The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes Bash and python3 to perform statistical calculations (such as Kolmogorov-Smirnov tests) that are not natively supported by the agent's core environment. This involves piping numeric data histograms into a Python interpreter via shell commands.
[DYNAMIC_EXECUTION]: The skill dynamically fetches and executes a Python script (scripts/ks2.py) using the llma-skill-file-get tool. The script is used for distribution-shift detection and is written to a temporary location (/tmp) before execution. The skill also provides inline Python snippets for statistical analysis.
[PERSISTENCE]: The skill maintains a durable "watchlist" and seasonality-matched "baselines" using the scout-scratchpad-remember and scout-scratchpad-search tools. This allows the agent to maintain state and round-robin checks across multiple scheduled runs, which is a standard operational pattern for this type of scout.
[INDIRECT_PROMPT_INJECTION]: The skill processes external data that could theoretically contain malicious instructions.
Ingestion points: Raw analytics series and event properties are ingested via insight-query and execute-sql. Historical context is retrieved via scout-scratchpad-search.
Boundary markers: The instructions do not define specific delimiters for the ingested analytics data, though the logic heavily prioritizes numeric interpretation for scoring.
Capability inventory: The skill can author inbox reports (scout-emit-report), create persistent notebooks (notebooks-create), and execute local scripts via Bash.
Sanitization: No explicit sanitization or filtering of non-numeric event property strings is documented before they are summarized into report prose.

signals-scout-anomaly-detection