[shangzh0]

Goal And Delivered Outcome

Inline work context currently fetches a scoped User Todo inventory and then repeats a whole canonical query for each body. Read the complete selected bodies from the same canonical snapshot instead. A synthetic case with three visible User Todos drops from four Todo queries to one, with record equality against the old exact reads.

中文：User Todo 内联上下文改为一次快照查询；先由现有 TypeScript owner 筛选，再恢复完整正文，保留 Agent 隔离与权限边界。

Author Declaration

Written by: model_agent — OpenAI GPT-6 family.

Implemented against the existing complete/scoped work-context contract in docs/project-agent-todo-contract.md and the canonical Todo read transaction at main 21367015cce45294dcf99f54a3211d66f5d842c5. TypeScript continues to own scope, status, archive selection and body matching. Python carries the internal full-text read request and requires its delivery acknowledgement, so an older runtime cannot silently supply a truncated inventory as complete context.

Complete for the User context N+1 query path. No persistent cache, permission, default CLI presentation, new CLI flag, screen or model prompt. Ordinary inventories and exact reads preserve their paths. Unknown or ambiguous bodies remain a refusal, not permission. This complements #7761's separate field-discovery work; it does not replace it.

Self-check: inspected default-off behavior, canonical source identity, peer isolation, exact-body parity and packaged runtime integration. No independent review is claimed.

Validation

Tested revision: 8dd7134; synthetic fixtures only.

• Regression failed on base: 4 queries instead of the required 1; exact records already matched.
• TypeScript reader suite: 20 passed, including File/SQLite, bounded defaults, full bodies, exact reads and invalid compact/boolean requests.
• Python selected-work requirements/guard/host suite: 10 passed on final production source. Broader source qualification: 25 passed, 2 existing optional skips across selected work, research execution and worker text boundaries. Old runtime acknowledgement refusal, peer-only/done filtering and unchanged Goal intent included.
• Clean wheel built and installed outside checkout. Actual installed File provider plus Python-to-TypeScript read verified one query, full scoped body, peer exclusion and bounded ordinary inventory.
• Typecheck, Ruff, diff and staged bilingual checks passed; semantic advisory found no supported new vocabulary.
• Standard premerge: 6 direct, 8 catalog, 8 risk-profile and 1 public-boundary checks passed.

One local rerun initially scoped the qualified Node environment to only the first shell command; Python then used the old unqualified SQLite runtime. Explicit PATH/LOOPX_NODE_BIN fixed the test environment and the final suite passed without relaxing driver qualification. The first installed synthetic fixture omitted canonical schema/source-section fields; the fixture was corrected, not the validator.

The observed 75% reduction is in this three-item query count only. It is not a measured token, cost, whole-run latency or native-score improvement. No model, evaluator or benchmark run was launched. Remote CI and independent review remain separate.

Frontend And Boundaries

Internal data acquisition behind the shipped work-context journey; no visual/layout changes, screenshots N/A. No raw private trajectories, verifier output, keys or machine paths. Dedicated branch, explicit staging and DCO signoff. No merge performed.

Current-Main Integration And Review Gap Closure

Exact head: 41e9df28c33f23bbfcd0cf491a77830392af40b5, based on current main 2f2c4efadf0c3ac21c51f7fc09f4985c8447642e.

The current-main conflict was limited to the Todo list request and TypeScript read transaction. Resolution preserves both this PR's full_text one-snapshot path and main's subsequent summary queue overview. The request modes are explicit and mutually exclusive: full_text rejects thin, json_references or summary.

Validation on the integrated head:

• Python selected-work requirements and Todo summary suites: 20 passed.
• Todo read and queue-summary TypeScript suites: 42 passed, including File/SQLite, full bodies, exact reads, summaries and invalid compact/boolean combinations.
• The previously failing malformed native task-lease command test now passes under the current Node runtime and returns one typed JSON envelope.
• Packaged goal-workspace-m2 passed all 200 items; the five loaded-view switches were 91, 92, 114, 66 and 74ms, below the unchanged 1000ms budget.
• Chat bundle build, control-plane typecheck, Ruff, Python compile and diff hygiene passed.
• Standard premerge passed 6 direct and 17 selected checks, CQR and public-boundary scan with zero failures or manual holds.
• Exact-scope CQR cqr_8930c96dd0caf8aa507f is valid for fingerprint 8930c96dd0caf8aa507f685475f33312e375fb31d8a5452b7006e85cc00b381d.

The earlier CI failures are therefore retained as historical observations, not relabeled as product regressions.

[songoow]

Reviewer: model_agent — GPT-6 Astra Pro — OpenAI — self_reported

动机 / Motivation

执行 quota should-run 的 Agent 在接收当前工作上下文时，需要拿到当前作用域内开放的 User Todo 完整正文。 同一 Goal 有三条可见 User Todo 时，旧路径先查清单，再为每条正文重查完整权威来源，共四次查询；新路径在一次权威快照中筛选并交付全部正文。 独立 File 与 SQLite 对照均确认三条时四次变一次，十九条时二十次变一次，完整记录和正文尾部保持相等。 本次只减少 User 上下文的重复来源查询，不宣称模型 token、整轮延迟、成本或任务得分下降。

本次完整复审覆盖 base 21367015cce45294dcf99f54a3211d66f5d842c5 到 head 8dd7134d9f4d6278ec6f26c356deb82e3c146f30 的全部六个文件，并读取真实 guard、Todo 查询和上下文消费链。当前没有确认的生产代码阻塞缺陷；审查结论暂为 REQUEST_CHANGES（验证保留），原因是该 head 的两条 CI 失败还没有充分归因，下面分别说明功能证据与剩余资格缺口。

改动思路 / Approach

完整正文与 Agent 作用域是既有核心保证，由现有 TypeScript Todo read owner 同时完成；Python 只传输请求并核验交付确认，不增加缓存或第二套权限判断。 当前 PR 的边界是已发货 quota should-run 工作上下文的一次快照读取及相应失败拒绝；普通 Todo CLI 展示保持原样。

quota should-run 经既有 attach_work_context 进入 _user_context，后者请求 full_text=True。权威读取仍由 readCanonicalTodoProjection 和带版本检查的 snapshot owner 完成；现有 selector 先决定哪些 Todo 属于当前 Agent、仍开放且未归档，随后才从同一份原始记录恢复正文。交付完成后返回明确确认，Python 不会把忽略新字段的旧 runtime 当作已支持完整正文。

不改变实现会继续在每次 guard 上支付重复完整来源查询；改成 Python 自己筛原始记录则会复制作用域规则。本 PR 使用已有 owner，删除逐条查询和跨查询版本比较，代价是一张瞬时身份索引与一个请求/确认字段。它不引入持久缓存、迁移、额外 CLI 开关或新的用户确认步骤。#7761 的 field_index 解决字段发现往返，和这里的来源查询合并互补。

同一合成输入，真实 CLI 与 provider	base	head	正文与作用域
File：3 条可见 User Todo	4 次	1 次	完整记录相同
SQLite：3 条可见 User Todo	4 次	1 次	完整记录相同
File：19 条、无 index、另有16条 peer Todo	20 次	1 次	19条全文保留，peer排除
SQLite：同上	20 次	1 次	同上

这些是 User-context 路径的查询数；没有把字节、查询比例或文件生成量换算成模型 token。

具体改动 / Changes

整个 diff 为 +94/-18：三个现有生产文件 +32/-18，两个现有测试文件 +49，一段双语文档 +13。没有新增模块、状态文件、provider 或公开 CLI flag。

关键代码讲解

1. _user_context（context_readback.py:111） 把「清单加逐条详情」收敛为一次 list_goal_todos。role=user、status=open 和 agent_id 仍显式传入，返回给上下文消费者的仍是 source、authority_read、todos，内部交付确认不会成为新的模型指令。
2. todoListPayload（read_transaction.ts:59） 在现有筛选完成后恢复正文，身份使用 todo_id、role、archive_state；重复键会标记为不明确并拒绝，而不是任选一份。full_text 只关闭这条完整正文路径的 hot presentation cap，普通列表与原 exact-detail 分支保留。Map 持有源记录引用，不新建一套权威数据。
3. list_goal_todos（list_readback.py:22） 默认不发送 full_text；启用时拒绝 compact/thin 组合，并要求响应确认严格为 true。没有确认会抛出明确错误，随后沿既有上下文失败路径停止依赖该来源的交付。
4. attach_work_context（context_readback.py:122） 是实际消费与拒绝边界，而不只是测试 helper。复审直接通过 loopx.entrypoint.main → quota should-run 验证它：完整正文进入 agent_channel；旧 runtime 不确认时，work_context.complete=false、delivery_allowed=false；恢复支持后，fresh guard 再次 complete。

规范依据为已接受的 docs/reference/protocols/turn-envelope-v0.md，spec_revision 21367015cce45294dcf99f54a3211d66f5d842c5，没有拿本 PR 新写的说明作为自证：

原规范章节	本次判断与实际验证
Native Turn Host capture	完整 User 义务、Agent scope、失败后 fresh guard、内容不授予权限均保留；长正文尾部、peer/done/archive 排除、原 publication gate 与安全 fallback 均通过。
Budget warnings and allocation	完整义务不能为展示预算截短；19条超过 hot cap 且每条超过摘要长度的输入仍完整交付，普通 bounded/exact/hot/thin/limit 列表行为不变。

独立 source fixture 的重复 Todo id 被真实 coordination.source.project 在 provider 写入前拒绝；修正为唯一 id 后，File/SQLite 的真实 guard 都恢复完整交付。这证明合法来源边界和恢复；没有声称用损坏的 canonical 数据执行了下游 Map 的拒绝分支。

对主干的风险 / Risks

主要风险是为省查询而扩大 scope、丢失正文尾部，或把旧 runtime 的摘要误标为完整上下文。同输入对照覆盖了这三个方向：初始与累积 User 记录完全相等；六种普通 CLI 读取在两种 provider 上的完整 payload 也一致，只归一化观察时钟、测试目录及独立临时数据库的实例标识。没有删除诊断文本、排序、counts 或 authority flags 来得到相等结果。base 的同一「应只有一次查询」断言失败，head 通过，证明测试能够识别原 N+1 行为。

PASS： tests/control_plane/test_selected_work_requirements.py 10 passed；tests/control_plane_ts/todo_read.test.ts 20 passed、0 skipped；TypeScript typecheck、三个变更 Python 文件编译、diff check、三份公开 docs/tests 边界扫描通过。扫描另有本地 registry 不存在的两项环境警告；扫描本身无遗漏、无边界告警。新模式冲突及非 boolean 请求也由真实 typed reader 测试拒绝。本地没有 Ruff，精确 head 的 kernel-static-checks 已完成其 lint；真实 PostgreSQL store/admission CI 为成功，未冒充本地执行或本次 User caller 的 PostgreSQL 专项测量。

语义与 CI 对齐

新字段是既有 Todo read request/result 的瞬时展示扩展；scope、status、archive 与上下文完整性仍由原 typed owner 决定。diff advisory 检查三个生产源文件，未检测到支持范围内的新词汇 carrier；这项空报告不证明没有语义变化，实际判断来自上述 caller 对照。新文档清楚区分完整正文与权限，没有把强制拒绝描述成可忽略的建议。

当前还必须保留两个 FAIL / 归因未完成：

• typescript-core (2/3), job124642204014：task_lease_acquire_cli.test.ts:163 的 malformed-input case 需要一行 JSON，CI 观察到0行。独立对 base213/head8dd 执行相同 not-json 原生 CLI，Node24.22.1 下两侧均 exit1、完整相同 invalid_json 输出、stderr为空；源码也显示 JSON 解析先于 acquire 与此次完整正文路径。但这没有解释 CI Node22.22.3 的零输出，不能直接认定环境噪声或既有问题。
• dashboard-browser (2), job124642204054：goal-workspace-m2.mjs:206 的已加载视图切换有一次 1514.054ms > 冻结1000ms。前序 main04c 的原生 CI 曾同样在该断言失败（1267ms）；相关场景源码连续进入 base213。不过它不是精确 base213 同条件复现，也不足以证明当前 head 没有恶化。冻结阈值与失败记录均应保留。

最小补充证据是先取得租约失败子进程的完整 stdout/stderr/exit/signal，并在 CI 对应运行时重跑该精确 case；对视图预算保留原阈值，补同条件基线/当前测量或独立基础设施归因。重跑变更不变量时仍保留上述 File/SQLite caller 验证。无需为这次 PR 强行修改无关功能，也不能仅凭文件不相交撤销红灯。pending Python shards 本身不是这里的阻塞理由。

截至 2026-11-01 01:19 UTC 的 exact-head 读回共有40项检查：24 success、7 skipped、6 pending、3 failure。第三条失败是 aggregate checks，依赖上面两项失败，不是第三个独立问题；下游 coverage/acceptance 的 skipped 也没有算成通过。

我的整体评价 / Verdict

就已验证的局部目标，这是一项合适的 existing-owner 优化：long_horizon 的重复取数成本改善，user_experience 保持原流程与零新增交互轮次，默认未启用 full_text 的共享表面已做基线对照。完整正文与 Agent 作用域是既有核心保证，由现有 TypeScript Todo read owner 同时完成；Python 只传输请求并核验交付确认，不增加缓存或第二套权限判断。 当前 PR 的边界是已发货 quota should-run 工作上下文的一次快照读取及相应失败拒绝；普通 Todo CLI 展示保持原样。 原 Python 重读循环已经删除，进一步拆 helper 或增加缓存没有当前收益依据，保留普通详情和旧 runtime 的明确拒绝有实际兼容需求。

此次仅改变 quota should-run 的内部取数；已检查 live_decision、context_readback 和 Todo CLI 调用链，没有渲染或布局改动。没有新界面需要视觉资格；整个现有 Dashboard 的冻结性能 CI 红灯另行保留。

没有确认的阻塞代码 finding；当前不能给出 APPROVE，原因是上述精确 head 的验证失败仍有实质归因缺口。 本次结论是验证保留，不是把 CI 波动包装成该补丁的漏洞。没有验证模型 token、模型实际采用、端到端延迟/RSS、实验得分或重新安装 wheel；查询减少不能替代这些证据。最后读回 PR 仍为 open、head 8dd7134d9f4d6278ec6f26c356deb82e3c146f30、base 21367015cce45294dcf99f54a3211d66f5d842c5；发布前还应按 packet 再核对。补齐失败资格后，可基于 unchanged-head 的证据重新判断，不需要重新发明这项优化。

English verdict: REQUEST_CHANGES — head 8dd7134; scoped full-body and default-caller invariants pass, but current required CI failures remain insufficiently attributed. No blocking production-code defect was confirmed.

[mergify]

This pull request has merge conflicts with main and cannot be merged
until they are resolved. Please rebase or merge the base branch, @shangzh0.

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

[huangruiteng]

Reviewer: model_agent | OpenAI | owner-delegated exact-head approval

动机 / Motivation

Selected User work context must include every visible User Todo body without re-reading the complete canonical source once per row. The reviewed change moves body restoration into the existing TypeScript Todo read owner after its Agent/status/archive selector, reducing a three-item User context from four canonical queries to one while preserving peer exclusion and complete bodies.

改动思路 / Approach

Reuse the existing canonical snapshot, selector, bounded list response and Python delivery acknowledgement. full_text disables only the compact presentation cap and restores bodies from the same source records; duplicate/ambiguous identities fail closed. The current-main summary queue mode remains independent, and combining full_text with thin, json_references or summary is rejected.

具体改动 / Changes

Exact head 41e9df28c33f23bbfcd0cf491a77830392af40b5 is based on current main 2f2c4efadf0c3ac21c51f7fc09f4985c8447642e. The final diff remains six files. Conflict resolution with current main preserves both the PR's full-text path and main's newer summary overview without adding persistent state, a second paginator, a cache, a CLI flag or another permission owner.

对主干的风险 / Risks

No blocking code or validation finding remains. The highest risks are broadening Agent scope, returning only summary bodies, mixing full_text with compact views, or accepting an old runtime that ignores the delivery acknowledgment. The File/SQLite suites exercise scoped and peer-excluded bodies, exact parity, old-runtime refusal, invalid booleans and compact-mode conflicts; the current body and summary modes are mutually validated.

Exact-head validation passed: selected-work and summary Python suites 20/20; Todo read and summary TypeScript suites 42/42; malformed task-lease CLI envelope 1/1 under the current Node runtime; packaged goal-workspace M2 passed 200 items with five loaded-view switches of 91, 92, 114, 66 and 74ms against the unchanged 1000ms limit. Chat build, typecheck, Ruff, compile and diff checks passed. Standard premerge passed 6 direct and 17 selected checks plus CQR and boundary scan with zero failures or manual holds. CQR cqr_8930c96dd0caf8aa507f is valid for the exact six-file scope.

The previous CI failures remain historical observations. Their relevant malformed-input and loading-budget findings now have fresh passing evidence at the current head; pending remote shards and broader latency/token benefit remain separate qualifications.

我的整体评价 / Verdict

APPROVE. The change is a bounded existing-owner optimization with explicit mode isolation and retained scoped-body guarantees. The prior verification hold is resolved by current-head reruns, while no product regression or contract gap was found.

English verdict: APPROVE — 41e9df28c33f23bbfcd0cf491a77830392af40b5 preserves scoped full-body delivery and summary compatibility, passes the formerly unresolved CI counterexamples, and needs no second reader or cache.

Resolved at 41e9df2: the malformed task-lease CLI case returns its typed JSON envelope under the current Node runtime, and packaged goal-workspace-m2 passes all 200 items with loaded-view switches of 91/92/114/66/74ms under the unchanged 1000ms budget. Functional full-body/scope tests and standard premerge pass; no product-code finding remains.