The Agent Skills Directory

[COMMAND_EXECUTION]: Multiple scripts, including setup_experiment.py, run_experiment.py, and built-in evaluators (benchmark_size.py, benchmark_speed.py, build_speed.py, memory_usage.py, test_pass_rate.py), utilize subprocess.run(shell=True) to execute evaluation commands stored in config.cfg. This provides a mechanism for arbitrary shell execution based on experiment settings.
[INDIRECT_PROMPT_INJECTION]: The LLM judge evaluators (llm_judge_content.py, llm_judge_copy.py, llm_judge_prompt.py) are vulnerable to indirect prompt injection. 1. Ingestion points: These scripts read the content of the file being optimized (TARGET_FILE). 2. Boundary markers: Absent; file content is directly interpolated into judge prompts. 3. Capability inventory: The skill can execute shell commands, modify the file system via git, and schedule tasks. 4. Sanitization: Absent; content is processed without filtering.
[PERSISTENCE]: The /ar:loop command schedules recurring background tasks using CronCreate, allowing the optimization loop to persist and execute logic without active user oversight.
[DYNAMIC_EXECUTION]: The skill manages its environment by copying evaluator scripts into a project-local directory and generating dynamic strategy documents (program.md) that influence the agent's behavior in subsequent loop iterations.
[EXTERNAL_DOWNLOADS]: The installation instructions direct users to download the skill's source code from a third-party GitHub repository (alirezarezvani/claude-skills).

autoresearch-agent