Claude Code launches cloud Auto-fix: automatic repair for CI failures, just walk away, and when you return, the PR is already green.

Anthropic engineer Noah Zweben（@noahzweben）announced that Claude Code has officially launched a cloud Auto-fix feature: web and mobile work sessions can now automatically track Pull Requests, continuously fix CI failures, and respond to code review comments—keeping the PRs in a green status at all times. The entire process runs fully remotely; you can completely step away, and when you come back, the PR is ready to merge.

Three ways to use it

According to the official instructions, there are three ways to enable Auto-fix:

For PRs created directly in the Claude Code web version, click “CI > auto-fix”

In the mobile version, directly tell the AI agent to run auto-fix (a quick action button is coming soon)

Paste any PR link you want to monitor, and tell the agent to run auto-fix

To use this feature, you need to install the Claude GitHub App. To use Auto-merge (automatic merging), you must enable it in the repo in advance.

“Engineers’ work shifts from writing code to reviewing AI decisions”

This update has sparked a deep discussion in the developer community about the essence of software development. Some people are excited, saying, “In the past, you had to keep your laptop running all night watching CI—now you can sleep.” Others have raised more fundamental questions: if AI agents handle everything from writing code, running tests, fixing errors, to merging PRs, what are engineers even reviewing?

Some developers have pointed to the risk: without human review, automatic fixes could push incorrect changes to production at 2 a.m., and nobody might notice. MergeShield’s observation hits the nail on the head: “Not all PRs that pass CI have the same risk. Auto-merging is fine for something like a dependency version upgrade, but rewriting auth middleware should still be reviewed by a person—even if it passes the tests.”

From “having AI write code” to “giving AI the entire feedback loop”

Observers say this update represents a real qualitative shift. In the past, the focus of AI tools was “generating code.” Now Claude Code is about “owning the feedback loop”—CI fails, the agent reads the errors, the agent patches the issue, CI reruns, until it passes. This closed-loop agent model shifts the basic unit of work from a “commit” to an “intention”: you state what you want to do, and the agent is responsible for seeing it through.

This also comes in sync with the launch of Claude Code Auto Mode. The direction of Claude Code overall is: let AI take over more execution-layer work, and the engineer’s role increasingly shifts from “executor” to “reviewer.”

This article Claude Code launches cloud Auto-fix: CI failure auto-repair, just walk away—the PR is already green when you return first appeared on Chain News ABMedia.