Qredence Documentation

POST /api/sessions/{session_id}/turns returns a text/event-stream response. Each frame sets the SSE event field to the event type and carries a JSON data payload with a matching type. The stream ends with a raw [DONE] frame. Payload fields use camelCase on the wire. The source of truth is src/fleet_rlm/api/ui_stream.py, and openapi.yaml publishes it as FleetUIMessageChunk.

Stream order

The stream opens with turn_status (phase: "preparation") and repeats it every runtime.heartbeat_seconds until Fleet claims and prepares the Turn.
turn_start marks the start of execution.
Step, reasoning, code, output, tool, and text events stream while the RLM works.
The stream ends with exactly one terminal event: turn_finish, turn_error, or turn_cancelled.

If claim or preparation fails, the stream emits turn_error and then turn_finish.

Lifecycle events

Type	Fields	Description
`turn_status`	`phase`, `status`, `message`	Preparation and progress status.
`turn_start`	`runId`, `sessionId`, `delivery` (`live` or `replay`), `traceId`	Execution started. `traceId` appears when MLflow tracing is enabled and `mlflow.expose_trace_id` is `true`.
`step_start`	`step`	An RLM iteration started.
`step_finish`	`step`, `durationMs`	An RLM iteration finished.
`turn_finish`	`finishReason`, `status`, `checkpointVersion`, `durationMs`, `traceId`	The Turn settled.
`turn_error`	`message`, `code`, `durationMs`	The Turn failed. `code` defaults to `execution_failed`.
`turn_cancelled`	`reason`, `durationMs`	The Run was cancelled.

Content events

Type	Fields	Description
`reasoning`	`streamId`, `step`, `text`, `delta`, `final`	Model reasoning for a step.
`code`	`streamId`, `step`, `code`, `isDelta`, `final`	Python the model wrote for the Sandbox.
`output`	`streamId`, `step`, `output`, `isDelta`, `final`	Sandbox execution output.
`text`	`streamId`, `delta`, `text`, `final`, `role`	Assistant answer text.
`structured_result`	`schemaId`, `schemaVersion`, `value`	A typed result payload.

Activity events

Type	Fields	Description
`tool_call`	`toolCallId`, `toolName`, `input`	A host tool call started.
`tool_result`	`toolCallId`, `toolName`, `output`, `error`	A host tool call returned.
`skill`	`skillId`, `name`, `phase`, `version`, `trust`, `affordances`	A Skill was selected or loaded.
`child_progress`	`childId`, `taskLabel`, `state`, `elapsedMs`, `outcome`, `evidence`, `gaps`, `resultFileCount`, `codeExcerpt`, `outputExcerpt`, `cleanupState`, `parentRunId`	Progress of a recursive child RLM. `state` is `not_started`, `running`, `completed`, `failed`, `cancelled`, or `timed_out`.
`attachment`	`attachmentId`, `phase`, `filename`, `byteSize`	An Attachment was staged for the Turn.
`artifact`	`artifactId`, `artifactKind`, `title`, `mediaType`, `byteSize`, `checksumSha256`	An Artifact was produced.
`usage`	`iterations`, `durationMs`, `usage`	Token and iteration usage.
`warning`	`message`, `code`	A non-fatal warning.

Compatibility chunks

The stream contract also admits Vercel AI SDK UI message chunk shapes, such as start, text-delta, tool-input-available, finish, and data-* parts. The response sets the x-vercel-ai-ui-message-stream header. New clients should consume the event types above.

Example

event: turn_status
data: {"type":"turn_status","phase":"preparation"}

event: turn_start
data: {"type":"turn_start","runId":"…","sessionId":"…","delivery":"live"}

event: text
data: {"type":"text","streamId":"text","delta":"The first 20 Fibonacci numbers are…","final":false,"role":"assistant"}

event: turn_finish
data: {"type":"turn_finish","finishReason":"stop","status":"completed","checkpointVersion":3}

data: [DONE]

Fleet-RLM

Fleet Prime Agent

Fleet Pi

GEPA Omni

Qredence Skills

Fleet Reasoner

Qredence Plugins

Stream events

Stream order

Lifecycle events

Content events

Activity events

Compatibility chunks

Example

​Stream order

​Lifecycle events

​Content events

​Activity events

​Compatibility chunks

​Example

Stream order

Lifecycle events

Content events

Activity events

Compatibility chunks

Example