The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill defines a workflow for the agent to ingest and evaluate external data, including project source code and user-provided evaluation definitions, which could be exploited to provide malicious instructions to the agent.
Ingestion points: The framework relies on reading files from the .claude/evals/ directory and various project source files (e.g., src/auth.ts) using the Read and Grep tools.
Boundary markers: The instructions lack explicit delimiters or specific directives for the agent to separate its operational instructions from the content being evaluated (e.g., when using the 'Model-Based Grader').
Capability inventory: The skill utilizes the Bash, Write, Edit, Read, Grep, and Glob tools to perform its tasks.
Sanitization: There are no defined mechanisms for sanitizing the input data or validating the integrity of the evaluation definitions before they are processed by the agent.
[COMMAND_EXECUTION]: The skill explicitly instructs the agent to use the Bash tool to perform deterministic checks. This includes running build scripts (npm run build), test suites (npm test), and pattern matching (grep) to verify the success of the implementation tasks.

eval-harness