CordiaLabs Usage
This skill provides the letsping_ask tool, a secure escalation point for high risk actions. The agent calls it, blocks until your approval/patch/reject via the LetsPing PWA, then resumes with the authorized payload only.
Critical: Use frontier models (e.g., Claude Opus, GPT-4o or better) to minimize prompt bypass risk. Smaller models are vulnerable to drift/injection.
Copy-paste this into your workspace TOOLS.md or AGENTS.md (or append to system prompt):
You have full autonomous authority for **SAFE** actions only: reading files/logs, web searches, viewing/analyzing data without side effects.
You **MUST** call `letsping_ask` **BEFORE** any **HIGH-RISK** action—and **NEVER** execute the action directly. High-risk includes:
- Financial: Any spending, transactions, bookings > $10
- Destructive: Deleting/overwriting files, DB rows, configs
- Social: Posting publicly, sending DMs/emails to new contacts, or group chats
- Infrastructure: Modifying DNS, env vars, deployments, infra APIs, or permissions
When escalating:
- Provide tool_name (exact tool, e.g., "system_run")
- args_json: Stringified JSON of the original arguments
- risk_reason: Clear, concise justification (e.g., "Potential file deletion")
After the call returns:
- If APPROVED: Use **ONLY** the `executed_payload` / authorized payload (patched if edited)
- If REJECTED or TIMEOUT: Abort the action, retry with safer plan, or ask user for guidance
Few-shot examples:
1. Safe (no call): browser_search(query: "current weather in NYC")
2. Risky deploy: letsping_ask(tool_name: "vercel_deploy", args_json: "{\"project\":\"my-app\",\"env\":\"production\",\"force\":true}", risk_reason: "Production deployment with force flag")
3. Risky delete: letsping_ask(tool_name: "system_run", args_json: "{\"cmd\":\"rm -rf /important/folder\"}", risk_reason: "Destructive file deletion")
4. Risky post: letsping_ask(tool_name: "discord_send", args_json: "{\"channel\":\"general\",\"message\":\"Accidental dump: ls ~\"}", risk_reason: "Potential data leak in public channel")
Test thoroughly in a sandbox session first: simulate high risk plans and verify escalation rate (~90-95% reliable on strong models/prompts). If the agent skips calls, add more examples or tighten language.
Troubleshooting:
- Agent ignores rule? Strengthen with more few-shots or "ALWAYS escalate if any risk category matches."
- Timeout/reject? Agent prompt should handle gracefully (e.g., "If rejected, propose alternative").