Merge pull request #256 from CJackHwang/dev

全模型全渠道附件上传deepseek功能
全接口兼容性待测试

.github/workflows/release-artifacts.yml

@@ -79,7 +79,7 @@ jobs:
             CGO_ENABLED=0 GOOS="${GOOS}" GOARCH="${GOARCH}" \
               go build -trimpath -ldflags="-s -w -X ds2api/internal/version.BuildVersion=${BUILD_VERSION}" -o "${STAGE}/${BIN}" ./cmd/ds2api
 
-            cp config.example.json .env.example internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm LICENSE README.MD README.en.md "${STAGE}/"
+            cp config.example.json .env.example LICENSE README.MD README.en.md "${STAGE}/"
             cp -R static/admin "${STAGE}/static/admin"
 
             if [ "${GOOS}" = "windows" ]; then

AGENTS.md

@@ -0,0 +1,23 @@
+# AGENTS.md
+
+These rules apply to all agent-made changes in this repository.
+
+## PR Gate
+
+- Before opening or updating a PR, run the same local gates as `.github/workflows/quality-gates.yml`.
+- Required commands:
+  - `./scripts/lint.sh`
+  - `./tests/scripts/check-refactor-line-gate.sh`
+  - `./tests/scripts/run-unit-all.sh`
+  - `npm run build --prefix webui`
+
+## Go Lint Rules
+
+- Run `gofmt -w` on every changed Go file before commit or push.
+- Do not ignore error returns from I/O-style cleanup calls such as `Close`, `Flush`, `Sync`, or similar methods.
+- If a cleanup error cannot be returned, log it explicitly.
+
+## Change Scope
+
+- Keep changes additive and tightly scoped to the requested feature or bugfix.
+- Do not mix unrelated refactors into feature PRs unless they are required to make the change pass gates.

API.en.md

@@ -4,6 +4,8 @@ Language: [中文](API.md) | [English](API.en.md)
 
 This document describes the actual behavior of the current Go codebase.
 
+Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Deployment](docs/DEPLOY.en.md) / [Testing](docs/TESTING.md)
+
 ---
 
 ## Table of Contents
@@ -171,7 +173,7 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 
 ### `GET /v1/models`
 
-No auth required. Returns supported models.
+No auth required. Returns the currently supported DeepSeek native model list.
 
 **Response**:
 
@@ -182,11 +184,21 @@ No auth required. Returns supported models.
     {"id": "deepseek-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
 
+> Note: `/v1/models` returns normalized DeepSeek native model IDs. Common aliases are accepted only as request input and are not expanded as separate items in this endpoint.
+
 ### Model Alias Resolution
 
 For `chat` / `responses` / `embeddings`, DS2API follows a wide-input/strict-output policy:
@@ -209,7 +221,7 @@ Content-Type: application/json
 
 | Field | Type | Required | Notes |
 | --- | --- | --- | --- |
-| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-4o`, `gpt-5-codex`, `o3`, `claude-sonnet-4-5`, etc.) |
+| `model` | string | ✅ | DeepSeek native models + common aliases (`gpt-4o`, `gpt-5-codex`, `o3`, `claude-sonnet-4-5`, `gemini-2.5-pro`, etc.) |
 | `messages` | array | ✅ | OpenAI-style messages |
 | `stream` | boolean | ❌ | Default `false` |
 | `tools` | array | ❌ | Function calling schema |
@@ -267,6 +279,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` models emit `delta.reasoning_content`
 - Text emits `delta.content`
 - Last chunk includes `finish_reason` and `usage`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 #### Tool Calls
 
@@ -383,6 +396,7 @@ Business auth required. Returns OpenAI-compatible embeddings shape.
 ## Claude-Compatible API
 
 Besides `/anthropic/v1/*`, DS2API also supports shortcut paths: `/v1/messages`, `/messages`, `/v1/messages/count_tokens`, `/messages/count_tokens`.
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `GET /anthropic/v1/models`
 
@@ -404,7 +418,7 @@ No auth required.
 }
 ```
 
-> Note: the example is partial; the real response includes historical Claude 1.x/2.x/3.x/4.x IDs and common aliases.
+> Note: the example is partial; besides the current primary aliases, the real response also includes Claude 4.x snapshots plus historical 3.x / 2.x / 1.x IDs and common aliases.
 
 ### `POST /anthropic/v1/messages`
 
@@ -517,6 +531,7 @@ Supported paths:
 - `/v1/models/{model}:streamGenerateContent` (compat path)
 
 Authentication is the same as other business routes (`Authorization: Bearer <token>` or `x-api-key`).
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -535,6 +550,7 @@ Returns SSE (`text/event-stream`), each chunk as `data: <json>`:
 - regular text: incremental text chunks
 - `tools` mode: buffered and emitted as `functionCall` at finalize phase
 - final chunk: includes `finishReason: "STOP"` and `usageMetadata`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 ---
 

API.md

@@ -4,6 +4,8 @@
 
 本文档描述当前 Go 代码库的实际 API 行为。
 
+文档导航：[总览](README.MD) / [架构说明](docs/ARCHITECTURE.md) / [部署指南](docs/DEPLOY.md) / [测试指南](docs/TESTING.md)
+
 ---
 
 ## 目录
@@ -171,7 +173,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 
 ### `GET /v1/models`
 
-无需鉴权。返回当前支持的模型列表。
+无需鉴权。返回当前支持的 DeepSeek 原生模型列表。
 
 **响应示例**：
 
@@ -182,11 +184,21 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
     {"id": "deepseek-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-expert-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-chat-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-vision-reasoner-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
 
+> 说明：`/v1/models` 返回的是规范化后的 DeepSeek 原生模型 ID；常见 alias 仅用于请求入参解析，不会在该接口中单独展开返回。
+
 ### 模型 alias 解析策略
 
 对 `chat` / `responses` / `embeddings` 的 `model` 字段采用“宽进严出”：
@@ -209,7 +221,7 @@ Content-Type: application/json
 
 | 字段 | 类型 | 必填 | 说明 |
 | --- | --- | --- | --- |
-| `model` | string | ✅ | 支持 DeepSeek 原生模型 + 常见 alias（如 `gpt-4o`、`gpt-5-codex`、`o3`、`claude-sonnet-4-5`） |
+| `model` | string | ✅ | 支持 DeepSeek 原生模型 + 常见 alias（如 `gpt-4o`、`gpt-5-codex`、`o3`、`claude-sonnet-4-5`、`gemini-2.5-pro` 等） |
 | `messages` | array | ✅ | OpenAI 风格消息数组 |
 | `stream` | boolean | ❌ | 默认 `false` |
 | `tools` | array | ❌ | Function Calling 定义 |
@@ -267,6 +279,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` 模型输出 `delta.reasoning_content`
 - 普通文本输出 `delta.content`
 - 最后一段包含 `finish_reason` 和 `usage`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 #### Tool Calls
 
@@ -389,6 +402,7 @@ data: [DONE]
 ## Claude 兼容接口
 
 除标准路径 `/anthropic/v1/*` 外，还支持快捷路径 `/v1/messages`、`/messages`、`/v1/messages/count_tokens`、`/messages/count_tokens`。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `GET /anthropic/v1/models`
 
@@ -410,7 +424,7 @@ data: [DONE]
 }
 ```
 
-> 说明：示例仅展示部分模型；实际返回包含 Claude 1.x/2.x/3.x/4.x 历史模型 ID 与常见别名。
+> 说明：示例仅展示部分模型；实际返回除当前主别名外，还包含 Claude 4.x snapshots，以及 3.x / 2.x / 1.x 历史模型 ID 与常见别名。
 
 ### `POST /anthropic/v1/messages`
 
@@ -523,6 +537,7 @@ data: {"type":"message_stop"}
 - `/v1/models/{model}:streamGenerateContent`（兼容路径）
 
 鉴权方式同业务接口（`Authorization: Bearer <token>` 或 `x-api-key`）。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -541,6 +556,7 @@ data: {"type":"message_stop"}
 - 常规文本：持续返回增量文本 chunk
 - `tools` 场景：会缓冲并在结束时输出 `functionCall` 结构
 - 结束 chunk：包含 `finishReason: "STOP"` 与 `usageMetadata`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 ---
 

Dockerfile

@@ -34,7 +34,7 @@ CMD ["/usr/local/bin/ds2api"]
 
 FROM runtime-base AS runtime-from-source
 COPY --from=go-builder /out/ds2api /usr/local/bin/ds2api
-COPY --from=go-builder /app/internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
+
 COPY --from=go-builder /app/config.example.json /app/config.example.json
 COPY --from=webui-builder /app/static/admin /app/static/admin
 
@@ -53,13 +53,13 @@ RUN set -eux; \
     test -n "${PKG_DIR}"; \
     mkdir -p /out/static; \
     cp "${PKG_DIR}/ds2api" /out/ds2api; \
-    cp "${PKG_DIR}/sha3_wasm_bg.7b9ca65ddd.wasm" /out/sha3_wasm_bg.7b9ca65ddd.wasm; \
+
     cp "${PKG_DIR}/config.example.json" /out/config.example.json; \
     cp -R "${PKG_DIR}/static/admin" /out/static/admin
 
 FROM runtime-base AS runtime-from-dist
 COPY --from=dist-extract /out/ds2api /usr/local/bin/ds2api
-COPY --from=dist-extract /out/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
+
 COPY --from=dist-extract /out/config.example.json /app/config.example.json
 COPY --from=dist-extract /out/static/admin /app/static/admin
 

README.MD

@@ -16,6 +16,10 @@
 
 将 DeepSeek Web 对话能力转换为 OpenAI、Claude 与 Gemini 兼容 API。后端为 **Go 全量实现**，前端为 React WebUI 管理台（源码在 `webui/`，部署时自动构建到 `static/admin`）。
 
+文档入口：[文档导航](docs/README.md) / [架构说明](docs/ARCHITECTURE.md) / [接口文档](API.md)
+
+【感谢Linux.do社区及GitHub社区各位开发者对项目的支持与贡献】
+
 > **重要免责声明**
 >
 > 本仓库仅供学习、研究、个人实验和内部验证使用，不提供任何形式的商业授权、适用性保证或结果保证。
@@ -24,7 +28,7 @@
 >
 > 请勿将本项目用于违反服务条款、协议、法律法规或平台规则的场景。商业使用前请自行确认 `LICENSE`、相关协议以及你是否获得了作者的书面许可。
 
-## 架构概览
+## 架构概览（摘要）
 
 ```mermaid
 flowchart LR
@@ -48,7 +52,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(并发槽位 + 等待队列)"]
             DSClient["DeepSeek Client\n(Session / Auth / HTTP)"]
-            Pow["PoW WASM\n(wazero 预加载)"]
+            Pow["PoW 实现\n(纯 Go 毫秒级)"]
             Tool["Tool Sieve\n(Go/Node 语义对齐)"]
         end
     end
@@ -72,6 +76,8 @@ flowchart LR
     Bridge --> Client
 ```
 
+详细架构拆分与目录职责见 [docs/ARCHITECTURE.md](docs/ARCHITECTURE.md)。
+
 - **后端**：Go（`cmd/ds2api/`、`api/`、`internal/`），不依赖 Python 运行时
 - **前端**：React 管理台（`webui/`），运行时托管静态构建产物
 - **部署**：本地运行、Docker、Vercel Serverless、Linux systemd
@@ -95,7 +101,7 @@ flowchart LR
 | Gemini 兼容 | `POST /v1beta/models/{model}:generateContent`、`POST /v1beta/models/{model}:streamGenerateContent`（及 `/v1/models/{model}:*` 路径） |
 | 多账号轮询 | 自动 token 刷新、邮箱/手机号双登录方式 |
 | 并发队列控制 | 每账号 in-flight 上限 + 等待队列，动态计算建议并发值 |
-| DeepSeek PoW | WASM 计算（`wazero`），无需外部 Node.js 依赖 |
+| DeepSeek PoW | 纯 Go 高性能实现（DeepSeekHashV1），毫秒级响应 |
 | Tool Calling | 防泄漏处理：非代码块高置信特征识别、`delta.tool_calls` 早发、结构化增量输出 |
 | Admin API | 配置管理、运行时设置热更新、账号测试 / 批量测试、会话清理、导入导出、Vercel 同步、版本检查 |
 | WebUI 管理台 | `/admin` 单页应用（中英文双语、深色模式） |
@@ -114,26 +120,35 @@ flowchart LR
 
 ## 模型支持
 
-### OpenAI 接口
+### OpenAI 接口（`GET /v1/models`）
 
-| 模型 | thinking | search |
-| --- | --- | --- |
-| `deepseek-chat` | ❌ | ❌ |
-| `deepseek-reasoner` | ✅ | ❌ |
-| `deepseek-chat-search` | ❌ | ✅ |
-| `deepseek-reasoner-search` | ✅ | ✅ |
+| 模型类型 | 模型 ID | thinking | search |
+| --- | --- | --- | --- |
+| default | `deepseek-chat` | ❌ | ❌ |
+| default | `deepseek-reasoner` | ✅ | ❌ |
+| default | `deepseek-chat-search` | ❌ | ✅ |
+| default | `deepseek-reasoner-search` | ✅ | ✅ |
+| expert | `deepseek-expert-chat` | ❌ | ❌ |
+| expert | `deepseek-expert-reasoner` | ✅ | ❌ |
+| expert | `deepseek-expert-chat-search` | ❌ | ✅ |
+| expert | `deepseek-expert-reasoner-search` | ✅ | ✅ |
+| vision | `deepseek-vision-chat` | ❌ | ❌ |
+| vision | `deepseek-vision-reasoner` | ✅ | ❌ |
+| vision | `deepseek-vision-chat-search` | ❌ | ✅ |
+| vision | `deepseek-vision-reasoner-search` | ✅ | ✅ |
 
-### Claude 接口
+除原生模型外，也支持常见 alias 输入（如 `gpt-4o`、`gpt-5-codex`、`o3`、`claude-sonnet-4-5`、`gemini-2.5-pro` 等），但 `/v1/models` 返回的是规范化后的 DeepSeek 原生模型 ID。
 
-| 模型 | 默认映射 |
+### Claude 接口（`GET /anthropic/v1/models`）
+
+| 当前常用模型 | 默认映射 |
 | --- | --- |
 | `claude-sonnet-4-5` | `deepseek-chat` |
 | `claude-haiku-4-5`（兼容 `claude-3-5-haiku-latest`） | `deepseek-chat` |
 | `claude-opus-4-6` | `deepseek-reasoner` |
 
 可通过配置中的 `claude_mapping` 或 `claude_model_mapping` 覆盖映射关系。
-另外，`/anthropic/v1/models` 现已包含 Claude 1.x/2.x/3.x/4.x 历史模型 ID 与常见别名，便于旧客户端直接兼容。
-
+`/anthropic/v1/models` 除上述当前主别名外，还会返回 Claude 4.x snapshots，以及 3.x / 2.x / 1.x 历史模型 ID 与常见 alias，便于旧客户端直接兼容。
 
 #### Claude Code 接入避坑（实测）
 
@@ -148,6 +163,15 @@ Gemini 适配器将模型名通过 `model_aliases` 或内置规则映射到 Deep
 
 ## 快速开始
 
+### 部署方式优先级建议
+
+推荐按以下顺序选择部署方式：
+
+1. **下载 Release 构建包运行**：最省事，产物已编译完成，最适合大多数用户。
+2. **Docker / GHCR 镜像部署**：适合需要容器化、编排或云环境部署。
+3. **Vercel 部署**：适合已有 Vercel 环境且接受其平台约束的场景。
+4. **本地源码运行 / 自行编译**：适合开发、调试或需要自行修改代码的场景。
+
 ### 通用第一步（所有部署方式）
 
 把 `config.json` 作为唯一配置源（推荐做法）：
@@ -161,29 +185,19 @@ cp config.example.json config.json
 - 本地运行：直接读取 `config.json`
 - Docker / Vercel：由 `config.json` 生成 `DS2API_CONFIG_JSON`（Base64）注入环境变量，也可以直接写原始 JSON
 
-### 方式一：本地运行
+### 方式一：下载 Release 构建包
 
-**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）
+每次发布 Release 时，GitHub Actions 会自动构建多平台二进制包：
 
 ```bash
-# 1. 克隆仓库
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
-
-# 2. 配置
+# 下载对应平台的压缩包后
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 cp config.example.json config.json
-# 编辑 config.json，填入你的 DeepSeek 账号信息和 API key
-
-# 3. 启动
-go run ./cmd/ds2api
+# 编辑 config.json
+./ds2api
 ```
 
-默认本地访问地址：`http://127.0.0.1:5001`
-
-服务实际绑定：`0.0.0.0:5001`，因此同一局域网设备通常也可以通过你的内网 IP 访问。
-
-> **WebUI 自动构建**：本地首次启动时，若 `static/admin` 不存在，会自动尝试执行 `npm ci`（仅在缺少依赖时）和 `npm run build -- --outDir static/admin --emptyOutDir`（需要本机有 Node.js）。你也可以手动构建：`./scripts/build-webui.sh`
-
 ### 方式二：Docker 运行
 
 ```bash
@@ -237,35 +251,28 @@ base64 < config.json | tr -d '\n'
 
 详细部署说明请参阅 [部署指南](docs/DEPLOY.md)。
 
-### 方式四：下载 Release 构建包
+### 方式四：本地源码运行
 
-每次发布 Release 时，GitHub Actions 会自动构建多平台二进制包：
+**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）
 
 ```bash
-# 下载对应平台的压缩包后
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 1. 克隆仓库
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
+
+# 2. 配置
 cp config.example.json config.json
-# 编辑 config.json
-./ds2api
+# 编辑 config.json，填入你的 DeepSeek 账号信息和 API key
+
+# 3. 启动
+go run ./cmd/ds2api
 ```
 
-### 方式五：OpenCode CLI 接入
+默认本地访问地址：`http://127.0.0.1:5001`
 
-1. 复制示例配置：
+服务实际绑定：`0.0.0.0:5001`，因此同一局域网设备通常也可以通过你的内网 IP 访问。
 
-```bash
-cp opencode.json.example opencode.json
-```
-
-2. 编辑 `opencode.json`：
-- 将 `baseURL` 改为你的 DS2API 地址（例如 `https://your-domain.com/v1`）
-- 将 `apiKey` 改为你的 DS2API key（对应 `config.keys`）
-
-3. 在项目目录启动 OpenCode CLI（按你的安装方式运行 `opencode`）。
-
-> 建议优先使用 OpenAI 兼容路径（`/v1/*`），即示例里的 `@ai-sdk/openai-compatible` provider。
-> 若客户端支持 `wire_api`，可分别测试 `responses` 与 `chat`，DS2API 两条链路都兼容。
+> **WebUI 自动构建**：本地首次启动时，若 `static/admin` 不存在，会自动尝试执行 `npm ci`（仅在缺少依赖时）和 `npm run build -- --outDir static/admin --emptyOutDir`（需要本机有 Node.js）。你也可以手动构建：`./scripts/build-webui.sh`
 
 ## 配置说明
 
@@ -344,7 +351,6 @@ cp opencode.json.example opencode.json
 | `DS2API_CONFIG_PATH` | 配置文件路径 | `config.json` |
 | `DS2API_CONFIG_JSON` | 直接注入配置（JSON 或 Base64） | — |
 | `DS2API_ENV_WRITEBACK` | 环境变量模式下自动写回配置文件并切换文件模式（`1/true/yes/on`） | 关闭 |
-| `DS2API_WASM_PATH` | PoW WASM 文件路径 | 自动查找 |
 | `DS2API_STATIC_ADMIN_DIR` | 管理台静态文件目录 | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | 启动时自动构建 WebUI | 本地开启，Vercel 关闭 |
 | `DS2API_DEV_PACKET_CAPTURE` | 本地开发抓包开关（记录最近会话请求/响应体） | 本地非 Vercel 默认开启 |
@@ -432,72 +438,6 @@ go run ./cmd/ds2api
 {"query":"广州天气","sample_id":"gz-weather-from-memory"}
 ```
 
-## 项目结构
-
-```text
-ds2api/
-├── app/                     # 统一 HTTP Handler 组装层（供本地与 Serverless 复用）
-├── cmd/
-│   ├── ds2api/              # 本地 / 容器启动入口
-│   └── ds2api-tests/        # 端到端测试集入口
-├── api/
-│   ├── index.go             # Vercel Serverless Go 入口
-│   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # 账号池与并发队列
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI 兼容适配器（含 Tool Call 解析、Vercel 流式 prepare/release）
-│   │   ├── claude/          # Claude 兼容适配器
-│   │   └── gemini/          # Gemini 兼容适配器（generateContent / streamGenerateContent）
-│   ├── admin/               # Admin API handlers（含 Settings 热更新）
-│   ├── auth/                # 鉴权与 JWT
-│   ├── claudeconv/          # Claude 消息格式转换
-│   ├── compat/              # Go 版本兼容与回归测试辅助
-│   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek API 客户端、PoW WASM
-│   ├── js/                  # Node 运行时流式处理与兼容逻辑
-│   ├── devcapture/          # 开发抓包模块
-│   ├── rawsample/           # 原始流样本可见文本提取与回放辅助
-│   ├── format/              # 输出格式化
-│   ├── prompt/              # Prompt 构建
-│   ├── server/              # HTTP 路由与中间件（chi router）
-│   ├── sse/                 # SSE 解析工具
-│   ├── stream/              # 统一流式消费引擎
-│   ├── testsuite/           # 端到端测试框架与用例编排
-│   ├── translatorcliproxy/  # CLIProxy 桥接与流写入组件
-│   ├── toolcall/            # Tool Call 解析、修复与格式化（核心业务逻辑）
-│   ├── util/                # 通用工具函数（Token 估算、JSON 辅助等）
-│   ├── version/             # 版本解析 / 比较与 tag 规范化
-│   └── webui/               # WebUI 静态文件托管与自动构建
-├── webui/                   # React WebUI 源码（Vite + Tailwind）
-│   └── src/
-│       ├── app/             # 路由、鉴权、配置状态管理
-│       ├── features/        # 业务功能模块（account/settings/vercel/apiTester）
-│       ├── components/      # 登录/落地页等通用组件
-│       └── locales/         # 中英文语言包（zh.json / en.json）
-├── scripts/
-│   └── build-webui.sh       # WebUI 手动构建脚本
-├── tests/
-│   ├── compat/              # 兼容性测试夹具与期望输出
-│   ├── node/                # Node 侧单元测试（chat-stream / tool-sieve）
-│   ├── raw_stream_samples/  # 原始 SSE 样本与回放元数据
-│   └── scripts/             # 统一测试脚本入口（unit/e2e）
-├── docs/                    # 部署 / 贡献 / 测试等辅助文档
-├── static/admin/            # WebUI 构建产物（不提交到 Git）
-├── .github/
-│   ├── workflows/           # GitHub Actions（质量门禁 + Release 自动构建）
-│   ├── ISSUE_TEMPLATE/      # Issue 模板
-│   └── PULL_REQUEST_TEMPLATE.md
-├── config.example.json      # 配置文件示例
-├── .env.example             # 环境变量示例
-├── Dockerfile               # 多阶段构建（WebUI + Go）
-├── docker-compose.yml       # 生产环境 Docker Compose
-├── docker-compose.dev.yml   # 开发环境 Docker Compose
-├── vercel.json              # Vercel 路由与构建配置
-└── go.mod / go.sum          # Go 模块依赖
-```
-
 ## 文档索引
 
 | 文档 | 说明 |

README.en.md

@@ -16,6 +16,8 @@ Language: [中文](README.MD) | [English](README.en.md)
 
 DS2API converts DeepSeek Web chat capability into OpenAI-compatible, Claude-compatible, and Gemini-compatible APIs. The backend is a **pure Go implementation**, with a React WebUI admin panel (source in `webui/`, build output auto-generated to `static/admin` during deployment).
 
+Documentation entry: [Docs Index](docs/README.md) / [Architecture](docs/ARCHITECTURE.en.md) / [API Reference](API.en.md)
+
 > **Important Disclaimer**
 >
 > This repository is provided for learning, research, personal experimentation, and internal validation only. It does not grant any commercial authorization and comes with no warranty of fitness, stability, or results.
@@ -24,7 +26,7 @@ DS2API converts DeepSeek Web chat capability into OpenAI-compatible, Claude-comp
 >
 > Do not use this project in ways that violate service terms, agreements, laws, or platform rules. Before any commercial use, review the `LICENSE`, the relevant terms, and confirm that you have the author's written permission.
 
-## Architecture Overview
+## Architecture Overview (Summary)
 
 ```mermaid
 flowchart LR
@@ -48,7 +50,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(in-flight slots + wait queue)"]
             DSClient["DeepSeek Client\n(session / auth / HTTP)"]
-            Pow["PoW WASM\n(wazero preload)"]
+            Pow["PoW Solver\n(Pure Go ms-level)"]
             Tool["Tool Sieve\n(Go/Node semantic parity)"]
         end
     end
@@ -72,6 +74,8 @@ flowchart LR
     Bridge --> Client
 ```
 
+For the full module-by-module architecture and directory responsibilities, see [docs/ARCHITECTURE.en.md](docs/ARCHITECTURE.en.md).
+
 - **Backend**: Go (`cmd/ds2api/`, `api/`, `internal/`), no Python runtime
 - **Frontend**: React admin panel (`webui/`), served as static build at runtime
 - **Deployment**: local run, Docker, Vercel serverless, Linux systemd
@@ -95,7 +99,7 @@ flowchart LR
 | Gemini compatible | `POST /v1beta/models/{model}:generateContent`, `POST /v1beta/models/{model}:streamGenerateContent` (plus `/v1/models/{model}:*` paths) |
 | Multi-account rotation | Auto token refresh, email/mobile dual login |
 | Concurrency control | Per-account in-flight limit + waiting queue, dynamic recommended concurrency |
-| DeepSeek PoW | WASM solving via `wazero`, no external Node.js dependency |
+| DeepSeek PoW | Pure Go high-performance solver (DeepSeekHashV1), ms-level response |
 | Tool Calling | Anti-leak handling: non-code-block feature match, early `delta.tool_calls`, structured incremental output |
 | Admin API | Config management, runtime settings hot-reload, account testing/batch test, session cleanup, import/export, Vercel sync, version check |
 | WebUI Admin Panel | SPA at `/admin` (bilingual Chinese/English, dark mode) |
@@ -114,26 +118,35 @@ flowchart LR
 
 ## Model Support
 
-### OpenAI Endpoint
+### OpenAI Endpoint (`GET /v1/models`)
 
-| Model | thinking | search |
-| --- | --- | --- |
-| `deepseek-chat` | ❌ | ❌ |
-| `deepseek-reasoner` | ✅ | ❌ |
-| `deepseek-chat-search` | ❌ | ✅ |
-| `deepseek-reasoner-search` | ✅ | ✅ |
+| Family | Model ID | thinking | search |
+| --- | --- | --- | --- |
+| default | `deepseek-chat` | ❌ | ❌ |
+| default | `deepseek-reasoner` | ✅ | ❌ |
+| default | `deepseek-chat-search` | ❌ | ✅ |
+| default | `deepseek-reasoner-search` | ✅ | ✅ |
+| expert | `deepseek-expert-chat` | ❌ | ❌ |
+| expert | `deepseek-expert-reasoner` | ✅ | ❌ |
+| expert | `deepseek-expert-chat-search` | ❌ | ✅ |
+| expert | `deepseek-expert-reasoner-search` | ✅ | ✅ |
+| vision | `deepseek-vision-chat` | ❌ | ❌ |
+| vision | `deepseek-vision-reasoner` | ✅ | ❌ |
+| vision | `deepseek-vision-chat-search` | ❌ | ✅ |
+| vision | `deepseek-vision-reasoner-search` | ✅ | ✅ |
 
-### Claude Endpoint
+Besides native IDs, DS2API also accepts common aliases as input (for example `gpt-4o`, `gpt-5-codex`, `o3`, `claude-sonnet-4-5`, `gemini-2.5-pro`), but `/v1/models` returns normalized DeepSeek native model IDs.
 
-| Model | Default Mapping |
+### Claude Endpoint (`GET /anthropic/v1/models`)
+
+| Current common model | Default Mapping |
 | --- | --- |
 | `claude-sonnet-4-5` | `deepseek-chat` |
 | `claude-haiku-4-5` (compatible with `claude-3-5-haiku-latest`) | `deepseek-chat` |
 | `claude-opus-4-6` | `deepseek-reasoner` |
 
 Override mapping via `claude_mapping` or `claude_model_mapping` in config.
-In addition, `/anthropic/v1/models` now includes historical Claude 1.x/2.x/3.x/4.x IDs and common aliases for legacy client compatibility.
-
+Besides the current primary aliases above, `/anthropic/v1/models` also returns Claude 4.x snapshots plus historical 3.x / 2.x / 1.x IDs and common aliases for legacy client compatibility.
 
 #### Claude Code integration pitfalls (validated)
 
@@ -148,6 +161,15 @@ The Gemini adapter maps model names to DeepSeek native models via `model_aliases
 
 ## Quick Start
 
+### Recommended deployment priority
+
+Recommended order when choosing a deployment method:
+
+1. **Download and run release binaries**: the easiest path for most users because the artifacts are already built.
+2. **Docker / GHCR image deployment**: suitable for containerized, orchestrated, or cloud environments.
+3. **Vercel deployment**: suitable if you already use Vercel and accept its platform constraints.
+4. **Run from source / build locally**: suitable for development, debugging, or when you need to modify the code yourself.
+
 ### Universal First Step (all deployment modes)
 
 Use `config.json` as the single source of truth (recommended):
@@ -161,47 +183,37 @@ Recommended per deployment mode:
 - Local run: read `config.json` directly
 - Docker / Vercel: generate Base64 from `config.json` and inject as `DS2API_CONFIG_JSON`, or paste raw JSON directly
 
-### Option 1: Local Run
+### Option 1: Download Release Binaries
 
-**Prerequisites**: Go 1.26+, Node.js `20.19+` or `22.12+` (only if building WebUI locally)
+GitHub Actions automatically builds multi-platform archives on each Release:
 
 ```bash
-# 1. Clone
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
-
-# 2. Configure
+# After downloading the archive for your platform
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 cp config.example.json config.json
-# Edit config.json with your DeepSeek account info and API keys
-
-# 3. Start
-go run ./cmd/ds2api
+# Edit config.json
+./ds2api
 ```
 
-Default local URL: `http://127.0.0.1:5001`
-
-The server actually binds to `0.0.0.0:5001`, so devices on the same LAN can usually reach it through your private IP as well.
-
-> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm ci` (only when dependencies are missing) and `npm run build -- --outDir static/admin --emptyOutDir` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
-
-### Option 2: Docker
+### Option 2: Docker / GHCR
 
 ```bash
-# 1. Prepare env file and config file
+# Pull prebuilt image
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
+# Or run a pinned version
+# docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+
+# Prepare env file and config file
 cp .env.example .env
 cp config.example.json config.json
 
-# 2. Edit .env (at least set DS2API_ADMIN_KEY; optionally set DS2API_HOST_PORT to change the host port)
-#    DS2API_ADMIN_KEY=replace-with-a-strong-secret
-
-# 3. Start
+# Start with compose
 docker-compose up -d
-
-# 4. View logs
-docker-compose logs -f
 ```
 
-The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
+The default `docker-compose.yml` uses `ghcr.io/cjackhwang/ds2api:latest` and maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
 
 Rebuild after updates: `docker-compose up -d --build`
 
@@ -237,35 +249,28 @@ base64 < config.json | tr -d '\n'
 
 For detailed deployment instructions, see the [Deployment Guide](docs/DEPLOY.en.md).
 
-### Option 4: Download Release Binaries
+### Option 4: Local Run
 
-GitHub Actions automatically builds multi-platform archives on each Release:
+**Prerequisites**: Go 1.26+, Node.js `20.19+` or `22.12+` (only if building WebUI locally)
 
 ```bash
-# After downloading the archive for your platform
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 1. Clone
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
+
+# 2. Configure
 cp config.example.json config.json
-# Edit config.json
-./ds2api
+# Edit config.json with your DeepSeek account info and API keys
+
+# 3. Start
+go run ./cmd/ds2api
 ```
 
-### Option 5: OpenCode CLI
+Default local URL: `http://127.0.0.1:5001`
 
-1. Copy the example config:
+The server actually binds to `0.0.0.0:5001`, so devices on the same LAN can usually reach it through your private IP as well.
 
-```bash
-cp opencode.json.example opencode.json
-```
-
-2. Edit `opencode.json`:
-- Set `baseURL` to your DS2API endpoint (for example, `https://your-domain.com/v1`)
-- Set `apiKey` to your DS2API key (from `config.keys`)
-
-3. Start OpenCode CLI in the project directory (run `opencode` using your installed method).
-
-> Recommended: use the OpenAI-compatible path (`/v1/*`) via `@ai-sdk/openai-compatible` as shown in the example.
-> If your client supports `wire_api`, test both `responses` and `chat`; DS2API supports both paths.
+> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm ci` (only when dependencies are missing) and `npm run build -- --outDir static/admin --emptyOutDir` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
 
 ## Configuration
 
@@ -344,7 +349,6 @@ cp opencode.json.example opencode.json
 | `DS2API_CONFIG_PATH` | Config file path | `config.json` |
 | `DS2API_CONFIG_JSON` | Inline config (JSON or Base64) | — |
 | `DS2API_ENV_WRITEBACK` | Auto-write env-backed config to file and transition to file mode (`1/true/yes/on`) | Disabled |
-| `DS2API_WASM_PATH` | PoW WASM file path | Auto-detect |
 | `DS2API_STATIC_ADMIN_DIR` | Admin static assets dir | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | Auto-build WebUI on startup | Enabled locally, disabled on Vercel |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | Max in-flight requests per account | `2` |
@@ -430,72 +434,6 @@ The save endpoint can target a chain by `query`, `chain_key`, or `capture_id`. E
 {"query":"Guangzhou weather","sample_id":"gz-weather-from-memory"}
 ```
 
-## Project Structure
-
-```text
-ds2api/
-├── app/                     # Unified HTTP handler assembly (shared by local + serverless)
-├── cmd/
-│   ├── ds2api/              # Local / container entrypoint
-│   └── ds2api-tests/        # End-to-end testsuite entrypoint
-├── api/
-│   ├── index.go             # Vercel Serverless Go entry
-│   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # Account pool and concurrency queue
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI adapter (incl. tool call parsing, Vercel stream prepare/release)
-│   │   ├── claude/          # Claude adapter
-│   │   └── gemini/          # Gemini adapter (generateContent / streamGenerateContent)
-│   ├── admin/               # Admin API handlers (incl. Settings hot-reload)
-│   ├── auth/                # Auth and JWT
-│   ├── claudeconv/          # Claude message format conversion
-│   ├── compat/              # Go-version compatibility and regression helpers
-│   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek API client, PoW WASM
-│   ├── js/                  # Node runtime stream/compat logic
-│   ├── devcapture/          # Dev packet capture module
-│   ├── rawsample/           # Visible-text extraction and replay helpers for raw stream samples
-│   ├── format/              # Output formatting
-│   ├── prompt/              # Prompt construction
-│   ├── server/              # HTTP routing and middleware (chi router)
-│   ├── sse/                 # SSE parsing utilities
-│   ├── stream/              # Unified stream consumption engine
-│   ├── testsuite/           # End-to-end testsuite framework and case orchestration
-│   ├── translatorcliproxy/  # CLIProxy bridge and stream writer components
-│   ├── toolcall/            # Tool Call parsing, repair, and formatting (core business logic)
-│   ├── util/                # Common utilities (Token estimation, JSON helpers, etc.)
-│   ├── version/             # Version parsing/comparison and tag normalization
-│   └── webui/               # WebUI static file serving and auto-build
-├── webui/                   # React WebUI source (Vite + Tailwind)
-│   └── src/
-│       ├── app/             # Routing, auth, config state
-│       ├── features/        # Feature modules (account/settings/vercel/apiTester)
-│       ├── components/      # Shared UI pieces (login/landing, etc.)
-│       └── locales/         # Language packs (zh.json / en.json)
-├── scripts/
-│   └── build-webui.sh       # Manual WebUI build script
-├── tests/
-│   ├── compat/              # Compatibility fixtures and expected outputs
-│   ├── node/                # Node-side unit tests (chat-stream / tool-sieve)
-│   ├── raw_stream_samples/  # Raw SSE samples and replay metadata
-│   └── scripts/             # Unified test script entrypoints (unit/e2e)
-├── docs/                    # Deployment / contributing / testing docs
-├── static/admin/            # WebUI build output (not committed to Git)
-├── .github/
-│   ├── workflows/           # GitHub Actions (quality gates + release automation)
-│   ├── ISSUE_TEMPLATE/      # Issue templates
-│   └── PULL_REQUEST_TEMPLATE.md
-├── config.example.json      # Config file template
-├── .env.example             # Environment variable template
-├── Dockerfile               # Multi-stage build (WebUI + Go)
-├── docker-compose.yml       # Production Docker Compose
-├── docker-compose.dev.yml   # Development Docker Compose
-├── vercel.json              # Vercel routing and build config
-└── go.mod / go.sum          # Go module dependencies
-```
-
 ## Documentation Index
 
 | Document | Description |

VERSION

@@ -1 +1 @@
-3.1.2
+3.4.0

docs/ARCHITECTURE.en.md

@@ -0,0 +1,136 @@
+# DS2API Architecture & Project Layout
+
+Language: [中文](ARCHITECTURE.md) | [English](ARCHITECTURE.en.md)
+
+> This file is the single architecture source for directory layout, module boundaries, and execution flow.
+
+## 1. Top-level Layout (expanded)
+
+> Notes: this is the **fully expanded** project directory list (excluding metadata/dependency dirs such as `.git/` and `webui/node_modules/`), with each folder annotated by purpose.
+
+```text
+ds2api/
+├── .github/                              # GitHub collaboration and CI config
+│   ├── ISSUE_TEMPLATE/                   # Issue templates
+│   └── workflows/                        # GitHub Actions workflows
+├── api/                                  # Serverless entrypoints (Vercel Go/Node)
+├── app/                                  # Application-level handler assembly
+├── cmd/                                  # Executable entrypoints
+│   ├── ds2api/                           # Main service bootstrap
+│   └── ds2api-tests/                     # E2E testsuite CLI bootstrap
+├── docs/                                 # Project documentation
+├── internal/                             # Core implementation (non-public packages)
+│   ├── account/                          # Account pool, inflight slots, waiting queue
+│   ├── adapter/                          # Multi-protocol adapters
+│   │   ├── claude/                       # Claude protocol adapter
+│   │   ├── gemini/                       # Gemini protocol adapter
+│   │   └── openai/                       # OpenAI adapter and shared execution core
+│   ├── admin/                            # Admin API (config/accounts/ops)
+│   ├── auth/                             # Auth/JWT/credential resolution
+│   ├── claudeconv/                       # Claude message conversion helpers
+│   ├── compat/                           # Compatibility and regression helpers
+│   ├── config/                           # Config loading/validation/hot reload
+│   ├── deepseek/                         # DeepSeek upstream client capabilities
+│   │   └── transport/                    # DeepSeek transport details
+│   ├── devcapture/                       # Dev capture and troubleshooting
+│   ├── format/                           # Response formatting layer
+│   │   ├── claude/                       # Claude output formatting
+│   │   └── openai/                       # OpenAI output formatting
+│   ├── js/                               # Node runtime related logic
+│   │   ├── chat-stream/                  # Node streaming bridge
+│   │   ├── helpers/                      # JS helper modules
+│   │   │   └── stream-tool-sieve/        # JS implementation of tool sieve
+│   │   └── shared/                       # Shared semantics between Go/Node
+│   ├── prompt/                           # Prompt composition
+│   ├── rawsample/                        # Raw sample read/write and management
+│   ├── server/                           # Router and middleware assembly
+│   ├── sse/                              # SSE parsing utilities
+│   ├── stream/                           # Unified stream consumption engine
+│   ├── testsuite/                        # Testsuite execution framework
+│   ├── textclean/                        # Text cleanup
+│   ├── toolcall/                         # Tool-call parsing and repair
+│   ├── translatorcliproxy/               # Cross-protocol translation bridge
+│   ├── util/                             # Shared utility helpers
+│   ├── version/                          # Version query/compare
+│   └── webui/                            # WebUI static hosting logic
+├── plans/                                # Stage plans and manual QA records
+├── pow/                                  # PoW standalone implementation + benchmarks
+├── scripts/                              # Build/release helper scripts
+├── tests/                                # Test assets and scripts
+│   ├── compat/                           # Compatibility fixtures + expected outputs
+│   │   ├── expected/                     # Expected output samples
+│   │   └── fixtures/                     # Fixture inputs
+│   │       ├── sse_chunks/               # SSE chunk fixtures
+│   │       └── toolcalls/                # Tool-call fixtures
+│   ├── node/                             # Node unit tests
+│   ├── raw_stream_samples/               # Upstream raw SSE samples
+│   │   ├── content-filter-trigger-20260405-jwt3/          # Content-filter terminal sample
+│   │   ├── continue-thinking-snapshot-replay-20260405/    # Continue-thinking sample
+│   │   ├── guangzhou-weather-reasoner-search-20260404/    # Search/reference sample
+│   │   ├── markdown-format-example-20260405/              # Markdown sample
+│   │   └── markdown-format-example-20260405-spacefix/     # Space-fix sample
+│   ├── scripts/                          # Test entry scripts
+│   └── tools/                            # Testing helper tools
+└── webui/                                # React admin console source
+    ├── public/                           # Static assets
+    └── src/                              # Frontend source code
+        ├── app/                          # Routing/state scaffolding
+        ├── components/                   # Shared UI components
+        ├── features/                     # Feature modules
+        │   ├── account/                  # Account management page
+        │   ├── apiTester/                # API tester page
+        │   ├── settings/                 # Settings page
+        │   └── vercel/                   # Vercel sync page
+        ├── layout/                       # Layout components
+        ├── locales/                      # i18n strings
+        └── utils/                        # Frontend utilities
+```
+
+## 2. Primary Request Flow
+
+```mermaid
+flowchart LR
+    C[Client/SDK] --> R[internal/server/router.go]
+    R --> OA[OpenAI Adapter]
+    R --> CA[Claude Adapter]
+    R --> GA[Gemini Adapter]
+    R --> AD[Admin API]
+
+    CA --> BR[translatorcliproxy]
+    GA --> BR
+    BR --> CORE[internal/adapter/openai ChatCompletions]
+    OA --> CORE
+
+    CORE --> AUTH[internal/auth + config key/account resolver]
+    CORE --> POOL[internal/account queue + concurrency]
+    CORE --> TOOL[internal/toolcall parser + sieve]
+    CORE --> DS[internal/deepseek client]
+    DS --> U[DeepSeek upstream]
+```
+
+## 3. Responsibilities in `internal/`
+
+- `internal/server`: router tree + middlewares (health, protocol routes, Admin/WebUI).
+- `internal/adapter/openai`: shared execution core (chat/responses/embeddings + tool semantics).
+- `internal/adapter/{claude,gemini}`: protocol wrappers only (no duplicated upstream execution).
+- `internal/translatorcliproxy`: structure translation between Claude/Gemini and OpenAI.
+- `internal/deepseek`: upstream request/session/PoW/SSE handling.
+- `internal/stream` + `internal/sse`: stream parsing and incremental assembly.
+- `internal/toolcall`: JSON/XML/invoke/text-kv tool-call parsing + anti-leak sieve.
+- `internal/admin`: config/accounts/vercel sync/version/dev-capture endpoints.
+- `internal/config`: config loading/validation + runtime settings hot-reload.
+- `internal/account`: managed account pool, inflight slots, waiting queue.
+
+## 4. WebUI Runtime Relation
+
+- `webui/` stores frontend source (Vite + React).
+- Runtime serves static output from `static/admin`.
+- On first local startup, if `static/admin` is missing, DS2API may auto-build it (Node.js required).
+
+## 5. Documentation Split Strategy
+
+- Onboarding & quick start: `README.MD` / `README.en.md`
+- Architecture & layout: `docs/ARCHITECTURE*.md` (this file)
+- API contracts: `API.md` / `API.en.md`
+- Deployment/testing/contributing: `docs/DEPLOY*`, `docs/TESTING.md`, `docs/CONTRIBUTING*`
+- Deep topics: `docs/toolcall-semantics.md`, `docs/DeepSeekSSE行为结构说明-2026-04-05.md`

docs/ARCHITECTURE.md

@@ -0,0 +1,136 @@
+# DS2API 架构与项目结构说明
+
+语言 / Language: [中文](ARCHITECTURE.md) | [English](ARCHITECTURE.en.md)
+
+> 本文档用于集中维护“代码目录结构 + 模块边界 + 主链路调用关系”。
+
+## 1. 顶层目录结构（展开）
+
+> 说明：以下为仓库内业务相关目录的**完整展开**（排除 `.git/` 与 `webui/node_modules/` 这类依赖/元数据目录），并标注每个文件夹作用。
+
+```text
+ds2api/
+├── .github/                              # GitHub 协作与 CI 配置
+│   ├── ISSUE_TEMPLATE/                   # Issue 模板
+│   └── workflows/                        # GitHub Actions 工作流
+├── api/                                  # Serverless 入口（Vercel Go/Node）
+├── app/                                  # 应用级 handler 装配层
+├── cmd/                                  # 可执行程序入口
+│   ├── ds2api/                           # 主服务启动入口
+│   └── ds2api-tests/                     # E2E 测试集 CLI 入口
+├── docs/                                 # 项目文档目录
+├── internal/                             # 核心业务实现（不对外暴露）
+│   ├── account/                          # 账号池、并发槽位、等待队列
+│   ├── adapter/                          # 多协议适配层
+│   │   ├── claude/                       # Claude 协议适配
+│   │   ├── gemini/                       # Gemini 协议适配
+│   │   └── openai/                       # OpenAI 协议与统一执行核心
+│   ├── admin/                            # Admin API（配置/账号/运维）
+│   ├── auth/                             # 鉴权/JWT/凭证解析
+│   ├── claudeconv/                       # Claude 消息格式转换工具
+│   ├── compat/                           # 兼容性辅助与回归支持
+│   ├── config/                           # 配置加载、校验、热更新
+│   ├── deepseek/                         # DeepSeek 上游客户端能力
+│   │   └── transport/                    # DeepSeek 传输层细节
+│   ├── devcapture/                       # 开发抓包与调试采集
+│   ├── format/                           # 响应格式化层
+│   │   ├── claude/                       # Claude 输出格式化
+│   │   └── openai/                       # OpenAI 输出格式化
+│   ├── js/                               # Node Runtime 相关逻辑
+│   │   ├── chat-stream/                  # Node 流式输出桥接
+│   │   ├── helpers/                      # JS 辅助函数
+│   │   │   └── stream-tool-sieve/        # Tool sieve JS 实现
+│   │   └── shared/                       # Go/Node 共用语义片段
+│   ├── prompt/                           # Prompt 组装
+│   ├── rawsample/                        # raw sample 读写与管理
+│   ├── server/                           # 路由与中间件装配
+│   ├── sse/                              # SSE 解析工具
+│   ├── stream/                           # 统一流式消费引擎
+│   ├── testsuite/                        # 测试集执行框架
+│   ├── textclean/                        # 文本清洗
+│   ├── toolcall/                         # 工具调用解析与修复
+│   ├── translatorcliproxy/               # 多协议互转桥
+│   ├── util/                             # 通用工具函数
+│   ├── version/                          # 版本查询/比较
+│   └── webui/                            # WebUI 静态托管相关逻辑
+├── plans/                                # 阶段计划与人工验收记录
+├── pow/                                  # PoW 独立实现与基准
+├── scripts/                              # 构建/发布/辅助脚本
+├── tests/                                # 测试资源与脚本
+│   ├── compat/                           # 兼容性夹具与期望输出
+│   │   ├── expected/                     # 预期结果样本
+│   │   └── fixtures/                     # 测试输入夹具
+│   │       ├── sse_chunks/               # SSE chunk 夹具
+│   │       └── toolcalls/                # toolcall 夹具
+│   ├── node/                             # Node 单元测试
+│   ├── raw_stream_samples/               # 上游原始 SSE 样本
+│   │   ├── content-filter-trigger-20260405-jwt3/          # 风控终态样本
+│   │   ├── continue-thinking-snapshot-replay-20260405/    # continue 样本
+│   │   ├── guangzhou-weather-reasoner-search-20260404/    # 搜索+引用样本
+│   │   ├── markdown-format-example-20260405/              # Markdown 样本
+│   │   └── markdown-format-example-20260405-spacefix/     # 空格修复样本
+│   ├── scripts/                          # 测试脚本入口
+│   └── tools/                            # 测试辅助工具
+└── webui/                                # React 管理台源码
+    ├── public/                           # 静态资源
+    └── src/                              # 前端源码
+        ├── app/                          # 路由/状态框架
+        ├── components/                   # 共享组件
+        ├── features/                     # 功能模块
+        │   ├── account/                  # 账号管理页面
+        │   ├── apiTester/                # API 测试页面
+        │   ├── settings/                 # 设置页面
+        │   └── vercel/                   # Vercel 同步页面
+        ├── layout/                       # 布局组件
+        ├── locales/                      # 国际化文案
+        └── utils/                        # 前端工具函数
+```
+
+## 2. 请求主链路
+
+```mermaid
+flowchart LR
+    C[Client/SDK] --> R[internal/server/router.go]
+    R --> OA[OpenAI Adapter]
+    R --> CA[Claude Adapter]
+    R --> GA[Gemini Adapter]
+    R --> AD[Admin API]
+
+    CA --> BR[translatorcliproxy]
+    GA --> BR
+    BR --> CORE[internal/adapter/openai ChatCompletions]
+    OA --> CORE
+
+    CORE --> AUTH[internal/auth + config key/account resolver]
+    CORE --> POOL[internal/account queue + concurrency]
+    CORE --> TOOL[internal/toolcall parser + sieve]
+    CORE --> DS[internal/deepseek client]
+    DS --> U[DeepSeek upstream]
+```
+
+## 3. internal/ 子模块职责
+
+- `internal/server`：路由树和中间件挂载（健康检查、协议入口、Admin/WebUI）。
+- `internal/adapter/openai`：统一执行内核（chat/responses/embeddings 与 tool calling 语义）。
+- `internal/adapter/{claude,gemini}`：协议输入输出适配，不重复实现上游调用逻辑。
+- `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转。
+- `internal/deepseek`：上游请求、会话、PoW、SSE 消费。
+- `internal/stream` + `internal/sse`：流式解析与增量处理。
+- `internal/toolcall`：JSON/XML/invoke/text-kv 工具调用解析及防泄漏筛分。
+- `internal/admin`：配置管理、账号管理、Vercel 同步、版本检查、开发抓包。
+- `internal/config`：配置加载、校验、运行时 settings 热更新。
+- `internal/account`：托管账号池、并发槽位、等待队列。
+
+## 4. WebUI 与运行时关系
+
+- `webui/` 是前端源码（Vite + React）。
+- 运行时托管目录是 `static/admin`（构建产物）。
+- 本地首次启动若 `static/admin` 缺失，会尝试自动构建（依赖 Node.js）。
+
+## 5. 文档拆分策略
+
+- 总览与快速开始：`README.MD` / `README.en.md`
+- 架构与目录：`docs/ARCHITECTURE*.md`（本文件）
+- 接口协议：`API.md` / `API.en.md`
+- 部署、测试、贡献：`docs/DEPLOY*`、`docs/TESTING.md`、`docs/CONTRIBUTING*`
+- 专题：`docs/toolcall-semantics.md`、`docs/DeepSeekSSE行为结构说明-2026-04-05.md`

docs/CONTRIBUTING.en.md

@@ -94,58 +94,12 @@ Manually build WebUI to `static/admin/`:
 
 ## Project Structure
 
-```text
-ds2api/
-├── app/                     # Shared HTTP handler assembly (local + serverless)
-├── cmd/
-│   ├── ds2api/              # Local/container entrypoint
-│   └── ds2api-tests/        # End-to-end testsuite entrypoint
-├── api/
-│   ├── index.go             # Vercel Serverless Go entry
-│   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # Account pool and concurrency queue
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI adapter
-│   │   ├── claude/          # Claude adapter
-│   │   └── gemini/          # Gemini adapter
-│   ├── admin/               # Admin API handlers
-│   ├── auth/                # Auth and JWT
-│   ├── claudeconv/          # Claude message conversion
-│   ├── compat/              # Go-version compatibility and regression helpers
-│   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek client, PoW WASM
-│   ├── js/                  # Node runtime stream/compat logic
-│   ├── devcapture/          # Dev packet capture
-│   ├── format/              # Output formatting
-│   ├── prompt/              # Prompt building
-│   ├── server/              # HTTP routing (chi router)
-│   ├── sse/                 # SSE parsing utilities
-│   ├── stream/              # Unified stream consumption engine
-│   ├── testsuite/           # Testsuite framework and scenario orchestration
-│   ├── translatorcliproxy/  # CLIProxy bridge and stream writer
-│   ├── util/                # Common utilities
-│   ├── version/             # Version parsing and comparison
-│   └── webui/               # WebUI static hosting
-├── webui/                   # React WebUI source
-│   └── src/
-│       ├── app/             # Routing, auth, config state
-│       ├── features/        # Feature modules
-│       ├── components/      # Shared components
-│       └── locales/         # Language packs
-├── scripts/                 # Build and test scripts
-├── tests/
-│   ├── compat/              # Compatibility fixtures and expected outputs
-│   ├── node/                # Node-side unit tests
-│   └── scripts/             # Test script entrypoints (unit/e2e)
-├── plans/                   # Plans, gates, and manual smoke-test records
-├── static/admin/            # WebUI build output (not committed)
-├── Dockerfile               # Multi-stage build
-├── docker-compose.yml       # Production
-├── docker-compose.dev.yml   # Development
-└── vercel.json              # Vercel config
-```
+To avoid documentation drift, directory layout and module responsibilities were moved to:
+
+- [docs/ARCHITECTURE.en.md](./ARCHITECTURE.en.md)
+- [docs/README.md](./README.md)
+
+Before contributing, review the architecture doc sections for request flow and `internal/` module boundaries.
 
 ## Reporting Issues
 

docs/CONTRIBUTING.md

@@ -94,58 +94,12 @@ docker-compose -f docker-compose.dev.yml up
 
 ## 项目结构
 
-```text
-ds2api/
-├── app/                     # 统一 HTTP Handler 装配（本地 + Serverless）
-├── cmd/
-│   ├── ds2api/              # 本地/容器启动入口
-│   └── ds2api-tests/        # 端到端测试集入口
-├── api/
-│   ├── index.go             # Vercel Serverless Go 入口
-│   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── (rewrite targets in vercel.json)
-├── internal/
-│   ├── account/             # 账号池与并发队列
-│   ├── adapter/
-│   │   ├── openai/          # OpenAI 兼容适配器
-│   │   ├── claude/          # Claude 兼容适配器
-│   │   └── gemini/          # Gemini 兼容适配器
-│   ├── admin/               # Admin API handlers
-│   ├── auth/                # 鉴权与 JWT
-│   ├── claudeconv/          # Claude 消息格式转换
-│   ├── compat/              # Go 版本兼容与回归测试辅助
-│   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek 客户端、PoW WASM
-│   ├── js/                  # Node 运行时流式/兼容逻辑
-│   ├── devcapture/          # 开发抓包
-│   ├── format/              # 输出格式化
-│   ├── prompt/              # Prompt 构建
-│   ├── server/              # HTTP 路由（chi router）
-│   ├── sse/                 # SSE 解析工具
-│   ├── stream/              # 统一流式消费引擎
-│   ├── testsuite/           # 测试集框架与场景编排
-│   ├── translatorcliproxy/  # CLIProxy 桥接与流式写入
-│   ├── util/                # 通用工具
-│   ├── version/             # 版本解析与比较
-│   └── webui/               # WebUI 静态托管
-├── webui/                   # React WebUI 源码
-│   └── src/
-│       ├── app/             # 路由、鉴权、配置状态
-│       ├── features/        # 业务功能模块
-│       ├── components/      # 通用组件
-│       └── locales/         # 语言包
-├── scripts/                 # 构建与测试脚本
-├── tests/
-│   ├── compat/              # 兼容夹具与期望输出
-│   ├── node/                # Node 侧单元测试
-│   └── scripts/             # 测试脚本入口（unit/e2e）
-├── plans/                   # 计划、门禁和手工烟测记录
-├── static/admin/            # WebUI 构建产物（不提交）
-├── Dockerfile               # 多阶段构建
-├── docker-compose.yml       # 生产环境
-├── docker-compose.dev.yml   # 开发环境
-└── vercel.json              # Vercel 配置
-```
+为避免与其他文档重复维护，目录结构与模块职责已迁移到：
+
+- [docs/ARCHITECTURE.md](./ARCHITECTURE.md)
+- [docs/README.md](./README.md)
+
+贡献前建议先阅读架构文档中的“请求主链路”和 `internal/` 模块职责，再定位改动范围。
 
 ## 问题反馈
 

docs/DEPLOY.en.md

@@ -4,15 +4,18 @@ Language: [中文](DEPLOY.md) | [English](DEPLOY.en.md)
 
 This guide covers all deployment methods for the current Go-based codebase.
 
+Doc map: [Index](./README.md) | [Architecture](./ARCHITECTURE.en.md) | [API](../API.en.md) | [Testing](./TESTING.md)
+
 ---
 
 ## Table of Contents
 
+- [Recommended deployment priority](#recommended-deployment-priority)
 - [Prerequisites](#0-prerequisites)
-- [1. Local Run](#1-local-run)
-- [2. Docker Deployment](#2-docker-deployment)
+- [1. Download Release Binaries](#1-download-release-binaries)
+- [2. Docker / GHCR Deployment](#2-docker--ghcr-deployment)
 - [3. Vercel Deployment](#3-vercel-deployment)
-- [4. Download Release Binaries](#4-download-release-binaries)
+- [4. Local Run from Source](#4-local-run-from-source)
 - [5. Reverse Proxy (Nginx)](#5-reverse-proxy-nginx)
 - [6. Linux systemd Service](#6-linux-systemd-service)
 - [7. Post-Deploy Checks](#7-post-deploy-checks)
@@ -20,6 +23,17 @@ This guide covers all deployment methods for the current Go-based codebase.
 
 ---
 
+## Recommended deployment priority
+
+Recommended order when choosing a deployment method:
+
+1. **Download and run release binaries**: the easiest path for most users because the artifacts are already built.
+2. **Docker / GHCR image deployment**: suitable for containerized, orchestrated, or cloud environments.
+3. **Vercel deployment**: suitable if you already use Vercel and accept its platform constraints.
+4. **Run from source / build locally**: suitable for development, debugging, or when you need to modify the code yourself.
+
+---
+
 ## 0. Prerequisites
 
 | Dependency | Minimum Version | Notes |
@@ -46,70 +60,59 @@ Use `config.json` as the single source of truth:
 
 ---
 
-## 1. Local Run
+## 1. Download Release Binaries
 
-### 1.1 Basic Steps
+Built-in GitHub Actions workflow: `.github/workflows/release-artifacts.yml`
+
+- **Trigger**: only on Release `published` (no build on normal push)
+- **Outputs**: multi-platform binary archives + `sha256sums.txt`
+- **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
+
+| Platform | Architecture | Format |
+| --- | --- | --- |
+| Linux | amd64, arm64 | `.tar.gz` |
+| macOS | amd64, arm64 | `.tar.gz` |
+| Windows | amd64 | `.zip` |
+
+Each archive includes:
+
+- `ds2api` executable (`ds2api.exe` on Windows)
+- `static/admin/` (built WebUI assets)
+- `config.example.json`, `.env.example`
+- `README.MD`, `README.en.md`, `LICENSE`
+
+### Usage
 
 ```bash
-# Clone
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
+# 1. Download the archive for your platform
+# 2. Extract
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 
-# Copy and edit config
+# 3. Configure
 cp config.example.json config.json
-# Open config.json and fill in:
-#   - keys: your API access keys
-#   - accounts: DeepSeek accounts (email or mobile + password)
+# Edit config.json
 
-# Start
-go run ./cmd/ds2api
-```
-
-Default local access URL: `http://127.0.0.1:5001`; the server actually binds to `0.0.0.0:5001` (override with `PORT`).
-
-### 1.2 WebUI Build
-
-On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm; when dependencies are missing it runs `npm ci` first, then `npm run build -- --outDir static/admin --emptyOutDir`).
-
-Manual build:
-
-```bash
-./scripts/build-webui.sh
-```
-
-Or step by step:
-
-```bash
-cd webui
-npm install
-npm run build
-# Output goes to static/admin/
-```
-
-Control auto-build via environment variable:
-
-```bash
-# Disable auto-build
-DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
-
-# Force enable auto-build
-DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
-```
-
-### 1.3 Compile to Binary
-
-```bash
-go build -o ds2api ./cmd/ds2api
+# 4. Start
 ./ds2api
 ```
 
+### Maintainer Release Flow
+
+1. Create and publish a GitHub Release (with tag, for example `vX.Y.Z`)
+2. Wait for the `Release Artifacts` workflow to complete
+3. Download the matching archive from Release Assets
+
 ---
 
-## 2. Docker Deployment
+## 2. Docker / GHCR Deployment
 
 ### 2.1 Basic Steps
 
 ```bash
+# Pull prebuilt image
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
 # Copy env template and config file
 cp .env.example .env
 cp config.example.json config.json
@@ -126,7 +129,13 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
-The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
+The default `docker-compose.yml` directly uses `ghcr.io/cjackhwang/ds2api:latest` and maps host port `6011` to container port `5001`. If you want `5001` exposed directly, set `DS2API_HOST_PORT=5001` (or adjust the `ports` mapping).
+
+If you want a pinned version instead of `latest`, you can also pull a specific tag directly:
+
+```bash
+docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+```
 
 ### 2.2 Update
 
@@ -348,58 +357,61 @@ If API responses return Vercel HTML `Authentication Required`:
 
 ---
 
-## 4. Download Release Binaries
+## 4. Local Run from Source
 
-Built-in GitHub Actions workflow: `.github/workflows/release-artifacts.yml`
-
-- **Trigger**: only on Release `published` (no build on normal push)
-- **Outputs**: multi-platform binary archives + `sha256sums.txt`
-- **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
-
-| Platform | Architecture | Format |
-| --- | --- | --- |
-| Linux | amd64, arm64 | `.tar.gz` |
-| macOS | amd64, arm64 | `.tar.gz` |
-| Windows | amd64 | `.zip` |
-
-Each archive includes:
-
-- `ds2api` executable (`ds2api.exe` on Windows)
-- `static/admin/` (built WebUI assets)
-- `sha3_wasm_bg.7b9ca65ddd.wasm` (optional; binary has embedded fallback)
-- `config.example.json`, `.env.example`
-- `README.MD`, `README.en.md`, `LICENSE`
-
-### Usage
+### 4.1 Basic Steps
 
 ```bash
-# 1. Download the archive for your platform
-# 2. Extract
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# Clone
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
 
-# 3. Configure
+# Copy and edit config
 cp config.example.json config.json
-# Edit config.json
+# Open config.json and fill in:
+#   - keys: your API access keys
+#   - accounts: DeepSeek accounts (email or mobile + password)
 
-# 4. Start
-./ds2api
+# Start
+go run ./cmd/ds2api
 ```
 
-### Maintainer Release Flow
+Default local access URL: `http://127.0.0.1:5001`; the server actually binds to `0.0.0.0:5001` (override with `PORT`).
 
-1. Create and publish a GitHub Release (with tag, for example `vX.Y.Z`)
-2. Wait for the `Release Artifacts` workflow to complete
-3. Download the matching archive from Release Assets
+### 4.2 WebUI Build
 
-### Pull from GHCR (Optional)
+On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm; when dependencies are missing it runs `npm ci` first, then `npm run build -- --outDir static/admin --emptyOutDir`).
+
+Manual build:
 
 ```bash
-# latest
-docker pull ghcr.io/cjackhwang/ds2api:latest
+./scripts/build-webui.sh
+```
 
-# specific version (example)
-docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+Or step by step:
+
+```bash
+cd webui
+npm install
+npm run build
+# Output goes to static/admin/
+```
+
+Control auto-build via environment variable:
+
+```bash
+# Disable auto-build
+DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
+
+# Force enable auto-build
+DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
+```
+
+### 4.3 Compile to Binary
+
+```bash
+go build -o ds2api ./cmd/ds2api
+./ds2api
 ```
 
 ---
@@ -456,8 +468,6 @@ server {
 # Copy compiled binary and related files to target directory
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# Optional: if you want to use an external WASM file (override the embedded one, from a release package or build output)
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DEPLOY.md

@@ -4,15 +4,18 @@
 
 本指南基于当前 Go 代码库，详细说明各种部署方式。
 
+本页导航：[文档总索引](./README.md)｜[架构说明](./ARCHITECTURE.md)｜[接口文档](../API.md)｜[测试指南](./TESTING.md)
+
 ---
 
 ## 目录
 
+- [部署方式优先级建议](#部署方式优先级建议)
 - [前置要求](#0-前置要求)
-- [一、本地运行](#一本地运行)
-- [二、Docker 部署](#二docker-部署)
+- [一、下载 Release 构建包](#一下载-release-构建包)
+- [二、Docker / GHCR 部署](#二docker--ghcr-部署)
 - [三、Vercel 部署](#三vercel-部署)
-- [四、下载 Release 构建包](#四下载-release-构建包)
+- [四、本地源码运行](#四本地源码运行)
 - [五、反向代理（Nginx）](#五反向代理nginx)
 - [六、Linux systemd 服务化](#六linux-systemd-服务化)
 - [七、部署后检查](#七部署后检查)
@@ -20,6 +23,17 @@
 
 ---
 
+## 部署方式优先级建议
+
+推荐按以下顺序选择部署方式：
+
+1. **下载 Release 构建包运行**：最省事，产物已编译完成，最适合大多数用户。
+2. **Docker / GHCR 镜像部署**：适合需要容器化、编排或云环境部署。
+3. **Vercel 部署**：适合已有 Vercel 环境且接受其平台约束的场景。
+4. **本地源码运行 / 自行编译**：适合开发、调试或需要自行修改代码的场景。
+
+---
+
 ## 0. 前置要求
 
 | 依赖 | 最低版本 | 说明 |
@@ -46,70 +60,59 @@ cp config.example.json config.json
 
 ---
 
-## 一、本地运行
+## 一、下载 Release 构建包
 
-### 1.1 基本步骤
+仓库内置 GitHub Actions 工作流：`.github/workflows/release-artifacts.yml`
+
+- **触发条件**：仅在 Release `published` 时触发（普通 push 不会构建）
+- **构建产物**：多平台二进制压缩包 + `sha256sums.txt`
+- **容器镜像发布**：仅发布到 GHCR（`ghcr.io/cjackhwang/ds2api`）
+
+| 平台 | 架构 | 文件格式 |
+| --- | --- | --- |
+| Linux | amd64, arm64 | `.tar.gz` |
+| macOS | amd64, arm64 | `.tar.gz` |
+| Windows | amd64 | `.zip` |
+
+每个压缩包包含：
+
+- `ds2api` 可执行文件（Windows 为 `ds2api.exe`）
+- `static/admin/`（WebUI 构建产物）
+- `config.example.json`、`.env.example`
+- `README.MD`、`README.en.md`、`LICENSE`
+
+### 使用步骤
 
 ```bash
-# 克隆仓库
-git clone https://github.com/CJackHwang/ds2api.git
-cd ds2api
+# 1. 下载对应平台的压缩包
+# 2. 解压
+tar -xzf ds2api_<tag>_linux_amd64.tar.gz
+cd ds2api_<tag>_linux_amd64
 
-# 复制并编辑配置
+# 3. 配置
 cp config.example.json config.json
-# 使用你喜欢的编辑器打开 config.json，填入：
-#   - keys: 你的 API 访问密钥
-#   - accounts: DeepSeek 账号（email 或 mobile + password）
+# 编辑 config.json
 
-# 启动服务
-go run ./cmd/ds2api
-```
-
-默认本地访问地址是 `http://127.0.0.1:5001`；服务实际绑定 `0.0.0.0:5001`，可通过 `PORT` 环境变量覆盖。
-
-### 1.2 WebUI 构建
-
-本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm；缺依赖时会先执行 `npm ci`，再执行 `npm run build -- --outDir static/admin --emptyOutDir`）。
-
-你也可以手动构建：
-
-```bash
-./scripts/build-webui.sh
-```
-
-或手动执行：
-
-```bash
-cd webui
-npm install
-npm run build
-# 产物输出到 static/admin/
-```
-
-通过环境变量控制自动构建行为：
-
-```bash
-# 强制关闭自动构建
-DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
-
-# 强制开启自动构建
-DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
-```
-
-### 1.3 编译为二进制文件
-
-```bash
-go build -o ds2api ./cmd/ds2api
+# 4. 启动
 ./ds2api
 ```
 
+### 维护者发布步骤
+
+1. 在 GitHub 创建并发布 Release（带 tag，如 `vX.Y.Z`）
+2. 等待 Actions 工作流 `Release Artifacts` 完成
+3. 在 Release 的 Assets 下载对应平台压缩包
+
 ---
 
-## 二、Docker 部署
+## 二、Docker / GHCR 部署
 
 ### 2.1 基本步骤
 
 ```bash
+# 拉取预编译镜像
+docker pull ghcr.io/cjackhwang/ds2api:latest
+
 # 复制环境变量模板和配置文件
 cp .env.example .env
 cp config.example.json config.json
@@ -126,7 +129,13 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
-默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请设置 `DS2API_HOST_PORT=5001`（或者手动调整 `ports` 配置）。
+默认 `docker-compose.yml` 直接使用 `ghcr.io/cjackhwang/ds2api:latest`，并把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请设置 `DS2API_HOST_PORT=5001`（或者手动调整 `ports` 配置）。
+
+如需固定版本，也可以直接拉取指定 tag：
+
+```bash
+docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+```
 
 ### 2.2 更新
 
@@ -348,58 +357,61 @@ No Output Directory named "public" found after the Build completed.
 
 ---
 
-## 四、下载 Release 构建包
+## 四、本地源码运行
 
-仓库内置 GitHub Actions 工作流：`.github/workflows/release-artifacts.yml`
-
-- **触发条件**：仅在 Release `published` 时触发（普通 push 不会构建）
-- **构建产物**：多平台二进制压缩包 + `sha256sums.txt`
-- **容器镜像发布**：仅发布到 GHCR（`ghcr.io/cjackhwang/ds2api`）
-
-| 平台 | 架构 | 文件格式 |
-| --- | --- | --- |
-| Linux | amd64, arm64 | `.tar.gz` |
-| macOS | amd64, arm64 | `.tar.gz` |
-| Windows | amd64 | `.zip` |
-
-每个压缩包包含：
-
-- `ds2api` 可执行文件（Windows 为 `ds2api.exe`）
-- `static/admin/`（WebUI 构建产物）
-- `sha3_wasm_bg.7b9ca65ddd.wasm`（可选；程序内置 embed fallback）
-- `config.example.json`、`.env.example`
-- `README.MD`、`README.en.md`、`LICENSE`
-
-### 使用步骤
+### 4.1 基本步骤
 
 ```bash
-# 1. 下载对应平台的压缩包
-# 2. 解压
-tar -xzf ds2api_<tag>_linux_amd64.tar.gz
-cd ds2api_<tag>_linux_amd64
+# 克隆仓库
+git clone https://github.com/CJackHwang/ds2api.git
+cd ds2api
 
-# 3. 配置
+# 复制并编辑配置
 cp config.example.json config.json
-# 编辑 config.json
+# 使用你喜欢的编辑器打开 config.json，填入：
+#   - keys: 你的 API 访问密钥
+#   - accounts: DeepSeek 账号（email 或 mobile + password）
 
-# 4. 启动
-./ds2api
+# 启动服务
+go run ./cmd/ds2api
 ```
 
-### 维护者发布步骤
+默认本地访问地址是 `http://127.0.0.1:5001`；服务实际绑定 `0.0.0.0:5001`，可通过 `PORT` 环境变量覆盖。
 
-1. 在 GitHub 创建并发布 Release（带 tag，如 `vX.Y.Z`）
-2. 等待 Actions 工作流 `Release Artifacts` 完成
-3. 在 Release 的 Assets 下载对应平台压缩包
+### 4.2 WebUI 构建
 
-### 拉取 GHCR 镜像（可选）
+本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm；缺依赖时会先执行 `npm ci`，再执行 `npm run build -- --outDir static/admin --emptyOutDir`）。
+
+你也可以手动构建：
 
 ```bash
-# latest
-docker pull ghcr.io/cjackhwang/ds2api:latest
+./scripts/build-webui.sh
+```
 
-# 指定版本（示例）
-docker pull ghcr.io/cjackhwang/ds2api:v3.0.0
+或手动执行：
+
+```bash
+cd webui
+npm install
+npm run build
+# 产物输出到 static/admin/
+```
+
+通过环境变量控制自动构建行为：
+
+```bash
+# 强制关闭自动构建
+DS2API_AUTO_BUILD_WEBUI=false go run ./cmd/ds2api
+
+# 强制开启自动构建
+DS2API_AUTO_BUILD_WEBUI=true go run ./cmd/ds2api
+```
+
+### 4.3 编译为二进制文件
+
+```bash
+go build -o ds2api ./cmd/ds2api
+./ds2api
 ```
 
 ---
@@ -456,8 +468,6 @@ server {
 # 将编译好的二进制文件和相关文件复制到目标目录
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# 可选：若你希望使用外置 WASM 文件（覆盖内置版本，来自 release 包或构建产物）
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DeepSeekSSE行为结构说明-2026-04-05.md

@@ -4,6 +4,8 @@
 > 当前 corpus 由 4 份原始流组成，覆盖搜索+引用、风控终态、Markdown 输出和空格敏感输出等行为。
 > 补充：文末还会注明少量“当前实现已确认、但 corpus 尚未完整覆盖”的行为，例如长思考场景下的自动续写状态。
 
+文档导航：[文档总索引](./README.md) / [测试指南](./TESTING.md) / [样本目录说明](../tests/raw_stream_samples/README.md)
+
 ## 1. 样本覆盖
 
 下列样本共同构成了本文的观察基础：

docs/README.md

@@ -0,0 +1,53 @@
+# DS2API 文档导航 | Documentation Index
+
+语言 / Language: [中文](README.md) | [English](README.md#english)
+
+## 中文
+
+为减少重复维护，本仓库文档按“入口文档 + 专题文档”拆分。建议从下列顺序阅读：
+
+1. [项目总览（README）](../README.MD)
+2. [架构与目录说明](./ARCHITECTURE.md)
+3. [接口文档（API）](../API.md)
+4. [部署指南](./DEPLOY.md)
+5. [测试指南](./TESTING.md)
+6. [贡献指南](./CONTRIBUTING.md)
+
+### 专题文档
+
+- [Tool Calling 统一语义](./toolcall-semantics.md)
+- [DeepSeek SSE 行为结构说明（逆向观察）](./DeepSeekSSE行为结构说明-2026-04-05.md)
+
+### 文档维护约定
+
+- `README.MD` / `README.en.md`：面向首次接触用户，保留“是什么 + 怎么快速跑起来”。
+- `docs/ARCHITECTURE*.md`：面向开发者，集中维护项目结构、模块职责与调用链。
+- `API*.md`：面向客户端接入者，聚焦接口行为、鉴权和示例。
+- 其他 `docs/*.md`：主题化说明，避免在多个文档重复粘贴同一段内容。
+
+---
+
+## English
+
+To reduce maintenance drift, docs are split into an “entry doc + topical docs” layout.
+
+Recommended reading order:
+
+1. [Project overview (README)](../README.en.md)
+2. [Architecture and project layout](./ARCHITECTURE.en.md)
+3. [API reference](../API.en.md)
+4. [Deployment guide](./DEPLOY.en.md)
+5. [Testing guide](./TESTING.md)
+6. [Contributing guide](./CONTRIBUTING.en.md)
+
+### Topical docs
+
+- [Tool-calling unified semantics](./toolcall-semantics.md)
+- [DeepSeek SSE behavior notes (reverse-engineered)](./DeepSeekSSE行为结构说明-2026-04-05.md)
+
+### Maintenance conventions
+
+- `README.MD` / `README.en.md`: onboarding-oriented (“what + quick start”).
+- `docs/ARCHITECTURE*.md`: developer-oriented source of truth for module boundaries and execution flow.
+- `API*.md`: integration-oriented behavior/contracts.
+- Other `docs/*.md`: focused topics, avoid copy-pasting the same section into multiple files.

docs/TESTING.md

@@ -2,6 +2,8 @@
 
 语言 / Language: 中文 + English（同页）
 
+文档导航： [总览](../README.MD) / [架构说明](./ARCHITECTURE.md) / [部署指南](./DEPLOY.md) / [接口文档](../API.md)
+
 ## 概述 | Overview
 
 DS2API 提供两个层级的测试：
@@ -235,6 +237,7 @@ go run ./cmd/ds2api-tests --no-preflight
 说明：
 - 该工具默认重放 `tests/raw_stream_samples/manifest.json` 声明的 canonical 样本，按上游 SSE 顺序做 1:1 仿真解析。
 - 默认校验不出现 `FINISHED` 文本泄露，并要求存在结束信号。
+- 默认**不**把 `raw accumulated_token_usage` 与本地解析 token 做强一致校验（当前实现以内容估算为准）；如需强校验可显式加 `--fail-on-token-mismatch`。
 - 每次运行都会把本地派生结果写入 `artifacts/raw-stream-sim/<run-id>/<sample-id>/replay.output.txt`，并输出结构化报告。
 - 如果你有历史基线目录，可以通过 `--baseline-root` 让工具直接做文本对比。
 - 更完整的协议级行为结构说明见 [DeepSeekSSE行为结构说明-2026-04-05.md](./DeepSeekSSE行为结构说明-2026-04-05.md)。

docs/toolcall-semantics.md

@@ -2,6 +2,8 @@
 
 本文档描述当前代码中 `ParseToolCallsDetailed` / `parseToolCallsDetailed` 的**实际行为**，用于对齐 Go 与 Node Runtime。
 
+文档导航：[总览](../README.MD) / [架构说明](./ARCHITECTURE.md) / [测试指南](./TESTING.md)
+
 ## 1) 输出结构（当前实现）
 
 - `calls`：解析得到的工具调用列表（`name` + `input`）。

go.mod

@@ -8,7 +8,6 @@ require (
 	github.com/google/uuid v1.6.0
 	github.com/refraction-networking/utls v1.8.2
 	github.com/router-for-me/CLIProxyAPI/v6 v6.9.14
-	github.com/tetratelabs/wazero v1.11.0
 )
 
 require (
@@ -19,7 +18,7 @@ require (
 	github.com/tidwall/pretty v1.2.1 // indirect
 	github.com/tidwall/sjson v1.2.5 // indirect
 	golang.org/x/crypto v0.49.0 // indirect
-	golang.org/x/net v0.52.0 // indirect
+	golang.org/x/net v0.52.0
 	golang.org/x/sys v0.42.0 // indirect
 	gopkg.in/yaml.v3 v3.0.1 // indirect
 )

go.sum

@@ -18,8 +18,6 @@ github.com/sirupsen/logrus v1.9.4 h1:TsZE7l11zFCLZnZ+teH4Umoq5BhEIfIzfRDZ1Uzql2w
 github.com/sirupsen/logrus v1.9.4/go.mod h1:ftWc9WdOfJ0a92nsE2jF5u5ZwH8Bv2zdeOC42RjbV2g=
 github.com/stretchr/testify v1.10.0 h1:Xv5erBjTwe/5IxqUQTdXv5kgmIvbHo3QQyRwhJsOfJA=
 github.com/stretchr/testify v1.10.0/go.mod h1:r2ic/lqez/lEtzL7wO/rwa5dbSLXVDPFyf8C91i36aY=
-github.com/tetratelabs/wazero v1.11.0 h1:+gKemEuKCTevU4d7ZTzlsvgd1uaToIDtlQlmNbwqYhA=
-github.com/tetratelabs/wazero v1.11.0/go.mod h1:eV28rsN8Q+xwjogd7f4/Pp4xFxO7uOGbLcD/LzB1wiU=
 github.com/tidwall/gjson v1.14.2/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=
 github.com/tidwall/gjson v1.18.0 h1:FIDeeyB800efLX89e5a8Y0BNH+LOngJyGrIWxG2FKQY=
 github.com/tidwall/gjson v1.18.0/go.mod h1:/wbyibRr2FHMks5tjHJ5F8dMZh3AcwJEMf5vlfC0lxk=

internal/adapter/claude/proxy_vercel_test.go

@@ -34,11 +34,13 @@ func (s openAIProxyStub) ChatCompletions(w http.ResponseWriter, _ *http.Request)
 
 type openAIProxyCaptureStub struct {
 	seenModel string
+	seenReq   map[string]any
 }
 
 func (s *openAIProxyCaptureStub) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 	var req map[string]any
 	_ = json.NewDecoder(r.Body).Decode(&req)
+	s.seenReq = req
 	if m, ok := req["model"].(string); ok {
 		s.seenModel = m
 	}
@@ -84,3 +86,33 @@ func TestClaudeProxyViaOpenAIPreservesClaudeMapping(t *testing.T) {
 		t.Fatalf("expected mapped proxy model deepseek-reasoner, got %q", got)
 	}
 }
+
+func TestClaudeProxyTranslatesInlineImageToOpenAIDataURL(t *testing.T) {
+	openAI := &openAIProxyCaptureStub{}
+	h := &Handler{OpenAI: openAI}
+	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", strings.NewReader(`{"model":"claude-sonnet-4-5","messages":[{"role":"user","content":[{"type":"text","text":"hello"},{"type":"image","source":{"type":"base64","media_type":"image/png","data":"QUJDRA=="}}]}],"stream":false}`))
+	rec := httptest.NewRecorder()
+
+	h.Messages(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	messages, _ := openAI.seenReq["messages"].([]any)
+	if len(messages) != 1 {
+		t.Fatalf("expected one translated message, got %#v", openAI.seenReq)
+	}
+	msg, _ := messages[0].(map[string]any)
+	content, _ := msg["content"].([]any)
+	if len(content) != 2 {
+		t.Fatalf("expected translated content blocks, got %#v", msg)
+	}
+	imageBlock, _ := content[1].(map[string]any)
+	if strings.TrimSpace(asString(imageBlock["type"])) != "image_url" {
+		t.Fatalf("expected image_url block, got %#v", imageBlock)
+	}
+	imageURL, _ := imageBlock["image_url"].(map[string]any)
+	if !strings.HasPrefix(strings.TrimSpace(asString(imageURL["url"])), "data:image/png;base64,") {
+		t.Fatalf("expected translated data url, got %#v", imageBlock)
+	}
+}

internal/adapter/claude/standard_request.go

@@ -36,7 +36,7 @@ func normalizeClaudeRequest(store ConfigReader, req map[string]any) (claudeNorma
 		thinkingEnabled = false
 		searchEnabled = false
 	}
-	finalPrompt := deepseek.MessagesPrepare(toMessageMaps(dsPayload["messages"]))
+	finalPrompt := deepseek.MessagesPrepareWithThinking(toMessageMaps(dsPayload["messages"]), thinkingEnabled)
 	toolNames := extractClaudeToolNames(toolsRequested)
 	if len(toolNames) == 0 && len(toolsRequested) > 0 {
 		toolNames = []string{"__any_tool__"}

internal/adapter/claude/stream_runtime_core.go

@@ -24,10 +24,9 @@ type claudeStreamRuntime struct {
 	bufferToolContent     bool
 	stripReferenceMarkers bool
 
-	messageID    string
-	thinking     strings.Builder
-	text         strings.Builder
-	outputTokens int
+	messageID string
+	thinking  strings.Builder
+	text      strings.Builder
 
 	nextBlockIndex     int
 	thinkingBlockOpen  bool
@@ -70,9 +69,6 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ErrorMessage != "" {
 		s.upstreamErr = parsed.ErrorMessage
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("upstream_error")}

internal/adapter/claude/stream_runtime_finalize.go

@@ -109,9 +109,6 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 	}
 
 	outputTokens := util.EstimateTokens(finalThinking) + util.EstimateTokens(finalText)
-	if s.outputTokens > 0 {
-		outputTokens = s.outputTokens
-	}
 	s.send("message_delta", map[string]any{
 		"type": "message_delta",
 		"delta": map[string]any{

internal/adapter/gemini/convert_request.go

@@ -28,7 +28,7 @@ func normalizeGeminiRequest(store ConfigReader, routeModel string, req map[strin
 	}
 
 	toolsRaw := convertGeminiTools(req["tools"])
-	finalPrompt, toolNames := openai.BuildPromptForAdapter(messagesRaw, toolsRaw, "")
+	finalPrompt, toolNames := openai.BuildPromptForAdapter(messagesRaw, toolsRaw, "", thinkingEnabled)
 	passThrough := collectGeminiPassThrough(req)
 
 	return util.StandardRequest{

internal/adapter/gemini/handler_generate.go

@@ -149,14 +149,13 @@ func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *ht
 		cleanVisibleOutput(result.Thinking, stripReferenceMarkers),
 		cleanVisibleOutput(result.Text, stripReferenceMarkers),
 		toolNames,
-		result.OutputTokens,
 	))
 }
 
 //nolint:unused // retained for native Gemini non-stream handling path.
-func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string, outputTokens int) map[string]any {
+func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string) map[string]any {
 	parts := buildGeminiPartsFromFinal(finalText, finalThinking, toolNames)
-	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText, outputTokens)
+	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText)
 	return map[string]any{
 		"candidates": []map[string]any{
 			{
@@ -174,14 +173,10 @@ func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, final
 }
 
 //nolint:unused // retained for native Gemini non-stream handling path.
-func buildGeminiUsage(finalPrompt, finalThinking, finalText string, outputTokens int) map[string]any {
+func buildGeminiUsage(finalPrompt, finalThinking, finalText string) map[string]any {
 	promptTokens := util.EstimateTokens(finalPrompt)
 	reasoningTokens := util.EstimateTokens(finalThinking)
 	completionTokens := util.EstimateTokens(finalText)
-	if outputTokens > 0 {
-		completionTokens = outputTokens
-		reasoningTokens = 0
-	}
 	return map[string]any{
 		"promptTokenCount":     promptTokens,
 		"candidatesTokenCount": reasoningTokens + completionTokens,

internal/adapter/gemini/handler_stream_runtime.go

@@ -65,9 +65,8 @@ type geminiStreamRuntime struct {
 	stripReferenceMarkers bool
 	toolNames             []string
 
-	thinking     strings.Builder
-	text         strings.Builder
-	outputTokens int
+	thinking strings.Builder
+	text     strings.Builder
 }
 
 //nolint:unused // retained for native Gemini stream handling path.
@@ -112,9 +111,6 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" || parsed.Stop {
 		return streamengine.ParsedDecision{Stop: true}
 	}
@@ -198,6 +194,6 @@ func (s *geminiStreamRuntime) finalize() {
 			},
 		},
 		"modelVersion":  s.model,
-		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText, s.outputTokens),
+		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText),
 	})
 }

internal/adapter/gemini/handler_test.go

@@ -82,11 +82,17 @@ func (s geminiOpenAIErrorStub) ChatCompletions(w http.ResponseWriter, _ *http.Re
 }
 
 type geminiOpenAISuccessStub struct {
-	stream bool
-	body   string
+	stream  bool
+	body    string
+	seenReq map[string]any
 }
 
-func (s geminiOpenAISuccessStub) ChatCompletions(w http.ResponseWriter, _ *http.Request) {
+func (s *geminiOpenAISuccessStub) ChatCompletions(w http.ResponseWriter, r *http.Request) {
+	if r != nil {
+		var req map[string]any
+		_ = json.NewDecoder(r.Body).Decode(&req)
+		s.seenReq = req
+	}
 	if s.stream {
 		w.Header().Set("Content-Type", "text/event-stream")
 		w.WriteHeader(http.StatusOK)
@@ -144,7 +150,7 @@ func TestGeminiRoutesRegistered(t *testing.T) {
 func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 	h := &Handler{
 		Store: testGeminiConfig{},
-		OpenAI: geminiOpenAISuccessStub{
+		OpenAI: &geminiOpenAISuccessStub{
 			body: `{"id":"chatcmpl-1","object":"chat.completion","choices":[{"index":0,"message":{"role":"assistant","tool_calls":[{"id":"call_1","type":"function","function":{"name":"eval_javascript","arguments":"{\"code\":\"1+1\"}"}}]},"finish_reason":"tool_calls"}]}`,
 		},
 	}
@@ -184,7 +190,7 @@ func TestGenerateContentReturnsFunctionCallParts(t *testing.T) {
 }
 
 func TestGenerateContentMixedToolSnippetAlsoTriggersFunctionCall(t *testing.T) {
-	h := &Handler{Store: testGeminiConfig{}, OpenAI: geminiOpenAISuccessStub{}}
+	h := &Handler{Store: testGeminiConfig{}, OpenAI: &geminiOpenAISuccessStub{}}
 	r := chi.NewRouter()
 	RegisterRoutes(r, h)
 
@@ -217,7 +223,7 @@ func TestGenerateContentMixedToolSnippetAlsoTriggersFunctionCall(t *testing.T) {
 func TestStreamGenerateContentEmitsSSE(t *testing.T) {
 	h := &Handler{
 		Store:  testGeminiConfig{},
-		OpenAI: geminiOpenAISuccessStub{stream: true},
+		OpenAI: &geminiOpenAISuccessStub{stream: true},
 	}
 	r := chi.NewRouter()
 	RegisterRoutes(r, h)
@@ -251,6 +257,39 @@ func TestStreamGenerateContentEmitsSSE(t *testing.T) {
 	}
 }
 
+func TestGeminiProxyTranslatesInlineImageToOpenAIDataURL(t *testing.T) {
+	openAI := &geminiOpenAISuccessStub{}
+	h := &Handler{Store: testGeminiConfig{}, OpenAI: openAI}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	body := `{"contents":[{"role":"user","parts":[{"text":"hello"},{"inlineData":{"mimeType":"image/png","data":"QUJDRA=="}}]}]}`
+	req := httptest.NewRequest(http.MethodPost, "/v1beta/models/gemini-2.5-pro:generateContent", strings.NewReader(body))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	messages, _ := openAI.seenReq["messages"].([]any)
+	if len(messages) != 1 {
+		t.Fatalf("expected one translated message, got %#v", openAI.seenReq)
+	}
+	msg, _ := messages[0].(map[string]any)
+	content, _ := msg["content"].([]any)
+	if len(content) != 2 {
+		t.Fatalf("expected translated content blocks, got %#v", msg)
+	}
+	imageBlock, _ := content[1].(map[string]any)
+	if strings.TrimSpace(asString(imageBlock["type"])) != "image_url" {
+		t.Fatalf("expected image_url block, got %#v", imageBlock)
+	}
+	imageURL, _ := imageBlock["image_url"].(map[string]any)
+	if !strings.HasPrefix(strings.TrimSpace(asString(imageURL["url"])), "data:image/png;base64,") {
+		t.Fatalf("expected translated data url, got %#v", imageBlock)
+	}
+}
+
 func TestGenerateContentOpenAIProxyErrorUsesGeminiEnvelope(t *testing.T) {
 	h := &Handler{
 		Store: testGeminiConfig{},

internal/adapter/openai/chat_stream_runtime.go

@@ -37,7 +37,6 @@ type chatStreamRuntime struct {
 	streamToolNames   map[int]string
 	thinking          strings.Builder
 	text              strings.Builder
-	outputTokens      int
 }
 
 func newChatStreamRuntime(
@@ -99,6 +98,19 @@ func (s *chatStreamRuntime) sendDone() {
 	}
 }
 
+func (s *chatStreamRuntime) sendFailedChunk(status int, message, code string) {
+	s.sendChunk(map[string]any{
+		"status_code": status,
+		"error": map[string]any{
+			"message": message,
+			"type":    openAIErrorType(status),
+			"code":    code,
+			"param":   nil,
+		},
+	})
+	s.sendDone()
+}
+
 func (s *chatStreamRuntime) finalize(finishReason string) {
 	finalThinking := s.thinking.String()
 	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
@@ -169,13 +181,22 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 	if len(detected.Calls) > 0 || s.toolCallsEmitted {
 		finishReason = "tool_calls"
 	}
-	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
-	if s.outputTokens > 0 {
-		usage["completion_tokens"] = s.outputTokens
-		if prompt, ok := usage["prompt_tokens"].(int); ok {
-			usage["total_tokens"] = prompt + s.outputTokens
+	if len(detected.Calls) == 0 && !s.toolCallsEmitted && strings.TrimSpace(finalText) == "" {
+		status := http.StatusTooManyRequests
+		message := "Upstream model returned empty output."
+		code := "upstream_empty_output"
+		if strings.TrimSpace(finalThinking) != "" {
+			message = "Upstream model returned reasoning without visible output."
 		}
+		if finishReason == "content_filter" {
+			status = http.StatusBadRequest
+			message = "Upstream content filtered the response and returned no output."
+			code = "content_filter"
+		}
+		s.sendFailedChunk(status, message, code)
+		return
 	}
+	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
 		s.completionID,
 		s.created,
@@ -190,10 +211,10 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter {
+		if strings.TrimSpace(s.text.String()) == "" {
+			return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("content_filter")}
+		}
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReasonHandlerRequested}
 	}
 	if parsed.ErrorMessage != "" {
@@ -243,7 +264,7 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 						if !s.emitEarlyToolDeltas {
 							continue
 						}
-						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.toolNames, s.streamToolNames)
+						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.streamToolNames)
 						if len(filtered) == 0 {
 							continue
 						}

internal/adapter/openai/deps.go

@@ -18,6 +18,7 @@ type AuthResolver interface {
 type DeepSeekCaller interface {
 	CreateSession(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error)
 	GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error)
+	UploadFile(ctx context.Context, a *auth.RequestAuth, req deepseek.UploadFileRequest, maxAttempts int) (*deepseek.UploadFileResult, error)
 	CallCompletion(ctx context.Context, a *auth.RequestAuth, payload map[string]any, powResp string, maxAttempts int) (*http.Response, error)
 	DeleteSessionForToken(ctx context.Context, token string, sessionID string) (*deepseek.DeleteSessionResult, error)
 	DeleteAllSessionsForToken(ctx context.Context, token string) error

internal/adapter/openai/embeddings_handler.go

@@ -26,8 +26,13 @@ func (h *Handler) Embeddings(w http.ResponseWriter, r *http.Request) {
 	}
 	defer h.Auth.Release(a)
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}

internal/adapter/openai/file_inline_upload.go

@@ -0,0 +1,382 @@
+package openai
+
+import (
+	"context"
+	"crypto/sha256"
+	"encoding/base64"
+	"fmt"
+	"mime"
+	"net/http"
+	"net/url"
+	"path/filepath"
+	"strings"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+const maxInlineFilesPerRequest = 50
+
+type inlineFileUploadError struct {
+	status  int
+	message string
+	err     error
+}
+
+func (e *inlineFileUploadError) Error() string {
+	if e == nil {
+		return ""
+	}
+	if strings.TrimSpace(e.message) != "" {
+		return e.message
+	}
+	if e.err != nil {
+		return e.err.Error()
+	}
+	return "inline file processing failed"
+}
+
+type inlineUploadState struct {
+	ctx          context.Context
+	handler      *Handler
+	auth         *auth.RequestAuth
+	uploadedByID map[string]string
+	uploadCount  int
+}
+
+type inlineDecodedFile struct {
+	Data            []byte
+	ContentType     string
+	Filename        string
+	ReplacementType string
+}
+
+func (h *Handler) preprocessInlineFileInputs(ctx context.Context, a *auth.RequestAuth, req map[string]any) error {
+	if h == nil || h.DS == nil || len(req) == 0 {
+		return nil
+	}
+	state := &inlineUploadState{
+		ctx:          ctx,
+		handler:      h,
+		auth:         a,
+		uploadedByID: map[string]string{},
+	}
+	for _, key := range []string{"messages", "input", "attachments"} {
+		if raw, ok := req[key]; ok {
+			updated, err := state.walk(raw)
+			if err != nil {
+				return err
+			}
+			req[key] = updated
+		}
+	}
+	if refIDs := collectOpenAIRefFileIDs(req); len(refIDs) > 0 {
+		req["ref_file_ids"] = stringsToAnySlice(refIDs)
+	}
+	return nil
+}
+
+func writeOpenAIInlineFileError(w http.ResponseWriter, err error) {
+	inlineErr, ok := err.(*inlineFileUploadError)
+	if !ok || inlineErr == nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "Failed to process file input.")
+		return
+	}
+	status := inlineErr.status
+	if status == 0 {
+		status = http.StatusInternalServerError
+	}
+	message := strings.TrimSpace(inlineErr.message)
+	if message == "" {
+		message = "Failed to process file input."
+	}
+	writeOpenAIError(w, status, message)
+}
+
+func (s *inlineUploadState) walk(raw any) (any, error) {
+	switch x := raw.(type) {
+	case []any:
+		out := make([]any, len(x))
+		for i, item := range x {
+			updated, err := s.walk(item)
+			if err != nil {
+				return nil, err
+			}
+			out[i] = updated
+		}
+		return out, nil
+	case map[string]any:
+		if replacement, replaced, err := s.tryUploadBlock(x); replaced || err != nil {
+			return replacement, err
+		}
+		for _, key := range []string{"messages", "input", "attachments", "content", "files", "items", "data", "source", "file", "image_url"} {
+			if nested, ok := x[key]; ok {
+				updated, err := s.walk(nested)
+				if err != nil {
+					return nil, err
+				}
+				x[key] = updated
+			}
+		}
+		return x, nil
+	default:
+		return raw, nil
+	}
+}
+
+func (s *inlineUploadState) tryUploadBlock(block map[string]any) (map[string]any, bool, error) {
+	decoded, ok, err := decodeOpenAIInlineFileBlock(block)
+	if err != nil {
+		return nil, true, &inlineFileUploadError{status: http.StatusBadRequest, message: err.Error(), err: err}
+	}
+	if !ok {
+		return nil, false, nil
+	}
+	if s.uploadCount >= maxInlineFilesPerRequest {
+		return nil, true, fmt.Errorf("exceeded maximum of %d inline files per request", maxInlineFilesPerRequest)
+	}
+	fileID, err := s.uploadInlineFile(decoded)
+	if err != nil {
+		return nil, true, &inlineFileUploadError{status: http.StatusInternalServerError, message: "Failed to upload inline file.", err: err}
+	}
+	s.uploadCount++
+	replacement := map[string]any{
+		"type":    decoded.ReplacementType,
+		"file_id": fileID,
+	}
+	if decoded.Filename != "" {
+		replacement["filename"] = decoded.Filename
+	}
+	if decoded.ContentType != "" {
+		replacement["mime_type"] = decoded.ContentType
+	}
+	return replacement, true, nil
+}
+
+func (s *inlineUploadState) uploadInlineFile(file inlineDecodedFile) (string, error) {
+	sum := sha256.Sum256(append([]byte(file.ContentType+"\x00"+file.Filename+"\x00"), file.Data...))
+	cacheKey := fmt.Sprintf("%x", sum[:])
+	if fileID, ok := s.uploadedByID[cacheKey]; ok && strings.TrimSpace(fileID) != "" {
+		return fileID, nil
+	}
+	contentType := strings.TrimSpace(file.ContentType)
+	if contentType == "" {
+		contentType = http.DetectContentType(file.Data)
+	}
+	result, err := s.handler.DS.UploadFile(s.ctx, s.auth, deepseek.UploadFileRequest{
+		Filename:    file.Filename,
+		ContentType: contentType,
+		Data:        file.Data,
+	}, 3)
+	if err != nil {
+		return "", err
+	}
+	fileID := strings.TrimSpace(result.ID)
+	if fileID == "" {
+		return "", fmt.Errorf("upload succeeded without file id")
+	}
+	s.uploadedByID[cacheKey] = fileID
+	return fileID, nil
+}
+
+func decodeOpenAIInlineFileBlock(block map[string]any) (inlineDecodedFile, bool, error) {
+	if block == nil {
+		return inlineDecodedFile{}, false, nil
+	}
+	if strings.TrimSpace(asString(block["file_id"])) != "" {
+		return inlineDecodedFile{}, false, nil
+	}
+	if nested, ok := block["file"].(map[string]any); ok {
+		decoded, matched, err := decodeOpenAIInlineFileBlock(nested)
+		if err != nil || !matched {
+			return decoded, matched, err
+		}
+		if decoded.Filename == "" {
+			decoded.Filename = pickInlineFilename(block, decoded.ContentType, defaultInlinePrefix(decoded.ReplacementType))
+		}
+		return decoded, true, nil
+	}
+	blockType := strings.ToLower(strings.TrimSpace(asString(block["type"])))
+	if raw, matched := extractInlineImageDataURL(block); matched {
+		data, contentType, err := decodeInlinePayload(raw, contentTypeFromMap(block))
+		if err != nil {
+			return inlineDecodedFile{}, true, fmt.Errorf("invalid image input")
+		}
+		return inlineDecodedFile{
+			Data:            data,
+			ContentType:     contentType,
+			Filename:        pickInlineFilename(block, contentType, "image"),
+			ReplacementType: "input_image",
+		}, true, nil
+	}
+	if raw, matched := extractInlineFilePayload(block, blockType); matched {
+		data, contentType, err := decodeInlinePayload(raw, contentTypeFromMap(block))
+		if err != nil {
+			return inlineDecodedFile{}, true, fmt.Errorf("invalid file input")
+		}
+		return inlineDecodedFile{
+			Data:            data,
+			ContentType:     contentType,
+			Filename:        pickInlineFilename(block, contentType, defaultInlinePrefix(blockType)),
+			ReplacementType: "input_file",
+		}, true, nil
+	}
+	return inlineDecodedFile{}, false, nil
+}
+
+func extractInlineImageDataURL(block map[string]any) (string, bool) {
+	imageURL := block["image_url"]
+	switch x := imageURL.(type) {
+	case string:
+		if isDataURL(x) {
+			return strings.TrimSpace(x), true
+		}
+	case map[string]any:
+		if raw := strings.TrimSpace(asString(x["url"])); isDataURL(raw) {
+			return raw, true
+		}
+	}
+	if raw := strings.TrimSpace(asString(block["url"])); isDataURL(raw) {
+		return raw, true
+	}
+	return "", false
+}
+
+func extractInlineFilePayload(block map[string]any, blockType string) (string, bool) {
+	for _, value := range []any{block["file_data"], block["base64"], block["data"]} {
+		if raw := strings.TrimSpace(asString(value)); raw != "" {
+			if strings.Contains(blockType, "file") || block["file_data"] != nil || block["filename"] != nil || block["file_name"] != nil || block["name"] != nil {
+				return raw, true
+			}
+		}
+	}
+	return "", false
+}
+
+func decodeInlinePayload(raw string, explicitContentType string) ([]byte, string, error) {
+	raw = strings.TrimSpace(raw)
+	if raw == "" {
+		return nil, "", fmt.Errorf("empty payload")
+	}
+	if isDataURL(raw) {
+		return decodeDataURL(raw, explicitContentType)
+	}
+	decoded, err := decodeBase64Flexible(raw)
+	if err != nil {
+		return nil, "", err
+	}
+	contentType := strings.TrimSpace(explicitContentType)
+	if contentType == "" && len(decoded) > 0 {
+		contentType = http.DetectContentType(decoded)
+	}
+	return decoded, contentType, nil
+}
+
+func decodeDataURL(raw string, explicitContentType string) ([]byte, string, error) {
+	raw = strings.TrimSpace(raw)
+	if !isDataURL(raw) {
+		return nil, "", fmt.Errorf("unsupported data url")
+	}
+	header, payload, ok := strings.Cut(raw, ",")
+	if !ok {
+		return nil, "", fmt.Errorf("invalid data url")
+	}
+	meta := strings.TrimSpace(strings.TrimPrefix(header, "data:"))
+	contentType := strings.TrimSpace(explicitContentType)
+	if contentType == "" {
+		contentType = "application/octet-stream"
+		if meta != "" {
+			parts := strings.Split(meta, ";")
+			if len(parts) > 0 && strings.TrimSpace(parts[0]) != "" {
+				contentType = strings.TrimSpace(parts[0])
+			}
+		}
+	}
+	if strings.Contains(strings.ToLower(meta), ";base64") {
+		decoded, err := decodeBase64Flexible(payload)
+		if err != nil {
+			return nil, "", err
+		}
+		return decoded, contentType, nil
+	}
+	decoded, err := url.PathUnescape(payload)
+	if err != nil {
+		return nil, "", err
+	}
+	return []byte(decoded), contentType, nil
+}
+
+func decodeBase64Flexible(raw string) ([]byte, error) {
+	raw = strings.TrimSpace(raw)
+	for _, enc := range []*base64.Encoding{base64.StdEncoding, base64.RawStdEncoding, base64.URLEncoding, base64.RawURLEncoding} {
+		decoded, err := enc.DecodeString(raw)
+		if err == nil {
+			return decoded, nil
+		}
+	}
+	return nil, fmt.Errorf("invalid base64 payload")
+}
+
+func contentTypeFromMap(block map[string]any) string {
+	for _, value := range []any{block["mime_type"], block["mimeType"], block["content_type"], block["contentType"], block["media_type"], block["mediaType"]} {
+		if contentType := strings.TrimSpace(asString(value)); contentType != "" {
+			return contentType
+		}
+	}
+	if imageURL, ok := block["image_url"].(map[string]any); ok {
+		for _, value := range []any{imageURL["mime_type"], imageURL["mimeType"], imageURL["content_type"], imageURL["contentType"]} {
+			if contentType := strings.TrimSpace(asString(value)); contentType != "" {
+				return contentType
+			}
+		}
+	}
+	return ""
+}
+
+func pickInlineFilename(block map[string]any, contentType string, prefix string) string {
+	for _, value := range []any{block["filename"], block["file_name"], block["name"]} {
+		if name := strings.TrimSpace(asString(value)); name != "" {
+			return filepath.Base(name)
+		}
+	}
+	if prefix == "" {
+		prefix = "upload"
+	}
+	ext := ".bin"
+	if parsedType := strings.TrimSpace(contentType); parsedType != "" {
+		if comma := strings.Index(parsedType, ";"); comma >= 0 {
+			parsedType = strings.TrimSpace(parsedType[:comma])
+		}
+		if exts, err := mime.ExtensionsByType(parsedType); err == nil && len(exts) > 0 && strings.TrimSpace(exts[0]) != "" {
+			ext = exts[0]
+		}
+	}
+	return prefix + ext
+}
+
+func defaultInlinePrefix(blockType string) string {
+	blockType = strings.ToLower(strings.TrimSpace(blockType))
+	if strings.Contains(blockType, "image") {
+		return "image"
+	}
+	return "upload"
+}
+
+func isDataURL(raw string) bool {
+	return strings.HasPrefix(strings.ToLower(strings.TrimSpace(raw)), "data:")
+}
+
+func stringsToAnySlice(items []string) []any {
+	out := make([]any, 0, len(items))
+	for _, item := range items {
+		trimmed := strings.TrimSpace(item)
+		if trimmed == "" {
+			continue
+		}
+		out = append(out, trimmed)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}

internal/adapter/openai/file_inline_upload_test.go

@@ -0,0 +1,274 @@
+package openai
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+type inlineUploadDSStub struct {
+	uploadCalls    []deepseek.UploadFileRequest
+	lastCtx        context.Context
+	completionReq  map[string]any
+	createSession  string
+	uploadErr      error
+	completionResp *http.Response
+}
+
+func (m *inlineUploadDSStub) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	if strings.TrimSpace(m.createSession) == "" {
+		return "session-id", nil
+	}
+	return m.createSession, nil
+}
+
+func (m *inlineUploadDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "pow", nil
+}
+
+func (m *inlineUploadDSStub) UploadFile(ctx context.Context, _ *auth.RequestAuth, req deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	m.lastCtx = ctx
+	m.uploadCalls = append(m.uploadCalls, req)
+	if m.uploadErr != nil {
+		return nil, m.uploadErr
+	}
+	return &deepseek.UploadFileResult{
+		ID:       "file-inline-1",
+		Filename: req.Filename,
+		Bytes:    int64(len(req.Data)),
+		Status:   "uploaded",
+		Purpose:  req.Purpose,
+	}, nil
+}
+
+func (m *inlineUploadDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
+	m.completionReq = payload
+	if m.completionResp != nil {
+		return m.completionResp, nil
+	}
+	return makeOpenAISSEHTTPResponse(
+		`data: {"p":"response/content","v":"ok"}`,
+		`data: [DONE]`,
+	), nil
+}
+
+func (m *inlineUploadDSStub) DeleteSessionForToken(_ context.Context, _ string, _ string) (*deepseek.DeleteSessionResult, error) {
+	return &deepseek.DeleteSessionResult{Success: true}, nil
+}
+
+func (m *inlineUploadDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	return nil
+}
+
+func TestPreprocessInlineFileInputsReplacesDataURLAndCollectsRefFileIDs(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{DS: ds}
+	req := map[string]any{
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{
+						"type":      "image_url",
+						"image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="},
+					},
+				},
+			},
+		},
+	}
+	ctx, cancel := context.WithCancel(context.Background())
+	defer cancel()
+
+	if err := h.preprocessInlineFileInputs(ctx, &auth.RequestAuth{DeepSeekToken: "token"}, req); err != nil {
+		t.Fatalf("preprocess failed: %v", err)
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload, got %d", len(ds.uploadCalls))
+	}
+	if ds.lastCtx != ctx {
+		t.Fatalf("expected upload to use request context")
+	}
+	if ds.uploadCalls[0].ContentType != "image/png" {
+		t.Fatalf("expected image/png, got %q", ds.uploadCalls[0].ContentType)
+	}
+	if ds.uploadCalls[0].Filename != "image.png" {
+		t.Fatalf("expected inferred filename image.png, got %q", ds.uploadCalls[0].Filename)
+	}
+	messages, _ := req["messages"].([]any)
+	first, _ := messages[0].(map[string]any)
+	content, _ := first["content"].([]any)
+	block, _ := content[0].(map[string]any)
+	if block["type"] != "input_image" {
+		t.Fatalf("expected input_image replacement, got %#v", block)
+	}
+	if block["file_id"] != "file-inline-1" {
+		t.Fatalf("expected file-inline-1 replacement id, got %#v", block)
+	}
+	refIDs, _ := req["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected ref_file_ids: %#v", req["ref_file_ids"])
+	}
+}
+
+func TestPreprocessInlineFileInputsDeduplicatesIdenticalPayloads(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{DS: ds}
+	req := map[string]any{
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="}},
+					map[string]any{"type": "image_url", "image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="}},
+				},
+			},
+		},
+	}
+
+	if err := h.preprocessInlineFileInputs(context.Background(), &auth.RequestAuth{DeepSeekToken: "token"}, req); err != nil {
+		t.Fatalf("preprocess failed: %v", err)
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected deduplicated single upload, got %d", len(ds.uploadCalls))
+	}
+	refIDs, _ := req["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected ref_file_ids after dedupe: %#v", req["ref_file_ids"])
+	}
+}
+
+func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	h.ChatCompletions(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	if ds.completionReq == nil {
+		t.Fatal("expected completion payload to be captured")
+	}
+	refIDs, _ := ds.completionReq["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected completion ref_file_ids: %#v", ds.completionReq["ref_file_ids"])
+	}
+}
+
+func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","input":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"input_image","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	refIDs, _ := ds.completionReq["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected completion ref_file_ids: %#v", ds.completionReq["ref_file_ids"])
+	}
+}
+
+func TestChatCompletionsInlineUploadFailureReturnsBadRequest(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,%%%"}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	h.ChatCompletions(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.completionReq != nil {
+		t.Fatalf("did not expect completion call on upload decode error")
+	}
+}
+
+func TestResponsesInlineUploadFailureReturnsInternalServerError(t *testing.T) {
+	ds := &inlineUploadDSStub{uploadErr: errors.New("boom")}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","input":[{"role":"user","content":[{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusInternalServerError {
+		t.Fatalf("expected 500, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.completionReq != nil {
+		t.Fatalf("did not expect completion call after upload failure")
+	}
+}
+
+func TestVercelPrepareUploadsInlineFilesBeforeLeasePayload(t *testing.T) {
+	t.Setenv("VERCEL", "1")
+	t.Setenv("DS2API_VERCEL_INTERNAL_SECRET", "stream-secret")
+	ds := &inlineUploadDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions?__stream_prepare=1", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("X-Ds2-Internal-Token", "stream-secret")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(ds.uploadCalls) != 1 {
+		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	payload, _ := out["payload"].(map[string]any)
+	if payload == nil {
+		t.Fatalf("expected payload in prepare response, got %#v", out)
+	}
+	refIDs, _ := payload["ref_file_ids"].([]any)
+	if len(refIDs) != 1 || refIDs[0] != "file-inline-1" {
+		t.Fatalf("unexpected payload ref_file_ids: %#v", payload["ref_file_ids"])
+	}
+}

internal/adapter/openai/file_refs.go

@@ -0,0 +1,94 @@
+package openai
+
+import "strings"
+
+func collectOpenAIRefFileIDs(req map[string]any) []string {
+	if len(req) == 0 {
+		return nil
+	}
+	out := make([]string, 0, 4)
+	seen := map[string]struct{}{}
+	for _, key := range []string{
+		"ref_file_ids",
+		"file_ids",
+		"attachments",
+		"messages",
+		"input",
+	} {
+		raw := req[key]
+		if raw == nil {
+			continue
+		}
+		// Skip top-level strings for 'messages' and 'input' as they are likely plain text content,
+		// not file IDs. String file IDs are expected in 'ref_file_ids' or 'file_ids'.
+		if key == "messages" || key == "input" {
+			if _, ok := raw.(string); ok {
+				continue
+			}
+		}
+		appendOpenAIRefFileIDs(&out, seen, raw)
+	}
+	if len(out) == 0 {
+		return nil
+	}
+	return out
+}
+
+func appendOpenAIRefFileIDs(out *[]string, seen map[string]struct{}, raw any) {
+	switch x := raw.(type) {
+	case string:
+		addOpenAIRefFileID(out, seen, x)
+	case []string:
+		for _, item := range x {
+			addOpenAIRefFileID(out, seen, item)
+		}
+	case []any:
+		for _, item := range x {
+			appendOpenAIRefFileIDs(out, seen, item)
+		}
+	case map[string]any:
+		if fileID := strings.TrimSpace(asString(x["file_id"])); fileID != "" {
+			addOpenAIRefFileID(out, seen, fileID)
+		}
+		if strings.Contains(strings.ToLower(strings.TrimSpace(asString(x["type"]))), "file") {
+			if fileID := strings.TrimSpace(asString(x["id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+		}
+		if fileMap, ok := x["file"].(map[string]any); ok {
+			if fileID := strings.TrimSpace(asString(fileMap["file_id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+			if fileID := strings.TrimSpace(asString(fileMap["id"])); fileID != "" {
+				addOpenAIRefFileID(out, seen, fileID)
+			}
+		}
+		// Recurse into potential containers. Note: we do NOT recurse into 'content' or 'input'
+		// if they are plain strings (handled by the top-level switch), but they are usually
+		// nested inside the map branch anyway.
+		// To be safe, we only recurse into these known container keys.
+		for _, key := range []string{"ref_file_ids", "file_ids", "attachments", "messages", "input", "content", "files", "items", "data", "source"} {
+			if nested, ok := x[key]; ok {
+				// If it's a message content that is a string, we must NOT treat it as an ID.
+				if key == "content" || key == "input" {
+					if _, ok := nested.(string); ok {
+						continue
+					}
+				}
+				appendOpenAIRefFileIDs(out, seen, nested)
+			}
+		}
+	}
+}
+
+func addOpenAIRefFileID(out *[]string, seen map[string]struct{}, fileID string) {
+	fileID = strings.TrimSpace(fileID)
+	if fileID == "" {
+		return
+	}
+	if _, ok := seen[fileID]; ok {
+		return
+	}
+	seen[fileID] = struct{}{}
+	*out = append(*out, fileID)
+}

internal/adapter/openai/files_route_test.go

@@ -0,0 +1,202 @@
+package openai
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"mime/multipart"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+type managedFilesAuthStub struct{}
+
+func (managedFilesAuthStub) Determine(_ *http.Request) (*auth.RequestAuth, error) {
+	return &auth.RequestAuth{
+		UseConfigToken: true,
+		DeepSeekToken:  "managed-token",
+		CallerID:       "caller:test",
+		AccountID:      "acct-123",
+		TriedAccounts:  map[string]bool{},
+	}, nil
+}
+
+func (managedFilesAuthStub) DetermineCaller(_ *http.Request) (*auth.RequestAuth, error) {
+	return &auth.RequestAuth{
+		UseConfigToken: true,
+		DeepSeekToken:  "managed-token",
+		CallerID:       "caller:test",
+		AccountID:      "acct-123",
+		TriedAccounts:  map[string]bool{},
+	}, nil
+}
+
+func (managedFilesAuthStub) Release(_ *auth.RequestAuth) {}
+
+type filesRouteDSStub struct {
+	lastReq deepseek.UploadFileRequest
+	upload  *deepseek.UploadFileResult
+	err     error
+}
+
+func (m *filesRouteDSStub) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "", nil
+}
+
+func (m *filesRouteDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "", nil
+}
+
+func (m *filesRouteDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, req deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	m.lastReq = req
+	if m.err != nil {
+		return nil, m.err
+	}
+	if m.upload != nil {
+		return m.upload, nil
+	}
+	return &deepseek.UploadFileResult{ID: "file-123", Filename: req.Filename, Bytes: int64(len(req.Data)), Purpose: req.Purpose, Status: "uploaded"}, nil
+}
+
+func (m *filesRouteDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
+	return nil, errors.New("not implemented")
+}
+
+func (m *filesRouteDSStub) DeleteSessionForToken(_ context.Context, _ string, _ string) (*deepseek.DeleteSessionResult, error) {
+	return &deepseek.DeleteSessionResult{Success: true}, nil
+}
+
+func (m *filesRouteDSStub) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	return nil
+}
+
+func newMultipartUploadRequest(t *testing.T, purpose string, filename string, data []byte) *http.Request {
+	t.Helper()
+	var body bytes.Buffer
+	writer := multipart.NewWriter(&body)
+	if purpose != "" {
+		if err := writer.WriteField("purpose", purpose); err != nil {
+			t.Fatalf("write purpose failed: %v", err)
+		}
+	}
+	part, err := writer.CreateFormFile("file", filename)
+	if err != nil {
+		t.Fatalf("create form file failed: %v", err)
+	}
+	if _, err := part.Write(data); err != nil {
+		t.Fatalf("write file failed: %v", err)
+	}
+	if err := writer.Close(); err != nil {
+		t.Fatalf("close writer failed: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", &body)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", writer.FormDataContentType())
+	return req
+}
+
+func TestFilesRouteUploadSuccess(t *testing.T) {
+	ds := &filesRouteDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if ds.lastReq.Filename != "notes.txt" {
+		t.Fatalf("expected filename notes.txt, got %q", ds.lastReq.Filename)
+	}
+	if ds.lastReq.Purpose != "assistants" {
+		t.Fatalf("expected purpose assistants, got %q", ds.lastReq.Purpose)
+	}
+	if string(ds.lastReq.Data) != "hello world" {
+		t.Fatalf("unexpected uploaded data: %q", string(ds.lastReq.Data))
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	if out["object"] != "file" {
+		t.Fatalf("expected file object, got %#v", out)
+	}
+	if out["id"] != "file-123" {
+		t.Fatalf("expected file id file-123, got %#v", out["id"])
+	}
+	if out["filename"] != "notes.txt" {
+		t.Fatalf("expected filename notes.txt, got %#v", out["filename"])
+	}
+}
+
+func TestFilesRouteUploadIncludesAccountIDForManagedAccount(t *testing.T) {
+	ds := &filesRouteDSStub{}
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: managedFilesAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	if out["account_id"] != "acct-123" {
+		t.Fatalf("expected account_id acct-123, got %#v", out["account_id"])
+	}
+}
+
+func TestFilesRouteRejectsNonMultipart(t *testing.T) {
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", bytes.NewBufferString(`{"purpose":"assistants"}`))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}
+
+func TestFilesRouteRequiresFileField(t *testing.T) {
+	h := &Handler{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
+	r := chi.NewRouter()
+	RegisterRoutes(r, h)
+
+	var body bytes.Buffer
+	writer := multipart.NewWriter(&body)
+	if err := writer.WriteField("purpose", "assistants"); err != nil {
+		t.Fatalf("write field failed: %v", err)
+	}
+	if err := writer.Close(); err != nil {
+		t.Fatalf("close writer failed: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/v1/files", &body)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", writer.FormDataContentType())
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}

internal/adapter/openai/handler_chat.go

@@ -5,6 +5,7 @@ import (
 	"encoding/json"
 	"io"
 	"net/http"
+	"strings"
 	"time"
 
 	"ds2api/internal/auth"
@@ -43,11 +44,20 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 
 	r = r.WithContext(auth.WithAuth(r.Context(), a))
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	stdReq, err := normalizeOpenAIChatRequest(h.Store, req, requestTraceID(r))
 	if err != nil {
 		writeOpenAIError(w, http.StatusBadRequest, err.Error())
@@ -127,18 +137,10 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, re
 	stripReferenceMarkers := h.compatStripReferenceMarkers()
 	finalThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
 	finalText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if writeUpstreamEmptyOutputError(w, finalThinking, finalText, result.ContentFilter) {
+	if writeUpstreamEmptyOutputError(w, finalText, result.ContentFilter) {
 		return
 	}
 	respBody := openaifmt.BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText, toolNames)
-	if result.OutputTokens > 0 {
-		if usage, ok := respBody["usage"].(map[string]any); ok {
-			usage["completion_tokens"] = result.OutputTokens
-			if prompt, ok := usage["prompt_tokens"].(int); ok {
-				usage["total_tokens"] = prompt + result.OutputTokens
-			}
-		}
-	}
 	writeJSON(w, http.StatusOK, respBody)
 }
 

internal/adapter/openai/handler_chat_auto_delete_test.go

@@ -27,6 +27,10 @@ func (m *autoDeleteModeDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _
 	return "pow", nil
 }
 
+func (m *autoDeleteModeDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, _ deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	return &deepseek.UploadFileResult{ID: "file-id", Filename: "file.txt", Bytes: 1, Status: "uploaded"}, nil
+}
+
 func (m *autoDeleteModeDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
 	return m.resp, nil
 }

internal/adapter/openai/handler_files.go

@@ -0,0 +1,104 @@
+package openai
+
+import (
+	"io"
+	"net/http"
+	"strings"
+	"time"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/deepseek"
+)
+
+const openAIUploadMaxMemory = 32 << 20
+
+func (h *Handler) UploadFile(w http.ResponseWriter, r *http.Request) {
+	a, err := h.Auth.Determine(r)
+	if err != nil {
+		status := http.StatusUnauthorized
+		detail := err.Error()
+		if err == auth.ErrNoAccount {
+			status = http.StatusTooManyRequests
+		}
+		writeOpenAIError(w, status, detail)
+		return
+	}
+	defer h.Auth.Release(a)
+	if !strings.HasPrefix(strings.ToLower(strings.TrimSpace(r.Header.Get("Content-Type"))), "multipart/form-data") {
+		writeOpenAIError(w, http.StatusBadRequest, "content-type must be multipart/form-data")
+		return
+	}
+	// Enforce a hard cap on the total request body size to prevent OOM
+	r.Body = http.MaxBytesReader(w, r.Body, openAIUploadMaxSize)
+	if err := r.ParseMultipartForm(openAIUploadMaxMemory); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "file size exceeds limit")
+			return
+		}
+		writeOpenAIError(w, http.StatusBadRequest, "invalid multipart form")
+		return
+	}
+	if r.MultipartForm != nil {
+		defer func() { _ = r.MultipartForm.RemoveAll() }()
+	}
+	r = r.WithContext(auth.WithAuth(r.Context(), a))
+	file, header, err := r.FormFile("file")
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, "file is required")
+		return
+	}
+	defer func() { _ = file.Close() }()
+	data, err := io.ReadAll(file)
+	if err != nil {
+		writeOpenAIError(w, http.StatusBadRequest, "failed to read uploaded file")
+		return
+	}
+	contentType := strings.TrimSpace(header.Header.Get("Content-Type"))
+	if contentType == "" && len(data) > 0 {
+		contentType = http.DetectContentType(data)
+	}
+	result, err := h.DS.UploadFile(r.Context(), a, deepseek.UploadFileRequest{
+		Filename:    header.Filename,
+		ContentType: contentType,
+		Purpose:     strings.TrimSpace(r.FormValue("purpose")),
+		Data:        data,
+	}, 3)
+	if err != nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "Failed to upload file.")
+		return
+	}
+	if result != nil && result.AccountID == "" {
+		result.AccountID = a.AccountID
+	}
+	writeJSON(w, http.StatusOK, buildOpenAIFileObject(result))
+}
+
+func buildOpenAIFileObject(result *deepseek.UploadFileResult) map[string]any {
+	if result == nil {
+		obj := map[string]any{
+			"id":             "",
+			"object":         "file",
+			"bytes":          0,
+			"created_at":     time.Now().Unix(),
+			"filename":       "",
+			"purpose":        "",
+			"status":         "uploaded",
+			"status_details": nil,
+		}
+		return obj
+	}
+	obj := map[string]any{
+		"id":             result.ID,
+		"object":         "file",
+		"bytes":          result.Bytes,
+		"created_at":     time.Now().Unix(),
+		"filename":       result.Filename,
+		"purpose":        result.Purpose,
+		"status":         result.Status,
+		"status_details": nil,
+	}
+	if result.AccountID != "" {
+		obj["account_id"] = result.AccountID
+	}
+	return obj
+}

internal/adapter/openai/handler_routes.go

@@ -13,6 +13,13 @@ import (
 	"ds2api/internal/util"
 )
 
+const (
+	// openAIUploadMaxSize limits total multipart request body size (100 MiB).
+	openAIUploadMaxSize = 100 << 20
+	// openAIGeneralMaxSize limits total JSON request body size (100 MiB).
+	openAIGeneralMaxSize = 100 << 20
+)
+
 // writeJSON is a package-internal alias kept to avoid mass-renaming across
 // every call-site in this package.
 var writeJSON = util.WriteJSON
@@ -46,6 +53,7 @@ func RegisterRoutes(r chi.Router, h *Handler) {
 	r.Post("/v1/chat/completions", h.ChatCompletions)
 	r.Post("/v1/responses", h.Responses)
 	r.Get("/v1/responses/{response_id}", h.GetResponseByID)
+	r.Post("/v1/files", h.UploadFile)
 	r.Post("/v1/embeddings", h.Embeddings)
 }
 

internal/adapter/openai/handler_toolcall_format.go

@@ -113,7 +113,7 @@ func formatIncrementalStreamToolCallDeltas(deltas []toolCallDelta, ids map[int]s
 	return out
 }
 
-func filterIncrementalToolCallDeltasByAllowed(deltas []toolCallDelta, allowedNames []string, seenNames map[int]string) []toolCallDelta {
+func filterIncrementalToolCallDeltasByAllowed(deltas []toolCallDelta, seenNames map[int]string) []toolCallDelta {
 	if len(deltas) == 0 {
 		return nil
 	}

internal/adapter/openai/handler_toolcall_test.go

@@ -275,7 +275,7 @@ func TestHandleNonStreamFencedToolCallExamplePromotesToolCall(t *testing.T) {
 	TestHandleNonStreamFencedToolCallExampleDoesNotPromoteToolCall(t)
 }
 
-func TestHandleNonStreamReturns502WhenUpstreamOutputEmpty(t *testing.T) {
+func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		`data: {"p":"response/content","v":""}`,
@@ -284,8 +284,8 @@ func TestHandleNonStreamReturns502WhenUpstreamOutputEmpty(t *testing.T) {
 	rec := httptest.NewRecorder()
 
 	h.handleNonStream(rec, context.Background(), resp, "cid-empty", "deepseek-chat", "prompt", false, nil)
-	if rec.Code != http.StatusBadGateway {
-		t.Fatalf("expected status 502 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected status 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
 	out := decodeJSONBody(t, rec.Body.String())
 	errObj, _ := out["error"].(map[string]any)
@@ -313,6 +313,25 @@ func TestHandleNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutp
 	}
 }
 
+func TestHandleNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
+	h := &Handler{}
+	resp := makeSSEHTTPResponse(
+		`data: {"p":"response/thinking_content","v":"Only thinking"}`,
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+
+	h.handleNonStream(rec, context.Background(), resp, "cid-thinking-only", "deepseek-reasoner", "prompt", true, nil)
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected status 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
+	}
+}
+
 func TestHandleStreamToolCallInterceptsWithoutRawContentLeak(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(

internal/adapter/openai/leaked_output_sanitize.go

@@ -2,13 +2,21 @@ package openai
 
 import (
 	"regexp"
+	"strings"
 )
 
 var emptyJSONFencePattern = regexp.MustCompile("(?is)```json\\s*```")
 var leakedToolCallArrayPattern = regexp.MustCompile(`(?is)\[\{\s*"function"\s*:\s*\{[\s\S]*?\}\s*,\s*"id"\s*:\s*"call[^"]*"\s*,\s*"type"\s*:\s*"function"\s*}\]`)
 var leakedToolResultBlobPattern = regexp.MustCompile(`(?is)<\s*\|\s*tool\s*\|\s*>\s*\{[\s\S]*?"tool_call_id"\s*:\s*"call[^"]*"\s*}`)
 
-// leakedMetaMarkerPattern matches DeepSeek special tokens in BOTH forms:
+var leakedThinkTagPattern = regexp.MustCompile(`(?is)</?\s*think\s*>`)
+
+// leakedBOSMarkerPattern matches DeepSeek BOS markers in BOTH forms:
+//   - ASCII underscore: <｜begin_of_sentence｜>
+//   - U+2581 variant:   <｜begin▁of▁sentence｜>
+var leakedBOSMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*begin[_▁]of[_▁]sentence\s*[｜\|]>`)
+
+// leakedMetaMarkerPattern matches the remaining DeepSeek special tokens in BOTH forms:
 //   - ASCII underscore: <｜end_of_sentence｜>, <｜end_of_toolresults｜>, <｜end_of_instructions｜>
 //   - U+2581 variant:   <｜end▁of▁sentence｜>, <｜end▁of▁toolresults｜>, <｜end▁of▁instructions｜>
 var leakedMetaMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*(?:assistant|tool|end[_▁]of[_▁]sentence|end[_▁]of[_▁]thinking|end[_▁]of[_▁]toolresults|end[_▁]of[_▁]instructions)\s*[｜\|]>`)
@@ -35,11 +43,48 @@ func sanitizeLeakedOutput(text string) string {
 	out := emptyJSONFencePattern.ReplaceAllString(text, "")
 	out = leakedToolCallArrayPattern.ReplaceAllString(out, "")
 	out = leakedToolResultBlobPattern.ReplaceAllString(out, "")
+	out = stripDanglingThinkSuffix(out)
+	out = leakedThinkTagPattern.ReplaceAllString(out, "")
+	out = leakedBOSMarkerPattern.ReplaceAllString(out, "")
 	out = leakedMetaMarkerPattern.ReplaceAllString(out, "")
 	out = sanitizeLeakedAgentXMLBlocks(out)
 	return out
 }
 
+func stripDanglingThinkSuffix(text string) string {
+	matches := leakedThinkTagPattern.FindAllStringIndex(text, -1)
+	if len(matches) == 0 {
+		return text
+	}
+	depth := 0
+	lastOpen := -1
+	for _, loc := range matches {
+		tag := strings.ToLower(text[loc[0]:loc[1]])
+		compact := strings.ReplaceAll(strings.ReplaceAll(strings.TrimSpace(tag), " ", ""), "\t", "")
+		if strings.HasPrefix(compact, "</") {
+			if depth > 0 {
+				depth--
+				if depth == 0 {
+					lastOpen = -1
+				}
+			}
+			continue
+		}
+		if depth == 0 {
+			lastOpen = loc[0]
+		}
+		depth++
+	}
+	if depth == 0 || lastOpen < 0 {
+		return text
+	}
+	prefix := text[:lastOpen]
+	if strings.TrimSpace(prefix) == "" {
+		return ""
+	}
+	return prefix
+}
+
 func sanitizeLeakedAgentXMLBlocks(text string) string {
 	out := text
 	for _, pattern := range leakedAgentXMLBlockPatterns {

internal/adapter/openai/leaked_output_sanitize_test.go

@@ -26,6 +26,22 @@ func TestSanitizeLeakedOutputRemovesStandaloneMetaMarkers(t *testing.T) {
 	}
 }
 
+func TestSanitizeLeakedOutputRemovesThinkAndBosMarkers(t *testing.T) {
+	raw := "A<think>B</think>C<｜begin▁of▁sentence｜>D<| begin_of_sentence |>E<｜begin_of_sentence｜>F"
+	got := sanitizeLeakedOutput(raw)
+	if got != "ABCDEF" {
+		t.Fatalf("unexpected sanitize result for think/BOS markers: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesDanglingThinkBlock(t *testing.T) {
+	raw := "Answer prefix<think>internal reasoning that never closes"
+	got := sanitizeLeakedOutput(raw)
+	if got != "Answer prefix" {
+		t.Fatalf("unexpected sanitize result for dangling think block: %q", got)
+	}
+}
+
 func TestSanitizeLeakedOutputRemovesAgentXMLLeaks(t *testing.T) {
 	raw := "Done.<attempt_completion><result>Some final answer</result></attempt_completion>"
 	got := sanitizeLeakedOutput(raw)

internal/adapter/openai/models_route_test.go

@@ -22,6 +22,24 @@ func TestGetModelRouteDirectAndAlias(t *testing.T) {
 		}
 	})
 
+	t.Run("direct_expert", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/v1/models/deepseek-expert-chat", nil)
+		rec := httptest.NewRecorder()
+		r.ServeHTTP(rec, req)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+		}
+	})
+
+	t.Run("direct_vision", func(t *testing.T) {
+		req := httptest.NewRequest(http.MethodGet, "/v1/models/deepseek-vision-chat", nil)
+		rec := httptest.NewRecorder()
+		r.ServeHTTP(rec, req)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+		}
+	})
+
 	t.Run("alias", func(t *testing.T) {
 		req := httptest.NewRequest(http.MethodGet, "/v1/models/gpt-4.1", nil)
 		rec := httptest.NewRecorder()

internal/adapter/openai/prompt_build.go

@@ -5,22 +5,22 @@ import (
 	"ds2api/internal/util"
 )
 
-func buildOpenAIFinalPrompt(messagesRaw []any, toolsRaw any, traceID string) (string, []string) {
-	return buildOpenAIFinalPromptWithPolicy(messagesRaw, toolsRaw, traceID, util.DefaultToolChoicePolicy())
+func buildOpenAIFinalPrompt(messagesRaw []any, toolsRaw any, traceID string, thinkingEnabled bool) (string, []string) {
+	return buildOpenAIFinalPromptWithPolicy(messagesRaw, toolsRaw, traceID, util.DefaultToolChoicePolicy(), thinkingEnabled)
 }
 
-func buildOpenAIFinalPromptWithPolicy(messagesRaw []any, toolsRaw any, traceID string, toolPolicy util.ToolChoicePolicy) (string, []string) {
+func buildOpenAIFinalPromptWithPolicy(messagesRaw []any, toolsRaw any, traceID string, toolPolicy util.ToolChoicePolicy, thinkingEnabled bool) (string, []string) {
 	messages := normalizeOpenAIMessagesForPrompt(messagesRaw, traceID)
 	toolNames := []string{}
 	if tools, ok := toolsRaw.([]any); ok && len(tools) > 0 {
 		messages, toolNames = injectToolPrompt(messages, tools, toolPolicy)
 	}
-	return deepseek.MessagesPrepare(messages), toolNames
+	return deepseek.MessagesPrepareWithThinking(messages, thinkingEnabled), toolNames
 }
 
 // BuildPromptForAdapter exposes the OpenAI-compatible prompt building flow so
 // other protocol adapters (for example Gemini) can reuse the same tool/history
 // normalization logic and remain behavior-compatible with chat/completions.
-func BuildPromptForAdapter(messagesRaw []any, toolsRaw any, traceID string) (string, []string) {
-	return buildOpenAIFinalPrompt(messagesRaw, toolsRaw, traceID)
+func BuildPromptForAdapter(messagesRaw []any, toolsRaw any, traceID string, thinkingEnabled bool) (string, []string) {
+	return buildOpenAIFinalPrompt(messagesRaw, toolsRaw, traceID, thinkingEnabled)
 }

internal/adapter/openai/prompt_build_test.go

@@ -40,7 +40,7 @@ func TestBuildOpenAIFinalPrompt_HandlerPathIncludesToolRoundtripSemantics(t *tes
 		},
 	}
 
-	finalPrompt, toolNames := buildOpenAIFinalPrompt(messages, tools, "")
+	finalPrompt, toolNames := buildOpenAIFinalPrompt(messages, tools, "", false)
 	if len(toolNames) != 1 || toolNames[0] != "get_weather" {
 		t.Fatalf("unexpected tool names: %#v", toolNames)
 	}
@@ -73,17 +73,14 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 		},
 	}
 
-	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "")
-	if !strings.Contains(finalPrompt, "After receiving a tool result, use it directly.") {
-		t.Fatalf("vercel prepare finalPrompt missing final-answer instruction: %q", finalPrompt)
-	}
-	if !strings.Contains(finalPrompt, "Only call another tool if the result is insufficient.") {
-		t.Fatalf("vercel prepare finalPrompt missing retry guard instruction: %q", finalPrompt)
+	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "", false)
+	if !strings.Contains(finalPrompt, "Remember: Output ONLY the <tool_calls>...</tool_calls> XML block when calling tools.") {
+		t.Fatalf("vercel prepare finalPrompt missing final tool-call anchor instruction: %q", finalPrompt)
 	}
 	if !strings.Contains(finalPrompt, "TOOL CALL FORMAT") {
 		t.Fatalf("vercel prepare finalPrompt missing xml format instruction: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "Do NOT wrap the XML in markdown code fences") {
+	if !strings.Contains(finalPrompt, "Do NOT wrap XML in markdown fences") {
 		t.Fatalf("vercel prepare finalPrompt missing no-fence xml instruction: %q", finalPrompt)
 	}
 	if strings.Contains(finalPrompt, "```json") {

internal/adapter/openai/responses_embeddings_test.go

@@ -156,6 +156,33 @@ func TestNormalizeResponsesInputAsMessagesFunctionCallItemPreservesConcatenatedA
 	}
 }
 
+func TestCollectOpenAIRefFileIDs(t *testing.T) {
+	got := collectOpenAIRefFileIDs(map[string]any{
+		"ref_file_ids": []any{"file-top", "file-dup"},
+		"attachments": []any{
+			map[string]any{"file_id": "file-attachment"},
+		},
+		"input": []any{
+			map[string]any{
+				"type": "message",
+				"content": []any{
+					map[string]any{"type": "input_file", "file_id": "file-input"},
+					map[string]any{"type": "input_file", "id": "file-dup"},
+				},
+			},
+		},
+	})
+	want := []string{"file-top", "file-dup", "file-attachment", "file-input"}
+	if len(got) != len(want) {
+		t.Fatalf("expected %d file ids, got %#v", len(want), got)
+	}
+	for i, id := range want {
+		if got[i] != id {
+			t.Fatalf("unexpected file ids at %d: got=%#v want=%#v", i, got, want)
+		}
+	}
+}
+
 func TestExtractEmbeddingInputs(t *testing.T) {
 	got := extractEmbeddingInputs([]any{"a", "b"})
 	if len(got) != 2 || got[0] != "a" || got[1] != "b" {

internal/adapter/openai/responses_handler.go

@@ -65,11 +65,20 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
+	r.Body = http.MaxBytesReader(w, r.Body, openAIGeneralMaxSize)
 	var req map[string]any
 	if err := json.NewDecoder(r.Body).Decode(&req); err != nil {
+		if strings.Contains(strings.ToLower(err.Error()), "too large") {
+			writeOpenAIError(w, http.StatusRequestEntityTooLarge, "request body too large")
+			return
+		}
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	traceID := requestTraceID(r)
 	stdReq, err := normalizeOpenAIResponsesRequest(h.Store, req, traceID)
 	if err != nil {
@@ -117,7 +126,7 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 	stripReferenceMarkers := h.compatStripReferenceMarkers()
 	sanitizedThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
 	sanitizedText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if writeUpstreamEmptyOutputError(w, sanitizedThinking, sanitizedText, result.ContentFilter) {
+	if writeUpstreamEmptyOutputError(w, sanitizedText, result.ContentFilter) {
 		return
 	}
 	textParsed := toolcall.ParseStandaloneToolCallsDetailed(sanitizedText, toolNames)
@@ -130,14 +139,6 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 	}
 
 	responseObj := openaifmt.BuildResponseObject(responseID, model, finalPrompt, sanitizedThinking, sanitizedText, toolNames)
-	if result.OutputTokens > 0 {
-		if usage, ok := responseObj["usage"].(map[string]any); ok {
-			usage["output_tokens"] = result.OutputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + result.OutputTokens
-			}
-		}
-	}
 	h.getResponseStore().put(owner, responseID, responseObj)
 	writeJSON(w, http.StatusOK, responseObj)
 }

internal/adapter/openai/responses_stream_runtime_core.go

@@ -51,7 +51,6 @@ type responsesStreamRuntime struct {
 	messagePartAdded  bool
 	sequence          int
 	failed            bool
-	outputTokens      int
 
 	persistResponse func(obj map[string]any)
 }
@@ -100,6 +99,30 @@ func newResponsesStreamRuntime(
 	}
 }
 
+func (s *responsesStreamRuntime) failResponse(message, code string) {
+	s.failed = true
+	failedResp := map[string]any{
+		"id":          s.responseID,
+		"type":        "response",
+		"object":      "response",
+		"model":       s.model,
+		"status":      "failed",
+		"output":      []any{},
+		"output_text": "",
+		"error": map[string]any{
+			"message": message,
+			"type":    "invalid_request_error",
+			"code":    code,
+			"param":   nil,
+		},
+	}
+	if s.persistResponse != nil {
+		s.persistResponse(failedResp)
+	}
+	s.sendEvent("response.failed", openaifmt.BuildResponsesFailedPayload(s.responseID, s.model, message, code))
+	s.sendDone()
+}
+
 func (s *responsesStreamRuntime) finalize() {
 	finalThinking := s.thinking.String()
 	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
@@ -122,41 +145,21 @@ func (s *responsesStreamRuntime) finalize() {
 	s.closeMessageItem()
 
 	if s.toolChoice.IsRequired() && len(detected) == 0 {
-		s.failed = true
-		message := "tool_choice requires at least one valid tool call."
-		failedResp := map[string]any{
-			"id":          s.responseID,
-			"type":        "response",
-			"object":      "response",
-			"model":       s.model,
-			"status":      "failed",
-			"output":      []any{},
-			"output_text": "",
-			"error": map[string]any{
-				"message": message,
-				"type":    "invalid_request_error",
-				"code":    "tool_choice_violation",
-				"param":   nil,
-			},
+		s.failResponse("tool_choice requires at least one valid tool call.", "tool_choice_violation")
+		return
+	}
+	if len(detected) == 0 && strings.TrimSpace(finalText) == "" {
+		code := "upstream_empty_output"
+		message := "Upstream model returned empty output."
+		if finalThinking != "" {
+			message = "Upstream model returned reasoning without visible output."
 		}
-		if s.persistResponse != nil {
-			s.persistResponse(failedResp)
-		}
-		s.sendEvent("response.failed", openaifmt.BuildResponsesFailedPayload(s.responseID, s.model, message, "tool_choice_violation"))
-		s.sendDone()
+		s.failResponse(message, code)
 		return
 	}
 	s.closeIncompleteFunctionItems()
 
 	obj := s.buildCompletedResponseObject(finalThinking, finalText, detected)
-	if s.outputTokens > 0 {
-		if usage, ok := obj["usage"].(map[string]any); ok {
-			usage["output_tokens"] = s.outputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + s.outputTokens
-			}
-		}
-	}
 	if s.persistResponse != nil {
 		s.persistResponse(obj)
 	}
@@ -185,9 +188,6 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
-	if parsed.OutputTokens > 0 {
-		s.outputTokens = parsed.OutputTokens
-	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" || parsed.Stop {
 		return streamengine.ParsedDecision{Stop: true}
 	}

internal/adapter/openai/responses_stream_runtime_events.go

@@ -48,7 +48,7 @@ func (s *responsesStreamRuntime) processToolStreamEvents(events []toolStreamEven
 			if !s.emitEarlyToolDeltas {
 				continue
 			}
-			filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.toolNames, s.functionNames)
+			filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.functionNames)
 			if len(filtered) == 0 {
 				continue
 			}

internal/adapter/openai/responses_stream_test.go

@@ -518,6 +518,44 @@ func TestHandleResponsesStreamRequiredMalformedToolPayloadFails(t *testing.T) {
 	}
 }
 
+func TestHandleResponsesStreamFailsWhenUpstreamHasOnlyThinking(t *testing.T) {
+	h := &Handler{}
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	rec := httptest.NewRecorder()
+
+	sseLine := func(path, value string) string {
+		b, _ := json.Marshal(map[string]any{
+			"p": path,
+			"v": value,
+		})
+		return "data: " + string(b) + "\n"
+	}
+
+	streamBody := sseLine("response/thinking_content", "Only thinking") + "data: [DONE]\n"
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(streamBody)),
+	}
+
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, false, nil, util.DefaultToolChoicePolicy(), "")
+
+	body := rec.Body.String()
+	if !strings.Contains(body, "event: response.failed") {
+		t.Fatalf("expected response.failed event, body=%s", body)
+	}
+	if strings.Contains(body, "event: response.completed") {
+		t.Fatalf("did not expect response.completed, body=%s", body)
+	}
+	payload, ok := extractSSEEventPayload(body, "response.failed")
+	if !ok {
+		t.Fatalf("expected response.failed payload, body=%s", body)
+	}
+	errObj, _ := payload["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", payload)
+	}
+}
+
 func TestHandleResponsesStreamAllowsUnknownToolName(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -627,7 +665,7 @@ func TestHandleResponsesNonStreamToolChoiceNoneStillAllowsFunctionCall(t *testin
 	}
 }
 
-func TestHandleResponsesNonStreamReturns502WhenUpstreamOutputEmpty(t *testing.T) {
+func TestHandleResponsesNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	h := &Handler{}
 	rec := httptest.NewRecorder()
 	resp := &http.Response{
@@ -639,8 +677,8 @@ func TestHandleResponsesNonStreamReturns502WhenUpstreamOutputEmpty(t *testing.T)
 	}
 
 	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-chat", "prompt", false, nil, util.DefaultToolChoicePolicy(), "")
-	if rec.Code != http.StatusBadGateway {
-		t.Fatalf("expected 502 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
 	out := decodeJSONBody(t, rec.Body.String())
 	errObj, _ := out["error"].(map[string]any)
@@ -671,6 +709,28 @@ func TestHandleResponsesNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWi
 	}
 }
 
+func TestHandleResponsesNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
+	h := &Handler{}
+	rec := httptest.NewRecorder()
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body: io.NopCloser(strings.NewReader(
+			`data: {"p":"response/thinking_content","v":"Only thinking"}` + "\n" +
+				`data: [DONE]` + "\n",
+		)),
+	}
+
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-reasoner", "prompt", true, nil, util.DefaultToolChoicePolicy(), "")
+	if rec.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	out := decodeJSONBody(t, rec.Body.String())
+	errObj, _ := out["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", out)
+	}
+}
+
 func extractSSEEventPayload(body, targetEvent string) (map[string]any, bool) {
 	scanner := bufio.NewScanner(strings.NewReader(body))
 	matched := false

internal/adapter/openai/standard_request.go

@@ -24,9 +24,10 @@ func normalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID
 		responseModel = resolvedModel
 	}
 	toolPolicy := util.DefaultToolChoicePolicy()
-	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy)
+	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy, thinkingEnabled)
 	toolNames = ensureToolDetectionEnabled(toolNames, req["tools"])
 	passThrough := collectOpenAIChatPassThrough(req)
+	refFileIDs := collectOpenAIRefFileIDs(req)
 
 	return util.StandardRequest{
 		Surface:        "openai_chat",
@@ -40,6 +41,7 @@ func normalizeOpenAIChatRequest(store ConfigReader, req map[string]any, traceID
 		Stream:         util.ToBool(req["stream"]),
 		Thinking:       thinkingEnabled,
 		Search:         searchEnabled,
+		RefFileIDs:     refFileIDs,
 		PassThrough:    passThrough,
 	}, nil
 }
@@ -74,12 +76,13 @@ func normalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 	if err != nil {
 		return util.StandardRequest{}, err
 	}
-	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy)
+	finalPrompt, toolNames := buildOpenAIFinalPromptWithPolicy(messagesRaw, req["tools"], traceID, toolPolicy, thinkingEnabled)
 	toolNames = ensureToolDetectionEnabled(toolNames, req["tools"])
 	if !toolPolicy.IsNone() {
 		toolPolicy.Allowed = namesToSet(toolNames)
 	}
 	passThrough := collectOpenAIChatPassThrough(req)
+	refFileIDs := collectOpenAIRefFileIDs(req)
 
 	return util.StandardRequest{
 		Surface:        "openai_responses",
@@ -93,6 +96,7 @@ func normalizeOpenAIResponsesRequest(store ConfigReader, req map[string]any, tra
 		Stream:         util.ToBool(req["stream"]),
 		Thinking:       thinkingEnabled,
 		Search:         searchEnabled,
+		RefFileIDs:     refFileIDs,
 		PassThrough:    passThrough,
 	}, nil
 }

internal/adapter/openai/standard_request_test.go

@@ -41,6 +41,36 @@ func TestNormalizeOpenAIChatRequest(t *testing.T) {
 	}
 }
 
+func TestNormalizeOpenAIChatRequestCollectsRefFileIDs(t *testing.T) {
+	store := newEmptyStoreForNormalizeTest(t)
+	req := map[string]any{
+		"model": "gpt-5-codex",
+		"messages": []any{
+			map[string]any{
+				"role": "user",
+				"content": []any{
+					map[string]any{"type": "input_text", "text": "hello"},
+					map[string]any{"type": "input_file", "file_id": "file-msg"},
+				},
+			},
+		},
+		"attachments": []any{
+			map[string]any{"file_id": "file-attachment"},
+		},
+		"ref_file_ids": []any{"file-top", "file-attachment"},
+	}
+	n, err := normalizeOpenAIChatRequest(store, req, "")
+	if err != nil {
+		t.Fatalf("normalize failed: %v", err)
+	}
+	if len(n.RefFileIDs) != 3 {
+		t.Fatalf("expected 3 distinct file ids, got %#v", n.RefFileIDs)
+	}
+	if n.RefFileIDs[0] != "file-top" || n.RefFileIDs[1] != "file-attachment" || n.RefFileIDs[2] != "file-msg" {
+		t.Fatalf("unexpected file ids: %#v", n.RefFileIDs)
+	}
+}
+
 func TestNormalizeOpenAIResponsesRequestInput(t *testing.T) {
 	store := newEmptyStoreForNormalizeTest(t)
 	req := map[string]any{

internal/adapter/openai/stream_status_test.go

@@ -50,6 +50,10 @@ func (m streamStatusDSStub) GetPow(_ context.Context, _ *auth.RequestAuth, _ int
 	return "pow", nil
 }
 
+func (m streamStatusDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, _ deepseek.UploadFileRequest, _ int) (*deepseek.UploadFileResult, error) {
+	return &deepseek.UploadFileResult{ID: "file-id", Filename: "file.txt", Bytes: 1, Status: "uploaded"}, nil
+}
+
 func (m streamStatusDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
 	return m.resp, nil
 }
@@ -238,3 +242,140 @@ func TestChatCompletionsStreamContentFilterStopsNormallyWithoutLeak(t *testing.T
 		t.Fatalf("expected finish_reason=stop for content-filter upstream stop, got %#v", choice["finish_reason"])
 	}
 }
+
+func TestChatCompletionsStreamEmitsFailureFrameWhenUpstreamOutputEmpty(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS:    streamStatusDSStub{resp: makeOpenAISSEHTTPResponse("data: [DONE]")},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","messages":[{"role":"user","content":"hi"}],"stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	if len(frames) != 1 {
+		t.Fatalf("expected one failure frame, got %#v body=%s", frames, rec.Body.String())
+	}
+	last := frames[0]
+	statusCode, ok := last["status_code"].(float64)
+	if !ok || int(statusCode) != http.StatusTooManyRequests {
+		t.Fatalf("expected status_code=429, got %#v body=%s", last["status_code"], rec.Body.String())
+	}
+	errObj, _ := last["error"].(map[string]any)
+	if asString(errObj["code"]) != "upstream_empty_output" {
+		t.Fatalf("expected code=upstream_empty_output, got %#v", last)
+	}
+}
+
+func TestResponsesStreamUsageIgnoresBatchAccumulatedTokenUsage(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
+			`data: {"p":"response/content","v":"hello"}`,
+			`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":190},{"p":"quasi_status","v":"FINISHED"}]}`,
+		)},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","input":"hi","stream":true}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	if len(frames) == 0 {
+		t.Fatalf("expected at least one json frame, body=%s", rec.Body.String())
+	}
+	last := frames[len(frames)-1]
+	resp, _ := last["response"].(map[string]any)
+	if resp == nil {
+		t.Fatalf("expected response payload in final frame, got %#v", last)
+	}
+	usage, _ := resp["usage"].(map[string]any)
+	if usage == nil {
+		t.Fatalf("expected usage in response payload, got %#v", resp)
+	}
+	if got, _ := usage["output_tokens"].(float64); int(got) == 190 {
+		t.Fatalf("expected upstream accumulated token usage to be ignored, got %#v", usage["output_tokens"])
+	}
+}
+
+func TestResponsesNonStreamUsageIgnoresPromptAndOutputTokenUsage(t *testing.T) {
+	statuses := make([]int, 0, 1)
+	h := &Handler{
+		Store: mockOpenAIConfig{wideInput: true},
+		Auth:  streamStatusAuthStub{},
+		DS: streamStatusDSStub{resp: makeOpenAISSEHTTPResponse(
+			`data: {"p":"response/content","v":"ok"}`,
+			`data: {"p":"response","o":"BATCH","v":[{"p":"token_usage","v":{"prompt_tokens":11,"completion_tokens":29}},{"p":"quasi_status","v":"FINISHED"}]}`,
+		)},
+	}
+	r := chi.NewRouter()
+	r.Use(captureStatusMiddleware(&statuses))
+	RegisterRoutes(r, h)
+
+	reqBody := `{"model":"deepseek-chat","input":"hi","stream":false}`
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", strings.NewReader(reqBody))
+	req.Header.Set("Authorization", "Bearer direct-token")
+	req.Header.Set("Content-Type", "application/json")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if len(statuses) != 1 || statuses[0] != http.StatusOK {
+		t.Fatalf("expected captured status 200, got %#v", statuses)
+	}
+	var out map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &out); err != nil {
+		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
+	}
+	usage, _ := out["usage"].(map[string]any)
+	if usage == nil {
+		t.Fatalf("expected usage object, got %#v", out)
+	}
+	input, _ := usage["input_tokens"].(float64)
+	output, _ := usage["output_tokens"].(float64)
+	total, _ := usage["total_tokens"].(float64)
+	if int(output) == 29 {
+		t.Fatalf("expected upstream completion token usage to be ignored, got %#v", usage["output_tokens"])
+	}
+	if int(total) != int(input)+int(output) {
+		t.Fatalf("expected total_tokens=input_tokens+output_tokens, usage=%#v", usage)
+	}
+}

internal/adapter/openai/upstream_empty.go

@@ -2,14 +2,14 @@ package openai
 
 import "net/http"
 
-func writeUpstreamEmptyOutputError(w http.ResponseWriter, thinking, text string, contentFilter bool) bool {
-	if thinking != "" || text != "" {
+func writeUpstreamEmptyOutputError(w http.ResponseWriter, text string, contentFilter bool) bool {
+	if text != "" {
 		return false
 	}
 	if contentFilter {
 		writeOpenAIErrorWithCode(w, http.StatusBadRequest, "Upstream content filtered the response and returned no output.", "content_filter")
 		return true
 	}
-	writeOpenAIErrorWithCode(w, http.StatusBadGateway, "Upstream model returned empty output.", "upstream_empty_output")
+	writeOpenAIErrorWithCode(w, http.StatusTooManyRequests, "Upstream model returned empty output.", "upstream_empty_output")
 	return true
 }

internal/adapter/openai/vercel_stream.go

@@ -52,6 +52,10 @@ func (h *Handler) handleVercelStreamPrepare(w http.ResponseWriter, r *http.Reque
 		writeOpenAIError(w, http.StatusBadRequest, "invalid json")
 		return
 	}
+	if err := h.preprocessInlineFileInputs(r.Context(), a, req); err != nil {
+		writeOpenAIInlineFileError(w, err)
+		return
+	}
 	if !util.ToBool(req["stream"]) {
 		writeOpenAIError(w, http.StatusBadRequest, "stream must be true")
 		return

internal/admin/handler.go

@@ -26,9 +26,15 @@ func RegisterRoutes(r chi.Router, h *Handler) {
 		pr.Get("/config/export", h.configExport)
 		pr.Post("/keys", h.addKey)
 		pr.Delete("/keys/{key}", h.deleteKey)
+		pr.Get("/proxies", h.listProxies)
+		pr.Post("/proxies", h.addProxy)
+		pr.Put("/proxies/{proxyID}", h.updateProxy)
+		pr.Delete("/proxies/{proxyID}", h.deleteProxy)
+		pr.Post("/proxies/test", h.testProxy)
 		pr.Get("/accounts", h.listAccounts)
 		pr.Post("/accounts", h.addAccount)
 		pr.Delete("/accounts/{identifier}", h.deleteAccount)
+		pr.Put("/accounts/{identifier}/proxy", h.updateAccountProxy)
 		pr.Get("/queue/status", h.queueStatus)
 		pr.Post("/accounts/test", h.testSingleAccount)
 		pr.Post("/accounts/test-all", h.testAllAccounts)

internal/admin/handler_accounts_crud.go

@@ -68,6 +68,7 @@ func (h *Handler) listAccounts(w http.ResponseWriter, r *http.Request) {
 			"identifier":    acc.Identifier(),
 			"email":         acc.Email,
 			"mobile":        acc.Mobile,
+			"proxy_id":      acc.ProxyID,
 			"has_password":  acc.Password != "",
 			"has_token":     token != "",
 			"token_preview": preview,
@@ -86,6 +87,11 @@ func (h *Handler) addAccount(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 	err := h.Store.Update(func(c *config.Config) error {
+		if acc.ProxyID != "" {
+			if _, ok := findProxyByID(*c, acc.ProxyID); !ok {
+				return fmt.Errorf("代理不存在")
+			}
+		}
 		mobileKey := config.CanonicalMobileKey(acc.Mobile)
 		for _, a := range c.Accounts {
 			if acc.Email != "" && a.Email == acc.Email {

internal/admin/handler_accounts_testing.go

@@ -15,8 +15,17 @@ import (
 	"ds2api/internal/config"
 	"ds2api/internal/deepseek"
 	"ds2api/internal/sse"
+	"ds2api/internal/util"
 )
 
+type modelAliasSnapshotReader struct {
+	aliases map[string]string
+}
+
+func (m modelAliasSnapshotReader) ModelAliases() map[string]string {
+	return m.aliases
+}
+
 func (h *Handler) testSingleAccount(w http.ResponseWriter, r *http.Request) {
 	var req map[string]any
 	_ = json.NewDecoder(r.Body).Decode(&req)
@@ -115,10 +124,11 @@ func (h *Handler) testAccount(ctx context.Context, acc config.Account, model, me
 		result["message"] = "登录成功但写入运行时 token 失败: " + err.Error()
 		return result
 	}
-	authCtx := &authn.RequestAuth{UseConfigToken: false, DeepSeekToken: token}
-	sessionID, err := h.DS.CreateSession(ctx, authCtx, 1)
+	authCtx := &authn.RequestAuth{UseConfigToken: false, DeepSeekToken: token, AccountID: identifier, Account: acc}
+	proxyCtx := authn.WithAuth(ctx, authCtx)
+	sessionID, err := h.DS.CreateSession(proxyCtx, authCtx, 1)
 	if err != nil {
-		newToken, loginErr := h.DS.Login(ctx, acc)
+		newToken, loginErr := h.DS.Login(proxyCtx, acc)
 		if loginErr != nil {
 			result["message"] = "创建会话失败: " + err.Error()
 			return result
@@ -129,7 +139,7 @@ func (h *Handler) testAccount(ctx context.Context, acc config.Account, model, me
 			result["message"] = "刷新 token 成功但写入运行时 token 失败: " + err.Error()
 			return result
 		}
-		sessionID, err = h.DS.CreateSession(ctx, authCtx, 1)
+		sessionID, err = h.DS.CreateSession(proxyCtx, authCtx, 1)
 		if err != nil {
 			result["message"] = "创建会话失败: " + err.Error()
 			return result
@@ -137,7 +147,7 @@ func (h *Handler) testAccount(ctx context.Context, acc config.Account, model, me
 	}
 
 	// 获取会话数量
-	sessionStats, sessionErr := h.DS.GetSessionCountForToken(ctx, token)
+	sessionStats, sessionErr := h.DS.GetSessionCountForToken(proxyCtx, token)
 	if sessionErr == nil && sessionStats != nil {
 		result["session_count"] = sessionStats.FirstPageCount
 	}
@@ -149,17 +159,28 @@ func (h *Handler) testAccount(ctx context.Context, acc config.Account, model, me
 		return result
 	}
 	thinking, search, ok := config.GetModelConfig(model)
+	resolvedModel, resolved := config.ResolveModel(modelAliasSnapshotReader{
+		aliases: h.Store.Snapshot().ModelAliases,
+	}, model)
+	if resolved {
+		model = resolvedModel
+		thinking, search, ok = config.GetModelConfig(model)
+	}
 	if !ok {
 		thinking, search = false, false
 	}
-	_ = search
-	pow, err := h.DS.GetPow(ctx, authCtx, 1)
+	pow, err := h.DS.GetPow(proxyCtx, authCtx, 1)
 	if err != nil {
 		result["message"] = "获取 PoW 失败: " + err.Error()
 		return result
 	}
-	payload := map[string]any{"chat_session_id": sessionID, "prompt": deepseek.MessagesPrepare([]map[string]any{{"role": "user", "content": message}}), "ref_file_ids": []any{}, "thinking_enabled": thinking, "search_enabled": search}
-	resp, err := h.DS.CallCompletion(ctx, authCtx, payload, pow, 1)
+	payload := util.StandardRequest{
+		ResolvedModel: model,
+		FinalPrompt:   deepseek.MessagesPrepare([]map[string]any{{"role": "user", "content": message}}),
+		Thinking:      thinking,
+		Search:        search,
+	}.CompletionPayload(sessionID)
+	resp, err := h.DS.CallCompletion(proxyCtx, authCtx, payload, pow, 1)
 	if err != nil {
 		result["message"] = "请求失败: " + err.Error()
 		return result
@@ -244,25 +265,29 @@ func (h *Handler) deleteAllSessions(w http.ResponseWriter, r *http.Request) {
 	}
 
 	// 每次先登录刷新一次 token，避免使用过期 token。
-	token, err := h.DS.Login(r.Context(), acc)
+	authCtx := &authn.RequestAuth{UseConfigToken: false, AccountID: acc.Identifier(), Account: acc}
+	proxyCtx := authn.WithAuth(r.Context(), authCtx)
+	token, err := h.DS.Login(proxyCtx, acc)
 	if err != nil {
 		writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "登录失败: " + err.Error()})
 		return
 	}
 	_ = h.Store.UpdateAccountToken(acc.Identifier(), token)
+	authCtx.DeepSeekToken = token
 
 	// 删除所有会话
-	err = h.DS.DeleteAllSessionsForToken(r.Context(), token)
+	err = h.DS.DeleteAllSessionsForToken(proxyCtx, token)
 	if err != nil {
 		// token 可能过期，尝试重新登录并重试一次
-		newToken, loginErr := h.DS.Login(r.Context(), acc)
+		newToken, loginErr := h.DS.Login(proxyCtx, acc)
 		if loginErr != nil {
 			writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "删除失败: " + err.Error()})
 			return
 		}
 		token = newToken
 		_ = h.Store.UpdateAccountToken(acc.Identifier(), token)
-		if retryErr := h.DS.DeleteAllSessionsForToken(r.Context(), token); retryErr != nil {
+		authCtx.DeepSeekToken = token
+		if retryErr := h.DS.DeleteAllSessionsForToken(proxyCtx, token); retryErr != nil {
 			writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "删除失败: " + retryErr.Error()})
 			return
 		}

internal/admin/handler_accounts_testing_test.go

@@ -5,6 +5,7 @@ import (
 	"context"
 	"encoding/json"
 	"errors"
+	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
@@ -133,3 +134,78 @@ func TestDeleteAllSessions_RetryWithReloginOnDeleteFailure(t *testing.T) {
 		t.Fatalf("expected refreshed token persisted, got %q", updated.Token)
 	}
 }
+
+type completionPayloadDSMock struct {
+	payload map[string]any
+}
+
+func (m *completionPayloadDSMock) Login(_ context.Context, _ config.Account) (string, error) {
+	return "new-token", nil
+}
+
+func (m *completionPayloadDSMock) CreateSession(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "session-id", nil
+}
+
+func (m *completionPayloadDSMock) GetPow(_ context.Context, _ *auth.RequestAuth, _ int) (string, error) {
+	return "pow-ok", nil
+}
+
+func (m *completionPayloadDSMock) CallCompletion(_ context.Context, _ *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
+	m.payload = payload
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader("data: {\"v\":\"ok\"}\n\ndata: [DONE]\n\n")),
+	}, nil
+}
+
+func (m *completionPayloadDSMock) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	return nil
+}
+
+func (m *completionPayloadDSMock) GetSessionCountForToken(_ context.Context, _ string) (*deepseek.SessionStats, error) {
+	return &deepseek.SessionStats{Success: true}, nil
+}
+
+func TestTestAccount_MessageModeUsesExpertModelTypeForExpertModel(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{"accounts":[{"email":"batch@example.com","password":"pwd","token":"seed-token"}]}`)
+	store := config.LoadStore()
+	ds := &completionPayloadDSMock{}
+	h := &Handler{Store: store, DS: ds}
+	acc, ok := store.FindAccount("batch@example.com")
+	if !ok {
+		t.Fatal("expected test account")
+	}
+
+	result := h.testAccount(context.Background(), acc, "deepseek-expert-chat", "hello")
+
+	if ok, _ := result["success"].(bool); !ok {
+		t.Fatalf("expected success=true, got %#v", result)
+	}
+	if got := ds.payload["model_type"]; got != "expert" {
+		t.Fatalf("expected model_type expert, got %#v", got)
+	}
+	if got := ds.payload["chat_session_id"]; got != "session-id" {
+		t.Fatalf("unexpected chat_session_id: %#v", got)
+	}
+}
+
+func TestTestAccount_MessageModeUsesVisionModelTypeForVisionModel(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{"accounts":[{"email":"batch@example.com","password":"pwd","token":"seed-token"}]}`)
+	store := config.LoadStore()
+	ds := &completionPayloadDSMock{}
+	h := &Handler{Store: store, DS: ds}
+	acc, ok := store.FindAccount("batch@example.com")
+	if !ok {
+		t.Fatal("expected test account")
+	}
+
+	result := h.testAccount(context.Background(), acc, "deepseek-vision-chat", "hello")
+
+	if ok, _ := result["success"].(bool); !ok {
+		t.Fatalf("expected success=true, got %#v", result)
+	}
+	if got := ds.payload["model_type"]; got != "vision" {
+		t.Fatalf("expected model_type vision, got %#v", got)
+	}
+}

internal/admin/handler_config_read.go

@@ -3,6 +3,8 @@ package admin
 import (
 	"net/http"
 	"strings"
+
+	"ds2api/internal/config"
 )
 
 func (h *Handler) getConfig(w http.ResponseWriter, _ *http.Request) {
@@ -10,6 +12,7 @@ func (h *Handler) getConfig(w http.ResponseWriter, _ *http.Request) {
 	safe := map[string]any{
 		"keys":                  snap.Keys,
 		"accounts":              []map[string]any{},
+		"proxies":               []map[string]any{},
 		"env_backed":            h.Store.IsEnvBacked(),
 		"env_source_present":    h.Store.HasEnvConfigSource(),
 		"env_writeback_enabled": h.Store.IsEnvWritebackEnabled(),
@@ -36,12 +39,27 @@ func (h *Handler) getConfig(w http.ResponseWriter, _ *http.Request) {
 			"identifier":    acc.Identifier(),
 			"email":         acc.Email,
 			"mobile":        acc.Mobile,
+			"proxy_id":      acc.ProxyID,
 			"has_password":  strings.TrimSpace(acc.Password) != "",
 			"has_token":     token != "",
 			"token_preview": preview,
 		})
 	}
 	safe["accounts"] = accounts
+	proxies := make([]map[string]any, 0, len(snap.Proxies))
+	for _, proxy := range snap.Proxies {
+		proxy = config.NormalizeProxy(proxy)
+		proxies = append(proxies, map[string]any{
+			"id":           proxy.ID,
+			"name":         proxy.Name,
+			"type":         proxy.Type,
+			"host":         proxy.Host,
+			"port":         proxy.Port,
+			"username":     proxy.Username,
+			"has_password": strings.TrimSpace(proxy.Password) != "",
+		})
+	}
+	safe["proxies"] = proxies
 	writeJSON(w, http.StatusOK, safe)
 }
 

internal/admin/handler_proxies.go

@@ -0,0 +1,202 @@
+package admin
+
+import (
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/url"
+	"strings"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/config"
+	"ds2api/internal/deepseek"
+)
+
+var proxyConnectivityTester = func(ctx context.Context, proxy config.Proxy) map[string]any {
+	return deepseek.TestProxyConnectivity(ctx, proxy)
+}
+
+func validateProxyMutation(cfg *config.Config) error {
+	if cfg == nil {
+		return nil
+	}
+	if err := config.ValidateProxyConfig(cfg.Proxies); err != nil {
+		return err
+	}
+	return config.ValidateAccountProxyReferences(cfg.Accounts, cfg.Proxies)
+}
+
+func proxyResponse(proxy config.Proxy) map[string]any {
+	proxy = config.NormalizeProxy(proxy)
+	return map[string]any{
+		"id":           proxy.ID,
+		"name":         proxy.Name,
+		"type":         proxy.Type,
+		"host":         proxy.Host,
+		"port":         proxy.Port,
+		"username":     proxy.Username,
+		"has_password": strings.TrimSpace(proxy.Password) != "",
+	}
+}
+
+func (h *Handler) listProxies(w http.ResponseWriter, _ *http.Request) {
+	proxies := h.Store.Snapshot().Proxies
+	items := make([]map[string]any, 0, len(proxies))
+	for _, proxy := range proxies {
+		proxy = config.NormalizeProxy(proxy)
+		items = append(items, map[string]any{
+			"id":           proxy.ID,
+			"name":         proxy.Name,
+			"type":         proxy.Type,
+			"host":         proxy.Host,
+			"port":         proxy.Port,
+			"username":     proxy.Username,
+			"has_password": strings.TrimSpace(proxy.Password) != "",
+		})
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"items": items, "total": len(items)})
+}
+
+func (h *Handler) addProxy(w http.ResponseWriter, r *http.Request) {
+	var req map[string]any
+	_ = json.NewDecoder(r.Body).Decode(&req)
+	proxy := toProxy(req)
+	err := h.Store.Update(func(c *config.Config) error {
+		c.Proxies = append(c.Proxies, proxy)
+		return validateProxyMutation(c)
+	})
+	if err != nil {
+		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"success": true, "proxy": proxyResponse(proxy)})
+}
+
+func (h *Handler) updateProxy(w http.ResponseWriter, r *http.Request) {
+	proxyID := chi.URLParam(r, "proxyID")
+	if decoded, err := url.PathUnescape(proxyID); err == nil {
+		proxyID = decoded
+	}
+	var req map[string]any
+	_ = json.NewDecoder(r.Body).Decode(&req)
+	proxy := toProxy(req)
+	proxy.ID = strings.TrimSpace(proxyID)
+
+	err := h.Store.Update(func(c *config.Config) error {
+		for i, existing := range c.Proxies {
+			existing = config.NormalizeProxy(existing)
+			if existing.ID != proxy.ID {
+				continue
+			}
+			if proxy.Password == "" {
+				proxy.Password = existing.Password
+			}
+			c.Proxies[i] = proxy
+			return validateProxyMutation(c)
+		}
+		return newRequestError("代理不存在")
+	})
+	if err != nil {
+		if detail, ok := requestErrorDetail(err); ok {
+			writeJSON(w, http.StatusNotFound, map[string]any{"detail": detail})
+			return
+		}
+		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"success": true, "proxy": proxyResponse(proxy)})
+}
+
+func (h *Handler) deleteProxy(w http.ResponseWriter, r *http.Request) {
+	proxyID := chi.URLParam(r, "proxyID")
+	if decoded, err := url.PathUnescape(proxyID); err == nil {
+		proxyID = decoded
+	}
+	err := h.Store.Update(func(c *config.Config) error {
+		idx := -1
+		for i, existing := range c.Proxies {
+			existing = config.NormalizeProxy(existing)
+			if existing.ID == strings.TrimSpace(proxyID) {
+				idx = i
+				break
+			}
+		}
+		if idx < 0 {
+			return newRequestError("代理不存在")
+		}
+		c.Proxies = append(c.Proxies[:idx], c.Proxies[idx+1:]...)
+		for i := range c.Accounts {
+			if strings.TrimSpace(c.Accounts[i].ProxyID) == strings.TrimSpace(proxyID) {
+				c.Accounts[i].ProxyID = ""
+			}
+		}
+		return validateProxyMutation(c)
+	})
+	if err != nil {
+		if detail, ok := requestErrorDetail(err); ok {
+			writeJSON(w, http.StatusNotFound, map[string]any{"detail": detail})
+			return
+		}
+		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
+		return
+	}
+	writeJSON(w, http.StatusOK, map[string]any{"success": true})
+}
+
+func (h *Handler) testProxy(w http.ResponseWriter, r *http.Request) {
+	var req map[string]any
+	_ = json.NewDecoder(r.Body).Decode(&req)
+	proxyID := fieldString(req, "proxy_id")
+
+	var proxy config.Proxy
+	if proxyID != "" {
+		var ok bool
+		proxy, ok = findProxyByID(h.Store.Snapshot(), proxyID)
+		if !ok {
+			writeJSON(w, http.StatusNotFound, map[string]any{"detail": "代理不存在"})
+			return
+		}
+	} else {
+		proxy = toProxy(req)
+	}
+
+	result := proxyConnectivityTester(r.Context(), proxy)
+	writeJSON(w, http.StatusOK, result)
+}
+
+func (h *Handler) updateAccountProxy(w http.ResponseWriter, r *http.Request) {
+	identifier := chi.URLParam(r, "identifier")
+	if decoded, err := url.PathUnescape(identifier); err == nil {
+		identifier = decoded
+	}
+	var req map[string]any
+	_ = json.NewDecoder(r.Body).Decode(&req)
+	proxyID := fieldString(req, "proxy_id")
+
+	err := h.Store.Update(func(c *config.Config) error {
+		if proxyID != "" {
+			if _, ok := findProxyByID(*c, proxyID); !ok {
+				return newRequestError("代理不存在")
+			}
+		}
+		for i, acc := range c.Accounts {
+			if !accountMatchesIdentifier(acc, identifier) {
+				continue
+			}
+			c.Accounts[i].ProxyID = proxyID
+			return validateProxyMutation(c)
+		}
+		return newRequestError("账号不存在")
+	})
+	if err != nil {
+		if detail, ok := requestErrorDetail(err); ok {
+			writeJSON(w, http.StatusBadRequest, map[string]any{"detail": detail})
+			return
+		}
+		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
+		return
+	}
+	h.Pool.Reset()
+	writeJSON(w, http.StatusOK, map[string]any{"success": true, "proxy_id": proxyID})
+}

internal/admin/handler_proxies_test.go

@@ -0,0 +1,227 @@
+package admin
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+
+	"github.com/go-chi/chi/v5"
+
+	"ds2api/internal/account"
+	"ds2api/internal/config"
+)
+
+func newAdminProxyTestHandler(t *testing.T, raw string) *Handler {
+	t.Helper()
+	t.Setenv("DS2API_CONFIG_JSON", raw)
+	store := config.LoadStore()
+	return &Handler{
+		Store: store,
+		Pool:  account.NewPool(store),
+	}
+}
+
+func TestAddProxyPersistsNormalizedProxy(t *testing.T) {
+	h := newAdminProxyTestHandler(t, `{"accounts":[]}`)
+
+	r := chi.NewRouter()
+	r.Post("/admin/proxies", h.addProxy)
+
+	req := httptest.NewRequest(http.MethodPost, "/admin/proxies", bytes.NewBufferString(`{
+		"name":"  HK Exit  ",
+		"type":" SOCKS5H ",
+		"host":" 127.0.0.1 ",
+		"port":1081,
+		"username":" user ",
+		"password":" pass "
+	}`))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	proxies := h.Store.Snapshot().Proxies
+	if len(proxies) != 1 {
+		t.Fatalf("expected 1 proxy, got %d", len(proxies))
+	}
+	if proxies[0].Name != "HK Exit" {
+		t.Fatalf("unexpected proxy name: %#v", proxies[0])
+	}
+	if proxies[0].Type != "socks5h" {
+		t.Fatalf("unexpected proxy type: %#v", proxies[0])
+	}
+	if proxies[0].Username != "user" || proxies[0].Password != "pass" {
+		t.Fatalf("expected trimmed credentials, got %#v", proxies[0])
+	}
+	if proxies[0].ID == "" {
+		t.Fatalf("expected generated proxy id, got %#v", proxies[0])
+	}
+}
+
+func TestAddProxyDoesNotFailOnUnrelatedInvalidRuntimeConfig(t *testing.T) {
+	router := newHTTPAdminHarness(t, `{
+		"keys":["k1"],
+		"runtime":{
+			"account_max_inflight":8,
+			"global_max_inflight":4
+		}
+	}`, &testingDSMock{})
+
+	rec := httptest.NewRecorder()
+	router.ServeHTTP(rec, adminReq(http.MethodPost, "/proxies", []byte(`{
+		"name":"HK Exit",
+		"type":"socks5h",
+		"host":"127.0.0.1",
+		"port":1080
+	}`)))
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected add proxy success despite unrelated runtime issue, got %d body=%s", rec.Code, rec.Body.String())
+	}
+
+	readRec := httptest.NewRecorder()
+	router.ServeHTTP(readRec, adminReq(http.MethodGet, "/config", nil))
+	if readRec.Code != http.StatusOK {
+		t.Fatalf("config read status=%d body=%s", readRec.Code, readRec.Body.String())
+	}
+	var payload map[string]any
+	if err := json.Unmarshal(readRec.Body.Bytes(), &payload); err != nil {
+		t.Fatalf("decode config response: %v", err)
+	}
+	proxies, _ := payload["proxies"].([]any)
+	if len(proxies) != 1 {
+		t.Fatalf("expected proxy to be persisted, got %#v", payload["proxies"])
+	}
+}
+
+func TestDeleteProxyClearsAssignedAccountProxyID(t *testing.T) {
+	h := newAdminProxyTestHandler(t, `{
+		"proxies":[{"id":"proxy-1","name":"Node 1","type":"socks5","host":"127.0.0.1","port":1080}],
+		"accounts":[{"email":"u@example.com","password":"pwd","proxy_id":"proxy-1"}]
+	}`)
+
+	r := chi.NewRouter()
+	r.Delete("/admin/proxies/{proxyID}", h.deleteProxy)
+
+	req := httptest.NewRequest(http.MethodDelete, "/admin/proxies/proxy-1", nil)
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	snap := h.Store.Snapshot()
+	if len(snap.Proxies) != 0 {
+		t.Fatalf("expected proxy removed, got %#v", snap.Proxies)
+	}
+	if len(snap.Accounts) != 1 {
+		t.Fatalf("expected account kept, got %#v", snap.Accounts)
+	}
+	if snap.Accounts[0].ProxyID != "" {
+		t.Fatalf("expected proxy assignment cleared, got %#v", snap.Accounts[0])
+	}
+}
+
+func TestUpdateProxyResponseDoesNotExposeStoredPassword(t *testing.T) {
+	h := newAdminProxyTestHandler(t, `{
+		"proxies":[{"id":"proxy-1","name":"Node 1","type":"socks5h","host":"127.0.0.1","port":1080,"username":"u","password":"secret"}]
+	}`)
+
+	r := chi.NewRouter()
+	r.Put("/admin/proxies/{proxyID}", h.updateProxy)
+
+	req := httptest.NewRequest(http.MethodPut, "/admin/proxies/proxy-1", bytes.NewBufferString(`{
+		"name":"Node 1",
+		"type":"socks5h",
+		"host":"127.0.0.2",
+		"port":1081,
+		"username":"u2"
+	}`))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	var payload map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &payload); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	proxy, _ := payload["proxy"].(map[string]any)
+	if _, exists := proxy["password"]; exists {
+		t.Fatalf("response should not expose password, got %#v", proxy)
+	}
+	if hasPassword, _ := proxy["has_password"].(bool); !hasPassword {
+		t.Fatalf("expected has_password=true, got %#v", proxy)
+	}
+}
+
+func TestUpdateAccountProxyAssignsProxyID(t *testing.T) {
+	h := newAdminProxyTestHandler(t, `{
+		"proxies":[{"id":"proxy-1","name":"Node 1","type":"socks5h","host":"127.0.0.1","port":1080}],
+		"accounts":[{"email":"u@example.com","password":"pwd"}]
+	}`)
+
+	r := chi.NewRouter()
+	r.Put("/admin/accounts/{identifier}/proxy", h.updateAccountProxy)
+
+	req := httptest.NewRequest(http.MethodPut, "/admin/accounts/u@example.com/proxy", bytes.NewBufferString(`{"proxy_id":"proxy-1"}`))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	acc, ok := h.Store.FindAccount("u@example.com")
+	if !ok {
+		t.Fatal("expected account")
+	}
+	if acc.ProxyID != "proxy-1" {
+		t.Fatalf("expected proxy assigned, got %#v", acc)
+	}
+}
+
+func TestTestProxyUsesStoredProxy(t *testing.T) {
+	h := newAdminProxyTestHandler(t, `{
+		"proxies":[{"id":"proxy-1","name":"Node 1","type":"socks5h","host":"127.0.0.1","port":1080}]
+	}`)
+
+	original := proxyConnectivityTester
+	defer func() { proxyConnectivityTester = original }()
+
+	var got config.Proxy
+	proxyConnectivityTester = func(_ context.Context, proxy config.Proxy) map[string]any {
+		got = proxy
+		return map[string]any{
+			"success":       true,
+			"proxy_id":      proxy.ID,
+			"proxy_type":    proxy.Type,
+			"response_time": 12,
+		}
+	}
+
+	r := chi.NewRouter()
+	r.Post("/admin/proxies/test", h.testProxy)
+
+	req := httptest.NewRequest(http.MethodPost, "/admin/proxies/test", bytes.NewBufferString(`{"proxy_id":"proxy-1"}`))
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("unexpected status: %d body=%s", rec.Code, rec.Body.String())
+	}
+	if got.ID != "proxy-1" || got.Type != "socks5h" {
+		t.Fatalf("expected stored proxy passed to tester, got %#v", got)
+	}
+	var payload map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &payload); err != nil {
+		t.Fatalf("decode response: %v", err)
+	}
+	if ok, _ := payload["success"].(bool); !ok {
+		t.Fatalf("expected success payload, got %#v", payload)
+	}
+}

internal/admin/helpers.go

@@ -65,6 +65,7 @@ func toAccount(m map[string]any) config.Account {
 		Email:    email,
 		Mobile:   mobile,
 		Password: fieldString(m, "password"),
+		ProxyID:  fieldString(m, "proxy_id"),
 	}
 }
 
@@ -100,9 +101,36 @@ func accountMatchesIdentifier(acc config.Account, identifier string) bool {
 func normalizeAccountForStorage(acc config.Account) config.Account {
 	acc.Email = strings.TrimSpace(acc.Email)
 	acc.Mobile = config.NormalizeMobileForStorage(acc.Mobile)
+	acc.ProxyID = strings.TrimSpace(acc.ProxyID)
 	return acc
 }
 
+func toProxy(m map[string]any) config.Proxy {
+	return config.NormalizeProxy(config.Proxy{
+		ID:       fieldString(m, "id"),
+		Name:     fieldString(m, "name"),
+		Type:     fieldString(m, "type"),
+		Host:     fieldString(m, "host"),
+		Port:     intFrom(m["port"]),
+		Username: fieldString(m, "username"),
+		Password: fieldString(m, "password"),
+	})
+}
+
+func findProxyByID(c config.Config, proxyID string) (config.Proxy, bool) {
+	id := strings.TrimSpace(proxyID)
+	if id == "" {
+		return config.Proxy{}, false
+	}
+	for _, proxy := range c.Proxies {
+		proxy = config.NormalizeProxy(proxy)
+		if proxy.ID == id {
+			return proxy, true
+		}
+	}
+	return config.Proxy{}, false
+}
+
 func accountDedupeKey(acc config.Account) string {
 	if email := strings.TrimSpace(acc.Email); email != "" {
 		return "email:" + email

internal/compat/go_compat_test.go

@@ -37,7 +37,6 @@ func TestGoCompatSSEFixtures(t *testing.T) {
 			Finished      bool             `json:"finished"`
 			NewType       string           `json:"new_type"`
 			ContentFilter bool             `json:"content_filter"`
-			OutputTokens  int              `json:"output_tokens"`
 			ErrorMessage  string           `json:"error_message"`
 		}
 		mustLoadJSON(t, expectedPath, &expected)
@@ -58,11 +57,10 @@ func TestGoCompatSSEFixtures(t *testing.T) {
 			res.Stop != expected.Finished ||
 			res.NextType != expected.NewType ||
 			res.ContentFilter != expected.ContentFilter ||
-			res.OutputTokens != expected.OutputTokens ||
 			res.ErrorMessage != expected.ErrorMessage {
-			t.Fatalf("fixture %s mismatch:\n got parts=%#v finished=%v newType=%q contentFilter=%v outputTokens=%d errorMessage=%q\nwant parts=%#v finished=%v newType=%q contentFilter=%v outputTokens=%d errorMessage=%q",
-				name, gotParts, res.Stop, res.NextType, res.ContentFilter, res.OutputTokens, res.ErrorMessage,
-				expected.Parts, expected.Finished, expected.NewType, expected.ContentFilter, expected.OutputTokens, expected.ErrorMessage)
+			t.Fatalf("fixture %s mismatch:\n got parts=%#v finished=%v newType=%q contentFilter=%v errorMessage=%q\nwant parts=%#v finished=%v newType=%q contentFilter=%v errorMessage=%q",
+				name, gotParts, res.Stop, res.NextType, res.ContentFilter, res.ErrorMessage,
+				expected.Parts, expected.Finished, expected.NewType, expected.ContentFilter, expected.ErrorMessage)
 		}
 	}
 }

internal/config/codec.go

@@ -20,6 +20,9 @@ func (c Config) MarshalJSON() ([]byte, error) {
 	if len(c.Accounts) > 0 {
 		m["accounts"] = c.Accounts
 	}
+	if len(c.Proxies) > 0 {
+		m["proxies"] = c.Proxies
+	}
 	if len(c.ClaudeMapping) > 0 {
 		m["claude_mapping"] = c.ClaudeMapping
 	}
@@ -70,6 +73,10 @@ func (c *Config) UnmarshalJSON(b []byte) error {
 			if err := json.Unmarshal(v, &c.Accounts); err != nil {
 				return fmt.Errorf("invalid field %q: %w", k, err)
 			}
+		case "proxies":
+			if err := json.Unmarshal(v, &c.Proxies); err != nil {
+				return fmt.Errorf("invalid field %q: %w", k, err)
+			}
 		case "claude_mapping":
 			if err := json.Unmarshal(v, &c.ClaudeMapping); err != nil {
 				return fmt.Errorf("invalid field %q: %w", k, err)
@@ -130,6 +137,7 @@ func (c Config) Clone() Config {
 	clone := Config{
 		Keys:           slices.Clone(c.Keys),
 		Accounts:       slices.Clone(c.Accounts),
+		Proxies:        slices.Clone(c.Proxies),
 		ClaudeMapping:  cloneStringMap(c.ClaudeMapping),
 		ClaudeModelMap: cloneStringMap(c.ClaudeModelMap),
 		ModelAliases:   cloneStringMap(c.ModelAliases),

internal/config/config.go

@@ -1,8 +1,16 @@
 package config
 
+import (
+	"crypto/sha1"
+	"encoding/hex"
+	"fmt"
+	"strings"
+)
+
 type Config struct {
 	Keys             []string          `json:"keys,omitempty"`
 	Accounts         []Account         `json:"accounts,omitempty"`
+	Proxies          []Proxy           `json:"proxies,omitempty"`
 	ClaudeMapping    map[string]string `json:"claude_mapping,omitempty"`
 	ClaudeModelMap   map[string]string `json:"claude_model_mapping,omitempty"`
 	ModelAliases     map[string]string `json:"model_aliases,omitempty"`
@@ -22,6 +30,38 @@ type Account struct {
 	Mobile   string `json:"mobile,omitempty"`
 	Password string `json:"password,omitempty"`
 	Token    string `json:"token,omitempty"`
+	ProxyID  string `json:"proxy_id,omitempty"`
+}
+
+type Proxy struct {
+	ID       string `json:"id,omitempty"`
+	Name     string `json:"name,omitempty"`
+	Type     string `json:"type,omitempty"`
+	Host     string `json:"host,omitempty"`
+	Port     int    `json:"port,omitempty"`
+	Username string `json:"username,omitempty"`
+	Password string `json:"password,omitempty"`
+}
+
+func NormalizeProxy(p Proxy) Proxy {
+	p.ID = strings.TrimSpace(p.ID)
+	p.Name = strings.TrimSpace(p.Name)
+	p.Type = strings.ToLower(strings.TrimSpace(p.Type))
+	p.Host = strings.TrimSpace(p.Host)
+	p.Username = strings.TrimSpace(p.Username)
+	p.Password = strings.TrimSpace(p.Password)
+	if p.ID == "" {
+		p.ID = StableProxyID(p)
+	}
+	if p.Name == "" && p.Host != "" && p.Port > 0 {
+		p.Name = fmt.Sprintf("%s:%d", p.Host, p.Port)
+	}
+	return p
+}
+
+func StableProxyID(p Proxy) string {
+	sum := sha1.Sum([]byte(strings.ToLower(strings.TrimSpace(p.Type)) + "|" + strings.ToLower(strings.TrimSpace(p.Host)) + "|" + fmt.Sprintf("%d", p.Port) + "|" + strings.TrimSpace(p.Username)))
+	return "proxy_" + hex.EncodeToString(sum[:6])
 }
 
 func (c *Config) ClearAccountTokens() {

internal/config/config_edge_test.go

@@ -49,6 +49,51 @@ func TestGetModelConfigDeepSeekReasonerSearch(t *testing.T) {
 	}
 }
 
+func TestGetModelConfigDeepSeekExpertChat(t *testing.T) {
+	thinking, search, ok := GetModelConfig("deepseek-expert-chat")
+	if !ok {
+		t.Fatal("expected ok for deepseek-expert-chat")
+	}
+	if thinking || search {
+		t.Fatalf("expected no thinking/search for deepseek-expert-chat, got thinking=%v search=%v", thinking, search)
+	}
+}
+
+func TestGetModelConfigDeepSeekExpertReasonerSearch(t *testing.T) {
+	thinking, search, ok := GetModelConfig("deepseek-expert-reasoner-search")
+	if !ok {
+		t.Fatal("expected ok for deepseek-expert-reasoner-search")
+	}
+	if !thinking || !search {
+		t.Fatalf("expected both true, got thinking=%v search=%v", thinking, search)
+	}
+}
+
+func TestGetModelConfigDeepSeekVisionReasonerSearch(t *testing.T) {
+	thinking, search, ok := GetModelConfig("deepseek-vision-reasoner-search")
+	if !ok {
+		t.Fatal("expected ok for deepseek-vision-reasoner-search")
+	}
+	if !thinking || !search {
+		t.Fatalf("expected both true, got thinking=%v search=%v", thinking, search)
+	}
+}
+
+func TestGetModelTypeDefaultExpertAndVision(t *testing.T) {
+	defaultType, ok := GetModelType("deepseek-chat")
+	if !ok || defaultType != "default" {
+		t.Fatalf("expected default model_type, got ok=%v model_type=%q", ok, defaultType)
+	}
+	expertType, ok := GetModelType("deepseek-expert-chat")
+	if !ok || expertType != "expert" {
+		t.Fatalf("expected expert model_type, got ok=%v model_type=%q", ok, expertType)
+	}
+	visionType, ok := GetModelType("deepseek-vision-chat")
+	if !ok || visionType != "vision" {
+		t.Fatalf("expected vision model_type, got ok=%v model_type=%q", ok, visionType)
+	}
+}
+
 func TestGetModelConfigCaseInsensitive(t *testing.T) {
 	thinking, search, ok := GetModelConfig("DeepSeek-Chat")
 	if !ok {
@@ -551,6 +596,30 @@ func TestOpenAIModelsResponse(t *testing.T) {
 	if len(data) == 0 {
 		t.Fatal("expected non-empty models list")
 	}
+	expected := map[string]bool{
+		"deepseek-chat":                   false,
+		"deepseek-reasoner":               false,
+		"deepseek-chat-search":            false,
+		"deepseek-reasoner-search":        false,
+		"deepseek-expert-chat":            false,
+		"deepseek-expert-reasoner":        false,
+		"deepseek-expert-chat-search":     false,
+		"deepseek-expert-reasoner-search": false,
+		"deepseek-vision-chat":            false,
+		"deepseek-vision-reasoner":        false,
+		"deepseek-vision-chat-search":     false,
+		"deepseek-vision-reasoner-search": false,
+	}
+	for _, model := range data {
+		if _, ok := expected[model.ID]; ok {
+			expected[model.ID] = true
+		}
+	}
+	for id, seen := range expected {
+		if !seen {
+			t.Fatalf("expected OpenAI model list to include %s", id)
+		}
+	}
 }
 
 func TestClaudeModelsResponse(t *testing.T) {

internal/config/config_test.go

@@ -32,6 +32,47 @@ func TestLoadStoreClearsTokensFromConfigInput(t *testing.T) {
 	}
 }
 
+func TestLoadStorePreservesProxiesAndAccountProxyAssignment(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"proxies":[
+			{
+				"id":"proxy-sh-1",
+				"name":"Shanghai Exit",
+				"type":"socks5h",
+				"host":"127.0.0.1",
+				"port":1080,
+				"username":"demo",
+				"password":"secret"
+			}
+		],
+		"accounts":[
+			{
+				"email":"u@example.com",
+				"password":"p",
+				"proxy_id":"proxy-sh-1"
+			}
+		]
+	}`)
+
+	store := LoadStore()
+	snap := store.Snapshot()
+	if len(snap.Proxies) != 1 {
+		t.Fatalf("expected 1 proxy, got %d", len(snap.Proxies))
+	}
+	if snap.Proxies[0].ID != "proxy-sh-1" {
+		t.Fatalf("unexpected proxy id: %#v", snap.Proxies[0])
+	}
+	if snap.Proxies[0].Type != "socks5h" {
+		t.Fatalf("unexpected proxy type: %#v", snap.Proxies[0])
+	}
+	if len(snap.Accounts) != 1 {
+		t.Fatalf("expected 1 account, got %d", len(snap.Accounts))
+	}
+	if snap.Accounts[0].ProxyID != "proxy-sh-1" {
+		t.Fatalf("expected account proxy assignment preserved, got %#v", snap.Accounts[0])
+	}
+}
+
 func TestLoadStoreDropsLegacyTokenOnlyAccounts(t *testing.T) {
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"accounts":[

internal/config/model_alias_test.go

@@ -2,6 +2,10 @@ package config
 
 import "testing"
 
+type mockModelAliasReader map[string]string
+
+func (m mockModelAliasReader) ModelAliases() map[string]string { return m }
+
 func TestResolveModelDirectDeepSeek(t *testing.T) {
 	got, ok := ResolveModel(nil, "deepseek-chat")
 	if !ok || got != "deepseek-chat" {
@@ -30,6 +34,31 @@ func TestResolveModelUnknown(t *testing.T) {
 	}
 }
 
+func TestResolveModelDirectDeepSeekExpert(t *testing.T) {
+	got, ok := ResolveModel(nil, "deepseek-expert-chat")
+	if !ok || got != "deepseek-expert-chat" {
+		t.Fatalf("expected deepseek-expert-chat, got ok=%v model=%q", ok, got)
+	}
+}
+
+func TestResolveModelCustomAliasToExpert(t *testing.T) {
+	got, ok := ResolveModel(mockModelAliasReader{
+		"my-expert-model": "deepseek-expert-reasoner-search",
+	}, "my-expert-model")
+	if !ok || got != "deepseek-expert-reasoner-search" {
+		t.Fatalf("expected alias -> deepseek-expert-reasoner-search, got ok=%v model=%q", ok, got)
+	}
+}
+
+func TestResolveModelCustomAliasToVision(t *testing.T) {
+	got, ok := ResolveModel(mockModelAliasReader{
+		"my-vision-model": "deepseek-vision-chat-search",
+	}, "my-vision-model")
+	if !ok || got != "deepseek-vision-chat-search" {
+		t.Fatalf("expected alias -> deepseek-vision-chat-search, got ok=%v model=%q", ok, got)
+	}
+}
+
 func TestClaudeModelsResponsePaginationFields(t *testing.T) {
 	resp := ClaudeModelsResponse()
 	if _, ok := resp["first_id"]; !ok {

internal/config/models.go

@@ -19,6 +19,14 @@ var DeepSeekModels = []ModelInfo{
 	{ID: "deepseek-reasoner", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 	{ID: "deepseek-chat-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 	{ID: "deepseek-reasoner-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-expert-chat", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-expert-reasoner", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-expert-chat-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-expert-reasoner-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-vision-chat", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-vision-reasoner", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-vision-chat-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
+	{ID: "deepseek-vision-reasoner-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 }
 
 var ClaudeModels = []ModelInfo{
@@ -72,11 +80,40 @@ func GetModelConfig(model string) (thinking bool, search bool, ok bool) {
 		return false, true, true
 	case "deepseek-reasoner-search":
 		return true, true, true
+	case "deepseek-expert-chat":
+		return false, false, true
+	case "deepseek-expert-reasoner":
+		return true, false, true
+	case "deepseek-expert-chat-search":
+		return false, true, true
+	case "deepseek-expert-reasoner-search":
+		return true, true, true
+	case "deepseek-vision-chat":
+		return false, false, true
+	case "deepseek-vision-reasoner":
+		return true, false, true
+	case "deepseek-vision-chat-search":
+		return false, true, true
+	case "deepseek-vision-reasoner-search":
+		return true, true, true
 	default:
 		return false, false, false
 	}
 }
 
+func GetModelType(model string) (modelType string, ok bool) {
+	switch lower(model) {
+	case "deepseek-chat", "deepseek-reasoner", "deepseek-chat-search", "deepseek-reasoner-search":
+		return "default", true
+	case "deepseek-expert-chat", "deepseek-expert-reasoner", "deepseek-expert-chat-search", "deepseek-expert-reasoner-search":
+		return "expert", true
+	case "deepseek-vision-chat", "deepseek-vision-reasoner", "deepseek-vision-chat-search", "deepseek-vision-reasoner-search":
+		return "vision", true
+	default:
+		return "", false
+	}
+}
+
 func IsSupportedDeepSeekModel(model string) bool {
 	_, _, ok := GetModelConfig(model)
 	return ok

internal/config/paths.go

@@ -33,10 +33,6 @@ func ConfigPath() string {
 	return ResolvePath("DS2API_CONFIG_PATH", "config.json")
 }
 
-func WASMPath() string {
-	return ResolvePath("DS2API_WASM_PATH", "sha3_wasm_bg.7b9ca65ddd.wasm")
-}
-
 func RawStreamSampleRoot() string {
 	return ResolvePath("DS2API_RAW_STREAM_SAMPLE_ROOT", "tests/raw_stream_samples")
 }

internal/config/validation.go

@@ -6,6 +6,9 @@ import (
 )
 
 func ValidateConfig(c Config) error {
+	if err := ValidateProxyConfig(c.Proxies); err != nil {
+		return err
+	}
 	if err := ValidateAdminConfig(c.Admin); err != nil {
 		return err
 	}
@@ -21,6 +24,55 @@ func ValidateConfig(c Config) error {
 	if err := ValidateAutoDeleteConfig(c.AutoDelete); err != nil {
 		return err
 	}
+	if err := ValidateAccountProxyReferences(c.Accounts, c.Proxies); err != nil {
+		return err
+	}
+	return nil
+}
+
+func ValidateProxyConfig(proxies []Proxy) error {
+	seen := make(map[string]struct{}, len(proxies))
+	for _, proxy := range proxies {
+		proxy = NormalizeProxy(proxy)
+		if err := ValidateTrimmedString("proxies.id", proxy.ID, true); err != nil {
+			return err
+		}
+		switch proxy.Type {
+		case "socks5", "socks5h":
+		default:
+			return fmt.Errorf("proxies.type must be one of socks5, socks5h")
+		}
+		if err := ValidateTrimmedString("proxies.host", proxy.Host, true); err != nil {
+			return err
+		}
+		if err := ValidateIntRange("proxies.port", proxy.Port, 1, 65535, true); err != nil {
+			return err
+		}
+		if _, ok := seen[proxy.ID]; ok {
+			return fmt.Errorf("duplicate proxy id: %s", proxy.ID)
+		}
+		seen[proxy.ID] = struct{}{}
+	}
+	return nil
+}
+
+func ValidateAccountProxyReferences(accounts []Account, proxies []Proxy) error {
+	if len(accounts) == 0 {
+		return nil
+	}
+	ids := make(map[string]struct{}, len(proxies))
+	for _, proxy := range proxies {
+		ids[NormalizeProxy(proxy).ID] = struct{}{}
+	}
+	for _, acc := range accounts {
+		proxyID := strings.TrimSpace(acc.ProxyID)
+		if proxyID == "" {
+			continue
+		}
+		if _, ok := ids[proxyID]; !ok {
+			return fmt.Errorf("account proxy_id references unknown proxy: %s", proxyID)
+		}
+	}
 	return nil
 }
 

internal/deepseek/client_auth.go

@@ -13,6 +13,7 @@ import (
 )
 
 func (c *Client) Login(ctx context.Context, acc config.Account) (string, error) {
+	clients := c.requestClientsForAccount(acc)
 	payload := map[string]any{
 		"password":  strings.TrimSpace(acc.Password),
 		"device_id": "deepseek_to_api",
@@ -27,7 +28,7 @@ func (c *Client) Login(ctx context.Context, acc config.Account) (string, error)
 	} else {
 		return "", errors.New("missing email/mobile")
 	}
-	resp, err := c.postJSON(ctx, c.regular, DeepSeekLoginURL, BaseHeaders, payload)
+	resp, err := c.postJSON(ctx, clients.regular, clients.fallback, DeepSeekLoginURL, BaseHeaders, payload)
 	if err != nil {
 		return "", err
 	}
@@ -52,11 +53,12 @@ func (c *Client) CreateSession(ctx context.Context, a *auth.RequestAuth, maxAtte
 	if maxAttempts <= 0 {
 		maxAttempts = c.maxRetries
 	}
+	clients := c.requestClientsForAuth(ctx, a)
 	attempts := 0
 	refreshed := false
 	for attempts < maxAttempts {
 		headers := c.authHeaders(a.DeepSeekToken)
-		resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekCreateSessionURL, headers, map[string]any{"agent": "chat"})
+		resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekCreateSessionURL, headers, map[string]any{"agent": "chat"})
 		if err != nil {
 			config.Logger.Warn("[create_session] request error", "error", err, "account", a.AccountID)
 			attempts++
@@ -64,9 +66,7 @@ func (c *Client) CreateSession(ctx context.Context, a *auth.RequestAuth, maxAtte
 		}
 		code, bizCode, msg, bizMsg := extractResponseStatus(resp)
 		if status == http.StatusOK && code == 0 && bizCode == 0 {
-			data, _ := resp["data"].(map[string]any)
-			bizData, _ := data["biz_data"].(map[string]any)
-			sessionID, _ := bizData["id"].(string)
+			sessionID := extractCreateSessionID(resp)
 			if sessionID != "" {
 				return sessionID, nil
 			}
@@ -91,16 +91,25 @@ func (c *Client) CreateSession(ctx context.Context, a *auth.RequestAuth, maxAtte
 }
 
 func (c *Client) GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts int) (string, error) {
+	return c.GetPowForTarget(ctx, a, DeepSeekCompletionTargetPath, maxAttempts)
+}
+
+func (c *Client) GetPowForTarget(ctx context.Context, a *auth.RequestAuth, targetPath string, maxAttempts int) (string, error) {
 	if maxAttempts <= 0 {
 		maxAttempts = c.maxRetries
 	}
+	targetPath = strings.TrimSpace(targetPath)
+	if targetPath == "" {
+		targetPath = DeepSeekCompletionTargetPath
+	}
+	clients := c.requestClientsForAuth(ctx, a)
 	attempts := 0
 	refreshed := false
 	for attempts < maxAttempts {
 		headers := c.authHeaders(a.DeepSeekToken)
-		resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekCreatePowURL, headers, map[string]any{"target_path": "/api/v0/chat/completion"})
+		resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekCreatePowURL, headers, map[string]any{"target_path": targetPath})
 		if err != nil {
-			config.Logger.Warn("[get_pow] request error", "error", err, "account", a.AccountID)
+			config.Logger.Warn("[get_pow] request error", "error", err, "account", a.AccountID, "target_path", targetPath)
 			attempts++
 			continue
 		}
@@ -109,14 +118,14 @@ func (c *Client) GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts in
 			data, _ := resp["data"].(map[string]any)
 			bizData, _ := data["biz_data"].(map[string]any)
 			challenge, _ := bizData["challenge"].(map[string]any)
-			answer, err := c.powSolver.Compute(ctx, challenge)
+			answer, err := ComputePow(ctx, challenge)
 			if err != nil {
 				attempts++
 				continue
 			}
 			return BuildPowHeader(challenge, answer)
 		}
-		config.Logger.Warn("[get_pow] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "use_config_token", a.UseConfigToken, "account", a.AccountID)
+		config.Logger.Warn("[get_pow] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "use_config_token", a.UseConfigToken, "account", a.AccountID, "target_path", targetPath)
 		if a.UseConfigToken {
 			if !refreshed && shouldAttemptRefresh(status, code, bizCode, msg, bizMsg) {
 				if c.Auth.RefreshToken(ctx, a) {
@@ -201,6 +210,22 @@ func isAuthIndicativeBizFailure(msg string, bizMsg string) bool {
 	return false
 }
 
+// DeepSeek has returned create-session ids in both biz_data.id and
+// biz_data.chat_session.id across observed response variants; accept either.
+func extractCreateSessionID(resp map[string]any) string {
+	data, _ := resp["data"].(map[string]any)
+	bizData, _ := data["biz_data"].(map[string]any)
+	if sessionID, _ := bizData["id"].(string); strings.TrimSpace(sessionID) != "" {
+		return strings.TrimSpace(sessionID)
+	}
+	if chatSession, ok := bizData["chat_session"].(map[string]any); ok {
+		if sessionID, _ := chatSession["id"].(string); strings.TrimSpace(sessionID) != "" {
+			return strings.TrimSpace(sessionID)
+		}
+	}
+	return ""
+}
+
 func extractResponseStatus(resp map[string]any) (code int, bizCode int, msg string, bizMsg string) {
 	code = intFrom(resp["code"])
 	msg, _ = resp["msg"].(string)

internal/deepseek/client_auth_test.go

@@ -0,0 +1,34 @@
+package deepseek
+
+import "testing"
+
+func TestExtractCreateSessionIDSupportsLegacyShape(t *testing.T) {
+	resp := map[string]any{
+		"data": map[string]any{
+			"biz_data": map[string]any{
+				"id": "legacy-session-id",
+			},
+		},
+	}
+
+	if got := extractCreateSessionID(resp); got != "legacy-session-id" {
+		t.Fatalf("expected legacy session id, got %q", got)
+	}
+}
+
+func TestExtractCreateSessionIDSupportsNestedChatSessionShape(t *testing.T) {
+	resp := map[string]any{
+		"data": map[string]any{
+			"biz_data": map[string]any{
+				"chat_session": map[string]any{
+					"id":         "nested-session-id",
+					"model_type": "default",
+				},
+			},
+		},
+	}
+
+	if got := extractCreateSessionID(resp); got != "nested-session-id" {
+		t.Fatalf("expected nested session id, got %q", got)
+	}
+}

internal/deepseek/client_completion.go

@@ -10,18 +10,20 @@ import (
 
 	"ds2api/internal/auth"
 	"ds2api/internal/config"
+	trans "ds2api/internal/deepseek/transport"
 )
 
 func (c *Client) CallCompletion(ctx context.Context, a *auth.RequestAuth, payload map[string]any, powResp string, maxAttempts int) (*http.Response, error) {
 	if maxAttempts <= 0 {
 		maxAttempts = c.maxRetries
 	}
+	clients := c.requestClientsForAuth(ctx, a)
 	headers := c.authHeaders(a.DeepSeekToken)
 	headers["x-ds-pow-response"] = powResp
 	captureSession := c.capture.Start("deepseek_completion", DeepSeekCompletionURL, a.AccountID, payload)
 	attempts := 0
 	for attempts < maxAttempts {
-		resp, err := c.streamPost(ctx, DeepSeekCompletionURL, headers, payload)
+		resp, err := c.streamPost(ctx, clients.stream, DeepSeekCompletionURL, headers, payload)
 		if err != nil {
 			attempts++
 			time.Sleep(time.Second)
@@ -44,11 +46,13 @@ func (c *Client) CallCompletion(ctx context.Context, a *auth.RequestAuth, payloa
 	return nil, errors.New("completion failed")
 }
 
-func (c *Client) streamPost(ctx context.Context, url string, headers map[string]string, payload any) (*http.Response, error) {
+func (c *Client) streamPost(ctx context.Context, doer trans.Doer, url string, headers map[string]string, payload any) (*http.Response, error) {
 	b, err := json.Marshal(payload)
 	if err != nil {
 		return nil, err
 	}
+	headers = c.jsonHeaders(headers)
+	clients := c.requestClientsFromContext(ctx)
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(b))
 	if err != nil {
 		return nil, err
@@ -56,7 +60,7 @@ func (c *Client) streamPost(ctx context.Context, url string, headers map[string]
 	for k, v := range headers {
 		req.Header.Set(k, v)
 	}
-	resp, err := c.stream.Do(req)
+	resp, err := doer.Do(req)
 	if err != nil {
 		config.Logger.Warn("[deepseek] fingerprint stream request failed, fallback to std transport", "url", url, "error", err)
 		req2, reqErr := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(b))
@@ -66,7 +70,7 @@ func (c *Client) streamPost(ctx context.Context, url string, headers map[string]
 		for k, v := range headers {
 			req2.Header.Set(k, v)
 		}
-		return c.fallbackS.Do(req2)
+		return clients.fallbackS.Do(req2)
 	}
 	return resp, nil
 }

internal/deepseek/client_continue.go

@@ -51,6 +51,7 @@ func (c *Client) callContinue(ctx context.Context, a *auth.RequestAuth, sessionI
 	if strings.TrimSpace(sessionID) == "" || responseMessageID <= 0 {
 		return nil, errors.New("missing continue identifiers")
 	}
+	clients := c.requestClientsForAuth(ctx, a)
 	headers := c.authHeaders(a.DeepSeekToken)
 	headers["x-ds-pow-response"] = powResp
 	payload := map[string]any{
@@ -60,7 +61,7 @@ func (c *Client) callContinue(ctx context.Context, a *auth.RequestAuth, sessionI
 	}
 	config.Logger.Info("[auto_continue] calling continue", "session_id", sessionID, "message_id", responseMessageID)
 	captureSession := c.capture.Start("deepseek_continue", DeepSeekContinueURL, a.AccountID, payload)
-	resp, err := c.streamPost(ctx, DeepSeekContinueURL, headers, payload)
+	resp, err := c.streamPost(ctx, clients.stream, DeepSeekContinueURL, headers, payload)
 	if err != nil {
 		return nil, err
 	}

internal/deepseek/client_core.go

@@ -3,6 +3,7 @@ package deepseek
 import (
 	"context"
 	"net/http"
+	"sync"
 	"time"
 
 	"ds2api/internal/auth"
@@ -23,24 +24,27 @@ type Client struct {
 	stream     trans.Doer
 	fallback   *http.Client
 	fallbackS  *http.Client
-	powSolver  *PowSolver
 	maxRetries int
+
+	proxyClientsMu sync.RWMutex
+	proxyClients   map[string]requestClients
 }
 
 func NewClient(store *config.Store, resolver *auth.Resolver) *Client {
 	return &Client{
-		Store:      store,
-		Auth:       resolver,
-		capture:    devcapture.Global(),
-		regular:    trans.New(60 * time.Second),
-		stream:     trans.New(0),
-		fallback:   &http.Client{Timeout: 60 * time.Second},
-		fallbackS:  &http.Client{Timeout: 0},
-		powSolver:  NewPowSolver(config.WASMPath()),
-		maxRetries: 3,
+		Store:        store,
+		Auth:         resolver,
+		capture:      devcapture.Global(),
+		regular:      trans.New(60 * time.Second),
+		stream:       trans.New(0),
+		fallback:     &http.Client{Timeout: 60 * time.Second},
+		fallbackS:    &http.Client{Timeout: 0},
+		maxRetries:   3,
+		proxyClients: map[string]requestClients{},
 	}
 }
 
-func (c *Client) PreloadPow(ctx context.Context) error {
-	return c.powSolver.init(ctx)
+// PreloadPow 保留兼容接口，纯 Go 实现无需预加载。
+func (c *Client) PreloadPow(_ context.Context) error {
+	return nil
 }

internal/deepseek/client_file_status.go

@@ -0,0 +1,188 @@
+package deepseek
+
+import (
+	"context"
+	"errors"
+	"fmt"
+	"net/http"
+	"net/url"
+	"strings"
+	"time"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+)
+
+const (
+	fileReadyPollAttempts = 60
+	fileReadyPollInterval = time.Second
+	fileReadyPollTimeout  = 65 * time.Second
+)
+
+var fileReadySleep = time.Sleep
+
+func (c *Client) waitForUploadedFile(ctx context.Context, a *auth.RequestAuth, result *UploadFileResult) error {
+	if result == nil || strings.TrimSpace(result.ID) == "" {
+		return nil
+	}
+	if isReadyUploadFileStatus(result.Status) {
+		return nil
+	}
+
+	pollCtx, cancel := context.WithTimeout(ctx, fileReadyPollTimeout)
+	defer cancel()
+
+	var lastErr error
+	for attempt := 0; attempt < fileReadyPollAttempts; attempt++ {
+		if err := pollCtx.Err(); err != nil {
+			if lastErr != nil {
+				return fmt.Errorf("waiting for file %s to become ready: %w", result.ID, lastErr)
+			}
+			return fmt.Errorf("waiting for file %s to become ready: %w", result.ID, err)
+		}
+
+		fetched, err := c.fetchUploadedFile(pollCtx, a, result.ID)
+		if err == nil && fetched != nil {
+			mergeUploadFileResults(result, fetched)
+			if isReadyUploadFileStatus(result.Status) {
+				return nil
+			}
+			lastErr = fmt.Errorf("status=%s", strings.TrimSpace(result.Status))
+		} else if err != nil {
+			lastErr = err
+			config.Logger.Debug("[upload_file] waiting for file readiness", "file_id", result.ID, "attempt", attempt+1, "error", err)
+		}
+
+		if attempt < fileReadyPollAttempts-1 {
+			fileReadySleep(fileReadyPollInterval)
+		}
+	}
+
+	if lastErr == nil {
+		lastErr = fmt.Errorf("status=%s", strings.TrimSpace(result.Status))
+	}
+	return fmt.Errorf("file %s did not become ready: %w", result.ID, lastErr)
+}
+
+func (c *Client) fetchUploadedFile(ctx context.Context, a *auth.RequestAuth, fileID string) (*UploadFileResult, error) {
+	fileID = strings.TrimSpace(fileID)
+	if fileID == "" {
+		return nil, errors.New("file id is required")
+	}
+	clients := c.requestClientsForAuth(ctx, a)
+	reqURL := DeepSeekFetchFilesURL + "?file_ids=" + url.QueryEscape(fileID)
+	headers := c.authHeaders(a.DeepSeekToken)
+
+	resp, status, err := c.getJSONWithStatus(ctx, clients.regular, reqURL, headers)
+	if err != nil {
+		return nil, err
+	}
+
+	code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+	if status != http.StatusOK || code != 0 || bizCode != 0 {
+		if strings.TrimSpace(bizMsg) != "" {
+			msg = bizMsg
+		}
+		if msg == "" {
+			msg = http.StatusText(status)
+		}
+		return nil, fmt.Errorf("request failed: status=%d, code=%d, msg=%s", status, code, msg)
+	}
+
+	result := extractFetchedUploadFileResult(resp, fileID)
+	if result == nil || strings.TrimSpace(result.ID) == "" {
+		return nil, errors.New("fetch files succeeded without matching file data")
+	}
+	result.Raw = resp
+	return result, nil
+}
+
+func extractFetchedUploadFileResult(resp map[string]any, targetID string) *UploadFileResult {
+	targetID = strings.TrimSpace(targetID)
+	if resp == nil || targetID == "" {
+		return nil
+	}
+
+	var walk func(any) *UploadFileResult
+	walk = func(v any) *UploadFileResult {
+		switch x := v.(type) {
+		case map[string]any:
+			if result := buildUploadFileResultFromMap(x, targetID); result != nil {
+				return result
+			}
+			for _, nested := range x {
+				if result := walk(nested); result != nil {
+					return result
+				}
+			}
+		case []any:
+			for _, item := range x {
+				if result := walk(item); result != nil {
+					return result
+				}
+			}
+		}
+		return nil
+	}
+
+	if result := walk(resp); result != nil {
+		return result
+	}
+	return nil
+}
+
+func buildUploadFileResultFromMap(m map[string]any, targetID string) *UploadFileResult {
+	fileID := strings.TrimSpace(firstNonEmptyString(m, "id", "file_id"))
+	if fileID == "" || !strings.EqualFold(fileID, targetID) {
+		return nil
+	}
+	result := &UploadFileResult{
+		ID:       fileID,
+		Filename: firstNonEmptyString(m, "name", "filename", "file_name"),
+		Status:   firstNonEmptyString(m, "status", "file_status"),
+		Purpose:  firstNonEmptyString(m, "purpose"),
+		IsImage:  firstBool(m, "is_image", "isImage"),
+		Bytes:    firstPositiveInt64(m, "bytes", "size", "file_size"),
+	}
+	if result.Status == "" {
+		result.Status = "uploaded"
+	}
+	return result
+}
+
+func mergeUploadFileResults(dst, src *UploadFileResult) {
+	if dst == nil || src == nil {
+		return
+	}
+	if strings.TrimSpace(src.ID) != "" {
+		dst.ID = strings.TrimSpace(src.ID)
+	}
+	if strings.TrimSpace(src.Filename) != "" {
+		dst.Filename = strings.TrimSpace(src.Filename)
+	}
+	if src.Bytes > 0 {
+		dst.Bytes = src.Bytes
+	}
+	if strings.TrimSpace(src.Status) != "" {
+		dst.Status = strings.TrimSpace(src.Status)
+	}
+	if strings.TrimSpace(src.Purpose) != "" {
+		dst.Purpose = strings.TrimSpace(src.Purpose)
+	}
+	dst.IsImage = src.IsImage
+	if len(src.Raw) > 0 {
+		dst.Raw = src.Raw
+	}
+	if src.RawHeaders != nil {
+		dst.RawHeaders = src.RawHeaders.Clone()
+	}
+}
+
+func isReadyUploadFileStatus(status string) bool {
+	switch strings.ToLower(strings.TrimSpace(status)) {
+	case "processed", "ready", "done", "available", "success", "completed", "finished":
+		return true
+	default:
+		return false
+	}
+}

internal/deepseek/client_http_helpers.go

@@ -35,6 +35,12 @@ func preview(b []byte) string {
 	return s
 }
 
+func (c *Client) jsonHeaders(headers map[string]string) map[string]string {
+	out := cloneStringMap(headers)
+	out["Content-Type"] = "application/json"
+	return out
+}
+
 func ScanSSELines(resp *http.Response, onLine func([]byte) bool) error {
 	scanner := bufio.NewScanner(resp.Body)
 	buf := make([]byte, 0, 64*1024)

internal/deepseek/client_http_json.go

@@ -11,8 +11,8 @@ import (
 	trans "ds2api/internal/deepseek/transport"
 )
 
-func (c *Client) postJSON(ctx context.Context, doer trans.Doer, url string, headers map[string]string, payload any) (map[string]any, error) {
-	body, status, err := c.postJSONWithStatus(ctx, doer, url, headers, payload)
+func (c *Client) postJSON(ctx context.Context, doer trans.Doer, fallback trans.Doer, url string, headers map[string]string, payload any) (map[string]any, error) {
+	body, status, err := c.postJSONWithStatus(ctx, doer, fallback, url, headers, payload)
 	if err != nil {
 		return nil, err
 	}
@@ -22,11 +22,12 @@ func (c *Client) postJSON(ctx context.Context, doer trans.Doer, url string, head
 	return body, nil
 }
 
-func (c *Client) postJSONWithStatus(ctx context.Context, doer trans.Doer, url string, headers map[string]string, payload any) (map[string]any, int, error) {
+func (c *Client) postJSONWithStatus(ctx context.Context, doer trans.Doer, fallback trans.Doer, url string, headers map[string]string, payload any) (map[string]any, int, error) {
 	b, err := json.Marshal(payload)
 	if err != nil {
 		return nil, 0, err
 	}
+	headers = c.jsonHeaders(headers)
 	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(b))
 	if err != nil {
 		return nil, 0, err
@@ -44,7 +45,7 @@ func (c *Client) postJSONWithStatus(ctx context.Context, doer trans.Doer, url st
 		for k, v := range headers {
 			req2.Header.Set(k, v)
 		}
-		resp, err = c.fallback.Do(req2)
+		resp, err = fallback.Do(req2)
 		if err != nil {
 			return nil, 0, err
 		}
@@ -64,6 +65,7 @@ func (c *Client) postJSONWithStatus(ctx context.Context, doer trans.Doer, url st
 }
 
 func (c *Client) getJSONWithStatus(ctx context.Context, doer trans.Doer, url string, headers map[string]string) (map[string]any, int, error) {
+	clients := c.requestClientsFromContext(ctx)
 	req, err := http.NewRequestWithContext(ctx, http.MethodGet, url, nil)
 	if err != nil {
 		return nil, 0, err
@@ -81,7 +83,7 @@ func (c *Client) getJSONWithStatus(ctx context.Context, doer trans.Doer, url str
 		for k, v := range headers {
 			req2.Header.Set(k, v)
 		}
-		resp, err = c.fallback.Do(req2)
+		resp, err = clients.fallback.Do(req2)
 		if err != nil {
 			return nil, 0, err
 		}

internal/deepseek/client_http_json_test.go

@@ -0,0 +1,52 @@
+package deepseek
+
+import (
+	"context"
+	"errors"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+func TestPostJSONWithStatusUsesProvidedFallbackClient(t *testing.T) {
+	var fallbackCalled bool
+	client := &Client{}
+	primary := failingDoer{err: errors.New("primary failed")}
+	fallbackDoer := doerFunc(func(req *http.Request) (*http.Response, error) {
+		fallbackCalled = true
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     make(http.Header),
+			Body:       io.NopCloser(strings.NewReader(`{"ok":true}`)),
+			Request:    req,
+		}, nil
+	})
+
+	resp, status, err := client.postJSONWithStatus(
+		context.Background(),
+		primary,
+		fallbackDoer,
+		"https://example.com/api",
+		map[string]string{"x-test": "1"},
+		map[string]any{"foo": "bar"},
+	)
+	if err != nil {
+		t.Fatalf("postJSONWithStatus error: %v", err)
+	}
+	if status != http.StatusOK {
+		t.Fatalf("status=%d want=%d", status, http.StatusOK)
+	}
+	if !fallbackCalled {
+		t.Fatal("expected provided fallback doer to be called")
+	}
+	if ok, _ := resp["ok"].(bool); !ok {
+		t.Fatalf("unexpected response body: %#v", resp)
+	}
+}
+
+type doerFunc func(*http.Request) (*http.Response, error)
+
+func (f doerFunc) Do(req *http.Request) (*http.Response, error) {
+	return f(req)
+}

internal/deepseek/client_session.go

@@ -36,6 +36,7 @@ func (c *Client) GetSessionCount(ctx context.Context, a *auth.RequestAuth, maxAt
 	if maxAttempts <= 0 {
 		maxAttempts = c.maxRetries
 	}
+	clients := c.requestClientsForAuth(ctx, a)
 
 	stats := &SessionStats{
 		AccountID: a.AccountID,
@@ -50,7 +51,7 @@ func (c *Client) GetSessionCount(ctx context.Context, a *auth.RequestAuth, maxAt
 		// 构建请求 URL
 		reqURL := DeepSeekFetchSessionURL + "?lte_cursor.pinned=false"
 
-		resp, status, err := c.getJSONWithStatus(ctx, c.regular, reqURL, headers)
+		resp, status, err := c.getJSONWithStatus(ctx, clients.regular, reqURL, headers)
 		if err != nil {
 			config.Logger.Warn("[get_session_count] request error", "error", err, "account", a.AccountID)
 			attempts++
@@ -106,10 +107,11 @@ func (c *Client) GetSessionCount(ctx context.Context, a *auth.RequestAuth, maxAt
 
 // GetSessionCountForToken 直接使用 token 获取会话数量（直通模式）
 func (c *Client) GetSessionCountForToken(ctx context.Context, token string) (*SessionStats, error) {
+	clients := c.requestClientsFromContext(ctx)
 	headers := c.authHeaders(token)
 	reqURL := DeepSeekFetchSessionURL + "?lte_cursor.pinned=false"
 
-	resp, status, err := c.getJSONWithStatus(ctx, c.regular, reqURL, headers)
+	resp, status, err := c.getJSONWithStatus(ctx, clients.regular, reqURL, headers)
 	if err != nil {
 		return nil, err
 	}
@@ -160,7 +162,7 @@ func (c *Client) GetSessionCountAll(ctx context.Context) []*SessionStats {
 		// 如果没有 token，尝试登录获取
 		if token == "" {
 			var err error
-			token, err = c.Login(ctx, acc)
+			token, err = c.Login(auth.WithAuth(ctx, &auth.RequestAuth{AccountID: acc.Identifier(), Account: acc}), acc)
 			if err != nil {
 				results = append(results, &SessionStats{
 					AccountID:    accountID,
@@ -171,7 +173,8 @@ func (c *Client) GetSessionCountAll(ctx context.Context) []*SessionStats {
 			}
 		}
 
-		stats, err := c.GetSessionCountForToken(ctx, token)
+		ctxWithAuth := auth.WithAuth(ctx, &auth.RequestAuth{AccountID: acc.Identifier(), Account: acc, DeepSeekToken: token})
+		stats, err := c.GetSessionCountForToken(ctxWithAuth, token)
 		if err != nil {
 			results = append(results, &SessionStats{
 				AccountID:    accountID,
@@ -190,6 +193,7 @@ func (c *Client) GetSessionCountAll(ctx context.Context) []*SessionStats {
 
 // FetchSessionPage 获取会话列表（支持分页）
 func (c *Client) FetchSessionPage(ctx context.Context, a *auth.RequestAuth, cursor string) ([]SessionInfo, bool, error) {
+	clients := c.requestClientsForAuth(ctx, a)
 	headers := c.authHeaders(a.DeepSeekToken)
 
 	// 构建请求 URL
@@ -200,7 +204,7 @@ func (c *Client) FetchSessionPage(ctx context.Context, a *auth.RequestAuth, curs
 	}
 	reqURL := DeepSeekFetchSessionURL + "?" + params.Encode()
 
-	resp, status, err := c.getJSONWithStatus(ctx, c.regular, reqURL, headers)
+	resp, status, err := c.getJSONWithStatus(ctx, clients.regular, reqURL, headers)
 	if err != nil {
 		return nil, false, err
 	}

internal/deepseek/client_session_delete.go

@@ -22,6 +22,7 @@ func (c *Client) DeleteSession(ctx context.Context, a *auth.RequestAuth, session
 	if maxAttempts <= 0 {
 		maxAttempts = c.maxRetries
 	}
+	clients := c.requestClientsForAuth(ctx, a)
 
 	result := &DeleteSessionResult{
 		SessionID: sessionID,
@@ -42,7 +43,7 @@ func (c *Client) DeleteSession(ctx context.Context, a *auth.RequestAuth, session
 			"chat_session_id": sessionID,
 		}
 
-		resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekDeleteSessionURL, headers, payload)
+		resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekDeleteSessionURL, headers, payload)
 		if err != nil {
 			config.Logger.Warn("[delete_session] request error", "error", err, "session_id", sessionID)
 			attempts++
@@ -81,6 +82,7 @@ func (c *Client) DeleteSession(ctx context.Context, a *auth.RequestAuth, session
 
 // DeleteSessionForToken 直接使用 token 删除会话（直通模式）
 func (c *Client) DeleteSessionForToken(ctx context.Context, token string, sessionID string) (*DeleteSessionResult, error) {
+	clients := c.requestClientsFromContext(ctx)
 	result := &DeleteSessionResult{
 		SessionID: sessionID,
 	}
@@ -95,7 +97,7 @@ func (c *Client) DeleteSessionForToken(ctx context.Context, token string, sessio
 		"chat_session_id": sessionID,
 	}
 
-	resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekDeleteSessionURL, headers, payload)
+	resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekDeleteSessionURL, headers, payload)
 	if err != nil {
 		result.ErrorMessage = err.Error()
 		return result, err
@@ -114,10 +116,11 @@ func (c *Client) DeleteSessionForToken(ctx context.Context, token string, sessio
 
 // DeleteAllSessions 删除所有会话（谨慎使用）
 func (c *Client) DeleteAllSessions(ctx context.Context, a *auth.RequestAuth) error {
+	clients := c.requestClientsForAuth(ctx, a)
 	headers := c.authHeaders(a.DeepSeekToken)
 	payload := map[string]any{}
 
-	resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekDeleteAllSessionsURL, headers, payload)
+	resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekDeleteAllSessionsURL, headers, payload)
 	if err != nil {
 		config.Logger.Warn("[delete_all_sessions] request error", "error", err)
 		return err
@@ -135,10 +138,11 @@ func (c *Client) DeleteAllSessions(ctx context.Context, a *auth.RequestAuth) err
 
 // DeleteAllSessionsForToken 直接使用 token 删除所有会话（直通模式）
 func (c *Client) DeleteAllSessionsForToken(ctx context.Context, token string) error {
+	clients := c.requestClientsFromContext(ctx)
 	headers := c.authHeaders(token)
 	payload := map[string]any{}
 
-	resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekDeleteAllSessionsURL, headers, payload)
+	resp, status, err := c.postJSONWithStatus(ctx, clients.regular, clients.fallback, DeepSeekDeleteAllSessionsURL, headers, payload)
 	if err != nil {
 		config.Logger.Warn("[delete_all_sessions_for_token] request error", "error", err)
 		return err

internal/deepseek/client_upload.go

@@ -0,0 +1,282 @@
+package deepseek
+
+import (
+	"bytes"
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"mime/multipart"
+	"net/http"
+	"net/textproto"
+	"path/filepath"
+	"strconv"
+	"strings"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+	trans "ds2api/internal/deepseek/transport"
+)
+
+type UploadFileRequest struct {
+	Filename    string
+	ContentType string
+	Purpose     string
+	Data        []byte
+}
+
+type UploadFileResult struct {
+	ID         string
+	Filename   string
+	Bytes      int64
+	Status     string
+	Purpose    string
+	AccountID  string
+	IsImage    bool
+	Raw        map[string]any
+	RawHeaders http.Header
+}
+
+func (c *Client) UploadFile(ctx context.Context, a *auth.RequestAuth, req UploadFileRequest, maxAttempts int) (*UploadFileResult, error) {
+	if maxAttempts <= 0 {
+		maxAttempts = c.maxRetries
+	}
+	if len(req.Data) == 0 {
+		return nil, errors.New("file is required")
+	}
+	filename := strings.TrimSpace(req.Filename)
+	if filename == "" {
+		filename = "upload.bin"
+	}
+	contentType := strings.TrimSpace(req.ContentType)
+	if contentType == "" {
+		contentType = "application/octet-stream"
+	}
+	purpose := strings.TrimSpace(req.Purpose)
+	body, contentTypeHeader, err := buildUploadMultipartBody(filename, contentType, req.Data)
+	if err != nil {
+		return nil, err
+	}
+	capturePayload := map[string]any{
+		"filename":     filename,
+		"content_type": contentType,
+		"purpose":      purpose,
+		"bytes":        len(req.Data),
+	}
+	captureSession := c.capture.Start("deepseek_upload_file", DeepSeekUploadFileURL, a.AccountID, capturePayload)
+	attempts := 0
+	refreshed := false
+	powHeader := ""
+	for attempts < maxAttempts {
+		clients := c.requestClientsForAuth(ctx, a)
+		if strings.TrimSpace(powHeader) == "" {
+			powHeader, err = c.GetPowForTarget(ctx, a, DeepSeekUploadTargetPath, maxAttempts)
+			if err != nil {
+				return nil, err
+			}
+			clients = c.requestClientsForAuth(ctx, a)
+		}
+		headers := c.authHeaders(a.DeepSeekToken)
+		headers["Content-Type"] = contentTypeHeader
+		headers["x-ds-pow-response"] = powHeader
+		headers["x-file-size"] = strconv.Itoa(len(req.Data))
+		headers["x-thinking-enabled"] = "1"
+		resp, err := c.doUpload(ctx, clients.regular, clients.fallback, DeepSeekUploadFileURL, headers, body)
+		if err != nil {
+			config.Logger.Warn("[upload_file] request error", "error", err, "account", a.AccountID, "filename", filename)
+			powHeader = ""
+			attempts++
+			continue
+		}
+		if captureSession != nil {
+			resp.Body = captureSession.WrapBody(resp.Body, resp.StatusCode)
+		}
+		payloadBytes, readErr := readResponseBody(resp)
+		_ = resp.Body.Close()
+		if readErr != nil {
+			powHeader = ""
+			attempts++
+			continue
+		}
+		parsed := map[string]any{}
+		if len(payloadBytes) > 0 {
+			if err := json.Unmarshal(payloadBytes, &parsed); err != nil {
+				config.Logger.Warn("[upload_file] json parse failed", "status", resp.StatusCode, "preview", preview(payloadBytes))
+			}
+		}
+		code, bizCode, msg, bizMsg := extractResponseStatus(parsed)
+		if resp.StatusCode == http.StatusOK && code == 0 && bizCode == 0 {
+			result := extractUploadFileResult(parsed)
+			result.Raw = parsed
+			result.RawHeaders = resp.Header.Clone()
+			if result.Filename == "" {
+				result.Filename = filename
+			}
+			if result.Bytes == 0 {
+				result.Bytes = int64(len(req.Data))
+			}
+			if result.Purpose == "" {
+				result.Purpose = purpose
+			}
+			if result.AccountID == "" {
+				result.AccountID = a.AccountID
+			}
+			if result.ID == "" {
+				return nil, errors.New("upload file succeeded without file id")
+			}
+			if err := c.waitForUploadedFile(ctx, a, result); err != nil {
+				return nil, err
+			}
+			return result, nil
+		}
+		config.Logger.Warn("[upload_file] failed", "status", resp.StatusCode, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "account", a.AccountID, "filename", filename)
+		powHeader = ""
+		if a.UseConfigToken {
+			if !refreshed && shouldAttemptRefresh(resp.StatusCode, code, bizCode, msg, bizMsg) {
+				if c.Auth.RefreshToken(ctx, a) {
+					refreshed = true
+					attempts++
+					continue
+				}
+			}
+			if c.Auth.SwitchAccount(ctx, a) {
+				refreshed = false
+				attempts++
+				continue
+			}
+		}
+		attempts++
+	}
+	return nil, errors.New("upload file failed")
+}
+
+func buildUploadMultipartBody(filename, contentType string, data []byte) ([]byte, string, error) {
+	var buf bytes.Buffer
+	writer := multipart.NewWriter(&buf)
+	partHeader := textproto.MIMEHeader{}
+	partHeader.Set("Content-Disposition", fmt.Sprintf(`form-data; name="file"; filename=%q`, escapeMultipartFilename(filename)))
+	partHeader.Set("Content-Type", contentType)
+	part, err := writer.CreatePart(partHeader)
+	if err != nil {
+		return nil, "", err
+	}
+	if _, err := part.Write(data); err != nil {
+		return nil, "", err
+	}
+	if err := writer.Close(); err != nil {
+		return nil, "", err
+	}
+	return buf.Bytes(), writer.FormDataContentType(), nil
+}
+
+func escapeMultipartFilename(filename string) string {
+	filename = filepath.Base(strings.TrimSpace(filename))
+	filename = strings.ReplaceAll(filename, `\`, "_")
+	filename = strings.ReplaceAll(filename, `"`, "_")
+	if filename == "." || filename == "" {
+		return "upload.bin"
+	}
+	return filename
+}
+
+func (c *Client) doUpload(ctx context.Context, doer trans.Doer, fallback trans.Doer, url string, headers map[string]string, body []byte) (*http.Response, error) {
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return nil, err
+	}
+	for k, v := range headers {
+		req.Header.Set(k, v)
+	}
+	resp, err := doer.Do(req)
+	if err == nil {
+		return resp, nil
+	}
+	config.Logger.Warn("[deepseek] fingerprint upload request failed, fallback to std transport", "url", url, "error", err)
+	req2, reqErr := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if reqErr != nil {
+		return nil, reqErr
+	}
+	for k, v := range headers {
+		req2.Header.Set(k, v)
+	}
+	return fallback.Do(req2)
+}
+
+func extractUploadFileResult(resp map[string]any) *UploadFileResult {
+	result := &UploadFileResult{Status: "uploaded"}
+	data, _ := resp["data"].(map[string]any)
+	bizData, _ := data["biz_data"].(map[string]any)
+	searchMaps := []map[string]any{resp, data, bizData}
+	for _, parent := range []map[string]any{resp, data, bizData} {
+		if parent == nil {
+			continue
+		}
+		for _, key := range []string{"file", "biz_data", "data"} {
+			if nested, ok := parent[key].(map[string]any); ok {
+				searchMaps = append(searchMaps, nested)
+			}
+		}
+	}
+	for _, m := range searchMaps {
+		if m == nil {
+			continue
+		}
+		if result.ID == "" {
+			result.ID = firstNonEmptyString(m, "id", "file_id")
+		}
+		if result.Filename == "" {
+			result.Filename = firstNonEmptyString(m, "name", "filename", "file_name")
+		}
+		if result.Status == "uploaded" {
+			if status := firstNonEmptyString(m, "status", "file_status"); status != "" {
+				result.Status = status
+			}
+		}
+		if !result.IsImage {
+			result.IsImage = firstBool(m, "is_image", "isImage")
+		}
+		if result.Purpose == "" {
+			result.Purpose = firstNonEmptyString(m, "purpose")
+		}
+		if result.AccountID == "" {
+			result.AccountID = firstNonEmptyString(m, "account_id", "accountId", "owner_account_id", "ownerAccountId")
+		}
+		if result.Bytes == 0 {
+			result.Bytes = firstPositiveInt64(m, "bytes", "size", "file_size")
+		}
+	}
+	return result
+}
+
+func firstBool(m map[string]any, keys ...string) bool {
+	for _, key := range keys {
+		switch v := m[key].(type) {
+		case bool:
+			return v
+		case string:
+			switch strings.ToLower(strings.TrimSpace(v)) {
+			case "true", "1", "yes", "y":
+				return true
+			}
+		}
+	}
+	return false
+}
+
+func firstNonEmptyString(m map[string]any, keys ...string) string {
+	for _, key := range keys {
+		if v, _ := m[key].(string); strings.TrimSpace(v) != "" {
+			return strings.TrimSpace(v)
+		}
+	}
+	return ""
+}
+
+func firstPositiveInt64(m map[string]any, keys ...string) int64 {
+	for _, key := range keys {
+		if v := toInt64(m[key], 0); v > 0 {
+			return v
+		}
+	}
+	return 0
+}

internal/deepseek/client_upload_test.go

@@ -0,0 +1,216 @@
+package deepseek
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/hex"
+	"encoding/json"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+	"time"
+
+	"ds2api/internal/auth"
+	powpkg "ds2api/pow"
+)
+
+func TestBuildUploadMultipartBodyOmitsPurposeAndIncludesFilePart(t *testing.T) {
+	body, contentType, err := buildUploadMultipartBody(`../demo.txt`, "text/plain", []byte("hello"))
+	if err != nil {
+		t.Fatalf("buildUploadMultipartBody error: %v", err)
+	}
+	if !strings.HasPrefix(contentType, "multipart/form-data; boundary=") {
+		t.Fatalf("unexpected content type: %q", contentType)
+	}
+	payload := string(body)
+	if strings.Contains(payload, `name="purpose"`) || strings.Contains(payload, "assistants") {
+		t.Fatalf("expected purpose to be omitted from payload: %q", payload)
+	}
+	if !strings.Contains(payload, `name="file"; filename="demo.txt"`) {
+		t.Fatalf("expected sanitized filename in payload: %q", payload)
+	}
+	if !strings.Contains(payload, "Content-Type: text/plain") {
+		t.Fatalf("expected file content type in payload: %q", payload)
+	}
+	if !strings.Contains(payload, "hello") {
+		t.Fatalf("expected file content in payload: %q", payload)
+	}
+}
+
+func TestExtractUploadFileResultSupportsNestedShapes(t *testing.T) {
+	got := extractUploadFileResult(map[string]any{
+		"data": map[string]any{
+			"biz_data": map[string]any{
+				"file": map[string]any{
+					"file_id":   "file_123",
+					"file_name": "report.pdf",
+					"file_size": 99,
+					"status":    "processed",
+					"purpose":   "assistants",
+					"is_image":  true,
+				},
+			},
+		},
+	})
+	if got.ID != "file_123" {
+		t.Fatalf("expected id file_123, got %#v", got)
+	}
+	if got.Filename != "report.pdf" {
+		t.Fatalf("expected filename report.pdf, got %#v", got)
+	}
+	if got.Bytes != 99 {
+		t.Fatalf("expected bytes 99, got %#v", got)
+	}
+	if got.Status != "processed" {
+		t.Fatalf("expected status processed, got %#v", got)
+	}
+	if got.Purpose != "assistants" {
+		t.Fatalf("expected purpose assistants, got %#v", got)
+	}
+	if !got.IsImage {
+		t.Fatalf("expected image flag true, got %#v", got)
+	}
+}
+
+func TestUploadFileUsesUploadTargetPowAndMultipartHeaders(t *testing.T) {
+	challengeHash := powpkg.DeepSeekHashV1([]byte(powpkg.BuildPrefix("salt", 1712345678) + "42"))
+	powResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"challenge":{"algorithm":"DeepSeekHashV1","challenge":"` + hex.EncodeToString(challengeHash[:]) + `","salt":"salt","expire_at":1712345678,"difficulty":1000,"signature":"sig","target_path":"` + DeepSeekUploadTargetPath + `"}}}}`
+	uploadResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"file":{"file_id":"file_789","filename":"demo.txt","bytes":5,"status":"processed","purpose":"assistants","is_image":false}}}}`
+	var seenPow string
+	var seenTargetPath string
+	var seenContentType string
+	var seenFileSize string
+	var seenBody string
+	call := 0
+	client := &Client{
+		regular: doerFunc(func(req *http.Request) (*http.Response, error) {
+			call++
+			bodyBytes, _ := io.ReadAll(req.Body)
+			switch call {
+			case 1:
+				seenTargetPath = string(bodyBytes)
+				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(powResponse)), Request: req}, nil
+			case 2:
+				seenPow = req.Header.Get("x-ds-pow-response")
+				seenContentType = req.Header.Get("Content-Type")
+				seenFileSize = req.Header.Get("x-file-size")
+				seenBody = string(bodyBytes)
+				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(uploadResponse)), Request: req}, nil
+			default:
+				t.Fatalf("unexpected request count %d", call)
+				return nil, nil
+			}
+		}),
+		fallback: &http.Client{Transport: roundTripperFunc(func(req *http.Request) (*http.Response, error) {
+			return nil, nil
+		})},
+		maxRetries: 1,
+	}
+	result, err := client.UploadFile(context.Background(), &auth.RequestAuth{DeepSeekToken: "token", TriedAccounts: map[string]bool{}}, UploadFileRequest{
+		Filename:    "demo.txt",
+		ContentType: "text/plain",
+		Purpose:     "assistants",
+		Data:        []byte("hello"),
+	}, 1)
+	if err != nil {
+		t.Fatalf("UploadFile error: %v", err)
+	}
+	if result.ID != "file_789" {
+		t.Fatalf("expected uploaded file id file_789, got %#v", result)
+	}
+	if !strings.Contains(seenTargetPath, `"target_path":"`+DeepSeekUploadTargetPath+`"`) {
+		t.Fatalf("expected upload target_path in pow request, got %q", seenTargetPath)
+	}
+	if strings.TrimSpace(seenPow) == "" {
+		t.Fatal("expected x-ds-pow-response header")
+	}
+	rawPow, err := base64.StdEncoding.DecodeString(seenPow)
+	if err != nil {
+		t.Fatalf("decode pow header failed: %v", err)
+	}
+	var powHeader map[string]any
+	if err := json.Unmarshal(rawPow, &powHeader); err != nil {
+		t.Fatalf("unmarshal pow header failed: %v", err)
+	}
+	if powHeader["target_path"] != DeepSeekUploadTargetPath {
+		t.Fatalf("expected pow target_path %q, got %#v", DeepSeekUploadTargetPath, powHeader["target_path"])
+	}
+	if seenFileSize != "5" {
+		t.Fatalf("expected x-file-size=5, got %q", seenFileSize)
+	}
+	if !strings.HasPrefix(seenContentType, "multipart/form-data; boundary=") {
+		t.Fatalf("expected multipart content type, got %q", seenContentType)
+	}
+	if !strings.Contains(seenBody, `name="file"; filename="demo.txt"`) {
+		t.Fatalf("expected file part in upload body: %q", seenBody)
+	}
+}
+
+func TestUploadFileWaitsForProcessedFetchFiles(t *testing.T) {
+	oldSleep := fileReadySleep
+	fileReadySleep = func(time.Duration) {}
+	defer func() { fileReadySleep = oldSleep }()
+
+	challengeHash := powpkg.DeepSeekHashV1([]byte(powpkg.BuildPrefix("salt", 1712345678) + "42"))
+	powResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"challenge":{"algorithm":"DeepSeekHashV1","challenge":"` + hex.EncodeToString(challengeHash[:]) + `","salt":"salt","expire_at":1712345678,"difficulty":1000,"signature":"sig","target_path":"` + DeepSeekUploadTargetPath + `"}}}}`
+	uploadResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"file":{"file_id":"file_789","filename":"demo.txt","bytes":5,"status":"PENDING","purpose":"assistants","is_image":false}}}}`
+	pendingFetchResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"files":[{"file_id":"file_789","filename":"demo.txt","bytes":5,"status":"PENDING","purpose":"assistants","is_image":false}]}}}`
+	processedFetchResponse := `{"code":0,"msg":"ok","data":{"biz_code":0,"biz_data":{"files":[{"file_id":"file_789","filename":"demo.txt","bytes":5,"status":"processed","purpose":"assistants","is_image":true}]}}}`
+
+	var call int
+	client := &Client{
+		regular: doerFunc(func(req *http.Request) (*http.Response, error) {
+			call++
+			switch call {
+			case 1:
+				bodyBytes, _ := io.ReadAll(req.Body)
+				if !strings.Contains(string(bodyBytes), `"target_path":"`+DeepSeekUploadTargetPath+`"`) {
+					t.Fatalf("expected pow target path request, got %s", string(bodyBytes))
+				}
+				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(powResponse)), Request: req}, nil
+			case 2:
+				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(uploadResponse)), Request: req}, nil
+			case 3, 4:
+				if req.Method != http.MethodGet {
+					t.Fatalf("expected GET fetch request, got %s", req.Method)
+				}
+				if req.URL.Path != "/api/v0/file/fetch_files" {
+					t.Fatalf("expected fetch files path /api/v0/file/fetch_files, got %q", req.URL.Path)
+				}
+				if got := req.URL.Query().Get("file_ids"); got != "file_789" {
+					t.Fatalf("expected file_ids=file_789, got %q", got)
+				}
+				respBody := pendingFetchResponse
+				if call == 4 {
+					respBody = processedFetchResponse
+				}
+				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(respBody)), Request: req}, nil
+			default:
+				t.Fatalf("unexpected request count %d", call)
+				return nil, nil
+			}
+		}),
+		fallback:   &http.Client{Transport: roundTripperFunc(func(req *http.Request) (*http.Response, error) { return nil, nil })},
+		maxRetries: 1,
+	}
+
+	result, err := client.UploadFile(context.Background(), &auth.RequestAuth{DeepSeekToken: "token", TriedAccounts: map[string]bool{}}, UploadFileRequest{
+		Filename:    "demo.txt",
+		ContentType: "text/plain",
+		Purpose:     "assistants",
+		Data:        []byte("hello"),
+	}, 1)
+	if err != nil {
+		t.Fatalf("UploadFile error: %v", err)
+	}
+	if result.ID != "file_789" {
+		t.Fatalf("expected uploaded file id file_789, got %#v", result)
+	}
+	if result.Status != "processed" {
+		t.Fatalf("expected final status processed, got %#v", result.Status)
+	}
+	if call != 4 {
+		t.Fatalf("expected 4 requests, got %d", call)
+	}
+}

internal/deepseek/constants.go

@@ -12,18 +12,22 @@ const (
 	DeepSeekCreatePowURL         = "https://chat.deepseek.com/api/v0/chat/create_pow_challenge"
 	DeepSeekCompletionURL        = "https://chat.deepseek.com/api/v0/chat/completion"
 	DeepSeekContinueURL          = "https://chat.deepseek.com/api/v0/chat/continue"
+	DeepSeekUploadFileURL        = "https://chat.deepseek.com/api/v0/file/upload_file"
+	DeepSeekFetchFilesURL        = "https://chat.deepseek.com/api/v0/file/fetch_files"
 	DeepSeekFetchSessionURL      = "https://chat.deepseek.com/api/v0/chat_session/fetch_page"
 	DeepSeekDeleteSessionURL     = "https://chat.deepseek.com/api/v0/chat_session/delete"
 	DeepSeekDeleteAllSessionsURL = "https://chat.deepseek.com/api/v0/chat_session/delete_all"
+	DeepSeekCompletionTargetPath = "/api/v0/chat/completion"
+	DeepSeekUploadTargetPath     = "/api/v0/file/upload_file"
 )
 
 var defaultBaseHeaders = map[string]string{
 	"Host":              "chat.deepseek.com",
-	"User-Agent":        "DeepSeek/1.6.11 Android/35",
+	"User-Agent":        "DeepSeek/1.8.0 Android/35",
 	"Accept":            "application/json",
 	"Content-Type":      "application/json",
 	"x-client-platform": "android",
-	"x-client-version":  "1.6.11",
+	"x-client-version":  "1.8.0",
 	"x-client-locale":   "zh_CN",
 	"accept-charset":    "UTF-8",
 }

internal/deepseek/constants_shared.json

@@ -1,18 +1,16 @@
 {
   "base_headers": {
     "Host": "chat.deepseek.com",
-    "User-Agent": "DeepSeek/1.6.11 Android/35",
+    "User-Agent": "DeepSeek/1.8.0 Android/35",
     "Accept": "application/json",
-    "Content-Type": "application/json",
     "x-client-platform": "android",
-    "x-client-version": "1.6.11",
+    "x-client-version": "1.8.0",
     "x-client-locale": "zh_CN",
     "accept-charset": "UTF-8"
   },
   "skip_contains_patterns": [
     "quasi_status",
     "elapsed_secs",
-    "token_usage",
     "pending_fragment",
     "conversation_mode",
     "fragments/-1/status",

internal/deepseek/deepseek_edge_test.go

@@ -105,43 +105,16 @@ func TestBuildPowHeaderEmptyChallenge(t *testing.T) {
 	}
 }
 
-// ─── PowSolver pool size ─────────────────────────────────────────────
-
-func TestPowPoolSizeFromEnvDefault(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "")
-	got := powPoolSizeFromEnv()
-	if got < 1 {
-		t.Fatalf("expected positive default pool size, got %d", got)
-	}
-}
-
-func TestPowPoolSizeFromEnvInvalid(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "abc")
-	got := powPoolSizeFromEnv()
-	if got < 1 {
-		t.Fatalf("expected positive default for invalid, got %d", got)
-	}
-}
-
-func TestPowPoolSizeFromEnvSpecificValue(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "5")
-	got := powPoolSizeFromEnv()
-	if got != 5 {
-		t.Fatalf("expected 5, got %d", got)
-	}
-}
-
 // ─── NewClient ───────────────────────────────────────────────────────
 
 func TestNewClientInitialState(t *testing.T) {
 	client := NewClient(nil, nil)
-	if client.powSolver == nil {
-		t.Fatal("expected powSolver to be initialized")
+	if client == nil {
+		t.Fatal("expected non-nil client")
 	}
 }
 
 func TestNewClientPreloadPowIdempotent(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "1")
 	client := NewClient(nil, nil)
 	if err := client.PreloadPow(context.Background()); err != nil {
 		t.Fatalf("first preload failed: %v", err)
@@ -150,16 +123,3 @@ func TestNewClientPreloadPowIdempotent(t *testing.T) {
 		t.Fatalf("second preload failed: %v", err)
 	}
 }
-
-// ─── PowSolver init and module pool ──────────────────────────────────
-
-func TestPowSolverPoolSizeMatchesEnv(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "2")
-	solver := NewPowSolver("test.wasm")
-	if err := solver.init(context.Background()); err != nil {
-		t.Fatalf("init failed: %v", err)
-	}
-	if cap(solver.pool) != 2 {
-		t.Fatalf("expected pool capacity 2, got %d", cap(solver.pool))
-	}
-}

internal/deepseek/embedded_pow.go

@@ -1,6 +0,0 @@
-package deepseek
-
-import _ "embed"
-
-//go:embed assets/sha3_wasm_bg.7b9ca65ddd.wasm
-var embeddedWASM []byte

internal/deepseek/pow.go

@@ -3,218 +3,27 @@ package deepseek
 import (
 	"context"
 	"encoding/base64"
-	"encoding/binary"
 	"encoding/json"
 	"errors"
-	"math"
-	"os"
-	stdruntime "runtime"
-	"strconv"
-	"sync"
 
-	"ds2api/internal/config"
-
-	"github.com/tetratelabs/wazero"
-	"github.com/tetratelabs/wazero/api"
+	"ds2api/pow"
 )
 
-type PowSolver struct {
-	wasmPath string
-	once     sync.Once
-	err      error
-
-	runtime  wazero.Runtime
-	compiled wazero.CompiledModule
-	pool     chan *pooledModule
-	poolSize int
-}
-
-type pooledModule struct {
-	mod     api.Module
-	stackFn api.Function
-	allocFn api.Function
-	freeFn  api.Function
-	solveFn api.Function
-}
-
-func NewPowSolver(wasmPath string) *PowSolver {
-	return &PowSolver{wasmPath: wasmPath}
-}
-
-func (p *PowSolver) init(ctx context.Context) error {
-	p.once.Do(func() {
-		wasmBytes, err := os.ReadFile(p.wasmPath)
-		if err != nil {
-			if len(embeddedWASM) == 0 {
-				p.err = err
-				return
-			}
-			wasmBytes = embeddedWASM
-		}
-		p.runtime = wazero.NewRuntime(ctx)
-		p.compiled, p.err = p.runtime.CompileModule(ctx, wasmBytes)
-		if p.err == nil {
-			p.poolSize = powPoolSizeFromEnv()
-			p.pool = make(chan *pooledModule, p.poolSize)
-			for range p.poolSize {
-				inst, err := p.createModule(ctx)
-				if err != nil {
-					p.err = err
-					return
-				}
-				p.pool <- inst
-			}
-		}
-	})
-	return p.err
-}
-
-func (p *PowSolver) Compute(ctx context.Context, challenge map[string]any) (int64, error) {
-	if err := p.init(ctx); err != nil {
-		return 0, err
-	}
+// ComputePow 使用纯 Go 实现求解 PoW challenge (DeepSeekHashV1)。
+func ComputePow(ctx context.Context, challenge map[string]any) (int64, error) {
 	algo, _ := challenge["algorithm"].(string)
 	if algo != "DeepSeekHashV1" {
 		return 0, errors.New("unsupported algorithm")
 	}
 	challengeStr, _ := challenge["challenge"].(string)
 	salt, _ := challenge["salt"].(string)
-	signature, _ := challenge["signature"].(string)
-	targetPath, _ := challenge["target_path"].(string)
-	_ = signature
-	_ = targetPath
-
-	difficulty := toFloat64(challenge["difficulty"], 144000)
 	expireAt := toInt64(challenge["expire_at"], 1680000000)
-	prefix := salt + "_" + itoa(expireAt) + "_"
+	difficulty := toInt64FromFloat(challenge["difficulty"], 144000)
 
-	pm, err := p.acquireModule(ctx)
-	if err != nil {
-		return 0, err
-	}
-	defer p.releaseModule(pm)
-
-	mem := pm.mod.Memory()
-	if mem == nil {
-		return 0, errors.New("wasm memory missing")
-	}
-	retPtrs, err := pm.stackFn.Call(ctx, uint64(uint32(^uint32(15)))) // -16 i32
-	if err != nil || len(retPtrs) == 0 {
-		return 0, errors.New("stack alloc failed")
-	}
-	retptr := uint32(retPtrs[0])
-	defer func() {
-		_, _ = pm.stackFn.Call(context.Background(), 16)
-	}()
-
-	chPtr, chLen, err := writeUTF8(ctx, pm.allocFn, mem, challengeStr)
-	if err != nil {
-		return 0, err
-	}
-	defer freeUTF8(pm.freeFn, chPtr, chLen)
-
-	prefixPtr, prefixLen, err := writeUTF8(ctx, pm.allocFn, mem, prefix)
-	if err != nil {
-		return 0, err
-	}
-	defer freeUTF8(pm.freeFn, prefixPtr, prefixLen)
-
-	if _, err := pm.solveFn.Call(ctx,
-		uint64(retptr),
-		uint64(chPtr), uint64(chLen),
-		uint64(prefixPtr), uint64(prefixLen),
-		math.Float64bits(difficulty),
-	); err != nil {
-		return 0, err
-	}
-
-	statusBytes, ok := mem.Read(retptr, 4)
-	if !ok {
-		return 0, errors.New("read status failed")
-	}
-	status := int32(binary.LittleEndian.Uint32(statusBytes))
-	valueBytes, ok := mem.Read(retptr+8, 8)
-	if !ok {
-		return 0, errors.New("read value failed")
-	}
-	value := math.Float64frombits(binary.LittleEndian.Uint64(valueBytes))
-	if status == 0 {
-		return 0, errors.New("pow solve failed")
-	}
-	return int64(value), nil
-}
-
-func (p *PowSolver) createModule(ctx context.Context) (*pooledModule, error) {
-	mod, err := p.runtime.InstantiateModule(ctx, p.compiled, wazero.NewModuleConfig())
-	if err != nil {
-		return nil, err
-	}
-	stackFn := mod.ExportedFunction("__wbindgen_add_to_stack_pointer")
-	allocFn := mod.ExportedFunction("__wbindgen_export_0")
-	solveFn := mod.ExportedFunction("wasm_solve")
-	if stackFn == nil || allocFn == nil || solveFn == nil {
-		_ = mod.Close(context.Background())
-		return nil, errors.New("required wasm exports missing")
-	}
-	return &pooledModule{
-		mod:     mod,
-		stackFn: stackFn,
-		allocFn: allocFn,
-		freeFn:  mod.ExportedFunction("__wbindgen_export_2"),
-		solveFn: solveFn,
-	}, nil
-}
-
-func (p *PowSolver) acquireModule(ctx context.Context) (*pooledModule, error) {
-	if p.pool != nil {
-		for {
-			select {
-			case pm := <-p.pool:
-				if pm != nil {
-					return pm, nil
-				}
-			case <-ctx.Done():
-				return nil, ctx.Err()
-			}
-		}
-	}
-	return p.createModule(ctx)
-}
-
-func (p *PowSolver) releaseModule(pm *pooledModule) {
-	if pm == nil || pm.mod == nil {
-		return
-	}
-	if p.pool != nil {
-		select {
-		case p.pool <- pm:
-			return
-		default:
-		}
-	}
-	_ = pm.mod.Close(context.Background())
-}
-
-func writeUTF8(ctx context.Context, allocFn api.Function, mem api.Memory, text string) (uint32, uint32, error) {
-	data := []byte(text)
-	res, err := allocFn.Call(ctx, uint64(len(data)), 1)
-	if err != nil || len(res) == 0 {
-		return 0, 0, errors.New("alloc failed")
-	}
-	ptr := uint32(res[0])
-	if !mem.Write(ptr, data) {
-		return 0, 0, errors.New("mem write failed")
-	}
-	return ptr, uint32(len(data)), nil
-}
-
-func freeUTF8(freeFn api.Function, ptr, size uint32) {
-	if freeFn == nil || ptr == 0 || size == 0 {
-		return
-	}
-	_, _ = freeFn.Call(context.Background(), uint64(ptr), uint64(size), 1)
+	return pow.SolvePow(ctx, challengeStr, salt, expireAt, difficulty)
 }
 
+// BuildPowHeader 序列化 {algorithm,challenge,salt,answer,signature,target_path} 为 base64(JSON)。
 func BuildPowHeader(challenge map[string]any, answer int64) (string, error) {
 	payload := map[string]any{
 		"algorithm":   challenge["algorithm"],
@@ -257,32 +66,7 @@ func toInt64(v any, d int64) int64 {
 	}
 }
 
-func itoa(n int64) string {
-	return strconv.FormatInt(n, 10)
-}
-
-func powPoolSizeFromEnv() int {
-	const fallback = 4
-	n := fallback
-	if cpus := stdruntime.GOMAXPROCS(0); cpus > 0 {
-		n = cpus
-	}
-	if raw := os.Getenv("DS2API_POW_POOL_SIZE"); raw != "" {
-		if v, err := strconv.Atoi(raw); err == nil && v > 0 {
-			n = v
-		}
-	}
-	if n > 64 {
-		return 64
-	}
-	return n
-}
-
-func PreloadWASM(wasmPath string) {
-	solver := NewPowSolver(wasmPath)
-	if err := solver.init(context.Background()); err != nil {
-		config.Logger.Warn("[WASM] preload failed", "error", err)
-		return
-	}
-	config.Logger.Info("[WASM] module preloaded", "path", wasmPath)
+// toInt64FromFloat 与 toInt64 等价，仅名称区分用途。
+func toInt64FromFloat(v any, d int64) int64 {
+	return toInt64(v, d)
 }

internal/deepseek/pow_test.go

@@ -3,66 +3,18 @@ package deepseek
 import (
 	"context"
 	"testing"
-	"time"
 )
 
-func TestPowPoolSizeFromEnv(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "3")
-	if got := powPoolSizeFromEnv(); got != 3 {
-		t.Fatalf("expected pool size 3, got %d", got)
-	}
-}
-
-func TestPowSolverAcquireReleaseReusesModule(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "1")
-	solver := NewPowSolver("missing-file.wasm")
-	if err := solver.init(context.Background()); err != nil {
-		t.Fatalf("init failed: %v", err)
-	}
-
-	pm1, err := solver.acquireModule(context.Background())
-	if err != nil {
-		t.Fatalf("acquire first module failed: %v", err)
-	}
-	solver.releaseModule(pm1)
-
-	pm2, err := solver.acquireModule(context.Background())
-	if err != nil {
-		t.Fatalf("acquire second module failed: %v", err)
-	}
-	if pm1 != pm2 {
-		t.Fatalf("expected pooled module reuse, got different instances")
-	}
-	solver.releaseModule(pm2)
-}
-
-func TestPowSolverAcquireHonorsContextWhenPoolExhausted(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "1")
-	solver := NewPowSolver("missing-file.wasm")
-	if err := solver.init(context.Background()); err != nil {
-		t.Fatalf("init failed: %v", err)
-	}
-
-	held, err := solver.acquireModule(context.Background())
-	if err != nil {
-		t.Fatalf("acquire held module failed: %v", err)
-	}
-	defer solver.releaseModule(held)
-
-	ctx, cancel := context.WithTimeout(context.Background(), 20*time.Millisecond)
-	defer cancel()
-	if _, err := solver.acquireModule(ctx); err == nil {
-		t.Fatalf("expected context cancellation while pool is exhausted")
-	}
-}
-
-func TestClientPreloadPowUsesClientSolver(t *testing.T) {
-	t.Setenv("DS2API_POW_POOL_SIZE", "1")
+func TestPreloadPowNoOp(t *testing.T) {
 	client := NewClient(nil, nil)
 	if err := client.PreloadPow(context.Background()); err != nil {
-		t.Fatalf("preload failed: %v", err)
-	}
-	if client.powSolver.runtime == nil || client.powSolver.compiled == nil {
-		t.Fatalf("expected client pow solver to be initialized")
+		t.Fatalf("PreloadPow should be no-op, got error: %v", err)
+	}
+}
+
+func TestComputePowUnsupportedAlgorithm(t *testing.T) {
+	_, err := ComputePow(context.Background(), map[string]any{"algorithm": "unknown"})
+	if err == nil {
+		t.Fatal("expected error for unsupported algorithm")
 	}
 }

internal/deepseek/prompt.go

@@ -5,3 +5,7 @@ import "ds2api/internal/prompt"
 func MessagesPrepare(messages []map[string]any) string {
 	return prompt.MessagesPrepare(messages)
 }
+
+func MessagesPrepareWithThinking(messages []map[string]any, thinkingEnabled bool) string {
+	return prompt.MessagesPrepareWithThinking(messages, thinkingEnabled)
+}

internal/deepseek/proxy.go

@@ -0,0 +1,239 @@
+package deepseek
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"net/http"
+	"strconv"
+	"strings"
+	"time"
+
+	"golang.org/x/net/proxy"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+	trans "ds2api/internal/deepseek/transport"
+)
+
+type requestClients struct {
+	regular   trans.Doer
+	stream    trans.Doer
+	fallback  *http.Client
+	fallbackS *http.Client
+}
+
+type hostLookupFunc func(ctx context.Context, network, host string) ([]string, error)
+
+var proxyConnectivityTestURL = "https://chat.deepseek.com/"
+
+var defaultHostLookup hostLookupFunc = func(ctx context.Context, _ string, host string) ([]string, error) {
+	return net.DefaultResolver.LookupHost(ctx, host)
+}
+
+func proxyDialAddress(ctx context.Context, proxyType, address string, lookup hostLookupFunc) (string, error) {
+	proxyType = strings.ToLower(strings.TrimSpace(proxyType))
+	if proxyType != "socks5" {
+		return address, nil
+	}
+	host, port, err := net.SplitHostPort(address)
+	if err != nil {
+		return "", err
+	}
+	if net.ParseIP(host) != nil {
+		return address, nil
+	}
+	if lookup == nil {
+		lookup = defaultHostLookup
+	}
+	addrs, err := lookup(ctx, "ip", host)
+	if err != nil {
+		return "", err
+	}
+	if len(addrs) == 0 {
+		return "", fmt.Errorf("no ip address resolved for %s", host)
+	}
+	return net.JoinHostPort(addrs[0], port), nil
+}
+
+func proxyCacheKey(proxyCfg config.Proxy) string {
+	proxyCfg = config.NormalizeProxy(proxyCfg)
+	return strings.Join([]string{
+		proxyCfg.ID,
+		proxyCfg.Type,
+		strings.ToLower(proxyCfg.Host),
+		strconv.Itoa(proxyCfg.Port),
+		proxyCfg.Username,
+		proxyCfg.Password,
+	}, "|")
+}
+
+func proxyDialContext(proxyCfg config.Proxy) (trans.DialContextFunc, error) {
+	proxyCfg = config.NormalizeProxy(proxyCfg)
+	var authCfg *proxy.Auth
+	if proxyCfg.Username != "" || proxyCfg.Password != "" {
+		authCfg = &proxy.Auth{User: proxyCfg.Username, Password: proxyCfg.Password}
+	}
+	forward := &net.Dialer{Timeout: 15 * time.Second, KeepAlive: 30 * time.Second}
+	dialer, err := proxy.SOCKS5("tcp", net.JoinHostPort(proxyCfg.Host, strconv.Itoa(proxyCfg.Port)), authCfg, forward)
+	if err != nil {
+		return nil, err
+	}
+	return func(ctx context.Context, network, address string) (net.Conn, error) {
+		target, err := proxyDialAddress(ctx, proxyCfg.Type, address, defaultHostLookup)
+		if err != nil {
+			return nil, err
+		}
+		if ctxDialer, ok := dialer.(proxy.ContextDialer); ok {
+			return ctxDialer.DialContext(ctx, network, target)
+		}
+		return dialer.Dial(network, target)
+	}, nil
+}
+
+func (c *Client) defaultRequestClients() requestClients {
+	return requestClients{
+		regular:   c.regular,
+		stream:    c.stream,
+		fallback:  c.fallback,
+		fallbackS: c.fallbackS,
+	}
+}
+
+func (c *Client) resolveProxyForAccount(acc config.Account) (config.Proxy, bool) {
+	if c == nil || c.Store == nil {
+		return config.Proxy{}, false
+	}
+	proxyID := strings.TrimSpace(acc.ProxyID)
+	if proxyID == "" {
+		return config.Proxy{}, false
+	}
+	snap := c.Store.Snapshot()
+	for _, proxyCfg := range snap.Proxies {
+		proxyCfg = config.NormalizeProxy(proxyCfg)
+		if proxyCfg.ID == proxyID {
+			return proxyCfg, true
+		}
+	}
+	return config.Proxy{}, false
+}
+
+func (c *Client) requestClientsFromContext(ctx context.Context) requestClients {
+	if a, ok := auth.FromContext(ctx); ok {
+		return c.requestClientsForAccount(a.Account)
+	}
+	return c.defaultRequestClients()
+}
+
+func (c *Client) requestClientsForAuth(ctx context.Context, a *auth.RequestAuth) requestClients {
+	if a != nil {
+		return c.requestClientsForAccount(a.Account)
+	}
+	return c.requestClientsFromContext(ctx)
+}
+
+func (c *Client) requestClientsForAccount(acc config.Account) requestClients {
+	proxyCfg, ok := c.resolveProxyForAccount(acc)
+	if !ok {
+		return c.defaultRequestClients()
+	}
+
+	key := proxyCacheKey(proxyCfg)
+	c.proxyClientsMu.RLock()
+	cached, ok := c.proxyClients[key]
+	c.proxyClientsMu.RUnlock()
+	if ok {
+		return cached
+	}
+
+	dialContext, err := proxyDialContext(proxyCfg)
+	if err != nil {
+		config.Logger.Warn("[proxy] build dialer failed", "proxy_id", proxyCfg.ID, "error", err)
+		return c.defaultRequestClients()
+	}
+
+	bundle := requestClients{
+		regular:   trans.NewWithDialContext(60*time.Second, dialContext),
+		stream:    trans.NewWithDialContext(0, dialContext),
+		fallback:  trans.NewFallbackClient(60*time.Second, dialContext),
+		fallbackS: trans.NewFallbackClient(0, dialContext),
+	}
+
+	c.proxyClientsMu.Lock()
+	if c.proxyClients == nil {
+		c.proxyClients = make(map[string]requestClients)
+	}
+	c.proxyClients[key] = bundle
+	c.proxyClientsMu.Unlock()
+	return bundle
+}
+
+func applyProxyConnectivityHeaders(req *http.Request) {
+	if req == nil {
+		return
+	}
+	for key, value := range BaseHeaders {
+		key = strings.TrimSpace(key)
+		value = strings.TrimSpace(value)
+		if key == "" || value == "" {
+			continue
+		}
+		req.Header.Set(key, value)
+	}
+}
+
+func proxyConnectivityStatus(statusCode int) (bool, string) {
+	switch {
+	case statusCode >= 200 && statusCode < 300:
+		return true, fmt.Sprintf("代理可达，目标返回 HTTP %d", statusCode)
+	case statusCode >= 300 && statusCode < 500:
+		return true, fmt.Sprintf("代理可达，但目标返回 HTTP %d（可能是风控或挑战）", statusCode)
+	default:
+		return false, fmt.Sprintf("目标返回 HTTP %d", statusCode)
+	}
+}
+
+func TestProxyConnectivity(ctx context.Context, proxyCfg config.Proxy) map[string]any {
+	start := time.Now()
+	proxyCfg = config.NormalizeProxy(proxyCfg)
+	result := map[string]any{
+		"success":       false,
+		"proxy_id":      proxyCfg.ID,
+		"proxy_type":    proxyCfg.Type,
+		"response_time": 0,
+	}
+
+	if err := config.ValidateProxyConfig([]config.Proxy{proxyCfg}); err != nil {
+		result["message"] = "代理配置无效: " + err.Error()
+		return result
+	}
+	dialContext, err := proxyDialContext(proxyCfg)
+	if err != nil {
+		result["message"] = "代理拨号器初始化失败: " + err.Error()
+		return result
+	}
+
+	client := trans.NewFallbackClient(15*time.Second, dialContext)
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, proxyConnectivityTestURL, nil)
+	if err != nil {
+		result["message"] = err.Error()
+		return result
+	}
+	applyProxyConnectivityHeaders(req)
+
+	resp, err := client.Do(req)
+	result["response_time"] = int(time.Since(start).Milliseconds())
+	if err != nil {
+		result["message"] = err.Error()
+		return result
+	}
+	defer func() {
+		if closeErr := resp.Body.Close(); closeErr != nil {
+			config.Logger.Warn("[proxy] close response body failed", "proxy_id", proxyCfg.ID, "error", closeErr)
+		}
+	}()
+
+	result["status_code"] = resp.StatusCode
+	result["success"], result["message"] = proxyConnectivityStatus(resp.StatusCode)
+	return result
+}

internal/deepseek/proxy_test.go

@@ -0,0 +1,85 @@
+package deepseek
+
+import (
+	"context"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+func TestProxyDialAddressUsesLocalResolutionForSocks5(t *testing.T) {
+	ctx := context.Background()
+	resolved, err := proxyDialAddress(ctx, "socks5", "example.com:443", func(_ context.Context, network, host string) ([]string, error) {
+		if network != "ip" {
+			t.Fatalf("unexpected lookup network: %q", network)
+		}
+		if host != "example.com" {
+			t.Fatalf("unexpected lookup host: %q", host)
+		}
+		return []string{"203.0.113.10"}, nil
+	})
+	if err != nil {
+		t.Fatalf("proxyDialAddress returned error: %v", err)
+	}
+	if resolved != "203.0.113.10:443" {
+		t.Fatalf("expected locally resolved address, got %q", resolved)
+	}
+}
+
+func TestProxyDialAddressKeepsHostnameForSocks5h(t *testing.T) {
+	ctx := context.Background()
+	lookups := 0
+	resolved, err := proxyDialAddress(ctx, "socks5h", "example.com:443", func(_ context.Context, network, host string) ([]string, error) {
+		lookups++
+		return []string{"203.0.113.10"}, nil
+	})
+	if err != nil {
+		t.Fatalf("proxyDialAddress returned error: %v", err)
+	}
+	if resolved != "example.com:443" {
+		t.Fatalf("expected hostname preserved for remote DNS, got %q", resolved)
+	}
+	if lookups != 0 {
+		t.Fatalf("expected no local DNS lookup for socks5h, got %d", lookups)
+	}
+}
+
+func TestApplyProxyConnectivityHeadersUsesBaseHeaders(t *testing.T) {
+	req, err := http.NewRequest(http.MethodGet, "https://chat.deepseek.com/", nil)
+	if err != nil {
+		t.Fatalf("http.NewRequest returned error: %v", err)
+	}
+
+	applyProxyConnectivityHeaders(req)
+
+	for key, want := range BaseHeaders {
+		if got := req.Header.Get(key); got != want {
+			t.Fatalf("expected header %q=%q, got %q", key, want, got)
+		}
+	}
+}
+
+func TestProxyConnectivityStatus(t *testing.T) {
+	cases := []struct {
+		name       string
+		statusCode int
+		success    bool
+		wantText   string
+	}{
+		{name: "ok", statusCode: 200, success: true, wantText: "HTTP 200"},
+		{name: "challenge", statusCode: 403, success: true, wantText: "风控或挑战"},
+		{name: "upstream error", statusCode: 502, success: false, wantText: "HTTP 502"},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			success, message := proxyConnectivityStatus(tc.statusCode)
+			if success != tc.success {
+				t.Fatalf("expected success=%v, got %v", tc.success, success)
+			}
+			if message == "" || !strings.Contains(message, tc.wantText) {
+				t.Fatalf("expected message to contain %q, got %q", tc.wantText, message)
+			}
+		})
+	}
+}

internal/deepseek/transport/transport.go

@@ -15,21 +15,33 @@ type Doer interface {
 	Do(req *http.Request) (*http.Response, error)
 }
 
+type DialContextFunc func(ctx context.Context, network, addr string) (net.Conn, error)
+
 type Client struct {
 	http *http.Client
 }
 
 func New(timeout time.Duration) *Client {
+	return NewWithDialContext(timeout, nil)
+}
+
+func NewWithDialContext(timeout time.Duration, dialContext DialContextFunc) *Client {
+	useEnvProxy := dialContext == nil
+	if dialContext == nil {
+		dialContext = (&net.Dialer{Timeout: 15 * time.Second, KeepAlive: 30 * time.Second}).DialContext
+	}
 	base := &http.Transport{
-		Proxy:               http.ProxyFromEnvironment,
 		ForceAttemptHTTP2:   false,
 		MaxIdleConns:        200,
 		MaxIdleConnsPerHost: 100,
 		IdleConnTimeout:     90 * time.Second,
-		DialContext:         (&net.Dialer{Timeout: 15 * time.Second, KeepAlive: 30 * time.Second}).DialContext,
-		DialTLSContext:      safariTLSDialer(),
+		DialContext:         dialContext,
+		DialTLSContext:      safariTLSDialer(dialContext),
 		TLSClientConfig:     &tls.Config{MinVersion: tls.VersionTLS12},
 	}
+	if useEnvProxy {
+		base.Proxy = http.ProxyFromEnvironment
+	}
 	return &Client{http: &http.Client{Timeout: timeout, Transport: base}}
 }
 
@@ -37,10 +49,31 @@ func (c *Client) Do(req *http.Request) (*http.Response, error) {
 	return c.http.Do(req)
 }
 
-func safariTLSDialer() func(ctx context.Context, network, addr string) (net.Conn, error) {
-	var dialer net.Dialer
+func NewFallbackClient(timeout time.Duration, dialContext DialContextFunc) *http.Client {
+	useEnvProxy := dialContext == nil
+	if dialContext == nil {
+		dialContext = (&net.Dialer{Timeout: 15 * time.Second, KeepAlive: 30 * time.Second}).DialContext
+	}
+	base := &http.Transport{
+		ForceAttemptHTTP2:   false,
+		MaxIdleConns:        200,
+		MaxIdleConnsPerHost: 100,
+		IdleConnTimeout:     90 * time.Second,
+		DialContext:         dialContext,
+		TLSClientConfig:     &tls.Config{MinVersion: tls.VersionTLS12},
+	}
+	if useEnvProxy {
+		base.Proxy = http.ProxyFromEnvironment
+	}
+	return &http.Client{Timeout: timeout, Transport: base}
+}
+
+func safariTLSDialer(dialContext DialContextFunc) func(ctx context.Context, network, addr string) (net.Conn, error) {
+	if dialContext == nil {
+		dialContext = (&net.Dialer{Timeout: 15 * time.Second, KeepAlive: 30 * time.Second}).DialContext
+	}
 	return func(ctx context.Context, network, addr string) (net.Conn, error) {
-		plainConn, err := dialer.DialContext(ctx, network, addr)
+		plainConn, err := dialContext(ctx, network, addr)
 		if err != nil {
 			return nil, err
 		}

internal/js/chat-stream/sse_parse_impl.js

@@ -20,7 +20,9 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
     };
   }
 
-  const outputTokens = extractAccumulatedTokenUsage(chunk);
+  const usage = extractAccumulatedTokenUsage(chunk);
+  const promptTokens = usage.prompt;
+  const outputTokens = usage.output;
 
   if (Object.prototype.hasOwnProperty.call(chunk, 'error')) {
     return {
@@ -29,7 +31,8 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: true,
       contentFilter: false,
       errorMessage: formatErrorMessage(chunk.error),
-      outputTokens: 0,
+      promptTokens,
+      outputTokens,
       newType: currentType,
     };
   }
@@ -43,6 +46,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: true,
       contentFilter: true,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType: currentType,
     };
@@ -55,6 +59,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: false,
       contentFilter: false,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType: currentType,
     };
@@ -67,6 +72,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
         finished: true,
         contentFilter: false,
         errorMessage: '',
+        promptTokens,
         outputTokens,
         newType: currentType,
       };
@@ -77,6 +83,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: false,
       contentFilter: false,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType: currentType,
     };
@@ -89,6 +96,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: false,
       contentFilter: false,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType: currentType,
     };
@@ -157,6 +165,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
         finished: true,
         contentFilter: false,
         errorMessage: '',
+        promptTokens,
         outputTokens,
         newType,
       };
@@ -168,6 +177,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
         finished: false,
         contentFilter: false,
         errorMessage: '',
+        promptTokens,
         outputTokens,
         newType,
       };
@@ -182,6 +192,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: false,
       contentFilter: false,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType,
     };
@@ -196,6 +207,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
         finished: true,
         contentFilter: false,
         errorMessage: '',
+        promptTokens,
         outputTokens,
         newType,
       };
@@ -207,6 +219,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
       finished: false,
       contentFilter: false,
       errorMessage: '',
+      promptTokens,
       outputTokens,
       newType,
     };
@@ -242,6 +255,7 @@ function parseChunkForContent(chunk, thinkingEnabled, currentType, stripReferenc
     finished: false,
     contentFilter: false,
     errorMessage: '',
+    promptTokens,
     outputTokens,
     newType,
   };
@@ -429,56 +443,10 @@ function hasContentFilterStatusValue(v) {
 }
 
 function extractAccumulatedTokenUsage(chunk) {
-  return findAccumulatedTokenUsage(chunk);
-}
-
-function findAccumulatedTokenUsage(v) {
-  if (Array.isArray(v)) {
-    for (const item of v) {
-      const n = findAccumulatedTokenUsage(item);
-      if (n > 0) {
-        return n;
-      }
-    }
-    return 0;
-  }
-  if (!v || typeof v !== 'object') {
-    return 0;
-  }
-  const pathValue = asString(v.p);
-  if (pathValue && pathValue.toLowerCase().includes('accumulated_token_usage')) {
-    const n = toInt(v.v);
-    if (n > 0) {
-      return n;
-    }
-  }
-  const direct = toInt(v.accumulated_token_usage);
-  if (direct > 0) {
-    return direct;
-  }
-  for (const value of Object.values(v)) {
-    const n = findAccumulatedTokenUsage(value);
-    if (n > 0) {
-      return n;
-    }
-  }
-  return 0;
-}
-
-function toInt(v) {
-  if (typeof v === 'number' && Number.isFinite(v)) {
-    return Math.trunc(v);
-  }
-  if (typeof v === 'string' && v.trim() !== '') {
-    const n = Number(v);
-    if (Number.isFinite(n)) {
-      return Math.trunc(n);
-    }
-  }
-  if (typeof v !== 'number') {
-    return 0;
-  }
-  return Number.isFinite(v) ? Math.trunc(v) : 0;
+  // 临时策略：忽略上游 usage 字段（accumulated_token_usage / token_usage），
+  // 统一使用内部估算计数，避免上下文累计口径误差。
+  void chunk;
+  return { prompt: 0, output: 0 };
 }
 
 function formatErrorMessage(v) {

internal/js/chat-stream/token_usage.js

@@ -1,15 +1,17 @@
 'use strict';
 
-function buildUsage(prompt, thinking, output, outputTokens = 0) {
-  const promptTokens = estimateTokens(prompt);
+function buildUsage(prompt, thinking, output, outputTokens = 0, providedPromptTokens = 0) {
   const reasoningTokens = estimateTokens(thinking);
   const completionTokens = estimateTokens(output);
+
+  const finalPromptTokens = Number.isFinite(providedPromptTokens) && providedPromptTokens > 0 ? Math.trunc(providedPromptTokens) : estimateTokens(prompt);
+
   const overriddenCompletionTokens = Number.isFinite(outputTokens) && outputTokens > 0 ? Math.trunc(outputTokens) : 0;
   const finalCompletionTokens = overriddenCompletionTokens > 0 ? overriddenCompletionTokens : reasoningTokens + completionTokens;
   return {
-    prompt_tokens: promptTokens,
+    prompt_tokens: finalPromptTokens,
     completion_tokens: finalCompletionTokens,
-    total_tokens: promptTokens + finalCompletionTokens,
+    total_tokens: finalPromptTokens + finalCompletionTokens,
     completion_tokens_details: {
       reasoning_tokens: reasoningTokens,
     },

internal/js/chat-stream/vercel_stream_impl.js

@@ -125,7 +125,6 @@ async function handleVercelStream(req, res, rawBody, payload) {
     let currentType = thinkingEnabled ? 'thinking' : 'text';
     let thinkingText = '';
     let outputText = '';
-    let outputTokens = 0;
     const toolSieveEnabled = toolPolicy.toolSieveEnabled;
     const toolSieveState = createToolSieveState();
     let toolCallsEmitted = false;
@@ -178,7 +177,7 @@ async function handleVercelStream(req, res, rawBody, payload) {
         created,
         model,
         choices: [{ delta: {}, index: 0, finish_reason: reason }],
-        usage: buildUsage(finalPrompt, thinkingText, outputText, outputTokens),
+        usage: buildUsage(finalPrompt, thinkingText, outputText),
       });
       if (!res.writableEnded && !res.destroyed) {
         res.write('data: [DONE]\n\n');
@@ -227,9 +226,6 @@ async function handleVercelStream(req, res, rawBody, payload) {
           if (!parsed.parsed) {
             continue;
           }
-          if (parsed.outputTokens > 0) {
-            outputTokens = parsed.outputTokens;
-          }
           currentType = parsed.newType;
           if (parsed.errorMessage) {
             await finish('content_filter');