What you need
PullYard
Any current version, on macOS, Windows or Linux. Install it here.
Claude Code CLI
Installed and signed in with your own account. PullYard runs the CLI; it doesn't ship one.
The repository on your computer
The agent edits a local copy, so it needs the project cloned.
Step by step
From a red check to a pushed fix.
- Open the pull request with the red checkPullYard shows a banner with the number of failing checks and a Fix with Claude Code button. The same button is on every failing check and in the CI log view.
- Click Fix with Claude CodePullYard finds the failing step in each GitHub Actions log, keeps the error and the lines leading up to it, and adds the check's annotations.
- Read the promptThe Fix sheet opens with a prompt ready to go. Edit it if you know something the log doesn't.
- Click LaunchNothing runs until you do. The agent then works in a session next to the pull request, and you approve its tool calls and see each edit as a diff.
- Push, then re-runCommit and push the fix. If the failure was a flake, Re-run failed jobs restarts the GitHub Actions jobs after a confirmation.
What the agent actually sees
A CI log can run to tens of thousands of lines, and almost none of them matter. PullYard looks for the step that failed, keeps the error and the lines right before it, and adds the annotations the check reported.
That's what goes into the prompt, along with the check's name and the pull request it belongs to. All of it is wrapped in marked blocks and labelled as data, so text inside a log can't pass itself off as an instruction.
Your GitHub token is never sent to the log storage host.
Questions
Does it work with CI other than GitHub Actions?
Partly. For GitHub Actions PullYard reads the log and trims it to the failure. Checks from other CI services contribute their title, summary and details link.
Can a pull request trick the agent through the log?
PullYard guards against it. Logs, check names and pull request titles go to the agent inside clearly marked blocks with a random marker, labelled as data and not instructions.
Will a huge log slow things down?
No. Logs are streamed and trimmed to the failure, so a very large log doesn't fill memory.
Does the agent start on its own when CI fails?
No. PullYard prepares the prompt, and nothing runs until you click Launch.
Next red check, try the button.
Free and open source, for macOS, Windows and Linux. On macOS: brew install --cask pullyard/tap/pullyard