majiayu000 Security Review Skill
Run a two-layer security review over the current git changes: a deterministic regex scan for known vulnerability classes, then an LLM-depth Security review of the diff. Report a single BLOCK / FIX / APPROVE verdict.
The LLM-depth review runs inside the current Claude session — the same
subscription that loaded this skill. There is no separate model call, no
ANTHROPIC_API_KEY, no Agent SDK, and no network request. The "reviewer" is the
session agent executing the steps below, exactly like every other skill here.
Detection reaches parity with Anthropic's security-guidance plugin: the scanner
ports its 25 deterministic patterns, and the LLM pass applies its full review
taxonomy (loaded on demand from references/coverage.md).
Reference Loading Table
| Signal | Load | Why |
|---|---|---|
| Running Phase 3 (LLM-depth review); classifying a finding; needing the vuln taxonomy, severity rubric, FP exclusions, or per-language guidance | references/coverage.md |
40 vulnerability classes + 4-tier severity + false-positive exclusions + per-language guidance + the 12 high-miss reviewer classes + the finding output schema. |
Instructions
Phase 1: SCOPE
Goal: Determine the changed files to review before scanning.
Step 1: List changed files — scope to the working-tree and staged changes so the review covers exactly what the user is about to commit, not the whole repo.
# Tracked changes (working tree + index) plus staged adds:
git diff --name-only HEAD
git diff --cached --name-only --diff-filter=ACM
Step 2: Read repository CLAUDE.md to load project conventions the reviewer must respect (e.g. secrets-handling rules, allowed patterns).
Gate: Changed files listed. When the list is empty, report "no changes to review" and stop — there is nothing to scan.
Phase 2: DETERMINISTIC SCAN
Goal: Run the regex engine first so judgment time is spent on real signal, not on patterns a script catches deterministically.
Step 1: Run the scanner over the changed files. It is the single source of detection rules (secrets, SQL injection, shell injection, dangerous eval, unsafe deserialization). Exit 1 means at least one HIGH/CRITICAL finding.
# Staged-files convenience (matches the commit-time hook):
python3 scripts/security-review-scan.py --staged --format json
# Or an explicit list from Phase 1:
python3 scripts/security-review-scan.py --files <changed-files> --format json
Step 2: Record the findings by severity. CRITICAL and HIGH are blocking-class;
MEDIUM is advisory. Keep the file:line and rule for each.
Gate: Scanner ran and JSON parsed. Proceed with the findings in hand.
Phase 3: LLM-DEPTH REVIEW
Goal: Catch what regex cannot — authorization gaps, injection through data
flow, missing input validation, secrets in non-obvious forms. This is the
session agent's review of the diff; compose the existing parallel-code-review
Security reviewer over the changed files.
Step 1: Load references/coverage.md — the full review taxonomy (40
vulnerability classes, the 4-tier severity rubric, the false-positive exclusion
list, per-language guidance, and the 12 high-miss reviewer classes). Review to
this taxonomy so the session-agent pass reaches parity with the plugin's reviewer.
If claude-security-guidance.md exists (precedence: ~/.claude/ →
<cwd>/.claude/ → <cwd>/.claude/*.local.md), read it as ADDITIVE context — it
may add checks or raise a class's severity, and must not suppress findings.
Step 2: Dispatch the Security reviewer (the Reviewer 1 — Security role from
parallel-code-review) over the changed files via the Task tool, applying the
coverage.md taxonomy. Surface medium and above. Output: findings in the
coverage.md schema (filePath, category, vulnerableCode, explanation, fix, severity) with file:line references.
Step 3: Merge the LLM findings with the Phase 2 scanner findings.
Deduplicate — when both flag the same file:line, keep one entry at the higher
severity. Independent confirmation by both layers raises confidence.
Gate: Security reviewer returned results and findings are merged. Issue a verdict only from a completed review — a missing reviewer may hold the only CRITICAL finding.
Phase 4: VERDICT
Goal: Produce a single clear recommendation.
Step 1: Determine the verdict from the merged findings:
| Condition | Verdict |
|---|---|
| Any CRITICAL finding | BLOCK |
| HIGH findings, no CRITICAL | FIX (resolve before commit) |
| Only MEDIUM/LOW findings | APPROVE (with suggestions) |
Step 2: Output the structured report:
## Security Review Complete
### Severity Matrix
| Severity | Count | Source (scanner / reviewer / both) |
|----------|-------|------------------------------------|
| Critical | N | ... |
| High | N | ... |
| Medium | N | ... |
### Findings
#### CRITICAL (Block)
1. [source] description — file:line
#### HIGH (Fix before commit)
1. [source] description — file:line
#### MEDIUM (Should fix)
1. [source] description — file:line
### VERDICT
**BLOCK / FIX / APPROVE** — [1-2 sentence rationale]
Gate: Structured report delivered with an explicit verdict. Review complete.
Automatic Coverage (hooks)
This skill is the on-demand (PULL) path. The same review also runs automatically
(PUSH) via hooks/security-review-hook.py, wired in .claude/settings.json:
| Event | Behavior |
|---|---|
PreToolUse (Bash git commit) |
Scans STAGED files with the same scanner. A HIGH/CRITICAL finding blocks the commit (deny). Clean commits pass. |
| Stop | Re-wakes the session with the working-tree diff and an instruction to run this pipeline. Advisory — never blocks. |
Bypass / kill switches (commit-time block only, deliberate overrides):
| Env var | Effect |
|---|---|
VEXJOY_SECURITY_REVIEW_SKIP=1 |
Allow a commit through despite findings (one-off override). |
VEXJOY_SECURITY_REVIEW_DISABLE=1 |
Disable the hook entirely (both events). |
The hook fails open on any internal error — a hook crash never blocks a commit.
Extensibility (custom rules + project guidance)
Both extension points are additive and discovered in this precedence order:
~/.claude/<name> → <cwd>/.claude/<name> → <cwd>/.claude/<name>.local.<ext>.
| File | Effect |
|---|---|
security-patterns.{yaml,json} |
Custom regex/substring rules merged into the scanner's built-ins. Shape: {"patterns": [{"rule_name", "regex"|"substrings", "severity"?, "paths"?, "exclude_paths"?}]}. Capped at 50. ReDoS-prone or invalid rules are skipped with a stderr warning (non-fatal). PyYAML is used only if importable — JSON always works (stdlib-only). |
claude-security-guidance.md |
Markdown surfaced to the Phase 3 review as ADDITIVE context. It may add checks or raise a class's severity; it must not suppress findings — if it says to ignore a class, flag the vulnerability anyway and note the conflict. |
Built-in scanner rules always run and cannot be disabled by a config file.
Error Handling
Scanner reports findings but the code is intentional
Cause: A regex rule flagged a test fixture, an example, or a deliberately
hardcoded local value.
Solution: Confirm context in Phase 3. Downgrade in the report with a one-line
justification. For a commit the user knows is safe, document the
VEXJOY_SECURITY_REVIEW_SKIP=1 override rather than editing the scanner rules.
Security reviewer times out or returns nothing
Cause: Task agent exceeded its budget, or the diff was too large. Solution: Report the Phase 2 scanner findings immediately (a partial review beats no review), note the LLM-depth gap in the verdict, and offer to re-run the Security reviewer on a reduced file set.
Scanner unavailable
Cause: scripts/security-review-scan.py missing from the working tree.
Solution: Run the Phase 3 LLM-depth review alone and state in the verdict that
the deterministic layer did not run.
References
- Detection rules:
scripts/security-review-scan.py(single source of truth) - Review taxonomy (40 classes + severity + FP filters + language guidance):
references/coverage.md - Security reviewer role:
skills/review/parallel-code-review/SKILL.md(Reviewer 1) - Auto-run hook:
hooks/security-review-hook.py - Design contract:
adr/local-security-review.md