What problem does it solve? Teams shipping AI delivery plugins previously evaluated skills and MCP servers separately, so whole-plugin questions like stage recovery and actual tool use went unmeasured, and opinions about plugin quality were settled by whoever read the artifact last rather than by an agreed ruler. ## Core Features & Use Cases - Five-stage evaluation flow: locate the plugin at a released version, profile trace evidence, define what good means, judge against the approved ruler, and report findings. - Gated scoring: nothing is scored before a person approves rulers.md, and nothing closes before a person reads findings.md; scoring runs through a pinned model via round_judge, never the orchestrating agent. - Honest insufficiency handling: thin trace history yields an explicit not measured rather than a fabricated score, with document- and initiative-scoped rulers as fallbacks. - Use case: A team lead asks whether a plugin's flow is going in circles. Instead of agreeing from a personal read, the flow profiles five usable runs, applies the pre-agreed ruler, and produces a gated findings document. ## Quick Start Invoke the /zz-plugin-eval:eval command to start a gated evaluation of one installed plugin at a released version.
