The Agent Skills Directory

[COMMAND_EXECUTION]: The skill relies on the Bash tool to execute multiple local Python scripts for agent orchestration, including scripts/run_agent.py, scripts/agent_orchestrator.py, and run_tests.py. While these appear to be internal project tools, shell execution carries risks if input parameters (such as the <INSTRUCTION> or <FEEDBACK> strings) are not properly sanitized against command injection.
[PROMPT_INJECTION]: The skill is susceptible to indirect prompt injection (Category 8). The Supervisor agent (Claude) is instructed to read, audit, and approve design specifications and implementation plans generated by an Executor sub-agent (Gemini). If the sub-agent is compromised via malicious data in its workspace, it could generate instructions designed to manipulate the Supervisor's subsequent actions.
Ingestion points: The Supervisor reads docs/superpowers/specs/*.md and implementation_plan.md created by the sub-agent.
Boundary markers: The skill defines three "Supervision & Review Gates" (Design Spec, Implementation Plan, QA/Verification) to inspect sub-agent output.
Capability inventory: The agent possesses Bash, Read, and Write capabilities, which are used to execute tests and merge code.
Sanitization: The instructions require the Supervisor to manually audit the content for placeholders and architectural alignment before approval.

co-pilot-loop