Skip to main content
POST
Create message and stream response

Authorizations

Authorization
string
header
required

The access token received from the authorization server in the OAuth 2.0 flow.

Path Parameters

project_id
string
required
chat_id
string
required

Body

application/json

A follow-up message in an existing conversation.

on_busy and idempotency_key are reply-only: they matter when a conversation may already be running a turn. Creating a conversation always dispatches its first turn (a new chat is never busy), and chat creation is not itself idempotency-keyed, so a fused create cannot honor a retry key — those fields would be a false promise there and live on the reply request only.

message
string
required

User message to send to the Summation agent.

Minimum string length: 1
display_content
string | null

Optional alternate rendering of the message for display; the agent still receives message.

file_ids
string[]

Project file ids to attach to the message as context.

mentioned_file_ids
string[]

Project file ids referenced inline in the message.

model
string | null

Chat model id to use for this turn, from GET /v1/chat-models. Per-request model selection is honored where enabled for the tenant; otherwise the default model applies. The reply metadata reports the model that ran.

Minimum string length: 1
effort
enum<string> | null

Reasoning effort for this turn (low, medium, high, xhigh, max). Honored under the same tenant setting as model selection; otherwise the default effort applies.

Available options:
low,
medium,
high,
xhigh,
max
on_busy
enum<string>
default:queue

What to do when the conversation is already working on a previous message: 'queue' durably enqueues this turn to run next (streamed once it dispatches); 'reject' returns a conversation_busy error instead.

Available options:
queue,
reject
idempotency_key
string | null

Optional client-supplied key that makes this reply retry-safe: replaying the same key returns the same accepted turn instead of creating a duplicate. Supply a stable key per logical message when retrying.

Required string length: 1 - 128

Response

Server-Sent Events stream of agent activity.

Each event is a UTF-8 line prefixed with event:, id:, or data: per the SSE spec. The data: payload is a JSON object with the shape {"type": <event_type>, "sequence": <int>, "data": <object>}.

Event types emitted by the agent:

  • status — high-level progress; data.message (queued/running text) or data.summary (short status line). When the conversation was busy and the message was queued (request field on_busy: "queue"), the first event is status with data.message: "queued" (or "held" if a Stop has already paused the queue), data.queuedMessageId, data.position, data.behindMessageId, data.held, and data.revision; the reply then streams once the queued turn dispatches. A held queue does not run until it is resumed.
  • queue.updated — the waiting message's queue state changed; data.position, data.held, data.revision, data.queuedMessageId. held: true means a Stop paused the queue: the message will not run until the queue is resumed (POST .../queue/resume). revision only moves forward; ignore any snapshot older than one you hold.
  • tool.started — agent invoked a tool; data.id, data.name
  • tool.input — tool input arguments; data.id, data.name, data.input
  • tool.completed — tool finished; data.id
  • message.delta — incremental assistant text; data.text (append in order)
  • done — terminal event; data.messageId, data.content (final assistant message), data.status (complete / error / cancelled)
  • error — terminal error; data.code, data.message. A queued message may end with code: "queue_wait_timeout" after 30 minutes of waiting — the message is still queued and can be observed again via the queue endpoints.

Resumability: when the connection drops before done, the same stream can be resumed via GET /v1/projects/{project_id}/conversations/{chat_id}/messages/{message_id}/events. Before a queued message has bound an assistant message id, reconnect by re-sending with the same idempotency_key, or poll GET /v1/projects/{project_id}/conversations/{chat_id}/queue/{queued_message_id}.