looplisted

# loop ## Cycle 1. Inspect the smallest missing evidence. **Done when:** smallest gap identified, next action clear 2. Act on the smallest coherent change. **Done when:** artifact created and narrow surface-area check passed 3. Verify with narrow checks first. **Done when:** narrow checks pass (lint/unit test/spot-check only) 4. Checkpoint the result. 5. Repeat until done or blocked. ## Stop conditions Stop and re-route if: - the evidence no longer supports the current path - the diff grows beyond the intended scope - the same failure repeats without progress - the work becomes multi-step enough to require `plan` - context budget exceeds 75% → emit `handoff`, then stop the loop - cumulative tool-result bytes >50KB without progress → summarize + drop noise before next iteration ## Stuck protocol If the same task has been attempted >2 times without measurable progress (same failure repeating, diff not advancing), stop the loop and emit: "Stuck: [task], attempt N, last blocker: [X]." Switch to a different approach or tool. After 2 approach switches fail without progress, escalate to the user — do not keep cycling silently. ## Output (per checkpoint) One line per completed step: verdict + artifact produced. Report only what changed or failed. Do not narrate steps that passed cleanly. ## Parallel mode If a `Monitor` task is running (CI settle, build, long test), do NOT busy-wait. Pick up an independent priority — refactor a different file, draft a memory note, answer