dsh-vision-router 3.0.5 → 3.1.0
This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.
- package/README.md +8 -21
- package/README.zh.md +8 -21
- package/docs/architecture/model-catalog-ownership.md +41 -0
- package/docs/development/dsh-rc3-adaptation.md +29 -0
- package/docs/doctor.md +12 -0
- package/docs/free-models.zh-CN.md +5 -162
- package/docs/http-provider-compatibility.md +5 -3
- package/docs/releases/v3.1.0.md +34 -0
- package/index.js +21 -0
- package/lib/catalog-corrections.js +67 -8
- package/lib/doctor-cli.js +66 -4
- package/lib/http-compat.js +37 -4
- package/lib/live-model-discovery.js +10 -7
- package/lib/settings-ia-client-prelude.js +89 -1
- package/lib/trusted-vision-hints.js +5 -4
- package/package.json +4 -4
- package/presets/README.md +5 -32
- package/presets/dashscope.yaml +0 -18
- package/presets/openrouter.yaml +0 -17
- package/presets/ovh.yaml +0 -14
- package/presets/siliconflow.yaml +0 -18
- package/presets/zhipu.yaml +0 -18
package/README.md
CHANGED
|
@@ -22,7 +22,7 @@
|
|
|
22
22
|
> **Data leaves your machine by default.** When a cloud vision model is used, Vision Router sends the image (or a derived crop), the vision prompt and related request metadata to that provider. A fresh install has an anonymous OVHcloud vision fallback enabled, so "free" and "no key" do **not** mean offline. Do not use the cloud chain for confidential, regulated or classified material. See [Data flow and strict local-only use](#data-flow-and-strict-local-only-use).
|
|
23
23
|
|
|
24
24
|
> [!WARNING]
|
|
25
|
-
> **v3.0
|
|
25
|
+
> **v3.1.0 release notes:** simpler image-input setup and clear cloud-upload privacy controls, plus DSH compatibility, HTTP and dependency fixes. See [v3.1.0 release notes](docs/releases/v3.1.0.md). For the installable version, check [GitHub Releases](https://github.com/ysr666/dsh-vision-router/releases/latest) or npm for the installable version.
|
|
26
26
|
|
|
27
27
|
## Start here
|
|
28
28
|
|
|
@@ -61,7 +61,7 @@ These are ordinary chat requests: your normal model reasons through the task and
|
|
|
61
61
|
- [Tools](#tools)
|
|
62
62
|
- [Quick start](#quick-start)
|
|
63
63
|
- [Data flow and strict local-only use](#data-flow-and-strict-local-only-use)
|
|
64
|
-
- [
|
|
64
|
+
- [Configure your own vision models](#configure-your-own-vision-models)
|
|
65
65
|
- [Why this exists](#why-this-exists)
|
|
66
66
|
- [How it compares](#how-it-compares)
|
|
67
67
|
- [Configuration](#configuration)
|
|
@@ -232,26 +232,13 @@ Vision Router does not operate fully offline in its default configuration. The e
|
|
|
232
232
|
|
|
233
233
|
To stop **Vision Router itself** from sending image data to remote vision endpoints, enable **Settings → Vision Router → General → Local-only vision**. This is a runtime policy, not a destructive rewrite: saved cloud rows and `freeFallback` stay configured but cannot execute while the switch is on. Only loopback visual endpoints (`localhost`, `127.0.0.0/8`, `::1`) are eligible, including Ollama, LM Studio, and custom local HTTP backends. For a **strict local-only workflow**, also use a local chat model because vision result text still returns to the current chat model. Test the final network boundary in your own environment; Vision Router cannot make a remote chat model, proxy or Host integration local.
|
|
234
234
|
|
|
235
|
-
##
|
|
235
|
+
## Configure your own vision models
|
|
236
236
|
|
|
237
|
-
|
|
237
|
+
**DSH owns providers, credentials and the global model catalog; Vision Router uses configured Host models for visual orchestration, execution, testing and diagnostics.** No plugin-specific commercial vendor YAML is required. Vision Router can privately supplement its own backend picker with live `/models` evidence for **already configured** Host providers and endpoint-scoped image-capability hints; these never register a new provider or mutate DSH's global model directory. Cached/saved-only IDs do not themselves authorize a direct model bridge. See [architecture provenance](docs/architecture/model-catalog-ownership.md).
|
|
238
238
|
|
|
239
|
-
|
|
240
|
-
|---|---|---|---|---|
|
|
241
|
-
| OVHcloud AI Endpoints (access key) | `Qwen2.5-VL-72B-Instruct` — the same endpoint the built-in fallback uses | **400 req/min** per project per model (vs 2 anonymous) | ✅ | OVH account → Public Cloud project (attach a payment method; free models are not charged) → AI Endpoints access key |
|
|
242
|
-
| Zhipu (bigmodel.cn) | `glm-4.6v-flash` · `glm-4.1v-thinking-flash` · `glm-4v-flash` — three permanently free models; chaining them triples capacity | uncapped tokens | ✅ | open.bigmodel.cn → API keys |
|
|
243
|
-
| DashScope (Aliyun) | `qwen3-vl-flash` (limited-time free) and the Qwen-VL series | new users: 1M tokens per model series / 90 days | ✅ | bailian.console.aliyun.com |
|
|
244
|
-
| Intern AI (Shanghai AI Lab) | `internvl-latest` · `internvl3.5-latest` | 30 RPM, **90M tokens/month** | ✅ | chat.intern-ai.org.cn |
|
|
245
|
-
| Groq | `meta-llama/llama-4-scout-17b-16e-instruct` (native multimodal, up to 5 images) | 30 RPM / 14,400 req/day, no card | ❌ proxy | console.groq.com |
|
|
246
|
-
| Google AI Studio | `gemini-2.5-flash` · `gemini-2.5-flash-lite` | 10–30 RPM / 500–1,500 req/day | ❌ proxy | aistudio.google.com |
|
|
247
|
-
| NVIDIA NIM | `meta/llama-3.2-11b-vision-instruct` · `nvidia/nemotron-nano-12b-v2-vl` | 40 RPM, no card | ⚠️ | build.nvidia.com |
|
|
248
|
-
| OpenCode Zen | `mimo-v2.5-free` (vision + code) | 30 RPM / 500 req/day | ⚠️ | opencode.ai/zen |
|
|
249
|
-
| OpenRouter | `google/gemma-4-26b-a4b-it:free` · `google/gemma-4-31b-it:free` | 50 req/day on unpaid accounts | ❌ proxy | openrouter.ai |
|
|
239
|
+
Manage provider credentials and models in **DSH Settings → Models**, then pick those models from the same Host catalog in **Settings → Vision Router**. Use **Test Vision** to verify that the exact model actually reads image content. If Host metadata incorrectly marks a real vision model text-only, prefer DSH **Model input types → Image**; on older Hosts, declare `input: [text, image]` on a **verified image-capable** `llm-pi-ai` model and reload configuration. Never fabricate image support for text-only models.
|
|
250
240
|
|
|
251
|
-
|
|
252
|
-
|
|
253
|
-
> [!NOTE]
|
|
254
|
-
> Free-tier policies change without notice — Cerebras retired its free tier in July 2026 (now a one-time $5 credit), SambaNova's free tier is down to 20 requests/day, and Hugging Face's is $0.10/month. Third-party “`:free` relay” aggregators are deliberately not listed: they rotate quickly, lack SLAs, and some resell quota in ways that violate upstream terms.
|
|
241
|
+
The built-in anonymous OVH fallback is a separate Vision Router `httpProviders` runtime and not a DSH provider preset. Strictly local workflows also require a local chat model and the Local-only vision setting. Vendor price, quota and model IDs change; this project does not keep an ongoing marketing or free-tier list. The [August 2026 survey](docs/free-models.zh-CN.md) remains available solely as historical research.
|
|
255
242
|
|
|
256
243
|
## Why this exists
|
|
257
244
|
|
|
@@ -302,7 +289,7 @@ The vision tools try backends in order and surface an error only after all of th
|
|
|
302
289
|
2. **Local Ollama (optional, off by default)**: `localOllama.enabled` adds keyless, offline recognition through your local Ollama (for example qwen2.5vl);
|
|
303
290
|
3. **Local LM Studio (optional, off by default)**: `localLmStudio.enabled` follows Ollama and requires the real model identifier shown in LM Studio Developer or returned by `/v1/models`;
|
|
304
291
|
4. **Advanced custom HTTP vision endpoints**: legacy/advanced `httpProviders`, when present, run after the local backends;
|
|
305
|
-
5. **Built-in anonymous OVH fallback**: always last and never exposed in a model picker. The current quality-first chain is `Qwen3.5-397B-A17B` → `Qwen2.5-VL-72B-Instruct` → `Qwen3.6-27B` → `Qwen3.5-9B`. OVH anonymous limits are **2 requests/minute per IP per model**. The four models have independent buckets, so spreading requests across them is about **8 RPM in theory**, subject to OVH's actual rate limiting. No signup or API key is required. OVH has scheduled the previously bundled Mistral Small 3.2 model for decommission on November 4, 2026; it is no longer part of the automatic fallback, but user-managed rows are not changed.
|
|
292
|
+
5. **Built-in anonymous OVH fallback**: always last and never exposed in a model picker. The current quality-first chain is `Qwen3.5-397B-A17B` → `Qwen2.5-VL-72B-Instruct` → `Qwen3.6-27B` → `Qwen3.5-9B`. OVH anonymous limits are **2 requests/minute per IP per model**. The four models have independent buckets, so spreading requests across them is about **8 RPM in theory**, subject to OVH's actual rate limiting. No signup or API key is required. OVH has scheduled the previously bundled Mistral Small 3.2 model for decommission on November 4, 2026; it is no longer part of the automatic fallback, but user-managed rows are not changed. For more capacity, configure your own model in DSH; see [Configure your own vision models](#configure-your-own-vision-models).
|
|
306
293
|
|
|
307
294
|
> [!IMPORTANT]
|
|
308
295
|
> This “vision chain” is the **eyes** used by Vision Router: each settings row selects one user vision model, while the lower-right chat picker selects the **brain/conversation model**. The two are deliberately separate. Text-only DeepSeek/opencode models are filtered out of the vision-backend dropdown, and the internal `Vision HTTP` transport route is no longer exposed to users.
|
|
@@ -622,7 +609,7 @@ exploration (LLM7.io and OVHcloud anonymous tiers) by
|
|
|
622
609
|
|
|
623
610
|
## Release notes and community
|
|
624
611
|
|
|
625
|
-
Release details: [v3.0
|
|
612
|
+
Release details: [v3.1.0 release notes](docs/releases/v3.1.0.md). Check GitHub Releases or npm for the published version.
|
|
626
613
|
|
|
627
614
|
<p align="center">
|
|
628
615
|
<sub>Ecosystem:</sub>
|
package/README.zh.md
CHANGED
|
@@ -22,7 +22,7 @@
|
|
|
22
22
|
> **默认配置会让数据出网。** 使用云端视觉模型时,Vision Router 会将图片(或由它裁剪出的局部)、识图提示词及相关请求元数据发送给对应供应商。新安装默认开启 OVHcloud 匿名视觉兜底,因此“免费”和“免 Key”不等于离线。涉密、涉案、内部或其他受限材料请勿使用云端链路。详见[数据流向与严格纯本地配置](#数据流向与严格纯本地配置)。
|
|
23
23
|
|
|
24
24
|
> [!WARNING]
|
|
25
|
-
> **v3.0
|
|
25
|
+
> **v3.1.0 更新说明:** 简化识图模型设置,增加云端出网关闭指引与「不再提醒」,并修正 DSH 兼容、HTTP 和依赖安全问题。参见[v3.1.0 更新说明](docs/releases/v3.1.0.md)。实际可安装版本请以 [GitHub Releases](https://github.com/ysr666/dsh-vision-router/releases/latest)或 npm 为准。
|
|
26
26
|
|
|
27
27
|
## 先用起来
|
|
28
28
|
|
|
@@ -61,7 +61,7 @@ npx @deepseek-ai/dsh plugin --profile web add dsh-vision-router
|
|
|
61
61
|
- [工具](#工具)
|
|
62
62
|
- [快速开始](#快速开始)
|
|
63
63
|
- [数据流向与严格纯本地配置](#数据流向与严格纯本地配置)
|
|
64
|
-
- [
|
|
64
|
+
- [配置自己的视觉模型](#配置自己的视觉模型)
|
|
65
65
|
- [为什么做这个](#为什么做这个)
|
|
66
66
|
- [对比同类插件](#对比同类插件)
|
|
67
67
|
- [配置项](#配置项)
|
|
@@ -232,26 +232,13 @@ Vision Router 在默认配置下不是完全离线工具。具体边界取决于
|
|
|
232
232
|
|
|
233
233
|
若要阻止 **Vision Router 本身**把图片发送给远程视觉端点,请开启 **设置 → Vision Router → 常规 → 仅本地视觉**。这是运行时策略,不会破坏性修改配置:已保存的云端识图行和 `freeFallback` 会保留,但开关开启期间无法执行;只有回环地址(`localhost`、`127.0.0.0/8`、`::1`)上的视觉端点可以运行,包括 Ollama、LM Studio 和自定义本机 HTTP 后端。若要求**整个工作流严格纯本地**,还必须同时使用本地聊天模型,因为识图结果文本仍会交给当前聊天模型。请在自己的环境中验证最终网络边界;Vision Router 无法把远程聊天模型、代理或 Host 集成变成本地服务。
|
|
234
234
|
|
|
235
|
-
##
|
|
235
|
+
## 配置自己的视觉模型
|
|
236
236
|
|
|
237
|
-
|
|
237
|
+
**供应商、凭据和全局模型目录由DSH负责;Vision Router使用DSH已配置模型,负责视觉编排、执行、测试与诊断。** 无需安装插件专属的厂商YAML。对**已在DSH配置**的供应商,插件可在自己的视觉后端选择器内补充实时`/models`发现结果及严格限定端点的视觉能力提示;这些内容不会新增供应商或修改DSH全局目录,过期缓存或仅保存的模型ID也不能单独授权直连。参见[历史架构依据](docs/architecture/model-catalog-ownership.md)。
|
|
238
238
|
|
|
239
|
-
|
|
240
|
-
|---|---|---|---|---|
|
|
241
|
-
| OVHcloud AI Endpoints(access key) | `Qwen2.5-VL-72B-Instruct`——与内置兜底同一个端点 | **400 次/分钟**/项目/模型(对比匿名 2 次/分钟) | ✅ | 注册 OVH 账号 → Public Cloud 项目(需挂支付方式;免费模型不扣费)→ AI Endpoints access key |
|
|
242
|
-
| 智谱(bigmodel.cn) | `glm-4.6v-flash` · `glm-4.1v-thinking-flash` · `glm-4v-flash`——三个永久免费模型,串起来容量 ×3 | token 不限量 | ✅ | open.bigmodel.cn → API keys |
|
|
243
|
-
| 阿里云百炼 | `qwen3-vl-flash`(限免)与 Qwen-VL 系列 | 新用户每模型系列 100 万 token / 90 天 | ✅ | bailian.console.aliyun.com |
|
|
244
|
-
| Intern AI(上海AI实验室) | `internvl-latest` · `internvl3.5-latest` | 30 RPM,**9000 万 token/月** | ✅ | chat.intern-ai.org.cn |
|
|
245
|
-
| Groq | `meta-llama/llama-4-scout-17b-16e-instruct`(原生多模态,最多 5 张图) | 30 RPM / 14,400 次/天,免卡 | ❌ 需代理 | console.groq.com |
|
|
246
|
-
| Google AI Studio | `gemini-2.5-flash` · `gemini-2.5-flash-lite` | 10–30 RPM / 500–1,500 次/天 | ❌ 需代理 | aistudio.google.com |
|
|
247
|
-
| NVIDIA NIM | `meta/llama-3.2-11b-vision-instruct` · `nvidia/nemotron-nano-12b-v2-vl` | 40 RPM,免卡 | ⚠️ | build.nvidia.com |
|
|
248
|
-
| OpenCode Zen | `mimo-v2.5-free`(视觉 + 代码) | 30 RPM / 500 次/天 | ⚠️ | opencode.ai/zen |
|
|
249
|
-
| OpenRouter | `google/gemma-4-26b-a4b-it:free` · `google/gemma-4-31b-it:free` | 未充值账户 50 次/天 | ❌ 需代理 | openrouter.ai |
|
|
239
|
+
在DSH「设置 → 模型」维护供应商、模型与密钥,在Vision Router视觉后端列表中选择同一Host目录里的模型,并用「测试识图」核实精确模型确实读取图片。若Host把真正的视觉模型误报为仅文本,优先使用DSH模型设置的「输入类型 → 图片」;旧Host可在**确认模型确实接受图片**后配置`llm-pi-ai`的`input: [text, image]`,再重载配置,不得对纯文本模型伪造能力。
|
|
250
240
|
|
|
251
|
-
|
|
252
|
-
|
|
253
|
-
> [!NOTE]
|
|
254
|
-
> 免费政策随时可能调整——Cerebras 已在 2026 年 7 月取消免费档(改为一次性 $5 赠金),SambaNova 免费档收紧到 20 次/天,Hugging Face 只剩 $0.10/月。第三方“`:free` 中转”聚合站刻意不列入:轮换频繁、无 SLA,部分还存在违反上游条款的转售行为。
|
|
241
|
+
内置OVH匿名兜底属于Vision Router工具的`httpProviders`运行时,不是DSH厂商目录条目。严格纯本地使用需要同时使用本地聊天模型并开启「仅本地视觉」。商用API的价格、免费额度及模型ID动态变化,本项目不再维护推荐表。[2026年8月历史调研](docs/free-models.zh-CN.md)仅供追溯。
|
|
255
242
|
|
|
256
243
|
## 为什么做这个
|
|
257
244
|
|
|
@@ -301,7 +288,7 @@ Vision Router 在默认配置下不是完全离线工具。具体边界取决于
|
|
|
301
288
|
2. **本地 Ollama(可选,默认关)**:`localOllama.enabled` 开启后,通过本机 Ollama 做免 Key、离线识别(例如 qwen2.5vl);
|
|
302
289
|
3. **本地 LM Studio(可选,默认关)**:`localLmStudio.enabled` 排在 Ollama 之后,模型名必须填写 LM Studio Developer 页或 `/v1/models` 返回的真实标识;
|
|
303
290
|
4. **高级自定义 HTTP 视觉端点**:旧配置/高级配置中的 `httpProviders` 排在本地后端之后;
|
|
304
|
-
5. **内置 OVH 匿名免费兜底**:固定最后尝试,不需要出现在任何模型选择器里。当前内置链按质量优先为 `Qwen3.5-397B-A17B` → `Qwen2.5-VL-72B-Instruct` → `Qwen3.6-27B` → `Qwen3.5-9B`。OVH 匿名限额为 **每 IP、每模型 2 次/分钟**;4 个模型是独立限额,因此理论上分散请求可到约 **8 次/分钟**,实际仍以 OVH 当时的限流为准。OVH 已公告原有 Mistral Small 3.2 模型将在 2026 年 11 月 4 日下线,因此已从自动兜底移除,但不会修改用户手动保存的配置。免注册、免 Key
|
|
291
|
+
5. **内置 OVH 匿名免费兜底**:固定最后尝试,不需要出现在任何模型选择器里。当前内置链按质量优先为 `Qwen3.5-397B-A17B` → `Qwen2.5-VL-72B-Instruct` → `Qwen3.6-27B` → `Qwen3.5-9B`。OVH 匿名限额为 **每 IP、每模型 2 次/分钟**;4 个模型是独立限额,因此理论上分散请求可到约 **8 次/分钟**,实际仍以 OVH 当时的限流为准。OVH 已公告原有 Mistral Small 3.2 模型将在 2026 年 11 月 4 日下线,因此已从自动兜底移除,但不会修改用户手动保存的配置。免注册、免 Key。如需额外容量,请通过DSH配置自备模型;参见[配置自己的视觉模型](#配置自己的视觉模型)。
|
|
305
292
|
|
|
306
293
|
> [!IMPORTANT]
|
|
307
294
|
> 这里的“视觉链”是 Vision Router 调用的**眼睛**:设置页里每一行只选一个用户视觉模型;聊天页右下角选择的是**脑子/会话模型**,两者完全分开。纯文本 DeepSeek / opencode 不会出现在视觉后端下拉里;内部 `Vision HTTP` 也不会再暴露给用户。
|
|
@@ -616,7 +603,7 @@ npx dsh-vision-router doctor --profile web
|
|
|
616
603
|
|
|
617
604
|
## 版本公告与社区链接
|
|
618
605
|
|
|
619
|
-
版本更新详情:[v3.0
|
|
606
|
+
版本更新详情:[v3.1.0 更新说明](docs/releases/v3.1.0.md)。正式可安装版本请以 GitHub Releases 或 npm 为准。
|
|
620
607
|
|
|
621
608
|
<p align="center">
|
|
622
609
|
<sub>生态收录:</sub>
|
|
@@ -0,0 +1,41 @@
|
|
|
1
|
+
# 模型目录所有权与预设退役|架构决策
|
|
2
|
+
|
|
3
|
+
> **DSH负责模型供应商及目录;Vision Router负责视觉能力编排、执行与诊断。只有解决真实兼容问题所需的代码或配置,才应由Vision Router长期维护。**
|
|
4
|
+
|
|
5
|
+
DSH管理用户供应商、模型、凭据、Host目录及输入模态权限;Vision Router的设置页经`connection.api.llm.models({})`读取相同目录,视觉模式用`modelDirectories.directoryFor(sessionId)`读取Host权威选择状态。Router继续管理视觉工具、像素处理、能力评估、执行、回退和诊断;不另造商用供应商注册表。Host声明的输入模态不能取代真实图片投递验证。
|
|
6
|
+
|
|
7
|
+
## 私有视觉模型发现不是第二套供应商目录(#223、#225、#227、#230)
|
|
8
|
+
|
|
9
|
+
原则中的“DSH负责目录”,指**DSH拥有供应商注册、已配置路由、全局模型目录及其最终权限**;不等于Vision Router只能展示静态DSH目录里当下已有的ID。历史需求已有真实证据:
|
|
10
|
+
|
|
11
|
+
- [PR #223](https://github.com/ysr666/dsh-vision-router/pull/223)把被误标仅文本的模型直连桥限制到DSH/pi-ai的**精确本地预线拒绝**,不会用网络、认证或其他失败当作绕过Host的理由。
|
|
12
|
+
- [PR #225](https://github.com/ysr666/dsh-vision-router/pull/225)针对DSH/pi-ai静态目录遗漏动态新型号,为**DSH已配置**的供应商端点增加独立的`/models`查询。结果只进入Vision Router自己的选择器,不写入DSH全局目录;单凭模型ID在`/models`出现不代表支持图片。
|
|
13
|
+
- [PR #227](https://github.com/ysr666/dsh-vision-router/pull/227)让用户已经保存的模型ID继续可见、可编辑;**仅保存历史不授权UNKNOWN_MODEL直连**。
|
|
14
|
+
- [PR #230](https://github.com/ysr666/dsh-vision-router/pull/230)来自官方BigModel端点`/models`漏列能真实调用的视觉型号。只对端点与协议完全匹配的官方路由添加少数视觉能力提示;缓存重启后仅可展示,必须重新验证当前路由才可授权执行。
|
|
15
|
+
|
|
16
|
+
现有实现位于`src/lib/live-model-discovery.js`、`src/lib/vision-model-registry.js`、`src/lib/trusted-vision-hints.js`及相关Browser私有预加载层。它们承担的是**对用户已配置模型进行补充发现、显示来源并核验能力**,不是在DVR注册新供应商或替代DSH的全局模型目录。属于用户可用性与兼容层的有条件例外,**不能与静态商业厂商YAML一起删除**。
|
|
17
|
+
|
|
18
|
+
未来如DSH原生提供同等的活模型发现,应逐条核对:当前支持的每个Host是否覆盖静态目录缺项、模型身份和图片输入真实性、缓存失效、已保存项显示、凭据权限、未知型号失败闭合及浏览器只在插件私域展示。所有这些条件满足后,再独立讨论退役此例外。
|
|
19
|
+
|
|
20
|
+
## 逐项来源核查(清理前)
|
|
21
|
+
|
|
22
|
+
| 来源 | 当初的目的 | 本轮结论 |
|
|
23
|
+
| --- | --- | --- |
|
|
24
|
+
| [提交16e29d07](https://github.com/ysr666/dsh-vision-router/commit/16e29d07b943d2b2e1dfb615683ab07f24cc96af),2026-08-13(无关联PR) | 降低用户手动写`llm-pi-ai`配置的门槛,添加五份供应商YAML和免费模型调研。 | 退役五份静态示例;原始报告可通过Git历史追溯。 |
|
|
25
|
+
| [提交cf4eb6b3](https://github.com/ysr666/dsh-vision-router/commit/cf4eb6b35fe1005dcf9fd46f840a6688ef60372e) | `zai.yaml`改名为`zhipu.yaml`,避免自定义bigmodel.cn provider key与内置`zai`冲突。 | 不删除已有用户的手动配置,只退役仓库示例。 |
|
|
26
|
+
| [PR #89](https://github.com/ysr666/dsh-vision-router/pull/89),2026-08-16 | DSH Web自定义视觉模型未保存`image`能力声明,曾触发下拉过滤及难以理解的空列表。 | **保留通用图片能力修复指引**和实测提示,不延续过时的“全部隐藏”描述。 |
|
|
27
|
+
| [PR #676](https://github.com/ysr666/dsh-vision-router/pull/676),2026-10-09 | DSH 0.2.1-alpha.2变更Web信任初始化所有权,新增`presets/web-connection-ready/`私有运行时桥接。 | **不能删**,仍由`cordis.patch.yml`加载、`package.json.files`打包,必须保留Host跨代回归门。 |
|
|
28
|
+
| [PR #729](https://github.com/ysr666/dsh-vision-router/pull/729),2026-10-19 | 提交收费Atlas Cloud网关示例及文档;贡献者披露商业关联。 | 按同一职责原则关闭而未合并,不针对特定厂商区别对待。 |
|
|
29
|
+
|
|
30
|
+
## 清理边界
|
|
31
|
+
|
|
32
|
+
只删除仓库发布包中需人工复制的`dashscope.yaml`、`zhipu.yaml`、`siliconflow.yaml`、`openrouter.yaml`、`ovh.yaml`。已有用户的`settings.yaml`、凭据、模型注册一概不动。中英文README改为DSH原生模型配置路径;原动态价格清单仅由不可变历史快照承载。
|
|
33
|
+
|
|
34
|
+
**必须保留:** `presets/web-connection-ready/`、`cordis.patch.yml`引用、`package.json`中的`presets`打包项、`src/lib/core-primitives.js`运行时匿名OVH兜底;`src/lib/http-compat.js`的协议兼容预设及catalog corrections不能与手工厂商YAML混淆。后续清理每条真正运行时兼容规则前,也必须查Issue/PR、复现原因、受支持Host替代能力及负向测试。
|
|
35
|
+
|
|
36
|
+
## 验收门
|
|
37
|
+
|
|
38
|
+
- `pnpm test`、`pnpm typecheck`、`pnpm build`及`pnpm pack`。tarball须含`presets/web-connection-ready/index.mjs`和`package.json`,不含五份厂商YAML。
|
|
39
|
+
- 原有Windows/macOS/Linux、Node22/24与支持窗口内的DSH稳定版/preview Host门不得回归。
|
|
40
|
+
- DSH模型目录选择、#89图片能力提示、真实识图校验、运行时OVH兜底与连接信任权限必须保持原状。
|
|
41
|
+
- 真正移除`web-connection-ready`或供应商协议兼容逻辑,另开PR证明上游问题已不再存在,并提供回滚策略。
|
|
@@ -0,0 +1,29 @@
|
|
|
1
|
+
# DSH 0.2.1-rc.3 next-channel compatibility qualification
|
|
2
|
+
|
|
3
|
+
- **Channel:** next / RC only; not stable, and not alpha.
|
|
4
|
+
- **Exact upstream tag:** `dsh-v0.2.1-rc.3`.
|
|
5
|
+
- **Immutable upstream commit:** `1f81c047fc48fa190a683e469daa699857e9afe8`.
|
|
6
|
+
- **GitHub release:** 2026-10-23 15:26:39 UTC (2026-10-23 23:26:39 Asia/Shanghai).
|
|
7
|
+
- **Previous qualified next:** RC2 `687f282e8625c6b8d11ff58211a90524be2e4317`, covered by merged PR #724.
|
|
8
|
+
- **Baseline DVR:** latest `main` at branch creation (`6bce1d87bacb472b84e30b281ee573d9862d679e`).
|
|
9
|
+
- **Upstream compare:** https://github.com/deepseek-ai/deepseek-harness/compare/dsh-v0.2.1-rc.2...dsh-v0.2.1-rc.3
|
|
10
|
+
|
|
11
|
+
## Upstream changes and potential compatibility impact
|
|
12
|
+
|
|
13
|
+
Six upstream commits: navigation-entry customization (including More menus and pinned launcher state), navigation/preset analytics, menu and tooltip fixes, official DeepSeek API-key model listing refresh, release metadata, and release merge. The published RC3 notes explicitly call out More menus, restoration of an unsent New Session draft, and API-key adapter model list parity with account login.
|
|
14
|
+
|
|
15
|
+
| Surface | RC3 review / regression focus |
|
|
16
|
+
| --- | --- |
|
|
17
|
+
| Host / peer admission | Pin exact source and run official Host evaluator. Do not change DVR `engines.dsh`, optional peer ranges, or the historical minimum supported Host. |
|
|
18
|
+
| Session | New Session now reuses an eligible blank session; verify Session identity, draft/attachment ownership, cold resume and lifecycle boundaries. |
|
|
19
|
+
| Settings | New entry pin persistence and provider model listing must not alter DVR Settings mount, save, reload or readback. |
|
|
20
|
+
| Bundle | `web-app/cordis.patch.yml` adds `ui-entry-customization`; verify DVR overlay ordering and three actual recompositions. |
|
|
21
|
+
| Web/client | New `sidebar.panellist.actions`, `conversation.view.actions`, and guide actions plus pinned state/menus; exercise real Chromium, Vision toggle and settings UI. |
|
|
22
|
+
| Model/attachments | API-key official `/models` discovery changes model catalog freshness; preserve explicit image-capability checks and mixed attachment routing, not model-name assumptions. |
|
|
23
|
+
| Service injection | New UI plugin and model-listing services must not interfere with existing DVR client inject, Web/connection overlay, Desktop, or Host-Sharp lifecycle. |
|
|
24
|
+
|
|
25
|
+
## Qualification gates (not yet claimed passed)
|
|
26
|
+
|
|
27
|
+
The dedicated `.github/workflows/dsh-021-rc3-validation.yml` duplicates the proven RC2 matrix but pins RC3's **exact SHA** and retains RC2, alpha and stable lanes separately. It must execute on an ordinary PR, including Node 22/24, Linux/macOS/Windows exact-source and Host-Sharp, Windows Node 22/24 + macOS Desktop, real Web + Chromium cold Vision, mixed attachments, Settings mount/save/reload/readback, three bundle recomposes, source/peer/proxy contracts and existing adversarial/security regression suites.
|
|
28
|
+
|
|
29
|
+
**State:** Await final-head CI and actual Host/Chromium evidence. Any failure must be investigated and fixed at the root before merge. No Copilot review gate; no workflow_dispatch substitute; no automatic DVR release.
|
package/docs/doctor.md
CHANGED
|
@@ -112,6 +112,18 @@ npx dsh-vision-router doctor --profile web --sessions --json
|
|
|
112
112
|
|
|
113
113
|
The JSON report is schema-versioned and includes the running Doctor/Vision Router version. It is intentionally minimized for issue reports: it does not contain the raw profile manifest, raw log lines, API keys, raw dependency specs, local dependency paths, URL credentials/query strings/fragments, session ids, or raw runtime error text. It keeps only structured diagnostic facts such as install mode, installed version, route status, failure counts, optional capability versions and known session-repair kinds.
|
|
114
114
|
|
|
115
|
+
### Paste-ready GitHub Issue template
|
|
116
|
+
|
|
117
|
+
For a short, **read-only** troubleshooting report you can paste into a GitHub Issue without exporting raw logs or your DSH profile, run:
|
|
118
|
+
|
|
119
|
+
```sh
|
|
120
|
+
npx dsh-vision-router doctor --profile web --issue-template
|
|
121
|
+
```
|
|
122
|
+
|
|
123
|
+
Add `--no-runtime` for offline checks and `--sessions` only if historical session corruption is relevant. The template contains a whitelist of structural facts (Doctor version, OS, Node, installation-mode counts, coarse runtime status and failure counts) and blank fields for the exact provider/model, HTTP status and reproduction steps. It **does not send anything** or include local paths, profile names, raw runtime URLs, settings values, credentials, raw logs or images. Always review your filled-in answers before sharing.
|
|
124
|
+
|
|
125
|
+
`--issue-template` is intentionally incompatible with `--fix` and `--json`; use normal `repair` separately if needed.
|
|
126
|
+
|
|
115
127
|
## Repair the UTF-8 BOM startup failure
|
|
116
128
|
|
|
117
129
|
If DSH fails before plugins can load with an error such as:
|
|
@@ -1,166 +1,9 @@
|
|
|
1
|
-
|
|
2
|
-
# dsh-vision-router 免费/内置视觉模型方案调研报告
|
|
1
|
+
# 免费视觉API调研(2026年8月)——历史资料
|
|
3
2
|
|
|
4
|
-
>
|
|
3
|
+
> 本文件仅为历史来源索引,**不再是当前厂商、模型、额度或价格推荐表**。原报告随[2026-08-13提交16e29d07](https://github.com/ysr666/dsh-vision-router/commit/16e29d07b943d2b2e1dfb615683ab07f24cc96af)产生,帮助当时需要手工配置多模态模型的用户;其模型ID、免费额度与地区可用性可能已过期。
|
|
5
4
|
|
|
6
|
-
|
|
5
|
+
[查看原始报告的不可变历史版本](https://github.com/ysr666/dsh-vision-router/blob/16e29d07b943d2b2e1dfb615683ab07f24cc96af/docs/free-models.zh-CN.md)
|
|
7
6
|
|
|
8
|
-
|
|
9
|
-
2. **大陆直连最省事**:**阿里云百炼 DashScope**(新用户每个模型系列 100 万 token / 90 天,含 Qwen-VL 系列,OpenAI 兼容端点)> **智谱 Z.ai / bigmodel.cn**(`glm-4.6v-flash` 永久免费视觉)> **SiliconFlow**(¥14 赠金覆盖便宜 VL)。
|
|
10
|
-
3. **完全免 key 的视觉理解 API 目前不存在**:pollinations.ai 现在只有文本模型 `openai-fast`(`vision: false`),无 vision 模型。
|
|
11
|
-
4. **插件不可内置第三方密钥**(安全 + ToS),正确做法是仓库随附 **provider 预设 yaml**(baseURL + 免费 model id + `apiKeyEnv`),用户只填一个 key。
|
|
7
|
+
当前在DSH「设置 → 模型」配置供应商、模型和凭据,再在Vision Router中选择DSH模型目录里的视觉后端,使用「测试识图」核验实际图片投递。真实视觉模型若被Host错误标为纯文本,请优先在DSH模型选项中声明图片输入;旧Host可按[PR #89](https://github.com/ysr666/dsh-vision-router/pull/89)对已验证视觉模型使用`input: [text, image]`。匿名OVH默认兜底仍是插件运行时工具后端,并未删除。
|
|
12
8
|
|
|
13
|
-
|
|
14
|
-
|
|
15
|
-
## 1. OpenRouter 免费视觉模型
|
|
16
|
-
|
|
17
|
-
**是否需要 key**:需要(用户已有 OpenRouter key 即可直接用)。
|
|
18
|
-
**免费额度/限速**:所有免费模型共享一个总配额 —— **20 请求/分钟**;免费账户 **50 请求/天**;账户历史累计充值 ≥ $10 后提升到 **最高 1,000 请求/天**(20 req/min 不变)。上游供应商可能施加更紧的限制。(来源:mintlify free-llm-api-resources / teamday.ai / OpenRouter 官方模型 API)
|
|
19
|
-
|
|
20
|
-
**当前支持 image 输入的免费模型**(来源:OpenRouter 实时 API + teamday.ai 2026-08-03 快照):
|
|
21
|
-
|
|
22
|
-
| 精确 model id | 模态(输入) | 上下文 | 最大输出 | 用途/备注 |
|
|
23
|
-
|---|---|---|---|---|
|
|
24
|
-
| `google/gemma-4-26b-a4b-it:free` | image + text + video | 262,144 | 32,768 | **通用视觉理解,首选免费 VLM** |
|
|
25
|
-
| `google/gemma-4-31b-it:free` | image + text + video | 262,144 | 32,768 | 通用视觉理解,质量更高 |
|
|
26
|
-
| `nvidia/nemotron-nano-12b-v2-vl:free` | image + text + video | 128,000 | 128,000 | 视觉语言模型,输出上限大 |
|
|
27
|
-
| `nvidia/nemotron-3-nano-omni-30b-a3b-reasoning:free` | text + audio + image + video | 256,000 | 65,536 | 全模态,含推理 |
|
|
28
|
-
| `nvidia/nemotron-3.5-content-safety:free` | text + image | 128,000 | 8,192 | ⚠️ 内容安全分类器,非通用 VLM,勿用 |
|
|
29
|
-
| `google/lyria-3-pro-preview` / `google/lyria-3-clip-preview` | text + image | 1,048,576 | 65,536 | ⚠️ 音乐/音视频类,非通用图片理解(不确定) |
|
|
30
|
-
| `openrouter/free` | text + image | 200,000 | — | 免费模型路由(可过滤 image 能力),随机选模型 |
|
|
31
|
-
|
|
32
|
-
**关于用户点名的几个系列**(结论均为“免费变体已下架”或“不存在”):
|
|
33
|
-
- **qwen-vl 系列**:`qwen/qwen-vl-plus:free` 不存在了;付费的 `qwen/qwen2.5-vl-72b-instruct`、`qwen/qwen3-vl-8b/32b/235b-*` 仍在(付费)。
|
|
34
|
-
- **llama-4 系列**:`meta-llama/llama-4-scout:free` 不存在;`meta-llama/llama-4-scout`、`llama-4-maverick`(image 输入)为付费。
|
|
35
|
-
- **gemma-3 系列**:`google/gemma-3-27b-it:free` 不存在;付费 `gemma-3-4b/12b/27b-it` 仍在。免费档已升级为 **gemma-4**。
|
|
36
|
-
|
|
37
|
-
> ⚠️ 注意:OpenRouter 免费档近期频繁轮换(teamday 记录:一周内 15→14,前一周 20 个,数周内下架了 8 个)。不要把生产/长任务硬编码到某个 `:free` 端点。
|
|
38
|
-
|
|
39
|
-
**信息来源**:[OpenRouter models API](https://openrouter.ai/api/v1/models)、[teamday.ai: Best Free OpenRouter Models 2026](https://www.teamday.ai/blog/best-free-ai-models-openrouter-2026)、[mintlify: OpenRouter Free](https://mintlify.wiki/cheahjs/free-llm-api-resources/providers/free/openrouter)、[OpenRouter qwen-vl-plus:free 定价页](https://openrouter.ai/qwen/qwen-vl-plus:free/providers)
|
|
40
|
-
|
|
41
|
-
---
|
|
42
|
-
|
|
43
|
-
## 2. 其他免费视觉 API(带 key 但免费额度)
|
|
44
|
-
|
|
45
|
-
### 2.1 SiliconFlow 硅基流动(api.siliconflow.cn)—— 大陆直连 ✅
|
|
46
|
-
|
|
47
|
-
| 项 | 内容 |
|
|
48
|
-
|---|---|
|
|
49
|
-
| OpenAI 兼容 baseURL | `https://api.siliconflow.cn/v1` |
|
|
50
|
-
| 免费额度 | 新用户 **¥14 赠金**(约 2000 万 token 量级,具体以控制台为准);另有少量 `$0` 模型(当前为文本 Qwen3-8B、DeepSeek-R1-Distill-Qwen-7B、**DeepSeek-OCR** 等) |
|
|
51
|
-
| 免费视觉模型 | **不确定**:当前免费列表里未见“通用 VL”永久免费项;`DeepSeek-OCR` 是图像输入(文档 OCR,非通用理解)。历史上 `Qwen/Qwen2.5-VL-7B-Instruct` 曾免费,现需在控制台核对 |
|
|
52
|
-
| 推荐视觉 model id(便宜,赠金可覆盖) | `Qwen/Qwen2.5-VL-7B-Instruct`、`Qwen/Qwen2.5-VL-32B-Instruct`、`Qwen/Qwen3-VL-8B-Instruct`、`Qwen/Qwen3-VL-32B-Instruct` |
|
|
53
|
-
| 直连性 | 中国大陆**直连**,无需代理 |
|
|
54
|
-
|
|
55
|
-
**信息来源**:[pricepertoken: SiliconFlow Free Tier](https://pricepertoken.com/endpoints/siliconflow/free)、[yangmao.ai: SiliconFlow Free API](https://yangmao.ai/en/providers/siliconflow/free-api/)、[SiliconFlow 限速文档](https://docs.siliconflow.cn/cn/userguide/rate-limits/rate-limit-and-upgradation)
|
|
56
|
-
|
|
57
|
-
### 2.2 Z.ai / 智谱 GLM —— 大陆直连(bigmodel.cn)✅
|
|
58
|
-
|
|
59
|
-
| 项 | 内容 |
|
|
60
|
-
|---|---|
|
|
61
|
-
| OpenAI 兼容 baseURL | 国际版 `https://api.z.ai/v1`(可能需代理);大陆版 `https://open.bigmodel.cn/api/paas/v4`(直连) |
|
|
62
|
-
| 免费视觉模型(永久免费) | **`glm-4.6v-flash`(GLM-4.6V-Flash,Free)** |
|
|
63
|
-
| 限时免费视觉模型 | `glm-4.5v`(GLM-4.5V)、`glm-4.6v`(GLM-4.6V)、`glm-5v-turbo`(GLM-5V-Turbo)标记为 “Limited-time Free / 限时免费”(非永久,可能过期) |
|
|
64
|
-
| 其它视觉 model id | `glm-ocr`(OCR,极便宜 $0.03/M) |
|
|
65
|
-
| 关于 GLM-4V | 用户提到的老 **GLM-4V 已停用/被替代**,现行为 GLM-4.5V / 4.6V / 5V-Turbo |
|
|
66
|
-
| 是否需要 key | 需要(Z.ai 或 bigmodel.cn 的 API key) |
|
|
67
|
-
|
|
68
|
-
**信息来源**:[Z.AI 官方定价页](https://docs.z.ai/guides/overview/pricing)、[models.dev: Z.ai](https://models.dev/providers/zai/)、[Z.AI 概览文档](https://docs.z.ai/guides/overview/overview)
|
|
69
|
-
|
|
70
|
-
### 2.3 阿里云百炼 DashScope —— 大陆直连 ✅(综合最优)
|
|
71
|
-
|
|
72
|
-
| 项 | 内容 |
|
|
73
|
-
|---|---|
|
|
74
|
-
| OpenAI 兼容 baseURL | `https://dashscope.aliyuncs.com/compatible-mode/v1`(国际版 `https://dashscope-intl.aliyuncs.com/compatible-mode/v1`) |
|
|
75
|
-
| 免费额度 | 新用户开通后 **90 天**内,**每个模型系列 100 万 token**;覆盖「通义千问基础 / 垂直(OCR·Audio·Coder) / **多模态 Omni+VL 系列** / 第三方(DeepSeek·Kimi·GLM)」 |
|
|
76
|
-
| 视觉 model id | `qwen-vl-plus`、`qwen-vl-max`、`qwen2.5-vl-7b-instruct`、`qwen2.5-vl-72b-instruct`、`qwen3-vl-plus`、`qwen-vl-ocr`(OCR)、`qwen-omni-turbo`(全模态) |
|
|
77
|
-
| 是否需要 key | 需要(阿里云百炼 API key,控制台开通即得) |
|
|
78
|
-
| 直连性 | 中国大陆**直连** |
|
|
79
|
-
|
|
80
|
-
**信息来源**:[阿里云百炼新人免费额度(官方)](https://help.aliyun.com/zh/model-studio/new-free-quota)、[掘金:百炼新用户免费额度机制](https://juejin.cn/post/7594315259463516195)、[阿里云 Qwen-VL OpenAI 兼容(官方)](https://help.aliyun.com/zh/model-studio/qwen-vl-compatible-with-openai)
|
|
81
|
-
|
|
82
|
-
### 2.4 Groq —— 需代理,视觉可用性不确定
|
|
83
|
-
|
|
84
|
-
| 项 | 内容 |
|
|
85
|
-
|---|---|
|
|
86
|
-
| OpenAI 兼容 baseURL | `https://api.groq.com/openai/v1` |
|
|
87
|
-
| 免费额度 | **永久免费档**,限速约 **30 RPM / 6,000 TPM**(部分模型 15,000 TPM),无总量上限,免卡 |
|
|
88
|
-
| 免费模型 | Llama 3.3 70B、Llama 4 Scout 17B、Llama 4 Maverick 17B、Mixtral 8x7B、Gemma 2 9B、DeepSeek-R1-Distill-Llama-70B 等 |
|
|
89
|
-
| 视觉支持 | Groq 有 vision 支持(历史上 `llama-3.2-11b-vision-preview`、`llama-3.2-90b-vision-preview`;Llama 4 Scout/Maverick 原生多模态)。但**免费档哪个视觉模型当前可用 = 不确定**,需用 Groq 模型列表核对 |
|
|
90
|
-
| 直连性 | 中国大陆**通常需代理**,直连不稳定 |
|
|
91
|
-
|
|
92
|
-
**信息来源**:[yangmao.ai: Groq Free Tier](https://yangmao.ai/en/providers/groq/free-tier/)、[Groq 限速文档](https://console.groq.com/docs/rate-limits)、[langchain PR #34620(Groq vision 支持)](https://github.com/langchain-ai/langchain/pull/34620)
|
|
93
|
-
|
|
94
|
-
### 2.5 Cloudflare Workers AI —— 免费额度大,但接入较绕、大陆不友好
|
|
95
|
-
|
|
96
|
-
| 项 | 内容 |
|
|
97
|
-
|---|---|
|
|
98
|
-
| 免费额度 | 每个账户 **每天 10,000 免费 Neurons**(Workers Free 计划),每日 00:00 UTC 重置,免卡;超额按 $0.011/1K Neurons 计费 |
|
|
99
|
-
| 视觉模型 | `@cf/meta/llama-3.2-11b-vision-instruct`(首次使用需先向 Meta 同意 License)、Llama 4 Scout 17B 16E Instruct 等 |
|
|
100
|
-
| 认证 | 需要 **Cloudflare 账户 + Account ID + API Token(Bearer)** |
|
|
101
|
-
| OpenAI 兼容端点 | 原生 REST `https://api.cloudflare.com/client/v4/accounts/{ACCOUNT_ID}/ai/run/{MODEL}`;也有 OpenAI 兼容风格 `/ai/v1/chat/completions`(**不确定**是否与 pi-ai 完全兼容) |
|
|
102
|
-
| 直连性 | 中国大陆访问 Cloudflare API 不稳定,接入配置(Account ID + Token + License 同意)较绕 |
|
|
103
|
-
|
|
104
|
-
**信息来源**:[pricepertoken: Cloudflare Workers AI Free Tier](https://pricepertoken.com/endpoints/cloudflare/free)、[Cloudflare llama-vision 官方教程](https://github.com/cloudflare/cloudflare-docs/blob/production/src/content/docs/workers-ai/guides/tutorials/llama-vision-tutorial.mdx)
|
|
105
|
-
|
|
106
|
-
### 2.6 GitHub Models —— 免费档已收窄,需 Copilot、大陆需代理
|
|
107
|
-
|
|
108
|
-
| 项 | 内容 |
|
|
109
|
-
|---|---|
|
|
110
|
-
| OpenAI 兼容 baseURL | `https://models.github.ai/inference`(旧 `https://models.inference.ai.azure.com` 已于 2025-07 弃用) |
|
|
111
|
-
| 免费额度 | **免费档请求数很有限**;限速按 GitHub **Copilot 订阅档位**(Free / Pro / Pro+ / Business / Enterprise)递增;单请求 token 上限**极严** |
|
|
112
|
-
| 视觉模型 | `gpt-4o`、`gpt-4o-mini`、`gpt-4.1`、`gpt-4.1-mini`(均支持 image);`Llama-3.2-11B-Vision-Instruct`、`Llama-3.2-90B-Vision-Instruct`、`Llama-4-Scout`、`Llama-4-Maverick` |
|
|
113
|
-
| token 获取 | GitHub **Personal Access Token(PAT)** 或 Copilot token |
|
|
114
|
-
| 直连性 | 中国大陆**需代理** |
|
|
115
|
-
|
|
116
|
-
> ⚠️ 注意:GitHub Models 的免费/开放 API 路线在 2025 年中已被 GitHub Copilot 体系吸收,`gpt-4o` 在 GitHub Models 上被弃用(转向 gpt-4.1 系)。它不是“开箱即用免费”的首选。
|
|
117
|
-
|
|
118
|
-
**信息来源**:[GitHub Changelog:Azure endpoint 弃用](https://github.blog/changelog/2025-07-17-deprecation-of-azure-endpoint-for-github-models/)、[mintlify: GitHub Models](https://mintlify.wiki/cheahjs/free-llm-api-resources/providers/free/github-models)、[GitHub Models 迁移说明](https://www.nocode.tech/article/github-models-dies-july-30-heres-where-to-migrate-your-no-code-ai-backend)
|
|
119
|
-
|
|
120
|
-
---
|
|
121
|
-
|
|
122
|
-
## 3. 完全免 key 的视觉 API
|
|
123
|
-
|
|
124
|
-
**结论:目前不存在可稳定使用的“零注册/零密钥”视觉理解 API。**
|
|
125
|
-
|
|
126
|
-
- **pollinations.ai**:实测其 OpenAI 兼容模型列表(`https://text.pollinations.ai/openai/models`)目前**只有一个文本模型 `openai-fast`(GPT-OSS 20B,`vision: false`,`input_modalities: ["text"]`)**。它免费做的是**文生图**(`image.pollinations.ai`)和文本生成,**没有视觉理解(image input → 描述)模型**。
|
|
127
|
-
- 其余“免 key”聚合端点多为文本/生图,且**稳定性差、无 SLA、随时变更**,不适合做插件默认视觉后端。
|
|
128
|
-
|
|
129
|
-
**信息来源**:[pollinations models 端点实测](https://text.pollinations.ai/openai/models)、[pollinations README](https://github.com/learn05/pollinations-AI/blob/master/pollinator-agent/README-pollinations.md)
|
|
130
|
-
|
|
131
|
-
---
|
|
132
|
-
|
|
133
|
-
## 4. “预先内置”的落地建议(provider 预设,不内置密钥)
|
|
134
|
-
## 4. “预先内置”的落地建议(provider 预设,不内置密钥)
|
|
135
|
-
|
|
136
|
-
**原则**:绝不把第三方 key 写进仓库(安全 + 违反 ToS)。本仓库随附 [`presets/`](../presets/) 目录:
|
|
137
|
-
每个供应商一份按 llm-pi-ai 实际 schema 校正过的 `settings.yaml` 片段(`api: openai-completions` +
|
|
138
|
-
`baseURL` + 免费/便宜 model id + `apiKeyEnv`,`input: [text, image]`),用户只填一个 Key 即用。
|
|
139
|
-
|
|
140
|
-
当前内置清单:
|
|
141
|
-
|
|
142
|
-
| 预设 | 平台 | 免费视觉模型 | 直连 | 说明 |
|
|
143
|
-
|---|---|---|---|---|
|
|
144
|
-
| `presets/dashscope.yaml` | 阿里云百炼 | `qwen-vl-plus`、`qwen2.5-vl-7b-instruct` | ✅ | 新用户每系列 100 万 token/90 天(综合最优) |
|
|
145
|
-
| `presets/zai.yaml` | 智谱 bigmodel.cn | `glm-4.6v-flash` | ✅ | **永久免费** |
|
|
146
|
-
| `presets/siliconflow.yaml` | 硅基流动 | `Qwen/Qwen2.5-VL-7B/32B-Instruct` | ✅ | ¥14 赠金覆盖 |
|
|
147
|
-
| `presets/openrouter.yaml` | OpenRouter | `google/gemma-4-31b-it:free` 等 | 需代理 | 50 次/天;免费名单轮换频繁 |
|
|
148
|
-
| `presets/ovh.yaml` | OVHcloud AI Endpoints | `Qwen2.5-VL-72B-Instruct` | ✅ | **免注册免 Key**(2 次/分钟/IP);已内置为插件默认 httpProviders |
|
|
149
|
-
|
|
150
|
-
## 5. 结论与推荐顺序
|
|
151
|
-
|
|
152
|
-
① 百炼 DashScope(额度最大 + 直连)> ② 智谱 bigmodel.cn `glm-4.6v-flash`(永久免费 + 直连)>
|
|
153
|
-
③ SiliconFlow(直连 + 赠金)> ④ OpenRouter(已有 key,`gemma-4:free`,需代理)>
|
|
154
|
-
⑤ Cloudflare(额度大但接入绕)> ⑥ GitHub Models(已收窄)> ⑦ Groq(需代理,视觉可用性不确定)>
|
|
155
|
-
⑧ pollinations(无视觉理解)。
|
|
156
|
-
|
|
157
|
-
## 6. 不确定项
|
|
158
|
-
|
|
159
|
-
- SiliconFlow 通用 VL 是否永久免费:官方未明示,赠金用完后以账单为准;
|
|
160
|
-
- Groq 免费档视觉模型可用性:官方免费层主要是文本模型;
|
|
161
|
-
- Cloudflare Workers AI 的 OpenAI 兼容端点兼容性:官方端点非 OpenAI 形状,需网关转换;
|
|
162
|
-
- Z.ai 除 `glm-4.6v-flash` 外各 VL 的限时免费截止时间;
|
|
163
|
-
- OpenRouter 免费名单轮换风险:数周内已下架 `qwen-vl-plus:free`、`llama-4-scout:free`、
|
|
164
|
-
`gemma-3-27b-it:free`、`llama-3.2-11b-vision:free` 等,本节数据以报告生成当日为准。
|
|
165
|
-
|
|
166
|
-
> 所有来源链接见上文各节;免费额度、模型名称、地区可用性均可能变化,使用前请以官方页面为准。
|
|
9
|
+
见[架构决策和历史回溯](architecture/model-catalog-ownership.md)。
|
|
@@ -12,9 +12,11 @@ This means a newly released Qwen/GLM/Gemini/OpenAI-compatible model normally nee
|
|
|
12
12
|
|
|
13
13
|
## Current preset
|
|
14
14
|
|
|
15
|
-
### `glm-4v-flash*`
|
|
15
|
+
### `glm-4v-flash*` on Zhipu's official endpoint
|
|
16
16
|
|
|
17
|
-
`glm-4v-flash`
|
|
17
|
+
[Issue #53](https://github.com/ysr666/dsh-vision-router/issues/53) reproduced error 1210 (output token limit `[1,1024]`) when calling `glm-4v-flash` on `https://open.bigmodel.cn/api/paas/v4/chat/completions`. For that **exact HTTPS origin and path**, the compatibility layer proactively caps requests at 1024, preserving stricter user values.
|
|
18
|
+
|
|
19
|
+
A self-hosted or third-party gateway may use the same model ID but a different output limit. Vision Router does **not** impose the official endpoint's 1024 cap on such a DSH-configured gateway. Instead, a clear HTTP 400/422 token-limit response can still trigger the existing bounded, error-driven correction. This narrows the earlier model-name-only rule; deployments that depended on automatic 1024 capping at a custom endpoint should specify `maxTokens: 1024` explicitly or return a parseable limit error.
|
|
18
20
|
|
|
19
21
|
A stricter user setting is preserved:
|
|
20
22
|
|
|
@@ -27,7 +29,7 @@ httpProviders:
|
|
|
27
29
|
maxTokens: 768
|
|
28
30
|
```
|
|
29
31
|
|
|
30
|
-
|
|
32
|
+
On the official endpoint, the request remains at 768. If `maxTokens` is 4096 (including the generic default), the GLM preset sends 1024; it does not rewrite an unrelated endpoint.
|
|
31
33
|
|
|
32
34
|
## Image prompt compatibility
|
|
33
35
|
|
|
@@ -0,0 +1,34 @@
|
|
|
1
|
+
# v3.1.0 — release notes / 更新说明
|
|
2
|
+
|
|
3
|
+
These notes describe the v3.1.0 source changes. The published package and its provenance must be verified separately through the protected release workflow, GitHub Releases and npm.
|
|
4
|
+
|
|
5
|
+
本文描述 v3.1.0 源码改动。实际已发布的包体及来源证明需通过受保护流程、GitHub Releases 和 npm 另行核实。
|
|
6
|
+
|
|
7
|
+
## Simpler setup / 更容易使用
|
|
8
|
+
|
|
9
|
+
- **#754**: The Vision Router General page no longer treats merely choosing a model as proof that images can reach it. When DSH declares a model text-only, a concise bilingual explanation points to DSH Settings → Models → Image input, with a Recheck action and optional collapsed legacy Host help. The plugin never silently changes Host model metadata, and enabling a flag does not grant multimodal capabilities to a text-only model.
|
|
10
|
+
- **#754**:不再把“选中了模型”误报成“肯定能看图”。DSH 明确声明纯文本时,默认界面用简短中文/英文提示用户在 DSH「设置 → 模型」开启图片输入,并提供「重新检测」;旧版 YAML 字段折叠在可选帮助中。插件不会擅自修改 Host 模型配置,也不会把纯文本模型变为多模态模型。
|
|
11
|
+
- **#756**: Cloud-image privacy guidance now tells the user exactly how to stop **Vision Router image uploads**: on the DSH machine, enable Local-only vision in Vision Router → General and **Save**. A browser-local “Do not remind again” preference minimizes repeated explanation, while retaining a short egress status and a “How to stop uploads” control. Dismissing the notice **does not consent to uploads or change settings**. Unsaved changes never claim that the local-only policy has taken effect. A cloud chat model can still receive **derived vision-result text**; fully local use also requires selecting a local chat model.
|
|
12
|
+
- **#756**:清楚告诉用户在运行 DSH 的电脑上打开 Vision Router「常规」,开启「仅本地视觉」并**保存**,即可阻止**本插件向云端视觉服务上传图片**。可勾选「不再提醒」,该浏览器仍保留一行提醒和「查看如何关闭出网」入口;收起提醒**不等于授权上传,也不会修改配置**。未保存的草稿不会误报为已生效。如果聊天模型在云端,**识图结果文字仍可能出网**;完全本地使用还需在聊天页选择本地聊天模型。
|
|
13
|
+
|
|
14
|
+
## Compatibility, fixes and maintenance / 兼容与修复
|
|
15
|
+
|
|
16
|
+
- **#751**: Qualified DSH `0.2.1-rc.3` against exact upstream source and real Host/Chromium checks, without raising the supported DSH/Node floor. / 固定上游精确源码与真实 Host/Chromium 验收 `0.2.1-rc.3`,不抬高 DSH/Node 最低支持要求。
|
|
17
|
+
- **#745, #747, #748, #750**: Scoped HTTP 400 token-limit parsing, official route matching and trusted model-capability hints; do not infer that every vendor-specific HTTP error is fixed. / 收敛 HTTP 400 token 限额解析、官方路由匹配和模型能力提示的信任边界;不代表所有供应商 HTTP 错误均已解决。
|
|
18
|
+
- **#746**: Restored vendor catalog ownership to DSH while retaining the packaged Web sidecar. / 将供应商目录归属交还 DSH,保留随包发布的 Web sidecar。
|
|
19
|
+
- **#740**: Improved `doctor`'s safe, paste-ready GitHub issue template. / 优化 `doctor` 生成可安全粘贴的 GitHub Issue 模板。
|
|
20
|
+
- **#752 and #753**: Updated the lockfile to patched `follow-redirects@1.16.1` and `zod@4.6.6`, respectively. Existing GitHub advisory states may take time to reconcile; changes do not claim an independent full security audit. / 锁文件纳入修复后的 `follow-redirects@1.16.1` 与 `zod@4.6.6`。GitHub 告警状态可能延迟同步;本次升级不代表完成额外全面安全审计。
|
|
21
|
+
- **#758**: Patch `undici` to >=7.30.1 (staying in major 7), and `puppeteer-core` to >=25.16.0 (staying in major 25). The latter removes the vulnerable `chromium-bidi -> zod@3.24.1` production dependency; the direct Zod minimum is raised to `^4.6.6`. The **combined v3.1.0 source candidate** passed `pnpm audit --prod --audit-level high --registry https://registry.npmjs.org` with **no known vulnerabilities**. This is not a guarantee that future advisories cannot appear. / 修复 `undici` 与 `puppeteer-core` 的生产依赖漏洞,消除 Chromium BiDi 引入的旧 Zod 3,并将直接 Zod 最低版本提高至 `4.6.6`。合并预演候选在官方 npm 注册表执行生产依赖审计显示“未发现已知漏洞”;这不等于保证未来不会出现新漏洞。
|
|
22
|
+
- CI and dependency Action pins were refreshed where validated. / 更新已验收的 CI 与依赖 Action 固定版本。
|
|
23
|
+
|
|
24
|
+
## Compatibility and known boundaries / 支持窗口与边界
|
|
25
|
+
|
|
26
|
+
- Node `^22.19.0` or `>=24.0.0` (as stated in package metadata). DSH `>=0.1.5-rc.1 <0.2.0-0 || >=0.2.0-rc.2 <0.3.0-0`. No new Host privileges, automatic Host configuration rewrite, Session schema migration, or public API breaking change is intended.
|
|
27
|
+
- 继续支持上述 Node 和 DSH 版本范围;不新增 Host 权限、不擅自改写 Host 设置、不迁移 Session 格式、不故意引入公共 API 破坏性变更。
|
|
28
|
+
- #756 only affects **notice presentation**; it does not introduce a new consent system or alter network enforcement. A browser-only localStorage preference is not synchronized across browsers or devices. / #756 只改变**提示展示**,不新增隐私授权机制,不改变网络拦截逻辑。浏览器提示偏好不会跨设备同步。
|
|
29
|
+
- Historical intermittent Desktop selection/RPC issues (e.g. #738), vendor-specific HTTP reports and upstream installer behavior are not claimed resolved solely by this minor release. / 历史偶发 Desktop 选择/RPC、特定供应商 HTTP 以及上游安装器问题,不能仅因本版本发布就宣称已全部根修。
|
|
30
|
+
|
|
31
|
+
## Release evidence / 发布验收
|
|
32
|
+
|
|
33
|
+
- #754 and #756 completed their exact PR-head cross-platform CI with no failures (#754: 105 successes, 1 conditional skip; #756: 105 successes, 1 conditional skip), including Node 22/24 and real Host gates. / #754 与 #756 的精确 PR-head 跨平台 CI 均零失败,Node 22/24 与真实 Host 门禁通过。
|
|
34
|
+
- The final combined release-preparation commit, npm tarball, OIDC provenance and published artifact are separate gates and **must be checked before calling v3.1.0 released**. / 最终合并后的候选提交、npm 包体、OIDC 来源证明及发布结果仍需受保护流程核验,未完成前不得称正式发布。
|
package/index.js
CHANGED
|
@@ -58,9 +58,11 @@ import {
|
|
|
58
58
|
import { streamWithVisionSessionAffinity } from './lib/session-affinity-runtime.js'
|
|
59
59
|
import {
|
|
60
60
|
routingCorrectionFor,
|
|
61
|
+
correctionPreservesRuntimeTransport,
|
|
61
62
|
toAnthropicMessages,
|
|
62
63
|
callAnthropicCompatible,
|
|
63
64
|
} from './lib/catalog-corrections.js'
|
|
65
|
+
import { providerTransportFor } from './lib/live-model-discovery.js'
|
|
64
66
|
import { getOfficialDeepSeekCatalog } from './lib/official-deepseek-catalog.js'
|
|
65
67
|
import {
|
|
66
68
|
classifyVisionFailure,
|
|
@@ -1644,6 +1646,25 @@ export function apply(ctx, config = {}, runtime = {}) {
|
|
|
1644
1646
|
if (correction === undefined) return undefined
|
|
1645
1647
|
const rawProfile = rawChannelProfileOf(pair.provider)
|
|
1646
1648
|
const resolvedProfile = resolvedPiAiProfileOf(pair.provider)
|
|
1649
|
+
// A built-in pi-ai catalog fact is not proof that the user's actual
|
|
1650
|
+
// DSH provider config still uses the official gateway. Refuse to reroute
|
|
1651
|
+
// another gateway's credentials into the hard-coded official endpoint.
|
|
1652
|
+
let activeHostTransport
|
|
1653
|
+
let strictHostTransport
|
|
1654
|
+
try {
|
|
1655
|
+
activeHostTransport = providerTransportFor(ctx, pair.provider)
|
|
1656
|
+
strictHostTransport = providerTransportFor(ctx, pair.provider, { strictBaseURL: true })
|
|
1657
|
+
} catch { return undefined }
|
|
1658
|
+
if (!correctionPreservesRuntimeTransport(
|
|
1659
|
+
correction,
|
|
1660
|
+
activeHostTransport,
|
|
1661
|
+
strictHostTransport,
|
|
1662
|
+
rawProfile,
|
|
1663
|
+
resolvedProfile,
|
|
1664
|
+
// piProvider.api is a registry of wire handlers, not a resolved
|
|
1665
|
+
// protocol string. Only use its baseUrl as optional transport evidence.
|
|
1666
|
+
{ baseUrl: resolvedProfile?.piProvider?.baseUrl },
|
|
1667
|
+
)) return undefined
|
|
1647
1668
|
const apiKeyEnv = [rawProfile, resolvedProfile]
|
|
1648
1669
|
.map((profile) => (profile && typeof profile.apiKeyEnv === 'string' ? profile.apiKeyEnv : ''))
|
|
1649
1670
|
.find((ref) => ref !== '')
|