The Agent Skills Directory

[COMMAND_EXECUTION]: The skill provides explicit shell commands to compile C source files and execute Python scripts to verify data race detection logic.
SKILL.md and evals/README.md contain commands such as gcc -Wall -Wextra -Werror -O2 -pthread examples/good/race_free_counter.c -o /tmp/c1.exe && /tmp/c1.exe.
Verification steps involve running python examples/good/kcsan_model.py and other internal scripts.
[DYNAMIC_EXECUTION]: The skill uses Python models and compiled C binaries to demonstrate and verify KCSAN behavior at runtime.
Scripts like kcsan_model.py and race_fix_demo.py are executed to simulate memory interleaving and verify fix efficacy.
The reliance on local execution of provided code is a functional requirement but represents a dynamic execution surface.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to triage KCSAN reports and review kernel patches, which are untrusted external data sources.
Ingestion points: The skill instructions in SKILL.md describe processing user-provided KCSAN report text ("BUG: KCSAN: data-race in ...") and kernel patch files.
Boundary markers: There are no explicit instructions for the agent to use delimiters or ignore potentially embedded instructions in the report data.
Capability inventory: The agent has the capability to execute shell commands (gcc, python) as part of its verification workflow.
Sanitization: No sanitization or validation of the incoming report/patch content is described before the agent processes it.

data-race-kernel-detection