domain Model
-
-Key domain types:
-
-- `PromptDefinition`: Defines the "what" of the task: templates, inputs, output contract, validation settings, repair settings, and optional `default_profile`.
-- `ExecutionProfile`: Defines the "how" of execution: endpoint, model, generation parameters, timeout, reasoning effort, and `api_key_env`.
-- `RunRequest`: The intent to execute or prepare a prompt, including `prompt_id`, optional `profile_id`, inputs, variables, and optional runtime overrides.
-- `PreparedRun`: The result of the prepare/render phase. Contains rendered messages, effective runtime settings, selected profile ID, prompt metadata, input hashes, prompt hash, and other pre-LLM metadata.
-- `RunResult`: The result of a full run. Contains the generated `Artifact`, `ValidationResult`, raw model output, token usage, and `RunMetadata`.
-- `RunMetadata`: Detailed tracing information, including prompt ID/version, selected profile ID, effective model parameters, usage tokens, hashes, timestamps, validation status, and repair attempts where applicable.
-- `RenderedPrompt`: Provider-neutral rendered prompt structure.
-- `RenderedMessage`: Provider-neutral rendered message with role and content.
-- `ArtifactRef`: A reference to an input artifact, such as `file` or `inline`.
-- `Artifact`: Loaded artifact content with name, content type, body, source URI, size, and hash.
-- `ValidationResult`: Validation status and details for full runs.
-
-`PreparedRun` should be serializable for JSON output and should also be representable in a human-readable text format.
-
-## 5. Interfaces and Adapters
-
-### Primary Ports
-
-- `promptdef.Repository`: Lookup for prompt definitions.
-- `profile.Repository`: Lookup for execution profiles.
-- `artifact.Reader`: Loading of artifact content.
-- `prompt.Renderer`: Template rendering.
-- `llm.Client`: Model generation.
-- `validate.Validator`: Output validation.
-- `format.PreparedRunFormatter` or equivalent: Optional formatting abstraction for rendered/prepared output.
-
-### Current Adapters
-
-- **Repositories**: Filesystem YAML loaders for both prompts and profiles.
-- **Artifact Reader**: Composite reader supporting `file` and `inline`.
-- **Prompt Renderer**: Go templates with a custom `input` helper.
-- **LLM Client**: OpenAI-compatible `/chat/completions` over HTTP.
-- **Validator**: Standard validator supporting `none`, `basic`, `json`, and `json_schema`.
-- **CLI Adapter**: Supports `run`, `render`, and `serve`.
-- **HTTP Adapter**: Supports full run execution through `POST /v1/runs`.
-
-## 6. Public Contracts
-
-### CLI
-
-Scriptorium should expose at least these commands:
-
-- `scriptorium run`: Executes a prompt by preparing it, calling the LLM, validating output, optionally repairing structured output, and returning an artifact.
-- `scriptorium render`: Prepares and renders a prompt without calling the LLM.
-- `scriptorium serve`: Starts the HTTP API using infrastructure-only flags.
-
-### `scriptorium run`
-
-`run` uses flags such as:
-
-- `--prompt-dir`
-- `--profile-dir`
-- `--prompt`
-- `--profile`
-- `--input`
-- `--var`
-- `--out`
-- runtime overrides such as `--model`, `--llm-base-url`, `--temperature`, `--max-tokens`, `--top-p`, `--timeout`, and `--api-key-env` if supported.
-
-`run` should produce the generated artifact as its primary output.
-
-### `scriptorium render`
-
-`render` prepares and renders a prompt without calling an LLM.
-
-It should use the same prompt/profile/input/variable/runtime override flags as `run` where applicable:
-
-- `--prompt-dir`
-- `--profile-dir`
-- `--prompt`
-- `--profile`
-- `--input`
-- `--var`
-- runtime overrides such as `--model`, `--llm-base-url`, `--temperature`, `--max-tokens`, `--top-p`, `--timeout`, and `--api-key-env` if supported.
-- `--format`, with initial support for `text` and `json`.
-
-Default render output format should be `text`.
-
-`render` must not call the LLM.
-
-`render` should show the same rendered messages and effective runtime settings that `run` would use.
-
-`render` should include enough information to debug:
-
-- prompt ID
-- prompt version
-- selected profile ID
-- effective runtime settings
-- input hashes
-- prompt hash
-- rendered messages
-
-`render` must not include resolved API key values.
-
-`render` may include the `api_key_env` name.
-
-### Render Output Formats
-
-Initial render output formats:
-
-- `text`: Human-readable default format.
-- `json`: Machine-readable structured representation of the prepared run.
-
-Additional formats, such as `markdown`, may be added later.
-
-Render output formatting should be modular. Adding a new output format should not require changing the prepare/run core logic.
-
-The output formatting layer should consume a `PreparedRun` and produce bytes or text for the adapter. It should not reload prompts, re-resolve artifacts, re-render templates, call the LLM, or perform validation.
-
-### `scriptorium serve`
-
-`serve` starts the HTTP API.
-
-It should use infrastructure-only flags such as:
-
-- `--addr`
-- `--prompt-dir`
-- `--profile-dir`
-- `--schema-dir`
-
-`serve` should not introduce a server-level model/runtime precedence layer unless explicitly documented and intentionally implemented.
-
-### HTTP API
-
-Current HTTP API:
-
-- `POST /v1/runs`: Accepts `RunRequest` JSON and returns `RunResponse` JSON. No built-in auth.
-
-Request may include runtime overrides under `model` and an `include_raw_output` boolean.
-
-`raw_model_output` is exposed only when explicitly requested with `include_raw_output=true`.
-
-A future HTTP prepare/render endpoint may be added, such as `POST /v1/renders` or `POST /v1/runs/prepare`, but the initial render feature may be CLI-only. If added later, it should call the same usecase-level prepare path as `scriptorium render`.
-
-### YAML Shapes
-
-Prompt YAML includes:
-
-- `id`
-- `version`
-- optional `default_profile`
-- `inputs`
-- `messages`
-- `output`
-
-Inputs support:
-
-- `name`
-- `required`
-- optional `content_type`
-- `description`
-
-Messages require:
-
-- `role`
-- exactly one of `content` or `content_file`
-
-`content_file` resolves relative to the prompt YAML location.
-
-Profile YAML includes:
-
-- `id`
-- `endpoint`
-- `model`
-- generation parameters
-- timeout settings
-- reasoning settings
-- `api_key_env`
-
-Prompt content must not appear in profile YAML.
-
-Model/runtime/API-key settings must not appear in prompt YAML, except that prompt YAML may specify `default_profile`.
-
-## 7. Render Feature Design
-
-The render feature is a first-class use case, not a CLI-only shortcut.
-
-### Goals
-
-The render feature should help users:
-
-- inspect fully rendered prompt messages
-- debug missing inputs
-- verify template variable substitution
-- verify selected profile resolution
-- verify runtime override precedence
-- verify file-backed prompt loading
-- inspect input hashes and prompt hashes
-- prepare for future token budgeting and prompt-size inspection
-
-### Non-Goals
-
-The render feature should not:
-
-- call an LLM
-- validate model output
-- repair structured output
-- resolve or print API key values
-- mutate artifacts
-- save outputs to artifact storage unless a future explicit output option is added
-- become an orchestration step manager
-
-### Usecase Shape
-
-The preferred usecase shape is:
-
-- `Runner.Prepare(ctx, RunRequest) (*PreparedRun, error)`
-- `Runner.Run(ctx, RunRequest) (*RunResult, error)`
-
-`Runner.Run` should call `Runner.Prepare`.
-
-The prepare flow should be the only implementation of:
-
-- prompt loading
-- profile selection
-- runtime override resolution
-- artifact resolution
-- prompt rendering
-- pre-run metadata/hash calculation
-
-### CLI Shape
-
-The preferred command name is `render`.
-
-The command should support:
-
-- `--format text`
-- `--format json`
-
-Default format:
-
-- `text`
-
-Unknown formats should produce a clear error.
-
-Formatting should be centralized through a small formatter registry, strategy, switch, or interface so new formats can be added without modifying usecase logic.
-
-### Text Output Expectations
-
-Text output should be optimized for human inspection.
-
-It should include, at minimum:
-
-- prompt ID and version
-- selected profile ID
-- model name
-- endpoint
-- effective generation settings
-- input hashes
-- rendered prompt hash
-- rendered messages grouped by role
-
-Text output should be readable and deterministic enough for tests.
-
-It should not include raw API key values.
-
-### JSON Output Expectations
-
-JSON output should be a structured representation of `PreparedRun` or a DTO derived from it.
-
-It should include, at minimum:
-
-- prompt ID and version
-- selected profile ID
-- effective runtime settings
-- input hashes
-- rendered prompt hash
-- rendered messages
-
-JSON output should not include raw API key values.
-
-JSON output should remain stable enough to be useful for automation and integration tests.
-
-## 8. Guardrails
-
-- **Separation of Concerns**: Prompt content must not belong in execution profiles; model/API settings must not belong in prompt definitions.
-- **Security**: Raw API keys are unsupported in all configuration and transport layers. Only `api_key_env` is used.
-- **Secret Handling**: Resolved API key values must never appear in rendered output, metadata, logs, HTTP responses, or CLI output.
-- **Path Resolution**: `content_file` paths in prompt definitions resolve relative to the prompt YAML file.
-- **Integrity**: No silent prompt truncation or omission of content.
-- **Reliability**: Repair loops are strictly bounded by `repair_attempts`.
-- **No Orchestration Creep**: Scriptorium prepares and executes a single prompt request. It does not coordinate multi-stage workflows.
-- **Render Reuse**: The full run path must reuse the prepare/render path to avoid divergent behavior.
-- **Formatter Isolation**: Render output formatters must not perform usecase work. They only format a completed `PreparedRun`.
-
-## 9. Testing Strategy
-
-Tests should protect both the run path and the prepare/render path.
-
-### Prepare / Render Tests
-
-Add tests for:
-
-- preparing a prompt with explicit profile selection
-- preparing a prompt using `default_profile`
-- failing when no explicit profile and no `default_profile` exist
-- runtime overrides beating profile values
-- profile values beating application defaults
-- file-backed prompt bodies rendering correctly
-- required inputs failing when missing
-- optional inputs being absent when not referenced
-- unknown input references failing
-- input hashes being included
-- rendered prompt hash being included
-- effective runtime settings being included
-- `api_key_env` name being included where appropriate
-- resolved API key values never appearing in `PreparedRun`
-- prepare path not calling the LLM
-
-### CLI Render Tests
-
-Add tests for:
-
-- `scriptorium render` mapping flags into `RunRequest`
-- default text output
-- explicit `--format text`
-- explicit `--format json`
-- unknown format failure
-- text output includes prompt/profile/messages
-- JSON output includes prompt/profile/messages
-- rendered output never includes resolved API key values
-
-### Run Reuse Tests
-
-Add tests proving:
-
-- `Runner.Run` reuses prepare behavior
-- run and render resolve the same prompt/profile/runtime settings for equivalent inputs
-- run still validates output
-- run still performs bounded repair where configured
-
-### Existing Tests
-
-Continue testing:
-
-- prompt definition loading
-- execution profile loading
-- artifact loading
-- prompt rendering
-- LLM adapter behavior
-- validation behavior
-- HTTP request/response mapping
-
-## 10. Extension Points
-
-Future work should remain grounded in the current architecture:
-
-- **Artifacts**: Add S3 artifact references via a new `artifact.Reader`.
-- **LLM**: Implement additional provider adapters, such as Anthropic or Google.
-- **Execution**: Add token budgeting, streaming generation, and batch execution capabilities.
-- **Prepare/Render**: Add token estimates, prompt-size summaries, or additional render output formats.
-- **Repositories**: Implement database-backed repositories for prompts and profiles.
-- **Profiles**: Support more granular profile versioning and environment-specific profiles.
-- **HTTP**: Add an HTTP prepare/render endpoint if Narratio or another caller needs it.
-
-Future render formats should plug into the formatter layer and should not require changes to the usecase layer.

convert.go

@@ -0,0 +1,406 @@
+package scriptorium
+
+import (
+	"reflect"
+
+	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+)
+
+func toDomainRunRequest(req RunRequest) (domain.RunRequest, error) {
+	execution, err := toDomainExecutionTargetOverride(req.Execution)
+	if err != nil {
+		return domain.RunRequest{}, err
+	}
+	return domain.RunRequest{
+		PromptID:      req.PromptID,
+		PromptVersion: req.PromptVersion,
+		ProfileID:     req.ProfileID,
+		APIKey:        req.APIKey,
+		Inputs:        toDomainArtifactRefMap(req.Inputs),
+		Vars:          copyStringMap(req.Vars),
+		Execution:     execution,
+		Validation:    toDomainOutputContractPtr(req.Validation),
+		Metadata:      copyStringMap(req.Metadata),
+	}, nil
+}
+
+func fromDomainPreparedRun(prepared *domain.PreparedRun) *PreparedRun {
+	if prepared == nil {
+		return nil
+	}
+	return &PreparedRun{
+		PromptID:             prepared.PromptID,
+		PromptVersion:        prepared.PromptVersion,
+		PromptHash:           prepared.PromptHash,
+		SelectedProfileID:    prepared.SelectedProfileID,
+		EffectiveModelParams: fromDomainExecutionTarget(prepared.EffectiveModelParams),
+		OutputContract:       fromDomainOutputContract(prepared.OutputContract),
+		StructuredOutput:     fromDomainStructuredOutputSpec(prepared.StructuredOutput),
+		InputHashes:          copyStringMap(prepared.InputHashes),
+		SessionID:            prepared.SessionID,
+		RenderedPromptHash:   prepared.RenderedPromptHash,
+		Messages:             fromDomainRenderedMessages(prepared.Messages),
+		StartTime:            prepared.StartTime,
+		EndTime:              prepared.EndTime,
+		DurationMS:           prepared.DurationMS,
+	}
+}
+
+func fromDomainRunResult(result *domain.RunResult) *RunResult {
+	if result == nil {
+		return nil
+	}
+	return &RunResult{
+		RunID:                result.RunID,
+		Artifact:             fromDomainArtifact(result.Artifact),
+		RawOutput:            result.RawOutput,
+		Validation:           fromDomainValidationResult(result.Validation),
+		PromptID:             result.PromptID,
+		PromptVersion:        result.PromptVersion,
+		PromptHash:           result.PromptHash,
+		RenderedPromptHash:   result.RenderedPromptHash,
+		SelectedProfileID:    result.SelectedProfileID,
+		ModelName:            result.ModelName,
+		Endpoint:             result.Endpoint,
+		EffectiveModelParams: fromDomainExecutionTarget(result.EffectiveModelParams),
+		InputHashes:          copyStringMap(result.InputHashes),
+		Usage:                fromDomainTokenUsage(result.Usage),
+		StartTime:            result.StartTime,
+		EndTime:              result.EndTime,
+		Duration:             result.Duration,
+	}
+}
+
+func fromDomainGenerateRequest(req domain.GenerateRequest) GenerateRequest {
+	return GenerateRequest{
+		Prompt:           fromDomainRenderedPrompt(req.Prompt),
+		Target:           fromDomainExecutionTarget(req.Target),
+		TargetPresence:   fromDomainExecutionTargetPresence(req.TargetPresence),
+		StructuredOutput: fromDomainStructuredOutputSpec(req.StructuredOutput),
+		APIKey:           req.Target.APIKey,
+	}
+}
+
+func toDomainGenerateResponse(resp *GenerateResponse) *domain.GenerateResponse {
+	if resp == nil {
+		return nil
+	}
+	return &domain.GenerateResponse{
+		Content: resp.Content,
+		Usage:   toDomainTokenUsage(resp.Usage),
+	}
+}
+
+func fromDomainRenderedPrompt(prompt domain.RenderedPrompt) RenderedPrompt {
+	return RenderedPrompt{
+		SessionID: prompt.SessionID,
+		Messages:  fromDomainRenderedMessages(prompt.Messages),
+	}
+}
+
+func toDomainArtifactRefMap(src map[string]ArtifactRef) map[string]domain.ArtifactRef {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]domain.ArtifactRef, len(src))
+	for k, v := range src {
+		out[k] = toDomainArtifactRef(v)
+	}
+	return out
+}
+
+func toDomainArtifactRef(ref ArtifactRef) domain.ArtifactRef {
+	return domain.ArtifactRef{
+		Type: domain.ArtifactRefType(ref.Type),
+		URI:  ref.URI,
+		Body: ref.Body,
+	}
+}
+
+func fromDomainArtifact(artifact domain.Artifact) Artifact {
+	return Artifact{
+		Name:        artifact.Name,
+		ContentType: artifact.ContentType,
+		Body:        copyBytes(artifact.Body),
+		URI:         artifact.URI,
+		Size:        artifact.Size,
+		Hash:        artifact.Hash,
+	}
+}
+
+func toDomainExecutionTargetOverride(override *ExecutionTargetOverride) (*domain.ExecutionTargetOverride, error) {
+	if override == nil {
+		return nil, nil
+	}
+	extraParams, err := copyPublicJSONMap(override.ExtraParams)
+	if err != nil {
+		return nil, err
+	}
+	return &domain.ExecutionTargetOverride{
+		Endpoint:        override.Endpoint,
+		Model:           override.Model,
+		Temperature:     copyFloat64Ptr(override.Temperature),
+		MaxTokens:       copyIntPtr(override.MaxTokens),
+		TopP:            copyFloat64Ptr(override.TopP),
+		TimeoutSeconds:  copyIntPtr(override.TimeoutSeconds),
+		ServiceTier:     override.ServiceTier,
+		ReasoningEffort: override.ReasoningEffort,
+		APIKeyEnv:       override.APIKeyEnv,
+		ExtraParams:     extraParams,
+	}, nil
+}
+
+func fromDomainExecutionTarget(target domain.ExecutionTarget) ExecutionTarget {
+	return ExecutionTarget{
+		Endpoint:        target.Endpoint,
+		Model:           target.Model,
+		Temperature:     target.Temperature,
+		MaxTokens:       target.MaxTokens,
+		TopP:            target.TopP,
+		TimeoutSeconds:  target.TimeoutSeconds,
+		ServiceTier:     target.ServiceTier,
+		ReasoningEffort: target.ReasoningEffort,
+		APIKeyEnv:       target.APIKeyEnv,
+		ExtraParams:     copyAnyMap(target.ExtraParams),
+	}
+}
+
+func fromDomainExecutionTargetPresence(presence domain.ExecutionTargetPresence) ExecutionTargetPresence {
+	return ExecutionTargetPresence{
+		Temperature:    presence.Temperature,
+		MaxTokens:      presence.MaxTokens,
+		TopP:           presence.TopP,
+		TimeoutSeconds: presence.TimeoutSeconds,
+	}
+}
+
+func toDomainOutputContractPtr(contract *OutputContract) *domain.OutputContract {
+	if contract == nil {
+		return nil
+	}
+	out := toDomainOutputContract(*contract)
+	return &out
+}
+
+func toDomainOutputContract(contract OutputContract) domain.OutputContract {
+	return domain.OutputContract{
+		Format:         domain.OutputFormat(contract.Format),
+		ValidationMode: domain.ValidationMode(contract.ValidationMode),
+		SchemaPath:     contract.SchemaPath,
+		RepairAttempts: contract.RepairAttempts,
+	}
+}
+
+func fromDomainOutputContract(contract domain.OutputContract) OutputContract {
+	return OutputContract{
+		Format:         OutputFormat(contract.Format),
+		ValidationMode: ValidationMode(contract.ValidationMode),
+		SchemaPath:     contract.SchemaPath,
+		RepairAttempts: contract.RepairAttempts,
+	}
+}
+
+func fromDomainValidationResult(result domain.ValidationResult) ValidationResult {
+	return ValidationResult{
+		Status:         ValidationStatus(result.Status),
+		Mode:           ValidationMode(result.Mode),
+		Errors:         copyStringSlice(result.Errors),
+		SchemaPath:     result.SchemaPath,
+		RepairAttempts: result.RepairAttempts,
+		IsValid:        result.IsValid,
+	}
+}
+
+func fromDomainTokenUsage(usage domain.TokenUsage) TokenUsage {
+	return TokenUsage{
+		PromptTokens:     usage.PromptTokens,
+		CompletionTokens: usage.CompletionTokens,
+		TotalTokens:      usage.TotalTokens,
+		CachedTokens:     usage.CachedTokens,
+		CacheWriteTokens: usage.CacheWriteTokens,
+	}
+}
+
+func toDomainTokenUsage(usage TokenUsage) domain.TokenUsage {
+	return domain.TokenUsage{
+		PromptTokens:     usage.PromptTokens,
+		CompletionTokens: usage.CompletionTokens,
+		TotalTokens:      usage.TotalTokens,
+		CachedTokens:     usage.CachedTokens,
+		CacheWriteTokens: usage.CacheWriteTokens,
+	}
+}
+
+func fromDomainRenderedMessages(messages []domain.RenderedMessage) []RenderedMessage {
+	if messages == nil {
+		return nil
+	}
+	out := make([]RenderedMessage, len(messages))
+	for i, msg := range messages {
+		out[i] = RenderedMessage{
+			Role:         msg.Role,
+			Content:      msg.Content,
+			CacheControl: fromDomainCacheControl(msg.CacheControl),
+		}
+	}
+	return out
+}
+
+func fromDomainCacheControl(cacheControl *domain.CacheControl) *CacheControl {
+	if cacheControl == nil {
+		return nil
+	}
+	return &CacheControl{
+		Type: CacheControlType(cacheControl.Type),
+		TTL:  cacheControl.TTL,
+	}
+}
+
+func fromDomainStructuredOutputSpec(spec *domain.StructuredOutputSpec) *StructuredOutputSpec {
+	if spec == nil {
+		return nil
+	}
+	out := &StructuredOutputSpec{
+		Type: StructuredOutputType(spec.Type),
+	}
+	if spec.JSONSchema != nil {
+		out.JSONSchema = &StructuredOutputJSONSpec{
+			Name:   spec.JSONSchema.Name,
+			Strict: spec.JSONSchema.Strict,
+			Schema: copyAny(spec.JSONSchema.Schema),
+		}
+	}
+	return out
+}
+
+func copyStringMap(src map[string]string) map[string]string {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]string, len(src))
+	for k, v := range src {
+		out[k] = v
+	}
+	return out
+}
+
+func copyAnyMap(src map[string]any) map[string]any {
+	if src == nil {
+		return nil
+	}
+	out := make(map[string]any, len(src))
+	for k, v := range src {
+		out[k] = copyAny(v)
+	}
+	return out
+}
+
+func copyAny(value any) any {
+	if value == nil {
+		return nil
+	}
+	switch v := value.(type) {
+	case map[string]any:
+		return copyAnyMap(v)
+	case []any:
+		out := make([]any, len(v))
+		for i, item := range v {
+			out[i] = copyAny(item)
+		}
+		return out
+	case []string:
+		return copyStringSlice(v)
+	case []byte:
+		return copyBytes(v)
+	default:
+		return copyReflectValue(reflect.ValueOf(value)).Interface()
+	}
+}
+
+func copyReflectValue(value reflect.Value) reflect.Value {
+	if !value.IsValid() {
+		return value
+	}
+
+	switch value.Kind() {
+	case reflect.Interface:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		copied := copyReflectValue(value.Elem())
+		if copied.IsValid() && copied.Type().AssignableTo(value.Type()) {
+			return copied
+		}
+		out := reflect.New(value.Type()).Elem()
+		out.Set(copied)
+		return out
+	case reflect.Pointer:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.New(value.Type().Elem())
+		out.Elem().Set(copyReflectValue(value.Elem()))
+		return out
+	case reflect.Map:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.MakeMapWithSize(value.Type(), value.Len())
+		iter := value.MapRange()
+		for iter.Next() {
+			out.SetMapIndex(copyReflectValue(iter.Key()), copyReflectValue(iter.Value()))
+		}
+		return out
+	case reflect.Slice:
+		if value.IsNil() {
+			return reflect.Zero(value.Type())
+		}
+		out := reflect.MakeSlice(value.Type(), value.Len(), value.Cap())
+		for i := 0; i < value.Len(); i++ {
+			out.Index(i).Set(copyReflectValue(value.Index(i)))
+		}
+		return out
+	case reflect.Array:
+		out := reflect.New(value.Type()).Elem()
+		for i := 0; i < value.Len(); i++ {
+			out.Index(i).Set(copyReflectValue(value.Index(i)))
+		}
+		return out
+	default:
+		return value
+	}
+}
+
+func copyStringSlice(src []string) []string {
+	if src == nil {
+		return nil
+	}
+	out := make([]string, len(src))
+	copy(out, src)
+	return out
+}
+
+func copyBytes(src []byte) []byte {
+	if src == nil {
+		return nil
+	}
+	out := make([]byte, len(src))
+	copy(out, src)
+	return out
+}
+
+func copyFloat64Ptr(src *float64) *float64 {
+	if src == nil {
+		return nil
+	}
+	v := *src
+	return &v
+}
+
+func copyIntPtr(src *int) *int {
+	if src == nil {
+		return nil
+	}
+	v := *src
+	return &v
+}

docs/api.md

@@ -0,0 +1,297 @@
+# HTTP API Reference
+
+This is the canonical public HTTP contract for Scriptorium.
+
+Implemented route:
+
+- `POST /v1/runs`
+
+For CLI behavior, see [CLI reference](cli.md). For config and prompt/profile
+file formats, see [Configuration reference](config.md).
+
+The maintained request-shape example is `examples/http-run.json`. It requires a
+running `serve` process with an artifact root that can read the referenced
+files, plus a reachable model endpoint for full execution.
+
+## Base URL And Deployment
+
+`scriptorium serve` listens on `server.addr` or `serve --addr`. The default is
+`:8080`.
+
+The route path is always:
+
+```text
+/v1/runs
+```
+
+The HTTP adapter has no built-in authentication or authorization. Deploy it
+behind trusted network and authentication controls.
+
+## Media Types
+
+- Request body: JSON object.
+- Response body: JSON object.
+- Response `Content-Type`: `application/json`.
+
+Requests are decoded as JSON regardless of the request `Content-Type` header.
+There are no shared query parameters.
+
+## Request Limits
+
+HTTP limits are configured through `server.*` config fields or `serve` flags:
+
+- `server.max_request_bytes`: encoded JSON request body limit, including inline input bodies.
+- `server.max_artifact_bytes`: file artifact limit for HTTP `file` input references.
+- `server.max_response_bytes`: encoded JSON response limit, including artifact body and optional raw output.
+
+Each limit defaults to `16777216` bytes. `0` disables that limit.
+
+## `POST /v1/runs`
+
+Runs one prompt request and returns the generated artifact, validation result,
+and metadata.
+
+### Request Body
+
+```json
+{
+  "prompt_id": "generic.markdown_summary",
+  "profile_id": "local-fast",
+  "prompt_version": "1.0.0",
+  "inputs": {
+    "transcript": {
+      "type": "file",
+      "uri": "./examples/fixtures/transcript.md"
+    },
+    "glossary": {
+      "type": "inline",
+      "body": "party:\n  - Rin"
+    }
+  },
+  "vars": {
+    "session_date": "2026-05-04"
+  },
+  "model": {
+    "endpoint": "http://localhost:8000/v1",
+    "model": "gpt-4o-mini",
+    "temperature": 0,
+    "max_tokens": 800,
+    "top_p": 1,
+    "timeout_seconds": 120,
+    "service_tier": "priority",
+    "reasoning_effort": "medium",
+    "api_key_env": "SCRIPTORIUM_API_KEY",
+    "extra_params": {
+      "provider_option": "enabled"
+    }
+  },
+  "include_raw_output": false
+}
+```
+
+Request fields:
+
+| Field | Required | Description |
+| --- | --- | --- |
+| `prompt_id` | yes | Prompt ID. Must not be blank. |
+| `prompt_version` | no | Prompt version filter. |
+| `profile_id` | no | Execution profile ID. If omitted, the prompt must define `default_profile`. |
+| `inputs` | yes | Object mapping prompt input names to input references. Must contain at least one entry. |
+| `vars` | no | Object mapping template variable names to string values. |
+| `model` | no | Runtime model override object. |
+| `include_raw_output` | no | When `true`, include `raw_model_output` in the response. |
+
+Input reference fields:
+
+| Field | Required | Description |
+| --- | --- | --- |
+| `type` | yes | `file` or `inline`. |
+| `uri` | for `file` | File URI/path. |
+| `body` | for `inline` | Inline artifact body. |
+
+HTTP `file` references require `server.artifact_root` or `serve
+--artifact-root`. Relative file URIs resolve against that root. Absolute file
+URIs are accepted only when lexically inside the root. Relative traversal and
+absolute paths outside the root return `400 artifact_not_allowed`.
+
+The containment check is lexical and does not resolve symlinks. Symlinks inside
+the artifact root are followed by the operating system, including symlinks that
+point outside the root. Keep the artifact root narrow and not writable by
+untrusted users.
+
+Model override fields:
+
+| Field | Description |
+| --- | --- |
+| `endpoint` | Runtime endpoint override. |
+| `model` | Runtime model override. |
+| `temperature` | Number in range `0..2`. Explicit `0` is an override. |
+| `max_tokens` | Integer greater than or equal to `0`. Explicit `0` is an override. |
+| `top_p` | Number in range `0..1`. Explicit `0` is an override. |
+| `timeout_seconds` | Integer greater than or equal to `0`. Explicit `0` disables the outbound client timeout. |
+| `service_tier` | Provider-specific request tier. |
+| `reasoning_effort` | Provider-specific reasoning setting. |
+| `api_key_env` | Name of an environment variable containing the API key. |
+| `extra_params` | JSON-compatible provider-specific top-level request fields. |
+
+Raw API-key values are not accepted in HTTP payloads. A field such as
+`api_key` is rejected as unknown JSON.
+
+`extra_params` keys must not be empty and must not collide with reserved
+outbound fields: `model`, `session_id`, `messages`, `temperature`,
+`max_tokens`, `top_p`, `service_tier`, `reasoning_effort`, or
+`response_format`.
+
+### Strict JSON Rules
+
+Request decoding is strict:
+
+- malformed JSON returns `400 invalid_json`
+- unknown request fields return `400 invalid_json`
+- unknown `inputs` item fields return `400 invalid_json`
+- unknown `model` fields return `400 invalid_json`
+- trailing JSON tokens after the request object return `400 invalid_json`
+- request bodies above the configured limit return `413 request_too_large`
+
+### Success Response
+
+Status: `200 OK`
+
+```json
+{
+  "artifact": {
+    "name": "output",
+    "content_type": "text/markdown",
+    "body": "Generated content",
+    "size": 17,
+    "hash": "..."
+  },
+  "validation": {
+    "status": "passed",
+    "mode": "basic",
+    "repair_attempts": 0,
+    "is_valid": true
+  },
+  "metadata": {
+    "run_id": "...",
+    "prompt_id": "generic.markdown_summary",
+    "prompt_version": "1.0.0",
+    "prompt_hash": "...",
+    "rendered_prompt_hash": "...",
+    "selected_profile_id": "local-fast",
+    "model_name": "gpt-4o-mini",
+    "endpoint": "http://localhost:8000/v1",
+    "model_params": {
+      "endpoint": "http://localhost:8000/v1",
+      "model": "gpt-4o-mini",
+      "temperature": 0.2,
+      "max_tokens": 500,
+      "top_p": 1,
+      "timeout_seconds": 90
+    },
+    "input_hashes": {
+      "transcript": "..."
+    },
+    "usage": {
+      "prompt_tokens": 11,
+      "completion_tokens": 22,
+      "total_tokens": 33,
+      "cached_tokens": 0,
+      "cache_write_tokens": 0
+    },
+    "start_time": "2026-05-04T12:00:00Z",
+    "end_time": "2026-05-04T12:00:01Z",
+    "duration_ms": 1000,
+    "validation_mode": "basic",
+    "validation_status": "passed",
+    "repair_attempts_used": 0
+  }
+}
+```
+
+Response fields:
+
+- `artifact`: generated output artifact.
+- `validation`: validation result for the generated artifact.
+- `metadata`: run and effective runtime metadata.
+- `raw_model_output`: omitted unless `include_raw_output` is `true`.
+
+`artifact.uri` is omitted when empty. `validation.errors` and
+`validation.schema_path` are omitted when empty. `model_params.service_tier`,
+`model_params.reasoning_effort`, `model_params.api_key_env`, and
+`model_params.extra_params` are omitted when empty.
+
+`metadata.usage.cached_tokens` and `metadata.usage.cache_write_tokens` are
+always present as numbers. They are `0` when the provider omits compatible cache
+usage fields or reports no cache activity.
+
+### Validation Failure Response
+
+Generated-content validation failures still return `200 OK`.
+
+```json
+{
+  "validation": {
+    "status": "failed",
+    "mode": "json",
+    "errors": ["invalid JSON: ..."],
+    "repair_attempts": 0,
+    "is_valid": false
+  }
+}
+```
+
+The response still includes `artifact` and `metadata`.
+
+## Error Responses
+
+Error body shape:
+
+```json
+{
+  "error": {
+    "code": "invalid_request",
+    "message": "prompt_id is required"
+  }
+}
+```
+
+Current status/code mapping:
+
+| Status | Code | Meaning |
+| --- | --- | --- |
+| `400` | `invalid_json` | Malformed JSON, unknown JSON field, or trailing JSON token. |
+| `400` | `invalid_request` | Missing/invalid request fields or invalid runtime overrides. |
+| `400` | `profile_required` | No `profile_id` and prompt has no `default_profile`. |
+| `400` | `prompt_load_failed` | Prompt definition YAML/contract failed to load. |
+| `400` | `profile_load_failed` | Profile YAML/contract failed to load, including raw `api_key`. |
+| `400` | `artifact_not_allowed` | HTTP file refs are disabled or requested path is outside artifact root. |
+| `400` | `artifact_read_failed` | Input artifact could not be read or input ref was unsupported/invalid. |
+| `400` | `prompt_render_failed` | Prompt template rendering failed. |
+| `400` | `api_key_env_missing` | Selected `api_key_env` variable is unset or empty. |
+| `404` | `not_found` | Route path is unknown. |
+| `404` | `prompt_not_found` | Prompt ID/version was not found. |
+| `404` | `profile_not_found` | Profile ID was not found. |
+| `405` | `method_not_allowed` | Method is not `POST` on `/v1/runs`. |
+| `413` | `request_too_large` | Encoded JSON request body exceeds configured request limit. |
+| `413` | `artifact_too_large` | HTTP file input artifact exceeds configured artifact limit. |
+| `413` | `response_too_large` | Encoded JSON response exceeds configured response limit. |
+| `500` | `validation_runtime_failed` | Validator runtime/schema loading failed. |
+| `500` | `internal_error` | Unclassified server error. |
+| `502` | `llm_failed` | Outbound model request failed. |
+
+HTTP error messages are intentionally concise and do not include sensitive
+internal causes.
+
+## Retry And Idempotency
+
+Scriptorium does not provide idempotency keys, pagination, caching headers, or
+rate limiting.
+
+Clients may retry transport failures or `5xx` responses when their surrounding
+workflow can tolerate another model call. A retry can generate different output
+and incur another provider request.
+
+## Example File
+
+- `examples/http-run.json`

docs/cli.md

@@ -0,0 +1,233 @@
+# CLI Reference
+
+## Shortest Useful Command
+
+```bash
+go run ./cmd/scriptorium render \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml
+```
+
+`render` prepares the prompt, loads input artifacts, resolves the execution
+profile, and prints the prepared request without calling an LLM.
+
+## Command Overview
+
+- `scriptorium run`: prepare a prompt, call the configured LLM, write generated output, and print a run summary.
+- `scriptorium render`: prepare a prompt only; write prepared-run output as `text` or `json`.
+- `scriptorium serve`: start the HTTP server for `POST /v1/runs`.
+
+Canonical related references:
+
+- [Configuration reference](config.md)
+- [HTTP API reference](api.md)
+- [Subprocess integration](integrations/subprocess.md)
+
+## Common Rules
+
+- `--config` is supported by `run`, `render`, and `serve`.
+- Positional arguments are rejected.
+- `run` and `render` require `--prompt`, at least one `--input`, and an effective `prompt_dir`.
+- `serve` requires an effective `prompt_dir`.
+- `profile_dir` is optional. Without it, only built-in profiles are available.
+- If `profile_dir` is set, custom profiles override built-in profiles with the same ID.
+- Prompt cache control, `session_id`, structured output, and provider-specific profile fields are configured in YAML, not with CLI flags.
+
+Config precedence is:
+
+1. built-in defaults
+2. config file values
+3. CLI flags
+
+## Flag Reference
+
+### `scriptorium run`
+
+```bash
+scriptorium run [flags]
+```
+
+Required through flags or config:
+
+- `--prompt-dir <dir>`: prompt definition directory.
+
+Required as flags:
+
+- `--prompt <id>`: prompt ID to execute.
+- `--input name=path`: input file mapping. Repeat or use comma-separated mappings.
+
+Optional flags:
+
+- `--config <path>`: application config file.
+- `--profile-dir <dir>`: custom profile definition directory.
+- `--schema-dir <dir>`: schema base directory for `json_schema` validation.
+- `--profile <id>`: execution profile override. If omitted, the prompt `default_profile` is used.
+- `--var name=value`: template variable mapping. Repeat or use comma-separated mappings.
+- `--out <path>`: write generated artifact body to a file instead of stdout.
+- `--llm-base-url <url>`: runtime endpoint override.
+- `--model <name>`: runtime model override.
+- `--api-key-env <name>`: runtime API-key environment variable name override.
+- `--temperature <float>`: runtime temperature override.
+- `--max-tokens <int>`: runtime max tokens override.
+- `--top-p <float>`: runtime top-p override.
+- `--timeout <duration>`: runtime timeout override using Go duration syntax, such as `30s` or `2m`.
+
+Deprecated aliases:
+
+- `--prompt-id <id>`: alias for `--prompt`.
+- `--profile-id <id>`: alias for `--profile`.
+
+Runtime override notes:
+
+- Omitted numeric override flags preserve the selected profile/default value.
+- Explicit zero values override the selected profile/default value.
+- `--timeout 0s` disables the outbound HTTP client timeout for that request.
+- There is no raw API-key flag; use `--api-key-env`.
+
+### `scriptorium render`
+
+```bash
+scriptorium render [flags]
+```
+
+Required through flags or config:
+
+- `--prompt-dir <dir>`: prompt definition directory.
+
+Required as flags:
+
+- `--prompt <id>`: prompt ID to render.
+- `--input name=path`: input file mapping. Repeat or use comma-separated mappings.
+
+Optional flags:
+
+- `--config <path>`: application config file.
+- `--prompt-dir <dir>`: prompt definition directory.
+- `--profile-dir <dir>`: custom profile definition directory.
+- `--profile <id>`: execution profile override.
+- `--var name=value`: template variable mapping. Repeat or use comma-separated mappings.
+- `--out <path>`: write prepared-run output to a file instead of stdout.
+- `--llm-base-url <url>`: runtime endpoint override for the prepared request.
+- `--model <name>`: runtime model override for the prepared request.
+- `--api-key-env <name>`: runtime API-key environment variable name override.
+- `--temperature <float>`: runtime temperature override.
+- `--max-tokens <int>`: runtime max tokens override.
+- `--top-p <float>`: runtime top-p override.
+- `--timeout <duration>`: runtime timeout override using Go duration syntax.
+- `--format text|json`: prepared-run output format. Defaults to config `defaults.render_format`, then `text`.
+
+Deprecated aliases:
+
+- `--prompt-id <id>`: alias for `--prompt`.
+- `--profile-id <id>`: alias for `--profile`.
+
+Notes:
+
+- `render` resolves profiles, loads schemas for `json_schema` prompts, and validates `api_key_env`.
+- `render` does not accept `--schema-dir`; use config `schema_dir` for render-time schema lookup.
+- `render` does not call the LLM.
+
+### `scriptorium serve`
+
+```bash
+scriptorium serve [flags]
+```
+
+Required through flags or config:
+
+- `--prompt-dir <dir>`: prompt definition directory.
+
+Optional flags:
+
+- `--config <path>`: application config file.
+- `--addr <listen-address>`: HTTP listen address.
+- `--prompt-dir <dir>`: prompt definition directory.
+- `--profile-dir <dir>`: custom profile definition directory.
+- `--schema-dir <dir>`: schema base directory for `json_schema` validation.
+- `--artifact-root <dir>`: base directory for HTTP `file` input references.
+- `--max-request-bytes <n>`: maximum HTTP request body bytes; `0` disables the limit.
+- `--max-artifact-bytes <n>`: maximum HTTP file artifact bytes; `0` disables the limit.
+- `--max-response-bytes <n>`: maximum encoded HTTP response body bytes; `0` disables the limit.
+
+Notes:
+
+- `serve` does not accept runtime model override flags such as `--model` or `--llm-base-url`.
+- HTTP request fields and error codes are documented in the [HTTP API reference](api.md).
+- HTTP `file` input references are rejected unless an artifact root is configured.
+- HTTP size-limit flags affect only `serve`.
+
+## Input And Variable Syntax
+
+- `--input name=path` maps prompt input names to local file paths.
+- `--var name=value` maps prompt template variables to string values.
+- Both flags can be repeated.
+- Both flags also accept comma-separated mappings, such as `--input transcript=./t.md,glossary=./g.yml`.
+- Values may contain `=` after the first separator, such as `--var note=a=b=c`.
+- Empty names and empty values are rejected.
+
+CLI `run` and `render` convert every `--input` mapping to a `file` artifact
+reference. HTTP also supports `inline` input references; see [HTTP API
+reference](api.md).
+
+## Output Behavior
+
+`run`:
+
+- Writes generated artifact content to stdout by default.
+- Writes generated artifact content to `--out` when provided.
+- Prints a success summary to stderr.
+- Prints errors to stderr on failure.
+
+`render`:
+
+- Writes prepared-run output to stdout by default.
+- Writes prepared-run output to `--out` when provided.
+- Does not print a success summary.
+
+`serve`:
+
+- Logs startup and server errors to stderr.
+
+## Exit Codes
+
+- `0`: success.
+- `1`: parse, config, load, render, generation, output-write, or runtime error.
+- `2`: `run` completed and wrote output, but validation status is `failed`.
+
+## Common Workflows
+
+Render prompt inputs and variables as JSON:
+
+```bash
+go run ./cmd/scriptorium render \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --var session_date=2026-05-04 \
+  --format json
+```
+
+Run a prompt with an explicit profile and file output:
+
+```bash
+go run ./cmd/scriptorium run \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --profile local-fast \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --out ./summary.md
+```
+
+Start the HTTP server with example config:
+
+```bash
+go run ./cmd/scriptorium serve --config ./examples/config.yml
+```
+
+Copyable maintained script:
+
+- `examples/render-markdown-summary.sh`

docs/config.md

@@ -0,0 +1,324 @@
+# Configuration Reference
+
+## Config Discovery And Precedence
+
+Application settings are resolved in this order:
+
+1. built-in defaults
+2. `config.yml` values
+3. CLI overrides
+
+When `--config` is omitted, Scriptorium searches:
+
+1. `/usr/local/etc/scriptorium/config.yml`
+2. `/etc/scriptorium/config.yml`
+
+If neither file exists, Scriptorium uses built-in defaults. When
+`--config <path>` is provided, that file must exist and decode successfully.
+
+## Minimal Working Config
+
+```yaml
+prompt_dir: ./examples/prompts
+```
+
+This is enough for `run` and `render` when selected prompts use built-in
+profiles. Set `profile_dir` when prompts or requests use custom profiles.
+
+The maintained repository example is `examples/config.yml`.
+
+## Production-Oriented Config
+
+```yaml
+prompt_dir: /opt/scriptorium/prompts
+profile_dir: /opt/scriptorium/profiles
+schema_dir: /opt/scriptorium/schemas
+
+server:
+  addr: 127.0.0.1:8080
+  artifact_root: /var/lib/scriptorium/artifacts
+  max_request_bytes: 16777216
+  max_artifact_bytes: 16777216
+  max_response_bytes: 16777216
+
+defaults:
+  render_format: text
+```
+
+The maintained full example is `examples/config.full.yml`.
+
+## App Config Reference
+
+Top-level fields:
+
+| Field | Default | Description |
+| --- | --- | --- |
+| `prompt_dir` | unset | Directory containing prompt definition YAML files. Required effectively by `run`, `render`, and `serve`. |
+| `profile_dir` | unset | Directory containing custom profile YAML files. Built-in profiles remain available when unset. |
+| `schema_dir` | `.` | Base directory for relative JSON Schema paths. |
+| `server` | `{}` | HTTP service settings used by `serve`. |
+| `defaults` | `{}` | Adapter defaults. |
+
+`server` fields:
+
+| Field | Default | Description |
+| --- | --- | --- |
+| `server.addr` | `:8080` | Listen address for `serve`. |
+| `server.artifact_root` | unset | Base directory for HTTP `file` input references. Without it, HTTP file refs are rejected. |
+| `server.max_request_bytes` | `16777216` | Maximum encoded HTTP request body bytes. `0` disables the limit. |
+| `server.max_artifact_bytes` | `16777216` | Maximum HTTP file artifact bytes. `0` disables the limit. |
+| `server.max_response_bytes` | `16777216` | Maximum encoded HTTP response bytes. `0` disables the limit. |
+
+`defaults` fields:
+
+| Field | Default | Description |
+| --- | --- | --- |
+| `defaults.render_format` | `text` | Default `render` output format: `text` or `json`. |
+
+Config rules:
+
+- YAML decoding is strict; unknown fields are rejected.
+- HTTP size limits must be greater than or equal to `0`.
+- Empty string config values are ignored.
+- Raw API key fields are not supported in app config.
+
+## Prompt Definition Files
+
+Prompt definitions are YAML files anywhere under `prompt_dir`. Nested
+directories are organizational; callers select prompts by YAML `id`, not file
+path.
+
+Example:
+
+```yaml
+id: generic.structured_events
+version: "1.0.0"
+default_profile: local-quality
+description: Produce structured event JSON from a transcript.
+
+inputs:
+  - name: transcript
+    required: true
+    content_type: text/markdown
+    description: Source transcript content
+  - name: glossary
+    required: false
+    content_type: text/yaml
+    description: Optional glossary context
+
+messages:
+  - role: system
+    content_file: ./generic.structured_events.system.md
+  - role: user
+    content_file: ./generic.structured_events.user.md
+
+output:
+  format: json
+  validation_mode: json_schema
+  schema_path: structured_events.schema.json
+  repair_attempts: 0
+```
+
+Prompt fields:
+
+| Field | Required | Description |
+| --- | --- | --- |
+| `id` | yes | Prompt identifier used by `--prompt` and HTTP `prompt_id`. |
+| `version` | yes | Prompt version. |
+| `default_profile` | no | Profile ID used when a request does not provide a profile. |
+| `description` | no | Human-readable description. |
+| `session_id` | no | Go-template string rendered from request vars and forwarded as provider `session_id` when non-empty. |
+| `inputs` | no | Named input declarations. |
+| `messages` | yes | Chat message templates. |
+| `output` | yes | Output format and validation contract. |
+
+`inputs[]` fields:
+
+- `name` (required)
+- `required` (optional boolean)
+- `content_type` (optional metadata)
+- `description` (optional)
+
+`messages[]` fields:
+
+- `role` (required)
+- exactly one of `content` or `content_file`
+- `cache_control` (optional)
+
+Message rules:
+
+- `content_file` resolves relative to the prompt YAML file location.
+- Repeated roles are allowed.
+- Prompt YAML decoding is strict.
+- Duplicate input names are invalid.
+- Duplicate prompt IDs are invalid for a requested ID/version.
+
+`messages[].cache_control` fields:
+
+| Field | Required | Supported values |
+| --- | --- | --- |
+| `type` | yes | `ephemeral` |
+| `ttl` | no | `1h` |
+
+`session_id` behavior:
+
+- Rendered with the same variable context as message templates.
+- Trimmed and omitted when empty.
+- Rejected when longer than 256 Unicode code points.
+- CLI callers pass variables with `--var`; HTTP callers use `vars`.
+
+`output` fields:
+
+| Field | Required | Supported values |
+| --- | --- | --- |
+| `format` | yes | `text`, `markdown`, `json` |
+| `validation_mode` | yes | `none`, `basic`, `json`, `json_schema` |
+| `schema_path` | only for `json_schema` | Relative to `schema_dir` unless absolute. |
+| `repair_attempts` | yes | Integer greater than or equal to `0`. |
+
+Repair boundary:
+
+- `repair_attempts` is part of the prompt contract.
+- The current CLI and HTTP wiring constructs the runner without a repairer, so normal `run` and `serve` execution does not perform repair attempts.
+
+## Profile Definition Files
+
+Execution profiles are YAML files anywhere under `profile_dir`. Nested
+directories are organizational; callers select profiles by YAML `id`, not file
+path.
+
+Scriptorium also ships built-in profiles. Custom profiles override built-ins
+with the same ID.
+
+Example:
+
+```yaml
+id: local-fast
+endpoint: http://localhost:8000/v1
+model: gpt-4o-mini
+temperature: 0.2
+max_tokens: 500
+top_p: 1.0
+timeout_seconds: 90
+api_key_env: SCRIPTORIUM_API_KEY
+service_tier: priority
+reasoning_effort: medium
+extra_params:
+  provider_route: primary
+```
+
+Profile fields:
+
+| Field | Required | Description |
+| --- | --- | --- |
+| `id` | yes | Profile identifier. |
+| `endpoint` | yes | OpenAI-compatible base URL including `/v1`. |
+| `model` | yes | Provider model name. |
+| `temperature` | no | Range `0..2`. |
+| `max_tokens` | no | Integer greater than or equal to `0`. |
+| `top_p` | no | Range `0..1`. |
+| `timeout_seconds` | no | Integer greater than or equal to `0`. |
+| `service_tier` | no | Provider-specific request tier. |
+| `reasoning_effort` | no | Provider-specific reasoning setting. |
+| `api_key_env` | no | Environment variable name containing the API key. |
+| `extra_params` | no | JSON-compatible provider-specific top-level request fields. |
+
+Execution defaults before profile/request overrides:
+
+| Field | Default |
+| --- | --- |
+| `temperature` | `0.0` |
+| `max_tokens` | `0` |
+| `top_p` | `1.0` |
+| `timeout_seconds` | `600` |
+
+Profile rules:
+
+- Profile YAML decoding is strict.
+- Duplicate custom profile IDs are invalid.
+- Matching custom and built-in IDs are valid override behavior.
+- Raw `api_key` is rejected; use `api_key_env`.
+- If `api_key_env` is set, the named environment variable must be set before `run`, `render`, or HTTP execution can prepare the request.
+- Profile numeric fields merge by non-zero value. Request overrides are presence-aware, so explicit zero values are supported through CLI flags or HTTP model overrides.
+- `extra_params` keys must not be empty and must not collide with reserved outbound fields: `model`, `session_id`, `messages`, `temperature`, `max_tokens`, `top_p`, `service_tier`, `reasoning_effort`, or `response_format`.
+
+Built-in profile catalog:
+
+| Provider | ID | Model | API key env |
+| --- | --- | --- | --- |
+| aion-labs | `aion-2` | `aion-labs/aion-2.0` | `OPENROUTER_API_KEY` |
+| anthropic | `claude-fable-latest` | `~anthropic/claude-fable-latest` | `OPENROUTER_API_KEY` |
+| anthropic | `claude-haiku-latest` | `~anthropic/claude-haiku-latest` | `OPENROUTER_API_KEY` |
+| anthropic | `claude-opus-latest` | `~anthropic/claude-opus-latest` | `OPENROUTER_API_KEY` |
+| anthropic | `claude-sonnet-latest` | `~anthropic/claude-sonnet-latest` | `OPENROUTER_API_KEY` |
+| deepseek | `deepseek-3-2` | `deepseek/deepseek-v3.2` | `OPENROUTER_API_KEY` |
+| deepseek | `deepseek-4-pro` | `deepseek/deepseek-v4-pro` | `OPENROUTER_API_KEY` |
+| google | `gemini-2-flash` | `google/gemini-2.5-flash` | `OPENROUTER_API_KEY` |
+| google | `gemini-2-flash-lite` | `google/gemini-2.5-flash-lite` | `OPENROUTER_API_KEY` |
+| google | `gemini-2-pro` | `google/gemini-2.5-pro` | `OPENROUTER_API_KEY` |
+| google | `gemini-3-flash-lite` | `google/gemini-3.1-flash-lite` | `OPENROUTER_API_KEY` |
+| google | `gemini-flash-latest` | `~google/gemini-flash-latest` | `OPENROUTER_API_KEY` |
+| google | `gemini-pro-latest` | `~google/gemini-pro-latest` | `OPENROUTER_API_KEY` |
+| google | `gemma-4-31b` | `google/gemma-4-31b-it:exacto` | `OPENROUTER_API_KEY` |
+| minimax | `minimax-m2` | `minimax/minimax-m2.5` | `OPENROUTER_API_KEY` |
+| minimax | `minimax-m3` | `minimax/minimax-m3` | `OPENROUTER_API_KEY` |
+| mistral | `mistral-large-2512` | `mistralai/mistral-large-2512` | `OPENROUTER_API_KEY` |
+| mistral | `mistral-medium-3-5` | `mistralai/mistral-medium-3-5` | `OPENROUTER_API_KEY` |
+| mistral | `mistral-small-3` | `mistralai/mistral-small-3.2-24b-instruct` | `OPENROUTER_API_KEY` |
+| mistral | `mistral-small-4` | `mistralai/mistral-small-2603` | `OPENROUTER_API_KEY` |
+| nvidia | `nemotron-3-ultra` | `nvidia/nemotron-3-ultra-550b-a55b` | `OPENROUTER_API_KEY` |
+| openai | `gpt-5-mini` | `openai/gpt-5.4-mini` | `OPENROUTER_API_KEY` |
+| openai | `gpt-5-nano` | `openai/gpt-5.4-nano` | `OPENROUTER_API_KEY` |
+
+## Schema Behavior
+
+Schemas are JSON files, typically under `schema_dir`.
+
+Rules:
+
+- `output.validation_mode: json_schema` requires `output.schema_path`.
+- Relative `schema_path` values resolve from `schema_dir`.
+- Absolute `schema_path` values are used directly.
+- Nested schemas must be referenced by relative path; schemas are not searched recursively by basename.
+- Missing or invalid schema documents are runtime validation errors.
+- Invalid generated JSON produces validation status `failed`, not a runtime error.
+
+## Artifact References
+
+Supported request input artifact reference types are:
+
+- `file`
+- `inline`
+
+CLI `run` and `render` create `file` references from `--input name=path`.
+
+HTTP `file` references require `server.artifact_root` or `serve
+--artifact-root`. Relative file URIs resolve under that root. Absolute paths
+and relative traversal outside the root are rejected by lexical checks. Symlinks
+inside the root are followed by the operating system, including symlinks that
+point outside the root.
+
+HTTP `inline` references do not require an artifact root.
+
+## Secrets Handling
+
+- Keep secret values in environment variables.
+- Store only environment-variable names in `api_key_env`.
+- Do not put raw API keys in config, prompts, profiles, CLI arguments, examples, or HTTP request bodies.
+
+## Maintained Examples
+
+- Minimal app config: `examples/config.yml`
+- Full app config: `examples/config.full.yml`
+- Prompt examples: `examples/prompts/`
+- Custom profile examples: `examples/profiles/`
+- Schema examples: `examples/schemas/`
+- Input fixtures: `examples/fixtures/`
+- Render script: `examples/render-markdown-summary.sh`
+- HTTP request-shape example: `examples/http-run.json`
+
+## Integration References
+
+- [CLI reference](cli.md)
+- [HTTP API reference](api.md)
+- [Outbound OpenAI-compatible contract](integrations/openai-compatible-chat.md)

docs/config/config-yml.md

@@ -1,48 +0,0 @@
-# Main `config.yml`
-
-`config.yml` defines application-level defaults used by CLI commands.
-
-By default, Scriptorium looks for `/usr/local/etc/scriptorium/config.yml` and, if not present, then for `/etc/scriptorium/config.yml`. You can also pass `--config PATH`.
-
-## Complete Example
-
-```yaml
-prompt_dir: ./prompts
-profile_dir: ./profiles
-schema_dir: ./schemas
-
-server:
-  addr: :8080
-
-defaults:
-  render_format: text
-```
-
-## Available Options
-
-- `prompt_dir` (optional): Default directory for prompt definition YAML files.
-- `profile_dir` (optional): Default directory for execution profile YAML files.
-- `schema_dir` (optional): Base directory for JSON schema files used by `json_schema` validation.
-
-### `server`
-
-- `addr` (optional): HTTP server listen address for `scriptorium serve`.
-
-### `defaults`
-
-- `render_format` (optional): Default output format for `scriptorium render`.
-- Allowed values: `text`, `json`.
-
-## Precedence
-
-For run/render/serve settings, precedence is:
-
-1. Explicit CLI flags
-2. `config.yml`
-3. Built-in defaults
-
-## Notes and Rules
-
-- Unknown YAML fields fail to load (strict decoding).
-- This file does not accept API keys.
-- `config.yml` sets directory/server defaults only; prompt/profile content remains in their own files.

docs/config/profile-definitions.md

@@ -1,41 +0,0 @@
-# Execution Profile Definitions
-
-Execution Profiles define **how** Scriptorium calls an LLM endpoint.
-
-A profile file is YAML, typically stored under `profiles/`, for example `profiles/local-quality.yaml`.
-
-## Complete Example
-
-```yaml
-id: local-quality
-endpoint: http://localhost:8000/v1
-model: gpt-4.1
-temperature: 0.0
-max_tokens: 1200
-top_p: 1.0
-timeout_seconds: 180
-reasoning_effort: medium
-api_key_env: SCRIPTORIUM_API_KEY
-extra_params:
-  provider: openrouter
-  route: fallback
-```
-
-## Available Options
-
-- `id` (required): Unique profile identifier used by `--profile` or prompt `default_profile`.
-- `endpoint` (required): OpenAI-compatible base URL, usually ending in `/v1`.
-- `model` (required): Model name to request at that endpoint.
-- `temperature` (optional): Sampling temperature. Valid range is `0` to `2`.
-- `max_tokens` (optional): Max completion tokens. Must be `>= 0`.
-- `top_p` (optional): Nucleus sampling parameter. Valid range is `0` to `1`.
-- `timeout_seconds` (optional): Request timeout in seconds. Must be `>= 0`.
-- `reasoning_effort` (optional): Provider/model-specific reasoning level string.
-- `api_key_env` (optional): Environment variable name that holds the API key.
-- `extra_params` (optional): String key/value map for provider-specific parameters.
-
-## Notes and Rules
-
-- Raw API keys are not supported. Do **not** add `api_key` fields.
-- Unknown YAML fields fail to load (strict decoding).
-- If `api_key_env` is set, the environment variable must be present when the run executes.

docs/config/prompt-definitions.md

@@ -1,73 +0,0 @@
-# Prompt Definition Files
-
-Prompt Definitions define **what** Scriptorium should do.
-
-A prompt file is YAML, typically stored under `prompts/`, for example `prompts/generic.structured_events.yaml`.
-
-## Complete Example
-
-```yaml
-id: generic.structured_events
-version: "1.0.0"
-default_profile: local-quality
-description: Extract events from a transcript into structured JSON.
-
-inputs:
-  - name: transcript
-    required: true
-    content_type: text/markdown
-    description: Source transcript
-  - name: glossary
-    required: false
-    content_type: text/yaml
-    description: Optional glossary context
-
-messages:
-  - role: system
-    content: |
-      You are a structured extraction assistant.
-      Return only JSON.
-  - role: user
-    content_file: ./generic.structured_events.user.md
-
-output:
-  format: json
-  validation_mode: json_schema
-  schema_path: structured_events.schema.json
-  repair_attempts: 1
-```
-
-## Available Options
-
-- `id` (required): Prompt identifier used by `--prompt` / `prompt_id`.
-- `version` (required): Prompt version string.
-- `default_profile` (optional): Execution profile ID used when no explicit profile is provided.
-- `description` (optional): Human-readable description.
-
-### `inputs[]`
-
-- `name` (required): Logical input name referenced in templates via `{{input "name"}}`.
-- `required` (optional): If `true`, run fails when input is missing.
-- `content_type` (optional): Metadata only (not enforced yet).
-- `description` (optional): Human-readable input description.
-
-### `messages[]`
-
-- `role` (required): Message role such as `system` or `user`.
-- `content` (optional): Inline Go-template message body.
-- `content_file` (optional): Path to a template file.
-
-Each message must set **exactly one** of `content` or `content_file`.
-
-### `output`
-
-- `format` (required): One of `text`, `markdown`, `json`.
-- `validation_mode` (required): One of `none`, `basic`, `json`, `json_schema`.
-- `schema_path` (required when `validation_mode: json_schema`): Path to JSON Schema file.
-- `repair_attempts` (required): Number of bounded repair retries (`>= 0`).
-
-## Notes and Rules
-
-- Unknown YAML fields fail to load (strict decoding).
-- `content_file` paths are resolved relative to the prompt YAML file.
-- For `json_schema` validation mode, Scriptorium also sends provider-level structured output requests automatically.

docs/config/schema-definitions.md

@@ -1,82 +0,0 @@
-# JSON Schema Definition Files
-
-Schema definition files describe the expected JSON output contract for prompts that use:
-
-- `output.format: json`
-- `output.validation_mode: json_schema`
-
-Schema files are JSON, typically stored under `schemas/`, for example `schemas/structured_events.schema.json`.
-
-## Complete Example
-
-```json
-{
-  "$schema": "https://json-schema.org/draft/2020-12/schema",
-  "$id": "https://example.com/schemas/structured-events.schema.json",
-  "title": "Structured Events",
-  "description": "Expected shape for extracted event output",
-  "type": "object",
-  "properties": {
-    "summary": {
-      "type": "string",
-      "minLength": 1,
-      "description": "High-level session summary"
-    },
-    "events": {
-      "type": "array",
-      "items": {
-        "type": "object",
-        "properties": {
-          "title": { "type": "string" },
-          "type": {
-            "type": "string",
-            "enum": ["discovery", "combat", "social", "travel", "downtime", "other"]
-          },
-          "notes": { "type": "string" }
-        },
-        "required": ["title", "type"],
-        "additionalProperties": false
-      },
-      "minItems": 0
-    }
-  },
-  "required": ["summary", "events"],
-  "additionalProperties": false,
-  "$defs": {
-    "nonEmptyString": {
-      "type": "string",
-      "minLength": 1
-    }
-  }
-}
-```
-
-## Available Options
-
-Scriptorium does not define custom schema keywords. It expects a valid JSON Schema document and passes it to the validator/provider.
-
-Commonly used JSON Schema options include:
-
-- `$schema`: Draft identifier URI.
-- `$id`: Schema identifier URI.
-- `title`: Human-readable schema title.
-- `description`: Human-readable schema description.
-- `type`: Expected JSON type (`object`, `array`, `string`, etc.).
-- `properties`: Object field definitions.
-- `required`: Required object fields.
-- `additionalProperties`: Whether undeclared fields are allowed.
-- `items`: Array item schema.
-- `enum`: Allowed literal values.
-- `const`: Single allowed literal value.
-- `oneOf`, `anyOf`, `allOf`: Composition rules.
-- `minimum`, `maximum`: Numeric bounds.
-- `minLength`, `maxLength`, `pattern`: String constraints.
-- `minItems`, `maxItems`: Array constraints.
-- `$defs`: Reusable local definitions.
-- `$ref`: Reference to another schema/definition.
-
-## Notes and Rules
-
-- Schema path comes from prompt `output.schema_path` and is resolved relative to `schema_dir`.
-- If schema loading fails for `json_schema` mode, the run fails before the LLM request.
-- Keep schemas strict (`additionalProperties: false`) when you want predictable output shape.

docs/consumers/api.md

@@ -0,0 +1,122 @@
+# Consumer Integration Overview
+
+This guide is for applications that call Scriptorium from another codebase.
+
+Scriptorium exposes three integration surfaces:
+
+| Surface | Use when |
+| --- | --- |
+| Go package | The consumer is Go, needs typed requests/results, or wants injected LLM clients for tests. |
+| CLI subprocess | The consumer wants process isolation or is not written in Go. |
+| HTTP API | The consumer needs a service boundary or remote access to `POST /v1/runs`. |
+
+Canonical references:
+
+- Go package: [Package scriptorium](pkg-scriptorium.md)
+- CLI subprocess: [Subprocess integration](../integrations/subprocess.md)
+- HTTP: [HTTP API reference](../api.md)
+- File formats: [Configuration reference](../config.md)
+
+## Required Deployment Inputs
+
+Every integration needs operators to provide:
+
+- prompt definitions;
+- profile definitions or built-in profile IDs;
+- schema files when prompts use `json_schema`;
+- input artifacts or inline input bodies;
+- API-key environment variables or direct per-request keys where supported.
+
+Raw API keys do not belong in config, prompt files, profile YAML, CLI
+arguments, or HTTP request bodies.
+
+## Recommended Workflow
+
+Use the Go package when:
+
+- the consumer is a Go application;
+- the application needs `context.Context` cancellation;
+- repeated calls should avoid subprocess startup;
+- tests need a fake LLM client;
+- direct per-request `RunRequest.APIKey` is required.
+
+Use the CLI subprocess when:
+
+- the consumer is not Go;
+- process isolation is useful;
+- stdout/stderr separation and exit codes are enough;
+- the consumer already manages local files and environment variables.
+
+Use HTTP when:
+
+- Scriptorium should run as a service;
+- multiple clients need a shared prompt/profile deployment;
+- clients can reach a trusted, protected HTTP boundary.
+
+## Minimal Go Example
+
+```go
+engine, err := scriptorium.NewEngine(scriptorium.Config{
+	PromptDir:  "./examples/prompts",
+	ProfileDir: "./examples/profiles",
+	SchemaDir:  "./examples/schemas",
+})
+if err != nil {
+	return err
+}
+
+prepared, err := engine.Prepare(ctx, scriptorium.RunRequest{
+	PromptID: "generic.markdown_summary",
+	Inputs: map[string]scriptorium.ArtifactRef{
+		"transcript": scriptorium.File("./examples/fixtures/transcript.md"),
+		"glossary":   scriptorium.File("./examples/fixtures/glossary.yml"),
+	},
+})
+if err != nil {
+	return err
+}
+_ = prepared.Messages
+```
+
+Run the maintained package example:
+
+```bash
+go run ./examples/go-library/prepare
+```
+
+## Subprocess Workflow
+
+Invoke `scriptorium render` for preflight and `scriptorium run` for generation.
+Capture stdout and stderr separately. Treat exit code `2` from `run` as a
+completed generation with failed validation.
+
+See [Subprocess integration](../integrations/subprocess.md) for the stable
+invocation contract.
+
+## HTTP Workflow
+
+Run `scriptorium serve` behind trusted controls and send JSON requests to
+`POST /v1/runs`.
+
+Do not duplicate endpoint schemas in consumers. Use the [HTTP API
+reference](../api.md) as the authoritative contract.
+
+## Consumer Responsibilities
+
+Consumers are responsible for:
+
+- selecting prompt/profile IDs as deployment configuration;
+- supplying all required inputs and vars;
+- protecting generated artifacts and rendered prompts as sensitive data;
+- deciding whether to keep output when validation fails;
+- implementing retries only when another model call is acceptable.
+
+Scriptorium does not persist run state. Retrying a failed or timed-out request
+can produce different output and can incur another provider request.
+
+## Status Behavior
+
+- Go package methods return typed results or errors that support `errors.Is`.
+- CLI `run` exits `2` when generation succeeds but validation fails.
+- HTTP returns `200 OK` for generated-content validation failures and exposes the failed status in the response body.
+- Runtime validation failures are errors.

docs/consumers/pkg-scriptorium.md

@@ -0,0 +1,284 @@
+# Package scriptorium
+
+Import path:
+
+```go
+import "gitea.maximumdirect.net/eric/scriptorium"
+```
+
+The root package is the public Go facade for Scriptorium's prompt prepare/run
+workflow. It exposes typed requests, results, source options, injected LLM
+clients, and stable public errors while keeping `internal/*` packages private.
+
+## Intended Use Cases
+
+Use the package when a Go application needs:
+
+- in-process prompt preparation or execution;
+- typed request/result structs;
+- direct `context.Context` cancellation;
+- injected/fake LLM clients for tests;
+- direct per-request `RunRequest.APIKey`.
+
+Use [Subprocess integration](../integrations/subprocess.md) or the [HTTP API](../api.md)
+when a process or service boundary is preferred.
+
+## Construct An Engine
+
+```go
+engine, err := scriptorium.NewEngine(scriptorium.Config{
+	PromptDir:  "./examples/prompts",
+	ProfileDir: "./examples/profiles",
+	SchemaDir:  "./examples/schemas",
+})
+if err != nil {
+	return err
+}
+```
+
+`Config` fields:
+
+| Field | Description |
+| --- | --- |
+| `PromptDir` | Prompt definition directory. Required unless `WithPromptFS` or `WithPromptFile` is used. |
+| `ProfileDir` | Optional custom profile directory overlaid above built-in profiles. |
+| `SchemaDir` | Schema directory. Defaults to `.` when empty. |
+| `Timeout` | Default timeout for the built-in OpenAI-compatible client. |
+| `HTTPClient` | Optional HTTP client for the built-in OpenAI-compatible client. |
+
+`NewEngine` accepts `nil` options and ignores them. Invalid construction wraps
+`ErrInvalidConfig`.
+
+## Source Options
+
+Directory fields are the compatibility path. Explicit source options override
+the matching directory field.
+
+Prompt sources:
+
+- `WithPromptFS(fsys, root)`
+- `WithPromptFile(path)`
+
+Profile sources:
+
+- `WithProfileFS(fsys, root)`
+- `WithProfileFile(path)`
+- `WithProfiles(profiles...)`
+
+Schema sources:
+
+- `WithSchemaFS(fsys, root)`
+- `WithSchemaFile(path)`
+
+LLM source:
+
+- `WithLLMClient(client)`
+
+Source behavior:
+
+- Prompt and profile YAML use the same strict rules as directory loading.
+- Prompt `content_file` values resolve relative to the prompt file.
+- `fs.FS` roots are containment boundaries for prompt content files and schema paths.
+- File options expose the selected file by its base name.
+- Profile source precedence is in-memory profiles, then explicit profile file/FS/directory source, then built-ins.
+- `WithLLMClient(nil)` returns `ErrInvalidConfig`.
+
+## In-Memory Profiles
+
+Use `WithProfiles` when the application already has typed model settings:
+
+```go
+profile := scriptorium.OpenAICompatibleProfile(scriptorium.OpenAICompatibleProfileConfig{
+	ID:             "app.default",
+	Endpoint:       "https://openrouter.ai/api/v1",
+	Model:          "mistralai/mistral-small-3.2-24b-instruct",
+	APIKeyRequired: true,
+})
+
+engine, err := scriptorium.NewEngine(cfg, scriptorium.WithProfiles(profile))
+```
+
+`Profile` and `OpenAICompatibleProfileConfig` include:
+
+- `ID`
+- `Endpoint`
+- `Model`
+- `Temperature`
+- `MaxTokens`
+- `TopP`
+- `TimeoutSeconds`
+- `ServiceTier`
+- `ReasoningEffort`
+- `APIKeyRequired`
+- `ExtraParams`
+
+`WithProfiles` rejects duplicate IDs in one call. In-memory profiles do not
+store raw keys. When `APIKeyRequired` is true, pass the secret on each request
+with `RunRequest.APIKey`.
+
+`ExtraParams` must be JSON-compatible: strings, booleans, finite numbers,
+objects with string keys, arrays/slices, and nil. Unsupported values, non-string
+map keys, non-finite floats, and cycles return `ErrInvalidConfig` for profiles
+or `ErrInvalidRequest` for request overrides.
+
+## Prepare Workflow
+
+`Prepare` resolves prompt/profile/input/schema state and renders messages
+without calling an LLM.
+
+```go
+prepared, err := engine.Prepare(ctx, scriptorium.RunRequest{
+	PromptID: "generic.markdown_summary",
+	Inputs: map[string]scriptorium.ArtifactRef{
+		"transcript": scriptorium.File("./examples/fixtures/transcript.md"),
+		"glossary":   scriptorium.File("./examples/fixtures/glossary.yml"),
+	},
+})
+if err != nil {
+	return err
+}
+_ = prepared.EffectiveModelParams
+```
+
+`PreparedRun` includes prompt ID/version/hash, selected profile, effective
+model params, output contract, structured-output metadata, input hashes,
+rendered prompt hash, rendered messages, and timing fields. It does not include
+raw API-key values, model output, validation results, or internal target
+presence metadata.
+
+## Run Workflow
+
+`Run` calls `Prepare`, invokes the configured LLM client, builds the output
+artifact, and validates the output.
+
+```go
+result, err := engine.Run(ctx, scriptorium.RunRequest{
+	PromptID: "generic.markdown_summary",
+	APIKey:   apiKey,
+	Inputs: map[string]scriptorium.ArtifactRef{
+		"transcript": scriptorium.File("./examples/fixtures/transcript.md"),
+		"glossary":   scriptorium.File("./examples/fixtures/glossary.yml"),
+	},
+})
+if err != nil {
+	return err
+}
+_ = result.Artifact
+```
+
+`RunResult` includes run ID, output artifact, raw output, validation result,
+prompt/profile/model metadata, effective model params, input hashes, usage, and
+timing fields.
+
+Generated-content validation failures return a successful `RunResult` with
+`Validation.Status == ValidationFailed`. Runtime/schema validation errors
+return an error that matches `ErrValidation`.
+
+## Inputs
+
+Input helpers:
+
+- `File(path)`: file-backed artifact reference.
+- `Inline(body)`: inline artifact body.
+- `InlineWithURI(uri, body)`: inline artifact body with URI metadata.
+
+Input map keys must match the prompt's expected input names.
+
+## Injected LLM Clients
+
+Use `WithLLMClient` for tests or custom model integrations:
+
+```go
+type fakeLLM struct{}
+
+func (fakeLLM) Generate(ctx context.Context, req scriptorium.GenerateRequest) (*scriptorium.GenerateResponse, error) {
+	return &scriptorium.GenerateResponse{
+		Content: "generated text",
+		Usage:   scriptorium.TokenUsage{TotalTokens: 12},
+	}, nil
+}
+
+engine, err := scriptorium.NewEngine(cfg, scriptorium.WithLLMClient(fakeLLM{}))
+```
+
+Injected clients receive:
+
+- rendered prompt;
+- effective execution target;
+- numeric target presence metadata;
+- structured-output spec when applicable;
+- direct request API key when provided.
+
+Custom clients should not log raw prompts or API keys by default.
+
+## Overrides And API Keys
+
+`RunRequest` fields:
+
+| Field | Description |
+| --- | --- |
+| `PromptID` | Prompt ID. |
+| `PromptVersion` | Optional prompt version filter. |
+| `ProfileID` | Optional profile override. |
+| `APIKey` | Direct per-request API key. |
+| `Inputs` | Input artifact references. |
+| `Vars` | Template variables. |
+| `Execution` | Per-request model overrides. |
+| `Validation` | Per-request output contract override. |
+| `Metadata` | Request metadata reserved for callers. |
+
+`RunRequest.Execution` uses pointer fields for numeric values so explicit zero
+overrides are preserved:
+
+```go
+zero := 0
+req.Execution = &scriptorium.ExecutionTargetOverride{
+	MaxTokens: &zero,
+}
+```
+
+Direct `RunRequest.APIKey` takes precedence over profile `api_key_env` for the
+default OpenAI-compatible client. It is request-scoped, uses `json:"-"`, and is
+not included in `PreparedRun` or `RunResult` JSON. Normal Go string formatting
+of `RunRequest` and `GenerateRequest` reports only whether a direct key is set.
+
+Raw API keys do not belong in profile YAML, in-memory profiles, or app config.
+Avoid reflection-based debug dumps of request structs because exported fields
+remain visible to tools that bypass `String` and `GoString`.
+
+## Errors
+
+Public methods wrap context while preserving stable sentinel checks with
+`errors.Is`:
+
+- `ErrInvalidConfig`
+- `ErrInvalidRequest`
+- `ErrPromptNotFound`
+- `ErrProfileNotFound`
+- `ErrPromptLoad`
+- `ErrProfileLoad`
+- `ErrArtifactLoad`
+- `ErrPromptRender`
+- `ErrLLMGenerate`
+- `ErrValidation`
+
+Example:
+
+```go
+if errors.Is(err, scriptorium.ErrPromptNotFound) {
+	return err
+}
+```
+
+## Examples
+
+Run the maintained prepare-only example from the repository root:
+
+```bash
+go run ./examples/go-library/prepare
+```
+
+See also:
+
+- [Configuration reference](../config.md)
+- [Consumer integration overview](api.md)

docs/development.md

@@ -0,0 +1,60 @@
+# Development
+
+This is the contributor entry point for Scriptorium. Use the task-specific
+reading guide below before making changes. Canonical architecture, contracts,
+component behavior, and policies remain in their owning documents.
+
+## Initial Orientation
+
+Before starting work:
+
+1. inspect the working tree and preserve unrelated changes;
+2. read the architecture policy for code or design work;
+3. read the policy, contract, and internal documents listed for the task;
+4. inspect the relevant implementation and tests before deciding how to change
+   them.
+
+Start with:
+
+- [Architecture policy](policy/architecture.md) for system boundaries,
+  invariants, and non-goals;
+- [Documentation policy](policy/documentation.md) before changing
+  documentation;
+- [Testing policy](policy/testing.md) before adding, rewriting, or deleting
+  tests.
+
+## Task-Specific Reading Guide
+
+| Task | Read before changing |
+| --- | --- |
+| Public Go package or engine behavior | [Go package consumer contract](consumers/pkg-scriptorium.md), [runner internals](internal/runner.md), [adapter internals](internal/adapters.md), and [source internals](internal/sources.md) |
+| CLI commands, flags, output, or exit behavior | [CLI contract](cli.md) and [adapter internals](internal/adapters.md) |
+| HTTP routes, DTOs, limits, or status mapping | [HTTP API contract](api.md), [adapter internals](internal/adapters.md), and [source internals](internal/sources.md) |
+| Application configuration | [Configuration contract](config.md), [adapter internals](internal/adapters.md), and [source internals](internal/sources.md) |
+| Prompt, profile, schema, or artifact loading | [Configuration contract](config.md) and [source internals](internal/sources.md) |
+| Runner orchestration, rendering, validation, or repair | [Runner internals](internal/runner.md) and [source internals](internal/sources.md) |
+| OpenAI-compatible request or response behavior | [OpenAI-compatible integration](integrations/openai-compatible-chat.md), [runner internals](internal/runner.md), and [adapter internals](internal/adapters.md) |
+| Subprocess behavior | [Subprocess integration](integrations/subprocess.md) and [CLI contract](cli.md) |
+| Runtime operation, recovery, or troubleshooting | [Operations](operations.md) and [troubleshooting](troubleshooting.md) |
+| Examples or copyable assets | The owning contract for the demonstrated behavior and the related files under `examples/` |
+| Architecture decisions or future work | The [documentation policy](policy/documentation.md), relevant accepted ADRs under `adr/`, and relevant roadmap documents under `roadmap/` |
+
+For cross-cutting changes, follow every applicable row. Internal component
+documents own detailed subsystem change recipes.
+
+## Baseline Validation
+
+Use focused checks while iterating, then run validation proportionate to the
+change and the risks described by the testing policy.
+
+The repository-level baseline for code changes is:
+
+```bash
+go test ./...
+go vet ./...
+go build ./cmd/scriptorium
+```
+
+Documentation-only work does not require the full Go suite unless it changes
+commands, examples, generated output, or another behavior that the suite
+validates. Always check changed links, paths, examples, and canonical ownership.

docs/integrations/narratio.md

@@ -1,322 +0,0 @@
-# Narratio -> Scriptorium CLI Integration
-
-## 1. Purpose
-
-This document defines how Narratio should invoke Scriptorium through the **public CLI**.
-
-This is a **subprocess integration contract**, not an internal Go API contract.
-
-## 2. Assumptions
-
-- `scriptorium` is installed and available on `PATH`.
-- Scriptorium is configured with `config.yml`.
-- `config.yml` provides `prompt_dir`, `profile_dir`, and `schema_dir` as needed.
-- Prompt and profile libraries are already deployed for the environment.
-- Narratio provides prepared artifact files (for example polished transcript, glossary, previous recap, campaign notes).
-- Initial integration is synchronous subprocess execution.
-- Narratio remains the orchestrator.
-
-In normal operation, Narratio does not need to pass `--prompt-dir` and `--profile-dir` if they are supplied by Scriptorium config.
-
-Narratio may pass `--config <PATH>` when it must use a non-default Scriptorium config file.
-
-## 3. Core Commands Narratio May Call
-
-Primary commands for subprocess integration:
-
-- `scriptorium run`
-- `scriptorium render`
-
-For production generation, use `scriptorium run`.
-
-`scriptorium render` is for debugging, dry-runs, test assertions, and validating command construction without LLM execution.
-
-Note: `scriptorium serve` and HTTP API exist, but they are not the initial integration path.
-
-## 4. Command Selection Guidance
-
-- Use `run` to generate an output artifact.
-- Use `render` to inspect the prepared prompt and effective settings without calling the LLM.
-- Use `render --format json` when Narratio/tests need structured prepare output.
-
-## 5. Recommended `run` Invocation Shape
-
-Production shape:
-
-```bash
-scriptorium run \
-  --prompt <prompt_id> \
-  --input transcript=<processed-transcript-path> \
-  --out <output-artifact-path>
-```
-
-Common optional additions:
-
-- `--config <path>`: use a specific Scriptorium config file.
-- `--profile <profile_id>`: override prompt default profile.
-- `--var name=value` (repeatable): small metadata values.
-- `--input name=path` (repeatable): additional named artifacts.
-- `--timeout <duration>`: per-run timeout override.
-- Runtime model override flags (`--llm-base-url`, `--model`, etc.) only for exceptional/operator-directed cases.
-
-## 6. Recommended `render` Invocation Shape
-
-Human-readable debug shape:
-
-```bash
-scriptorium render \
-  --prompt <prompt_id> \
-  --input transcript=<processed-transcript-path> \
-  --format text
-```
-
-Structured debug/test shape:
-
-```bash
-scriptorium render \
-  --prompt <prompt_id> \
-  --input transcript=<processed-transcript-path> \
-  --format json \
-  --out <render-debug-path>
-```
-
-`render` does **not** call the LLM, does **not** validate model output, and does **not** perform repair.
-
-## 7. Inputs
-
-- Pass inputs as repeated `--input name=path` flags.
-- `name` must match the Prompt Definition input name.
-- Prefer absolute paths, or paths relative to a working directory controlled by Narratio.
-- Pass Audita output as the primary transcript input.
-- Additional inputs may include glossary, previous recap, campaign notes, event logs, final state maps, or other prompt-specific artifacts.
-- Scriptorium reads input files directly; Narratio does not need to inline file content for CLI use.
-
-## 8. Variables
-
-Use repeated `--var name=value` for small metadata values.
-
-Typical examples:
-
-- `session_date`
-- `session_id`
-- `campaign_name`
-- `previous_session_id`
-- `output_kind`
-
-Large content belongs in input files, not `--var` values.
-
-## 9. Prompt IDs and Output Artifact Types
-
-Narratio should treat prompt IDs as configuration, not hardcoded business logic.
-
-Narratio config may map stage/output names to prompt IDs, for example:
-
-- session recap prompt
-- structured event extraction prompt
-- glossary suggestion prompt
-- player-facing summary prompt
-
-Prompt IDs used by Narratio should come from the deployed Scriptorium prompt library.
-
-## 10. Profiles
-
-- Prompts may declare `default_profile`.
-- Narratio may omit `--profile` to use prompt default profile.
-- Narratio may pass `--profile` to force profile selection.
-- This enables environment/profile selection like `local-fast`, `local-quality`, `frontier`, `batch`, or test profiles.
-- Profile names should generally be Narratio configuration values.
-
-## 11. Runtime Overrides
-
-Supported runtime override flags:
-
-- `--llm-base-url`
-- `--model`
-- `--api-key-env`
-- `--temperature`
-- `--max-tokens`
-- `--top-p`
-- `--timeout`
-
-Guidance:
-
-- Keep normal model/runtime settings in Execution Profiles.
-- Use runtime overrides only for explicit per-run exceptions, tests, or operator overrides.
-- Never pass raw API keys on the command line.
-- `--api-key-env` names an environment variable; Narratio must ensure that variable is set in subprocess environment.
-
-## 12. Config Behavior
-
-- Default config path: `/etc/scriptorium/config.yml`.
-- `--config <PATH>` overrides default path.
-- Missing default config is allowed by Scriptorium.
-- If `--config` is provided explicitly, the file must exist and be valid.
-- CLI flags override `config.yml`.
-- `config.yml` overrides built-in application defaults.
-
-Narratio can either:
-
-- rely on system default config path, or
-- carry an explicit config path and pass `--config`.
-
-## 13. Environment Handling
-
-Subprocess environment recommendations:
-
-- Pass through required API-key environment variables referenced by `api_key_env`.
-- Do not pass raw API keys as CLI arguments.
-- Avoid logging full environment dumps.
-- Capture stdout and stderr separately.
-- Use a controlled working directory.
-- Prefer absolute artifact paths.
-
-## 14. Output Handling
-
-For `scriptorium run`:
-
-- Use `--out` when Narratio needs durable artifact files.
-- Without `--out`, artifact content is written to stdout.
-- Preferred orchestration pattern: always use `--out`, then treat the file as stage output artifact.
-- Capture stderr for diagnostics.
-
-For `scriptorium render`:
-
-- Use `--out` to store render diagnostics.
-- Use `--format json` when tests need to inspect selected profile, effective runtime settings, input hashes, prompt hash, and rendered messages.
-
-## 15. Exit Status and Errors
-
-Current CLI behavior (verified from implementation/tests):
-
-- `0`: success.
-- `1`: runtime/parse/config/load/render/generation/IO error.
-- `2`: run completed but output validation failed (`ValidationFailed`).
-
-Additional details:
-
-- On `run`, output artifact write happens before exit code selection. If validation fails, artifact may still be written and exit code is `2`.
-- `stderr` carries both errors and normal run summary output; non-empty stderr alone does not imply failure.
-- `render` returns `0` on success and `1` on failures.
-
-Narratio should treat non-zero exit codes as failed stage execution, but may record generated artifact paths if a run exited `2` and output file exists.
-
-## 16. Recommended Narratio Integration Pattern
-
-1. Build CLI args from Narratio stage configuration.
-2. Use subprocess context cancellation/timeout.
-3. Pass absolute input paths.
-4. Pass `--out` to a session-scoped artifact path.
-5. Add `--var` metadata values.
-6. Optionally add `--config`.
-7. Optionally add `--profile`.
-8. Ensure required API-key env vars are present.
-9. Run subprocess synchronously.
-10. Capture stdout/stderr separately.
-11. On success, store output artifact path and invocation metadata in stage artifacts.
-12. On failure, store exit code and stderr diagnostics in stage status.
-
-## 17. Suggested Narratio Configuration Shape
-
-Illustrative (not required schema):
-
-```yaml
-scriptorium:
-  config_path: /etc/scriptorium/config.yml
-  stages:
-    session_recap:
-      prompt_id: dnd.session_recap
-      profile_id: local-quality   # optional
-      inputs: [transcript, glossary, previous_recap]
-      vars: [session_id, session_date, campaign_name]
-      output_path_template: artifacts/{session_id}/session_recap.md
-      timeout: 2m
-      render_debug: false
-```
-
-The key idea: map Narratio stage/artifact names to prompt ID, optional profile, expected inputs, and output destination.
-
-## 18. Testing Strategy for Narratio Integration
-
-- Use `scriptorium render --format json` to verify command construction without LLM calls.
-- Use dedicated test prompt/profile libraries for integration tests.
-- Use small fixture transcripts.
-- Verify missing-input failure behavior.
-- Verify prompt `default_profile` behavior.
-- Verify explicit `--profile` override behavior.
-- Verify `--config` behavior (default and explicit).
-- Verify output file creation when `--out` is used.
-- Verify stderr capture on failures.
-- Avoid real API keys in tests.
-
-## 19. Security and Privacy Notes
-
-- Never pass raw API keys on command line.
-- Do not log full rendered prompts by default; transcripts may contain sensitive content.
-- Avoid logging prompt content unless explicit debug mode is enabled.
-- Treat generated artifacts as potentially sensitive.
-- Use session-scoped, access-controlled output paths.
-- `api_key_env` names should come from environment management, not embedded secrets.
-
-## 20. Initial D&D Artifact Generation Examples
-
-These are examples only. Use prompt IDs from the deployed prompt library.
-
-Session recap:
-
-```bash
-scriptorium run \
-  --prompt dnd.session_recap \
-  --input transcript=/work/session-42/transcript.polished.md \
-  --input glossary=/work/session-42/glossary.yml \
-  --out /work/session-42/artifacts/session_recap.md
-```
-
-Structured events:
-
-```bash
-scriptorium run \
-  --prompt dnd.structured_events \
-  --input transcript=/work/session-42/transcript.polished.md \
-  --out /work/session-42/artifacts/structured_events.json
-```
-
-Glossary suggestions:
-
-```bash
-scriptorium run \
-  --prompt dnd.glossary_suggestions \
-  --input transcript=/work/session-42/transcript.polished.md \
-  --input previous_recap=/work/session-41/artifacts/session_recap.md \
-  --out /work/session-42/artifacts/glossary_suggestions.md
-```
-
-Player-facing summary:
-
-```bash
-scriptorium run \
-  --prompt dnd.player_summary \
-  --input transcript=/work/session-42/transcript.polished.md \
-  --input structured_events=/work/session-42/artifacts/structured_events.json \
-  --out /work/session-42/artifacts/player_summary.md
-```
-
-## 21. Non-Goals
-
-Initial Narratio integration should not:
-
-- call Scriptorium internal Go packages
-- use HTTP API as the primary path
-- expect Scriptorium to read S3 refs directly
-- make Scriptorium responsible for Narratio stage state
-- make Scriptorium responsible for notification
-- require Scriptorium to understand D&D workflow semantics beyond prompt definitions
-
-## 22. Future Extension Notes
-
-Possible later extensions:
-
-- HTTP API integration
-- S3 artifact references if Scriptorium adds S3 reader support
-- storing render diagnostics alongside generated artifacts
-- token budgeting/prompt-size checks
-- batch execution if Scriptorium later adds batch support

docs/integrations/openai-compatible-chat.md

@@ -0,0 +1,193 @@
+# OpenAI-Compatible Chat Integration
+
+## Scope
+
+This document defines the outbound LLM contract implemented by `internal/llm/openai_compatible_client.go`.
+
+It documents only fields and behaviors currently serialized by code.
+
+## Endpoint Construction
+
+Request endpoint is built as:
+
+1. choose base URL:
+   - `GenerateRequest.Target.Endpoint` if set
+   - otherwise client config `BaseURL`
+2. trim trailing slash
+3. append `/chat/completions`
+
+Example:
+
+- base URL: `http://localhost:8000/v1`
+- final URL: `http://localhost:8000/v1/chat/completions`
+
+## Request Fields Sent
+
+Serialized JSON fields:
+
+- `model` (required after fallback resolution)
+- `session_id` (only when the rendered prompt includes a non-empty session ID)
+- `messages` (rendered prompt messages)
+- `temperature` (when non-zero, or when explicitly overridden to zero)
+- `max_tokens` (when non-zero, or when explicitly overridden to zero)
+- `top_p` (when non-zero, or when explicitly overridden to zero)
+- `service_tier` (only when non-empty)
+- `reasoning_effort` (only when non-empty)
+- `response_format` (only when structured output is provided)
+- profile/request `extra_params` as additional provider-specific top-level fields
+
+`service_tier` is provider-specific. OpenRouter currently documents request values such as `flex` and `priority`; Scriptorium forwards any non-empty configured value and lets the backend validate support.
+
+`reasoning_effort` is provider-specific. Scriptorium forwards any non-empty configured value as top-level `reasoning_effort` and lets the backend validate support.
+
+`extra_params` are flattened into the outbound JSON object. They are not wrapped in an `extra_params` object:
+
+```json
+{
+  "model": "gpt-4o-mini",
+  "messages": [
+    {
+      "role": "user",
+      "content": "rendered text"
+    }
+  ],
+  "provider_route": "primary",
+  "provider_options": {
+    "retry_budget": 2
+  }
+}
+```
+
+`extra_params` values must be JSON-compatible. Supported value shapes include strings, numbers, booleans, objects, and arrays.
+
+Reserved `extra_params` keys are rejected before the HTTP request is made:
+
+- `model`
+- `session_id`
+- `messages`
+- `temperature`
+- `max_tokens`
+- `top_p`
+- `service_tier`
+- `reasoning_effort`
+- `response_format`
+
+Empty `extra_params` keys and values that cannot be encoded as JSON are also rejected before the HTTP request is made.
+
+`session_id` is rendered from prompt YAML using request variables and serialized as a top-level JSON request field. Scriptorium does not send an `x-session-id` header. Empty rendered session IDs are omitted, and values longer than 256 characters are rejected before the HTTP request.
+
+Messages without prompt cache control serialize with string `content`:
+
+```json
+{
+  "role": "system",
+  "content": "rendered text"
+}
+```
+
+Messages with prompt cache control serialize as a single text content-block array:
+
+```json
+{
+  "role": "system",
+  "content": [
+    {
+      "type": "text",
+      "text": "rendered text",
+      "cache_control": {
+        "type": "ephemeral",
+        "ttl": "1h"
+      }
+    }
+  ]
+}
+```
+
+When cache-control `ttl` is unset in the prompt definition, `ttl` is omitted from the outbound payload.
+
+Structured output is currently `json_schema` only, serialized as:
+
+```json
+{
+  "response_format": {
+    "type": "json_schema",
+    "json_schema": {
+      "name": "...",
+      "strict": true,
+      "schema": {"type": "object"}
+    }
+  }
+}
+```
+
+## Authentication Header
+
+If `Target.APIKey` is set:
+
+- set `Authorization: Bearer <value>`
+- do not read `Target.APIKeyEnv`
+
+If `Target.APIKey` is empty and `Target.APIKeyEnv` is set:
+
+- resolve environment variable value at request time
+- set `Authorization: Bearer <value>`
+
+If the environment variable is unset/empty:
+
+- request fails before HTTP call (`ErrInvalidRequest`)
+
+If both `Target.APIKey` and `Target.APIKeyEnv` are empty:
+
+- no `Authorization` header is sent
+
+## Timeout Behavior
+
+Base timeout comes from client configuration.
+
+Per-request override:
+
+- if `Target.TimeoutSeconds > 0`, use that value for request timeout
+- if `Target.TimeoutSeconds == 0` and the value came from an explicit request override, disable the HTTP client timeout
+- if `Target.TimeoutSeconds < 0`, request is rejected (`ErrInvalidRequest`)
+
+## Response Expectations
+
+Expected successful response shape (subset used):
+
+- `choices[0].message.content`
+- `usage.prompt_tokens`
+- `usage.completion_tokens`
+- `usage.total_tokens`
+- `usage.prompt_tokens_details.cached_tokens` (optional)
+- `usage.cache_write_tokens` (optional)
+
+Absent cache usage fields are treated as zero. Parsed cache usage is exposed through run results and adapter response surfaces as:
+
+- `cached_tokens`
+- `cache_write_tokens`
+
+Malformed response conditions include:
+
+- invalid JSON
+- empty `choices`
+- empty `choices[0].message.content`
+
+Malformed responses return `ErrMalformedResponse`.
+
+## Error Handling
+
+- network/request-construction failures: `ErrRequestFailed`
+- non-2xx HTTP status: `ErrUnexpectedStatus` (includes status code; provider response bodies are not included)
+- malformed response shape/content: `ErrMalformedResponse`
+
+## Unsupported Or Non-Serialized Fields
+
+The client does not serialize top-level `cache_control`.
+
+No built-in retries, tool-calls, or multi-request payload modes are implemented in this client.
+
+## Relationship To Runner
+
+When prompt validation mode is `json_schema`, runner prepares a structured-output schema spec and passes it to the client as `StructuredOutput`.
+
+The client only serializes the provider request payload; it does not load schema files itself.

docs/integrations/subprocess.md

@@ -0,0 +1,130 @@
+# Subprocess Integration
+
+This document defines the supported subprocess contract for downstream
+applications invoking Scriptorium through the public CLI.
+
+This is a CLI contract. Go callers that want an in-process typed API should use
+the [package guide](../consumers/pkg-scriptorium.md).
+
+## Supported Commands
+
+Downstream applications should invoke:
+
+- `scriptorium render` for preflight/debug output without LLM execution.
+- `scriptorium run` for generation.
+
+`scriptorium serve` is an HTTP service command, not the recommended subprocess
+contract for per-request execution.
+
+## Recommended Invocation Shapes
+
+Render:
+
+```bash
+scriptorium render \
+  --config <config_path> \
+  --prompt <prompt_id> \
+  --input transcript=<path> \
+  --format json
+```
+
+Run:
+
+```bash
+scriptorium run \
+  --config <config_path> \
+  --prompt <prompt_id> \
+  --input transcript=<path> \
+  --out <artifact_path>
+```
+
+Callers may add:
+
+- `--profile <profile_id>`
+- repeatable `--input name=path`
+- repeatable `--var name=value`
+- runtime overrides when explicitly needed, such as `--model`, `--llm-base-url`, `--api-key-env`, and `--timeout`
+
+Do not pass raw API keys as command arguments.
+
+## Config And Directory Behavior
+
+Callers can rely on resolved app config or pass explicit paths.
+
+Default config search order:
+
+1. `/usr/local/etc/scriptorium/config.yml`
+2. `/etc/scriptorium/config.yml`
+
+Rules:
+
+- Explicit `--config` requires file existence and valid syntax.
+- CLI flags override config values.
+- `run` and `render` require an effective `prompt_dir`.
+- `profile_dir` is optional because built-in profiles are available.
+
+## Profile Selection
+
+Profile selection follows runner behavior:
+
+1. explicit `--profile`
+2. prompt `default_profile`
+3. error if neither is available
+
+Treat prompt and profile IDs as deployment configuration, not hardcoded business
+logic.
+
+## Input And Variable Contract
+
+- Inputs use repeated `--input name=path`.
+- Input names must match prompt definition input names.
+- Variables use repeated `--var name=value`.
+- Both flags also accept comma-separated mappings.
+- Prefer file inputs for large content.
+
+CLI inputs are file references. HTTP-only `inline` references are documented in
+the [HTTP API reference](../api.md).
+
+## Environment Contract
+
+- Pass through required API-key environment variables referenced by `api_key_env`.
+- Keep subprocess environments scoped to required variables.
+- Use `--api-key-env` only to name an environment variable.
+- Never pass raw API keys via argv.
+
+## Stdout And Stderr
+
+`run`:
+
+- stdout: generated artifact body unless `--out` is used.
+- stderr: success summary and errors.
+
+`render`:
+
+- stdout: prepared-run output unless `--out` is used.
+- stderr: errors.
+
+Capture stdout and stderr separately. Do not parse stderr as a stable data
+format beyond exit status handling.
+
+## Exit Status Contract
+
+- `0`: success.
+- `1`: parse, config, load, render, generation, IO, or runtime error.
+- `2`: `run` completed and output was written, but validation failed.
+
+A `run` exit code `2` can still produce output on stdout or at `--out`.
+Consumers must decide whether to keep or discard that output.
+
+## Security Notes
+
+- Treat generated artifacts, rendered prompts, stdout, and stderr as potentially sensitive.
+- Use controlled output paths and access controls for persisted artifacts.
+- Avoid logging full rendered prompts or generated artifacts by default.
+
+## Canonical References
+
+- CLI behavior: [CLI reference](../cli.md)
+- Config and file formats: [Configuration reference](../config.md)
+- Operations: [Operations guide](../operations.md)
+- Troubleshooting: [Troubleshooting](../troubleshooting.md)

docs/internal/adapters.md

@@ -0,0 +1,155 @@
+# Adapter Internals
+
+## Purpose
+
+Adapters translate external interfaces into domain requests and translate domain results back out. They wire dependencies, apply app config, and own IO concerns, but they do not make runner decisions.
+
+Source-loading behavior belongs in `docs/internal/sources.md`. User-facing CLI, HTTP, and package contracts belong in `docs/cli.md`, `docs/api.md`, and `docs/consumers/pkg-scriptorium.md`.
+
+## Adapter Map
+
+- `cmd/scriptorium`: process entrypoint.
+- `internal/adapter/cli`: command parsing, config handoff, runner construction, stdout/stderr, exit codes.
+- `internal/adapter/http`: `POST /v1/runs` request/response mapping and HTTP error/status mapping.
+- root package `scriptorium`: public Go facade over internal runner types and dependencies.
+
+Supporting implementation packages used during adapter wiring:
+
+- `internal/config`
+- `internal/defaults`
+- `internal/format`
+- `internal/llm`
+- `internal/prompt`
+
+## Inputs And Outputs
+
+CLI adapter:
+
+- Input: process args, optional config file, filesystem sources, environment variables.
+- Output: process exit code, stdout artifact/prepared output, stderr summaries and errors.
+
+HTTP adapter:
+
+- Input: HTTP request method/path/headers/body for `POST /v1/runs`.
+- Output: JSON success or error body with mapped status code.
+
+Public Go facade:
+
+- Input: typed `scriptorium.Config`, `Option`, and `RunRequest` values.
+- Output: typed `PreparedRun` and `RunResult` values plus public sentinel errors.
+
+## Boundaries
+
+- Adapters convert external shapes to `domain.RunRequest` and back.
+- Runner orchestration remains in `internal/usecase`.
+- Prompt/profile/schema/artifact source rules remain in repository, validator, and artifact packages.
+- LLM provider request serialization remains in `internal/llm`.
+- Public package types are facade types; internal domain types do not leak across the package boundary.
+
+## Config Fields Used
+
+Adapter app settings:
+
+- `prompt_dir`
+- `profile_dir`
+- `schema_dir`
+- `server.addr`
+- `server.artifact_root`
+- `server.max_request_bytes`
+- `server.max_artifact_bytes`
+- `server.max_response_bytes`
+- `defaults.render_format`
+
+Execution request/profile settings passed through the runner:
+
+- `endpoint`
+- `model`
+- `temperature`
+- `max_tokens`
+- `top_p`
+- `timeout_seconds`
+- `service_tier`
+- `api_key_env`
+- `reasoning_effort`
+- `extra_params`
+
+CLI and HTTP preserve numeric override presence so omitted values and explicit zero values remain distinct.
+
+## CLI Adapter
+
+Implemented commands:
+
+- `run`
+- `render`
+- `serve`
+
+Behavior:
+
+- `run` constructs a runner with direct filesystem artifact reading and calls `Runner.Run`.
+- `render` constructs a runner and calls `Runner.Prepare`; it does not call the LLM.
+- `serve` constructs a restricted artifact reader and HTTP handler, then starts an unauthenticated HTTP server.
+- `run` exits `2` when generation succeeds but validation fails.
+- parse, runtime, and output-write errors exit `1`.
+- deprecated `--prompt-id` and `--profile-id` aliases are accepted.
+
+## HTTP Adapter
+
+Behavior:
+
+- Accepts only `POST /v1/runs`.
+- Decodes JSON strictly and rejects unknown fields and trailing JSON tokens.
+- Rejects empty `prompt_id` and empty `inputs` before calling the runner.
+- Does not accept raw API key values in the request body.
+- Returns validation failures as `200` responses with failed validation details.
+- Maps request-body, artifact, and encoded-response size failures to `413`.
+- Maps domain and repository errors to stable error codes without returning wrapped internal cause text.
+
+The HTTP adapter has no built-in authentication or authorization. Deployment controls must be provided outside the process.
+
+## Public Go Facade
+
+Behavior:
+
+- `NewEngine` wires the same default runner components as CLI/HTTP unless options override them.
+- Prompt, profile, and schema sources may come from directories, single files, or `fs.FS` roots.
+- `WithProfiles` adds in-memory profiles ahead of file-backed and built-in profiles.
+- `WithLLMClient` injects custom model behavior.
+- `RunRequest.APIKey` is request-scoped and direct; it is used only for generation and is stripped from public results.
+- internal errors are mapped to public sentinels in `errors.go`.
+
+## Failure Behavior
+
+Adapters should:
+
+- keep external error payloads concise and stable.
+- avoid leaking raw secret values.
+- use sentinels and typed errors for mapping.
+- preserve strict external input decoding.
+- keep validation content failures distinct from runtime errors.
+
+CLI writes human-readable summaries to stderr. HTTP writes JSON error envelopes. The public Go facade returns typed errors.
+
+## State And Manifests
+
+Adapters do not add durable run state.
+
+- No adapter writes run manifests.
+- No adapter implements checkpoint, skip, or resume behavior.
+- CLI output files are caller-selected artifacts, not internal state.
+
+## Tests To Inspect
+
+- `internal/adapter/cli/run_test.go`
+- `internal/adapter/http/handler_test.go`
+- `engine_test.go`
+- `internal/format/prepared_run_test.go`
+- `internal/llm/openai_compatible_client_test.go`
+
+## Architectural Invariants
+
+- Adapter packages stay thin and translation-focused.
+- App config is resolved before dependency construction.
+- External input strictness is part of contract stability.
+- CLI and HTTP construct runners without a repairer.
+- HTTP endpoint details remain canonical in `docs/api.md`.
+- Public Go package details remain canonical in `docs/consumers/pkg-scriptorium.md`.

docs/internal/runner.md

@@ -0,0 +1,146 @@
+# Runner Internals
+
+## Purpose
+
+`internal/usecase.Runner` is the core prompt-execution orchestrator. It prepares prompt requests, calls the configured LLM client for `Run`, validates generated output, and returns domain results.
+
+Transport parsing, DTOs, CLI output, HTTP status mapping, and public package type conversion belong outside the runner.
+
+## Inputs And Outputs
+
+Primary inputs:
+
+- `domain.RunRequest`
+- repositories/readers/renderers/validators injected at construction
+- `context.Context` for cancellation
+
+Primary outputs:
+
+- `domain.PreparedRun` from `Prepare`
+- `domain.RunResult` from `Run`
+- wrapped sentinel errors for adapter mapping
+
+LLM boundary types:
+
+- `domain.GenerateRequest`
+- `domain.GenerateResponse`
+
+## Dependencies
+
+`Runner` depends on package interfaces instead of concrete adapter types:
+
+- `promptdef.Repository`
+- `profile.Repository`
+- `artifact.Reader`
+- `prompt.Renderer`
+- `llm.Client`
+- `validate.Validator`
+- optional `usecase.OutputRepairer`
+
+The CLI, HTTP adapter, and public Go package construct these dependencies and pass them in.
+
+## Config Fields
+
+`Runner` does not read app config files. Effective behavior is determined by injected dependencies and the `domain.RunRequest`.
+
+Adapter wiring commonly reflects these app config fields:
+
+- `prompt_dir`
+- `profile_dir`
+- `schema_dir`
+- `server.artifact_root`
+- HTTP request/artifact/response size limits
+
+Runtime model settings are resolved from the selected profile plus request overrides.
+
+## Prepare Flow
+
+`Prepare`:
+
+1. requires a non-empty prompt ID.
+2. loads the prompt definition and computes its hash.
+3. selects the profile from request `profile_id`, then prompt `default_profile`.
+4. loads the selected execution profile.
+5. merges built-in execution defaults, profile values, and request overrides.
+6. applies request-scoped direct API key values for public Go callers.
+7. validates endpoint, model, and credential requirements.
+8. resolves the output contract and JSON Schema document when required.
+9. reads input artifacts.
+10. renders prompt messages and hashes the rendered prompt.
+11. returns a prepared run without calling the LLM.
+
+Numeric request overrides are presence-aware: omitted values preserve the current effective value, while explicit zero values are real overrides.
+
+## Run Flow
+
+`Run`:
+
+1. creates a run ID and start timestamp.
+2. calls `Prepare`.
+3. calls the injected LLM client with rendered messages, effective target, target presence, and structured-output settings.
+4. builds the output artifact.
+5. validates the output.
+6. optionally attempts bounded repair when a repairer is injected and the contract permits repair.
+7. returns the run result with artifact, raw output, validation, hashes, selected profile/model metadata, usage, and timing.
+
+`Run` must reuse `Prepare`; prepare logic should not be duplicated elsewhere.
+
+## Validation And Repair
+
+Validation content failures are returned as successful run results with `Validation.Status == failed`. They are not runtime errors.
+
+Validation runtime failures, such as schema load or compile errors, return `ErrValidation`.
+
+Repair attempts occur only when all conditions are true:
+
+- a repairer is injected
+- `repair_attempts` is greater than zero
+- validation status is `failed`
+- validation mode is `json` or `json_schema`
+
+CLI and HTTP wiring call `usecase.NewRunner(...)`, which does not inject a repairer. Normal CLI and HTTP execution therefore does not repair invalid output.
+
+## Failure Behavior
+
+Stable runner sentinels include:
+
+- `ErrInvalidRequest`
+- `ErrProfileRequired`
+- `ErrAPIKeyEnvMissing`
+- `ErrAPIKeyRequired`
+- `ErrPromptLoad`
+- `ErrProfileLoad`
+- `ErrArtifactLoad`
+- `ErrPromptRender`
+- `ErrLLMGenerate`
+- `ErrValidation`
+
+Adapters should use `errors.Is` against sentinels and lower-level repository errors instead of matching message text.
+
+Secret values must not appear in prepared output, run results, logs, HTTP responses, or serialized public package results. The effective API-key environment-variable name may appear.
+
+## State And Manifests
+
+The runner is stateless across requests.
+
+- No durable run store.
+- No manifest files.
+- No checkpoint, skip, or resume behavior.
+- Recovery is a new request after correcting inputs, config, or environment.
+
+## Tests To Inspect
+
+- `internal/usecase/runner_test.go`
+- `internal/usecase/integration_test.go`
+- `engine_test.go`
+- `internal/adapter/cli/run_test.go`
+- `internal/adapter/http/handler_test.go`
+
+## Architectural Invariants
+
+- Use-case decisions stay in `internal/usecase`.
+- `Run` reuses `Prepare`.
+- Prompt/profile/artifact/schema loading remains behind injected boundaries.
+- Validation content failures are result state; validation runtime failures are errors.
+- Repair loops are bounded by `repair_attempts` and repairer presence.
+- Resolved secret values are never serialized or emitted.

docs/internal/sources.md

@@ -0,0 +1,157 @@
+# Source Internals
+
+## Purpose
+
+This document covers implemented prompt, profile, schema, artifact, and catalog source behavior. It is for developers changing loaders or source wiring.
+
+Full user-facing YAML and config reference material belongs in `docs/config.md`.
+
+## Prompt Definition Sources
+
+`internal/promptdef` provides directory-backed and `fs.FS` repositories.
+
+Behavior:
+
+- recursively scans `.yaml` and `.yml` files.
+- decodes YAML with known-fields checking.
+- looks up prompts by YAML `id`, not by path.
+- optionally filters by prompt `version`.
+- rejects duplicate matching prompt IDs.
+- requires `id`, `version`, and at least one message.
+- requires each message to set exactly one of `content` or `content_file`.
+- resolves filesystem `content_file` values relative to the prompt YAML file.
+- resolves `fs.FS` `content_file` values inside the configured source root.
+- permits prompt subdirectories only as organization; they are not part of prompt identity.
+
+For `fs.FS` roots, absolute paths and relative traversal outside the source root are rejected by catalog path helpers.
+
+## Profile Sources
+
+`internal/profile` provides directory-backed, `fs.FS`, and overlay repositories. `internal/profile/builtin` embeds built-in profile YAML assets and exposes them through the same repository interface.
+
+Behavior:
+
+- recursively scans `.yaml` and `.yml` files.
+- decodes YAML with known-fields checking.
+- looks up profiles by YAML `id`, not by path.
+- rejects duplicate IDs inside the same source.
+- rejects raw `api_key` fields in YAML; file-backed profiles must use `api_key_env`.
+- validates required `endpoint` and `model` values.
+- validates numeric profile ranges.
+
+Overlay behavior:
+
+- custom profiles are primary.
+- built-in profiles are fallback.
+- fallback occurs only after a primary `ErrProfileNotFound`.
+- primary validation, YAML, duplicate, and raw-key errors are returned directly.
+- duplicate IDs across custom and built-in sources are allowed because the custom profile overrides the built-in one.
+
+The public Go facade can add in-memory profiles ahead of file-backed and built-in profiles.
+
+## Schema Sources
+
+`internal/validate` provides:
+
+- `StandardValidator` for filesystem paths.
+- `FSValidator` for `fs.FS` roots and single-file public schema sources.
+
+Behavior:
+
+- `json_schema` validation requires a non-empty `schema_path`.
+- filesystem schema paths resolve relative to `schema_dir` unless absolute.
+- directory-backed schema lookup uses the explicit `schema_path`; it does not search recursively by basename.
+- `fs.FS` schema paths must remain inside the configured source root.
+- single-file schema sources match by the configured file base name.
+- schema documents are loaded before the LLM call for structured output.
+- JSON parse failures are validation content failures.
+- schema access, decode, registration, and compile failures are runtime validation errors.
+
+## Artifact Sources
+
+`internal/artifact` supports two input artifact reference types:
+
+- `inline`
+- `file`
+
+Inline behavior:
+
+- requires a non-empty body.
+- produces text/plain artifacts.
+- hashes the body bytes.
+
+Direct file behavior:
+
+- used by CLI `run`, CLI `render`, and the public Go facade.
+- requires a non-empty URI.
+- reads from the process filesystem without HTTP artifact-root restrictions.
+- infers content type from file extension, defaulting to text/plain.
+
+Restricted file behavior:
+
+- used by HTTP `serve`.
+- allows inline artifacts even when no artifact root is configured.
+- denies file artifacts when no artifact root is configured.
+- resolves relative file URIs against `server.artifact_root`.
+- accepts absolute file URIs only when they pass containment checks.
+- applies `server.max_artifact_bytes` when configured.
+
+Restricted containment is lexical. It cleans paths and checks the relative path against the configured root; it does not resolve symlinks. Symlinks inside the root are followed by the operating system, including symlinks that target files outside the root.
+
+## Catalog Helpers
+
+`internal/filecatalog` centralizes shared source helpers:
+
+- recursive YAML discovery for filesystem and `fs.FS` roots.
+- deterministic sorting.
+- `.yaml` and `.yml` filtering.
+- display paths for diagnostics.
+- YAML file stems.
+- `fs.FS` root cleaning and containment checks.
+
+Repository code should use these helpers instead of reimplementing path traversal and containment rules.
+
+## Failure Behavior
+
+Common source failures:
+
+- missing prompt/profile/schema/artifact files.
+- invalid YAML or JSON.
+- unknown YAML fields.
+- duplicate prompt or profile IDs.
+- prompt/profile validation errors.
+- raw API key fields in profile YAML.
+- unsupported artifact reference type.
+- missing inline body or file URI.
+- artifact outside HTTP root.
+- artifact exceeding HTTP size limit.
+- schema load or compile failure.
+
+Prompt/profile repository lookup errors are mapped by adapters separately from runtime runner errors. Validation content failures remain result state; source and schema runtime failures return errors.
+
+## State And Manifests
+
+Source packages do not persist run state.
+
+- No manifests are read or written.
+- No source package implements skip or resume behavior.
+- Source reads reflect the current filesystem or `fs.FS` state for each request.
+
+## Tests To Inspect
+
+- `internal/promptdef/repository_test.go`
+- `internal/profile/repository_test.go`
+- `internal/profile/builtin/repository_test.go`
+- `internal/artifact/reader_test.go`
+- `internal/validate/standard_validator_test.go`
+- `internal/usecase/integration_test.go`
+- `engine_test.go`
+
+## Architectural Invariants
+
+- Prompt/profile identity comes from YAML `id`.
+- External YAML decoding remains strict.
+- File-backed profile YAML never accepts raw API key values.
+- Built-in profiles are fallback, not a replacement for custom source validation.
+- HTTP file artifacts remain rooted by lexical containment.
+- Schema runtime failures remain errors, while JSON/schema content mismatches remain validation results.

docs/operations.md

@@ -0,0 +1,164 @@
+# Operations Guide
+
+## Scope
+
+This guide covers operating the implemented CLI commands and HTTP service. It
+does not replace the [CLI reference](cli.md), [Configuration reference](config.md),
+or [HTTP API reference](api.md).
+
+## Operational Model
+
+Scriptorium executes one prompt request per CLI invocation or HTTP request.
+
+Important boundaries:
+
+- No durable run state is stored.
+- No manifest, archive, checkpoint, or built-in backup workflow is written.
+- No built-in resume behavior exists.
+- Recovery is rerun-based: correct inputs, config, or environment, then run again.
+
+## Filesystem Layout
+
+Operational deployments usually provide:
+
+- `prompt_dir`: prompt definition YAML files and adjacent `content_file` templates.
+- `profile_dir`: optional custom profile YAML files.
+- `schema_dir`: optional JSON Schema files.
+- `server.artifact_root`: optional HTTP file-input root for `serve`.
+
+Keep these directories readable by the Scriptorium process. Keep
+`server.artifact_root` narrow and not writable by untrusted users.
+
+## Normal CLI Workflow
+
+Use `render` before `run` when changing prompt/profile/input wiring:
+
+```bash
+go run ./cmd/scriptorium render \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --format json
+```
+
+Use `run` for generation after preflight:
+
+```bash
+go run ./cmd/scriptorium run \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --out ./summary.md
+```
+
+Before production runs, confirm:
+
+- the effective config path is the intended one;
+- prompt/profile/schema directories are readable;
+- input file paths exist and match prompt input names;
+- required API-key environment variables are set;
+- the selected model endpoint is reachable from the process environment.
+
+## HTTP Service Operation
+
+Start the service with:
+
+```bash
+go run ./cmd/scriptorium serve --config ./examples/config.yml
+```
+
+The implemented HTTP route is `POST /v1/runs`; request and response fields are
+defined in the [HTTP API reference](api.md).
+
+The maintained HTTP request-shape example is `examples/http-run.json`.
+
+HTTP service notes:
+
+- Unknown JSON fields are rejected.
+- `inline` input references work without an artifact root.
+- `file` input references require `server.artifact_root` or `serve --artifact-root`.
+- Request bodies, HTTP file input artifacts, and encoded JSON responses are size-limited.
+- Validation content failures return `200 OK` with `validation.status: "failed"`.
+
+Security boundary:
+
+- `serve` has no built-in authentication or authorization.
+- Put it behind trusted controls such as a private network, authenticated reverse proxy, or API gateway.
+- Do not expose an artifact root containing unrelated sensitive files.
+- Symlinks inside the artifact root are followed by the operating system.
+
+## Secrets Handling
+
+Raw API keys are not accepted in app config, profiles, CLI flags, or HTTP
+request bodies.
+
+Use this pattern:
+
+1. Set an environment variable containing the secret value.
+2. Store only the variable name in profile `api_key_env` or request override `api_key_env`.
+3. Scope the process environment to the minimum required variables.
+
+## Output, Logs, And Exit Codes
+
+`run`:
+
+- stdout: generated artifact body unless `--out` is used.
+- stderr: summary on success, errors on failure.
+- exit `2`: generation completed and output was written, but validation failed.
+
+`render`:
+
+- stdout: prepared-run output unless `--out` is used.
+- stderr: errors.
+- exit `0` on success, `1` on failure.
+
+`serve`:
+
+- stderr: startup and server errors.
+- HTTP response body: JSON success or error envelope.
+
+## Validation Behavior
+
+Prompt `output.validation_mode` controls validation:
+
+- `none`: skipped.
+- `basic`: output body must not be empty.
+- `json`: output body must parse as JSON.
+- `json_schema`: output body must parse as JSON and satisfy the configured schema.
+
+Runtime/schema failures are hard failures (`run` exit `1`, HTTP error).
+Generated-content validation failures are soft failures (`run` exit `2`, HTTP
+`200 OK` with failed validation status).
+
+## Size Limits
+
+Defaults are documented in [Configuration reference](config.md). Operationally:
+
+- Keep default HTTP limits unless larger payloads are measured and expected.
+- Prefer `inline` HTTP inputs for small payloads.
+- Prefer `file` HTTP inputs for larger local artifacts under a controlled artifact root.
+- Increase `server.max_response_bytes` when generated artifacts or requested raw output are expected to be large.
+- Use `0` only when another trusted layer enforces size limits.
+
+## Maintained Examples
+
+- `examples/config.yml`
+- `examples/config.full.yml`
+- `examples/render-markdown-summary.sh`
+- `examples/http-run.json`
+
+## Safe Recovery
+
+For failed CLI commands or HTTP requests:
+
+1. Capture stderr or the HTTP error `code` and `message`.
+2. Confirm config path and effective directory settings.
+3. Verify prompt ID, profile ID, schema path, and input mappings.
+4. Verify required API-key environment variables.
+5. Reproduce with `render --format json` when pre-LLM resolution is uncertain.
+6. Rerun after correction.
+
+Because Scriptorium does not persist run state, rerun is the supported recovery
+path.

docs/policy/architecture.md

@@ -0,0 +1,118 @@
+# Architecture
+
+This document is the development architecture policy for Scriptorium.
+
+It is for developers and LLM coding agents. User-facing behavior belongs in `README.md` and the docs under `docs/` that target operators/users.
+
+## Project Shape
+
+Scriptorium is a narrow prompt-execution application with three entry paths:
+
+- CLI `run`
+- CLI `render`
+- HTTP `POST /v1/runs` through `serve`
+- public Go package `gitea.maximumdirect.net/eric/scriptorium`
+
+Domain behavior is centralized in `internal/usecase` and `internal/domain`.
+
+## Core Principles
+
+- Keep orchestration narrow: Scriptorium executes one prompt request; it is not a multi-step workflow engine.
+- Keep adapter logic thin: adapters map external shapes to domain requests/results and should not hold domain decisions.
+- Keep boundaries explicit: repositories/loaders/renderers/validators/LLM client stay behind package interfaces.
+- Keep config strict: YAML/JSON decoding for external inputs should reject unknown fields.
+- Keep secrets out of payloads: raw API key values must not be accepted or emitted.
+
+## Package Boundaries
+
+Current package map:
+
+- root package `scriptorium`: public Go facade over engine construction, source options, request/result types, and error mapping.
+- `cmd/scriptorium`: process entrypoint.
+- `internal/adapter/cli`: command parsing, app wiring for CLI commands, output behavior.
+- `internal/adapter/http`: HTTP DTO mapping and error/status mapping.
+- `internal/config`: application settings loading and CLI override precedence.
+- `internal/defaults`: compile-time default constants.
+- `internal/domain`: core request/result and contract types.
+- `internal/usecase`: `Runner` prepare/run orchestration and repair-hook boundary.
+- `internal/promptdef`: filesystem prompt-definition repository.
+- `internal/profile`: filesystem, `fs.FS`, and overlay execution-profile repositories.
+- `internal/profile/builtin`: embedded built-in execution profiles.
+- `internal/filecatalog`: shared YAML discovery and `fs.FS` source helpers.
+- `internal/artifact`: artifact reference readers.
+- `internal/prompt`: template renderer.
+- `internal/llm`: provider-neutral LLM client interface and OpenAI-compatible implementation.
+- `internal/validate`: validator interfaces and standard implementation.
+- `internal/format`: prepared-run output formatting.
+
+Detailed component behavior is documented in:
+
+- `docs/internal/runner.md`
+- `docs/internal/adapters.md`
+- `docs/internal/sources.md`
+
+## Configuration And Precedence
+
+Application settings are resolved as:
+
+1. built-in defaults
+2. config file values
+3. CLI overrides
+
+`config.yml` is for application wiring (directories, server address, render default format), not prompt/profile runtime execution settings.
+
+Profile selection and runtime model resolution remain use-case concerns.
+
+## State And Persistence Policy
+
+Scriptorium has no durable run-state store.
+
+- No built-in resume/checkpoint/archive behavior.
+- Recovery model is rerun after correcting inputs/config/environment.
+
+## External Integration Policy
+
+Current external contracts:
+
+- inbound HTTP contract: `POST /v1/runs`, documented canonically in `docs/api.md`
+- outbound model contract: OpenAI-compatible chat completions subset
+- subprocess contract for integrators: CLI `run`/`render`
+- public Go package contract: `docs/consumers/pkg-scriptorium.md`
+
+Integration docs belong under `docs/integrations/`.
+
+## Error Handling And Logging
+
+- Wrap errors with domain/operation context.
+- Map domain errors to adapter-appropriate statuses/codes without leaking sensitive internals.
+- Keep stderr summaries concise for CLI success/error paths.
+- Never emit raw secret values.
+
+## Testing Expectations
+
+- Core runner behavior should be covered with isolated unit tests and fixture-based integration tests.
+- Adapter behavior should be tested for parse/mapping/error semantics.
+- Config parsing, prompt/profile loading, validator behavior, and LLM client error handling should remain covered by package tests.
+- Repository-level docs/examples that claim runnable behavior should be validated by tests or smoke commands.
+
+## Documentation Expectations
+
+- Document implemented behavior only outside `docs/roadmap/`.
+- Keep canonical reference locations stable (`docs/cli.md`, `docs/config.md`, `docs/operations.md`, `docs/troubleshooting.md`, `docs/internal/`).
+- Update docs in the same change when architecture-relevant behavior changes.
+
+## Architectural Invariants
+
+- `Runner.Run` reuses `Runner.Prepare` flow.
+- CLI and HTTP currently instantiate `Runner` without a repairer.
+- Artifact reading supports `inline` and `file` references.
+- Unknown input fields in config/prompt/profile/http JSON should be rejected by strict decoding.
+- Raw API key values must not be accepted through config/HTTP payloads.
+
+## Non-Goals
+
+- Do not move orchestration responsibilities from external callers into Scriptorium.
+- Do not add adapter-specific business logic in `internal/adapter/*` packages.
+- Do not bypass repository/renderer/validator/LLM boundaries by introducing cross-package coupling.
+
+Work that is not implemented belongs in `docs/roadmap/`.

docs/policy/documentation.md

@@ -0,0 +1,163 @@
+# Documentation Policy
+
+## Purpose
+
+This policy assigns each documentation topic to one canonical owner. Its goal is
+to keep this repository's documentation accurate, concise, discoverable, and
+resistant to drift for users, operators, developers, integrators, and LLM
+coding agents.
+
+## Core Rules
+
+### One Canonical Owner
+
+Each authoritative fact belongs in one document. A non-owning document may give
+a short, stable summary for orientation, but it must link to the canonical owner
+instead of repeating volatile details.
+
+Volatile details include commands, flags, configuration fields and defaults,
+module keys, schemas, file names, paths, status codes, retry behavior, and
+runtime guarantees. If readers could reasonably treat a statement as a
+contract, maintain it only in the owning document.
+
+Minimal tested usage examples are allowed outside the owning contract when this
+policy assigns them an orientation or instructional purpose. They must link to
+the canonical contract and must not redefine complete syntax, defaults, or
+semantics.
+
+### Current And Future Behavior
+
+Outside `docs/roadmap/`, documentation describes implemented behavior only.
+Partial features may be described only to their implemented boundary.
+
+ADRs are the narrow exception: an ADR may record an accepted architectural
+decision before implementation, but acceptance must not be presented as proof
+that the behavior exists. The roadmap owns implementation status and sequencing
+until the decision is implemented. Current architecture, user, operator,
+integration, and internal documentation are updated when the behavior lands.
+
+### Audience And Detail
+
+Write for the document's stated audience and include only the detail needed for
+its owned topic. User and operator docs should not expose implementation detail.
+Developer docs should link to user-facing and external contracts rather than
+restate them.
+
+### Examples
+
+Complete copyable files belong in `examples/`. Documentation may use the
+smallest illustrative snippet needed to explain its owned topic, but should link
+to maintained examples instead of embedding a second complete copy.
+
+Examples must be valid, secret-free, and tested where practical. Commands and
+configuration used in documentation should match the application.
+
+### Security And Privacy
+
+Documentation and examples must not contain real credentials, private keys,
+private environment dumps, sensitive source material, or private infrastructure
+details unless intentionally public. Document secret-handling mechanisms, not
+secret values.
+
+## Canonical Ownership
+
+| Topic | Canonical owner | Owned content | Content owned elsewhere |
+| --- | --- | --- | --- |
+| Product orientation and minimal end-to-end quickstart | `README.md` | What this project is, why it is useful, one shortest successful invocation, and links onward. | Complete command reference, configuration reference, operational procedures, implementation detail. |
+| Contributor entry point | `docs/development.md` | Task-oriented reading guide, minimal contributor orientation, baseline validation commands, and links to canonical docs. | Package inventory, architecture rules, subsystem behavior, and detailed change recipes, which belong in the relevant internal component document. |
+| Current application architecture | `docs/policy/architecture.md` | System shape, normative ownership, dependency direction, architectural boundaries, invariants, safety properties, and non-goals. | Concrete package inventory, implementation mechanics, contributor procedures, decision history, future work. |
+| Documentation organization | `docs/policy/documentation.md` | Documentation ownership, audience boundaries, maintenance rules, and ADR/document lifecycle. | Application architecture or product behavior. |
+| Testing policy | `docs/policy/testing.md` | Test philosophy, risk-based sufficiency, test boundaries, doubles, coverage guidance, regression-test policy, and criteria for adding, rewriting, or deleting tests. | Subsystem behavior, application contracts, subsystem-specific test inventories, and implementation plans. |
+| CLI contract | `docs/cli.md` | Commands, arguments, flags, invocation semantics, and exit codes. | End-to-end operating procedures, configuration field definitions, runtime filesystem layout, module implementation details. |
+| Configuration contract | `docs/config.md` | Discovery and precedence, file schema, fields, defaults, environment overrides, validation rules, and user-selectable module or validator keys. | Complete example files, CLI syntax, runtime state lifecycle, module implementation details. |
+| Operations | `docs/operations.md` | Runtime workflows, physical filesystem and state layout, output, cache, and debug handling, resume, cleanup, permissions, recovery, and operational limits. | CLI flag syntax, configuration field definitions, logical output schemas, implementation mechanics. |
+| Public HTTP contract | `docs/api.md` | Routes, authentication, media types, request and response schemas, status codes, pagination, caching, idempotency, rate limits, and HTTP retry semantics. | Client walkthroughs, upstream or downstream integration internals, implementation detail. |
+| Consumer guidance | `docs/consumers/` | Task-oriented use of the public interface, minimal client examples, and consumer responsibilities. | HTTP wire semantics, external protocol contracts, internal implementation detail. |
+| External and durable integration contracts | `docs/integrations/` | External file formats and protocols, upstream and downstream contracts, logical output bundle paths and schemas, media types, and compatibility behavior. | Physical runtime placement and lifecycle, internal transformations, CLI syntax, configuration defaults. |
+| Implemented component inventory | `docs/internal/overview.md` | Current packages and components, their implemented responsibilities, and links to focused internal docs. | Normative architecture, contributor reading policy, external contracts. |
+| Internal component behavior | Other files under `docs/internal/` | Implementation flow, internal collaborators and state transitions, package-local guarantees and failures, and relevant tests. | Global architecture invariants, configuration definitions and defaults, external schemas, operator procedures. |
+| Architectural decision history | `docs/adr/` | Significant decisions, context, alternatives, rationale, consequences, and supersession history. | Current behavior reference, implementation status, task sequencing. |
+| Future work and implementation status | `docs/roadmap/` | Proposed, accepted, deferred, or rejected work; implementation status; sequencing; and task breakdowns. | Implemented behavior reference and architectural decision rationale. |
+| Complete copyable artifacts | `examples/` | Maintained configuration, inputs, and other files intended to be copied or run. | Field-by-field reference, command reference, prose explanation. |
+
+Documents that do not exist are required only when the corresponding interface
+or responsibility exists. Do not create placeholder API, consumer, integration,
+or operations documents for behavior the application does not have.
+
+## Boundary Rules
+
+### Orientation
+
+The README owns product orientation. The developer guide routes contributors.
+Architecture owns normative structure. Internal overview owns the current
+concrete component map. These documents may link to one another but should not
+maintain parallel package or behavior descriptions.
+
+### Commands, Configuration, And Operations
+
+CLI documentation answers how to invoke the application. Configuration
+documentation answers what settings mean. Operations answers what happens to
+runtime state and how to operate or recover the application. When a workflow
+crosses these topics, choose the document that owns the task and link to the
+other contracts.
+
+### Contracts And Implementation
+
+Integration and API documents define externally observable shapes and
+semantics. Internal documents explain how this project implements or consumes
+those contracts. Internal docs may name a field, file, or protocol to identify
+a dependency, but must link to its canonical contract for the definition.
+
+### Security Topics
+
+This policy owns what documentation and examples may contain. Architecture owns
+application security invariants. Configuration owns credential-supply
+mechanisms. Operations owns permissions and handling of sensitive runtime
+artifacts. Internal docs own implementation mechanisms only.
+
+## Architecture Decision Records
+
+Use sequentially numbered ADR filenames such as
+`0001-record-architecture-decisions.md`. Follow the lightweight Nygard format:
+
+1. title;
+2. status;
+3. date;
+4. context;
+5. decision;
+6. alternatives considered;
+7. consequences.
+
+Use one of these statuses:
+
+- **Proposed:** the decision is under consideration and may change;
+- **Accepted:** the decision is approved, whether or not implementation is
+  complete;
+- **Rejected:** the proposed decision was considered and not adopted;
+- **Superseded:** a later ADR replaces the accepted decision.
+
+A proposed ADR transitions to accepted or rejected. An accepted ADR transitions
+to superseded only when a later accepted ADR replaces it. An ADR may be created
+as accepted when the decision has already been made.
+
+Treat the decision content of an accepted ADR as immutable. Its status and
+supersession metadata may be updated, but a changed decision requires a new ADR.
+A superseded ADR must link to its replacement, and the replacement must link
+back to the superseded ADR. Rejected architectural alternatives belong in the
+ADR; rejected product ideas belong in the roadmap.
+
+## Maintenance
+
+When behavior changes, update its canonical owner in the same change. If
+ownership moves, remove the old definition and replace it with a link where
+navigation remains useful.
+
+Before completing documentation work:
+
+- verify affected behavior and examples;
+- check commands, flags, fields, defaults, schemas, and paths against their
+  implementation;
+- keep unimplemented behavior in the roadmap, subject to the ADR exception;
+- remove stale references and validate links;
+- confirm that non-owning documents summarize and link rather than redefine;
+- confirm that no secrets or sensitive private data were added.

docs/policy/testing.md

@@ -0,0 +1,296 @@
+# Testing Policy
+
+## Purpose
+
+Our tests exist to make **incorrect changes expensive and correct changes cheap**.
+
+We do not optimize for test count, line coverage, exhaustive isolation, or the fewest possible tests. We optimize for sufficient confidence in important behavior while imposing as little unnecessary friction as possible on future development.
+
+## Every test has a cost
+
+Testing is not an unqualified good. Every test imposes both an immediate cost and a continuing lifetime cost.
+
+A test must be:
+
+- written and reviewed;
+- understood by future maintainers and coding agents;
+- executed in local and CI workflows;
+- diagnosed when it fails;
+- updated when legitimate behavior changes;
+- maintained as fixtures, APIs, and dependencies evolve; and
+- removed or rewritten when it becomes redundant, brittle, misleading, or obsolete.
+
+Tests also create cognitive and architectural friction. They can constrain refactoring, duplicate policy, slow feedback loops, add noise to failures, and cause harmless implementation changes to require unrelated edits across the suite.
+
+A test is warranted only when the confidence it provides justifies these costs.
+
+Apply this cost-benefit analysis at two levels:
+
+1. **Per test:** What realistic defect does this test detect, how consequential would that defect be, and is that protection worth the test's lifetime cost?
+2. **Across the suite:** Does this collection provide materially more confidence than a smaller, simpler suite would?
+
+The preferred test suite is a **lean suite that provides sufficient confidence in the risks that matter, without redundant or low-value tests**. We seek sufficient confidence with the least unnecessary testing friction, not the fewest possible tests.
+
+Some friction is intentional. Tests should make dangerous changes—such as breaking compatibility, corrupting data, violating security boundaries, or reintroducing subtle bugs—require deliberate review. They should not make ordinary internal changes needlessly expensive.
+
+The cost of a test is not a reason to omit testing by default. Do not cite maintenance cost abstractly. When omitting a plausible test, be able to state why the protected failure is low-risk, already covered, obvious, reversible, or cheaper to detect elsewhere. For consequential, subtle, or difficult-to-observe behavior, the presumption should favor testing.
+
+## Default testing style
+
+Use a **classical/Detroit-style** approach:
+
+- Test observable behavior, resulting state, contracts, and invariants.
+- Use real internal collaborators when they are fast and deterministic.
+- Use fakes, stubs, or mocks primarily at expensive, nondeterministic, destructive, or external boundaries.
+- Prefer package-level behavioral tests over tests coupled to private helpers or internal call sequences.
+- Treat exact collaborator interactions as testable behavior only when the interaction itself is a requirement.
+
+Examples of appropriate seams include clocks, randomness, subprocesses, remote APIs, object storage, email, and paid LLM calls.
+
+## Test execution requirements
+
+Tests in the default suite must be deterministic, offline, and independent of real credentials. They must not invoke paid APIs or depend on mutable external services. Tests that require live infrastructure must be explicitly opt-in and clearly separated from the default suite.
+
+Control clocks, randomness, environment variables, and other process-global or machine-specific state when they affect behavior. Tests should be safe to run repeatedly and alongside other tests without depending on execution order or state left by an earlier test.
+
+## What deserves tests
+
+Prioritize tests for:
+
+1. Public and package-level contracts.
+2. Domain rules and important invariants.
+3. Boundary conditions and malformed input.
+4. Failure handling, cancellation, retries, recovery, and partial success.
+5. Serialization, schemas, compatibility, and round trips.
+6. Previously observed or plausible regressions.
+7. Representative integration and end-to-end workflows.
+
+A package-level contract is behavior relied upon by another package or major collaborator, not every observable detail of a package implementation.
+
+For behavior involving **data integrity, destructive operations, compatibility, security, concurrency, idempotency, or recovery**, presume that durable tests are required unless the behavior is already credibly protected at another layer.
+
+Do not add tests merely because a function, branch, or line exists. Do not add a test when the same meaningful risk is already adequately protected elsewhere.
+
+## Choose the right test boundary
+
+Test through the narrowest stable boundary that expresses the behavior clearly.
+
+This is often the package API, but it may instead be:
+
+- a smaller pure function when dense domain logic is most clearly isolated there;
+- a package-level operation when several internal collaborators jointly produce the behavior; or
+- a larger integration boundary when correctness emerges from interaction with a real dependency.
+
+Do not force all behavior through oversized end-to-end tests. Do not test every private helper merely because it exists. Choose the boundary that gives durable confidence with the least incidental coupling.
+
+## Test behavior, not implementation
+
+A test should protect a decision, contract, or invariant—not memorialize the current implementation.
+
+Before adding or retaining a test, ask:
+
+> What realistic defect would this test catch?
+
+A test is suspect when its main purpose is to detect that someone:
+
+- changed an internal constant;
+- renamed or split a private helper;
+- reordered equivalent internal operations;
+- changed incidental formatting;
+- replaced one correct algorithm with another; or
+- refactored internal object structure without changing behavior.
+
+Refactoring should normally require no test edits unless the refactored structure is itself part of the contract.
+
+A test can be factually correct and still have negative value. Accurately describing current behavior is not enough; the protected behavior must be important enough to justify the future friction.
+
+## Expected effects of different changes
+
+Use the following expectations when evaluating test failures and test maintenance:
+
+| Change | Expected effect on tests |
+|---|---|
+| Internal refactor that preserves behavior | Existing tests should normally remain unchanged and continue to pass. |
+| Change to an internal default with no contractual significance | Behavioral tests should normally remain unchanged; tests should derive expectations from configuration or relationships rather than duplicate the old value. |
+| Intentional change to public behavior, policy, schema, or compatibility guarantees | The relevant tests should be reviewed and changed deliberately. |
+| Accidental violation of a contract or invariant | Tests should fail; fix the production code rather than rewriting the tests to accept the defect. |
+
+A test failing is not the same as a test needing to be edited. Many tests may correctly fail because of one production defect. The maintenance smell is a correct internal change that requires unrelated expectation updates throughout the suite.
+
+## Separate mechanism from policy
+
+Configurable thresholds and defaults must not be duplicated throughout the test suite.
+
+For example, do not encode an internal concurrency limit indirectly:
+
+```go
+// Production policy:
+const maxConcurrency = 4
+
+// Brittle test:
+err := startProcesses(5)
+require.Error(t, err)
+```
+
+Instead, test the mechanism relationally:
+
+```go
+const limit = 2
+runner := NewRunner(limit)
+
+require.NoError(t, runner.Start(limit))
+require.ErrorIs(t, runner.Start(limit+1), ErrTooMuchConcurrency)
+```
+
+The test should prove:
+
+- the configured limit is accepted; and
+- one beyond the configured limit is rejected.
+
+The production default should be tested exactly only when its literal value is itself a public, operational, safety, protocol, or compatibility requirement.
+
+Apply the same rule to limits, timeouts, capacities, retry counts, and ranges: test relationships and behavior, not duplicated literals.
+
+For concurrency limits, test both kinds of behavior when relevant:
+
+1. **Configuration enforcement:** invalid or excessive requested values are handled correctly.
+2. **Runtime enforcement:** observed peak concurrency never exceeds the configured limit.
+
+Use a test-controlled limit and measure the behavior relative to that limit. Do not merely assert today's default value.
+
+## Avoid semantic duplication across layers
+
+Each behavior should have a clear test owner.
+
+- Parser tests own parsing cases.
+- Validator tests own validation rules.
+- Domain tests own transformations and invariants.
+- Adapter tests own external integration behavior.
+- Orchestrator tests own coordination and failure propagation.
+- CLI tests own argument and configuration mapping.
+- End-to-end tests prove that representative assembled workflows work.
+
+Higher-level tests should not repeat every lower-level case. A single intentional policy change should not require unrelated edits across many test files.
+
+Tests that are individually reasonable may still be collectively redundant. Evaluate the marginal value of each additional test in light of the protection already provided by the rest of the suite.
+
+## Use test doubles deliberately
+
+Choose the least elaborate test double that provides the required control or observation.
+
+As a default:
+
+1. Prefer real collaborators when they are fast and deterministic.
+2. Use small in-memory fakes when realistic stateful behavior is helpful.
+3. Use stubs when a dependency only needs to provide controlled responses.
+4. Use mocks when the interaction itself is contractual.
+
+Mocks are appropriate when the contract includes facts such as:
+
+- a notification is sent exactly once;
+- a transaction is committed only after successful writes;
+- cancellation reaches a subprocess;
+- an expensive API is called no more than once; or
+- a security audit event is emitted.
+
+Do not use mocks merely to isolate every object or reproduce the implementation's call graph.
+
+## Go-specific guidance
+
+Use:
+
+- table-driven tests for meaningful behavioral categories and boundaries;
+- `t.TempDir()` for real filesystem behavior;
+- `httptest.Server` for realistic HTTP interactions;
+- fuzz tests for parsers, normalization, path handling, and broad input spaces;
+- golden files only when the complete output is intentionally stable;
+- integration tests where correctness depends on component interaction; and
+- a small number of representative end-to-end tests.
+
+Avoid exact error-string assertions unless the wording is itself contractual. Prefer `errors.Is`, `errors.As`, typed errors, or structured error fields.
+
+At CLI boundaries, prefer exit classifications, structured output, and the smallest stable semantic fragment needed to identify the error. Do not snapshot complete diagnostic wording unless it is contractual.
+
+Golden-file updates must require an explicit local flag. CI must not update golden files automatically, and reviewers must inspect the semantic diff before accepting an update.
+
+Keep tests readable and direct. Test helpers and fixture frameworks must earn their own maintenance cost; do not build elaborate test infrastructure for small or isolated needs.
+
+## Coverage
+
+Coverage is a diagnostic, not a target.
+
+Use it to find untested critical branches and unexpectedly weak packages. Do not write low-value tests solely to increase a percentage, and do not infer test quality from coverage alone.
+
+Pure domain logic will often warrant higher coverage than CLI wiring or external adapters. Uneven coverage is acceptable when it reflects risk.
+
+Increasing coverage is valuable only when the newly covered behavior protects a meaningful risk at an acceptable cost.
+
+## Regression tests
+
+A bug fix should normally include a regression test that fails before the fix and passes afterward.
+
+Retain the test when the defect could realistically recur and its consequences justify the ongoing cost. Prefer the narrowest durable test of the violated contract or invariant; do not preserve accidental implementation details from the original bug.
+
+Not every historical bug requires a permanent test. If the underlying design has made recurrence impossible, the test has become redundant, or a stronger invariant test now subsumes it, remove or consolidate it.
+
+## Deleting or rewriting tests
+
+Tests are maintained code, not permanent historical artifacts.
+
+Delete or rewrite a test when its maintenance cost exceeds the confidence it provides.
+
+Strong candidates include tests that:
+
+- require updates after harmless internal changes;
+- directly assert private constants without protecting a real contract;
+- duplicate the same policy across several layers;
+- verify mock choreography rather than outcomes;
+- snapshot large amounts of incidental output;
+- test trivial private helpers already exercised through stable package behavior;
+- protect risks already covered more effectively elsewhere;
+- are flaky, misleading, obsolete, or disproportionately expensive to diagnose; or
+- no longer correspond to a plausible failure mode.
+
+Several brittle tests may encode one genuine requirement. Replace them with one durable behavior-level or invariant test rather than preserving all of them.
+
+Deleting a low-value test can improve the quality of the suite by reducing noise, maintenance burden, and friction around legitimate change.
+
+## Reviewing a proposed test
+
+Use the following questions when the value, boundary, or durability of a proposed test is not self-evident. Significant test additions should be reviewable against them, but written answers are not required for every routine test.
+
+1. What realistic defect would it catch?
+2. How likely is that defect?
+3. How consequential would it be?
+4. Is the behavior already protected elsewhere?
+5. At which layer should this behavior be owned?
+6. Does the test assert a durable contract or an incidental implementation detail?
+7. Could the implementation be refactored without changing the behavior and without editing this test?
+8. What should cause this test to fail?
+9. What legitimate changes should not cause this test to fail?
+10. What ongoing maintenance, execution, and diagnostic cost will the test impose?
+11. Is there a smaller or more direct test that protects the same risk?
+
+Do not add the test when its expected lifetime cost exceeds its expected protective value.
+
+When deciding not to test plausible behavior, record or be able to explain why the risk is low, already protected, obvious, reversible, or cheaper to detect elsewhere.
+
+## Definition of sufficient
+
+A test suite is sufficient when:
+
+- important contracts and invariants are protected;
+- meaningful boundaries and failure modes are exercised;
+- realistic and consequential regressions are credibly protected against silent recurrence;
+- behavior involving data integrity, destructive operations, compatibility, security, concurrency, idempotency, and recovery is credibly protected;
+- important external boundaries have realistic integration coverage;
+- representative complete workflows are tested;
+- failures provide useful signal rather than redundant noise;
+- legitimate internal changes usually do not require test edits; and
+- additional tests would mostly repeat existing protection or preserve inconsequential implementation details.
+
+Sufficiency is a risk judgment, not a coverage percentage or test count. Reassess it as the application, its users, and the consequences of failure evolve.
+
+The governing rule is:
+
+> Test heavily where failure is consequential, subtle, or difficult to detect after the fact. Test lightly where failure is obvious, reversible, and inexpensive—and retain no test whose lifetime cost exceeds the confidence it provides.

docs/roadmap/documentation.md

@@ -0,0 +1,510 @@
+# Documentation Compliance Roadmap
+
+## Status
+
+Proposed implementation plan. This document records the findings of the
+documentation audit performed after adoption of the canonical-ownership policy.
+The revisions described here are not yet implemented.
+
+## Objective
+
+Bring the current Scriptorium documentation into compliance with
+`docs/policy/documentation.md` before beginning the Promptkit migration.
+
+The refresh should:
+
+- give every topic one canonical owner;
+- remove parallel definitions of volatile contracts;
+- correct current factual discrepancies;
+- preserve useful contributor and operational guidance in the appropriate
+  documents;
+- establish the missing internal overview and ADR history;
+- leave Promptkit and other future behavior in roadmap and ADR documents until
+  implemented.
+
+This roadmap covers documentation organization and current-behavior accuracy.
+A substantive review of `docs/policy/testing.md` remains separate.
+
+## Audit Baseline
+
+The audit covered:
+
+- `README.md`;
+- all Markdown files under `docs/`;
+- all maintained files under `examples/`;
+- the CLI parser and output behavior;
+- app-config shapes and defaults;
+- HTTP DTOs, limits, and error mappings;
+- the public Go facade;
+- prompt, profile, schema, artifact, renderer, runner, and LLM behavior;
+- the embedded built-in profile assets.
+
+At the time of the audit:
+
+- all local Markdown link targets existed;
+- `go test ./...` passed;
+- the README/CLI render command completed successfully;
+- `examples/render-markdown-summary.sh` completed successfully;
+- `go run ./examples/go-library/prepare` completed successfully.
+
+The dominant problem is duplicated ownership rather than broad factual
+staleness. CLI, configuration, HTTP, validation, state, and security behavior
+are repeated across contracts, operations, troubleshooting, integrations,
+consumer guides, architecture, and internal documents.
+
+## Required Accuracy Corrections
+
+Make these corrections while revising the owning documents:
+
+1. `docs/policy/architecture.md` says Scriptorium has three entry paths but
+   lists four. Distinguish the three executable entry paths from the public Go
+   package, or describe four total entry paths.
+2. The built-in profile catalog in `docs/config.md` omits
+   `deepseek-4-flash`. Reconcile the complete catalog with
+   `internal/profile/builtin/assets/`.
+3. `docs/config.md` marks prompt `output.repair_attempts` as required. The
+   loader permits omission and resolves it to zero; document it as optional
+   with an effective default of zero.
+4. `docs/consumers/pkg-scriptorium.md` says every input map key must match a
+   declared prompt input. The renderer enforces required declared inputs and
+   resolves names actually referenced by templates, but it does not reject
+   every undeclared extra input. Describe the implemented boundary.
+5. `docs/troubleshooting.md` says validation errors can be inspected in CLI
+   stderr. The CLI success summary reports the number of validation errors, not
+   their detailed messages. Correct the diagnostic guidance or separately
+   change the product before documenting richer output.
+6. `docs/integrations/openai-compatible-chat.md` describes the `session_id`
+   limit as characters. The implementation counts Unicode code points; use the
+   canonical terminology consistently.
+7. The same integration document makes a time-sensitive statement about
+   provider-supported service-tier values even though Scriptorium accepts and
+   forwards any non-empty value. Remove the provider catalog claim or cite and
+   version an intentionally maintained external contract.
+8. The `render` and `serve` flag sections in `docs/cli.md` list
+   `--prompt-dir` both as an effective requirement and again as an optional
+   flag. Keep one complete flag entry and separately explain how the
+   requirement may be satisfied.
+9. Troubleshooting command examples containing placeholders such as
+   `<prompt-id>` are not directly copyable shell commands. Use clearly defined
+   shell variables, concrete maintained examples, or prose diagnostic steps.
+
+## Canonical Structure Changes
+
+### Create `docs/internal/overview.md`
+
+Create the canonical implemented-component inventory and move the concrete
+package map out of architecture.
+
+The overview should:
+
+- list current public, command, adapter, domain, use-case, source, format,
+  validation, and LLM components;
+- give each component a short implemented responsibility;
+- link to focused internal documents and relevant external contracts;
+- avoid restating global architecture rules or user-facing behavior.
+
+Update `docs/development.md` to route general repository-orientation work
+through the overview once it exists.
+
+### Establish `docs/adr/`
+
+Create the ADR directory under the policy already defined in
+`docs/policy/documentation.md`.
+
+Record the significant accepted documentation-ownership decision in an initial
+ADR if historical rationale is useful. The Promptkit split ADR remains part of
+Step 2 in `docs/roadmap/migration.md` and should not be pulled ahead of its
+gate merely to populate the directory.
+
+ADRs must own decision context, alternatives, rationale, and consequences.
+Roadmaps must continue to own implementation status and sequencing.
+
+### Consolidate Troubleshooting Into Operations
+
+The ownership table assigns recovery to `docs/operations.md` and does not give
+troubleshooting a separate canonical owner. Fold the useful
+symptom/diagnosis/safe-recovery material from `docs/troubleshooting.md` into a
+concise operations section, then remove `docs/troubleshooting.md`.
+
+The consolidated material should:
+
+- organize failures by operational task rather than duplicate every contract
+  field and status;
+- link exact CLI syntax and exit codes to `docs/cli.md`;
+- link config fields and defaults to `docs/config.md`;
+- link HTTP codes and schemas to `docs/api.md`;
+- describe only diagnostic and recovery actions locally.
+
+Update README, development, integration, and other links after the removal.
+
+## File-By-File Revision Catalog
+
+### `README.md`
+
+- Retain the product description and one tested end-to-end quickstart.
+- Keep the quickstart linked to the complete CLI contract.
+- Replace the repeated example inventory with a short link to the maintained
+  examples section in the appropriate canonical contract, or make each retained
+  path a direct link.
+- Keep the documentation list navigational; do not summarize contracts there.
+- Remove the troubleshooting link if troubleshooting is consolidated into
+  operations.
+
+### `docs/development.md`
+
+- Retain only contributor orientation, task-specific routing, and baseline
+  validation.
+- Add `docs/internal/overview.md` to the appropriate reading paths once it
+  exists.
+- Update the operations/troubleshooting row after consolidation.
+- Link to actual ADRs when created; do not duplicate their decisions.
+- Confirm that all detailed recipes removed from the former development guide
+  have an internal owner before declaring the refresh complete.
+
+### `docs/policy/architecture.md`
+
+- Fix the entry-path count.
+- Replace the concrete package inventory with a link to
+  `docs/internal/overview.md`.
+- Remove config precedence, field-level behavior, exact routes, and exact
+  import-path contracts except for the smallest orientation summary and links
+  to their owners.
+- Remove the local testing checklist and link to
+  `docs/policy/testing.md`. Component-specific test inventories remain in
+  internal docs.
+- Remove the local documentation checklist and link to
+  `docs/policy/documentation.md`.
+- Keep normative system shape, ownership, dependency direction, state
+  philosophy, security properties, error-handling principles, invariants, and
+  non-goals.
+- Review the coding and dependency rules removed from the former development
+  guide. Preserve still-valid normative rules here without listing the current
+  dependency inventory, which is concrete implementation information.
+- Move implementation facts such as current repairer wiring to the relevant
+  internal component document unless they are intentionally elevated to
+  architecture invariants.
+
+### `docs/policy/documentation.md`
+
+- No structural rewrite is currently required.
+- Recheck its ownership table after the troubleshooting consolidation and ADR
+  creation.
+- Keep future policy refinements in this document rather than distributing
+  documentation rules across architecture or development.
+
+### `docs/policy/testing.md`
+
+- It is the sole owner of global testing philosophy and sufficiency rules.
+- Remove competing global testing guidance from architecture and other docs.
+- Leave the planned substantive testing-policy review for its separate work
+  cycle.
+
+### `docs/cli.md`
+
+- Keep the complete command, argument, flag, output, and exit-code contract.
+- Remove the repeated definition of config precedence and link to
+  `docs/config.md`.
+- Move profile source and override semantics that are not CLI invocation
+  semantics to `docs/config.md`.
+- Resolve duplicate `--prompt-dir` entries in the `render` and `serve`
+  sections.
+- Verify every registered flag, deprecated alias, requirement, zero-value
+  behavior, and output destination against `internal/adapter/cli/run.go`.
+- Clarify timeout conversion for subsecond Go durations if that behavior is
+  intended to remain public.
+- Keep only compact workflow examples and link to maintained scripts under
+  `examples/`.
+
+### `docs/config.md`
+
+- Keep app-config discovery, precedence, fields, defaults, validation,
+  credential-supply mechanisms, prompt/profile formats, and selectable profile
+  catalog as the canonical contract.
+- Replace complete inline config, prompt, and profile files with the smallest
+  useful snippets and links to copyable files under `examples/`.
+- If a production-oriented complete config remains useful, move it into
+  `examples/` and link it rather than maintaining a second complete copy.
+- Correct `repair_attempts` optionality.
+- Add `deepseek-4-flash` and reconcile every built-in profile ID, model, and
+  credential-variable name with embedded assets.
+- Establish a low-friction way to prevent catalog drift, such as a generated
+  catalog section or a focused consistency check, if its maintenance value
+  justifies the cost.
+- Keep schema configuration and prompt/profile file-format rules here; move
+  implementation mechanics to internal source docs.
+- Reduce the HTTP artifact-reference section to configuration meaning and link
+  request shapes/statuses to `docs/api.md` and runtime security handling to
+  `docs/operations.md`.
+- Keep the secret-supply mechanism here; move deployment permissions and
+  sensitive-runtime handling to operations.
+
+### `docs/api.md`
+
+- Keep routes, media types, request/response schemas, strict JSON behavior,
+  status codes, HTTP retry semantics, and HTTP artifact-access outcomes.
+- Replace repeated app-config defaults and CLI flag syntax with links to
+  `docs/config.md` and `docs/cli.md`.
+- Retain HTTP limit effects but let configuration own field defaults and
+  precedence.
+- Keep the lexical containment and symlink behavior as externally observable
+  API/security behavior; operations may summarize its deployment consequence
+  and link back.
+- Reduce request and response examples to compact contract-bearing shapes and
+  link to `examples/http-run.json`.
+- Reverify all DTO fields, omission behavior, limit responses, and error mappings
+  against `internal/adapter/http/`.
+
+### `docs/operations.md`
+
+- Rewrite as a task-oriented runbook rather than a secondary CLI, config, and
+  API reference.
+- Keep deployment layout, process permissions, sensitive runtime artifact
+  handling, normal workflow, service exposure, capacity planning, recovery,
+  and the stateless rerun model.
+- Link exact commands and exit codes to `docs/cli.md`.
+- Link config fields, defaults, validation modes, and credential-supply
+  mechanics to `docs/config.md`.
+- Link HTTP route, status, schema, and limit semantics to `docs/api.md`.
+- Remove the repeated maintained-example inventory and link to the owning
+  examples section.
+- Incorporate the useful troubleshooting material and then delete
+  `docs/troubleshooting.md`.
+
+### `docs/troubleshooting.md`
+
+- Correct the CLI validation-error diagnostic during migration.
+- Move useful symptom, diagnostic, and safe-fix material into
+  `docs/operations.md`.
+- Remove repeated definitions of flags, fields, defaults, HTTP codes, routes,
+  and validation semantics; link to their canonical contracts.
+- Delete this file after all useful recovery guidance and inbound links have
+  been handled.
+
+### `docs/consumers/api.md`
+
+- Keep integration-surface selection, minimal consumer workflow, and consumer
+  responsibilities.
+- Retain one minimal Go example as permitted instructional content and link to
+  the complete package contract and maintained example.
+- Replace repeated CLI exit and HTTP status definitions with links to their
+  canonical contracts.
+- Replace repeated deployment input and credential definitions with links to
+  configuration and operations.
+- Keep retry and artifact-retention decisions as consumer responsibilities
+  without restating wire semantics.
+
+### `docs/consumers/pkg-scriptorium.md`
+
+- Keep the import path, public constructors, options, types, errors, workflows,
+  and public security boundary as the canonical Go package contract.
+- Correct the input-name validation statement.
+- Reverify public fields, JSON behavior, option precedence, nil behavior,
+  source containment, direct-key handling, and sentinel errors against the root
+  package.
+- Keep a minimal package example and link to
+  `examples/go-library/prepare`; avoid duplicating complete maintained source.
+- Link prompt/profile/schema file formats to `docs/config.md` rather than
+  redefining them.
+
+### `docs/integrations/subprocess.md`
+
+- Narrow this document to subprocess-specific compatibility and process
+  integration concerns.
+- Link command syntax, flags, stdout/stderr behavior, and exit codes to
+  `docs/cli.md` rather than maintaining parallel definitions.
+- Link config search and field semantics to `docs/config.md`.
+- Retain process isolation, environment propagation, stream capture, output
+  ownership, cancellation/termination expectations, and security guidance when
+  these are specific to subprocess consumers.
+- Move task-oriented surface-selection advice to `docs/consumers/api.md` if it
+  is currently duplicated.
+
+### `docs/integrations/openai-compatible-chat.md`
+
+- Keep only the outbound HTTP wire contract: endpoint construction, request
+  payload, authentication header, timeouts, response subset, and unsupported
+  protocol behavior.
+- Remove the concrete implementation-file introduction and internal error
+  sentinel catalog.
+- Correct the `session_id` unit to Unicode code points.
+- Remove or deliberately source/version the provider service-tier claim.
+- Reverify reserved `extra_params` keys, structured-output envelope,
+  cache-control encoding, usage mapping, and response requirements against
+  `internal/llm/openai_compatible_client.go`.
+- Link runner schema preparation to internal runner documentation rather than
+  explaining orchestration locally.
+
+### `docs/internal/overview.md` (new)
+
+- Own the complete current component and package inventory.
+- Link each component to focused internal docs and external contracts.
+- Include the command entry point and public facade without redefining their
+  external behavior.
+- Absorb the still-useful concrete repository-layout material removed from the
+  old development guide and architecture policy.
+
+### `docs/internal/adapters.md`
+
+- Keep implementation flow, collaborators, translation boundaries, wiring,
+  error-mapping mechanisms, and relevant tests.
+- Remove exact flag lists, config field definitions, route schemas, response
+  codes, and exit-code definitions; link their canonical owners.
+- Keep only package-local invariants rather than repeating global architecture.
+- Incorporate the former contributor recipes for adding app-config fields, CLI
+  flags, and adapter capabilities.
+- Describe how to verify affected adapter contracts without duplicating the
+  global testing policy.
+
+### `docs/internal/runner.md`
+
+- Keep prepare/run flow, dependency boundaries, state transitions, failure
+  categories, repair mechanics, hashing/validation coordination, and relevant
+  tests.
+- Remove the app-config field list and link to adapter/config documentation.
+- Keep internal sentinel and collaborator information, but link public error
+  behavior to the package/API contracts.
+- Distinguish package-local guarantees from global architectural invariants.
+- Add a focused change recipe for runner orchestration when useful.
+
+### `docs/internal/sources.md`
+
+- Keep loader implementations, source precedence mechanics, containment
+  implementation, failure categories, collaborators, and relevant tests.
+- Link YAML/JSON field definitions and defaults to `docs/config.md`.
+- Link HTTP-visible artifact outcomes to `docs/api.md` and deployment handling
+  to operations.
+- Incorporate the former recipe for updating prompt, profile, schema, and
+  built-in-profile assets.
+- Distinguish implementation mechanics from external source contracts.
+
+### `docs/internal/llm.md` (new)
+
+Create a focused internal document if the implementation material removed from
+the outbound integration contract remains useful.
+
+It should own:
+
+- client construction and internal collaborator boundaries;
+- internal request mapping and timeout selection;
+- internal error categories;
+- relevant tests and package-local change guidance.
+
+It must link to `docs/integrations/openai-compatible-chat.md` for the wire
+contract rather than repeat payload definitions.
+
+### `docs/roadmap/migration.md`
+
+- Retain migration status, gates, sequencing, and task breakdowns.
+- Mark the documentation-refresh gate complete only after this roadmap's
+  completion criteria are satisfied.
+- After the Promptkit split ADR is accepted, replace duplicated decision
+  rationale and architectural ownership detail with a concise summary and ADR
+  link where practical.
+- Update paths and document ownership affected by this refresh.
+
+### `examples/`
+
+- Keep complete copyable artifacts here rather than in reference documents.
+- Verify `config.yml` and `config.full.yml` through the real config loader.
+- Keep prompt/profile/schema examples covered by representative repository and
+  engine tests.
+- Keep the render script and Go package example runnable from the repository
+  root.
+- Validate `http-run.json` structurally against the HTTP DTO contract without
+  requiring a live model endpoint.
+- Add a separate production-oriented config example only if that maintained
+  artifact has clear value; otherwise remove the duplicate production block
+  from config documentation.
+- Preserve the fixtures as non-sensitive synthetic data.
+
+## Implementation Sequence
+
+### Stage 1: Establish Canonical Developer Structure
+
+1. Create the initial documentation ADR if retained.
+2. Create `docs/internal/overview.md`.
+3. Rewrite architecture around normative ownership and invariants.
+4. Update `docs/development.md` routing.
+5. Rescue still-valid development rules and assign every removed recipe.
+
+**Gate:** Contributor orientation, architecture, component inventory, policies,
+and decision history have distinct owners and no stale links.
+
+### Stage 2: Correct And Consolidate External Contracts
+
+1. Revise `docs/config.md` and correct its known factual issues.
+2. Revise `docs/cli.md`.
+3. Revise `docs/api.md`.
+4. Revise both consumer guides.
+5. Revise both integration contracts.
+6. Verify each volatile contract against code and maintained assets.
+
+**Gate:** Each externally observable field, default, flag, route, status, import
+path, and wire behavior has one authoritative definition.
+
+### Stage 3: Refocus Internal Documentation
+
+1. Revise adapter, runner, and source docs.
+2. Add the internal LLM document if warranted.
+3. Incorporate the former subsystem change recipes.
+4. Remove global policy and external-contract duplication.
+
+**Gate:** Internal docs explain implementation and package-local guarantees,
+link external contracts, and contain the detailed contributor recipes needed
+for safe changes.
+
+### Stage 4: Rewrite Operations And Consolidate Recovery Guidance
+
+1. Rewrite operations as a runbook.
+2. Migrate useful troubleshooting content.
+3. Delete `docs/troubleshooting.md`.
+4. Update all inbound links and the development reading guide.
+
+**Gate:** Operations owns deployment and recovery without serving as a second
+CLI, config, or API reference.
+
+### Stage 5: Normalize Examples And Orientation
+
+1. Move any remaining complete copyable artifacts into `examples/`.
+2. Update README navigation and example links.
+3. Run maintained smoke examples.
+4. Add only high-value automated checks justified by the testing policy.
+
+**Gate:** Examples are canonical, runnable, secret-free, linked from their
+owning references, and not duplicated as complete files in prose docs.
+
+### Stage 6: Final Compliance Pass
+
+1. Validate all local links and paths.
+2. Search for duplicated volatile values across non-owning documents.
+3. Reconcile flags, config fields/defaults, API DTOs/codes, public types/errors,
+   and built-in profiles with implementation.
+4. Run `go test ./...`, `go vet ./...`, and the documented build command.
+5. Run the maintained render script and Go package example.
+6. Confirm that non-roadmap current docs contain implemented behavior only.
+7. Record completion in this roadmap and update the documentation gate in
+   `docs/roadmap/migration.md`.
+
+**Gate:** Code, tests, examples, contracts, internal documentation, operations,
+policies, and roadmap status agree.
+
+## Completion Criteria
+
+The documentation refresh is complete when:
+
+- every topic in the ownership table has one canonical owner;
+- `docs/internal/overview.md` exists and architecture no longer owns the package
+  inventory;
+- significant accepted decisions have an ADR owner;
+- all known factual discrepancies in this audit are corrected;
+- subsystem recipes from the former development guide have been preserved in
+  relevant internal docs;
+- operations contains the retained recovery guidance and the standalone
+  troubleshooting document is removed;
+- contracts and internal docs link to one another instead of maintaining
+  parallel volatile definitions;
+- maintained examples are runnable and not duplicated as complete prose
+  examples;
+- local links, validation commands, tests, and smoke checks pass;
+- the documentation-refresh gate in the Promptkit migration roadmap is marked
+  complete.

docs/roadmap/migration.md

@@ -0,0 +1,292 @@
+# Promptkit Migration Roadmap
+
+## Status
+
+Accepted plan. This document describes proposed work that is not yet
+implemented.
+
+## Objective
+
+Split the current repository into two projects:
+
+- **Promptkit**: the reusable Go framework, public Go facade, execution engine,
+  source and validation support, OpenAI-compatible client, extension
+  interfaces, and built-in execution-profile registry.
+- **Scriptorium**: a slim runnable application that imports Promptkit and
+  provides the CLI and HTTP interfaces.
+
+Scriptorium will become another downstream Promptkit consumer rather than the
+owner of the framework.
+
+## Compatibility And Migration Policy
+
+This is an intentionally breaking change.
+
+- New and migrated Go consumers must import Promptkit instead of Scriptorium.
+- Scriptorium will not retain type aliases, forwarding packages, deprecated
+  facade APIs, or other source-compatibility shims.
+- Existing consumers may continue using a previously tagged Scriptorium module
+  version until they are migrated.
+- The migration does not need to preserve compatibility between intermediate
+  development states. Each completed phase must instead leave the affected
+  repository internally consistent and tested.
+- Promptkit should initially preserve the useful shape and behavior of the
+  current public Go facade where doing so reduces extraction risk. Broader API
+  redesign should follow the split unless required to establish the new
+  boundary.
+
+## Target Ownership
+
+Promptkit should own application-neutral framework behavior:
+
+- public engine, request, result, option, extension, and error APIs;
+- prompt-definition loading and rendering;
+- execution profiles, overlays, and the built-in profile registry;
+- artifact-loading interfaces and general-purpose `file` and `inline` support;
+- schema loading and output validation;
+- LLM client boundaries and the OpenAI-compatible implementation;
+- preparation and execution orchestration;
+- framework and execution defaults.
+
+Scriptorium should own executable and transport concerns:
+
+- the `scriptorium` command and its `run`, `render`, and `serve` interfaces;
+- CLI parsing, output formatting, exit codes, and process behavior;
+- application-config discovery and CLI precedence;
+- HTTP routing, request and response DTOs, limits, and error/status mapping;
+- HTTP artifact-root and deployment security policy;
+- server and adapter defaults;
+- executable examples, operations guidance, and transport documentation.
+
+The intended dependency direction is:
+
+```text
+Scriptorium CLI and HTTP adapters
+                |
+                v
+             Promptkit
+                |
+                v
+     consumer-supplied sources and clients
+```
+
+Scriptorium must use Promptkit's public API. It must not depend on Promptkit
+implementation packages or reproduce framework orchestration.
+
+## Migration Steps
+
+### Step 1: Refresh And Synchronize Documentation
+
+Perform a repository-wide documentation refresh before migration development.
+
+At minimum:
+
+- reconcile all current-behavior documentation with the code, tests, examples,
+  defaults, and current public contracts;
+- introduce the planned documentation-policy updates;
+- establish an architecture decision record policy and canonical ADR location;
+- resolve stale, duplicated, or misplaced material;
+- validate documentation links and maintained examples;
+- leave future migration behavior in `docs/roadmap/` until implemented.
+
+**Gate:** Do not begin architectural migration work until the documentation
+refresh and policy updates are merged and the repository has an agreed,
+accurate baseline.
+
+### Step 2: Record The Architectural Decision And Detailed Boundary
+
+Create an ADR, under the policy established in Step 1, that records:
+
+- the decision to split Promptkit from Scriptorium;
+- the target ownership and dependency direction;
+- the selected Promptkit repository and Go module paths;
+- the breaking-change and versioning policy;
+- ownership of configuration fields and defaults;
+- artifact-reader and HTTP containment responsibilities;
+- local multi-repository development and release coordination;
+- documentation ownership after the split.
+
+Use the ADR to resolve any remaining public-boundary decisions before code is
+moved.
+
+**Gate:** The ADR is accepted, and every existing package, public contract,
+configuration category, and maintained asset has a target owner.
+
+### Step 3: Characterize Existing Framework Behavior
+
+Strengthen or add contract-focused tests where needed so extraction can be
+verified without relying on package placement.
+
+Preserve coverage of:
+
+- `Prepare` and `Run` behavior;
+- prompt, profile, execution-default, and request-override precedence;
+- presence-aware numeric overrides;
+- built-in profile fallback and custom-profile overlays;
+- strict YAML and JSON decoding;
+- prompt, profile, schema, and artifact source behavior;
+- structured-output requests and output validation;
+- validation failures versus validation runtime errors;
+- secret handling and redaction;
+- public error classification;
+- HTTP artifact restrictions and transport mappings.
+
+**Gate:** Current framework and adapter contracts are represented by passing
+tests sufficient to detect behavioral regressions during the split.
+
+### Step 4: Make Scriptorium Adapters Consume The Public Facade
+
+Within the current repository, refactor the CLI and HTTP adapters to use the
+public framework facade rather than constructing or importing internal runner
+components directly.
+
+Add only the minimum public capabilities needed to support this boundary. These
+may include:
+
+- a small `Run`/`Prepare` consumer interface;
+- injectable artifact-reading behavior for Scriptorium's restricted HTTP
+  policy;
+- source options currently available only through internal constructors;
+- prepared-run formatting based on public types;
+- stable public error classification required by CLI and HTTP mappings.
+
+Do not broadly export internal repositories, domain types, or use-case
+implementations.
+
+**Gate:** The CLI and HTTP adapters use only the public framework API for
+framework behavior, and all tests and documented smoke commands pass.
+
+### Step 5: Create The Promptkit Repository
+
+Create the Promptkit repository and Go module as an explicit out-of-band
+operation.
+
+Establish:
+
+- repository access, ownership, and branch protections;
+- the module path selected by the ADR;
+- baseline development, architecture, documentation, and release policies;
+- CI for build, test, vet, and other agreed checks;
+- an initial package layout centered on a small public facade with internal
+  implementation packages;
+- a local development workflow for coordinated Promptkit and Scriptorium
+  changes, using a workspace or temporary uncommitted module replacement where
+  appropriate.
+
+Do not commit local filesystem `replace` directives to release branches.
+
+**Gate:** The Promptkit repository exists, is accessible to maintainers, has
+working CI and policy scaffolding, and can receive the extracted framework.
+Do not begin cross-repository extraction until this out-of-band work is
+confirmed complete.
+
+### Step 6: Extract And Stabilize Promptkit
+
+Move the application-neutral framework and built-in profile assets into
+Promptkit. Preserve implementation packages as internal where practical.
+
+The initial public API should remain focused on the established engine workflow
+and the source and client extension points required by real consumers. Avoid
+combining the extraction with unrelated API redesign.
+
+Move or recreate the relevant:
+
+- framework implementation;
+- public package tests and framework contract tests;
+- built-in profile assets and registry tests;
+- Go consumer examples;
+- framework, consumer, configuration-format, and integration documentation.
+
+Verify that Promptkit can be built, tested, and consumed independently of the
+Scriptorium repository.
+
+**Gate:** Promptkit independently provides the agreed framework contract,
+passes its CI checks, and has a tagged version that Scriptorium and other
+consumers can import.
+
+### Step 7: Slim Scriptorium And Adopt Promptkit
+
+Update Scriptorium to import the tagged Promptkit module and remove the
+framework implementation and public Go facade that Promptkit replaces.
+
+Retain only Scriptorium-owned executable and transport behavior. In particular:
+
+- wire CLI and HTTP requests through Promptkit's public API;
+- keep application config and transport defaults in Scriptorium;
+- keep restricted HTTP artifact policy in Scriptorium while injecting it
+  through Promptkit's supported boundary;
+- remove obsolete framework packages, tests, and documentation;
+- update Scriptorium examples and docs to describe the CLI and HTTP application;
+- direct Go framework consumers to Promptkit without providing compatibility
+  aliases or forwarding APIs.
+
+**Gate:** Scriptorium builds and passes all tests using a tagged Promptkit
+dependency, contains no duplicate framework implementation, and its current
+documentation describes only the slimmed application.
+
+### Step 8: Migrate Downstream Consumers To Promptkit
+
+Inventory downstream Go consumers and migrate each from the Scriptorium package
+to Promptkit. This work may occur in external repositories and must be tracked
+explicitly.
+
+For each consumer:
+
+- update module imports and dependencies;
+- adapt to any intentionally changed public API;
+- run its tests and relevant integration or smoke checks;
+- confirm configuration, source, validation, and error behavior;
+- release or deploy the migrated consumer through its normal process.
+
+Consumers that cannot migrate immediately may remain pinned to the last
+framework-bearing Scriptorium tag. No compatibility work is required in the new
+Scriptorium project for those consumers.
+
+**Gate:** All in-scope downstream consumers are either migrated and verified or
+explicitly recorded as remaining on the previous Scriptorium version with an
+owner and follow-up plan. Do not declare the ecosystem migration complete until
+the required out-of-band consumer changes are confirmed.
+
+### Step 9: Complete Release And Documentation Cutover
+
+Complete the coordinated project transition:
+
+- publish Promptkit before dependent Scriptorium releases;
+- release the breaking Scriptorium version against the tagged Promptkit
+  dependency;
+- publish migration guidance that maps the former Scriptorium Go API to
+  Promptkit;
+- update cross-project links, examples, package documentation, and release
+  notes;
+- verify that no release artifact depends on local workspaces or replacements;
+- archive completed roadmap material according to the documentation policy in
+  effect at that time.
+
+**Gate:** Promptkit and Scriptorium are independently releasable, their
+documentation has distinct and accurate ownership, and the migration status of
+all identified downstream consumers is recorded.
+
+## Cross-Cutting Constraints
+
+- Preserve the invariant that execution orchestration remains narrow and
+  application-neutral.
+- Keep adapter-specific decisions out of Promptkit.
+- Keep Scriptorium dependent only on Promptkit's supported public API.
+- Preserve strict external decoding, error classification, validation
+  semantics, and secret redaction throughout the migration.
+- Keep each repository buildable and testable at merged phase boundaries.
+- Coordinate cross-repository changes through tagged dependencies and explicit
+  gates rather than assuming atomic commits across repositories.
+- Document only implemented behavior outside roadmap files.
+
+## Completion Criteria
+
+The migration is complete when:
+
+- Promptkit is the independent owner of the reusable framework and built-in
+  profiles;
+- Scriptorium is a slim CLI and HTTP consumer of Promptkit;
+- Scriptorium no longer exposes or maintains the former public Go framework;
+- all required downstream migrations and external repository work have been
+  completed or explicitly dispositioned;
+- both repositories build, test, document, version, and release independently.

docs/troubleshooting.md

@@ -0,0 +1,361 @@
+# Troubleshooting
+
+This guide lists common implemented failure modes and safe fixes.
+
+Canonical references:
+
+- [CLI reference](cli.md)
+- [Configuration reference](config.md)
+- [HTTP API reference](api.md)
+- [Operations guide](operations.md)
+
+## Missing Or Invalid Config
+
+Symptom:
+
+- CLI error includes `application config error`, `config file not found`, `invalid config YAML`, or `invalid config`.
+
+Likely cause:
+
+- `--config` points to a missing file.
+- YAML syntax is invalid.
+- Config contains unknown fields or negative HTTP size limits.
+
+Diagnostic step:
+
+```bash
+go run ./cmd/scriptorium render --config /path/to/config.yml --prompt generic.markdown_summary --input transcript=./examples/fixtures/transcript.md --input glossary=./examples/fixtures/glossary.yml
+```
+
+Safe fix:
+
+- Correct the config path.
+- Fix YAML syntax.
+- Remove unknown fields.
+- Keep raw secrets out of config.
+
+Relevant links: [Configuration reference](config.md), [CLI reference](cli.md)
+
+## Missing Prompt Directory
+
+Symptom:
+
+- CLI parse error says the prompt directory is required.
+
+Likely cause:
+
+- Neither config nor CLI flags provide an effective `prompt_dir`.
+
+Diagnostic step:
+
+- Re-run once with explicit `--prompt-dir`.
+
+Safe fix:
+
+- Set `prompt_dir` in config or pass `--prompt-dir`.
+
+Relevant links: [Configuration reference](config.md), [CLI reference](cli.md)
+
+## Unknown Flags
+
+Symptom:
+
+- CLI parse error for an unknown flag.
+
+Likely cause:
+
+- Typo.
+- Flag is valid for another command.
+- `serve` was given runtime model override flags.
+
+Diagnostic step:
+
+- Compare the command with the command-specific flag list.
+
+Safe fix:
+
+- Remove unsupported flags.
+- Use `run` or `render` for runtime model overrides.
+
+Relevant links: [CLI reference](cli.md)
+
+## Prompt Load Failures
+
+Symptom:
+
+- CLI run/render fails during prompt loading.
+- HTTP returns `404 prompt_not_found` or `400 prompt_load_failed`.
+
+Likely cause:
+
+- Prompt ID/version does not exist.
+- Prompt YAML is invalid or has unknown fields.
+- Prompt contract is invalid, such as missing messages, invalid output mode, bad `content_file`, or missing `schema_path` for `json_schema`.
+
+Diagnostic step:
+
+```bash
+go run ./cmd/scriptorium render --config ./examples/config.yml --prompt <prompt-id> --input transcript=./examples/fixtures/transcript.md --input glossary=./examples/fixtures/glossary.yml --format json
+```
+
+Safe fix:
+
+- Correct prompt ID/version.
+- Fix prompt YAML and referenced `content_file` paths.
+- Fix output contract fields.
+
+Relevant links: [Configuration reference](config.md), [CLI reference](cli.md)
+
+## Profile Load Failures
+
+Symptom:
+
+- CLI run/render fails during profile loading.
+- HTTP returns `404 profile_not_found`, `400 profile_load_failed`, or `400 profile_required`.
+
+Likely cause:
+
+- Profile ID does not exist.
+- Request omitted profile and prompt has no `default_profile`.
+- Profile YAML is invalid or has unknown fields.
+- Profile contains raw `api_key`.
+
+Diagnostic step:
+
+```bash
+go run ./cmd/scriptorium render --config ./examples/config.yml --prompt generic.markdown_summary --profile <profile-id> --input transcript=./examples/fixtures/transcript.md --input glossary=./examples/fixtures/glossary.yml
+```
+
+Safe fix:
+
+- Correct profile ID or prompt `default_profile`.
+- Fix profile YAML and value ranges.
+- Replace raw `api_key` with `api_key_env`.
+
+Relevant links: [Configuration reference](config.md), [CLI reference](cli.md)
+
+## Input Artifact Failures
+
+Symptom:
+
+- CLI run/render fails while reading inputs.
+- HTTP returns `400 artifact_read_failed`, `400 artifact_not_allowed`, or `413 artifact_too_large`.
+
+Likely cause:
+
+- Input file path is missing or unreadable.
+- HTTP input type is unsupported or missing required fields.
+- HTTP file refs are disabled because no artifact root is configured.
+- HTTP file path is lexically outside the artifact root.
+- HTTP file input exceeds `server.max_artifact_bytes`.
+
+Diagnostic step:
+
+- Verify each input path exists and is readable by the process.
+- For HTTP, verify input refs use `file` or `inline`.
+- For HTTP file refs, verify the artifact root and compare file size to `server.max_artifact_bytes`.
+
+Safe fix:
+
+- Correct paths and permissions.
+- Configure a narrow artifact root for HTTP file refs.
+- Use relative paths under the artifact root or switch to `inline`.
+- Increase `server.max_artifact_bytes` only for expected larger inputs.
+
+Relevant links: [HTTP API reference](api.md), [Configuration reference](config.md)
+
+## Missing API-Key Environment Variable
+
+Symptom:
+
+- CLI render/run fails with an API-key environment error.
+- HTTP returns `400 api_key_env_missing`.
+
+Likely cause:
+
+- Selected profile or runtime override sets `api_key_env`, but the environment variable is unset or empty.
+
+Diagnostic step:
+
+```bash
+printenv SCRIPTORIUM_API_KEY
+```
+
+Safe fix:
+
+- Set the required environment variable before starting the CLI command or HTTP service.
+- Or use a profile that does not require provider API-key auth.
+
+Relevant links: [Configuration reference](config.md), [Operations guide](operations.md)
+
+## Prompt Template Render Failures
+
+Symptom:
+
+- CLI render/run fails during prompt rendering.
+- HTTP returns `400 prompt_render_failed`.
+
+Likely cause:
+
+- Template references an input that was not supplied.
+- Template syntax or variable reference is invalid.
+
+Diagnostic step:
+
+- Run `render --format json` with the same prompt, inputs, vars, and profile.
+
+Safe fix:
+
+- Align `{{input "name"}}` references with request input names.
+- Fix template syntax and variable names.
+
+Relevant links: [Configuration reference](config.md), [CLI reference](cli.md)
+
+## LLM Request Failures
+
+Symptom:
+
+- CLI `run` fails during generation.
+- HTTP returns `502 llm_failed`.
+
+Likely cause:
+
+- Endpoint is unreachable.
+- Provider returns non-2xx.
+- Request times out.
+- Provider response is malformed.
+
+Diagnostic step:
+
+- Run `render` first to confirm pre-LLM preparation works.
+- Check selected endpoint/model in prepared output.
+- Check network/provider logs for timeout or non-2xx details.
+
+Safe fix:
+
+- Correct endpoint/model/profile settings.
+- Adjust timeout when appropriate.
+- Resolve provider or network issue.
+
+Relevant links: [Operations guide](operations.md), [Configuration reference](config.md)
+
+## Validation Failed
+
+Symptom:
+
+- CLI `run` exits `2`.
+- HTTP returns `200 OK` with `validation.status` set to `failed`.
+
+Likely cause:
+
+- Generated output failed `basic`, `json`, or `json_schema` content validation.
+
+Diagnostic step:
+
+- Inspect validation errors in CLI stderr or the HTTP response.
+
+Safe fix:
+
+- Refine prompt instructions.
+- Adjust schema or model/profile settings.
+- Rerun after correction.
+
+Relevant links: [Operations guide](operations.md), [HTTP API reference](api.md)
+
+## Validation Runtime Failure
+
+Symptom:
+
+- CLI `run` fails with validation runtime error.
+- HTTP returns `500 validation_runtime_failed`.
+
+Likely cause:
+
+- `json_schema` schema file is missing or unreadable.
+- Schema JSON is invalid.
+
+Diagnostic step:
+
+- Verify `schema_dir` and prompt `output.schema_path`.
+- Check schema file readability and JSON syntax.
+
+Safe fix:
+
+- Correct schema path or permissions.
+- Fix schema JSON.
+- Rerun.
+
+Relevant links: [Configuration reference](config.md), [Operations guide](operations.md)
+
+## HTTP JSON Or Request Contract Errors
+
+Symptom:
+
+- HTTP returns `400 invalid_json` or `400 invalid_request`.
+
+Likely cause:
+
+- JSON body is malformed.
+- Request has unknown fields or trailing JSON tokens.
+- Required `prompt_id` or `inputs` is missing.
+- Runtime override values are out of range.
+- `extra_params` collides with reserved outbound fields.
+
+Diagnostic step:
+
+- Revalidate request JSON and compare fields with the API reference.
+
+Safe fix:
+
+- Send one JSON object with only supported fields.
+- Include `prompt_id` and at least one input.
+- Use valid model override ranges.
+- Remove reserved `extra_params` keys.
+
+Relevant links: [HTTP API reference](api.md)
+
+## HTTP Size Limit Errors
+
+Symptom:
+
+- HTTP returns `413 request_too_large`, `413 artifact_too_large`, or `413 response_too_large`.
+
+Likely cause:
+
+- JSON request body exceeds `server.max_request_bytes`.
+- HTTP file input exceeds `server.max_artifact_bytes`.
+- Encoded JSON response exceeds `server.max_response_bytes`.
+
+Diagnostic step:
+
+- Compare request, file input, and expected response sizes with configured limits.
+
+Safe fix:
+
+- Use smaller inline inputs or switch to file inputs under the artifact root.
+- Reduce generated output size.
+- Omit `include_raw_output`.
+- Increase limits only when the deployment expects larger payloads.
+
+Relevant links: [HTTP API reference](api.md), [Operations guide](operations.md)
+
+## HTTP Route Or Method Errors
+
+Symptom:
+
+- HTTP returns `404 not_found` or `405 method_not_allowed`.
+
+Likely cause:
+
+- Path is not `/v1/runs`.
+- Method on `/v1/runs` is not `POST`.
+
+Diagnostic step:
+
+- Check the request URL and method.
+
+Safe fix:
+
+- Send `POST /v1/runs`.
+
+Relevant links: [HTTP API reference](api.md)

engine.go

@@ -0,0 +1,318 @@
+package scriptorium
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"io/fs"
+	"net/http"
+	"os"
+	"path/filepath"
+	"strings"
+	"time"
+
+	artifactadapter "gitea.maximumdirect.net/eric/scriptorium/internal/artifact"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/llm"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/profile/builtin"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/prompt"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/usecase"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/validate"
+)
+
+// ErrInvalidConfig indicates invalid public engine configuration.
+var ErrInvalidConfig = errors.New("invalid engine configuration")
+
+var (
+	ErrInvalidRequest  = errors.New("invalid run request")
+	ErrPromptNotFound  = errors.New("prompt not found")
+	ErrProfileNotFound = errors.New("profile not found")
+	ErrPromptLoad      = errors.New("failed to load prompt definition")
+	ErrProfileLoad     = errors.New("failed to load execution profile")
+	ErrArtifactLoad    = errors.New("failed to load artifact")
+	ErrPromptRender    = errors.New("failed to render prompt")
+	ErrLLMGenerate     = errors.New("failed to generate output")
+	ErrValidation      = errors.New("failed to validate output")
+)
+
+// Engine prepares and runs Scriptorium prompt requests.
+type Engine struct {
+	runner *usecase.Runner
+}
+
+// Config configures a public Scriptorium engine.
+type Config struct {
+	PromptDir  string
+	ProfileDir string
+	SchemaDir  string
+	Timeout    time.Duration
+	HTTPClient *http.Client
+}
+
+// Option customizes engine construction.
+type Option interface {
+	apply(*engineOptions) error
+}
+
+type optionFunc func(*engineOptions) error
+
+func (f optionFunc) apply(options *engineOptions) error {
+	return f(options)
+}
+
+type engineOptions struct {
+	llmClient       llm.Client
+	promptDefs      promptdef.Repository
+	profiles        profile.Repository
+	memoryProfiles  profile.Repository
+	validator       validate.Validator
+	promptSource    bool
+	profileSource   bool
+	memorySource    bool
+	validatorSource bool
+}
+
+// WithLLMClient injects a custom LLM client for execution.
+func WithLLMClient(client LLMClient) Option {
+	return optionFunc(func(options *engineOptions) error {
+		if client == nil {
+			return ErrInvalidConfig
+		}
+		options.llmClient = publicLLMClientAdapter{client: client}
+		return nil
+	})
+}
+
+// WithPromptFS loads prompt definitions from fsys under root.
+//
+// The source uses the same strict prompt YAML rules as configured prompt
+// directories, and prompt content_file paths resolve within this source.
+func WithPromptFS(fsys fs.FS, root string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		if fsys == nil {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.promptDefs = promptdef.NewFSRepository(fsys, root)
+		options.promptSource = true
+		return nil
+	})
+}
+
+// WithPromptFile loads prompt definitions from the single prompt file at path.
+//
+// Relative prompt content_file paths resolve from the file's directory.
+func WithPromptFile(path string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		fsys, root, err := fileSource(path)
+		if err != nil {
+			return err
+		}
+		options.promptDefs = promptdef.NewFSRepository(fsys, root)
+		options.promptSource = true
+		return nil
+	})
+}
+
+// WithProfileFS loads execution profiles from fsys under root.
+//
+// Profiles from this source overlay built-in profiles. Profile YAML must use
+// api_key_env for environment-based credentials; raw API keys are rejected.
+func WithProfileFS(fsys fs.FS, root string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		if fsys == nil {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.profiles = profile.NewFSRepository(fsys, root)
+		options.profileSource = true
+		return nil
+	})
+}
+
+// WithProfileFile loads execution profiles from the single profile file at path.
+//
+// The profile overlays built-in profiles. Profile YAML must use api_key_env for
+// environment-based credentials; raw API keys are rejected.
+func WithProfileFile(path string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		fsys, root, err := fileSource(path)
+		if err != nil {
+			return err
+		}
+		options.profiles = profile.NewFSRepository(fsys, root)
+		options.profileSource = true
+		return nil
+	})
+}
+
+// WithProfiles configures in-memory profiles that take precedence over
+// configured profile files and built-in profiles.
+func WithProfiles(profiles ...Profile) Option {
+	return optionFunc(func(options *engineOptions) error {
+		repo, err := newMemoryProfileRepository(profiles)
+		if err != nil {
+			return err
+		}
+		options.memoryProfiles = repo
+		options.memorySource = true
+		return nil
+	})
+}
+
+// WithSchemaFS loads JSON Schema documents from fsys under root.
+//
+// Prompt schema_path values resolve within this source when schema validation
+// or structured output is requested.
+func WithSchemaFS(fsys fs.FS, root string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		if fsys == nil {
+			return ErrInvalidConfig
+		}
+		if strings.TrimSpace(root) == "" {
+			return ErrInvalidConfig
+		}
+		options.validator = validate.NewFSValidator(fsys, root)
+		options.validatorSource = true
+		return nil
+	})
+}
+
+// WithSchemaFile loads JSON Schema documents from the single schema file at path.
+//
+// Prompt schema_path values refer to the file's base name.
+func WithSchemaFile(path string) Option {
+	return optionFunc(func(options *engineOptions) error {
+		fsys, root, err := fileSource(path)
+		if err != nil {
+			return err
+		}
+		options.validator = validate.NewFSValidator(fsys, root)
+		options.validatorSource = true
+		return nil
+	})
+}
+
+// NewEngine constructs an Engine using the same default internal components as
+// the CLI and HTTP adapters.
+func NewEngine(cfg Config, opts ...Option) (*Engine, error) {
+	var options engineOptions
+	for _, opt := range opts {
+		if opt == nil {
+			continue
+		}
+		if err := opt.apply(&options); err != nil {
+			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
+		}
+	}
+
+	promptDefs := options.promptDefs
+	if !options.promptSource {
+		if strings.TrimSpace(cfg.PromptDir) == "" {
+			return nil, fmt.Errorf("%w: prompt directory is required", ErrInvalidConfig)
+		}
+		promptDefs = promptdef.NewFilesystemRepository(cfg.PromptDir)
+	}
+
+	profiles := builtin.NewRepositoryWithDirectory(cfg.ProfileDir)
+	if options.profileSource {
+		profiles = builtin.NewRepositoryWithPrimary(options.profiles)
+	}
+	if options.memorySource {
+		profiles = profile.NewOverlayRepository(options.memoryProfiles, profiles)
+	}
+
+	validator := options.validator
+	if !options.validatorSource {
+		schemaDir := cfg.SchemaDir
+		if strings.TrimSpace(schemaDir) == "" {
+			schemaDir = defaults.SchemaDirDefault
+		}
+		validator = validate.NewStandardValidator(schemaDir)
+	}
+
+	llmClient := options.llmClient
+	if llmClient == nil {
+		var err error
+		llmClient, err = llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
+			Timeout:    cfg.Timeout,
+			HTTPClient: cfg.HTTPClient,
+		})
+		if err != nil {
+			return nil, fmt.Errorf("%w: %v", ErrInvalidConfig, err)
+		}
+	}
+
+	return &Engine{
+		runner: usecase.NewRunner(
+			promptDefs,
+			profiles,
+			artifactadapter.NewCompositeReader(),
+			prompt.NewGoRenderer(),
+			llmClient,
+			validator,
+		),
+	}, nil
+}
+
+func fileSource(name string) (fs.FS, string, error) {
+	cleanName := strings.TrimSpace(name)
+	if cleanName == "" {
+		return nil, "", ErrInvalidConfig
+	}
+	dir := filepath.Dir(cleanName)
+	base := filepath.Base(cleanName)
+	if base == "." || base == string(filepath.Separator) || strings.TrimSpace(base) == "" {
+		return nil, "", ErrInvalidConfig
+	}
+	info, err := os.Stat(cleanName)
+	if err != nil {
+		return nil, "", fmt.Errorf("%w: failed to access source file %q: %v", ErrInvalidConfig, cleanName, err)
+	}
+	if info.IsDir() {
+		return nil, "", fmt.Errorf("%w: source path %q must be a file", ErrInvalidConfig, cleanName)
+	}
+	return os.DirFS(dir), filepath.ToSlash(base), nil
+}
+
+// Prepare resolves a prompt request without calling an LLM.
+func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, error) {
+	if e == nil || e.runner == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
+	}
+
+	domainReq, err := toDomainRunRequest(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	prepared, err := e.runner.Prepare(ctx, domainReq)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return fromDomainPreparedRun(prepared), nil
+}
+
+// Run executes a prompt request and returns the generated artifact and metadata.
+func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error) {
+	if e == nil || e.runner == nil {
+		return nil, fmt.Errorf("%w: engine is nil", ErrInvalidConfig)
+	}
+
+	domainReq, err := toDomainRunRequest(req)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	result, err := e.runner.Run(ctx, domainReq)
+	if err != nil {
+		return nil, mapPublicError(err)
+	}
+	return fromDomainRunResult(result), nil
+}

errors.go

@@ -0,0 +1,79 @@
+package scriptorium
+
+import (
+	"errors"
+	"fmt"
+
+	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/usecase"
+)
+
+func mapPublicError(err error) error {
+	if err == nil {
+		return nil
+	}
+	if hasPublicError(err) {
+		return err
+	}
+	publicErr := publicErrorFor(err)
+	if publicErr == nil {
+		return err
+	}
+	return fmt.Errorf("%w: %w", publicErr, err)
+}
+
+func hasPublicError(err error) bool {
+	for _, publicErr := range []error{
+		ErrInvalidConfig,
+		ErrInvalidRequest,
+		ErrPromptNotFound,
+		ErrProfileNotFound,
+		ErrPromptLoad,
+		ErrProfileLoad,
+		ErrArtifactLoad,
+		ErrPromptRender,
+		ErrLLMGenerate,
+		ErrValidation,
+	} {
+		if errors.Is(err, publicErr) {
+			return true
+		}
+	}
+	return false
+}
+
+func publicErrorFor(err error) error {
+	switch {
+	case errors.Is(err, promptdef.ErrPromptDefinitionNotFound):
+		return ErrPromptNotFound
+	case errors.Is(err, profile.ErrProfileNotFound):
+		return ErrProfileNotFound
+	case errors.Is(err, usecase.ErrPromptLoad):
+		return ErrPromptLoad
+	case errors.Is(err, usecase.ErrProfileLoad):
+		return ErrProfileLoad
+	case errors.Is(err, promptdef.ErrInvalidYAML), errors.Is(err, promptdef.ErrInvalidPromptDefinition):
+		return ErrPromptLoad
+	case isProfileLoadCause(err):
+		return ErrProfileLoad
+	case errors.Is(err, usecase.ErrArtifactLoad):
+		return ErrArtifactLoad
+	case errors.Is(err, usecase.ErrPromptRender):
+		return ErrPromptRender
+	case errors.Is(err, usecase.ErrLLMGenerate):
+		return ErrLLMGenerate
+	case errors.Is(err, usecase.ErrValidation):
+		return ErrValidation
+	case errors.Is(err, usecase.ErrInvalidRequest):
+		return ErrInvalidRequest
+	default:
+		return nil
+	}
+}
+
+func isProfileLoadCause(err error) bool {
+	return errors.Is(err, profile.ErrInvalidYAML) ||
+		errors.Is(err, profile.ErrInvalidProfile) ||
+		errors.Is(err, profile.ErrRawAPIKeyNotAllowed)
+}

examples/config.full.yml

@@ -0,0 +1,13 @@
+prompt_dir: ./examples/prompts
+profile_dir: ./examples/profiles
+schema_dir: ./examples/schemas
+
+server:
+  addr: 127.0.0.1:8080
+  artifact_root: .
+  max_request_bytes: 16777216
+  max_artifact_bytes: 16777216
+  max_response_bytes: 16777216
+
+defaults:
+  render_format: text

examples/config.yml

@@ -1,9 +1,10 @@
-prompt_dir: ./prompts
-profile_dir: ./profiles
-schema_dir: ./schemas
+prompt_dir: ./examples/prompts
+profile_dir: ./examples/profiles
+schema_dir: ./examples/schemas
 
 server:
   addr: :8080
+  artifact_root: .
 
 defaults:
   render_format: text

examples/go-library/prepare/main.go

@@ -0,0 +1,50 @@
+package main
+
+import (
+	"context"
+	"encoding/json"
+	"log"
+	"os"
+
+	"gitea.maximumdirect.net/eric/scriptorium"
+)
+
+func main() {
+	engine, err := scriptorium.NewEngine(scriptorium.Config{
+		PromptDir:  "./examples/prompts",
+		ProfileDir: "./examples/profiles",
+		SchemaDir:  "./examples/schemas",
+	})
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	prepared, err := engine.Prepare(context.Background(), scriptorium.RunRequest{
+		PromptID: "generic.markdown_summary",
+		Inputs: map[string]scriptorium.ArtifactRef{
+			"transcript": scriptorium.File("./examples/fixtures/transcript.md"),
+			"glossary":   scriptorium.File("./examples/fixtures/glossary.yml"),
+		},
+	})
+	if err != nil {
+		log.Fatal(err)
+	}
+
+	summary := struct {
+		PromptID          string            `json:"prompt_id"`
+		SelectedProfileID string            `json:"selected_profile_id"`
+		Model             string            `json:"model"`
+		MessageCount      int               `json:"message_count"`
+		InputHashes       map[string]string `json:"input_hashes"`
+	}{
+		PromptID:          prepared.PromptID,
+		SelectedProfileID: prepared.SelectedProfileID,
+		Model:             prepared.EffectiveModelParams.Model,
+		MessageCount:      len(prepared.Messages),
+		InputHashes:       prepared.InputHashes,
+	}
+
+	if err := json.NewEncoder(os.Stdout).Encode(summary); err != nil {
+		log.Fatal(err)
+	}
+}

examples/http-run.json

@@ -0,0 +1,18 @@
+{
+  "prompt_id": "generic.markdown_summary",
+  "profile_id": "local-fast",
+  "inputs": {
+    "transcript": {
+      "type": "file",
+      "uri": "./examples/fixtures/transcript.md"
+    },
+    "glossary": {
+      "type": "file",
+      "uri": "./examples/fixtures/glossary.yml"
+    }
+  },
+  "vars": {
+    "session_date": "2026-05-04"
+  },
+  "include_raw_output": false
+}

examples/render-markdown-summary.sh

@@ -0,0 +1,13 @@
+#!/usr/bin/env bash
+set -euo pipefail
+
+repo_root="$(cd "$(dirname "${BASH_SOURCE[0]}")/.." && pwd)"
+
+cd "$repo_root"
+
+go run ./cmd/scriptorium render \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --format text

formatting.go

@@ -0,0 +1,51 @@
+package scriptorium
+
+import "fmt"
+
+// String returns a concise request summary without exposing direct API keys.
+func (r RunRequest) String() string {
+	return r.redactedString()
+}
+
+// GoString returns a concise request summary without exposing direct API keys.
+func (r RunRequest) GoString() string {
+	return r.redactedString()
+}
+
+func (r RunRequest) redactedString() string {
+	return fmt.Sprintf(
+		"scriptorium.RunRequest{PromptID:%q PromptVersion:%q ProfileID:%q APIKeySet:%t Inputs:%d Vars:%d ExecutionSet:%t ValidationSet:%t Metadata:%d}",
+		r.PromptID,
+		r.PromptVersion,
+		r.ProfileID,
+		r.APIKey != "",
+		len(r.Inputs),
+		len(r.Vars),
+		r.Execution != nil,
+		r.Validation != nil,
+		len(r.Metadata),
+	)
+}
+
+// String returns a concise request summary without exposing direct API keys or
+// rendered prompt content.
+func (r GenerateRequest) String() string {
+	return r.redactedString()
+}
+
+// GoString returns a concise request summary without exposing direct API keys or
+// rendered prompt content.
+func (r GenerateRequest) GoString() string {
+	return r.redactedString()
+}
+
+func (r GenerateRequest) redactedString() string {
+	return fmt.Sprintf(
+		"scriptorium.GenerateRequest{Messages:%d Model:%q APIKeySet:%t StructuredOutputSet:%t ExtraParams:%d}",
+		len(r.Prompt.Messages),
+		r.Target.Model,
+		r.APIKey != "",
+		r.StructuredOutput != nil,
+		len(r.Target.ExtraParams),
+	)
+}

internal/adapter/cli/run.go

@@ -19,7 +19,7 @@ import (
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
 	renderformat "gitea.maximumdirect.net/eric/scriptorium/internal/format"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/llm"
-	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/profile/builtin"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/prompt"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/usecase"
@@ -34,7 +34,6 @@ const (
 
 const (
 	errPromptDirRequired = "prompt directory is required; provide --prompt-dir or config.yml prompt_dir"
-	errProfileDirRequired = "profile directory is required; provide --profile-dir or config.yml profile_dir"
 )
 
 type runConfig struct {
@@ -79,6 +78,22 @@ type serveConfig struct {
 	promptDir        string
 	profileDir       string
 	schemaDir        string
+	artifactRoot     string
+	maxRequestBytes  int64
+	maxArtifactBytes int64
+	maxResponseBytes int64
+}
+
+type commonCommandSettings struct {
+	promptDir           string
+	profileDir          string
+	schemaDir           string
+	serverAddr          string
+	artifactRoot        string
+	maxRequestBytes     int64
+	maxArtifactBytes    int64
+	maxResponseBytes    int64
+	defaultRenderFormat renderformat.PreparedRunOutputFormat
 }
 
 type listFlag []string
@@ -125,22 +140,13 @@ func runCommand(args []string, stdout, stderr io.Writer) int {
 		return ExitRuntimeError
 	}
 
-	llmClient, err := llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
-		Timeout: defaults.LLMRequestTimeoutDefault,
-	})
+	llmClient, err := newOpenAIClient()
 	if err != nil {
 		fmt.Fprintf(stderr, "llm client error: %v\n", err)
 		return ExitRuntimeError
 	}
 
-	runner := usecase.NewRunner(
-		promptdef.NewFilesystemRepository(cfg.promptDir),
-		profile.NewFilesystemRepository(cfg.profileDir),
-		artifactadapter.NewCompositeReader(),
-		prompt.NewGoRenderer(),
-		llmClient,
-		validate.NewStandardValidator(cfg.schemaDir),
-	)
+	runner := newRunner(cfg.promptDir, cfg.profileDir, cfg.schemaDir, llmClient)
 
 	res, runErr := runner.Run(context.Background(), req)
 	if runErr != nil {
@@ -170,14 +176,7 @@ func renderCommand(args []string, stdout, stderr io.Writer) int {
 		return ExitRuntimeError
 	}
 
-	runner := usecase.NewRunner(
-		promptdef.NewFilesystemRepository(cfg.promptDir),
-		profile.NewFilesystemRepository(cfg.profileDir),
-		artifactadapter.NewCompositeReader(),
-		prompt.NewGoRenderer(),
-		nil,
-		validate.NewStandardValidator(cfg.schemaDir),
-	)
+	runner := newRunner(cfg.promptDir, cfg.profileDir, cfg.schemaDir, nil)
 
 	prepared, prepErr := runner.Prepare(context.Background(), req)
 	if prepErr != nil {
@@ -205,24 +204,24 @@ func serveCommand(args []string, stderr io.Writer) int {
 		return ExitRuntimeError
 	}
 
-	llmClient, err := llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
-		Timeout: defaults.LLMRequestTimeoutDefault,
-	})
+	llmClient, err := newOpenAIClient()
 	if err != nil {
 		fmt.Fprintf(stderr, "llm client error: %v\n", err)
 		return ExitRuntimeError
 	}
 
-	runner := usecase.NewRunner(
-		promptdef.NewFilesystemRepository(cfg.promptDir),
-		profile.NewFilesystemRepository(cfg.profileDir),
-		artifactadapter.NewCompositeReader(),
-		prompt.NewGoRenderer(),
-		llmClient,
-		validate.NewStandardValidator(cfg.schemaDir),
-	)
+	artifactReader, err := artifactadapter.NewRestrictedCompositeReaderWithLimit(cfg.artifactRoot, cfg.maxArtifactBytes)
+	if err != nil {
+		fmt.Fprintf(stderr, "artifact root error: %v\n", err)
+		return ExitRuntimeError
+	}
 
-	h := httpadapter.NewHandler(runner)
+	runner := newRunnerWithArtifactReader(cfg.promptDir, cfg.profileDir, cfg.schemaDir, llmClient, artifactReader)
+
+	h := httpadapter.NewHandlerWithOptions(runner, httpadapter.HandlerOptions{
+		MaxRequestBytes:  cfg.maxRequestBytes,
+		MaxResponseBytes: cfg.maxResponseBytes,
+	})
 	srv := &http.Server{
 		Addr:              cfg.addr,
 		Handler:           h,
@@ -299,6 +298,10 @@ func parseServeArgs(args []string) (*serveConfig, error) {
 	fs.StringVar(&cfg.promptDir, "prompt-dir", "", "directory containing prompt definition YAML files")
 	fs.StringVar(&cfg.profileDir, "profile-dir", "", "directory containing execution profile YAML files")
 	fs.StringVar(&cfg.schemaDir, "schema-dir", "", "base directory for validation schemas")
+	fs.StringVar(&cfg.artifactRoot, "artifact-root", "", "base directory for HTTP file input artifacts")
+	fs.Int64Var(&cfg.maxRequestBytes, "max-request-bytes", 0, "maximum HTTP request body bytes; 0 disables the limit")
+	fs.Int64Var(&cfg.maxArtifactBytes, "max-artifact-bytes", 0, "maximum HTTP file artifact bytes; 0 disables the limit")
+	fs.Int64Var(&cfg.maxResponseBytes, "max-response-bytes", 0, "maximum HTTP response body bytes; 0 disables the limit")
 
 	if err := fs.Parse(args); err != nil {
 		return nil, err
@@ -307,31 +310,41 @@ func parseServeArgs(args []string) (*serveConfig, error) {
 		return nil, fmt.Errorf("unexpected positional args: %v", fs.Args())
 	}
 
-	settings, err := resolveAppSettings(fs, cfg.configPath, appconfig.CLIOverrides{
+	settings, err := resolveCommonSettings(fs, cfg.configPath, appconfig.CLIOverrides{
 		PromptDir:        cfg.promptDirIfSet(fs),
 		ProfileDir:       cfg.profileDirIfSet(fs),
 		SchemaDir:        cfg.schemaDirIfSet(fs),
 		ServerAddr:       cfg.addrIfSet(fs),
+		ArtifactRoot:     cfg.artifactRootIfSet(fs),
+		MaxRequestBytes:  cfg.maxRequestBytesIfSet(fs),
+		MaxArtifactBytes: cfg.maxArtifactBytesIfSet(fs),
+		MaxResponseBytes: cfg.maxResponseBytesIfSet(fs),
 	})
 	if err != nil {
 		return nil, err
 	}
 
-	cfg.promptDir = settings.PromptDir
-	cfg.profileDir = settings.ProfileDir
-	cfg.schemaDir = settings.SchemaDir
-	cfg.addr = settings.ServerAddr
+	cfg.promptDir = settings.promptDir
+	cfg.profileDir = settings.profileDir
+	cfg.schemaDir = settings.schemaDir
+	cfg.addr = settings.serverAddr
+	cfg.artifactRoot = settings.artifactRoot
+	cfg.maxRequestBytes = settings.maxRequestBytes
+	cfg.maxArtifactBytes = settings.maxArtifactBytes
+	cfg.maxResponseBytes = settings.maxResponseBytes
 
-	if strings.TrimSpace(cfg.promptDir) == "" {
-		return nil, errors.New(errPromptDirRequired)
-	}
-	if strings.TrimSpace(cfg.profileDir) == "" {
-		return nil, errors.New(errProfileDirRequired)
+	if err := validateRequiredLibraryDirs(cfg.promptDir); err != nil {
+		return nil, err
 	}
 
 	cfg.promptDir = filepath.Clean(cfg.promptDir)
+	if strings.TrimSpace(cfg.profileDir) != "" {
 		cfg.profileDir = filepath.Clean(cfg.profileDir)
+	}
 	cfg.schemaDir = filepath.Clean(cfg.schemaDir)
+	if strings.TrimSpace(cfg.artifactRoot) != "" {
+		cfg.artifactRoot = filepath.Clean(cfg.artifactRoot)
+	}
 	return cfg, nil
 }
 
@@ -359,7 +372,7 @@ func finalizeExecutionRequestConfig(fs *flag.FlagSet, cfg *runConfig) error {
 		return fmt.Errorf("unexpected positional args: %v", fs.Args())
 	}
 
-	settings, err := resolveAppSettings(fs, cfg.configPath, appconfig.CLIOverrides{
+	settings, err := resolveCommonSettings(fs, cfg.configPath, appconfig.CLIOverrides{
 		PromptDir:  cfg.promptDirIfSet(fs),
 		ProfileDir: cfg.profileDirIfSet(fs),
 		SchemaDir:  cfg.schemaDirIfSet(fs),
@@ -368,16 +381,13 @@ func finalizeExecutionRequestConfig(fs *flag.FlagSet, cfg *runConfig) error {
 		return err
 	}
 
-	cfg.promptDir = settings.PromptDir
-	cfg.profileDir = settings.ProfileDir
-	cfg.schemaDir = settings.SchemaDir
-	cfg.defaultRenderFormat = settings.DefaultRenderFormat
+	cfg.promptDir = settings.promptDir
+	cfg.profileDir = settings.profileDir
+	cfg.schemaDir = settings.schemaDir
+	cfg.defaultRenderFormat = settings.defaultRenderFormat
 
-	if strings.TrimSpace(cfg.promptDir) == "" {
-		return errors.New(errPromptDirRequired)
-	}
-	if strings.TrimSpace(cfg.profileDir) == "" {
-		return errors.New(errProfileDirRequired)
+	if err := validateRequiredLibraryDirs(cfg.promptDir); err != nil {
+		return err
 	}
 	if strings.TrimSpace(cfg.promptID) == "" {
 		return errors.New("--prompt is required")
@@ -386,7 +396,9 @@ func finalizeExecutionRequestConfig(fs *flag.FlagSet, cfg *runConfig) error {
 		return errors.New("at least one --input is required")
 	}
 	cfg.promptDir = filepath.Clean(cfg.promptDir)
+	if strings.TrimSpace(cfg.profileDir) != "" {
 		cfg.profileDir = filepath.Clean(cfg.profileDir)
+	}
 	if cfg.outputPath != "" {
 		cfg.outputPath = filepath.Clean(cfg.outputPath)
 	}
@@ -449,6 +461,34 @@ func (c *serveConfig) addrIfSet(fs *flag.FlagSet) string {
 	return ""
 }
 
+func (c *serveConfig) artifactRootIfSet(fs *flag.FlagSet) string {
+	if flagWasSet(fs, "artifact-root") {
+		return c.artifactRoot
+	}
+	return ""
+}
+
+func (c *serveConfig) maxRequestBytesIfSet(fs *flag.FlagSet) *int64 {
+	if flagWasSet(fs, "max-request-bytes") {
+		return &c.maxRequestBytes
+	}
+	return nil
+}
+
+func (c *serveConfig) maxArtifactBytesIfSet(fs *flag.FlagSet) *int64 {
+	if flagWasSet(fs, "max-artifact-bytes") {
+		return &c.maxArtifactBytes
+	}
+	return nil
+}
+
+func (c *serveConfig) maxResponseBytesIfSet(fs *flag.FlagSet) *int64 {
+	if flagWasSet(fs, "max-response-bytes") {
+		return &c.maxResponseBytes
+	}
+	return nil
+}
+
 func registerConfigPathFlag(fs *flag.FlagSet, target *string) {
 	fs.StringVar(
 		target,
@@ -476,6 +516,55 @@ func resolveAppSettings(fs *flag.FlagSet, configPath string, overrides appconfig
 	return merged, nil
 }
 
+func resolveCommonSettings(fs *flag.FlagSet, configPath string, overrides appconfig.CLIOverrides) (commonCommandSettings, error) {
+	settings, err := resolveAppSettings(fs, configPath, overrides)
+	if err != nil {
+		return commonCommandSettings{}, err
+	}
+	return commonCommandSettings{
+		promptDir:           settings.PromptDir,
+		profileDir:          settings.ProfileDir,
+		schemaDir:           settings.SchemaDir,
+		serverAddr:          settings.ServerAddr,
+		artifactRoot:        settings.ArtifactRoot,
+		maxRequestBytes:     settings.MaxRequestBytes,
+		maxArtifactBytes:    settings.MaxArtifactBytes,
+		maxResponseBytes:    settings.MaxResponseBytes,
+		defaultRenderFormat: settings.DefaultRenderFormat,
+	}, nil
+}
+
+func validateRequiredLibraryDirs(promptDir string) error {
+	if strings.TrimSpace(promptDir) == "" {
+		return errors.New(errPromptDirRequired)
+	}
+	return nil
+}
+
+func newRunner(promptDir, profileDir, schemaDir string, llmClient llm.Client) *usecase.Runner {
+	return newRunnerWithArtifactReader(promptDir, profileDir, schemaDir, llmClient, artifactadapter.NewCompositeReader())
+}
+
+func newRunnerWithArtifactReader(promptDir, profileDir, schemaDir string, llmClient llm.Client, artifactReader artifactadapter.Reader) *usecase.Runner {
+	if artifactReader == nil {
+		artifactReader = artifactadapter.NewCompositeReader()
+	}
+	return usecase.NewRunner(
+		promptdef.NewFilesystemRepository(promptDir),
+		builtin.NewRepositoryWithDirectory(profileDir),
+		artifactReader,
+		prompt.NewGoRenderer(),
+		llmClient,
+		validate.NewStandardValidator(schemaDir),
+	)
+}
+
+func newOpenAIClient() (*llm.OpenAICompatibleClient, error) {
+	return llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{
+		Timeout: defaults.LLMRequestTimeoutDefault,
+	})
+}
+
 func buildRunRequestFromConfig(cfg *runConfig) (domain.RunRequest, error) {
 	inputMappings, err := parseMappings(cfg.inputRaw, false)
 	if err != nil {
@@ -495,18 +584,25 @@ func buildRunRequestFromConfig(cfg *runConfig) (domain.RunRequest, error) {
 		inputs[name] = domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: path}
 	}
 
-	var modelOverride *domain.ExecutionTarget
+	var modelOverride *domain.ExecutionTargetOverride
 	if cfg.llmBaseURLSet || cfg.modelSet || cfg.temperatureSet || cfg.maxTokensSet || cfg.topPSet || cfg.apiKeyEnvSet || cfg.timeoutSet {
-		modelOverride = &domain.ExecutionTarget{
+		modelOverride = &domain.ExecutionTargetOverride{
 			Endpoint:  cfg.llmBaseURL,
 			Model:     cfg.model,
-			Temperature: cfg.temperature,
-			MaxTokens:   cfg.maxTokens,
-			TopP:        cfg.topP,
 			APIKeyEnv: cfg.apiKeyEnv,
 		}
+		if cfg.temperatureSet {
+			modelOverride.Temperature = &cfg.temperature
+		}
+		if cfg.maxTokensSet {
+			modelOverride.MaxTokens = &cfg.maxTokens
+		}
+		if cfg.topPSet {
+			modelOverride.TopP = &cfg.topP
+		}
 		if cfg.timeoutSet {
-			modelOverride.TimeoutSeconds = int(cfg.timeout.Seconds())
+			timeoutSeconds := int(cfg.timeout.Seconds())
+			modelOverride.TimeoutSeconds = &timeoutSeconds
 		}
 	}
 
@@ -588,7 +684,7 @@ func printSummary(stderr io.Writer, res *domain.RunResult) {
 	if res == nil {
 		return
 	}
-	fmt.Fprintf(stderr, "prompt=%s@%s selected_profile=%s model=%s validation=%s mode=%s validation_errors=%d prompt_hash=%s inputs=%d usage=%d/%d/%d\n",
+	fmt.Fprintf(stderr, "prompt=%s@%s selected_profile=%s model=%s validation=%s mode=%s validation_errors=%d prompt_hash=%s inputs=%d usage=%d/%d/%d",
 		res.PromptID,
 		res.PromptVersion,
 		res.SelectedProfileID,
@@ -602,11 +698,15 @@ func printSummary(stderr io.Writer, res *domain.RunResult) {
 		res.Usage.CompletionTokens,
 		res.Usage.TotalTokens,
 	)
+	if res.Usage.CachedTokens != 0 || res.Usage.CacheWriteTokens != 0 {
+		fmt.Fprintf(stderr, " cached_tokens=%d cache_write_tokens=%d", res.Usage.CachedTokens, res.Usage.CacheWriteTokens)
+	}
+	fmt.Fprintln(stderr)
 }
 
 func printUsage(w io.Writer) {
 	fmt.Fprintln(w, "usage: scriptorium <run|render|serve> ...")
 	fmt.Fprintln(w, "  run:   scriptorium run [--config PATH] [--prompt-dir DIR] [--profile-dir DIR] --prompt ID --input name=path [--input ...] [--profile ID] [--llm-base-url URL] [--model NAME] [--api-key-env ENV] [--temperature N] [--max-tokens N] [--top-p N] [--var k=v] [--out path] [--timeout 10m]")
 	fmt.Fprintln(w, "  render: scriptorium render [--config PATH] [--prompt-dir DIR] [--profile-dir DIR] --prompt ID --input name=path [--input ...] [--profile ID] [--llm-base-url URL] [--model NAME] [--api-key-env ENV] [--temperature N] [--max-tokens N] [--top-p N] [--var k=v] [--format text|json] [--out path] [--timeout 10m]")
-	fmt.Fprintf(w, "  serve: scriptorium serve [--config PATH] [--addr %s] [--prompt-dir DIR] [--profile-dir DIR] [--schema-dir DIR]\n", defaults.HTTPAddrDefault)
+	fmt.Fprintf(w, "  serve: scriptorium serve [--config PATH] [--addr %s] [--prompt-dir DIR] [--profile-dir DIR] [--schema-dir DIR] [--artifact-root DIR] [--max-request-bytes N] [--max-artifact-bytes N] [--max-response-bytes N]\n", defaults.HTTPAddrDefault)
 }

internal/adapter/http/dto.go

@@ -23,13 +23,14 @@ type inputRefDTO struct {
 type modelOverrideRequestDTO struct {
 	Endpoint        string         `json:"endpoint,omitempty"`
 	Model           string         `json:"model,omitempty"`
-	Temperature     float64           `json:"temperature,omitempty"`
-	MaxTokens       int               `json:"max_tokens,omitempty"`
-	TopP            float64           `json:"top_p,omitempty"`
-	TimeoutSeconds  int               `json:"timeout_seconds,omitempty"`
+	Temperature     *float64       `json:"temperature,omitempty"`
+	MaxTokens       *int           `json:"max_tokens,omitempty"`
+	TopP            *float64       `json:"top_p,omitempty"`
+	TimeoutSeconds  *int           `json:"timeout_seconds,omitempty"`
+	ServiceTier     string         `json:"service_tier,omitempty"`
 	ReasoningEffort string         `json:"reasoning_effort,omitempty"`
 	APIKeyEnv       string         `json:"api_key_env,omitempty"`
-	ExtraParams     map[string]string `json:"extra_params,omitempty"`
+	ExtraParams     map[string]any `json:"extra_params,omitempty"`
 }
 
 type runResponseDTO struct {
@@ -75,15 +76,18 @@ type modelParamsDTO struct {
 	MaxTokens       int            `json:"max_tokens"`
 	TopP            float64        `json:"top_p"`
 	TimeoutSeconds  int            `json:"timeout_seconds"`
+	ServiceTier     string         `json:"service_tier,omitempty"`
 	ReasoningEffort string         `json:"reasoning_effort,omitempty"`
 	APIKeyEnv       string         `json:"api_key_env,omitempty"`
-	ExtraParams     map[string]string `json:"extra_params,omitempty"`
+	ExtraParams     map[string]any `json:"extra_params,omitempty"`
 }
 
 type tokenUsageDTO struct {
 	PromptTokens     int `json:"prompt_tokens"`
 	CompletionTokens int `json:"completion_tokens"`
 	TotalTokens      int `json:"total_tokens"`
+	CachedTokens     int `json:"cached_tokens"`
+	CacheWriteTokens int `json:"cache_write_tokens"`
 }
 
 type validationDTO struct {

internal/adapter/http/handler.go

@@ -4,9 +4,12 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"io"
 	"net/http"
 	"strings"
 
+	"gitea.maximumdirect.net/eric/scriptorium/internal/artifact"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
@@ -19,10 +22,23 @@ type Runner interface {
 
 type Handler struct {
 	runner  Runner
+	options HandlerOptions
+}
+
+type HandlerOptions struct {
+	MaxRequestBytes  int64
+	MaxResponseBytes int64
 }
 
 func NewHandler(runner Runner) *Handler {
-	return &Handler{runner: runner}
+	return NewHandlerWithOptions(runner, HandlerOptions{
+		MaxRequestBytes:  defaults.HTTPMaxRequestBytesDefault,
+		MaxResponseBytes: defaults.HTTPMaxResponseBytesDefault,
+	})
+}
+
+func NewHandlerWithOptions(runner Runner, options HandlerOptions) *Handler {
+	return &Handler{runner: runner, options: options}
 }
 
 func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
@@ -36,9 +52,26 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	}
 
 	var req runRequestDTO
-	dec := json.NewDecoder(r.Body)
+	body := r.Body
+	if h.options.MaxRequestBytes > 0 {
+		body = http.MaxBytesReader(w, r.Body, h.options.MaxRequestBytes)
+	}
+	dec := json.NewDecoder(body)
 	dec.DisallowUnknownFields()
 	if err := dec.Decode(&req); err != nil {
+		if isRequestTooLarge(err) {
+			writeError(w, http.StatusRequestEntityTooLarge, "request_too_large", "request body is too large")
+			return
+		}
+		writeError(w, http.StatusBadRequest, "invalid_json", "invalid JSON request body")
+		return
+	}
+	var trailing any
+	if err := dec.Decode(&trailing); err != io.EOF {
+		if isRequestTooLarge(err) {
+			writeError(w, http.StatusRequestEntityTooLarge, "request_too_large", "request body is too large")
+			return
+		}
 		writeError(w, http.StatusBadRequest, "invalid_json", "invalid JSON request body")
 		return
 	}
@@ -61,19 +94,9 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
-	var model *domain.ExecutionTarget
+	var model *domain.ExecutionTargetOverride
 	if req.Model != nil {
-		model = &domain.ExecutionTarget{
-			Endpoint:        req.Model.Endpoint,
-			Model:           req.Model.Model,
-			Temperature:     req.Model.Temperature,
-			MaxTokens:       req.Model.MaxTokens,
-			TopP:            req.Model.TopP,
-			TimeoutSeconds:  req.Model.TimeoutSeconds,
-			ReasoningEffort: req.Model.ReasoningEffort,
-			APIKeyEnv:       req.Model.APIKeyEnv,
-			ExtraParams:     req.Model.ExtraParams,
-		}
+		model = executionTargetOverrideFromModelOverrideDTO(req.Model)
 	}
 
 	res, err := h.runner.Run(r.Context(), domain.RunRequest{
@@ -109,22 +132,14 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 			SelectedProfileID:  res.SelectedProfileID,
 			ModelName:          res.ModelName,
 			Endpoint:           res.Endpoint,
-			ModelParams: modelParamsDTO{
-				Endpoint:        res.EffectiveModelParams.Endpoint,
-				Model:           res.EffectiveModelParams.Model,
-				Temperature:     res.EffectiveModelParams.Temperature,
-				MaxTokens:       res.EffectiveModelParams.MaxTokens,
-				TopP:            res.EffectiveModelParams.TopP,
-				TimeoutSeconds:  res.EffectiveModelParams.TimeoutSeconds,
-				ReasoningEffort: res.EffectiveModelParams.ReasoningEffort,
-				APIKeyEnv:       res.EffectiveModelParams.APIKeyEnv,
-				ExtraParams:     res.EffectiveModelParams.ExtraParams,
-			},
+			ModelParams:        modelParamsDTOFromExecutionTarget(res.EffectiveModelParams),
 			InputHashes:        res.InputHashes,
 			Usage: tokenUsageDTO{
 				PromptTokens:     res.Usage.PromptTokens,
 				CompletionTokens: res.Usage.CompletionTokens,
 				TotalTokens:      res.Usage.TotalTokens,
+				CachedTokens:     res.Usage.CachedTokens,
+				CacheWriteTokens: res.Usage.CacheWriteTokens,
 			},
 			StartTime:          res.StartTime,
 			EndTime:            res.EndTime,
@@ -138,7 +153,40 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		raw := res.RawOutput
 		resp.RawModelOutput = &raw
 	}
-	writeJSON(w, http.StatusOK, resp)
+	writeLimitedJSON(w, http.StatusOK, resp, h.options.MaxResponseBytes)
+}
+
+func executionTargetOverrideFromModelOverrideDTO(dto *modelOverrideRequestDTO) *domain.ExecutionTargetOverride {
+	if dto == nil {
+		return nil
+	}
+	return &domain.ExecutionTargetOverride{
+		Endpoint:        dto.Endpoint,
+		Model:           dto.Model,
+		Temperature:     dto.Temperature,
+		MaxTokens:       dto.MaxTokens,
+		TopP:            dto.TopP,
+		TimeoutSeconds:  dto.TimeoutSeconds,
+		ServiceTier:     dto.ServiceTier,
+		ReasoningEffort: dto.ReasoningEffort,
+		APIKeyEnv:       dto.APIKeyEnv,
+		ExtraParams:     dto.ExtraParams,
+	}
+}
+
+func modelParamsDTOFromExecutionTarget(target domain.ExecutionTarget) modelParamsDTO {
+	return modelParamsDTO{
+		Endpoint:        target.Endpoint,
+		Model:           target.Model,
+		Temperature:     target.Temperature,
+		MaxTokens:       target.MaxTokens,
+		TopP:            target.TopP,
+		TimeoutSeconds:  target.TimeoutSeconds,
+		ServiceTier:     target.ServiceTier,
+		ReasoningEffort: target.ReasoningEffort,
+		APIKeyEnv:       target.APIKeyEnv,
+		ExtraParams:     target.ExtraParams,
+	}
 }
 
 func mapValidation(v domain.ValidationResult) validationDTO {
@@ -160,16 +208,22 @@ func mapRunError(err error) (int, string, string) {
 		return http.StatusNotFound, "profile_not_found", "execution profile not found"
 	case errors.Is(err, promptdef.ErrInvalidYAML), errors.Is(err, promptdef.ErrInvalidPromptDefinition):
 		return http.StatusBadRequest, "prompt_load_failed", "failed to load prompt definition"
-	case errors.Is(err, profile.ErrInvalidYAML), errors.Is(err, profile.ErrInvalidProfile):
+	case errors.Is(err, profile.ErrInvalidYAML), errors.Is(err, profile.ErrInvalidProfile), errors.Is(err, profile.ErrRawAPIKeyNotAllowed):
 		return http.StatusBadRequest, "profile_load_failed", "failed to load execution profile"
-	case errors.Is(err, usecase.ErrInvalidRequest) && strings.Contains(err.Error(), "profile id is required either in request or prompt default_profile"):
+	case errors.Is(err, usecase.ErrProfileRequired):
 		return http.StatusBadRequest, "profile_required", "profile_id is required when prompt default_profile is not set"
-	case errors.Is(err, usecase.ErrInvalidRequest) && strings.Contains(err.Error(), "api key environment variable"):
+	case errors.Is(err, usecase.ErrAPIKeyEnvMissing):
 		return http.StatusBadRequest, "api_key_env_missing", "api_key_env is set but the environment variable is missing"
 	case errors.Is(err, usecase.ErrInvalidRequest):
 		return http.StatusBadRequest, "invalid_request", "invalid run request"
-	case errors.Is(err, usecase.ErrProfileLoad):
+	case errors.Is(err, usecase.ErrPromptLoad):
 		return http.StatusBadRequest, "prompt_load_failed", "failed to load prompt definition"
+	case errors.Is(err, usecase.ErrProfileLoad):
+		return http.StatusBadRequest, "profile_load_failed", "failed to load execution profile"
+	case errors.Is(err, artifact.ErrFileNotAllowed), errors.Is(err, artifact.ErrFileOutsideRoot):
+		return http.StatusBadRequest, "artifact_not_allowed", "file input artifact is not allowed"
+	case errors.Is(err, artifact.ErrFileTooLarge):
+		return http.StatusRequestEntityTooLarge, "artifact_too_large", "file input artifact is too large"
 	case errors.Is(err, usecase.ErrArtifactLoad):
 		return http.StatusBadRequest, "artifact_read_failed", "failed to read input artifact"
 	case errors.Is(err, usecase.ErrPromptRender):
@@ -184,9 +238,23 @@ func mapRunError(err error) (int, string, string) {
 }
 
 func writeJSON(w http.ResponseWriter, status int, v any) {
+	writeLimitedJSON(w, status, v, 0)
+}
+
+func writeLimitedJSON(w http.ResponseWriter, status int, v any, maxBytes int64) {
+	data, err := json.Marshal(v)
+	if err != nil {
+		writeError(w, http.StatusInternalServerError, "internal_error", "internal server error")
+		return
+	}
+	data = append(data, '\n')
+	if maxBytes > 0 && int64(len(data)) > maxBytes {
+		writeError(w, http.StatusRequestEntityTooLarge, "response_too_large", "response body is too large")
+		return
+	}
 	w.Header().Set("Content-Type", "application/json")
 	w.WriteHeader(status)
-	_ = json.NewEncoder(w).Encode(v)
+	_, _ = w.Write(data)
 }
 
 func writeError(w http.ResponseWriter, status int, code, message string) {
@@ -197,3 +265,8 @@ func writeError(w http.ResponseWriter, status int, code, message string) {
 		},
 	})
 }
+
+func isRequestTooLarge(err error) bool {
+	var maxBytesErr *http.MaxBytesError
+	return errors.As(err, &maxBytesErr)
+}

internal/adapter/http/handler_test.go

@@ -4,15 +4,19 @@ import (
 	"bytes"
 	"context"
 	"encoding/json"
-	"errors"
 	"fmt"
 	"net/http"
 	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"reflect"
 	"strings"
 	"testing"
 	"time"
 
+	"gitea.maximumdirect.net/eric/scriptorium/internal/artifact"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/llm"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/usecase"
@@ -32,6 +36,40 @@ func (f *fakeRunner) Run(ctx context.Context, req domain.RunRequest) (*domain.Ru
 	return f.result, nil
 }
 
+type handlerPromptRepo struct {
+	def *domain.PromptDefinition
+}
+
+func (r handlerPromptRepo) GetPromptDefinition(ctx context.Context, id string, version string) (*domain.PromptDefinition, error) {
+	return r.def, nil
+}
+
+type handlerProfileRepo struct {
+	profile *domain.ExecutionProfile
+}
+
+func (r handlerProfileRepo) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	return r.profile, nil
+}
+
+type handlerArtifactReader struct{}
+
+func (handlerArtifactReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	return &domain.Artifact{Name: "input", Body: []byte("input"), Hash: "hash"}, nil
+}
+
+type handlerRenderer struct{}
+
+func (handlerRenderer) Render(ctx context.Context, definition *domain.PromptDefinition, inputs map[string]*domain.Artifact, vars map[string]string) (*domain.RenderedPrompt, error) {
+	return &domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}}, nil
+}
+
+type handlerLLMClient struct{}
+
+func (handlerLLMClient) Generate(ctx context.Context, req domain.GenerateRequest) (*domain.GenerateResponse, error) {
+	return &domain.GenerateResponse{Content: "ok"}, nil
+}
+
 func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 	start := time.Now().UTC()
 	end := start.Add(2 * time.Second)
@@ -62,10 +100,17 @@ func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 			MaxTokens:      42,
 			TopP:           0.9,
 			TimeoutSeconds: 120,
+			ServiceTier:    "priority",
 			APIKeyEnv:      envName,
 		},
 		InputHashes: map[string]string{"transcript": "h1"},
-		Usage:       domain.TokenUsage{PromptTokens: 1, CompletionTokens: 2, TotalTokens: 3},
+		Usage: domain.TokenUsage{
+			PromptTokens:     1,
+			CompletionTokens: 2,
+			TotalTokens:      3,
+			CachedTokens:     4,
+			CacheWriteTokens: 5,
+		},
 		StartTime: start,
 		EndTime:   end,
 		Duration:  2 * time.Second,
@@ -81,7 +126,7 @@ func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
     "transcript": {"type": "file", "uri": "./t.md"}
   },
   "vars": {"k": "v"},
-  "model": {"model": "gpt-x", "timeout_seconds": 120, "api_key_env": "SCRIPTORIUM_API_KEY"}
+  "model": {"model": "gpt-x", "timeout_seconds": 120, "service_tier": "flex", "api_key_env": "SCRIPTORIUM_API_KEY"}
 }`)
 	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewReader(body))
 	w := httptest.NewRecorder()
@@ -110,10 +155,20 @@ func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 	if metadata["model_name"] != "m1" || metadata["endpoint"] != "http://llm/v1" {
 		t.Fatalf("unexpected model metadata: name=%#v endpoint=%#v", metadata["model_name"], metadata["endpoint"])
 	}
+	usage := metadata["usage"].(map[string]any)
+	if usage["prompt_tokens"] != float64(1) || usage["completion_tokens"] != float64(2) || usage["total_tokens"] != float64(3) {
+		t.Fatalf("unexpected base usage metadata: %#v", usage)
+	}
+	if usage["cached_tokens"] != float64(4) || usage["cache_write_tokens"] != float64(5) {
+		t.Fatalf("unexpected cache usage metadata: %#v", usage)
+	}
 	modelParams := metadata["model_params"].(map[string]any)
 	if modelParams["api_key_env"] != envName {
 		t.Fatalf("expected model_params.api_key_env=%q, got %#v", envName, modelParams["api_key_env"])
 	}
+	if modelParams["service_tier"] != "priority" {
+		t.Fatalf("expected model_params.service_tier=priority, got %#v", modelParams["service_tier"])
+	}
 	if strings.Contains(w.Body.String(), secret) {
 		t.Fatalf("response leaked raw API key value: %s", w.Body.String())
 	}
@@ -130,9 +185,111 @@ func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 	if r.last.Execution == nil || r.last.Execution.Model != "gpt-x" {
 		t.Fatalf("expected model override, got %#v", r.last.Execution)
 	}
-	if r.last.Execution.TimeoutSeconds != 120 {
+	if r.last.Execution.TimeoutSeconds == nil || *r.last.Execution.TimeoutSeconds != 120 {
 		t.Fatalf("expected timeout_seconds override 120, got %#v", r.last.Execution)
 	}
+	if r.last.Execution.ServiceTier != "flex" {
+		t.Fatalf("expected service_tier override flex, got %#v", r.last.Execution)
+	}
+}
+
+func TestHandlerInlineRefsWorkWithoutArtifactRoot(t *testing.T) {
+	h := newArtifactRootHandler(t, "")
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"inline","body":"inline body"}}
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+}
+
+func TestHandlerFileRefsWithoutArtifactRootAreRejected(t *testing.T) {
+	h := newArtifactRootHandler(t, "")
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"file","uri":"input.txt"}}
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusBadRequest, "artifact_not_allowed")
+}
+
+func TestHandlerFileRefsUnderArtifactRootWork(t *testing.T) {
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "input.txt"), []byte("allowed"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	h := newArtifactRootHandler(t, root)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"file","uri":"input.txt"}}
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+}
+
+func TestHandlerFileRefsAboveArtifactLimitAreRejected(t *testing.T) {
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "large.txt"), []byte("123456"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	h := newArtifactRootHandlerWithLimit(t, root, 5)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"file","uri":"large.txt"}}
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusRequestEntityTooLarge, "artifact_too_large")
+}
+
+func TestHandlerFileRefsOutsideArtifactRootAreRejected(t *testing.T) {
+	root := t.TempDir()
+	outside := t.TempDir()
+	if err := os.WriteFile(filepath.Join(outside, "secret.txt"), []byte("denied"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	h := newArtifactRootHandler(t, root)
+
+	tests := []struct {
+		name string
+		uri  string
+	}{
+		{name: "relative traversal", uri: filepath.Join("..", filepath.Base(outside), "secret.txt")},
+		{name: "absolute outside root", uri: filepath.Join(outside, "secret.txt")},
+	}
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			body := fmt.Sprintf(`{
+				"prompt_id":"p",
+				"inputs":{"x":{"type":"file","uri":%q}}
+			}`, tc.uri)
+			req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(body))
+			w := httptest.NewRecorder()
+
+			h.ServeHTTP(w, req)
+
+			assertHTTPErrorCode(t, w, http.StatusBadRequest, "artifact_not_allowed")
+		})
+	}
 }
 
 func TestHandlerPostRunsSuccessUsingPromptDefaultProfile(t *testing.T) {
@@ -164,6 +321,265 @@ func TestHandlerPostRunsSuccessUsingPromptDefaultProfile(t *testing.T) {
 	if metadata["selected_profile_id"] != "prompt-default" {
 		t.Fatalf("expected selected_profile_id from result, got %#v", metadata["selected_profile_id"])
 	}
+	usage := metadata["usage"].(map[string]any)
+	if usage["cached_tokens"] != float64(0) || usage["cache_write_tokens"] != float64(0) {
+		t.Fatalf("expected zero cache usage fields to be included, got %#v", usage)
+	}
+}
+
+func TestHandlerModelOverrideMapsAllSupportedExecutionFields(t *testing.T) {
+	r := &fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte("ok")},
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1"},
+	}}
+	h := NewHandler(r)
+
+	reqBody := `{
+		"prompt_id": "prompt-1",
+		"inputs": {"transcript": {"type": "file", "uri": "./t.md"}},
+		"model": {
+			"endpoint": "http://override/v1",
+			"model": "override-model",
+			"temperature": 0.6,
+			"max_tokens": 250,
+			"top_p": 0.85,
+			"timeout_seconds": 33,
+			"service_tier": "flex",
+			"reasoning_effort": "medium",
+			"api_key_env": "SCRIPTORIUM_API_KEY",
+			"extra_params": {"provider_option":"on"}
+		}
+	}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(reqBody))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+	if r.last.Execution == nil {
+		t.Fatalf("expected execution override in run request")
+	}
+	got := r.last.Execution
+	if got.Endpoint != "http://override/v1" ||
+		got.Model != "override-model" ||
+		got.ServiceTier != "flex" ||
+		got.ReasoningEffort != "medium" ||
+		got.APIKeyEnv != "SCRIPTORIUM_API_KEY" {
+		t.Fatalf("unexpected mapped execution target: %+v", got)
+	}
+	if got.Temperature == nil || *got.Temperature != 0.6 {
+		t.Fatalf("unexpected mapped temperature: %#v", got.Temperature)
+	}
+	if got.MaxTokens == nil || *got.MaxTokens != 250 {
+		t.Fatalf("unexpected mapped max_tokens: %#v", got.MaxTokens)
+	}
+	if got.TopP == nil || *got.TopP != 0.85 {
+		t.Fatalf("unexpected mapped top_p: %#v", got.TopP)
+	}
+	if got.TimeoutSeconds == nil || *got.TimeoutSeconds != 33 {
+		t.Fatalf("unexpected mapped timeout_seconds: %#v", got.TimeoutSeconds)
+	}
+	if !reflect.DeepEqual(got.ExtraParams, map[string]any{"provider_option": "on"}) {
+		t.Fatalf("unexpected mapped extra_params: %#v", got.ExtraParams)
+	}
+}
+
+func TestHandlerModelOverrideAcceptsJSONCompatibleExtraParams(t *testing.T) {
+	r := &fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte("ok")},
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1"},
+	}}
+	h := NewHandler(r)
+
+	reqBody := `{
+		"prompt_id": "prompt-1",
+		"inputs": {"transcript": {"type": "file", "uri": "./t.md"}},
+		"model": {
+			"extra_params": {
+				"string_value": "enabled",
+				"number_value": 42,
+				"boolean_value": true,
+				"object_value": {"nested": "value", "count": 2},
+				"array_value": ["first", 3, false]
+			}
+		}
+	}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(reqBody))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+	if r.last.Execution == nil {
+		t.Fatal("expected execution override in run request")
+	}
+	want := map[string]any{
+		"string_value":  "enabled",
+		"number_value":  float64(42),
+		"boolean_value": true,
+		"object_value":  map[string]any{"nested": "value", "count": float64(2)},
+		"array_value":   []any{"first", float64(3), false},
+	}
+	if !reflect.DeepEqual(r.last.Execution.ExtraParams, want) {
+		t.Fatalf("unexpected mapped extra_params:\ngot=%#v\nwant=%#v", r.last.Execution.ExtraParams, want)
+	}
+}
+
+func TestHandlerModelOverrideExplicitZeroTemperatureMapsAsPresent(t *testing.T) {
+	r := &fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte("ok")},
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1", Temperature: 0},
+	}}
+	h := NewHandler(r)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id": "prompt-1",
+		"inputs": {"transcript": {"type": "file", "uri": "./t.md"}},
+		"model": {"temperature": 0}
+	}`))
+	w := httptest.NewRecorder()
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+	if r.last.Execution == nil || r.last.Execution.Temperature == nil {
+		t.Fatalf("expected temperature override to be present, got %#v", r.last.Execution)
+	}
+	if *r.last.Execution.Temperature != 0 {
+		t.Fatalf("expected zero temperature override, got %v", *r.last.Execution.Temperature)
+	}
+}
+
+func TestHandlerModelOverrideOmittedTemperatureMapsAsAbsent(t *testing.T) {
+	r := &fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte("ok")},
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1", Temperature: 0.7},
+	}}
+	h := NewHandler(r)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id": "prompt-1",
+		"inputs": {"transcript": {"type": "file", "uri": "./t.md"}},
+		"model": {"model": "override-model"}
+	}`))
+	w := httptest.NewRecorder()
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+	if r.last.Execution == nil {
+		t.Fatal("expected model override")
+	}
+	if r.last.Execution.Temperature != nil {
+		t.Fatalf("expected omitted temperature to remain absent, got %#v", r.last.Execution.Temperature)
+	}
+
+	var resp map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("invalid JSON response: %v", err)
+	}
+	metadata := resp["metadata"].(map[string]any)
+	params := metadata["model_params"].(map[string]any)
+	if params["temperature"] != 0.7 {
+		t.Fatalf("expected effective profile/default temperature in response, got %#v", params["temperature"])
+	}
+}
+
+func TestHandlerResponseMetadataModelParamsIncludesAllSupportedFields(t *testing.T) {
+	r := &fakeRunner{result: &domain.RunResult{
+		Artifact: domain.Artifact{
+			Name:        "output",
+			ContentType: "text/plain",
+			Body:        []byte("ok"),
+			Size:        2,
+			Hash:        "abc",
+		},
+		Validation: domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{
+			Endpoint:        "http://llm/v1",
+			Model:           "gpt-test",
+			Temperature:     0.4,
+			MaxTokens:       321,
+			TopP:            0.7,
+			TimeoutSeconds:  45,
+			ServiceTier:     "priority",
+			ReasoningEffort: "high",
+			APIKeyEnv:       "SCRIPTORIUM_API_KEY",
+			ExtraParams: map[string]any{
+				"provider_option": "on",
+				"number_value":    42,
+				"object_value":    map[string]any{"nested": "value"},
+			},
+		},
+	}}
+	h := NewHandler(r)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{"prompt_id":"p","inputs":{"x":{"type":"file","uri":"a"}}}`))
+	w := httptest.NewRecorder()
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", w.Code, w.Body.String())
+	}
+
+	var resp map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("invalid JSON response: %v", err)
+	}
+	metadata := resp["metadata"].(map[string]any)
+	params := metadata["model_params"].(map[string]any)
+
+	if params["endpoint"] != "http://llm/v1" {
+		t.Fatalf("unexpected endpoint: %#v", params["endpoint"])
+	}
+	if params["model"] != "gpt-test" {
+		t.Fatalf("unexpected model: %#v", params["model"])
+	}
+	if params["temperature"] != 0.4 {
+		t.Fatalf("unexpected temperature: %#v", params["temperature"])
+	}
+	if params["max_tokens"] != float64(321) {
+		t.Fatalf("unexpected max_tokens: %#v", params["max_tokens"])
+	}
+	if params["top_p"] != 0.7 {
+		t.Fatalf("unexpected top_p: %#v", params["top_p"])
+	}
+	if params["timeout_seconds"] != float64(45) {
+		t.Fatalf("unexpected timeout_seconds: %#v", params["timeout_seconds"])
+	}
+	if params["service_tier"] != "priority" {
+		t.Fatalf("unexpected service_tier: %#v", params["service_tier"])
+	}
+	if params["reasoning_effort"] != "high" {
+		t.Fatalf("unexpected reasoning_effort: %#v", params["reasoning_effort"])
+	}
+	if params["api_key_env"] != "SCRIPTORIUM_API_KEY" {
+		t.Fatalf("unexpected api_key_env: %#v", params["api_key_env"])
+	}
+	extraParams, ok := params["extra_params"].(map[string]any)
+	if !ok {
+		t.Fatalf("expected extra_params object, got %#v", params["extra_params"])
+	}
+	if extraParams["provider_option"] != "on" {
+		t.Fatalf("unexpected extra_params.provider_option: %#v", extraParams["provider_option"])
+	}
+	if extraParams["number_value"] != float64(42) {
+		t.Fatalf("unexpected extra_params.number_value: %#v", extraParams["number_value"])
+	}
+	objectValue, ok := extraParams["object_value"].(map[string]any)
+	if !ok || objectValue["nested"] != "value" {
+		t.Fatalf("unexpected extra_params.object_value: %#v", extraParams["object_value"])
+	}
 }
 
 func TestHandlerInvalidJSON(t *testing.T) {
@@ -178,6 +594,69 @@ func TestHandlerInvalidJSON(t *testing.T) {
 	}
 }
 
+func TestHandlerRejectsTrailingJSON(t *testing.T) {
+	h := NewHandler(&fakeRunner{})
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{"prompt_id":"p","inputs":{"x":{"type":"file","uri":"a"}}} {}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusBadRequest, "invalid_json")
+}
+
+func TestHandlerRequestTooLarge(t *testing.T) {
+	h := NewHandlerWithOptions(&fakeRunner{}, HandlerOptions{MaxRequestBytes: 12})
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{"prompt_id":"p","inputs":{"x":{"type":"file","uri":"a"}}}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusRequestEntityTooLarge, "request_too_large")
+}
+
+func TestHandlerMalformedJSONBelowLimitStillBadRequest(t *testing.T) {
+	h := NewHandlerWithOptions(&fakeRunner{}, HandlerOptions{MaxRequestBytes: 1024})
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString("{"))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusBadRequest, "invalid_json")
+}
+
+func TestHandlerResponseTooLarge(t *testing.T) {
+	h := NewHandlerWithOptions(&fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte(strings.Repeat("x", 128))},
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1"},
+	}}, HandlerOptions{MaxRequestBytes: 1024, MaxResponseBytes: 64})
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{"prompt_id":"p","inputs":{"x":{"type":"file","uri":"a"}}}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusRequestEntityTooLarge, "response_too_large")
+}
+
+func TestHandlerRawOutputDoesNotBypassResponseLimit(t *testing.T) {
+	h := NewHandlerWithOptions(&fakeRunner{result: &domain.RunResult{
+		Artifact:             domain.Artifact{Body: []byte("ok")},
+		RawOutput:            strings.Repeat("raw", 80),
+		Validation:           domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic, IsValid: true},
+		EffectiveModelParams: domain.ExecutionTarget{Endpoint: "http://llm/v1", Model: "m1"},
+	}}, HandlerOptions{MaxRequestBytes: 1024, MaxResponseBytes: 128})
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"file","uri":"a"}},
+		"include_raw_output":true
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	assertHTTPErrorCode(t, w, http.StatusRequestEntityTooLarge, "response_too_large")
+}
+
 func TestHandlerMissingPromptID(t *testing.T) {
 	h := NewHandler(&fakeRunner{})
 	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{"inputs":{"x":{"type":"file","uri":"a"}}}`))
@@ -198,6 +677,54 @@ func TestHandlerMissingPromptID(t *testing.T) {
 	}
 }
 
+func TestHandlerReservedExtraParamsThroughRunnerMapsToInvalidRequest(t *testing.T) {
+	llmClient, err := llm.NewOpenAICompatibleClient(llm.OpenAICompatibleConfig{})
+	if err != nil {
+		t.Fatal(err)
+	}
+	runner := usecase.NewRunner(
+		handlerPromptRepo{def: &domain.PromptDefinition{
+			ID:             "p",
+			Version:        "1",
+			DefaultProfile: "exec",
+			Templates:      []domain.PromptMessageTemplate{{Role: "user", Content: "hi"}},
+			OutputFormat:   domain.FormatText,
+			Validation:     domain.OutputContract{Format: domain.FormatText, ValidationMode: domain.ValidationNone},
+		}},
+		handlerProfileRepo{profile: &domain.ExecutionProfile{
+			ID:       "exec",
+			Endpoint: "http://example.invalid/v1",
+			Model:    "model",
+		}},
+		handlerArtifactReader{},
+		handlerRenderer{},
+		llmClient,
+		nil,
+	)
+	h := NewHandler(runner)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/runs", bytes.NewBufferString(`{
+		"prompt_id":"p",
+		"inputs":{"x":{"type":"file","uri":"a"}},
+		"model":{"extra_params":{"model":"collision"}}
+	}`))
+	w := httptest.NewRecorder()
+
+	h.ServeHTTP(w, req)
+
+	if w.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", w.Code, w.Body.String())
+	}
+	var resp map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("invalid JSON response: %v", err)
+	}
+	errBody := resp["error"].(map[string]any)
+	if errBody["code"] != "invalid_request" {
+		t.Fatalf("expected invalid_request code, got %#v", errBody["code"])
+	}
+}
+
 func TestHandlerUsecaseErrorMapping(t *testing.T) {
 	tests := []struct {
 		name       string
@@ -207,12 +734,14 @@ func TestHandlerUsecaseErrorMapping(t *testing.T) {
 		message    string
 		avoidCause string
 	}{
-		{name: "prompt not found", err: wrap(usecase.ErrProfileLoad, promptdef.ErrPromptDefinitionNotFound), status: http.StatusNotFound, code: "prompt_not_found", message: "prompt definition not found"},
-		{name: "prompt load invalid", err: wrap(usecase.ErrProfileLoad, promptdef.ErrInvalidPromptDefinition), status: http.StatusBadRequest, code: "prompt_load_failed", message: "failed to load prompt definition"},
-		{name: "missing profile/default", err: wrap(usecase.ErrInvalidRequest, errors.New("profile id is required either in request or prompt default_profile")), status: http.StatusBadRequest, code: "profile_required", message: "profile_id is required when prompt default_profile is not set"},
+		{name: "prompt not found", err: wrap(usecase.ErrPromptLoad, promptdef.ErrPromptDefinitionNotFound), status: http.StatusNotFound, code: "prompt_not_found", message: "prompt definition not found"},
+		{name: "prompt load invalid", err: wrap(usecase.ErrPromptLoad, promptdef.ErrInvalidPromptDefinition), status: http.StatusBadRequest, code: "prompt_load_failed", message: "failed to load prompt definition"},
+		{name: "prompt load generic", err: wrap(usecase.ErrPromptLoad, fmt.Errorf("read failed")), status: http.StatusBadRequest, code: "prompt_load_failed", message: "failed to load prompt definition", avoidCause: "read failed"},
+		{name: "missing profile/default", err: wrap(usecase.ErrInvalidRequest, usecase.ErrProfileRequired), status: http.StatusBadRequest, code: "profile_required", message: "profile_id is required when prompt default_profile is not set"},
 		{name: "profile not found", err: wrap(usecase.ErrProfileLoad, profile.ErrProfileNotFound), status: http.StatusNotFound, code: "profile_not_found", message: "execution profile not found"},
 		{name: "profile invalid", err: wrap(usecase.ErrProfileLoad, profile.ErrInvalidProfile), status: http.StatusBadRequest, code: "profile_load_failed", message: "failed to load execution profile"},
-		{name: "api key env missing", err: wrap(usecase.ErrInvalidRequest, errors.New(`api key environment variable "SCRIPTORIUM_API_KEY" is not set`)), status: http.StatusBadRequest, code: "api_key_env_missing", message: "api_key_env is set but the environment variable is missing"},
+		{name: "profile load generic", err: wrap(usecase.ErrProfileLoad, fmt.Errorf("read failed")), status: http.StatusBadRequest, code: "profile_load_failed", message: "failed to load execution profile", avoidCause: "read failed"},
+		{name: "api key env missing", err: wrap(usecase.ErrInvalidRequest, usecase.ErrAPIKeyEnvMissing), status: http.StatusBadRequest, code: "api_key_env_missing", message: "api_key_env is set but the environment variable is missing"},
 		{name: "artifact", err: wrap(usecase.ErrArtifactLoad, fmt.Errorf("read failed")), status: http.StatusBadRequest, code: "artifact_read_failed", message: "failed to read input artifact", avoidCause: "read failed"},
 		{name: "prompt render", err: wrap(usecase.ErrPromptRender, fmt.Errorf("render failed")), status: http.StatusBadRequest, code: "prompt_render_failed", message: "failed to render prompt", avoidCause: "render failed"},
 		{name: "llm", err: wrap(usecase.ErrLLMGenerate, fmt.Errorf("llm failed")), status: http.StatusBadGateway, code: "llm_failed", message: "model generation request failed", avoidCause: "llm failed"},
@@ -318,3 +847,54 @@ func TestHandlerValidationFailureStillSuccessAndRawOutputOptIn(t *testing.T) {
 func wrap(stage error, cause error) error {
 	return fmt.Errorf("%w: %w", stage, cause)
 }
+
+func newArtifactRootHandler(t *testing.T, root string) *Handler {
+	t.Helper()
+
+	return newArtifactRootHandlerWithLimit(t, root, 0)
+}
+
+func newArtifactRootHandlerWithLimit(t *testing.T, root string, maxArtifactBytes int64) *Handler {
+	t.Helper()
+
+	reader, err := artifact.NewRestrictedCompositeReaderWithLimit(root, maxArtifactBytes)
+	if err != nil {
+		t.Fatalf("expected restricted artifact reader: %v", err)
+	}
+	runner := usecase.NewRunner(
+		handlerPromptRepo{def: &domain.PromptDefinition{
+			ID:             "p",
+			Version:        "1",
+			DefaultProfile: "exec",
+			Templates:      []domain.PromptMessageTemplate{{Role: "user", Content: "hi"}},
+			OutputFormat:   domain.FormatText,
+			Validation:     domain.OutputContract{Format: domain.FormatText, ValidationMode: domain.ValidationNone},
+		}},
+		handlerProfileRepo{profile: &domain.ExecutionProfile{
+			ID:       "exec",
+			Endpoint: "http://example.invalid/v1",
+			Model:    "model",
+		}},
+		reader,
+		handlerRenderer{},
+		handlerLLMClient{},
+		nil,
+	)
+	return NewHandler(runner)
+}
+
+func assertHTTPErrorCode(t *testing.T, w *httptest.ResponseRecorder, status int, code string) {
+	t.Helper()
+
+	if w.Code != status {
+		t.Fatalf("expected %d, got %d body=%s", status, w.Code, w.Body.String())
+	}
+	var resp map[string]any
+	if err := json.Unmarshal(w.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("invalid JSON response: %v", err)
+	}
+	errBody := resp["error"].(map[string]any)
+	if errBody["code"] != code {
+		t.Fatalf("expected code %q, got %#v", code, errBody["code"])
+	}
+}

internal/artifact/reader.go

@@ -7,15 +7,20 @@ import (
 	"fmt"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+	"io"
 	"mime"
 	"os"
 	"path/filepath"
+	"strings"
 )
 
 var (
 	ErrUnsupportedRefType = errors.New("unsupported artifact reference type")
 	ErrMissingInlineBody  = errors.New("missing body for inline artifact")
 	ErrMissingFilePath    = errors.New("missing file path for file artifact")
+	ErrFileNotAllowed     = errors.New("file artifact references are not allowed")
+	ErrFileOutsideRoot    = errors.New("file artifact path is outside artifact root")
+	ErrFileTooLarge       = errors.New("file artifact exceeds size limit")
 )
 
 // Reader resolves artifact references into actual artifacts.
@@ -26,7 +31,7 @@ type Reader interface {
 // CompositeReader routes artifact resolution based on the reference type.
 type CompositeReader struct {
 	inlineReader *inlineReader
-	fileReader   *fileReader
+	fileReader   Reader
 }
 
 func NewCompositeReader() Reader {
@@ -36,6 +41,21 @@ func NewCompositeReader() Reader {
 	}
 }
 
+func NewRestrictedCompositeReader(root string) (Reader, error) {
+	return NewRestrictedCompositeReaderWithLimit(root, 0)
+}
+
+func NewRestrictedCompositeReaderWithLimit(root string, maxBytes int64) (Reader, error) {
+	fileReader, err := newRestrictedFileReader(root, maxBytes)
+	if err != nil {
+		return nil, err
+	}
+	return &CompositeReader{
+		inlineReader: &inlineReader{},
+		fileReader:   fileReader,
+	}, nil
+}
+
 func (c *CompositeReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
 	select {
 	case <-ctx.Done():
@@ -89,21 +109,133 @@ func (r *fileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.
 		return nil, ErrMissingFilePath
 	}
 
-	data, err := os.ReadFile(ref.URI)
-	if err != nil {
-		return nil, fmt.Errorf("failed to read file %s: %w", ref.URI, err)
+	return readFileArtifact(ref.URI)
+}
+
+type deniedFileReader struct{}
+
+func (r deniedFileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
 	}
 
-	contentType := mime.TypeByExtension(filepath.Ext(ref.URI))
+	if ref.URI == "" {
+		return nil, ErrMissingFilePath
+	}
+	return nil, ErrFileNotAllowed
+}
+
+type restrictedFileReader struct {
+	root     string
+	maxBytes int64
+}
+
+func newRestrictedFileReader(root string, maxBytes int64) (Reader, error) {
+	if maxBytes < 0 {
+		return nil, fmt.Errorf("artifact size limit must be greater than or equal to 0")
+	}
+	cleanRoot := strings.TrimSpace(root)
+	if cleanRoot == "" {
+		return deniedFileReader{}, nil
+	}
+	absRoot, err := filepath.Abs(filepath.Clean(cleanRoot))
+	if err != nil {
+		return nil, fmt.Errorf("resolve artifact root: %w", err)
+	}
+	return &restrictedFileReader{root: absRoot, maxBytes: maxBytes}, nil
+}
+
+func (r *restrictedFileReader) Read(ctx context.Context, ref domain.ArtifactRef) (*domain.Artifact, error) {
+	select {
+	case <-ctx.Done():
+		return nil, ctx.Err()
+	default:
+	}
+
+	if ref.URI == "" {
+		return nil, ErrMissingFilePath
+	}
+
+	path, err := r.resolveLexicalPath(ref.URI)
+	if err != nil {
+		return nil, err
+	}
+	return readFileArtifactWithLimit(path, r.maxBytes)
+}
+
+// resolveLexicalPath checks cleaned path containment without resolving symlinks.
+func (r *restrictedFileReader) resolveLexicalPath(rawPath string) (string, error) {
+	cleanPath := filepath.Clean(strings.TrimSpace(rawPath))
+	var candidate string
+	if filepath.IsAbs(cleanPath) {
+		candidate = cleanPath
+	} else {
+		candidate = filepath.Join(r.root, cleanPath)
+	}
+
+	absCandidate, err := filepath.Abs(candidate)
+	if err != nil {
+		return "", fmt.Errorf("resolve artifact path: %w", err)
+	}
+	absCandidate = filepath.Clean(absCandidate)
+
+	rel, err := filepath.Rel(r.root, absCandidate)
+	if err != nil {
+		return "", fmt.Errorf("compare artifact path to root: %w", err)
+	}
+	if rel == ".." || strings.HasPrefix(rel, ".."+string(filepath.Separator)) || filepath.IsAbs(rel) {
+		return "", ErrFileOutsideRoot
+	}
+	return absCandidate, nil
+}
+
+func readFileArtifact(path string) (*domain.Artifact, error) {
+	return readFileArtifactWithLimit(path, 0)
+}
+
+func readFileArtifactWithLimit(path string, maxBytes int64) (*domain.Artifact, error) {
+	if maxBytes < 0 {
+		return nil, fmt.Errorf("file size limit must be greater than or equal to 0")
+	}
+
+	file, err := os.Open(path)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+	}
+	defer file.Close()
+
+	info, err := file.Stat()
+	if err != nil {
+		return nil, fmt.Errorf("failed to stat file %s: %w", path, err)
+	}
+	if maxBytes > 0 && info.Size() > maxBytes {
+		return nil, ErrFileTooLarge
+	}
+
+	var reader io.Reader = file
+	if maxBytes > 0 {
+		reader = io.LimitReader(file, maxBytes+1)
+	}
+	data, err := io.ReadAll(reader)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read file %s: %w", path, err)
+	}
+	if maxBytes > 0 && int64(len(data)) > maxBytes {
+		return nil, ErrFileTooLarge
+	}
+
+	contentType := mime.TypeByExtension(filepath.Ext(path))
 	if contentType == "" {
 		contentType = defaults.ContentTypeTextPlain
 	}
 
 	return &domain.Artifact{
-		Name:        filepath.Base(ref.URI),
+		Name:        filepath.Base(path),
 		ContentType: contentType,
 		Body:        data,
-		URI:         ref.URI,
+		URI:         path,
 		Size:        int64(len(data)),
 		Hash:        fmt.Sprintf("%x", sha256.Sum256(data)),
 	}, nil

internal/artifact/reader_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"errors"
 	"os"
+	"path/filepath"
 	"testing"
 
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
@@ -46,8 +47,8 @@ func TestCompositeReader_Read(t *testing.T) {
 
 	t.Run("unsupported ref type", func(t *testing.T) {
 		ref := domain.ArtifactRef{
-			Type: domain.ArtifactRefS3,
-			URI:  "s3://bucket/key",
+			Type: domain.ArtifactRefType("unsupported"),
+			URI:  "unsupported://bucket/key",
 		}
 		_, err := reader.Read(ctx, ref)
 		if !errors.Is(err, ErrUnsupportedRefType) {
@@ -56,6 +57,157 @@ func TestCompositeReader_Read(t *testing.T) {
 	})
 }
 
+func TestRestrictedCompositeReader(t *testing.T) {
+	ctx := context.Background()
+	root := t.TempDir()
+	outside := t.TempDir()
+
+	if err := os.WriteFile(filepath.Join(root, "input.txt"), []byte("allowed"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.Mkdir(filepath.Join(root, "nested"), 0o755); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(outside, "secret.txt"), []byte("denied"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	reader, err := NewRestrictedCompositeReader(root)
+	if err != nil {
+		t.Fatalf("expected restricted reader construction, got %v", err)
+	}
+
+	t.Run("accepts relative contained path", func(t *testing.T) {
+		art, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "nested/../input.txt"})
+		if err != nil {
+			t.Fatalf("expected contained relative path to succeed, got %v", err)
+		}
+		if string(art.Body) != "allowed" {
+			t.Fatalf("unexpected artifact body: %q", string(art.Body))
+		}
+	})
+
+	t.Run("accepts absolute contained path", func(t *testing.T) {
+		art, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filepath.Join(root, "input.txt")})
+		if err != nil {
+			t.Fatalf("expected contained absolute path to succeed, got %v", err)
+		}
+		if art.Name != "input.txt" {
+			t.Fatalf("unexpected artifact name: %q", art.Name)
+		}
+	})
+
+	t.Run("rejects relative traversal outside root", func(t *testing.T) {
+		_, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filepath.Join("..", filepath.Base(outside), "secret.txt")})
+		if !errors.Is(err, ErrFileOutsideRoot) {
+			t.Fatalf("expected ErrFileOutsideRoot, got %v", err)
+		}
+	})
+
+	t.Run("rejects absolute path outside root", func(t *testing.T) {
+		_, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: filepath.Join(outside, "secret.txt")})
+		if !errors.Is(err, ErrFileOutsideRoot) {
+			t.Fatalf("expected ErrFileOutsideRoot, got %v", err)
+		}
+	})
+}
+
+func TestRestrictedCompositeReaderFollowsSymlinkInsideRoot(t *testing.T) {
+	ctx := context.Background()
+	root := t.TempDir()
+	outside := t.TempDir()
+
+	target := filepath.Join(outside, "linked.txt")
+	if err := os.WriteFile(target, []byte("linked outside root"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	link := filepath.Join(root, "linked.txt")
+	if err := os.Symlink(target, link); err != nil {
+		t.Skipf("symlink creation unavailable: %v", err)
+	}
+
+	reader, err := NewRestrictedCompositeReader(root)
+	if err != nil {
+		t.Fatalf("expected restricted reader construction, got %v", err)
+	}
+
+	art, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "linked.txt"})
+	if err != nil {
+		t.Fatalf("expected symlink inside root to be followed, got %v", err)
+	}
+	if string(art.Body) != "linked outside root" {
+		t.Fatalf("unexpected artifact body: %q", string(art.Body))
+	}
+}
+
+func TestRestrictedCompositeReaderWithoutRootDeniesFileRefs(t *testing.T) {
+	reader, err := NewRestrictedCompositeReader("")
+	if err != nil {
+		t.Fatalf("expected restricted reader construction, got %v", err)
+	}
+
+	art, err := reader.Read(context.Background(), domain.ArtifactRef{Type: domain.ArtifactRefInline, Body: "inline"})
+	if err != nil {
+		t.Fatalf("expected inline ref to work without artifact root, got %v", err)
+	}
+	if string(art.Body) != "inline" {
+		t.Fatalf("unexpected inline body: %q", string(art.Body))
+	}
+
+	_, err = reader.Read(context.Background(), domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "input.txt"})
+	if !errors.Is(err, ErrFileNotAllowed) {
+		t.Fatalf("expected ErrFileNotAllowed, got %v", err)
+	}
+}
+
+func TestRestrictedCompositeReaderFileSizeLimit(t *testing.T) {
+	ctx := context.Background()
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "exact.txt"), []byte("12345"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+	if err := os.WriteFile(filepath.Join(root, "large.txt"), []byte("123456"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	reader, err := NewRestrictedCompositeReaderWithLimit(root, 5)
+	if err != nil {
+		t.Fatalf("expected restricted reader construction, got %v", err)
+	}
+
+	art, err := reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "exact.txt"})
+	if err != nil {
+		t.Fatalf("expected file at limit to succeed, got %v", err)
+	}
+	if string(art.Body) != "12345" {
+		t.Fatalf("unexpected artifact body: %q", string(art.Body))
+	}
+
+	_, err = reader.Read(ctx, domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "large.txt"})
+	if !errors.Is(err, ErrFileTooLarge) {
+		t.Fatalf("expected ErrFileTooLarge, got %v", err)
+	}
+}
+
+func TestRestrictedCompositeReaderFileSizeLimitZeroDisablesLimit(t *testing.T) {
+	root := t.TempDir()
+	if err := os.WriteFile(filepath.Join(root, "large.txt"), []byte("123456"), 0o644); err != nil {
+		t.Fatal(err)
+	}
+
+	reader, err := NewRestrictedCompositeReaderWithLimit(root, 0)
+	if err != nil {
+		t.Fatalf("expected restricted reader construction, got %v", err)
+	}
+	art, err := reader.Read(context.Background(), domain.ArtifactRef{Type: domain.ArtifactRefFile, URI: "large.txt"})
+	if err != nil {
+		t.Fatalf("expected unlimited reader to succeed, got %v", err)
+	}
+	if string(art.Body) != "123456" {
+		t.Fatalf("unexpected artifact body: %q", string(art.Body))
+	}
+}
+
 func TestFileReader_Read(t *testing.T) {
 	content := []byte("test file content")
 	tmpFile, err := os.CreateTemp("", "artifact_test_*.txt")

internal/config/config.go

@@ -41,6 +41,10 @@ type Config struct {
 
 type ServerConfig struct {
 	Addr             string `yaml:"addr"`
+	ArtifactRoot     string `yaml:"artifact_root"`
+	MaxRequestBytes  *int64 `yaml:"max_request_bytes"`
+	MaxArtifactBytes *int64 `yaml:"max_artifact_bytes"`
+	MaxResponseBytes *int64 `yaml:"max_response_bytes"`
 }
 
 type DefaultsConfig struct {
@@ -53,6 +57,10 @@ type AppSettings struct {
 	ProfileDir          string
 	SchemaDir           string
 	ServerAddr          string
+	ArtifactRoot        string
+	MaxRequestBytes     int64
+	MaxArtifactBytes    int64
+	MaxResponseBytes    int64
 	DefaultRenderFormat renderformat.PreparedRunOutputFormat
 }
 
@@ -62,6 +70,10 @@ type CLIOverrides struct {
 	ProfileDir       string
 	SchemaDir        string
 	ServerAddr       string
+	ArtifactRoot     string
+	MaxRequestBytes  *int64
+	MaxArtifactBytes *int64
+	MaxResponseBytes *int64
 	RenderFormat     string
 }
 
@@ -70,6 +82,9 @@ func BuiltInDefaults() AppSettings {
 	return AppSettings{
 		SchemaDir:           defaults.SchemaDirDefault,
 		ServerAddr:          defaults.HTTPAddrDefault,
+		MaxRequestBytes:     defaults.HTTPMaxRequestBytesDefault,
+		MaxArtifactBytes:    defaults.HTTPMaxArtifactBytesDefault,
+		MaxResponseBytes:    defaults.HTTPMaxResponseBytesDefault,
 		DefaultRenderFormat: renderformat.DefaultPreparedRunOutputFormat,
 	}
 }
@@ -142,6 +157,27 @@ func ApplyCLIOverrides(base AppSettings, overrides CLIOverrides) (AppSettings, e
 	if v := strings.TrimSpace(overrides.ServerAddr); v != "" {
 		out.ServerAddr = v
 	}
+	if v := strings.TrimSpace(overrides.ArtifactRoot); v != "" {
+		out.ArtifactRoot = filepath.Clean(v)
+	}
+	if overrides.MaxRequestBytes != nil {
+		if *overrides.MaxRequestBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_request_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxRequestBytes = *overrides.MaxRequestBytes
+	}
+	if overrides.MaxArtifactBytes != nil {
+		if *overrides.MaxArtifactBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_artifact_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxArtifactBytes = *overrides.MaxArtifactBytes
+	}
+	if overrides.MaxResponseBytes != nil {
+		if *overrides.MaxResponseBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_response_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxResponseBytes = *overrides.MaxResponseBytes
+	}
 	if rawFormat := strings.TrimSpace(overrides.RenderFormat); rawFormat != "" {
 		parsed, err := renderformat.ParsePreparedRunOutputFormat(rawFormat)
 		if err != nil {
@@ -181,6 +217,27 @@ func applyConfig(base AppSettings, cfg Config) (AppSettings, error) {
 	if v := strings.TrimSpace(cfg.Server.Addr); v != "" {
 		out.ServerAddr = v
 	}
+	if v := strings.TrimSpace(cfg.Server.ArtifactRoot); v != "" {
+		out.ArtifactRoot = filepath.Clean(v)
+	}
+	if cfg.Server.MaxRequestBytes != nil {
+		if *cfg.Server.MaxRequestBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_request_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxRequestBytes = *cfg.Server.MaxRequestBytes
+	}
+	if cfg.Server.MaxArtifactBytes != nil {
+		if *cfg.Server.MaxArtifactBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_artifact_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxArtifactBytes = *cfg.Server.MaxArtifactBytes
+	}
+	if cfg.Server.MaxResponseBytes != nil {
+		if *cfg.Server.MaxResponseBytes < 0 {
+			return AppSettings{}, fmt.Errorf("%w: server.max_response_bytes must be greater than or equal to 0", ErrInvalidConfig)
+		}
+		out.MaxResponseBytes = *cfg.Server.MaxResponseBytes
+	}
 	if rawFormat := strings.TrimSpace(cfg.Defaults.RenderFormat); rawFormat != "" {
 		parsed, err := renderformat.ParsePreparedRunOutputFormat(rawFormat)
 		if err != nil {

internal/config/config_test.go

@@ -6,6 +6,7 @@ import (
 	"path/filepath"
 	"testing"
 
+	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
 	renderformat "gitea.maximumdirect.net/eric/scriptorium/internal/format"
 )
 
@@ -24,6 +25,20 @@ func TestLoadConfigMissingImplicitPathUsesBuiltInDefaults(t *testing.T) {
 	}
 }
 
+func TestBuiltInDefaultsIncludeHTTPSizeLimits(t *testing.T) {
+	got := BuiltInDefaults()
+
+	if got.MaxRequestBytes != defaults.HTTPMaxRequestBytesDefault {
+		t.Fatalf("unexpected max request bytes: %d", got.MaxRequestBytes)
+	}
+	if got.MaxArtifactBytes != defaults.HTTPMaxArtifactBytesDefault {
+		t.Fatalf("unexpected max artifact bytes: %d", got.MaxArtifactBytes)
+	}
+	if got.MaxResponseBytes != defaults.HTTPMaxResponseBytesDefault {
+		t.Fatalf("unexpected max response bytes: %d", got.MaxResponseBytes)
+	}
+}
+
 func TestLoadConfigMissingExplicitPathReturnsError(t *testing.T) {
 	tmp := t.TempDir()
 	missing := filepath.Join(tmp, "missing.yml")
@@ -92,6 +107,10 @@ profile_dir: ./profiles
 schema_dir: ./schemas
 server:
   addr: 127.0.0.1:9090
+  artifact_root: ./artifacts
+  max_request_bytes: 1024
+  max_artifact_bytes: 2048
+  max_response_bytes: 4096
 defaults:
   render_format: json
 `)
@@ -113,11 +132,61 @@ defaults:
 	if got.ServerAddr != "127.0.0.1:9090" {
 		t.Fatalf("unexpected server.addr: %q", got.ServerAddr)
 	}
+	if got.ArtifactRoot != filepath.Clean("./artifacts") {
+		t.Fatalf("unexpected server.artifact_root: %q", got.ArtifactRoot)
+	}
+	if got.MaxRequestBytes != 1024 {
+		t.Fatalf("unexpected server.max_request_bytes: %d", got.MaxRequestBytes)
+	}
+	if got.MaxArtifactBytes != 2048 {
+		t.Fatalf("unexpected server.max_artifact_bytes: %d", got.MaxArtifactBytes)
+	}
+	if got.MaxResponseBytes != 4096 {
+		t.Fatalf("unexpected server.max_response_bytes: %d", got.MaxResponseBytes)
+	}
 	if got.DefaultRenderFormat != renderformat.PreparedRunFormatJSON {
 		t.Fatalf("unexpected defaults.render_format: %q", got.DefaultRenderFormat)
 	}
 }
 
+func TestLoadConfigAcceptsZeroHTTPSizeLimits(t *testing.T) {
+	path := writeConfigFile(t, "config.yml", `
+server:
+  max_request_bytes: 0
+  max_artifact_bytes: 0
+  max_response_bytes: 0
+`)
+
+	got, err := LoadConfig(path, true)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if got.MaxRequestBytes != 0 || got.MaxArtifactBytes != 0 || got.MaxResponseBytes != 0 {
+		t.Fatalf("expected zero limits to be preserved, got request=%d artifact=%d response=%d", got.MaxRequestBytes, got.MaxArtifactBytes, got.MaxResponseBytes)
+	}
+}
+
+func TestLoadConfigRejectsNegativeHTTPSizeLimits(t *testing.T) {
+	tests := []struct {
+		name string
+		body string
+	}{
+		{name: "request", body: "server:\n  max_request_bytes: -1\n"},
+		{name: "artifact", body: "server:\n  max_artifact_bytes: -1\n"},
+		{name: "response", body: "server:\n  max_response_bytes: -1\n"},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			path := writeConfigFile(t, "config.yml", tc.body)
+			_, err := LoadConfig(path, true)
+			if !errors.Is(err, ErrInvalidConfig) {
+				t.Fatalf("expected ErrInvalidConfig, got %v", err)
+			}
+		})
+	}
+}
+
 func TestLoadConfigEmptyFileResolvesToBuiltInDefaults(t *testing.T) {
 	path := writeConfigFile(t, "config.yml", "")
 
@@ -185,14 +254,25 @@ func TestApplyCLIOverridesAppliesPrecedence(t *testing.T) {
 		ProfileDir:          "/from/config/profiles",
 		SchemaDir:           "/from/config/schemas",
 		ServerAddr:          ":1234",
+		ArtifactRoot:        "/from/config/artifacts",
+		MaxRequestBytes:     111,
+		MaxArtifactBytes:    222,
+		MaxResponseBytes:    333,
 		DefaultRenderFormat: renderformat.PreparedRunFormatJSON,
 	}
+	maxRequestBytes := int64(0)
+	maxArtifactBytes := int64(444)
+	maxResponseBytes := int64(555)
 
 	got, err := ApplyCLIOverrides(base, CLIOverrides{
 		PromptDir:        "./prompts-cli",
 		ProfileDir:       "./profiles-cli",
 		SchemaDir:        "./schemas-cli",
 		ServerAddr:       ":8081",
+		ArtifactRoot:     "./artifacts-cli",
+		MaxRequestBytes:  &maxRequestBytes,
+		MaxArtifactBytes: &maxArtifactBytes,
+		MaxResponseBytes: &maxResponseBytes,
 		RenderFormat:     "text",
 	})
 	if err != nil {
@@ -211,11 +291,45 @@ func TestApplyCLIOverridesAppliesPrecedence(t *testing.T) {
 	if got.ServerAddr != ":8081" {
 		t.Fatalf("unexpected server addr: %q", got.ServerAddr)
 	}
+	if got.ArtifactRoot != filepath.Clean("./artifacts-cli") {
+		t.Fatalf("unexpected artifact root: %q", got.ArtifactRoot)
+	}
+	if got.MaxRequestBytes != 0 {
+		t.Fatalf("unexpected max request bytes: %d", got.MaxRequestBytes)
+	}
+	if got.MaxArtifactBytes != 444 {
+		t.Fatalf("unexpected max artifact bytes: %d", got.MaxArtifactBytes)
+	}
+	if got.MaxResponseBytes != 555 {
+		t.Fatalf("unexpected max response bytes: %d", got.MaxResponseBytes)
+	}
 	if got.DefaultRenderFormat != renderformat.PreparedRunFormatText {
 		t.Fatalf("unexpected render format: %q", got.DefaultRenderFormat)
 	}
 }
 
+func TestApplyCLIOverridesRejectsNegativeHTTPSizeLimits(t *testing.T) {
+	negative := int64(-1)
+
+	tests := []struct {
+		name      string
+		overrides CLIOverrides
+	}{
+		{name: "request", overrides: CLIOverrides{MaxRequestBytes: &negative}},
+		{name: "artifact", overrides: CLIOverrides{MaxArtifactBytes: &negative}},
+		{name: "response", overrides: CLIOverrides{MaxResponseBytes: &negative}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			_, err := ApplyCLIOverrides(BuiltInDefaults(), tc.overrides)
+			if !errors.Is(err, ErrInvalidConfig) {
+				t.Fatalf("expected ErrInvalidConfig, got %v", err)
+			}
+		})
+	}
+}
+
 func TestApplyCLIOverridesInvalidRenderFormatReturnsError(t *testing.T) {
 	_, err := ApplyCLIOverrides(BuiltInDefaults(), CLIOverrides{RenderFormat: "yaml"})
 	if err == nil {

internal/defaults/defaults.go

@@ -14,6 +14,9 @@ const (
 	ContentTypeTextMarkdown     = "text/markdown"
 	ContentTypeApplicationJSON  = "application/json"
 	OpenAIChatCompletionsPath   = "/chat/completions"
+	HTTPMaxRequestBytesDefault  = 16 * 1024 * 1024
+	HTTPMaxArtifactBytesDefault = 16 * 1024 * 1024
+	HTTPMaxResponseBytesDefault = 16 * 1024 * 1024
 
 	ExecutionDefaultTemperature    = 0.0
 	ExecutionDefaultMaxTokens      = 0

internal/domain/domain.go

@@ -10,7 +10,6 @@ type ArtifactRefType string
 const (
 	ArtifactRefInline ArtifactRefType = "inline"
 	ArtifactRefFile   ArtifactRefType = "file"
-	ArtifactRefS3     ArtifactRefType = "s3"
 )
 
 // OutputFormat defines the desired format of the generated artifact.
@@ -41,14 +40,33 @@ const (
 	ValidationSkipped ValidationStatus = "skipped"
 )
 
+// CacheControlType defines provider cache behavior for prompt content.
+type CacheControlType string
+
+const (
+	CacheControlEphemeral CacheControlType = "ephemeral"
+)
+
+const (
+	// SessionIDMaxLength is OpenRouter's documented maximum session_id length.
+	SessionIDMaxLength = 256
+)
+
+// CacheControl describes provider cache metadata attached to prompt content.
+type CacheControl struct {
+	Type CacheControlType `yaml:"type" json:"type"`
+	TTL  string           `yaml:"ttl,omitempty" json:"ttl,omitempty"`
+}
+
 // RunRequest represents a request to generate a single artifact.
 type RunRequest struct {
 	PromptID      string
 	PromptVersion string
 	ProfileID     string
+	APIKey        string `json:"-" yaml:"-"`
 	Inputs        map[string]ArtifactRef
 	Vars          map[string]string
-	Execution     *ExecutionTarget
+	Execution     *ExecutionTargetOverride
 	Validation    *OutputContract
 	Metadata      map[string]string
 }
@@ -72,7 +90,6 @@ type RunResult struct {
 	StartTime            time.Time
 	EndTime              time.Time
 	Duration             time.Duration
-	Error                error
 }
 
 // PreparedRun contains pre-LLM execution state from the prepare/render phase.
@@ -83,9 +100,11 @@ type PreparedRun struct {
 	PromptHash           string                  `json:"prompt_hash,omitempty"`
 	SelectedProfileID    string                  `json:"selected_profile_id"`
 	EffectiveModelParams ExecutionTarget         `json:"effective_model_params"`
+	TargetPresence       ExecutionTargetPresence `json:"-"`
 	OutputContract       OutputContract          `json:"output_contract"`
 	StructuredOutput     *StructuredOutputSpec   `json:"structured_output,omitempty"`
 	InputHashes          map[string]string       `json:"input_hashes,omitempty"`
+	SessionID            string                  `json:"session_id,omitempty"`
 	RenderedPromptHash   string                  `json:"rendered_prompt_hash"`
 	Messages             []RenderedMessage       `json:"messages"`
 	StartTime            time.Time               `json:"start_time,omitempty"`
@@ -116,6 +135,7 @@ type PromptDefinition struct {
 	Version        string                  `yaml:"version"`
 	DefaultProfile string                  `yaml:"default_profile"`
 	Description    string                  `yaml:"description"`
+	SessionID      string                  `yaml:"session_id" json:"session_id,omitempty"`
 	Inputs         []PromptInput           `yaml:"inputs"`
 	Templates      []PromptMessageTemplate `yaml:"templates"`
 	OutputFormat   OutputFormat            `yaml:"output_format"`
@@ -135,6 +155,7 @@ type PromptMessageTemplate struct {
 	Role         string        `yaml:"role"`
 	Content      string        `yaml:"content"`
 	ContentFile  string        `yaml:"content_file"`
+	CacheControl *CacheControl `yaml:"cache_control,omitempty" json:"cache_control,omitempty"`
 }
 
 // ExecutionProfile describes how and where to execute a model.
@@ -146,9 +167,34 @@ type ExecutionProfile struct {
 	MaxTokens       int            `yaml:"max_tokens"`
 	TopP            float64        `yaml:"top_p"`
 	TimeoutSeconds  int            `yaml:"timeout_seconds"`
+	ServiceTier     string         `yaml:"service_tier"`
 	ReasoningEffort string         `yaml:"reasoning_effort"`
 	APIKeyEnv       string         `yaml:"api_key_env"`
-	ExtraParams     map[string]string `yaml:"extra_params"`
+	APIKeyRequired  bool           `yaml:"-" json:"-"`
+	ExtraParams     map[string]any `yaml:"extra_params"`
+}
+
+// ExecutionTargetOverride represents per-request runtime setting overrides.
+type ExecutionTargetOverride struct {
+	Endpoint        string         `json:"endpoint,omitempty"`
+	Model           string         `json:"model,omitempty"`
+	Temperature     *float64       `json:"temperature,omitempty"`
+	MaxTokens       *int           `json:"max_tokens,omitempty"`
+	TopP            *float64       `json:"top_p,omitempty"`
+	TimeoutSeconds  *int           `json:"timeout_seconds,omitempty"`
+	ServiceTier     string         `json:"service_tier,omitempty"`
+	ReasoningEffort string         `json:"reasoning_effort,omitempty"`
+	APIKeyEnv       string         `json:"api_key_env,omitempty"`
+	ExtraParams     map[string]any `json:"extra_params,omitempty"`
+}
+
+// ExecutionTargetPresence tracks which effective runtime fields came from an
+// explicit request override even when the resolved value is a zero value.
+type ExecutionTargetPresence struct {
+	Temperature    bool
+	MaxTokens      bool
+	TopP           bool
+	TimeoutSeconds bool
 }
 
 // ExecutionTarget represents effective model runtime settings for a run.
@@ -159,9 +205,12 @@ type ExecutionTarget struct {
 	MaxTokens       int            `yaml:"max_tokens" json:"max_tokens"`
 	TopP            float64        `yaml:"top_p" json:"top_p"`
 	TimeoutSeconds  int            `yaml:"timeout_seconds" json:"timeout_seconds"`
+	ServiceTier     string         `yaml:"service_tier" json:"service_tier"`
 	ReasoningEffort string         `yaml:"reasoning_effort" json:"reasoning_effort"`
 	APIKeyEnv       string         `yaml:"api_key_env" json:"api_key_env"`
-	ExtraParams     map[string]string `yaml:"extra_params" json:"extra_params"`
+	APIKey          string         `yaml:"-" json:"-"`
+	APIKeyRequired  bool           `yaml:"-" json:"-"`
+	ExtraParams     map[string]any `yaml:"extra_params" json:"extra_params"`
 }
 
 // OutputContract defines the requirements for the output artifact.
@@ -174,6 +223,7 @@ type OutputContract struct {
 
 // RenderedPrompt represents the prompt after template application.
 type RenderedPrompt struct {
+	SessionID string            `json:"session_id,omitempty"`
 	Messages  []RenderedMessage `json:"messages"`
 }
 
@@ -181,12 +231,14 @@ type RenderedPrompt struct {
 type RenderedMessage struct {
 	Role         string        `json:"role"`
 	Content      string        `json:"content"`
+	CacheControl *CacheControl `json:"cache_control,omitempty"`
 }
 
 // GenerateRequest is the internal request passed to the LLM client.
 type GenerateRequest struct {
 	Prompt           RenderedPrompt
 	Target           ExecutionTarget
+	TargetPresence   ExecutionTargetPresence
 	StructuredOutput *StructuredOutputSpec
 }
 
@@ -221,6 +273,8 @@ type TokenUsage struct {
 	PromptTokens     int
 	CompletionTokens int
 	TotalTokens      int
+	CachedTokens     int
+	CacheWriteTokens int
 }
 
 // ValidationResult represents the outcome of an output validation.
@@ -232,23 +286,3 @@ type ValidationResult struct {
 	RepairAttempts int
 	IsValid        bool
 }
-
-// RunMetadata contains auditing information for a run.
-type RunMetadata struct {
-	RunID              string
-	PromptID           string
-	PromptVersion      string
-	PromptHash         string
-	RenderedPromptHash string
-	SelectedProfileID  string
-	InputHashes        map[string]string
-	ModelEndpoint      string
-	ModelName          string
-	Params             ExecutionTarget
-	Timestamp          time.Time
-	Duration           time.Duration
-	Usage              TokenUsage
-	ValidationMode     ValidationMode
-	ValidationStatus   ValidationStatus
-	RepairAttempts     int
-}

internal/domain/prepared_run_test.go

@@ -20,6 +20,7 @@ func TestPreparedRunJSONDoesNotIncludeSecretValues(t *testing.T) {
 			Endpoint:  "http://llm/v1",
 			Model:     "gpt-test",
 			APIKeyEnv: envName,
+			APIKey:    secret,
 		},
 		InputHashes:        map[string]string{"transcript": "hash-1"},
 		RenderedPromptHash: "rendered-hash",
@@ -53,3 +54,88 @@ func TestPreparedRunJSONDoesNotIncludeSecretValues(t *testing.T) {
 		}
 	}
 }
+
+func TestPreparedRunJSONIncludesMessageCacheControlOnlyWhenPresent(t *testing.T) {
+	prepared := PreparedRun{
+		PromptID:          "prompt.id",
+		SelectedProfileID: "local-fast",
+		EffectiveModelParams: ExecutionTarget{
+			Endpoint: "http://llm/v1",
+			Model:    "gpt-test",
+		},
+		RenderedPromptHash: "rendered-hash",
+		Messages: []RenderedMessage{
+			{
+				Role:    "system",
+				Content: "You are helpful.",
+				CacheControl: &CacheControl{
+					Type: CacheControlEphemeral,
+					TTL:  "1h",
+				},
+			},
+			{Role: "user", Content: "Summarize this."},
+		},
+	}
+
+	b, err := json.Marshal(prepared)
+	if err != nil {
+		t.Fatalf("marshal failed: %v", err)
+	}
+
+	var decoded struct {
+		Messages []map[string]any `json:"messages"`
+	}
+	if err := json.Unmarshal(b, &decoded); err != nil {
+		t.Fatalf("unmarshal failed: %v", err)
+	}
+	if len(decoded.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(decoded.Messages))
+	}
+
+	cacheControl, ok := decoded.Messages[0]["cache_control"].(map[string]any)
+	if !ok {
+		t.Fatalf("expected cache_control on first message, got %#v", decoded.Messages[0])
+	}
+	if cacheControl["type"] != string(CacheControlEphemeral) || cacheControl["ttl"] != "1h" {
+		t.Fatalf("unexpected cache_control payload: %#v", cacheControl)
+	}
+	if _, ok := decoded.Messages[1]["cache_control"]; ok {
+		t.Fatalf("expected second message to omit cache_control, got %#v", decoded.Messages[1])
+	}
+}
+
+func TestPreparedRunJSONIncludesSessionIDOnlyWhenPresent(t *testing.T) {
+	prepared := PreparedRun{
+		PromptID:          "prompt.id",
+		SelectedProfileID: "local-fast",
+		EffectiveModelParams: ExecutionTarget{
+			Endpoint: "http://llm/v1",
+			Model:    "gpt-test",
+		},
+		SessionID:          "session-123",
+		RenderedPromptHash: "rendered-hash",
+		Messages:           []RenderedMessage{{Role: "user", Content: "Summarize this."}},
+	}
+
+	b, err := json.Marshal(prepared)
+	if err != nil {
+		t.Fatalf("marshal failed: %v", err)
+	}
+
+	var decoded map[string]any
+	if err := json.Unmarshal(b, &decoded); err != nil {
+		t.Fatalf("unmarshal failed: %v", err)
+	}
+	if decoded["session_id"] != "session-123" {
+		t.Fatalf("expected session_id in prepared run JSON, got %#v", decoded["session_id"])
+	}
+
+	prepared.SessionID = ""
+	b, err = json.Marshal(prepared)
+	if err != nil {
+		t.Fatalf("marshal failed: %v", err)
+	}
+	if strings.Contains(string(b), "session_id") {
+		t.Fatalf("expected empty session_id to be omitted, got %s", b)
+	}
+}

internal/filecatalog/catalog.go

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

internal/filecatalog/catalog_test.go

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

internal/format/prepared_run.go

@@ -96,6 +96,9 @@ func (textPreparedRunFormatter) Format(prepared *domain.PreparedRun) ([]byte, er
 	if prepared.PromptHash != "" {
 		fmt.Fprintf(&b, "prompt_hash: %s\n", prepared.PromptHash)
 	}
+	if prepared.SessionID != "" {
+		fmt.Fprintf(&b, "session_id: %s\n", prepared.SessionID)
+	}
 	fmt.Fprintf(&b, "rendered_prompt_hash: %s\n", prepared.RenderedPromptHash)
 
 	target := prepared.EffectiveModelParams
@@ -106,6 +109,9 @@ func (textPreparedRunFormatter) Format(prepared *domain.PreparedRun) ([]byte, er
 	fmt.Fprintf(&b, "  max_tokens: %d\n", target.MaxTokens)
 	fmt.Fprintf(&b, "  top_p: %g\n", target.TopP)
 	fmt.Fprintf(&b, "  timeout_seconds: %d\n", target.TimeoutSeconds)
+	if target.ServiceTier != "" {
+		fmt.Fprintf(&b, "  service_tier: %s\n", target.ServiceTier)
+	}
 	if target.ReasoningEffort != "" {
 		fmt.Fprintf(&b, "  reasoning_effort: %s\n", target.ReasoningEffort)
 	}
@@ -120,7 +126,11 @@ func (textPreparedRunFormatter) Format(prepared *domain.PreparedRun) ([]byte, er
 		}
 		sort.Strings(keys)
 		for _, k := range keys {
-			fmt.Fprintf(&b, "    %s: %s\n", k, target.ExtraParams[k])
+			renderedValue, err := formatExtraParamTextValue(target.ExtraParams[k])
+			if err != nil {
+				return nil, fmt.Errorf("failed to format extra_params.%s: %w", k, err)
+			}
+			fmt.Fprintf(&b, "    %s: %s\n", k, renderedValue)
 		}
 	}
 
@@ -148,6 +158,13 @@ func (textPreparedRunFormatter) Format(prepared *domain.PreparedRun) ([]byte, er
 		messages := byRole[role]
 		for i, msg := range messages {
 			fmt.Fprintf(&b, "    - message: %d\n", i+1)
+			if msg.CacheControl != nil {
+				fmt.Fprintf(&b, "      cache_control: %s", msg.CacheControl.Type)
+				if msg.CacheControl.TTL != "" {
+					fmt.Fprintf(&b, " ttl=%s", msg.CacheControl.TTL)
+				}
+				fmt.Fprintln(&b)
+			}
 			fmt.Fprintln(&b, "      content: |")
 			content := msg.Content
 			if content == "" {
@@ -162,3 +179,15 @@ func (textPreparedRunFormatter) Format(prepared *domain.PreparedRun) ([]byte, er
 
 	return b.Bytes(), nil
 }
+
+func formatExtraParamTextValue(value any) (string, error) {
+	if s, ok := value.(string); ok {
+		return s, nil
+	}
+
+	b, err := json.Marshal(value)
+	if err != nil {
+		return "", err
+	}
+	return string(b), nil
+}

internal/format/prepared_run_test.go

@@ -28,6 +28,7 @@ func TestTextFormatterIncludesPreparedRunDetails(t *testing.T) {
 		"max_tokens: 256",
 		"top_p: 0.8",
 		"timeout_seconds: 45",
+		"service_tier: priority",
 		"reasoning_effort: medium",
 		"api_key_env: SCRIPTORIUM_API_KEY",
 		"prompt_hash: prompt-hash",
@@ -48,6 +49,36 @@ func TestTextFormatterIncludesPreparedRunDetails(t *testing.T) {
 	}
 }
 
+func TestTextFormatterRendersExtraParamsDeterministically(t *testing.T) {
+	prepared := samplePreparedRun()
+	prepared.EffectiveModelParams.ExtraParams = map[string]any{
+		"z_string": "enabled",
+		"b_number": 42,
+		"a_object": map[string]any{
+			"nested": "value",
+			"count":  2,
+		},
+		"c_array": []any{"first", 3, false},
+	}
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatText)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	s := string(out)
+
+	want := strings.Join([]string{
+		"  extra_params:",
+		"    a_object: {\"count\":2,\"nested\":\"value\"}",
+		"    b_number: 42",
+		"    c_array: [\"first\",3,false]",
+		"    z_string: enabled",
+	}, "\n")
+	if !strings.Contains(s, want) {
+		t.Fatalf("expected deterministic extra_params block %q, got:\n%s", want, s)
+	}
+}
+
 func TestTextFormatterDoesNotIncludeResolvedAPIKeyValue(t *testing.T) {
 	const secret = "super-secret-api-key"
 	t.Setenv("SCRIPTORIUM_API_KEY", secret)
@@ -61,8 +92,94 @@ func TestTextFormatterDoesNotIncludeResolvedAPIKeyValue(t *testing.T) {
 	}
 }
 
+func TestTextFormatterDoesNotIncludeDirectAPIKeyValue(t *testing.T) {
+	const directKey = "direct-format-key"
+	prepared := samplePreparedRun()
+	prepared.EffectiveModelParams.APIKey = directKey
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatText)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if strings.Contains(string(out), directKey) {
+		t.Fatalf("text output should not include direct api key value: %s", out)
+	}
+}
+
+func TestTextFormatterIncludesMessageCacheControlBeforeContent(t *testing.T) {
+	prepared := samplePreparedRun()
+	prepared.Messages = []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "System guidance.",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+				TTL:  "1h",
+			},
+		},
+		{Role: "user", Content: "Summarize the transcript."},
+	}
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatText)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	s := string(out)
+	if !strings.Contains(s, "  system:\n    - message: 1\n      cache_control: ephemeral ttl=1h\n      content: |") {
+		t.Fatalf("expected system message cache control before content, got:\n%s", s)
+	}
+	if strings.Count(s, "cache_control:") != 1 {
+		t.Fatalf("expected exactly one cache_control line, got:\n%s", s)
+	}
+}
+
+func TestTextFormatterIncludesSessionIDWhenPresent(t *testing.T) {
+	prepared := samplePreparedRun()
+	prepared.SessionID = "session-123"
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatText)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if !strings.Contains(string(out), "session_id: session-123\n") {
+		t.Fatalf("expected session_id in text output, got:\n%s", out)
+	}
+}
+
+func TestTextFormatterOmitsEmptyCacheControlTTL(t *testing.T) {
+	prepared := samplePreparedRun()
+	prepared.Messages = []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "System guidance.",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+			},
+		},
+	}
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatText)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	s := string(out)
+	if !strings.Contains(s, "      cache_control: ephemeral\n") {
+		t.Fatalf("expected cache_control line without ttl, got:\n%s", s)
+	}
+	if strings.Contains(s, "ttl=") {
+		t.Fatalf("expected empty ttl to be omitted, got:\n%s", s)
+	}
+}
+
 func TestJSONFormatterEmitsValidJSONAndIncludesPreparedRunFields(t *testing.T) {
 	prepared := samplePreparedRun()
+	prepared.SessionID = "session-123"
+	prepared.EffectiveModelParams.ExtraParams = map[string]any{
+		"number": 42,
+		"nested": map[string]any{
+			"enabled": true,
+		},
+	}
 
 	out, err := FormatPreparedRun(prepared, PreparedRunFormatJSON)
 	if err != nil {
@@ -86,9 +203,24 @@ func TestJSONFormatterEmitsValidJSONAndIncludesPreparedRunFields(t *testing.T) {
 	if decoded["rendered_prompt_hash"] != "rendered-hash" {
 		t.Fatalf("expected rendered_prompt_hash in json output, got %#v", decoded["rendered_prompt_hash"])
 	}
-	if _, ok := decoded["effective_model_params"]; !ok {
+	if decoded["session_id"] != "session-123" {
+		t.Fatalf("expected session_id in json output, got %#v", decoded["session_id"])
+	}
+	modelParams, ok := decoded["effective_model_params"].(map[string]any)
+	if !ok {
 		t.Fatalf("expected effective_model_params in json output, got %#v", decoded)
 	}
+	extraParams, ok := modelParams["extra_params"].(map[string]any)
+	if !ok {
+		t.Fatalf("expected extra_params in json output, got %#v", modelParams["extra_params"])
+	}
+	if extraParams["number"] != float64(42) {
+		t.Fatalf("unexpected numeric extra param in json output: %#v", extraParams["number"])
+	}
+	nested, ok := extraParams["nested"].(map[string]any)
+	if !ok || nested["enabled"] != true {
+		t.Fatalf("unexpected nested extra param in json output: %#v", extraParams["nested"])
+	}
 	if _, ok := decoded["input_hashes"]; !ok {
 		t.Fatalf("expected input_hashes in json output, got %#v", decoded)
 	}
@@ -97,6 +229,47 @@ func TestJSONFormatterEmitsValidJSONAndIncludesPreparedRunFields(t *testing.T) {
 	}
 }
 
+func TestJSONFormatterIncludesMessageCacheControlOnlyWhenPresent(t *testing.T) {
+	prepared := samplePreparedRun()
+	prepared.Messages = []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "System guidance.",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+				TTL:  "1h",
+			},
+		},
+		{Role: "user", Content: "Summarize the transcript."},
+	}
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatJSON)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+
+	var decoded struct {
+		Messages []map[string]any `json:"messages"`
+	}
+	if err := json.Unmarshal(out, &decoded); err != nil {
+		t.Fatalf("expected valid json output, got %v", err)
+	}
+	if len(decoded.Messages) != 2 {
+		t.Fatalf("expected 2 messages, got %d", len(decoded.Messages))
+	}
+
+	cacheControl, ok := decoded.Messages[0]["cache_control"].(map[string]any)
+	if !ok {
+		t.Fatalf("expected first message cache_control, got %#v", decoded.Messages[0])
+	}
+	if cacheControl["type"] != string(domain.CacheControlEphemeral) || cacheControl["ttl"] != "1h" {
+		t.Fatalf("unexpected cache_control payload: %#v", cacheControl)
+	}
+	if _, ok := decoded.Messages[1]["cache_control"]; ok {
+		t.Fatalf("expected second message to omit cache_control, got %#v", decoded.Messages[1])
+	}
+}
+
 func TestJSONFormatterDoesNotIncludeResolvedAPIKeyValue(t *testing.T) {
 	const secret = "super-secret-api-key"
 	t.Setenv("SCRIPTORIUM_API_KEY", secret)
@@ -110,6 +283,20 @@ func TestJSONFormatterDoesNotIncludeResolvedAPIKeyValue(t *testing.T) {
 	}
 }
 
+func TestJSONFormatterDoesNotIncludeDirectAPIKeyValue(t *testing.T) {
+	const directKey = "direct-format-key"
+	prepared := samplePreparedRun()
+	prepared.EffectiveModelParams.APIKey = directKey
+
+	out, err := FormatPreparedRun(prepared, PreparedRunFormatJSON)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if strings.Contains(string(out), directKey) {
+		t.Fatalf("json output should not include direct api key value: %s", out)
+	}
+}
+
 func TestParsePreparedRunOutputFormatRecognizesSupportedNames(t *testing.T) {
 	tests := []struct {
 		name  string
@@ -168,6 +355,7 @@ func samplePreparedRun() *domain.PreparedRun {
 			MaxTokens:       256,
 			TopP:            0.8,
 			TimeoutSeconds:  45,
+			ServiceTier:     "priority",
 			ReasoningEffort: "medium",
 			APIKeyEnv:       "SCRIPTORIUM_API_KEY",
 		},

internal/llm/openai_compatible_client.go

@@ -12,6 +12,7 @@ import (
 	"os"
 	"strings"
 	"time"
+	"unicode/utf8"
 
 	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
@@ -54,10 +55,11 @@ func NewOpenAICompatibleClient(cfg OpenAICompatibleConfig) (*OpenAICompatibleCli
 
 	var client *http.Client
 	if cfg.HTTPClient != nil {
-		client = cfg.HTTPClient
-		if client.Timeout == 0 {
-			client.Timeout = timeout
+		cloned := *cfg.HTTPClient
+		if cloned.Timeout == 0 {
+			cloned.Timeout = timeout
 		}
+		client = &cloned
 	} else {
 		client = &http.Client{Timeout: timeout}
 	}
@@ -75,14 +77,6 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 		return nil, fmt.Errorf("%w: timeout_seconds must be greater than or equal to 0", ErrInvalidRequest)
 	}
 
-	model := strings.TrimSpace(req.Target.Model)
-	if model == "" {
-		model = strings.TrimSpace(c.defaultModel)
-	}
-	if model == "" {
-		return nil, fmt.Errorf("%w: model is required", ErrInvalidRequest)
-	}
-
 	endpoint := strings.TrimSpace(req.Target.Endpoint)
 	if endpoint == "" {
 		endpoint = c.baseURL
@@ -92,36 +86,17 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 	}
 	endpoint = strings.TrimRight(endpoint, "/") + defaults.OpenAIChatCompletionsPath
 
-	wireReq := openAIChatRequest{
-		Model: model,
-	}
-
-	wireReq.Messages = make([]openAIChatMessage, 0, len(req.Prompt.Messages))
-	for _, msg := range req.Prompt.Messages {
-		wireReq.Messages = append(wireReq.Messages, openAIChatMessage{
-			Role:    msg.Role,
-			Content: msg.Content,
-		})
-	}
-
-	if req.Target.Temperature != 0 {
-		wireReq.Temperature = &req.Target.Temperature
-	}
-	if req.Target.MaxTokens != 0 {
-		wireReq.MaxTokens = &req.Target.MaxTokens
-	}
-	if req.Target.TopP != 0 {
-		wireReq.TopP = &req.Target.TopP
-	}
-	if req.StructuredOutput != nil {
-		responseFormat, err := toOpenAIResponseFormat(req.StructuredOutput)
+	wireReq, err := openAIChatRequestFromGenerateRequest(req, c.defaultModel)
 	if err != nil {
 		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
-		wireReq.ResponseFormat = responseFormat
+
+	wirePayload, err := openAIChatRequestPayload(wireReq)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
 
-	payload, err := json.Marshal(wireReq)
+	payload, err := json.Marshal(wirePayload)
 	if err != nil {
 		return nil, fmt.Errorf("%w: failed to encode request: %v", ErrRequestFailed, err)
 	}
@@ -131,7 +106,9 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 		return nil, fmt.Errorf("%w: failed to create request: %v", ErrRequestFailed, err)
 	}
 	httpReq.Header.Set("Content-Type", "application/json")
-	if envName := strings.TrimSpace(req.Target.APIKeyEnv); envName != "" {
+	if apiKey := strings.TrimSpace(req.Target.APIKey); apiKey != "" {
+		httpReq.Header.Set("Authorization", "Bearer "+apiKey)
+	} else if envName := strings.TrimSpace(req.Target.APIKeyEnv); envName != "" {
 		apiKey := strings.TrimSpace(os.Getenv(envName))
 		if apiKey == "" {
 			return nil, fmt.Errorf("%w: api key environment variable %q is not set", ErrInvalidRequest, envName)
@@ -142,6 +119,8 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 	effectiveTimeout := c.timeout
 	if req.Target.TimeoutSeconds > 0 {
 		effectiveTimeout = time.Duration(req.Target.TimeoutSeconds) * time.Second
+	} else if req.TargetPresence.TimeoutSeconds {
+		effectiveTimeout = 0
 	}
 
 	httpClient := c.httpClient
@@ -160,8 +139,8 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 	defer httpResp.Body.Close()
 
 	if httpResp.StatusCode < 200 || httpResp.StatusCode >= 300 {
-		body, _ := io.ReadAll(io.LimitReader(httpResp.Body, 4096))
-		return nil, fmt.Errorf("%w: status=%d body=%q", ErrUnexpectedStatus, httpResp.StatusCode, strings.TrimSpace(string(body)))
+		_, _ = io.Copy(io.Discard, io.LimitReader(httpResp.Body, 4096))
+		return nil, fmt.Errorf("%w: status=%d", ErrUnexpectedStatus, httpResp.StatusCode)
 	}
 
 	var wireResp openAIChatResponse
@@ -183,32 +162,166 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 			PromptTokens:     wireResp.Usage.PromptTokens,
 			CompletionTokens: wireResp.Usage.CompletionTokens,
 			TotalTokens:      wireResp.Usage.TotalTokens,
+			CachedTokens:     wireResp.Usage.PromptTokensDetails.CachedTokens,
+			CacheWriteTokens: wireResp.Usage.CacheWriteTokens,
 		},
 	}, nil
 }
 
+func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultModel string) (openAIChatRequest, error) {
+	model := strings.TrimSpace(req.Target.Model)
+	if model == "" {
+		model = strings.TrimSpace(defaultModel)
+	}
+	if model == "" {
+		return openAIChatRequest{}, errors.New("model is required")
+	}
+
+	wireReq := openAIChatRequest{
+		Model: model,
+	}
+	if sessionID := strings.TrimSpace(req.Prompt.SessionID); sessionID != "" {
+		if n := utf8.RuneCountInString(sessionID); n > domain.SessionIDMaxLength {
+			return openAIChatRequest{}, fmt.Errorf("session_id length %d exceeds maximum %d", n, domain.SessionIDMaxLength)
+		}
+		wireReq.SessionID = sessionID
+	}
+
+	wireReq.Messages = make([]openAIChatRequestMessage, 0, len(req.Prompt.Messages))
+	for _, msg := range req.Prompt.Messages {
+		wireReq.Messages = append(wireReq.Messages, openAIChatRequestMessageFromRenderedMessage(msg))
+	}
+
+	if req.Target.Temperature != 0 || req.TargetPresence.Temperature {
+		wireReq.Temperature = &req.Target.Temperature
+	}
+	if req.Target.MaxTokens != 0 || req.TargetPresence.MaxTokens {
+		wireReq.MaxTokens = &req.Target.MaxTokens
+	}
+	if req.Target.TopP != 0 || req.TargetPresence.TopP {
+		wireReq.TopP = &req.Target.TopP
+	}
+	if strings.TrimSpace(req.Target.ServiceTier) != "" {
+		wireReq.ServiceTier = req.Target.ServiceTier
+	}
+	if strings.TrimSpace(req.Target.ReasoningEffort) != "" {
+		wireReq.ReasoningEffort = req.Target.ReasoningEffort
+	}
+	if len(req.Target.ExtraParams) > 0 {
+		wireReq.ExtraParams = req.Target.ExtraParams
+	}
+	if req.StructuredOutput != nil {
+		responseFormat, err := toOpenAIResponseFormat(req.StructuredOutput)
+		if err != nil {
+			return openAIChatRequest{}, err
+		}
+		wireReq.ResponseFormat = responseFormat
+	}
+
+	return wireReq, nil
+}
+
 type openAIChatRequest struct {
 	Model           string                     `json:"model"`
-	Messages       []openAIChatMessage   `json:"messages"`
+	SessionID       string                     `json:"session_id,omitempty"`
+	Messages        []openAIChatRequestMessage `json:"messages"`
 	Temperature     *float64                   `json:"temperature,omitempty"`
 	MaxTokens       *int                       `json:"max_tokens,omitempty"`
 	TopP            *float64                   `json:"top_p,omitempty"`
+	ServiceTier     string                     `json:"service_tier,omitempty"`
+	ReasoningEffort string                     `json:"reasoning_effort,omitempty"`
 	ResponseFormat  *openAIResponseFormat      `json:"response_format,omitempty"`
+	ExtraParams     map[string]any             `json:"-"`
 }
 
-type openAIChatMessage struct {
+func openAIChatRequestPayload(req openAIChatRequest) (map[string]any, error) {
+	out := map[string]any{
+		"model":    req.Model,
+		"messages": req.Messages,
+	}
+	if req.SessionID != "" {
+		out["session_id"] = req.SessionID
+	}
+	if req.Temperature != nil {
+		out["temperature"] = *req.Temperature
+	}
+	if req.MaxTokens != nil {
+		out["max_tokens"] = *req.MaxTokens
+	}
+	if req.TopP != nil {
+		out["top_p"] = *req.TopP
+	}
+	if req.ServiceTier != "" {
+		out["service_tier"] = req.ServiceTier
+	}
+	if req.ReasoningEffort != "" {
+		out["reasoning_effort"] = req.ReasoningEffort
+	}
+	if req.ResponseFormat != nil {
+		out["response_format"] = req.ResponseFormat
+	}
+
+	for key, value := range req.ExtraParams {
+		if key == "" {
+			return nil, errors.New("extra_params key must not be empty")
+		}
+		if _, reserved := reservedOpenAIChatRequestFields[key]; reserved {
+			return nil, fmt.Errorf("extra_params key %q collides with reserved request field", key)
+		}
+		if _, err := json.Marshal(value); err != nil {
+			return nil, fmt.Errorf("extra_params.%s must be JSON-serializable: %w", key, err)
+		}
+		out[key] = value
+	}
+
+	return out, nil
+}
+
+var reservedOpenAIChatRequestFields = map[string]struct{}{
+	"model":            {},
+	"session_id":       {},
+	"messages":         {},
+	"temperature":      {},
+	"max_tokens":       {},
+	"top_p":            {},
+	"service_tier":     {},
+	"reasoning_effort": {},
+	"response_format":  {},
+}
+
+type openAIChatRequestMessage struct {
+	Role    string `json:"role"`
+	Content any    `json:"content"`
+}
+
+type openAIChatTextContentBlock struct {
+	Type         string              `json:"type"`
+	Text         string              `json:"text"`
+	CacheControl *openAICacheControl `json:"cache_control,omitempty"`
+}
+
+type openAICacheControl struct {
+	Type string `json:"type"`
+	TTL  string `json:"ttl,omitempty"`
+}
+
+type openAIChatResponseMessage struct {
 	Role    string `json:"role"`
 	Content string `json:"content"`
 }
 
 type openAIChatResponse struct {
 	Choices []struct {
-		Message openAIChatMessage `json:"message"`
+		Message openAIChatResponseMessage `json:"message"`
 	} `json:"choices"`
 	Usage struct {
 		PromptTokens        int `json:"prompt_tokens"`
 		CompletionTokens    int `json:"completion_tokens"`
 		TotalTokens         int `json:"total_tokens"`
+		PromptTokensDetails struct {
+			CachedTokens int `json:"cached_tokens"`
+		} `json:"prompt_tokens_details"`
+		CacheWriteTokens int `json:"cache_write_tokens"`
 	} `json:"usage"`
 }
 
@@ -223,6 +336,28 @@ type openAIJSONSchemaEnvelope struct {
 	Schema any    `json:"schema"`
 }
 
+func openAIChatRequestMessageFromRenderedMessage(msg domain.RenderedMessage) openAIChatRequestMessage {
+	wireMsg := openAIChatRequestMessage{
+		Role:    msg.Role,
+		Content: msg.Content,
+	}
+	if msg.CacheControl == nil {
+		return wireMsg
+	}
+
+	wireMsg.Content = []openAIChatTextContentBlock{
+		{
+			Type: "text",
+			Text: msg.Content,
+			CacheControl: &openAICacheControl{
+				Type: string(msg.CacheControl.Type),
+				TTL:  msg.CacheControl.TTL,
+			},
+		},
+	}
+	return wireMsg
+}
+
 func toOpenAIResponseFormat(spec *domain.StructuredOutputSpec) (*openAIResponseFormat, error) {
 	if spec == nil {
 		return nil, nil

internal/llm/openai_compatible_client_test.go

@@ -4,6 +4,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"math"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -13,6 +14,64 @@ import (
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
 )
 
+func TestNewOpenAICompatibleClientDoesNotMutateSuppliedZeroTimeoutClient(t *testing.T) {
+	transport := http.DefaultTransport
+	supplied := &http.Client{Transport: transport}
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		HTTPClient: supplied,
+	})
+	if err != nil {
+		t.Fatalf("unexpected constructor error: %v", err)
+	}
+
+	if supplied.Timeout != 0 {
+		t.Fatalf("expected supplied client timeout to remain zero, got %v", supplied.Timeout)
+	}
+	if client.httpClient == supplied {
+		t.Fatal("expected constructed client to use a cloned HTTP client")
+	}
+	if client.httpClient.Timeout != client.timeout {
+		t.Fatalf("expected cloned client timeout %v, got %v", client.timeout, client.httpClient.Timeout)
+	}
+	if client.httpClient.Timeout <= 0 {
+		t.Fatalf("expected constructed client to use a positive default timeout, got %v", client.httpClient.Timeout)
+	}
+	if client.httpClient.Transport != transport {
+		t.Fatal("expected cloned client to preserve the supplied transport")
+	}
+}
+
+func TestNewOpenAICompatibleClientDoesNotMutateSuppliedNonzeroTimeoutClient(t *testing.T) {
+	transport := http.DefaultTransport
+	suppliedTimeout := 37 * time.Second
+	supplied := &http.Client{
+		Timeout:   suppliedTimeout,
+		Transport: transport,
+	}
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		Timeout:    2 * time.Second,
+		HTTPClient: supplied,
+	})
+	if err != nil {
+		t.Fatalf("unexpected constructor error: %v", err)
+	}
+
+	if supplied.Timeout != suppliedTimeout {
+		t.Fatalf("expected supplied client timeout to remain %v, got %v", suppliedTimeout, supplied.Timeout)
+	}
+	if client.httpClient == supplied {
+		t.Fatal("expected constructed client to use a cloned HTTP client")
+	}
+	if client.httpClient.Timeout != suppliedTimeout {
+		t.Fatalf("expected cloned client timeout %v, got %v", suppliedTimeout, client.httpClient.Timeout)
+	}
+	if client.httpClient.Transport != transport {
+		t.Fatal("expected cloned client to preserve the supplied transport")
+	}
+}
+
 func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 	type observedRequest struct {
 		Authorization string
@@ -61,6 +120,7 @@ func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 			Temperature: 0.4,
 			MaxTokens:   123,
 			TopP:        0.7,
+			ServiceTier: "priority",
 			APIKeyEnv:   "SCRIPTORIUM_TEST_API_KEY",
 		},
 		StructuredOutput: &domain.StructuredOutputSpec{
@@ -88,6 +148,9 @@ func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 	if resp.Usage.PromptTokens != 11 || resp.Usage.CompletionTokens != 22 || resp.Usage.TotalTokens != 33 {
 		t.Fatalf("unexpected usage: %+v", resp.Usage)
 	}
+	if resp.Usage.CachedTokens != 0 || resp.Usage.CacheWriteTokens != 0 {
+		t.Fatalf("expected absent cache usage fields to remain zero, got %+v", resp.Usage)
+	}
 
 	if obs.Authorization != "Bearer secret-key" {
 		t.Fatalf("unexpected Authorization header: %q", obs.Authorization)
@@ -95,6 +158,18 @@ func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 	if got, ok := obs.Body["model"].(string); !ok || got != "gpt-test" {
 		t.Fatalf("unexpected model payload: %#v", obs.Body["model"])
 	}
+	if got, ok := obs.Body["temperature"].(float64); !ok || got != 0.4 {
+		t.Fatalf("unexpected temperature payload: %#v", obs.Body["temperature"])
+	}
+	if got, ok := obs.Body["max_tokens"].(float64); !ok || got != 123 {
+		t.Fatalf("unexpected max_tokens payload: %#v", obs.Body["max_tokens"])
+	}
+	if got, ok := obs.Body["top_p"].(float64); !ok || got != 0.7 {
+		t.Fatalf("unexpected top_p payload: %#v", obs.Body["top_p"])
+	}
+	if got, ok := obs.Body["service_tier"].(string); !ok || got != "priority" {
+		t.Fatalf("unexpected service_tier payload: %#v", obs.Body["service_tier"])
+	}
 
 	msgs, ok := obs.Body["messages"].([]any)
 	if !ok || len(msgs) != 2 {
@@ -131,6 +206,277 @@ func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 	}
 }
 
+func TestOpenAICompatibleClientDirectAPIKeyPreferredOverEnv(t *testing.T) {
+	const directKey = "direct-llm-key"
+	t.Setenv("SCRIPTORIUM_TEST_API_KEY", "env-key")
+
+	var gotAuth string
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		gotAuth = r.Header.Get("Authorization")
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{
+			Model:     "model",
+			APIKeyEnv: "SCRIPTORIUM_TEST_API_KEY",
+			APIKey:    directKey,
+		},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if gotAuth != "Bearer "+directKey {
+		t.Fatalf("unexpected Authorization header: %q", gotAuth)
+	}
+}
+
+func TestOpenAICompatibleClientSerializesCacheControlledMessageAsContentBlock(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+			{
+				Role:    "system",
+				Content: "Stable instructions.",
+				CacheControl: &domain.CacheControl{
+					Type: domain.CacheControlEphemeral,
+					TTL:  "1h",
+				},
+			},
+			{Role: "user", Content: "Dynamic request."},
+		}},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+
+	for _, forbidden := range []string{"cache_control", "extra_params"} {
+		if _, exists := observedBody[forbidden]; exists {
+			t.Fatalf("expected top-level %s to be omitted, got %#v", forbidden, observedBody[forbidden])
+		}
+	}
+
+	msgs, ok := observedBody["messages"].([]any)
+	if !ok || len(msgs) != 2 {
+		t.Fatalf("unexpected messages payload: %#v", observedBody["messages"])
+	}
+	msg0 := msgs[0].(map[string]any)
+	if msg0["role"] != "system" {
+		t.Fatalf("unexpected first message role: %#v", msg0["role"])
+	}
+	contentBlocks, ok := msg0["content"].([]any)
+	if !ok || len(contentBlocks) != 1 {
+		t.Fatalf("expected first message content block array, got %#v", msg0["content"])
+	}
+	block := contentBlocks[0].(map[string]any)
+	if block["type"] != "text" || block["text"] != "Stable instructions." {
+		t.Fatalf("unexpected text content block: %#v", block)
+	}
+	cacheControl, ok := block["cache_control"].(map[string]any)
+	if !ok {
+		t.Fatalf("expected cache_control on content block, got %#v", block)
+	}
+	if cacheControl["type"] != string(domain.CacheControlEphemeral) || cacheControl["ttl"] != "1h" {
+		t.Fatalf("unexpected cache_control payload: %#v", cacheControl)
+	}
+
+	msg1 := msgs[1].(map[string]any)
+	if msg1["role"] != "user" || msg1["content"] != "Dynamic request." {
+		t.Fatalf("expected uncached message to keep string content, got %#v", msg1)
+	}
+}
+
+func TestOpenAICompatibleClientOmitsEmptyCacheControlTTL(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+			{
+				Role:    "system",
+				Content: "Stable instructions.",
+				CacheControl: &domain.CacheControl{
+					Type: domain.CacheControlEphemeral,
+				},
+			},
+		}},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+
+	msgs := observedBody["messages"].([]any)
+	msg0 := msgs[0].(map[string]any)
+	contentBlocks := msg0["content"].([]any)
+	block := contentBlocks[0].(map[string]any)
+	cacheControl := block["cache_control"].(map[string]any)
+	if cacheControl["type"] != string(domain.CacheControlEphemeral) {
+		t.Fatalf("unexpected cache_control type: %#v", cacheControl)
+	}
+	if _, exists := cacheControl["ttl"]; exists {
+		t.Fatalf("expected empty ttl to be omitted, got %#v", cacheControl)
+	}
+}
+
+func TestOpenAICompatibleClientSerializesSessionID(t *testing.T) {
+	var observedBody map[string]any
+	var observedSessionHeader string
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		observedSessionHeader = r.Header.Get("x-session-id")
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{
+			SessionID: " session-123 ",
+			Messages:  []domain.RenderedMessage{{Role: "user", Content: "hi"}},
+		},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if observedBody["session_id"] != "session-123" {
+		t.Fatalf("expected top-level session_id, got %#v", observedBody["session_id"])
+	}
+	if observedSessionHeader != "" {
+		t.Fatalf("did not expect x-session-id header, got %q", observedSessionHeader)
+	}
+}
+
+func TestOpenAICompatibleClientOmitsEmptySessionID(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{
+			SessionID: "   ",
+			Messages:  []domain.RenderedMessage{{Role: "user", Content: "hi"}},
+		},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if _, exists := observedBody["session_id"]; exists {
+		t.Fatalf("expected empty session_id to be omitted, got %#v", observedBody["session_id"])
+	}
+}
+
+func TestOpenAICompatibleClientRejectsTooLongSessionID(t *testing.T) {
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: "http://example.com/v1",
+		Model:   "model",
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{
+			SessionID: strings.Repeat("x", domain.SessionIDMaxLength+1),
+			Messages:  []domain.RenderedMessage{{Role: "user", Content: "hi"}},
+		},
+	})
+	if err == nil {
+		t.Fatal("expected invalid request error")
+	}
+	if !errors.Is(err, ErrInvalidRequest) {
+		t.Fatalf("expected ErrInvalidRequest, got %v", err)
+	}
+}
+
+func TestOpenAICompatibleClientParsesCacheUsage(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		_, _ = w.Write([]byte(`{
+  "choices": [{"message": {"role": "assistant", "content": "ok"}}],
+  "usage": {
+    "prompt_tokens": 100,
+    "completion_tokens": 20,
+    "total_tokens": 120,
+    "prompt_tokens_details": {"cached_tokens": 80},
+    "cache_write_tokens": 60
+  }
+}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1", Model: "model"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	resp, err := client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if resp.Usage.PromptTokens != 100 || resp.Usage.CompletionTokens != 20 || resp.Usage.TotalTokens != 120 {
+		t.Fatalf("unexpected base usage fields: %+v", resp.Usage)
+	}
+	if resp.Usage.CachedTokens != 80 || resp.Usage.CacheWriteTokens != 60 {
+		t.Fatalf("unexpected cache usage fields: %+v", resp.Usage)
+	}
+}
+
 func TestOpenAICompatibleClientOmitsResponseFormatWhenNoStructuredOutput(t *testing.T) {
 	var observedBody map[string]any
 	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
@@ -157,6 +503,284 @@ func TestOpenAICompatibleClientOmitsResponseFormatWhenNoStructuredOutput(t *test
 	if _, exists := observedBody["response_format"]; exists {
 		t.Fatalf("expected response_format omitted, got %#v", observedBody["response_format"])
 	}
+	if _, exists := observedBody["service_tier"]; exists {
+		t.Fatalf("expected service_tier omitted, got %#v", observedBody["service_tier"])
+	}
+}
+
+func TestOpenAICompatibleClientSerializesReasoningEffortAndExtraParams(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{
+			Model:           "model",
+			ReasoningEffort: "high",
+			ExtraParams: map[string]any{
+				"string_value":  "on",
+				"number_value":  42,
+				"boolean_value": true,
+				"object_value":  map[string]any{"nested": "value", "count": 2},
+				"array_value":   []any{"first", 3, false},
+			},
+		},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if observedBody["reasoning_effort"] != "high" {
+		t.Fatalf("expected reasoning_effort high, got %#v", observedBody["reasoning_effort"])
+	}
+	if observedBody["string_value"] != "on" {
+		t.Fatalf("unexpected string extra param: %#v", observedBody["string_value"])
+	}
+	if observedBody["number_value"] != float64(42) {
+		t.Fatalf("unexpected number extra param: %#v", observedBody["number_value"])
+	}
+	if observedBody["boolean_value"] != true {
+		t.Fatalf("unexpected boolean extra param: %#v", observedBody["boolean_value"])
+	}
+	objectValue, ok := observedBody["object_value"].(map[string]any)
+	if !ok || objectValue["nested"] != "value" || objectValue["count"] != float64(2) {
+		t.Fatalf("unexpected object extra param: %#v", observedBody["object_value"])
+	}
+	if _, exists := observedBody["extra_params"]; exists {
+		t.Fatalf("expected extra_params wrapper omitted, got %#v", observedBody["extra_params"])
+	}
+	arrayValue, ok := observedBody["array_value"].([]any)
+	if !ok || len(arrayValue) != 3 || arrayValue[0] != "first" || arrayValue[1] != float64(3) || arrayValue[2] != false {
+		t.Fatalf("unexpected array extra param: %#v", observedBody["array_value"])
+	}
+}
+
+func TestOpenAICompatibleClientOmitsReasoningEffortWhenUnset(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if _, exists := observedBody["reasoning_effort"]; exists {
+		t.Fatalf("expected reasoning_effort omitted, got %#v", observedBody["reasoning_effort"])
+	}
+	if _, exists := observedBody["extra_params"]; exists {
+		t.Fatalf("expected extra_params wrapper omitted, got %#v", observedBody["extra_params"])
+	}
+}
+
+func TestOpenAICompatibleClientSerializesExplicitZeroNumericOverrides(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{Model: "model"},
+		TargetPresence: domain.ExecutionTargetPresence{
+			Temperature: true,
+			MaxTokens:   true,
+			TopP:        true,
+		},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if observedBody["temperature"] != float64(0) {
+		t.Fatalf("expected explicit zero temperature, got %#v", observedBody["temperature"])
+	}
+	if observedBody["max_tokens"] != float64(0) {
+		t.Fatalf("expected explicit zero max_tokens, got %#v", observedBody["max_tokens"])
+	}
+	if observedBody["top_p"] != float64(0) {
+		t.Fatalf("expected explicit zero top_p, got %#v", observedBody["top_p"])
+	}
+}
+
+func TestOpenAICompatibleClientOmitsImplicitZeroNumericFields(t *testing.T) {
+	var observedBody map[string]any
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		defer r.Body.Close()
+		if err := json.NewDecoder(r.Body).Decode(&observedBody); err != nil {
+			t.Fatalf("failed to decode request body: %v", err)
+		}
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{Model: "model"},
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	for _, field := range []string{"temperature", "max_tokens", "top_p"} {
+		if _, exists := observedBody[field]; exists {
+			t.Fatalf("expected implicit zero field %q to be omitted, got body %#v", field, observedBody)
+		}
+	}
+}
+
+func TestOpenAICompatibleClientExplicitZeroTimeoutDisablesClientTimeout(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		time.Sleep(20 * time.Millisecond)
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: ts.URL + "/v1",
+		Timeout: time.Nanosecond,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt:         domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target:         domain.ExecutionTarget{Model: "model", TimeoutSeconds: 0},
+		TargetPresence: domain.ExecutionTargetPresence{TimeoutSeconds: true},
+	})
+	if err != nil {
+		t.Fatalf("expected explicit zero timeout to disable client timeout, got %v", err)
+	}
+}
+
+func TestOpenAICompatibleClientOmittedTimeoutUsesClientTimeout(t *testing.T) {
+	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		time.Sleep(20 * time.Millisecond)
+		_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+	}))
+	defer ts.Close()
+
+	client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{
+		BaseURL: ts.URL + "/v1",
+		Timeout: time.Nanosecond,
+	})
+	if err != nil {
+		t.Fatal(err)
+	}
+
+	_, err = client.Generate(context.Background(), domain.GenerateRequest{
+		Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+		Target: domain.ExecutionTarget{Model: "model", TimeoutSeconds: 0},
+	})
+	if err == nil {
+		t.Fatal("expected omitted timeout to use client timeout")
+	}
+	if !errors.Is(err, ErrRequestFailed) {
+		t.Fatalf("expected ErrRequestFailed, got %v", err)
+	}
+}
+
+func TestOpenAICompatibleClientRejectsInvalidExtraParamsBeforeProviderCall(t *testing.T) {
+	tests := []struct {
+		name        string
+		extraParams map[string]any
+		want        string
+	}{
+		{name: "empty key", extraParams: map[string]any{"": "empty"}, want: "key must not be empty"},
+		{name: "unserializable value", extraParams: map[string]any{"bad": math.Inf(1)}, want: "JSON-serializable"},
+	}
+	for _, key := range []string{
+		"model",
+		"session_id",
+		"messages",
+		"temperature",
+		"max_tokens",
+		"top_p",
+		"service_tier",
+		"reasoning_effort",
+		"response_format",
+	} {
+		tests = append(tests, struct {
+			name        string
+			extraParams map[string]any
+			want        string
+		}{
+			name:        "reserved key " + key,
+			extraParams: map[string]any{key: "collision"},
+			want:        "reserved request field",
+		})
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			called := false
+			ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				called = true
+				_, _ = w.Write([]byte(`{"choices":[{"message":{"content":"ok"}}]}`))
+			}))
+			defer ts.Close()
+
+			client, err := NewOpenAICompatibleClient(OpenAICompatibleConfig{BaseURL: ts.URL + "/v1"})
+			if err != nil {
+				t.Fatal(err)
+			}
+
+			_, err = client.Generate(context.Background(), domain.GenerateRequest{
+				Prompt: domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "user", Content: "hi"}}},
+				Target: domain.ExecutionTarget{Model: "model", ExtraParams: tc.extraParams},
+			})
+			if err == nil {
+				t.Fatal("expected invalid request error")
+			}
+			if !errors.Is(err, ErrInvalidRequest) {
+				t.Fatalf("expected ErrInvalidRequest, got %v", err)
+			}
+			if !strings.Contains(err.Error(), tc.want) {
+				t.Fatalf("expected error to contain %q, got %v", tc.want, err)
+			}
+			if called {
+				t.Fatal("provider should not be called for invalid extra_params")
+			}
+		})
+	}
 }
 
 func TestOpenAICompatibleClientNoAuthorizationHeaderWhenNoAPIKey(t *testing.T) {
@@ -279,9 +903,10 @@ func TestOpenAICompatibleClientEndpointOverride(t *testing.T) {
 }
 
 func TestOpenAICompatibleClientNon2xxError(t *testing.T) {
+	const sensitiveBody = `provider-secret-fragment request_payload_details`
 	ts := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
 		w.WriteHeader(http.StatusBadRequest)
-		_, _ = w.Write([]byte(`{"error":"bad request payload"}`))
+		_, _ = w.Write([]byte(`{"error":"` + sensitiveBody + `"}`))
 	}))
 	defer ts.Close()
 
@@ -299,8 +924,11 @@ func TestOpenAICompatibleClientNon2xxError(t *testing.T) {
 	if !errors.Is(err, ErrUnexpectedStatus) {
 		t.Fatalf("expected ErrUnexpectedStatus, got %v", err)
 	}
-	if !strings.Contains(err.Error(), "400") || !strings.Contains(err.Error(), "bad request payload") {
-		t.Fatalf("expected status/body details, got %v", err)
+	if !strings.Contains(err.Error(), "status=400") {
+		t.Fatalf("expected status detail, got %v", err)
+	}
+	if strings.Contains(err.Error(), sensitiveBody) {
+		t.Fatalf("expected provider response body to be redacted, got %v", err)
 	}
 }
 

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

@@ -0,0 +1,9 @@
+id: aion-2
+endpoint: https://openrouter.ai/api/v1
+model: aion-labs/aion-2.0
+temperature: 0.72
+reasoning_effort: high
+top_p: 0.95
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: claude-fable-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~anthropic/claude-fable-latest"
+reasoning_effort: high
+timeout_seconds: 600
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: claude-haiku-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~anthropic/claude-haiku-latest"
+reasoning_effort: medium
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: claude-opus-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~anthropic/claude-opus-latest"
+reasoning_effort: high
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: claude-sonnet-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~anthropic/claude-sonnet-latest"
+reasoning_effort: high
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: deepseek-3-2
+endpoint: https://openrouter.ai/api/v1
+model: deepseek/deepseek-v3.2
+reasoning_effort: high
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: deepseek-4-flash
+endpoint: https://openrouter.ai/api/v1
+model: deepseek/deepseek-v4-flash
+#reasoning_effort: medium
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: deepseek-4-pro
+endpoint: https://openrouter.ai/api/v1
+model: deepseek/deepseek-v4-pro
+reasoning_effort: high
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-2-flash-lite
+endpoint: https://openrouter.ai/api/v1
+model: "google/gemini-2.5-flash-lite"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-2-flash
+endpoint: https://openrouter.ai/api/v1
+model: "google/gemini-2.5-flash"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-2-pro
+endpoint: https://openrouter.ai/api/v1
+model: "google/gemini-2.5-pro"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-3-flash-lite
+endpoint: https://openrouter.ai/api/v1
+model: "google/gemini-3.1-flash-lite"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-flash-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~google/gemini-flash-latest"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemini-pro-latest
+endpoint: https://openrouter.ai/api/v1
+model: "~google/gemini-pro-latest"
+#temperature: 0.15
+reasoning_effort: high
+#top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: gemma-4-31b
+endpoint: https://openrouter.ai/api/v1
+model: google/gemma-4-31b-it:exacto
+temperature: 0.15
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: minimax-m2
+endpoint: https://openrouter.ai/api/v1
+model: minimax/minimax-m2.5
+temperature: 0.5
+reasoning_effort: high
+top_p: 0.95
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,9 @@
+id: minimax-m3
+endpoint: https://openrouter.ai/api/v1
+model: minimax/minimax-m3
+#temperature: 0.5
+reasoning_effort: high
+#top_p: 0.95
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: mistral-large-2512
+endpoint: https://openrouter.ai/api/v1
+model: mistralai/mistral-large-2512
+temperature: 0.15
+top_p: 0.98
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY

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

@@ -0,0 +1,8 @@
+id: mistral-medium-3-5
+endpoint: https://openrouter.ai/api/v1
+model: mistralai/mistral-medium-3-5
+temperature: 0.15
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY

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

@@ -0,0 +1,7 @@
+id: mistral-small-3
+endpoint: https://openrouter.ai/api/v1
+model: mistralai/mistral-small-3.2-24b-instruct
+temperature: 0.05
+top_p: 1.0
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY

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

@@ -0,0 +1,8 @@
+id: mistral-small-4
+endpoint: https://openrouter.ai/api/v1
+model: mistralai/mistral-small-2603
+temperature: 0.1
+reasoning_effort: high
+top_p: 0.98
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY

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

@@ -0,0 +1,7 @@
+id: nemotron-3-ultra
+endpoint: https://openrouter.ai/api/v1
+model: nvidia/nemotron-3-ultra-550b-a55b
+reasoning_effort: high
+timeout_seconds: 180
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: gpt-5-mini
+endpoint: https://openrouter.ai/api/v1
+model: "openai/gpt-5.4-mini"
+reasoning_effort: high
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

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

@@ -0,0 +1,7 @@
+id: gpt-5-nano
+endpoint: https://openrouter.ai/api/v1
+model: "openai/gpt-5.4-nano"
+reasoning_effort: high
+timeout_seconds: 240
+api_key_env: OPENROUTER_API_KEY
+service_tier: flex

internal/profile/builtin/repository.go

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

internal/profile/builtin/repository_test.go

@@ -0,0 +1,127 @@
+package builtin
+
+import (
+	"context"
+	"errors"
+	"io/fs"
+	"strings"
+	"testing"
+
+	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/profile"
+	"gopkg.in/yaml.v3"
+)
+
+func TestBuiltInProfilesValidateThroughRepository(t *testing.T) {
+	repo := NewRepository()
+	ids := loadBuiltInProfileIDs(t)
+	if len(ids) == 0 {
+		t.Fatal("expected built-in profiles")
+	}
+
+	for id := range ids {
+		t.Run(id, func(t *testing.T) {
+			p, err := repo.GetProfile(context.Background(), id)
+			if err != nil {
+				t.Fatalf("expected built-in profile %q to load, got %v", id, err)
+			}
+			if p.ID != id {
+				t.Fatalf("expected profile id %q, got %q", id, p.ID)
+			}
+		})
+	}
+}
+
+func TestBuiltInProfilesDoNotContainDuplicateIDsOrRawAPIKeys(t *testing.T) {
+	loadBuiltInProfileIDs(t)
+}
+
+func loadBuiltInProfileIDs(t *testing.T) map[string]string {
+	t.Helper()
+
+	ids := map[string]string{}
+	err := fs.WalkDir(assets, assetRoot, func(name string, d fs.DirEntry, err error) error {
+		if err != nil {
+			return err
+		}
+		if d.IsDir() || !strings.HasSuffix(name, ".yml") {
+			return nil
+		}
+
+		data, err := assets.ReadFile(name)
+		if err != nil {
+			t.Fatalf("failed to read built-in profile %s: %v", name, err)
+		}
+
+		var raw map[string]any
+		if err := yaml.Unmarshal(data, &raw); err != nil {
+			t.Fatalf("failed to decode built-in profile %s: %v", name, err)
+		}
+		if _, ok := raw["api_key"]; ok {
+			t.Fatalf("built-in profile %s contains raw api_key", name)
+		}
+		id, ok := raw["id"].(string)
+		if !ok || strings.TrimSpace(id) == "" {
+			t.Fatalf("built-in profile %s has missing id", name)
+		}
+		if previous, ok := ids[id]; ok {
+			t.Fatalf("duplicate built-in profile id %q in %s and %s", id, previous, name)
+		}
+		ids[id] = name
+		return nil
+	})
+	if err != nil {
+		t.Fatalf("failed to walk built-in profiles: %v", err)
+	}
+	return ids
+}
+
+func TestRepositoryWithPrimaryUsesPrimaryBeforeBuiltIns(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{
+		profiles: map[string]string{"mistral-small-3": "custom-model"},
+	})
+
+	p, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if err != nil {
+		t.Fatalf("expected profile to load, got %v", err)
+	}
+	if p.Model != "custom-model" {
+		t.Fatalf("expected primary profile to override built-in, got %+v", p)
+	}
+}
+
+func TestRepositoryWithPrimaryFallsBackToBuiltIns(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{})
+
+	p, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if err != nil {
+		t.Fatalf("expected built-in profile to load, got %v", err)
+	}
+	if p.ID != "mistral-small-3" {
+		t.Fatalf("unexpected profile: %+v", p)
+	}
+}
+
+func TestRepositoryWithPrimaryDoesNotFallBackAfterPrimaryError(t *testing.T) {
+	repo := NewRepositoryWithPrimary(staticProfileRepo{err: profile.ErrInvalidProfile})
+
+	_, err := repo.GetProfile(context.Background(), "mistral-small-3")
+	if !errors.Is(err, profile.ErrInvalidProfile) {
+		t.Fatalf("expected primary error, got %v", err)
+	}
+}
+
+type staticProfileRepo struct {
+	profiles map[string]string
+	err      error
+}
+
+func (r staticProfileRepo) GetProfile(_ context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.err != nil {
+		return nil, r.err
+	}
+	if model, ok := r.profiles[id]; ok {
+		return &domain.ExecutionProfile{ID: id, Endpoint: "http://primary/v1", Model: model}, nil
+	}
+	return nil, profile.ErrProfileNotFound
+}

internal/profile/filesystem_repository.go

@@ -5,11 +5,13 @@ import (
 	"context"
 	"errors"
 	"fmt"
+	"io/fs"
 	"os"
-	"path/filepath"
+	"path"
 	"strings"
 
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+	"gitea.maximumdirect.net/eric/scriptorium/internal/filecatalog"
 	"gopkg.in/yaml.v3"
 )
 
@@ -29,44 +31,89 @@ func NewFilesystemRepository(dir string) Repository {
 }
 
 func (r *filesystemRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	return loadProfile(ctx, os.DirFS(r.dir), ".", id)
+}
+
+type fsRepository struct {
+	fsys fs.FS
+	root string
+}
+
+func NewFSRepository(fsys fs.FS, root string) Repository {
+	return &fsRepository{fsys: fsys, root: root}
+}
+
+func (r *fsRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	return loadProfile(ctx, r.fsys, r.root, id)
+}
+
+type overlayRepository struct {
+	primary  Repository
+	fallback Repository
+}
+
+func NewOverlayRepository(primary, fallback Repository) Repository {
+	return &overlayRepository{primary: primary, fallback: fallback}
+}
+
+func (r *overlayRepository) GetProfile(ctx context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.primary != nil {
+		prof, err := r.primary.GetProfile(ctx, id)
+		if err == nil {
+			return prof, nil
+		}
+		if !errors.Is(err, ErrProfileNotFound) {
+			return nil, err
+		}
+	}
+	if r.fallback == nil {
+		return nil, ErrProfileNotFound
+	}
+	return r.fallback.GetProfile(ctx, id)
+}
+
+func loadProfile(ctx context.Context, fsys fs.FS, root string, id string) (*domain.ExecutionProfile, error) {
 	if strings.TrimSpace(id) == "" {
 		return nil, fmt.Errorf("%w: profile id is required", ErrInvalidProfile)
 	}
+	if fsys == nil {
+		return nil, fmt.Errorf("failed to read profile directory: filesystem is nil")
+	}
 
-	files, err := os.ReadDir(r.dir)
+	files, err := filecatalog.FindFSYAMLFiles(ctx, fsys, root)
 	if err != nil {
 		return nil, fmt.Errorf("failed to read profile directory: %w", err)
 	}
 
-	for _, file := range files {
+	var matches []profileMatch
+	for _, fullPath := range files {
 		select {
 		case <-ctx.Done():
 			return nil, ctx.Err()
 		default:
 		}
 
-		if file.IsDir() || (!strings.HasSuffix(file.Name(), ".yaml") && !strings.HasSuffix(file.Name(), ".yml")) {
-			continue
-		}
-
-		fullPath := filepath.Join(r.dir, file.Name())
-		data, err := os.ReadFile(fullPath)
+		relPath := filecatalog.DisplayPath(root, fullPath)
+		fileMatch := filecatalog.Stem(path.Base(fullPath)) == id
+		data, err := fs.ReadFile(fsys, fullPath)
 		if err != nil {
-			return nil, fmt.Errorf("failed to read profile file %s: %w", file.Name(), err)
+			return nil, fmt.Errorf("failed to read profile file %s: %w", relPath, err)
+		}
+		metadata := readProfileFileMetadata(data)
+		idMatch := fileMatch || metadata.id == id
+		if metadata.hasRawAPIKey {
+			if idMatch {
+				return nil, fmt.Errorf("%w: %s", ErrRawAPIKeyNotAllowed, relPath)
+			}
+			continue
 		}
 
 		var prof domain.ExecutionProfile
 		decoder := yaml.NewDecoder(bytes.NewReader(data))
 		decoder.KnownFields(true)
 		if err := decoder.Decode(&prof); err != nil {
-			if strings.Contains(err.Error(), "field api_key not found") {
-				if strings.TrimSuffix(strings.TrimSuffix(file.Name(), ".yaml"), ".yml") == id {
-					return nil, fmt.Errorf("%w: %s", ErrRawAPIKeyNotAllowed, file.Name())
-				}
-				continue
-			}
-			if strings.TrimSuffix(strings.TrimSuffix(file.Name(), ".yaml"), ".yml") == id {
-				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidYAML, file.Name(), err)
+			if idMatch {
+				return nil, fmt.Errorf("%w: %s: %v", ErrInvalidYAML, relPath, err)
 			}
 			continue
 		}
@@ -76,16 +123,68 @@ func (r *filesystemRepository) GetProfile(ctx context.Context, id string) (*doma
 		}
 		if err := validateProfile(&prof); err != nil {
 			if errors.Is(err, ErrRawAPIKeyNotAllowed) {
-				return nil, fmt.Errorf("%w: %s", err, file.Name())
+				return nil, fmt.Errorf("%w: %s", err, relPath)
 			}
-			return nil, fmt.Errorf("%w: %s: %v", ErrInvalidProfile, file.Name(), err)
+			return nil, fmt.Errorf("%w: %s: %v", ErrInvalidProfile, relPath, err)
 		}
-		return &prof, nil
+		matches = append(matches, profileMatch{
+			profile: &prof,
+			path:    relPath,
+		})
+	}
+
+	if len(matches) > 1 {
+		paths := make([]string, 0, len(matches))
+		for _, match := range matches {
+			paths = append(paths, match.path)
+		}
+		return nil, fmt.Errorf("%w: duplicate execution profile id %q found in: %s", ErrInvalidProfile, id, strings.Join(paths, ", "))
+	}
+
+	if len(matches) == 1 {
+		return matches[0].profile, nil
 	}
 
 	return nil, ErrProfileNotFound
 }
 
+type profileMatch struct {
+	profile *domain.ExecutionProfile
+	path    string
+}
+
+type profileFileMetadata struct {
+	id           string
+	hasRawAPIKey bool
+}
+
+func readProfileFileMetadata(data []byte) profileFileMetadata {
+	var node yaml.Node
+	if err := yaml.NewDecoder(bytes.NewReader(data)).Decode(&node); err != nil {
+		return profileFileMetadata{}
+	}
+	if node.Kind != yaml.DocumentNode || len(node.Content) == 0 {
+		return profileFileMetadata{}
+	}
+	mapping := node.Content[0]
+	if mapping.Kind != yaml.MappingNode {
+		return profileFileMetadata{}
+	}
+
+	var metadata profileFileMetadata
+	for i := 0; i+1 < len(mapping.Content); i += 2 {
+		key := mapping.Content[i]
+		value := mapping.Content[i+1]
+		switch key.Value {
+		case "id":
+			metadata.id = strings.TrimSpace(value.Value)
+		case "api_key":
+			metadata.hasRawAPIKey = true
+		}
+	}
+	return metadata
+}
+
 func validateProfile(p *domain.ExecutionProfile) error {
 	if strings.TrimSpace(p.ID) == "" {
 		return errors.New("id is required")

internal/profile/repository_test.go

@@ -2,10 +2,15 @@ package profile
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"os"
 	"path/filepath"
+	"strings"
 	"testing"
+	"testing/fstest"
+
+	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
 )
 
 func TestFilesystemRepository_GetProfile(t *testing.T) {
@@ -58,6 +63,149 @@ func TestFilesystemRepository_GetProfile(t *testing.T) {
 		if p.ReasoningEffort != "medium" {
 			t.Fatalf("unexpected reasoning_effort: %q", p.ReasoningEffort)
 		}
+		if p.ServiceTier != "priority" {
+			t.Fatalf("unexpected service_tier: %q", p.ServiceTier)
+		}
+	})
+
+	t.Run("valid nested profile", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "local")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "nested-local.yaml"), `
+id: nested-local
+endpoint: http://localhost:8000/v1
+model: nested-model
+temperature: 0.1
+`)
+
+		p, err := repo.GetProfile(ctx, "nested-local")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "nested-model" {
+			t.Fatalf("unexpected model: %q", p.Model)
+		}
+	})
+
+	t.Run("valid profile with JSON-compatible extra params", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "json-extra-params.yaml"), `
+id: json-extra-params
+endpoint: http://localhost:8000/v1
+model: nested-model
+extra_params:
+  string_value: enabled
+  number_value: 42
+  boolean_value: true
+  object_value:
+    nested: value
+    count: 2
+  array_value:
+    - first
+    - 3
+    - false
+`)
+
+		p, err := repo.GetProfile(ctx, "json-extra-params")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+
+		var got map[string]any
+		encoded, err := json.Marshal(p.ExtraParams)
+		if err != nil {
+			t.Fatalf("expected extra_params to marshal as JSON, got %v", err)
+		}
+		if err := json.Unmarshal(encoded, &got); err != nil {
+			t.Fatalf("expected extra_params JSON to decode, got %v", err)
+		}
+
+		if got["string_value"] != "enabled" {
+			t.Fatalf("unexpected string extra param: %#v", got["string_value"])
+		}
+		if got["number_value"] != float64(42) {
+			t.Fatalf("unexpected number extra param: %#v", got["number_value"])
+		}
+		if got["boolean_value"] != true {
+			t.Fatalf("unexpected boolean extra param: %#v", got["boolean_value"])
+		}
+		objectValue, ok := got["object_value"].(map[string]any)
+		if !ok {
+			t.Fatalf("expected object extra param, got %#v", got["object_value"])
+		}
+		if objectValue["nested"] != "value" || objectValue["count"] != float64(2) {
+			t.Fatalf("unexpected object extra param: %#v", objectValue)
+		}
+		arrayValue, ok := got["array_value"].([]any)
+		if !ok {
+			t.Fatalf("expected array extra param, got %#v", got["array_value"])
+		}
+		if len(arrayValue) != 3 || arrayValue[0] != "first" || arrayValue[1] != float64(3) || arrayValue[2] != false {
+			t.Fatalf("unexpected array extra param: %#v", arrayValue)
+		}
+	})
+
+	t.Run("duplicate profile IDs fail as ambiguous", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "duplicate-profile-a.yaml"), `
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: first-model
+`)
+		nestedDir := filepath.Join(tmpDir, "duplicates")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "duplicate-profile-b.yaml"), `
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: second-model
+`)
+
+		_, err := repo.GetProfile(ctx, "duplicate-profile")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected duplicate profile to return ErrInvalidProfile, got %v", err)
+		}
+		for _, want := range []string{"duplicate execution profile id", "duplicate-profile-a.yaml", filepath.Join("duplicates", "duplicate-profile-b.yaml")} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+
+	t.Run("nested raw api_key rejected for likely target file", func(t *testing.T) {
+		nestedDir := filepath.Join(tmpDir, "secure")
+		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
+			t.Fatal(err)
+		}
+		writeProfileTestFile(t, filepath.Join(nestedDir, "not_named_like_id.yaml"), `
+id: nested_raw_api_key
+endpoint: http://localhost:8000/v1
+model: m
+api_key: secret
+`)
+
+		_, err := repo.GetProfile(ctx, "nested_raw_api_key")
+		if !errors.Is(err, ErrRawAPIKeyNotAllowed) {
+			t.Fatalf("expected ErrRawAPIKeyNotAllowed, got %v", err)
+		}
+		if !strings.Contains(err.Error(), filepath.Join("secure", "not_named_like_id.yaml")) {
+			t.Fatalf("expected nested path in error, got %v", err)
+		}
+	})
+
+	t.Run("raw api_key in non-target profile is ignored", func(t *testing.T) {
+		writeProfileTestFile(t, filepath.Join(tmpDir, "raw-api-key-non-target.yaml"), `
+id: raw-api-key-non-target
+endpoint: http://localhost:8000/v1
+model: m
+api_key: secret
+`)
+
+		_, err := repo.GetProfile(ctx, "does-not-exist-with-raw-key-nearby")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound for non-target raw api_key file, got %v", err)
+		}
 	})
 
 	t.Run("invalid yaml", func(t *testing.T) {
@@ -109,3 +257,223 @@ func TestFilesystemRepository_GetProfile(t *testing.T) {
 		}
 	})
 }
+
+func writeProfileTestFile(t *testing.T, path string, content string) {
+	t.Helper()
+	if err := os.WriteFile(path, []byte(strings.TrimLeft(content, "\n")), 0o644); err != nil {
+		t.Fatalf("failed to write profile test file %q: %v", path, err)
+	}
+}
+
+func TestFSRepository(t *testing.T) {
+	ctx := context.Background()
+
+	t.Run("loads valid profiles from nested directories", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/provider/nested.yaml": profileMapFile(`
+id: nested-profile
+endpoint: http://localhost:8000/v1
+model: nested-model
+temperature: 0.1
+`),
+		}, "profiles")
+
+		p, err := repo.GetProfile(ctx, "nested-profile")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "nested-profile" || p.Model != "nested-model" {
+			t.Fatalf("unexpected profile: %+v", p)
+		}
+	})
+
+	t.Run("rejects unknown YAML fields", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/unknown.yaml": profileMapFile(`
+id: unknown-profile
+endpoint: http://localhost:8000/v1
+model: model
+unknown: value
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "unknown-profile")
+		if !errors.Is(err, ErrInvalidYAML) {
+			t.Fatalf("expected ErrInvalidYAML, got %v", err)
+		}
+	})
+
+	t.Run("rejects raw api_key in selected profile", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/raw.yaml": profileMapFile(`
+id: raw-profile
+endpoint: http://localhost:8000/v1
+model: model
+api_key: secret
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "raw-profile")
+		if !errors.Is(err, ErrRawAPIKeyNotAllowed) {
+			t.Fatalf("expected ErrRawAPIKeyNotAllowed, got %v", err)
+		}
+	})
+
+	t.Run("ignores raw api_key in non-selected profiles", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/raw.yaml": profileMapFile(`
+id: raw-profile
+endpoint: http://localhost:8000/v1
+model: model
+api_key: secret
+`),
+			"profiles/valid.yaml": profileMapFile(`
+id: valid-profile
+endpoint: http://localhost:8000/v1
+model: model
+`),
+		}, "profiles")
+
+		p, err := repo.GetProfile(ctx, "valid-profile")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.ID != "valid-profile" {
+			t.Fatalf("unexpected profile: %+v", p)
+		}
+	})
+
+	t.Run("rejects duplicate IDs within one source", func(t *testing.T) {
+		repo := NewFSRepository(fstest.MapFS{
+			"profiles/a.yaml": profileMapFile(`
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: first
+`),
+			"profiles/nested/b.yaml": profileMapFile(`
+id: duplicate-profile
+endpoint: http://localhost:8000/v1
+model: second
+`),
+		}, "profiles")
+
+		_, err := repo.GetProfile(ctx, "duplicate-profile")
+		if !errors.Is(err, ErrInvalidProfile) {
+			t.Fatalf("expected ErrInvalidProfile, got %v", err)
+		}
+		for _, want := range []string{"duplicate execution profile id", "a.yaml", "nested/b.yaml"} {
+			if !strings.Contains(err.Error(), want) {
+				t.Fatalf("expected error to contain %q, got %v", want, err)
+			}
+		}
+	})
+}
+
+func TestOverlayRepository(t *testing.T) {
+	ctx := context.Background()
+	primaryProfile := &domain.ExecutionProfile{ID: "shared", Endpoint: "http://primary", Model: "primary"}
+	fallbackProfile := &domain.ExecutionProfile{ID: "shared", Endpoint: "http://fallback", Model: "fallback"}
+
+	t.Run("returns primary matches before fallback matches", func(t *testing.T) {
+		repo := NewOverlayRepository(
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": primaryProfile}},
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+		)
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "primary" {
+			t.Fatalf("expected primary profile, got %+v", p)
+		}
+	})
+
+	t.Run("falls back on primary not found", func(t *testing.T) {
+		repo := NewOverlayRepository(
+			staticProfileRepo{},
+			staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+		)
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "fallback" {
+			t.Fatalf("expected fallback profile, got %+v", p)
+		}
+	})
+
+	t.Run("does not fall back after primary load errors", func(t *testing.T) {
+		for _, tc := range []struct {
+			name string
+			err  error
+		}{
+			{name: "invalid yaml", err: ErrInvalidYAML},
+			{name: "invalid profile", err: ErrInvalidProfile},
+			{name: "raw api key", err: ErrRawAPIKeyNotAllowed},
+		} {
+			t.Run(tc.name, func(t *testing.T) {
+				repo := NewOverlayRepository(
+					staticProfileRepo{err: tc.err},
+					staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}},
+				)
+
+				_, err := repo.GetProfile(ctx, "shared")
+				if !errors.Is(err, tc.err) {
+					t.Fatalf("expected %v, got %v", tc.err, err)
+				}
+			})
+		}
+	})
+
+	t.Run("returns not found when both sources miss", func(t *testing.T) {
+		repo := NewOverlayRepository(staticProfileRepo{}, staticProfileRepo{})
+
+		_, err := repo.GetProfile(ctx, "missing")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+
+	t.Run("nil primary uses fallback", func(t *testing.T) {
+		repo := NewOverlayRepository(nil, staticProfileRepo{profiles: map[string]*domain.ExecutionProfile{"shared": fallbackProfile}})
+
+		p, err := repo.GetProfile(ctx, "shared")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.Model != "fallback" {
+			t.Fatalf("expected fallback profile, got %+v", p)
+		}
+	})
+
+	t.Run("nil fallback returns not found after primary miss", func(t *testing.T) {
+		repo := NewOverlayRepository(staticProfileRepo{}, nil)
+
+		_, err := repo.GetProfile(ctx, "missing")
+		if !errors.Is(err, ErrProfileNotFound) {
+			t.Fatalf("expected ErrProfileNotFound, got %v", err)
+		}
+	})
+}
+
+func profileMapFile(content string) *fstest.MapFile {
+	return &fstest.MapFile{Data: []byte(strings.TrimLeft(content, "\n"))}
+}
+
+type staticProfileRepo struct {
+	profiles map[string]*domain.ExecutionProfile
+	err      error
+}
+
+func (r staticProfileRepo) GetProfile(_ context.Context, id string) (*domain.ExecutionProfile, error) {
+	if r.err != nil {
+		return nil, r.err
+	}
+	if p, ok := r.profiles[id]; ok {
+		cp := *p
+		return &cp, nil
+	}
+	return nil, ErrProfileNotFound
+}

internal/profile/testdata/valid_with_api_key_env.yaml

@@ -2,6 +2,7 @@ id: local-secure
 endpoint: http://localhost:8000/v1
 model: gpt-4o-mini
 api_key_env: SCRIPTORIUM_API_KEY
+service_tier: priority
 reasoning_effort: medium
 extra_params:
   provider: local

internal/prompt/go_renderer.go

@@ -6,7 +6,9 @@ import (
 	"errors"
 	"fmt"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
+	"strings"
 	"text/template"
+	"unicode/utf8"
 )
 
 var (
@@ -50,6 +52,11 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 		},
 	}
 
+	sessionID, err := renderSessionID(definition.SessionID, funcs, vars)
+	if err != nil {
+		return nil, err
+	}
+
 	var renderedMessages []domain.RenderedMessage
 
 	for i, tmplMsg := range definition.Templates {
@@ -77,10 +84,42 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 		renderedMessages = append(renderedMessages, domain.RenderedMessage{
 			Role:         tmplMsg.Role,
 			Content:      buf.String(),
+			CacheControl: cloneCacheControl(tmplMsg.CacheControl),
 		})
 	}
 
 	return &domain.RenderedPrompt{
+		SessionID: sessionID,
 		Messages:  renderedMessages,
 	}, nil
 }
+
+func renderSessionID(raw string, funcs template.FuncMap, vars map[string]string) (string, error) {
+	if strings.TrimSpace(raw) == "" {
+		return "", nil
+	}
+
+	tmpl, err := template.New("session_id").Funcs(funcs).Option("missingkey=error").Parse(raw)
+	if err != nil {
+		return "", fmt.Errorf("%w: session_id: %v", ErrInvalidTemplate, err)
+	}
+
+	var buf bytes.Buffer
+	if err := tmpl.Execute(&buf, vars); err != nil {
+		return "", fmt.Errorf("%w: session_id: %w", ErrRenderFailure, err)
+	}
+
+	sessionID := strings.TrimSpace(buf.String())
+	if n := utf8.RuneCountInString(sessionID); n > domain.SessionIDMaxLength {
+		return "", fmt.Errorf("%w: session_id length %d exceeds maximum %d", ErrRenderFailure, n, domain.SessionIDMaxLength)
+	}
+	return sessionID, nil
+}
+
+func cloneCacheControl(in *domain.CacheControl) *domain.CacheControl {
+	if in == nil {
+		return nil
+	}
+	out := *in
+	return &out
+}

internal/prompt/renderer_test.go

@@ -3,6 +3,7 @@ package prompt
 import (
 	"context"
 	"errors"
+	"strings"
 	"testing"
 
 	"gitea.maximumdirect.net/eric/scriptorium/internal/domain"
@@ -78,6 +79,66 @@ func TestGoRenderer_Render(t *testing.T) {
 		}
 	})
 
+	t.Run("copying cache control to rendered messages", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{
+					Role:    "system",
+					Content: "You are concise.",
+					CacheControl: &domain.CacheControl{
+						Type: domain.CacheControlEphemeral,
+						TTL:  "1h",
+					},
+				},
+				{Role: "user", Content: "Analyze this: {{input \"transcript\"}}"},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if len(res.Messages) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(res.Messages))
+		}
+		if res.Messages[0].CacheControl == nil {
+			t.Fatal("expected rendered cache control")
+		}
+		if res.Messages[0].CacheControl.Type != domain.CacheControlEphemeral {
+			t.Fatalf("unexpected cache control type: %q", res.Messages[0].CacheControl.Type)
+		}
+		if res.Messages[0].CacheControl.TTL != "1h" {
+			t.Fatalf("unexpected cache control ttl: %q", res.Messages[0].CacheControl.TTL)
+		}
+		if res.Messages[1].CacheControl != nil {
+			t.Fatalf("expected no cache control on second message, got %#v", res.Messages[1].CacheControl)
+		}
+	})
+
+	t.Run("rendered cache control does not alias source template", func(t *testing.T) {
+		source := &domain.CacheControl{Type: domain.CacheControlEphemeral, TTL: "1h"}
+		def := &domain.PromptDefinition{
+			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "You are concise.", CacheControl: source},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.Messages[0].CacheControl == source {
+			t.Fatal("expected rendered cache control to be cloned")
+		}
+
+		res.Messages[0].CacheControl.TTL = ""
+		if source.TTL != "1h" {
+			t.Fatalf("source cache control was mutated, ttl=%q", source.TTL)
+		}
+	})
+
 	t.Run("accessing vars", func(t *testing.T) {
 		def := &domain.PromptDefinition{
 			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},
@@ -95,6 +156,78 @@ func TestGoRenderer_Render(t *testing.T) {
 		}
 	})
 
+	t.Run("rendering session id from vars", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "  {{ .session_id }}  ",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, map[string]string{
+			"tone":       "concise",
+			"session_id": "agent-session-123",
+		})
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.SessionID != "agent-session-123" {
+			t.Fatalf("unexpected session id: %q", res.SessionID)
+		}
+	})
+
+	t.Run("empty rendered session id is omitted", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "   ",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		res, err := renderer.Render(ctx, def, inputs, vars)
+		if err != nil {
+			t.Fatalf("unexpected error: %v", err)
+		}
+		if res.SessionID != "" {
+			t.Fatalf("expected empty session id, got %q", res.SessionID)
+		}
+	})
+
+	t.Run("missing session id var fails rendering", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "{{ .session_id }}",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, vars)
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+	})
+
+	t.Run("too long rendered session id fails rendering", func(t *testing.T) {
+		def := &domain.PromptDefinition{
+			SessionID: "{{ .session_id }}",
+			Inputs:    []domain.PromptInput{{Name: "transcript", Required: true}},
+			Templates: []domain.PromptMessageTemplate{
+				{Role: "system", Content: "Speak in a {{.tone}} tone."},
+			},
+		}
+
+		_, err := renderer.Render(ctx, def, inputs, map[string]string{
+			"tone":       "concise",
+			"session_id": strings.Repeat("x", domain.SessionIDMaxLength+1),
+		})
+		if !errors.Is(err, ErrRenderFailure) {
+			t.Fatalf("expected ErrRenderFailure, got %v", err)
+		}
+	})
+
 	t.Run("inserting required input artifact", func(t *testing.T) {
 		def := &domain.PromptDefinition{
 			Inputs: []domain.PromptInput{{Name: "transcript", Required: true}},