Summary

Fixes #37044.

The OpenAI-compatible API server adapter was silently dropping the last_reasoning field from the agent result, so downstream UIs (Open WebUI, LobeChat, etc.) connected to the Hermes gateway never saw the model's chain-of-thought even when the model produced a visible reasoning block in the TUI / CLI.

This change adds an opt-in X-Hermes-Expose-Reasoning request header. When set to a truthy value, the gateway surfaces the reasoning/thinking content in BOTH paths:

• Non-streaming /v1/chat/completions: message.reasoning_content and message.reasoning are added to the assistant message
• Streaming /v1/chat/completions SSE: delta.reasoning_content and delta.reasoning chunks on the chat.completion.chunk stream

reasoning_content is the de facto standard used by Open WebUI / DeepSeek / OpenRouter / Nous Portal; reasoning is the OpenAI-native field name. Both are emitted for maximum client compatibility.

Wire-format compatibility

Default behaviour is unchanged — no reasoning fields are emitted unless the client opts in via the header, so strict OpenAI parsers that don't know about the extension won't break.

POST /v1/chat/completions
X-Hermes-Expose-Reasoning: true

Changes

• gateway/platforms/api_server.py:

  • Added _parse_expose_reasoning_header helper (reuses _coerce_request_bool so true|false|1|0|yes|no|on|off all work)
  • Non-streaming: when opted in, add reasoning_content + reasoning to the assistant message
  • Streaming: wired a tool_progress_callback that captures reasoning.available events from the agent and emits them as delta.reasoning_content chunks on the SSE stream
  • Added MAX_REASONING_CHUNK_BYTES (256 KB) defensive cap on per-chunk reasoning payload to prevent memory/bandwidth abuse

• tests/gateway/test_api_server.py: 6 new tests under TestChatCompletionsEndpoint:

  • test_non_streaming_omits_reasoning_by_default
  • test_non_streaming_exposes_reasoning_when_header_set
  • test_non_streaming_expose_reasoning_header_false_omits
  • test_streaming_omits_reasoning_chunks_by_default
  • test_streaming_exposes_reasoning_chunks_when_header_set
  • test_streaming_reasoning_chunk_capped_at_max_size

Test plan

• Full tests/gateway/test_api_server.py suite: 162/162 pass
• tests/run_agent/test_run_agent.py, test_partial_stream_finish_reason.py, test_streaming.py: 401/401 pass
• Independent reviewer subagent: no security or logic defects (4 non-blocking suggestions, all addressed)
• Static security scan: clean

Risk & scope

• Low risk — additive change, default behavior preserved
• No breaking changes — existing clients see identical responses
• Scope is contained — touches only gateway/platforms/api_server.py + matching tests
• No config schema change — opt-in via request header instead of a new gateway.expose_reasoning config flag (per-client granularity without a global config change)

🤖 Generated with [Claude Code]