The Agent Skills Directory

[SAFE]: The skill implements strict refusal rules for high-risk operations. It explicitly forbids the execution of sudo, destructive commands like rm -rf, ad-hoc package installations via pip/conda/apt, and bypassing CI checks with --no-verify. These rules ensure the agent remains within a safe, non-privileged workspace.
[EXTERNAL_DOWNLOADS]: All referenced repositories and domains, such as github.com/NVIDIA/cuopt.git and docs.rapids.ai, are official vendor-controlled or trusted sources. The skill correctly identifies that external scripts (e.g., curl|bash) are a security risk and instructs the agent to refuse them in favor of canonical setup procedures.
[COMMAND_EXECUTION]: The skill utilizes standard development tools including ./build.sh, ctest, pytest, pre-commit, and git. These are used in a manner consistent with its stated purpose as a developer assistant, with specific guidance on safe usage and mandatory user confirmation for write operations like git commit or git push.
[SAFE]: The static detection of code execution patterns in evals/evals.json is a false positive in the context of a security threat. The detected eval(user_input) string is part of a test case (ID dev-038) designed to verify that the agent correctly identifies and refuses to implement insecure code patterns.

cuopt-developer