You push a branch, open a PR, and walk away. CI fails on a flaky lint rule twenty minutes later. Normally that means coming back, reading logs, fixing one line, pushing again. With auto-fix turned on, Claude reads the failure, pushes the fix, and explains what it did — while you're still at lunch.
Key Takeaways
- Auto-fix lets Claude watch a PR and respond to CI failures and review comments on its own
- Turn it on from your terminal with one command:
/autofix-pr - Clear fixes get pushed automatically; ambiguous review comments get escalated to you instead
- It runs on Claude Code on the web (cloud), so your laptop can be closed
- Requires the Claude GitHub App installed on the repo
What auto-fix actually does
Once active, Claude subscribes to GitHub activity on your PR. Two kinds of events wake it up:
CI check failures. Claude reads the failing check, investigates the cause in the code, and pushes a fix if one is clear.
Review comments. When a reviewer leaves a comment, Claude evaluates it. If the request is unambiguous ("rename this variable", "add a null check"), it makes the change and replies in the session explaining what it did.
The part I appreciate most: it knows when not to act. If a comment could be read multiple ways, or touches something architecturally significant, Claude asks you before doing anything instead of guessing.
How to turn it on
Two ways, depending on where you are:
From your terminal — while on the PR's branch, just run:
/autofix-pr
This infers the open PR for your current branch and enables auto-fix for it on Claude Code on the web, in one step. No browser needed.
From Claude Code on the web — after creating a PR there, open the CI status bar and toggle Auto-fix in the CI panel.
What you need
- The Claude GitHub App installed on your repository (the setup flow prompts you if it isn't)
- Claude Code on the web enabled on your plan (Pro, Max, Team, or Enterprise)
- An open PR — auto-fix attaches to a specific pull request, not the whole repo
Things worth knowing before you rely on it
💡 The fix happens in a cloud session, not on your machine. You can open the session afterwards to read exactly what Claude changed and why — every auto-fix comes with an explanation, not just a silent push.
💡 It respects earlier instructions. If you told Claude "don't touch the test fixtures" in the session, a CI failure in fixtures gets surfaced to you instead of auto-fixed.
⚠️ 🔍 From the docs: auto-fix responds to events on that PR. Pushing a new branch with new failures does nothing until you enable auto-fix for that PR too. It's per-PR, not per-repo.
What it won't do
❌ It won't merge for you. Auto-fix pushes commits and replies to comments, but merge decisions stay human.
❌ It can't fix what it can't see. If your CI failure comes from a secret or an internal service the cloud environment can't reach, Claude will tell you rather than push a blind guess.
When this changes your workflow
The real shift isn't any single fix — it's that PR babysitting disappears. Push before a meeting, and the lint failures, type errors, and "please add a test" comments are handled by the time you're back. You review the final state instead of shepherding every red X.
Details live in the official docs: Claude Code on the web — auto-fix pull requests.