Automations with human review

1. Choose the host and trigger

An automation runs on the selected computer. Its runtime must remain running for schedules, webhooks, and tasks. Closing a client does not stop a run. A schedule stores its own time zone.

GitHub event triggers use the host’s GitHub CLI login and poll for changes. They work without a public incoming webhook. Webhook credentials are configured on the host desktop.

2. Define the task and review gate

A task step specifies instructions, project, harness, model, permissions, and checkout. Add a review step before a decision that needs human judgment. Inspect the preceding task’s conversation and changes before approving or rejecting the run.

• Use a new worktree to isolate a Git task.
• Specify the expected result and required checks.
• Edit branching graphs on desktop or web; mobile supports an ordered path.

3. Test manually, then enable triggers

New automations have automatic triggers paused. Save the workflow, run it manually, and inspect tasks and review gates. Enable its schedule or webhook once the result and permissions are right.

Only one run of an automation can be active or awaiting review at a time. Tasks sharing a local checkout wait for one another. Separate worktrees provide independent checkouts.

4. Inspect a failure before retrying

Retry preserves completed steps and resumes the unfinished step’s task, conversation, and checkout. It uses the original run configuration. Start a new run to use edited instructions.

Cancel stops the current task and prevents later steps. Waiting review gates survive restarts. Schedules do not replay downtime, so a sleeping host does not catch up every missed tick.