The Agent Skills Directory

[PERSISTENCE]: The skill uses the 'CronCreate' tool to schedule recurring autonomous execution of instructions at user-selected intervals. This establishes a persistent mechanism that continues across sessions until it expires after three days or is manually stopped.\n- [COMMAND_EXECUTION]: The scheduled task performs several shell operations, including 'git checkout', 'git add', and 'git commit' to manage experiment states. It also executes a local Python script ('run_experiment.py') to evaluate results.\n- [DYNAMIC_EXECUTION]: The skill involves dynamic modification and evaluation of files. The agent is instructed to edit a 'target' file and then run a Python command ('python {skill_path}/scripts/run_experiment.py') to process the changes, effectively executing code logic it has just modified.\n- [INDIRECT_PROMPT_INJECTION]: The skill processes instructions from external files which constitutes an attack surface.\n
Ingestion points: The agent reads '.autoresearch/.../config.cfg' and 'program.md' for configuration details and strategy instructions.\n
Boundary markers: No explicit delimiters or instructions are provided to the agent to prevent it from following malicious instructions if they were embedded in these files.\n
Capability inventory: The agent has filesystem write access, git execution rights, and the ability to run Python scripts.\n
Sanitization: No evidence of validation or sanitization exists for the content read from the configuration and strategy files before it is used to influence agent actions.

loop