[loopx-agent]

A healthy original conversation could become unavailable because its entry also waited for the full connection-settings inventory. The Session read now returns a limited display source under the same Core verification and binding fence; the conversation entry reads that descriptor directly. Settings keep their full inventory, and input retains fresh Core admission and the existing five-second read deadline.

The descriptor carries the exact binding, private-owner audience, workspace availability, title and executor. The frontend refuses missing or mismatched descriptors, group audiences and executor changes. Ordinary Session lists retain their previous shape. The compatibility local_sessions caller and both RFC mirrors are retained.

Validation:

• 20 focused Python checks passed; the exact-head Chat build passes TypeScript checking.
• The bound-conversation and App-steward smokes pass against real HTTP/Core and disposable storage, including SQLite authority. Provider identity and model protocol are synthetic.
• Packaged browser comparison: with settings returning 503, the baseline cannot open the original conversation; the candidate reads all 14 synthetic messages and long replies with zero settings requests. Source denial fences a nonempty draft from sending; recovery returns to the same source. The 390px conversation and complete-answer reader have no horizontal overflow. No real model or messaging provider was called.
• Premerge: six direct checks and sixteen selected checks passed with zero executed failures. The risk gate correctly retains four manual-review holds pending exact-head review and merge readiness.

Existing limitation: an unsent draft entered while the executor is unresolved becomes hidden when source recovery resolves the executor. The same counterexample reproduces on baseline adbcb08f2; its executor-keyed draft code is unchanged. This source-read fix does not qualify draft continuity, complete historical/media synchronization or an installed provider roundtrip.

Visual evidence uses disposable synthetic content. Before is adbcb08f2; after uses the identical functional source diff committed in dd838ba955193ee6cb4a9b8709ce21ebe8532fd9, then rebuilt at that exact head. The rebase adds only unrelated browser-fixture changes from current main.

Before — connection settings unavailable:

After — same injected settings failure, original conversation readable:

390px conversation and complete-answer reader:

[loopx-agent]

Reviewer: model_agent | GPT-6 | OpenAI | self_reported

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

动机

无阻塞发现。已绑定飞书原私聊的桌面或网页用户在打开原对话时受影响。原对话和权限仍健康，但连接设置清单失败时，旧入口把历史显示为不可用；新入口可读取同一原会话，真正的来源撤销仍禁止发送。打包实测中，设置接口返回 503 时，候选读取全部 14 条合成原消息及长答复，并且没有请求设置清单。本次不宣称完整历史与媒体同步、草稿恢复或真实 provider 往返已验收。既有未发送草稿在执行器从未解析状态恢复时暂不可见；完整同源旅程仍未完成。

改动思路

复用已有的来源资格决策和存储读取，在同一个 binding 锁内投影显示信息；设置健康不再成为打开原对话的隐含前置，也没有建立第二份权限或会话权威。本 PR 只交付可独立验收和回滚的来源读取修复；完整历史、媒体、原结论回传及草稿连续性继续由既有同源旅程验收承担。资格决策继续在既有 TypeScript Core，Python 只负责读取与显示投影，前端继续用准确的 binding 和 Session 冻结目标。把等待时间改长只能延迟失败，缓存资格会破坏新鲜权限；新增独立入口也会复制来源所有权。因此复用同一次已核验读取是更小的实际修复。

下表是本次 dd838ba 源码和执行观察的说明，不另作执行证据：

条件	基线	本次
原来源健康，设置503	来源不可用、历史空白	同一来源可读，零设置请求
原来源或权限失效	拒绝	仍拒绝，无替代会话
原来源恢复	原目标重试	同一原目标恢复；既有草稿问题保留

具体改动

评审依据是改动前已接受的 docs/architecture/rfcs/app-conversation-and-async-inbox-v0.md，spec_revision 65450d6。本 PR 的 RFC 新段落只披露读取修复，不能反过来降低既有验收。original-private-target 已实现：准确回显 binding、固定原 Session，不回退普通历史或新建。fresh-source-and-grant 已实现：provider 身份、当前工作区、binding 和 Core 资格仍新核验，群受众不纳入。single-service-desktop-continuity 延后：这是文档明确保持开放的安装态连续性要求，草稿、全历史、媒体和实际原消息回传仍由原有同源旅程验收负责。

关键代码讲解

1. ConversationBindings.local_session_snapshot 保留原 provider 观察、Session 枚举和 binding 锁，先调用不变的 listLocalBoundSessions 决策，再投影有限来源。local_sessions 是真实旧调用者的薄包装，既有管家入口和 managed-mode 不承担第二套资格规则。
2. resolveBoundConversation 复用 exact Session 索引，并由现有 schema 挑选有限来源字段；缺失、错绑、群受众、工作区不可用、执行器不符都拒绝。HTTP 只有绑定分支增加来源，普通 Session 列表和全量设置读取不变。
3. useBoundConversation 删除全设置清单的并行依赖，继续保留原五秒 abort、旧完成丢弃及重试。工作区类型只缩到实际消费的有限来源，没有新增布局、状态存储或迁移。

其余变化是现有真实 HTTP/Core smoke 的设置503控制、错包负例、三类权限失效用例、双语说明与四张公开合成内容截图。整个差异14文件 +130/-16，生产变化集中在五条既有路径；没有新 provider、队列、消费者、持久化权威或额外用户配置。

对主干的风险

最危险的回归是把显示来源误当权限证明，或者原来源失败时误读另一个 Session。现有 Core 私聊、当前授权、准确上下文及重复身份检查保持原位；binding 在观察期间变化仍拒绝，前端缺少新描述时明确要求更新并重试，绝不降级到普通列表。代价是新前端需要配套服务，已通过同包构建和拒绝旧/错包用例限定。

20个聚焦 Python 用例通过，bound-conversation 与 app-steward 两项真实 HTTP/Core smoke 通过，精确提交 Chat TypeScript 构建通过。预合入执行的6项直接检查和16项选中检查均通过，无执行失败；总门禁仍有4项人工评审 hold，本评审不能代替随后准确 head 的 approval closeout 与 merge readiness。当前 Goal 为 wait_for_ci=false，没有查询、轮询或等待远端 CI。

语义与 CI 对齐

此次复用既有 binding/Session 词汇，来源描述是衍生显示数据，不增加行动授权、义务或生命周期。语义探针0个受支持新载体只是有界辅助；实际判断依据是现有调用者、Core 规则和基线/候选同输入比较。设置503反例在基线失败、候选读取14消息；来源撤销时非空草稿仍不能发送。相同草稿恢复旅程在不可变基线和候选均丢失可见草稿，其执行器分键代码未改，作为既有缺口保留，未删除断言或提高限额来制造通过。真实 provider/model 尚未由这些合成用例证明。

我的整体评价

在 dd838ba 上，problem_context 是 justified_increment，long_horizon 和 user_experience 均 improved：重复打开原对话少了一项无关服务依赖，不再要求重新配置或额外确认；真正来源失效仍由原权限所有者恢复并重试。已检查打包页面的长答复、错误恢复和完整阅读页；桌面来源恢复后目标与执行者清楚，390px 正文和完整答复仍可读，未引入布局或样式迁移。未在本次打包合成场景中证明安装态 provider 往返或完整媒体同步；既有恢复草稿丢失在基线与候选均复现并保留。

这个切口有可用的实际调用者，架构归属明确，兼容包装有真实消费者；已选择的实现比添加缓存、服务或延长等待更小。本 PR 只交付可独立验收和回滚的来源读取修复；完整历史、媒体、原结论回传及草稿连续性继续由既有同源旅程验收承担。因此批准这一有限来源读取修复，完整同源体验保持未闭。回滚不涉及数据格式变化。自有 PR 以 COMMENTED 记录本批准结论，正式自批准受 GitHub 限制；合并仍以独立授权和准确 head readiness 为准。

English verdict: APPROVE dd838ba. The original-source entry now remains readable when unrelated settings fail, while fresh Core admission and exact-target refusal are preserved. Focused runtime checks, both real HTTP/Core smokes, exact-head build and 6+16 local checks passed; four manual holds still require separate closeout/readiness. Existing base/head draft loss and wider history/media/provider acceptance remain open.