[loopx-agent]

Goal And Delivered Outcome

A recently closed Codex Session could precede an older usable Session in the same Goal channel. The UI kept displaying the older transcript while binding current controls to the closed row: model settings disappeared and LoopX mode reported conversation unavailable.

Current controls now select the newest non-closed Session for the selected executor. Closed transcripts remain readable, and resume_failed remains visible. Goal discovery reuses the same selector. Related to #6934 and #6937; intended base: main.

Author Declaration

• Written by: model_agent — OpenAI Codex, GPT-6 family.
• Specification and revision: reproduced recovery defect described above; existing Session identity contract in docs/integrations/session-runtime-control-plane-adapter.md at ef77d906.

Criterion	Disposition	Path	Validation
Closed history cannot own current controls	implemented	src/data/use-conversation-history.ts	closed-first / closed-only browser cases
Keep historical messages and visible recovery failures	implemented	same history owner	retained transcript and failed-current cases
Use the same rule for Goal discovery	implemented	src/views/dashboard-page.tsx	TypeScript/build and packaged navigation

Self-check: reviewed all seven changed paths, repository instructions, existing history/recovery consumers and actual packaged behavior. No server lifecycle, model dispatch, profile permissions or account configuration changes.

Scope And Continuation

Complete for this selection defect. This is a frontend consumer fix in the existing native Chat capability. The broader desktop activation and installed-account recovery work remains separate. Bounded refactoring applied: reuse the existing selector in discovery through a sessions-only input type; no additional state or abstraction.

Validation

• Tested revision: 42fb84defdc4a09802b44685027169bd44f2baac for the production source and packaged Core/frontend; final head dae52135a46dd2cbf706191528b42c9ccb8e81fd adds only the after/mobile screenshots.
• Run state: finished
• Input classes: synthetic, authorized_private_read_only

Check kind	Result	Evidence / limitation
static	passed	pnpm run build: dashboard TypeScript, Vite and verified Chat bundle.
regression_parity	passed	node smoke/conversation-history-order-browser-smoke.mjs: all 11 scenarios, including the original eight history races, closed-first, closed-only and failed-current.
unit	passed	pnpm run test:conversation-returns: source isolation, late return, deduplication and transport uncertainty.
real_backend	passed	Disposable native profiles with real Codex app-server and actual Chat HTTP handler: original Session/thread/home recovery, coexistence and backend reopen; seven refusal cases without admitting a Turn. The native profile owner is unchanged by this PR.
real_entrypoint	passed	Debug packaged native App with the exact committed Core/frontend: retained history, accurate model/effort, editable unsent composer, switching profiles, normal Quit/reopen. No model inference or business-message replay. This is QA, not formal installed-App adoption.
static	passed	Final-head standard canary premerge: four direct checks, five catalog checks, eight risk-profile smokes and public-boundary scan. Exact-diff change-quality receipt recorded and verified.
manual	passed	Desktop and 390px views; no horizontal overflow. All public images use synthetic data. Original owner-local Session content was inspected read-only and stayed unchanged; no private details are published.
integration	not_run	Remote PR CI has not been claimed green. Production signed OTA activation and original-account execution are outside this frontend fix.

Frontend / Visual Evidence

• UI impact: changed
• Source data: synthetic
• Boundary: browser first-viewport captures of the packaged Core/frontend used by the debug native App. Before: Core/frontend 04c147c6. After: Core/frontend 42fb84def; final head adds images only. Native QA binary is unchanged. No public preview deployment; local QA services are stopped after validation.
• Attention review: removes the false availability error and restores existing model and runtime controls, with no extra labels or controls. Closed history remains in the transcript; actual recovery failures remain visible.
• Craft review: existing layout, tokens and state treatments retained; desktop and mobile capture the changed conversation state.

Before:

After:

Mobile:

Type of Change

• Bug fix
• Test update

LoopX Area

• Public docs or presentation surface (dashboard)
• Host or runtime integration

Technical Direction

Native conversation recovery; no new roadmap claim.

Shared-authority RFC fixture impact

N/A: no authority backend, routing or promotion changes.

Boundary Checklist

• No private state, credentials, raw traces, internal links or local machine paths in the diff or public evidence.
• No duplicated benchmark execution.
• Scoped to the reproduced conversation recovery defect.
• No paired RFC changes; bilingual diff check passed.
• First viewport evidence is embedded above.
• Every commit has a DCO sign-off.

[loopx-agent]

Reviewer: model_agent; model=gpt-6.1-sol; provider=OpenAI; declaration_source=runtime_reported; reasoning_effort=max

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

APPROVE — dae52135a46dd2cbf706191528b42c9ccb8e81fd。完整七路径评审通过，未发现阻塞问题。代码结论、合并资格及合并权限分别核验。

动机

用户回到一个已有 Goal 时，希望读到原对话、看到它的实际模型，并继续在原上下文中工作。较新的关闭会话原先会遮蔽较早的可用会话：正文仍显示旧历史，当前控制却绑定关闭行，模型消失，LoopX 模式出现 conversation unavailable。不修复会增加人工排查和重开会话的成本。

当前版本在既有选择器中排除 closed，保留选定执行器的第一个非关闭会话；关闭历史仍进入正文，resume_failed 仍明确可见。隔离的真实 Chat 存储和 HTTP 读取证实关闭行先于原 ready 行；实际构建的界面恢复原模型、两段历史和可编辑的未发送草稿，390px 下没有横向溢出。原两段消息逐项不变，未产生 Turn 或模型调用。该检查使用合成 adapter，不代表正式安装或原账号执行已验收。

改动思路

修复现有消费者即可解决这个选择错误。只改排序会保留错误的关闭行；另设可用标记或新会话管理器会复制权威。现在 Goal 发现也复用同一个选择器，避免概览与对话采用不同规则。

Session 是原生对话及其执行身份，历史消息与当前可操作身份是两个读取用途。排除关闭行只影响当前控制，不删除其历史，不更换账号、原生目录、线程或已选择的执行器。服务端仍拥有恢复、创建会话和发送准入；前端一次观察不授予执行权。普通路径无需增加用户确认或配置步骤。

具体改动

依据为 docs/integrations/session-runtime-control-plane-adapter.md，固定规范版本 ef77d9064607dfc69edd3f21926dc4c22126c44b。以下标识映射原文未编号条款：original-session-identity 保留原 Session/线程/配置身份；observed-model-readback 使用实际 Session 模型而非目录猜测；history-read-recovery 保留历史及不可读状态，原当前来源未读到时仍禁止发送；separate-activation 保持普通对话、显式模型选择与 Agent Loop 启动的既有边界。前三项由选择器、现有历史 owner 和真实 HTTP/构建界面验证；最后一项由代码与无 Turn 读回验证。正式 OTA/安装、真实模型调用和完整跨渠道工作不在本次选择缺陷的交付范围内，仍由既有恢复工作承担。

七个路径共 +30/-8：两个生产文件、两个既有回归文件和三张合成截图。新增量主要是三个关闭/失败场景，没有新模块、配置、协议字段或持久化状态。已逐张检查 PR 正文中的前后及窄屏首屏，画面符合声明的合成数据边界。

关键代码讲解

1. currentChannelSession 在服务端原有顺序上，按精确 agent_id 和 status !== "closed" 选择当前行。它不重排、不改写输入，也不跳过恢复失败；sessions-only 类型让发现入口直接复用。实际基线/当前函数对相同 20 组完整输入执行，16 组完整输出相同，4 组关闭会话变化符合预期。
2. useConversationHistory 继续保管 scope、快照、重试及发送门。当前来源必须有对应快照；较早历史读失败不能重启当前流。它使用新的选择结果决定当前身份，同时保留完整历史列表，因而选取与正文不会互相覆盖。
3. PersonalGoalHome 的发现分支在 L2087 复用该选择器；原恢复分支也调用同一 owner，再交既有 createChatSession(..., "resume_latest") 路径处理。发现不再直接取列表第一行。实际构建界面从 Goal 进入对话，原模型、历史和草稿均可读。

对主干的风险

最强反例是关闭行后面虽有旧 ready 行，但真正最新的非关闭行是 resume_failed，或者所选执行器的当前快照读失败。把这些也当历史略过，会掩盖真实失败并可能创建错误替代会话。20 组对照保留 failed/busy/未知状态、不可恢复标记、执行器隔离及原输入顺序；只剩关闭行时返回空当前身份。11 个完整浏览器场景覆盖这三个新增分支及原有八种迟到、重试、刷新、切换竞争。

本轮独立验证：TypeScript/Vite 与 Chat bundle 构建通过；串行历史浏览器场景 11/11；会话返回检查通过；真实 HTTP/store 历史恢复、真实 HTTP/SQLite/Core 原会话绑定检查通过；19 个原生目录/HTTP 绑定测试通过；模型菜单 64 项双语/可访问断言通过（SSR）。干净精确七路径 premerge 为 4 项直接检查、5 项 catalog、8 项风险检查及 1 项公私边界检查通过，无 manual hold。未查询远端 CI。

保留早先失败：首次与构建并行的开发服务器检查中，两个场景发生 Vite 浮层/等待超时，其余九项通过；没有删除断言或放宽超时。构建结束后完整串行重跑、独立函数对照、真实 HTTP 边界及实际打包界面补验所改不变量。开发服务器首次启动的间歇原因未宣称已修复。私有 UI 夹具还曾因模式响应漏字段进入页面兜底；修正夹具响应后在原浏览器任务重载完成检查，产品源码未因此修改。后续通过不会改写这些历史失败。

隔离浏览器会沿既有入口恢复会话，并可能更新会话元数据；整个浏览器过程不称为纯只读。原消息保持不变，业务发送为零、Turn 为零。模型/native adapter 与若干无关工作区投影为合成输入，部分无关服务故意不可用；不据此声称全 App、正式安装、真实账号或 live Lark 资格。主干较新的 completed-history 后端整理未改本次客户端选择 owner；完整合并组合和安装采用仍是后续独立边界。

我的整体评价

本次选择缺陷达到 goal_achieved。long_horizon 为 improved，限于保留可继续使用的原会话身份；user_experience 为 improved，限于从原 Goal 直接读回模型和历史、编辑草稿，无新增必需交互。两者都不推断模型完成了更多工作或总成本收益。

在既有选择器中修复条件，并让发现入口复用它，是相称且可回滚的边界。未引入新的会话事实、状态同步、默认执行、指令义务或权限。当前完整代码可批准；实际安装、原生执行与受保护合并继续由原 owner 的资格及授权决定。

English verdict: APPROVE — dae5213. Full seven-path review. The shared selector excludes closed Sessions from current controls while retaining their history and visible resume failures. Twenty paired inputs, all eleven isolated browser cases, real HTTP/store/binding checks, the built desktop/narrow journey and local premerge support this bounded repair. Earlier development-server and private-fixture failures remain recorded. No installed/native-model, CI, merge-authority or complete recovery-program claim.