The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The scripts/detect_slow_rank.py script fetches the mstt tool source code from the official repository at gitee.com/ascend/mstt.git to facilitate the local analysis environment.
[COMMAND_EXECUTION]: The skill executes various system commands including git, pip, and msprof-analyze via the subprocess module. These calls use specific argument lists, ensuring that external processes are invoked securely without shell expansion risks.
[DYNAMIC_EXECUTION]: The analysis workflow involves compiling and installing the msprof_analyze package from source at runtime, allowing the skill to configure its own diagnostic environment as needed.
[INDIRECT_PROMPT_INJECTION]: The skill analyzes external profiling data provided by users to generate performance reports.
Ingestion points: The detection script accepts directories or ZIP archives containing NPU profiling data via the --input parameter in scripts/detect_slow_rank.py.
Boundary markers: Present. The output is structured with clear text-based separators (e.g., '📊 快慢卡检测结果分析') to define analysis sections.
Capability inventory: The script executes NPU-specific analysis binaries and queries resulting SQLite databases for diagnostic metrics.
Sanitization: Absent. No explicit sanitization or filtering is performed on the contents of the ingested profiling data before statistical processing.

profiling-analysis-communication