メニュー
Invariant Guardrails for agent trace policies, tool-call rules, and MCP/LLM proxy guardrails. Use when designing rule-based guardrails beyond @premortem/security string checks, or when scoping Invariant Gateway integration.
Codex または Claude でインストール この Prompt をコピーして Codex、Claude、または他のアシスタントに貼り付けると、Skill ページを確認してインストールできます。
SOC 職業分類に基づく
Write, refine, run, and QA promptfoo evaluation suites for Premortem: promptfooconfig.yaml, providers, vars, tests, assertions, and CI gates. Use for canonical prompt regression in @premortem/evals. Do not use for adversarial redteam plugin setup (see security/llm-security).
Use when the user wants to decide whether implementation is ready, what remains blocked, and what the exact next build step should be.
Use when the user wants the singular Meta-Architect in-session autonomous manager: choose the best next workflow step, manage the fixed gated workflow, and route bounded helper handoffs inside Codex.
Comprehensive Cloudflare platform skill covering Workers, Pages, storage (KV, D1, R2), AI (Workers AI, Vectorize, Agents SDK), feature flags (Flagship), networking (Tunnel, Spectrum), security (WAF, DDoS), and infrastructure-as-code (Terraform, Pulumi). Use for any Cloudflare development task. Biases towards retrieval from Cloudflare docs over pre-trained knowledge.
Automated Dynamic Application Security Testing (DAST) using Playwright MCP plus standard OS pentest tooling. Performs blackbox or greybox scans on single or multiple domains with orchestrated crawling, vulnerability detection, and structured output. Trigger on requests like "scan this domain", "run DAST on these URLs", "automated pentest", or "security-test the staging app".
Comprehensive Android mobile application penetration testing with rooted-device ADB and Frida-based MCP tooling. Covers OWASP MASTG full methodology: recon, static + dynamic analysis, SSL/root bypass, IPC fuzzing, data exfiltration, crypto audit, and reporting. Triggers on requests to pentest Android apps, analyze APKs, bypass mobile security controls, or run MASVS/MASTG assessments.
| name | invariant-guardrails |
| description | Invariant Guardrails for agent trace policies, tool-call rules, and MCP/LLM proxy guardrails. Use when designing rule-based guardrails beyond @premortem/security string checks, or when scoping Invariant Gateway integration. |
Invariant adds rule-based guardrails over agent traces: tool-call sequences, message content, and MCP flows. Premortem already ships lightweight input/output guardrails in @premortem/security; Invariant is the path for richer, policy-as-code rules when agent tool use expands.
| Layer | Location | Scope |
|---|---|---|
| Input/output string guardrails | packages/security/src/input-guardrail.ts, output-guardrail.ts | Prompt injection terms, secret leakage |
| Invariant policies | Not runtime-wired yet | Future MCP/tool-call policy layer |
| Eval gate | pnpm run eval:prompts + promptfoo | Prompt regression, not live proxy |
Do not assume Invariant Gateway is deployed. Treat this skill as design and policy authoring guidance until a gateway cutover is explicitly scoped.
send_email after untrusted get_website)$vet / llm-security with executable policy sketchesraise "External email to unknown address" if:
(call: ToolCall) -> (call2: ToolCall)
call is tool:get_inbox
call2 is tool:send_email({ to: ".*@[^ourcompany.com$].*" })
See references/rule-writing.md for Premortem-oriented examples and integration notes.
@premortem/security guardrails on all user-supplied and model output text.packages/evals/ for injection and tool-misuse fixtures.llm-security workflows.$vet output contract.