feat(core,cli,web,docs,skills): AgentHub 0.4.6 with the DeepSeek vision model (#388)

Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com>

Author: hiyouga

changelog/unreleased/2026-08-21-agenthub-0.4.6.md

@@ -0,0 +1,26 @@
+# AgentHub 0.4.6: a DeepSeek vision model, and DeepSeek on the Responses protocol
+
+- **Date:** 2026-08-21
+- **Type:** feature
+- **Scope:** `core`, `cli`, `web`, `model-catalog`, `docs`
+- **PR:** [#388](https://github.com/Prism-Shadow/penguin-harness/pull/388)
+- **Breaking:** yes — a `deepseek-v4*` model pointed at a Chat Completions-only endpoint now needs `client_type = "openai-chat"` set by hand
+
+[中文版](2026-08-21-agenthub-0.4.6.zh.md)
+
+`@prismshadow/agenthub` moved from 0.4.5 to 0.4.6, which adds `deepseek-v4-flash-vision-exp` and moves the first-party DeepSeek client from Chat Completions to the OpenAI Responses API (agenthub [#185](https://github.com/Prism-Shadow/agenthub/pull/185)). The model catalog gained the new model in two groups, and the Web model dialog's base URL hint now reads `/responses` for DeepSeek.
+
+## Details
+
+- `deepseek-v4-flash-vision-exp` joined the DeepSeek group: image input on top of V4 Flash's text capabilities, a 1M context window, and V4 Flash's own price. It is the only vision-capable model in that group: a session running it reads images directly, instead of routing them through the `vision_model` proxy `describe_image` uses for text-only models. A fresh Project still defaults to `deepseek-v4-flash`.
+- `deepseek/deepseek-v4-flash-vision-exp` joined the OpenRouter group at that gateway's published rates ($0.22 input / $0.66 output / $0.007 cache read per million tokens) and its 1,048,576-token context window. DeepSeek is the only provider serving it there.
+- Both rows record the official **off-peak** tier, the same convention the existing DeepSeek rows follow: peak hours (Beijing 9:00–12:00 and 14:00–18:00) bill exactly double, and the cost centre uses one rate, so peak usage is under-counted 2x.
+- The protocol-path hint at the right edge of the base URL field reads `/responses` for the DeepSeek group and for an explicit `deepseek-v4` client type. The gateway rows reselling DeepSeek pin `openai-chat` and keep `/chat/completions`.
+- The shipped `agenthub-models` skill lists the new id in both its official and OpenRouter columns, and records which wire protocol the `deepseek-v4` client speaks.
+- `pnpm-workspace.yaml`'s `minimumReleaseAgeExclude` entry names 0.4.6 and drops 0.4.5, as that file's own note instructed.
+
+## Compatibility
+
+Existing Projects do not pick up the two new catalog rows on their own: presets are copied into `.project_config.toml` when a Project is created, and nothing rewrites them afterwards. Use **sync presets** on the models page to append them; the stored default model is never touched.
+
+A model whose id contains `deepseek-v4` and which carries no `client_type` is routed to the DeepSeek client by its id alone, and that client now posts to `{base_url}/responses`. Entries pointed at `https://api.deepseek.com` are unaffected — DeepSeek serves both protocols. An entry pointed at an endpoint that serves Chat Completions only (a self-hosted server, a relay, a third-party DeepSeek-compatible gateway) starts failing after the upgrade: set `client_type = "openai-chat"` on it, in `.project_config.toml` or through the models page's protocol selector, and the request shape returns to what it was.

changelog/unreleased/2026-08-21-agenthub-0.4.6.zh.md

@@ -0,0 +1,26 @@
+# AgentHub 0.4.6：DeepSeek 视觉模型，以及 DeepSeek 改走 Responses 协议
+
+- **Date:** 2026-08-21
+- **Type:** feature
+- **Scope:** `core`, `cli`, `web`, `model-catalog`, `docs`
+- **PR:** [#388](https://github.com/Prism-Shadow/penguin-harness/pull/388)
+- **Breaking:** yes — 指向仅提供 Chat Completions 端点的 `deepseek-v4*` 模型，现在需要手工设置 `client_type = "openai-chat"`
+
+[English](2026-08-21-agenthub-0.4.6.md)
+
+`@prismshadow/agenthub` 从 0.4.5 升到 0.4.6。新版本加入了 `deepseek-v4-flash-vision-exp`，并把第一方 DeepSeek client 从 Chat Completions 改为 OpenAI Responses API（agenthub [#185](https://github.com/Prism-Shadow/agenthub/pull/185)）。模型目录在两个分组中收录了该模型，Web 模型对话框中 base URL 的协议路径提示对 DeepSeek 改为 `/responses`。
+
+## 细节
+
+- `deepseek-v4-flash-vision-exp` 进入 DeepSeek 分组：在 V4 Flash 的文本能力之上支持图像输入，上下文窗口 100 万 Token，价格与 V4 Flash 相同。它是该分组中唯一支持图像输入的模型：以它运行的会话直接读图，无需再经 `describe_image` 为 text-only 模型准备的 `vision_model` 代读路径。新建 Project 的默认模型仍是 `deepseek-v4-flash`。
+- `deepseek/deepseek-v4-flash-vision-exp` 进入 OpenRouter 分组，按该网关公布的价格记录（每百万 Token 输入 $0.22 / 输出 $0.66 / 缓存命中 $0.007），上下文窗口为 1,048,576 Token。该模型在 OpenRouter 上只有 DeepSeek 一家提供。
+- 两条记录都采用官方**低谷时段**档，与既有 DeepSeek 记录一致：高峰时段（北京时间 9:00–12:00、14:00–18:00）按整整双倍计费，而费用中心只用单一价格，因此高峰用量会被低估一半。
+- base URL 输入框右端的协议路径提示，对 DeepSeek 分组以及显式的 `deepseek-v4` client type 均显示 `/responses`。转售 DeepSeek 的网关分组固定 `openai-chat`，仍显示 `/chat/completions`。
+- 随附的 `agenthub-models` skill 在官方与 OpenRouter 两列中都列出了新 id，并记录了 `deepseek-v4` client 使用的传输协议。
+- `pnpm-workspace.yaml` 中 `minimumReleaseAgeExclude` 的条目改为 0.4.6；按该文件自身注释的要求，移除了 0.4.5 的条目。
+
+## 兼容性
+
+既有 Project 不会自动获得这两条新目录记录：预置模型在创建 Project 时写入 `.project_config.toml`，此后不再被改写。在模型页点击**同步预置**即可追加，存储的默认模型不受影响。
+
+model id 中含 `deepseek-v4`、且未填写 `client_type` 的条目，仅凭 id 就会被路由到 DeepSeek client，而该 client 现在向 `{base_url}/responses` 发请求。指向 `https://api.deepseek.com` 的条目不受影响——DeepSeek 两种协议都提供。指向仅提供 Chat Completions 端点的条目（自建服务、中转站、第三方 DeepSeek 兼容网关）在升级后会开始失败：为其设置 `client_type = "openai-chat"` 即可——在 `.project_config.toml` 中填写，或用模型页的协议选择器选择——请求形态即恢复原样。

packages/cli/package.json

@@ -24,7 +24,7 @@
   },
   "dependencies": {
     "@modelcontextprotocol/client": "^2.0.0",
-    "@prismshadow/agenthub": "^0.4.5",
+    "@prismshadow/agenthub": "^0.4.6",
     "@prismshadow/penguin-core": "workspace:*",
     "@prismshadow/penguin-server": "workspace:*",
     "@prismshadow/penguin-skills": "workspace:*",

packages/core/package.json

@@ -49,7 +49,7 @@
   },
   "dependencies": {
     "@modelcontextprotocol/client": "^2.0.0",
-    "@prismshadow/agenthub": "^0.4.5",
+    "@prismshadow/agenthub": "^0.4.6",
     "@prismshadow/penguin-skills": "workspace:*",
     "arktype": "^2.2.3",
     "smol-toml": "^1.3.0",

packages/core/src/state/model-catalog.ts

@@ -3,7 +3,8 @@
  * auto-route, shared by core's default config, server's initial config, and web/cli display.
  * Data verified as of 2026-07-10 (Qwen Token Plan entries: 2026-07-20; MiniMax: 2026-08-03;
  * DeepSeek, Gemini 3.7, GLM-5.3 and the whole OpenAI line-up (direct + OpenRouter):
- * 2026-08-18; the direct Anthropic group: 2026-08-20 — per each provider's docs).
+ * 2026-08-18; the direct Anthropic group: 2026-08-20; the DeepSeek V4 Flash Vision Exp rows:
+ * 2026-08-21 — per each provider's docs).
  * Docs: packages/docs/content/models.{zh,en}.md (site path /docs/models) documents the
  * provider groups and credential resolution described here.
  *
@@ -247,6 +248,16 @@ export const MODEL_CATALOG: ModelCatalogEntry[] = [
     pricing: cny(0.05, 1.5, 4.5),
     supportsVision: false,
   },
+  {
+    // The experimental vision revision of V4 Flash (added 2026-08-21): image input on top of
+    // the base model's text capabilities, at the same published price.
+    modelId: "deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "deepseek",
+    contextWindow: 1000000,
+    pricing: cny(0.05, 1.5, 4.5),
+    supportsVision: true,
+  },
   {
     modelId: "deepseek-v4-pro",
     displayName: "DeepSeek V4 Pro",
@@ -358,6 +369,20 @@ export const MODEL_CATALOG: ModelCatalogEntry[] = [
     clientType: "openai-chat",
     baseUrl: OPENROUTER_BASE_URL,
   },
+  {
+    // DeepSeek serves this one alone on OpenRouter, so the stored rates are its own published
+    // USD list ($0.22 / $0.66 / $0.007 cache read), and the context window is that endpoint's
+    // 1,048,576. Like the direct group, the price is the OFF-PEAK tier: the models API exposes
+    // the peak windows as `pricing.overrides` billing exactly double.
+    modelId: "deepseek/deepseek-v4-flash-vision-exp",
+    displayName: "DeepSeek V4 Flash Vision Exp",
+    provider: "openrouter",
+    contextWindow: 1048576,
+    pricing: usd(0.007, 0.22, 0.66),
+    supportsVision: true,
+    clientType: "openai-chat",
+    baseUrl: OPENROUTER_BASE_URL,
+  },
   {
     // The 0813 general-availability release of DeepSeek V4 Pro (OpenRouter listing dated
     // 2026-08-12); the default routed endpoint is DeepSeek's own API, so the price matches

packages/core/test/model-catalog.test.ts

@@ -131,6 +131,13 @@ describe("model-catalog", () => {
     expect(catalogEntryFor("zhipu", "glm-5.2")?.contextWindow).toBe(1000000);
     expect(catalogEntryFor("qwen-token-plan", "glm-5.2")?.contextWindow).toBe(1048576);
     expect(catalogEntryFor("deepseek", "deepseek-v4-pro")?.provider).toBe("deepseek");
+    // The vision revision is a model of its own in both the direct group and on OpenRouter,
+    // and it is the only vision-capable DeepSeek row in either.
+    expect(catalogEntryFor("deepseek", "deepseek-v4-flash-vision-exp")?.supportsVision).toBe(true);
+    expect(
+      catalogEntryFor("openrouter", "deepseek/deepseek-v4-flash-vision-exp")?.supportsVision,
+    ).toBe(true);
+    expect(catalogEntryFor("deepseek", "deepseek-v4-flash")?.supportsVision).toBe(false);
     expect(catalogEntryFor("qwen-token-plan", "deepseek-v4-pro")?.provider).toBe("qwen-token-plan");
     expect(catalogEntryFor("minimax", "MiniMax-M3")?.displayName).toBe("MiniMax M3");
   });
@@ -167,6 +174,7 @@ describe("model-catalog", () => {
       "anthropic/claude-sonnet-5",
       "deepseek/deepseek-v4-flash-0731",
       "deepseek/deepseek-v4-flash",
+      "deepseek/deepseek-v4-flash-vision-exp",
       "deepseek/deepseek-v4-pro-0813",
       "deepseek/deepseek-v4-pro",
       "google/gemini-3.7-flash",

packages/docs/content/models.en.md

@@ -77,7 +77,7 @@ The gateway groups (openrouter / fireworks / siliconflow / qwen-token-plan / qwe
 
 The preset catalog also carries OpenRouter's free tier: the `:free` model variant `nvidia/nemotron-3-ultra-550b-a55b:free` and the `openrouter/free` unified Free Models Router. They cost nothing, but are subject to OpenRouter's free-tier rate limits and data policy.
 
-Some models in the preset catalog: deepseek-v4-pro / deepseek-v4-flash, MiniMax-M3, gemini-3.7-flash, claude-opus-5 / claude-opus-4-8 / claude-sonnet-5, gpt-5.6 / gpt-5.5, glm-5.3, kimi-k3, qwen3.8-max (not exhaustive). The whole OpenAI line-up is listed twice — directly (your own OpenAI key, list prices) and on OpenRouter as `openai/<id>` (the gateway's rates, which follow its running promotions). DeepSeek's direct-group prices record the official off-peak tier (peak hours, Beijing 9:00–12:00 and 14:00–18:00, bill double).
+Some models in the preset catalog: deepseek-v4-pro / deepseek-v4-flash / deepseek-v4-flash-vision-exp (the DeepSeek group's only vision-capable model), MiniMax-M3, gemini-3.7-flash, claude-opus-5 / claude-opus-4-8 / claude-sonnet-5, gpt-5.6 / gpt-5.5, glm-5.3, kimi-k3, qwen3.8-max (not exhaustive). The whole OpenAI line-up is listed twice — directly (your own OpenAI key, list prices) and on OpenRouter as `openai/<id>` (the gateway's rates, which follow its running promotions). DeepSeek's direct-group prices record the official off-peak tier (peak hours, Beijing 9:00–12:00 and 14:00–18:00, bill double).
 
 ## Local / self-hosted OpenAI-compatible endpoints (e.g. vLLM)
 

packages/docs/content/models.zh.md

@@ -77,7 +77,7 @@ api_key = "sk-..."
 
 预置目录还收录了 OpenRouter 的免费档：`:free` 模型变体 `nvidia/nemotron-3-ultra-550b-a55b:free` 与统一路由 `openrouter/free`(Free Models Router)，零成本可用，但受 OpenRouter 免费档速率限制与数据政策约束。
 
-预置目录中的部分模型：deepseek-v4-pro / deepseek-v4-flash、MiniMax-M3、gemini-3.7-flash、claude-opus-5 / claude-opus-4-8 / claude-sonnet-5、gpt-5.6 / gpt-5.5、glm-5.3、kimi-k3、qwen3.8-max 等(非完整清单)。OpenAI 全系列都收录了两份——直连(用自己的 OpenAI Key，记牌价)与 OpenRouter 上的 `openai/<id>`(记网关实际计费价，会随其促销浮动)。DeepSeek 直连分组的价格记录官方低谷时段档(高峰时段——北京时间 9:00–12:00、14:00–18:00——按双倍计费)。
+预置目录中的部分模型：deepseek-v4-pro / deepseek-v4-flash / deepseek-v4-flash-vision-exp(DeepSeek 分组中唯一支持图像输入的模型)、MiniMax-M3、gemini-3.7-flash、claude-opus-5 / claude-opus-4-8 / claude-sonnet-5、gpt-5.6 / gpt-5.5、glm-5.3、kimi-k3、qwen3.8-max 等(非完整清单)。OpenAI 全系列都收录了两份——直连(用自己的 OpenAI Key，记牌价)与 OpenRouter 上的 `openai/<id>`(记网关实际计费价，会随其促销浮动)。DeepSeek 直连分组的价格记录官方低谷时段档(高峰时段——北京时间 9:00–12:00、14:00–18:00——按双倍计费)。
 
 ## 本地 / 自建 OpenAI 兼容端点（如 vLLM）
 

packages/skills/skills/agenthub-models/SKILL.md

@@ -3,7 +3,7 @@ name: agenthub-models
 description: Call model APIs through @prismshadow/agenthub — streaming text generation, image generation, speech synthesis, embeddings and the supported-model registry with one client.
 short_description: Call model APIs with one AgentHub client.
 short_description_zh: 用一个 AgentHub 客户端调用模型 API。
-version: 14
+version: 15
 updated: 2026-08-21T00:00:00Z
 ---
 
@@ -64,7 +64,7 @@ Use exact model ids. If an id is not in the table below and the user has not giv
 | Kimi K3          | `kimi-k3`                                                             | OpenRouter `moonshotai/kimi-k3`                                                                                                                 |
 | Kimi K2.7 Code   | —                                                                     | SiliconFlow `moonshotai/Kimi-K2.7-Code`; Fireworks AI `accounts/fireworks/models/kimi-k2p7-code`                                                |
 | Kimi K2.6        | `kimi-k2.6`                                                           | OpenRouter `moonshotai/kimi-k2.6`; SiliconFlow `Pro/moonshotai/Kimi-K2.6`                                                                       |
-| DeepSeek V4      | `deepseek-v4-pro`, `deepseek-v4-flash`                                | OpenRouter `deepseek/deepseek-v4-pro-0813`, `deepseek/deepseek-v4-pro`, `deepseek/deepseek-v4-flash`, `deepseek/deepseek-v4-flash-0731`; Fireworks AI `accounts/fireworks/models/deepseek-v4-flash-0731`; SiliconFlow `deepseek-ai/DeepSeek-V4-Pro`, `deepseek-ai/DeepSeek-V4-Flash` |
+| DeepSeek V4      | `deepseek-v4-pro`, `deepseek-v4-flash`, `deepseek-v4-flash-vision-exp` | OpenRouter `deepseek/deepseek-v4-pro-0813`, `deepseek/deepseek-v4-pro`, `deepseek/deepseek-v4-flash`, `deepseek/deepseek-v4-flash-0731`, `deepseek/deepseek-v4-flash-vision-exp`; Fireworks AI `accounts/fireworks/models/deepseek-v4-flash-0731`; SiliconFlow `deepseek-ai/DeepSeek-V4-Pro`, `deepseek-ai/DeepSeek-V4-Flash` |
 | GLM 5.3          | `glm-5.3`                                                             | OpenRouter `z-ai/glm-5.3`                                                                                                                       |
 | GLM 5.2          | `glm-5.2`                                                             | OpenRouter `z-ai/glm-5.2`; SiliconFlow `zai-org/GLM-5.2`                                                                                        |
 | GLM 5.1          | `glm-5.1`                                                             | —                                                                                                                                               |
@@ -108,6 +108,7 @@ The registry is the curated current line-up, so prefer it when picking a model o
   - `clientType: "ant-messages"` — Anthropic Messages-compatible endpoints (Anthropic, OpenRouter `https://openrouter.ai/api`, DeepSeek `https://api.deepseek.com/anthropic`, Z.AI, MiniMax).
 - Exception: an id served by an OpenAI-compatible gateway that still matches a first-party substring (e.g. OpenRouter's `google/gemini-3.7-flash`, `anthropic/claude-sonnet-5` or `openai/gpt-5.6-sol` on the `/api/v1` endpoint) would auto-route to the vendor protocol client — and a dotted id like `anthropic/claude-opus-4.8` matches nothing and throws. Always pass an explicit `clientType` for gateway ids; never rely on the id. Routing reads `clientType` (or the model id) as a plain lowercased string and never looks at `baseUrl`, so the vendor prefix gives no protection.
 - OpenRouter serves both protocols at `https://openrouter.ai/api/v1`, so its `openai/*` ids work with `clientType: "openai-responses"` as well as `"openai-chat"`; use Responses when you want reasoning items round-tripped.
+- The first-party `deepseek-v4` client posts to `{baseUrl}/responses` (AgentHub 0.4.6 moved it off Chat Completions). A self-hosted endpoint serving a `deepseek-v4*` id over Chat Completions must therefore pass `clientType: "openai-chat"` explicitly rather than rely on id routing.
 - API key: constructor parameter first, then the provider environment variable — `DEEPSEEK_API_KEY`, `ANTHROPIC_API_KEY` (also for `ant-messages`), `OPENAI_API_KEY` (also for `openai-chat`/`openai-responses`), `GEMINI_API_KEY`, `ZAI_API_KEY`, `MOONSHOT_API_KEY`, `MINIMAX_API_KEY`. Base URLs read the same names with `_BASE_URL`.
 
 ## Streaming text