loop-engineerlisted

# Loop Engineer Turn a coding task into a **closed loop** that runs until a goal is verifiably met, instead of a single best-effort pass. Based on the four-loop stack from LangChain's *The Art of Loop Engineering* (agent → verification → event-driven → hill-climbing). The leverage is not the model — it's the loop you wrap around it. The single rule: **never declare done without the checker passing the done-condition.** ## Session topology For the four-stage fleet, run the maker and checker in separate top-level stage-parent sessions when possible. The current default parents are Opus 5/Claude Code for planning, Kimi K3/Pi for coding, GPT-5.6 Luna/Codex at max effort for debugging, and Grok 4.5/Pi for reviewing. A native child may not spawn nested children; use the GitHub Project item, GitHub issue, git/PR artifacts, and handoffs to bridge sessions. ## Workflow Run these in order. Loops 1–2 are the core; 3–4 are escalations. ### 1. Lock the goal (do this before any code) Write a short **loop contract** at the top of your working notes: - **Goal** — one sentence, the user-visible outcome. - **Done-condition** — a *machine-checkable* gate. Prefer a command: `npm test`, `pytest -k feature`, `tsc --noEmit`, a curl returning 200, a screenshot diff. If it can't be checked by running something, rewrite it until it can. - **Rubric** — 3–6 bullets the checker reviews against (correctness, edge cases, no regressions, matches existing patterns, no debug cruft). See [REFERENCE.md