feat(api-server): surface reasoning_content on /v1/chat/completions

[adaofeliz]

Problem

/v1/runs and /v1/responses already relay reasoning.available previews as structured events, but /v1/chat/completions never wired a callback to receive them — so reasoning/thinking text is silently dropped for that endpoint, even though it's the one most OpenAI-SDK clients (Open WebUI, plain openai SDK, etc.) actually use.

Fix

Add a reasoning.available-filtered tool_progress_callback to both the streaming and non-streaming paths of _handle_chat_completions:

• Streaming: emits delta.reasoning_content chunks — the same field name DeepSeek/Moonshot/OpenRouter already use for thinking traces — as a separate SSE data: line, distinct from delta.content and the existing event: hermes.tool.progress tool-lifecycle events.
• Non-streaming: accumulates previews into message.reasoning_content on the final JSON response.

Tool-call events are unaffected: the new callback only forwards reasoning.available and ignores every other tool_progress_callback event type, so tool_start_callback/tool_complete_callback still own the tool-lifecycle channel exactly as before (no duplicate emits, per the prior comment this PR updates).

Testing

• Added test_stream_includes_reasoning_content and test_non_streaming_includes_reasoning_content to tests/gateway/test_api_server.py::TestChatCompletionsEndpoint, mirroring the existing test_stream_includes_tool_progress pattern.
• python3 -m pytest tests/gateway/test_api_server.py -q → 87 passed (same 7 pre-existing failures as unmodified main, unrelated to this change — missing dotenv/network fixtures in a fresh checkout).
• ruff check clean on both changed files.
• Manually verified end-to-end against a running gateway (/v1/runs, which shares _create_agent/_run_agent) that reasoning surfaces without any tool-call leakage.

Note

On an Idempotency-Key cache hit (non-streaming), reasoning_content will be absent even if the original computation had reasoning — the idempotency cache only stores (result, usage), not the side-channel reasoning previews. This mirrors how streaming chunks also aren't replayed on cache hit, so it's a pre-existing limitation of that cache's shape rather than something this PR needs to solve.

[teknium1]

Thanks for tracing the missing chat-completions callback path. The endpoint gap is real, but this implementation captures only the fallback preview channel rather than provider reasoning deltas.

Problems

• gateway/platforms/api_server.py proposed lines 3932 and 3964 wire tool_progress_callback only. Provider reasoning_content is dispatched via agent/chat_completion_helpers.py:3210-3214 → AIAgent._fire_reasoning_delta(), which invokes only reasoning_callback in run_agent.py:5831-5842. The API adapter does not currently thread a reasoning_callback through _create_agent (gateway/platforms/api_server.py:2348-2355, 2642-2659), so actual provider thinking deltas remain dropped.
• The added tests manually emit reasoning.available through tool_progress_callback; they do not cover the reasoning_callback transport that supplies reasoning_content.

Suggested changes

• Thread reasoning_callback through _run_agent and _create_agent, then use it for the SSE chunks and non-streaming accumulation.
• Test that callback directly and retain a combined lifecycle assertion so tool events remain single-emitted.

Automated hermes-sweeper review.

[teknium1]

This callback only receives reasoning.available previews. Provider reasoning deltas flow through AIAgent._fire_reasoning_delta() → reasoning_callback (run_agent.py:5831-5842), so please thread and use reasoning_callback here as well; otherwise standard reasoning_content streaming remains absent.