The Agent Skills Directory

[COMMAND_EXECUTION]: The skill generates Python introspection scripts in the local scratch/ directory and executes them using the pixi package manager (Shape 1 and Shape 2 lookups). This allows the agent to extract exact signatures and help text from currently installed library versions.
[EXTERNAL_DOWNLOADS]: Through 'Shape 3' lookups, the skill performs web searches and fetches narrative documentation from external URLs, which are then cached as markdown files in the workspace. While it targets documentation, this involves network operations to non-whitelisted domains.
[PROMPT_INJECTION]: The skill contains explicit instructions to override the agent's default behavior, such as 'No symbols from memory' and 'Recognition is not a lookup,' to prevent hallucinations based on training data. Additionally, the ingestion and processing of external documentation for caching creates a potential surface for indirect prompt injection if the source content is untrusted.

python-api