Preserve provider reasoning and Codex summaries in buffered responses

[acastellana]

Buffered responses discarded reasoning fields returned by providers, preventing downstream agents from capturing their complete response. Preserve provider-exposed reasoning and usage in buffered chat, including stream-backed aggregation, and retain Codex summaries and opaque reasoning items in chat and Responses output.

• Forward explicit Codex reasoning effort/summary controls through the existing OAuth backend. No new authentication flow or default reasoning enablement.
• Keep truly empty responses eligible for fallback. Empty/whitespace reasoning and metadata-only details do not satisfy the first-output deadline; nonempty summaries, text and opaque payloads remain intact.
• Replay Responses reasoning items and summary events before subsequent message/tool items, with contiguous output indexes and increasing sequence numbers.
• Keep response content out of bounded diagnostics and operational ledger summaries. Tests verify both preservation in the response and exclusion from diagnostics.

Validation: 152 focused tests passed locally, including empty-response/deadline regressions and reconstruction of the actual /v1/responses SSE output. Full CI on 4c2364b: 1,131 Python tests passed (2 skipped), 729 core tests passed, and both images built with startup/import smoke checks passing. CI run. Branch includes current main and the PR119 probability-rounding fix.

Scope: upstream providers are mocked in tests; no live provider/OAuth request or deployment performed. Public chat reasoning deltas, native non-compatible adapters and downstream trace persistence remain outside this change. Compatible SSE reasoning-details fragments are preserved as received; failed Codex streams may still discard partial summaries. This preserves only provider-exposed data, not hidden internal reasoning.

Rollback: revert this PR; no database or configuration migration.

Summary by CodeRabbit

• New Features

  • Provider reasoning summaries are now preserved in buffered and streaming chat responses.
  • Reasoning-only responses are supported without being treated as empty assistant messages.
  • Responses API streams reasoning items and summary events before subsequent text or tool-call output.
  • Reasoning token usage is included when provided by the upstream model.
  • Codex reasoning controls, including effort and summary settings, are forwarded.

• Documentation

  • Added documentation covering reasoning behavior, supported controls, usage reporting, and scope limitations.

[coderabbitai]

Note

Currently processing new changes in this PR. This may take a few minutes, please wait...

⚙️ Run configuration

Configuration used: defaults

Review profile: CHILL

Plan: Advanced

Run ID: 4fef68a6-5f0e-478e-bc5a-abf853e86702

📥 Commits

Reviewing files that changed from the base of the PR and between a045d19 and 4c2364b.

📒 Files selected for processing (9)

• README.md
• codex_backend.py
• provider_adapters/openai_compatible.py
• responses_api.py
• shim.py
• tests/test_provider_diagnostics.py
• tests/test_reasoning_capture.py
• tests/test_responses_api.py
• tests/test_responses_shim.py

 ________________________________________________
< PENDU: Prompt Engineered Nifty Debugging Unit. >
 ------------------------------------------------
  \
   \   \
        \ /\
        ( )
      .( o ).

✨ Finishing Touches

📝 Generate docstrings

• Commit to this branch
• Create a new PR

🧪 Generate unit tests (beta)

• Commit to this branch
• Create a new PR

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}