Loop

Loop

Run iterative verification loops and hillclimbs on Antigravity without manual polling.

How It Works

The /loop skill drives discrete edits against a verification command (--verify "<command>"). It maintains iteration state in <appDataDir>/brain/<conversation-id>/scratch/loop/state.json. It uses the native Antigravity schedule tool to manage reactive iteration delays.

For open-ended autonomous tasks where no single bash exit code determines completion, use the /goal slash command or the Autonomous run playbook instead. For standing background jobs or periodic status checks, use the schedule tool directly.

Invocation

Format: /loop <goal> --verify "<command>" [--max-iterations N] [--interval S]

Default values:

• max-iterations: 10
• interval: 30 seconds

State Machine

The loop maintains durability through structured state transitions. The file state.json records git revision, active task ID, iteration counters, and status. During execution, state.json records git revision for each step.

1. Frame. Read target files. Generate baseline evidence by executing the verify command via run_command or bash. Record initial state in <appDataDir>/brain/<conversation-id>/scratch/loop/state.json. Record the current git_revision from git rev-parse HEAD.
2. Step. Apply one discrete edit or hypothesis. Update git_revision if a commit or checkpoint is created.
3. Verify. Run the verification command. Capture stdout, stderr, and exit code.
4. Decide.
   • When verify passes, log success, mark status CONVERGED, and finish.
   • When verify fails and iterations remain, record the delta, revert regressions if hillclimbing, and set a reactive timer using schedule(DurationSeconds=interval, Prompt="Loop iteration wakeup for <goal>", TimerCondition="never"). Update state.json with the returned task_id and next iteration count.
   • Stop calling tools immediately after scheduling to end the turn and let the timer fire.
   • When budget is reached, mark status EXHAUSTED and hand back to the user with a summary table.
5. Wakeup. On schedule trigger, the agent reactively wakes up, reads state.json, and executes the next step.

Crash and Restart Recovery

The loop recovers gracefully across crashes, context compactions, or server restarts:

1. State restoration. On session startup or resume, read <appDataDir>/brain/<conversation-id>/scratch/loop/state.json.
2. Revision verification. Compare git rev-parse HEAD with the recorded git_revision. If the working tree has drifted, reset to the recorded revision before proceeding.
3. Deduplication. Inspect the recorded task_id in state.json. Use manage_task with Action: "list" or Action: "status" to inspect pending scheduled tasks.
4. Orphan cleanup. Kill stale or orphaned timers using manage_task with Action: "kill" to prevent duplicate wakeups.
5. Execution resumption. Resume iteration from the restored state machine position without repeating finished work.

Arguments