feat(api-server): stream model reasoning as reasoning_content in chat completions

[ladyada-piclaw]

Summary

The /v1/chat/completions streaming path wired stream_delta_callback, tool_start_callback, and tool_complete_callback, but never reasoning_callback. As a result, a model's reasoning stream (e.g. Claude extended thinking, thinking_delta → agent._fire_reasoning_delta) was silently dropped for OpenAI-compatible frontends.

Open WebUI (and other OpenAI-compatible UIs) render a collapsible "thinking" block from delta.reasoning_content chunks — but nothing on this path emitted them.

Change

Thread reasoning_callback through the existing call chain and route reasoning deltas to a dedicated SSE chunk:

• _create_agent and _run_agent gain a reasoning_callback parameter, forwarded to AIAgent (which already accepts and fires it via _fire_reasoning_delta).
• In the streaming branch of _handle_chat_completions, a new _on_reasoning(text) enqueues a tagged ("__reasoning__", text) item (mirroring the existing ("__tool_progress__", …) pattern).
• _emit gains an elif that renders that tagged item as a chat.completion.chunk carrying delta.reasoning_content.

Visible delta.content, the hermes.tool.progress events, and the end-of-stream sentinel are all unchanged. Empty/None reasoning chunks are skipped.

Real behavior / verification

• Module imports cleanly; both _create_agent and _run_agent signatures carry reasoning_callback (verified via inspect.signature).
• End-to-end chain confirmed in source: Anthropic streaming thinking_delta (chat_completion_helpers.py) → agent._fire_reasoning_delta → self.reasoning_callback (assigned in agent_init.py) → _on_reasoning → _stream_q → _emit.
• Deterministic unit-level reproduction of the _emit branch confirms a tagged reasoning item serializes to {"delta": {"reasoning_content": "..."}} while normal content still serializes to {"delta": {"content": "..."}}.

Scope / safety

• No new model tools; agent loop and prompt caching untouched.
• Additive only (+23 lines, one file); content/tool-progress paths byte-identical when no reasoning is emitted.

[alt-glitch]

This was generated by AI during triage.
Duplicate of #13401 — same mechanism and same file (gateway/platforms/api_server.py): thread reasoning_callback through the /v1/chat/completions streaming path and emit delta.reasoning_content SSE chunks. #13401 is the earlier (2026-04-21) canonical open PR. Broader siblings #39006 and #48024 are related. Maintainer picks one.

[teknium1]

Thanks for the focused Chat Completions reasoning-streaming fix. The premise is confirmed on current main: _handle_chat_completions forwards only text and tool callbacks at gateway/platforms/api_server.py:2311-2321, while _write_sse_chat_completion only special-cases tool-progress tuples at gateway/platforms/api_server.py:2499-2510. The agent-side callback is live: Anthropic thinking_delta reaches _fire_reasoning_delta in agent/chat_completion_helpers.py:2675-2679.

Problems

• This PR changes only gateway/platforms/api_server.py; gh pr diff 52606 contains no regression test. Add coverage in tests/gateway/test_api_server.py for callback forwarding and the emitted delta.reasoning_content chunk.
• The supplied discussion points to open #57094 as the broader active implementation, covering this path plus non-streaming Chat and Responses streaming. This PR remains a valid narrow fix, but consolidation should avoid landing overlapping callback plumbing independently.

Suggested changes

• Mock _run_agent, invoke its received reasoning_callback, and assert reasoning and visible text serialize into their respective delta fields.
• Add a forwarding-seam test through _run_agent and _create_agent.

This is an automated hermes-sweeper review.

[teknium1]

Please add an API-server SSE regression test that invokes this callback through the mocked _run_agent call and verifies it serializes as delta.reasoning_content rather than delta.content; the current PR changes no test files.

[Addonniss]

Thanks for the review. #57094 (disclosure: my PR) already has tests covering both streaming delta.reasoning_content and non-streaming reasoning_content in the response body, plus Responses API support. It's been rebased onto latest main and all 198 tests pass. Happy to close this in favor of #57094 if that helps consolidate.

One ask: could the duplicate label be removed from #57094 so it gets a proper review? #13401 was already closed by its author in favor of #57094.