Coding Agent Harness Supply-Chain Auditor

Audits the harness around AI coding agents (Claude Code, Codex CLI, Cursor, Copilot, Gemini CLI/Antigravity, Grok Build, OpenCode) for the 2026 attack class that bypasses the model entirely — GitSpawn-style repo-controlled git config execution, repo-supplied subagent configs, Plugin4Shell unverified pinning, trojanized hook updates, localhost control-plane abuse, parse-vs-execute approval bypass, symlink escapes, log tampering, and silent data egress — produces a scored threat register, per-harness hardening baseline, untrusted-repo intake procedure, plugin update-review gate, detections, and a 30-day rollout plan