The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell commands and Python scripts to manage project state and evaluate code experiments.
Evidence: The instructions use git checkout, git add, and git commit to manage branches and commits, and execute setup_experiment.py and run_experiment.py located in the skill's local scripts directory.
[INDIRECT_PROMPT_INJECTION]: The skill reads history and configuration files from the project directory to decide on future code changes, creating a surface for untrusted data to influence the agent.
Ingestion points: The skill reads results.tsv, program.md, and config.cfg from the .autoresearch directory (SKILL.md).
Boundary markers: Absent. The agent is instructed to review the raw content of these files to determine its next experiment iteration.
Capability inventory: The agent can edit target files, commit changes to the repository, and execute evaluation scripts.
Sanitization: Absent. The agent relies on its internal reasoning to interpret the potentially untrusted file content.

ar-run