v0.7.18

LLM Council 0.7.18 Release Notes

Highlights

• Added provider-specific prompt-cache support without pretending every provider
  uses the same cache model.
• Added Gemini API and Vertex Gemini cached-content lifecycle handling:
  create, TTL refresh, expiry recreation, best-effort cleanup, billing warnings,
  and response metadata.
• Added OpenRouter route-gated prompt-cache passthrough for eligible Anthropic
  routes.
• Added OpenAI cache telemetry parsing without sending unsupported request
  controls.

Live Proof

• OpenRouter live proof returned cache_read_tokens=8116 on the second
  identical call through an Anthropic route.
• Gemini API live proof created cachedContents/..., returned
  cache_read_tokens=12603, and deleted the resource successfully.
• Vertex Gemini live proof created projects/.../cachedContents/..., returned
  cache_read_tokens=14002, and deleted the resource successfully.

Live proof currently covers non-streaming generation paths. Streaming
cached-content lifecycle behavior is covered by regression tests and documented
with a stricter safety contract: retry expiry only before any chunk is yielded,
then emit final cleanup metadata on successful streams.

Upgrade Notes

• Install all provider SDKs with:

pip install 'the-llm-council[all]>=0.7.18'

• Gemini/Vertex cached-content creation requires explicit stable source_text;
  Council does not infer cacheable source content from dynamic prompts.
• Created or refreshed cached-content resources can incur storage-duration
  billing until TTL expiry or successful cleanup.

Validation

• Focused cached-content lifecycle tests: 28 passed.
• Broader prompt-cache suite: 157 passed.
• Full suite: 431 passed.
• Final council review: approved with no blocking issues.