lsm / neokai / 25473190811

81%

push

github

Add OpenAI reasoning.effort support to anthropic-codex bridge (#1810)

* Add OpenAI reasoning.effort support to anthropic-codex bridge

- Enable extendedThinking and granular thinkingModes for anthropic-codex provider
- Update PROVIDER_THINKING_MODES to 'granular' for anthropic-codex
- Map ThinkingLevel budget_tokens → OpenAI reasoning.effort (low/medium/high/xhigh)
- Add reasoning.encrypted_content inclusion for multi-turn with store:false
- Handle 6 new OpenAI SSE reasoning events, translating to Anthropic thinking blocks
- Pass reasoning items back into subsequent turns' input array
- Map output_tokens_details.reasoning_tokens to thinking_tokens in usage
- Add SSE helpers for thinking blocks in translator
- Add tests for reasoning parameter mapping, SSE translation, multi-turn pass-through, and token accounting

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(openai-responses-bridge): keep prior reasoning items until stream replaces them

Remove the premature sessionReasoningItems.delete(sessionId) on HTTP OK
so that reasoning items survive SDK-level retries and are only overwritten
when the stream actually emits new reasoning items via onReasoningItems.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(openai-responses-bridge): always update reasoning cache and include encrypted_content when reusing

- Always call onReasoningItems after response.completed, even with an empty
  array, so stale reasoning is cleared when a turn returns no reasoning.
- Include reasoning.encrypted_content in the request whenever cached reasoning
  items are present, even if the current turn has no thinking config.

Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>

* fix(openai-responses-bridge): align token estimates with forced full-history mode

When cached reasoning items exist the bridge disables previous_response_id
and sends full history plus injected reasoning items, but token estimation
still followed the old continuation/bod... (continued)

8199 of 11342 branches covered (72.29%)

Branch coverage included in aggregate %.

189 of 192 new or added lines in 6 files covered. (98.44%)

63163 of 75174 relevant lines covered (84.02%)

287.55 hits per line