fix(chat): consolidate model controls and retain native reply parity

[loopx-agent]

A Goal conversation showed both a model menu and a separate defaults button, while a host that had not reported a model left a permanent warning in the header. Defaults now live in the existing model disclosure, with explicit missing-readback detail inside it. Reported model/effort/tier remain separate from choices for the next conversation; opening defaults and returning restores keyboard focus to the model menu.

Refresh errors remain visible and successful refreshes retain an accessible announcement. The runtime selector, status-only mode and Agent Loop configuration keep their separate owners. A second focused commit repairs eight native Lark test regressions after quoted reply context was introduced: requests are addressed by canonical provider/event identity, and only the current canonical message selects a control command.

Validation

• 44 model/default/effort tests passed through real HTTP/File and native stdio paths; 71 original-source, reply-context, exact-stop/background recovery and Session-isolation tests passed.
• Compiled UI exercised in Ego at 1512×982 and 390×844, Chinese/English, keyboard/Escape/return focus, one composer and no horizontal overflow. Defaults preview/apply/rollback reached the production configuration API in a disposable runtime. Conversation/catalog data was synthetic; no model request was sent to a real account.
• TypeScript/Vite build, source-bundle verification, 44 bilingual model/header assertions plus the complete workspace contract and two keyboard-composition checks, focused Ruff, full semantic bindings and standard premerge passed (four direct checks, compilation and all 17 selected catalog/risk/public checks). Exact-scope change-quality receipt is valid.
• The hosted build exposed an incidental source-order assertion for refresh states. It is replaced with bilingual SSR checks of loading, success announcement and retryable errors. The new test-only head passed quick premerge; other hosted jobs remain pending. Independent code review approved this exact head; merge readiness, merge authority and installed adoption remain separate.
• test(lark): select reply requests by provider message identity #6754 independently fixes related Lark request lookups, including four paths shared here. Retain its broader completed/background coverage and reconcile those identical canonical-identity edits when it lands; this proposal also fixes the untrusted control-source tests.
• Physical macOS IME, an installed desktop upgrade and actual Lark user adoption remain separate qualifications. No merge, installation or release is claimed.

Visual evidence

Tested source tree: ad057d8b2f671fcb0bafd890750c07db8eaa4c33. These are actual packaged-UI captures with public synthetic data. Before assets are the verified 90932447eb5ed38a0d3fec8c57e21a62f57f849c bundle; after assets match this source tree. Whole-viewport review keeps ordinary chat focused on one model entry and the composer, while diagnostics remain available on demand.

View	Before	After
Desktop
Narrow

Local disposable preview: http://127.0.0.1:58080/chat/?goalId=progress-projection&view=conversation&statusUrl=%2Fstatus.json (available on the author's machine while the preview process runs).

[loopx-agent]

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

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

精确审查 6756@ad057d8b2f671fcb0bafd890750c07db8eaa4c33；合并基点 90932447eb5ed38a0d3fec8c57e21a62f57f849c。APPROVE，未发现当前阻塞问题；代码结论、合并资格与合并权限分开。

动机

在 Goal 对话里查看当前模型、选择下次对话设置或维护默认配置的用户。 原头部同时放模型菜单和默认配置按钮，模型尚未回传时还长期显示缺失警告；现在默认配置进入同一披露菜单，未知观测在展开后明确说明，返回工作区恢复到模型入口。 独立打包页面通过同一模型入口进入原配置编辑器，预览不写入，应用后真实 File/HTTP 读回独立四项默认值，回滚恢复原先的缺省状态，返回焦点及390px无横向溢出通过。 只验证候选源码打包页面与可丢弃配置后端；会话和目录为合成数据，不发送真实模型或Lark消息，不代表已安装桌面、原生IME或本人实际Lark采纳。

改动思路

沿现有 ChatModelMenu、ChannelHeader 与机器配置 preview/apply/rollback owner 整合入口；不引入新配置、对话身份或前端写入决策。 当前 PR 收敛对话头部模型入口、缺失观测披露、成功播报和焦点返回，并修正引用消息上下文引入后的Lark测试定位；运行时、状态只读及Agent Loop仍归各自既有owner。 不改变现有状态和权限owner。保留两个入口会继续消耗头部注意力；另开设置页又会复制机器默认配置责任。普通用户继续聊天无需额外确认；可选默认配置入口多一次展开，换取单一入口和清晰的当前/下次语义。

触发	当前结果
查看/编辑默认模型	同一菜单进入原事务，预览不写，应用可回读
目录失败/退出默认编辑	显式重试；返回现存模型入口焦点

此表只概括本头源码与下述执行结果，不证明安装或模型采纳。

具体改动

规格 docs/integrations/session-runtime-control-plane-adapter.md，不可变版本 90932447eb5ed38a0d3fec8c57e21a62f57f849c。Projection boundary：Missing model readback stays unavailable; a catalog or requested choice is not the current observation.；New conversation settings：Choices preserve existing history and active work; ordinary defaults use existing revision-checked machine preview/apply/readback/rollback.；Conversation presentation example：One composer and preserved keyboard/native delivery boundaries; messaging quoted context grants no authority.

完整14路径包含模型菜单/头部/焦点三个生产文件、原生适配说明、两个浏览器scenario、四张合成首屏图和Lark/SSR断言。236→ad只改两个测试文件：将刷新源码顺序正则替换为双语SSR的idle/loading/成功播报/失败重试断言。生产UI、Python规则和图片blob均未变；旧头133项Python结果在逐文件失效检查后沿用，本头重新构建并验证bundle、44SSR和workspace contract。Lark改用provider/message identity仍验证背景停止、同Session与非法来源；引用文本保留且不授权控制命令。相关#6754有重叠，集成时应保留其更全的completed/background断言并去除重复helper，不在本头增设共享测试框架。

关键代码讲解

• ChatModelMenu：当前实际观测与下次对话选择分开，默认配置复用同一菜单。
• ChannelHeader：移除重复默认按钮，成功反馈只做可访问播报，错误仍可见。
• openSettings / closeSettings：默认编辑器仍用旧事务，关闭后定位现存模型summary。
• _request_for_event：根据provider/message身份读取原请求，不再依赖带引用上下文的消息全文相等。

对主干的风险

未发现阻塞问题。重点反例是把未知实际模型误显示为待选值、把默认编辑误当成当前会话变更，以及关闭编辑器后焦点落到已移除按钮。已在真实打包页通过配置 API 的预览、应用、回滚和焦点读回验证，并用目录503验证可见错误、显式重试及Escape。浏览器 Session/目录数据是合成的，不能作为真实模型运行资格。

本轮验证：133 Python tests通过；44双语SSR、workspace contract、TypeScript/Vite build及source bundle verify通过；当前ad头premerge5直接、8目录、8风险和1公开边界检查通过。 最初选了不存在的test_chat_model_selection.py，pytest未执行；纠正为实际7文件后133通过。直接Node运行SSR脚本不支持tsx，改用仓库规定vite SSR命令后44通过。一次浏览器等待input过宽歧义，未重发动作，读当前快照后继续。 只验证候选源码打包页面与可丢弃配置后端；会话和目录为合成数据，不发送真实模型或Lark消息，不代表已安装桌面、原生IME或本人实际Lark采纳。 独立浏览器宽屏及390px双语保持一个composer，实际API四字段分开；原配置从absent经预览/应用成为configured，再经显式回滚恢复absent。首屏更安静且仍能发现默认设置；未知观测留在披露内，不假定模型运行。真实IME、已安装App与实际Lark采纳未认证。

我的整体评价

APPROVE。独立打包页面通过同一模型入口进入原配置编辑器，预览不写入，应用后真实 File/HTTP 读回独立四项默认值，回滚恢复原先的缺省状态，返回焦点及390px无横向溢出通过。 沿现有 ChatModelMenu、ChannelHeader 与机器配置 preview/apply/rollback owner 整合入口；不引入新配置、对话身份或前端写入决策。 当前 PR 收敛对话头部模型入口、缺失观测披露、成功播报和焦点返回，并修正引用消息上下文引入后的Lark测试定位；运行时、状态只读及Agent Loop仍归各自既有owner。 long_horizon 与 user_experience 在此有界流程 improved；不以UI读回推导真实模型或长期效果。已有owner内的入口收敛和按身份定位测试是合适的前瞻整理，无需增加框架。 没有执行合并、安装或正式运行时升级，没有查询、轮询或等待CI。

English verdict: APPROVE. The exact head consolidates existing model/default controls without changing native binding or write authority. Real packaged defaults preview/apply/rollback, failure/retry, bilingual narrow layout and focus pass; installed/live-model adoption remains unqualified.