Add a roadmap and implementation plan for an initial public library package

.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 +1,4 @@
+Please carefully review the relevant documents in `docs/policy` before making any changes to this repository.
+ - `development.md` defines the contributor workflow for this application.
+ - `architecture.md` provides the canonical high-level architecture policy for this repository, and should be reviewed before writing or changing any code.
+ - `documentation.md` provides the canonical documentation policy for this repository, and should be reviewed before writing or changing any documentation.

docs/cli.md

@@ -32,6 +32,8 @@ Integration references:
   - an effective `prompt_dir` and `profile_dir` (from flags or config)
 - `serve` requires an effective `prompt_dir` and `profile_dir` (from flags or config).
 - Positional arguments are rejected.
+- Prompt cache control is configured in prompt YAML (`messages[].cache_control`), not with CLI flags.
+- Provider-specific `reasoning_effort` and `extra_params` are configured in profile YAML or HTTP model overrides, not with CLI flags.
 
 ## Flag Reference
 
@@ -56,6 +58,11 @@ Integration references:
 - `--top-p <float>`: runtime top-p override.
 - `--timeout <duration>`: runtime timeout override (Go duration syntax, for example `30s`, `2m`).
 
+Numeric runtime override flags are presence-aware:
+
+- omitted numeric flags preserve the selected profile/default value
+- explicit zero values override the selected profile/default value (`--temperature 0`, `--max-tokens 0`, `--top-p 0`, `--timeout 0s`)
+
 ### `scriptorium render`
 
 - Supports the same flags as `run`, except:
@@ -82,6 +89,7 @@ Notes:
 
 - `--input name=path` maps prompt input names to local file paths.
 - `--var name=value` maps template variable names to values.
+- If a prompt defines `session_id: "{{ .session_id }}"`, pass the OpenRouter sticky-routing value with `--var session_id=<value>`.
 - Both flags can be repeated.
 - Both flags also support comma-separated batches, for example:
   - `--input transcript=./t.md,glossary=./g.yml`
@@ -93,6 +101,7 @@ Notes:
 - Writes generated artifact content to stdout by default.
 - Writes generated artifact content to `--out` when provided.
 - Prints run summary metadata to stderr on success.
+- Appends `cached_tokens=<n> cache_write_tokens=<n>` to the summary only when the provider reports non-zero cache usage.
 - Prints errors to stderr on failure.
 
 `render`:

docs/config.md

@@ -104,6 +104,7 @@ Field reference:
 - `version` (required): prompt version.
 - `default_profile` (optional): profile ID used when request does not provide `profile_id`.
 - `description` (optional): prompt description.
+- `session_id` (optional): Go-template string for OpenRouter sticky-routing `session_id`; rendered from request vars.
 - `inputs` (optional list): expected named inputs.
 - `messages` (required list): prompt message templates.
 - `output` (required object): output contract.
@@ -119,6 +120,7 @@ Field reference:
 
 - `role` (required)
 - `content` or `content_file` (exactly one is required)
+- `cache_control` (optional object): provider prompt-cache metadata for this message
 
 Message rules:
 
@@ -128,6 +130,35 @@ Message rules:
 - Prompt decoding is strict; unknown YAML fields are rejected.
 - Duplicate prompt IDs are invalid. If multiple files declare the requested prompt ID, Scriptorium fails instead of choosing one.
 
+`messages[].cache_control` fields:
+
+- `type` (required when `cache_control` is present): currently only `ephemeral`.
+- `ttl` (optional): currently only `1h`; omitted from outbound requests when unset.
+
+Example cache-controlled message:
+
+```yaml
+messages:
+  - role: system
+    content_file: ./stable_context.md
+    cache_control:
+      type: ephemeral
+      ttl: 1h
+  - role: user
+    content: |
+      {{input "transcript"}}
+```
+
+Use cache control on stable reusable prompt content. Dynamic per-run inputs before the cache-controlled message change the provider cache key.
+
+Example prompt-level session ID:
+
+```yaml
+session_id: "{{ .session_id }}"
+```
+
+When configured, `session_id` is rendered with the same variable context as messages. The rendered value is trimmed, omitted when empty, and rejected if longer than 256 characters. CLI callers pass the value through `--var session_id=<value>`; HTTP callers pass it through `"vars": {"session_id": "<value>"}`.
+
 `output` fields:
 
 - `format` (required): `text`, `markdown`, or `json`.
@@ -158,6 +189,11 @@ top_p: 1.0
 timeout_seconds: 90
 api_key_env: SCRIPTORIUM_API_KEY
 service_tier: priority
+reasoning_effort: medium
+extra_params:
+  provider_route: primary
+  provider_options:
+    retry_budget: 2
 ```
 
 Field reference:
@@ -170,9 +206,9 @@ Field reference:
 - `top_p` (optional): range `0..1`
 - `timeout_seconds` (optional): `>= 0`
 - `service_tier` (optional): provider-specific request tier such as OpenRouter `flex` or `priority`
-- `reasoning_effort` (optional)
+- `reasoning_effort` (optional): serialized as top-level `reasoning_effort` in outbound chat-completions requests
 - `api_key_env` (optional)
-- `extra_params` (optional map of strings)
+- `extra_params` (optional map): JSON-compatible provider-specific parameters. Values may be strings, numbers, booleans, objects, or arrays.
 
 Profile rules:
 
@@ -180,11 +216,14 @@ Profile rules:
 - Raw `api_key` is rejected; use `api_key_env`.
 - If `api_key_env` is set, that environment variable must be set when preparing/running.
 - Duplicate profile IDs are invalid. If multiple files declare the requested profile ID, Scriptorium fails instead of choosing one.
+- `extra_params` keys must not be empty and must not collide with reserved outbound request fields: `model`, `session_id`, `messages`, `temperature`, `max_tokens`, `top_p`, `service_tier`, `reasoning_effort`, or `response_format`.
 
 Current outbound request behavior:
 
-- The OpenAI-compatible client currently serializes: `model`, `messages`, `temperature`, `max_tokens`, `top_p`, `service_tier`, and optional `response_format` for `json_schema` prompts.
-- `reasoning_effort` and `extra_params` are parsed and carried in effective settings, but are not currently serialized into outbound chat-completions requests.
+- The OpenAI-compatible client currently serializes: `model`, optional `session_id`, `messages`, `temperature`, `max_tokens`, `top_p`, `service_tier`, `reasoning_effort`, optional `response_format` for `json_schema` prompts, and `extra_params`.
+- `extra_params` are flattened into provider-specific top-level JSON request fields. They are not wrapped in an `extra_params` object on the outbound provider request.
+- Messages without `cache_control` serialize with string `content`.
+- Messages with `cache_control` serialize as a single text content-block array containing `cache_control`.
 
 ## Schema Behavior
 

docs/integrations/http-api.md

@@ -50,7 +50,10 @@ Copyable request example file:
     "reasoning_effort": "medium",
     "api_key_env": "SCRIPTORIUM_API_KEY",
     "extra_params": {
-      "route": "primary"
+      "route": "primary",
+      "provider_options": {
+        "retry_budget": 2
+      }
     }
   },
   "include_raw_output": false
@@ -67,6 +70,14 @@ Input reference types currently supported by runtime artifact loading:
 - `file`
 - `inline`
 
+Model override notes:
+
+- Numeric model override fields distinguish omitted values from explicit zero values. For example, omitting `temperature` preserves the selected profile/default value, while `"temperature": 0` explicitly sets the effective temperature to zero.
+- `extra_params` accepts JSON-compatible values: strings, numbers, booleans, objects, and arrays.
+- `extra_params` are passed through effective model metadata and flattened into top-level provider request fields by the OpenAI-compatible client.
+- `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`.
+- Raw API-key values are not accepted. Use `api_key_env` to name an environment variable.
+
 ## Strict JSON Rules
 
 Request decoding uses strict JSON field checks:
@@ -119,7 +130,10 @@ Response shape:
       "reasoning_effort": "medium",
       "api_key_env": "SCRIPTORIUM_API_KEY",
       "extra_params": {
-        "route": "primary"
+        "route": "primary",
+        "provider_options": {
+          "retry_budget": 2
+        }
       }
     },
     "input_hashes": {
@@ -128,7 +142,9 @@ Response shape:
     "usage": {
       "prompt_tokens": 11,
       "completion_tokens": 22,
-      "total_tokens": 33
+      "total_tokens": 33,
+      "cached_tokens": 0,
+      "cache_write_tokens": 0
     },
     "start_time": "2026-05-04T12:00:00Z",
     "end_time": "2026-05-04T12:00:01Z",
@@ -142,6 +158,8 @@ Response shape:
 
 `raw_model_output` is omitted by default.
 
+`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.
+
 To include it, send:
 
 - `"include_raw_output": true`

docs/integrations/openai-compatible-chat.md

@@ -26,15 +26,85 @@ Example:
 Serialized JSON fields:
 
 - `model` (required after fallback resolution)
-- `messages` (role/content pairs from rendered prompt)
-- `temperature` (only when non-zero)
-- `max_tokens` (only when non-zero)
-- `top_p` (only when non-zero)
+- `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
@@ -72,6 +142,7 @@ 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
@@ -82,6 +153,13 @@ Expected successful response shape (subset used):
 - `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:
 
@@ -99,10 +177,7 @@ Malformed responses return `ErrMalformedResponse`.
 
 ## Unsupported Or Non-Serialized Fields
 
-The following fields may exist in profile/effective settings but are not currently serialized into outbound chat-completions payloads:
-
-- `reasoning_effort`
-- `extra_params`
+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.
 

docs/internal/adapters.md

@@ -22,11 +22,13 @@ CLI adapter:
 
 - Input: process args, filesystem config/assets, environment.
 - Output: exit code, stdout artifact/prepared output, stderr summaries/errors.
+- `run` summaries include cache usage counters only when either parsed cache counter is non-zero.
 
 HTTP adapter:
 
 - Input: JSON request body (`runRequestDTO`).
 - Output: JSON success/error body with mapped status codes.
+- Success metadata includes token usage plus cache usage counters.
 
 Filesystem repositories:
 
@@ -67,6 +69,8 @@ Primary app settings consumed by adapters:
 Execution profile/request settings used through runner:
 
 - `endpoint`, `model`, `temperature`, `max_tokens`, `top_p`, `timeout_seconds`, `service_tier`, `api_key_env`, `reasoning_effort`, `extra_params`
+- CLI and HTTP request adapters preserve caller intent for numeric runtime overrides. Omitted values remain absent; explicit zero values are mapped as explicit overrides.
+- HTTP `extra_params` accepts JSON-compatible values and maps them to domain request overrides without provider-specific adapter logic.
 
 ## External Dependencies
 
@@ -93,6 +97,13 @@ Artifact refs:
 LLM adapter:
 
 - endpoint appends `/chat/completions`.
+- rendered messages without cache control serialize with string `content`.
+- rendered messages with cache control serialize as one text content block with `cache_control`.
+- non-empty `reasoning_effort` serializes as a top-level provider request field.
+- `extra_params` flatten into provider-specific top-level JSON request fields.
+- reserved `extra_params` keys are rejected before the provider call: `model`, `session_id`, `messages`, `temperature`, `max_tokens`, `top_p`, `service_tier`, `reasoning_effort`, and `response_format`.
+- empty `extra_params` keys and values that cannot be JSON-encoded are rejected before the provider call.
+- compatible cache usage response fields are parsed into domain token usage.
 - non-2xx responses map to request failure errors.
 - malformed responses (including missing/empty first choice content) are errors.
 
@@ -141,4 +152,5 @@ Behavior highlights:
 - Adapter packages do not own runner decision logic.
 - External request/response strictness is part of contract stability.
 - Prepared-render output never includes resolved API key values.
-- Outbound OpenAI-compatible request includes only currently serialized fields (`model`, `messages`, optional `temperature`, `max_tokens`, `top_p`, optional `service_tier`, optional `response_format`).
+- Outbound OpenAI-compatible request includes currently serialized first-class fields (`model`, optional `session_id`, `messages`, optional `temperature`, `max_tokens`, `top_p`, optional `service_tier`, optional `reasoning_effort`, optional `response_format`) plus validated `extra_params` flattened as provider-specific top-level fields.
+- Outbound cache control is message-level only; no top-level cache-control field is serialized.

docs/internal/runner.md

@@ -103,16 +103,27 @@ Validation content failures are not run errors:
    - built-in execution defaults
    - selected profile values
    - request overrides
+   - request numeric overrides are presence-aware, so omitted values preserve the current effective value and explicit zero values override it
 6. verify required `api_key_env` environment variable:
    - missing/empty env value returns an invalid request with `ErrAPIKeyEnvMissing`
    - only the environment-variable name is retained; secret value is never returned
 7. resolve output contract and structured-output schema payload when `json_schema` mode is active.
 8. read input artifacts.
-9. render prompt messages.
+9. render prompt messages, including any normalized message cache-control metadata.
 10. compute prompt/input/render hashes and return `PreparedRun`.
 
+`rendered_prompt_hash` includes cache-control metadata when present because it affects the outbound provider request. Prompts without cache control keep the role/content hash behavior.
+
 `Prepare` does not call the LLM.
 
+Runtime target notes:
+
+- Profile `extra_params` and request `extra_params` carry JSON-compatible values through prepared output, run metadata, and `domain.GenerateRequest.Target`.
+- The OpenAI-compatible client serializes non-empty `reasoning_effort` as a top-level provider request field.
+- The OpenAI-compatible client flattens `extra_params` into provider-specific top-level JSON request fields.
+- Empty `extra_params` keys, reserved outbound field names, and values that cannot be JSON-encoded fail before the provider request.
+- Resolved API-key values are never stored in `PreparedRun`, `RunResult`, logs, or HTTP responses.
+
 ## Run Flow
 
 `Run` performs:
@@ -123,7 +134,7 @@ Validation content failures are not run errors:
 4. build output artifact content type from output format.
 5. validate output.
 6. optionally attempt bounded repair when repairer is injected and contract allows it.
-7. return `RunResult` with artifact, raw output, validation, hashes, profile/model metadata, usage, and timestamps.
+7. return `RunResult` with artifact, raw output, validation, hashes, profile/model metadata, token/cache usage, and timestamps.
 
 ## Repair Hook Boundary
 

docs/policy/documentation.md

@@ -2,12 +2,13 @@
 
 ## Purpose
 
-Project documentation must help four audiences:
+Project documentation must help five audiences:
 
 1. users who need to run the application;
 2. administrators/operators who need to configure and operate it;
 3. developers who need to understand and change it safely;
-4. LLM coding agents that need clear scope, boundaries, and invariants.
+4. LLM coding agents that need clear scope, boundaries, and invariants;
+5. developers and LLM coding agents integrating this project from another codebase.
 
 Docs should be accurate, concise, task-oriented, and organized by audience. Prefer links to canonical docs over repetition.
 
@@ -42,11 +43,14 @@ Canonical homes:
 
 - project purpose and quickstart: `README.md`
 - development principles: `docs/policy/architecture.md`
+- public HTTP API reference: `docs/api.md`
 - configuration reference: `docs/config.md`
 - CLI reference: `docs/cli.md`
 - operations and recovery: `docs/operations.md`
 - troubleshooting: `docs/troubleshooting.md`
+- public API/package consumer guidance: `docs/consumers/`
 - implemented internals: `docs/internal/`
+- external protocol, service, and file-format contracts: `docs/integrations/`
 - future work: `docs/roadmap/`
 - contributor workflow: `docs/policy/development.md`
 - copyable examples: `examples/`
@@ -106,7 +110,7 @@ Recommended:
 - `examples/`
 - `docs/policy/development.md`
 
-### Modular, staged, service-oriented, or orchestration application
+### Modular, service-oriented, or orchestration application
 
 Required:
 - `docs/cli.md`, if CLI-based
@@ -119,6 +123,31 @@ Recommended:
 - `docs/troubleshooting.md`
 - validated examples under `examples/`
 
+### Public HTTP API service
+
+Required:
+- `docs/api.md`
+- `docs/cli.md`, if CLI-based
+- `docs/config.md`, if config-driven
+- `docs/operations.md`
+- `docs/internal/`
+- `docs/policy/development.md`
+
+Recommended:
+- `docs/troubleshooting.md`
+- `docs/consumers/`, for task-oriented client integration guides
+- `docs/integrations/`, for upstream/downstream service contracts
+- validated examples under `examples/`
+
+### Project with public packages or consumer APIs
+
+Required:
+- `docs/consumers/api.md`
+- one `docs/consumers/pkg-<name>.md` file per public package, if public packages exist
+
+Recommended:
+- copyable consumer examples under `examples/`, if practical
+
 ## Required Documents
 
 ### README.md
@@ -159,7 +188,35 @@ It should include:
 - architectural invariants;
 - explicit non-goals, if useful.
 
-For small projects, this file may be brief. It may simply state that the project is intentionally narrow, monolithic, and dependency-light.
+Notably, this file should prescribe a core development *policy* that should remain unchanged as the application evolves. It is not a place for details (e.g., CLI flags) that could change over time.
+
+The contents of `architecture.md` should be trim and concise. LLMs may be directed to review it routinely via AGENTS.md, CLAUDE.md, or similar.
+
+### docs/api.md
+
+**Audience:** external HTTP API consumers, developers, LLM coding agents integrating by HTTP
+
+Required for projects whose primary public interface is HTTP.
+
+`docs/api.md` is the canonical public HTTP API contract. It should be normative for external consumers and should not be duplicated by README, operations docs, consumer guides, or integration docs.
+
+It should include:
+
+1. base URL conventions;
+2. authentication and authorization behavior, if implemented;
+3. response envelope;
+4. supported media types and content negotiation behavior;
+5. shared query parameters;
+6. endpoint reference grouped by route family;
+7. request parameters and validation rules;
+8. response fields, units, nullability, and optionality;
+9. error response shape and status codes;
+10. pagination, caching, rate-limit, idempotency, and retry behavior, if implemented;
+11. compact request and response examples.
+
+It must document only implemented endpoints and behavior. Planned endpoints, proposed fields, future filters, and experimental response shapes belong only under `docs/roadmap/`.
+
+For HTTP API projects, `docs/consumers/` may provide task-oriented client integration guides, but those guides should link to `docs/api.md` for the authoritative endpoint contract.
 
 ### docs/policy/development.md
 
@@ -175,7 +232,7 @@ It should include:
 - dependency policy;
 - how to add config fields;
 - how to add CLI flags;
-- how to add stages/modules/adapters, if applicable;
+- how to add modules or adapters, if applicable;
 - how to update examples;
 - documentation update expectations.
 
@@ -216,7 +273,7 @@ Explain when commands are useful, not just their syntax.
 
 **Audience:** administrators, operators
 
-Required for applications that maintain state, support resume behavior, run multiple stages, write durable artifacts, use remote storage, or require recovery procedures.
+Required for applications that maintain state, support resume behavior, run multi-step workflows, write durable artifacts, use remote storage, or require recovery procedures.
 
 It should cover:
 
@@ -244,11 +301,40 @@ Each entry should include:
 - safe fix;
 - relevant links.
 
+### docs/consumers/
+
+**Audience:** developers and LLM coding agents integrating this project from another codebase
+
+Required for projects with public packages, SDKs, client APIs, plugin APIs, or other application-facing integration surfaces.
+
+This directory describes how an external codebase should consume the project's public API. It should be task-oriented and copyable where useful. It is not the place for internal implementation details or operator procedures.
+
+For projects whose public API is HTTP, `docs/consumers/` is not required, and it should not duplicate the endpoint reference in `docs/api.md`. If present, it may provide practical integration workflows, client-specific examples, or migration notes that link back to `docs/api.md`.
+
+`docs/consumers/api.md` should provide the consumer-facing overview and primary implementation workflow. It should include:
+
+1. intended consumer audience and use cases;
+2. required inputs supplied by operators or deployment configuration;
+3. recommended public package or API workflow;
+4. minimal copyable example;
+5. consumer responsibilities and boundaries;
+6. retry, idempotency, or status behavior, if applicable;
+7. links to package-specific docs and canonical integration contracts.
+
+Package-specific docs should be named `pkg-<name>.md` and should include:
+
+1. import path;
+2. intended use cases;
+3. primary types and functions needed by consumers;
+4. minimal examples;
+5. validation, error, retry, and boundary behavior;
+6. links to canonical file-format or wire-protocol contracts.
+
 ### docs/internal/
 
 **Audience:** developers, LLM coding agents
 
-Required for modular, staged, service-oriented, or orchestration projects.
+Required for modular, service-oriented, or orchestration projects.
 
 This directory describes implemented internal components. It is not the roadmap.
 
@@ -289,7 +375,9 @@ Roadmap docs should not be confused with current behavior.
 
 Required for projects that depend on external CLIs, APIs, services, protocols, or file formats where the integration contract is important to maintain.
 
-This directory contains concise, versioned reference notes for external integration contracts. It should document only the parts of the external system that this project actually uses.
+This directory contains concise, versioned reference notes for external integration contracts. It should document only the parts of the external system that this project actually uses or exposes.
+
+For public HTTP API services, `docs/integrations/` should document upstream, downstream, storage, protocol, or runtime contracts that the service depends on or bridges. It should not become a second copy of the public HTTP endpoint reference; that belongs in `docs/api.md`.
 
 Use one file per integration where useful.
 
@@ -346,8 +434,10 @@ Before merging documentation changes, verify:
 
 - README is concise and orientation-focused.
 - `docs/policy/architecture.md` describes development principles.
+- `docs/api.md` is the canonical HTTP contract for HTTP API services.
 - Future work appears only under `docs/roadmap/`.
 - User-facing docs avoid unnecessary internals.
+- Consumer-facing docs explain public APIs without duplicating HTTP endpoint or integration contracts.
 - Developer-facing docs preserve boundaries and invariants.
 - Config examples match the schema.
 - CLI examples match real commands and flags.

docs/roadmap/implementation.md

@@ -0,0 +1,321 @@
+# Library API Implementation Plan
+
+This plan implements the target state in `docs/roadmap/library.md`.
+
+Audience: LLM coding agents implementing the feature in order. Follow `docs/policy/architecture.md`, `docs/policy/development.md`, and `docs/policy/documentation.md` before changing code.
+
+## Constraints
+
+- Add a public root package named `scriptorium`; keep existing `internal/*` packages internal.
+- Define public facade types and convert to/from internal domain types. Do not alias internal domain types as the public API.
+- Do not rewire CLI or HTTP through the public facade in this implementation.
+- Preserve current CLI, HTTP, prompt/profile loading, validation, secret-handling, and outbound LLM behavior.
+- Do not add dependencies.
+- Keep each stage passing `go test ./...` before moving to the next stage.
+
+## Stage 1: Public Types, Engine Construction, And Prepare
+
+Goal: make prompt preparation usable from an imported root package without calling an LLM.
+
+### Public Package
+
+Create Go files at the module root using:
+
+```go
+package scriptorium
+```
+
+Expose:
+
+```go
+type Engine struct { /* unexported fields */ }
+
+type Config struct {
+	PromptDir  string
+	ProfileDir string
+	SchemaDir  string
+	Timeout    time.Duration
+	HTTPClient *http.Client
+}
+
+type Option func(*engineOptions) error
+
+func NewEngine(cfg Config, opts ...Option) (*Engine, error)
+func (e *Engine) Prepare(ctx context.Context, req RunRequest) (*PreparedRun, error)
+```
+
+Construction rules:
+
+- `PromptDir` and `ProfileDir` are required.
+- `SchemaDir` defaults to the same built-in default used by app config.
+- `Timeout`, when non-zero, configures the default OpenAI-compatible client timeout.
+- `HTTPClient`, when non-nil, is used by the default OpenAI-compatible client.
+- `NewEngine` wires the same internal components used by CLI/HTTP: filesystem prompt/profile repositories, composite artifact reader, Go template renderer, standard validator, and OpenAI-compatible LLM client.
+- Return public `ErrInvalidConfig` for invalid engine configuration.
+
+### Public Types
+
+Define public facade types with exported fields:
+
+- `RunRequest`
+- `PreparedRun`
+- `ArtifactRef`
+- `Artifact`
+- `ExecutionTarget`
+- `ExecutionTargetOverride`
+- `ExecutionTargetPresence`
+- `OutputContract`
+- `ValidationResult`
+- `TokenUsage`
+- `RenderedMessage`
+- `CacheControl`
+- `StructuredOutputSpec`
+
+Use the same enum string values as internal domain types for formats, validation modes, validation statuses, artifact ref types, cache-control type, and structured-output type.
+
+Required request shape:
+
+```go
+type RunRequest struct {
+	PromptID      string
+	PromptVersion string
+	ProfileID     string
+	Inputs        map[string]ArtifactRef
+	Vars          map[string]string
+	Execution     *ExecutionTargetOverride
+	Validation    *OutputContract
+	Metadata      map[string]string
+}
+```
+
+`ExecutionTargetOverride` must preserve numeric override presence using pointer fields:
+
+```go
+Temperature    *float64
+MaxTokens      *int
+TopP           *float64
+TimeoutSeconds *int
+```
+
+`PreparedRun` should include the same user-observable fields as internal `domain.PreparedRun`, but should not expose internal-only target presence metadata.
+
+### Input Helpers
+
+Expose:
+
+```go
+func File(path string) ArtifactRef
+func Inline(body string) ArtifactRef
+func InlineWithURI(uri string, body string) ArtifactRef
+```
+
+Mapping:
+
+- `File(path)` maps to artifact type `file` with `URI: path`.
+- `Inline(body)` maps to artifact type `inline` with `Body: body`.
+- `InlineWithURI(uri, body)` maps to artifact type `inline` with both fields set.
+
+### Conversion Layer
+
+Implement unexported conversion helpers in the public package:
+
+- public run request to internal `domain.RunRequest`
+- internal `domain.PreparedRun` to public `PreparedRun`
+- internal artifacts/messages/contracts/validation/usage/structured-output to public equivalents
+- public execution override to internal `domain.ExecutionTargetOverride`
+
+Conversions must deep-copy maps and slices that cross the public/internal boundary.
+
+### Tests
+
+Add root package tests.
+
+Required tests:
+
+- `NewEngine` rejects missing `PromptDir`.
+- `NewEngine` rejects missing `ProfileDir`.
+- `Prepare` works with `examples/config.yml` directories when passed directly through `Config`.
+- `Prepare` works with `File` input refs.
+- `Prepare` works with `Inline` input refs.
+- `Prepare` output does not expose raw API-key values or internal target presence metadata in JSON.
+- Explicit zero execution overrides survive into prepared effective settings.
+
+### Verification
+
+Run:
+
+```bash
+go test ./...
+```
+
+## Stage 2: Run, LLM Injection, And Public Errors
+
+Goal: make full execution usable and testable without real provider credentials.
+
+### Public Run Method
+
+Expose:
+
+```go
+func (e *Engine) Run(ctx context.Context, req RunRequest) (*RunResult, error)
+```
+
+`RunResult` should expose:
+
+- run ID
+- artifact
+- raw output
+- validation result
+- prompt/profile/model metadata
+- effective model params
+- input hashes
+- token/cache usage
+- start/end/duration timing
+
+Do not expose raw API-key values.
+
+### Public LLM Injection
+
+Expose:
+
+```go
+type LLMClient interface {
+	Generate(context.Context, GenerateRequest) (*GenerateResponse, error)
+}
+
+func WithLLMClient(client LLMClient) Option
+```
+
+Public `GenerateRequest` must include:
+
+- rendered prompt
+- effective execution target
+- execution target presence
+- structured-output spec
+
+Public `GenerateResponse` must include:
+
+- content
+- token usage
+
+Implementation rule:
+
+- `WithLLMClient` wraps the public client in an unexported adapter that satisfies `internal/llm.Client`.
+- The adapter converts internal generate requests to public generate requests and converts public generate responses back to internal responses.
+- A nil client passed to `WithLLMClient` returns `ErrInvalidConfig`.
+
+Default behavior:
+
+- If no custom LLM client is supplied, `NewEngine` uses `internal/llm.NewOpenAICompatibleClient`.
+- `Config.Timeout` and `Config.HTTPClient` apply only to the default OpenAI-compatible client.
+
+### Public Errors
+
+Define public sentinel errors:
+
+- `ErrInvalidConfig`
+- `ErrInvalidRequest`
+- `ErrPromptNotFound`
+- `ErrProfileNotFound`
+- `ErrPromptLoad`
+- `ErrProfileLoad`
+- `ErrArtifactLoad`
+- `ErrPromptRender`
+- `ErrLLMGenerate`
+- `ErrValidation`
+
+Public methods must map internal errors to public sentinels while preserving wrapped context. Callers must be able to use `errors.Is`.
+
+Mapping rules:
+
+- missing/invalid public engine config -> `ErrInvalidConfig`
+- internal `usecase.ErrInvalidRequest` -> `ErrInvalidRequest`
+- internal prompt not found -> `ErrPromptNotFound`
+- internal profile not found -> `ErrProfileNotFound`
+- internal prompt load errors -> `ErrPromptLoad`
+- internal profile load errors -> `ErrProfileLoad`
+- internal artifact load errors -> `ErrArtifactLoad`
+- internal prompt render errors -> `ErrPromptRender`
+- internal LLM generate errors -> `ErrLLMGenerate`
+- internal validation runtime errors -> `ErrValidation`
+
+Do not expose internal sentinel values as public API.
+
+### Tests
+
+Required tests:
+
+- `Run` succeeds with `WithLLMClient` fake and returns typed artifact, raw output, validation, metadata, and usage.
+- `Run` passes rendered prompt, effective execution target, and target presence to the injected LLM client.
+- `Run` validation failure returns a successful result with failed validation, not an error.
+- public errors support `errors.Is` for invalid request, prompt not found, profile not found, artifact load, render failure, LLM failure, and validation runtime failure.
+- nil `WithLLMClient(nil)` returns `ErrInvalidConfig`.
+- default OpenAI-compatible client can still be constructed without real provider credentials.
+
+### Verification
+
+Run:
+
+```bash
+go test ./...
+```
+
+## Stage 3: Public Documentation And Consumer Examples
+
+Goal: document implemented library behavior in canonical public-consumer docs.
+
+### Docs
+
+After Stages 1 and 2 are implemented, update:
+
+- `README.md`: add a short link to library usage without turning the README into a manual.
+- `docs/internal/adapters.md`: list the public library facade as an implemented adapter surface.
+- `docs/consumers/api.md`: describe the public consumer API at a high level.
+- `docs/consumers/pkg-scriptorium.md`: document the root package usage, types, errors, and examples.
+
+Create `docs/consumers/` if it does not exist.
+
+Do not document unimplemented future library features outside `docs/roadmap/`.
+
+### Examples
+
+Add copyable library examples only if they can be tested without real credentials.
+
+Recommended example:
+
+- `examples/go-library/prepare/main.go` or equivalent prepare-only example using `examples/` prompt/profile/fixture assets.
+
+If adding a run example, it must use an injected fake LLM client and must not require provider credentials.
+
+### Tests
+
+Required tests:
+
+- doc/example smoke coverage for any added Go example using `go test` or `go test ./...`.
+- existing CLI/HTTP tests continue to pass unchanged.
+
+### Verification
+
+Run:
+
+```bash
+go test ./...
+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
+```
+
+## Final Checks
+
+Before considering the feature complete:
+
+1. Confirm the root package can be imported as `gitea.maximumdirect.net/eric/scriptorium`.
+2. Confirm public package tests do not require real provider credentials.
+3. Confirm `go test ./...` passes.
+4. Confirm the render smoke command passes.
+5. Confirm non-roadmap docs describe only implemented behavior.
+6. Confirm no public result or rendered/prepared output exposes raw API-key values.
+7. Confirm `git diff` does not include unrelated CLI/HTTP behavior changes.

docs/roadmap/library.md

@@ -0,0 +1,134 @@
+# Library API Roadmap
+
+This roadmap defines the target behavior for making Scriptorium usable as an imported Go library while retaining the current standalone CLI and HTTP application behavior.
+
+The implementation plan for this feature lives in `docs/roadmap/implementation.md`.
+
+## Motivation
+
+Scriptorium is currently optimized for subprocess use by other applications. That contract remains useful because it is language-neutral, operationally simple, and process-isolated.
+
+For Go callers, an imported library should provide:
+
+- typed requests and results instead of stdout/stderr parsing;
+- direct `context.Context` cancellation;
+- lower overhead for repeated calls;
+- easier test integration through injected clients or fixtures;
+- direct access to prepared-run data without process management;
+- fewer integration points where secrets or output metadata can be mishandled.
+
+The library is an additional adapter surface, not a replacement for the CLI or HTTP API.
+
+## Target State
+
+Scriptorium should expose a small public Go API suitable for common embedding use cases:
+
+- construct an engine from app-level settings such as prompt, profile, and schema directories;
+- prepare a prompt request without calling an LLM;
+- run a prompt request and receive a typed result;
+- pass file and inline artifacts;
+- apply profile selection, runtime overrides, vars, validation behavior, cache-control behavior, and structured-output behavior consistently with CLI/HTTP;
+- inject a custom LLM client or HTTP client where needed;
+- preserve existing CLI and HTTP behavior by continuing to route all entry paths through the same use-case layer.
+
+The public library API should be stable, narrow, and intentionally higher-level than the current `internal/*` package layout.
+
+## Public Package Policy
+
+The public package should be the module root:
+
+```go
+import "gitea.maximumdirect.net/eric/scriptorium"
+```
+
+Recommended usage shape:
+
+```go
+engine, err := scriptorium.NewEngine(scriptorium.Config{
+	PromptDir:  "./prompts",
+	ProfileDir: "./profiles",
+	SchemaDir:  "./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("./transcript.md"),
+	},
+})
+
+result, err := engine.Run(ctx, scriptorium.RunRequest{
+	PromptID: "generic.markdown_summary",
+	Inputs: map[string]scriptorium.ArtifactRef{
+		"transcript": scriptorium.File("./transcript.md"),
+	},
+})
+```
+
+## Policy Decisions
+
+### Public Package Scope
+
+Expose a narrow root facade package and keep existing `internal/*` packages internal.
+
+Reasoning:
+
+This gives callers the workflow they need without freezing the internal architecture as public API. It also preserves the current package-boundary policy and keeps future refactoring possible.
+
+### Public Type Strategy
+
+Define public facade types and map them to internal domain types.
+
+Reasoning:
+
+Public types can be designed around caller needs and long-term stability. Internal types can continue to evolve with implementation details such as adapter metadata, validation internals, and provider-specific behavior.
+
+### CLI And HTTP Reuse
+
+Keep CLI and HTTP on current internal wiring for the initial library release. Consider migrating them to the public facade only after the facade proves stable.
+
+Reasoning:
+
+This minimizes risk to the existing subprocess and HTTP contracts while adding the new API. It also avoids forcing the first public facade to satisfy every adapter edge case immediately.
+
+### Error Surface
+
+Expose public sentinel errors or typed error categories and map internal errors to them while preserving wrapped context.
+
+Reasoning:
+
+Library callers need stable, idiomatic error checks. Mapping internal errors avoids exposing internal package paths as public compatibility promises.
+
+## Scope
+
+In scope:
+
+- Public facade package for library consumers.
+- Public request, result, prepared-run, artifact reference, execution override, validation, and config types.
+- Public constructors for common file and inline input references.
+- Public engine methods for `Prepare` and `Run`.
+- Optional dependency injection for LLM behavior and HTTP behavior.
+- Stable error behavior suitable for `errors.Is` and `errors.As`.
+- Tests proving public API behavior matches CLI/use-case behavior.
+- Documentation and examples for library usage after implementation.
+
+Out of scope for the first library release:
+
+- Making every `internal/*` package public.
+- Replacing or rewiring the CLI or HTTP adapters.
+- Adding a durable run store or workflow engine.
+- Adding broad provider-specific SDK surfaces.
+- Adding non-Go language bindings.
+- Adding global mutable configuration.
+
+## Acceptance Criteria
+
+- A Go caller can import the root module and run a prompt without invoking a subprocess.
+- A Go caller can prepare a prompt without invoking an LLM.
+- Public library behavior matches current CLI/HTTP use-case semantics for prompt/profile loading, artifact reading, rendering, validation, and model invocation.
+- Existing CLI and HTTP behavior remains unchanged.
+- Library tests use injected/fake LLM behavior and do not require real provider credentials.
+- Public documentation is concise and limited to implemented behavior once code exists.

docs/troubleshooting.md

@@ -252,6 +252,39 @@ Relevant links:
 - [Configuration reference](config.md)
 - [Operations guide](operations.md)
 
+## Prompt Cache Misses Or No Cache Usage
+
+Symptom:
+
+- CLI run summary omits `cached_tokens` / `cache_write_tokens`.
+- HTTP `metadata.usage.cached_tokens` and `metadata.usage.cache_write_tokens` are both `0`.
+- Provider cost or latency does not improve after repeated similar runs.
+
+Likely cause:
+
+- The selected prompt has no `messages[].cache_control`.
+- Dynamic per-run input appears before the cache-controlled message and changes the provider cache key.
+- The provider does not support the serialized cache-control shape for the selected model.
+- The provider imposes minimum token thresholds or cache-breakpoint limits.
+
+Diagnostic step:
+
+- Run `render --format json` and verify the intended rendered message includes `cache_control`.
+- Confirm stable reusable context appears before the cache-controlled message, with dynamic input after it.
+- Check provider docs/logs for model support, minimum token thresholds, and breakpoint limits.
+
+Safe fix:
+
+- Move stable reusable context before the cache-controlled message.
+- Move highly dynamic input after the cache breakpoint.
+- Keep `cache_control.type: ephemeral` and, when using `ttl`, set `ttl: 1h`.
+- Use CLI cache counters or HTTP cache usage fields to verify cache reads/writes after rerunning.
+
+Relevant links:
+
+- [Configuration reference](config.md)
+- [OpenAI-compatible chat integration](integrations/openai-compatible-chat.md)
+
 ## Validation Status Failed (`run` Exit 2 Or HTTP 200 With Failed Status)
 
 Symptom:

internal/adapter/cli/run.go

@@ -513,18 +513,25 @@ 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
 		}
 	}
 
@@ -606,7 +613,7 @@ 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,
@@ -620,6 +627,10 @@ 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) {

internal/adapter/cli/run_test.go

@@ -747,6 +747,35 @@ func TestRenderCommandDefaultFormatTextIncludesPreparedDetailsAndNoSecrets(t *te
 	}
 }
 
+func TestRenderCommandExplicitZeroTemperatureReachesEffectiveSettings(t *testing.T) {
+	lib := newCLITestLibrary(t)
+	inputPath := lib.writeInputFile(t, "transcript.md", "hello transcript")
+
+	writePromptFile(t, lib.promptDir, "prompt.render", "local-default")
+	profile := `id: local-default
+endpoint: http://127.0.0.1:1/v1
+model: profile-model
+temperature: 0.7
+`
+	if err := os.WriteFile(filepath.Join(lib.profileDir, "local-default.yaml"), []byte(profile), 0o644); err != nil {
+		t.Fatalf("failed to write profile fixture: %v", err)
+	}
+
+	code, stdout, stderr := runCLICommand(t, renderCommand, []string{
+		"--prompt-dir", lib.promptDir,
+		"--profile-dir", lib.profileDir,
+		"--prompt", "prompt.render",
+		"--input", "transcript=" + inputPath,
+		"--temperature", "0",
+	})
+	if code != ExitOK {
+		t.Fatalf("expected ExitOK, got %d stderr=%q", code, stderr)
+	}
+	if !strings.Contains(stdout, "\n  temperature: 0\n") {
+		t.Fatalf("expected explicit zero temperature in effective settings, got:\n%s", stdout)
+	}
+}
+
 func TestRenderCommandSucceedsWithPromptAndProfileDirsFromConfig(t *testing.T) {
 	lib := newCLITestLibrary(t)
 	inputPath := lib.writeInputFile(t, "transcript.md", "hello transcript")
@@ -1064,6 +1093,38 @@ func TestWriteOutputAndSummaryUseSeparateWriters(t *testing.T) {
 	if !strings.Contains(stderr.String(), "prompt=p@1") {
 		t.Fatalf("expected summary on stderr, got %q", stderr.String())
 	}
+	if strings.Contains(stderr.String(), "cached_tokens=") || strings.Contains(stderr.String(), "cache_write_tokens=") {
+		t.Fatalf("expected zero cache usage to be omitted from summary, got %q", stderr.String())
+	}
+}
+
+func TestPrintSummaryIncludesCacheUsageWhenPresent(t *testing.T) {
+	var stderr bytes.Buffer
+
+	printSummary(&stderr, &domain.RunResult{
+		PromptID:           "p",
+		PromptVersion:      "1",
+		SelectedProfileID:  "exec",
+		ModelName:          "m",
+		Validation:         domain.ValidationResult{Status: domain.ValidationPassed, Mode: domain.ValidationBasic},
+		RenderedPromptHash: "h",
+		InputHashes:        map[string]string{"in": "x"},
+		Usage: domain.TokenUsage{
+			PromptTokens:     10,
+			CompletionTokens: 5,
+			TotalTokens:      15,
+			CachedTokens:     0,
+			CacheWriteTokens: 3,
+		},
+	})
+
+	summary := stderr.String()
+	if !strings.Contains(summary, "usage=10/5/15") {
+		t.Fatalf("expected base usage summary, got %q", summary)
+	}
+	if !strings.Contains(summary, "cached_tokens=0 cache_write_tokens=3") {
+		t.Fatalf("expected cache usage in summary, got %q", summary)
+	}
 }
 
 type cliTestLibrary struct {

internal/adapter/http/dto.go

@@ -23,14 +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 {
@@ -79,13 +79,15 @@ type modelParamsDTO struct {
 	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

@@ -61,9 +61,9 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		}
 	}
 
-	var model *domain.ExecutionTarget
+	var model *domain.ExecutionTargetOverride
 	if req.Model != nil {
-		model = executionTargetFromModelOverrideDTO(req.Model)
+		model = executionTargetOverrideFromModelOverrideDTO(req.Model)
 	}
 
 	res, err := h.runner.Run(r.Context(), domain.RunRequest{
@@ -105,6 +105,8 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 				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,
@@ -121,11 +123,11 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 	writeJSON(w, http.StatusOK, resp)
 }
 
-func executionTargetFromModelOverrideDTO(dto *modelOverrideRequestDTO) *domain.ExecutionTarget {
+func executionTargetOverrideFromModelOverrideDTO(dto *modelOverrideRequestDTO) *domain.ExecutionTargetOverride {
 	if dto == nil {
 		return nil
 	}
-	return &domain.ExecutionTarget{
+	return &domain.ExecutionTargetOverride{
 		Endpoint:        dto.Endpoint,
 		Model:           dto.Model,
 		Temperature:     dto.Temperature,

internal/adapter/http/handler_test.go

@@ -13,6 +13,7 @@ import (
 	"time"
 
 	"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 +33,34 @@ 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
+}
+
 func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 	start := time.Now().UTC()
 	end := start.Add(2 * time.Second)
@@ -66,7 +95,13 @@ func TestHandlerPostRunsSuccessWithExplicitProfileID(t *testing.T) {
 			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,
@@ -111,6 +146,13 @@ 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"])
@@ -134,7 +176,7 @@ 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" {
@@ -171,6 +213,10 @@ 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) {
@@ -211,20 +257,136 @@ func TestHandlerModelOverrideMapsAllSupportedExecutionFields(t *testing.T) {
 	got := r.last.Execution
 	if got.Endpoint != "http://override/v1" ||
 		got.Model != "override-model" ||
-		got.Temperature != 0.6 ||
-		got.MaxTokens != 250 ||
-		got.TopP != 0.85 ||
-		got.TimeoutSeconds != 33 ||
 		got.ServiceTier != "flex" ||
 		got.ReasoningEffort != "medium" ||
 		got.APIKeyEnv != "SCRIPTORIUM_API_KEY" {
 		t.Fatalf("unexpected mapped execution target: %+v", got)
 	}
-	if !reflect.DeepEqual(got.ExtraParams, map[string]string{"provider_option": "on"}) {
+	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{
@@ -245,8 +407,10 @@ func TestHandlerResponseMetadataModelParamsIncludesAllSupportedFields(t *testing
 			ServiceTier:     "priority",
 			ReasoningEffort: "high",
 			APIKeyEnv:       "SCRIPTORIUM_API_KEY",
-			ExtraParams: map[string]string{
+			ExtraParams: map[string]any{
 				"provider_option": "on",
+				"number_value":    42,
+				"object_value":    map[string]any{"nested": "value"},
 			},
 		},
 	}}
@@ -301,6 +465,13 @@ func TestHandlerResponseMetadataModelParamsIncludesAllSupportedFields(t *testing
 	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) {
@@ -335,6 +506,54 @@ 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

internal/domain/domain.go

@@ -40,6 +40,24 @@ 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
@@ -47,7 +65,7 @@ type RunRequest struct {
 	ProfileID     string
 	Inputs        map[string]ArtifactRef
 	Vars          map[string]string
-	Execution     *ExecutionTarget
+	Execution     *ExecutionTargetOverride
 	Validation    *OutputContract
 	Metadata      map[string]string
 }
@@ -82,9 +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"`
@@ -115,6 +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"`
@@ -134,6 +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.
@@ -148,7 +170,30 @@ type ExecutionProfile struct {
 	ServiceTier     string         `yaml:"service_tier"`
 	ReasoningEffort string         `yaml:"reasoning_effort"`
 	APIKeyEnv       string         `yaml:"api_key_env"`
-	ExtraParams     map[string]string `yaml:"extra_params"`
+	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.
@@ -162,7 +207,7 @@ type ExecutionTarget struct {
 	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"`
+	ExtraParams     map[string]any `yaml:"extra_params" json:"extra_params"`
 }
 
 // OutputContract defines the requirements for the output artifact.
@@ -175,6 +220,7 @@ type OutputContract struct {
 
 // RenderedPrompt represents the prompt after template application.
 type RenderedPrompt struct {
+	SessionID string            `json:"session_id,omitempty"`
 	Messages  []RenderedMessage `json:"messages"`
 }
 
@@ -182,12 +228,14 @@ 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
 }
 
@@ -222,6 +270,8 @@ type TokenUsage struct {
 	PromptTokens     int
 	CompletionTokens int
 	TotalTokens      int
+	CachedTokens     int
+	CacheWriteTokens int
 }
 
 // ValidationResult represents the outcome of an output validation.

internal/domain/prepared_run_test.go

@@ -53,3 +53,88 @@ 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/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
@@ -123,7 +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)
 		}
 	}
 
@@ -151,6 +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 == "" {
@@ -165,3 +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

@@ -49,6 +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)
@@ -62,8 +92,80 @@ func TestTextFormatterDoesNotIncludeResolvedAPIKeyValue(t *testing.T) {
 	}
 }
 
+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 {
@@ -87,9 +189,24 @@ 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)
 	}
@@ -98,6 +215,47 @@ 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)

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"
@@ -89,7 +90,12 @@ func (c *OpenAICompatibleClient) Generate(ctx context.Context, req domain.Genera
 		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
 	}
 
-	payload, err := json.Marshal(wireReq)
+	wirePayload, err := openAIChatRequestPayload(wireReq)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %v", ErrInvalidRequest, err)
+	}
+
+	payload, err := json.Marshal(wirePayload)
 	if err != nil {
 		return nil, fmt.Errorf("%w: failed to encode request: %v", ErrRequestFailed, err)
 	}
@@ -110,6 +116,8 @@ 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
@@ -151,6 +159,8 @@ 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
 }
@@ -167,27 +177,36 @@ func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultMod
 	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 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
 	}
 
-	if req.Target.Temperature != 0 {
+	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 {
+	if req.Target.MaxTokens != 0 || req.TargetPresence.MaxTokens {
 		wireReq.MaxTokens = &req.Target.MaxTokens
 	}
-	if req.Target.TopP != 0 {
+	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 {
@@ -201,27 +220,105 @@ func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultMod
 
 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"`
 }
 
@@ -236,6 +333,28 @@ 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"
@@ -89,6 +90,9 @@ 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)
@@ -144,6 +148,245 @@ func TestOpenAICompatibleClientGenerateSuccess(t *testing.T) {
 	}
 }
 
+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) {
@@ -175,7 +418,7 @@ func TestOpenAICompatibleClientOmitsResponseFormatWhenNoStructuredOutput(t *test
 	}
 }
 
-func TestOpenAICompatibleClientOmitsReasoningEffortAndExtraParams(t *testing.T) {
+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()
@@ -196,19 +439,257 @@ func TestOpenAICompatibleClientOmitsReasoningEffortAndExtraParams(t *testing.T)
 		Target: domain.ExecutionTarget{
 			Model:           "model",
 			ReasoningEffort: "high",
-			ExtraParams: map[string]string{
-				"provider_option": "on",
+			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 omitted, got %#v", observedBody["extra_params"])
+		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")
+			}
+		})
 	}
 }
 

internal/profile/repository_test.go

@@ -2,6 +2,7 @@ package profile
 
 import (
 	"context"
+	"encoding/json"
 	"errors"
 	"os"
 	"path/filepath"
@@ -85,6 +86,63 @@ temperature: 0.1
 		}
 	})
 
+	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

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}},

internal/promptdef/filesystem_repository.go

@@ -29,6 +29,7 @@ type promptDefinitionFile struct {
 	Version        string                   `yaml:"version"`
 	DefaultProfile *string                  `yaml:"default_profile"`
 	Description    string                   `yaml:"description"`
+	SessionID      string                   `yaml:"session_id"`
 	Inputs         []promptInputFile        `yaml:"inputs"`
 	Messages       []promptMessageFile      `yaml:"messages"`
 	Output         promptOutputContractFile `yaml:"output"`
@@ -45,6 +46,12 @@ type promptMessageFile struct {
 	Role         string            `yaml:"role"`
 	Content      string            `yaml:"content"`
 	ContentFile  string            `yaml:"content_file"`
+	CacheControl *cacheControlFile `yaml:"cache_control"`
+}
+
+type cacheControlFile struct {
+	Type string `yaml:"type"`
+	TTL  string `yaml:"ttl"`
 }
 
 type promptOutputContractFile struct {
@@ -212,6 +219,11 @@ func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*d
 			return nil, fmt.Errorf("message %d (%s) must set exactly one of content or content_file", i, role)
 		}
 
+		cacheControl, err := normalizeCacheControl(msg.CacheControl)
+		if err != nil {
+			return nil, fmt.Errorf("message %d (%s) cache_control: %w", i, role, err)
+		}
+
 		templateContent := msg.Content
 		resolvedContentFile := ""
 		if hasContentFile {
@@ -233,6 +245,7 @@ func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*d
 			Role:         role,
 			Content:      templateContent,
 			ContentFile:  resolvedContentFile,
+			CacheControl: cacheControl,
 		})
 	}
 
@@ -262,6 +275,7 @@ func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*d
 		Version:        version,
 		DefaultProfile: defaultProfile,
 		Description:    strings.TrimSpace(raw.Description),
+		SessionID:      strings.TrimSpace(raw.SessionID),
 		Inputs:         inputs,
 		Templates:      templates,
 		OutputFormat:   raw.Output.Format,
@@ -274,6 +288,30 @@ func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*d
 	}, nil
 }
 
+func normalizeCacheControl(raw *cacheControlFile) (*domain.CacheControl, error) {
+	if raw == nil {
+		return nil, nil
+	}
+
+	cacheType := strings.TrimSpace(raw.Type)
+	if cacheType == "" {
+		return nil, errors.New("type is required")
+	}
+	if domain.CacheControlType(cacheType) != domain.CacheControlEphemeral {
+		return nil, fmt.Errorf("unsupported type %q", cacheType)
+	}
+
+	ttl := strings.TrimSpace(raw.TTL)
+	if ttl != "" && ttl != "1h" {
+		return nil, fmt.Errorf("unsupported ttl %q", ttl)
+	}
+
+	return &domain.CacheControl{
+		Type: domain.CacheControlType(cacheType),
+		TTL:  ttl,
+	}, nil
+}
+
 func isValidOutputFormat(f domain.OutputFormat) bool {
 	switch f {
 	case domain.FormatText, domain.FormatMarkdown, domain.FormatJSON:

internal/promptdef/repository_test.go

@@ -68,6 +68,44 @@ func TestFilesystemRepository_GetPromptDefinition(t *testing.T) {
 		}
 	})
 
+	t.Run("valid cache control with ttl", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-cache-control-ttl", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		assertCacheControl(t, p.Templates[0].CacheControl, domain.CacheControlEphemeral, "1h")
+		if p.Templates[1].CacheControl != nil {
+			t.Fatalf("expected second message cache control to be nil, got %#v", p.Templates[1].CacheControl)
+		}
+	})
+
+	t.Run("valid cache control without ttl", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-cache-control-without-ttl", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if len(p.Templates) != 2 {
+			t.Fatalf("expected 2 messages, got %d", len(p.Templates))
+		}
+		assertCacheControl(t, p.Templates[0].CacheControl, domain.CacheControlEphemeral, "")
+		if p.Templates[1].CacheControl != nil {
+			t.Fatalf("expected second message cache control to be nil, got %#v", p.Templates[1].CacheControl)
+		}
+	})
+
+	t.Run("valid session id template", func(t *testing.T) {
+		p, err := repo.GetPromptDefinition(ctx, "valid-session-id", "")
+		if err != nil {
+			t.Fatalf("expected no error, got %v", err)
+		}
+		if p.SessionID != "{{ .session_id }}" {
+			t.Fatalf("expected trimmed session_id template, got %q", p.SessionID)
+		}
+	})
+
 	t.Run("valid nested file-backed prompt resolves content file relative to nested YAML", func(t *testing.T) {
 		nestedDir := filepath.Join(tmpDir, "dnd", "recap")
 		if err := os.MkdirAll(nestedDir, 0o755); err != nil {
@@ -258,6 +296,10 @@ output:
 		{name: "invalid validation mode", id: "invalid_validation_mode", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"invalid validation mode"}},
 		{name: "json_schema without schema_path", id: "json_schema_without_schema_path", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"schema_path"}},
 		{name: "unknown input field", id: "unknown_input_field", targetErr: ErrInvalidYAML, errSubstrs: []string{"field unknown_input_setting not found"}},
+		{name: "empty cache control type", id: "empty_cache_control_type", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "type is required"}},
+		{name: "unsupported cache control type", id: "unsupported_cache_control_type", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "unsupported type"}},
+		{name: "unsupported cache control ttl", id: "unsupported_cache_control_ttl", targetErr: ErrInvalidPromptDefinition, errSubstrs: []string{"cache_control", "unsupported ttl"}},
+		{name: "unknown cache control field", id: "unknown_cache_control_field", targetErr: ErrInvalidYAML, errSubstrs: []string{"field unexpected not found"}},
 	}
 
 	for _, tc := range cases {
@@ -282,6 +324,19 @@ output:
 	})
 }
 
+func assertCacheControl(t *testing.T, got *domain.CacheControl, wantType domain.CacheControlType, wantTTL string) {
+	t.Helper()
+	if got == nil {
+		t.Fatal("expected cache control, got nil")
+	}
+	if got.Type != wantType {
+		t.Fatalf("unexpected cache control type: got %q want %q", got.Type, wantType)
+	}
+	if got.TTL != wantTTL {
+		t.Fatalf("unexpected cache control ttl: got %q want %q", got.TTL, wantTTL)
+	}
+}
+
 func writePromptTestFile(t *testing.T, path string, content string) {
 	t.Helper()
 	if err := os.WriteFile(path, []byte(strings.TrimLeft(content, "\n")), 0o644); err != nil {

internal/promptdef/testdata/empty_cache_control_type.yaml

@@ -0,0 +1,10 @@
+id: empty-cache-control-type
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control: {}
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/unknown_cache_control_field.yaml

@@ -0,0 +1,12 @@
+id: unknown-cache-control-field
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control:
+      type: ephemeral
+      unexpected: true
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/unsupported_cache_control_ttl.yaml

@@ -0,0 +1,12 @@
+id: unsupported-cache-control-ttl
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control:
+      type: ephemeral
+      ttl: 5m
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/unsupported_cache_control_type.yaml

@@ -0,0 +1,11 @@
+id: unsupported-cache-control-type
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control:
+      type: persistent
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/valid_cache_control_ttl.yaml

@@ -0,0 +1,14 @@
+id: valid-cache-control-ttl
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control:
+      type: ephemeral
+      ttl: 1h
+  - role: user
+    content: "Summarize the input."
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/valid_cache_control_without_ttl.yaml

@@ -0,0 +1,13 @@
+id: valid-cache-control-without-ttl
+version: "1.0.0"
+messages:
+  - role: system
+    content: "Use cached instructions."
+    cache_control:
+      type: ephemeral
+  - role: user
+    content: "Summarize the input."
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/promptdef/testdata/valid_session_id.yaml

@@ -0,0 +1,10 @@
+id: valid-session-id
+version: "1.0.0"
+session_id: "  {{ .session_id }}  "
+messages:
+  - role: user
+    content: Hello.
+output:
+  format: markdown
+  validation_mode: basic
+  repair_attempts: 0

internal/usecase/runner.go

@@ -90,11 +90,15 @@ func (r *Runner) Run(ctx context.Context, req domain.RunRequest) (*domain.RunRes
 	}
 
 	genResp, err := r.llm.Generate(ctx, domain.GenerateRequest{
-		Prompt:           domain.RenderedPrompt{Messages: prepared.Messages},
+		Prompt:           domain.RenderedPrompt{SessionID: prepared.SessionID, Messages: prepared.Messages},
 		Target:           prepared.EffectiveModelParams,
+		TargetPresence:   prepared.TargetPresence,
 		StructuredOutput: prepared.StructuredOutput,
 	})
 	if err != nil {
+		if errors.Is(err, llm.ErrInvalidRequest) {
+			return nil, fmt.Errorf("%w: %w", ErrInvalidRequest, err)
+		}
 		return nil, fmt.Errorf("%w: %w", ErrLLMGenerate, err)
 	}
 
@@ -187,7 +191,10 @@ func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.Pr
 		return nil, fmt.Errorf("%w: %w", ErrProfileLoad, err)
 	}
 
-	effectiveModel := resolveExecutionTarget(execProfile, req.Execution)
+	effectiveModel, targetPresence, err := resolveExecutionTarget(execProfile, req.Execution)
+	if err != nil {
+		return nil, fmt.Errorf("%w: %w", ErrInvalidRequest, err)
+	}
 	if strings.TrimSpace(effectiveModel.Endpoint) == "" {
 		return nil, fmt.Errorf("%w: execution endpoint is required", ErrInvalidRequest)
 	}
@@ -230,9 +237,11 @@ func (r *Runner) Prepare(ctx context.Context, req domain.RunRequest) (*domain.Pr
 		PromptHash:           promptDefinitionHash,
 		SelectedProfileID:    selectedProfileID,
 		EffectiveModelParams: effectiveModel,
+		TargetPresence:       targetPresence,
 		OutputContract:       effectiveContract,
 		StructuredOutput:     structuredOutput,
 		InputHashes:          inputHashes,
+		SessionID:            renderedPrompt.SessionID,
 		RenderedPromptHash:   hashRenderedPrompt(*renderedPrompt),
 		Messages:             renderedPrompt.Messages,
 		StartTime:            start,
@@ -354,22 +363,75 @@ func mergeExecutionTarget(base domain.ExecutionTarget, override domain.Execution
 		out.APIKeyEnv = override.APIKeyEnv
 	}
 	if len(override.ExtraParams) > 0 {
-		cp := make(map[string]string, len(override.ExtraParams))
-		for k, v := range override.ExtraParams {
-			cp[k] = v
-		}
-		out.ExtraParams = cp
+		out.ExtraParams = copyExtraParams(override.ExtraParams)
 	}
 	return out
 }
 
-func resolveExecutionTarget(profileValue *domain.ExecutionProfile, override *domain.ExecutionTarget) domain.ExecutionTarget {
+func mergeExecutionTargetOverride(base domain.ExecutionTarget, override domain.ExecutionTargetOverride) (domain.ExecutionTarget, domain.ExecutionTargetPresence, error) {
+	out := base
+	var presence domain.ExecutionTargetPresence
+	if override.Endpoint != "" {
+		out.Endpoint = override.Endpoint
+	}
+	if override.Model != "" {
+		out.Model = override.Model
+	}
+	if override.Temperature != nil {
+		if *override.Temperature < 0 || *override.Temperature > 2 {
+			return domain.ExecutionTarget{}, domain.ExecutionTargetPresence{}, errors.New("temperature must be between 0 and 2")
+		}
+		out.Temperature = *override.Temperature
+		presence.Temperature = true
+	}
+	if override.MaxTokens != nil {
+		if *override.MaxTokens < 0 {
+			return domain.ExecutionTarget{}, domain.ExecutionTargetPresence{}, errors.New("max_tokens must be greater than or equal to 0")
+		}
+		out.MaxTokens = *override.MaxTokens
+		presence.MaxTokens = true
+	}
+	if override.TopP != nil {
+		if *override.TopP < 0 || *override.TopP > 1 {
+			return domain.ExecutionTarget{}, domain.ExecutionTargetPresence{}, errors.New("top_p must be between 0 and 1")
+		}
+		out.TopP = *override.TopP
+		presence.TopP = true
+	}
+	if override.TimeoutSeconds != nil {
+		if *override.TimeoutSeconds < 0 {
+			return domain.ExecutionTarget{}, domain.ExecutionTargetPresence{}, errors.New("timeout_seconds must be greater than or equal to 0")
+		}
+		out.TimeoutSeconds = *override.TimeoutSeconds
+		presence.TimeoutSeconds = true
+	}
+	if strings.TrimSpace(override.ServiceTier) != "" {
+		out.ServiceTier = override.ServiceTier
+	}
+	if strings.TrimSpace(override.ReasoningEffort) != "" {
+		out.ReasoningEffort = override.ReasoningEffort
+	}
+	if strings.TrimSpace(override.APIKeyEnv) != "" {
+		out.APIKeyEnv = override.APIKeyEnv
+	}
+	if len(override.ExtraParams) > 0 {
+		out.ExtraParams = copyExtraParams(override.ExtraParams)
+	}
+	return out, presence, nil
+}
+
+func resolveExecutionTarget(profileValue *domain.ExecutionProfile, override *domain.ExecutionTargetOverride) (domain.ExecutionTarget, domain.ExecutionTargetPresence, error) {
 	out := defaults.ExecutionTargetDefault()
 	out = mergeExecutionTarget(out, executionProfileToTarget(profileValue))
+	var presence domain.ExecutionTargetPresence
 	if override != nil {
-		out = mergeExecutionTarget(out, *override)
+		var err error
+		out, presence, err = mergeExecutionTargetOverride(out, *override)
+		if err != nil {
+			return domain.ExecutionTarget{}, domain.ExecutionTargetPresence{}, err
 		}
-	return out
+	}
+	return out, presence, nil
 }
 
 func validateAPIKeyEnv(apiKeyEnv string) error {
@@ -387,13 +449,6 @@ func executionProfileToTarget(p *domain.ExecutionProfile) domain.ExecutionTarget
 	if p == nil {
 		return domain.ExecutionTarget{}
 	}
-	cp := map[string]string(nil)
-	if len(p.ExtraParams) > 0 {
-		cp = make(map[string]string, len(p.ExtraParams))
-		for k, v := range p.ExtraParams {
-			cp[k] = v
-		}
-	}
 	return domain.ExecutionTarget{
 		Endpoint:        p.Endpoint,
 		Model:           p.Model,
@@ -404,10 +459,21 @@ func executionProfileToTarget(p *domain.ExecutionProfile) domain.ExecutionTarget
 		ServiceTier:     p.ServiceTier,
 		ReasoningEffort: p.ReasoningEffort,
 		APIKeyEnv:       p.APIKeyEnv,
-		ExtraParams:     cp,
+		ExtraParams:     copyExtraParams(p.ExtraParams),
 	}
 }
 
+func copyExtraParams(src map[string]any) map[string]any {
+	if len(src) == 0 {
+		return nil
+	}
+	cp := make(map[string]any, len(src))
+	for k, v := range src {
+		cp[k] = v
+	}
+	return cp
+}
+
 func resolveOutputContract(def *domain.PromptDefinition, override *domain.OutputContract) domain.OutputContract {
 	contract := def.Validation
 	if contract.Format == "" {
@@ -424,10 +490,23 @@ func resolveOutputContract(def *domain.PromptDefinition, override *domain.Output
 
 func hashRenderedPrompt(p domain.RenderedPrompt) string {
 	var b strings.Builder
+	if p.SessionID != "" {
+		b.WriteString("session_id=")
+		b.WriteString(p.SessionID)
+		b.WriteString("\n---\n")
+	}
 	for _, msg := range p.Messages {
 		b.WriteString(msg.Role)
 		b.WriteByte('\n')
 		b.WriteString(msg.Content)
+		if msg.CacheControl != nil {
+			b.WriteString("\ncache_control.type=")
+			b.WriteString(string(msg.CacheControl.Type))
+			if msg.CacheControl.TTL != "" {
+				b.WriteString("\ncache_control.ttl=")
+				b.WriteString(msg.CacheControl.TTL)
+			}
+		}
 		b.WriteString("\n---\n")
 	}
 	h := sha256.Sum256([]byte(b.String()))

internal/usecase/runner_test.go

@@ -14,6 +14,7 @@ import (
 
 	"gitea.maximumdirect.net/eric/scriptorium/internal/defaults"
 	"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/prompt"
 	"gitea.maximumdirect.net/eric/scriptorium/internal/promptdef"
@@ -160,7 +161,7 @@ func TestRunnerPrepareWithExplicitProfileSelection(t *testing.T) {
 		"a://t": {Body: []byte("transcript"), Hash: hashString("transcript")},
 		"a://g": {Body: []byte("glossary"), Hash: hashString("glossary")},
 	}}
-	renderer := &fakeRenderer{rendered: &domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "system", Content: "sys"}, {Role: "user", Content: "usr"}}}}
+	renderer := &fakeRenderer{rendered: &domain.RenderedPrompt{SessionID: "session-123", Messages: []domain.RenderedMessage{{Role: "system", Content: "sys"}, {Role: "user", Content: "usr"}}}}
 	llmClient := &fakeLLM{forbid: true}
 
 	runner := NewRunner(promptRepo, execRepo, reader, renderer, llmClient, nil)
@@ -172,7 +173,7 @@ func TestRunnerPrepareWithExplicitProfileSelection(t *testing.T) {
 			"transcript": {Type: domain.ArtifactRefFile, URI: "a://t"},
 			"glossary":   {Type: domain.ArtifactRefFile, URI: "a://g"},
 		},
-		Execution: &domain.ExecutionTarget{Endpoint: "http://override/v1", Model: "m", Temperature: 0.3, TimeoutSeconds: 90},
+		Execution: &domain.ExecutionTargetOverride{Endpoint: "http://override/v1", Model: "m", Temperature: float64Ptr(0.3), TimeoutSeconds: intPtr(90)},
 	})
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
@@ -198,6 +199,9 @@ func TestRunnerPrepareWithExplicitProfileSelection(t *testing.T) {
 	if len(prepared.Messages) != 2 {
 		t.Fatalf("expected two messages, got %d", len(prepared.Messages))
 	}
+	if prepared.SessionID != "session-123" {
+		t.Fatalf("expected prepared session id, got %q", prepared.SessionID)
+	}
 	if llmClient.calls != 0 {
 		t.Fatalf("prepare should not call llm, calls=%d", llmClient.calls)
 	}
@@ -269,11 +273,11 @@ func TestRunnerPrepareRuntimeOverrideBeatsSelectedProfileValue(t *testing.T) {
 		PromptID:  "p",
 		ProfileID: "exec",
 		Inputs:    singleInputRef(),
-		Execution: &domain.ExecutionTarget{
+		Execution: &domain.ExecutionTargetOverride{
 			Endpoint:       "http://override/v1",
 			Model:          "override-model",
-			Temperature:    0.7,
-			TimeoutSeconds: 30,
+			Temperature:    float64Ptr(0.7),
+			TimeoutSeconds: intPtr(30),
 			ServiceTier:    "flex",
 		},
 	})
@@ -291,6 +295,143 @@ func TestRunnerPrepareRuntimeOverrideBeatsSelectedProfileValue(t *testing.T) {
 	}
 }
 
+func TestRunnerPrepareRequestNumericOverridePresence(t *testing.T) {
+	tests := []struct {
+		name            string
+		override        *domain.ExecutionTargetOverride
+		wantTemperature float64
+		wantMaxTokens   int
+		wantTopP        float64
+		wantTimeoutSecs int
+		wantPresence    domain.ExecutionTargetPresence
+	}{
+		{
+			name:            "omitted preserves profile values",
+			override:        &domain.ExecutionTargetOverride{},
+			wantTemperature: 0.7,
+			wantMaxTokens:   321,
+			wantTopP:        0.8,
+			wantTimeoutSecs: 45,
+		},
+		{
+			name:            "explicit zero temperature",
+			override:        &domain.ExecutionTargetOverride{Temperature: float64Ptr(0)},
+			wantTemperature: 0,
+			wantMaxTokens:   321,
+			wantTopP:        0.8,
+			wantTimeoutSecs: 45,
+			wantPresence:    domain.ExecutionTargetPresence{Temperature: true},
+		},
+		{
+			name:            "explicit zero max tokens",
+			override:        &domain.ExecutionTargetOverride{MaxTokens: intPtr(0)},
+			wantTemperature: 0.7,
+			wantMaxTokens:   0,
+			wantTopP:        0.8,
+			wantTimeoutSecs: 45,
+			wantPresence:    domain.ExecutionTargetPresence{MaxTokens: true},
+		},
+		{
+			name:            "explicit zero top p",
+			override:        &domain.ExecutionTargetOverride{TopP: float64Ptr(0)},
+			wantTemperature: 0.7,
+			wantMaxTokens:   321,
+			wantTopP:        0,
+			wantTimeoutSecs: 45,
+			wantPresence:    domain.ExecutionTargetPresence{TopP: true},
+		},
+		{
+			name:            "explicit zero timeout",
+			override:        &domain.ExecutionTargetOverride{TimeoutSeconds: intPtr(0)},
+			wantTemperature: 0.7,
+			wantMaxTokens:   321,
+			wantTopP:        0.8,
+			wantTimeoutSecs: 0,
+			wantPresence:    domain.ExecutionTargetPresence{TimeoutSeconds: true},
+		},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			runner := NewRunner(
+				&fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)},
+				&fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{
+					"exec": {
+						ID:             "exec",
+						Endpoint:       "http://profile/v1",
+						Model:          "profile-model",
+						Temperature:    0.7,
+						MaxTokens:      321,
+						TopP:           0.8,
+						TimeoutSeconds: 45,
+					},
+				}},
+				defaultArtifactReader(),
+				defaultRenderer(),
+				&fakeLLM{forbid: true},
+				nil,
+			)
+
+			prepared, err := runner.Prepare(context.Background(), domain.RunRequest{
+				PromptID:  "p",
+				ProfileID: "exec",
+				Inputs:    singleInputRef(),
+				Execution: tc.override,
+			})
+			if err != nil {
+				t.Fatalf("expected no error, got %v", err)
+			}
+			got := prepared.EffectiveModelParams
+			if got.Temperature != tc.wantTemperature ||
+				got.MaxTokens != tc.wantMaxTokens ||
+				got.TopP != tc.wantTopP ||
+				got.TimeoutSeconds != tc.wantTimeoutSecs {
+				t.Fatalf("unexpected effective numeric settings: %+v", got)
+			}
+			if prepared.TargetPresence != tc.wantPresence {
+				t.Fatalf("unexpected target presence: got %+v want %+v", prepared.TargetPresence, tc.wantPresence)
+			}
+		})
+	}
+}
+
+func TestRunnerPrepareInvalidRequestNumericOverridesFail(t *testing.T) {
+	tests := []struct {
+		name     string
+		override *domain.ExecutionTargetOverride
+	}{
+		{name: "temperature below range", override: &domain.ExecutionTargetOverride{Temperature: float64Ptr(-0.1)}},
+		{name: "temperature above range", override: &domain.ExecutionTargetOverride{Temperature: float64Ptr(2.1)}},
+		{name: "max tokens below range", override: &domain.ExecutionTargetOverride{MaxTokens: intPtr(-1)}},
+		{name: "top p below range", override: &domain.ExecutionTargetOverride{TopP: float64Ptr(-0.1)}},
+		{name: "top p above range", override: &domain.ExecutionTargetOverride{TopP: float64Ptr(1.1)}},
+		{name: "timeout below range", override: &domain.ExecutionTargetOverride{TimeoutSeconds: intPtr(-1)}},
+	}
+
+	for _, tc := range tests {
+		t.Run(tc.name, func(t *testing.T) {
+			runner := NewRunner(
+				&fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)},
+				&fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}},
+				defaultArtifactReader(),
+				defaultRenderer(),
+				&fakeLLM{forbid: true},
+				nil,
+			)
+
+			_, err := runner.Prepare(context.Background(), domain.RunRequest{
+				PromptID:  "p",
+				ProfileID: "exec",
+				Inputs:    singleInputRef(),
+				Execution: tc.override,
+			})
+			if !errors.Is(err, ErrInvalidRequest) {
+				t.Fatalf("expected ErrInvalidRequest, got %v", err)
+			}
+		})
+	}
+}
+
 func TestRunnerPrepareSelectedProfileBeatsBuiltInDefault(t *testing.T) {
 	promptRepo := &fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)}
 	execRepo := &fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{
@@ -577,6 +718,100 @@ func TestDeriveStructuredSchemaName(t *testing.T) {
 	}
 }
 
+func TestHashRenderedPromptIncludesCacheControlWhenPresent(t *testing.T) {
+	uncached := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+		{Role: "system", Content: "sys"},
+		{Role: "user", Content: "usr"},
+	}}
+	wantLegacyHash := hashString("system\nsys\n---\nuser\nusr\n---\n")
+	if got := hashRenderedPrompt(uncached); got != wantLegacyHash {
+		t.Fatalf("expected no-cache hash to preserve legacy input, got %q want %q", got, wantLegacyHash)
+	}
+
+	withCache := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "sys",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+				TTL:  "1h",
+			},
+		},
+		{Role: "user", Content: "usr"},
+	}}
+	alsoWithCache := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "sys",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+				TTL:  "1h",
+			},
+		},
+		{Role: "user", Content: "usr"},
+	}}
+	withoutTTL := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+		{
+			Role:    "system",
+			Content: "sys",
+			CacheControl: &domain.CacheControl{
+				Type: domain.CacheControlEphemeral,
+			},
+		},
+		{Role: "user", Content: "usr"},
+	}}
+
+	cachedHash := hashRenderedPrompt(withCache)
+	if cachedHash == hashRenderedPrompt(uncached) {
+		t.Fatal("expected cache control to change rendered prompt hash")
+	}
+	if cachedHash != hashRenderedPrompt(alsoWithCache) {
+		t.Fatal("expected identical cache control metadata to produce stable hash")
+	}
+	if cachedHash == hashRenderedPrompt(withoutTTL) {
+		t.Fatal("expected ttl changes to affect rendered prompt hash")
+	}
+}
+
+func TestHashRenderedPromptIncludesSessionIDWhenPresent(t *testing.T) {
+	withoutSession := domain.RenderedPrompt{Messages: []domain.RenderedMessage{
+		{Role: "system", Content: "sys"},
+		{Role: "user", Content: "usr"},
+	}}
+	withSession := domain.RenderedPrompt{
+		SessionID: "session-123",
+		Messages: []domain.RenderedMessage{
+			{Role: "system", Content: "sys"},
+			{Role: "user", Content: "usr"},
+		},
+	}
+	alsoWithSession := domain.RenderedPrompt{
+		SessionID: "session-123",
+		Messages: []domain.RenderedMessage{
+			{Role: "system", Content: "sys"},
+			{Role: "user", Content: "usr"},
+		},
+	}
+	otherSession := domain.RenderedPrompt{
+		SessionID: "session-456",
+		Messages: []domain.RenderedMessage{
+			{Role: "system", Content: "sys"},
+			{Role: "user", Content: "usr"},
+		},
+	}
+
+	sessionHash := hashRenderedPrompt(withSession)
+	if sessionHash == hashRenderedPrompt(withoutSession) {
+		t.Fatal("expected session_id to change rendered prompt hash")
+	}
+	if sessionHash != hashRenderedPrompt(alsoWithSession) {
+		t.Fatal("expected identical session_id to produce stable hash")
+	}
+	if sessionHash == hashRenderedPrompt(otherSession) {
+		t.Fatal("expected session_id value changes to affect rendered prompt hash")
+	}
+}
+
 func TestRunnerRunSuccessful(t *testing.T) {
 	promptRepo := &fakePromptRepo{def: promptDef(domain.FormatMarkdown, domain.ValidationBasic, 0)}
 	execRepo := &fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}}
@@ -584,7 +819,7 @@ func TestRunnerRunSuccessful(t *testing.T) {
 		"a://t": {Body: []byte("transcript"), Hash: hashString("transcript")},
 		"a://g": {Body: []byte("glossary"), Hash: hashString("glossary")},
 	}}
-	renderer := &fakeRenderer{rendered: &domain.RenderedPrompt{Messages: []domain.RenderedMessage{{Role: "system", Content: "sys"}, {Role: "user", Content: "usr"}}}}
+	renderer := &fakeRenderer{rendered: &domain.RenderedPrompt{SessionID: "session-123", Messages: []domain.RenderedMessage{{Role: "system", Content: "sys"}, {Role: "user", Content: "usr"}}}}
 	llmClient := &fakeLLM{resp: &domain.GenerateResponse{Content: "# recap", Usage: domain.TokenUsage{TotalTokens: 7}}}
 
 	runner := NewRunner(promptRepo, execRepo, reader, renderer, llmClient, nil)
@@ -596,7 +831,7 @@ func TestRunnerRunSuccessful(t *testing.T) {
 			"transcript": {Type: domain.ArtifactRefFile, URI: "a://t"},
 			"glossary":   {Type: domain.ArtifactRefFile, URI: "a://g"},
 		},
-		Execution: &domain.ExecutionTarget{Endpoint: "http://override/v1", Model: "m", Temperature: 0.3, TimeoutSeconds: 90},
+		Execution: &domain.ExecutionTargetOverride{Endpoint: "http://override/v1", Model: "m", Temperature: float64Ptr(0.3), TimeoutSeconds: intPtr(90)},
 	})
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
@@ -631,6 +866,48 @@ func TestRunnerRunSuccessful(t *testing.T) {
 	if llmClient.lastReq.Target.TimeoutSeconds != 90 {
 		t.Fatalf("expected timeout propagation, got %d", llmClient.lastReq.Target.TimeoutSeconds)
 	}
+	if !llmClient.lastReq.TargetPresence.Temperature || !llmClient.lastReq.TargetPresence.TimeoutSeconds {
+		t.Fatalf("expected numeric override presence to be sent to llm, got %+v", llmClient.lastReq.TargetPresence)
+	}
+	if llmClient.lastReq.Prompt.SessionID != "session-123" {
+		t.Fatalf("expected session id to be sent to llm, got %q", llmClient.lastReq.Prompt.SessionID)
+	}
+}
+
+func TestRunnerRunPassesExtraParamsToGenerateRequestTarget(t *testing.T) {
+	extraParams := map[string]any{
+		"string_value":  "enabled",
+		"number_value":  42,
+		"boolean_value": true,
+		"object_value":  map[string]any{"nested": "value"},
+		"array_value":   []any{"first", 3, false},
+	}
+	promptRepo := &fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)}
+	execRepo := &fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{
+		"exec": {
+			ID:          "exec",
+			Endpoint:    "http://profile/v1",
+			Model:       "profile-model",
+			ExtraParams: extraParams,
+		},
+	}}
+	llmClient := &fakeLLM{resp: &domain.GenerateResponse{Content: "ok"}}
+	runner := NewRunner(promptRepo, execRepo, defaultArtifactReader(), defaultRenderer(), llmClient, nil)
+
+	res, err := runner.Run(context.Background(), domain.RunRequest{
+		PromptID:  "p",
+		ProfileID: "exec",
+		Inputs:    singleInputRef(),
+	})
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if !reflect.DeepEqual(res.EffectiveModelParams.ExtraParams, extraParams) {
+		t.Fatalf("expected run result extra_params to match profile values, got %#v", res.EffectiveModelParams.ExtraParams)
+	}
+	if !reflect.DeepEqual(llmClient.lastReq.Target.ExtraParams, extraParams) {
+		t.Fatalf("expected generate request extra_params to match profile values, got %#v", llmClient.lastReq.Target.ExtraParams)
+	}
 }
 
 func TestRunnerRunAndPrepareResolveSameProfileAndEffectiveSettings(t *testing.T) {
@@ -649,7 +926,7 @@ func TestRunnerRunAndPrepareResolveSameProfileAndEffectiveSettings(t *testing.T)
 		Inputs: map[string]domain.ArtifactRef{
 			"transcript": {Type: domain.ArtifactRefFile, URI: "a://t"},
 		},
-		Execution: &domain.ExecutionTarget{Endpoint: "http://override/v1", Model: "m", Temperature: 0.3, TimeoutSeconds: 90},
+		Execution: &domain.ExecutionTargetOverride{Endpoint: "http://override/v1", Model: "m", Temperature: float64Ptr(0.3), TimeoutSeconds: intPtr(90)},
 	}
 
 	prepared, err := runner.Prepare(context.Background(), req)
@@ -777,11 +1054,11 @@ func TestRunnerRunExplicitRuntimeOverrideBeatsSelectedProfileValue(t *testing.T)
 		PromptID:  "p",
 		ProfileID: "exec",
 		Inputs:    singleInputRef(),
-		Execution: &domain.ExecutionTarget{
+		Execution: &domain.ExecutionTargetOverride{
 			Endpoint:       "http://override/v1",
 			Model:          "override-model",
-			Temperature:    0.7,
-			TimeoutSeconds: 30,
+			Temperature:    float64Ptr(0.7),
+			TimeoutSeconds: intPtr(30),
 			ServiceTier:    "flex",
 		},
 	})
@@ -916,7 +1193,7 @@ func TestRunnerRunRuntimeAPIKeyEnvOverrideWorks(t *testing.T) {
 		PromptID:  "p",
 		ProfileID: "exec",
 		Inputs:    singleInputRef(),
-		Execution: &domain.ExecutionTarget{APIKeyEnv: envName},
+		Execution: &domain.ExecutionTargetOverride{APIKeyEnv: envName},
 	})
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
@@ -941,7 +1218,7 @@ func TestRunnerRunRuntimeAPIKeyEnvOverrideBeatsProfile(t *testing.T) {
 		PromptID:  "p",
 		ProfileID: "exec",
 		Inputs:    singleInputRef(),
-		Execution: &domain.ExecutionTarget{APIKeyEnv: runtimeEnv},
+		Execution: &domain.ExecutionTargetOverride{APIKeyEnv: runtimeEnv},
 	})
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
@@ -1040,6 +1317,28 @@ func TestRunnerRunLLMFailure(t *testing.T) {
 	}
 }
 
+func TestRunnerRunLLMInvalidRequestMapsToUsecaseInvalidRequest(t *testing.T) {
+	runner := NewRunner(
+		&fakePromptRepo{def: promptDef(domain.FormatText, domain.ValidationNone, 0)},
+		&fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}},
+		defaultArtifactReader(),
+		defaultRenderer(),
+		&fakeLLM{err: llm.ErrInvalidRequest},
+		nil,
+	)
+	_, err := runner.Run(context.Background(), domain.RunRequest{
+		PromptID:  "p",
+		ProfileID: "exec",
+		Inputs:    singleInputRef(),
+	})
+	if !errors.Is(err, ErrInvalidRequest) {
+		t.Fatalf("expected ErrInvalidRequest, got %v", err)
+	}
+	if errors.Is(err, ErrLLMGenerate) {
+		t.Fatalf("did not expect ErrLLMGenerate, got %v", err)
+	}
+}
+
 func TestRunnerRunValidationStillWorks(t *testing.T) {
 	validator := &fakeValidator{result: domain.ValidationResult{Status: domain.ValidationFailed, Mode: domain.ValidationBasic, Errors: []string{"bad"}, IsValid: false}}
 	runner := NewRunner(
@@ -1082,7 +1381,7 @@ func TestRunnerRunStructuredRepairRemainsBoundedAndUsesEffectiveModelSettings(t
 		PromptID:  "p",
 		ProfileID: "exec",
 		Inputs:    singleInputRef(),
-		Execution: &domain.ExecutionTarget{Endpoint: "http://override/v1", Model: "override-model", TimeoutSeconds: 22},
+		Execution: &domain.ExecutionTargetOverride{Endpoint: "http://override/v1", Model: "override-model", TimeoutSeconds: intPtr(22)},
 	})
 	if err != nil {
 		t.Fatalf("expected no error, got %v", err)
@@ -1169,7 +1468,7 @@ func TestExecutionProfileToTargetPopulatesAllFieldsAndCopiesExtraParams(t *testi
 		ServiceTier:     "priority",
 		ReasoningEffort: "medium",
 		APIKeyEnv:       "SCRIPTORIUM_API_KEY",
-		ExtraParams: map[string]string{
+		ExtraParams: map[string]any{
 			"provider_option": "on",
 		},
 	}
@@ -1208,12 +1507,18 @@ func TestResolveExecutionTargetProfileValuesPopulateAllSupportedFields(t *testin
 		ServiceTier:     "priority",
 		ReasoningEffort: "low",
 		APIKeyEnv:       "PROFILE_KEY",
-		ExtraParams: map[string]string{
+		ExtraParams: map[string]any{
 			"profile_option": "enabled",
 		},
 	}
 
-	target := resolveExecutionTarget(profileValue, nil)
+	target, presence, err := resolveExecutionTarget(profileValue, nil)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if presence != (domain.ExecutionTargetPresence{}) {
+		t.Fatalf("expected no request override presence, got %+v", presence)
+	}
 	if target.Endpoint != profileValue.Endpoint ||
 		target.Model != profileValue.Model ||
 		target.Temperature != profileValue.Temperature ||
@@ -1242,32 +1547,38 @@ func TestResolveExecutionTargetRuntimeOverridesBeatProfileForAllOverrideableFiel
 		ServiceTier:     "priority",
 		ReasoningEffort: "medium",
 		APIKeyEnv:       "PROFILE_KEY",
-		ExtraParams: map[string]string{
+		ExtraParams: map[string]any{
 			"profile_only": "yes",
 		},
 	}
-	override := &domain.ExecutionTarget{
+	override := &domain.ExecutionTargetOverride{
 		Endpoint:        "http://override/v1",
 		Model:           "override-model",
-		Temperature:     0.9,
-		MaxTokens:       111,
-		TopP:            0.5,
-		TimeoutSeconds:  30,
+		Temperature:     float64Ptr(0.9),
+		MaxTokens:       intPtr(111),
+		TopP:            float64Ptr(0.5),
+		TimeoutSeconds:  intPtr(30),
 		ServiceTier:     "flex",
 		ReasoningEffort: "high",
 		APIKeyEnv:       "RUNTIME_KEY",
-		ExtraParams: map[string]string{
+		ExtraParams: map[string]any{
 			"runtime_only": "yes",
 		},
 	}
 
-	target := resolveExecutionTarget(profileValue, override)
+	target, presence, err := resolveExecutionTarget(profileValue, override)
+	if err != nil {
+		t.Fatalf("expected no error, got %v", err)
+	}
+	if presence != (domain.ExecutionTargetPresence{Temperature: true, MaxTokens: true, TopP: true, TimeoutSeconds: true}) {
+		t.Fatalf("unexpected override presence: %+v", presence)
+	}
 	if target.Endpoint != override.Endpoint ||
 		target.Model != override.Model ||
-		target.Temperature != override.Temperature ||
-		target.MaxTokens != override.MaxTokens ||
-		target.TopP != override.TopP ||
-		target.TimeoutSeconds != override.TimeoutSeconds ||
+		target.Temperature != *override.Temperature ||
+		target.MaxTokens != *override.MaxTokens ||
+		target.TopP != *override.TopP ||
+		target.TimeoutSeconds != *override.TimeoutSeconds ||
 		target.ServiceTier != override.ServiceTier ||
 		target.ReasoningEffort != override.ReasoningEffort ||
 		target.APIKeyEnv != override.APIKeyEnv {
@@ -1311,12 +1622,12 @@ func TestMergeExecutionTargetEmptyStringOverridesDoNotErase(t *testing.T) {
 
 func TestMergeExecutionTargetEmptyExtraParamsDoesNotErase(t *testing.T) {
 	base := domain.ExecutionTarget{
-		ExtraParams: map[string]string{
+		ExtraParams: map[string]any{
 			"keep": "value",
 		},
 	}
 	override := domain.ExecutionTarget{
-		ExtraParams: map[string]string{},
+		ExtraParams: map[string]any{},
 	}
 
 	merged := mergeExecutionTarget(base, override)
@@ -1392,6 +1703,14 @@ func singleInputRef() map[string]domain.ArtifactRef {
 	return map[string]domain.ArtifactRef{"transcript": {Type: domain.ArtifactRefFile, URI: "a://ok"}}
 }
 
+func float64Ptr(v float64) *float64 {
+	return &v
+}
+
+func intPtr(v int) *int {
+	return &v
+}
+
 func newMinimalRunner(promptRepo *fakePromptRepo, execRepo *fakeExecutionProfileRepo) *Runner {
 	return NewRunner(
 		promptRepo,