The Agent Skills Directory

[DATA_EXFILTRATION]: The skill accesses sensitive local file paths containing agent interaction history and session logs.
Evidence: references/extraction-patterns.md provides SQL queries for ~/.claude/projects/*/*.jsonl and /tmp/claude-*/tasks/*.output.
Context: While no network exfiltration tools are used, the exposure of these sensitive logs to the agent's context is a significant data access surface.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted historical data (agent transcripts) which could contain adversarial instructions intended to influence the current session.
Ingestion points: The skill is specifically designed to ingest JSONL session files and agent output transcripts as defined in references/extraction-patterns.md.
Boundary markers: Instructions in SKILL.md attempt to limit the agent to "raw facts only," but there are no technical boundary markers or delimiters provided for the data being parsed.
Capability inventory: The skill utilizes mcp__plugin_agentskill-kaizen_kaizen-duckdb__execute_query, Read, Grep, and Glob tools to retrieve information.
Sanitization: There is no evidence of content sanitization or instruction filtering for the data ingested from transcripts or reports.

meta-inspector