Run and review a task

Follow execution, answer questions and accept the result against the original criteria.

When a decision is needed

Open a stage to read its result. A gate waits for your approval or rejection. A stopped stage requires you to inspect the cause and decide how work should continue.

Approve or return

At a gate, choose acceptance or rejection. Rejection opens the feedback field. Its text goes to the agent at the configured return destination.

For example: changing the warehouse leaves products from the previous warehouse visible. Ask the agent to update the result on warehouse changes and check products stocked only elsewhere.

Approval does not send a comment to the runner. Record additional requirements in the task description.

Acceptance evidence

A successful stage outcome alone does not prove completion of the whole task. Before acceptance, compare requirements, scenarios, results and evidence. The report format and rules for unchecked criteria are collected in the recommendations for profiles and routes.

Inspect the preview

Open the preview and verify acceptance criteria. The new feature example helps select scenarios. If autonomous processing is required, check it with the normal processes using the environment readiness guidance.

Acceptance and delivery are separate actions. Merging or publishing changes depends on the route and project rules.

Return to a stage

The Go here action on an ordinary node returns execution to that point. Its explanation is included in the entering agent's task. Explain what new information or change makes the repeat useful.

Investigate a stop

• Environment unavailable: restore the preview or required tools.
• Check failed: inspect the finding and return work to the stage that can fix it.
• Retries exhausted: identify why repetition is not producing progress.
• Human answer needed: respond through the relevant question form.

Reaching the attempt limit means stopping with a description of remaining work. Record open blockers, unchecked criteria and why progress stopped before choosing a supported continuation or return action.

Stall arbiter

The arbiter applies to ordinary routes and does not analyze loop attempts, automatically or manually. It is a separate agent that reads the attempt log, investigates the stop and proposes a next step.

Automatic initiation of analysis is disabled by default. Enable it under Settings → Runs → Stall arbiter, or start analysis manually at a stop. The human chooses to apply a proposal or their own instruction in the stop dialog; analysis alone does not resume the pipeline. The arbiter does not change task scope, remove acceptance criteria or run commands to repair a preview. Review its decisions before acceptance.

Loop attempts

The lead assigns work, checks rework and starts subsequent stages. Worker results go to the lead. You accept the whole task, and rejection goes back to the lead without resetting the round count.

At loop-ceiling, assess remaining work. To continue, increase PROFILE_LOOP_MAX_ROUNDS and choose Continue. Without an increased limit no model starts. A progress failure needs investigation; increasing a limit alone does not resolve it.

Read the loop guide for configuration and stop types.

Improve the next revision

Update instructions, checks or gate placement in a new draft, then validate and publish.