Anthropic prompt caching is opt-in per request: without explicit
ephemeral cache_control breakpoints in the body, no cache writes or
reads ever happen. The OpenAI-compatible codec never emitted them, so
every run on openrouter Claude models billed the full conversation at
the uncached input rate on every turn (0 cache tokens on the billing
page, confirmed by OpenRouter's activity portal).
- Add a `cache_control_breakpoints` model feature declaring that a
route only caches when the request marks the cacheable prefix; set it
on the builtin OpenRouter Claude rows. Catalog build rejects the flag
without `prompt_cache`.
- Teach the Chat Completions wire shape a parts-form content variant so
a message can carry the annotation; unmarked messages keep the
plain-string form for compatibility with strict servers.
- Mark the last system message (covers tools + system upstream) and the
second-to-last user turn, counting tool results as user turns —
mirroring the anthropic codec's placement so agent loops get
incremental cache hits.
- Extract the shared placement/opt-out policy into codec::cache and
refactor the anthropic codec onto it; anthropic wire snapshots are
unchanged.
- Honor `provider_options.<name>.auto_cache = false` as an opt-out and
consume the control key instead of merging it into the body.
- Mirror the new feature through settings (fabro-config), the OpenAPI
schema, and the generated TypeScript client.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
e5f0290229