agentleFS
Alexander-TyagunovAlexander-Tyagunov/magician/skills/sentinel/SKILL.md

sentinel

Skill14 starsChanged 49 days agoMITView on GitHub

Read-only security scan; OWASP Top 10, secrets, injection, dependencies, git history, auth. Use to audit a codebase for vulnerabilities.

  • Reads credentials

14 other projects use a version of this.

Adopted by

Most recently taken up by

The file

We split this skill into the rules it gives an agent

5 rules. 0 are common, 2 are distinctive, and 3 are very specific to this project.

  • 0

    Common

    Most of the 12 edited versions we read keep these.

  • 2

    Distinctive

    This author's own choices. Worth a look.

  • 3

    Very specific

    Tied to their project's names, files or tools. Probably skip.

  • Common
  • Distinctive
  • Very specific
  1. Very specific:

    Print the active model at start and end: If a fallback notice appears mid-scan, say so in the report and offer /model to return.

    line 25
  2. Distinctive:

    Run for detected stack:

    line 57
  3. Distinctive:

    Check for secrets in git history:

    line 67
  4. Very specific:

    Update the ledger before the report, and put the delta (new, fixed, still open) first (references/cadence-ledger.md, steps 2–3).

    line 101
  5. Very specific:

    Remove the marker first (rm -f .workspace/local/security-scan.active), so model switches are no longer held.

    line 118

“Common” is measured: kept by at least 6 in 10 of the 12 edited versions of this file we could read. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.

Keep what fits and drop the rest. Your version, ready for your agents.
Files
SKILL.md

/sentinel — Security Scan

Bundled command: if magician-scan is not found, run it as ${CLAUDE_PLUGIN_ROOT}/tools/magician-scan. If that is missing too (Claude chat ships no plugin tools), skip the static-analysis step, say so in the report, and run the rest.

For very large repos, raise /effort so the analysis stays thorough across the codebase (xhigh; max only as an opt-in for a frontier problem). See lore/models.md.

When to use

To audit a codebase for vulnerabilities: OWASP Top 10, credential and secret detection, injection surfaces, a dependency audit, a git-history secret scan and an auth spot-check. It never modifies code and ends with a severity-ranked report. Trigger phrases: "run a security scan", "audit for vulnerabilities", "check for leaked secrets". Not for fixing findings (/scrutinize) or a general code review (/divine).

The scan can change models under you

Cyber classifiers can move a flagged request to a fallback model, and the session continues there (lore/models.md).

  • Frame the work defensively. Audit and report. Never ask for working exploits, attack tooling or evasion techniques: they trip the classifier and are out of scope.
  • Print the active model at start and end. If a fallback notice appears mid-scan, say so in the report and offer /model to return.
  • A flag on the first request usually comes from workspace context (CLAUDE.md, skills, directory names). claude --safe-mode isolates that. switchModelsOnFlag: false asks before a switch.

Destructive-command hard gate (always on, not part of a scan)

Independent of any scan, magician ships a PreToolUse(Bash|PowerShell) hook, scripts/destructive-guard.sh (self-contained plain bash — no other interpreter), that unconditionally blocks catastrophic commands: recursive deletes of system or home roots, raw writes to disk devices, fork bombs, piping a network download straight into a shell or interpreter, decoding base64 straight into a shell, evaluating a downloaded command substitution, and a Bash command that could expose the secrets saved in magician's plugin options (one that names any of those options or the session environment file, reads another process's environment through the proc filesystem, or prints the whole environment with env, set or one of the other listing commands the guard checks). It exits 2, so the block lands before permission rules are evaluated — it overrides allow rules and fires in every mode (default/acceptEdits/auto/bypass), with no escape hatch. Common command wrappers (sudo, timeout, nohup, command, ...), ( )/{ } grouping and inline sh -c payloads (other shell options before the -c included) are unwrapped before the check. If a Bash or PowerShell call is refused with a magician destructive-guard: refused message, do not retry, rephrase, or obfuscate — the human must run it themselves outside the agent. Honest limit (CWE-78): a denylist can't catch every obfuscation, so this is a deterministic net layered under OS sandboxing + auto-mode's classifier + model judgment, not a complete sandbox.

Cadence and ledger

Findings persist in .workspace/local/security/findings.jsonl, never committed: open rows map unfixed vulnerabilities. A dismiss reason keeps a closed item closed. The first run is the baseline. Scan weekly: Claude Security scheduled scans where available, else /loop 1w /sentinel. Fixes are bounded patches in a reviewed PR. A fixed finding class becomes a scan rule or an eval. Format and steps: references/cadence-ledger.md. The ledger and the scan marker are the only files this skill writes.

Process

0. Mark the scan active

mkdir -p .workspace/local/security && touch .workspace/local/security-scan.active

While this marker exists, magician's PreModelSwitch hook (scripts/pre-model-switch.sh) holds any model switch for confirmation, because a safety-classifier fallback would otherwise change the model tier in the middle of the scan. Remove it at the end (rm -f .workspace/local/security-scan.active), and also when the scan is abandoned or fails; .workspace/local/ is never committed.

Then read the ledger (references/cadence-ledger.md, step 1).

1. Static Analysis (via magician-scan)

magician-scan .

magician-scan is plugin-provided (on PATH when the plugin is enabled) and makes no network calls. If magician-scan can't be run, note that in the report, skip this step, and continue with the remaining checks.

Reports: hardcoded credentials, private keys, eval() calls, SQL injection via % formatting, innerHTML XSS, dangerouslySetInnerHTML, os.system calls, shell=True subprocess.

2. Dependency Audit

Run for detected stack:

  • Node.js: npm audit
  • Python: pip-audit (if installed) or safety check
  • Go: govulncheck ./... (if installed)
  • Rust: cargo audit
  • Java: OWASP dependency-check (if configured)

The audit tools reach the network, and step 2.5 checks the install-time supply chain. What each tool sends, and the supply-chain checks, are in references/dependencies.md. Skip any tool the user doesn't want run.

3. Secret Detection

Check for secrets in git history:

git log --all -i -G'(password|secret|token|api[_-]?key)[[:space:]]*[:=]' --pretty=format:'%h %ad %an' --date=short --name-only -- '*.env' '*.key' '*.pem' 2>/dev/null | head -40

This lists commits and files only (it never prints values). Open a specific commit to confirm, but never paste secret values into the report; tell the user to rotate anything found.

4. Auth/Authz Spot Check

For web archetypes: identify all API endpoints and verify auth middleware is applied.

5. Input Validation Check

Scan for user input without sanitization.

Report Format

=== SENTINEL SECURITY REPORT ===
Date: <timestamp>
Target: <path>

CRITICAL: N
HIGH:     N
MEDIUM:   N
LOW:      N

[CRITICAL] src/auth.ts:45 — Hardcoded API key
[HIGH] src/db.ts:12 — SQL query built with string concatenation
...

DEPENDENCY AUDIT: N vulnerabilities found

OVERALL POSTURE: Clean | Needs attention | Requires immediate action

Update the ledger before the report, and put the delta (new, fixed, still open) first (references/cadence-ledger.md, steps 2–3).

6. CI Integration Note

For CI pipeline use: magician-scan exits 0 (clean), 1 (issues) or 2 (unknown option or missing path); magician-scan --help shows the usage.

# .github/workflows/security.yml
- name: Security scan
  run: magician-scan .

Obstacles

If this skill runs as a dispatched unit and hits something that blocks or degrades the work, do not wait for a human who is not there and do not silently ship a degraded result — return the Obstacles block from lore/obstacles.md (fields: STATUS, OBSTACLE, BLOCKER, SEVERITY, WORKAROUND, RECURRENCE, SCOPE, NEXT) to the caller alongside whatever you completed; interactively, say the same in prose.

Completion Signal

Remove the marker first (rm -f .workspace/local/security-scan.active), so model switches are no longer held; then:

"Sentinel complete. <N total findings>. Review report above and run /scrutinize for systematic remediation."

Download now (ZIP)

Does it work?

Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.

No reviews yet

People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.

  • It worked
  • Partly
  • It didn't
What did you use it for? e.g. Reviewing pull requests in a Rails app

Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.

Your trail

Files you open will collect here, so you can find them again.

Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.

Now do this for your own.

Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.