The Agent Skills Directory

[COMMAND_EXECUTION]: The skill executes local shell scripts within the repository (e.g., '.devcontainer/test-on-daytona.sh') and uses 'pnpm' for task execution and dependency management. These operations are constrained to the local project environment and are standard for development and testing workflows.
[DATA_EXPOSURE_AND_EXFILTRATION]: The skill provides instructions for managing credentials using Daytona secrets volumes and '.env' files. It explicitly cautions against printing keys to console logs, demonstrating an adherence to secure secret management practices.
[INDIRECT_PROMPT_INJECTION]: The skill ingests data from evaluation flows and writes the results to reports. While this creates a potential surface for indirect injection, the risk is negligible as it involves processing local test results.
Ingestion points: Evaluation results generated in 'evals/results/' and coded flows in 'evals/flows/'.
Boundary markers: None explicitly specified in the markdown instructions.
Capability inventory: Local command execution (bash, pnpm), Daytona CLI interactions, and file writes to the local results directory.
Sanitization: The runner generates machine-checkable assertions, which typically involve structured data validation rather than unrestricted natural language processing.

run-evals