| name | immunogen |
| description | Scan an AI-built ("vibe-coded") repo for launch-blocking issues — leaked secrets, exposed API keys, broken auth, missing row-level security, unsafe payment wiring, vulnerable dependencies, and risky AI-layer surface (skills/MCP servers/agent prompts). Returns a 0-100 Ship Score with a banded verdict and ranked findings. Trigger this skill when the user asks to "scan this repo for security issues before shipping", "check if this app is safe to launch", "audit this AI-generated codebase", or similar pre-launch security review requests. |
immunogen
immunogen is a static-analysis scanner purpose-built for repos built (or partly
built) by coding agents — Cursor, Claude Code, Lovable, Bolt, v0, Replit, and
similar tools quietly ship the same handful of launch-blocking mistakes, and
immunogen finds them before a user finds them in production.
What this skill does
Running the wrapper script performs one deterministic scan of a target
directory and returns:
- A headline A-F health grade (
Grade: F — Critical, etc. — see the
grade legend in the immunogen repo's README) layered on top of the
numeric score below it.
- A Ship Score (0-100) with a band (
Ready to ship / Low risk /
Moderate risk / High risk) and a pass/blocked verdict.
- Ranked findings — each with severity, a stable rule id, the exact
file:line location, a one-line message, and a confidence tag.
The scan is static only — it reads and parses source files, it never
executes, imports, or evals anything from the scanned repo (this is a hard
correctness/safety guarantee of the underlying library, not just this skill).
How to invoke it
Run the wrapper script against the path the user wants scanned (defaults to
the current directory if no path is given), or against a live http(s)://
URL to screen a deployed site (see "Live URL support" below):
node .claude/skills/immunogen/scan.mjs <path-to-repo>
node .claude/skills/immunogen/scan.mjs https://myapp.com --authorize
The first non-flag argument is the target (a local path or an http(s)://
URL). --authorize and --allow-private are forwarded to the scan when
present.
Live URL support
This wrapper now forwards a live-URL target and the --authorize /
--allow-private flags, so it can screen a deployed website directly:
node .claude/skills/immunogen/scan.mjs https://myapp.com --authorize
See the immunogen repo's README "Screen a live URL" section for what it
checks (security headers, CORS misconfig, exposed .env/.git, secrets in
served bundles, cookie flags). This makes REAL bounded network requests
against the target host, so it requires the caller's explicit authorization
(--authorize) — only screen sites the user owns or is explicitly
authorized to test. A URL passed without --authorize is refused by the
consent gate before any request is made (clean error, exit code 2); do not
work around that — surface the message and confirm the user is authorized.
--allow-private additionally permits a loopback/private host (local
dev/test only). Live-URL screening is deterministic and never invokes an LLM.
The wrapper resolves the immunogen library automatically — from the built
repo (pnpm build) when it lives inside it, or from an installed immunogen
package. If it can find neither it prints one clear error asking you to build
the repo or install the package; if you see that, do that first — do not
reimplement the scan logic yourself.
The script prints markdown to stdout. Present that markdown to the user
verbatim — do not summarize, truncate, or rephrase the headline grade line,
the Ship Score header, or the findings list; the exact grade, score, band,
verdict, and ranked findings are the product, and altering them would
misrepresent the scan result.
Determinism and offline behavior
This scan is deterministic and fully offline by default — no API key is
required, and none is read or forwarded by this wrapper. The same directory
always produces the same Ship Score and finding count (parity with the
immunogen CLI and the scan() library function is proven by an automated
test — tests/dist/skill-parity.test.mjs in the immunogen repo). Do not pass
any LLM/provider flags or environment variables when invoking this skill; the
zero-setup, offline path is the one this skill is wired to.
What NOT to do
- Do not read the scanned repo's files yourself and hand-write findings —
always run the wrapper and use its output.
- Do not execute,
require(), or eval any file from the scanned repo for
any reason, even to "double check" a finding.
- Do not invent or adjust the Ship Score or the A-F grade — report exactly
what the wrapper returned.
