In a project scratch dialog, send a correction while the agent is working. MAIster saves the message immediately and either steers the active turn or queues it for a later turn. Reloading the page does not discard queued messages.

​
Send a correction in a scratch dialog

  1. Open the project’s scratch Run and confirm that it is Running.
  2. Send the instruction through its message composer.
  3. Read the delivery label in the transcript and inspect the agent’s response.
Steering depends on the capability advertised by this specific adapter session. An unsupported session or a refused steering attempt uses the queue. Steered does not promise immediate compliance or cancellation of a tool call already in progress; check the transcript and resulting work. The scratch composer refuses a message during initial setup, while waiting for a permission answer, or after the dialog ends. Answer the permission through the HITL controls. Studio and evaluator assistant dialogs accept their next message only after their current turn finishes.

​
Recover queued work

If the scratch dialog crashes, Recover preserves earlier queued messages and places the recovery message after them. If a web restart or a retryable admission failure leaves messages queued, the server automatically retries the oldest message once the dialog and host session are ready. A Queued receipt means the message is saved; do not resend it just to restart dispatch. Stopping, completing, or terminally failing a dialog leaves undelivered queued rows visible as Not sent. Recover admits a crashed dialog; it does not force a live or idle Run into a new attempt. If a session has disappeared but the Run is still live, allow automatic reconciliation to classify it first. For NeedsInputIdle, answer the pending request to resume the saved session; continuation may wait for capacity or for the execution host to recover. After a host or session failure, recovery can deliver an instruction again. Inspect the transcript before resending instructions with external side effects. The personal librarian can send an operator message to an eligible scratch or persistent agent Run using your project permissions. Check its operation receipt for delivery. This is separate from the coordinator token interface below.

​
Message a persistent child agent

A coordinator using the external API or MCP run_message can address its own persistent agent child with exactly one of addressableKey or childRunId. childRunId must be a UUID. Set mode: "steer" to attempt delivery to its running turn; mode: "queue" is the default. The response’s delivery field reports steered or queued. Reuse requestKey for retries of the same message. This requires the coordinator’s run-bound token with runs:delegate authority. Flow children do not expose an addressable agent session for this operation. For result collection, run_collect accepts exactly one of childRunId or all: true; collection itself does not send a message.