dsh-cline-pass 0.2.8 → 0.2.9

package/README.en.md

@@ -1,61 +1,41 @@
 # dsh-cline-pass
 
-Connect [Cline Pass](https://cline.bot/cline-pass) subscription models to the [DeepSeek Harness](https://github.com/deepseek-ai). Install it as a dsh plugin and the models appear in the model list without running a proxy service.
+[![npm](https://img.shields.io/npm/v/dsh-cline-pass)](https://www.npmjs.com/package/dsh-cline-pass)
+[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 
-## Features
-
-- Use Cline Pass subscription models from dsh
-- Streaming, tool calls, image input, and reasoning
-- Multiple accounts with optional round-robin rotation
-- Probe, test, pin, or exclude upstream channels
-- Automatic failover before the first token
-- Web settings panel and `cline_pass_*` tools
+[中文](README.md) | English
 
-## Requirements
+Use [Cline Pass](https://cline.bot/cline-pass) subscription models directly in [DeepSeek Harness (dsh)](https://github.com/deepseek-ai), without running a separate proxy service.
 
-- dsh `>= 0.1.2-alpha.3 < 0.3.0`
-- Node.js `>= 20.3`
-- A Cline Pass API key
+## Features
 
-## Install
+- Streaming, tool calls, image input, and reasoning, depending on the model.
+- Multiple accounts with single-account and round-robin modes.
+- Discover, validate, pin, or exclude upstream channels, with failover before the first token.
+- Web settings panel and optional `cline_pass_*` management tools.
 
-```bash
-dsh plugin --profile web add dsh-cline-pass
-```
+## Quick start
 
-You can also install a local directory or tarball:
+Requires Node.js `>= 20.3.0`, dsh `>= 0.1.2-alpha.3 < 0.3.0`, and a valid Cline Pass API key.
 
 ```bash
-dsh plugin --profile web add /path/to/dsh-cline-pass
-dsh plugin --profile web add ./dsh-cline-pass-<version>.tgz
+dsh plugin --profile web add dsh-cline-pass
 ```
 
-Restart dsh after installation.
-
-## Configure
+Restart dsh, open **Settings → Built-in Plugins → Cline Pass** to save and test your API key, then select a Cline Pass model in the model picker. Keys are stored in the dsh credential store and displayed only in masked form.
 
-With the Web profile, open **Settings → Cline Pass** to save and test a key. Keys are stored in the dsh credential store and shown only in masked form.
-
-You can also set an environment variable before starting dsh:
+Alternatively, configure the key through an environment variable:
 
 ```bash
-export CLINE_PASS_API_KEY=sk_xxx
+export CLINE_PASS_API_KEY="your-cline-pass-api-key"
 dsh --profile web
 ```
 
-### Multiple accounts
-
-Add accounts in the settings panel and select single-account or round-robin mode. You can also use the tool interface:
-
-```text
-cline_pass_accounts action=add name=main key=sk_xxx
-cline_pass_accounts action=add name=backup key=sk_yyy
-cline_pass_accounts action=mode mode=roundrobin
-```
+These commands use the `web` profile; replace the name if you use another profile.
 
-### Configuration file
+## Configuration
 
-To override defaults, add a `cline-pass` section to your profile's `$DSH_HOME/profiles/<name>/cordis.patch.yml`:
+Use the Web panel for everyday configuration. For file-based configuration, edit your profile's `$DSH_HOME/profiles/<name>/cordis.patch.yml`, for example:
 
 ```yaml
 - id: cline-pass
@@ -66,77 +46,49 @@ To override defaults, add a `cline-pass` section to your profile's `$DSH_HOME/pr
         apiKeyEnv: CLINE_PASS_MAIN_KEY
       backup:
         apiKeyEnv: CLINE_PASS_BACKUP_KEY
+    exposeTools: false
 ```
 
-Common settings include `baseURL`, `knownModels`, `models`, `perModel`, `exposeTools`, `exposeCatalog`, and `historyLimit`. The default gateway is `https://api.cline.bot/api/v1`. `maxConcurrentRequests` caps how many streams each account keeps open to the gateway at once (default 16); further requests wait for a free slot, and `0` lifts the cap.
-
-### Management tools
-
-The `cline_pass_*` tools are **not registered by default**. Their eight definitions total roughly 5.5k characters of schema and description that ride in every request's tool list, and managing the route is an occasional need, so the default keeps that cost out of every conversation.
-
-Turn them on with **Expose the tools (takes effect after a restart)** in **Settings → Cline Pass**, or in the profile configuration:
-
-```yaml
-- id: cline-pass
-  config:
-    exposeTools: true
-```
-
-**The value is read when the plugin activates, so a change needs a dsh restart** — a tool already registered cannot be pulled back out of a conversation in progress.
-
-The provider route and the setup panel are unaffected — the panel is a browser page rather than a tool, and it stays available with the tools off.
+Each `apiKeyEnv` refers to a credential name or environment variable; configure a key for each account. Management tools are disabled by default. Set `exposeTools` to `true` and restart dsh to enable them.
 
-> A profile upgraded from 0.2.1 or earlier has no `exposeTools` key, so it adopts the new default of `false` and its `cline_pass_*` tools stop being registered. The panel reports this; tick the checkbox and restart to bring them back.
+The default gateway is `https://api.cline.bot/api/v1`. Each account allows up to `16` concurrent requests (`maxConcurrentRequests`; `0` removes the limit). See [`lib/index.js`](lib/index.js) for more options and [`lib/tools.js`](lib/tools.js) for management tools.
 
 ## Upstream channels
 
-For a new model, expand its row and click **Probe** to discover channels, then **Validate** to check which channels actually honor a pin. Choose a channel order or exclusions as needed. Validation sends one real request per channel and does not change the stored pin; **Test** checks the saved pin rules.
-
-You can also run the tools in this order:
-
-```text
-cline_pass_probe    model=cline-pass/glm-5.2
-cline_pass_validate model=cline-pass/glm-5.2
-cline_pass_pin      model=cline-pass/glm-5.2 upstreams=["alibaba","baseten"] pinMode=preferred
-```
+Expand a model row in the settings panel, click **Probe → Validate**, save your channel order or exclusions, then click **Test** to verify.
 
-`pinMode` supports `strict` and `preferred`. `preferred` tries channels in order and fails over before the first token when needed. Use `exclude` to block specific channels.
+- `strict`: Restrict each attempt to its specified candidate channel.
+- `preferred`: Send a preferred channel order to the gateway and try subsequent candidates when needed.
 
-## Tools
+Validation sends real requests to each channel without changing saved rules. Failover for retryable errors takes place before the first output token.
 
-| Tool | Purpose |
-| --- | --- |
-| `cline_pass_status` | Show route and account status |
-| `cline_pass_models` | List models and channels |
-| `cline_pass_probe` | Probe channels |
-| `cline_pass_validate` | Test each channel |
-| `cline_pass_test` | Verify the current settings |
-| `cline_pass_pin` | Save channel settings |
-| `cline_pass_accounts` | Manage accounts |
-| `cline_pass_history` | Read request history |
+## Development and contributing
 
-## Development
+The project uses JavaScript ES Modules with no build step. Install dependencies and run:
 
 ```bash
+npm install
 npm test
 npm run test:client
 npm run test:mount
 ```
 
-Live gateway checks are also available:
-
-```bash
-npm run test:live
-npm run test:live:image
-npm run test:live:reasoning
-```
+Tests require compatible dsh packages; the mount test also requires a complete dsh installation. `npm run test:live`, `npm run test:live:image`, and `npm run test:live:reasoning` access the live gateway, require a valid key, and consume quota or incur charges.
 
-These commands need a valid Cline Pass key and make a small number of paid requests.
+[Issues](https://github.com/yhshzh/dsh-cline-pass/issues) and Pull Requests are welcome. Include versions, reproduction steps, and sanitized logs in bug reports, and run relevant tests before submitting code.
 
 ## Acknowledgements
 
 Upstream channel and failover behavior was informed by the MIT-licensed [`cline-pass-switcher`](https://github.com/munmunjaklin458-afk/cline-pass-switcher).
 
+## Community links
+
+[LinuxDo](https://linux.do/) — A community for open source and technology, built on sincerity, kindness, unity, and professionalism.
+
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=yhshzh/dsh-cline-pass&type=Date)](https://www.star-history.com/#yhshzh/dsh-cline-pass&Date)
+
 ## License
 
 [MIT](LICENSE)

package/README.md

@@ -1,61 +1,41 @@
 # dsh-cline-pass
 
-将 [Cline Pass](https://cline.bot/cline-pass) 订阅模型接入 [DeepSeek Harness](https://github.com/deepseek-ai)。安装为 dsh 插件后，模型会直接出现在模型列表中，无需运行代理服务。
+[![npm](https://img.shields.io/npm/v/dsh-cline-pass)](https://www.npmjs.com/package/dsh-cline-pass)
+[![License: MIT](https://img.shields.io/badge/License-MIT-blue.svg)](LICENSE)
 
-## 功能
-
-- 使用 Cline Pass 订阅模型
-- 支持流式输出、工具调用、图片和 reasoning
-- 管理多个账号并轮询使用
-- 探测、测试、钉住或排除上游渠道
-- 在首个 token 前自动故障转移
-- 提供 Web 设置面板和 `cline_pass_*` 工具
+中文 | [English](README.en.md)
 
-## 要求
+将 [Cline Pass](https://cline.bot/cline-pass) 订阅模型接入 [DeepSeek Harness（dsh）](https://github.com/deepseek-ai)，直接在模型列表中使用，无需部署代理服务。
 
-- dsh `>= 0.1.2-alpha.3 < 0.3.0`
-- Node.js `>= 20.3`
-- Cline Pass API Key
+## 功能
 
-## 安装
+- 支持流式输出、工具调用、图片输入和推理，具体能力取决于模型。
+- 多账号管理，支持单账号与轮询模式。
+- 探测、校验、指定或排除上游渠道，支持首个 token 前的故障转移。
+- 提供 Web 设置面板和可选的 `cline_pass_*` 管理工具。
 
-```bash
-dsh plugin --profile web add dsh-cline-pass
-```
+## 快速开始
 
-也可以安装本地目录或 tarball：
+需要 Node.js `>= 20.3.0`、dsh `>= 0.1.2-alpha.3 < 0.3.0` 和有效的 Cline Pass API Key。
 
 ```bash
-dsh plugin --profile web add /path/to/dsh-cline-pass
-dsh plugin --profile web add ./dsh-cline-pass-<version>.tgz
+dsh plugin --profile web add dsh-cline-pass
 ```
 
-安装后重启 dsh。
-
-## 配置
-
-Web profile 可以在 **设置 → Cline Pass** 中保存和测试 Key。Key 会存入 dsh 凭据库，界面只显示掩码。
+安装后重启 dsh，在 **设置 → 内置插件 → Cline Pass** 中保存并测试 API Key，然后在模型选择器中选择 Cline Pass 模型。Key 存入 dsh 凭据库，界面仅显示掩码。
 
-也可以在启动 dsh 前设置环境变量：
+也可以通过环境变量配置：
 
 ```bash
-export CLINE_PASS_API_KEY=sk_xxx
+export CLINE_PASS_API_KEY="your-cline-pass-api-key"
 dsh --profile web
 ```
 
-### 多账号
-
-在设置面板中添加账号，并选择单账号或轮询模式。也可以调用工具：
-
-```text
-cline_pass_accounts action=add name=main key=sk_xxx
-cline_pass_accounts action=add name=backup key=sk_yyy
-cline_pass_accounts action=mode mode=roundrobin
-```
+以上使用 `web` profile，如需使用其他 profile，请替换对应名称。
 
-### 配置文件
+## 配置
 
-需要覆盖默认值时，在 profile 的 `$DSH_HOME/profiles/<name>/cordis.patch.yml` 中添加：
+日常配置可在 Web 面板完成。通过文件配置时，修改对应 profile 的 `$DSH_HOME/profiles/<name>/cordis.patch.yml`，例如：
 
 ```yaml
 - id: cline-pass
@@ -66,72 +46,36 @@ cline_pass_accounts action=mode mode=roundrobin
         apiKeyEnv: CLINE_PASS_MAIN_KEY
       backup:
         apiKeyEnv: CLINE_PASS_BACKUP_KEY
+    exposeTools: false
 ```
 
-### 管理工具
-
-`cline_pass_*` 工具**默认不注册**。8 个工具定义合计约 5.5k 字符，会出现在每次请求的工具列表里，而管理路由只是偶尔需要，所以默认让所有对话都不必付这份开销。
-
-需要时在**设置 → Cline Pass** 中勾选「开启工具注入（重启后生效）」，或在 profile 配置里打开：
-
-```yaml
-- id: cline-pass
-  config:
-    exposeTools: true
-```
-
-**该值在插件激活时读取，改动后需重启 dsh 才生效** —— 已经注册的工具无法从进行中的对话里撤回。
-
-提供商路由和设置面板不受影响 —— 面板是浏览器页面而非工具，工具关闭时仍然可用。
-
-> 从 0.2.1 及更早版本升级的 profile 里没有 `exposeTools` 键，因此会采用新的默认值 `false`：升级后 `cline_pass_*` 工具将不再注册。面板会提示，勾选该复选框并在重启后恢复。
+`apiKeyEnv` 指向凭据名称或环境变量，请分别配置两个账号的 Key。管理工具默认关闭，需要时将 `exposeTools` 设为 `true`，并重启 dsh。
 
-常用设置包括 `baseURL`、`knownModels`、`models`、`perModel`、`exposeCatalog` 和 `historyLimit`。默认网关地址为 `https://api.cline.bot/api/v1`。`maxConcurrentRequests` 限制每个账号同时发往网关的流数，默认 16，超出的请求排队等待空位，设为 `0` 表示不限。
+默认网关为 `https://api.cline.bot/api/v1`，每个账号的并发请求上限为 `16`（`maxConcurrentRequests`，`0` 表示不限）。更多选项见 [`lib/index.js`](lib/index.js)，管理工具见 [`lib/tools.js`](lib/tools.js)。
 
 ## 上游渠道
 
-首次使用模型时，展开面板中的模型行，先点击 **探测** 获取渠道，再点击 **校验** 检查各渠道是否确实采用了钉住规则，最后按需要选择渠道顺序或排除渠道。校验会向每个渠道发出一次真实请求，不会自动修改钉住配置；**测试** 可验证已保存的钉住规则。
+在设置面板中展开模型行，依次点击 **探测 → 校验**，保存渠道顺序或排除列表后，点击 **测试** 验证。
 
-也可以按以下顺序调用工具：
+- `strict`：每次尝试限制到指定的候选渠道。
+- `preferred`：向网关传递渠道偏好顺序，必要时尝试后续候选渠道。
 
-```text
-cline_pass_probe    model=cline-pass/glm-5.2
-cline_pass_validate model=cline-pass/glm-5.2
-cline_pass_pin      model=cline-pass/glm-5.2 upstreams=["alibaba","baseten"] pinMode=preferred
-```
-
-`pinMode` 支持 `strict` 和 `preferred`。`preferred` 会按顺序尝试渠道，并在首个 token 前失败时自动切换。`exclude` 可排除指定渠道。
+校验会向每个渠道发送真实请求，不会修改已保存的规则；可重试的故障转移发生在首个 token 输出前。
 
-## 工具
+## 开发与贡献
 
-| 工具 | 用途 |
-| --- | --- |
-| `cline_pass_status` | 查看路由和账号状态 |
-| `cline_pass_models` | 查看模型和渠道 |
-| `cline_pass_probe` | 探测渠道 |
-| `cline_pass_validate` | 测试每个渠道 |
-| `cline_pass_test` | 验证当前设置 |
-| `cline_pass_pin` | 保存渠道设置 |
-| `cline_pass_accounts` | 管理账号 |
-| `cline_pass_history` | 查看请求历史 |
-
-## 开发
+项目使用 JavaScript ES Modules，无需编译。安装依赖后运行：
 
 ```bash
+npm install
 npm test
 npm run test:client
 npm run test:mount
 ```
 
-访问真实网关的测试还包括：
-
-```bash
-npm run test:live
-npm run test:live:image
-npm run test:live:reasoning
-```
+测试需要兼容版本的 dsh 相关包，挂载测试还需要完整的 dsh 安装。`npm run test:live`、`npm run test:live:image` 和 `npm run test:live:reasoning` 会访问真实网关，需要有效 Key 并消耗额度或产生费用。
 
-这些命令需要有效的 Cline Pass Key，并会产生少量请求费用。
+欢迎提交 [Issue](https://github.com/yhshzh/dsh-cline-pass/issues) 或 Pull Request。报告问题时请附版本、复现步骤及脱敏日志；提交代码前请运行相关测试。
 
 ## 致谢
 
@@ -141,6 +85,10 @@ npm run test:live:reasoning
 
 [LinuxDo](https://linux.do/) — 真诚、友善、团结、专业，你的品质开源与技术社区
 
-## License
+## Star History
+
+[![Star History Chart](https://api.star-history.com/svg?repos=yhshzh/dsh-cline-pass&type=Date)](https://www.star-history.com/#yhshzh/dsh-cline-pass&Date)
+
+## 许可证
 
 [MIT](LICENSE)

package/lib/catalog.js

@@ -36,6 +36,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 384000,
     reasoning: true,
+    efforts: ['none', 'low', 'high', 'max'],
     input: ['text', 'image'],
   },
   'cline-pass/deepseek-v4-flash': {
@@ -43,6 +44,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 384000,
     reasoning: true,
+    efforts: ['none', 'low', 'medium', 'high'],
     input: ['text'],
   },
   'cline-pass/deepseek-v4-pro': {
@@ -50,6 +52,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 384000,
     reasoning: true,
+    efforts: ['none', 'low', 'high', 'max'],
     input: ['text'],
   },
   'cline-pass/glm-5.3': {
@@ -57,6 +60,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 131072,
     reasoning: true,
+    efforts: ['low', 'high', 'max'],
     input: ['text'],
   },
   'cline-pass/glm-5.3-flash': {
@@ -64,6 +68,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 131072,
     reasoning: true,
+    efforts: ['low', 'high', 'max'],
     input: ['text', 'image'],
   },
   'cline-pass/glm-5.2': {
@@ -71,6 +76,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 131072,
     reasoning: true,
+    efforts: ['none', 'high', 'max'],
     input: ['text'],
   },
   'cline-pass/kimi-k3': {
@@ -78,6 +84,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1048576,
     maxTokens: 131072,
     reasoning: true,
+    efforts: ['low', 'high', 'max'],
     input: ['text', 'image'],
   },
   'cline-pass/kimi-k2.7-code': {
@@ -85,6 +92,7 @@ export const MODEL_CATALOG = {
     contextWindow: 262144,
     maxTokens: 262144,
     reasoning: true,
+    efforts: ['none', 'low', 'medium', 'high'],
     input: ['text', 'image'],
   },
   'cline-pass/kimi-k2.6': {
@@ -106,6 +114,7 @@ export const MODEL_CATALOG = {
     contextWindow: 1000000,
     maxTokens: 131072,
     reasoning: true,
+    efforts: ['none', 'low', 'medium', 'xhigh'],
     input: ['text', 'image'],
   },
   'cline-pass/qwen3.7-max': {
@@ -165,6 +174,24 @@ function pick(configured, published, fallback) {
  */
 const SEAM_MODALITIES = new Set(['text', 'image'])
 
+/**
+ * The reasoning levels one model actually offers, as the seam wants them.
+ *
+ * Every reasoning model used to advertise the whole vocabulary, which the picker
+ * then showed — seven levels for a model that answers to four. The levels are a
+ * per-model fact, and the gateway publishes them (see `cline.js`), so an entry
+ * carries the list it was published with.
+ *
+ * An absent or empty list keeps the whole vocabulary, which is the honest answer
+ * for the sources that name none: a model published with only an on/off toggle, or
+ * with only a token budget, has no levels to narrow to.
+ */
+function reasonEfforts(ids) {
+  if (!Array.isArray(ids) || ids.length === 0) return REASONING_EFFORTS
+  const byId = new Map(REASONING_EFFORTS.map((effort) => [effort.id, effort]))
+  return ids.map((id) => byId.get(id) ?? { id, name: id, description: '' })
+}
+
 /**
  * The output budget to advertise for one model.
  *
@@ -233,6 +260,6 @@ export function resolveModelMetadata(provider, model, override, fallback, discov
     inputModalities: offered.length > 0 ? offered : ['text'],
     context: { contextWindow },
     defaultMaxTokens: maxTokens,
-    ...(reasoning === true ? { reasoning: { efforts: REASONING_EFFORTS } } : {}),
+    ...(reasoning === true ? { reasoning: { efforts: reasonEfforts(published.efforts) } } : {}),
   }
 }

package/lib/cline.js

@@ -8,10 +8,13 @@
  * - `GET  {baseURL}/users/{id}/usages` — the per-request usage history behind them
  * - `GET  api.cline.bot/.../recommended-models` — official subscription list
  * - `GET  models.dev/api.json` — community registry, a backstop
+ * - `GET  ai-gateway.vercel.sh/v1/models` — the upstream's own catalog
  * - `GET  openrouter.ai/api/v1/...` — endpoint detail for direct-pipeline models
  *
  * @module dsh-cline-pass/cline
  */
+
+import { REASONING_EFFORTS } from './catalog.js'
 /** The public Cline endpoint that lists subscription models without auth. */
 export const RECOMMENDED_MODELS_URL = 'https://api.cline.bot/api/v1/ai/cline/recommended-models'
 /** The authenticated sub-path that reports quota consumption for the account. */
@@ -36,6 +39,17 @@ const DAY_MS = 86_400_000
 export const MODEL_ROWS_MAX = 8
 /** The community model registry. */
 export const MODELS_DEV_URL = 'https://models.dev/api.json'
+/**
+ * The upstream gateway's own model catalog.
+ *
+ * It is the only source that publishes the reasoning levels behind a Cline Pass
+ * model. Cline Pass does not: its `/models` carries ids only, and its gateway
+ * accepts every level in the harness vocabulary for every model, so an unsupported
+ * one cannot even be probed out — a level outside the vocabulary is the only
+ * rejection it makes. The route is named in the gateway's own failures
+ * (`… from Vercel`), and that catalog is open.
+ */
+export const VERCEL_MODELS_URL = 'https://ai-gateway.vercel.sh/v1/models'
 /** OpenRouter's public API, used only for direct-pipeline endpoint detail. */
 export const OPENROUTER_API = 'https://openrouter.ai/api/v1'
 
@@ -147,6 +161,38 @@ function modelMetadata(entry) {
   return meta
 }
 
+/**
+ * The reasoning levels one model offers, from the upstream's `reasoning_options`.
+ *
+ * An `effort` entry names the levels; a `toggle` entry is the model's on/off switch,
+ * which in the harness vocabulary is `none`. Returned in escalating order so the
+ * picker reads in one direction whatever order the catalog happened to list.
+ *
+ * Models published with only a toggle, or only a token budget, name no levels at
+ * all. Those return undefined rather than an invented list, and the caller keeps the
+ * full vocabulary — narrowing a model to a single `none` would take away the switch
+ * the catalog just described.
+ *
+ * @returns the level ids, or undefined when the entry names none.
+ */
+function effortsFromOptions(options) {
+  if (!Array.isArray(options)) return undefined
+  const levels = new Set()
+  let toggle = false
+  for (const option of options) {
+    if (option?.type === 'toggle') toggle = true
+    if (option?.type !== 'effort' || !Array.isArray(option.values)) continue
+    for (const value of option.values) if (typeof value === 'string' && value.length > 0) levels.add(value)
+  }
+  // A toggle on its own names nothing to choose from — it says the switch exists,
+  // not what to set it to — so it is not narrowed to a lone `none`, which would
+  // offer "off" and take away the switch the catalog just described.
+  if (levels.size === 0) return undefined
+  if (toggle) levels.add('none')
+  const ordered = REASONING_EFFORTS.map((effort) => effort.id).filter((id) => levels.has(id))
+  return ordered.length > 0 ? ordered : undefined
+}
+
 /**
  * List the gateway's own catalog ids.
  * @returns the model ids; an empty array when the call fails or returns nothing.
@@ -307,11 +353,56 @@ export async function fetchOfficialModels({ signal, timeoutMs = 30000, fetchImpl
       }
     }
   } catch { /* source unavailable */ }
+  // The upstream's own catalog, matched by the id's tail: `cline-pass/glm-5.3` and
+  // `zai/glm-5.3` are the same model reached through the same gateway. It is also
+  // the only source that describes a model published after this release — Cline Pass
+  // lists ids, names and descriptions, and says nothing about what a model can do,
+  // so without this a new model falls back to the route's conservative window.
+  //
+  // Its output cap is deliberately NOT taken. It disagrees with the published one
+  // (32k where the shipped table says 384k) and it moves: the same model reported a
+  // different figure a week later. Adopting it would silently resize every request's
+  // completion reserve, which is what decides where compaction fires.
+  const upstreamByTail = new Map()
+  try {
+    const response = await fetchImpl(VERCEL_MODELS_URL, { signal: withTimeout(signal, timeoutMs) })
+    if (response.ok) {
+      const json = await response.json()
+      const rows = Array.isArray(json?.data) ? json.data : Array.isArray(json?.models) ? json.models : []
+      for (const row of rows) {
+        if (typeof row?.id !== 'string') continue
+        // The shared extractor already reads the display name and `modalities.input`,
+        // and it deliberately misses this catalog's snake_case output cap — which is
+        // exactly the field that must not be adopted.
+        const meta = modelMetadata(row)
+        const contextWindow = row?.context_window ?? row?.contextWindow
+        if (Number.isSafeInteger(contextWindow) && contextWindow > 0) meta.contextWindow = contextWindow
+        const efforts = effortsFromOptions(row.reasoning_options)
+        if (efforts !== undefined) meta.efforts = efforts
+        if (Object.keys(meta).length > 0) upstreamByTail.set(row.id.split('/').pop(), meta)
+      }
+      if (upstreamByTail.size > 0) sources.push('vercel')
+    }
+  } catch { /* source unavailable */ }
   const models = [...found.keys()].filter((id) => /^cline-pass\/[a-z0-9._-]+$/.test(id))
   const catalog = {}
   for (const id of models) {
-    const meta = found.get(id)
-    if (meta !== undefined && Object.keys(meta).length > 0) catalog[id] = meta
+    const meta = found.get(id) ?? {}
+    const upstream = upstreamByTail.get(id.slice('cline-pass/'.length))
+    if (upstream === undefined) {
+      if (Object.keys(meta).length > 0) catalog[id] = meta
+      continue
+    }
+    // Fills what the other sources leave out, so an id list that carries only a name
+    // still yields a real window, real levels and real modalities. A source that did
+    // describe a field keeps it: this is a gap filler, not an override.
+    const entry = { ...upstream, ...meta }
+    // The exception is the name, where the id list's "name" is the id itself and
+    // says nothing — the upstream publishes the label a person reads.
+    if (upstream.name !== undefined && (entry.name === undefined || entry.name === id || entry.name === id.slice('cline-pass/'.length))) {
+      entry.name = upstream.name
+    }
+    catalog[id] = entry
   }
   return { models, catalog, sources }
 }

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "dsh-cline-pass",
-  "version": "0.2.8",
+  "version": "0.2.9",
   "description": "Self-contained Cline Pass provider for dsh: register Cline Pass subscription models on the harness LLM seam, enumerate and validate the upstream channels behind each model, pin or exclude them with failover, and drive an account pool. No proxy process and no external cline-pass-switcher install required.",
   "type": "module",
   "main": "lib/index.js",