The Agent Skills Directory

[COMMAND_EXECUTION]: The skill uses a suite of local Python scripts to orchestrate tasks. bootstrap.py manages agent profiles, audit_routing.py verifies model configurations, and run_codex_agent.py executes subtasks. These tools use the Python subprocess module to interact with the local git and codex binaries using secure argument passing to prevent shell injection.
[REMOTE_CODE_EXECUTION]: Subagents are executed using the codex CLI with restricted sandboxes. The implementation in run_codex_agent.py enforces a read-only or workspace-write sandbox, disables network access for writers, and performs all modifications in an isolated candidate worktree, ensuring changes are reviewed before being integrated into the main repository.
[SAFE]: A specialized safety hook (gpt_engineer_guard.py) is included to intercept and block destructive commands such as git reset --hard or forced pushes. The skill also incorporates strict routing checks to ensure that tasks are only performed by specific authorized model profiles, preventing accidental or malicious model substitution.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted repository data but mitigates risk by wrapping subagent execution with explicit instructions and boundary markers. All delegated results are validated against a strict JSON schema (handoff.schema.json) before being processed by the lead agent.

gpt-engineer