Codex PROVE v1.1.0 — GPT-6 分工与轻量编排
这次主要更新模型分工,并减少之前反馈较多的流程开销和反复批准。
更新内容
- Astra 主控、GPT-6 Sol 专项、Terra 常规、GPT-6 Luna Max 批处理。 小任务仍然直接做,不要求四个模型每次全上。
- 有明确调用记录就直接开工,减少身份自证空回合;已授权的本地修复和测试继续进行,需要新权限或关键决定时才问用户。
- 独立任务并行,有依赖的分 Wave,同一写入范围只有一个 Owner。
- 借鉴 Ponytail,优先复用现有实现、标准库和平台能力。
- 完善跨平台安装、备份和恢复;更新中英文 README、配图及费用说明。
本次验证
- 114 项本地测试及 Skill Creator 校验通过。
- 四模型真实 Compatibility 流程通过:Astra 规划 → Sol/Luna 独立执行 → Terra 完成依赖任务 → 同一个 Astra 终审 PASS。隔离项目的 14 项检查通过,预置用户修改保持不变。
- macOS 安装完成,11 个运行文件与源码逐字节一致,config.toml 未变,旧版本已备份。
- 全新 CLI 会话发现已安装 Skill,简单文案修改零子代理;另一次新会话发现四个角色的正确工具声明。
- 发布提交
a30e84c的 Linux / macOS CI 与 Windows CI 均通过。
这次联跑使用明确指定模型和推理等级的 Compatibility 启动,不冒充旧 Desktop 会话已经加载了新 custom-role。Native Nested、实体 Windows 11 四模型运行仍未单独验证。本次 smoke 不代表普遍质量、速度或成本收益。
更新方法
更新仓库,重新运行对应平台的验证和安装脚本,再打开新的 Codex 会话。只执行 git pull 不会更新全局安装。
附带的 prove-v1.1.0-smoke.zip 可以复现本次合成产物的验收;解压后运行:
python3 -m unittest -v test_acceptance test_scheduler_worker它不会自动调用模型。SHA-256:dd5d81934f06b333401d19678c9277d9ea088f3cac4aa431fa9b48617d312e12。
English
GPT-6 routing with less ceremony: Astra leads, Sol handles specialist work, Terra handles regular implementation, and Luna Max handles mechanical batches. Small tasks stay Direct. The release includes recoverable local-work continuation, scoped ownership, dependency waves, safer installers, and refreshed bilingual documentation.
Verified: 114 local tests, Skill Creator validation, a live four-model Compatibility smoke with 14 passing checks and Astra's final PASS, a backed-up macOS installation, and fresh CLI discovery/Direct checks. Native Nested and physical Windows 11 four-model execution remain unverified. This smoke is not a performance or savings benchmark. Reinstall after updating and start a new Codex session.