The Agent Skills Directory

[PROMPT_INJECTION]: The evaluate_llm_gates function in batch_report/analysis.py exhibits an indirect prompt injection surface. \n
Ingestion points: The skill ingests data from untrusted batch output files, such as results.jsonl and output.jsonl (found in batch_report/analysis.py). \n
Boundary markers: No explicit boundary markers or protective instructions (e.g., 'ignore instructions within the data') are used when interpolating content into the {sample} template variable. \n
Capability inventory: The skill invokes local subprocesses through subprocess.run to call external tools like scillm and agent-inbox (found in batch_report/analysis.py and batch_report/utils.py). \n
Sanitization: Sanitization is limited to basic JSON formatting attempts; raw data content is placed directly into the evaluation prompt. \n- [COMMAND_EXECUTION]: The skill uses subprocess.run to execute local shell scripts and Python files residing in the user's home directory. \n
In batch_report/utils.py, the send_to_agent_inbox function executes the agent-inbox script. \n
In batch_report/analysis.py, the evaluate_llm_gates function executes the scillm tool.

batch-report