The Agent Skills Directory

[SAFE]: The skill is a legitimate performance engineering tool for AMD GPUs. It operates by executing local commands through the magpie CLI or Python module, which is standard for developer-oriented agents.
[EXTERNAL_DOWNLOADS]: The skill references repositories and documentation from trusted organizations, specifically github.com/AMD-AGI/Magpie. These references are informative and used to provide context for kernel source mapping and configuration templates.
[COMMAND_EXECUTION]: The skill documents how to execute benchmark and kernel analysis commands (e.g., magpie analyze, magpie benchmark). These commands are typical for performance profiling and involve local hardware interaction (GPU) and containerized workloads (Docker/Ray). The skill correctly instructs the agent to favor structured YAML configs and verify tool availability before execution.
[DATA_EXPOSURE_AND_EXFILTRATION]: No evidence of unauthorized data access or exfiltration was found. The skill explicitly warns against printing secret tokens (e.g., model-access tokens) in its reproducibility checklist. Network operations are scoped to legitimate benchmark framework needs (e.g., downloading model weights if configured by the user).
[PROMPT_INJECTION]: The instructions do not contain bypass markers, role-play injections, or attempts to override system safety filters. It maintains a technical, professional tone focused on GPU optimization.
[REMOTE_CODE_EXECUTION]: While the skill involves compiling and running GPU kernels (HIP, CUDA, Triton), this is the primary purpose of the tool. The instructions emphasize correctness gates (testcases) and local source files, avoiding piping unknown remote scripts into shells.

magpie-kernel-evaluator