fix(api-server): forward Codex commentary in streaming SSE endpoint

[webtecnica]

Summary

Wire interim_assistant_callback through APIServerAdapter._create_agent() and _run_agent() so Codex commentary (phase=commentary) is forwarded to API Server streaming clients.

Add assistant.commentary SSE event to the Hermes session stream endpoint (POST /api/sessions/{session_id}/chat/stream), gated behind display.show_commentary (defaults to true).

Fixes #67580

Problem

When Hermes uses the openai-codex backend through API Server, Codex can generate user-facing progress messages as assistant output items with phase="commentary". Hermes preserves those items in codex_message_items, but API Server streaming clients do not receive them.

Long-running, tool-heavy turns remain silent until the final answer. The user cannot tell whether the agent is making progress, stalled, or has misunderstood the task.

Root Cause

APIServerAdapter._create_agent() forwards stream_delta_callback, tool_progress_callback, tool_start_callback, and tool_complete_callback to AIAgent, but it does not accept or forward interim_assistant_callback. _run_agent() has the same omission. The Hermes session SSE handler wires assistant deltas and tool progress without a commentary path.

The core commentary behavior was added in #66115 and defaults to visible, but gateway/platforms/api_server.py was not part of that callback wiring.

Changes

1. _create_agent(): Added interim_assistant_callback=None parameter and passes it to AIAgent()
2. _run_agent(): Added interim_assistant_callback=None parameter and forwards it to _create_agent()
3. _handle_session_chat_stream(): Added _commentary callback that emits assistant.commentary SSE events with message_id, text, and already_streamed fields. Gated behind display.show_commentary config.

Safety

• Commentary only — phase="analysis" / raw chain-of-thought / reasoning summaries are NOT exposed through this path
• Final content remains unchanged — commentary is emitted as a separate event type and is never concatenated into assistant.completed content
• display.show_commentary: false preserves the quiet behavior (no commentary events emitted)
• Non-streaming endpoints continue to return only the final answer (no change)
• /v1/chat/completions delta content is unaffected (commentary does not go into choices[].delta.content)

[alt-glitch]

This was generated by AI during triage.

Duplicate of #67593: both wire interim_assistant_callback into the API Server session SSE path to emit Codex commentary.

[teknium1]

Thanks for tracing the existing callback boundary and keeping commentary separate from final response content.

Problems

• The new callback ignores already_streamed (gateway/platforms/api_server.py:2572 in this diff). Core can call this callback with already_streamed=True when the same visible text already went through normal deltas (run_agent.py:4893-4900), so this would duplicate commentary in the session stream.
• /v1/responses streaming still calls _run_agent without an interim callback on current main (gateway/platforms/api_server.py:3841-3853). The linked issue describes API Server streaming endpoints, and this supported stream would remain silent.
• This adds a public SSE event but changes neither regression tests nor the documented session-stream event contract (website/docs/user-guide/features/api-server.md:346).

Suggested changes

• Drop already_streamed=True events and add regression coverage for duplicate suppression, final-content separation, and the config gate.
• Extend the Responses stream with a distinct phase="commentary" message item, then document both stream contracts.

Automated hermes-sweeper review.

[teknium1]

Honor already_streamed here. Core invokes this callback with already_streamed=True when the same visible text was already sent through the delta stream (run_agent.py:4893-4900); emitting it again as assistant.commentary duplicates client-visible narration.