Claude Code’s Auto‑Fix PR: Sleep While It Repairs Your Pull Requests (At High Token Cost)

Claude Code has launched a new feature called Auto‑fix PR that lets developers submit a pull request and then step away while the AI monitors the PR, fixes CI failures, and addresses reviewer comments.

The service subscribes to all GitHub events for the PR. When a CI job fails, Claude fetches the failure logs, diagnoses the cause, and if a fix is clear, edits the code and pushes the change, notifying you of the action. For example, a test expected the route /login but the code had been changed to /signin; Claude corrected the test file and pushed the fix, causing CI to rerun.

If a reviewer leaves a comment, Claude will automatically resolve it when the intent is unambiguous; for ambiguous feedback or architectural decisions, it asks the developer instead of acting on its own.

Claude can also handle flaky tests as well as genuine failures, according to a response from Anthropic engineer Noah Zweben.

"This happens remotely so you can fully walk away and come back to a ready‑to‑go PR." – Noah Zweben, Anthropic

All replies posted by Claude on GitHub appear under your account but are clearly labeled "from Claude Code," providing transparency.

There are three ways to enable Auto‑fix:

From the Claude Code web UI: after creating the PR, open the CI status bar, select Auto‑fix, and enable it.

From the mobile app: tell Claude "watch this PR and fix any CI failures or review comments"; a quick‑action shortcut is also available.

For any existing PR: paste the PR link into a Claude conversation and ask to enable auto‑fix, which works regardless of where the PR originated.

Before using the feature, you must install the Claude GitHub App (github.com/apps/claude) in your repository.

The interface also offers an Auto‑merge option, disabled by default, which automatically merges the PR after all required checks pass. Enabling it can save effort but carries the risk of merging changes you haven’t personally reviewed.

Running a Claude Code session in the cloud consumes a notable amount of tokens, especially when continuously listening, analyzing, and pushing fixes for a PR. Parallel tasks increase the rate‑limit proportionally, and enabling auto‑fix on multiple PRs can quickly become expensive.

The feature is currently in a research‑preview stage and is available to premium‑seat users on Pro, Max, Team, and Enterprise plans.