# Loop Engineer

> Runs recurring agent loops: fixed or dynamic intervals, CI/PR babysitting, and re-invoking prompts/skills until stop criteria. Use for watch-until-done work. Emits LOOP_PLAN / LOOP_STATUS. Never starts unbounded loops without stop conditions or spam-notifies without signal.

- Skill: `willianbs/loop-engineer` (Agent Skill, multi-file: 2 files)
- Install (CLI): `npx skillmds@latest add willianbs/loop-engineer`
- Raw SKILL.md: https://api.skillmd.com/api/skills/willianbs/loop-engineer/raw
- Safety review: pending
- Works with: Claude Code, Claude.ai, OpenAI Codex
- Category: AI & ML
- Author: willianbs (https://skillmd.com/u/willianbs)
- Updated: 2026-09-17
- Page: https://skillmd.com/skills/willianbs/loop-engineer

---


# Purpose

Arm a controlled loop that re-runs a prompt or skill path on an interval (or dynamically) until success, failure, timeout, or user stop.

# When to Use / When NOT to Use

**Use when:** babysit CI/PR checks; watch deploys; poll until green/red; recurring `/loop` style work; long-running verification.

**Do not use when:** one-shot tasks; tight interactive coding sessions where a loop adds noise; production mutate-on-timer without explicit approval.

# Preconditions

Clear work unit (prompt or skill + args) and stop criteria. If interval omitted, use dynamic mode with rationale.

# Inputs / Outputs

**Inputs:** prompt/skill to run, interval or dynamic, stop criteria, optional gh/CI identifiers.

**Outputs:** `LOOP_PLAN` then recurring `LOOP_STATUS` updates.

# Upstream / Downstream

**Upstream:** engineering-os, quality-gate, user.

**Downstream:** any skill named in the loop body; pr-generator; incident-commander if loop detects sev.

# Core Principles

1. Every loop needs stop criteria (success/fail/timeout/max ticks).
2. Prefer signal-based wakes (CI status, log sentinel) over blind sleep when possible.
3. One loop per purpose; do not duplicate existing loops.
4. First run immediate; then interval — avoid double-fire on start.
5. Status updates are short; full work only on tick.
6. Destructive actions require explicit user approval each time (or pre-approval scope).
7. Stop cleanly and report why.

# Process

1. Parse: interval (`30s`/`5m`/`2h`) or dynamic; body prompt/skills; stop criteria.
2. Emit LOOP_PLAN — purpose, interval, stop, max ticks, sentinel.
3. Check for existing matching loop; reuse or stop duplicates.
4. Arm background waiter / poller as environment allows; or instruct host loop tool.
5. Run body once immediately; record LOOP_STATUS.
6. On each tick: run body; evaluate stop; update LOOP_STATUS.
7. On stop: final status, summary of changes, artifacts produced.

## Presets
- `babysit-pr` — poll checks until success/fail
- `babysit-ci` — watch workflow run
- `until-qg-pass` — re-run quality-gate path
- `custom` — user prompt

# Evidence Requirements

Tick times, commands, CI URLs, stop reason. No fake “still running” without checking.

# Stop Conditions / Failure Modes

| Condition | Action |
|-----------|--------|
| No stop criteria | Refuse to arm |
| Max ticks / timeout | Stop; LOOP_STATUS failed/timeout |
| User says stop | Stop immediately |
| Body returns Block Critical | Stop and escalate |

# Severity + Confidence

Escalate to incident-commander if loop observes prod SEV signals.

# Output Contract

```
## LOOP_PLAN
Purpose: ...
Interval: ... | dynamic
Body: skill/prompt
Stop criteria: ...
Max ticks: ...
## LOOP_STATUS
Tick: N
Observed: ...
Decision: continue | stop
Stop reason: ...
```

# Handoffs

**engineering-os**, **quality-gate**, **pr-generator**, **incident-commander**.

# Never

- Never arm infinite loops without max ticks/timeout.
- Never force-push or prod-mutate on a timer without explicit scope.
- Never spam full reports every tick — delta only.