claude-test

A scratch repository for exercising the Claude Code review workflow in isolation, so the poupe-ui repositories that ship code do not collect diagnostic commits.

Why it exists

In poupe-ui/eslint-plugin-tailwindcss the review job exited success on eight pull requests without ever posting a comment. The one run whose log survived recorded two permission denials and an empty inline-comment buffer, and the SDK output is suppressed by default, so the denied tools could not be named.

Two things make that hard to investigate in place:

• The action refuses to run when the workflow file on the branch differs from the copy on the default branch, so a pull request that edits the review workflow can never exercise it.
• A skipped action still reports success, so a green check does not mean a review happened.

Here the workflow ships on main from the first commit with show_full_output: true, so any pull request opened against it runs the real thing with its output visible.

Layout

• .github/workflows/claude-code-review.yml — the workflow under test, matching the sibling repositories apart from the diagnostic settings.
• src/ — throwaway code that exists only to give the reviewer something to look at.
• AGENTS.md and src/AGENTS.md — guidance for agents, laid out as in darvaza-proxy/x, each with one rule the review can check.