Iterating To Completion

Purpose

This skill prevents two failure modes:

1. Premature exit — agent stops before success criteria are met
2. Infinite loop — agent spins endlessly with no progress

Mechanism 1: Completion Promises

Every agent task has an explicit completion signal. The agent outputs this string ONLY when success criteria are truly met:

Rule: Never output the completion signal unless the criteria are truly met. The orchestrator blocks on this signal.

Mechanism 2: Scratchpad Protocol

Maintain a scratchpad file at .jonggrang/.ephemeral/scratchpad-{task_id}.md:

# Task: {task_id} — {title}
## Completed Steps
- [x] Step 1 — what was done
- [x] Step 2 — outcome

## Current Focus
- [ ] Step 3 — what you're working on now

## Failures & Learnings
- Attempt 2: Failed because X. Will try Y instead.

## Next Step
Clear description of what you do next iteration.

Update this EVERY iteration. If you're stuck, the scratchpad shows why.

Mechanism 3: Loop Detection

The orchestrator reads your last 3 outputs. If similarity > 90%, you're stuck.

Signs you're stuck:

• Repeating the same command outputs
• The same error message appearing 3+ times
• Trying the same approach with minor variation

When stuck:

1. Write to scratchpad: "STUCK: [reason]"
2. Try a fundamentally different approach (not a variation)
3. If still stuck after 2 more tries, output BLOCKED: [reason] and stop

Iteration Limit

Max 10 iterations per agent spawn. After 10:

1. Write current state to scratchpad
2. Output ITERATION_LIMIT_REACHED: [what's left]
3. Orchestrator will re-spawn you with fresh context and the scratchpad

Integration with Hooks

The feedback-loop.sh / OpenCode session.idle hook enforces this:

• If dirty bit is set and you output a completion signal without going through reviewer/tester, the hook blocks your exit.
• You must demonstrate the full cycle: implement → review → test → signal.