Clawsweeper Until Clean

Clawsweeper until clean

Use this skill when the source of truth is the Clawsweeper bot acting on the GitHub PR, not a local review command. Trigger Clawsweeper, wait for the fresh review, fix actionable findings, push, and repeat.

First establish three consecutive ClawSweeper re-reviews with zero actionable findings on one head and a platinum-or-better rating label. Then use the completed reviews to make up to three honest attempts at diamond. Platinum is successful; diamond is a bounded stretch goal, not a second gate.

Do not use this skill for a one-shot trigger. A single /clawsweeper re-review comment is enough for that. Use this skill when you want the loop run to convergence.

Non-negotiables

trigger: comment exactly `/clawsweeper re-review` on an eligible PR via `gh pr comment`; no other comment text is authorized
review_source: must be a clawsweeper review/comment posted AFTER your trigger comment
fix_tool: apply targeted fixes directly, or use the repo-specific fix workflow when one exists
stop_condition: 3 consecutive clawsweeper re-reviews with zero actionable findings
minimum_readiness: one of `rating: 🐚 platinum hermit`, `rating: 🦞 diamond lobster`, or `rating: 🦀 challenger crab` must be present after the clean streak
diamond_timing: inspect rank-up moves only after the clean platinum-or-better baseline
diamond_attempt_limit: 3 across the entire workflow, including any resumed run
platinum_result: success only with a concrete explanation of why diamond was not reached
counter_reset: any clawsweeper re-review that lists actionable findings resets the consecutive-clean counter to 0
no_early_exit: do not stop on 1 or 2 clean re-reviews
no_self_review: do not skip a re-trigger because you "already know it's clean"
no_self_label: never add, edit, or preserve the readiness label on Clawsweeper's behalf

Pre-flight

1. Confirm the target PR number and repo. Pin those values; reuse them every iteration.
2. Confirm gh is authenticated against the right account and gh pr view <pr> returns the expected PR. The command-comment exception applies only when the authenticated user authored the PR or the task contains explicit evidence that the user substantially contributed to it. If neither can be verified, ask before commenting.
3. Capture a baseline timestamp and head SHA using polling-and-freshness.md.
4. Recover diamond_attempts from the current workflow state and PR history. Count each prior clean platinum baseline followed by an author-controlled rank-up change. If the history is ambiguous, use the highest plausible count instead of risking a fourth attempt.
5. Decide a safety cap. Defaults: 6 re-review cycles per clean-convergence phase, at most 4 phases, and 20 min wall-clock per wait for any single Clawsweeper response. The first phase establishes the baseline; each later phase exists only after one diamond attempt changes review input.
6. Confirm the working tree is clean enough that fixes will land on the right branch. Note any untracked/local changes that might confuse targeted fixes.

Loop

Maintain across the whole session:

consecutive_clean = 0
phase_iterations = 0
total_iterations = 0
diamond_attempts = <recovered integer from 0 through 3>
last_trigger_at  = <ISO timestamp of your most recent /clawsweeper re-review comment>
last_head_sha    = <PR head SHA at trigger time>

Repeat:

1. Increment phase_iterations and total_iterations.
2. Trigger Clawsweeper with gh pr comment <pr> --body "/clawsweeper re-review".
3. Record last_trigger_at and last_head_sha.
4. Monitor for a fresh, finished response using polling-and-freshness.md.
5. Classify the response with verdicts.md.
6. If clean, increment consecutive_clean; stop only at 3/3.
7. If it has findings, reset consecutive_clean to 0, fix only those findings with fixing.md, push, and re-trigger.
8. If ambiguous, do not count it as clean. Re-trigger once; if ambiguity remains, stop and report.
9. If a safety cap is hit, stop and report unresolved state honestly.
10. After 3/3, wait for ClawSweeper's label update and verify a platinum-or-better label with polling-and-freshness.md. If it is absent at the wall-clock cap, stop incomplete. Do not apply it yourself.
11. Continue to After clean: try for diamond. Do not report final success before that step returns a terminal outcome.

Between consecutive clean re-reviews, do not edit code or push commits. The streak is only meaningful when Clawsweeper agrees with itself three times on the same tree.

After clean: try for diamond

Read rank-up.md and rating-rubric.md, then apply the post-clean decision. If it selects an author-controlled improvement, increment diamond_attempts before the mutation, reset consecutive_clean and phase_iterations, and return to the loop for another clean-convergence phase. After that phase, use the newest result to decide whether to stop or spend another attempt. Preserve diamond_attempts across any caller-driven resume so the workflow can never start a fourth rank-up cycle.

This step finishes as already-diamond-or-better, diamond-achieved, or platinum-with-explanation. A platinum result must name the concrete evidence, environment, scope, residual-risk, or owner-decision ceiling. Do not rerun an unchanged review merely in hope of a different rating.

Completion criteria

• Final stop reason is already-diamond-or-better, diamond-achieved, platinum-with-explanation, safety-cap-hit, or wall-clock-cap-hit.
• Success requires three consecutive fresh Clawsweeper verdicts with zero actionable findings on the same PR head SHA and one platinum-or-better rating label after that streak.
• Every actionable finding resets the clean counter, even on the third re-review.
• Every counted response is fresh, finished, and from Clawsweeper.
• The readiness label was observed rather than added by the agent, and no push occurred after the final counted review.
• At most three post-clean diamond attempts occurred across the complete workflow. Each attempted a distinct concrete improvement, and a final platinum result includes the specific reason it did not reach diamond.
• The final report follows reporting.md.

Context pointers

• Use polling-and-freshness.md for the baseline command, polling command, response identity, finished-response rules, and head-SHA discipline.
• Use verdicts.md for clean, findings, and ambiguous classification.
• Use fixing.md for targeted fix and push rules.
• Use rank-up.md only after the first clean platinum-or-better baseline, for the bounded diamond attempts and their stopping explanation.
• Use rating-rubric.md with rank-up.md to read the awarded label, tiers, and rank-up moves.
• Use reporting.md for iteration narration, final report fields, hard rules, and common mistakes.