The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructions specify installing the agentic-jujutsu package from the NPM registry using npx agentic-jujutsu, which fetches and executes external code at runtime.\n- [COMMAND_EXECUTION]: The tool provides an execute() method used to run shell commands. While documentation shows standard usage like git push, the method provides an unrestricted interface for executing arbitrary system commands.\n- [REMOTE_CODE_EXECUTION]: The skill implements logic (e.g., in the SelfImprovingAgent class) that receives a list of recommendedOperations from an AI model and executes them sequentially. This pattern represents dynamic execution of instructions generated at runtime.\n- [PROMPT_INJECTION]: The skill possesses a surface for indirect prompt injection, where malicious input data can influence the commands suggested by the AI.\n
Ingestion points: User-supplied task descriptions passed to jj.getSuggestion() and jj.startTrajectory() in SKILL.md.\n
Boundary markers: None identified; inputs are interpolated directly into the tool's logic.\n
Capability inventory: Arbitrary shell execution via jj.execute() across all documented examples.\n
Sanitization: No evidence of sanitization, validation, or escaping of task descriptions before they are used to generate suggested operations.

agentic-jujutsu