The Agent Skills Directory

[COMMAND_EXECUTION]: The skill utilizes the subprocess module to run git commands for repository management and commit analysis. All shell calls are implemented with list-style arguments, which is a secure method that avoids command injection vulnerabilities.
[EXTERNAL_DOWNLOADS]: Repository cloning and updating are performed targeting official project sources on GitHub and GitCode. These activities are consistent with the skill's purpose and do not involve untrusted or suspicious hosts.
[PROMPT_INJECTION]: The skill processes log data from external sources (regression and training logs) to extract failure contexts. While technically an indirect injection surface, the risk is handled via safe parsing methods.
Ingestion points: Reads comparison and training logs from /home/st_daily_verl/ via scripts/parse_result_log.py and scripts/extract_failure_tail.py.
Boundary markers: No explicit boundary markers are used when extracting log segments.
Capability inventory: File system access for reading logs and Git-based repository management via subprocess.run.
Sanitization: Uses ast.literal_eval for safe evaluation of Python-like lists found in logs and regex for pattern identification.

msverl-daily-regression-triage