Merge remote-tracking branch 'origin/master' into mergebot/pr998

# Conflicts: # packages/client/ui-conversation/README.i18n.yaml
2026-08-15 21:04:50 +00:00 · 2026-07-31 12:23:03 +08:00
parent 01ccfc0ad3 4a061f33d0
commit 5afdb5c235
50 changed files with 1240 additions and 102 deletions
--- a/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.i18n.yaml
+++ b/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write .agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md
-2026-06-21-bounded-llm-request-recovery.md: 83d47e3a7d91bbcd2ceaf7b11cf13316142eb3ed
-2026-06-21-bounded-llm-request-recovery.zh.md: 00dcbad3d1023ad33a22297bfe938b94bce839d4
+2026-06-21-bounded-llm-request-recovery.md: 5c76ed5d754ea40f41dff78cb56ee7fc139a32b1
+2026-06-21-bounded-llm-request-recovery.zh.md: 1fa56f3fe0405cab663c2843d423a78d910170dd
--- a/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md
+++ b/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.md
@@ -60,11 +60,11 @@ For an eligible failure with budget remaining, the one-based transient retry cou

 The plugin owns a lifetime `AbortController` and tracks every active recovery callback, including delegated waterfall work and backoff. Effect cleanup first unregisters the listener, then aborts and awaits the active callbacks; abort wins over a late delegated retry decision, and a captured callback can neither retry nor enter the rest of its waterfall after disposal. This makes HMR disposal quiescent even though Cordis has already captured the listener.

-Before sleeping, `dsh-llm-retry` appends one non-surface `llm/retry` session event containing the turn, failed step, provider, policy mode, complete resolved-policy key, provider-policy retry number, mode-specific finite maximum when present, scheduled delay, and `LlmFailure`. The key sorts the code set and separates retry histories when a provider route is replaced by a behaviorally different same-mode policy. The plugin owns the `SessionEventMap` augmentation; `dsh-session` remains generic persistence and does not absorb the optional policy's vocabulary. The event says what was scheduled, not that the next request completed; cancellation during the delay is subsequently visible on `turn/end`. The event ships only with a production renderer and replay/snapshot coverage, because its purpose is operational state rather than trace collection.
+Before sleeping, `dsh-llm-retry` appends one non-surface `llm/retry` session event containing the turn, failed step, provider, policy mode, complete resolved-policy key, provider-policy retry number, mode-specific finite maximum when present, scheduled delay, and `LlmFailure`. The key sorts the code set and separates retry histories when a provider route is replaced by a behaviorally different same-mode policy. The plugin owns the `SessionEventMap` augmentation and exports the payload through its browser-safe `./types` subpath; `dsh-session` remains generic persistence and does not absorb the optional policy's vocabulary. The event says what was scheduled, not that the next request completed; cancellation during the delay is subsequently visible on `turn/end`. The event ships with production renderers and replay/snapshot coverage, because its purpose is operational state rather than trace collection.

 The listener calls `next()` for a non-transient code, an exhausted policy budget, or an over-cap provider delay. This preserves composition with context-overflow recovery and later policy plugins. For an owned failure it records and awaits the delay, then returns `{ kind: 'retry' }` without delegating. Turn cancellation and plugin disposal end the wait without returning a retry; the loop's cancellation/disposal checks remain authoritative.

-The agent-spine demo bundle loads the plugin so the shared stdio/TUI, one-shot CLI, and ACP example compositions use the same provider-routed policy. Library consumers retain explicit plugin composition: omitting the plugin leaves request failures terminal.
+The agent-spine demo bundle loads the plugin so the shared stdio/TUI, one-shot CLI, ACP, and headless example compositions use the same provider-routed policy. The shipped Web composition also loads it, so browser and command-line requests use the same provider defaults. Library consumers retain explicit plugin composition: omitting the plugin leaves request failures terminal.

 ### Make one layer own visible attempts

@@ -82,7 +82,7 @@ Boundary tests prove termination at both actual transports. The hand-written ada

 ### Keep attempts separate in the existing log

-A failed attempt may leave `assistant/chunk` events in its closed step, but it never appends `assistant/message` and never dispatches a tool. A retry closes the failed turn, opens the next numbered turn, reconstructs the request from the durable surface, and produces its own chunks. UIs may render live chunks while a step is open, then mark or clear that transient view when `llm/retry` identifies the failed step or `turn/end` records failure; message derivation continues to ignore the failed chunks.
+A failed attempt may leave `assistant/chunk` events in its closed step, but it never appends `assistant/message` and never dispatches a tool. A retry closes the failed turn, opens the next numbered turn, reconstructs the request from the durable surface, and produces its own chunks. UIs may render live chunks while a step is open, then mark or clear that transient view when `llm/retry` identifies the failed step or `turn/end` records failure. Web validates the complete retry payload contract, clears the failed partial at `llm/retry`, projects consecutive retry-turn events into one stable row updated to the latest attempt, and derives scheduled, started, or cancelled status from subsequent turn facts. Its countdown anchors the scheduled delay to browser receipt rather than the Host event clock, uses ceiling-rounded seconds with a one-second floor, animates only while unresolved, and keeps exact latest failure details collapsed behind the row. Retry nodes anchor their own trajectory turn even when the failed attempt has no assistant node. Message derivation continues to ignore the failed chunks, and Web applies the same projection during history rebuild so refreshing cannot resurrect discarded partials or duplicate retry rows.

 If recovery is exhausted, the final failure is stored once on `turn/end.reason` with the structured facts. If transient recovery continues, `llm/retry` is the durable home for that attempt's failure and delay. No standalone final-error event or response-id vocabulary is added.

@@ -116,7 +116,7 @@ If recovery is exhausted, the final failure is stored once on `turn/end.reason`
 - Pure unit tests cover transient-code selection, exponential backoff and jitter bounds, valid and over-cap `Retry-After`, exhausted budgets, deterministic timer/random seams, and abort during backoff.
 - Real agent-loop tests cover failure before chunks, partial chunks then failure, thrown and in-band failures, retry to success in a new turn, exhaustion to structured `turn/end.reason`, and composition with `dsh-compact-basic` context-overflow recovery.
 - The partial-chunk integration test proves failed chunks remain attributed to the failed step, no assistant message or tool side effect is committed for that step, and the successful retry has distinct provenance.
- The plugin-owned `llm/retry` event is non-surface, survives JSONL and SQLite round trips, is ignored by message derivation, and drives TUI retraction plus scheduled-retry rendering. Keyless snapshots cover scheduling, cancellation, success, and exhaustion; ACP automation snapshots confirm that a discarded attempt stays off the wire while the recovered reply is emitted.
+- The plugin-owned `llm/retry` event is non-surface, survives JSONL and SQLite round trips, is ignored by message derivation, and drives TUI and Web retraction plus scheduled-retry rendering. Client tests cover complete wire validation, clock-independent countdown, cancellation versus completed retry labels, and trajectory attribution; keyless UI snapshots cover Web scheduling and success, a real Web composition test covers partial transport failure through recovery, and ACP automation snapshots confirm that a discarded attempt stays off the wire while the recovered reply is emitted.
 - Idle-watchdog tests prove the stable signal is rearmed only while `next()` is outstanding, disarmed during consumer think time and in `finally`, and classified separately from a total-call deadline and an earlier caller abort; adapter tests prove the signal stops the underlying request rather than merely detaching it.
 - Direct `ctx.llm.stream()` callers remain single-attempt and receive the same structured failure facts.

--- a/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.zh.md
+++ b/.agents/notes/implemented/architecture/2026-06-21-bounded-llm-request-recovery.zh.md
@@ -60,11 +60,11 @@ agent loop（智能体循环）会保留 `RequestError` 作为该精确的错误

 插件拥有一个全生命期 `AbortController`，并跟踪每个活跃的恢复回调，包括委托的 waterfall（瀑布式事件）工作与退避。effect 清理会先注销监听器，再中止并等待活跃回调；中止会胜过较晚到达的委托重试决策，被捕获的回调在插件释放后既不能重试，也不能进入其 waterfall 的剩余部分。尽管 Cordis 已捕获该监听器，此设计仍能使 HMR（热模块替换）释放达到完全停稳。

-休眠前，`dsh-llm-retry` 会追加一条不进入表层的 `llm/retry` 会话事件，其中包含轮次、失败步骤、提供方、策略 mode、完整的解析策略 key、提供方策略重试编号、该 mode 存在时的有限上限、计划延迟和 `LlmFailure`。该 key 会对 code 集排序，并在提供方路由被行为不同但 mode 相同的策略替换时分隔重试历史。该插件拥有 `SessionEventMap` 声明合并；`dsh-session` 继续负责通用持久化，不会吸收可选策略的词汇。事件记录已安排的内容，而不是下一个请求已完成；延迟期间取消随后会在 `turn/end` 中可见。因为该事件的目的是表示运行状态，而不是收集跟踪数据，所以它仅与生产渲染器及回放／快照覆盖一起交付。
+休眠前，`dsh-llm-retry` 会追加一条不进入表层的 `llm/retry` 会话事件，其中包含轮次、失败步骤、提供方、策略 mode、完整的解析策略 key、提供方策略重试编号、该 mode 存在时的有限上限、计划延迟和 `LlmFailure`。该 key 会对 code 集排序，并在提供方路由被行为不同但 mode 相同的策略替换时分隔重试历史。该插件拥有 `SessionEventMap` 声明合并，并通过其浏览器安全的 `./types` 子路径导出载荷；`dsh-session` 继续负责通用持久化，不会吸收可选策略的词汇。事件记录已安排的内容，而不是下一个请求已完成；延迟期间取消随后会在 `turn/end` 中可见。因为该事件的目的是表示运行状态，而不是收集跟踪数据，所以它会与生产渲染器及回放／快照覆盖一起交付。

 对非暂时性 code、耗尽的策略预算或超出上限的提供方延迟，监听器会调用 `next()`。这保留了与上下文溢出恢复及后续策略插件的组合能力。对自身处理的失败，它会记录并等待延迟，然后在不委托的情况下返回 `{ kind: 'retry' }`。轮次取消和插件释放会结束等待且不返回重试动作，此后仍以循环的取消／释放检查为准。

-agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次性 CLI（命令行界面）和 ACP（Agent Client Protocol）示例组合使用同一套按提供方路由的策略。库消费方仍需显式组合插件：省略该插件时，请求失败保持终态。
+agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次性 CLI（命令行界面）、ACP（Agent Client Protocol）和 headless 示例组合使用同一套按提供方路由的策略。随产品交付的 Web 组合也会加载该插件，因此浏览器请求与命令行请求使用相同的提供方默认值。库消费方仍需显式组合插件：省略该插件时，请求失败保持终态。

 ### 由单一层负责可见的尝试

@@ -82,7 +82,7 @@ agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次

 ### 在现有日志中分隔尝试

-一次失败尝试可以在已关闭的步骤中留下 `assistant/chunk` 事件，但绝不会追加 `assistant/message`，也不会分发工具。重试会关闭失败轮次，开启下一个编号轮次，从持久表层重建请求，并生成自己的分片。步骤仍处于打开状态时，UI 可以渲染实时分片；当 `llm/retry` 标识失败步骤，或 `turn/end` 记录失败时，UI 再标记或清除这份暂时视图。消息派生仍会忽略失败分片。
+一次失败尝试可以在已关闭的步骤中留下 `assistant/chunk` 事件，但绝不会追加 `assistant/message`，也不会分发工具。重试会关闭失败轮次，开启下一个编号轮次，从持久表层重建请求，并生成自己的分片。步骤仍处于打开状态时，UI 可以渲染实时分片；当 `llm/retry` 标识失败步骤，或 `turn/end` 记录失败时，UI 再标记或清除这份暂时视图。Web 会验证完整的重试载荷契约，在 `llm/retry` 到达时清除失败的部分输出，将连续重试轮次的事件投影为稳定的一行，并用最新一次尝试更新该行，再从后续轮次事实派生 scheduled、started 或 cancelled 状态。倒计时以浏览器收到事件的时刻为计划延迟的起点，而不是使用 Host 事件时钟；它按向上取整且不低于 1 秒的秒数显示，仅在重试尚未结束时显示动画，并把最近一次失败的准确详情折叠在该行之后。即使失败尝试没有 assistant 节点，重试节点也会锚定自身的轨迹轮次。消息派生仍会忽略失败分片；Web 在重建历史时也会应用同一投影，因此刷新页面不会让已丢弃的部分输出重新出现，也不会生成重复的重试行。

 如果恢复预算耗尽，最终失败会连同结构化事实在 `turn/end.reason` 中存储一次。如果暂时性恢复继续，`llm/retry` 就是该次尝试的失败与延迟的持久归属位置。本决策不增加独立的最终错误事件或响应 id 词汇。

@@ -116,7 +116,7 @@ agent-spine 演示组合包加载该插件，因此共享的 stdio/TUI、一次
 - 纯单元测试覆盖暂时性 code 选择、指数退避和抖动边界、有效及超出上限的 `Retry-After`、耗尽的预算、确定性定时器／随机数 seam，以及退避期间中止。
 - 真实 agent-loop 测试覆盖分片前失败、部分分片后失败、抛出及带内失败、在新轮次中重试至成功、耗尽后写入结构化 `turn/end.reason`，以及与 `dsh-compact-basic` 上下文溢出恢复的组合。
 - 部分分片集成测试证明：失败分片仍归属于失败步骤，该步骤不会提交 assistant 消息或工具副作用，成功的重试具有不同的来源信息。
- 插件拥有的不进入表层的 `llm/retry` 事件可在 JSONL 和 SQLite 往返后保留，被消息派生忽略，并驱动 TUI 撤回和计划重试渲染。无密钥快照覆盖调度、取消、成功和耗尽；ACP 自动化快照确认，被丢弃的尝试不会通过协议发出，而恢复后的回复会正常发出。
+- 插件拥有的不进入表层的 `llm/retry` 事件可在 JSONL 和 SQLite 往返后保留，被消息派生忽略，并驱动 TUI 和 Web 撤回及计划重试渲染。客户端测试覆盖完整的 wire 验证、独立于时钟的倒计时、已取消与已完成重试标签的区别以及轨迹归属；无密钥 UI 快照覆盖 Web 的调度与成功，真实 Web 组合测试覆盖部分传输失败直至恢复，ACP 自动化快照确认，被丢弃的尝试不会通过协议发出，而恢复后的回复会正常发出。
 - 空闲看门狗测试证明：只有 `next()` 尚未完成时才会重新布防稳定信号；在消费方思考期间及 `finally` 中会解除布防；它与总调用 deadline 以及更早发生的调用方中止分开分类。适配器测试证明该信号会终止底层请求，而不只是与其脱离。
 - `ctx.llm.stream()` 的直接调用方仍只尝试一次，并收到相同的结构化失败事实。

--- a/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.i18n.yaml
+++ b/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.i18n.yaml
@@ -0,0 +1,6 @@
+# Bilingual-pair consistency record (docs/i18n/README.md): the git blob hash of each
+# side as of the last confirmed-consistent state. Both languages carry equal authority;
+# after editing either side, bring the other along and re-record with:
+#   pnpm run verify-translation-pairing --write .agents/notes/implemented/process/2026-07-22-product-first-root-readme.md
+2026-07-22-product-first-root-readme.md: 32542a45019d64ed1826d4eb21e68c67c3c3d52e
+2026-07-22-product-first-root-readme.zh.md: 1c4d5fa53854bfcade9742da1fb74d9636909f84
--- a/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.md
+++ b/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.md
@@ -0,0 +1,33 @@
+# Agent Note: Product-first root README
+
+Status: implemented
+
+English | [中文](2026-07-22-product-first-root-readme.zh.md)
+
+## Problem
+
+The root README is the repository's product front door. Its product-first structure and established voice remain useful, but concrete entry points and capability claims drift as the runtime grows. Rewriting sections whose facts remain correct increases the review surface and discards language that already works.
+
+## Decision
+
+The root README preserves its existing structure, order, and wording wherever the underlying fact remains correct. A refresh changes only stale claims and adds material needed to represent shipped surfaces; it does not use repository growth as a reason to reframe the whole page.
+
+A note before installation thanks internal testers, states that features and experience remain unfinished, and asks for direct reports of failures, confusion, and friction through the WeCom group. The existing development-stage statement identifies DeepSeek Harness as being in internal testing.
+
+The user-surface section adds the ACP automation server and Python/JSON-RPC SDK beside the existing Web, TUI, and headless entries. The installed TUI remains the single `dsh` command; the Web instructions build the active checkout before running `dsh web`, and custom or reused checkout paths stay explicit. These launch paths must remain executable through a real PTY and a production build/HTTP smoke, respectively. The capability paragraph keeps its compact inventory style while adding the shipped PTY, LSP, web, goal, planning, task, sandbox, approval, settings, credentials, session-query, and telemetry families and stating that compositions select subsets. One adjacent bullet records the authoritative-session-log rule because persistence, replay, queries, telemetry, and interfaces depend on it.
+
+Detailed package and service inventories remain at their owning documentation. The English and Chinese README sides share the same technical structure, while their community sections continue to point to the primary channel for each language audience. The documentation website keeps its separate user-guide landing page.
+
+## Alternatives considered
+
+**Rewrite the README around a new product narrative.** A complete rewrite can make every current surface prominent, but it replaces accurate, reviewed copy and creates unnecessary churn. Current facts fit the established product-first structure.
+
+**Present the repository as an SDK and package catalog.** This exposes implementation breadth immediately but makes a new reader reconstruct the product from package names. The package map and generated capability graph remain the authoritative inventories.
+
+**Use a long marketing page with screenshots, badges, and duplicated tutorials.** Rich media can demonstrate a stable product journey, but it ages separately from commands and source contracts. The root stays compact and links to runnable examples and owned guides.
+
+**Project the root README as the documentation website home page.** A single landing page avoids two narratives, but the website's user guide and the repository's product/developer front door have different navigation and maintenance needs.
+
+## Consequences
+
+Reviewers can distinguish factual refreshes from editorial rewrites, and future updates retain established wording unless its meaning becomes false or incomplete. The README must still change with affected commands, entry points, release-stage claims, or high-level capability families, while exhaustive detail remains linked rather than copied.
--- a/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.zh.md
+++ b/.agents/notes/implemented/process/2026-07-22-product-first-root-readme.zh.md
@@ -0,0 +1,33 @@
+# Agent Note: 产品优先的根 README
+
+Status: implemented
+
+[English](2026-07-22-product-first-root-readme.md) | 中文
+
+## 问题
+
+根 README 是仓库的产品入口。其产品优先的结构和既有语气仍然有效，但随着运行时不断扩展，具体入口和能力声明会逐渐陈旧。重写事实仍然正确的章节，会扩大评审范围，也会丢弃已经行之有效的措辞。
+
+## 决策
+
+只要背后的事实仍然正确，根 README 就保留既有结构、顺序和措辞。刷新时只修正陈旧声明，并补充呈现已交付内容所需的信息；不会因为仓库规模增长就重构整篇叙事。
+
+安装说明之前的一则文字感谢内测用户，说明功能和体验仍待完善，并邀请大家通过企业微信群直接反馈失败、困惑和不顺手之处。既有的开发阶段声明明确说明 DeepSeek Harness 处于内测阶段。
+
+用户入口章节在已有的 Web、TUI 和 Headless 入口旁补充 ACP（Agent Client Protocol）自动化服务器和 Python/JSON-RPC SDK。安装后的 TUI 仍只需执行一条 `dsh` 命令；Web 说明要求先构建当前检出，再运行 `dsh web`，并明确处理自定义或复用的检出路径。这两条启动路径必须分别能在真实 PTY 与生产构建／HTTP 冒烟中原样执行。能力段落沿用简洁清单的写法，补充已经交付的 PTY、LSP、Web、目标、规划、任务、沙箱、审批、设置、凭据、会话查询和遥测等能力类别，并说明不同组合只选用其中一部分。相邻的一条列表项说明权威会话日志规则，因为持久化、回放、查询、遥测和各类接口都依赖它。
+
+包（package）与服务的完整清单仍由各自的归属文档维护。中英文 README 采用相同的技术结构，但社区章节仍分别指向各自语言受众的主要交流渠道。文档网站继续使用独立的用户指南首页。
+
+## 考虑过的替代方案
+
+**围绕新的产品叙事重写 README。** 完整重写能够突出所有现有入口和能力，但也会替换准确且已经过评审的文案，造成不必要的变动。现有事实能够纳入既有的产品优先结构。
+
+**将仓库呈现为 SDK 和包清单。** 这样能立即展现实现广度，却会迫使新读者从包名反推出产品。包索引与生成的能力图仍是权威清单。
+
+**使用包含截图、徽章和重复教程的长篇营销页面。** 富媒体能够展示稳定的产品使用路径，但其内容会独立于命令和源码契约而逐渐陈旧。根 README 保持紧凑，并链接到可运行示例和各自维护的指南。
+
+**将根 README 投影为文档网站首页。** 使用同一个首页可以避免两套叙事，但文档网站的用户指南与仓库面向产品和开发者的入口在导航和维护需求上并不相同。
+
+## 结果
+
+评审者可以区分事实更新与编辑性重写；今后的更新会保留既有措辞，除非其含义已经不再正确或完整。受影响的命令、入口、发布阶段声明或高层能力类别发生变化时，README 仍须同步更新；完整细节则继续以链接方式提供，而不是复制到正文。
--- a/README.i18n.yaml
+++ b/README.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write README.md
-README.md: f9f7294b42e29132d5cd46c0ab6a5f5265a1d8f3
-README.zh.md: 88cbf8522d8f1a183a48dc7e80858d1a0ced8f0f
+README.md: baf5d79b157ae845cc837261452853afd48dbe46
+README.zh.md: 57d7bcf44cda36b37ae233754dbfba4ead2204fd
--- a/README.md
+++ b/README.md
@@ -6,6 +6,16 @@ DeepSeek Harness (`dsh`) is an open-source coding agent built on the DeepSeek Ha

 It uses an architecture where **everything is a plugin**.

+## Internal testing notice
+
+Thank you for making time to try DeepSeek Harness.
+
+This version is still in internal testing. Some features remain unfinished, and parts of the experience may feel rough.
+
+“As one cuts and files, as one carves and polishes.” Products grow through repeated encounters with real use and candid feedback. The problems you uncover in practice may lead us to re-examine, or even discard, existing designs.
+
+We especially want to hear about moments of failure, confusion, or friction. If DeepSeek Harness does not help—or instead makes your work harder—please leave a message in our <a href="https://wj.qq.com/s2/27234598/03eb/">WeCom group</a> and tell us about your experience. Every report will help us refine it.
+
 ## Install

 Install `dsh` with one command:
@@ -22,20 +32,14 @@ The installer keeps every checkout under `~/.dsh/source`: the master clone at `~

 ### Web UI

-For the recommended local interface, build the frontend after installation and after each update, then start the Web UI. Resolve the running checkout from the `dsh` launcher so the command holds regardless of which staging worktree is current (the launcher resolves through the stable `current` symlink):
+For the recommended local interface, build the active checkout after installation and after each update, then start the Web UI:

 ```sh
-dsh_bin=$(cd "$(dirname "$(command -v dsh)")" && pwd -P)/$(basename "$(command -v dsh)")
-while [ -L "$dsh_bin" ]; do
-  link=$(readlink "$dsh_bin")
-  case $link in /*) dsh_bin=$link ;; *) dsh_bin=$(cd "$(dirname "$dsh_bin")" && cd "$(dirname "$link")" && pwd -P)/$(basename "$link") ;; esac
-done
-dsh_dir=$(cd "$(dirname "$dsh_bin")/.." && pwd -P)
-pnpm --dir "$dsh_dir" run build && pnpm --dir "$dsh_dir" run build:web
+(cd ~/.dsh/source/current && pnpm run build)
 dsh web
 ```

-The Web UI is served at `http://127.0.0.1:3080` by default.
+The full build produces the library and client bundles plus the frontend dist. The path above is the installer's default. If you set `DSH_SOURCE` or `DSH_CURRENT`, or reused an existing checkout, replace `~/.dsh/source/current` with that checkout path; see [`scripts/install.sh`](scripts/install.sh) for details. The Web UI is served at `http://127.0.0.1:3080` by default.

 ### TUI

@@ -53,11 +57,22 @@ Run one task, print the final answer, and exit:
 dsh -p "summarize this workspace"
 ```

+### Automation and SDKs
+
+From a source checkout with `DEEPSEEK_API_KEY` in the environment or its root `.env`, start the ACP automation server:
+
+```sh
+pnpm run demo:acp
+```
+
+The [Python SDK](python/README.md) drives a bundled JSON-RPC runtime. The [examples](examples/README.md) cover the runnable headless, ACP, JSON-RPC, Code Mode, and self-referential compositions.
+
 ## Why DeepSeek Harness

-Built-in capabilities cover file reading, editing, and search; shell execution; reusable skills; task tracking; subagents and workflows; persistent sessions; and context compaction. The TUI also includes Plan Mode.
+Built-in capabilities cover file reading, editing, and search; shell and persistent PTY execution; reusable skills; task tracking, goals, plans, todos, and background tasks; subagents and workflows; sandboxing and approvals; settings and credentials; persistent, resumable, forkable, and queryable sessions; LSP and web access; context compaction; and telemetry. Each composition selects the subset appropriate to its surface. The TUI and Web UI both include Plan Mode.

 - **Everything is a plugin.** Models, tools, policies, storage, context management, and interfaces are composable [Cordis plugins](docs/user/develop/basic/index.md), so deployments can extend or replace behavior without forking the agent loop. See the [architecture](docs/architecture.md) for the underlying design.
+- **Runs are reconstructable.** Anything visible to the model is logged in the authoritative session stream; persistence, resume/fork/query, replay, telemetry, and UIs derive from the same events. See the [session-log architecture](docs/architecture.md#session-log).
 - **Code Mode (opt-in).** It exposes a `run_code` tool and a generated TypeScript SDK; only program output re-enters model context. See [Code Mode](packages/core/tools/README.md#code-mode).
 - **Self-referential Cordis tools are opt-in.** They let the agent inspect its live runtime and mount or unmount plugins while it runs. See the [Cordis tools](packages/cordis/tool-cordis/README.md).

@@ -76,7 +91,7 @@ Start with the [development guide](docs/development.md) and read the [architectu

 For agents, follow [AGENTS.md](AGENTS.md).

-DeepSeek Harness is currently pre-release.
+DeepSeek Harness is currently in internal testing.

 ## License

--- a/README.zh.md
+++ b/README.zh.md
@@ -6,6 +6,16 @@ DeepSeek Harness（`dsh`）是一款基于 DeepSeek Harness SDK 构建的开源

 它采用了**一切皆插件**的架构。

+## 内测声明
+
+感谢您愿意拨冗试用 DeepSeek Harness。
+
+目前的版本仍处于内部测试阶段，有些功能仍待完善，有些体验难免粗粝。
+
+“如切如磋，如琢如磨。”产品的成长，离不开一次次真实的碰撞与坦诚的反馈。您在真实使用中暴露的问题，也可能促使我们重新审视，甚至推翻已有的设计。
+
+我们尤其希望听见那些失败、困惑与不顺手的时刻——如果它未能帮到您，甚至反而为工作平添了麻烦，请在企业微信群中留言，将使用感受告诉我们。每一条反馈，都会帮助我们把它打磨得更好。
+
 ## 安装

 使用一条命令安装 `dsh`：
@@ -22,20 +32,14 @@ curl -fsSL https://raw.githubusercontent.com/deepseek-harness/deepseek-harness/m

 ### Web UI

-推荐在本地使用 Web UI。安装完成后以及每次更新后，请先构建前端，再启动 Web UI。通过 `dsh` 启动器解析当前运行的检出，这样无论当前是哪个 staging worktree，命令都成立（启动器会经由稳定的 `current` 符号链接解析）：
+推荐在本地使用 Web UI。安装完成后以及每次更新后，请先构建当前生效的检出，再启动 Web UI：

 ```sh
-dsh_bin=$(cd "$(dirname "$(command -v dsh)")" && pwd -P)/$(basename "$(command -v dsh)")
-while [ -L "$dsh_bin" ]; do
-  link=$(readlink "$dsh_bin")
-  case $link in /*) dsh_bin=$link ;; *) dsh_bin=$(cd "$(dirname "$dsh_bin")" && cd "$(dirname "$link")" && pwd -P)/$(basename "$link") ;; esac
-done
-dsh_dir=$(cd "$(dirname "$dsh_bin")/.." && pwd -P)
-pnpm --dir "$dsh_dir" run build && pnpm --dir "$dsh_dir" run build:web
+(cd ~/.dsh/source/current && pnpm run build)
 dsh web
 ```

-Web UI 默认通过 `http://127.0.0.1:3080` 提供服务。
+完整构建会生成库与客户端 bundle，以及前端 dist。上述路径是安装器的默认位置。如果你设置过 `DSH_SOURCE` 或 `DSH_CURRENT`，或者复用了已有检出，请把 `~/.dsh/source/current` 换成该检出路径；详情见 [`scripts/install.sh`](scripts/install.sh)。Web UI 默认通过 `http://127.0.0.1:3080` 提供服务。

 ### TUI

@@ -53,11 +57,22 @@ dsh
 dsh -p "summarize this workspace"
 ```

+### 自动化与 SDK
+
+在源码检出中通过环境变量或根目录 `.env` 设置 `DEEPSEEK_API_KEY`，然后启动 ACP（Agent Client Protocol）自动化服务器：
+
+```sh
+pnpm run demo:acp
+```
+
+[Python SDK](python/README.md) 驱动随附的 JSON-RPC 运行时。[示例](examples/README.md)涵盖可运行的 headless、ACP、JSON-RPC、Code Mode 和自指组合。
+
 ## 为什么选择 DeepSeek Harness

-内置功能涵盖文件读取、编辑与搜索、shell 执行、可复用 skill（技能）、任务跟踪、subagent 与工作流、持久化会话，以及上下文压缩（context compaction）。TUI 还包含 Plan Mode。
+内置功能涵盖文件读取、编辑与搜索、shell 和持久 PTY 执行、可复用 skill（技能）、任务跟踪、目标、计划、待办事项与后台任务、subagent 与工作流、沙箱与审批、设置与凭据、可持久化、恢复、fork 与查询的会话、LSP 与 Web 访问、上下文压缩（context compaction），以及遥测。每个组合只选用适合其使用方式的能力子集。TUI 与 Web UI 均包含 Plan Mode。

 - **一切皆插件。** 模型、工具、策略、存储、上下文管理和界面均可组合为 [Cordis 插件](docs/user/develop/basic/index.md)，部署方无需 fork agent loop（智能体循环）即可扩展或替换行为。底层设计见[架构文档](docs/architecture.md)。
+- **运行可重建。** 凡是模型可见的内容，都会记录在权威会话流中；持久化、恢复／fork／查询、回放、遥测和 UI 均从同一组事件派生。参见[会话日志架构](docs/architecture.md#session-log)。
 - **Code Mode（需显式启用）。** 它会提供 `run_code` 工具和生成的 TypeScript SDK，只有程序输出会重新进入模型上下文。参见 [Code Mode](packages/core/tools/README.md#code-mode)。
 - **自指 Cordis 工具需显式启用。** 这些工具可让 agent 检查自身的实时运行时，并在运行中挂载或卸载插件。参见 [Cordis 工具](packages/cordis/tool-cordis/README.md)。

@@ -80,7 +95,7 @@ pnpm run test:coverage

 面向 agent：遵循 [AGENTS.md](AGENTS.md)。

-DeepSeek Harness 目前处于预发布阶段。
+DeepSeek Harness 目前处于内测阶段。

 ## 许可证

--- a/apps/cli/README.i18n.yaml
+++ b/apps/cli/README.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write apps/cli/README.md
-README.md: e4b34c11d5deb722caed199d6350f7931092a636
-README.zh.md: 5701bc8b6d99f00e68db572a58a0b6d520d67f08
+README.md: c36a75fc61fd7118f48c9b68be3144177df19534
+README.zh.md: e926fa99c4e483351f52ca4e76b668e26b34d02f
--- a/apps/cli/README.md
+++ b/apps/cli/README.md
@@ -17,8 +17,7 @@ The TUI surface:

 `dsh upgrade` is a guided fresh-session entry over the default TUI surface: it mints a fresh session in the invoking directory and seeds its first turn with the bundled `dsh-upgrade` skill, exactly as if the user typed `/skill:<name>`. The launcher passes the skill name on the boot context ([`INITIAL_SKILL_KEY`](../../packages/ui/tui/README.md)), which the TUI auto-invokes once the chat is live. Both take no options — `--config`, `-p`, and `--resume` fail loud — and seed only on this first launch, so a later `dsh --resume <id>` of the session is an ordinary TUI session with no re-injection.

-
-The Web and headless surfaces boot `base.cordis.yml` plus `web.cordis.yml`, then apply `$DSH_HOME/config.yaml`; an explicit `--config <path>` replaces that personal overlay. Both surfaces otherwise share the same composition: both treat the invoking directory as the default project and Workspace root, create named Workspaces beneath that root unless `--workspace-root <path>` overrides it, load applicable `AGENTS.md`/`CLAUDE.md` instructions into each agent-loop request prefix with a 65,536-byte render budget, opt into first-message model titles, and mount a disposable in-memory SQLite content-index service. That service is ACTIVE at boot, while its `node:sqlite` module and database handle open only on the first content search. This keeps Node 22 startup output free of SQLite's experimental warning before search is used; the first actual search may still emit the runtime warning. Each service instance owns its database, so parallel invocations neither share unsupported SQLite state nor leave derived index files behind, and the first search lazily reconciles live and persisted logs. Headless differs only in listening on an OS-assigned port (parallel `dsh -p` runs never collide; the stderr-printed URL opens the live session in a browser). Both need the frontend dist and client bundles built (`pnpm run build && pnpm run build:web`).
+The Web and headless surfaces boot `base.cordis.yml` plus `web.cordis.yml`, then apply `$DSH_HOME/config.yaml`; an explicit `--config <path>` replaces that personal overlay. Both surfaces otherwise share the same composition: both treat the invoking directory as the default project and Workspace root, create named Workspaces beneath that root unless `--workspace-root <path>` overrides it, load applicable `AGENTS.md`/`CLAUDE.md` instructions into each agent-loop request prefix with a 65,536-byte render budget, opt into first-message model titles, use the same bounded transient model-request retry policy as the TUI, and mount a disposable in-memory SQLite content-index service. That service is ACTIVE at boot, while its `node:sqlite` module and database handle open only on the first content search. This keeps Node 22 startup output free of SQLite's experimental warning before search is used; the first actual search may still emit the runtime warning. Each service instance owns its database, so parallel invocations neither share unsupported SQLite state nor leave derived index files behind, and the first search lazily reconciles live and persisted logs. Headless differs only in listening on an OS-assigned port (parallel `dsh -p` runs never collide; the stderr-printed URL opens the live session in a browser). Both need the frontend dist and client bundles built (`pnpm run build && pnpm run build:web`).

 The shipped TUI and Web compositions register the native DeepSeek adapter plus pi-ai OpenAI and Anthropic profiles. Credentials and endpoint overrides come from the provider-standard `DEEPSEEK_API_KEY` / `DEEPSEEK_BASE_URL`, `OPENAI_API_KEY` / `OPENAI_BASE_URL`, and `ANTHROPIC_API_KEY` / `ANTHROPIC_BASE_URL` pairs in the boot's layered environment.

--- a/apps/cli/README.zh.md
+++ b/apps/cli/README.zh.md
@@ -17,8 +17,7 @@ TUI 界面：

 `dsh upgrade` 是默认 TUI 界面之上的引导式全新会话入口：它在调用目录中创建一个全新会话，并以内置 `dsh-upgrade` skill 播种其首轮，效果等同于用户手动键入 `/skill:<name>`。启动器将 skill 名称提供到启动上下文（[`INITIAL_SKILL_KEY`](../../packages/ui/tui/README.md)），TUI 在聊天就绪后自动调用它。两者都不接受任何选项——`--config`、`-p`、`--resume` 都会明确报错——且仅在首次启动时播种，因此之后 `dsh --resume <id>` 恢复该会话时是普通 TUI 会话，不会重复注入。

-
-Web 和无头界面启动 `base.cordis.yml` 与 `web.cordis.yml`，随后应用 `$DSH_HOME/config.yaml`；显式的 `--config <path>` 会替代该个人覆盖。除此之外，两者共享同一套组合：两者都将调用目录视为默认项目和 Workspace 根目录，除非通过 `--workspace-root <path>` 覆盖，否则会在该根目录下创建具名 Workspace；它们会把适用的 `AGENTS.md`/`CLAUDE.md` 指令加载到每个 agent-loop 请求前缀中，渲染预算为 65,536 字节，并选用首条消息模型标题，且挂载一个可丢弃的内存 SQLite 内容索引服务。该服务在启动时处于 ACTIVE 状态，但其 `node:sqlite` 模块与数据库句柄分别要到首次内容搜索才会导入和打开。这样可使 Node 22 在尚未使用搜索时的启动输出不出现 SQLite 实验性警告；首次实际搜索仍可能发出运行时警告。每个服务实例独占自己的数据库，因此并行调用既不会共享不受支持的 SQLite 状态，也不会留下派生索引文件，首次搜索还会惰性对账实时日志与持久化日志。无头界面唯一的差异是监听操作系统分配的端口（并行 `dsh -p` 运行绝不冲突；stderr 打印的 URL 会在浏览器中打开实时会话）。两者都需要先构建前端 dist 和客户端 bundle（`pnpm run build && pnpm run build:web`）。
+Web 和无头界面启动 `base.cordis.yml` 与 `web.cordis.yml`，随后应用 `$DSH_HOME/config.yaml`；显式的 `--config <path>` 会替代该个人覆盖。除此之外，两者共享同一套组合：两者都将调用目录视为默认项目和 Workspace 根目录，除非通过 `--workspace-root <path>` 覆盖，否则会在该根目录下创建具名 Workspace；它们会把适用的 `AGENTS.md`/`CLAUDE.md` 指令加载到每个 agent-loop 请求前缀中，渲染预算为 65,536 字节，选用首条消息模型标题，采用与 TUI 相同的有界暂时性模型请求重试策略，并挂载一个可丢弃的内存 SQLite 内容索引服务。该服务在启动时处于 ACTIVE 状态，但其 `node:sqlite` 模块与数据库句柄分别要到首次内容搜索才会导入和打开。这样可使 Node 22 在尚未使用搜索时的启动输出不出现 SQLite 实验性警告；首次实际搜索仍可能发出运行时警告。每个服务实例独占自己的数据库，因此并行调用既不会共享不受支持的 SQLite 状态，也不会留下派生索引文件，首次搜索还会惰性对账实时日志与持久化日志。无头界面唯一的差异是监听操作系统分配的端口（并行 `dsh -p` 运行绝不冲突；stderr 打印的 URL 会在浏览器中打开实时会话）。两者都需要先构建前端 dist 和客户端 bundle（`pnpm run build && pnpm run build:web`）。

 已交付的 TUI 和 Web 组合会注册原生 DeepSeek 适配器，以及 pi-ai 的 OpenAI 和 Anthropic 提供方配置。凭据和端点覆盖来自启动分层环境中的提供方标准变量对：`DEEPSEEK_API_KEY` / `DEEPSEEK_BASE_URL`、`OPENAI_API_KEY` / `OPENAI_BASE_URL` 和 `ANTHROPIC_API_KEY` / `ANTHROPIC_BASE_URL`。

--- a/apps/web/tests/live-interactions.e2e.ts
+++ b/apps/web/tests/live-interactions.e2e.ts
@@ -221,8 +221,8 @@ describe('web e2e: live-turn interactions (cancel / error / retry)', () => {
    // only on change, so attempt count is invisible there).
    expect(sessionEvents.filter(e => e.type === 'llm/retry').length).toBeGreaterThanOrEqual(1)
    await expect.poll(() => page.getByText('event sourcing', { exact: false }).count(), { timeout: 10_000 }).toBeGreaterThan(0)
-    // Golden of the recovered end-state: indistinguishable from a clean
-    // completion — retries are deliberately invisible in the transcript.
+    // Golden of the recovered end-state: the discarded partial stays absent,
+    // while the settled retry row remains as durable recovery context.
    const snapshot = await captureStableAria(page, '[class*="centerCol"]', scaffold!.workspaceCwd)
    await compareOrRefreshGolden(RETRY_EXPECTED, snapshot, MODE)
    expect(tripwire.pageErrors).toEqual([])
--- a/apps/web/tests/smoke-real.e2e.ts
+++ b/apps/web/tests/smoke-real.e2e.ts
@@ -267,6 +267,98 @@ describe('dsh web keyless CLI smoke', () => {
    }
  })

+  it('retries a partial transport failure through the shipped Web composition', async () => {
+    requireDist()
+    const workspace = mkdtempSync(join(tmpdir(), 'dsh-web-retry-'))
+    const promptMarker = 'WEB_RETRY_REQUEST'
+    const recoveredMarker = 'WEB_RETRY_RECOVERED'
+    let mainAttempts = 0
+    const provider = createServer((request, response) => {
+      let body = ''
+      request.setEncoding('utf8')
+      request.on('data', (chunk: string) => { body += chunk })
+      request.on('end', () => {
+        const parsed = JSON.parse(body) as { max_tokens?: number; messages?: unknown[] }
+        const titleRequest = parsed.max_tokens === 64
+        const mainRequest = !titleRequest && body.includes(promptMarker)
+        response.writeHead(200, { 'content-type': 'text/event-stream' })
+        if (!mainRequest) {
+          response.end([
+            'data: {"choices":[{"delta":{"content":"Web retry title"}}]}',
+            'data: {"choices":[{"delta":{"content":""},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}',
+            'data: [DONE]',
+            '',
+          ].join('\n\n'))
+          return
+        }
+        mainAttempts++
+        if (mainAttempts === 1) {
+          response.write('data: {"choices":[{"delta":{"content":"WEB_RETRY_DISCARDED"}}]}\n\n')
+          setTimeout(() => { response.destroy() }, 20)
+          return
+        }
+        response.end([
+          `data: {"choices":[{"delta":{"content":"${recoveredMarker}"}}]}`,
+          'data: {"choices":[{"delta":{"content":""},"finish_reason":"stop"}],"usage":{"prompt_tokens":3,"completion_tokens":1}}',
+          'data: [DONE]',
+          '',
+        ].join('\n\n'))
+      })
+    })
+    await new Promise<void>(resolve => provider.listen(0, '127.0.0.1', resolve))
+    const address = provider.address()
+    if (address === null || typeof address === 'string') throw new Error('mock provider did not bind a TCP port')
+    const tsxLoader = pathToFileURL(createRequire(join(REPO_ROOT, 'package.json')).resolve('tsx')).href
+    const child = spawn(
+      process.execPath,
+      ['--import', tsxLoader, join(REPO_ROOT, 'apps/cli/src/bin.ts'), 'web', '--port', '0'],
+      {
+        cwd: workspace,
+        env: {
+          ...process.env,
+          DEEPSEEK_API_KEY: 'keyless-web-retry',
+          DEEPSEEK_BASE_URL: `http://127.0.0.1:${address.port}`,
+          DSH_HOME: join(workspace, '.dsh'),
+          TSX_TSCONFIG_PATH: join(REPO_ROOT, 'tsconfig.json'),
+        },
+        stdio: ['ignore', 'pipe', 'pipe'],
+      },
+    )
+    try {
+      const baseUrl = await waitForReadyLine(child)
+      const created = await rpc<{ sessionId: string }>(baseUrl, 'session.create', {})
+      await rpc<{ accepted: true }>(baseUrl, 'session.prompt', {
+        sessionId: created.sessionId,
+        mode: 'queue',
+        content: [{ type: 'text', text: promptMarker }],
+      })
+      let page: HistoryPage | undefined
+      await expect.poll(async () => {
+        page = await history(baseUrl, created.sessionId)
+        return hasAssistantMarker(page, recoveredMarker)
+      }, { timeout: 20_000 }).toBe(true)
+      if (page === undefined) throw new Error('retry history was not observed')
+      const retry = page.events.find(({ event }) => event.type === 'llm/retry')?.event
+      expect(mainAttempts).toBe(2)
+      expect(retry?.data).toMatchObject({
+        turn: 1,
+        step: 1,
+        retry: 1,
+        maxRetries: 2,
+        failure: { code: 'TRANSPORT' },
+      })
+      expect(JSON.stringify(page.events)).toContain('WEB_RETRY_DISCARDED')
+    } finally {
+      const closed = child.exitCode === null
+        ? new Promise<void>((resolveClose) => { child.once('close', () => { resolveClose() }) })
+        : Promise.resolve()
+      if (child.exitCode === null) child.kill('SIGTERM')
+      await closed
+      await new Promise<void>(resolveClose => provider.close(() => { resolveClose() }))
+      rmSync(workspace, { recursive: true, force: true })
+    }
+  }, 30_000)
+
  it('DSH_TOOLS_MODE=code collapses the provider wire tools to run_code with the SDK prompt section', async () => {
    requireDist()
    const workspace = mkdtempSync(join(tmpdir(), 'dsh-web-code-mode-'))
--- a/apps/web/tests/snapshots/live-interactions/retry.expected.md
+++ b/apps/web/tests/snapshots/live-interactions/retry.expected.md
@@ -11,6 +11,8 @@
  - img
 - button "Edit":
  - img
+- group:
+  - status: Retried model request (1/2) · {{duration}}
 - button "Think The user is asking for a one-sentence description of event sourcing. This is a straightforward knowledge question that doesn't require any skill loading or tool calls.":
  - img
  - img
--- a/docs/config-catalog.md
+++ b/docs/config-catalog.md
@@ -776,7 +776,7 @@ Requires: `agents`
 export type Config = Readonly<Record<string, never>>
 ```

-Source: [`packages/llm/llm-retry/src/index.ts:45`](../packages/llm/llm-retry/src/index.ts)
+Source: [`packages/llm/llm-retry/src/index.ts:47`](../packages/llm/llm-retry/src/index.ts)

 ## `@deepseek-ai/dsh-lsp-local`

--- a/packages/client/connection/src/client/fixture.ts
+++ b/packages/client/connection/src/client/fixture.ts
@@ -1126,6 +1126,8 @@ export function createFixtureApi(options: FixtureOptions = {}): ApiProxy {
  let failNextHistory = false
  /** Force-enders for currently open stream generators (timing hook: simulated connection loss). */
  const streamBreakers = new Set<() => void>()
+  /** Retry scenarios opened by timing hooks and completed in a later browser assertion phase. */
+  const retryScenarios = new Map<SessionId, { turn: number; stepStarted: boolean }>()

  // Timing-acceptance hooks (browser test backdoor): the in-memory fixture is ideally timed, which
  // is exactly what masked the open-window and reconnect-gap bugs (audit S1/S3). These let
@@ -1149,6 +1151,89 @@ export function createFixtureApi(options: FixtureOptions = {}): ApiProxy {
      const messageSeqs = log.filter(event => event.type === 'user/message').map(event => event.seq)
      append(sid(id), { type: 'session/title', data: { title, messageSeqs, source: { kind: 'provider', provider: 'fixture' } } })
    },
+    /** Open one failed model step whose partial remains visible until llm/retry arrives. */
+    beginModelRetry(id: string): void {
+      const sessionId = sid(id)
+      const turn = nextTurn.get(sessionId) ?? 0
+      nextTurn.set(sessionId, turn + 1)
+      retryScenarios.set(sessionId, { turn, stepStarted: true })
+      setRunning(sessionId, true)
+      append(sessionId, { type: 'turn/start', data: { turn, trigger: { kind: 'message', source: { kind: 'user' } } } })
+      append(sessionId, { type: 'user/message', surfaceOp: 'append', data: { content: text('请重试这个请求'), source: { kind: 'user' } } })
+      append(sessionId, { type: 'step/start', data: { turn, step: 1 } })
+      append(sessionId, { type: 'assistant/chunk', data: { turn, step: 1, chunk: { type: 'block-start', index: 0, blockType: 'text' } } })
+      append(sessionId, { type: 'assistant/chunk', data: { turn, step: 1, chunk: { type: 'text-delta', index: 0, text: '应撤回的半截回复' } } })
+      append(sessionId, { type: 'step/end', data: { turn, step: 1 } })
+    },
+    /** Record one retry decision, then open the next retry turn. */
+    scheduleModelRetry(id: string, retry = 1, delayMs = 450): void {
+      const sessionId = sid(id)
+      const scenario = retryScenarios.get(sessionId)
+      if (scenario === undefined) throw new Error(`fixture: no model retry scenario for ${id}`)
+      if (!scenario.stepStarted) {
+        append(sessionId, { type: 'step/start', data: { turn: scenario.turn, step: 1 } })
+        append(sessionId, { type: 'assistant/chunk', data: { turn: scenario.turn, step: 1, chunk: { type: 'block-start', index: 0, blockType: 'text' } } })
+        append(sessionId, { type: 'assistant/chunk', data: { turn: scenario.turn, step: 1, chunk: { type: 'text-delta', index: 0, text: `第 ${String(retry)} 次应撤回的回复` } } })
+        append(sessionId, { type: 'step/end', data: { turn: scenario.turn, step: 1 } })
+        scenario.stepStarted = true
+      }
+      const failure = { code: 'TRANSPORT', message: '连接被重置' }
+      append(sessionId, {
+        type: 'llm/retry',
+        data: {
+          turn: scenario.turn, step: 1,
+          provider: 'fixture', mode: 'normal', policyKey: 'fixture-normal',
+          retry, maxRetries: 2, delayMs, failure,
+        },
+      })
+      append(sessionId, {
+        type: 'turn/end',
+        data: { turn: scenario.turn, reason: { kind: 'error', step: 1, failure } },
+      })
+      const next = nextTurn.get(sessionId) ?? scenario.turn + 1
+      nextTurn.set(sessionId, next + 1)
+      append(sessionId, { type: 'turn/start', data: { turn: next, trigger: { kind: 'retry' } } })
+      scenario.turn = next
+      scenario.stepStarted = false
+    },
+    /** Record one retry decision, then cancel its source turn before the retry starts. */
+    cancelModelRetryDuringBackoff(id: string, delayMs = 450): void {
+      const sessionId = sid(id)
+      const scenario = retryScenarios.get(sessionId)
+      if (scenario === undefined) throw new Error(`fixture: no model retry scenario for ${id}`)
+      const failure = { code: 'TRANSPORT', message: '连接被重置' }
+      append(sessionId, {
+        type: 'llm/retry',
+        data: {
+          turn: scenario.turn, step: 1,
+          provider: 'fixture', mode: 'normal', policyKey: 'fixture-normal',
+          retry: 1, maxRetries: 2, delayMs, failure,
+        },
+      })
+      append(sessionId, { type: 'turn/end', data: { turn: scenario.turn, reason: { kind: 'aborted' } } })
+      retryScenarios.delete(sessionId)
+      setRunning(sessionId, false)
+    },
+    /** Finish the timing-hook retry with a finalized response in the open retry turn. */
+    completeModelRetry(id: string): void {
+      const sessionId = sid(id)
+      const scenario = retryScenarios.get(sessionId)
+      if (scenario === undefined) throw new Error(`fixture: no model retry scenario for ${id}`)
+      retryScenarios.delete(sessionId)
+      append(sessionId, { type: 'step/start', data: { turn: scenario.turn, step: 1 } })
+      append(sessionId, {
+        type: 'assistant/message',
+        surfaceOp: 'append',
+        data: {
+          turn: scenario.turn,
+          step: 1,
+          message: assistantMessage(text('重试后的完整回复')),
+        },
+      })
+      append(sessionId, { type: 'step/end', data: { turn: scenario.turn, step: 1 } })
+      append(sessionId, { type: 'turn/end', data: { turn: scenario.turn, reason: { kind: 'completed' } } })
+      setRunning(sessionId, false)
+    },
    /** Log append WITHOUT the mux emit: a frame lost in transit — history still serves it, the client must repull. */
    appendSilent(id: string, msg: string): void {
      const log = logOf(sid(id))
--- a/packages/client/connection/tests/fixture.spec.ts
+++ b/packages/client/connection/tests/fixture.spec.ts
@@ -19,6 +19,10 @@ interface TimingHooks {
  failNextHistory(): void
  appendUser(id: string, msg: string): void
  appendTitle(id: string, title: string): void
+  beginModelRetry(id: string): void
+  scheduleModelRetry(id: string, retry?: number, delayMs?: number): void
+  cancelModelRetryDuringBackoff(id: string, delayMs?: number): void
+  completeModelRetry(id: string): void
  appendSilent(id: string, msg: string): void
  breakStreams(): void
 }
@@ -814,8 +818,18 @@ describe('createFixtureApi', () => {
    hooks.appendSilent('fx-alpha', '静默丢帧')
    hooks.appendUser('fx-alpha', '正常直播')
    hooks.appendTitle('fx-alpha', 'Fixture 修订标题')
+    hooks.beginModelRetry('fx-alpha')
+    hooks.scheduleModelRetry('fx-alpha')
+    hooks.completeModelRetry('fx-alpha')
+    hooks.beginModelRetry('fx-alpha')
+    hooks.cancelModelRetryDuringBackoff('fx-alpha')
    await vi.waitFor(() => {
      expect(seen.some(f => f.type === 'session/event' && JSON.stringify(f.event.data).includes('正常直播'))).toBe(true)
+      expect(seen.some(f => f.type === 'session/event' && (f.event as { type: string }).type === 'llm/retry')).toBe(true)
+      expect(seen.some(f => f.type === 'session/event' && JSON.stringify(f.event.data).includes('重试后的完整回复'))).toBe(true)
+      expect(seen.some(f => f.type === 'session/event'
+        && f.event.type === 'turn/end'
+        && f.event.data.reason.kind === 'aborted')).toBe(true)
      expect(seen.some(f => f.type === 'session/projection' && f.key === 'title' && f.value === 'Fixture 修订标题')).toBe(true)
    })
    expect(seen.some(f => f.type === 'session/event' && JSON.stringify(f.event.data).includes('静默丢帧'))).toBe(false)
--- a/packages/client/runtime/README.i18n.yaml
+++ b/packages/client/runtime/README.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/runtime/README.md
-README.md: 12023868c577ebcae6898d13358a2456295496c2
-README.zh.md: 7ef4c93d36b3f0b32c0bfcf8a38892260240c74f
+README.md: 9f2b165f1a98dcecfa3ab82386da9b094cfd2f54
+README.zh.md: 3ed047e65d3bddc14c3b6b84f327bbeebf805d4b
--- a/packages/client/runtime/README.md
+++ b/packages/client/runtime/README.md
@@ -30,6 +30,10 @@ SlotsService gives the renderer separate bare observables for `useSessions` and

 `SessionManager` retains the latest validated `session/title` control snapshot independently of list and session-instance arrival. Newer event seqs replace older snapshots, title timestamps contribute to list recency, and a subscription baseline discards any retained title beyond its `lastSeq` before the optional folded title arrives. Explicit session removal also clears the retained title. The client-facing `SessionSummary.title` is therefore only the actual durable title; `displayTitle` is always present and falls back through the cwd basename and session id. A cold persisted session keeps that fallback until opening or resuming it causes the host to fold and project its log-backed title. `ISession.rename` settles the `title` projection cell directly from the unary response's `{title, seq}` under the same higher-seq-wins rule — the list row and every `useProjection('title')` reader update ahead of the push frame, whose later replay of the same seq is a no-op.

+## Model retry projection
+
+The Session object validates plugin-owned, provider-routed `llm/retry` payloads at the event wire boundary against the producer's complete field contract, including timer, integer, status, provider-delay, and non-empty diagnostic bounds. A valid event removes the matching failed step's streaming partial and inserts a durable retry notice at the event's sequence position. The notice is `scheduled` until a following retry turn starts; an aborted or disposed source turn marks it `cancelled`, while the retry turn marks it `started`. Normal-mode notices carry their finite maximum; always-mode notices remain explicitly unbounded. Window rebuild and history replay apply the same projection, so logged chunks from the discarded attempt never reappear as an interrupted reply after refresh. A terminal turn without `llm/retry` retains the existing behavior: visible unfinalized output is frozen as an interrupted assistant node.
+
 ## Session forking

 `ISessions.fork({sessionId, atSeq?, increaseTitle?})` resolves only after the child summary is locally addressable, carrying source lineage and cwd with `blank: false`; callers choose whether to open it. With `increaseTitle: true`, the client renames the child from the source session's persisted title: a trailing `(N)` or `（N）` is incremented without changing bracket style, while any other title gets ` (1)` appended; the rename is skipped when the source has no persisted title, and a rename failure rejects the promise but leaves the created child in place. This option is not sent in the Host fork request. A `workspace-attach-failed` response still identifies a child already published by the Host, so `SessionManager` reconciles that partial success before `SessionForkError` reaches the caller instead of making a retry create a duplicate child.
--- a/packages/client/runtime/README.zh.md
+++ b/packages/client/runtime/README.zh.md
@@ -30,6 +30,10 @@ SlotsService 分别为 renderer 提供 `useSessions` 与 `useWorkspaces` 的裸

 `SessionManager` 独立于列表和 Session 实例到达情况，保留最近一次通过验证的 `session/title` 控制快照。seq 更高的事件会替换旧快照，标题时间戳计入列表新近程度；订阅基线会先丢弃 seq 超过其 `lastSeq` 的任何已保留标题，再接收可选的折叠标题。显式移除 Session 也会清除已保留标题。因此，面向客户端的 `SessionSummary.title` 只包含实际的持久化标题；`displayTitle` 始终存在，并依次回退到 cwd basename 和 Session id。冷态持久化会话会保持该回退值，直到打开或恢复会话，促使主机折叠并投影由日志支撑的标题。`ISession.rename` 用 unary 响应中的 `{title, seq}` 直接结算 `title` 投影格，遵循同一 seq 高者胜规则——列表行和所有 `useProjection('title')` 读者在推送帧到达前即更新；推送帧随后重放同一 seq 时为无操作。

+## 模型重试投影
+
+Session 对象会在事件 wire 边界依据生产方的完整字段契约，验证由插件负责、按提供方路由的 `llm/retry` 载荷，包括计时器、整数、状态、提供方延迟和非空诊断字段的边界。有效事件会移除对应失败步骤的流式输出片段，并在该事件的序列位置插入一条持久的重试提示。该提示在后续重试轮次开始前为 `scheduled`；源轮次中止或释放会将其标记为 `cancelled`，重试轮次则会将其标记为 `started`。normal mode 提示携带其有限上限；always mode 提示则保持显式无界。窗口重建与历史回放应用相同的投影，因此刷新后，来自已丢弃尝试的日志分片绝不会重新显示为中断回复。没有 `llm/retry` 的终止轮次保留现有行为：可见但尚未定稿的输出会冻结为中断的 assistant 节点。
+
 ## 会话 fork

 `ISessions.fork({sessionId, atSeq?, increaseTitle?})` 只在子会话摘要已能在本地寻址后才完成；该摘要携带源会话的谱系和 cwd，且 `blank: false`，由调用方决定是否打开。`increaseTitle: true` 会在 client 端把源会话的持久化标题改名到子会话：尾部 `(N)` 或 `（N）` 递增并保留括号样式，其余标题追加 ` (1)`；源会话没有持久化标题时跳过改名，改名失败时拒绝 promise 但保留已创建的子会话。该选项不会进入 Host fork 请求。即使响应为 `workspace-attach-failed`，其中仍会标识 Host 已发布的子会话，因此 `SessionManager` 会先将这一部分成功对账，再让 `SessionForkError` 到达调用方，避免重试创建重复的子会话。
--- a/packages/client/runtime/package.json
+++ b/packages/client/runtime/package.json
@@ -36,6 +36,7 @@
    "@deepseek-ai/dsh-client-ui-slots": "workspace:^",
    "@deepseek-ai/dsh-host-apiproxy": "workspace:^",
    "@deepseek-ai/dsh-llm": "workspace:^",
+    "@deepseek-ai/dsh-llm-retry": "workspace:^",
    "@deepseek-ai/dsh-session": "workspace:^",
    "@deepseek-ai/dsh-session-projection": "workspace:^",
    "@deepseek-ai/dsh-session-title": "workspace:^",
@@ -49,6 +50,7 @@
  },
  "devDependencies": {
    "@deepseek-ai/dsh-invariants": "workspace:^",
+    "@deepseek-ai/dsh-timeout": "workspace:^",
    "@types/react": "~18.3.1",
    "cordis": "^4.0.0-rc.7"
  },
--- a/packages/client/runtime/src/client/index.ts
+++ b/packages/client/runtime/src/client/index.ts
@@ -45,7 +45,7 @@ export type {
 export type {
  AssistantBlock, AssistantMessageNode, AssistantProvenanceView, AssistantRequestConfig,
  AssistantTiming, CodeSubCall, CommandNode, ComposerPhase, ContextMessageNode, ConversationNode,
-  ConversationSnapshot, QueuedMessage, RunningToolCall,
+  ConversationSnapshot, ModelRetryNode, QueuedMessage, RunningToolCall,
  SteeringMessageNode, TodoItem, ToolResultNode, UnknownSurfaceNode, UserMessageNode,
 } from './sessions/conversation.ts'
 export type {
--- a/packages/client/runtime/src/client/sessions/conversation.ts
+++ b/packages/client/runtime/src/client/sessions/conversation.ts
@@ -5,6 +5,7 @@

 import type { CommandId } from '@deepseek-ai/dsh-commands/brand'
 import type { ContentBlock } from '@deepseek-ai/dsh-llm/types'
+import type { LlmRetryEventData } from '@deepseek-ai/dsh-llm-retry/types'
 import type { TodoItem } from '@deepseek-ai/dsh-session/types'
 import type {
  InboxItemId, RpcError, SessionId, ToolCallView, ToolResultView,
@@ -121,6 +122,19 @@ export interface ContextMessageNode {
  source: unknown
 }

+/** Durable notice that a closed failed step is waiting for a model-request retry. */
+export type ModelRetryNode = LlmRetryEventData & {
+  kind: 'model-retry'
+  seq: number
+  /** Unix epoch ms from the llm/retry session event. */
+  time: number
+  /**
+   * Client-derived lifecycle: scheduled until a retry turn starts, started
+   * once it does, or cancelled when the failed turn aborts first.
+   */
+  retryState: 'scheduled' | 'started' | 'cancelled'
+}
+
 /** A tool result paired (when in-window) with its call head. */
 export interface ToolResultNode {
  kind: 'tool-result'
@@ -183,6 +197,7 @@ export type ConversationNode =
  | AssistantMessageNode
  | SteeringMessageNode
  | ContextMessageNode
+  | ModelRetryNode
  | ToolResultNode
  | CommandNode
  | UnknownSurfaceNode
@@ -265,7 +280,7 @@ export interface PromptError {
 /** The immutable snapshot contract Session hands to uSES (see the web client architecture RFC). */
 export interface ConversationSnapshot {
  sessionId: SessionId
-  /** Surface fold product (finalized conversation nodes in surface order). */
+  /** Finalized surface events and durable operational notices in event order. */
  nodes: readonly ConversationNode[]
  /** Fold degradation flag (cross-window replace defense): when true, nodes come from the lenient linear scan. */
  foldDegraded: boolean
--- a/packages/client/runtime/src/client/sessions/session.ts
+++ b/packages/client/runtime/src/client/sessions/session.ts
@@ -2,6 +2,7 @@

 import type { Context } from 'cordis'
 import type { ContentBlock } from '@deepseek-ai/dsh-llm/types'
+import type { LlmRetryEventData } from '@deepseek-ai/dsh-llm-retry/types'
 import type { SessionEvent } from '@deepseek-ai/dsh-session/types'
 import type {
  HistoryEntry, IApiClient, InboxItemId, MuxFrame, QueueAction, RpcError,
@@ -12,8 +13,8 @@ import type {
 import { transportError } from '@deepseek-ai/dsh-host-apiproxy/api'
 import type { SessionFace } from '../contract/session.ts'
 import type {
-  CodeSubCall, ComposerPhase, ConversationNode, ConversationSnapshot, OpenState,
-  PromptError, QueuedMessage, RunningToolCall,
+  CodeSubCall, ComposerPhase, ConversationNode, ConversationSnapshot, ModelRetryNode,
+  OpenState, PromptError, QueuedMessage, RunningToolCall,
 } from './conversation.ts'
 import type { PendingInteraction } from './pending.ts'
 import { PendingWait } from './pending.ts'
@@ -26,6 +27,10 @@ import type { ProjectionsBaseline } from './projection-store.ts'
 /** Messages requested per history page. */
 export const PAGE_MESSAGES = 50

+// Browser bundles cannot value-import the host timeout library. This protocol
+// bound is pinned to @deepseek-ai/dsh-timeout's MAX_TIMER_DELAY_MS in tests.
+const MAX_RETRY_DELAY_MS = 2_147_483_647
+
 /** Manager-owned observers of a Session object's local state edges. */
 export interface SessionOptions {
  /**
@@ -88,9 +93,9 @@ export class Session implements SessionFace {
  private readonly foldAdapter = new FoldAdapter()
  private partial: PartialAccumulator | null = null
  private openCalls = new Map<string, RunningToolCall>()
-  /** Interrupted-turn terminal nodes (frozen partial text / aborted tool cards), merged into the flow by seq.
-   *  Derived from window events (turn/end sweep) — rebuilt by rebuildDerivedFromWindow like partial/openCalls. */
-  private frozenNodes: ConversationNode[] = []
+  /** Operational notices and interrupted-turn terminal nodes merged into the flow by seq.
+   *  Derived from window events — rebuilt by rebuildDerivedFromWindow like partial/openCalls. */
+  private derivedNodes: ConversationNode[] = []
  private pending = new Map<string, PendingInteraction>()
  // Revision counters preserve array identity when derived content is unchanged, so
  // React.memo children survive unrelated snapshot swaps (chunk storms must not re-render every
@@ -100,12 +105,12 @@ export class Session implements SessionFace {
  private callsCache: { rev: number; value: RunningToolCall[] } | null = null
  private pendingRev = 0
  private pendingCache: { rev: number; value: PendingInteraction[] } | null = null
+  private derivedRev = 0
+  private nodesCache: { folded: readonly ConversationNode[]; derivedRev: number; value: readonly ConversationNode[] } | null = null
  /** Authoritative stream-only inbox snapshot; pending work never hits history. */
  private queued: QueuedMessage[] = []
  private queueRev = 0
  private queueCache: { rev: number; value: QueuedMessage[] } | null = null
-  private frozenRev = 0
-  private nodesCache: { folded: readonly ConversationNode[]; frozenRev: number; value: readonly ConversationNode[] } | null = null
  /** `run_code` sub-dispatches by parent callId (window-derived, like openCalls). Appends
   *  copy-on-write the per-parent array so published snapshot references never mutate. */
  private codeDispatches = new Map<string, readonly CodeSubCall[]>()
@@ -625,8 +630,28 @@ export class Session implements SessionFace {
  }

  /** Per-event side effects (right column of the §A.9 dispatch table):
-   *  chunk accumulation / partial clear on finalize / openCalls add-remove. */
+   *  chunk/retry projection and openCalls add-remove. */
  private applyEventSideEffects(event: SessionEvent, view?: ToolEventView): void {
+    const eventType = event.type as string
+    if (eventType === 'llm/retry') {
+      const data = parseRetryEventData(event.data)
+      if (data === null) {
+        console.error(`[web-runtime] ignored malformed llm/retry event at seq ${event.seq}`)
+        return
+      }
+      if (this.partial !== null && this.partial.turn === data.turn && this.partial.step === data.step) {
+        this.partial = null
+      }
+      this.derivedNodes.push({
+        kind: 'model-retry',
+        seq: event.seq,
+        time: event.time,
+        retryState: 'scheduled',
+        ...data,
+      })
+      this.derivedRev++
+      return
+    }
    // The `tool/code-dispatch-start`/`tool/code-dispatch` pair is declared by
    // the host-side dsh-tools plugin whose types cannot enter the client
    // program (its host Context merges collide with the client's), so this
@@ -687,6 +712,10 @@ export class Session implements SessionFace {
      return
    }
    switch (event.type) {
+      case 'turn/start': {
+        if (event.data.trigger.kind === 'retry') this.settleScheduledRetry('started')
+        return
+      }
      case 'assistant/chunk': {
        const { turn, step, chunk } = event.data
        if (this.partial === null || this.partial.turn !== turn || this.partial.step !== step) {
@@ -715,6 +744,9 @@ export class Session implements SessionFace {
        return
      }
      case 'turn/end': {
+        if (event.data.reason.kind === 'aborted' || event.data.reason.kind === 'disposed') {
+          this.settleScheduledRetry('cancelled', event.data.turn)
+        }
        // Aborted turns never finalize. The accumulated partial is VALUE, not residue: freeze it
        // into an interrupted terminal node (pulse stops, text survives) instead of deleting it.
        // Shared by live and window-replay paths, so a refresh reconstructs the same frozen node
@@ -724,12 +756,12 @@ export class Session implements SessionFace {
          const visible = blocks.some(b => (b.kind === 'text' || b.kind === 'reasoning' ? b.text !== '' : true))
          if (visible) {
            // Fractional seq: strictly after every event of this turn (all < turn/end seq), before the next turn.
-            this.frozenNodes.push({
+            this.derivedNodes.push({
              kind: 'assistant', seq: event.seq - 0.9, time: event.time,
              turn: this.partial.turn, step: this.partial.step,
              blocks, interrupted: true,
            })
-            this.frozenRev++
+            this.derivedRev++
          }
          this.partial = null
        }
@@ -739,7 +771,7 @@ export class Session implements SessionFace {
          this.openCalls.delete(callId)
          this.callsRev++
          // The spinner card becomes an interrupted terminal card (never vanishes mid-flow).
-          this.frozenNodes.push({
+          this.derivedNodes.push({
            kind: 'tool-result', seq: event.seq - 0.8 + callOffset++ * 0.01, time: event.time,
            callId,
            call: { name: call.name, argsRaw: call.argsRaw },
@@ -747,7 +779,7 @@ export class Session implements SessionFace {
            content: [], isError: true, error: { name: 'Interrupted', code: 'interrupted' },
            callView: call.callView, resultView: null,
          })
-          this.frozenRev++
+          this.derivedRev++
        }
        return
      }
@@ -756,15 +788,36 @@ export class Session implements SessionFace {
    }
  }

-  /** Re-derive state (partial/openCalls/frozenNodes) from raw window events after a rebuild — keeps
-   *  paging/stitching consistent, and makes the live freeze and the history replay converge on the
-   *  same interrupted nodes (chunks are logged, so the replayed sweep re-freezes identical text). */
+  /**
+   * Settle the newest scheduled retry, optionally restricted to its failed turn.
+   * @param retryState - next client projection state to publish.
+   * @param turn - failed turn required for cancellation; omitted for the next retry turn start.
+   */
+  private settleScheduledRetry(
+    retryState: Exclude<ModelRetryNode['retryState'], 'scheduled'>,
+    turn?: number,
+  ): void {
+    const index = this.derivedNodes.findLastIndex(node =>
+      node.kind === 'model-retry'
+      && node.retryState === 'scheduled'
+      && (turn === undefined || node.turn === turn))
+    if (index < 0) return
+    const node = this.derivedNodes[index]
+    /* v8 ignore next -- findLastIndex's predicate narrows the indexed node only at runtime. */
+    if (node?.kind !== 'model-retry') return
+    this.derivedNodes[index] = { ...node, retryState }
+    this.derivedRev++
+  }
+
+  /** Re-derive state (partial/openCalls/derivedNodes) from raw window events after a rebuild — keeps
+   *  paging/stitching consistent, and makes live handling and history replay converge on the same
+   *  retry notices and interrupted nodes. */
  private rebuildDerivedFromWindow(): void {
    this.partial = null
    this.openCalls.clear()
    this.callsRev++
-    this.frozenNodes = []
-    this.frozenRev++
+    this.derivedNodes = []
+    this.derivedRev++
    this.codeDispatches = new Map()
    this.dispatchesRev++
    for (let i = 0; i < this.events.length; i++) {
@@ -781,17 +834,17 @@ export class Session implements SessionFace {

  private buildSnapshot(): ConversationSnapshot {
    const { nodes: folded, degraded } = this.foldAdapter.nodes()
-    // Frozen interrupted nodes ride fractional seqs: a stable merge keeps them in flow order.
-    // The merged array is cached on (folded reference, frozenRev) so an unchanged flow keeps its
+    // Derived nodes use their event seq or a nearby fractional seq: a stable merge keeps flow order.
+    // The merged array is cached on (folded reference, derivedRev) so an unchanged flow keeps its
    // reference across snapshot swaps (§A.9.4).
    let nodes: readonly ConversationNode[]
-    if (this.nodesCache !== null && this.nodesCache.folded === folded && this.nodesCache.frozenRev === this.frozenRev) {
+    if (this.nodesCache !== null && this.nodesCache.folded === folded && this.nodesCache.derivedRev === this.derivedRev) {
      nodes = this.nodesCache.value
    } else {
-      nodes = this.frozenNodes.length === 0
+      nodes = this.derivedNodes.length === 0
        ? folded
-        : [...folded, ...this.frozenNodes].sort((a, b) => a.seq - b.seq)
-      this.nodesCache = { folded, frozenRev: this.frozenRev, value: nodes }
+        : [...folded, ...this.derivedNodes].sort((a, b) => a.seq - b.seq)
+      this.nodesCache = { folded, derivedRev: this.derivedRev, value: nodes }
    }
    if (this.callsCache === null || this.callsCache.rev !== this.callsRev) {
      this.callsCache = { rev: this.callsRev, value: [...this.openCalls.values()] }
@@ -835,6 +888,58 @@ export class Session implements SessionFace {
  }
 }

+/** Validate the plugin-owned payload at the session-event wire boundary. */
+function parseRetryEventData(value: unknown): LlmRetryEventData | null {
+  if (value === null || typeof value !== 'object') return null
+  const data = value as Record<string, unknown>
+  const failure = data.failure
+  if (failure === null || typeof failure !== 'object') return null
+  const failureData = failure as Record<string, unknown>
+  if (!nonNegativeSafeInteger(data.turn)
+    || !nonNegativeSafeInteger(data.step)
+    || typeof data.provider !== 'string'
+    || data.provider.length === 0
+    || typeof data.policyKey !== 'string'
+    || data.policyKey.length === 0
+    || !positiveSafeInteger(data.retry)
+    || typeof data.delayMs !== 'number'
+    || !Number.isFinite(data.delayMs)
+    || data.delayMs < 0
+    || data.delayMs > MAX_RETRY_DELAY_MS
+    || typeof failureData.message !== 'string'
+    || failureData.message.length === 0
+    || typeof failureData.code !== 'string'
+    || failureData.code.length === 0) return null
+  if (data.mode === 'normal') {
+    if (!positiveSafeInteger(data.maxRetries) || data.retry > data.maxRetries) return null
+  } else if (data.mode === 'always') {
+    if ('maxRetries' in data) return null
+  } else {
+    return null
+  }
+  if (failureData.status !== undefined
+    && (typeof failureData.status !== 'number'
+      || !Number.isInteger(failureData.status)
+      || failureData.status < 100
+      || failureData.status > 599)) return null
+  if (failureData.providerRetryAfterMs !== undefined
+    && (typeof failureData.providerRetryAfterMs !== 'number'
+      || !Number.isFinite(failureData.providerRetryAfterMs)
+      || failureData.providerRetryAfterMs <= 0)) return null
+  if (failureData.requestId !== undefined
+    && (typeof failureData.requestId !== 'string'
+      || failureData.requestId.length === 0)) return null
+  return data as unknown as LlmRetryEventData
+}
+
+function nonNegativeSafeInteger(value: unknown): value is number {
+  return typeof value === 'number' && Number.isSafeInteger(value) && value >= 0
+}
+
+function positiveSafeInteger(value: unknown): value is number {
+  return nonNegativeSafeInteger(value) && value > 0
+}
+
 /**
 * The composerPhase judgment — the single site that knows the predicate
 * (consumers switch on the result, never re-derive). Monotone per session
--- a/packages/client/runtime/tests/event-script.ts
+++ b/packages/client/runtime/tests/event-script.ts
@@ -63,7 +63,25 @@ export const ev = {
    }),
  stepEnd: (seq: number, turn: number, step = 0): SessionEvent =>
    at(seq, { type: 'step/end', data: { turn, step } }),
-  turnEnd: (seq: number, turn: number, reason: 'completed' | 'cancelled' = 'completed'): SessionEvent =>
+  retry: (
+    seq: number,
+    turn: number,
+    step = 0,
+    retry = 1,
+    maxRetries = 2,
+    delayMs = 500,
+    message = 'temporary transport failure',
+  ): SessionEvent =>
+    at(seq, {
+      type: 'llm/retry',
+      data: {
+        turn, step,
+        provider: 'fake', mode: 'normal', policyKey: 'fake-normal',
+        retry, maxRetries, delayMs,
+        failure: { code: 'TRANSPORT', message },
+      },
+    }),
+  turnEnd: (seq: number, turn: number, reason: 'completed' | 'aborted' | 'disposed' = 'completed'): SessionEvent =>
    at(seq, { type: 'turn/end', data: { turn, reason: { kind: reason } } }),
  commandRun: (seq: number, commandId: string, name: string, args = ''): SessionEvent =>
    at(seq, { type: 'command/run', data: { commandId, name, args, source: { kind: 'user' } } }),
--- a/packages/client/runtime/tests/session.spec.ts
+++ b/packages/client/runtime/tests/session.spec.ts
@@ -8,6 +8,7 @@

 import { describe, expect, it, vi } from 'vitest'
 import type { SessionEvent } from '@deepseek-ai/dsh-session/types'
+import { MAX_TIMER_DELAY_MS } from '@deepseek-ai/dsh-timeout'
 import type { SessionId } from '@deepseek-ai/dsh-client-connection/client'
 import { Session } from '../src/client/sessions/session.ts'
 import { FakeApiClient, deferred, err, ok } from './fake-api.ts'
@@ -161,6 +162,214 @@ describe('live event path', () => {
    expect((last as { interrupted?: true }).interrupted).toBeUndefined()
  })

+  it('retracts the failed step partial on retry and keeps a replayable notice before the recovered response', async () => {
+    const { session } = await opened()
+    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
+    const retryTurn = [
+      ev.turnStart(6, 1),
+      ev.user(7, '请重试'),
+      ev.stepStart(8, 1),
+      ev.chunkStart(9, 1),
+      ev.chunkText(10, 1, '不完整回复'),
+      ev.stepEnd(11, 1),
+      ev.retry(12, 1, 0, 1, 2, 450, '连接被重置'),
+      at(13, {
+        type: 'turn/end',
+        data: {
+          turn: 1,
+          reason: {
+            kind: 'error', step: 0,
+            failure: { code: 'TRANSPORT', message: '连接被重置' },
+          },
+        },
+      }),
+      at(14, { type: 'turn/start', data: { turn: 2, trigger: { kind: 'retry' } } }),
+      ev.stepStart(15, 2),
+      ev.assistant(16, 2, '完整回复'),
+      ev.stepEnd(17, 2),
+      ev.turnEnd(18, 2),
+    ]
+    for (const event of retryTurn.slice(0, 7)) feed(event)
+
+    let snapshot = session.getSnapshot()
+    expect(snapshot.partial).toBeNull()
+    expect(snapshot.nodes.at(-1)).toMatchObject({
+      kind: 'model-retry',
+      retryState: 'scheduled',
+      turn: 1,
+      step: 0,
+      provider: 'fake',
+      mode: 'normal',
+      policyKey: 'fake-normal',
+      retry: 1,
+      maxRetries: 2,
+      delayMs: 450,
+      failure: { code: 'TRANSPORT', message: '连接被重置' },
+    })
+    expect(JSON.stringify(snapshot.nodes)).not.toContain('不完整回复')
+
+    for (const event of retryTurn.slice(7)) feed(event)
+    snapshot = session.getSnapshot()
+    expect(snapshot.nodes.slice(-2).map(node => node.kind)).toEqual(['model-retry', 'assistant'])
+    expect(snapshot.nodes.at(-2)).toMatchObject({ kind: 'model-retry', retryState: 'started' })
+    expect(snapshot.nodes.at(-1)).toMatchObject({ kind: 'assistant', blocks: [{ kind: 'text', text: '完整回复' }] })
+
+    const replay = makeSession()
+    replay.api.onHistory = () => histResponse([...plainTurn(0, 0, 'a', 'b'), ...retryTurn])
+    await replay.session.open()
+    expect(replay.session.getSnapshot().nodes).toEqual(snapshot.nodes)
+    expect(replay.session.getSnapshot().partial).toBeNull()
+  })
+
+  it('rejects retry payloads outside the producer contract without retracting the current partial', async () => {
+    const { session } = await opened()
+    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
+    feed(ev.turnStart(6, 1))
+    feed(ev.chunkStart(7, 1))
+    feed(ev.chunkText(8, 1, '仍在生成'))
+    const valid = {
+      turn: 1, step: 0,
+      provider: 'fake', mode: 'normal', policyKey: 'fake-normal',
+      retry: 1, maxRetries: 2, delayMs: 500,
+      failure: { code: 'TRANSPORT', message: 'temporary failure' },
+    }
+    const invalid = [
+      { ...valid, turn: Number.MAX_SAFE_INTEGER + 1 },
+      { ...valid, step: Number.MAX_SAFE_INTEGER + 1 },
+      { ...valid, provider: '' },
+      { ...valid, policyKey: '' },
+      { ...valid, retry: Number.MAX_SAFE_INTEGER + 1 },
+      { ...valid, maxRetries: Number.MAX_SAFE_INTEGER + 1 },
+      { ...valid, delayMs: -1 },
+      { ...valid, delayMs: Number.POSITIVE_INFINITY },
+      { ...valid, delayMs: MAX_TIMER_DELAY_MS + 1 },
+      { ...valid, failure: { ...valid.failure, message: '' } },
+      { ...valid, failure: { ...valid.failure, code: '' } },
+      { ...valid, failure: { ...valid.failure, status: '429' } },
+      { ...valid, failure: { ...valid.failure, status: 99 } },
+      { ...valid, failure: { ...valid.failure, status: 429.5 } },
+      { ...valid, failure: { ...valid.failure, status: 600 } },
+      { ...valid, failure: { ...valid.failure, providerRetryAfterMs: 0 } },
+      { ...valid, failure: { ...valid.failure, providerRetryAfterMs: Number.POSITIVE_INFINITY } },
+      { ...valid, failure: { ...valid.failure, requestId: 1 } },
+      { ...valid, failure: { ...valid.failure, requestId: '' } },
+    ]
+    const errorSpy = vi.spyOn(console, 'error').mockImplementation(() => undefined)
+    try {
+      for (const [index, data] of invalid.entries()) {
+        feed(at(9 + index, { type: 'llm/retry', data }))
+      }
+      expect(session.getSnapshot().partial?.blocks).toEqual([{ kind: 'text', text: '仍在生成' }])
+      expect(session.getSnapshot().nodes.filter(node => node.kind === 'model-retry')).toEqual([])
+      expect(errorSpy).toHaveBeenCalledTimes(invalid.length)
+      expect(errorSpy).toHaveBeenCalledWith('[web-runtime] ignored malformed llm/retry event at seq 9')
+    } finally {
+      errorSpy.mockRestore()
+    }
+  })
+
+  it('accepts complete retry payloads at the producer field boundaries', async () => {
+    const { session } = await opened()
+    session.handleMuxEnvelope('r' as never, {
+      type: 'session/event',
+      sessionId: SID,
+      event: at(6, {
+        type: 'llm/retry',
+        data: {
+          turn: Number.MAX_SAFE_INTEGER,
+          step: Number.MAX_SAFE_INTEGER,
+          provider: 'fake',
+          mode: 'normal',
+          policyKey: 'fake-normal',
+          retry: Number.MAX_SAFE_INTEGER,
+          maxRetries: Number.MAX_SAFE_INTEGER,
+          delayMs: MAX_TIMER_DELAY_MS,
+          failure: {
+            code: 'RATE_LIMIT',
+            message: 'provider busy',
+            status: 599,
+            providerRetryAfterMs: Number.MIN_VALUE,
+            requestId: 'req-1',
+          },
+        },
+      }),
+    })
+    expect(session.getSnapshot().nodes.at(-1)).toMatchObject({
+      kind: 'model-retry',
+      retryState: 'scheduled',
+      retry: Number.MAX_SAFE_INTEGER,
+      delayMs: MAX_TIMER_DELAY_MS,
+      failure: { status: 599, providerRetryAfterMs: Number.MIN_VALUE, requestId: 'req-1' },
+    })
+  })
+
+  it('projects always-mode retries and rejects mode-specific maximums or unknown modes', async () => {
+    const { session } = await opened()
+    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
+    const errorSpy = vi.spyOn(console, 'error').mockImplementation(() => undefined)
+    try {
+      feed(at(6, {
+        type: 'llm/retry',
+        data: {
+          turn: 1, step: 0,
+          provider: 'fake', mode: 'always', policyKey: 'fake-always',
+          retry: 3, delayMs: 500,
+          failure: { code: 'TRANSPORT', message: 'retry forever' },
+        },
+      }))
+      expect(session.getSnapshot().nodes.at(-1)).toMatchObject({
+        kind: 'model-retry',
+        retryState: 'scheduled',
+        mode: 'always',
+        retry: 3,
+      })
+
+      feed(at(7, {
+        type: 'llm/retry',
+        data: {
+          turn: 2, step: 0,
+          provider: 'fake', mode: 'always', policyKey: 'fake-always',
+          retry: 4, maxRetries: 4, delayMs: 500,
+          failure: { code: 'TRANSPORT', message: 'unexpected maximum' },
+        },
+      }))
+      feed(at(8, {
+        type: 'llm/retry',
+        data: {
+          turn: 2, step: 0,
+          provider: 'fake', mode: 'sometimes', policyKey: 'fake-unknown',
+          retry: 4, delayMs: 500,
+          failure: { code: 'TRANSPORT', message: 'unknown mode' },
+        },
+      }))
+      expect(session.getSnapshot().nodes.filter(node => node.kind === 'model-retry')).toHaveLength(1)
+      expect(errorSpy).toHaveBeenCalledTimes(2)
+    } finally {
+      errorSpy.mockRestore()
+    }
+  })
+
+  it.each(['aborted', 'disposed'] as const)(
+    'marks a scheduled retry as cancelled when its failed turn ends %s',
+    async (reason) => {
+      const { session } = await opened()
+      const feed = (event: SessionEvent) => {
+        session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event })
+      }
+      feed(ev.turnStart(6, 1))
+      feed(ev.retry(7, 1))
+      expect(session.getSnapshot().nodes.at(-1)).toMatchObject({
+        kind: 'model-retry',
+        retryState: 'scheduled',
+      })
+      feed(ev.turnEnd(8, 1, reason))
+      expect(session.getSnapshot().nodes.at(-1)).toMatchObject({
+        kind: 'model-retry',
+        retryState: 'cancelled',
+      })
+    },
+  )
+
  it('freezes an unfinalized partial into an interrupted node on turn/end (cancel path)', async () => {
    const { session } = await opened()
    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
@@ -168,7 +377,7 @@ describe('live event path', () => {
    feed(ev.user(7, '要被打断的'))
    feed(ev.chunkStart(8, 1))
    feed(ev.chunkText(9, 1, '说到一半'))
-    feed(ev.turnEnd(10, 1, 'cancelled')) // no assistant/message ever arrives
+    feed(ev.turnEnd(10, 1, 'aborted')) // no assistant/message ever arrives
    const snapshot = session.getSnapshot()
    expect(snapshot.partial).toBeNull()
    const frozen = snapshot.nodes.at(-1)
@@ -187,7 +396,7 @@ describe('live event path', () => {
    expect(session.getSnapshot().runningCalls).toEqual([])
    // Second call never resolves: turn/end freezes it as an error card.
    feed(ev.toolCall(9, 1, 'c2', 'slow_tool', '{}'))
-    feed(ev.turnEnd(10, 1, 'cancelled'))
+    feed(ev.turnEnd(10, 1, 'aborted'))
    const snapshot = session.getSnapshot()
    expect(snapshot.runningCalls).toEqual([])
    expect(snapshot.nodes.at(-1)).toMatchObject({
@@ -529,7 +738,7 @@ describe('remaining branches', () => {
    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
    feed(ev.turnStart(6, 1))
    feed(ev.chunkStart(7, 1)) // empty text block only, no delta
-    feed(ev.turnEnd(8, 1, 'cancelled'))
+    feed(ev.turnEnd(8, 1, 'aborted'))
    const snapshot = session.getSnapshot()
    expect(snapshot.partial).toBeNull()
    expect(snapshot.nodes.filter(n => n.kind === 'assistant' && (n as { interrupted?: true }).interrupted)).toEqual([])
@@ -543,7 +752,7 @@ describe('remaining branches', () => {
    feed(ev.turnStart(6, 1))
    feed(ev.toolCall(7, 1, 'turn1-call', 'echo', '{}'))
    feed(ev.toolCall(8, 2, 'turn2-call', 'echo', '{}')) // stray call attributed to a later turn
-    feed(ev.turnEnd(9, 1, 'cancelled'))
+    feed(ev.turnEnd(9, 1, 'aborted'))
    const snapshot = session.getSnapshot()
    expect(snapshot.runningCalls.map(c => c.callId)).toEqual(['turn2-call'])
    expect(snapshot.nodes.at(-1)).toMatchObject({ kind: 'tool-result', callId: 'turn1-call', isError: true })
@@ -637,7 +846,7 @@ describe('remaining branches', () => {
    const feed = (event: SessionEvent) => { session.handleMuxEnvelope('r' as never, { type: 'session/event', sessionId: SID, event }) }
    feed(ev.turnStart(6, 1))
    feed(at(7, { type: 'assistant/chunk', data: { turn: 1, step: 0, chunk: { type: 'tool-call-delta', index: 0, id: 'c1', name: 'echo', argumentsDelta: '{' } } }))
-    feed(ev.turnEnd(8, 1, 'cancelled'))
+    feed(ev.turnEnd(8, 1, 'aborted'))
    const frozen = session.getSnapshot().nodes.at(-1)
    expect(frozen).toMatchObject({ kind: 'assistant', interrupted: true, blocks: [{ kind: 'tool-call', callId: 'c1' }] })
  })
--- a/packages/client/runtime/tsconfig.json
+++ b/packages/client/runtime/tsconfig.json
@@ -35,6 +35,9 @@
    {
      "path": "../../llm/llm"
    },
+    {
+      "path": "../../llm/llm-retry"
+    },
    {
      "path": "../../support/invariants"
    }
--- a/packages/client/ui-conversation/README.i18n.yaml
+++ b/packages/client/ui-conversation/README.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/ui-conversation/README.md
-README.md: 275b08ae891579a6865b8ce2185028ccb88f0a4c
-README.zh.md: c8b2540ecfc6966b5b03e80ae45fa320ad1fcc23
+README.md: 3e0bca6610e5503e4c2c1f9fe5ad7a07bbcbdd2a
+README.zh.md: 1606461fc1060825133bb0c3f3f6381bda3926ad
--- a/packages/client/ui-conversation/README.md
+++ b/packages/client/ui-conversation/README.md
@@ -18,6 +18,8 @@ A tool call declaring the `terminal` render intent renders its command output in

 A tool call declaring the `web` render intent renders its web retrieval inline, at both conversation render sites, through ui-primitives' `WebBlock`. `contract/web-card-model.ts` is the single derivation from the snapshot's `resultView`, mirroring the terminal card, so the sites cannot disagree about what a web call shows; it yields null — the generic path — for a running call, a non-web result view, a generic result view, a `card` tag this client version does not know, or a web card whose `kind` this client version does not know (a newer host's value, which the wire cannot be trusted to be `search` or `fetch`). The keyed `WebRow` registers one component under both `web_search` and `web_fetch`, discriminating on the tool name only for its icon and title; a web-declaring tool without a keyed row lands on the `GenericToolCard` fallback, which grows the same resident card, and the details panel renders it at the primitive's full source allowance and, below the card, the flattened model-visible result content — a fetch body is readable only there, since its card carries only the URL and status. Rows cap at `CHAT_WEB_MAX_SOURCES` (8) against the panel's 16, the same summary-versus-reading split the terminal card draws ([decision](../../../.agents/notes/implemented/feature/2026-07-30-web-result-card-frontend.md)).

+The chat flow projects consecutive model-retry nodes across retry turns into one stable, muted status row updated to the latest attempt; every retry event remains in the runtime snapshot and session log. Its frontend countdown anchors the scheduled delay to client receipt, avoiding host/browser clock skew, rounds remaining time up to seconds, and has a one-second floor. The latest unresolved retry uses a left-to-right text shimmer. Subsequent turn facts distinguish an attempt that started from one cancelled during backoff, while the Host running bit only controls the live animation; the row then shows a static completed or cancelled label. Normal policy rows show the finite retry maximum; always policy rows show `∞`. Activating the row reveals the latest exact retry delay and failure message. The client runtime removes each failed step's streaming tail before its retry node arrives, while the status remains visible after a later attempt succeeds.
+
 Tool rows are slots too — the standalone tool ring (`ToolViewRegistry`/`ctx.toolviews`/outlet) is retired. The chat entry declares the keyed `'conversation.chat.toolview'` hole (session scope; the key space is runtime-open); its render site dispatches per row via `entryKey: toolName` with `GenericToolCard` as the call-site `fallback`. The owner payload is the uniform `ToolRowOwnerProps` (`callId`/`toolName`/`block`/`openFile`) and `ToolRowProps` pre-composes it with the session standard kit. A registrant is a plain plugin: `ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)` with `inject: ['slots', 'conversation']` as the load-order seam (apply mounts ConversationService after the chat registration, so the service being present guarantees the slot is declared); session differentiation happens inside the component (`useSessions` reading `parentId` — the bash sample is the third-party-posture exemplar). Trajectory/waterfall toolview slots share this shape and land with their own render sites (RendersCheck rejects a declaration nobody renders).

 The todo surfaces are two registrations over that shape, both plain registrant plugins with `inject: ['slots', 'conversation']`. `TodoRow` takes the `'conversation.chat.toolview'` key `todo_write` and summarizes what the call attempted (`<done>/<total> 已完成 · <active item>` parsed from its args, falling back to the generic summary on malformed or wrongly-shaped model JSON, and keeping the generic dot for non-ok execution states so a cancelled call never reads as a completed update). `TodoDock` takes the `'conversation.input.dock'` list slot at `order: -1` — above the queue rows — and is the plan strip: it reads the host-computed `todos` projection via `useProjection` (standing plan: latest `todo/write` with no later `turn/start`) and renders `TodoPanel`, which takes the plain list, hides itself while the list is empty, and starts collapsed as a header of title plus `"<done>/<total> tasks · <n> in progress"` (status glyphs are the figma check / progress / dashed-pending set). The dock adapter owns the selection so the panel stays a pure function of its props; the standing list lives here rather than in the row so the row stays one line. Anything the input-zone composer chain hides (a `conversation.composer` takeover such as ui-question's) hides the whole dock, this strip included.
--- a/packages/client/ui-conversation/README.zh.md
+++ b/packages/client/ui-conversation/README.zh.md
@@ -16,6 +16,8 @@

 声明 `web` 渲染意图的工具调用，会在两个对话渲染点上都通过 ui-primitives 的 `WebBlock` 内联渲染其 web 检索。`contract/web-card-model.ts` 是从快照的 `resultView` 推导的唯一位置，镜像终端卡片，因此两个渲染点不可能对一次 web 调用的显示产生分歧；对运行中的调用、非 web 的 result view、generic result view、本客户端版本不认识的 `card` 标签，或本客户端版本不认识 `kind` 的 web 卡片（更新的 host 发来的值，wire 上不可信其为 `search` 或 `fetch`），它返回 null，落回通用路径。键控的 `WebRow` 把一个组件注册在 `web_search` 与 `web_fetch` 两个键下，仅根据工具名判别以选取图标与标题；没有自己键控行的 web 声明工具落到 `GenericToolCard` 兜底，它长出同一张常驻卡片，详情面板则以原语的完整 source 额度渲染它，并在卡片下方渲染摊平的模型可见结果内容——fetch 正文只在此处可读，因为其卡片只携带 URL 和状态。行的上限是 `CHAT_WEB_MAX_SOURCES`（8），面板为 16，与终端卡片所画的摘要面对阅读面的同一划分（[决策](../../../.agents/notes/implemented/feature/2026-07-30-web-result-card-frontend.md)）。

+聊天流会将跨重试轮次连续出现的模型重试节点投影为一个稳定的弱化状态行，并用最新一次尝试更新该行；每个重试事件仍保留在运行时快照与会话日志中。前端倒计时以客户端收到事件的时刻为计划延迟的起点，避免 Host 与浏览器的时钟偏差；剩余时间向上取整到秒，且下限为 1 秒。最近一次尚未完成的重试会显示从左到右的文字渐变动画。后续轮次事实用于区分已开始的尝试与在退避期间取消的尝试，Host 的 running 位只控制实时动画；随后该行会显示静态的已完成或已取消标签。normal 策略行显示有限重试上限；always 策略行显示 `∞`。激活该行会显示最近一次重试的精确延迟和失败消息。客户端运行时会在相应重试节点到达前移除每个失败步骤的流式输出尾部；后续某次尝试成功后，该状态仍保持可见。
+
 工具行同样是 slot：独立工具环（`ToolViewRegistry`／`ctx.toolviews`／outlet）已经退役。聊天配置项声明键控的 `'conversation.chat.toolview'` 空位（Session scope；key 空间在运行时开放）；其渲染点逐行通过 `entryKey: toolName` 分发，并以 `GenericToolCard` 作为调用点 `fallback`。owner 载荷是统一的 `ToolRowOwnerProps`（`callId`／`toolName`／`block`／`openFile`），`ToolRowProps` 则预先将其与 Session 标准工具包组合。注册方只是普通插件：`ctx.slots.register({ name: 'conversation.chat.toolview', key: '<tool>', inject? }, Row)`，以 `inject: ['slots', 'conversation']` 作为加载顺序 seam（apply 在聊天注册后挂载 ConversationService，因此服务存在即可保证 slot 已声明）；Session 区分在组件内部完成（`useSessions` 读取 `parentId`，bash 示例是第三方姿态的范例）。Trajectory/waterfall 工具视图 slot 共享此形状，并随各自的渲染点落地（RendersCheck 会拒绝没有任何渲染方的声明）。

 审批经由本包声明的链接管编辑器：`ApprovalPanel` 注册为按选择器路由的 `'conversation.composer'` 配置项（ui-question 模式），在审批等待未决期间取代 InputBar 占据编辑器（琥珀色条、理由标题、来自运行中调用参数的配对命令行、一次性的拒绝／允许）。`contract/slots.ts` 中的 `PendingApproval` 领域面在运行时 `PendingWait` 载体之上拥有 wire 编码——带审计关联的 `ApprovalResponsePayload` 值；广播的 `approval/resolved` 帧使等待落定并恢复编辑器。侧边栏通过 manager 跟踪的 `waitingApproval` 列表位（未实例化会话同样点亮）镜像该阻塞状态，其优先级高于运行中圆环，直至问题解决。未决等待完全离开消息流：问题（ui-question）与审批（ApprovalPanel）都经编辑器接管作答，不再保留只读占位卡。编辑器底行的 Access 席位挂载 `PermissionSelect`，由 host 计算的 `permissions` 投影经标准工具包 `useProjection` 供数（key 缺席即隐藏 chip）；chip 打开 Menu 原语下拉，普通安全预设会立即经输入栏注入的 `command` 回调提交 `/permission <preset>`，而 `danger-full-access` 在界面中显示为 `Full access`，选择后先打开页面内的 Modal 风险确认。用户勾选确认项前启用按钮始终不可用；取消、Escape、关闭按钮与点击遮罩都不会提交命令。
--- a/packages/client/ui-conversation/src/client/chat/ChatView.tsx
+++ b/packages/client/ui-conversation/src/client/chat/ChatView.tsx
@@ -55,6 +55,17 @@ type RenderToolRow = ChatViewSlotProps['renderSlot']
 *  chat view narrows once to the runtime snapshot the binding actually feeds. */
 type UseConversation = SnapshotSelectorHook<ConversationSnapshot>

+function activeRetrySeq(nodes: readonly ConversationNode[], running: boolean): number | null {
+  if (!running) return null
+  for (let index = nodes.length - 1; index >= 0; index -= 1) {
+    const node = nodes[index]
+    if (node === undefined) continue
+    if (node.kind === 'model-retry') return node.retryState === 'cancelled' ? null : node.seq
+    if (node.kind === 'assistant' || node.kind === 'user') return null
+  }
+  return null
+}
+
 /** One `run_code` sub-dispatch row: the identical keyed-slot dispatch as a
 *  top-level call (same registrations, same fallback), nested by the parent.
 *  A started-but-unsettled sub-call arrives as the RunningToolCall shape and
@@ -262,6 +273,7 @@ export function ChatView({
  const selectedCallId = useStore(s => s.selection?.callId)

  const items = useMemo(() => deriveChatFlow(nodes), [nodes])
+  const activeRetry = useMemo(() => activeRetrySeq(nodes, running), [nodes, running])
  // Only the last content assistant of each turn owns IconActions; mid-turn
  // text (before tools) omits `time` so AssistantMarkdown stays chrome-free.
  const actionSeqs = useMemo(() => assistantActionsSeqs(nodes), [nodes])
@@ -424,7 +436,15 @@ export function ChatView({
    }
    /* v8 ignore next -- tool-result never reaches here: deriveChatFlow folds them into groups. */
    if (node.kind === 'tool-result') return null
-    return <MessageItem key={item.key} node={node} onFork={forkAt} t={t} />
+    return (
+      <MessageItem
+        key={item.key}
+        node={node}
+        retryActive={node.kind === 'model-retry' && node.seq === activeRetry}
+        onFork={forkAt}
+        t={t}
+      />
+    )
  }

  return (
--- a/packages/client/ui-conversation/src/client/chat/MessageItem.module.css
+++ b/packages/client/ui-conversation/src/client/chat/MessageItem.module.css
@@ -34,6 +34,106 @@
  padding: 2px 0;
 }

+.retryRow {
+  color: var(--dsw-alias-label-tertiary);
+  font-size: 13px;
+  line-height: 20px;
+}
+
+.retrySummary {
+  display: inline-flex;
+  align-items: center;
+  width: fit-content;
+  padding: 2px 0;
+  gap: 7px;
+  border-radius: 3px;
+  color: inherit;
+  cursor: pointer;
+  list-style: none;
+  user-select: none;
+}
+
+.retrySummary::-webkit-details-marker {
+  display: none;
+}
+
+.retrySummary::after {
+  width: 6px;
+  height: 6px;
+  border-right: 1.5px solid currentcolor;
+  border-bottom: 1.5px solid currentcolor;
+  content: '';
+  opacity: 0.8;
+  transform: rotate(-45deg);
+  transition: transform 120ms ease;
+}
+
+.retrySummary:hover {
+  color: var(--dsw-alias-label-secondary);
+}
+
+.retrySummary:focus-visible {
+  outline: 1.5px solid var(--dsw-alias-button-info-fill);
+  outline-offset: 2px;
+}
+
+.retryText {
+  color: inherit;
+}
+
+.retryRow[data-active] .retryText {
+  background:
+    linear-gradient(
+      90deg,
+      var(--dsw-alias-label-tertiary) 0%,
+      var(--dsw-alias-label-tertiary) 40%,
+      var(--dsw-alias-label-secondary) 50%,
+      var(--dsw-alias-label-tertiary) 60%,
+      var(--dsw-alias-label-tertiary) 100%
+    );
+  background-position: 100% 50%;
+  background-size: 200% 100%;
+  background-clip: text;
+  color: transparent;
+  animation: retry-shimmer 1.6s ease-in-out infinite;
+}
+
+.retryRow[open] .retrySummary::after {
+  transform: rotate(45deg);
+}
+
+.retryDetails {
+  display: grid;
+  gap: 2px;
+  margin-top: 3px;
+  padding-left: 14px;
+  overflow-wrap: anywhere;
+  font-size: 12px;
+  line-height: 18px;
+}
+
+.retryDetailLabel {
+  color: var(--dsw-alias-label-secondary);
+}
+
+@keyframes retry-shimmer {
+  from {
+    background-position: 100% 50%;
+  }
+
+  to {
+    background-position: 0 50%;
+  }
+}
+
+@media (prefers-reduced-motion: reduce) {
+  .retryRow[data-active] .retryText {
+    background: none;
+    color: inherit;
+    animation: none;
+  }
+}
+
 /* Reference chip projection inside a user bubble (`<skill>name</skill>` model
   spans render as chips; free geometry — no textarea pairing here). */
 .refChip {
--- a/packages/client/ui-conversation/src/client/chat/MessageItem.tsx
+++ b/packages/client/ui-conversation/src/client/chat/MessageItem.tsx
@@ -1,13 +1,11 @@
-// MessageItem: the four simple node kinds — user bubble (right-aligned, with
+// MessageItem: simple chat nodes — user bubble (right-aligned, with
 // clock + copy / branch / edit IconActions), steering (badged bubble), context
-// injection and unknown-surface JSON rows. Props are frozen node slices off
-// the snapshot cache; memo holds across streaming because unchanged nodes
-// keep their references.
+// injection, retry disclosure, and unknown-surface JSON rows.

-import { memo } from 'react'
+import { memo, useEffect, useMemo, useState } from 'react'
 import type { ReactNode } from 'react'
 import type {
-  ContextMessageNode, SteeringMessageNode, UnknownSurfaceNode, UserMessageNode,
+  ContextMessageNode, ModelRetryNode, SteeringMessageNode, UnknownSurfaceNode, UserMessageNode,
 } from '@deepseek-ai/dsh-client-runtime/client'
 import { JsonBlock, MessageText } from '@deepseek-ai/dsh-client-ui-primitives'
 import type { ChatViewSlotProps } from '../contract/slots.ts'
@@ -16,7 +14,8 @@ import { MessageIconActions } from './MessageIconActions.tsx'
 import css from './MessageItem.module.css'

 export interface MessageItemProps {
-  node: UserMessageNode | SteeringMessageNode | ContextMessageNode | UnknownSurfaceNode
+  node: UserMessageNode | SteeringMessageNode | ContextMessageNode | ModelRetryNode | UnknownSurfaceNode
+  retryActive?: boolean
  /** Fork the session through the turn containing this message (user-bubble branch action). */
  onFork?: (seq: number) => void
  /** The owning view's locale seat, passed down as a plain prop. */
@@ -34,6 +33,80 @@ function contentText(content: readonly unknown[]): { text: string; rest: unknown
  return { text: texts.join(''), rest }
 }

+function retrySeconds(milliseconds: number): number {
+  return Math.max(1, Math.ceil(milliseconds / 1_000))
+}
+
+interface RetryCountdown {
+  deadline: number
+  seconds: number
+}
+
+function ModelRetryItem({ node, active, t }: {
+  node: ModelRetryNode
+  active: boolean
+  t: ChatViewSlotProps['t']
+}) {
+  // Anchor the host-scheduled delay to this browser's first render of the
+  // retry node. Host event time and Date.now() may belong to different clocks.
+  const deadline = useMemo(() => Date.now() + node.delayMs, [node.delayMs, node.seq])
+  const scheduledSeconds = retrySeconds(node.delayMs)
+  const maximum = node.mode === 'normal' ? node.maxRetries : '∞'
+  const [countdown, setCountdown] = useState<RetryCountdown>(() => ({
+    deadline,
+    seconds: retrySeconds(deadline - Date.now()),
+  }))
+  const remainingSeconds = countdown.deadline === deadline
+    ? countdown.seconds
+    : retrySeconds(deadline - Date.now())
+
+  useEffect(() => {
+    if (!active) return
+    const updateCountdown = (): number => {
+      const next = retrySeconds(deadline - Date.now())
+      setCountdown(current => (
+        current.deadline === deadline && current.seconds === next
+          ? current
+          : { deadline, seconds: next }
+      ))
+      return next
+    }
+    if (updateCountdown() === 1) return
+    const timer = window.setInterval(() => {
+      if (updateCountdown() === 1) window.clearInterval(timer)
+    }, 250)
+    return () => { window.clearInterval(timer) }
+  }, [active, deadline])
+
+  const label = active
+    ? t('message.retry.active')
+    : node.retryState === 'cancelled'
+      ? t('message.retry.cancelled')
+      : node.retryState === 'started'
+        ? t('message.retry.started')
+        : t('message.retry.scheduled')
+  const seconds = active ? remainingSeconds : scheduledSeconds
+
+  return (
+    <details className={css.retryRow} data-active={active || undefined}>
+      <summary className={css.retrySummary}>
+        <span className={css.retryText} role="status">
+          {t('message.retry.status', { label, retry: node.retry, maximum, seconds })}
+        </span>
+      </summary>
+      <div className={css.retryDetails}>
+        <div>
+          <span className={css.retryDetailLabel}>{t('message.retry.delay')}</span>
+          {Math.round(node.delayMs)}ms
+        </div>
+        <div>
+          <span className={css.retryDetailLabel}>{t('message.retry.failure')}</span>
+          {node.failure.message}
+        </div>
+      </div>
+    </details>
+  )
+}
 /**
 * Display projection of reference forms in a user bubble (free geometry — no
 * textarea alignment constraint here); everything else stays plain text. The
@@ -66,7 +139,9 @@ function projectUserText(text: string): ReactNode {
  return <>{parts}</>
 }

-export const MessageItem = memo(function MessageItem({ node, onFork, t }: MessageItemProps) {
+export const MessageItem = memo(function MessageItem({
+  node, retryActive = false, onFork, t,
+}: MessageItemProps) {
  const truncated = (total: number): string => t('json.truncated', { total })
  switch (node.kind) {
    case 'user': {
@@ -105,6 +180,8 @@ export const MessageItem = memo(function MessageItem({ node, onFork, t }: Messag
      return (
        <ContextInjectionRow content={node.content} source={node.source} t={t} />
      )
+    case 'model-retry':
+      return <ModelRetryItem node={node} active={retryActive} t={t} />
    default:
      return (
        <div className={css.contextRow}>
--- a/packages/client/ui-conversation/src/client/chat/chat-flow.ts
+++ b/packages/client/ui-conversation/src/client/chat/chat-flow.ts
@@ -1,7 +1,8 @@
 /**
 * Chat flow derivation: ConversationSnapshot nodes -> render items. Tool
 * results group into consecutive-run tool groups (figma step-summary flow,
- * VERTICAL gap10) alternating with narration; everything else passes through.
+ * VERTICAL gap10) alternating with narration. Consecutive retry notices
+ * reuse the first notice's row while projecting the latest retry turn.
 * Item identity keys are stable across snapshots so the list parent can
 * subscribe to keys only while rows subscribe to content. IconActions ownership
 * (last content assistant per turn) is derived here too so ChatView and the
@@ -49,7 +50,7 @@ export function assistantActionsSeqs(nodes: readonly ConversationNode[]): Readon
 /**
 * Group finalized nodes into the step-summary flow.
 * @param nodes - snapshot nodes (surface order).
- * @returns flow items; consecutive tool-results merged into one group keyed by the first seq.
+ * @returns flow items; consecutive tool results and retry notices reuse their first key.
 */
 export function deriveChatFlow(nodes: readonly ConversationNode[]): ChatFlowItem[] {
  const items: ChatFlowItem[] = []
@@ -63,6 +64,17 @@ export function deriveChatFlow(nodes: readonly ConversationNode[]): ChatFlowItem
      } else {
        group.push(node)
      }
+    } else if (node.kind === 'model-retry') {
+      group = null
+      const previous = items[items.length - 1]
+      if (
+        previous?.kind === 'node'
+        && previous.node.kind === 'model-retry'
+      ) {
+        items[items.length - 1] = { ...previous, node }
+      } else {
+        items.push({ kind: 'node', key: `n${node.seq}`, node })
+      }
    } else {
      group = null
      items.push({ kind: 'node', key: `n${node.seq}`, node })
--- a/packages/client/ui-conversation/src/client/locales.ts
+++ b/packages/client/ui-conversation/src/client/locales.ts
@@ -53,6 +53,13 @@ export const zh = {
  'message.unknownBlock': '未知内容块',
  'message.stopped': '已停止',
  'message.branch': '在新对话中分支',
+  'message.retry.active': '正在重试模型请求',
+  'message.retry.cancelled': '模型请求重试已取消',
+  'message.retry.started': '已重试模型请求',
+  'message.retry.scheduled': '等待重试模型请求',
+  'message.retry.status': '{label}（{retry}/{maximum}） · {seconds}s',
+  'message.retry.delay': '重试延迟：',
+  'message.retry.failure': '失败原因：',
  'command.running': '执行中…',
  'command.failed': '命令失败',
  'command.done': '已完成',
@@ -140,6 +147,13 @@ export const en = {
  'message.unknownBlock': 'Unknown content block',
  'message.stopped': 'Stopped',
  'message.branch': 'Branch into a new conversation',
+  'message.retry.active': 'Retrying model request',
+  'message.retry.cancelled': 'Model request retry cancelled',
+  'message.retry.started': 'Retried model request',
+  'message.retry.scheduled': 'Waiting to retry model request',
+  'message.retry.status': '{label} ({retry}/{maximum}) · {seconds}s',
+  'message.retry.delay': 'Retry delay: ',
+  'message.retry.failure': 'Failure reason: ',
  'command.running': 'Running…',
  'command.failed': 'Command failed',
  'command.done': 'Completed',
--- a/packages/client/ui-conversation/tests/chat-branch-tails.spec.tsx
+++ b/packages/client/ui-conversation/tests/chat-branch-tails.spec.tsx
@@ -18,7 +18,10 @@ import { AssistantMarkdown } from '../src/client/chat/AssistantMarkdown.tsx'
 import { StatsLine, type StatsLineProps } from '../src/client/chat/StatsLine.tsx'
 import { zh } from '../src/client/locales.ts'

-afterEach(cleanup)
+afterEach(() => {
+  cleanup()
+  vi.useRealTimers()
+})

 // Mirrors the real lookup chain (conversation namespace, then common).
 const t: MessageItemProps['t'] = makeTranslate(zh, commonZh)
@@ -160,6 +163,157 @@ describe('MessageItem arms', () => {
    )
    expect(unknownView.getByText(/未知 surface 事件：surface\/next/)).toBeTruthy()
  })
+
+  it('collapses retry details behind the durable model retry status', () => {
+    vi.useFakeTimers()
+    vi.setSystemTime(10_000)
+    const view = render(
+      <MessageItem
+        t={t}
+        retryActive
+        node={{
+          kind: 'model-retry',
+          seq: 5,
+          time: 10_000,
+          retryState: 'scheduled',
+          turn: 1,
+          step: 0,
+          provider: 'mock',
+          mode: 'normal',
+          policyKey: 'mock-normal',
+          retry: 1,
+          maxRetries: 2,
+          delayMs: 2_500.4,
+          failure: { code: 'TRANSPORT', message: '连接被重置' },
+        }}
+      />,
+    )
+    const details = view.container.querySelector('details')
+    const summary = view.container.querySelector('summary')
+    expect(details?.open).toBe(false)
+    expect(details?.dataset.active).toBe('true')
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 3s')
+    expect(view.getByText('重试延迟：').parentElement?.textContent).toBe('重试延迟：2500ms')
+    expect(view.getByText('失败原因：').parentElement?.textContent).toBe('失败原因：连接被重置')
+
+    act(() => { vi.advanceTimersByTime(1_100) })
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 2s')
+    act(() => { vi.advanceTimersByTime(1_000) })
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 1s')
+
+    view.rerender(
+      <MessageItem
+        t={t}
+        retryActive
+        node={{
+          kind: 'model-retry',
+          seq: 6,
+          time: 12_100,
+          retryState: 'scheduled',
+          turn: 2,
+          step: 0,
+          provider: 'mock',
+          mode: 'normal',
+          policyKey: 'mock-normal',
+          retry: 2,
+          maxRetries: 2,
+          delayMs: 3_500.4,
+          failure: { code: 'TRANSPORT', message: '再次断开' },
+        }}
+      />,
+    )
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（2/2） · 4s')
+
+    if (summary === null) throw new Error('retry summary missing')
+    fireEvent.click(summary)
+    expect(details?.open).toBe(true)
+
+    view.rerender(
+      <MessageItem t={t} node={{
+        kind: 'model-retry',
+        seq: 6,
+        time: 12_100,
+        retryState: 'started',
+        turn: 2,
+        step: 0,
+        provider: 'mock',
+        mode: 'normal',
+        policyKey: 'mock-normal',
+        retry: 2,
+        maxRetries: 2,
+        delayMs: 3_500.4,
+        failure: { code: 'TRANSPORT', message: '再次断开' },
+      }}
+      />,
+    )
+    expect(details?.dataset.active).toBeUndefined()
+    expect(view.getByRole('status').textContent).toBe('已重试模型请求（2/2） · 4s')
+
+    view.rerender(
+      <MessageItem t={t} node={{
+        kind: 'model-retry',
+        seq: 7,
+        time: 12_100,
+        retryState: 'started',
+        turn: 3,
+        step: 0,
+        provider: 'mock',
+        mode: 'always',
+        policyKey: 'mock-always',
+        retry: 3,
+        delayMs: 3_500.4,
+        failure: { code: 'TRANSPORT', message: '继续重试' },
+      }}
+      />,
+    )
+    expect(view.getByRole('status').textContent).toBe('已重试模型请求（3/∞） · 4s')
+
+    view.rerender(
+      <MessageItem t={t} node={{
+        kind: 'model-retry',
+        seq: 8,
+        time: 12_100,
+        retryState: 'cancelled',
+        turn: 4,
+        step: 0,
+        provider: 'mock',
+        mode: 'normal',
+        policyKey: 'mock-normal',
+        retry: 1,
+        maxRetries: 2,
+        delayMs: 3_500.4,
+        failure: { code: 'TRANSPORT', message: '用户取消' },
+      }}
+      />,
+    )
+    expect(view.getByRole('status').textContent).toBe('模型请求重试已取消（1/2） · 4s')
+  })
+
+  it('synchronizes the countdown when an inactive retry becomes active at the one-second floor', () => {
+    vi.useFakeTimers()
+    vi.setSystemTime(10_000)
+    const node = {
+      kind: 'model-retry',
+      seq: 5,
+      time: 10_000,
+      retryState: 'scheduled',
+      turn: 1,
+      step: 0,
+      provider: 'mock',
+      mode: 'normal',
+      policyKey: 'mock-normal',
+      retry: 1,
+      maxRetries: 2,
+      delayMs: 5_000,
+      failure: { code: 'TRANSPORT', message: '连接被重置' },
+    } as const
+    const view = render(<MessageItem t={t} node={node} />)
+    expect(view.getByRole('status').textContent).toBe('等待重试模型请求（1/2） · 5s')
+
+    act(() => { vi.advanceTimersByTime(4_200) })
+    view.rerender(<MessageItem t={t} node={node} retryActive />)
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 1s')
+  })
 })

 describe('formatMessageClock', () => {
--- a/packages/client/ui-conversation/tests/chat-view.spec.tsx
+++ b/packages/client/ui-conversation/tests/chat-view.spec.tsx
@@ -7,8 +7,9 @@ import { afterEach, beforeEach, describe, expect, it, vi } from 'vitest'
 import { Profiler } from 'react'
 import { act, cleanup, fireEvent, render } from '@testing-library/react'
 import type {
-  AssistantMessageNode, CommandNode, ConversationNode, ConversationSnapshot, RunningToolCall, SessionId,
-  SessionListState, ToolResultNode, UserMessageNode, WorkspaceListState,
+  AssistantMessageNode, CommandNode, ConversationNode, ConversationSnapshot,
+  ModelRetryNode, RunningToolCall, SessionId, SessionListState, ToolResultNode,
+  UserMessageNode, WorkspaceListState,
 } from '@deepseek-ai/dsh-client-runtime/client'
 import { bindSnapshotSelector } from '@deepseek-ai/dsh-client-web-react'
 import { createSnapshotStore, PendingWait } from '@deepseek-ai/dsh-client-runtime/client'
@@ -67,6 +68,13 @@ const user = (seq: number, text: string): UserMessageNode => ({
 const assistant = (seq: number, text: string, turn = 1): AssistantMessageNode => ({
  kind: 'assistant', seq, time: seq * 1_000, turn, step: 1, blocks: [{ kind: 'text', text }],
 })
+const retry = (seq: number): ModelRetryNode => ({
+  kind: 'model-retry', seq, time: seq * 1_000, turn: 1, step: 0,
+  retryState: 'scheduled',
+  provider: 'mock', mode: 'normal', policyKey: 'mock-normal',
+  retry: 1, maxRetries: 2, delayMs: 450,
+  failure: { code: 'TRANSPORT', message: '连接被重置' },
+})
 const toolResult = (seq: number, callId: string, name = 'bash'): ToolResultNode => ({
  kind: 'tool-result', seq, time: seq * 1_000, callId,
  call: { name, argsRaw: `{"command":"cmd-${callId}","description":"run ${callId}"}` },
@@ -155,6 +163,17 @@ describe('chat-flow derivation', () => {
    expect(flowKeys(deriveChatFlow([...nodes, toolResult(7, 'd')]))).toBe('n1|n2|g3|n5|g6')
  })

+  it('reuses one stable row for consecutive retry turns', () => {
+    const first = retry(2)
+    const second = { ...retry(3), turn: 2, retry: 2 }
+    const initial = deriveChatFlow([user(1, 'try'), first])
+    const updated = deriveChatFlow([user(1, 'try'), first, second])
+    expect(flowKeys(initial)).toBe('n1|n2')
+    expect(flowKeys(updated)).toBe('n1|n2')
+    expect(updated).toHaveLength(2)
+    expect(updated[1]?.kind === 'node' && updated[1].node).toBe(second)
+  })
+
  it('skips render-nothing assistant nodes so tool runs stay one group', () => {
    // A tool-call-only step message (and blank text/reasoning) renders nothing:
    // it must not split the run into two groups with an empty line between.
@@ -227,6 +246,47 @@ describe('ChatView', () => {
    expect(view.getByText('run a')).toBeTruthy()
  })

+  it('animates only the latest unresolved model retry', () => {
+    const retryNode = retry(2)
+    const nextRetry = { ...retry(3), turn: 2, retry: 2 }
+    const context = {
+      kind: 'context', seq: 4, time: 4_000, content: [], source: null,
+    } as const satisfies ConversationNode
+    const h = makeHarness({ nodes: [user(1, 'try'), retryNode], running: true })
+    const view = render(<h.ChatView {...h.props} />)
+    const disclosure = view.container.querySelector('details')
+    expect(disclosure?.dataset.active).toBe('true')
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（1/2） · 1s')
+
+    act(() => {
+      h.set({ nodes: [user(1, 'try'), retryNode, nextRetry] })
+    })
+    expect(view.getAllByRole('status')).toHaveLength(1)
+    expect(view.container.querySelector('details')).toBe(disclosure)
+    expect(view.getByRole('status').textContent).toBe('正在重试模型请求（2/2） · 1s')
+
+    act(() => {
+      h.set({
+        nodes: [
+          user(1, 'try'),
+          retryNode,
+          { ...nextRetry, retryState: 'started' },
+          context,
+          assistant(5, 'done'),
+        ],
+        running: false,
+      })
+    })
+    expect(disclosure?.dataset.active).toBeUndefined()
+    expect(view.getByRole('status').textContent).toBe('已重试模型请求（2/2） · 1s')
+
+    act(() => {
+      h.set({ nodes: [user(1, 'try'), { ...retry(6), retryState: 'cancelled' }], running: true })
+    })
+    expect(disclosure?.dataset.active).toBeUndefined()
+    expect(view.getByRole('status').textContent).toContain('重试已取消')
+  })
+
  it('the expanded row Inspect pill hands the call id to inspectCall', () => {
    const h = makeHarness({
      nodes: [toolResult(3, 'a')],
--- a/packages/client/ui-trajectory/README.i18n.yaml
+++ b/packages/client/ui-trajectory/README.i18n.yaml
@@ -3,4 +3,4 @@
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/client/ui-trajectory/README.md
 README.md: b9c8b849b3454fe46e1fc37713d9d3b9449734cf
-README.zh.md: 19ae5050a4c4f7dfe80de0ab58e772e9d26e3f6a
+README.zh.md: 6ddc32f2f27c93f8ccc80b3d9b31d56d3cf4dd94
--- a/packages/client/ui-trajectory/README.zh.md
+++ b/packages/client/ui-trajectory/README.zh.md
@@ -2,7 +2,7 @@

 [English](README.md) | 中文

-Trajectory 渲染按轮次组织的事件记录表，其中可选择用户、助手、工具和嵌套子工具记录。较粗的分割线标示轮次边界，紧凑的行内标记标识步骤，主记录表仅保留索引、事件和内容；选择记录则会打开局部检查器，查看 token 用量、耗时、输入、输出和计时。固定在记录表上方的 Overview 区域从左到右投影记录的真实开始时间与耗时；拖选一个区间会将记录表聚焦到活动区间与该闭区间有重叠的所有记录，清除选择则恢复完整分支。runtime 的独立历史数据源提供原始上下文谱系，并投影因取消而冻结的助手和工具记录，因此 Trajectory 既不读取也不改变 Chat 会话快照。该包（package）保持为纯消费方插件（向会话的 `'conversation.view'` slot 环注册一个视图标签页，不提供服务，也不声明 Context 合并）。契约：api-contracts v3 §8。
+Trajectory 渲染按轮次组织的事件记录表，其中可选择用户、助手、工具和嵌套子工具记录。较粗的分割线标示轮次边界，紧凑的行内标记标识步骤，主记录表仅保留索引、事件和内容；选择记录则会打开局部检查器，查看 token 用量、耗时、输入、输出和计时。固定在记录表上方的 Overview 区域从左到右投影记录的真实开始时间与耗时；拖选一个区间会将记录表聚焦到活动区间与该闭区间有重叠的所有记录，清除选择则恢复完整分支。运行时的独立历史数据源提供原始上下文谱系，并投影因取消而冻结的助手和工具记录，因此 Trajectory 既不读取也不改变 Chat 会话快照。该包（package）保持为纯消费方插件（向会话的 `'conversation.view'` slot 环注册一个视图标签页，不提供服务，也不声明 Context 合并）。契约：api-contracts v3 §8。

 ## 模型体验

@@ -10,7 +10,7 @@ Trajectory 渲染按轮次组织的事件记录表，其中可选择用户、助

 #### KV Cache 影响

-无；该包（package）既不组装也不发送提供方请求。
+无；该包既不组装也不发送提供方请求。

 ## 已知限制与暂缓事项

--- a/packages/llm/llm-retry/README.i18n.yaml
+++ b/packages/llm/llm-retry/README.i18n.yaml
@@ -2,5 +2,5 @@
 # side as of the last confirmed-consistent state. Both languages carry equal authority;
 # after editing either side, bring the other along and re-record with:
 #   pnpm run verify-translation-pairing --write packages/llm/llm-retry/README.md
-README.md: 7a86652a794e70c4dfd00ab7427730387e3ec949
-README.zh.md: c66c04806597c11b5c64dcb01443bb84f489e0f5
+README.md: 8de3ea8c9321f04f5af1b0d7ab361f73eaabc822
+README.zh.md: 978854e9466e271535a10fcea406d0dcb5607285
--- a/packages/llm/llm-retry/README.md
+++ b/packages/llm/llm-retry/README.md
@@ -8,7 +8,7 @@ Each provider adapter owns an optional nested `retryPolicy`, captured when its r

 Both modes use bounded exponential backoff with symmetric jitter. A valid `providerRetryAfterMs` at or below `maxDelayMs` replaces local backoff without jitter. An over-cap provider delay makes normal mode delegate, while always mode uses its configured local backoff so it cannot terminate on that instruction.

-Before waiting, the plugin appends a non-surface `llm/retry` event with the provider, mode, canonical resolved-policy key, failure, and scheduled delay. The key includes every behavior-affecting field and sorts normal-mode codes because eligibility uses set membership. Retry numbers continue only across events with the same provider and complete policy key, so a route replacement with different limits, code membership, or backoff starts its own history. Normal events include the finite maximum; always events omit it, and UIs render `∞`. After the wait, the listener returns `{ kind: 'retry' }`, and the loop closes the failed turn and opens a retry turn over the same durable history. Cancellation and plugin disposal abort active backoff, drain active delegated recovery before applying the abort, and make a callback captured before disposal fail closed.
+Before waiting, the plugin appends a non-surface `llm/retry` event with the provider, mode, canonical resolved-policy key, failure, and scheduled delay. Its payload is available from the browser-safe `@deepseek-ai/dsh-llm-retry/types` subpath, so remote renderers can consume the durable status without loading the policy runtime. The key includes every behavior-affecting field and sorts normal-mode codes because eligibility uses set membership. Retry numbers continue only across events with the same provider and complete policy key, so a route replacement with different limits, code membership, or backoff starts its own history. Normal events include the finite maximum; always events omit it, and UIs render `∞`. After the wait, the listener returns `{ kind: 'retry' }`, and the loop closes the failed turn and opens a retry turn over the same durable history. Cancellation and plugin disposal abort active backoff, drain active delegated recovery before applying the abort, and make a callback captured before disposal fail closed.

 The separately published `./invariant` companion checks that every retry record names the current open turn and latest closed step, matches the failed request's durable provider, carries non-empty provider and policy identities, has mode-specific bounds, a unique step record, the correct provider-policy retry number, and a bounded timer delay. Full jitter may schedule zero milliseconds at its lower boundary.

--- a/packages/llm/llm-retry/README.zh.md
+++ b/packages/llm/llm-retry/README.zh.md
@@ -8,7 +8,7 @@

 两种 mode 都使用带对称 jitter 的有界指数退避。有效 `providerRetryAfterMs` 不超过 `maxDelayMs` 时会替换本地退避，并且不加 jitter。超出上限的提供方延迟会使 normal mode 继续委托；always mode 则改用已配置的本地退避，避免该指令终止重试。

-等待前，插件会追加一条不进入表层的 `llm/retry` 事件，其中包含提供方、mode、已解析策略的规范 key、失败和计划延迟。该 key 包含所有影响行为的字段，并对 normal mode 的 code 排序，因为合格性采用集合成员关系判断。只有提供方与完整策略 key 都相同的事件才会延续重试编号；因此，用限制、code 成员关系或退避不同的路由替换后，会开始自己的历史。normal 事件包含有限上限；always 事件省略该上限，UI 会渲染 `∞`。等待结束后，监听器返回 `{ kind: 'retry' }`，循环关闭失败轮次，并在同一持久历史上开启重试轮次。取消与插件 dispose 会中止活跃退避，在应用中止前排空活跃的委托恢复，并使 dispose 前捕获的 callback 只能以失败结束。
+等待前，插件会追加一条不进入表层的 `llm/retry` 事件，其中包含提供方、mode、已解析策略的规范 key、失败和计划延迟。该载荷由可安全用于浏览器的 `@deepseek-ai/dsh-llm-retry/types` 子路径导出，因此远程渲染器无需加载策略运行时即可使用该持久状态。该 key 包含所有影响行为的字段，并对 normal mode 的 code 排序，因为合格性采用集合成员关系判断。只有提供方与完整策略 key 都相同的事件才会延续重试编号；因此，用限制、code 成员关系或退避不同的路由替换后，会开始自己的历史。normal 事件包含有限上限；always 事件省略该上限，UI 会渲染 `∞`。等待结束后，监听器返回 `{ kind: 'retry' }`，循环关闭失败轮次，并在同一持久历史上开启重试轮次。取消与插件 dispose 会中止活跃退避，在应用中止前排空活跃的委托恢复，并使 dispose 前捕获的 callback 只能以失败结束。

 单独发布的 `./invariant` 配套模块会检查每个重试记录是否指向当前开启轮次及其最新已关闭步骤，是否与失败请求的持久提供方匹配，是否携带非空的提供方与策略标识，是否满足 mode 特定边界，是否拥有唯一步骤记录和正确的提供方策略重试编号，以及是否携带有界定时器延迟。完整 jitter 可以在下界调度为零毫秒。

--- a/packages/llm/llm-retry/package.json
+++ b/packages/llm/llm-retry/package.json
@@ -15,11 +15,16 @@
      "types": "./lib/types/invariant.d.ts",
      "default": "./lib/invariant.js"
    },
+    "./types": {
+      "types": "./lib/types/types.d.ts",
+      "default": "./lib/types/types.js"
+    },
    "./package.json": "./package.json"
  },
  "files": [
    "lib/index.js",
    "lib/invariant.js",
+    "lib/types/**/*.js",
    "lib/types/**/*.d.ts",
    "lib/types/**/*.d.ts.map",
    "src"
--- a/packages/llm/llm-retry/src/index.ts
+++ b/packages/llm/llm-retry/src/index.ts
@@ -38,6 +38,8 @@ declare module '@deepseek-ai/dsh-session' {
  }
 }

+export type { LlmRetryEventData } from './types.ts'
+
 export const name = 'llm-retry'
 export const inject = ['agents']

--- a/packages/llm/llm-retry/src/types.ts
+++ b/packages/llm/llm-retry/src/types.ts
@@ -0,0 +1,25 @@
+import type { LlmFailure } from '@deepseek-ai/dsh-llm/types'
+
+/** Durable payload recorded before one provider-routed model-request retry wait. */
+export type LlmRetryEventData =
+  | {
+    turn: number
+    step: number
+    provider: string
+    mode: 'normal'
+    policyKey: string
+    retry: number
+    maxRetries: number
+    delayMs: number
+    failure: LlmFailure
+  }
+  | {
+    turn: number
+    step: number
+    provider: string
+    mode: 'always'
+    policyKey: string
+    retry: number
+    delayMs: number
+    failure: LlmFailure
+  }
--- a/packages/llm/llm-retry/tests/retry.spec.ts
+++ b/packages/llm/llm-retry/tests/retry.spec.ts
@@ -12,7 +12,8 @@ import type {
  StreamChunk,
 } from '@deepseek-ai/dsh-llm'
 import SessionStore, { SessionId } from '@deepseek-ai/dsh-session'
-import type { SessionEvent } from '@deepseek-ai/dsh-session'
+import type { SessionEvent, SessionEventMap } from '@deepseek-ai/dsh-session'
+import type { LlmRetryEventData } from '@deepseek-ai/dsh-llm-retry/types'
 import SystemPrompt from '@deepseek-ai/dsh-system-prompt'
 import ToolRegistry, { defineContentToolFixture } from '@deepseek-ai/dsh-tools'
 import AgentRegistry from '@deepseek-ai/dsh-agent'
@@ -22,6 +23,10 @@ import * as retry from '../src/index.ts'

 type ScriptEntry = Error | Iterable<StreamChunk> | AsyncIterable<StreamChunk>

+it('keeps the browser-safe retry payload identical to the session event', () => {
+  expectTypeOf<LlmRetryEventData>().toEqualTypeOf<SessionEventMap['llm/retry']>()
+})
+
 class ScriptedAdapter extends LlmAdapter {
  readonly requests: GenerateOptions[] = []
  private retryPolicies: Readonly<Record<string, ResolvedRetryPolicy | undefined>> = {}
--- a/pnpm-lock.yaml
+++ b/pnpm-lock.yaml
@@ -1072,6 +1072,9 @@ importers:
      '@deepseek-ai/dsh-llm':
        specifier: workspace:^
        version: link:../../llm/llm
+      '@deepseek-ai/dsh-llm-retry':
+        specifier: workspace:^
+        version: link:../../llm/llm-retry
      '@deepseek-ai/dsh-session':
        specifier: workspace:^
        version: link:../../core/session
@@ -1094,6 +1097,9 @@ importers:
      '@deepseek-ai/dsh-invariants':
        specifier: workspace:^
        version: link:../../support/invariants
+      '@deepseek-ai/dsh-timeout':
+        specifier: workspace:^
+        version: link:../../util/timeout
      '@types/react':
        specifier: ~18.3.1
        version: 18.3.31
--- a/scripts/snapshots/translation-prompt-v4/request-response.expected.json
+++ b/scripts/snapshots/translation-prompt-v4/request-response.expected.json
--- a/tsconfig.base.json
+++ b/tsconfig.base.json
@@ -56,6 +56,7 @@
      "@deepseek-ai/dsh-goal/client": ["./packages/goal/goal/src/client.ts"],
      "@deepseek-ai/dsh-llm/types": ["./packages/llm/llm/src/types.ts"],
      "@deepseek-ai/dsh-llm/brand": ["./packages/llm/llm/src/brand.ts"],
+      "@deepseek-ai/dsh-llm-retry/types": ["./packages/llm/llm-retry/src/types.ts"],
      "@deepseek-ai/dsh-llm/message": ["./packages/llm/llm/src/message.ts"],
      "@deepseek-ai/dsh-commands/brand": ["./packages/ui/commands/src/brand.ts"],
      "@deepseek-ai/dsh-tui/prompt": ["./packages/ui/tui/src/prompt.ts"],