The Agent Skills Directory

[COMMAND_EXECUTION]: The skill instructs the agent to execute specific shell commands to set up the environment and run performance profiling tools.
Evidence: Shell commands for LD_LIBRARY_PATH and set_env.sh to configure the Ascend environment in SKILL.md.
Evidence: Execution of msprof and pytest for performance and correctness testing of user-supplied code in SKILL.md.
[INDIRECT_PROMPT_INJECTION]: The skill processes and executes code supplied by the user, which presents a vulnerability surface where malicious instructions could be embedded in the operator source code.
Ingestion points: The skill ingests user-provided Triton operator code (<op_name>.py) and performance test files (test_<op_name>_perf.py) as specified in SKILL.md.
Boundary markers: Absent; there are no specific instructions or markers to treat the processed operator code as potentially untrusted or to ignore embedded instructions.
Capability inventory: The skill utilizes system-level commands (pytest, msprof) to execute the operator code and requires file system access to modify and iterate on the code during optimization.
Sanitization: No evidence of code sanitization, static analysis, or input validation before execution is provided in the skill instructions.

vector-triton-ascend-ops-optimizer