tools-registry-context Skill
tools-registry 上下文 + 文档维护。用于给新会话热身(了解架构和近期提交),或在处理 tools-registry 时使用——改代码、规则、内容、数据或流程(proxy · 认证/密钥 · API · CLI · registry skill)——会从 docs/context 加载相关片段,让你在准确、有出处的上下文中行动。可接受可选的焦点查询(如 `/tools-registry-context <领域>`)。也运行 `/tools-registry-context sync` 在改动后更新文档片段(展示 → 批准 → 应用)。快要推送 main 分支时务必提到它。
安装方式:把技能目录放入 ~/.claude/skills/(Claude Code)或在 claude.ai 设置中启用;也可复制右侧安装命令一键添加。
技能指令原文(SKILL.md)
tools-registry-context — load the right design fragment, keep docs honest
tools-registry's design docs are fragments under docs/context/ (one per subsystem), each declaring the
source files it covers in frontmatter. MAP.md (next to this file) is the generated reverse
index: source file → the fragment that documents it. This skill has two modes.
Mode A — LOAD context (default). Adapt to when you're called:
Cold start (a fresh session, little/no prior context) → warm up. Read
docs/context/README.md (the index), then the fragments that matter
(the foundation-style overview fragments plus whichever subsystems the query or repo state points at).
Run git log --oneline -15 and git status to catch recent commits + uncommitted work. Then give a
short orientation — what tools-registry is, the subsystems in play, what changed recently — and say you're ready.
Mid-chat (a task/topic is already in play) → stay targeted. Map the artifacts/topic at hand via
MAP.md's "Source file → fragment(s)" table, read just those fragment(s), and proceed.
Don't re-warm the whole tree.
A focus query (/tools-registry-context ) always wins — use it to pick the fragment(s) and
focus the warm-up on that area, in either case.
Always: read the data model / behavior / RCAs / symbol anchors before changing anything; load only
what's relevant (never dump the whole tree); if an artifact you touch has no fragment, note it as a
gap for Mode B.
Mode B — SYNC docs (/tools-registry-context sync, or before a push)
Before a push to the main branch, remind the user to run this; proceed only on their yes. Then
follow MAINTAINING.md — the short version:
- Detect drift:
bash .Codex/skills/tools-registry-context/scripts/drift.sh(defaults to
origin/main..HEAD). It prints, per changed source, which fragment(s) document it — plus gaps.
- Draft updates: for each affected fragment, read it + the diff; update prose to match changed
behavior and verify cited symbols still exist (no line-number chasing — symbols don't drift).
New subsystem with no fragment → draft a new fragment from fragment.md.tmpl.
- Show, then apply: present proposed changes and get approval before writing.
- Regenerate:
python3 .Codex/skills/tools-registry-context/scripts/build-map.py(rewrites README + MAP). - Commit together: doc updates ride with the code in the same commit/push.
Invariants
- Docs are the source of truth; this skill is a lens. Fragments live in
docs/context/; never
duplicate them into the skill. The skill holds only the generated MAP.md + scripts + config.
- Frontmatter drives everything. A fragment's
sources:feeds the index, the MAP, and drift. When a
fragment starts covering a new file, add it to sources: and rerun build-map.py.
- Cite stable symbols, not line numbers. Anchor every claim to a grep-able symbol; bare line
numbers drift on every edit and slow sync. Describe what shipped, not intent.
- No automation behind the user's back. Sync is reminder → approve → apply. There is no git hook.
- Handoffs and plans are NOT documentation. They live outside
docs/context/(e.g..context/) and
are out of scope — never read, fold in, or scan them as fragment input.