Data teardown · 🦞 openclaw/openclaw · day 222
100,000 contributions.
One flat line.
OpenClaw crossed 100,000 issues and pull requests in 222 days, built mostly by AI agents with no central team. In the replies, someone asked whether that number is real progress or just churn. I went through the repo to find out, and the number that stuck with me had nothing to do with merges.
The finding
A codebase that never sleeps
Human repos follow the clock. Commits climb through the working day and drop off overnight, wherever the contributors happen to live. Line each project up against its own busiest hour and the pattern is easy to see: React, VS Code and Linux all fall close to zero at night. OpenClaw barely moves — its quietest hour still runs at 60% of peak.
A human team’s busiest hour has up to 29× the commits of its quietest. OpenClaw’s is 1.67×. With agents doing most of the work, there’s no evening slowdown and no weekend, so activity holds steady straight through the night.
The churn question, answered
The firehose, and the filter
56,268 pull requests in 222 days, about 253 a day, and roughly one in five merged. A 22% merge rate looks rough on its own, but the month-by-month view is kinder: most of the rejected PRs came during the early rush, and the rate kept climbing after that.
The firehose, by outcome
The swarm matured
merge rate of PRs opened that month
Most of that 72% was PRs getting turned down before they ever landed, not merged work that broke and had to be pulled back out. Volume peaked in March at 16,668 PRs with the merge rate down at 12%; as the rush cooled it climbed past 41%, and almost nothing that did merge was later reverted.
A few more things in the data
Field notes
Governance
It rewrote its own constitution 442 times
The agents’ rulebook, AGENTS.md, was touched in 442 commits — roughly twice a day. The human-facing CLAUDE.md changed 5 times. Nearly all of that editing is the agents tuning the rules they run on.
Authorship tells
Chatbot politeness, fossilised in git
The conversational habits of the models leak into the permanent record. Count them across every issue, PR and comment:
- “I cannot…”1,014×
- “…as an AI…”533×
- “let me fix…”220×
- “…absolutely right”198×
- “I apologize / my apologies”130×
- “great question”26×
Durability
34 times the swarm undid itself
Of 12,399 merged PRs, only 34 were ever reverted — 0.27%. Small enough to read every one. A few of the more revealing:
- revert(cli): remove global root refusal
- Revert "Check ClawHub trust before plugin installs"
- Revert "fix(memory): warn on gateway watcher FD risk"
Vanity metric
The lobster index
commit messages include a literal lobster emoji — the project’s mascot, turning up right there in the git log.
Method & caveats
All figures pulled from the public GitHub REST & Search APIs on 4 July 2026. “Merged” is a cohort rate: of the PRs opened in a given month, the share later merged. July is month-to-date. Reverts count merged PRs whose title matches GitHub’s auto-generated Revert "…" form (an undercount if reverts landed another way). The heartbeat uses GitHub’s 52-week commit punch-card, normalised per repo to its own busiest hour. Not shown: keyword-based “which model wrote it” counts — too easily confused with mentions of the tools as product features to publish honestly.
Mined by @sushsrinivasan.
Home →