Skip to content

Reviewing Work

When an agent finishes a task, it moves to "review" status. A human reviewer evaluates the changes and decides to approve or reject.

Opening a review

Choose a task in "review" status. You see the following information.

  • Diff view showing all file changes the agent made
  • Terminal output from the agent session
  • Acceptance criteria with checkboxes for each criterion
  • Task summary written by the agent

Approving a task

  1. Review the diff and terminal output
  2. Check each acceptance criterion as met or not met
  3. Choose Approve.

The task moves to "done" and the cascade engine checks for newly unblocked tasks.

Rejecting a task

  1. Review the changes and mark criteria results
  2. Add notes explaining what needs to change
  3. Optionally add inline comments on specific lines
  4. Choose Reject.

What happens next:

  1. A review record is saved with your feedback
  2. The review iteration counter increments
  3. The task resets to "ready"
  4. A new agent session starts with your structured feedback in the prompt
  5. The agent reads your notes and criteria results and tries again

Review iterations

Each rejection creates a new iteration. The agent sees the full feedback history, including all previous iterations. This helps it converge on the right solution.

The review iteration number is tracked on the task, so you can see how many attempts it took.

Review policies

You can set a review policy per sprint:

Policy What happens
manual Work waits for human review (default)
auto_commit Agent output is committed automatically
auto_pr Agent output is pushed as a pull request automatically

For early development, manual is recommended. As you build trust in the agent's output, you can switch to auto_commit or auto_pr.