AI‑Powered Bot ClawSweeper Closes 4,000 OpenClaw Issues in One Day

Overview

ClawSweeper is an AI‑driven maintenance robot for the OpenClaw repository. It runs up to 50 parallel Codex (gpt‑5.5) instances to scan open issues and pull requests continuously. The project had a backlog of roughly 5,000 issues and 4,000+ pull requests . In a single day the bot closed about 4,000 issues , with several thousand more still being processed. Throughput is constrained by GitHub and OpenAI API rate limits.

Operation Model

For each open issue or PR ClawSweeper creates a permanent Markdown report. When the AI finds sufficient evidence it posts a durable Codex‑generated review comment and, only then, may close the item. The bot suggests closure only under the following explicit conditions:

Feature already implemented in the current main branch

Problem cannot be reproduced on the current main branch

Work is more appropriate for a ClawHub skill/plugin than core code

Issue/PR is a duplicate or superseded by an authoritative one

Content is concrete but not executable in the repository

Content is inconsistent and no action can be taken

Stale for over 60 days with insufficient data for verification

Items created by maintainers are never auto‑closed; all other items remain open unless the bot closes them.

Review Workflow

Planners scan all open issues and PRs and assign each item to a shard.

Each shard checks out the

openclaw/openclaw

main

branch.

Codex (gpt‑5.5) performs a high‑reasoning review, spending at most 10 minutes per item.

For each item a Markdown file items/<number>.md is generated, containing the decision, evidence, suggested comment, runtime metadata, and a GitHub snapshot hash.

High‑confidence closure suggestions are marked proposed_close.

Review cadence:

Hourly review of items with recent activity.

Hourly review of items created in the past 7 days.

Daily review of items inactive for 30 days.

Weekly review of older, inactive items.

New or active items are processed every 5 minutes .

Application Workflow

Read existing reports; if the stored review result is still valid, update GitHub.

Update the single tagged Codex auto‑review comment.

Close an item only when the review result is explicit and high‑confidence.

Reuse the same comment when closing to avoid duplicate messages.

Move closed or already‑closed reports to closed/<number>.md.

Move reopened archived reports back to items/<number>.md and mark them as outdated.

Perform checkpoint commits and dashboard heartbeats during long runs.

By default the application workflow closes only issues (no age limit), with a 5‑second closure delay and a maximum of 50 items per checkpoint . Excess requests are queued for the next run.

Performance and Limits

The only remaining bottleneck is not the AI model itself but the rate‑limit restrictions imposed by GitHub and OpenAI. Removing these limits could enable near‑complete removal of stale code across the platform.

Repository

Project source:

https://github.com/openclaw/clawsweeper