The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes evaluation logic using subprocess.run with shell=True in scripts/run_experiment.py (line 104) and scripts/setup_experiment.py (line 39). This allows the execution of complex shell commands, including pipes and redirects, based on the evaluate_cmd configuration which may be influenced by project files or user input.
[DYNAMIC_EXECUTION]: The agent operates an autonomous loop that dynamically executes commands and scripts defined in config.cfg and evaluate.py. These files reside within the project directory and are triggered automatically during the research cycle, potentially executing malicious logic if the local environment is compromised.
[INDIRECT_PROMPT_INJECTION]: The agent processes external project data which can influence its optimization strategy and execution.
Ingestion points: The agent reads optimization goals and strategy from program.md, historical metrics from results.tsv, and the contents of the optimization target file specified in SKILL.md.
Boundary markers: No explicit delimiters or boundary instructions are provided to the agent to differentiate between its core operating protocols and the data found in optimized files.
Capability inventory: The skill possesses the capability to modify the file system, perform git operations (commit/reset), and execute arbitrary shell commands via the experiment runner.
Sanitization: Content read from project files is not sanitized or validated before being incorporated into the agent's decision-making process for the next optimization iteration.

autoresearch-agent