/start

You are starting a loop-engineering run. A Stop hook re-invokes you every turn until the verify command passes, so once the loop is started you must keep working autonomously — do not stop to ask for confirmation unless you are genuinely blocked.

User request: $ARGUMENTS

Do the following:

1. Parse the goal (free text) and any options from the request:

   • --verify — a shell command that exits 0 only when the work is correct (tests/lint/build).
   • --max-iterations — hard cap (default 50).
   • --completion-promise — an exact phrase to emit when done (used only when no verify command fits).
   • --criteria — a ||-separated checklist of deterministically verifiable conditions.

2. Refine if needed. If no --verify command was supplied, briefly inspect the repo and invoke the loop-planner agent to produce a short plan, explicit verifiable success criteria, and a recommended verify command. Always prefer a real verify command over a completion-promise when the repo supports one.

3. Initialize the loop state by running this with the Bash tool (omit flags you have no value for):

   "${CLAUDE_PLUGIN_ROOT}/scripts/start-loop.sh" \
     --goal "<goal>" \
     --verify "<cmd>" \
     --max-iterations <N> \
     --completion-promise "<TEXT>" \
     --criteria "<a||b||c>"
   

4. Confirm the resolved goal, success criteria, and verify command back to the user in 2–3 lines.

5. Begin iteration 1. Make the first concrete change toward the goal, then run the verify command yourself. When you yield the turn, the Stop hook re-runs verification and either re-prompts you with the failure output or ends the loop on success.

Guardrails: respect the iteration cap, never repeat an identical failing approach, and if you hit a true blocker (ambiguous requirement, missing access, irreversible action) run /loop-engine:stop and explain rather than burning iterations.