The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill references GitHub Actions workflows and repositories under the google-gemini organization for the purpose of auditing nightly test results and promoting evaluation suites. These references target a well-known service and organization.
[COMMAND_EXECUTION]: The documentation provides standard shell commands for building the project (npm run build), managing environment variables from local .env files, and executing test suites using vitest and the gh CLI.
[INDIRECT_PROMPT_INJECTION]: The behavioral evaluation framework involves seeding an agent's workspace with files (e.g., package.json, tsconfig.json) to simulate realistic environments. This creates an ingestion point for data that could theoretically contain instructions, which is the intended surface for testing the agent's decision-making logic.
Ingestion points: Test cases utilize a files object to populate the agent's local environment during evaluation.
Boundary markers: The provided test templates do not explicitly define specific delimiters or instructions for the agent to ignore embedded content within the seeded workspace files.
Capability inventory: The framework supports subprocess execution for Trajectory testing (evalTest) and auditing of tool logs to verify agent actions.
Sanitization: No specific sanitization or filtering of the simulated workspace content is described, as the environment is designed for internal logic verification.

behavioral-evals