Claude Code auto‑fixes Playwright failures

Anthropic’s Claude Code Action now includes a public GitHub Actions example for automatically fixing failed continuous integration runs and pushing a patch branch. (github.com) The workflow is named “Auto Fix CI Failures,” triggers on `workflow_run` after a failed CI job, checks out the pull request branch, and grants `contents: write` and `pull-requests: write` permissions. (github.com) Anthropic’s repository says the action is a general-purpose Claude Code tool for GitHub pull requests and issues that can answer questions and implement code changes. The repo showed about 7,300 stars and 1,800 forks when fetched. (github.com) The Playwright angle comes from how teams use browser tests: when an end-to-end check fails, somebody has to decide whether the page changed, the test is flaky, or the product actually broke. A Playwright-based QA agent repo describes that decision as a repeatable pipeline step. (github.com) That repo runs Playwright in GitHub Actions, saves JSON reports, screenshots, traces, and HTML output, then has an agent read the failure context from `test-results/results.json`. It classifies failures into broken locators, real bugs, flaky tests, or environment issues. (github.com) In that setup, only some categories lead to write actions: broken locators can open an automation bug or a linked healing pull request, while flaky and environment issues are logged without automated code changes. (github.com) A separate “Self-Healing PR” guide built on Claude Code uses the same shape: a human push triggers CI, a failure triggers Claude, Claude pushes one fix commit, and the workflow stops after a single bot-authored attempt. (github.com) That guide says the default `GITHUB_TOKEN` will not retrigger workflows on bot commits because GitHub blocks infinite loops, so it uses a GitHub App token instead. It also documents an opt-out `no-autofix` label and a guard that skips reruns if the last commit author is a bot. (github.com) Anthropic’s own example carries a sharper warning: it runs build and test commands from the pull request branch in a trusted context with write access and secrets. The file says not to use it in repositories that accept contributions from untrusted or semi-trusted collaborators. (github.com) So the story is less that Claude Code “magically fixes Playwright,” and more that public workflows now show how teams are wiring a coding agent into the test-failure loop. The automation can draft patches fast, but the published examples are built around branch limits, bot checks, and human review. (github.com 1) (github.com 2)