Red-Team Suite

coding-ethos red-team tests exercise bypass attempts as first-class policy fixtures. They are not ordinary lint examples; they are regression tests for the guardrail mission.

The initial harness lives in go/internal/redteam. It defines reusable scenarios and runs them against the same compiled policy bundle used by hooks, lint, Git wrappers, CI, SARIF, and MCP.

Current Scenarios

• raw Git hook bypass with --no-verify;
• absolute Git binary invocation;
• nested shell Git execution;
• protected hook runtime writes;
• protected hook runtime path traversal;
• protected hook runtime symlink targets;
• hook runtime deletion attempts;
• managed toolchain or PATH evasion;
• generated config drift;
• Git-wrapper bypass checks;
• lint preflight bypass checks.

Each scenario records the bypass class, enforcement surface, status, policy IDs, and whether the expected policy failed to fire. A missed scenario is a test failure, not an advisory warning.

Expansion Path

The next scenarios should cover live Claude/Codex/Gemini prompt attempts in disposable repositories. Live agent runs should reuse the same scenario definitions and report their results through the same result model instead of creating a separate prompt-only test format.