skill-comply

To verify agent compliance, you can generate behavioral specs from Markdown rule files, execute testing scenarios, and classify tool-call traces to measure whether the agent follows documented rules and skills.

Tool-call trace classification evaluates agent behaviors by capturing stream-json outputs from executed scenarios, classifying tool events with an LLM, and enforcing deterministic temporal ordering to generate compliance reports.

You test prompt independence by generating supportive, neutral, and competing scenarios from documented rules, executing them against an agent, and comparing the resulting tool-call traces for consistent behavioral compliance.

Yes, skill-comply integrates directly with claude -p stream-json outputs, capturing tool-call traces during scenario execution to classify events and enforce temporal ordering for compliance measurement.

You generate compliance reports by extracting expected behavioral sequences from TDD or linting rule Markdown files, executing pressure-testing scenarios, and emitting a self-contained Markdown report summarizing compliance rates.

LLM-based classification depends on stream-json tool-call traces and deterministic temporal ordering checks, meaning compliance measurement is limited to observable tool events rather than internal agent reasoning processes.