Lean Sec

Defensive review with a hard context budget. Scripts summarize; you read only hotspots; you patch only when asked.

Let SKILL_ROOT be the folder that contains this SKILL.md. Scan target defaults to the current workspace. Prefer python3; on Windows use py -3 if needed.

python3 "$SKILL_ROOT/scripts/lean_sec.py" scope .
python3 "$SKILL_ROOT/scripts/lean_sec.py" secrets .
python3 "$SKILL_ROOT/scripts/lean_sec.py" deps .

Do not copy these scripts into the target repo. Do not cat lockfiles, SARIF, coverage, or tool logs into the conversation; pipe bulky output through trim.

Mode

Pick one from the user request. Default is lean.

Never "go deep" because the repo is large. Large repos make lean more important.

Workflow

1. Cheap pass (required). Run scope, secrets, and deps. Read only their stdout. If a command fails, say so and continue with rg over auth/crypto/sql/upload/ci paths; still do not dump matches in bulk.
2. Choose hotspots. Take the script's ranked paths plus the user's named files. Ignore node_modules, dist, build, target, .git, lockfile bodies, generated assets.
3. Load one reference that matches the hottest class:
   • Web/injection/upload/SSRF → references/web.md
   • Session/JWT/RBAC/IDOR → references/auth.md
   • Keys, .env, CI secrets → references/secrets.md
   • Dependencies, CI, containers, IaC → references/supply-chain.md
   • Need stricter budget rules → references/token-budget.md Load extra references only in standard/deep when a second class is clearly in play.
4. Confirm, don't tour. rg first. Read files with line offsets. Stop a file once the issue is confirmed or cleared.
5. Report in the template below. Then stop unless the user asked to patch.
6. Patch (only if asked). Smallest hardening change. Re-run secrets if you touched credentials handling. Do not add scanners, git hooks, or new deps unless asked.

Hard rules

• Defense only. Describe location, impact, and the fix. Do not write exploits, PoCs, payloads, bypasses, or reproduction attack steps. If asked for a fix and an exploit/PoC, ship the fix and refuse the exploit in one sentence.
• No secret values. Scripts already redact. If you see a live credential, write kind @ file:line and [REDACTED]. Tell the user to rotate; do not echo the token.
• No context stuffing. Do not paste full files, lockfiles, Dockerfiles, SARIF, npm ls, or CVE encyclopedias. Cap any quoted snippet at 12 lines.
• Authorized scope. Review code and config the user can already see in this workspace. Do not scan third-party systems, brute-force, or probe live hosts.
• Do not load this skill's README.md or LICENSE while reviewing another repo.

Report template

## Lean-sec (`<lean|standard|deep>`)

**Asset:** <app/lib/ci> | **Stack:** <one line>
**Budget:** scripts=<scope,secrets,deps> files_read=<n> refs=<names>

| Sev | Location | Issue | Fix |
|-----|----------|-------|-----|
| HIGH | file:line | one clause | one concrete patch |

**Also:** <optional 1-line residual risk>
**Not reviewed:** <areas skipped to keep the budget>

Severity: HIGH exploitable with default access or leaked secret; MED needs extra conditions; LOW hardening. Sort HIGH first. Skip stylistic nits. If there are no issues, say that in two sentences and list what was covered.

Threat model (only if asked)

Five bullets: assets, actors, trust boundaries, most likely abuse, cheapest control. No diagrams unless asked. Do not load extra references for this.