The Agent Skills Directory

[DYNAMIC_EXECUTION]: The skill provides instructions and code for monkey-patching library functions at runtime, specifically targeting the transformer_engine package to insert instrumentation probes. This technique modifies the execution flow of the application by replacing established entry points with custom wrappers.
[DYNAMIC_EXECUTION]: The documentation recommends injecting code via the PYTHONPATH environment variable. This allows a "bootstrap" tracer to be dynamically loaded into the execution environment from a user-specified path at interpreter startup, facilitating the execution of external instrumentation code.
[COMMAND_EXECUTION]: The workflow instructs the execution of several shell-based tools for profiling and analysis, including nsys profile, ncu (NVIDIA Compute Profiler), and grep. These tools are used to monitor GPU kernels, library calls, and log output.
[INDIRECT_PROMPT_INJECTION]: The skill defines a process for reading and comparing external JSONL trace files, which represents an attack surface for indirect prompt injection.
Ingestion points: The agent or associated scripts ingest JSONL files from local or shared filesystems (e.g., Lustre) to perform causal diffing between runs.
Boundary markers: There are no defined delimiters or instructions provided to ensure the agent ignores natural language instructions that might be embedded within the fingerprinted data records.
Capability inventory: The skill possesses the capability to write to the filesystem (via Path.open("w")) and execute external profiling commands.
Sanitization: The trace comparison logic does not include sanitization or validation steps to filter potential injection payloads from the ingested JSONL content.

mcore-determinism-debug