loopfixlisted

# Loopfix ## Overview Loopfix is an autonomous work-review-fix loop for the current approved goal. The main agent owns triage and momentum: reviewer subagents advise, but the main agent decides what to fix now, what to reject, and what to defer for final human audit. **Core rule:** meaningful in-scope changes reset the loop. Do not finish until a reviewer pass after the latest meaningful change finds no unresolved current-goal issue, verification is fresh, and the runtime-neutral completion criteria are satisfied. Completion is an evidence-backed judgment by the main agent, not a hook, state file, forced re-prompt, or arbitrary iteration count. Do not add runtime-specific loop mechanisms. Continue only when an in-scope issue remains. **Iteration budget.** If the loop reaches **5 iterations** on the same goal without converging — the same class of issue keeps recurring, or fixes are not reducing the reviewer's findings — stop, report the stall, and ask the user. A loop that disagrees with its reviewer forever is a signal to surface the conflict, not to grind. When you stop, summarize: how many iterations ran, what the recurring issue is, what you have tried, and why you cannot resolve it without a human decision. ## Completion Criteria Before the first fix, define a short checklist for this run with all five fields: | Criterion | What to state | |---|---| | Goal | The approved user goal in one sentence | | In-scope outcomes | Observable behavior, files, tests, or docs