← ClaudeAtlas

kya-wraplisted

Wrap an agent tool through Shield KYA evaluate before any irreversible side effect. Use when adding tools, MCP calls, or writes that need ALLOW | DENY | REQUIRE_APPROVE.
The-Pixel-Boys/shield-kya · ★ 3 · AI & Automation · score 67
Install: claude install-skill The-Pixel-Boys/shield-kya
# Wrap a tool through Shield KYA This plugin is a **thin MCP gate**. It does not decide policy. Shield KYA evaluate on the control plane is the **sole PEP**. ## Verdicts | Verdict | Meaning | |---------|---------| | `ALLOW` | Side effect may proceed | | `DENY` | Do not execute. Never retry around the gate | | `REQUIRE_APPROVE` | Wait for a human `APPROVED`. Missing approval ⇒ no irreversible effect | Risk scores may only **raise** severity. They never auto-ALLOW. ## Dual plane - `host=ide`: authoring / dry-run - `host=runtime`: production work Same agent id, policy, approval, and trail on both. Label the host. Do not invent a second enforcement path in the editor. ## How to wrap 1. Register the agent principal (`register-agent` or console). Creating an agent is itself a tool (`kya.agent.register`). It is not a free write. 2. Give the tool a stable `toolId`. No marketplace adapter required. 3. Call evaluate (MCP `evaluate` / `eval-tool`) **before** the side effect. 4. If `REQUIRE_APPROVE`, request approval and wait. Do not execute on a pending ticket. 5. Offline `--offline` is a **sample fixture**. Do not treat it as production PEP. ## Do not - Bypass evaluate because the IDE "already reviewed" the change - Treat scanners, ORR, or this skill as a second PEP - Position Shield as a chargeback, dispute, or payment-recovery product