The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses Node.js 'child_process' and Python 'subprocess' to execute Git commands and launch worker processes. This is core to its functionality as a development orchestrator. Safeguards include a dedicated 'sanitize_diagnostic' function in 'agf.js' that redacts credentials from output and explicit instructions to avoid shell interpolation in 'SKILL.md'.
[PERSISTENCE]: The skill installs Git hooks ('pre-commit') and host-specific hooks ('.claude/settings.json', '.codex/hooks.json') to automate turn monitoring. These installations are documented, idempotent, and include uninstallation commands ('agf uninstall'). Documentation also provides opt-in examples for system task schedulers (crontab).
[DYNAMIC_EXECUTION]: The 'looper.js' and 'external-runner.js' components launch sub-processes to execute developer-written plans. The skill includes a 'nested worker' detection system using process tree inspection ('process-tree.js') to contain and report unauthorized recursive agent calls, mitigating risk from unverified sub-agents.
[SAFE]: Hardcoded identifiers in test files (e.g., UUIDs and hex strings in 'scripts/notebook-owner.test.js') are used as mock session tokens for testing environments and do not represent real leaked secrets.

agentflow