Project baseline

A presence check, not a code review: each practice has a probe, and a missing practice becomes a backlog item. Read-only until the user approves.

1. Context

• Detect the stack, the lockfile-selected package manager, the CI workflows, and the validation commands.
• Find the backlog paths exactly as audit-to-backlog does: a Backlog: <todo path>, <done path> line in AGENTS.md or CLAUDE.md, otherwise TODO.md and TODO-done.md. Read both if they exist and continue their category numbering under the ID prefix BASE.
• Read the project's AGENTS.md, docs/pitfalls.md, and decision records: a practice the project rejected or deferred on purpose is reported as deliberate, never proposed.

2. Probe each practice

Practice	Present when	Missing →
Env read through one validated module	a single module parses process.env (or the stack's equivalent) with a schema (zod or similar), and rg finds no direct env reads elsewhere in app code	P2
Agent deny-list for secrets	.claude/settings.json permissions.deny covers .env* and key files (*.pem, *.key)	P2
Secret scanning in CI	a CI step runs gitleaks, trufflehog, or the host's secret scanning	P2
Coverage floor that only rises	the test config sets coverage thresholds and CI enforces them	P2
Additive-only migration guard	CI fails on destructive migrations (drop/rename) without an explicit marker	P2; skip without a database
Dependency/boundary rules	dependency-cruiser, eslint boundaries, or a layer test that fails CI	P3
Real-database test tier	tests run against a real database (Docker or a CI service), not only mocks	P2; skip without a database
Safe error responses	API errors return one envelope with a code and a message (details optional), as error-handling-conventions defines; no stack traces or internal fields	P1 when a stack trace or internal error reaches a response, otherwise P2; skip without an API
CHANGELOG	CHANGELOG.md in Keep a Changelog format	P3
Docs-vs-code consistency check	a script or CI step checks docs against code (commands, paths, versions)	P3
Design system file	a project with a UI (components, pages or stylesheets in app code) has a root DESIGN.md in the DESIGN.md format (YAML tokens, then Overview … Do's and Don'ts)	P3; skip without a UI. When present, suggest npx @google/design.md lint DESIGN.md for the user to run; never run or install it
No runtime dependency two majors behind	the package manager's outdated report (pnpm outdated --prod, npm outdated --omit=dev, yarn outdated, uv pip list --outdated limited to the pyproject.toml dependencies, bundle outdated) lists no runtime dependency two or more major versions behind its latest release	P2 for a framework or runtime dependency, otherwise P3; skip when the report needs credentials or the network is unavailable, and say so; a nonzero exit from the outdated command alone is not a failure. bundle outdated evaluates the Gemfile and yarn runs the file a .yarnrc* names in yarnPath/yarn-path: both run repository code, so ask before either. Each item's Done when names the target version; its Check fails until the lockfile resolves that major, written for the lockfile's format (for example rg -q '/next@15\.' pnpm-lock.yaml). Dispatch nothing during the probe: the report offers upgrade plans (step 5)
Dependency release-age cooldown	the update bot or package manager enforces a minimum release age before a new version is taken (Dependabot cooldown, Renovate minimumReleaseAge, pnpm minimumReleaseAge, uv exclude-newer)	P3; skip without dependencies
After-edit lint hook	.claude/settings.json has a PostToolUse hook on Edit|Write that reads tool_input.file_path from stdin, lints that file, and exits 2 with only the failures on stderr	P3; skip without a linter
Agent sandbox	.claude/settings.json sets sandbox.enabled: true and ~/.ssh and ~/.aws are denied to it, in sandbox.filesystem.denyRead or as sandbox.credentials.files entries with "mode": "deny" (the sandbox reads them by default); the guard and permission rules match text, the sandbox restricts every command	P2
Pinned MCP servers	every stdio server in .mcp.json runs a pinned version (npx -y pkg@1.2.3, uvx pkg==1.2.3, an image digest), with no @latest and no bare package name; a server's tool descriptions reach the model, so a version bump is reviewed like a dependency	P2; skip without .mcp.json
Gate verdict in the exit code	every CI and validation script exits non-zero on failure (no || true, no continue-on-error on gates)	P2

Record each probe and its result as the evidence: the command or file checked, and what it found. Unsure counts as unsure, not missing: list it under questions.

3. Decisions, not defaults

Never propose a default for these; when the repository is inconsistent (two lockfiles, two test runners, mixed commit languages), list the evidence and ask the user to choose:

• Package manager (npm vs pnpm)
• Test runner (Jest vs Vitest)
• Authentication library
• Monitoring stack
• Commit message language

4. Draft the backlog items

One item per missing practice, in audit-to-backlog's shape, filed under its severity section and deduplicated against both backlog files:

- [ ] BASE-03 · P2 · Secret scanning in CI
  - Where: `.github/workflows/ci.yml:1`
  - Problem: no CI step scans for committed secrets.
  - Evidence: `rg -l 'gitleaks|trufflehog' .github/` → no match.
  - Fix: add a gitleaks step to the CI workflow that fails the job on a finding.
  - Size: S
  - Done when: CI fails on a commit that contains a secret.
  - Check: `rg -q 'gitleaks' .github/workflows/`

5. Report and ask

Project baseline — <repo>
Present: 6   Missing: 4 (P2 3, P3 1)   Deliberate: 1   Not applicable: 1
Needs your decision: test runner — package.json has both jest and vitest

Write the 4 items to TODO.md? A: all (recommended)  B: list IDs  C: none
Plan the 2 upgrades with dependency-upgrader first (reads changelogs on the web)? yes / no

The second question appears only when the dependency probe found items. On a yes, dispatch developer-os:dependency-upgrader once per dependency, for at most the three furthest behind, and put its steps into the item's Fix and its estimate into its Size before writing; mark its ASSUMPTION: lines as such in the item. The install commands in a plan are text for the user to run, never commands this skill runs.

If the backlog files do not exist yet, ask for the paths and record them as a Backlog: <todo path>, <done path> line in AGENTS.md (or CLAUDE.md), with its git-ignored path warning, as audit-to-backlog does.

Stop point

Stop after the report and the question. After approval, merge the approved items into the backlog file with Session N: <ID>, … lines of 5–7 items each (fewer than 5 items: one session) under ## Implementation sessions, grouped as audit-to-backlog groups them, show git diff --stat, and stop; implementation is implement-backlog's job.

Never

• Write or create any file before the user approves.
• Pick a default for a "needs your decision" item.
• Install dependencies, add CI steps, or change configuration; this skill only proposes.
• Commit, push, or open a pull request.
• Propose a practice the project rejected on purpose.