ClawSweeper: An AI-Powered Open Source Issue Management Bot

This bot has AI review over 13,000+ issues and PRs accumulated in the openclaw/openclaw repository to sort out candidates for closing versus items to keep open.

• It operates on a 3-stage pipeline: in Plan, priorities are set; in Review, OpenAI Codex (gpt-5.4) analyzes code, commit history, and comments in read-only mode; and Apply is manually triggered by a human.

• Closing suggestions are limited to 5 conditions: already implemented, cannot reproduce, needs migration to a separate plugin, content is unclear, or abandoned for 60+ days with insufficient information.

• Processing runs across 40 parallel shards, with a 10-minute timeout per item. Review results are saved as markdown files, along with evidence such as decision rationale, file paths, and commit SHAs.

Safeguards were also built in.

• Issues and PRs authored by OWNER, MEMBER, or COLLABORATOR are excluded from auto-close targets.

• If content changes between the time of suggestion and the time of application, this is detected via a snapshot hash and skipped.

• Suggestion and application are separated, leaving room for final human intervention.

Results as of April 2025 were also disclosed.

• Reviews completed over 7 days: 8,419 issues, 5,026 PRs

• Close candidate ratio: 33.7% for issues, 11.4% for PRs

• Actually cleaned up: 3,907

• Failures and errors: 17, less than 0.1% of the total

• Daily cadence coverage: 98.2%

The setup is relatively simple. The entire logic is contained in a single TypeScript file of about 2,500 lines, with tsgo as the compiler and oxlint/oxfmt for linting and formatting. It runs on an hourly GitHub Actions cron, and includes throttling logic that automatically waits in accordance with API rate limits.

The core message is that this structure—where AI doesn't replace maintainers but instead filters out only items with clear evidence while returning the rest to humans—actually works in practice.