Create conversation with streaming response

Start a new conversation and stream the AI response over Server-Sent Events (SSE). Behaves like POST /conversations but emits tokens, tool activity, and status updates incrementally instead of returning a single JSON response at the end.

Lifecycle

1. The server validates query, persists an in-progress conversation, then opens the SSE stream with HTTP 200.
2. A connected event is emitted immediately with the new conversationId so the client can link the stream (sidebar, parallel tabs, deep links) without an extra request.
3. AI-backend events stream through (token chunks, tool calls, status, etc.).
4. On success a single complete event is emitted carrying the full persisted conversation.
5. On failure an error event is emitted and the conversation is marked FAILED before the stream closes.

Event vocabulary

Three events have stable, server-defined data shapes:

• connected — { "message": string, "conversationId": string, "title": string }
• complete — { "conversation": Conversation, "meta": { "requestId": string, "timestamp": string, "duration": number } }
• error — { "error": string, "details"?: string }

The forwarded events are status, answer_chunk, tool_calls, restreaming, metadata, and tool_execution_complete. Their payloads come from the Python query service and may evolve. Note that raw tool_call / tool_success / tool_error / tool_result events emitted by the LLM tool runtime are rewrapped as status by the upstream wrapper before they reach this route, so clients on /conversations/stream never see those names directly. Clients should ignore unknown event names rather than treating them as errors.

Agent mode

When chatMode selects an agent mode (for example agent:auto), the optional tools list restricts which tools the agent may invoke for this turn. Outside agent modes the tools field is ignored.