The Agent Skills Directory

[SAFE]: The skill uses local Python scripts for document processing and retrieval. These scripts do not perform network operations, use hardcoded credentials, or access sensitive system directories.
[COMMAND_EXECUTION]: The agent is instructed to execute local scripts (search_evidence.py) for search tasks. These scripts are self-contained and do not utilize unsafe dynamic execution functions like eval() or exec().
[DATA_EXFILTRATION]: No exfiltration vectors were found. The retrieval logic is strictly lexical and local, ensuring that document content remains within the user's workspace.
[PROMPT_INJECTION]: The skill processes untrusted user-provided document corpora. To mitigate potential indirect prompt injection, the skill enforces an "Evidence Contract" that requires the agent to classify claims (Direct, Synthesis, Inference) and map them to specific source locations (file, line, page).
Ingestion points: scripts/search_evidence.py reading .md, .markdown, and .txt files.
Boundary markers: The agent is instructed to use structured evidence objects and explicitly label assumptions in the final answer.
Capability inventory: Shell execution (local search script), file system write (via create_specialized_skill.py).
Sanitization: The retrieval script performs lexical matching; the agent is instructed to manually verify evidence against original sources.

z-grounded-source-qa