[berlineric]

Description

OpenAI-compatible vLLM backends can report prompt cache writes as usage.prompt_tokens_details.created_cache_tokens. PromptTokensDetails currently accepts cache_write_tokens and cache_creation_input_tokens, so these cache writes are not captured in gateway accounting.

Accept created_cache_tokens through the existing CacheWriteTokensValue accessor. Preserve the existing greatest-value normalization across aliases without summing duplicate counts. Keep outgoing JSON normalized to the existing official/legacy fields.

The shared accessor is used by streaming and non-streaming chat completion translators. Add schema tests and translator tests for cold prompts, warm prompts, and overlapping aliases.

Validation

• go test ./internal/apischema/openai ./internal/translator -count=1 passed.
• git diff --check passed.
• make precommit attempted; blocked at API documentation generation with make: go: Permission denied. Draft until the full precommit gate is resolved. Unrelated generated changes were discarded.
• No deployed gateway images or cluster configurations changed.

AI Assistance

GitHub Copilot assisted with implementation and regression tests.

Release Notes

Accept vLLM created_cache_tokens usage as cache-write accounting in OpenAI-compatible chat responses.

[netlify]

✅ Deploy Preview for theagentrouter canceled.

Name	Link
🔨 Latest commit	a4c2686
🔍 Latest deploy log	https://app.netlify.com/projects/theagentrouter/deploys/6af04418f938d9fb085ac06c

[berlineric]

Withdrawing this draft: the requesting deployment now needs only cached_tokens/cache-read accounting and does not need the proposed cache-creation alias support.