skill-comply

To verify agent compliance, you generate an expected tool-call sequence and grade observed agent traces against it. This workflow validation checks whether coding agents follow declared rules by testing task scenarios across different prompt strictness levels.

LLM classification for tool-call tracing maps observed agent tool calls to expected spec steps. It uses an LLM to classify calls, then enforces deterministic temporal ordering validation to ensure the agent followed the correct sequence.

You test agent compliance across varying prompt strictness by running the same task at three levels: supportive, neutral, and competing. This stress testing detects when compliance breaks down under loosely supporting or competing prompt conditions.

You generate a self-contained Markdown compliance report containing compliance rates and full tool-call timelines. This output summarizes whether agents followed declared behaviors and includes the complete trace history for review.

Yes, you need pyyaml installed as a dependency to run agent compliance checks. The skill requires this Python library to parse configuration files and execute the compliance verification scripts.

The best way to detect silent process drift is to measure agent compliance by generating expected behavioral sequences and grading observed tool-call traces against them. This reveals whether agents comply with declared skills even when prompts loosely support them.