The Agent Skills Directory

[Instructional Documentation]: The skill serves as a guide for developers to investigate Pyright type-inference regressions. It does not contain executable code, scripts, or automated commands that could pose a security risk.
[Untrusted Data Awareness]: The skill explicitly warns the agent and the user to treat captured metadata, such as labels, paths, and URLs, as untrusted data. It provides clear instructions not to execute captured text or follow captured URLs, which serves as a mitigation against indirect prompt injection or social engineering.
[Local Storage Practices]: The workflow recommends storing generated artifacts in a local, private directory (.generated/evaluation-traces/), emphasizing that these should not be published or overwritten, aligning with safe data handling practices.
[Execution Boundary Restrictions]: The documentation describes built-in restrictions, such as the rejection of dynamic imports in local modules, which limits the attack surface for dynamic code execution during the tracing process.

evaluation-tracing