Document prompt cache control behavior

docs/cli.md

@@ -32,6 +32,7 @@ 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.
 
 ## Flag Reference
 
@@ -93,6 +94,7 @@ 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

@@ -119,6 +119,7 @@ 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 +129,27 @@ 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.
+
 `output` fields:
 
 - `format` (required): `text`, `markdown`, or `json`.
@@ -184,6 +206,8 @@ Profile rules:
 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.
+- Messages without `cache_control` serialize with string `content`.
+- Messages with `cache_control` serialize as a single text content-block array containing `cache_control`.
 - `reasoning_effort` and `extra_params` are parsed and carried in effective settings, but are not currently serialized into outbound chat-completions requests.
 
 ## Schema Behavior

docs/integrations/http-api.md

@@ -128,7 +128,9 @@ 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 +144,8 @@ 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

@@ -26,7 +26,7 @@ Example:
 Serialized JSON fields:
 
 - `model` (required after fallback resolution)
-- `messages` (role/content pairs from rendered prompt)
+- `messages` (rendered prompt messages)
 - `temperature` (only when non-zero)
 - `max_tokens` (only when non-zero)
 - `top_p` (only when non-zero)
@@ -35,6 +35,35 @@ Serialized JSON 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.
 
+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
@@ -82,6 +111,13 @@ 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:
 
@@ -104,6 +140,8 @@ The following fields may exist in profile/effective settings but are not current
 - `reasoning_effort`
 - `extra_params`
 
+The client also does not serialize top-level `cache_control` or `session_id`.
+
 No built-in retries, tool-calls, or multi-request payload modes are implemented in this client.
 
 ## Relationship To Runner

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:
 
@@ -93,6 +95,9 @@ 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`.
+- 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.
 
@@ -142,3 +147,4 @@ Behavior highlights:
 - 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 cache control is message-level only; no top-level cache-control/session fields are serialized.

docs/internal/runner.md

@@ -108,9 +108,11 @@ Validation content failures are not run errors:
    - 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.
 
 ## Run Flow
@@ -123,7 +125,7 @@ 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/roadmap/cache.md

@@ -0,0 +1,136 @@
+# Prompt Cache Control Roadmap
+
+## Purpose
+
+Scriptorium should support provider prompt-cache controls for OpenAI-compatible gateways that expose Anthropic-style cache breakpoints, especially OpenRouter.
+
+The feature should preserve existing prompt definitions. Prompt authors should opt in with optional message-level cache metadata, and Scriptorium should report cache usage when compatible providers return it.
+
+Implementation steps belong in `docs/roadmap/implementation.md`.
+
+## Target State
+
+Prompt authors can mark a rendered message as an explicit cache breakpoint:
+
+```yaml
+messages:
+  - role: system
+    content_file: ../common/transcript.system.md
+    cache_control:
+      type: ephemeral
+      ttl: 1h
+```
+
+Existing messages without `cache_control` continue to render and serialize as string content:
+
+```json
+{
+  "role": "system",
+  "content": "rendered text"
+}
+```
+
+Messages with `cache_control` serialize as a single text content block:
+
+```json
+{
+  "role": "system",
+  "content": [
+    {
+      "type": "text",
+      "text": "rendered text",
+      "cache_control": {
+        "type": "ephemeral",
+        "ttl": "1h"
+      }
+    }
+  ]
+}
+```
+
+Scriptorium should parse and expose provider cache usage when compatible response fields are present, including cached prompt tokens and cache-write tokens.
+
+## Prompt Authoring Policy
+
+Cache breakpoints should be used for stable reusable prompt prefixes.
+
+Recommended ordering:
+
+1. Put stable, reusable context first.
+2. Put `cache_control` on the last stable message that should be part of the reusable prefix.
+3. Put per-run dynamic inputs after that breakpoint.
+
+Example:
+
+```yaml
+messages:
+  - role: system
+    content_file: ../common/transcript.system.md
+  - role: user
+    content_file: ./character_meta_analysis.task.md
+  - role: user
+    content_file: ./character_meta_analysis.instructions.md
+    cache_control:
+      type: ephemeral
+      ttl: 1h
+  - role: user
+    content: |
+      <<<PREVIOUS_SESSION_RECAP
+      {{input "recap"}}
+      PREVIOUS_SESSION_RECAP>>>
+  - role: user
+    content: |
+      <<<CURRENT_SESSION_TRANSCRIPT
+      {{input "transcript"}}
+      CURRENT_SESSION_TRANSCRIPT>>>
+```
+
+## Supported Cache-Control Shape
+
+Initial support is message-level only:
+
+```yaml
+cache_control:
+  type: ephemeral
+  ttl: 1h
+```
+
+Rules:
+
+- `cache_control` is optional on each message.
+- `cache_control.type` is required when `cache_control` is present.
+- The only supported `type` value is `ephemeral`.
+- `ttl` is optional.
+- When set, the only supported `ttl` value is `1h`.
+- Empty `ttl` is omitted from the outbound payload.
+- Prompt decoding remains strict; unknown cache-control fields are rejected.
+- Existing `content` and `content_file` rules remain unchanged.
+
+## Compatibility
+
+This feature should be backward compatible for existing prompt definitions.
+
+Compatibility requirements:
+
+- Existing prompt YAML without `cache_control` loads unchanged.
+- Existing `render` output remains valid.
+- Existing outbound request payloads remain string-content messages unless `cache_control` is configured.
+- Existing integrations do not need to send new request fields.
+- CLI and HTTP callers do not need new request options for the initial feature.
+
+The only intentional prompt-definition contract change is the new optional `messages[].cache_control` object.
+
+## Deferred Work
+
+These are intentionally out of scope for the initial feature:
+
+- `content_blocks` prompt syntax.
+- Multiple text blocks inside a single message.
+- Image, tool, or non-text content blocks.
+- Provider-specific automatic prompt caching toggles.
+- Top-level OpenRouter `cache_control`.
+- Top-level OpenRouter `session_id`.
+- General-purpose serialization of `extra_params`.
+- Provider-specific validation profiles for cache-control limits.
+
+These can be added later without changing the message-level cache-control contract.

docs/roadmap/implementation.md

@@ -0,0 +1,390 @@
+# Prompt Cache Control Implementation Plan
+
+## Scope
+
+This plan implements the feature defined in `docs/roadmap/cache.md`.
+
+The implementation target is message-level Anthropic/OpenRouter-style cache control for rendered prompt messages, plus cache usage observability. Do not implement deferred roadmap items such as `content_blocks`, top-level `cache_control`, `session_id`, or general `extra_params` serialization.
+
+Follow the policy documents under `docs/policy/` while implementing:
+
+- Keep prompt/profile/HTTP/config decoding strict.
+- Keep adapters thin.
+- Keep provider request serialization in `internal/llm`.
+- Keep prompt loading in `internal/promptdef`.
+- Update canonical non-roadmap docs only when the behavior is implemented.
+- Do not accept or emit raw secret values.
+
+## Stage 1: Domain Types And Prompt Loader
+
+Objective: make prompt YAML accept and validate optional `messages[].cache_control`, and carry normalized cache-control metadata in domain prompt definitions.
+
+Files to update:
+
+- `internal/domain/domain.go`
+- `internal/promptdef/filesystem_repository.go`
+- `internal/promptdef/repository_test.go`
+- `internal/promptdef/testdata/`
+
+Domain decisions:
+
+- Add a typed cache-control value in `internal/domain`.
+- Use a pointer on message structs so absence is distinguishable from an empty object.
+- Use `omitempty` on JSON tags for the new field so existing no-cache prompt and render hashes do not gain `null` fields.
+
+Use this domain shape:
+
+```go
+type CacheControlType string
+
+const (
+    CacheControlEphemeral CacheControlType = "ephemeral"
+)
+
+type CacheControl struct {
+    Type CacheControlType `yaml:"type" json:"type"`
+    TTL  string           `yaml:"ttl,omitempty" json:"ttl,omitempty"`
+}
+```
+
+Add cache control to:
+
+```go
+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"`
+}
+```
+
+Prompt loader changes:
+
+- Add this field to `promptMessageFile`:
+
+```go
+CacheControl *cacheControlFile `yaml:"cache_control"`
+```
+
+- Add a private `cacheControlFile` struct with `type` and `ttl`.
+- In `normalizePromptDefinition`, validate `cache_control` per message.
+- Trim whitespace from `type` and `ttl`.
+- Reject present `cache_control` when `type` is empty.
+- Reject any `type` other than `ephemeral`.
+- Accept omitted or empty `ttl`.
+- Reject any non-empty `ttl` other than `1h`.
+- Copy normalized values into `domain.PromptMessageTemplate.CacheControl`.
+- Preserve the existing rule that each message must set exactly one of `content` or `content_file`.
+
+Testing requirements:
+
+- Add a valid fixture with `cache_control: {type: ephemeral, ttl: 1h}`.
+- Add a valid fixture with `cache_control: {type: ephemeral}`.
+- Add invalid fixtures for:
+  - empty `cache_control.type`
+  - unsupported `cache_control.type`
+  - unsupported `cache_control.ttl`
+  - unknown field under `cache_control`
+- Assert valid fixtures load with the expected `CacheControl` values.
+- Assert invalid fixtures fail with `promptdef.ErrInvalidPromptDefinition` or `promptdef.ErrInvalidYAML` as appropriate.
+- Re-run:
+
+```bash
+go test ./internal/promptdef
+```
+
+Do not update renderer, formatter, LLM serialization, adapters, or user-facing docs in this stage.
+
+## Stage 2: Rendering, Prepared Output, And Render Hashing
+
+Objective: preserve cache-control metadata through rendering, expose it in prepared-run output, and make `rendered_prompt_hash` reflect outbound-affecting cache metadata.
+
+Files to update:
+
+- `internal/domain/domain.go`
+- `internal/prompt/go_renderer.go`
+- `internal/prompt/renderer_test.go`
+- `internal/usecase/runner.go`
+- `internal/usecase/runner_test.go`
+- `internal/domain/prepared_run_test.go`
+- `internal/format/prepared_run.go`
+- `internal/format/prepared_run_test.go`
+
+Domain changes:
+
+- Add cache control to rendered messages:
+
+```go
+type RenderedMessage struct {
+    Role         string        `json:"role"`
+    Content      string        `json:"content"`
+    CacheControl *CacheControl `json:"cache_control,omitempty"`
+}
+```
+
+Renderer changes:
+
+- When a `PromptMessageTemplate` has `CacheControl`, copy it to the returned `RenderedMessage`.
+- Do not mutate the template's pointer in place.
+- Add and use a small helper such as `cloneCacheControl` to avoid aliasing.
+- Do not change template execution behavior for `{{input "name"}}` or `{{.var}}`.
+
+Rendered hash decision:
+
+- `rendered_prompt_hash` must change when cache-control metadata changes, because the outbound provider request changes.
+- Preserve existing hashes for no-cache prompts by keeping the current role/content hash flow.
+- Append normalized cache-control metadata only when present, for example:
+  - `cache_control.type=<type>`
+  - `cache_control.ttl=<ttl>` only when non-empty
+- Include cache-control fields in a stable order.
+
+Prepared-run formatting:
+
+- JSON prepared-run output should include `cache_control` on messages only when present.
+- Text prepared-run output should show cache-control metadata under the specific message, before `content`, for example:
+
+```text
+      cache_control: ephemeral ttl=1h
+      content: |
+```
+
+- If `ttl` is empty, omit the `ttl=...` suffix.
+- Do not print cache-control lines for messages without cache control.
+
+Testing requirements:
+
+- Renderer test: cache control is copied to rendered messages.
+- Renderer test: changing returned rendered cache control does not mutate the source template.
+- Runner test: `rendered_prompt_hash` changes when cache control differs and remains stable for identical rendered messages.
+- Format JSON test: `cache_control` appears only when present.
+- Format text test: cache control appears under the correct message.
+- Existing render/format tests still pass.
+- Re-run:
+
+```bash
+go test ./internal/prompt ./internal/usecase ./internal/domain ./internal/format
+```
+
+Do not update outbound LLM serialization or adapter response surfaces in this stage.
+
+## Stage 3: Outbound Chat Serialization And Usage Parsing
+
+Objective: serialize cache-controlled messages as one-item text content-block arrays and parse cache usage fields from compatible provider responses.
+
+Files to update:
+
+- `internal/llm/openai_compatible_client.go`
+- `internal/llm/openai_compatible_client_test.go`
+- `internal/domain/domain.go`
+
+Request serialization decisions:
+
+- Keep string content for messages without cache control.
+- Use content blocks only when `RenderedMessage.CacheControl` is non-nil.
+- Do not introduce `content_blocks` prompt syntax.
+- Do not serialize top-level `cache_control`, top-level `session_id`, or `extra_params`.
+- Split request and response wire message structs. The current `openAIChatMessage` is used in both request and response paths; changing request content to `any` should not make response decoding less strict than necessary.
+
+Use request wire types equivalent to:
+
+```go
+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"`
+}
+```
+
+When cache control is present, content should serialize as:
+
+```json
+[
+  {
+    "type": "text",
+    "text": "rendered text",
+    "cache_control": {
+      "type": "ephemeral",
+      "ttl": "1h"
+    }
+  }
+]
+```
+
+When `ttl` is empty, omit `ttl`.
+
+Usage parsing decisions:
+
+- Extend `domain.TokenUsage`:
+
+```go
+type TokenUsage struct {
+    PromptTokens     int
+    CompletionTokens int
+    TotalTokens      int
+    CachedTokens     int
+    CacheWriteTokens int
+}
+```
+
+- Parse `usage.prompt_tokens_details.cached_tokens` into `CachedTokens`.
+- Parse `usage.cache_write_tokens` into `CacheWriteTokens`.
+- Treat absent cache fields as zero.
+- Do not fail response decoding if cache fields are absent.
+- If a provider later returns a different cache-write field, add it in a later change after observing a real response.
+
+Testing requirements:
+
+- Existing request serialization test still sees string message content for no-cache messages.
+- New request serialization test sees content-block array for cache-controlled messages.
+- New request serialization test verifies empty `ttl` is omitted.
+- Response parsing test verifies cached and cache-write tokens are mapped.
+- Response parsing test verifies absent cache usage fields remain zero.
+- Existing structured-output and service-tier behavior remains unchanged.
+- Re-run:
+
+```bash
+go test ./internal/llm
+```
+
+Do not update CLI or HTTP observability in this stage unless required by compilation after extending `TokenUsage`.
+
+## Stage 4: CLI And HTTP Cache Usage Observability
+
+Objective: surface parsed cache usage through existing CLI and HTTP result paths without adding request knobs.
+
+Files to update:
+
+- `internal/adapter/cli/run.go`
+- `internal/adapter/cli/run_test.go`
+- `internal/adapter/http/dto.go`
+- `internal/adapter/http/handler.go`
+- `internal/adapter/http/handler_test.go`
+
+CLI decisions:
+
+- Do not add CLI flags.
+- Keep success summaries concise.
+- Preserve the existing summary output when cache usage is zero.
+- When either cache field is non-zero, append:
+
+```text
+cached_tokens=<n> cache_write_tokens=<n>
+```
+
+HTTP decisions:
+
+- Do not add request fields.
+- Add numeric fields to `metadata.usage`:
+  - `cached_tokens`
+  - `cache_write_tokens`
+- Always include these fields as numbers, matching the current non-omitempty numeric usage fields.
+- Continue omitting raw model output unless `include_raw_output` is true.
+
+Testing requirements:
+
+- CLI summary omits cache fields when both are zero.
+- CLI summary includes cache fields when either is non-zero.
+- HTTP success response includes `metadata.usage.cached_tokens`.
+- HTTP success response includes `metadata.usage.cache_write_tokens`.
+- Existing HTTP request mapping is unchanged.
+- Re-run:
+
+```bash
+go test ./internal/adapter/cli ./internal/adapter/http
+```
+
+## Stage 5: Canonical Documentation
+
+Objective: update implemented-behavior docs after the feature works and tests pass.
+
+Files to update:
+
+- `docs/config.md`
+- `docs/integrations/openai-compatible-chat.md`
+- `docs/cli.md`
+- `docs/troubleshooting.md`
+- `docs/internal/runner.md`
+- `docs/internal/adapters.md`, if adapter response behavior needs an explicit note
+- `docs/roadmap/cache.md`, only if the accepted target changes during implementation
+
+Documentation requirements:
+
+- Keep all unimplemented ideas under `docs/roadmap/`.
+- Document `messages[].cache_control` in the prompt definition reference in `docs/config.md`.
+- Document the outbound message serialization rule in `docs/integrations/openai-compatible-chat.md`.
+- Document parsed cache usage fields in `docs/integrations/openai-compatible-chat.md`.
+- In `docs/cli.md`, note that prompt cache control is configured in prompt YAML, not with CLI flags.
+- In `docs/troubleshooting.md`, add cache miss guidance:
+  - stable context should appear before the cache breakpoint
+  - dynamic input before the breakpoint changes the cache key
+  - providers may impose minimum token thresholds and breakpoint limits
+  - use cached/cache-write token counts to verify behavior
+- Do not claim support for deferred roadmap items.
+
+Testing and verification:
+
+```bash
+go test ./...
+```
+
+If documentation examples include new YAML snippets, ensure they match the strict prompt loader schema.
+
+## Stage 6: Final Verification
+
+Objective: verify the whole change set end to end.
+
+Required commands:
+
+```bash
+go test ./...
+go build ./cmd/scriptorium
+```
+
+Recommended smoke checks:
+
+```bash
+go run ./cmd/scriptorium render \
+  --config ./examples/config.yml \
+  --prompt generic.markdown_summary \
+  --input transcript=./examples/fixtures/transcript.md \
+  --input glossary=./examples/fixtures/glossary.yml \
+  --format json
+```
+
+If adding a new example prompt with `cache_control`, also smoke-render that prompt with `--format json` and inspect that the rendered message includes `cache_control`.
+
+Final review checklist:
+
+- Existing prompt definitions without `cache_control` still load.
+- Prompt definitions with `cache_control` render successfully.
+- Unknown cache-control YAML fields are rejected.
+- No-cache outbound messages still serialize with string `content`.
+- Cache-controlled outbound messages serialize with text content-block arrays.
+- `rendered_prompt_hash` changes when cache-control metadata changes.
+- Cache usage fields are zero when providers omit them.
+- CLI and HTTP request surfaces have no new knobs.
+- CLI and HTTP response surfaces expose cache usage.
+- Non-roadmap docs describe only implemented behavior.
+- `docs/roadmap/cache.md` still defines target state and deferred work, not task steps.
+
+## Stage Boundaries
+
+Stages 1 through 4 can be implemented in separate prompts if needed. Stage 5 should happen only after behavior is implemented. Stage 6 should be performed after all implementation and documentation changes are complete.
+
+If a stage fails because prior stage changes were not implemented, stop and complete the earlier stage first rather than papering over missing domain fields or DTO mappings.
+
+## Open Questions
+
+None required for the initial implementation.
+
+The plan chooses message-level `cache_control`, string content for legacy messages, text-block array content only when cache control is present, and no new CLI/HTTP request knobs. These choices match `docs/roadmap/cache.md` and preserve existing prompt-definition compatibility.

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

@@ -606,7 +606,7 @@ 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 +620,10 @@ 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

@@ -1064,6 +1064,38 @@ 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

@@ -86,6 +86,8 @@ 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

@@ -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,

internal/adapter/http/handler_test.go

@@ -66,7 +66,13 @@ 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 +117,13 @@ 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"])
@@ -171,6 +184,10 @@ 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) {

internal/domain/domain.go

@@ -40,6 +40,19 @@ const (
 	ValidationSkipped ValidationStatus = "skipped"
 )
 
+// CacheControlType defines provider cache behavior for prompt content.
+type CacheControlType string
+
+const (
+	CacheControlEphemeral CacheControlType = "ephemeral"
+)
+
+// 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
@@ -134,6 +147,7 @@ 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.
@@ -182,6 +196,7 @@ 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.
@@ -222,6 +237,8 @@ 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 +53,52 @@ 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])
+	}
+}

internal/format/prepared_run.go

@@ -151,6 +151,13 @@ 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 == "" {

internal/format/prepared_run_test.go

@@ -62,6 +62,58 @@ 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 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()
 
@@ -98,6 +150,47 @@ 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

@@ -151,6 +151,8 @@ 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
 }
@@ -168,12 +170,9 @@ func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultMod
 		Model: model,
 	}
 
-	wireReq.Messages = make([]openAIChatMessage, 0, len(req.Prompt.Messages))
+	wireReq.Messages = make([]openAIChatRequestMessage, 0, len(req.Prompt.Messages))
 	for _, msg := range req.Prompt.Messages {
-		wireReq.Messages = append(wireReq.Messages, openAIChatMessage{
+		wireReq.Messages = append(wireReq.Messages, openAIChatRequestMessageFromRenderedMessage(msg))
-			Role:    msg.Role,
-			Content: msg.Content,
-		})
 	}
 
 	if req.Target.Temperature != 0 {
@@ -201,7 +200,7 @@ func openAIChatRequestFromGenerateRequest(req domain.GenerateRequest, defaultMod
 
 type openAIChatRequest struct {
 	Model          string                     `json:"model"`
-	Messages       []openAIChatMessage   `json:"messages"`
+	Messages       []openAIChatRequestMessage `json:"messages"`
 	Temperature    *float64                   `json:"temperature,omitempty"`
 	MaxTokens      *int                       `json:"max_tokens,omitempty"`
 	TopP           *float64                   `json:"top_p,omitempty"`
@@ -209,19 +208,39 @@ type openAIChatRequest struct {
 	ResponseFormat *openAIResponseFormat      `json:"response_format,omitempty"`
 }
 
-type openAIChatMessage struct {
+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 +255,28 @@ 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

@@ -89,6 +89,9 @@ 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 +147,155 @@ 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", "session_id", "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 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) {

internal/prompt/go_renderer.go

@@ -77,6 +77,7 @@ 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),
 		})
 	}
 
@@ -84,3 +85,11 @@ func (r *goRenderer) Render(ctx context.Context, definition *domain.PromptDefini
 		Messages: renderedMessages,
 	}, nil
 }
+
+func cloneCacheControl(in *domain.CacheControl) *domain.CacheControl {
+	if in == nil {
+		return nil
+	}
+	out := *in
+	return &out
+}

internal/prompt/renderer_test.go

@@ -78,6 +78,66 @@ 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}},

internal/promptdef/filesystem_repository.go

@@ -45,6 +45,12 @@ 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 +218,11 @@ 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 +244,7 @@ func normalizePromptDefinition(raw *promptDefinitionFile, sourcePath string) (*d
 			Role:         role,
 			Content:      templateContent,
 			ContentFile:  resolvedContentFile,
+			CacheControl: cacheControl,
 		})
 	}
 
@@ -274,6 +286,30 @@ 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 +68,34 @@ 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 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 +286,10 @@ 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 +314,19 @@ 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/usecase/runner.go

@@ -428,6 +428,14 @@ func hashRenderedPrompt(p domain.RenderedPrompt) string {
 		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

@@ -577,6 +577,61 @@ 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 TestRunnerRunSuccessful(t *testing.T) {
 	promptRepo := &fakePromptRepo{def: promptDef(domain.FormatMarkdown, domain.ValidationBasic, 0)}
 	execRepo := &fakeExecutionProfileRepo{profiles: map[string]*domain.ExecutionProfile{"exec": defaultExecutionProfile()}}