agent-enforcement

# Agent Enforcement Skill ## Overview Two-layer defense against orchestration violations in `/team`: 1. **PreToolUse hook** (`hooks/enforce-team-rules.sh`) — runtime violation blocker 2. **Model upgrade** (`team.md: model: opus`) — better instruction following ## How /team Invokes Models | Model Type | Method | Tool | Reliability | |------------|--------|------|-------------| | Internal (Claude) | Agent({resolved_agent}) | Agent tool | High (same process) | | External (Grok, Gemini, etc.) | claudish MCP tools (team/create_session) | MCP | 100% (deterministic) | External models are called via claudish MCP tools — no Bash invocation needed. ## Hook Rules (enforce-team-rules.sh) The PreToolUse hook intercepts Task and Bash tool calls at runtime: | Rule | Condition | Action | |------|-----------|--------| | 1 | /team Task with wrong agent | DENY (must be in agent whitelist) | | 2 | /tmp/ in Task prompt | DENY | ### Layer 1: PreToolUse Hook Detects /team workflows by vote template pattern in Task prompts. Blocks: - **Wrong agent for /team Tasks:** Only agents in the whitelist are allowed: `dev:researcher`, `dev:debugger`, `dev:developer`, `dev:architect`, `dev:qa-engineer`, `dev:devops`, `dev:frontend-developer` - **Insecure paths:** No `/tmp/` paths in Task prompts (use `ai-docs/sessions/`) ### Layer 2: model: opus The `/team` command uses `model: opus` (Opus 4.6) which follows complex XML instructions much more reliably than Sonnet (~90% vs ~33% compliance). ## Agent Selection for Task Delegation | Task Type | Primary Agent | Alternatives | |-----------|--------------|--------------| | Investigation | dev:researcher | dev:debugger | | Review | dev:reviewer | — | | Architecture | dev:architect | — | | Implementation | dev:developer | — | | Testing | dev:qa-engineer | — | | DevOps | dev:devops | — | | UI/Design | dev:frontend-developer | — | ## Agent Resolution The `/team` command resolves the agent dynamically from the context detection table in `team.md` based on task keywords (e.g., "debug" → `dev:debugger`, "review" → `dev:researcher`). The hook enforces a whitelist of all valid agents. Methods: - **Internal models** — Agent(agent) via Agent tool - **External models** — `team` MCP tool (handles parallel execution internally) ## Validation Run the test suite to verify enforcement: ```bash cd autotest/team && bash run-tests.sh ``` ## Troubleshooting **Hook blocking legitimate calls:** The hook triggers on /team vote template patterns in Task prompts and claudish in Bash commands. Normal Task usage (without vote templates) is never affected. **resolve-agents.sh not found:** The `/team` command uses its own built-in logic. The hook still provides runtime protection. **Hook behavior:** The hook skips existence checks (`which claudish`, `command -v claudish`).