The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external Torch Profiler JSON trace files. While the scripts treat these as structured data and do not execute content as code, processing untrusted external data presents a theoretical surface for indirect prompt injection.
Ingestion points: scripts/add_layer_track.py, scripts/compact_gpu_tracks.py, and scripts/serve_local_trace.py all ingest external trace files provided via the --trace argument.
Boundary markers: Absent. The data is parsed as JSON structures; no explicit prompt delimiters are used as the tool operates programmatically.
Capability inventory: The skill can read/write files (via json and gzip) and run a local loopback server.
Sanitization: Uses the standard Python json library for structure validation.
[DATA_EXPOSURE]: The serve_local_trace.py script starts a local HTTP server on 127.0.0.1 to serve trace data to the browser. The data is transmitted to the well-known Perfetto UI service (https://ui.perfetto.dev) via the postMessage API for visualization. This is a standard and safe integration pattern.
[SAFE]: The Python scripts use only the standard library and implement security best practices, such as using the 'x' mode in open() for exclusive creation to help prevent symlink-based file replacement attacks.

torch-profiler-layer-track