Lead ReviewCopy Anchor Link

Lead Review withholds a worker turn after that turn calls complete_task, and it blocks further external messages while a review is already open. Assigning a lead does not hide ordinary replies. A turn that never calls complete_task returns the assistant text.

Public list and get payloads do not include leadAgentId or leadReviewConfig. There is no public REST, SDK, or hosted MCP read for a work item.

ConfigurationCopy Anchor Link

A worker stores leadAgentId and leadReviewConfig:

Saving this in Action Agent Studio requires agent:update. The worker cannot be its own lead. An inactive or deleted lead is rejected at save time. A worker whose organization preset already has a lead cannot set its own.

Resolution is preset-first. An organization preset with a lead is the reviewer, and the worker columns are ignored. A platform preset does not carry an organization lead, so the worker columns apply. See Presets.

When a turn is gatedCopy Anchor Link

One active work item exists per worker session.

• The first external message is stored, truncated to 50,000 characters, as the task context.
• Further messages while the item is inProgress are appended.
• An inProgress item idle for 7 days moves to abandoned.
• Any other non-terminal status rejects the new message with LEAD_REVIEW_PENDING.

Terminal statuses are completed, rejected, failed, cancelled, and abandoned. approved is not terminal. It becomes completed after a successful internal delivery.

complete_task is available only while the trusted turn context matches an inProgress work item. The model supplies summary only: non-empty, at most 2,000 characters. A pending completion signal withholds the assistant text and stores that message as leadGated. Visible transcript reads skip leadGated rows. If the signal read fails, the text is withheld. If there is no signal, the reply is returned and then screened by the organization output policy when that stage has rules.

After the turn, a complete signal moves the item to reviewQueued and queues a Lead review. The submission snapshot is at most 100,000 characters. Incomplete provenance moves the item to awaitingHuman.

RESTCopy Anchor Link

Permission to send is agent:trigger. Review permission is not required to send.

Non-streaming POST /v1/chat/completions:

Streaming (stream: true):

• A busy intake fails before the stream and uses the same 409.
• A gated turn is OpenAI SSE: an opening assistant role chunk with empty content, no text deltas, then finish_reason: "stop" and data: [DONE]. The stream carries no workItemId.
• A mid-stream runtime error is one data: {"error": ...} frame and no [DONE].

The OpenAI chat route does not emit a review-status event. That event belongs to a separate public SSE format.

Hosted MCPCopy Anchor Link

send_action_agent_message uses the same non-streaming wait path.

A gated turn is a successful tool result:

{
  "success": true,
  "result": {
    "leadReviewPending": true,
    "text": "",
    "revision": null,
    "workItemId": null
  }
}

A busy intake is a tool error. The public text is The operation could not be completed. Try again. It does not include LEAD_REVIEW_PENDING.

Channel triggers use the same wait path. A gated or busy turn does not post an immediate channel reply. A turn with no completion signal is posted as usual.

Review, delivery, and ActionFlowsCopy Anchor Link

agent:review is granted to Owner, Admin, and Member. It gates the Lead Reviews queue, the review detail, approve, reject, return, cancel, delivery recovery, and the dashboard operator tools. Those operator tools are not on the public hosted MCP server.

The actionable queue is status awaitingHuman, or status approved with a gating delivery in deliveryUnknown. Agent Activity links those rows to the review detail. It does not render the approved text. reviewQueued and reviewing are outside that filter.

The Lead model sees a server-built envelope and three tools: submit_lead_review, inspect_lead_messages, and inspect_lead_artifacts. It does not receive HTTP, MCP, Knowledge, database tools, or a nested lead. Verdicts are approved, changesRequested, and escalated. Up to 4 lead executions run per attempt. The review lease is 20 minutes.

Human decisions require status awaitingHuman:

Approved output:

• api and Action Agent Studio reveal the stored assistant message in the transcript. The original HTTP call does not receive a second body. On success the work item moves from approved to completed.
• Slack, Discord, Teams, and email post the approved snapshot, or the summary when the snapshot is empty, through the active trigger binding. A silent binding does not post.
• scheduled is marked sent without a channel message.
• A recorded ActionFlow continuation is not delivered by this outbox.

An Action Agent node whose worker has a resolved lead does not return immediately. It registers a continuation, dispatches the prompt, then pauses the Flow. The stored node output is { data: {}, success: true }. Run history status is waitingForHitl. No HITL request is created. The run event is Flow run is waiting for Lead review.

Resume starts a new Flow run from the pause snapshot and continues after the agent node. The approved summary stays on the continuation row. It is not copied into the node output. See Cloud Action Agents.

RelatedCopy Anchor Link

• Sessions & chat
• REST API
• send_action_agent_message
• Guardrails
• Presets
• HITL Approvals
• Organization permissions