hermes - 💡(How to fix) Fix API server /v1/responses drops agent reasoning from output_items

Problem

The api_server platform's /v1/responses endpoint (both SSE streaming and batch paths) emits function_call, function_call_output, and message output items, but never emits the agent's reasoning as a reasoning output_item — even though run_agent.py already produces reasoning text and fires reasoning_callback during streaming.

This breaks downstream consumers that persist output[] from the Responses API and expect the OpenAI Responses spec shape {type:'reasoning', summary, content}. CLI/Telegram/Slack get a preview via _emit_reasoning_preview, but Responses-API clients lose the chain-of-thought entirely.

Reproduction

1. Run hermes-agent exposing /v1/responses.
2. Use a model/profile that emits reasoning (e.g. a thinking-enabled OpenAI/Anthropic model).
3. POST to /v1/responses with stream:true (or stream:false).
4. Inspect response.completed.output[] (or the GET /v1/responses/{id} snapshot).

Observed: output[] contains only function_call, function_call_output, message. No reasoning item.

Expected: A reasoning item (per OpenAI Responses spec) carrying summary[].summary_text and content[].reasoning_text, ordered before the final message.

Root cause

In gateway/platforms/api_server.py:

• _create_agent (api_server.py:830) builds AIAgent without passing reasoning_callback.
• _dispatch (api_server.py:1695) only routes __tool_started__ / __tool_completed__ tags + plain text deltas — no __reasoning__ branch.
• _extract_output_items (api_server.py:2563) (batch path) only emits function_call / function_call_output / message.

AIAgent._fire_reasoning_delta exists and is invoked during streaming in run_agent.py, but its target callback is never installed in the api_server platform.

Proposed fix

Wire reasoning_callback end-to-end:

1. _create_agent / _run_agent accept reasoning_callback kwarg, forward to AIAgent.
2. _handle_responses installs an _on_reasoning callback that pushes ("__reasoning__", text) onto the SSE queue.
3. _dispatch buffers reasoning deltas in reasoning_parts.
4. New _flush_reasoning() helper emits a single response.output_item.added/done pair (type:'reasoning', with summary + content text parts) before the assistant message item closes — preserving the canonical order function_calls → reasoning → message.
5. The incomplete-snapshot path (early disconnect) also appends the buffered reasoning, so GET /v1/responses/{id} still surfaces thinking.

Happy to send a PR (already have a working patch + 2 new tests in tests/gateway/test_api_server.py, 140 passed).