The Agent Skills Directory

[DYNAMIC_EXECUTION]: The benchmark runner scripts/benchmark.mjs uses dynamic imports to load fixture files provided via command-line arguments. It also utilizes createRequire and dynamic import() calls to resolve and load the webgpu and three libraries from the local environment at runtime.
[COMMAND_EXECUTION]: The skill involves running a Node.js utility that performs direct file system operations, including creating directories, reading local source files, and writing benchmark results and binary artifacts to the disk.
[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to analyze external code from target projects to construct benchmark fixtures. There are no boundary markers or sanitization steps to prevent instructions embedded in this external code from influencing the agent's behavior during fixture creation.
Ingestion points: Target project source files, fixtures, and code retrieved via the threenix MCP server.
Boundary markers: None identified in the provided instructions or templates.
Capability inventory: File system access (read/write), dynamic script execution (import), and hardware GPU access.
Sanitization: No evidence of input validation or sanitization for the ingested code or generated fixtures.
[DATA_EXFILTRATION]: The runner script automatically snapshots the fixture file and any declared source dependencies into the output directory. This mechanism involves reading local files and writing their content into new files, which could be leveraged to expose sensitive local files if they are incorrectly included in the fixture's source list.

threenix-benchmark-webgpu