SIsignificant-gravitas
v2
chat
chat

Stream Chat Post

Changed on

Start a new turn and return an AI SDK UI message stream.

Returns an SSE stream (text/event-stream) with Vercel AI SDK chunks (text fragments, tool-call UI, tool results). The generation runs in a background task that survives client disconnects; reconnect via GET /sessions/{session_id}/stream to resume.

Follow-up messages typed while a turn is already running should use POST /sessions/{session_id}/messages/pending. If an older client still posts that follow-up here, we queue it defensively but still return a valid empty UI-message stream so AI SDK transports never receive a JSON body from the stream endpoint.

Args: session_id: The chat session identifier. request: Request body with message, is_user_message, and optional context. user_id: Authenticated user ID.

post/api/chat/sessions/{session_id}/stream

Request

  • The document declares no server URL.
  • Auth
    HTTP bearer

Path parameters

session_idstring required

Request body

messagestring required
is_user_messageboolean
contextobject nullable
voiceboolean

Voice mode is waiting on speech. Adds one line asking the reply to open with a spoken acknowledgement before any tool call; text turns never pay for it.

file_idsstring[] nullable
folder_idsstring[] nullable

Workspace folder IDs attached to this message. Named in the message for the model to open, never expanded into their files.

model'standard' | 'advanced' nullable

Model tier: 'standard' for the default model, 'advanced' for the highest-capability model. If None, the server applies per-user LD targeting then falls back to config.

message_idstring nullable

Optional per-click UUID generated by the frontend. The server scopes it to the authenticated user and session before using the result as ChatMessage.id. Frontend / network / RMQ-redelivery retransmits of the same logical send reuse the key, and the server claims it once: a retransmit neither starts a parallel turn nor, while a turn is running, queues the message again, and gets a subscribe-only response. Distinct user clicks (even with identical text) MUST send different ids — the frontend's per-click crypto.randomUUID() guarantees that.

expert_kickoffboolean

Marks the hidden, once-per-expert day-one kickoff. The server derives its owner-scoped message ID and persistence metadata.

autopilot_mode'ask_first' | 'auto' | 'unsupervised' nullable

The chat's approval mode from this turn on; None keeps the mode it already has.

Response

Successful Response

unknown required

Changes