mirror of
https://github.com/got-feedBack/feedBack.git
synced 2026-09-13 08:29:28 +00:00
Miguel_LZPFandBret Mogilefsky
b45751164f
Addresses 12 of 13 review comments from Copilot and CodeRabbit on PR #332. One comment (no-manifests in validate-plugins.yml) is declined and answered inline; the rest are applied here. Substantive fixes: - .github/workflows/validate-plugins.yml — add --noconftest to the schema-tests step. tests/conftest.py imports structlog at module level, but the CI job only installs requirements-test.txt (pytest/httpx/jsonschema), so pytest collection would fail at conftest import. The schema tests don't use shared fixtures, so skipping conftest is safe and avoids dragging the full runtime requirements into a 2 KB validation job. (Copilot) - schema/plugin.schema.json — tighten the server_files regex on both settings.server_files and diagnostics.server_files to match the runtime _validate_relpath rules in plugins/__init__.py. The previous regex only blocked absolute paths, drive letters, backslashes, and "..". The runtime also rejects "//", "./", "/./", and leading-dotfile segments. Schema-valid manifests are now also load-time-valid. Verified the regex against 12 cases: the 3 in-tree manifests still validate. (Copilot) - .claude/skills/plugin-validate/SKILL.md — add a per-iteration plugin_ok flag so we no longer print "OK <path>" after an earlier FAIL in the same manifest. Schema-pass + id-mismatch previously produced both FAIL and OK lines for one plugin. (CodeRabbit) - docs/websocket-protocol.md — clarify song_info.tuning array length is source-dependent (typically 6 guitar, 4 bass, but extended-range GP imports can be 7/8/5/6). Recommend highway.getStringCount() for the authoritative count. Line 30 already said this; the table row on line 12 was the stale half. (CodeRabbit) Trivial fixes: - .claude/rules/plugin-author.md — "wants included" -> "wants to include" in the settings.server_files rule. (CodeRabbit) - Markdown MD040 — add `text` language tags to 7 bare-fence code blocks across AGENTS.md, docs/PLUGIN_AUTHORING.md, docs/testing-plugins.md, docs/plugin-logging.md, .claude/README.md, .claude/agents/slopsmith-reviewer.md, and .claude/skills/plugin-validate/SKILL.md (two fences). (CodeRabbit) Declined: - .github/workflows/validate-plugins.yml no-manifests -> exit 0 (CodeRabbit suggested exit 1). Plugins in this repo are in-tree, not submodules (no .gitmodules, git submodule status empty), and the workflow has a path filter on plugins/**/plugin.json so it only runs when a manifest actually changes. Exit 0 is correct. Answered inline on the PR. Verification: pytest tests/test_plugin_schema.py -v --noconftest # 8 passed python -c "import json,glob,jsonschema; s=json.load(open('schema/plugin.schema.json')); [jsonschema.validate(json.load(open(p)), s) for p in sorted(glob.glob('plugins/*/plugin.json'))]" # ok — all 3 in-tree manifests validate against tightened schema Signed-off-by: Miguel_LZPF <mgcdreamer@gmail.com>
43 lines
2.8 KiB
Markdown
43 lines
2.8 KiB
Markdown
# `.claude/` — Claude Code surfaces
|
|
|
|
This directory holds [Claude Code](https://claude.ai/code) artifacts: skills, rules, and subagents. **Other AI tools** (Cursor, Copilot, Codex, Aider) should read [`AGENTS.md`](../AGENTS.md) instead — `.claude/` is Claude-Code-specific.
|
|
|
|
## Layout
|
|
|
|
```text
|
|
.claude/
|
|
├── agents/ Subagents (invoked via @<name>)
|
|
│ └── slopsmith-reviewer.md Plugin-aware code review
|
|
├── rules/ Globs-scoped behaviour rules
|
|
│ └── plugin-author.md Rules that fire when editing plugins/**
|
|
├── skills/ Skills (description triggers auto-invocation)
|
|
│ ├── plugin-scaffold/SKILL.md Generate a new plugin skeleton
|
|
│ ├── plugin-validate/SKILL.md Validate a plugin.json against the schema
|
|
│ └── speckit-*/ Spec-kit skills (auto-generated; don't edit)
|
|
├── settings.json Repo defaults (no hooks enabled by default)
|
|
└── README.md You are here
|
|
```
|
|
|
|
## Conventions
|
|
|
|
- **Spec-kit owns `skills/speckit-*`.** Don't edit those manually — they're regenerated by the spec-kit installer (`.specify/`).
|
|
- **Repo skills live alongside spec-kit skills** in `skills/`. New skills should have a kebab-case directory name and a `SKILL.md` with YAML frontmatter.
|
|
- **Hooks are off by default.** Hooks in `.claude/settings.json` run in every contributor's Claude Code session — enabling one imposes work on everyone. The settings file ships with `hooks: {}` and a commented example so anyone who wants to opt in can copy it locally.
|
|
- **Subagent invocation.** Use `@slopsmith-reviewer` to invoke a subagent explicitly. Plain auto-routing isn't currently configured.
|
|
|
|
## Adding a new skill
|
|
|
|
1. Create `.claude/skills/<kebab-name>/SKILL.md`.
|
|
2. YAML frontmatter must include `name` (must match directory) and `description` (must read like a sentence — Claude Code matches the description against user intent for auto-invocation).
|
|
3. Body: instructions Claude follows when the skill triggers. Keep it concrete and short — long skills bit-rot.
|
|
|
|
## Adding a new rule
|
|
|
|
1. Create `.claude/rules/<topic>.md`.
|
|
2. Frontmatter may include `globs:` (array of glob patterns scoping when the rule applies). Without `globs`, the rule loads for every session.
|
|
3. Body: terse, imperative guidance — "always do X", "never do Y", "if you see Z, …".
|
|
|
|
## Why so little here?
|
|
|
|
The maintainer's personal Claude Code rules (context-mode routing, RTK, `@git-worker`, etc.) deliberately do **not** live in this repo. Those are user-global concerns in `~/.claude/` — keeping them out of the repo keeps the shared surface lean and makes the repo equally useful to contributors using Cursor, Copilot, plain editors, etc.
|