| name | byo-loop-task |
| description | Scaffold a bring-your-own-repo loop task: find a failing verifier, write tasks/<id>.yaml, and run the loop in the user's codebase. Use when applying Loop Engineering outside atlas-oms or when the user says BYO or "my own repo." |
BYO Loop Task
Apply Loop Engineering to the user's repo, not a workshop lab.
Prerequisites
- Loop harness installed:
loop-discipline, loop-termination, loop-safety rules
- User repo open as Cursor workspace
- A command that returns pass/fail on real pain
Procedure
1. Find the verifier candidate
Ask or discover:
- A failing test, lint error, or type error on current branch
- Or tech debt they want fixed with objective done
Run the candidate command. It must fail for the right reason before looping.
2. Scaffold task YAML
From workshop repo (if available):
./scripts/scaffold-byo-task.sh --id <kebab-id> --verifier "<command>"
Or copy tasks/_template.yaml → tasks/<id>.yaml in the workshop repo, or tasks/<id>.yaml in the user repo under a tasks/ folder they commit.
Minimum fields: task_id, rung, seeded_state, verifier, termination.max_iterations, common_fails.
3. Set rung
| Situation | Rung |
|---|
| User judges every output | 0 — comparison only |
| Agent runs verifier, user approves each step | 1 |
| Rules + verifier loop, user reviews at green | 2 (default) |
| Background agent + PR review | 3 |
4. Run the loop
@loop-discipline @loop-termination @loop-safety
/loop-engineering
Include full task spec. Use /verifier-crafting first if the verifier is weak.
5. Log (optional)
./scripts/loop-harness.sh --task <id> --method B --log-only --pass true --iterations N --log loop-runs/<id>.json
Anti-patterns
- Looping without a failing verifier on seed
- Verifier that only checks happy path
- Skipping regression gates on multi-file fixes
Reference
Workshop guide: BYO.md · Template: tasks/_template.yaml
