The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes standard shell utilities such as wc, ls, and grep to interact with its own logging infrastructure and verify the existence of the route-events.jsonl file.
Evidence: Found in SKILL.md (Phase 1) and references/error-handling.md, where commands like wc -l "$LOG" and ls ~/.claude/hooks/ are used for diagnostics.
[DYNAMIC_EXECUTION]: The reference materials include Python scripts designed to be executed via python3 for parsing JSONL data to perform health checks and event counting on local logs.
Evidence: references/error-handling.md and references/preferred-patterns.md contain hardcoded Python scripts that are piped to python3 to validate JSON consistency and count event types.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection because it ingests untrusted data from the web for fact-checking and reads log files containing user-provided snippets.
Ingestion points: Untrusted data enters the agent context through WebFetch (external URLs), WebSearch (search results), and route-events.jsonl (containing request_snippet fields).
Boundary markers: The skill employs a multi-agent review pattern, using a separate dispatch for self-review (skills/shared-patterns/report-self-review.md) to synthesize and verify synthesized findings.
Capability inventory: The skill possesses Bash (local file diagnostics), Write (report generation), Agent (sub-task dispatch), and WebFetch/WebSearch (network access).
Sanitization: During the "EXTRACT" phase, the skill performs structured extraction of specific fields (claims, statistics, quotes) from ingested source material to isolate facts from potentially malicious prose instructions.

research