git:20261009.a513934 to git:20261011.6165ab8

10 added, 8 removed. Audit A to A.

# AGENTS.md

A Claude Code plugin marketplace in ESM.

The runtime scripts in `plugins/` are `.mjs` files on Node.js 22.18 or later, so they use only `node:` modules.

A file in the import graph of a hooks module (`hooks/register.mjs`) has no `node:` import, because the module runs with no Node.

- Code that needs Python is Python, and the tests, `tools/`, and evals run on Bun.

- The core plugin is in `plugins/dotclaude/`, next to the `dotclaude-browser` and `dotclaude-jev` add-on plugins.

+ Code that needs Python is Python, and the tests and `tools/` run on Node.js 22.18 or later, with pnpm for the dev dependencies.

+ The core plugin is in `plugins/dotclaude/`, next to six add-on plugins: `dotclaude-codegraph`, `dotclaude-browser`, `dotclaude-jev`, `dotclaude-docs`, `dotclaude-sessions`, and `dotclaude-lab`.

Each add-on plugin works without the core plugin, so it imports nothing from another plugin.

Only `plugins/` ships to users.

`tests/` has one folder for each plugin, and `tools/` holds the repository scripts.

- `plugins/dotclaude/tests/` holds the hook lab, which `claude plugin test` runs against the hooks module.

+ `plugins/<name>/tests/` holds the hook lab of a plugin with a hooks module, which `claude plugin test` runs against that module.

Docs, design, and evidence: the [wiki](https://github.com/xsyetopz/dotclaude/wiki), with its source in `wiki/`.

## Commands (repository root)

- - `just check` runs lint, tests, validation, and the hook lab.

+ - `just check` runs lint, the type check, tests, validation, and the hook labs.

It must pass before done.

- - `bun test tests/dotclaude/guard.test.mjs` runs one test file.

- - `just lab` runs the hook lab: the tests of the hooks module with stubbed Claude Code events.

+ - `just test tests/dotclaude/guard/rules.test.mjs` runs one test file with `node --test`.

+ - `just lab` runs the hook lab of each plugin: the tests of its hooks module with stubbed Claude Code events.

Test each change to `hooks/register.mjs` there before a live session.

- - `bunx markdownlint-cli2 README.md` lints Markdown.

+ - `pnpm run lint:md` lints the Markdown files.

Headings and code blocks have a 100-column bound.

- `just sandbox` runs Claude Code with this checkout in a separate config.

- Read `wiki/Sandbox.md` first.

+ Read `wiki/plugins/Sandbox.md` first.

## Rules

+ - Use the env vars of Claude Code and flags.

+ Add a dotclaude env var only when neither can do the work, and give it the prefix `DOTCLAUDE_`.

- Keep each file at 300 lines or less, and each test file at 500 lines or less.

Split a larger file into files that each hold one concern, and split a long `.md` file into linked parts.

Generated files, such as `schemas/`, are out of scope.

- Write docs in STE-flavored ASD-STE100.

Use no semicolons.

- Use semantic line breaks in all text that is not code: Markdown, comments, messages to Claude, commit messages, pull requests, and chat replies.

Start each sentence on a new line, and break a long sentence only between clauses.

Never break a line at a column or in a word.

- Put each code item in inline code (single backticks) in all text that is not code: Markdown, comments, test names, messages to Claude, commit messages, pull requests, and chat replies.

A code item is a command, a tool name, a file path, an identifier, a settings key, an env var, or a literal value, such as `cat F`, `Read`, `hooks.json`, and `maxTurns`.

The backticks show a reader, a person or a model, which words are code and which words are prose.

- Write each message that goes to Claude (hook output, deny reasons, skill and agent prompts) in strict ASD-STE100.

Follow Claude's prompting best practices: give the reason, say what to do, and use no forceful words.

- Change a shared runtime bound only in `plugins/dotclaude/lib/budget.mjs`,

a feature bound only in `features/<name>/limits.mjs`,

and a repository bound only in `tools/limits.mjs`.

Tests pin the copies of each bound in code and config, not in prose.

- Do not write runtime JavaScript (`plugins/`) for a feature that the latest Claude Code has.

When Claude Code has a setting, a hook, or another extension point for a need, use it.

- In a plugin, `lib/` imports only `lib/`.

A folder in `features/` imports only `lib/` and itself, never another feature.

The entry points (`hooks/register.mjs`, `features/*/cli.mjs`, and the skill scripts) import `lib/` and features.

No file in `plugins/` imports another plugin, `tools/`, or `tests/`.

- Tests give commands to the guards as strings.

Never run a guarded command.

- Never patch the Claude Code CLI binary or its npm package, although it is minified JavaScript that you can read.

A patch violates Anthropic's terms of service.

Use only documented extension points.

Reading the bundle for evidence is permitted.

- Add CHANGELOG entries under `[Unreleased]`.

Change versions with `just bump`.

- Write public docs in `wiki/`, and run `just wiki` to publish them.

`docs/README.md` only links to the wiki.

`external/` holds local sources, and git ignores it.