agentic-loop

name	agentic-loop
description	Run a bounded change → verify → fix cycle until an objective check passes (tests green, lint clean, build/type-check ok) or a hard iteration cap is hit. Use for any task that needs repeated attempts against a verifiable signal instead of a single one-shot edit.
model	inherit
effort	medium
allowed-tools	["Read","Edit","Bash","Grep","Glob"]
triggers	["keep trying until","make the tests pass","fix until green","iterate on this","loop until"]
tags	["loop","iterate","verify","tests","autonomous"]

Agentic Loop

Drive work with a verifiable signal, not by eyeballing. Change, run the check, read the failure, fix, repeat — bounded by a hard cap so it can never spin forever.

When to use

Use this whenever there's an objective pass/fail signal to close on:

• Make failing tests pass.
• Get lint/format clean.
• Make the build or type-check succeed.
• Reach a coverage or benchmark threshold.

If there's no machine-checkable signal, don't loop — do the edit once and ask the user to confirm.

The loop

Read the cap and verify command from ${CLAUDE_PLUGIN_ROOT}/config.json (loop.max_iterations, default 5; loop.verify_command). If no command is configured, infer the project's own (e.g. npm test, pytest, cargo test, make check) and confirm it with the user before the first run.

iteration = 0
loop:
  1. Run the verify command. Capture full output.
  2. PASS  → stop, report success (iterations used).
  3. FAIL  → read the error. Form ONE concrete hypothesis for the top failure.
  4. Make the smallest change that addresses that hypothesis.
  5. iteration += 1
  6. iteration >= max_iterations  → STOP. Report status + remaining failures.
     else → loop.

Rules that keep it honest

• The check is the judge. A change isn't done because it looks right — it's done because the command passes. Re-run every iteration; never trust a prior run.
• One hypothesis per iteration. Fix the top failure, not five things at once — otherwise you can't tell what worked.
• Smallest change that could work. Big rewrites mid-loop hide regressions.
• Never weaken the check to pass it. Don't delete/skip tests, lower a threshold, or catch and swallow to force green. If the check itself is wrong, stop and say so.
• No progress after 2 iterations on the same failure → stop and escalate with what you tried and the current error. Don't burn the cap thrashing.
• Hard cap is absolute. At the cap, stop and report — success is not required to end cleanly.

Routing (optional)

For a heavier loop, pair with model-router: plan the fix at the plan role, apply it at execute/cheap. Keep the verify step deterministic (it's a command, not a model call).

Report on exit

Always end with: pass or fail, iterations used out of the cap, the verify command, and — if failed — the remaining errors and your best next step. If the outcome is a standing fact ("test suite is pnpm test"), store it with memory-keeper.