Overview

ClawSweeper is a maintenance bot that focuses on being conservative in its actions. It scans every issue, pull request, and commit in specified repositories and only proposes changes when there is strong evidence that a change is needed. The bot runs weekly and checks each item once a week.

The bot reviews items in a focused way. It can inspect a range of commits on a local branch before they are pushed, but that feature is a pre‑submission tool, not a hosted workflow lane. Each reviewed issue or PR results in a single markdown file saved in the report repository under records/<repo>/items/<n>.md. That file contains the decision, evidence, a proposed comment, runtime metadata, and a snapshot hash. The markdown keeps a durable record so the history of the decision can be audited later.

When the bot decides to act it does so in a single, guarded way. It never posts a new comment unless a marker is present; it never closes an item unless it is implemented, unreproducible, duplicate, incoherent, or obviously stale. If a maintainer has added their own comment, the bot leaves the item open. When the bot does decide to close or comment, it edits one “marker‑backed” comment per issue or PR rather than stacking new comments. This keeps the discussion clean and ensures that maintainers have a single source of truth.

The tool runs three separate hosted lanes: review, apply, and repair. The review lane triggers on events, hot intake, and normal backfill and creates the reports. The apply lane performs guarded close or comment mutations only after re‑fetching the live GitHub state in every write. The repair lane handles a loop of review, fix, re‑review, and merge for PRs that opt in or for strict reproducible bug issues that open a guarded generated pull request. There is also a local branch review lane that is used before a PR is created.

Key features

Scans all issues, pull requests, and commits weekly
Generates one markdown report per item with evidence, decision, and snapshot hash
Keeps a single marker‑backed comment per item, avoiding comment noise
Acts only when a decision is strongly supported by evidence
Separate hosted lanes for review, apply, and repair
Runnable locally with `pnpm run plan`, `review`, and `apply-decisions`