zz-plugin-eval

Plugin zz-plugin-eval from marketplace zz-stack in zhixuan312/zz-stack. Commands read from the repo's manifest — 4 days ago.

Skills (9)

• zz-plugin-define-qualify

  Stage 4 of zz-plugin-eval (DEFINE/QUALIFY), and the one gate that matters most. Derive what good means for THIS plugin from its own profile and DISCOVER's candidates, write it into protocol.md, get a person to agree it, then qualify every model-backed evaluator it names before anything is scored.

• zz-plugin-discover

  Stage 3 of zz-plugin-eval (DISCOVER). Mine one OBSERVE snapshot's own real refusals and stage returns for candidate failure modes, before any protocol exists — so DEFINE/QUALIFY freezes questions worth asking, not questions invented from nothing.

• eval

  Run the ZZ Plugin Evaluation flow for your team.

• zz-plugin-evaluate

  Stage 5 of zz-plugin-eval (EVALUATE). Bind an approved protocol version to a subject's own observation snapshot, route every measure the protocol names to the real evidence it judges, and reduce the result to one deterministic overall score with its status, coverage and guardrails. No recommendation

• zz-plugin-explain

  Stage 6 of zz-plugin-eval (EXPLAIN). Record what EVALUATE's own score and assessments actually found — strengths, defects, unknowns, each with an owner — and let findings.md regenerate itself from the current record. Ungated measurement output, never an approval gate.

• zz-plugin-identify

  Stage 1 of zz-plugin-eval (IDENTIFY). Settle which plugin is being evaluated, at which exact content — catalog release or third-party capture — before any other tool on the door will resolve anything against it. Writes an immutable subject_version.

• zz-plugin-improve

  Stage 7 of zz-plugin-eval (IMPROVE). Turn plugin-owned findings into one candidate patch, build and gate it locally, and hand it to promotion for an owned subject — or write an owner-facing proposal for one this team cannot release. A released improvement is judged on real use afterwards, never by r

• zz-plugin-observe

  Stage 2 of zz-plugin-eval (OBSERVE). Compute the pre-protocol observation snapshot — production facts from this subject's real runs, in one resolved window — with its sufficiency verdict and the coverage it was derived from. No model touches any of it.

• zz-plugin-promote-verify

  Stage 8 of zz-plugin-eval (PROMOTE/VERIFY), the promotion boundary. Once IMPROVE has a built and gated candidate of an owned subject, prepare and gate the exact patch, apply it only after every required owner approves, record what happened, then judge the release on real use — and roll it back if it

MCP servers (1)

• zz-plugin-eval

  connects to: https://api.zz-stack.com/eval/mcp