checkup — the canonical project checkup ("botte doctor")

So a "complete checkup" doesn't depend on you writing a good prompt: one command runs the standard sequence in the right order and prints a single verdict, all local / 0 cloud tokens.

python -m skills.checkup.cli            # current dir
python -m skills.checkup.cli <project> --json
python -m skills.checkup.cli . --pr-comment   # Markdown verdict for a PR comment
python -m skills.checkup.cli . --doctor       # + machine scan + ranked top-3 opportunities
python -m skills.checkup.cli . --doctor --fresh   # re-scan for local LLM backends (not cached)

--doctor — the one-verb assembly

--doctor is checkup plus two things it doesn't otherwise do:

• machine scan — is a local LLM backend reachable (LM Studio/Ollama)? via [[llm_backends]] audit(); reads the cached registry unless --fresh.
• ranked top-3 opportunities — fixes from [[fix]] find_fixes(), ranked by estimated token cost (highest first), plus any remaining drift items, capped at 3.

Both feed a one-line verdict: ✅ sain when there's no drift and a local backend is active, or ⚠️ N optimisation(s), ~X tokens d'opportunités otherwise. Pure assembly of existing modules — no new skill, no new report format.

Sequence (cheap → deeper)

1. policy — is .botte/policy.md committed? (shared rules for devs/agents)
2. directives — CLAUDE.md/AGENTS.md health + stale path refs
3. metrics — LOC per component + always-on cost + savings framing
4. infra — hardware/software/MCP cluster tips (+ ASCII diagram)
5. duplication — stdlib AST duplicate-function scan
6. security — taint / data-flow scan ([[fallow_like]] TaintAnalyzer), CWE-tagged; high-severity findings become drift. 0 cloud tokens.
7. drift — MCP wired? directives stale/oversized? policy missing? security?

Then points at the deep code audit (secrets/dead-code) for when you want it.

Why

A component update or another dev/agent introduces drift — stale directives, unwired MCP, an oversized CLAUDE.md, a missing policy. /checkup catches all of it in one pass, with a fixed optimal procedure, instead of an ad-hoc prompt that may be sub-optimal and may not prioritise local models.

On pull requests

--pr-comment prints a verdict-first Markdown comment (carrying a stable marker so a bot can edit it in place), including the security section — so the taint/data-flow scan rides into CI for free. The 🧦 Botte Checkup (PR) GitHub Action (.github/workflows/botte-pr-checkup.yml) runs this on every PR and posts/updates a single comment via gh — 0 cloud tokens, no extra dependencies. The workflow sets BOTTE_CHECKUP_CONTEXT=github-pr, so machine-local MCP wiring is reported as not applicable instead of impossible CI drift; local checkups still flag missing wiring. Any project that deployed botte-secrète can reuse the same workflow.

Related: [[preflight]] (enforces prefer-local every turn), [[infra_advisor]], [[metrics]], [[directives_audit]], [[bootstrap]].