The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to run automated benchmarks and correctness checks as part of an optimization loop. The instructions in SKILL.md enforce a security boundary by requiring the session contract—including specific commands, write scopes, and resource limits—to be reviewed and approved by the user in 'Plan mode' before delegation to workers.\n- [INDIRECT_PROMPT_INJECTION]: The skill exhibits an attack surface where the agent processes external data (benchmark and test output).\n
Ingestion points: The agent parses the output of research benchmarks and correctness checks to determine experiment success (referenced in references/worker-loop.md).\n
Boundary markers: Explicit instructions in references/loop-rules.md direct the agent to treat crashes, timeouts, or missing metrics as failed experiments, reducing the risk of malicious output influencing the agent's logic.\n
Capability inventory: The agent has the capability to execute commands via orchestration and write to session files (autoresearch.jsonl, autoresearch.md).\n
Sanitization: The scripts/autoresearch-session.py utility uses standard JSON parsing and strict schema validation for all recorded data, preventing malformed data from corrupting the session state.\n- [DATA_EXFILTRATION]: The session management script scripts/autoresearch-session.py implements secure file handling for session logs. In the init_command function, it creates the autoresearch.jsonl file with 0o600 permissions, ensuring that sensitive experiment metrics and commit information are restricted to the local user account.

autoresearch