The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes several entry points using the uv run command to manage local services and APIs, such as uv run fleet-rlm serve-api and uv run fleet-rlm chat, which execute local code for runtime management.\n- [REMOTE_CODE_EXECUTION]: The skill uses the DaytonaInterpreter class to execute Python code strings within a sandboxed environment via the execute() method. This is a primary feature for document analysis and automated reasoning tasks.\n- [PROMPT_INJECTION]: The skill includes functionality for processing external data (e.g., document chunks in rlm-subcall and documents in SummarizeLongDocument), which introduces an indirect prompt injection surface.\n
Ingestion points: Data enters the context through document inputs in the rlm object and file paths/queries in the rlm-subcall skill.\n
Boundary markers: The skill documentation does not mention the use of boundary markers or instructions to isolate the LLM from potential commands embedded in the processed data.\n
Capability inventory: The skill has access to Python execution via interp.execute() and local shell execution via uv run.\n
Sanitization: There is no evidence of data sanitization or validation shown before content is processed by the agent signatures.

rlm