Loop Engineering Skill
Grill (orchestrator interview)
- Goal: "What are you trying to accomplish?"
- Success criteria: "Describe how we know the loop succeeded."
- Success type: choose from
test,build,lint,command,fileExists,oracle,observer,manual. For CLI steps providesuccessCommand; for file detection providesuccessPath. - Execute agent: fixer / designer / explorer / librarian
- Verify agent: oracle / observer / test
- Max attempts (default 3)
- Optional context files: which files or directories should be read before execution?
Loop Monitor
- Listen to callbacks:
onLoopComplete(loopID, success)→ report final outcomeonEscalated(loopID, reason)→ escalate to humanonManualReview(loopID, reason)→ prompt human to approve/fail and callresolveManualReview(loopID, passed, reason)
- Show current state and attempt count on each callback
- For manual verification, present the failure reason before asking for pass/fail
- If human forces cancellation, call
cancel(loopID)through the orchestrator
Notes
- Manual verification is the minimal on-ramp (autoresearch pattern). It pauses the loop until
resolveManualReviewis called. Do not auto-resolve. - BackgroundJobBoard signals (totalErrors, timeoutCount) are already baked into the runtime.