audit-toolchain
Use when asked whether a repo's tooling is current, before a toolchain migration, or when a dependency or build step feels dated. Compares the repo to tech-radar.md and the stack profile and reports g
Файлы2
Открыть в GitHubSKILL.mdskills/core/audit-toolchain/SKILL.md
GitHubAudit Toolchain
Read-only report: where the repo differs from the Adopt rows of tech-radar.md and from its own stack profile. Migration cost scale and skill mapping: migration-map.md.
1. Audit (change nothing)
cat .claude/stack-profile.md 2>/dev/null
ls pnpm-lock.yaml package-lock.json bun.lock bun.lockb yarn.lock uv.lock go.mod Cargo.toml mise.toml .tool-versions .nvmrc justfile Makefile 2>/dev/null
cat package.json 2>/dev/null | head -80
grep -E '^(go|toolchain) ' go.mod 2>/dev/null; grep -E 'requires-python|\[tool\.(ruff|uv|mypy|pyright)' pyproject.toml 2>/dev/null; grep -E '^(edition|rust-version)' Cargo.toml 2>/dev/null
ls .github/workflows 2>/dev/null; grep -hE 'uses:|node-version|go-version|python-version' .github/workflows/*.y*ml 2>/dev/null | head -40
ls biome.json* eslint.config.* .eslintrc* .prettierrc* .oxlintrc.json .golangci.y*ml ruff.toml 2>/dev/null
Profile present: use it as the intended stack and report drift from it. Absent: detect per the table in stack-profile.md; if the answer is ambiguous and changes the report, ask one question and suggest set-up-stack-profile.
2. Decide
- Everything matches Adopt rows → print "already in place" with the checked date and stop.
- Else continue.
3. Verify the radar is fresh
The radar carries a date. If it is older than 90 days, or any row you will cite is unverified, re-check that row live per version-protocol.md first (npm view <pkg> version, curl PyPI JSON, gh api repos/<o>/<r>/releases/latest). Report the real latest, not the radar's copy.
4. Compare
For each tool the repo uses, find its row and classify:
| Finding | Meaning |
|---|---|
| Adopt-gap | The repo lacks the Adopt choice for a concern (no formatter; Node pinned two LTS lines behind; no runtime manager file). |
| Hold-in-use | The repo uses a Hold item (ESLint+Prettier in a new repo, Terraform BUSL, an end-of-life Node). |
| Behind | Right tool, old major (versions from step 3). |
| Drift | The repo differs from its own profile (profile says Biome; repo runs ESLint). |
| Fine | Matches. Not listed. |
| Unlisted | Tool not on the radar: report it with no ring; do not guess. |
Check these concerns, in order: runtime (Node/Go/Python/Rust line), runtime manager, package manager, bundler, linter/formatter, type checker, test runner, CI pins (actions by SHA), task runner, observability seam, IaC.
5. Report
One table, highest value first:
| # | Finding | Evidence (path:line) |
Radar ring | Cost S/M/L | Fixed by | Risk |
|---|
- Cost: S under half a day, M one to three days, L more than three days or touching many files (scale in migration-map.md).
- Fixed by: the skill that performs the migration, from the map. If no skill covers it, write "no skill" and offer
extend-skillset. - Risk: what could break and how to roll back.
Then a short "Do first" list (at most three) chosen by value over cost, and a "Leave alone" list for Hold items where migration cost outweighs the benefit (with the reason). Respect the principle "consistency beats local optimum" from engineering-principles.md: a half-migrated repo is worse than an old one.
6. Act only when asked
Changing anything needs a go-ahead from the user per item. When approved, run the mapped skill; one migration per PR. Never upgrade a user-pinned version unasked.
7. Verify
The report ran only read commands: git status --porcelain is unchanged. Every cited line number exists (sed -n '<n>p' <file>). Every "latest" has a command and a date behind it.