Releases: yehyakin/codex-prove
Release list
Codex PROVE v1.0.0
Codex PROVE is the model-neutral successor to Sol Control.
- New canonical Skill:
$codex-prove - PROVE: Planning, Routing, Ownership, Verification, Evidence
- Generic Controller, Complex Worker, and Efficient Worker roles
- Transactional macOS, Linux, and Windows migration from managed v0.1-v0.5 installs
- Temporary explicit
$sol-controlcompatibility alias for v1.0 - 115 local tests and 39 Forward scenarios
- POSIX and Windows PowerShell 5.1/7 CI passing
- Fresh-session Compatibility route and all three runtime role mappings verified
Native Nested with the new role names and physical Windows 11 runtime identity remain explicitly unverified. See the implementation report and runtime surface matrix for evidence boundaries.
Sol Control v0.5.0
Sol Control v0.5.0
中文
本版本把 Sol Control 升级为证据优先的单主控编排:Sol 继续负责理解、规划、分配与最终审核,Terra High / Luna Max 只执行有边界任务。
- 引入稳定 Requirement ID、产物优先审核和 verify-the-verifier。
- 高风险或证据冲突时最多增加一次只读 challenge;普通任务零 challenge。
- 长任务增加可恢复 Resume Packet,保留 ownership、候选身份和重试次数。
- 正式移除旧
$sol-luna兼容入口,统一使用$sol-control。 - v0.4.x 升级会先校验并备份旧入口,再事务性移除;
--restore-latest可恢复升级前状态。 - 本地
136/136测试通过;Ubuntu、macOS、Windows PowerShell 5.1 与 PowerShell 7 CI 全部通过。
限制:完整 48-cell × 3 次匹配基准尚未完成。现有一组真实 smoke pair 不能支持普遍的质量、成本或延迟提升结论。
完整证据见 SOL_CONTROL_V050_IMPLEMENTATION_REPORT.md。
English
This release makes Sol Control evidence-first while keeping one controller: Sol owns understanding, planning, assignment, and final review; Terra High and Luna Max execute bounded tasks only.
- Stable Requirement IDs, artifact-first review, and verify-the-verifier checks.
- At most one read-only challenge for high-consequence or conflicting evidence; zero challenge for ordinary work.
- A resumable packet for long tasks that preserves ownership, candidate identity, and attempt counts.
- The legacy
$sol-lunacompatibility entry is removed; use$sol-control. - v0.4.x upgrades verify and back up the old entry before transactional removal;
--restore-latestrestores the pre-upgrade state. 136/136local tests pass, with green Ubuntu, macOS, Windows PowerShell 5.1, and PowerShell 7 CI.
Limitation: the complete 48-cell × 3 matched benchmark is not finished. One live smoke pair does not justify a general quality, cost, or latency improvement claim.
See SOL_CONTROL_V050_IMPLEMENTATION_REPORT.md for the full evidence record.
Sol Control v0.4.1
修复内容 / Fix
修复 Codex Desktop 中的身份握手死锁。Desktop 子 Agent 无法观察自己的模型和 reasoning effort,v0.4.0 因要求 child 自证而正确返回 unproven 并 Fail Closed。
v0.4.1 改为由权威 Host/tool 角色映射和父线程启动记录证明 agent、model、effort 与 fork_turns; child 只证明权限边界、运行约束和零副作用。无法取得 Host 权威证明时仍然 Fail Closed。
This release fixes the Desktop identity-handshake deadlock. Exact agent/model/effort/fork proof now comes from the authoritative Host/tool mapping and parent launch record; the child reports only permissions, operational constraints, and zero side effects. Fail Closed remains mandatory when Host proof is unavailable.
验证 / Validation
- Local:
scripts/validate.shPASS; 113/113 tests PASS; installation consistency PASS - POSIX tag CI: Ubuntu + macOS, Python 3.11/3.13 PASS
- Windows tag CI: Windows Server 2022 + windows-latest, PowerShell 5.1/pwsh PASS
- Desktop Sol handshake and final review: PASS
边界 / Boundary
Desktop Sol handshake and final review are verified. Desktop nested worker dispatch was not exercised by this acceptance run and remains unproven.
Full diff: v0.4.0...v0.4.1