Threat Model Skill — Security skill for Claude Code

threat-model — a Claude Code skill: write the constitution before you run the scanner

[繁體中文說明 → README.zh-TW.md](README.zh-TW.md)

A security scanner without a threat model reports everything and nothing. This skill makes Claude Code first write down what must **never** happen in your project, then scan against that list, triage every finding by *Source → Control → Sink*, and fix at the boundary with a regression test that fails on the unmodified code.

It is the "five questions" method from OpenAI's Codex Security plugin and defuse's invariant-centric threat modeling, made tool-agnostic and packaged for Claude Code.

What it does

Phase	What Claude does	What you get
A · Constitution	Answers the five questions from the code (assets, roles × data, trust boundaries, invariants, entry points) and asks you only what code cannot tell	docs/security/THREAT_MODEL.md with numbered invariants, each with an enforcement level (L1 documented / L3 tested / L5 runtime-blocked) and a way to verify it
B · Scan, two layers	Deterministic tools first (gitleaks, pip-audit, bandit, npm audit; each one either RAN or is reported SKIPPED), then a semantic pass driven by the invariants: "show me a request that violates INV-3"	Raw outputs in a scratch dir; candidates, not verdicts
C · Triage	Every candidate gets source / control (missing, bypassed, mis-scoped, incomplete) / sink / the invariant it violates / an executed reproducer / a mechanical severity	docs/security/findings.md, a ledger with stable IDs and states; anything with no invariant is routed to its owner instead of buried
D · Fix	Fix at the boundary, write a test that fails before and passes after, run both ways, promote the invariant's level, record proof gaps honestly	Commits that carry the test and the updated model together

Design points that matter in practice:

• The roles matrix is a test, not a table. The skill has Claude walk the framewor