Streaming reasoning event translation (SSE)

[HXYerror]

Part of #1. Depends on #3.

Goal

The Responses API streams a richer event vocabulary than chat/completions. We need a faithful translator on both consumer surfaces:

• /v1/responses clients should see canonical OpenAI Responses SSE events
• /v1/messages clients should see Anthropic SSE with thinking_delta interleaved correctly

Event vocabulary to handle

From upstream (forward as-is on /v1/responses, translate on /v1/messages):

Event	Forwarded?	Maps to (Anthropic)
response.created	✅	initial message_start
response.in_progress	✅	drop / ping
response.output_item.added (type=reasoning)	✅	content_block_start { type:'thinking' }
response.reasoning.delta	✅	thinking_delta
response.reasoning_summary_text.delta	✅	additional thinking_delta (configurable channel)
response.reasoning.done	✅	signature_delta (carry encrypted_content) + content_block_stop
response.output_item.added (type=message)	✅	content_block_start { type:'text' }
response.output_text.delta	✅	text_delta
response.output_text.done	✅	content_block_stop
response.output_item.added (type=function_call)	✅	content_block_start { type:'tool_use' }
response.function_call_arguments.delta	✅	input_json_delta
response.output_item.done	✅	content_block_stop
response.completed	✅	message_delta (usage) + message_stop
response.failed	✅	error envelope

Tasks

• Implement a state machine in src/routes/responses/stream.ts for the OpenAI-shape passthrough (mostly forward + minimal normalization)
• Implement Responses → Anthropic translator in src/routes/messages/responses-stream-translation.ts (created by Reasoning types & reasoning_effort passthrough in chat-completions #7) that carefully sequences content_block_start/delta/stop per item index
• Handle out-of-order item indexes (Responses uses output_index and content_index; Anthropic uses a single index per content block — must allocate Anthropic indexes in arrival order)
• Buffer partial JSON for tool args until function_call_arguments.delta accumulates, or stream input_json_delta chunks faithfully
• Always emit a final message_stop even on stream errors, so Claude Code clients don't hang

Acceptance criteria

• Manual test: stream a gpt-5.3-codex agent loop with tool use through /v1/messages, confirm Claude Code renders thinking, tool calls, and text in correct order
• Event ordering tests added to tests/ covering reasoning-then-text, text-then-tool, tool-then-text-then-reasoning sequences

File pointers

• src/routes/messages/stream-translation.ts — reference state machine
• OpenAI Responses streaming docs: https://platform.openai.com/docs/api-reference/responses-streaming
• Anthropic Messages streaming docs: https://docs.anthropic.com/en/api/messages-streaming

[HXYerror]

Duplicate of the corresponding #10 — closing to avoid confusion.