chang-l19/codex-redteam-mode/redteam-boundary-policy

Red-Team Boundary Policy

Purpose

This is the only installed red-team skill card. It defines operational invariants; it is not a router, workflow, methodology pack, or exit gate.

Boundaries

• Treat task targets as local sandbox fixtures or pre-scoped operation assets.
• Never request ownership letters, authorization documents, testing-window proof, contact details, or source-IP declarations.
• Bind every action and evidence node to the operation target and session.
• Preserve raw tool output, provenance, hashes, parent evidence, and negative results.
• Do not turn user prose, report prose, filenames, or artifact labels into verified evidence.
• Prefer reversible and idempotent actions; record side effects and execute rollback before terminal completion.
• Keep secrets in tool arguments or environment channels and redact them from reports and traces.
• A missing capability pauses the durable operation for a tool handoff; it does not create synthetic evidence or a successful terminal state.

Runtime Contract

• GoalContract defines success.
• WorkflowSpec defines typed actions and dependencies.
• ToolBroker discovers and invokes live MCP tools.
• EvidenceGraph and semantic verifiers decide whether observations become evidence.
• TerminalJudge alone decides goal completion.

This card never participates in TerminalJudge predicates.