The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes potentially untrusted data from multiple sources.
Ingestion points: The skill reads from definition.json, workload.jsonl, memory/*.json, responses from gpu-wiki/tools/query_nl.py, and content fetched from the public_web.
Boundary markers: The instructions specify that the sandbox environment isolates inputs and the driver deletes input JSONs before importing candidate code.
Capability inventory: The skill utilizes tools/sandbox.py for code execution and records experiment results using <JOURNAL_CLI> and tools/iteration_trace.py.
Sanitization: Explicit instructions are provided to exclude credentials, private URLs, and raw tool output from telemetry.
[COMMAND_EXECUTION]: The skill defines a loop that requires executing multiple local scripts and shell commands.
Evidence: Usage of python tools/sandbox.py, bash tools/profile_nvidia.sh, python3 tools/iteration_trace.py, and the variable command <JOURNAL_CLI>. These are used for telemetry, profiling, and journaling.
[DYNAMIC_EXECUTION]: The skill facilitates the creation and execution of code at runtime as part of the optimization process.
Evidence: The agent is instructed to modify kernel.py and is permitted to create a local fallback driver in the <PROFILE_DIR>/harness/ directory, which is then executed by the profiling tools.

gpu-kernel-episode-loop