The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The script profiling-analysis-communication/scripts/detect_slow_rank.py performs runtime installation of the mstt toolset. It uses git clone to fetch source code from https://gitee.com/ascend/mstt.git, followed by pip install and setup.py bdist_wheel to compile and install the package on the local system.
[EXTERNAL_DOWNLOADS]: The skill downloads external software components from Gitee (gitee.com/ascend/mstt.git) to fulfill its communication analysis functionality. While Gitee is a standard hosting platform for the Ascend community, this behavior introduces supply chain risks if the remote repository is compromised.
[COMMAND_EXECUTION]: Multiple scripts utilize the subprocess module to orchestrate analysis workflows:
scripts/performance_analysis_main_process.py uses subprocess.Popen to launch sub-skills.
profiling-analysis-communication/scripts/detect_slow_rank.py uses subprocess.run to execute git, pip, and specialized analysis commands (msprof-analyze).
profiling-analysis-computing/scripts/op_perf_analysis_combine.py uses subprocess.run to chain multiple Python scripts together.
[INDIRECT_PROMPT_INJECTION]: The skill has a surface for indirect prompt injection via data processing:
Ingestion points: Reads and parses user-provided CSV files (step_trace_time.csv, op_statistic.csv, op_summary.csv, kernel_details.csv) and SQLite databases (analysis.db).
Boundary markers: Absent; the skill directly processes the contents of these files into internal data structures.
Capability inventory: The skill has capabilities to write files (reports), execute shell commands via subprocess, and install remote code.
Sanitization: Relies on standard library parsers (pandas, sqlite3) but does not specifically sanitize inputs against adversarial content designed to influence the agent's summary or logic flow.

profiling-analysis