Read images with deepseek-v4-flash-vision-exp, and move DeepSeek onto the Responses protocol

README.md

@@ -46,7 +46,7 @@ https://github.com/user-attachments/assets/c49a21a1-5bf9-4768-a76d-f73c9a03ca87
 | Claude 4.6-5   | Official/Amazon Bedrock/UModelVerse | `claude-opus-5`        | Text, Image      | Text                           |
 | GPT-5.4-5.6    | Official/OpenRouter/UModelVerse     | `gpt-5.6-sol`          | Text, Image      | Text, Embedding                |
 | Kimi-K2.5/K2.6/K3 | Official/OpenRouter/SiliconFlow  | `kimi-k3`              | Text, Image      | Text                           |
-| DeepSeek V4    | Official/OpenRouter/SiliconFlow     | `deepseek-v4-pro`      | Text             | Text                           |
+| DeepSeek V4    | Official/OpenRouter/SiliconFlow     | `deepseek-v4-pro`      | Text, Image      | Text                           |
 | GLM-5.1-5.3    | Official/OpenRouter/SiliconFlow     | `glm-5.3`              | Text             | Text                           |
 | MiniMax-M3     | Official                            | `MiniMax-M3`           | Text, Image      | Text                           |
 | Qwen3.6        | OpenRouter/SiliconFlow/vLLM         | `qwen/qwen3.6-35b-a3b` | Text, Image      | Text, Embedding                |

changelog/unreleased/2026-08-21-deepseek-responses-protocol.md

@@ -0,0 +1,36 @@
+# DeepSeek moves onto the OpenAI Responses protocol
+
+- **Date:** 2026-08-21
+- **Type:** refactor
+- **Scope:** `deepseek_v4`, `registry`, `tests`
+
+[中文版](2026-08-21-deepseek-responses-protocol.zh.md)
+
+## What changed
+
+- `DeepSeekV4Client` calls `/responses` instead of `/chat/completions`, reading the
+  `response.reasoning_text.delta`, `response.output_text.delta`,
+  `response.function_call_arguments.delta` and `response.completed` events, and rebuilding a
+  replayed chain of thought as a `reasoning` item whose `content` is `reasoning_text`.
+- The two SiliconFlow entries, `deepseek-ai/DeepSeek-V4-Flash` and
+  `deepseek-ai/DeepSeek-V4-Pro`, moved to the generic `openai-chat` client: SiliconFlow serves
+  Chat Completions only. The OpenRouter and official entries keep the `deepseek-v4` client.
+- The DeepSeek cases in the empty-response, tool-call-argument and unknown-event unit suites
+  moved to the Responses wire shape, and those first two suites gained a case for the generic
+  `openai-responses` client alongside it.
+
+## Configuration behavior
+
+| `UniConfig` key | On the wire |
+| --- | --- |
+| `thinking_level` | `reasoning.effort`, pre-mapped to what DeepSeek settles on: `none` / `low` / `high` / `high` / `high` / `max`. Effort `none` turns thinking off; the Chat Completions `thinking` toggle is ignored on this endpoint |
+| `thinking_summary` | left out: the endpoint accepts `summary` but never generates one |
+| `max_tokens` | `max_output_tokens` |
+| `system_prompt` | `instructions` |
+| `temperature` | `UnsupportedParameterError` unless it is `1.0` |
+| `tool_choice` | `auto` and `none` only; anything else raises `UnsupportedParameterError` |
+| `fast_mode` | `UnsupportedParameterError`: `service_tier` is not supported |
+| `prompt_caching` | `ENABLE` only; caching is automatic |
+
+Usage arrives as `input_tokens` with `input_tokens_details.cached_tokens`, and `output_tokens`
+with `output_tokens_details.reasoning_tokens`.

changelog/unreleased/2026-08-21-deepseek-responses-protocol.zh.md

@@ -0,0 +1,35 @@
+# DeepSeek 改用 OpenAI Responses 协议
+
+- **Date:** 2026-08-21
+- **Type:** refactor
+- **Scope:** `deepseek_v4`, `registry`, `tests`
+
+[English](2026-08-21-deepseek-responses-protocol.md)
+
+## 变更内容
+
+- `DeepSeekV4Client` 改为调用 `/responses` 而非 `/chat/completions`，解析
+  `response.reasoning_text.delta`、`response.output_text.delta`、
+  `response.function_call_arguments.delta` 与 `response.completed` 事件，回放思维链时重建成
+  `content` 为 `reasoning_text` 的 `reasoning` 条目。
+- SiliconFlow 的两个条目 `deepseek-ai/DeepSeek-V4-Flash` 与 `deepseek-ai/DeepSeek-V4-Pro` 改用通用
+  的 `openai-chat` 客户端：SiliconFlow 只提供 Chat Completions。OpenRouter 与官方条目仍走
+  `deepseek-v4` 客户端。
+- 空响应、工具参数与未知事件三个单测套件里的 DeepSeek 用例改为 Responses 线格式，前两个套件同时
+  新增了通用 `openai-responses` 客户端的用例。
+
+## 配置映射
+
+| `UniConfig` 键 | 线上字段 |
+| --- | --- |
+| `thinking_level` | `reasoning.effort`，按 DeepSeek 实际生效的值预先映射：`none` / `low` / `high` / `high` / `high` / `max`。effort 取 `none` 才能关闭思考，Chat Completions 的 `thinking` 开关在该端点上被忽略 |
+| `thinking_summary` | 不下发：该端点接受 `summary` 但从不生成 |
+| `max_tokens` | `max_output_tokens` |
+| `system_prompt` | `instructions` |
+| `temperature` | 非 `1.0` 时抛 `UnsupportedParameterError` |
+| `tool_choice` | 只接受 `auto` 与 `none`，其余抛 `UnsupportedParameterError` |
+| `fast_mode` | 抛 `UnsupportedParameterError`：不支持 `service_tier` |
+| `prompt_caching` | 只接受 `ENABLE`，缓存由服务端自动处理 |
+
+用量字段为 `input_tokens`（含 `input_tokens_details.cached_tokens`）与 `output_tokens`
+（含 `output_tokens_details.reasoning_tokens`）。

changelog/unreleased/2026-08-21-deepseek-vision.md

@@ -0,0 +1,24 @@
+# deepseek-v4-flash-vision-exp reads images, in a prompt and in a tool result
+
+- **Date:** 2026-08-21
+- **Type:** feature
+- **Scope:** `deepseek_v4`, `registry`, `llmsdk_docs`, `tests`
+
+[中文版](2026-08-21-deepseek-vision.zh.md)
+
+## What changed
+
+- `deepseek-v4-flash-vision-exp` joined the registry on the official API with `Text, Image`
+  input, a 1M context window, and the `deepseek-v4-flash` price (¥1.5 / ¥4.5 per million
+  tokens off-peak, ¥0.05 on a cache hit).
+- `DeepSeekV4Client` sends an `image_url` item as an `input_image` content part and a
+  `tool_result` image as an `input_image` inside `function_call_output`, so a tool can hand the
+  model a picture it produced. Both an HTTP(S) URL and a base64 data URL are accepted.
+- A model without vision refuses an image item with
+  `DeepSeek <model> does not support image inputs.` rather than sending one: Chat Completions
+  answers `400` and the Responses API silently substitutes placeholder text.
+- `deepseek-v4-flash-vision-exp` replaced `deepseek-v4-flash` as the official DeepSeek entry in
+  both E2E model lists, with image understanding enabled; the three protocol-mode entries stay
+  on `deepseek-v4-flash`.
+- `llmsdk_docs/deepseek_v4/` gained snapshots of the vision and Responses API guides, and its
+  quickstart lists the third model.

changelog/unreleased/2026-08-21-deepseek-vision.zh.md

@@ -0,0 +1,21 @@
+# deepseek-v4-flash-vision-exp 可读图片，提示词与工具返回都支持
+
+- **Date:** 2026-08-21
+- **Type:** feature
+- **Scope:** `deepseek_v4`, `registry`, `llmsdk_docs`, `tests`
+
+[English](2026-08-21-deepseek-vision.md)
+
+## 变更内容
+
+- 注册表新增官方 API 上的 `deepseek-v4-flash-vision-exp`，输入模态为 `Text, Image`，上下文窗口
+  1M，价格与 `deepseek-v4-flash` 相同（谷值每百万 token 输入 ¥1.5、输出 ¥4.5，命中缓存 ¥0.05）。
+- `DeepSeekV4Client` 把 `image_url` 条目发成 `input_image` 内容块，把 `tool_result` 里的图片发成
+  `function_call_output` 内的 `input_image`，于是工具可以把自己产出的图片交给模型。HTTP(S) 链接与
+  base64 data URL 都可用。
+- 不带视觉能力的模型遇到图片条目会直接报
+  `DeepSeek <model> does not support image inputs.` 而不是照发：Chat Completions 会返回 `400`，
+  Responses API 则会静默替换成占位文本。
+- 两份 E2E 模型清单里，官方 DeepSeek 条目由 `deepseek-v4-flash` 换成
+  `deepseek-v4-flash-vision-exp` 并打开图像理解；三个协议模式条目仍用 `deepseek-v4-flash`。
+- `llmsdk_docs/deepseek_v4/` 新增视觉与 Responses API 两份官方文档快照，quickstart 补上第三个模型。

changelog/unreleased/2026-08-21-responses-input-order.md

@@ -0,0 +1,15 @@
+# Responses input items keep the order the assistant produced them in
+
+- **Date:** 2026-08-21
+- **Type:** fix
+- **Scope:** `deepseek_v4`, `openai_responses`
+
+[中文版](2026-08-21-responses-input-order.zh.md)
+
+## What changed
+
+- `DeepSeekV4Client` and `OpenaiResponsesClient` flush the message text collected so far
+  before appending a reasoning, `function_call` or `function_call_output` item, so an assistant
+  turn that spoke before calling a tool replays as text, then call, then output.
+- Before the fix the assistant message was appended after the items it preceded, which DeepSeek
+  answered with `400 No tool output found for tool call <id>` on the next turn.

changelog/unreleased/2026-08-21-responses-input-order.zh.md

@@ -0,0 +1,15 @@
+# Responses 输入条目保持助手产出时的顺序
+
+- **Date:** 2026-08-21
+- **Type:** fix
+- **Scope:** `deepseek_v4`, `openai_responses`
+
+[English](2026-08-21-responses-input-order.md)
+
+## 变更内容
+
+- `DeepSeekV4Client` 与 `OpenaiResponsesClient` 在追加 reasoning、`function_call` 或
+  `function_call_output` 条目之前，先把已经收集的消息正文落盘，因此「先说话再调工具」的助手轮次
+  会按正文、调用、结果的顺序回放。
+- 修复前助手消息被排到它本应领先的条目之后，DeepSeek 会在下一轮返回
+  `400 No tool output found for tool call <id>`。

changelog/unreleased/README.md

@@ -2,5 +2,8 @@
 
 [中文版](README.zh.md)
 
+- [2026-08-21] `deepseek-v4-flash-vision-exp` reads images, in a prompt and in a tool result. ([details](2026-08-21-deepseek-vision.md))
+- [2026-08-21] DeepSeek moves onto the OpenAI Responses protocol. ([details](2026-08-21-deepseek-responses-protocol.md))
+- [2026-08-21] Responses input items keep the order the assistant produced them in. ([details](2026-08-21-responses-input-order.md))
 - [2026-08-21] The playground remembers its configuration across a reload. ([details](2026-08-21-playground-config-memory.md), [#184](https://github.com/Prism-Shadow/agenthub/pull/184))
 - [2026-08-20] Gemini TTS requests carry a single text turn, and the playground plays one audio clip. ([details](2026-08-20-gemini-tts-single-turn.md), [#184](https://github.com/Prism-Shadow/agenthub/pull/184))

changelog/unreleased/README.zh.md

@@ -2,5 +2,8 @@
 
 [English](README.md)
 
+- [2026-08-21] `deepseek-v4-flash-vision-exp` 可读图片，提示词与工具返回都支持。([详情](2026-08-21-deepseek-vision.zh.md))
+- [2026-08-21] DeepSeek 改用 OpenAI Responses 协议。([详情](2026-08-21-deepseek-responses-protocol.zh.md))
+- [2026-08-21] Responses 输入条目保持助手产出时的顺序。([详情](2026-08-21-responses-input-order.zh.md))
 - [2026-08-21] Playground 刷新后保留配置。([详情](2026-08-21-playground-config-memory.zh.md), [#184](https://github.com/Prism-Shadow/agenthub/pull/184))
 - [2026-08-20] Gemini TTS 请求只带一轮文本，Playground 播放单条音频。([详情](2026-08-20-gemini-tts-single-turn.zh.md), [#184](https://github.com/Prism-Shadow/agenthub/pull/184))

llmsdk_docs/deepseek_v4/README.md

@@ -11,8 +11,10 @@ This directory contains documentation for using the DeepSeek API with OpenAI-com
 The `docs/` folder contains focused guides for DeepSeek API features:
 
 - [chat_completion_api.md](./docs/chat_completion_api.md) - DeepSeek Chat Completion API reference
+- [responses-api.md](./docs/responses-api.md) - Responses API support: streaming events, image input, parameter and input-item coverage
 - [kv-cache.md](./docs/kv-cache.md) - Default context caching, hit rules, and usage counters
 - [multi-round-chat.md](./docs/multi-round-chat.md) - Stateless multi-turn conversation history
+- [vision.md](./docs/vision.md) - Image input on deepseek-v4-flash-vision-exp across the three API formats, with detail levels and limits
 - [thinking-mode.md](./docs/thinking-mode.md) - Thinking mode, reasoning content, effort control, and tool-call context rules
 - [tool-calls.md](./docs/tool-calls.md) - Tool calling, thinking-mode tool calls, and strict mode
 

llmsdk_docs/deepseek_v4/docs/responses-api.md

@@ -0,0 +1,156 @@
+# Using the Responses API
+
+> Source: https://api-docs.deepseek.com/guides/responses_api (snapshot 2026-08-21)
+
+DeepSeek serves the OpenAI Responses API format at base URL `https://api.deepseek.com`, so the
+OpenAI SDK reaches it with only a base URL and API key change.
+
+## Basic Usage
+
+```python
+from openai import OpenAI
+
+client = OpenAI(api_key="<your DeepSeek API Key>", base_url="https://api.deepseek.com")
+response = client.responses.create(
+    model="deepseek-v4-flash",
+    instructions="You are a helpful assistant.",
+    input="Hi, how are you?",
+)
+print(response.output_text)
+```
+
+## Streaming
+
+Set `stream: true` to receive server-sent events. Each event carries a `type` and a monotonically
+increasing sequence number, and the stream ends with a terminal event rather than `data: [DONE]`.
+
+```python
+stream = client.responses.create(
+    model="deepseek-v4-flash",
+    instructions="You are a helpful assistant.",
+    input="Hi, how are you?",
+    stream=True,
+)
+for event in stream:
+    if event.type == "response.output_text.delta":
+        print(event.delta, end="")
+```
+
+| Event | Purpose |
+| --- | --- |
+| `response.created` | Initial event; status is `in_progress` |
+| `response.in_progress` | Response generation ongoing |
+| `response.output_item.added` / `response.output_item.done` | Output item lifecycle (reasoning, message, function_call, ...) |
+| `response.content_part.added` / `response.content_part.done` | Content part lifecycle within an item |
+| `response.reasoning_text.delta` / `response.reasoning_text.done` | Chain-of-thought incremental / complete |
+| `response.output_text.delta` / `response.output_text.done` | Output text incremental / complete |
+| `response.function_call_arguments.delta` / `response.function_call_arguments.done` | Function arguments incremental / complete |
+| `response.custom_tool_call_input.delta` / `response.custom_tool_call_input.done` | Custom tool input incremental / complete |
+| `response.web_search_call.in_progress` / `.searching` / `.completed` | Web search status |
+| `response.completed` | Normal completion, with the full response object |
+| `response.incomplete` | Truncation (for example `max_output_tokens` reached) |
+| `response.failed` | Failure, with error details |
+
+## Image Input
+
+`deepseek-v4-flash-vision-exp` accepts images as `input_image` content parts:
+
+```python
+response = client.responses.create(
+    model="deepseek-v4-flash-vision-exp",
+    input=[
+        {
+            "role": "user",
+            "content": [
+                {"type": "input_text", "text": "What is in this image?"},
+                {"type": "input_image", "image_url": "https://example.com/image.jpg", "detail": "low"},
+            ],
+        }
+    ],
+)
+print(response.output_text)
+```
+
+Images may also ride in `function_call_output` and `custom_tool_call_output`, so a tool can hand
+the model an image it produced.
+
+### `input_image` fields
+
+- `image_url`: an HTTP(S) URL (max 8192 characters) or a base64 data URL. JPEG, PNG, GIF and WebP.
+- `file_id`: an image id from the Files API (`file-api-...`).
+- `detail`: `low` / `high` / `original` / `auto`. `low` downsamples to 512x512; the others keep the
+  original size. Ignored when `file_id` is set.
+
+`image_url` and `file_id` are mutually exclusive: passing both, or neither, returns `400`.
+
+### Image restrictions
+
+- Images are allowed only in `user` / `developer` messages and in `function_call_output` /
+  `custom_tool_call_output`. An image in a `system` or `assistant` message returns `400`.
+- Only vision models read `input_image` parts; other models replace them with placeholder text.
+- Limits: 32 MiB per inline image, 64 MiB per `file_id` image, 64 MiB total inline (up to 200 MiB
+  with `file_id`), and 600 images per request.
+
+## Top-level request parameters
+
+| Parameter | Status | Notes |
+| --- | --- | --- |
+| `model` | Supported | `deepseek-v4-flash` / `deepseek-v4-pro` / `deepseek-v4-flash-vision-exp` |
+| `input` | Supported | String or input item list; `input` or `instructions` is required |
+| `instructions` | Supported | Inserted as the first system message |
+| `stream` | Supported | — |
+| `temperature` | Supported | Range [0.0, 2.0]; no effect in thinking mode |
+| `top_p` | Supported | No effect in thinking mode |
+| `max_output_tokens` | Supported | — |
+| `top_logprobs` | Supported | Range [0, 20] |
+| `tools` | Partially supported | `function` / `web_search` supported; other types ignored |
+| `tool_choice` | Supported | `none` / `auto` / `required` / a specific tool |
+| `reasoning` | Partially supported | `effort` supported; `summary` accepted but never generated |
+| `text` | Partially supported | `format` supported; `verbosity` has no effect |
+| `user` | Supported | See Rate Limit & Isolation |
+| `parallel_tool_calls` | Ignored | Always enabled |
+| `max_tool_calls` | Ignored | — |
+| `previous_response_id` | Not supported | Stateless API |
+| `conversation` | Not supported | Stateless API |
+| `store` | Not supported | Always `store: false` |
+| `background` | Not supported | — |
+| `metadata` | Not supported | — |
+| `include` | Not supported | — |
+| `prompt` | Not supported | — |
+| `truncation` | Not supported | Exceeding the context window returns `400` |
+| `service_tier` | Not supported | — |
+| `safety_identifier` | Not supported | — |
+| `prompt_cache_key` / `prompt_cache_retention` | Not supported | Context caching is automatic |
+| `context_management` | Not supported | — |
+| `stream_options` | Not supported | — |
+
+Unsupported parameters are silently ignored rather than rejected.
+
+## Input item types
+
+| Type | Status | Notes |
+| --- | --- | --- |
+| `message` | Supported | Roles `user` / `assistant` / `system` / `developer` (developer is treated as user); content takes a string or `input_text` / `output_text` / `input_image` parts. The vision model reads images; other models substitute a placeholder. File inputs are unsupported |
+| `function_call` | Supported | Merged into the adjacent assistant message |
+| `function_call_output` | Supported | `output` takes a string or a content part list; `input_image` parts are read as real images by the vision model |
+| `reasoning` | Supported | Plain-text content is merged into the assistant message; `summary` and `encrypted_content` are unsupported |
+| `web_search_call` | Supported | Passed back as-is; the server restores its results |
+| `custom_tool_call` / `custom_tool_call_output` | Supported | Only the `apply_patch` tool; images in the output are read by the vision model |
+| Other types | Ignored | — |
+
+## Tools
+
+| Type | Status | Notes |
+| --- | --- | --- |
+| `function` | Supported | — |
+| `web_search` / `web_search_2025_08_26` | Supported | Executed server-side; `search_context_size` and `user_location` are ignored; auto-continuation is capped at 10 rounds |
+| `custom` | Partially supported | Only `{"type": "custom", "name": "apply_patch"}`; another name returns `400` |
+| `file_search` / `code_interpreter` / `computer_use` / `mcp` / other built-ins | Ignored | — |
+
+## Response fields and token usage
+
+The response object matches the OpenAI Responses API structure; unsupported capabilities come back
+as fixed values (`store: false`, `previous_response_id: null`, `parallel_tool_calls: true`).
+
+- `input_tokens`, with `input_tokens_details.cached_tokens` for context cache hits.
+- `output_tokens`, with `output_tokens_details.reasoning_tokens` for chain-of-thought tokens.