Merging an AI PR feels good. Reverting it three days later doesn’t.
agent-recall tracks AI-coauthored merged pull requests across the top public OSS repositories
and watches what happens next — reverts (within 14 days) and
hotfix-follows (within 7 days). Real GitHub data only. No mocks.
—
—
—
—
—
—
Recall Score leaderboard
Higher score = more post-merge regret. 0–20 green 21–50 yellow 51–100 red
| # |
Agent |
Recall |
Revert rate |
Hotfix rate |
MTTR |
Merges (30d) |
Reverts |
Hotfixes |
|
Biggest regret this week
loading…
Recent regrets
Most recent revert and hotfix events linked to AI-coauthored merges.
Monitored repositories
Top OSS repos that have produced AI-coauthored merges, ranked by AI merge volume.
| Repository |
Stars |
AI merges |
Reverts |
Hotfixes |
About agent-recall
agent-recall answers a question that every dashboard about AI coding agents
has so far ducked: when an AI-coauthored PR does merge, does it actually survive?
Methodology
- We discover AI-coauthored merged PRs across the top 200 OSS repositories on public GitHub using the GitHub Search API. Each agent is detected by its public
Co-Authored-By: trailer (Claude Code, GitHub Copilot, Cursor, OpenAI Codex, Devin, Aider, Jules/Gemini, Cline).
- We watch for revert PRs (GitHub’s standard
Revert "..." (#N) format) opened within 14 days of an AI merge, cross-referenced by PR number.
- We watch for hotfix-follow PRs — pull requests whose title matches
fix: / hotfix: / bug: / patch: and whose changed-file list overlaps with the AI merge’s changed files, opened within 7 days.
- Recall Score =
round(min(100, 40 · revert_rate + 30 · hotfix_rate + 30 · (1 − mttr hours/168))). Higher is worse.
- All numbers refresh every 30–60 minutes from cron. No mocks, no seeds, no synthetic fallbacks — if GitHub is unavailable we show the last cached value with a chip.
Caveats
- Only revert PRs that follow GitHub’s auto-generated
Revert "..." (#N) title are detected reliably. Manual squash-rollbacks are missed.
- Hotfix detection is a heuristic; a follow-up PR touching the same files for unrelated reasons can cause a false positive. We log files-overlap-count so you can audit.
- Attribution is at the agent boundary — the human author who pressed merge is not held accountable here.
- The 5,000-star floor and pushed-in-last-30-days filter limit the repo universe to roughly 200 active repos. We re-scan daily.
Sister products
merge-trail (merge rates),
agent-bloat (PR diff size),
pr-bounce (rejection reasons).