The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external Kitaru investigation sessions and evaluation results, creating a surface for indirect prompt injection.
Ingestion points: The skill reads investigation sessions, verdicts, and evaluation outcomes via MCP operations (e.g., kitaru_review_read) and CLI commands (e.g., investigation get) as documented in references/kitaru-contracts.md.
Boundary markers: No explicit boundary markers or instructions to ignore embedded instructions are specified for the external data being processed.
Capability inventory: The skill executes CLI commands, writes JSON snapshots to the local file system, and runs a Python-based summary script (scripts/summarize_validation.py).
Sanitization: The scripts/summarize_validation.py script implements strict structural validation using custom helper functions (e.g., _object, _text) to verify the integrity of the JSON snapshot, though it does not perform semantic sanitization of the natural language text within the records.
[COMMAND_EXECUTION]: The skill utilizes the Kitaru CLI and MCP tools to interact with the Kitaru platform. These operations are within the scope of the skill's stated purpose and follow best practices by requiring explicit user authorization for remote writes or paid execution.
[SAFE]: The skill follows secure credential management practices, explicitly instructing the agent to keep keys and raw traces out of committed files and snapshots, instead relying on environment variables or established provider connections.

kitaru-validate-evaluator