The Agent Skills Directory

[SAFE]: The skill operates entirely within the context of formal methods maintenance. It defines a structured workflow for comparing documentation, code, and formal model files to detect 'drift'.
[SAFE]: Instruction headers like Core Rule and Workflow are instructional and do not attempt to bypass agent safety filters or override system prompts. The directive 'Do not make the model green by silently changing the property' is a domain-specific best practice for formal verification, not a prompt injection attempt.
[SAFE]: The skill mentions various formal methods tools (Z3, TLA+, Alloy, Dafny, etc.) and CI environments, but it does not provide or execute arbitrary or hidden commands. It expects the agent to use standard tools already present in the environment or defined in existing CI workflows.
[SAFE]: Network and file access patterns are limited to the intended domain (reading source code, docs, and model files). No exfiltration patterns, hardcoded credentials, or suspicious remote downloads were detected.
[SAFE]: The evals and fixtures directories contain legitimate testing data for the skill, demonstrating expected behaviors like handling TLC (TLA+ Checker) log changes and payment idempotency traces.

formal-methods-drift-guard