菜单
Run the generic footgun scanners (secrets + structural) over the repo and return a ranked findings list. Language-agnostic, no deps. Use at session-start (as part of harden), before a checkpoint, and on demand to answer "is anything unsafe or fragile in here?".
用 Codex 或 Claude 帮你安装 复制这段 Prompt，粘贴到 Codex、Claude 或其他助手里，让它检查 Skill 页面并帮你完成安装。
基于 SOC 职业分类
Stand up Mission Control — the live, two-way ops board for a project (SSE-live, phone-friendly, token-gated, localhost-only) with the encrypted Secrets channel built in. The coordinator writes state via mc-update.py and the operator sees it + checks items off + hands over secrets, from their phone. Use to start MC for a project, or when the operator asks "where do I see what's happening / how do I give you a key".
Stand up Mission Control — the live, two-way ops board for a project (SSE-live, phone-friendly, token-gated, localhost-only) with the encrypted Secrets channel built in. The coordinator writes state via mc-update.py and the operator sees it + checks items off + hands over secrets, from their phone. Use to start MC for a project, or when the operator asks "where do I see what's happening / how do I give you a key".
Apply this orchestration system to an EXISTING (brownfield) repo. Read the codebase, draft a CLAUDE.md + specialist subagents that match its real domains, add a tracker + memory — additively, without touching its source. Use when the human wants to run the coordinator pattern on a repo they already have.
Report a semantic beat to Mission Control — a decision reached, a milestone hit, or a completion-with-artifact (what shipped + its link). Use when you agree on an approach with the operator, hit a real milestone, get blocked/unblocked, or finish a workstream with something to show — the "why" and the artifact the automatic lifecycle hook can't see. Complements (does not replace) the spawned/working/done spine the plugin's lifecycle hook streams for free.
Take a greenfield idea to a built, tested, running first slice. Propose a tech stack (don't assume one), scaffold it FRESH with the real CLIs, wire one proof test, run it. Use when the human asks to spin up / start / bootstrap a new project or service.
Show a running result on the human's phone — or share a PUBLIC link a friend can open. Start the project's service(s), expose them over Tailscale (serve = your devices only; funnel = public, human-gated), and report ONE link. Two tiers: a single service (tailscale serve/funnel) or ≥2 services sharing an origin (the box-ingress Caddy). Use when the human wants to SEE the thing running, or SHARE it, not just read that tests passed.
| name | scan |
| description | Run the generic footgun scanners (secrets + structural) over the repo and return a ranked findings list. Language-agnostic, no deps. Use at session-start (as part of harden), before a checkpoint, and on demand to answer "is anything unsafe or fragile in here?". |
A thin runner over the two shipped scanners (scripts/scan-secrets.sh + scripts/scan-structure.sh).
They are generic and language-agnostic (git + grep + coreutils, no dependencies), so they work the
moment a repo exists — that's the point: the secret/structural footguns are the same shapes in every stack,
so they don't wait on bootstrap/adopt to wire a toolchain.
Run both scanners from the repo root:
bash scripts/scan-secrets.sh — leaked credentials. A finding is a hard stop (exit 1): a committed
.env, a private key, a cloud/vendor token, a hardcoded credential literal, or a server-only service-role
key referenced from client code. The output redacts the secret value (never echo a secret into a log).bash scripts/scan-structure.sh — structural footguns (advisory): oversized files, DELETE/UPDATE without
a WHERE, broad RLS/GRANT ALL, a React app with no ErrorBoundary, promise chains with no .catch,
engine-fragile CSS. This feeds harden; it does not block on its own.--staged (only staged files, used by the pre-commit hook), --strict (structural scan
exits non-zero on a HIGH finding — use when you want it as a gate, not advice).If the scanners aren't present (a brownfield repo before adopt installed the machinery), say so and
offer to run adopt / drop them in — don't silently report "clean".
Rank + report. The scanners already rank by severity (CRITICAL → HIGH → MEDIUM → LOW). Relay the ranked list concisely. Don't dump raw output if it's long — summarise counts by severity + the top items.
Hand off. Secret findings → fix now (remove, rotate, move to env-loaded config) before anything else.
Structural findings → hand the list to the harden skill, which decides real-vs-false-positive and
closes the real ones. A genuine false positive gets a pragma: allowlist secret line marker or a path in
.scanignore — with a one-line reason, never to silence a real finding.
These are heuristic scanners tuned for low noise and velocity, not a security product — they catch the
footguns we've actually seen bite a build, not every possible one. A clean scan means "none of the known
footguns," not "provably safe." For security/money/irreversible paths, pair it with the review skill (an
adversarial agent briefed to break it) — the scanner finds shapes, the reviewer finds reasoning bugs.