The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill's evaluation and debugging workflows (documented in env-evals.md and eval-loop.md) involve ingesting production traces and external datasets to run agents in isolated containers. This creates an attack surface where instructions embedded in the processed data could influence the agent's behavior during the evaluation loop.
Ingestion points: Dataset files (JSONL/CSV), task instructions, and environment seed data located in the evals/ directory.
Boundary markers: The documentation does not specify the use of delimiters or specific safety instructions to mitigate instructions embedded within the ingested datasets.
Capability inventory: The skill facilitates shell command execution, Docker container management via lmnr-cli eval run, and execution of code within containers via self.exec_as_agent (as seen in references/env-evals.md).
Sanitization: While the skill recommends scrubbing PII/secrets from production data, it does not provide explicit guidance for sanitizing or escaping the content to prevent prompt injection.
[COMMAND_EXECUTION]: The skill relies on various command-line tools for its operations, including lmnr-cli, docker, uv, and various language-specific package managers (npm, pip, poetry). These tools are used for setting up the environment, running local evaluations, and instrumenting codebases.

laminar