The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted external data which may contain adversarial instructions targeting the agent.
Ingestion points: The skill ingests data using gh pr view --json title,body,comments and git log -p as specified in Step 1 of SKILL.md.
Boundary markers: No explicit boundary markers or safety instructions are provided to the agent to ignore instructions embedded within the PR description or comments.
Capability inventory: The skill has extensive shell capabilities, including running git, gh, and other AI model CLIs, and the ability to write and execute arbitrary test scripts.
Sanitization: There is no mention of sanitizing or escaping the PR data before it is processed or passed to other models in the "three-model panel".
[DYNAMIC_EXECUTION]: The skill's primary verification method involves runtime code generation and execution.
Evidence: Step 5 in SKILL.md explicitly instructs the agent to "Write a script or test that runs the real code, run it, and paste what happened" to provide proof of safety. While this is the intended purpose of the skill, it involves the execution of dynamically created code based on external context.
[COMMAND_EXECUTION]: The skill utilizes several CLI tools to gather information and communicate with external services.
Evidence: The skill body contains instructions for using git, gh, claude, and codex (via codex exec) to perform diff analysis and reach a panel of different AI models.

blast-radius