The Agent Skills Directory

[COMMAND_EXECUTION]: The skill performs environment discovery by executing shell commands to detect tools and project metadata.
Evidence: which codex, ls package.json, ls pyproject.toml.
[REMOTE_CODE_EXECUTION]: The skill invokes the 'codex' CLI to perform model-driven analysis. It explicitly directs the agent to use flags like --full-auto and --dangerously-bypass-approvals-and-sandbox. These flags are designed to remove security restrictions and user confirmation steps, allowing a remote-model-controlled agent to execute arbitrary code on the local machine.
[DYNAMIC_EXECUTION]: The skill dynamically generates and executes Bash commands in the background to handle parallel processing, which can be manipulated by varying the $ARGUMENTS provided to the skill.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external project source code across multiple files without defined boundary markers or sanitization, creating a vulnerability surface where malicious code in the analyzed repository could inject instructions into the sub-agents.
Ingestion points: File system exploration via ls and agent-led code reading.
Boundary markers: Absent; there are no instructions to the sub-agents to ignore embedded directives in the code they analyze.
Capability inventory: The skill has background shell execution, filesystem read access, and the ability to invoke external autonomous CLI tools.
Sanitization: Absent; the analyzed content is passed directly to sub-agents for reasoning.

product-analysis