# claude-playground A tiny zero-dependency Node.js todo CLI, used as a sandbox for trying Claude Code features. ## Commands - `npm test` — run the test suite (built-in `node:test`, no install needed) - `npm start -- add "something"` / `npm start -- ls` — run the CLI (data in `todos.json`, override with `TODO_FILE`) ## Conventions - ES modules, Node >= 20, no third-party dependencies. - `src/todo.js` holds pure functions; keep I/O in `load`/`save` and `src/cli.js`. - Every behavior change gets a test in `test/`. - Commit messages: short imperative subject line. ## Workflow (required) Never commit, merge, or push to `main`. Only a human (Christian) merges into `main`. These rules are instructions, not a lock. Gitea branch protection on `main` is the real guard; the `git push … main` deny rules in `.claude/settings.json` only catch the obvious spellings. 1. Start every change from a fresh `main`. Run `git fetch origin` as its own command, then `git switch --no-track -c / origin/main` (e.g. `fix/duplicate-ids`, `docs/readme`). `--no-track` keeps the branch from tracking `main`. 2. Commit on that branch; run `npm test` before every push. 3. Push with `git push -u origin HEAD`, then open a PR against `main` (see "Gitea API" below). 4. Review it with `/code-review `. Always name the branch, so the review diffs it against `main` whatever its upstream is. 5. Post the result as a PR comment. The first line is the verdict: **Needs fixes** if there is any correctness finding, otherwise **Approved** (cleanup-only findings are listed but don't block). The second line is `_Automated self-review by Claude; not an independent review._`, followed by the findings. 6. If it needs fixes, fix the correctness findings on the same branch and repeat steps 2–5. At most 2 fix rounds: after that, post what's left and hand off to the human. 7. Stop there. Do not merge the PR, even if approved; tell the human it's ready. ### Gitea API No `gh`/`tea`: use `curl` with `$GITEA_TOKEN`. One-time human setup in the git-ignored `.claude/settings.local.json`: set `env.GITEA_TOKEN` and add `code.westerkam.org` to `sandbox.network.allowedDomains`. Settings load at session start, so after adding them the session must be restarted (`claude --continue`). If the token is unset, stop and ask the human; don't work around it. Shell state doesn't persist between commands, so set every variable in the same command as the call. Write comment bodies to `$TMPDIR`, never into the repo. `-fsS` makes HTTP errors fail loudly; check that the PR response has a `.number`. ```sh API=https://code.westerkam.org/api/v1/repos/christian/claude-playground # Open a PR BRANCH=fix/x TITLE="Fix x" BODY="…" jq -n --arg head "$BRANCH" --arg title "$TITLE" --arg body "$BODY" '{head:$head, base:"main", title:$title, body:$body}' \ | curl -fsS -X POST -H "Authorization: token $GITEA_TOKEN" -H "Content-Type: application/json" -d @- "$API/pulls" | jq -e .number # Comment on PR number $N N=1 jq -n --rawfile body "$TMPDIR/comment.md" '{body:$body}' \ | curl -fsS -X POST -H "Authorization: token $GITEA_TOKEN" -H "Content-Type: application/json" -d @- "$API/issues/$N/comments" | jq -r .html_url ``` Run `git fetch`, `git commit` and `git push` as standalone commands (no `cd … &&` prefix, nothing chained after them): only then do they run outside the sandbox, which reaching the remote and commit signing need. ## Remote Gitea at code.westerkam.org — `origin` = `git@code.westerkam.org:christian/claude-playground.git`.