fix(gateway): forward reasoning in chat completions (#37044)

[rodboev]

Summary

/v1/chat/completions exposed assistant text but dropped reasoning generated by the provider. The streaming route had no path from AIAgent.reasoning_callback into its SSE writer, and the non-streaming message discarded the completed turn's last_reasoning value. Using the derived reasoning.available tool-progress event would have mislabeled final answer text and capped the output.

This rebuild forwards the existing provider reasoning callback through the API adapter and emits its deltas as OpenAI-compatible delta.reasoning_content. The callback uses current main's ThreadSafeAsyncQueue and shared _sse_frame() path. Non-streaming responses include the current turn's reasoning_content under the same display.show_reasoning policy. The display policy defaults to hidden and applies the existing API-server override precedence.

Fixes #37044

Changes

• gateway/platforms/api_server.py: pass provider reasoning through _run_agent() and _create_agent(), feed tagged deltas through ThreadSafeAsyncQueue, serialize them with _sse_frame(), and include non-empty last_reasoning in non-streaming messages when display is enabled.
• tests/gateway/test_api_server.py: cover callback construction and forwarding, full reasoning payloads, reasoning-only output, empty and disabled reasoning, tool lifecycle coexistence, and non-streaming response shapes.

Validation

Scenario	Behavior
Streaming provider reasoning with display enabled	Preserved byte-for-byte in delta.reasoning_content; assistant text remains in delta.content.
Reasoning-only output	Emits reasoning and normal SSE completion without fabricated assistant content.
Empty reasoning or display disabled	Omits reasoning_content while answer, tool lifecycle, and completion output continue.
Structured tool lifecycle	Retains one correlated running and completed event without using the tool-progress reasoning preview.
Non-streaming response	Adds current-turn message.reasoning_content only when non-empty and display is enabled.
/v1/responses, /v1/runs, session chat, messaging	Retain their existing endpoint-specific reasoning and event contracts.

Test plan

• bash scripts/run_tests.sh tests/gateway/test_api_server.py -k "reasoning or run_agent or stream_includes_tool_progress or stream_emits_tool_lifecycle or stream_tool_lifecycle" -v --timeout=0 — 15 passed.
• bash scripts/run_tests.sh tests/gateway/test_api_server.py -k "stream_includes_tool_progress or stream_emits_tool_lifecycle or stream_tool_lifecycle_skips_internal or stream_cancelled_persists_incomplete_snapshot or stream_client_disconnect_persists_incomplete_snapshot" -v --timeout=0 — 5 passed.
• ruff check gateway/platforms/api_server.py tests/gateway/test_api_server.py — passed.

Not in scope

/v1/responses, /v1/runs, native session chat, and messaging platforms keep their established response formats and callback contracts.

[teknium1]

Thanks for covering the non-streaming response shape; current main still drops last_reasoning from /v1/chat/completions at gateway/platforms/api_server.py:2417-2420.

Problems

• The streaming branch added at gateway/platforms/api_server.py:2027 uses tool_progress_callback and reasoning.available. On current main, that event is derived from stripped assistant_message.content and capped at 500 characters (agent/conversation_loop.py:4334-4350), not from provider reasoning. Native reasoning deltas use reasoning_callback (run_agent.py:4745-4752; agent/chat_completion_helpers.py:2057-2065). This can label answer text as delta.reasoning_content and misses structured/reasoning-only outputs.
• The new streaming test manually fires reasoning.available, so it does not cover the production reasoning callback chain.

Suggested changes

• Thread reasoning_callback through _run_agent and _create_agent, then queue and serialize its deltas as delta.reasoning_content.
• Update the stream test to invoke that callback and cover structured/reasoning-only output.

Automated hermes-sweeper review.

[rodboev]

Good catch. The streaming path was treating reasoning.available as provider reasoning even though current main synthesizes that event from final assistant text.

I removed that source from the SSE path and threaded a dedicated reasoning_callback through APIServerAdapter._run_agent() and _create_agent() into AIAgent. Streaming responses now forward provider reasoning through that callback as delta.reasoning_content, byte-for-byte, while assistant content stays in delta.content. Non-streaming reasoning_content still comes from last_reasoning and stays gated by display.show_reasoning.

The new regression enters _run_agent(), reaches _create_agent(), invokes the installed provider callback, and verifies that reasoning-only and longer-than-500-character output survives unchanged. /v1/runs, /v1/responses, and the structured tool lifecycle callbacks are unchanged.

[logabell]

Just want to comment here that this would be a much appreciated feature for my integration with OpenWebUI