feat(usage): surface Claude thinking token counts to clients

[yidecode]

Stacked on #2762 — that one has to land first. Adding reasoning_tokens for Claude while calculateCostFromTokens still adds reasoning on top of completion would make Claude join the double-billing #2762 removes. This branch contains #2762's commit; it will rebase to a single commit once that merges.

Anthropic reports thinking tokens on message_delta as usage.output_tokens_details.thinking_tokens. Nothing consumed it — the Claude branches of extractUsage, extractUsageFromResponse and toOpenAIUsage's claude extractor all stopped at input/output/cache — so the count never reached cost tracking, the request-detail record, or the client.

The non-obvious part

Extracting the count is not enough. claude-to-openai builds its own usage object and assigns it to the shared stream state, mixing Claude field names into it:

state.usage = { prompt_tokens, completion_tokens, total_tokens,
                input_tokens, output_tokens };
state.usage.output_tokens_details = { thinking_tokens };

stream.js then hands that object to filterUsageForFormat before emitting the final chunk — and the OpenAI allow-list has no output_tokens_details, because that is a Claude name. So a naive fix extracts the count correctly and still drops it on the way out. The stored request-detail record shows the leak exactly: a hybrid {prompt_tokens, …, input_tokens, output_tokens, output_tokens_details} object that no single format's allow-list fully matches.

The count is therefore mirrored under both OpenAI spellings (reasoning_tokens and completion_tokens_details.reasoning_tokens), and output_tokens_details is added to the Claude allow-list so Claude-format clients keep it too.

completion_tokens is deliberately left alone — thinking tokens are already inside output_tokens, so the reasoning-inclusive convention #2762 relies on still holds.

Why it matters

GitHub Copilot returns a signed but empty thinking block for its 4.7+ Claude shims (claude-sonnet-5, claude-opus-4.7, claude-opus-4.8, claude-fable-5) — the reasoning happens, the text is withheld server-side. Verified end-to-end against a live gateway:

sonnet-4.6   reasoning_content=1012ch   reasoning_tokens=426
sonnet-5     reasoning_content=   0ch   reasoning_tokens=886

Without the count, that second row is indistinguishable from a broken pipeline — which is exactly how it gets reported. With it, a client can tell "withheld upstream" from "the proxy ate my thinking".

Verification

• 4 new tests driving the real createSSETransformStreamWithLogger(CLAUDE → OPENAI) transform, including the withheld-text case and the Claude-format allow-list.
• Full suite: failure set identical to the master baseline, zero regression.

本 PR 由 Claude Code 辅助完成