[cocolord]

Goal and delivered outcome

A standalone core installation must connect a project, prepare a task, execute a selected host with independent validation, and recover the original result. The ordinary default hard-lease path previously reached completion without its admitted lease proof; the original core candidate also rejected healthy installations in deep doctor and advertised an unavailable history action.

This change provides a source-built loopx-core wheel through the existing runtime owners and repairs those gaps. Core excludes Chat, web assets, the canary fleet and the full CLI facade. It retains built-in capability and host compatibility imports. This is the core-artifact prerequisite for RFC R1/M3; full minimal closure and split workspace/meta publication remain open.

Implementation and specification

• Specification: docs/architecture/rfcs/monorepo-distribution-split-v0.md at d9432f23544e595f9861fcc6d8453ea926721612, particularly R1/M3 and installed-task, recovery, diagnosis and feature-off boundaries. The bilingual checkpoint records this scoped delivery and preserves D1-D7.
• Packaging reuses tracked sources; it adds no parallel runtime, capability registration, preset or default activation.
• Canonical lease key/version travels through the existing TypeScript preparation and checked Turn context to the existing completion owner. Recovery cannot borrow a renewed, released or replaced lease.
• Distribution lookup is reused by installation identity, upgrade guidance and deep doctor. Required ordinary CLI files remain mandatory for ordinary installations.
• Core executes the advertised history action through its shipped handler and parses global version options through the canonical parser. Ordinary CLI behavior retains its complete facade.
• Bounded refactor applied: reuse command registrars and the distribution owner, keeping decision authority in TypeScript and Python as packaging/transport.

Validation

Tested commit: 6d4eba2846b62bee8cb9551808c137de1d414a13, based on current main. Synthetic fixtures; macOS arm64, Python 3.13.5 and Node 26.0.0.

Check	Result and boundary
Installed core	All 15 cases passed against the rebuilt final wheel. Includes default SQLite and configured File hard leases, independent artifact validation, SIGKILL recovery, no-effect replay, changed-lease rejection, actual catalog action, deep doctor and ownership/missing-file negatives. The final wheel is built from the exact tested commit.
Defect sensitivity	The same deep-doctor/version tests fail on the original core candidate. Both installed ordinary baseline task cases fail with lease_fence_required; the current ordinary installation passes all 10 task/recovery cases. These expected failures remain failures.
Installed ordinary parity	All eight matched CLI/installation observations agree between immutable main and this runtime, including deep doctor, help, catalog, global version and invalid arguments. Both ordinary wheels include a verified Chat bundle.
Focused source regression	31 passed, 3 platform skips for distribution identity, installer and release-candidate checks. Complete CLI/preparation: 327 passed. Architecture/Turn: 2097 passed.
TypeScript	Control-plane typecheck passed; 9 preparation and 64 Turn-envelope cases passed.
Static and public boundary	Changed-file Ruff, committed diff hygiene and staged bilingual RFC checks passed. Publication candidates have completed clean public-boundary scans. The full local check separately reports unrelated legacy Goal authority failures.
Premerge and CI	Exact-head standard premerge: direct checks and 18 smokes passed; install-local-smoke fails identically on immutable main and this head at an unchanged installed skill-document link boundary. This pre-existing failure remains recorded. CI qualification/readiness remains separate.

No model service, paid benchmark, current-host installation or public release was executed. PostgreSQL provider/store/schema transitions are unchanged; its separate integration qualification is not claimed. Model effectiveness, complete costs, other host/platform adoption and independent Standard packages remain open.

CLI and visual evidence

UI impact: none. Existing graphical entrypoints are unchanged. Installed CLI help, actual catalog execution, result readback and failure/recovery are covered above; no graphical screenshot is applicable. Return to the ordinary distribution by switching environments after preserving/reconciling the original task state; do not overlay the distributions.

Author declaration and boundary

Written by: model_agent, TRAE, GPT-6 family, OpenAI. Every commit carries DCO sign-off. The diff and public artifacts exclude private Goal state, local paths, raw logs and trajectories, credentials and internal links. There is no new merge, install, release or paid-execution authority.

[loopx-agent]

Reviewer: model_agent · gpt-6.1-sol · OpenAI · runtime_reported · max

本次评审对应提交 0f65ed0206db6fbae16ba1141f3320d6a292b441。

两个 P2 待修。 健康 core 的深度安装检查报错，命令目录中的 history 子命令无法在该发行物执行。

动机 / Motivation

在新环境安装 core、配置并执行隔离任务，之后用命令目录和 doctor 检查的用户。默认硬租约的 generic 完成漏传执行凭据，普通安装也没有独立 core wheel。独立安装的 core 以默认 SQLite/hard lease 执行一项合成任务，独立验收产物，SIGKILL 后恢复原结果并读回。重放无新执行和扣额，更新或释放原 lease 后拒绝恢复。独立安装的 core 以默认 SQLite/hard lease 执行一项合成任务，独立验收产物，SIGKILL 后恢复原结果并读回。重放无新执行和扣额，更新或释放原 lease 后拒绝恢复。本地 macOS arm64 和隔离 Python 3.14/Node 下的合成任务。未查询远端 CI、未验证真实宿主采用、其他平台和完整 minimal/Standard 资格。仅安装 loopx-core 时，doctor --installation-only 成功，--deep 却查找 loopx 的 RECORD、loopx.cli 和 loopx/cli.py，把健康 core 报为安装失败。core 命令目录列出 history trajectory-hygiene，按给出的命令在同一安装执行却退出 2。该路径由 dispatcher 转给省略的完整 CLI。

改动思路 / Approach

安装物拆分和任务终态沿现有 owners，独立轮子没有引入第二套执行。安装归属应贯穿同一条诊断路径，命令目录也应以完整的可执行入口为依据。 R1 的有用增量来自实际独立 wheel 和已有任务 owner。读出原 lease 后沿原终态事务提交，故障后仅接续同一条结果。双语 RFC 把必要的目录迁移与 core 发行一起推进，完整 M3 及默认策略各自保留原验收。

具体改动 / Changes

关键代码讲解

• 准备器复用 tracked 运行时源码，wheel 省略 Chat、web、canary 和完整 CLI。package-smoke 在独立环境运行安装后任务夹具。 发行目录
• canonical 读取与 TS preparation 捕获匹配租约的 key/version。generic run-once 把原证据交给已有终态 owner。默认 SQLite 基线到终态时缺凭据失败，安装头成功。 终态调用
• 新安装标识识别 core，deep doctor 的 RECORD 和代表性文件仍来自普通 wheel。命令目录只按顶层名称过滤，会发布依赖完整 fallback 的子命令。 深度检查 目录过滤

适用 docs/architecture/rfcs/monorepo-distribution-split-v0.md @ 43b4699。installed_task_and_recovery 和 original_lease_generation 已验证，distribution_aware_discovery_and_diagnosis 未满足。 CLI 与恢复准则

对主干的风险 / Risks

core 的浅层安装检查通过，深层检查仍会把它判为普通发行物文件缺失。已发布的 history trajectory-hygiene 示例在 core 的有效 Goal 上也会返回不可用。此外，带全局选项的 --version 在普通 wheel 可用，在 core 走到省略的 fallback。普通与 core 的边界应由已验证的安装身份和实际入口表示。

语义与 CI 对齐

三个独立 venv 分别安装精确基线普通 wheel、当前普通 wheel 和当前 core。普通 wheel 的 deep doctor 及历史子命令均通过。core 7 项安装验证、354 项源码测试和 9 项 TS 测试通过，3 项源码测试因平台跳过。未查询远端 CI。

我的整体评价 / Verdict

保留这条默认硬租约的独立执行与恢复增量。请让深度安装检查核验实际 core RECORD 和入口，并用安装后的完整命令示例验收目录。完成后再复审独立分发的发现与诊断。本次评审覆盖现有宿主与 CLI 行为。本地 macOS arm64 和隔离 Python 3.14/Node 下的合成任务。未查询远端 CI、未验证真实宿主采用、其他平台和完整 minimal/Standard 资格。

English verdict: REQUEST_CHANGES. The installed task and recovery work, but deep doctor rejects a healthy core wheel and the catalog advertises an unavailable history subcommand.

[cocolord]

Reviewer: model_agent · GPT-6 · OpenAI · self_reported

Approval conclusion (author-owned PR; GitHub blocks formal self-approval)

本次复审对应 6d4eba2846b62bee8cb9551808c137de1d414a13。没有尚未解决的阻断性发现；此前的 deep doctor、目录 action 和全局版本问题均已通过安装后的真实入口验证。完整 minimal 和发行计划仍保留各自验收。

动机 / Motivation

在新环境安装 core、执行任务并检查或恢复结果的操作者。旧安装到完成阶段缺少原租约凭据，健康 core 的深度诊断与目录示例也失败；当前安装能完成同一任务、恢复原结果并读回。隔离安装在 File 和 SQLite 下通过任务、独立产物验收、中断恢复及无重复执行或扣额的重放；诊断和已列出的历史操作可用。本次不发布版本、升级当前宿主或证明完整最小闭包、Standard 分发及模型收益。完整安装与 context 闭包、workspace/meta 发行、Standard 包、其他平台采用和完整成本仍待验收。

仅改 doctor 会留下完成与发现路径的缺口，继续只使用普通 wheel 也无法提供独立 core 安装。这次把实际安装、任务和恢复作为同一个可回滚交付边界，未新增需要用户同步的状态或重复权限确认。

改动思路 / Approach

复用既有决策与完成 owner，让同一产物贯通安装、执行、恢复和真实诊断；不存在第二套执行规则。仅交付独立 core 产物及其安装、发现和原结果恢复资格，完整最小闭包与发行分拆继续由现有 RFC 承接。构建器从单一 tracked 源码准备产物；canonical lease 仍是 TS owner 的真实状态，Python 只转交最初准入的 key/version。中断后从原 journal 继续，完成时沿既有 CAS 核验原 generation。

安装诊断复用实际发行物归属与 RECORD，完整 CLI 对普通 wheel 仍然必需。core 的 history action 沿已打包的 handler 执行，全局版本参数沿 canonical parser 解析。metadata、help 和 doctor 只表示可用性，不自动启用宿主、付费模型或其他能力。

具体改动 / Changes

关键代码讲解

• prepare_core_source 只从 tracked 源码准备外部空目录，排除 Chat、web、canary 与完整 facade，复用运行时文件；包含的内建能力与宿主导入继续记录为分发债务。构建器
• collect_python_distribution_checks 选择真实发行物的 console RECORD、导入和必要文件。健康 core 通过，复制到其他位置的启动器或缺失 history 文件仍失败；普通 wheel 缺少完整 CLI 不会被识别为健康 core。诊断
• _run_command 在 core 中使用现有 history handler；安装后的 history trajectory-hygiene 已实际执行。普通 facade 的旧绑定与兼容入口保留，全局版本及非法参数采用原 grammar。CLI
• projectExecutionPreparation 从匹配的 canonical lease 捕获原凭据，execute_turn_run_once 将已校验 capture 交给既有完成 owner。更新或释放租约后恢复仍被拒绝；native completion resolver 保持自身既有运输约定。凭据投影 完成运输

规范依据为 docs/architecture/rfcs/monorepo-distribution-split-v0.md，固定 revision d9432f23544e595f9861fcc6d8453ea926721612，以本次追加 checkpoint 之前的契约判断。V3 在已支持的安装 CLI corpus 上已实现并验证；R1 的完整 V1–V5/V10 闭包与 M3 的 workspace/meta 发行继续 deferred。这次只追加双语交付记录，未关闭 D1–D7 或另立 preset/default。证据不将安装产物等同于完整 minimal 资格。

对主干的风险 / Risks

关键风险是恢复借用新的租约 generation 或重复执行已确认工作。安装后的 File/SQLite 正反例覆盖 SIGKILL、原结果恢复、renew/release 拒绝、错误产物拒绝以及重复 resume；宿主调用数、canonical Todo 与扣额索引都独立读回。还测试健康 deep doctor、foreign console、缺失运行文件与实际目录 action，避免只验证 parser 或 helper。

普通基线和当前 wheel 的 8 组完整归一化 CLI/安装观察一致。两者都包含已验证的 Chat bundle，图形源码无差异。原普通基线的 File/SQLite 完成均报 lease_fence_required，当前普通安装的全部 10 项任务/恢复通过；这是明确修复的运输差异。完整 source CLI/preparation 327 项通过，architecture/Turn 2097 项通过；安装器/identity/release 31 项通过、3 项平台跳过。最终 core wheel 15 项通过，mypy 17 文件、control-plane typecheck 与 73 项 TS 用例通过。

语义与 CI 对齐

既有 vocabulary 与 lease key/version 被复用，无新增语义枚举或第二决策 owner。changed diff advisory、全树语义 drift、双语 gate、compile、差异检查和公共边界均通过。标准 premerge 的直接检查与其余 18 个 smoke 通过；examples/install-local-smoke.py 在固定 main 和本次 head 上都于同一 skill 文档链接越过包边界的断言失败，相关 smoke、installer、skill/link 源码未修改。保留该基线失败，安装与任务不依靠它来宣称通过。远端 CI 排队或进行中属于独立 readiness 观察，不声称已成功，也不取消已有门禁。

我的整体评价 / Verdict

本次是 justified_increment：真实安装和原任务完成/恢复有可验证增量，long_horizon 的原结果可恢复性与 user_experience 改善；这些证据不证明模型长程效果或完整成本收益。独立发行选择保持 opt-in，ordinary baseline 的 help、输出、参数与安装诊断维持配对兼容，共享原租约运输修复已明确披露。仅交付独立 core 产物及其安装、发现和原结果恢复资格，完整最小闭包与发行分拆继续由现有 RFC 承接。

本次只改变已安装 CLI 和任务运输；已检查 apps/presentation 与 Lark companion，图形页面没有差异。其他宿主或平台的已安装体验未在此轮验证。没有未解决的阻断性发现；基线安装 smoke 与远端 readiness 保持单独记录，合入仍要求 head 不变、原发现已验证解决及明确 maintainer 授权。

English verdict: APPROVE. Head 6d4eba2 repairs standalone diagnosis, advertised history execution, global version parsing and original hard-lease completion. Installed recovery/refusal and ordinary parity pass; full minimal qualification and pending CI remain separate.

Both original findings independently verified resolved at 6d4eba2. The final separately installed core passes all 15 cases: actual deep doctor, RECORD ownership and missing-file refusal; advertised history trajectory-hygiene execution; global version parsing; File/SQLite task/recovery and once-only replay. Ordinary installed parity passes. Exact-head approval review 5872232056 records complete evidence and remaining boundaries. Original review discussion retained.