Skip to content

Latest commit

 

History

History
32 lines (25 loc) · 2.72 KB

File metadata and controls

32 lines (25 loc) · 2.72 KB

Changelog

All notable changes to this project are recorded in this file.

The format follows Keep a Changelog and this project adopts Semantic Versioning.

[Unreleased]

[0.1.0] - 2026-06-04

Added

  • First release of judge-codex — Codex-as-orthogonal-LLM-jury for the plan cycle ecosystem.
  • 8 slash commands (/judge-codex:setup, :discover, :plan, :implementation, :final, :auto, :status, :help).
  • 7 sub-agents: judge-codex-jury (thin forwarder) + 4 stage-judges (discover-judge, plan-judge, implementation-judge, final-judge) + quality-evaluator (Autoresearch keep/discard gate) + report-writer (auto-aggregator).
  • 3 skills: codex-cli-runtime, judge-prompting, cycle-plan-context.
  • 4 JSON schemas (one per stage) using plan's canonical verdict vocabulary plus final-stage meta-verdicts (META_DEFECT_FOUND, AGGREGATOR_BUG_SUSPECTED).
  • scripts/codex-companion-judge.mjs (~350 lines): single-entry runtime that locates artifacts in the consumer repo, assembles prompts (agent system + golden-rule + artifact + anti-anchoring), invokes codex exec --output-schema --output-last-message --skip-git-repo-check --cd, validates + persists output.
  • scripts/setup-check.sh — environment verification (Node ≥ 18.18, codex CLI, login).
  • hooks/hooks.json — optional Stop hook (off by default).
  • Marketplace manifest at .claude-plugin/marketplace.json for /plugin marketplace add install.
  • README documents cycle-aware verdict vocabulary distinct from the binary approve / needs-attention of generic code-review plugins.

Verified

  • Live integration test against plan/knowledge-base/plans/harden-fabrication-and-cq-gate-plan.md:
    • Codex elapsed 87993 ms.
    • Codex verdict INVALID (score 49) — caught fabricated ADR reference D9 that the Claude-side plan-confidence M3 v0.1 detector had missed (M3 only inspects #### Evidence blocks; D9 was in plan prose outside that scope).
    • Plus 2 medium-severity soft caps (goal_not_smart_timebound, risks_section_missing).
    • This is the disagreement-with-Claude scenario the plugin exists to surface.

Inspired by

  • openai/codex-plugin-cc (Apache 2.0) — adopted .claude-plugin/ layout, codex-companion subprocess pattern, disable-model-invocation: true convention, Stop hook. Distinct: cycle-aware (one command per plan stage with golden-rule injection), structured verdict vocabulary, review-of-review final stage.
  • paulohenriquevn/loop-code-review (MIT) — adopted per-specialist agent layout, Autoresearch keep/discard quality-evaluator gate, report-writer aggregator.