feat(history,gateway): D4 会话未读/已读游标 — unreadCount + POST /sessions/{id}/read + session_read 广播

- chat_sessions 增 read_at（毫秒，与 chat_messages.created_at 同单位）；迁移 0010 并把既有会话回填为各自最新消息时间（升级当天不误报）
- ListSessions 单查询派生 unreadCount（无 N+1，复用 idx_chat_msgs_session_created 覆盖索引）
- MarkSessionRead 只前进 MAX(read_at, ?)；未知会话 ErrSessionNotFound；游标方法挂在网关消费侧窄接口（沿用 GetMessage/messageLookup 先例，避免污染 runtime 接口与其 fake）
- POST /api/v1/sessions/{id}/read：writeAuth + 400/404/501，成功广播 session_read、失败路径不广播
- Docs/api-client v0.37（§1 unreadCount / 新行 8a / §3.9 session_read / §4.2 用法）
- 含迁移回填测试、store/handler 测试与变异验证；顺带把 migrate_pinned_test 的「回退一步=0009」硬编码改为对迁移数量免疫

Docs/api-client.md

@@ -26,13 +26,14 @@
 | 1b | `/api/v1/auth/logout` | POST | refreshToken 或 JWT | ✅ | **登出**（B2）：body `{refreshToken}` 撤**单个** grant（幂等）；带 `Authorization: Bearer <jwt>` 且无 body → 撤**该账号全部** grant（全设备登出，返回 `{revokedCount}`）。refresh 相关操作成功才审计 |
 | 2 | `/api/v1/health` | GET | 否 | ✅ | 健康检查 `{data:{status:"ok"}}` |
 | 3 | `/api/v1/chat` | POST | JWT | ✅ | `{text,image?,deviceId?,mode?,stream?}` → `{data:{reply}}`；`stream:true` 时经 WS 触发流式。`image` 为 base64 data: URI 或 http(s) URL，入视觉路由（design/06）——**真视觉需配 `LLM_VISION_MODEL`/凭据，无凭据时降级为主模型直接看图或视觉·OCR 工具文本描述**；`deviceId` 为**客户端生成的具体端标识**（design/16 §4.5，如 `web:chrome:a1b2c3`，可选），落库到消息、便于主人知道从哪个端发来；`mode` 为**行为模式**（design/01 §3 / design/16 §5.1，如 `chat`/`code`/`home`/`docs`/…，可选），随该轮次落库——IM 恒 `chat`、缺省回落 `chat`，供前端按模式渲染气泡（仅 `chat` 才解析 `<p>` 拆气泡）。**会话编排（design/15）**：非流式且该会话「正忙」时**不阻塞**，返回 `202` `{data:{queued:true,queueDepth}}`（稍后按上下文续接）；待处理队列满时返回 `409` `{error:{code:"chat_busy"}}` |
-| 4 | `/api/v1/sessions` | GET | JWT | ✅ | `{data:[session...]}`。**会话持久化（design/16）**：注入 HistoryStore 后返回持久化会话元数据 `{data:[{id,type("main"|"sub"),title?,createdAt,pinned,lastMessage?}]}`（**排序（D5）：`pinned=true` 优先，其次主会话 `id="web"` 恒在前，再按 `createdAt` 升序**；次级 `web:<uuid>`；`pinned` 恒返回 bool；`lastMessage` 为**最后一条消息预览** `{content,role(user|assistant),createdAt}`，content 折叠空白截前 60 rune，空会话省略）；未注入时退化回渠道会话列表 |
+| 4 | `/api/v1/sessions` | GET | JWT | ✅ | `{data:[session...]}`。**会话持久化（design/16）**：注入 HistoryStore 后返回持久化会话元数据 `{data:[{id,type("main"|"sub"),title?,createdAt,pinned,unreadCount,lastMessage?}]}`（**排序（D5）：`pinned=true` 优先，其次主会话 `id="web"` 恒在前，再按 `createdAt` 升序**；次级 `web:<uuid>`；`pinned` 恒返回 bool；**`unreadCount`（D4）恒返回整数**——该会话中 `createdAt >` 已读游标的消息条数（0 = 无未读），与 `lastMessage` **同一条 SQL 派生**、无 N+1；`lastMessage` 为**最后一条消息预览** `{content,role(user|assistant),createdAt}`，content 折叠空白截前 60 rune，空会话省略）；未注入时退化回渠道会话列表 |
 | 5 | `/api/v1/sessions` | POST | JWT(owner/admin) | ✅ | **新建次级会话** → `201 {data:{sessionId}}`（`sessionId="web:<uuid>"`）；未注入 HistoryStore 回 `501 not_implemented` |
 | 6 | `/api/v1/sessions/{id}/messages` | GET | JWT(owner/admin) | ✅ | **热分页拉历史**：`?limit=&before=&platform=&after=`（A5）默认 `limit=50`（上限 200）。**`before`**=上一页最旧 `createdAt`（毫秒），按时间**倒序**；**`after`**=<`createdAt` 毫秒>，按时间**升序**返回 `created_at > after` 的消息（**正向增量**，客户端把上次看到的 `createdAt` 当游标补断线期间漏的，解决重连只回放 50 条的缺口）；**`after` 与 `before` 互斥**，同传回 `400`；负数/非数字回 `400`。`platform` 为**可选**来源平台过滤（design/16 §4.4）——Web 端不传（=全量共享主会话），IM 适配器传**本平台**（如 `telegram`）只拉本渠道消息；`deviceId` 为**客户端具体端标识**；`senderName` 为**第三方发送者昵称**；**`q`**（D2）=**会话内内容搜索**（大小写不敏感 ASCII LIKE，`%`/`_`/`\` 已转义为字面量），仍按时间**倒序**、`before`/`limit` 同分页语义；`q` 与 `after` **互斥**（同传 `400`），空 `q` 视为未传（走普通分页），搜索**严格限定在 `{id}` 会话内**（绝不跨会话，群聊/陌生人隔离会话不会外泄）；非法 `limit`/`before`/`after` 回 `400`；未注入 HistoryStore 回 `501` |
 | 6a | `/api/v1/messages/{id}` | PATCH | JWT(owner/admin) | ✅ | **编辑单条消息**（D3）：body `{content}`（非空、≤4000 rune，空/纯空白/超长回 `400`）→ **`{data:<消息DTO>}`**（`data` **直接就是**更新后的完整消息 DTO，含 `id/sessionId/role/content/.../createdAt`，**无 `message` 包一层**；**只改 `content`**，role/时间戳/平台保持不变）；消息不存在回 `404 not_found`；未注入 HistoryStore 回 `501`。**成功后广播下行 `message_edited`**（§3.8，账号其它端免刷新同步）。**重发不单独设端点**——客户端用「编辑后的内容再 `POST /chat`」即可（避免重复轮次语义） |
 | 6b | `/api/v1/messages/{id}` | DELETE | JWT(owner/admin) | ✅ | **删除单条消息**（D3）→ `{data:{deleted:<id>}}`；不存在回 `404 not_found`；未注入 HistoryStore 回 `501`。**成功后广播下行 `message_deleted`**（§3.8，账号其它端按 `id` 移除） |
 | 7 | `/api/v1/sessions/{id}` | DELETE | JWT(owner/admin) | ✅ | **删除次级会话及其消息** → `{data:{deleted:<id>}}`；主会话不可删回 `403 forbidden`；未注入 HistoryStore 回 `501` |
 | 8 | `/api/v1/sessions/{id}` | PATCH | JWT(owner/admin) | ✅ | **会话元数据更新**（A4 改名 + D5 置顶）：body `{title?,pinned?}`（**至少一个字段**，都缺回 `400`）。`title` trim、≤80 rune，空/超长回 `400`；`pinned` 为 bool（置顶/取消，主会话与次级会话均可）。响应只回显**本次变更**的字段：`{data:{id,title?}}` / `{data:{id,pinned?}}` / 两者都改则都有（**title-only 保持 A4 原 `{id,title}` 形状**）；不存在回 `404`；主会话(`id="web"`)可改名/置顶但不可删；未注入 HistoryStore 回 `501`。另：会话**首条用户消息到达时自动命名** title（取前 20 rune 折叠空白；纯空白回落 `主会话`/`次级会话`），列表不再空名 |
+| 8a | `/api/v1/sessions/{id}/read` | POST | JWT(owner/admin) | ✅ | **推进已读游标**（D4 / Android 协同 B3）：body `{before?}`——`before` 为**可选**的毫秒 Unix 时间戳，**缺省（或空 body）= 现在**，即「我看到的都算已读」。→ `200 {data:{sessionId,readAt}}`（`readAt` 为推进后的游标毫秒）。**判定**：`createdAt <= readAt` 的消息算已读（列表 `unreadCount` 只数 `createdAt > readAt` 的）；**语义**：**会话级单一游标、单 owner 账号**——读一处 = **该账号所有端已读**（不做 per-device 游标，与「同一账号所有浏览器都属于同一 owner」一致）；**只前进不后退**（SQL `read_at = MAX(read_at, before)`），所以乱序/并发/重试请求幂等安全、绝不会把已读变回未读。`before` 为 `0`/负数/非数字、或 body 不是 JSON 对象 → `400`；未知会话（含隔离 `sess-xxx`，不在 `chat_sessions`）→ `404`；未注入 HistoryStore → `501 not_implemented`。**成功后广播下行 `session_read`**（§3.9）让同账号其它端清红点；**失败路径不广播**。**升级回填**：迁移 `0010` 把既有会话游标回填为「该会话现有最新消息时间」，升级当天不误报未读 |
 | 9 | `/api/v1/personas` | GET | JWT | ✅ | `{data:[persona...]}` |
 | 10 | `/api/v1/audio/transcribe` | POST | JWT(owner) | ✅ | **语音→文本**：body `{audio:"<base64>",format?,deviceId?}` → `{data:{text, provider?}}`（走后端 STT，百炼 `qwen-audio-3.0-asr-flash`；**当结果为空文本/silent 降级时响应体带 `provider`**（如 `"silent"`，非空文本省略），客户端据此提示「没识别到声音」） |
 | 11 | `/api/v1/audio/speech` | POST | JWT(owner) | ✅ | **文本→音频**：body `{text,voice?,format?,instruction?}` → `200` 音频字节（`Content-Type: audio/<format>`，非 JSON envelope，走后端 TTS，百炼 `cosyvoice-v3-flash`，复刻音色可带 `instruction` 情绪指令；**静默降级空音频时仍 `500 tts_empty`，并附 `X-Voice-Provider: silent` 响应头**，客户端据此提示「没生成声音」） |
@@ -144,7 +145,7 @@
 ## 3. WebSocket `/ws/chat`（实时）
 
 - **握手**: `ws://<host>:8080/ws/chat?token=<JWT>` 或 `Authorization: Bearer`；未认证在握手前回 `401` JSON。
-- **下行（服务端→客户端）**: 经 bus 广播 `{type:"<event>", data:..., turnId?, seq?, sessionId?}`（`turnId`/`seq` 为流式关联字段，见 §3.1；`sessionId` 为**会话持久化（design/16）新增**，能确定会话的事件带上，供客户端按激活会话过滤；**v0.36 起结构化流帧 `message.*` / `thinking.*` / `tool_call.*` / `tool_execution.update` 也带 `sessionId` + `turnId`**——多会话并发流式时客户端据此把 delta 路由到激活会话，见 §3.1）——`message.start/delta/end` / `thinking.start/delta/end` / `tool_call.start/delta/end` / `tool_execution.update`（结构化流，design/04 §1.1），以及向后兼容的 `stream_chunk` / `stream_end` / `thinking`(思维链) / `mood`(情绪变化，见 §3.7) / `reminder` / `device_update` / `tool_used`(工具回显) / `subagent_event` / `graph_update` / **`message_persisted`**（会话持久化增量，见 §3.3）/ **`message_edited`**、**`message_deleted`**（消息编辑/删除增量，见 §3.8）。**会话编排状态（design/15，均为广播，`data.sessionId` 区分会话）**：`session_busy` / `pending_ack` / `pending_drain` / `session_idle` / `session_interrupted` / `pending_full`（见 §3.2）。语音回复走 `voice_output`（定向回请求连接，非广播）。
+- **下行（服务端→客户端）**: 经 bus 广播 `{type:"<event>", data:..., turnId?, seq?, sessionId?}`（`turnId`/`seq` 为流式关联字段，见 §3.1；`sessionId` 为**会话持久化（design/16）新增**，能确定会话的事件带上，供客户端按激活会话过滤；**v0.36 起结构化流帧 `message.*` / `thinking.*` / `tool_call.*` / `tool_execution.update` 也带 `sessionId` + `turnId`**——多会话并发流式时客户端据此把 delta 路由到激活会话，见 §3.1）——`message.start/delta/end` / `thinking.start/delta/end` / `tool_call.start/delta/end` / `tool_execution.update`（结构化流，design/04 §1.1），以及向后兼容的 `stream_chunk` / `stream_end` / `thinking`(思维链) / `mood`(情绪变化，见 §3.7) / `reminder` / `device_update` / `tool_used`(工具回显) / `subagent_event` / `graph_update` / **`message_persisted`**（会话持久化增量，见 §3.3）/ **`message_edited`**、**`message_deleted`**（消息编辑/删除增量，见 §3.8）/ **`session_read`**（已读游标推进，见 §3.9）。**会话编排状态（design/15，均为广播，`data.sessionId` 区分会话）**：`session_busy` / `pending_ack` / `pending_drain` / `session_idle` / `session_interrupted` / `pending_full`（见 §3.2）。语音回复走 `voice_output`（定向回请求连接，非广播）。
 
   > **`history` 回放事件（design/16，新增）**：WS 握手升级成功后、该连接下一帧即收到 `{type:"history", sessionId:"web", data:{messages:[{id,sessionId,role,content,mode?,mood?,image?,platform?,deviceId?,createdAt}]}}`——回放主会话最新 50 条（`thinking` 已剥离，不泄给非 owner；默认主会话，次级会话由客户端按激活会话走 `GET /sessions/{id}/messages` 拉取）。客户端据此在重连/新开连接后补全上下文。消息体含 `senderName?`（第三方昵称，design/16 §4.6，主人私聊为空）。
 
@@ -218,6 +219,12 @@
 - **客户端处理**：与 `message_persisted` **同一过滤规则**——仅当 `sessionId === activeSessionId` 时处理（`message_deleted` 的 `sessionId` 缺失时按 `id` 兜底：消息 `id` 全局唯一，只会命中它自己那条）。`message_edited` → **按 `id` 原位更新该条的 `content`**（不新增行：找不到该 `id` 即忽略，避免把别的会话的消息注入当前列表）；`message_deleted` → **按 `id` 从列表移除**。两个操作都**不触发整页重拉**——这正是它相对「重拉绕行」的收益。
 - **失败不广播**：`404`（消息不存在）/ `400`（内容非法）/ `501`（未注入 HistoryStore）**不**产生任何事件，客户端不会因一次失败请求而丢掉本地消息。
 
+### 3.9 已读游标事件（`session_read`，D4）
+- **触发**：`POST /api/v1/sessions/{id}/read` **成功推进游标后**由网关广播（与 `message_persisted` 同一条 hub 广播路径：**全连接广播 + 客户端按 `sessionId` 过滤**，hub 无按会话订阅）。用途是**同账号多端清红点**：一端读了，其它端不用重拉列表就把该会话的未读清零。
+- **载荷**：`{type:"session_read", sessionId, data:{readAt}}`——`readAt` 为该会话推进后的**已读游标**（毫秒 Unix 时间戳，与该会话消息的 `createdAt` 同单位）。
+- **客户端处理**：仅当 `sessionId === activeSessionId`（或该会话在自己维护的会话列表里）时处理 → 把该会话 `unreadCount` 置 0（更精确的做法是按 `readAt` 重算本地未读；本地没有更完整的数据时直接清零也不会错，因为 `readAt` 只前进）。**它只驱动渲染**：唯一真源仍是后端 `chat_sessions.read_at`，漏收该帧的端在下一次 `GET /api/v1/sessions` 自然收敛（不需要补发、不需要重连）。
+- **失败不广播**：`400`（`before` 非法）/ `404`（会话不存在）/ `501`（未注入 HistoryStore）**不**产生事件，其它端不会误清红点。
+
 ## 4. 客户端接入指引（Web / 移动 / 原生 / IM bridge 通用）
 
 ### 4.1 一次性接入流程（三步走）
@@ -226,8 +233,9 @@
 3. **带端标识**（可选但推荐）：每条消息/请求可带 `deviceId`（见 §4.3），让她知道你从哪个端来。
 
 ### 4.2 会话与消息用法（聊天）
-- **列出会话**：`GET /api/v1/sessions` → `[{id,type:"main"|"sub",title?,createdAt,pinned}]`。**排序：置顶会话（`pinned=true`）最前，其次主会话 `id="web"`**（她与你的主线，最大化继承上下文），再按创建时间升序；次级 `web:<uuid>` 由 `POST /api/v1/sessions` 新建（临时小事/专业问题，独立上下文回溯）。
+- **列出会话**：`GET /api/v1/sessions` → `[{id,type:"main"|"sub",title?,createdAt,pinned,unreadCount,lastMessage?}]`。**排序：置顶会话（`pinned=true`）最前，其次主会话 `id="web"`**（她与你的主线，最大化继承上下文），再按创建时间升序；次级 `web:<uuid>` 由 `POST /api/v1/sessions` 新建（临时小事/专业问题，独立上下文回溯）。
 - **改名 / 置顶**：`PATCH /api/v1/sessions/{id}` body `{title?}`（≤80 rune）或 `{pinned?:bool}`（至少一个字段）→ `{data:{id,title?,pinned?}}`（只回显变更字段）。
+- **未读 / 已读游标（D4）**：列表项的 **`unreadCount`** 直接用于会话列表红点（恒返回整数，0 = 无未读）。客户端进入会话、或滚到底看到最新消息后调用 **`POST /api/v1/sessions/{id}/read`**：body `{"before":<毫秒>}`（如「本次已渲染的最新一条的 `createdAt`」），或**省略 body** = 标记到**现在**。后端**只前进**地推进该会话**唯一**游标（单 owner 账号：读一处 = 所有端已读），并广播 `session_read`（§3.9）——同账号其它端**免刷新**清红点。列表 → 读 → 再看列表即可自洽（漏收广播的端下次拉列表自然收敛）。
 - **拉取激活会话历史**：`GET /api/v1/sessions/{id}/messages?limit=50&before=<ms>&platform=&q=`。
   - **热分页**：`limit` 默认 50（上限 200），`before` = 当前最旧一条的 `createdAt`（毫秒），向上翻页。`platform` 可选（IM 适配器传本平台只拉本渠道；Web/通用客户端不传 = 全量共享主会话）。
   - **搜索**：`q` 非空 = **会话内**内容搜索（倒序，仍支持 `before`/`limit` 分页）；空 `q` = 普通分页；`q` 与 `after` 互斥。**不跨会话**。
@@ -253,6 +261,7 @@
 - 语音：上行 `voice_input`（base64 音频或已转写 text），下行 `voice_output`（合成音频 base64 + text）。
 
 ## 5. 变更记录
+- `v0.37`（**D4 未读状态 / 已读游标**，对客户端**新增一个语义字段、一个 REST 端点、一个下行事件**）：① `chat_sessions` 新增 **`read_at INTEGER NOT NULL DEFAULT 0`**（迁移 **0010**，down 可逆；**毫秒 Unix 时间戳**，与 `chat_messages.created_at` **同单位**）。② **语义**：**会话级单一已读游标 + 单 owner 账号**——同一账号所有端共用一个游标，**读一处 = 全端已读**（不做 per-device/per-user 游标，与「同一账号所有浏览器都属于同一 owner」一致）；`createdAt <= readAt` 的消息算已读，未读数 = `createdAt > readAt` 的条数。③ **`GET /api/v1/sessions` 每项新增恒返回的 `unreadCount`**（整数，0 = 无未读）：与 `lastMessage` **同一条 SQL 派生**（关联子查询 `COUNT(*) WHERE session_id = s.id AND created_at > s.read_at`，走已有 `idx_chat_msgs_session_created`，**无 N+1、不新增索引**）。④ **新增 `POST /api/v1/sessions/{id}/read`**（owner/admin，与 `PATCH /sessions/{id}` 同鉴权）：body `{before?}`（可选毫秒；缺省= 现在）→ `200 {data:{sessionId,readAt}}`；`before` 为 0/负数/非数字/body 非对象 → `400`，未知会话（含隔离 `sess-xxx`）→ `404`，未注入 HistoryStore → `501`。**游标只前进不后退**（`read_at = MAX(read_at, before)`），乱序/并发/重试请求幂等安全。⑤ **新增下行事件 `session_read`**：`{type:"session_read", sessionId, data:{readAt}}`，全连接广播 + 客户端按 `sessionId` 过滤，用于同账号其它端清红点；失败路径不广播（§3.9）。⑥ **升级回填**：0010 把既有会话游标回填为「该会话现有最新消息时间」，**升级当天不误报未读**（既有历史全部视为已读）。⑦ `HistoryStore`（网关消费侧契约）新增 `MarkSessionRead(ctx, sessionID, before)`；`SessionMeta` 新增 `UnreadCount`。**向后兼容**：仅新增字段/端点/事件，既有契约形状不变；不消费 `unreadCount`/`session_read` 的客户端行为不变。详见 §1 行 4/8a 与 §3.9。
 - `v0.36`（**A1 流式帧附加会话 / 回合标识**，对客户端**新增两个可选字段、无端点变更**）：结构化流帧 `message.start/delta/end`、`thinking.start/delta/end`、`tool_call.start/delta/end`、`tool_execution.update` 现在带 `sessionId` + `turnId`（**向后兼容**：只新增帧字段，`data` 结构与既有事件形状不变）。`sessionId` 由 loop 在**回合边界**从运行上下文注入（模型适配器在 loop 之下、天然不感知会话，保持零值；loop 外无会话上下文时该字段为空）；`turnId` 由网关**每回合**新生成（`newTurnID`）并经 `loop.WithTurnID` 挂进运行上下文，`runTurn` 是编排层与 legacy 直连路径**唯一**的回合收口点，两路都覆盖——同一会话同时只有一个 in-flight 回合，故同轮内所有帧的 `turnId` 一致。此前这几帧属于 design/16 §4.2 标注的「若知道就带上」缺口（v0.10 起记录）：**多会话并发流式时前端会把别的会话的 delta 接进当前会话气泡**，现在客户端按 `sessionId === activeSessionId` 过滤即可路由到激活会话。`stream_chunk` / `stream_end` 的 `sessionId`/`turnId` 行为不变。详见 §3.1。
 - `v0.35`（**可观测性收尾：后端日志 tail 端点**，对客户端**新增一个 admin 只读端点**）：`GET /api/v1/admin/observability/logs?level=&limit=`（owner/admin）→ `200 {data:[{time,level,source,msg}]}`，**最新在前**。数据源为 logger 新增的**有界环形缓冲**（`internal/logger`，容量 2000，满则丢最旧、并发安全、不阻塞、不改动既有控制台/文件输出），组合根在构建 logger 时挂上、在网关就绪后经 `SetLogBuffer` 注入；缓冲未注入回 `501 not_implemented`（不伪造空 tail）。`time` 为 **RFC3339 UTC（秒精度）**，与 `trace.startedAt`/`audit.when` 一致；`level ∈ debug|info|warn|error`（大小写不敏感）为**最低等级过滤**，空 = 全部，未知值 `400 invalid_level`；`limit` 缺省 `200`、上限 `1000`。`source` 取 `logger.WithSource` 的 `source=`（`Scoped` 记录回落 `module/sub`），无来源为空串。详见 §2「可观测性契约」。
 - `v0.34`（**D3 收尾：消息编辑 / 删除的 WS 实时广播**，对客户端**新增两个下行事件、无 REST 契约变更**）：`PATCH /api/v1/messages/{id}` 成功后广播 **`message_edited`**（`{type, sessionId, data:<更新后的完整消息DTO>}`，`thinking` 已剥离）、`DELETE /api/v1/messages/{id}` 成功后广播 **`message_deleted`**（`{type, sessionId, data:{id}}`）；两者复用 `message_persisted` 的同一 hub 广播路径（**全连接广播 + 客户端按 `sessionId` 过滤**，hub 无按会话订阅），客户端**按 `id` 原位改 `content` / 按 `id` 移除**，不再需要「改/删后重拉」的绕行。`message_deleted` 的 `sessionId` 由后端在删除前读取消息所属会话填入（`HistoryStore` 新增 `GetMessage`；读不到时客户端按全局唯一的 `id` 兜底）。`404`/`400`/`501` 等失败路径**不广播**。Web 端已消费（`lib/ws.ts` + `store/useAppStore` 新增 `updateMessageById`/`removeMessageById`）。详见 §3.8。

internal/db/migrate_pinned_test.go

@@ -40,9 +40,12 @@ func TestMigrate0009PinnedColumnIdempotent(t *testing.T) {
 		t.Fatal("chat_sessions.pinned should still exist after idempotent re-run")
 	}
 
-	// Down one step rolls back only 0009, dropping the column.
+	// Roll back every migration newer than 0008 (0009 itself plus any migration added
-	if err := MigrateDown(ctx, conn, 1, testLogger()); err != nil {
+	// after it, currently 0010), so this test keeps validating the 0009 down no matter
-		t.Fatalf("MigrateDown(1): %v", err)
+	// how many later migrations the repo has grown.
+	steps := len(upMigrationIDs(t)) - 8
+	if err := MigrateDown(ctx, conn, steps, testLogger()); err != nil {
+		t.Fatalf("MigrateDown(%d): %v", steps, err)
 	}
 	if columnExists(t, ctx, conn, "chat_sessions", "pinned") {
 		t.Fatal("chat_sessions.pinned should be dropped by the 0009 down")

internal/db/migrations/0010_chat_session_read_at.down.sql

@@ -0,0 +1,9 @@
+-- 0010_chat_session_read_at down: drop the read cursor added by the up migration.
+-- SQLite DROP COLUMN (3.35+) is already relied on by 0006/0007/0009, and read_at is
+-- not referenced by any index, view or constraint, so a plain drop is safe. The
+-- rolled-back state is the pre-D4 one: no unread badge data at all (the cursor is
+-- derived display state, not data a session needs to stay valid). The backfilled
+-- values cannot be restored by a re-up beyond the newest-message rule, which is
+-- exactly the upgrade default the up migration applies.
+
+ALTER TABLE chat_sessions DROP COLUMN read_at;

internal/db/migrations/0010_chat_session_read_at.up.sql

@@ -0,0 +1,34 @@
+-- 0010_chat_session_read_at: per-session read cursor backing the session-list
+-- unread badge (DevDocs/design/16 §2.2/§4.2, Android 协同批次 D4/B3). There is a
+-- single owner account and no per-device identity at this layer, so the cursor is
+-- session-scoped: reading on any one end marks the session read for every end of
+-- the account (one cursor, no per-user/per-device cursor table).
+--
+-- read_at is a Unix MILLISECOND timestamp in the same unit as
+-- chat_messages.created_at (both are INTEGER ms columns; the store assigns
+-- time.Now().UnixMilli()), so the unread count is a plain
+-- `created_at > read_at` comparison. 0 means "nothing read yet".
+--
+-- Backfill on upgrade: every existing session's cursor is set to its newest
+-- message timestamp, so an upgrade does not suddenly light up the whole session
+-- list as unread (a session with no messages — and a session whose messages do not
+-- exist — stays at 0, which is also 0 unread).
+--
+-- The 0003 chat_sessions table already exists at this point, so ALTER ADD COLUMN is
+-- the correct, non-destructive step. The canonical schema (including this column) is
+-- also re-issued idempotently by internal/runtime/history.NewStore so a fresh
+-- in-memory store used by tests stays self-contained; the migration is the
+-- authoritative path for an existing database.
+--
+-- No new index is needed: the unread count filters
+-- `session_id = ? AND created_at > ?`, which the existing
+-- idx_chat_msgs_session_created (session_id, created_at DESC) already covers as an
+-- equality + range scan.
+
+ALTER TABLE chat_sessions ADD COLUMN read_at INTEGER NOT NULL DEFAULT 0;
+
+UPDATE chat_sessions
+SET read_at = COALESCE(
+    (SELECT MAX(m.created_at) FROM chat_messages m WHERE m.session_id = chat_sessions.id),
+    0
+);

internal/runtime/history/session_preview.go

@@ -7,11 +7,12 @@ import (
 	"strings"
 )
 
-// This file owns the last-message preview carried in the session list
+// This file owns the two derived per-session display fields carried in the session
-// (GET /api/v1/sessions): a display-only summary of each session's newest message so a
+// list (GET /api/v1/sessions): the last-message preview and the unread count. Both are
-// client can render a conversation row without one request per session. The preview is
+// read-only projections so a client can render a conversation row (with its unread
-// truncated for display and never feeds identity or permission decisions — identity
+// badge) without one request per session. The preview is truncated for display, and
-// comes from channel actor metadata / JWT (project red line).
+// neither field ever feeds identity or permission decisions — identity comes from
+// channel actor metadata / JWT (project red line).
 
 // lastMessagePreviewRunes bounds the preview content so a long message cannot bloat the
 // session-list response. A longer message is whitespace-collapsed and truncated to this
@@ -25,14 +26,19 @@ type MessageSummary struct {
 	CreatedAt int64  // Unix milliseconds
 }
 
-// listSessionsQuery reads every session plus its newest message in one statement. The
+// listSessionsQuery reads every session plus its newest message and its unread count
-// LEFT JOIN's correlated subquery resolves each session's newest row via the
+// in one statement. The LEFT JOIN's correlated subquery resolves each session's newest
-// (session_id, created_at DESC) index (rowid DESC breaks a same-millisecond tie by
+// row via the (session_id, created_at DESC) index (rowid DESC breaks a same-millisecond
-// insertion order), so a session with no messages still appears with NULL preview
+// tie by insertion order), so a session with no messages still appears with NULL preview
-// columns. Ordering floats pinned sessions first, then keeps the main session before
+// columns. The unread count is a correlated COUNT(*) over the same index
-// unpinned subs, then subs oldest-first.
+// (session_id = s.id AND created_at > s.read_at is an equality + range scan), so no
+// per-session round trip (N+1) is needed and no extra index is required. Ordering floats
+// pinned sessions first, then keeps the main session before unpinned subs, then subs
+// oldest-first.
 const listSessionsQuery = `SELECT s.id, s.type, s.title, s.pinned, s.created_at,
-		m.role, m.content, m.created_at
+		m.role, m.content, m.created_at,
+		(SELECT COUNT(*) FROM chat_messages c
+			WHERE c.session_id = s.id AND c.created_at > s.read_at)
 	FROM chat_sessions s
 	LEFT JOIN chat_messages m ON m.rowid = (
 		SELECT rowid FROM chat_messages
@@ -44,8 +50,9 @@ const listSessionsQuery = `SELECT s.id, s.type, s.title, s.pinned, s.created_at,
 
 // ListSessions returns pinned sessions first, then the main session, then every sub
 // session, oldest first, each carrying a LastMessage preview of its newest message (nil
-// when the session has none). Sessions and previews are read in a single query, so the
+// when the session has none) and the UnreadCount derived from the session's read cursor.
-// query count does not grow with the number of sessions. The returned slice is non-nil.
+// Sessions, previews and unread counts are read in a single query, so the query count
+// does not grow with the number of sessions. The returned slice is non-nil.
 func (s *Store) ListSessions(ctx context.Context) ([]SessionMeta, error) {
 	rows, err := s.db.QueryContext(ctx, listSessionsQuery, SessionMain)
 	if err != nil {
@@ -67,9 +74,10 @@ func (s *Store) ListSessions(ctx context.Context) ([]SessionMeta, error) {
 	return sessions, nil
 }
 
-// scanSessionMeta reads one joined chat_sessions + last-message row. The three message
+// scanSessionMeta reads one joined chat_sessions + last-message + unread-count row. The
-// columns are NULL for a session that has never received a message, so they scan into
+// three message columns are NULL for a session that has never received a message, so they
-// nullable holders and only produce a LastMessage when a row is present.
+// scan into nullable holders and only produce a LastMessage when a row is present. The
+// unread count is a non-NULL COUNT(*), so it scans into a plain int.
 func scanSessionMeta(sc rowScanner) (SessionMeta, error) {
 	var (
 		meta    SessionMeta
@@ -79,7 +87,7 @@ func scanSessionMeta(sc rowScanner) (SessionMeta, error) {
 		content sql.NullString
 		lastAt  sql.NullInt64
 	)
-	if err := sc.Scan(&meta.ID, &typeStr, &meta.Title, &pinned, &meta.CreatedAt, &role, &content, &lastAt); err != nil {
+	if err := sc.Scan(&meta.ID, &typeStr, &meta.Title, &pinned, &meta.CreatedAt, &role, &content, &lastAt, &meta.UnreadCount); err != nil {
 		return SessionMeta{}, err
 	}
 	meta.Type = typeStr

internal/runtime/history/session_read.go

@@ -0,0 +1,52 @@
+package history
+
+import (
+	"context"
+	"fmt"
+	"time"
+)
+
+// This file owns the per-session read cursor (design/16 §4.2, Android 协同批次
+// D4/B3). There is a single owner account and no per-device identity at this
+// layer, so the cursor is session-scoped: reading on any one end marks the session
+// read for every end of the account. The cursor lives on chat_sessions.read_at in
+// the same millisecond unit as chat_messages.created_at, so the unread badge is a
+// plain `created_at > read_at` count (see listSessionsQuery).
+
+// MarkSessionRead advances the session's read cursor to before and returns the
+// resulting cursor as a Unix millisecond timestamp. A zero before means "now" — the
+// caller's "mark everything I have seen" case, which does not have to read the
+// clock itself.
+//
+// The cursor only ever moves forward: the UPDATE applies SQLite's two-argument
+// scalar MAX, so a stale or out-of-order request (two ends racing, a retried
+// request) can never un-read messages. A negative before therefore leaves the stored
+// value untouched rather than moving it backwards.
+//
+// An unknown session id returns ErrSessionNotFound, matching RenameSession/SetPinned
+// — isolated (group/stranger) sessions have no chat_sessions row and are not
+// addressable here.
+func (s *Store) MarkSessionRead(ctx context.Context, sessionID string, before time.Time) (int64, error) {
+	ms := before.UnixMilli()
+	if before.IsZero() {
+		ms = time.Now().UnixMilli()
+	}
+	res, err := s.db.ExecContext(ctx,
+		`UPDATE chat_sessions SET read_at = MAX(read_at, ?) WHERE id = ?`, ms, sessionID)
+	if err != nil {
+		return 0, fmt.Errorf("history: mark session read %s: %w", sessionID, err)
+	}
+	n, err := res.RowsAffected()
+	if err != nil {
+		return 0, fmt.Errorf("history: mark session read %s rows: %w", sessionID, err)
+	}
+	if n == 0 {
+		return 0, fmt.Errorf("history: mark session read %s: %w", sessionID, ErrSessionNotFound)
+	}
+	var readAt int64
+	if err := s.db.QueryRowContext(ctx,
+		`SELECT read_at FROM chat_sessions WHERE id = ?`, sessionID).Scan(&readAt); err != nil {
+		return 0, fmt.Errorf("history: read session cursor %s: %w", sessionID, err)
+	}
+	return readAt, nil
+}

internal/runtime/history/session_read_test.go

@@ -0,0 +1,351 @@
+package history
+
+import (
+	"context"
+	"database/sql"
+	"errors"
+	"fmt"
+	"os"
+	"path/filepath"
+	"testing"
+	"time"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/db"
+)
+
+// These tests cover the D4 per-session read cursor (design/16 §4.2, Android
+// collaboration B3): the unread count derived in ListSessions, the forward-only
+// cursor semantics of MarkSessionRead, and the 0010 migration's upgrade backfill.
+
+// TestListSessionsUnreadCountFromCursor pins the unread derivation: a message counts as
+// unread only when it is strictly newer than the session's cursor (created_at > read_at),
+// so the cursor is inclusive of the timestamp it names (a client that has rendered a
+// message sent at T marks T read). A fresh session starts at cursor 0 with everything
+// unread, and a cursor at the newest message clears the badge.
+func TestListSessionsUnreadCountFromCursor(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+	for _, ts := range []int64{1000, 2000, 3000} {
+		if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "m", CreatedAt: ts}); err != nil {
+			t.Fatalf("seed %d: %v", ts, err)
+		}
+	}
+	// Never read: all three messages are unread (a brand-new account/session).
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 3 {
+		t.Fatalf("initial unreadCount = %d, want 3", got)
+	}
+	// Cursor on a message's own timestamp counts that message as read.
+	readAt, err := st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(2000))
+	if err != nil {
+		t.Fatalf("mark read: %v", err)
+	}
+	if readAt != 2000 {
+		t.Fatalf("readAt = %d, want 2000", readAt)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 1 {
+		t.Fatalf("unreadCount after cursor 2000 = %d, want 1 (only 3000 newer)", got)
+	}
+	// A cursor between two messages reads the earlier one only.
+	if _, err := st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(2500)); err != nil {
+		t.Fatalf("mark read 2500: %v", err)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 1 {
+		t.Fatalf("unreadCount after cursor 2500 = %d, want 1", got)
+	}
+	// A cursor at the newest message clears the badge.
+	if _, err := st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(3000)); err != nil {
+		t.Fatalf("mark read 3000: %v", err)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 0 {
+		t.Fatalf("unreadCount after cursor 3000 = %d, want 0", got)
+	}
+}
+
+// TestListSessionsUnreadCountSessionIsolation pins the per-session scope: each session
+// carries its own cursor, so reading the main session must not clear (or inflate) a sub
+// session's badge, and vice versa.
+func TestListSessionsUnreadCountSessionIsolation(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+	sub, err := st.NewSubSession(ctx, "owner")
+	if err != nil {
+		t.Fatalf("new sub: %v", err)
+	}
+	for _, m := range []Message{
+		{SessionID: MainSessionID, Role: RoleUser, Content: "main-1", CreatedAt: 1000},
+		{SessionID: MainSessionID, Role: RoleAssistant, Content: "main-2", CreatedAt: 2000},
+		{SessionID: sub, Role: RoleUser, Content: "sub-1", CreatedAt: 1500},
+	} {
+		if err := st.Append(ctx, m); err != nil {
+			t.Fatalf("seed %q: %v", m.Content, err)
+		}
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 2 {
+		t.Fatalf("main unreadCount = %d, want 2", got)
+	}
+	if got := sessionOf(t, st, sub).UnreadCount; got != 1 {
+		t.Fatalf("sub unreadCount = %d, want 1", got)
+	}
+
+	// Reading the main session must leave the sub untouched.
+	if _, err := st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(2000)); err != nil {
+		t.Fatalf("mark main read: %v", err)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 0 {
+		t.Fatalf("main unreadCount after read = %d, want 0", got)
+	}
+	if got := sessionOf(t, st, sub).UnreadCount; got != 1 {
+		t.Fatalf("sub unreadCount after main read = %d, want 1 (cursors are per-session)", got)
+	}
+
+	// A message arriving after the read re-opens the main badge.
+	if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "main-3", CreatedAt: 3000}); err != nil {
+		t.Fatalf("append main-3: %v", err)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 1 {
+		t.Fatalf("main unreadCount after new message = %d, want 1", got)
+	}
+}
+
+// TestMarkSessionReadForwardOnly covers the racing/malformed-request side of the cursor
+// contract: the cursor never moves backwards, so a stale or repeated request is
+// idempotent and can never resurrect already-read messages.
+func TestMarkSessionReadForwardOnly(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+	for _, ts := range []int64{1000, 2000, 3000} {
+		if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "m", CreatedAt: ts}); err != nil {
+			t.Fatalf("seed %d: %v", ts, err)
+		}
+	}
+	// Advance to the newest message.
+	readAt, err := st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(3000))
+	if err != nil || readAt != 3000 {
+		t.Fatalf("mark 3000 = (%d, %v), want (3000, nil)", readAt, err)
+	}
+	// A stale request (an end that was behind, or a retry of an older read) must not
+	// move the cursor back and re-open the badge.
+	readAt, err = st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(1000))
+	if err != nil {
+		t.Fatalf("stale mark: %v", err)
+	}
+	if readAt != 3000 {
+		t.Fatalf("stale mark readAt = %d, want the cursor to stay at 3000", readAt)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 0 {
+		t.Fatalf("unreadCount after stale mark = %d, want 0 (never un-read)", got)
+	}
+	// Repeating the same cursor is a no-op, and a negative cursor is clamped by the
+	// same MAX (it cannot move the stored value at all).
+	if readAt, err = st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(3000)); err != nil || readAt != 3000 {
+		t.Fatalf("repeat mark = (%d, %v), want (3000, nil)", readAt, err)
+	}
+	if readAt, err = st.MarkSessionRead(ctx, MainSessionID, time.UnixMilli(-5000)); err != nil || readAt != 3000 {
+		t.Fatalf("negative mark = (%d, %v), want (3000, nil)", readAt, err)
+	}
+}
+
+// TestMarkSessionReadZeroTimeMeansNow covers the "mark everything I have seen" call the
+// HTTP layer makes when the body omits before: the zero time resolves to the wall clock,
+// which reads every message that already exists.
+func TestMarkSessionReadZeroTimeMeansNow(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	ctx := context.Background()
+	seedAt := time.Now().UnixMilli()
+	if err := st.Append(ctx, Message{SessionID: MainSessionID, Role: RoleUser, Content: "m", CreatedAt: seedAt}); err != nil {
+		t.Fatalf("seed: %v", err)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 1 {
+		t.Fatalf("unreadCount = %d, want 1", got)
+	}
+	readAt, err := st.MarkSessionRead(ctx, MainSessionID, time.Time{})
+	if err != nil {
+		t.Fatalf("mark zero time: %v", err)
+	}
+	if readAt < seedAt {
+		t.Fatalf("readAt = %d, want >= the seeded message time %d", readAt, seedAt)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 0 {
+		t.Fatalf("unreadCount after mark-now = %d, want 0", got)
+	}
+}
+
+// TestMarkSessionReadUnknownSession pins the error contract: an id with no chat_sessions
+// row (including an isolated group/stranger session) is ErrSessionNotFound, so the HTTP
+// layer can answer 404 instead of silently reporting success for a session that does not
+// exist.
+func TestMarkSessionReadUnknownSession(t *testing.T) {
+	t.Parallel()
+	st := newStore(t)
+	if _, err := st.MarkSessionRead(context.Background(), "web:missing", time.UnixMilli(1000)); !errors.Is(err, ErrSessionNotFound) {
+		t.Fatalf("unknown session err = %v, want ErrSessionNotFound", err)
+	}
+}
+
+// migrationReadAtSQL reads one of the 0010 migration files from the source tree, so the
+// test applies the same SQL text the embedded migration runs rather than a copy.
+func migrationReadAtSQL(t *testing.T, name string) string {
+	t.Helper()
+	b, err := os.ReadFile(filepath.Join("..", "..", "db", "migrations", name))
+	if err != nil {
+		t.Fatalf("read migration %s: %v", name, err)
+	}
+	return string(b)
+}
+
+// columnPresent reports whether table has a column named column.
+func columnPresent(t *testing.T, conn *sql.DB, table, column string) bool {
+	t.Helper()
+	var n int
+	if err := conn.QueryRowContext(context.Background(),
+		"SELECT count(*) FROM pragma_table_info(?) WHERE name = ?", table, column).Scan(&n); err != nil {
+		t.Fatalf("pragma_table_info(%s): %v", table, err)
+	}
+	return n == 1
+}
+
+// openLegacyDB opens an in-memory DB carrying the pre-0010 schema (chat_sessions with no
+// read_at column) plus two legacy sessions with prior history — the real fixed main
+// session "web" and a sub session — so the 0010 migration can be applied to a database
+// that really predates this feature.
+func openLegacyDB(t *testing.T) *sql.DB {
+	t.Helper()
+	ctx := context.Background()
+	conn, err := db.Open(ctx, ":memory:", testLogger())
+	if err != nil {
+		t.Fatalf("open memory db: %v", err)
+	}
+	t.Cleanup(func() { _ = conn.Close() })
+	ddls := []string{
+		`CREATE TABLE chat_sessions (
+			id         TEXT PRIMARY KEY,
+			type       TEXT NOT NULL,
+			title      TEXT NOT NULL DEFAULT '',
+			pinned     INTEGER NOT NULL DEFAULT 0,
+			created_at INTEGER NOT NULL)`,
+		`CREATE TABLE chat_messages (
+			id          TEXT PRIMARY KEY,
+			session_id  TEXT NOT NULL,
+			role        TEXT NOT NULL,
+			content     TEXT NOT NULL,
+			mode        TEXT,
+			mood        TEXT,
+			image       TEXT,
+			thinking    TEXT,
+			platform    TEXT NOT NULL DEFAULT 'web',
+			device_id   TEXT NOT NULL DEFAULT '',
+			sender_name TEXT NOT NULL DEFAULT '',
+			adapter_id  TEXT NOT NULL DEFAULT '',
+			created_at  INTEGER NOT NULL)`,
+		`CREATE INDEX idx_chat_msgs_session_created ON chat_messages (session_id, created_at DESC)`,
+	}
+	for _, ddl := range ddls {
+		if _, err := conn.ExecContext(ctx, ddl); err != nil {
+			t.Fatalf("legacy ddl: %v", err)
+		}
+	}
+	legacy := []struct {
+		sessionID   string
+		sessionType string
+		messageAt   []int64
+	}{
+		{sessionID: MainSessionID, sessionType: SessionMain, messageAt: []int64{1000, 4000}},
+		{sessionID: "web:legacy", sessionType: SessionSub, messageAt: []int64{1000, 2000}},
+	}
+	for _, l := range legacy {
+		if _, err := conn.ExecContext(ctx,
+			`INSERT INTO chat_sessions (id, type, title, pinned, created_at) VALUES (?, ?, '', 0, ?)`,
+			l.sessionID, l.sessionType, int64(100)); err != nil {
+			t.Fatalf("insert legacy session %s: %v", l.sessionID, err)
+		}
+		for i, ts := range l.messageAt {
+			if _, err := conn.ExecContext(ctx,
+				`INSERT INTO chat_messages (id, session_id, role, content, created_at) VALUES (?, ?, ?, ?, ?)`,
+				fmt.Sprintf("%s-%d", l.sessionID, i), l.sessionID, RoleUser, "legacy message", ts); err != nil {
+				t.Fatalf("insert legacy message: %v", err)
+			}
+		}
+	}
+	return conn
+}
+
+// TestMigration0010ReadAtBackfill exercises the real 0010 migration against a legacy
+// database: up adds the column and backfills each session's cursor to its newest message,
+// so upgrade day reports 0 unread for existing history — including the fixed main session,
+// whose existing history is the whole point of the upgrade — and the store (which
+// re-issues the canonical DDL idempotently) then works on top of it; a message that arrives
+// after the migration is unread until it is marked read; down drops the column again.
+func TestMigration0010ReadAtBackfill(t *testing.T) {
+	ctx := context.Background()
+	conn := openLegacyDB(t)
+	if columnPresent(t, conn, "chat_sessions", "read_at") {
+		t.Fatal("legacy schema should not have read_at before the migration")
+	}
+
+	if _, err := conn.ExecContext(ctx, migrationReadAtSQL(t, "0010_chat_session_read_at.up.sql")); err != nil {
+		t.Fatalf("apply 0010 up: %v", err)
+	}
+	if !columnPresent(t, conn, "chat_sessions", "read_at") {
+		t.Fatal("read_at should exist after the 0010 up migration")
+	}
+	// The backfill is each session's newest message timestamp, not a blanket 0 (which
+	// would mark the whole existing history unread) and not the first message's.
+	for _, tc := range []struct {
+		sessionID string
+		want      int64
+	}{
+		{sessionID: MainSessionID, want: 4000},
+		{sessionID: "web:legacy", want: 2000},
+	} {
+		var backfilled int64
+		if err := conn.QueryRowContext(ctx,
+			`SELECT read_at FROM chat_sessions WHERE id = ?`, tc.sessionID).Scan(&backfilled); err != nil {
+			t.Fatalf("read backfilled cursor %s: %v", tc.sessionID, err)
+		}
+		if backfilled != tc.want {
+			t.Fatalf("backfilled read_at for %s = %d, want its newest legacy message %d", tc.sessionID, backfilled, tc.want)
+		}
+	}
+
+	// The store comes up on the migrated schema and reports no unread for either
+	// pre-existing session (upgrade day must not light up every row, main session
+	// included); the store's INSERT OR IGNORE for the main session must not overwrite
+	// the backfilled cursor.
+	st, err := NewStore(ctx, conn, testLogger())
+	if err != nil {
+		t.Fatalf("new store on migrated db: %v", err)
+	}
+	for _, sessionID := range []string{MainSessionID, "web:legacy"} {
+		if got := sessionOf(t, st, sessionID).UnreadCount; got != 0 {
+			t.Fatalf("legacy session %s unreadCount = %d, want 0 after backfill", sessionID, got)
+		}
+	}
+	// A message written after the migration is unread, and marking it read clears it.
+	if err := st.Append(ctx, Message{SessionID: "web:legacy", Role: RoleAssistant, Content: "new", CreatedAt: 9000}); err != nil {
+		t.Fatalf("append after migration: %v", err)
+	}
+	if got := sessionOf(t, st, "web:legacy").UnreadCount; got != 1 {
+		t.Fatalf("unreadCount after post-migration message = %d, want 1", got)
+	}
+	if got := sessionOf(t, st, MainSessionID).UnreadCount; got != 0 {
+		t.Fatalf("main unreadCount = %d, want 0 (untouched by the sub session's new message)", got)
+	}
+	readAt, err := st.MarkSessionRead(ctx, "web:legacy", time.UnixMilli(9000))
+	if err != nil || readAt != 9000 {
+		t.Fatalf("mark read = (%d, %v), want (9000, nil)", readAt, err)
+	}
+	if got := sessionOf(t, st, "web:legacy").UnreadCount; got != 0 {
+		t.Fatalf("unreadCount after mark = %d, want 0", got)
+	}
+
+	if _, err := conn.ExecContext(ctx, migrationReadAtSQL(t, "0010_chat_session_read_at.down.sql")); err != nil {
+		t.Fatalf("apply 0010 down: %v", err)
+	}
+	if columnPresent(t, conn, "chat_sessions", "read_at") {
+		t.Fatal("read_at should be dropped by the 0010 down migration")
+	}
+}

internal/runtime/history/store.go

@@ -121,10 +121,20 @@ type SessionMeta struct {
 	// A client uses it to render the conversation row without a per-session request;
 	// identity and permissions never derive from it.
 	LastMessage *MessageSummary
+	// UnreadCount is the number of messages newer than the session's read cursor
+	// (design/16 §4.2, D4), computed by ListSessions in the same query. The cursor
+	// is session-scoped and single-owner: reading on any one end marks the session
+	// read for every end of the account. Like LastMessage it is display-only —
+	// identity and permissions never derive from it.
+	UnreadCount int
 }
 
 // HistoryStore abstracts the chat-history store so callers (the gateway) and tests can
-// hold an interface rather than a concrete *Store.
+// hold an interface rather than a concrete *Store. It is deliberately kept to the calls
+// the runtime-level consumers (channelbridge, backthink) actually make: the D4 read
+// cursor (MarkSessionRead) is declared on the consumer-side interface in
+// internal/services/gateway (contracts.go) instead, so a narrower store fake elsewhere
+// does not have to implement a cursor it never advances. *Store satisfies both.
 type HistoryStore interface {
 	// Append persists msg, assigning an id and timestamp when the caller omitted them.
 	Append(ctx context.Context, msg Message) error
@@ -579,6 +589,7 @@ CREATE TABLE IF NOT EXISTS chat_sessions (
     type       TEXT NOT NULL,
     title      TEXT NOT NULL DEFAULT '',
     pinned     INTEGER NOT NULL DEFAULT 0,
+    read_at    INTEGER NOT NULL DEFAULT 0,
     created_at INTEGER NOT NULL
 )`
 

internal/services/gateway/contracts.go

@@ -116,6 +116,11 @@ type HistoryStore interface {
 	// history.ErrSessionNotFound.
 	SetPinned(ctx context.Context, sessionID string, pinned bool) error
 	NewSubSession(ctx context.Context, ownerID string) (string, error)
+	// MarkSessionRead advances a session's read cursor (D4) to before and returns
+	// the resulting cursor as a Unix millisecond timestamp; a zero before means
+	// "now". The cursor only moves forward. An unknown id returns
+	// history.ErrSessionNotFound.
+	MarkSessionRead(ctx context.Context, sessionID string, before time.Time) (int64, error)
 	ListSessions(ctx context.Context) ([]history.SessionMeta, error)
 }
 

internal/services/gateway/history.go

@@ -26,6 +26,12 @@ type historySessionSummary struct {
 	// per-session request. It is omitted for a session with no messages. Content is
 	// already truncated by the store.
 	LastMessage *historyLastMessage `json:"lastMessage,omitempty"`
+	// UnreadCount is the number of messages newer than the session's read cursor
+	// (D4), derived in the same list query. It is always returned (0 when nothing
+	// is unread) so a client can render/clear the badge without a second call. The
+	// cursor is session-scoped and single-owner: reading on any one end clears the
+	// badge for every end of the account.
+	UnreadCount int `json:"unreadCount"`
 }
 
 // historyLastMessage is the wire shape of the last-message preview in a session summary.
@@ -79,6 +85,7 @@ func (s *Server) handleHistorySessions(w http.ResponseWriter, r *http.Request) {
 			CreatedAt:   m.CreatedAt,
 			Pinned:      m.Pinned,
 			LastMessage: toHistoryLastMessage(m.LastMessage),
+			UnreadCount: m.UnreadCount,
 		})
 	}
 	writeJSON(w, http.StatusOK, out)

internal/services/gateway/routes.go

@@ -83,6 +83,10 @@ func (s *Server) registerHistoryRoutes(mux *http.ServeMux, writeAuth func(http.H
 	mux.Handle("POST /api/v1/sessions", writeAuth(http.HandlerFunc(s.handleCreateSubSession)))
 	mux.Handle("DELETE /api/v1/sessions/{id}", writeAuth(http.HandlerFunc(s.handleDeleteSession)))
 	mux.Handle("PATCH /api/v1/sessions/{id}", writeAuth(http.HandlerFunc(s.handlePatchSession)))
+	// Read cursor (D4 / Android collaboration B3): advancing it clears the session's
+	// unread badge on every end of the account, so it is an owner/admin history write
+	// like the other session mutations. It broadcasts a `session_read` frame.
+	mux.Handle("POST /api/v1/sessions/{id}/read", writeAuth(http.HandlerFunc(s.handleMarkSessionRead)))
 	// Single-message edit/delete (Android collaboration batch ③a, D3). A message is
 	// addressed by its own opaque id; both are owner/admin-gated history writes.
 	mux.Handle("PATCH /api/v1/messages/{id}", writeAuth(http.HandlerFunc(s.handlePatchMessage)))

internal/services/gateway/session_orchestrator_test.go

@@ -231,6 +231,13 @@ func (f *fakeHistory) ListSessions(context.Context) ([]history.SessionMeta, erro
 	return nil, nil
 }
 
+// MarkSessionRead satisfies the HistoryStore contract for the read-cursor path (D4).
+// The orchestrator never advances a cursor, so the fake only reports the requested
+// cursor back without storing it.
+func (f *fakeHistory) MarkSessionRead(_ context.Context, _ string, before time.Time) (int64, error) {
+	return before.UnixMilli(), nil
+}
+
 func (f *fakeHistory) records() []history.Message {
 	f.mu.Lock()
 	defer f.mu.Unlock()

internal/services/gateway/session_read.go

@@ -0,0 +1,97 @@
+package gateway
+
+import (
+	"encoding/json"
+	"errors"
+	"io"
+	"net/http"
+	"time"
+
+	apperr "git.yeij.top/AskaEth/Cyrene_v2/internal/errors"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
+)
+
+// This file owns the D4 read-cursor surface (Android collaboration B3): the
+// POST /api/v1/sessions/{id}/read endpoint that advances a session's read cursor,
+// plus the `session_read` realtime frame it broadcasts so the account's other
+// browsers clear the unread badge without a re-read. The cursor itself lives in
+// chat_sessions.read_at (internal/runtime/history); GET /sessions derives the
+// unreadCount from it (history.go). The cursor is session-scoped and single-owner:
+// one account, one cursor per session — reading on any one end marks the session
+// read for every end.
+
+// wsSessionRead is the WS event type for a read-cursor advance.
+const wsSessionRead = "session_read"
+
+// sessionReadRequest is the POST /api/v1/sessions/{id}/read body. Before is a
+// pointer so an absent field (an empty body) is distinguishable from an explicit
+// value: absent means "now" — the client marks everything it has seen read.
+type sessionReadRequest struct {
+	Before *int64 `json:"before"`
+}
+
+// handleMarkSessionRead advances a session's read cursor and answers
+// {data:{sessionId, readAt}} with the resulting cursor in Unix milliseconds.
+// It is owner/admin-gated like the other history writes (writeAuth).
+//
+// body {before?} is an optional millisecond timestamp bounding what is marked
+// read — messages with created_at <= readAt become read. Absent (or an empty body)
+// means now. A present non-positive or non-numeric before answers 400: the cursor
+// is a real millisecond timestamp, and "unread by a negative cursor" is not a
+// meaningful client request. An unknown session id answers 404; a missing
+// HistoryStore answers 501.
+//
+// On success the new cursor is broadcast as a `session_read` frame so the account's
+// other browsers clear the badge too (reading on one end = read everywhere). A
+// failure path broadcasts nothing.
+func (s *Server) handleMarkSessionRead(w http.ResponseWriter, r *http.Request) {
+	if s.history == nil {
+		writeError(w, &apperr.AppError{Code: "not_implemented", Message: "会话历史服务未配置（未注入 HistoryStore）", HTTPCode: http.StatusNotImplemented})
+		return
+	}
+	id := r.PathValue("id")
+	if id == "" {
+		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "session id is required"))
+		return
+	}
+	var req sessionReadRequest
+	if err := json.NewDecoder(r.Body).Decode(&req); err != nil && !errors.Is(err, io.EOF) {
+		writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "invalid request body"))
+		return
+	}
+	var before time.Time
+	if req.Before != nil {
+		if *req.Before <= 0 {
+			writeError(w, apperr.Wrap(apperr.ErrBadRequest, "bad_request", "before must be a positive millisecond timestamp"))
+			return
+		}
+		before = time.UnixMilli(*req.Before)
+	}
+	readAt, err := s.history.MarkSessionRead(r.Context(), id, before)
+	if err != nil {
+		if errors.Is(err, history.ErrSessionNotFound) {
+			writeError(w, apperr.Wrap(apperr.ErrNotFound, "not_found", "session %q not found", id))
+			return
+		}
+		s.log.Error("history mark session read", "err", err, "session", id)
+		writeError(w, apperr.Wrap(apperr.ErrInternal, "session_failed", "failed to mark session read"))
+		return
+	}
+	s.hub.broadcastSessionRead(id, readAt)
+	writeJSON(w, http.StatusOK, map[string]any{"sessionId": id, "readAt": readAt})
+}
+
+// broadcastSessionRead fans a read-cursor advance out to every client as a
+// `session_read` frame carrying the session's new cursor, so the same account's
+// other ends clear the session's unread badge without a re-read. Like the other
+// history frames it is an all-connections broadcast (the hub has no per-session
+// subscription), and the client filters it by sessionId; the cursor's single source
+// of truth stays chat_sessions.read_at, so a client that missed the frame converges
+// on its next GET /sessions.
+func (h *hub) broadcastSessionRead(sessionID string, readAt int64) {
+	h.broadcast(wsEvent{
+		Type:      wsSessionRead,
+		SessionID: sessionID,
+		Data:      map[string]any{"readAt": readAt},
+	})
+}

internal/services/gateway/session_read_test.go

@@ -0,0 +1,233 @@
+package gateway
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/auth"
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/history"
+)
+
+// These tests cover the D4 read-cursor contract (Android collaboration B3): the
+// unreadCount carried by GET /api/v1/sessions, the advance endpoint
+// POST /api/v1/sessions/{id}/read, and the `session_read` frame that clears the badge
+// on the account's other browsers.
+
+// unreadBySession decodes GET /api/v1/sessions into a map of session id → unreadCount via
+// the wire key, so the test asserts the JSON contract rather than the in-package struct.
+func unreadBySession(t *testing.T, ts *httptest.Server, tok string) map[string]int {
+	t.Helper()
+	resp, body := doReq(t, http.MethodGet, ts.URL+"/api/v1/sessions", nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("list sessions status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	var sessions []struct {
+		ID          string `json:"id"`
+		UnreadCount int    `json:"unreadCount"`
+	}
+	decodeDataSlice(t, body, &sessions)
+	out := make(map[string]int, len(sessions))
+	for _, s := range sessions {
+		out[s.ID] = s.UnreadCount
+	}
+	return out
+}
+
+// TestSessionsUnreadCountAndMarkRead covers the happy path end to end: a session with
+// unread messages reports the count, POST /read moves the cursor and answers
+// {sessionId, readAt}, the list drops to the messages sent after the cursor, and a
+// body-less POST (mark everything seen) clears the badge.
+func TestSessionsUnreadCountAndMarkRead(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, st := newHistoryEnv(t)
+	ctx := context.Background()
+	sub, err := st.NewSubSession(ctx, "owner")
+	if err != nil {
+		t.Fatalf("new sub: %v", err)
+	}
+	for _, m := range []history.Message{
+		{SessionID: history.MainSessionID, Role: history.RoleUser, Content: "a", CreatedAt: 1000},
+		{SessionID: history.MainSessionID, Role: history.RoleAssistant, Content: "b", CreatedAt: 2000},
+		{SessionID: history.MainSessionID, Role: history.RoleUser, Content: "c", CreatedAt: 3000},
+		{SessionID: sub, Role: history.RoleUser, Content: "sub", CreatedAt: 1500},
+	} {
+		if err := st.Append(ctx, m); err != nil {
+			t.Fatalf("seed %q: %v", m.Content, err)
+		}
+	}
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+	base := ts.URL + "/api/v1/sessions/" + history.MainSessionID + "/read"
+
+	// Nothing read yet: the main session is fully unread, the sub session independently so.
+	if got := unreadBySession(t, ts, tok); got[history.MainSessionID] != 3 || got[sub] != 1 {
+		t.Fatalf("initial unread = %+v, want main 3 / sub 1", got)
+	}
+
+	// Advance the main cursor to 2000: messages at or before it are read.
+	resp, body := doReq(t, http.MethodPost, base, map[string]any{"before": 2000}, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("mark read status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	env := decodeBody(t, body)
+	if len(env) != 1 {
+		t.Fatalf("envelope = %v, want only a data field", env)
+	}
+	var data struct {
+		SessionID string `json:"sessionId"`
+		ReadAt    int64  `json:"readAt"`
+	}
+	if err := json.Unmarshal(env["data"], &data); err != nil {
+		t.Fatalf("decode read data: %v (body=%s)", err, body)
+	}
+	if data.SessionID != history.MainSessionID || data.ReadAt != 2000 {
+		t.Fatalf("read data = %+v, want sessionId %s readAt 2000", data, history.MainSessionID)
+	}
+
+	// Only the later message stays unread, and the sub session is untouched.
+	if got := unreadBySession(t, ts, tok); got[history.MainSessionID] != 1 || got[sub] != 1 {
+		t.Fatalf("unread after cursor 2000 = %+v, want main 1 / sub 1", got)
+	}
+
+	// No body = "mark everything I have seen read" (resolved to now server-side).
+	resp, body = doReq(t, http.MethodPost, base, nil, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("body-less mark status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	if err := json.Unmarshal(decodeBody(t, body)["data"], &data); err != nil {
+		t.Fatalf("decode body-less read data: %v (body=%s)", err, body)
+	}
+	if data.ReadAt < 3000 {
+		t.Fatalf("body-less readAt = %d, want >= the newest message 3000", data.ReadAt)
+	}
+	if got := unreadBySession(t, ts, tok); got[history.MainSessionID] != 0 || got[sub] != 1 {
+		t.Fatalf("unread after mark-now = %+v, want main 0 / sub 1", got)
+	}
+	markedNow := data.ReadAt
+
+	// The cursor only moves forward: re-posting a stale cursor must not re-open the badge.
+	resp, body = doReq(t, http.MethodPost, base, map[string]any{"before": 1000}, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("stale mark status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	if err := json.Unmarshal(decodeBody(t, body)["data"], &data); err != nil {
+		t.Fatalf("decode stale read data: %v", err)
+	}
+	if data.ReadAt != markedNow {
+		t.Fatalf("stale readAt = %d, want the cursor to stay at %d", data.ReadAt, markedNow)
+	}
+	if got := unreadBySession(t, ts, tok); got[history.MainSessionID] != 0 {
+		t.Fatalf("unread after stale mark = %d, want 0", got[history.MainSessionID])
+	}
+}
+
+// TestMarkSessionReadValidation covers the failure contract: a present but unusable
+// `before` answers 400, an unknown session 404, a viewer 403 and a missing HistoryStore
+// 501 — none of which may be mistaken for a successful read.
+func TestMarkSessionReadValidation(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt, _ := newHistoryEnv(t)
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+	base := ts.URL + "/api/v1/sessions/web/read"
+
+	for _, tc := range []struct {
+		name string
+		body any
+	}{
+		{name: "zero", body: map[string]any{"before": 0}},
+		{name: "negative", body: map[string]any{"before": -1}},
+		{name: "non-numeric", body: map[string]any{"before": "abc"}},
+		{name: "not an object", body: "not-an-object"},
+	} {
+		t.Run(tc.name, func(t *testing.T) {
+			resp, body := doReq(t, http.MethodPost, base, tc.body, tok)
+			if resp.StatusCode != http.StatusBadRequest {
+				t.Fatalf("status = %d, want 400 (body=%s)", resp.StatusCode, body)
+			}
+			wantErrorShape(t, body)
+		})
+	}
+
+	// An empty object means "now", not "no-op".
+	resp, body := doReq(t, http.MethodPost, base, map[string]any{}, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("empty object status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+
+	// Unknown session → 404; the session id is not auto-created.
+	resp, body = doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions/web:missing/read", map[string]any{"before": 1000}, tok)
+	if resp.StatusCode != http.StatusNotFound {
+		t.Fatalf("unknown session status = %d, want 404 (body=%s)", resp.StatusCode, body)
+	}
+
+	// A viewer cannot write the cursor (owner/admin history write, like PATCH /sessions).
+	viewer := signAny(t, jwt, "acct", auth.RoleViewer)
+	resp, body = doReq(t, http.MethodPost, base, map[string]any{"before": 1000}, viewer)
+	if resp.StatusCode != http.StatusForbidden {
+		t.Fatalf("viewer status = %d, want 403 (body=%s)", resp.StatusCode, body)
+	}
+
+	// And no token at all is 401, not a silent success.
+	resp, body = doReq(t, http.MethodPost, base, map[string]any{"before": 1000}, "")
+	if resp.StatusCode != http.StatusUnauthorized {
+		t.Fatalf("anonymous status = %d, want 401 (body=%s)", resp.StatusCode, body)
+	}
+}
+
+// TestMarkSessionReadNotWired pins the honest degradation: with no HistoryStore injected
+// the endpoint answers 501 not_implemented rather than pretending the cursor moved.
+func TestMarkSessionReadNotWired(t *testing.T) {
+	t.Parallel()
+	_, ts, jwt := buildEnv(t, cfgDefault(), Deps{})
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	resp, body := doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions/web/read", map[string]any{"before": 1000}, tok)
+	if resp.StatusCode != http.StatusNotImplemented {
+		t.Fatalf("status = %d, want 501 (body=%s)", resp.StatusCode, body)
+	}
+	wantErrorShape(t, body)
+}
+
+// TestMarkSessionReadBroadcasts asserts the cross-end sync frame: a successful advance
+// pushes exactly one `session_read` frame scoped to that session and carrying the new
+// cursor, while a failed one (404) pushes nothing.
+func TestMarkSessionReadBroadcasts(t *testing.T) {
+	t.Parallel()
+	srv, ts, jwt, _ := newHistoryEnv(t)
+	c := newFakeConn(srv.hub, 32)
+	srv.hub.register(c)
+	tok := signAny(t, jwt, "acct", auth.RoleOwner)
+
+	resp, body := doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions/"+history.MainSessionID+"/read", map[string]any{"before": 2000}, tok)
+	if resp.StatusCode != http.StatusOK {
+		t.Fatalf("status = %d, want 200 (body=%s)", resp.StatusCode, body)
+	}
+	ev := readFakeFrame(t, c)
+	if ev.Type != "session_read" {
+		t.Fatalf("frame type = %q, want session_read (frame=%+v)", ev.Type, ev)
+	}
+	if ev.SessionID != history.MainSessionID {
+		t.Fatalf("frame sessionId = %q, want %q", ev.SessionID, history.MainSessionID)
+	}
+	data, ok := ev.Data.(map[string]any)
+	if !ok {
+		t.Fatalf("frame data = %+v, want object", ev.Data)
+	}
+	if data["readAt"] != float64(2000) {
+		t.Fatalf("frame data.readAt = %v, want 2000", data["readAt"])
+	}
+	if err := connHasNoFrame(c); err != nil {
+		t.Fatal(err)
+	}
+
+	// A failed advance must not tell other ends the badge was cleared.
+	resp, body = doReq(t, http.MethodPost, ts.URL+"/api/v1/sessions/web:missing/read", map[string]any{"before": 2000}, tok)
+	if resp.StatusCode != http.StatusNotFound {
+		t.Fatalf("unknown status = %d, want 404 (body=%s)", resp.StatusCode, body)
+	}
+	if err := connHasNoFrame(c); err != nil {
+		t.Fatal(err)
+	}
+}