[cocolord]

Goal And Delivered Outcome

Pure feedback and work-declaration reads should not load mutation or lease execution adapters. This proposal combines the existing deferred feedback adapters (#8339) with the declaration codecs, on current target main efd49352c. The equivalent anchored conversation clock repair from #8343 is already present on main and is preserved.

A real quota replan-view --replan-stage explore read also exposed a transport defect: a scalar-only staged view was compressed into a root replacement with an empty property path, which the Python consumer rejects. Nested records remain compact; staged roots now use the existing property edits. The original context, source hash and absent Explore basis read back successfully.

纯反馈和工作声明读取现在复用已有按需加载边界，保留原导出、写入顺序及 lease 栅栏。本提案在同一版本整合声明 codec 和 #8339 反馈组件，同时修复可空 Explore staged 读取的空 root path 编码缺陷。既有 source／number lineage 与嵌套压缩保持。

Author Declaration

• Written by: model_agent — GPT-6 family via TRAE.
• Specification and revision: monorepo-distribution-split-v0, existing diagnostic_projection.ts property-edit transport, and repository real-path validation policy.

Criterion	Disposition	Owning path	Evidence
Reuse typed owners and preserve declaration values/exports	implemented	work_items/write_scope_codec.ts, original acquire exports	Node 22 cold-loading/digest/source-identity checks and complete lifecycle readback
Preserve feedback adapter identity, patch points and write/refusal ordering	implemented	loopx/feedback.py	Existing unavailable-adapter, write/recovery and policy checks; original star exports retained
Preserve exact captured reads and numeric/source lineage	implemented	quota/diagnostic_projection.ts	Root property-path oracle fails before the fix; real File/SQLite captured Explore reads now preserve an absent basis and original hash
Complete source/wheel File/SQLite correctness and supported backend recovery	implemented for this runtime slice	existing CLI and authority owners	Four original 101-Todo lifecycle oracle sets and isolated PostgreSQL store/service suites
Whole-path benefit, optional distribution rollout and installed adoption	deferred	existing RFC owners	No performance or full-program completion claim; original qualification gaps remain

Scope And Continuation

This is a combined runtime proposal on one tested head, retaining the existing components and their historical evidence. Pure helpers stay separate from execution adapters; the new root-path correction stays in the existing TS transport owner. No new capability, provider, vocabulary, decision owner or authority cache is added.

The bounded refactoring pass applies the existing declaration-codec separation and limits the existing scalar compaction to its supported transport shape. Report subscription/construction loading remains the next known boundary. Component approval does not approve this combined head; protected merge still requires independent review/readiness and maintainer authority.

Validation

• Tested revision: b26cf572c08429d662d02ab7a71176f0420d6f00 (runtime source bytes match the checked source and installed wheel).
• Run state: finished locally; new-head CI and review remain pending.
• Input classes: synthetic, public_fixture.

Check kind	Result	Evidence / limitation
regression_parity	passed	Independent root-path test fails on parent; preserved source-number identity cases and real captured reads pass after correction
unit / integration	passed	37 feedback/replan cases plus 2 newly extended absent-Explore CLI cases; 34 Node 22.22.3 loading/digest/root/identity cases; 15 focused Node 26 cases
real_entrypoint	passed	Four new source/wheel × File/SQLite 101-Todo episodes retain original artifact, full scopes, completion time, four receipts, one debit, refusal/restart/replay and both managed native exits
real_backend	passed	Isolated PostgreSQL 16: 391 store and 11 service cases, no skips, separate disposable databases; this is source-suite coverage
real_entrypoint	passed	Original research-record-search on development and source/wheel packaged paths retains the negative paragraph, two original history rows, Goal isolation and exports; packaged wheel conversation-activity also passes
static / packaging	passed with limits	Installed ownership first; 1,921 present tracked package entries match bytes, including every changed runtime file. Seventeen tracked doc/media/testing or source-only entries are omitted; broader distribution coverage is not inferred
static / premerge	passed with inherited advisory	Standard diff-derived canary: 6 direct checks, 19 selected checks, no blocking failure/hold. One unchanged private_conversations.py maintainability finding remains an inherited advisory. Staged/head bilingual checks, public-boundary scan and production semantic advisory pass
integration	failed, retained	Initial source-driver preparation used python while its console selected python3, producing distinct runtime fingerprints and failing warmup/exit coverage. Subsequent episodes use each actual launcher interpreter; the failed episode and original semantics are retained
manual / CI	not_run on new head	Earlier CI research-record timeout did not reproduce locally. Its platform/timing cause remains unknown; successful local original assertions do not certify new-head CI

Corrected preparation failures (merge markers, missing browser binary, source/installed interpreter selection and Unix socket path length) remain in private evidence. Frozen 20-cost samples and attribution/default-bytecode pairs are neither rerun nor relabeled. Whole-cost benefit, model/attention, installed adoption, PostgreSQL host/wheel and the retained I3/minor-release, M1/#8011, full R0/V10/R1, D2/D5/D6/D3 and context/model/split-install qualifications remain open.

四条完整流程与真实 PG 校验通过，原研究记录负结果、历史和导出断言保持。此前 CI 超时本地未复现，原因仍未知；新 head 的 CI／评审继续待验收。保留初始解释器准备失败及全部冻结成本／PID 观察失败，不据模块数、测试数或组合提案宣称整体收益。

Frontend / Visual Evidence

• UI impact: none; no presentation or interaction code changed. Existing populated research and conversation journeys were exercised against the real synthetic history backend in development/packaged paths.
• First viewport and preview: N/A for unchanged presentation.

Type of Change

• Bug fix
• Refactoring
• Documentation update
• Test update

LoopX Area

• Control plane
• Build and packaging validation

Shared-authority RFC fixture impact

Synthetic 101-Todo full-read/lease/replay fixtures retain original semantics and artifact. File/SQLite source and wheel arms and isolated real PostgreSQL source suites pass. No authority promotion or runtime routing change is introduced; PostgreSQL host/wheel rollout remains unqualified.

Boundary Checklist

• Public code and aggregate synthetic evidence only; no private state, credentials, raw logs, local paths or internal links.
• No duplicate benchmark work or job launch.
• Existing RFC mirrors both changed and staged/head bilingual checks pass.
• UI impact is none.
• Every commit includes DCO sign-off.

[loopx-agent]

Reviewer: model_agent | gpt-6.1-sol | OpenAI | runtime_reported | reasoning_effort=max

动机 / Motivation

普通反馈校验、工作声明和精确捕获读取的调用者，是本次组合修复的使用者。旧路径把纯校验或声明与写入、租约执行耦合；可空 Explore staged 视图还会生成消费者拒绝的空根路径。新路径按需加载执行适配器，并只在嵌套对象使用整块压缩。当前组合版本独立通过加载/值差分、根路径反例及真实 File/SQLite staged 读取；原副作用 owner、权限和 receipt 语义保持。本 PR 不新增能力、持久协议、权限、默认开关或安装入口。完整 distribution/preset、R0/V10、模型实际采用和长期净收益仍由父 RFC 的原资格负责，本次局部修复不能结清。

改动思路 / Approach

旧提交 57b 的纯 codec 评审未发布；作者更新为当前组合版本后，我重新核对全部差异，没有继承组件批准。work_requirements 直接消费纯词法 owner，反馈模块只在明确访问写入符号或调用 append 时解析既有适配器。租约、写入栅栏、锁、journal/receipt 和恢复仍归现有执行 owner。既有词法规则和反馈适配器复用原实现，根节点修复留在原 transport owner，不新增状态、执行器或并行决策。

最强反例是“普通读已不加载执行模块，但写入前置检查或精确捕获读取被破坏”。当前独立进程故障探针验证未选择租约 owner 时 ping 可用，选择后两次准确失败，后续 ping 可用；反馈缺 writer 在 registry 读取前拒绝。根路径测试按既有 Python 消费者的非空属性路径要求构造，基线失败、当前通过；真实 File/SQLite 的 Explore 读取保留原 source hash 和 null basis。当前边界是纯调用路径的依赖隔离及既有 staged 读取修复，不以模块数或组件批准代替组合验收。

具体改动 / Changes

关键代码讲解

准确 head b26cf572c08429d662d02ab7a71176f0420d6f00，真实 merge base efd49352cc087cb299d5f7194cefd95a64b83e95，完整范围 12 文件 +496/-108。packet 的 key_files 只列关键文件，完整 Git 差异还包括 digest guard 和新增 root oracle，均已检查。相对旧 57b，四个 codec/调用模块及 npm manifests 字节不变；反馈模块与先前组件 6d1 字节相同，但仍重跑当前组合的相关验证。

• write_scope_codec.ts:87 保留请求原拼写、首次出现顺序、coercion、字符类与 160 字符限制。task_lease_acquire.ts:214 的严格解码仍先拒绝非法成员，不能用过滤代替授权。原模块保留相同函数对象；没有第二份规则。
• todos/work_requirements.ts:93 仅改变词法导入，省略、null、空数组和非法成员拒绝保持；Goal 根路径投影不授予写权限。
• feedback.py 的 __getattr__ 与 _mutation_adapter 保留原符号对象、star exports 和可 patch 的模块 binding。append_human_reward 先解析两个 adapter，再读 registry 或追加 index；实际写入处仍重新读取 binding，原 TS 栅栏及 index/state 次序不变。模块存在或 dir 可见均不等于写入已获授权。
• quota/diagnostic_projection.ts:75 的 contextProjectionChanges 在标量对象压缩分支增加 path.length > 0。根对象继续用属性编辑，符合 turn_envelope.py:_apply_context_projection_changes 的非空路径契约；嵌套压缩、数字/source lineage 沿原路径。新增 root oracle 验证可重建视图且不修改 source；File/SQLite CLI 追加的 Explore 断言核实 basis 为 null、source hash 不变。
• digest guard 仍为既有 action wildcard 构造声明精确数量 1；可见 64-hex matcher 反测仍失败。双语 RFC 只追加非规范证据记录，保留失败、冻结成本和安装采用边界，没有改写规范以迎合实现。

规范依据：docs/architecture/rfcs/monorepo-distribution-split-v0.md，固定修订 efd49352cc087cb299d5f7194cefd95a64b83e95。I2 单 owner、I6 无顺带 schema/私密边界变化、I7 未选择不激活与选择失败闭合，在受影响切片 implemented；I8 以精确来源/数字测试与真实 staged 读取核验本次 transport 修复，不能推成全部模型采用。V1 绑定冷进程/故障注入，V2 绑定实际 File/SQLite 租约及反馈/重规划检查，V3 绑定值、拒绝、导出和 CLI 读回；这些只验收受影响路径。I3 公共 Python 模块搬迁 out_of_scope，旧模块仍是原 owner；原导出绑定保留。I9 的收益边界保留，完整 V10、preset/安装及实际模型采用 deferred 给原 RFC 资格。

当前 Node 24.21.0/macOS：加载/digest/root/source identity 34 PASS，完整类型检查 PASS；新 base/head 的 138 组完整值及错误结果一致。将新增 root oracle 对准未修改基线，正确 FAIL 于空属性路径；当前 PASS。当前反馈与真实 replan/lease/requirements 合并组 51 PASS、2 FAIL，相邻反馈 84 PASS。两项失败都是 selected-Todo 旧测试在第 407 行读取缺失的 selected_todo，不是 Explore 断言失败；同一 16 项命令在不可变 base 与当前 head 的完整详情分别核对并保留，下文没有将其改写为通过。

当前标准 premerge 六项直接检查、19 项选择检查全部 PASS，零失败、跳过或 manual hold。先前 57b 的结果、基线失败以及 root oracle 首次因基线没有新测试文件而未执行的准备结果均保留；后者已改用同一新 oracle 的独立私有测试副本执行，实际基线断言失败，不把“找不到文件”算作回归检出。作者的 PostgreSQL、wheel、最低 Node 和完整 browser 声明未在本轮独立复验，也未查询远端 CI。

对主干的风险 / Risks

未发现当前组合 head 的阻断问题。词法规则搬迁的主要风险是非法 scope 被丢弃后意外扩大授权；138 组差分、原严格调用分支及真实租约 replay/renew/release/complete/readback 支持保持。反馈的风险是纯路径解耦后发生半写入或错过 writer patch；现有 mutation preflight、原 binding 查找、当前 adapter 和相邻反馈测试验证了该边界。根路径修复只限制既有优化的适用位置，不增加 transport 类型或策略分支。

语义与 CI 对齐

复用既有词汇、owner 和属性编辑协议，没有新增状态分类、默认启用或 hidden obligation；原错误与恢复文字没有压缩。两项 selected-Todo KeyError 为同条件基线失败，保留完整身份、详情及环境/夹具证据；不删除断言、不放宽检查。Windows、最低 Node、独立 PostgreSQL/wheel、本机安装采用、模型及完整长期成本仍未验证。主干整合、CQR 和 merge readiness 分别读回；当前通过不代表获准合并。

我的整体评价 / Verdict

APPROVE。 这是 justified_increment：纯调用路径的隔离与可空 staged 读取已通过当前组合验证，父 RFC 的完整资格保持开放。long_horizon 为 preserved，真实 lease/replan 的重放、完成、终态拒绝、原 source hash 与后续 staged 读取保持；user_experience 为 improved，正常纯调用避免不相关适配器缺失的失败，Explore 读取恢复且没有新增确认轮次。视觉检查为 not_applicable，因为没有渲染页面变化；未验证范围为 Windows、最低 Node、独立 PostgreSQL/wheel、安装、模型采用和完整成本。既有词法规则和反馈适配器复用原实现，根节点修复留在原 transport owner，不新增状态、执行器或并行决策。当前边界是纯调用路径的依赖隔离及既有 staged 读取修复，不以模块数或组件批准代替组合验收。未合并。

English verdict: APPROVE — head b26cf57 preserves the selected owners and affected File/SQLite behavior while fixing nullable staged-root transport; inherited failures and broader qualification remain explicit.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @cocolord.

Choose the remote for the base repository, not an out-of-date fork.
For a fork clone, first inspect git remote -v; upstream must point
to https://github.com/loopx-project/loopx.git. If it is absent, add it
with git remote add upstream https://github.com/loopx-project/loopx.git.
Then run:

git fetch upstream
git rebase upstream/main
# Resolve each conflict, git add the resolved files, then git rebase --continue.
git push --force-with-lease origin HEAD

For a same-repository clone whose origin points to
https://github.com/loopx-project/loopx.git, use origin instead of
upstream for fetch/rebase. If you prefer merging the base, use
git merge <base-remote>/main and push normally.

Keep the DCO Signed-off-by trailer on every commit when you rebase.
https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork

[loopx-agent]

Self-review of exact head 6ca929ad885b0cda4edb9b2ac7409ed7a7d61fae: PASS.

The latest-main merge completed automatically. Main already contains the feedback-adapter loading changes, so the final PR delta is now 10 files: declaration codec separation, lease consumers, the scalar staged-root transport correction, focused tests, and bilingual RFC evidence.

Exact-head verification passed: 119 Node tests with 1 expected skip; full control-plane TypeScript typecheck; 100 Python feedback/replan/requirements/lease cases. Two selected-Todo cases fail identically on immutable main at the same missing selected_todo assertion, and are explicitly retained as inherited baseline failures. Quick premerge passed all direct checks, semantic-vocabulary, task-lease runtime and public-boundary checks.

[huangruiteng]

Independent approval of exact head 6ca929ad885b0cda4edb9b2ac7409ed7a7d61fae.

The latest-main result preserves the validated deferred loading and declaration-codec boundary while correcting scalar staged-root transport. Node, TypeScript, focused Python and quick premerge checks passed. The two selected-Todo failures reproduce unchanged on immutable main and remain explicitly identified rather than waived.