dsh-connect-doubao 0.4.0 → 0.4.1

package/CHANGELOG.md

@@ -4,6 +4,49 @@
 
 ---
 
+## 0.4.1 — 2026-11-05
+
+> 「工作任务」模型接入批次：把豆包客户端里那三个**只能在工作任务模式跑**的条目
+> （2.1 Lite `seed-lite-7b`、2.1 Pro `5`、自动 `9`）真正接到 DSH 里，走的正是会员/
+> 活动付费额度。上一版把它们整个藏起来，理由是「外部客户端注册不进 CUA 运行时」——
+> 那个理由对了一半：运行时确实接不进去，但它不是必需的，`tool_local` 的三条 HTTP
+> 接口就够把一轮任务回合走完。新增 4 例回归测试（另改写 2 例旧的「必须隐藏」断言），
+> 全仓 189 例。
+
+### 新增
+
+- **豆包 2.1 Lite / 2.1 Pro / 自动 进入模型选择器** —— 选中任一条，请求自动切到
+  「工作任务」模式（`mode_id: "3"`、`agent_mode: 1`、`need_deep_think: 10001`、
+  新会话 `conversation_mode: 2`、`model_extra_params.input_token_limit: "256k"`，
+  `seed-lite-7b` 另带 `reasoning_effort`），`general_task_param` 带 `task_input_json`。
+  实机跑通：Lite 与 Pro 各一次纯文本回合，约 5.8 s，`server_model_tag` 为 `38`。
+- **`tool_local` 应答器** —— 任务模式下服务端会按 `${unique_key}:${device_id}:${aid}:message:${seq}`
+  的顺序下发本地工作区指令；插件并发轮询 `msg/ack`，遇到 `command: 6003`（`cmd_call`，
+  如 `local.workspace.apply`）就回 `upload_command_result` + `msg/complete`，遇到
+  `command: 6004`（`agent_finished`）收尾。**这是解开 `710010702` 的全部原因。**
+- **`CHUNK_DELTA` 增量帧** —— 任务模式的正文增量以无 `block_type` 包裹的
+  `{ text, id }` 形式下发，按 `id` 分流到 `reasoning_content` / 正文。
+
+### 修复
+
+- **选 `seed-lite-7b` / `5` 会被静默换成「豆包 快速」** —— `findDoubaoModel` 只在内置
+  清单里查，查不到就回落到 `'0'`。现在三条任务模型都在清单里，不再回落。
+- **`msg/ack` 批量查询必失败** —— `{ msg_ids: [0..6] }` 会整体报
+  `{"code":710010702}`，因为批里只要含一个服务端还没生成的 `seq`，整批就废。
+  现在严格单条轮询，逐 `seq` 递进。
+- **自动发现把工作任务模型全部隐藏** —— 上一版按名称前缀一律排除。现在改为只排除
+  `DOUBAO_UNAVAILABLE_MODELS`（仅剩「即将下线」的 `4`），并按内置清单补齐，避免本机
+  IndexedDB 只缓存了对话子集时任务模型消失。
+
+### 已知边界
+
+- 纯文本、不触发本地工具的小回合，服务端会在 `ext` 里标 `general_agent_task_mode: "simple"`
+  与 `mode_downgrade_reason: "success"`，此时**不扣**活动赠送额度（实测调用前后
+  `usedAmount` 恒为 782）。要消耗付费额度，得让它真的执行多步工具/文档任务。
+- `msg/ack` 的 `seq` 没有可查询的上界，应答器在流结束后最多再补 8 次轮询即退出。
+
+---
+
 ## 0.4.0 — 2026-11-05
 
 > 稳定性与额度准确性批次：让「回答」在中断时说实话，让「停止」真的停下上游，让凭据不再经过

package/README.en.md

@@ -1,8 +1,10 @@
 # DSH Connect Doubao
 
 > Connect the sign-in **already held by the local Doubao desktop app** to DeepSeek
-> Harness as a Seed-family model provider, plus a **read-only** quota view
-> (5-hour window / weekly window, with reset countdowns).
+> Harness: **chat models** and **「工作任务」 (task-mode) models** (2.1 Lite / 2.1 Pro /
+> Auto, which spend the paid membership quota), **Seedream image generation**, plus a
+> **read-only** quota panel (5-hour window / weekly window / activity grant, each with
+> a reset countdown). Windows x64 only — credential decryption relies on Windows DPAPI.
 
 [![npm version](https://img.shields.io/npm/v/dsh-connect-doubao.svg)](https://www.npmjs.com/package/dsh-connect-doubao)
 [![license](https://img.shields.io/npm/l/dsh-connect-doubao.svg)](./LICENSE)
@@ -33,12 +35,14 @@ Without a signed-in Doubao app the plugin fails loudly rather than silently.
 
 ## Features
 
-- **Seed-family models** — the 6 entries the Doubao desktop app publishes (table below),
-  exposed as a DSH model provider.
+- **Two groups of Doubao models** — two 对话-mode entries (快速 / 2.1 Turbo·专家) and
+  three 「工作任务」-mode entries (2.1 Lite `seed-lite-7b`, 2.1 Pro `5`, Auto `9`).
+  The task models spend the **paid membership quota**; selecting one switches the
+  request into the full task-mode shape automatically.
 - **Three reasoning-effort levels** — low / mid / high, mapped to the measured upstream
   `reasoning_effort` values **3 / 4 / 5**.
-- **Read-only quota panel** — 5-hour and weekly windows as two bars, each with a
-  **reset countdown**; manual refresh plus a 5-minute automatic poll.
+- **Read-only quota panel** — 5-hour, weekly and activity-grant windows as bars, each
+  with a **reset countdown**; manual refresh plus a 5-minute automatic poll.
 - **Model checklist** — ticked entries appear in DSH's model picker and are sent as
   `model_item_key`.
 - **Account status line** — distinguishes signed in / not signed in / sign-in could
@@ -72,7 +76,10 @@ works everywhere):
 - ✅ **All six catalog entries render** — 豆包 快速 / 2.1 Lite (`0921 新版`) /
   2.1 Turbo (`专家`) / 2.1 Pro / 2.1 Turbo (`即将下线`) / 自动, with badges matching
   the desktop app and a count of `已启用 6 ↑`.
-- ✅ **A 「豆包（Doubao）」 group appears in DSH's model picker** with the six entries
+  > This screenshot is from **0.4.0**: back then the task-mode models were only
+  > rendered, not wired up, and the badges used the client's own names. Since 0.4.1
+  > they carry the `工作任务` badge and actually answer; the screenshot is not re-shot.
+- ✅ **A 「豆包（Doubao）」 group appears in DSH's model picker** with the entries
   listed individually. This is the **end-to-end evidence that provider registration
   and the catalog both work**.
 - ✅ **The checklist is interactive** — one checkbox per row plus `刷新` / `全选` /
@@ -84,7 +91,7 @@ works everywhere):
 
 - `doctor` passes all five checks: 29 cookies decrypted with 9/9 required session
   cookies present, the account accepted upstream, the shim bound to a loopback port with
-  a wrong bearer refused (401), and both quota windows readable.
+  a wrong bearer refused (401), all three quota windows readable, and 6 models offered.
 - Reasoning effort low=3 / mid=4 / high=5 was read by **diffing the configuration the
   client re-serialises when the level is switched**, then independently confirmed
   against `reasoning_effort_config` in the client's own config.
@@ -104,14 +111,18 @@ the document the server hands the app) — not reconstructed by guesswork. The
 `model_item_key` **is** the DSH model id, on purpose: one string instead of a
 translation layer that could only be wrong.
 
-| id | name | badge | reasoning |
-| --- | --- | --- | --- |
-| `0` | 豆包 快速 | — | ✅ |
-| `seed-lite-7b` | 豆包 2.1 Lite | `0921 新版` | ✅ |
-| `3` | 豆包 2.1 Turbo | `专家` | ✅ |
-| `5` | 豆包 2.1 Pro | — | ✅ |
-| `4` | 豆包 2.1 Turbo | `即将下线` | ✅ |
-| `9` | 自动 | — | ✅ |
+| id | name | badge | mode | reasoning |
+| --- | --- | --- | --- | --- |
+| `0` | 豆包 快速 | — | 对话 | ✅ |
+| `3` | 豆包 2.1 Turbo | `专家` | 对话 | ✅ |
+| `seed-lite-7b` | 豆包 2.1 Lite | `工作任务` | 工作任务 | ✅ |
+| `5` | 豆包 2.1 Pro | `工作任务` | 工作任务 | ✅ |
+| `9` | 自动 | `工作任务` | 工作任务 | ✅ |
+
+The three `工作任务` rows switch the request into task mode (`mode_id: "3"`,
+`agent_mode: 1`, `need_deep_think: 10001`) and are answered by the plugin's own
+`tool_local` responder — see 「工作任务」 mode below. They are the entries that spend
+the **paid membership quota**.
 
 **Reasoning effort**: low = `3`, mid = `4`, high = `5`. These values were read by
 **diffing the configuration the client re-serialises when the level is switched**,
@@ -119,7 +130,8 @@ then independently confirmed against `reasoning_effort_config` in the client's o
 config. They are observed, not inferred.
 
 > 「即将下线」 is not a property of the Turbo model: the app publishes **two** Turbo
-> entries and only the retiring one carries that badge.
+> entries and only the retiring one carries that badge. The retiring one (`4`) is
+> the only entry withheld from the picker.
 
 ---
 
@@ -139,7 +151,13 @@ DSH (pi-ai provider)
    (`bot_id` + `reasoning_effort`).
 3. **Streaming** — upstream's three-line frames are parsed into pi-ai's event stream;
    tool calls, compaction and permissions stay Harness-owned.
-4. **Read-only routes** — `/plugins/dsh-connect-doubao/quota` and `/catalog` serve quota
+4. **Task mode** — for the `工作任务` entries, `doubao-upstream` additionally runs a
+   `tool_local` responder: it polls the server's `msg/ack` **one `msg_id` at a time**
+   (a batch containing a not-yet-generated seq fails the whole batch with
+   `710010702`), answers `cmd_call` via `upload_command_result` + `msg/complete`, and
+   stops on `agent_finished`. Incremental `CHUNK_DELTA` frames are split into
+   thinking/answer by their block id.
+5. **Read-only routes** — `/plugins/dsh-connect-doubao/quota` and `/catalog` serve quota
    and catalog from an **explicit allow-list** (no upstream field is passed through),
    and refuse non-loopback origins.
 
@@ -257,22 +275,34 @@ npm run check       # all of the above
 1. **It uses a non-public interface.** Doubao publishes no third-party API; this
    plugin reverses the behaviour of its own client. Any upstream change can break it,
    without notice.
-2. **`'4'` (豆包 2.1 Turbo, "即将下线") is unverified.** The other 5 entries were
-   exercised in real conversations.
-3. **There is no response-side field that proves "the server really used that model".**
+2. **`'4'` (豆包 2.1 Turbo, "即将下线") is unverified** — and withheld from the
+   picker. The other 5 entries were exercised in real conversations, including both
+   task-mode models.
+3. **Task mode works, but whether a turn is billed depends on what it did.** In task
+   mode the server hands each turn to `Agent-GeneralTask`, and the plugin runs a
+   `tool_local` responder alongside (`msg/ack` → `upload_command_result` →
+   `msg/complete`) so the local workspace commands complete instead of failing with
+   `710010702`. **For a plain-text turn that triggers no local tool, the server marks
+   the round `general_agent_task_mode: "simple"` with `mode_downgrade_reason: "success"`,
+   and the activity-grant `usedAmount` does not move** (measured: unchanged at 782
+   across a call — i.e. downgraded, not billed). To actually spend the paid quota, the
+   turn has to do multi-step tool/document work. This is a measurement, not a promise.
+4. **There is no response-side field that proves "the server really used that model".**
    The upstream SSE carries an opaque marker in `message.ext` (observed as `38`); it is
    **not** a version number or a tier, and ByteDance has never published its meaning.
    What this plugin can establish is that it sends `model_item_key` with the request
    (confirmed by packet capture). How the server interprets that key is the server's
-   own business.
-4. **Windows only** — credential decryption relies on DPAPI. `User Data` does not
+   own business — sending an invented key still returns HTTP 200, so the field is not
+   validated. What 0.4.1 offers is **the whole task-mode request shape for a given
+   key** (verified to answer), not a guarantee that Pro rather than Lite did the work.
+5. **Windows only** — credential decryption relies on DPAPI. `User Data` does not
    have to live on `C:` (see the five layers above); a browser-only host has no
    native folder dialog, so the picker button reports that instead of failing
    silently — use `DSH_DOUBAO_PROFILE` there.
-5. **Risk-control sliders cannot be automated.** When one fires, send one message in
+6. **Risk-control sliders cannot be automated.** When one fires, send one message in
    the Doubao desktop app to clear it.
-6. **No image input.** The provider sends text blocks only.
-7. **Single account.** No account pool, no region switching, no check-in.
+7. **No image input.** The provider sends text blocks only.
+8. **Single account.** No account pool, no region switching, no check-in.
 
 ---
 

package/README.md

@@ -1,7 +1,9 @@
 # DSH Connect Doubao
 
-> 把**本机豆包（Doubao）桌面客户端已有的登录态**接入 DeepSeek Harness，作为 Seed 系列模型 provider，
-> 并在设置面板里提供一个**只读**的额度概览（5 小时限额 / 周额度 + 重置倒计时）。
+> 把**本机豆包（Doubao）桌面客户端已有的登录态**接入 DeepSeek Harness：**对话模型**与
+> **「工作任务」模型**（2.1 Lite / 2.1 Pro / 自动，走会员付费额度）、**Seedream 生图**，
+> 外加一块**只读**额度看板（5 小时限额 / 周额度 / 活动赠送额度，含重置倒计时）。
+> 仅 Windows x64 —— 凭据解密依赖 Windows DPAPI。
 
 [![npm version](https://img.shields.io/npm/v/dsh-connect-doubao.svg)](https://www.npmjs.com/package/dsh-connect-doubao)
 [![license](https://img.shields.io/npm/l/dsh-connect-doubao.svg)](./LICENSE)
@@ -30,10 +32,12 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
 
 ## 功能特性
 
-- **豆包对话接入** —— 对话模式真正可用的两个条目（见下表），作为 DSH 的模型 provider。
+- **两组豆包模型** —— 对话模式两条（快速 / 2.1 Turbo·专家）与「工作任务」模式三条
+  （2.1 Lite `seed-lite-7b`、2.1 Pro `5`、自动 `9`）。任务模型走的是**会员付费额度**，
+  选中即自动切到任务模式的完整请求形态。
 - **生图（Seedream 5.0）** —— 模型选择器里选「豆包 生图」，回复即以 Markdown 图片返回真实生成图，可直接下载落盘。2026-11-02 实测端到端通过。
 - **推理强度三档** —— 低 / 中 / 高，对应上游 `reasoning_effort` 实测值 **3 / 4 / 5**。
-- **只读额度面板** —— 5 小时限额与周额度两条进度条，各带**重置倒计时**；设置面板里手动刷新 + 5 分钟自动刷新。
+- **只读额度面板** —— 5 小时限额、周额度、活动赠送额度三条进度条，各带**重置倒计时**；设置面板里手动刷新 + 5 分钟自动刷新。
 - **模型勾选清单** —— 勾选的条目会出现在 DSH 的模型选择器中，请求时按 `model_item_key` 指定模型。
 - **账号状态行** —— 区分「已登录 / 未登录 / 登录信息读不出来」等状态，各自给出不同的处置建议。
 - **独立 loopback shim** —— 对话请求走一个只监听 `127.0.0.1` 的本地端点，带每进程共享密钥，任何非本机请求一律拒绝。
@@ -59,6 +63,8 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
   `周额度 已用 9% · 5d 21h 后重置`、`账号状态: 已登录`、`每 5 分钟自动刷新`、配置目录路径。
 - ✅ **模型清单六个条目全部渲染** —— 豆包 快速 / 2.1 Lite（`0921 新版`）/ 2.1 Turbo（`专家`）/
   2.1 Pro / 2.1 Turbo（`即将下线`）/ 自动，角标与桌面包一致，计数显示 `已启用 6 ↑`。
+  > 这张截图来自 **0.4.0**：当时工作任务模型只被渲染出来、还没接通，徽章取客户端原名。
+  > 0.4.1 起它们带 `工作任务` 徽章并且真的能对话，截图尚未重拍。
 - ✅ **模型选择器里出现「豆包（Doubao）」分组** —— 六个条目逐一列出。
   这是 **provider 注册成功 + catalog 生效** 的端到端证据。
 - ✅ **勾选交互可用** —— 每行一个勾选框，配 `刷新` / `全选` / `全不选` / `保存` 四个按钮。
@@ -67,8 +73,8 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
 
 ### 代码级 / 抓包级证据（非截图）
 
-- `doctor` 四段检查全绿：凭据 29 cookie 解密且 9/9 必需会话 cookie 齐备、账号被上游接受、
-  shim 绑定回环端口且错误 bearer 返回 401、两个额度窗口可读。
+- `doctor` 五项检查全绿：凭据 29 cookie 解密且 9/9 必需会话 cookie 齐备、账号被上游接受、
+  shim 绑定回环端口且错误 bearer 返回 401、三个额度窗口可读、模型发现开放 6 个。
 - 推理强度 低=3 / 中=4 / 高=5：由**抓取客户端切换档位时重新序列化的配置 diff**读出，
   并由客户端配置里的 `reasoning_effort_config` 独立确认。
 - 请求体带 `model_item_key` 与 `reasoning_effort`：抓包确认。
@@ -89,20 +95,40 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
 | 对话 | 豆包 快速（`0`）、豆包 2.1 Turbo「专家」（`3`） |
 | 工作任务 | 自动（`9`）、2.1 Lite（`seed-lite-7b`）、2.1 Pro（`5`）、Turbo（`3`） |
 
-**本插件只提供「对话」模式的两条**：
+**本插件提供这两组里的五条**（外加生图别名）：
 
-| id | 名称 | 徽章 | 推理强度 |
-| --- | --- | --- | --- |
-| `0` | 豆包 快速 | — | ✅ |
-| `3` | 豆包 2.1 Turbo | `专家` | ✅ |
+| id | 名称 | 徽章 | 模式 | 推理强度 |
+| --- | --- | --- | --- | --- |
+| `0` | 豆包 快速 | — | 对话 | ✅ |
+| `3` | 豆包 2.1 Turbo | `专家` | 对话 | ✅ |
+| `seed-lite-7b` | 豆包 2.1 Lite | `工作任务` | 工作任务 | ✅ |
+| `5` | 豆包 2.1 Pro | `工作任务` | 工作任务 | ✅ |
+| `9` | 自动 | `工作任务` | 工作任务 | ✅ |
 
-「工作任务」的四条**故意不列**：它们的执行侧是豆包客户端进程内的 CUA 运行时
-（`tasks-mode-resource` + `sbox`），外部客户端注册不进去，每一轮都会得到
-`710010702`。与其在模型选择器里摆出选不中的条目，不如不摆。
+**「工作任务」三条为什么能用了。** 上一版把它们整个隐藏，理由是执行侧是豆包客户端进程内的
+CUA 运行时（`tasks-mode-resource` + `sbox`），外部客户端注册不进去，每一轮必然得到
+`710010702`。那个判断只对了一半：CUA 进程确实接不进去，但**它不是必需的** —— 服务端
+真正等待的只是三条 HTTP 应答：
+
+```
+POST /alice/office/tool_local/msg/ack              # 轮询本地指令
+POST /alice/office/tool_local/upload_command_result # 回传执行结果
+POST /alice/office/tool_local/msg/complete          # 标记该指令完成
+```
+
+服务端按 `${unique_key}:${device_id}:${aid}:message:${seq}` 顺序下发指令（`seq = 0,1,2,…`）：
+`command: 6003`（`cmd_call`，如 `local.workspace.apply`）需要回传结果，
+`command: 6004`（`agent_finished`）表示这一轮结束。插件在流式生成期间并发跑一个应答器
+把这些走完，于是任务回合不再落到 `710010702`。**实测 2026-11-05 跑通，Lite 与 Pro
+各一次纯文本回合。**
+
+> ⚠️ **`msg/ack` 必须单条查询。** `{ msg_ids: [0,1,2,3,4,5,6] }` 会整体失败并返回
+> `{"code":710010702,"msg":"系统错误"}` —— 批里只要含一个服务端尚未生成的 `seq`，
+> 整批就废。这是逆向时最容易踩的坑，`tests/` 里没有钉它（需要真实上游才能复现）。
 
 **推理强度**：低 = `3`、中 = `4`、高 = `5`（由客户端切换档位时重新序列化的配置 diff 读出，
 并由配置里的 `reasoning_effort_config` 独立确认）。深度思考开关本身走
-`need_deep_think: 0`，档位只由 `reasoning_effort` 携带 —— 这是 2026-10-22 CDP 抓包更正的。
+`need_deep_think: 0`；任务模式下抓到的真实请求固定发 `10001`，档位仍只由 `reasoning_effort` 携带。
 
 ---
 
@@ -146,7 +172,9 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
   `0x22` 是 Latin-1、`0x63` 是 UTF-16LE。**实测确认 `0x63` 的长度字段是字节数而不是字符数**
   —— 按字符数会正好读多一倍，模型名后面会吃掉下一个字段的字节。
 - 客户端给「工作任务」专属模型的名称自带前缀（`工作任务 Pro`、`工作任务 Auto`），
-  所以模式归属直接从缓存里就能读出来：带前缀的一律不开放（它们需要 App 内的 CUA 运行时）。
+  所以模式归属直接从缓存里就能读出来。**但带前缀不再等于隐藏**：0.4.1 起它们会被
+  标记成 `taskMode` 并正常开放，只有 `DOUBAO_UNAVAILABLE_MODELS` 里的条目（目前仅
+  「即将下线」的 `4`）被排除。缓存里缺了某条任务模型时，会从内置清单补齐。
 - 缓存读不到、或解析结果不可信时，**回落到内置清单**，绝不清空用户的选择器。
 - 结果缓存 10 分钟；卡片上的「刷新」（POST）会丢弃记忆重新读盘。
 - `doctor` 第 5 项会报告这次到底是从缓存发现的、还是回落的。
@@ -154,7 +182,7 @@ Chromium cookie（通过 Windows DPAPI 在本进程内解密），把它当作
 当前这台机器上的实测输出：
 
 ```
-[ ok ] models  客户端缓存 5 个模型，开放 3 个（「工作任务」专属的按名字前缀排除）
+[ ok ] models  客户端缓存 5 个模型，开放 6 个（仅排除「即将下线」等不可用条目）
 ```
 
 ## 实验性：工具调用（默认关闭）
@@ -309,7 +337,7 @@ dsh-connect-doubao-probe --prompt "hi" --model 5 --thinking high --json
 --timeout <ms>               超时（默认 120000）
 ```
 
-`doctor` 输出四段独立结论：凭据 / 账号 / provider 装配 / 额度。
+`doctor` 输出五项独立结论：凭据 / 账号 / provider 装配 / 额度 / 模型发现。
 凭据只报**数量**，不报值。
 
 ---
@@ -350,21 +378,24 @@ npm run check       # 以上全部
 5. **风控滑块无法自动处理。** 触发后需去豆包客户端手动发一条消息完成验证。
 6. **图片输入未提供。** 本 provider 只发送文本块，不转发图片。
 7. **单账号。** 没有账号池、没有区域切换、没有签到。
-8. **「工作任务」跑不了。** `agent_mode` 的取值是 `2 = 普通对话`、`1 = 工作任务`（CDP 抓包实测，
-   2026-10-22）。任务模式下服务端会把每一轮交给 `Agent-GeneralTask`，而它的执行运行时是豆包
-   客户端进程里的 CUA（`tasks-mode-resource` + `sbox`），外部客户端注册不进去，必然回
-   `710010702`。所以插件固定走对话模式；请求体里的 `general_task_param` 发的是客户端在普通
-   对话下的默认值 `{"server_do_compare":true}`。
-9. **`model_item_key` 服务端根本不校验，所以选不了模型。** 三条证据（2026-10-23）：
+8. **「工作任务」能跑，但扣不扣费取决于这一轮做了什么。** `agent_mode` 的取值是
+   `2 = 普通对话`、`1 = 工作任务`（CDP 抓包实测，2026-10-22）。任务模式下服务端把每一轮交给
+   `Agent-GeneralTask`，并由插件并发跑一个 `tool_local` 应答器（`msg/ack` →
+   `upload_command_result` → `msg/complete`）把本地指令走完，所以不再必然回 `710010702`。
+   **但纯文本、不触发本地工具的小回合，服务端会在 `ext` 里标
+   `general_agent_task_mode: "simple"` 与 `mode_downgrade_reason: "success"`，此时活动赠送
+   额度的 `usedAmount` 不变**（实测调用前后恒为 782，即「走降级、不计费」）。想让任务模型
+   真正消耗付费额度，得让它执行多步工具/文档任务。这一条是实测结论，不是产品承诺。
+9. **`model_item_key` 服务端不校验，所以「选中哪个模型」只能靠模式与账号决定。** 三条证据
+   （2026-10-23）：
    1. CDP 抓到 App 选 2.1 Pro 时的真实请求：`model_item_key: "5"` + `mode_id: "3"` +
-      `agent_mode: 1` + `general_task_param.runtime_type: 2`（本地电脑）。前三项我们能复制，
-      最后一项是 App 进程内的 CUA 运行时，复制不了 —— 这正是 `710010702` 的来源。
+      `agent_mode: 1` + `general_task_param.runtime_type: 2`（本地电脑）。
    2. 用 `mode_id=3` + `key=5` 在对话模式发请求：正常回答，但与 `key=0` 无可测差异
       （耗时 15.1–17.6 s、长度 311–340 字，均在噪声范围内）。
    3. **真正的对照**：把三处 `model_item_key` 全部改成一个不存在的值直接发送，服务端仍返回
       200 与完整正常流，没有任何错误帧 —— 它不校验这个字段。
-   所以「在 DSH 里指定 2.1 Pro / 2.1 Lite」做不到：模型由服务端按产品/模式/账号决定。
-   真要用这两个模型，只能在豆包客户端自己的「工作任务」里用。
+   所以 0.4.1 能提供的是**「工作任务模式 + 该 key」这一整组请求形态**（实测能出答案），
+   而不是「服务端一定会用 Pro 而不是 Lite 来算」。要更强的模型，仍建议在豆包客户端里确认。
 10. **本插件的设置不走宿主 settings，存在自己的 JSON 文件里。**
     `~/.dsh/dsh-connect-doubao.settings.json`，由插件自己写入（先写临时文件再原子重命名），
     读取也不依赖任何宿主服务在线。原因是宿主这条路对本插件**完全关闭**：它的 fiber 永远停在

package/lib/cli-doctor.d.ts

@@ -1,4 +1,4 @@
-import { T as classifyDoubaoFailure, f as DoubaoUpstreamError } from "./doubao-upstream-BdIak5wD.js";
+import { O as classifyDoubaoFailure, f as DoubaoUpstreamError } from "./doubao-upstream-DEeYQzAG.js";
 //#region src/cli-doctor.d.ts
 type CheckStatus = 'ok' | 'fail' | 'skip';
 interface DoctorCheck {

package/lib/cli-doctor.js

@@ -1,6 +1,6 @@
 #!/usr/bin/env node
-import { H as PROFILE_ENV, I as REQUIRED_COOKIE_NAMES, K as listDoubaoProfiles, U as describeProfileLocation, V as loadDoubaoCredentials, _t as discoverDoubaoModels, ct as DOUBAO_PROVIDER_DISPLAY_NAME, i as formatDuration, j as probeAccount, m as DoubaoUpstreamError, n as fetchQuota, o as millisUntilReset, ot as DOUBAO_MODELS, r as findWindow, st as DOUBAO_PROVIDER, w as classifyDoubaoFailure } from "./doubao-quota-Do37dMMj.js";
-import { g as createDoubaoAdapter, i as doubaoQuotaView, m as createDoubaoShim, n as DOUBAO_QUOTA_PATH, p as liveModels, u as discoverAllCachedModels } from "./doubao-status-CeXUM4Cy.js";
+import { D as classifyDoubaoFailure, G as PROFILE_ENV, K as describeProfileLocation, P as probeAccount, W as loadDoubaoCredentials, Y as listDoubaoProfiles, bt as discoverDoubaoModels, dt as DOUBAO_PROVIDER_DISPLAY_NAME, i as formatDuration, lt as DOUBAO_MODELS, m as DoubaoUpstreamError, n as fetchQuota, o as millisUntilReset, r as findWindow, ut as DOUBAO_PROVIDER, z as REQUIRED_COOKIE_NAMES } from "./doubao-quota-BYF0-RMF.js";
+import { g as createDoubaoAdapter, i as doubaoQuotaView, m as createDoubaoShim, n as DOUBAO_QUOTA_PATH, p as liveModels, u as discoverAllCachedModels } from "./doubao-status-DI2Rj0U_.js";
 //#region src/cli-doctor.ts
 /**
 * `doctor` — verify the whole stack end to end without spending inference quota.
@@ -160,7 +160,7 @@ async function runDoctor() {
 	checks.push({
 		name: "models",
 		status: offered.length > 0 ? "ok" : "fail",
-		detail: cached.length > 0 ? `客户端缓存 ${cached.length} 个模型，开放 ${offered.length} 个（「工作任务」专属的按名字前缀排除）` : `读不到客户端缓存，回落到内置清单（${offered.length} 个）`,
+		detail: cached.length > 0 ? `客户端缓存 ${cached.length} 个模型，开放 ${offered.length} 个（仅排除「即将下线」等不可用条目）` : `读不到客户端缓存，回落到内置清单（${offered.length} 个）`,
 		data: {
 			source: cached.length > 0 ? "client-cache" : "static-fallback",
 			cached: cached.map((entry) => `${entry.key}=${entry.name}${entry.taskOnly ? "[工作任务]" : ""}`),

package/lib/cli-probe.js

@@ -1,5 +1,5 @@
 #!/usr/bin/env node
-import { M as streamChat, U as describeProfileLocation, V as loadDoubaoCredentials, bt as findDoubaoModel, i as formatDuration, j as probeAccount, m as DoubaoUpstreamError, n as fetchQuota, o as millisUntilReset, ot as DOUBAO_MODELS, r as findWindow, xt as findReasoningLevel } from "./doubao-quota-Do37dMMj.js";
+import { Ct as findDoubaoModel, F as streamChat, K as describeProfileLocation, P as probeAccount, W as loadDoubaoCredentials, i as formatDuration, lt as DOUBAO_MODELS, m as DoubaoUpstreamError, n as fetchQuota, o as millisUntilReset, r as findWindow, wt as findReasoningLevel } from "./doubao-quota-BYF0-RMF.js";
 //#region src/cli-probe.ts
 /**
 * Doubao smoke probe.

package/lib/client.js

@@ -1060,7 +1060,7 @@ window.__ModuleLoader__.load({
 			"account.profile": "配置目录",
 			"models.title": "模型",
 			"models.enabledCount": "已启用 {count} ↑",
-			"models.routingNote": "清单直接读自豆包客户端的缓存，豆包上线新模型时会自动出现，不需要等插件更新。「工作任务」专属的条目（2.1 Pro / Lite / Auto）由客户端内的运行时执行，外部客户端接不进去，已自动隐藏",
+			"models.routingNote": "清单直接读自豆包客户端的缓存，豆包上线新模型时会自动出现，不需要等插件更新。带「工作任务」徽章的条目（2.1 Lite / 2.1 Pro / 自动）走任务模式与会员付费额度，插件会自动完成本地指令应答",
 			"models.refresh": "刷新",
 			"models.refreshing": "刷新中…",
 			"models.save": "保存",
@@ -1132,7 +1132,7 @@ window.__ModuleLoader__.load({
 			"account.profile": "Profile path",
 			"models.title": "Models",
 			"models.enabledCount": "{count} enabled ↑",
-			"models.routingNote": "Only the 对话-mode entries are offered (快速 / 专家). The 「工作任务」-only models (2.1 Pro / 2.1 Lite) are executed by the Doubao desktop app runtime, which an external client cannot join",
+			"models.routingNote": "The list is read from the Doubao client cache, so new models appear without a plugin update. Entries badged 「工作任务」 (2.1 Lite / 2.1 Pro / Auto) run in task mode against the paid membership quota; the plugin answers the local workspace commands for you",
 			"models.refresh": "Refresh",
 			"models.refreshing": "Refreshing…",
 			"models.save": "Save",