The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill uses joblib.load() to dynamically load model artifacts from the filesystem for inference. This method uses pickle internally, which is inherently unsafe as it can execute arbitrary code during deserialization if the artifact files on disk are compromised or modified by an unauthorized user.
[INDIRECT_PROMPT_INJECTION]: The skill creates a self-improving service loop where untrusted external data influences the behavior of the system through retraining.
Ingestion points: The handle_ingest endpoint in SKILL.md accepts arbitrary JSON features and string label data from external API requests.
Boundary markers: None. The data is directly stored in SQLite and subsequently loaded for training without delimiters or instructions to ignore embedded malicious content.
Capability inventory: The skill possesses capabilities for writing files to the disk (joblib.dump), dynamically loading executable artifacts (joblib.load), and performing database operations via the honker library.
Sanitization: No sanitization, validation, or filtering is applied to the ingested features or labels before they are passed to the pipeline.fit() training routine.
[COMMAND_EXECUTION]: The Makefile and run.sh scripts utilize shell commands to manage environment variables, start background worker processes, and execute the service via uv run and litestream.

data-flywheel