[Epsilon006]

Summary

Fixes #8427. The OpenAI Responses plugin can deliver an assistant output message's phase before its first text delta when the provider supplies it at item start. The default remains completion-only on both HTTP and WebSocket.

llm = openai.responses.LLM(emit_phase_on_item_start=True)

With early delivery enabled:

• Phase notifications and text chunks expose the item's ID and any known phase in delta.extra["openai_item"]. Use the response and item IDs to associate late updates with previously streamed text.
• Repeated phase values for the same item produce one notification. A phase first supplied or changed at completion is emitted then. Text with an unknown phase streams immediately.
• Phase notifications also set delta.extra["openai"]["phase"] for chat history. Keeping per-item metadata separate preserves the last reported phase when a later item has no phase.

Related to #8429, which also proposes opt-in early phase delivery.

Verification

• 168 focused OpenAI and LiveKit Inference Responses tests passed. Four new cases reproduced the phase loss on the previous revision and now pass, covering LLMStream.collect() and voice aggregation over HTTP and WebSocket, including follow-up Responses formatting.
• Repository-wide Ruff lint and formatting, targeted mypy, and git diff --check passed.
• The Windows test harness stubbed the unused native audio-inference import and fails if a test calls it.

[CLAassistant]

All committers have signed the CLA.