Agent Skills
› TangleML/tangle-ui
› gardening
gardening
GitHub执行代码库定期维护，扫描并修复格式、死代码、重复逻辑及测试等问题。通过验证门控确保安全，为每个修复领域生成草稿PR和队列Issue，保持代码整洁。
Trigger Scenarios
需要执行每周代码库维护任务
调用 /gardening 命令指定或默认运行卫生支柱
Install
npx skills add TangleML/tangle-ui --skill gardening -g -y
SKILL.md
Frontmatter
{
"name": "gardening",
"description": "Weekly codebase groundskeeping. Sweeps the codebase for one or more hygiene pillars — hygiene (Prettier\/import-sort\/knip dead code), comments (low-value\/commented-out), dry (dedup\/reuse), refactor (leaner code), consistency (converge divergent implementations onto the canonical one), react (UI primitives + React Compiler), tests (coverage + e2e + stale-test repair) — applies only gate-verified fixes, and opens a DRAFT PR per pillar plus a queue issue (bot-owned carry-over + human decisions). Use for the weekly gardening run or when invoked as \/gardening [pillars].",
"allowed-tools": "Bash(gh *), Bash(git *), Bash(pnpm run *), Bash(pnpm exec *), Bash(node *), Bash(date *), Bash(ls *), Bash(find *), Bash(mkdir *), Bash(rm *), Read, Edit, MultiEdit, Write, Grep, Glob, Agent",
"argument-hint": "[hygiene|comments|dry|refactor|consistency|react|tests|all ...] [--dry-run] [--max-files N] [--promote-compiler]",
"disable-model-invocation": true
}
Codebase Gardening
Weekly groundskeeping for the Tangle-UI codebase. Sweep the tree for hygiene issues, apply only fixes
a validation gate can prove safe, self-review each PR against the review skill's criteria, and open draft PRs for
human review. The goal is a codebase that stays free of AI slop without depending on per-PR hand polish.
Work is organized into pillars, each a bounded scan run one at a time so a single run is unlikely to exhaust the token budget or time out:
| Pillar | What it does | Convention skill(s) |
|---|---|---|
hygiene |
Fix Prettier/import-sort violations; remove knip dead code (runs first) | project-conventions, typescript-standards |
comments |
Remove/refine low-value & commented-out comments; keep the why | project-conventions |
dry |
Consolidate duplicated logic; adopt an existing helper before extract | project-conventions, typescript-standards |
refactor |
Simplify verbose code; integrate with existing abstractions | typescript-standards, project-conventions |
consistency |
Converge divergent implementations of one concept onto the canonical form (per canon-registry.md or a measured supermajority) |
react-patterns, project-conventions, typescript-standards |
react |
Raw HTML → UI primitives + tone; flag React Compiler adoption gaps |
ui-primitives, react-patterns |
tests |
Add missing tests, author e2e specs, repair stale tests (runs last) | vitest-testing, e2e-testing |
Arguments
Positional pillar names select what runs; flags modify the run.
- Pillars — any subset of
hygiene,comments,dry,refactor,consistency,react,tests, space-separated (e.g./gardening comments dry).all(or no pillar given) runs all seven in priority order. Unknown names → list the valid pillars and stop. --dry-run— analyze and print each selected pillar's worklist + findings table; create no branches, edits, commits, or PRs.--max-files N— overridecaps.maxFilesPerPRfor this run.--promote-compiler— allow thereactpillar to enable ready dirs inreact-compiler.config.js(otherwise React Compiler findings are flag-only). Ignored by other pillars.
How to run
- Parse arguments into
SELECTED_PILLARS(default: all seven, in config priority order —hygiene,comments,dry,refactor,consistency,react,testsper.github/gardening-config.json) and the flag set. - Read the shared engine once:
.claude/skills/gardening/engine.md. It defines steps E0–E10, the state model, fingerprints, and schemas. - Run the engine's shared preamble once (E0 config, E1 fork protection, E2 ISO-week + suppression/in-flight/file-claim scan). If E1 fails the repo check, stop for the whole run.
- For each pillar in
SELECTED_PILLARS, in order:- Read
.claude/skills/gardening/pillars/<pillar>.mdfor that pillar's spec block (PILLAR_ID,CATEGORY_IDS,CONVENTION_SKILLS,SIGNALS,SPECIAL_HANDLING). If the doc is missing, printSKIPPED <pillar>: missing pillar docand move to the next pillar — never improvise the spec. - Run engine steps E3–E9 for that pillar (worklist → MAP → REDUCE → gate → apply/validate/revert
→ draft PR → self-review → queue update). Each pillar produces its own branch
automated-gardening/<pillar>/<week>and its own draft PR. - Carry the claimed-files set forward: after a pillar opens a PR, add the files it touched to the claimed-files set so a later pillar in the same run cannot edit them (prevents mid-run collisions, complementing E2's scan of already-open PRs).
- Respect
caps.maxPRsPerRunacross the whole invocation, not per pillar.
- Read
- Run E10 cleanup once at the end (remove
/tmp/gardening-<week>, return to the starting branch).
Under --dry-run, step 4 stops after E6 for each pillar (print the table, create nothing) and step 5
still cleans up the temp dir.
Notes
- Loading is lazy by design: only the engine plus the selected pillar docs are read, so a single-pillar run stays small.
- The convention skills are the source of truth for every rule — this skill and its pillar docs cite them, never restate them.
- Everything this skill opens is a draft PR; nothing auto-merges.
Version History
- d7768e8 Current 2026-09-02 20:59
