Agent Command Block Example

This example demonstrates a core coding-ethos behavior: unsafe agent tool use is blocked before the shell command runs.

Example Scenario

An agent attempts a raw Git command that would bypass the documented wrapper or tamper with protected hook behavior. The PreToolUse hook evaluates the command and returns a blocking policy decision.

Expected agent response:

1. Treat the block as policy feedback.
2. Stop the unsafe command path.
3. Use the documented coding-ethos Git wrapper.
4. Ask for admin approval only when the protected workflow requires it.
5. Report the policy ID and next safe command path.

Why This Matters

AI agents often try to recover from blocked commands by changing the shell shape: alternate shells, direct binary paths, subprocesses, PATH edits, aliases, or wrapper bypasses. coding-ethos treats those as the same class of risk. The policy follows intent, not just one command spelling.

MCP Follow-Up

Agents can use MCP to understand the block:

• policy_check_command: evaluate a proposed command before running it.
• policy_explain: explain the blocking policy ID.
• skill_lookup: load the safe Git workflow remediation skill.

The MCP response is advisory context. The actual enforcement remains in the agent hook and Git hook paths.

Related Docs

• MCP server
• Threat model
• Comparison