The Agent Skills Directory

[COMMAND_EXECUTION]: The skill contains Python scripts (flake_detector.py, regression_gate.py, trajectory_checker.py) that perform statistical analysis on JSONL result files. Analysis of the source code confirms they do not use subprocess, os.system, or any other shell execution utilities.
[DATA_EXFILTRATION]: The scripts are strictly deterministic and limited to the Python standard library. They do not perform network requests (requests, urllib, etc.) or access sensitive system paths. All operations are performed on local data files provided as arguments.
[REMOTE_CODE_EXECUTION]: There are no patterns of remote code execution. The skill does not download external scripts or packages. The CI example in assets/ci_eval_gate.example.yml demonstrates how to integrate the local scripts into a GitHub Actions workflow using standard checkout and execution patterns.
[DYNAMIC_EXECUTION]: All scripts use fixed logic for statistical calculations (e.g., Wilson score intervals, McNemar tests). There is no use of eval(), exec(), or dynamic module loading.
[INDIRECT_PROMPT_INJECTION]: The skill processes JSONL files containing agent trajectories. It acts as a code-based grader that compares strings and boolean values. Because the output of these tools is meant for human developers or CI logs and is not re-interpolated into an LLM prompt, the risk of indirect prompt injection is negligible.
[OBFUSCATION]: The scripts are written in clear, documented Python. No Base64, hex encoding, or character-level obfuscation was detected in the code or metadata.
[CREDENTIALS_UNSAFE]: No hardcoded API keys, tokens, or secrets were found. The sample data files contain only simulated agent performance metrics.

agent-regression-evals