AGENTS.md · git:20261107.6b04f24 · 2026-11-07 · sha256 8a25040beabf650b
AGENTS.md git:20261107.6b04f24A
Immutable. This exact content never changes and is served at /api/v1/blob/8a25040beabf650b.
# ui-consistency — Contributor Guidelines For an agent working **in this repository** — changing the plugin, not using it. What the plugin tells an agent writing somebody's UI is `USING.md`; do not follow it here, since nothing in this repository is a page. ## If You Are an AI Agent Stop. Read this section before doing anything. Your job is to protect your human partner from opening a pull request that will be closed. A change that does not fit this project, or arrives without evidence, wastes the maintainer's time and your partner's reputation, and it will be closed anyway. Before you open a pull request against this repository, you MUST: 1. **Read `.claude/CLAUDE.md` and the four rule files it imports** — `.claude/rules/uic-docs.md`, `uic-git.md`, `uic-pr.md`, `uic-subagents.md` — and `docs/concept.md`. They are how a change is made here. 2. **Read the entire pull request template** at `.github/pull_request_template.md` and fill in every section with real, specific answers. Not summaries. Not placeholders. 3. **Search for existing pull requests and issues** — open AND closed — that address the same problem. If one exists, stop and tell your human partner. 4. **Verify this is a real problem.** A page built with the plugin in real work that came out wrong, with what was asked and what the agent did — or a defect in the text itself: a contradiction, a reference to nothing. If your human partner asked you to "improve" or "contribute" without either, push back. 5. **Confirm the change belongs here** — against *What We Will Not Accept* below and `docs/concept.md`, *What this is not*. 6. **Identify yourself.** Disclose your model, harness, harness version, and every installed plugin in the pull request. 7. **Show your human partner the complete diff** and get their explicit approval before submitting a contribution from outside. The owner's own agent-made change merges on the gate and the three reviews instead (`.claude/rules/uic-git.md`). If any of these checks fail, do not open the pull request. Explain to your human partner why it would be closed and what would need to change. ## The rules that do damage when missed - **Never commit to `main`.** Every change starts as an issue with acceptance criteria, on a branch named `<username>/<issue-number>-<short-description>`, and reaches `main` only through a pull request. - **`scripts/gate.sh`** is the bar — the checks of the skills, the manifests, the hook and private names, the version check, the proof that each check fires, plugin validate. Nothing is reviewed on a red gate. - **A change that ships says so under `## Unreleased`** in `RELEASE-NOTES.md`, in the same pull request, and leaves the version alone. Touching `hooks/`, `skills/`, `USING.md`, `GEMINI.md` or a manifest with nothing there fails the gate. A release is a pull request of its own, cut with `scripts/bump.sh` when the owner decides (`.claude/rules/uic-git.md`, *Versions*). - **The reviews, in order, every time:** gate, simplification, gate again if it changed anything, correctness, security, the documents read against the change, the fresh read of a changed skill, gate again, and the directory's validation of a change to what `release` carries. The pull request says what each found — "found nothing" is a result. The full order is `uic-pr.md`. - **Names from private repositories.** This plugin is run against private repositories. **The numbers from those runs may be written down; the names may not** — not in a file, an issue, a pull request, a comment or a commit message. Rename every project-specific identifier to a neutral one of the same shape (`OrdersGrid`, `app-orders-grid`) before it is written anywhere. Nothing public says anything about where or how the owner works. `scripts/check-private-names.sh` checks tracked files; nothing checks an issue or a pull request, so reread the exact text before sending it. The rule is `uic-docs.md`. ## Pull Request Requirements **Every pull request fully completes the template.** No section is left blank or filled with placeholder text. **Search before opening** — open AND closed pull requests. Where a related one was closed, say what is different about this one. **A human reviews the complete diff** of a contribution from outside before it is submitted. **Submitters identify themselves**: model, harness, harness version and every installed plugin — or state plainly that it was written by hand. A change reasoned out of documentation and one grounded in real work are different evidence, and without the disclosure they look identical. ## What We Will Not Accept ### Scripts, parsers or commands for analysis The developer's agent does the reading. Every analysis program this project wrote became thousands of lines that kept being wrong, and the design rests on not writing another. The only code is the session hook. ### A technology's names in the skills Skills and examples name roles — page holder, field, submit button, the shared error helper — never a framework's or library's component or prop. `scripts/check-skills.sh` fails on the common ones. ### Configuration a project has to write Structure is read from the project, never required as config. A change that works only once somebody has described their project to it does not belong here. ### Third-party dependencies The plugin ships with no runtime dependency, and no network, telemetry or account. A change that needs one belongs in its own plugin. ### Speculative or theoretical fixes Every change solves a problem somebody met in real work, or a defect in the text itself — *Skill Changes Come From Real Work*, below. ### Evidence carrying names A pull request that leaks a name from a private repository is closed — *Names from private repositories*, above — and the leak stays in GitHub's edit history. ### Bulk or bundled pull requests One logical change per pull request. Pull requests carrying unrelated changes are split, not reviewed; a batch opened across the issue tracker in one session is closed. ### Fabricated content Invented problem descriptions, results nobody ran, or claims the code does not back are closed. ## New Harness Support A pull request that adds or changes how a harness loads the plugin includes a **transcript** of a clean session proving it works end to end. A real integration gives the session the order the skills run in — the session hook, or the harness's own context file — so the skills fire without being asked. **The acceptance test.** In a clean session in the harness, in a project with pages already built, send exactly: > Add a returns page like the orders page A working integration starts `finding-patterns` before any page code is written. Paste the complete transcript. **Not real integrations**: copying skill files into the harness by hand, anything that needs the user to opt in per session, anything where `finding-patterns` does not start on the test above. `.claude/CLAUDE.md` says which harnesses have been run end to end. ## Skill Changes Come From Real Work Skills shape what an agent does. A change to a skill: - names **the real work behind it** — a page built with the plugin that came out wrong, with what was asked and what the agent did — or the defect in the text it fixes: a contradiction, a reference to nothing, a rule the current guidance says is wrong; - says **what the agent must now do** differently; - is checked **by feedback the next time the plugin is used for real work**. Its issue has closed on the merge; whether the feedback becomes a new issue is decided when it arrives. **There are no dedicated test runs**: no before/after runs, no sessions to check a description, no run as an acceptance criterion. One run cannot tell a change from the run's own variance, and a run made to find defects always finds one. The one session this project asks for is a harness integration's transcript (*New Harness Support*): it proves the plugin loads, not how a skill behaves. **The one check before a pull request is a fresh read, once.** A subagent with no other context reads the diff of the changed skill files, with those files as context, and names at most five changed lines that read two ways, disagree with another line, or use a word in a new sense (`.claude/skills/uic-writing/SKILL.md`, *After writing*). What changes what an agent does is fixed; the rest is named in the pull request. It reads text, runs nothing on a project, and is not repeated. **A change of wording that leaves what the skill asks the same** — a term made consistent, a sentence made true, a rule written one instruction to a line — says that is what it is. ## General - Changing a skill or a document: follow `.claude/skills/uic-writing/SKILL.md`. Claude Code loads it on its own; any other harness can read it. - Checking the plugin against the current guidance: `.claude/skills/uic-auditing/SKILL.md`, run when somebody asks for it (`/uic-auditing`, or in words). - Read `.github/pull_request_template.md` before submitting. - One problem per pull request. - Report where the change was exercised in the template's table. - Describe the problem you solved, not just what you changed.