Skip to content

Codex PROVE v1.2.0 — Sol-first,按需分工

Latest

Choose a tag to compare

@yehyakin yehyakin released this 02 Oct 12:29
c54a592

v1.2.0:Sol 优先,按需分工

本版让 Sol 6.1 负责完整任务与有用的协调,Astra 只在难题或必要的独立评审中提供只读帮助,Luna 处理规则明确、可客观验收的机械批量。小任务保持 Direct;只有明确调用 $codex-prove 才启用,不引入后台进程、全局 hook 或新的运行依赖。

主要变化

  • 不设固定模型配额,不因 Host 身份未知而自动加一层协调者;只有能独立推进、值得交接成本的工作才拆分。
  • 明确唯一写入负责人、交接顺序、相关验证证据复用和外部动作恢复边界。
  • 加强开发计量:区分 root-only 与全程覆盖,识别重复、冲突、非法、缺失及失败的事件,不把已知成功子集包装成完整账单。
  • 更新中英文 README、安装/恢复说明、配图、同类项目研究与脱敏三组对照;历史失败与未验证项继续保留。

这些是工作约束,不是 OS 文件锁、持久 checkpoint、exactly-once 或硬计费上限。

发布与验证

  • 合并 PR:Release v1.2.0 #7。
  • 版本标签:v1.2.0;最终 main 提交:c54a59249a2560c4cb43dc8855cf160224121733。
  • 最终候选本地验证:169 项测试通过、0 跳过,源码校验及四组配图事实校验通过。合并后的 Git tree 与这一候选完全一致;207 个发布文件逐项核对通过。
  • main 的同一精确提交已通过全部 8 个矩阵 job:POSIX CI 与 Windows CI。完整日志确认实际检出的就是上述 main SHA,不借用 PR 合成合并提交的结果。
  • 标签推送后的同 SHA 复验也全部通过:tag POSIX CI 与 tag Windows CI,另 8 个矩阵 job。
main CI 环境 每个 job 的 unittest 结果 额外执行
Ubuntu,Python 3.11 / 3.13 169 项:167 通过、2 跳过 源码校验、shell 语法、安装只读预检
macOS,Python 3.11 / 3.13 169 项:168 通过、1 跳过 源码校验、shell 语法、安装只读预检
Windows 2025 / 2022,powershell / pwsh 169 项:132 通过、37 跳过 原生 PowerShell 解析、源码校验、Windows 安装生命周期

跳过项没有算作通过:所有 CI job 的浅克隆缺少历史提交,因此历史运行文件哈希复核跳过 1 项;Linux 另跳过 1 项 macOS sandbox 检查;Windows 另跳过 35 项 POSIX 专用测试及 1 项 macOS sandbox 检查。Windows 原生安装、迁移、恢复、拒绝覆盖与失败恢复测试另行真实执行并通过,不代表覆盖了 POSIX pilot。跳过总数来自完整日志,具体原因依据同版本测试源码条件核对。这些是托管 runner 验证,不是新模型实测或物理机器验收。

冻结运行内容为 11 文件,SHA-256 72d3f417ea4559bf1a5afa71448c79c6f5570b3f43a4ca1e3458909d60c674e0,与三组对照使用的运行内容相同。完整发布源码清单 SHA-256 为 c944a0a2f360050a0aa22e83fa18fb0a6eaa9182cef26e5b864202adbc0fdadf;这两个内容指纹不是 Git commit ID。

测量边界

三组对照报告只有一个任务、每组一次;三组独立检查均为 19/19,全部没有委派。PROVE 为 252.253 秒、统一 Standard API 等价估算 $0.1749800:相对普通 Codex 低 13.86% / 快 10.20%,相对定制 Host 的 da34 组高 0.61% / 慢 9.36%。这是固定顺序的单例观察,不是普遍性能、混合模型、订阅额度或提供方实际费用结论;外层协调与事后评分成本未计入。

82.8% 配图仍是预算示例,不是实测节省。配图档案中的 2026-10-01 候选状态是制作时快照,不代表当前发布状态。Native Nested、物理 Windows 机器和完整路由成功率仍不作已验证承诺。

安装与恢复

使用带 Git 元数据的完整 checkout,并固定版本。macOS / Linux:

git clone https://github.com/yehyakin/codex-prove.git
cd codex-prove
git checkout v1.2.0
bash scripts/validate.sh
bash scripts/install.sh

Windows 在同一版本的 checkout 中运行:

powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/validate.ps1
powershell.exe -NoProfile -ExecutionPolicy Bypass -File scripts/install.ps1

安装后重新打开 Codex 会话;模型可用性取决于账号。GitHub 自动生成的源码 ZIP/tarball、运行包片段及审计快照不是直接安装输入。恢复用 bash scripts/uninstall.sh --restore-latest,或 scripts/uninstall.ps1 -RestoreLatest;只恢复最近一次受管理备份,改过的受管理文件会触发拒绝,不覆盖其他 Skill/角色或整份共享配置。

English

Sol 6.1 now handles complete work and useful coordination; Astra provides scoped read-only advice, and Luna handles objectively checkable mechanical batches. Small work stays Direct. Invocation remains explicit, with no fixed model quota, daemon, global hook or new runtime dependency.

This release tightens write ownership, handoff, evidence reuse, recovery boundaries and usage accounting, and updates bilingual documentation, artwork and a redacted three-arm case. Workflow rules do not provide OS locks, durable checkpoints, exactly-once execution or hard billing limits.

The tag identifies main commit c54a59249a2560c4cb43dc8855cf160224121733, whose full source tree matches the locally verified candidate (169 tests, zero skips). All 8 main CI matrix jobs passed on that exact commit, as did all 8 additional tag-push jobs linked above. Each main job discovered 169 tests: Ubuntu passed 167 with 2 skips, macOS 168 with 1 skip, and Windows 132 with 37 skips. Skips reflect one unavailable historical revision in shallow clones plus platform-specific tests; native Windows lifecycle tests ran separately and passed. The linked full logs establish hosted-runner source/installer validation, not fresh model-runtime, physical-hardware or overall routing validation.

The single-task case passed 19/19 in every arm with no delegation. PROVE was 13.86% lower in normalized API-equivalent cost and 10.20% shorter in elapsed time than ordinary Codex, but 0.61% higher and 9.36% longer than the host-customized da34 arm. Fixed order, one run per arm, unknown provider charges and excluded outer coordination/grading costs limit this observation. It is not a general, mixed-model or subscription-savings claim. The 82.8% artwork remains illustrative budgeting.

Install from a complete Git checkout at v1.2.0, validate, run the platform installer, then reopen Codex. Source archives are not direct installer inputs. Rollback restores the latest managed backup rather than arbitrary versions or the user's whole configuration. Native Nested, physical Windows hardware and overall routing success rates remain unverified.

完整变更 / Full changelog · 版本说明 / Source release notes