ClawSweeper Files

Added

• Scaffolded ClawSweeper as a conservative OpenClaw maintainer bot that writes one markdown review record per open issue or pull request.
• Added proposal-only review flow plus an explicit apply mode for unchanged, high-confidence close proposals.
• Added targeted single-item review support.
• Added README dashboard links to generated item reports, fixed evidence, issue and PR close-rate metrics, cadence coverage, workflow status, and apply status.
• Added archived closed/ records so items/ can stay focused on open tracked items.
• Added a read-only audit command for checking live GitHub state against generated items/ and closed/ records. Thanks @stainlu.
• Added review runtime metadata to detail reports, including model and reasoning effort.
• Added MIT licensing.
• Added durable Codex automated review comments that are updated in place before any close action.
• Added a separate hourly apply/comment-sync workflow lane that can run alongside review work.
• Added a five-minute hot-intake review lane for new and recently active issues or pull requests, fanning out single-item review shards.
• Added targeted comment-sync mode so hot-intake reviews can publish durable Codex review comments immediately without closing items.
• Separated targeted comment-sync workflow concurrency from bulk apply so hot comment runs are not displaced by apply continuation backlog.
• Switched comment and close mutations to the openclaw-ci GitHub App installation token so GitHub attributes automated comments to the bot.
• Added Latest Run Activity dashboard counters for recent reviews, close decisions, comment syncs, apply skips, and close actions.
• Added a README Audit Health section plus a separate scheduled/manual workflow path to refresh it without making normal dashboard heartbeats scan GitHub. Thanks @stainlu.
• Added comma-separated targeted review dispatch so Audit Health findings can be reviewed together without waiting for normal batch selection. Thanks @stainlu.
• Added copyable targeted review inputs to Audit Health for reviewable drift findings. Thanks @stainlu.
• Added maintainer issue commands that let ClawSweeper create or update one guarded implementation pull request from an open issue.
• Added build as an issue implementation command alias.
• Added an automatic reproducible-bug implementation lane: strict bug reviews with high-confidence reproduction, no linked PR, and no feature/config scope can dispatch Codex to open an implementation PR.
• Added the clawsweeper:autogenerated label for PRs created by ClawSweeper's issue implementation lane.
• Added dedicated ClawSweeper event and merge notifications for OpenClaw agent hooks.
• Added automerge progress timelines that keep repair, review, wait, and merge events in one mutable status comment.
• Added automerge merge messages that summarize the reviewed PR change and any ClawSweeper repair/fixup work that was needed before merge.
• Added separate Codex debug artifacts for repair planning and repair execution so raw sessions and logs can be inspected without bloating normal published state.
• Added docs for scheduler capacity, automerge wait behavior, auto-update PRs, repair internals, and OpenClaw event hooks.

Changed

• Released ClawSweeper as 0.1.0.
• Let automerge fix execution run up to three Codex review-fix rounds by default, so new actionable findings found after validation feed back into the agent instead of stopping after one review-fix attempt.
• Updated repair workflow defaults to pass the four-attempt review loop through GitHub Actions instead of overriding the executor default with two attempts.
• Added bounded Git/GitHub network timeouts to repair execution so hung contributor-branch fetches fail with artifacts instead of exhausting the whole automerge job.
• Simplified substantive automerge repair so Codex owns the initial rebase, PR-comment review, CI inspection, and test/fix loop while the deterministic executor keeps GitHub mutations and final validation.
• Increased the repair executor budget inside the existing 45-minute Actions job so long Codex edit/test passes still have time for internal /review, post-flight, and artifact upload instead of wasting a retry on a 30-second end-of-budget review timeout; the workflow step timeout now leaves room for that larger internal budget to complete cleanly.
• Requeue repair runs immediately when a contributor branch advances during the safe push window, preserving the source-head race guard without waiting for a later sweep to retry against the latest head.
• Let scheduled comment-router sweeps re-enter labelled autofix/automerge PRs without a fresh comment, and dispatch repair when automerge activation sees a dirty or behind merge state.
• Filter routine GitHub activity before posting OpenClaw hook turns, retry transient hook failures with the same idempotency key, and document the retry controls for the activity lane.
• Switched review runs to GPT-5.5 with high reasoning.
• Limited protected-proposed audit failures to active item records so archived historical reports do not keep Audit Health in action-needed state.
• Increased sweep throughput over time with larger worker batches, 100 shards, chained continuation runs, and 50-review checkpoints.
• Renamed workflow run and job displays so review, apply, comment-sync, and audit runs are distinguishable in GitHub Actions.
• Made review cadence activity-aware: active items and items created in the last 7 days are checked hourly, older PRs and young issues are checked daily, and older inactive issues are checked weekly.
• Made policy changes force previously fresh reports back into review planning.
• Improved close evidence and comments with structured review notes, public docs links, ClawHub links, source links, fixed-version evidence, and nicer Markdown formatting.
• Added best-possible-solution review output so both close and keep-open comments explain the recommended path.
• Made review prompts acknowledge prior plugin links and prefer public docs.openclaw.ai links where appropriate.
• Clarified incoherent close-reason wording so rendered reports no longer collide with not_actionable_in_repo (#29). Thanks @xthunder0.
• Normalized repository profile lookup against configured target repos so mixed-case profile entries resolve correctly (#27). Thanks @xthunder0.
• Made apply runs issue-only by default, with no age floor, while still excluding maintainer-authored items.
• Made apply runs checkpoint their progress, publish dashboard heartbeats, and continue automatically while work remains.
• Made scheduled apply runs process both issues and pull requests by default, with manual apply_kind narrowing still available.
• Made apply checkpoint publish retries auto-resolve generated item/closed rename-delete conflicts from concurrent review publishes.
• Reduced the default apply close delay from 5 seconds to 2 seconds.
• Prioritized matching close proposals ahead of broad comment sync during apply runs so close batches do not stall on keep-open comment backfill.
• Increased scheduled apply wakeups to every 15 minutes and made idle apply runs exit after checking for close proposals instead of scanning keep-open records.
• Added a Recently Closed dashboard table with links to the target item and archived ClawSweeper report.
• Classified missing-open audit findings so strict mode reports only actionable missing-open drift while preserving total visibility. Thanks @stainlu.
• Added transient GitHub API/network retries with short backoff while preserving long secondary-rate-limit backoff and throttle heartbeats. Thanks @stainlu.
• Split the README dashboard into focused sections and collapsed the recent review table so the project page is easier to scan.
• Made PR review comments easier to scan with a compact summary, review details in collapsible sections, reproducibility surfaced for issues, and empty security sections omitted when there is nothing useful to say.
• Shortened review workflow startup and moved generated state to the state repo so review shards spend less time on setup.
• Kept repair workers on GPT-5.5 high reasoning with the fast service tier.
• Let trusted ClawSweeper verdicts with P0/P1/P2/P3 findings trigger repair even when the same review also contains a pass marker.
• Made repair label tagging non-blocking so label sync failures do not fail an otherwise useful repair worker.
• Capped final repair artifact debug copies to tail slices while keeping full Codex debug backups in dedicated debug artifacts.

Fixed

• Skipped missing or stale comment IDs in the comment router instead of failing the whole router on GitHub 404.
• Skipped replacement PR creation when a repair branch has no diff against the latest base branch, avoiding GitHub's "No commits between" failure.
• Prevented oversized executor JSONL/debug files from making final repair artifacts hundreds of megabytes.
• Emitted repair-worker heartbeats while Codex is running so GitHub Actions does not treat long silent model calls as stalled jobs before debug artifacts upload.
• Emitted execute-side Codex heartbeats during repair edit, review, and preflight subprocesses so automerge runs stay observable until debug artifacts upload.
• Kept final base-reconcile Codex workers from being squeezed down to the 30-second timeout floor by aligning the executor budget with the 40-minute repair step.
• Included ClawSweeper-captured codex exec --json outputs in Codex debug artifacts and kept execute-side logs under uploaded repair run artifacts.
• Kept substantive automerge repairs in the Codex edit loop after a clean rebase instead of treating base-sync head movement as the repair itself.
• Fed changed-surface validation failures back into Codex repair so automerge fixes can correct lint/typecheck fallout instead of stopping after the first failed pnpm check:changed.
• Passed the normalized changed-surface gate into Codex repair prompts so the agent runs, fixes, and reruns validation before returning to the deterministic executor.
• Backed up redacted Codex session/log artifacts from repair worker Actions runs so automerge stalls can be debugged from the raw model transcript.
• Prevented automerge repair workers from treating a clean rebase as a complete repair when the current ClawSweeper review still requires a substantive fix.
• Skipped event comment-router ledger publishes when a cancelled run exits before pnpm setup, avoiding noisy pnpm: command not found failures.
• Prevented duplicate automerge repair dispatches when the configured run-name prefix is trimmed but an active worker already exists for the same job path.
• Kept Codex review access read-only and verified the OpenClaw checkout before and after review.
• Authenticated Codex in CI without exposing GitHub write tokens to nested review sessions.
• Hardened strict review schema parsing and failure-evidence shape validation.
• Compacted related GitHub context for review prompts.
• Bounded shard runtime and continued after individual item review failures.
• Made review publishing reliable under concurrent workflow pushes.
• Reconciled tracked item folders when issues or PRs close or reopen.
• Hardened apply close safety with maintainer-author exclusions, protected-label checks, snapshot-change checks, idempotent reruns, and already-closed handling.
• Reduced apply snapshot API calls and added GitHub read/write retry backoff for long sweeps.
• Preserved close comment formatting and rendered applied comments from stored review evidence.
• Ensured README dashboard cadence metrics reflect the current review rules.
• Avoided duplicate close comments by adopting existing Codex review comments and adding a hidden marker for future updates.
• Corrected the GitHub Actions setup docs to describe app-token comment and close attribution.
• Documented the current bot/app operating model and the optional Actions write permission needed for app-token run cancellation.
• Cancelled stale pre-app apply run 24944438478 so it cannot keep posting maintainer-attributed comments.
• Guarded Codex process failure output so missing stdout/stderr does not hide the original review failure. Thanks @ZHOUKAILIAN.