refactor: replace bufio.Scanner with bufio.Reader for SSE stream parsing and track emitted text to prevent redundant output blocks

perf(streaming): optimize TTFT and reduce buffering latency

.github/workflows/release-artifacts.yml

@@ -47,7 +47,6 @@ jobs:
 
       - name: Release Blocking Gates
         run: |
-          ./tests/scripts/check-stage6-manual-smoke.sh
           ./tests/scripts/check-refactor-line-gate.sh
           ./tests/scripts/run-unit-all.sh
 

.gitignore

@@ -29,6 +29,7 @@ yarn.lock
 pnpm-lock.yaml
 
 # Build artifacts
+dist/
 *.tsbuildinfo
 .cache/
 .parcel-cache/

API.en.md

@@ -33,6 +33,8 @@ Docs: [Overview](README.en.md) / [Architecture](docs/ARCHITECTURE.en.md) / [Depl
 | Health probes | `GET /healthz`, `GET /readyz` |
 | CORS | Enabled (uniformly covers `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, and `/admin/*`; echoes the browser `Origin` when present, otherwise `*`; default allow-list includes `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`, and also accepts third-party preflight-requested headers such as `x-stainless-*`; `/v1/chat/completions` on Vercel Node Runtime matches the same behavior; internal-only `X-Ds2-Internal-Token` remains blocked) |
 
+- All JSON request bodies must be valid UTF-8; malformed byte sequences are rejected on ingress with `400 invalid json`.
+
 ### 3.0 Adapter-Layer Notes
 
 - OpenAI / Claude / Gemini protocols are now mounted on one shared `chi` router tree assembled in `internal/server/router.go`.
@@ -81,7 +83,7 @@ Two header formats accepted:
 - Token is in `config.keys` → **Managed account mode**: DS2API auto-selects an account via rotation
 - Token is not in `config.keys` → **Direct token mode**: treated as a DeepSeek token directly
 
-**Optional header**: `X-Ds2-Target-Account: <email_or_mobile>` — Pin a specific managed account.
+**Optional header**: `X-Ds2-Target-Account: <email_or_mobile>` — Pin a specific managed account; if the target account does not exist or the managed-account queue is exhausted, the request returns `429`, and current responses do not include `Retry-After`. If the account exists but login/refresh fails, the request returns the underlying `401` or upstream error.
 Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=` as the caller credential source.
 
 ### Admin Endpoints (`/admin/*`)
@@ -109,6 +111,7 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 | GET | `/v1/responses/{response_id}` | Business | Query stored response (in-memory TTL) |
 | POST | `/v1/embeddings` | Business | OpenAI Embeddings API |
 | POST | `/v1/files` | Business | OpenAI Files upload (multipart/form-data) |
+| GET | `/v1/files/{file_id}` | Business | Retrieve uploaded file status |
 | GET | `/anthropic/v1/models` | None | Claude model list |
 | POST | `/anthropic/v1/messages` | Business | Claude messages |
 | POST | `/anthropic/v1/messages/count_tokens` | Business | Claude token counting |
@@ -165,6 +168,8 @@ Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=
 | PUT | `/admin/chat-history/settings` | Admin | Update conversation history retention limit |
 | GET | `/admin/version` | Admin | Check current version and latest Release |
 
+OpenAI `/v1/*` paths are canonical. For clients configured with the bare DS2API service URL, the same OpenAI handlers are also exposed through root shortcuts: `/models`, `/models/{id}`, `/chat/completions`, `/responses`, `/responses/{response_id}`, `/embeddings`, `/files`, and `/files/{file_id}`.
+
 ---
 
 ## Health Endpoints
@@ -196,11 +201,15 @@ No auth required. Returns the currently supported DeepSeek native model list.
   "object": "list",
   "data": [
     {"id": "deepseek-v4-flash", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-v4-flash-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-pro", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-v4-pro-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-flash-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-v4-flash-search-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-pro-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
+    {"id": "deepseek-v4-pro-search-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-vision", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-v4-vision-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-v4-vision-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
@@ -224,6 +233,8 @@ Built-in aliases come from `internal/config/models.go`; `config.model_aliases` c
 - Gemini: `gemini-2.5-pro`, `gemini-2.5-flash`, `gemini-pro-vision`
 - Other compatibility families: `llama-*`, `qwen-*`, `mistral-*`, and `command-*` fall back through family heuristics
 
+Current vision support resolves only to `deepseek-v4-vision` and does not expose a separate `vision-search` variant.
+
 Retired historical families such as `claude-1.*`, `claude-2.*`, `claude-instant-*`, and `gpt-3.5*` are explicitly rejected.
 
 ### `POST /v1/chat/completions`
@@ -297,7 +308,7 @@ data: [DONE]
 - When thinking is enabled, the stream may 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
+- 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. Failed/interrupted endings (for example `response.failed`) may not include `usage`
 
 #### Tool Calls
 
@@ -413,7 +424,7 @@ Business auth required. Returns OpenAI-compatible embeddings shape.
 | `model` | string | ✅ | Supports native models + alias mapping |
 | `input` | string/array | ✅ | Supports string, string array, token array |
 
-> Requires `embeddings.provider`. Current supported values: `mock` / `deterministic` / `builtin`. If missing/unsupported, returns standard error shape with HTTP 501.
+> Requires `embeddings.provider`. Current supported values: `mock` / `deterministic` / `builtin` (all three use the same local deterministic implementation). If missing/unsupported, returns standard error shape with HTTP 501.
 
 ### `POST /v1/files`
 
@@ -427,9 +438,13 @@ Business auth required. OpenAI Files-compatible upload endpoint; currently only
 Constraints and behavior:
 
 - `Content-Type` must be `multipart/form-data` (otherwise `400`).
-- Total request size limit is `100 MiB` (over-limit returns `413`).
+- Total request size limit is **100 MiB** (over-limit returns `413`).
 - Success returns an OpenAI `file` object (`id/object/bytes/filename/purpose/status`, etc.) and includes `account_id` for source-account tracing.
 
+### `GET /v1/files/{file_id}`
+
+Business auth required. Retrieves the current DeepSeek upload status for a file and returns an OpenAI `file` object. Returns `404` when no matching file is found.
+
 ---
 
 ## Claude-Compatible API
@@ -481,6 +496,13 @@ anthropic-version: 2023-06-01
 | `stream` | boolean | ❌ | Default `false` |
 | `system` | string | ❌ | Optional system prompt |
 | `tools` | array | ❌ | Claude tool schema |
+| `thinking` | object | ❌ | Anthropic thinking config; translated into downstream reasoning control, and ignored by `-nothinking` models |
+| `temperature` | number | ❌ | Passed through to the downstream bridge; if `temperature` and `top_p` are both present, `temperature` wins |
+| `top_p` | number | ❌ | Passed through when `temperature` is absent |
+| `stop_sequences` | array | ❌ | Passed through as downstream stop sequences |
+| `tool_choice` | string/object | ❌ | Supports `auto` / `none` / `required` / `{"type":"function","name":"..."}` and is translated to downstream tool choice |
+
+> Note: `thinking`, `temperature`, `top_p`, `stop_sequences`, and `tool_choice` are translated through the compatibility bridge. Final behavior still depends on the selected model and upstream support. When both `temperature` and `top_p` are present, `temperature` takes precedence.
 
 #### Non-Stream Response
 
@@ -1209,7 +1231,7 @@ Clients should handle HTTP status code plus `error` / `detail` fields.
 | Code | Meaning |
 | --- | --- |
 | `401` | Authentication failed (invalid key/token, or expired admin JWT) |
-| `429` | Too many requests (exceeded inflight + queue capacity) |
+| `429` | Too many requests (exceeded inflight + queue capacity; current responses do not include `Retry-After`) |
 | `503` | Model unavailable or upstream error |
 
 ---

API.md

@@ -33,6 +33,8 @@
 | 健康检查 | `GET /healthz`、`GET /readyz` |
 | CORS | 已启用（统一覆盖 `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/admin/*`；浏览器有 `Origin` 时回显该 Origin，否则为 `*`；默认允许 `Content-Type`, `Authorization`, `X-API-Key`, `X-Ds2-Target-Account`, `X-Ds2-Source`, `X-Vercel-Protection-Bypass`, `X-Goog-Api-Key`, `Anthropic-Version`, `Anthropic-Beta`，并会放行预检里声明的第三方请求头，如 `x-stainless-*`；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同行为；内部专用头 `X-Ds2-Internal-Token` 仍被拦截） |
 
+- 所有 JSON 请求体都必须是合法 UTF-8；非法字节序列会在入站阶段被拒绝为 `400 invalid json`。
+
 ### 3.0 接口适配层说明
 
 - OpenAI / Claude / Gemini 三套协议已统一挂在同一 `chi` 路由树上，由 `internal/server/router.go` 负责装配。
@@ -81,7 +83,7 @@ Vercel 一键部署可先只填 `DS2API_ADMIN_KEY`，部署后在 `/admin` 导
 - token 在 `config.keys` 中 → **托管账号模式**，自动轮询选择账号
 - token 不在 `config.keys` 中 → **直通 token 模式**，直接作为 DeepSeek token 使用
 
-**可选请求头**：`X-Ds2-Target-Account: <email_or_mobile>` — 指定使用某个托管账号。
+**可选请求头**：`X-Ds2-Target-Account: <email_or_mobile>` — 指定使用某个托管账号；如果目标账号不存在，或管理账号队列已耗尽，相关业务请求会返回 `429`，当前不会附带 `Retry-After` 头。若账号存在但登录/刷新失败，则返回对应的 `401` 或上游错误。
 Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=` 作为凭据来源。
 
 ### Admin 接口（`/admin/*`）
@@ -109,6 +111,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | GET | `/v1/responses/{response_id}` | 业务 | 查询已生成 response（内存 TTL） |
 | POST | `/v1/embeddings` | 业务 | OpenAI Embeddings 接口 |
 | POST | `/v1/files` | 业务 | OpenAI Files 上传（multipart/form-data） |
+| GET | `/v1/files/{file_id}` | 业务 | 查询已上传文件状态 |
 | GET | `/anthropic/v1/models` | 无 | Claude 模型列表 |
 | POST | `/anthropic/v1/messages` | 业务 | Claude 消息接口 |
 | POST | `/anthropic/v1/messages/count_tokens` | 业务 | Claude token 计数 |
@@ -165,6 +168,8 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 | PUT | `/admin/chat-history/settings` | Admin | 更新对话记录保留条数 |
 | GET | `/admin/version` | Admin | 查询当前版本与最新 Release |
 
+OpenAI `/v1/*` 仍是规范路径。对于只配置 DS2API 根地址的客户端，同一套 OpenAI handler 也通过根路径快捷路由暴露：`/models`、`/models/{id}`、`/chat/completions`、`/responses`、`/responses/{response_id}`、`/embeddings`、`/files`、`/files/{file_id}`。
+
 ---
 
 ## 健康检查
@@ -204,9 +209,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
     {"id": "deepseek-v4-pro-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-pro-search-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
     {"id": "deepseek-v4-vision", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-v4-vision-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-v4-vision-search", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []},
-    {"id": "deepseek-v4-vision-search-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
+    {"id": "deepseek-v4-vision-nothinking", "object": "model", "created": 1677610602, "owned_by": "deepseek", "permission": []}
   ]
 }
 ```
@@ -232,6 +235,7 @@ Gemini 兼容客户端还可以使用 `x-goog-api-key`、`?key=` 或 `?api_key=`
 - 其他兼容族：`llama-*`、`qwen-*`、`mistral-*`、`command-*` 会按家族启发式回退
 
 上述 alias 若在请求名后追加 `-nothinking` 后缀，也会映射到对应的强制关闭 thinking 版本。
+当前视觉能力仅对应 `deepseek-v4-vision` / `deepseek-v4-vision-nothinking`，不会解析出独立的 `vision-search` 变体。
 
 退役历史模型（如 `claude-1.*`、`claude-2.*`、`claude-instant-*`、`gpt-3.5*`）会被显式拒绝。
 
@@ -306,7 +310,7 @@ data: [DONE]
 - 开启 thinking 时会输出 `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
+- 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。失败/中断型结束（例如 `response.failed`）可能不会携带 `usage`
 
 #### Tool Calls
 
@@ -423,7 +427,7 @@ data: [DONE]
 | `model` | string | ✅ | 支持原生模型 + alias 自动映射 |
 | `input` | string/array | ✅ | 支持字符串、字符串数组、token 数组 |
 
-> 需配置 `embeddings.provider`。当前支持：`mock` / `deterministic` / `builtin`。未配置或不支持时返回标准错误结构（HTTP 501）。
+> 需配置 `embeddings.provider`。当前支持：`mock` / `deterministic` / `builtin`（三者都走同一套本地确定性实现）。未配置或不支持时返回标准错误结构（HTTP 501）。
 
 ### `POST /v1/files`
 
@@ -437,9 +441,13 @@ data: [DONE]
 约束与行为：
 
 - 请求必须为 `multipart/form-data`，否则返回 `400`。
-- 请求体总大小上限 `100 MiB`（超限返回 `413`）。
+- 请求体总大小上限 **100 MiB**（超限返回 `413`）。
 - 成功返回 OpenAI `file` 对象（`id/object/bytes/filename/purpose/status` 等字段），并附带 `account_id` 便于定位来源账号。
 
+### `GET /v1/files/{file_id}`
+
+需要业务鉴权。查询 DeepSeek 上传文件的当前状态，并返回 OpenAI `file` 对象；未找到匹配文件时返回 `404`。
+
 ---
 
 ## Claude 兼容接口
@@ -494,6 +502,13 @@ anthropic-version: 2023-06-01
 | `stream` | boolean | ❌ | 默认 `false` |
 | `system` | string | ❌ | 可选系统提示 |
 | `tools` | array | ❌ | Claude tool 定义 |
+| `thinking` | object | ❌ | Anthropic thinking 配置；会转译为下游 reasoning 控制，`-nothinking` 模型会忽略 |
+| `temperature` | number | ❌ | 透传到下游；若同时提供 `top_p`，以 `temperature` 为准 |
+| `top_p` | number | ❌ | 当未提供 `temperature` 时透传到下游 |
+| `stop_sequences` | array | ❌ | 透传到下游停用序列 |
+| `tool_choice` | string/object | ❌ | 支持 `auto` / `none` / `required` / `{"type":"function","name":"..."}`，并会转译为下游工具选择 |
+
+> 说明：上述 `thinking`、`temperature`、`top_p`、`stop_sequences`、`tool_choice` 都会走兼容层转译；最终是否生效仍取决于当前模型和上游能力。`temperature` 与 `top_p` 同时存在时，`temperature` 优先。
 
 #### 非流式响应
 
@@ -1225,7 +1240,7 @@ Gemini 路由使用 Google 风格错误结构：
 | 状态码 | 说明 |
 | --- | --- |
 | `401` | 鉴权失败（key/token 无效，或 Admin JWT 过期） |
-| `429` | 请求过多（超出并发上限 + 等待队列） |
+| `429` | 请求过多（超出并发上限 + 等待队列；当前不附带 `Retry-After` 头） |
 | `503` | 模型不可用或上游服务异常 |
 
 ---

Dockerfile

@@ -29,7 +29,7 @@ WORKDIR /app
 RUN apt-get update \
     && apt-get install -y --no-install-recommends ca-certificates \
     && groupadd -r ds2api && useradd -r -g ds2api -d /app -s /sbin/nologin ds2api \
-    && mkdir -p /app/data && chown -R ds2api:ds2api /app \
+    && mkdir -p /app/data /data && chown -R ds2api:ds2api /app /data \
     && rm -rf /var/lib/apt/lists/*
 COPY --from=busybox-tools /bin/busybox /usr/local/bin/busybox
 EXPOSE 5001

README.MD

@@ -17,7 +17,7 @@
 
 语言 / Language: [中文](README.MD) | [English](README.en.md)
 
-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`）。
+将 DeepSeek Web 对话能力转换为 OpenAI、Claude 与 Gemini 兼容 API。核心后端以 **Go** 实现，Vercel 流式桥接额外使用少量 Node Runtime，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)
 
@@ -119,7 +119,7 @@ flowchart LR
 
 | 能力 | 说明 |
 | --- | --- |
-| OpenAI 兼容 | `GET /v1/models`、`GET /v1/models/{id}`、`POST /v1/chat/completions`、`POST /v1/responses`、`GET /v1/responses/{response_id}`、`POST /v1/embeddings`、`POST /v1/files` |
+| OpenAI 兼容 | `GET /v1/models`、`GET /v1/models/{id}`、`POST /v1/chat/completions`、`POST /v1/responses`、`GET /v1/responses/{response_id}`、`POST /v1/embeddings`、`POST /v1/files`、`GET /v1/files/{file_id}` |
 | Claude 兼容 | `GET /anthropic/v1/models`、`POST /anthropic/v1/messages`、`POST /anthropic/v1/messages/count_tokens`（及快捷路径 `/v1/messages`、`/messages`） |
 | Gemini 兼容 | `POST /v1beta/models/{model}:generateContent`、`POST /v1beta/models/{model}:streamGenerateContent`（及 `/v1/models/{model}:*` 路径） |
 | 统一 CORS 兼容 | `/v1/*`、`/anthropic/*`、`/v1beta/models/*`、`/admin/*` 统一走同一套 CORS 策略；Vercel 上 `/v1/chat/completions` 的 Node Runtime 也对齐相同放行规则，尽量减少第三方预检请求头限制 |
@@ -131,6 +131,8 @@ flowchart LR
 | WebUI 管理台 | `/admin` 单页应用（中英文双语、深色模式，支持查看服务器端对话记录） |
 | 运维探针 | `GET /healthz`（存活）、`GET /readyz`（就绪） |
 
+OpenAI `/v1/*` 仍是推荐的规范路径；同时支持 `/models`、`/chat/completions`、`/responses`、`/embeddings`、`/files`、`/files/{file_id}` 等根路径快捷路由，方便只配置 DS2API 根地址的第三方客户端。
+
 ## 平台兼容矩阵
 
 | 级别 | 平台 | 当前状态 |
@@ -158,10 +160,9 @@ flowchart LR
 | expert | `deepseek-v4-pro-search-nothinking` | 永久关闭，不受请求参数影响 | ✅ |
 | vision | `deepseek-v4-vision` | 默认开启，可由请求参数控制 | ❌ |
 | vision | `deepseek-v4-vision-nothinking` | 永久关闭，不受请求参数影响 | ❌ |
-| vision | `deepseek-v4-vision-search` | 默认开启，可由请求参数控制 | ✅ |
-| vision | `deepseek-v4-vision-search-nothinking` | 永久关闭，不受请求参数影响 | ✅ |
 
 除原生模型外，也支持常见 alias 输入（如 `gpt-4.1`、`gpt-5`、`gpt-5-codex`、`o3`、`claude-*`、`gemini-*` 等），但 `/v1/models` 返回的是规范化后的 DeepSeek 原生模型 ID。若 alias 名本身追加 `-nothinking` 后缀，也会映射到对应的强制关思考模型。完整 alias 行为以 [API.md](API.md#模型-alias-解析策略) 和 `config.example.json` 为准。
+当前上游视觉模型只暴露 `vision` 通道，不提供独立的联网搜索视觉变体。
 
 ### Claude 接口（`GET /anthropic/v1/models`）
 
@@ -246,6 +247,7 @@ 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）。
 同时默认把 `./config.json` 挂载到容器 `/data/config.json`，并设置 `DS2API_CONFIG_PATH=/data/config.json`，用于避免 `/app` 只读导致运行时 token 持久化失败。
+镜像会预创建 `/data` 并授权给非 root 的 `ds2api` 用户；如果使用单文件 bind mount，请确保宿主机 `config.json` 对容器用户可读写，例如 `chmod 644 config.json`。
 
 更新镜像：`docker-compose up -d --build`
 
@@ -255,6 +257,10 @@ docker-compose logs -f
 2. 部署完成后访问 `/admin`，使用 Zeabur 环境变量/模板指引中的 `DS2API_ADMIN_KEY` 登录。
 3. 在管理台导入/编辑配置（会写入并持久化到 `/data/config.json`）。
 
+Zeabur 首次空卷启动时可以没有 `/data/config.json`；DS2API 会先使用空的文件模式配置启动，并在管理台首次保存时创建该文件。
+
+不依赖模板手动部署时，在 Zeabur 中选择 GitHub 仓库服务，Root Directory 保持 `/`，使用仓库根目录 `Dockerfile` 构建；添加持久卷 `/data`，设置 `PORT=5001`、`DS2API_ADMIN_KEY=你的强密钥`、`DS2API_CONFIG_PATH=/data/config.json`，然后暴露 HTTP 端口 `5001`。更完整步骤见 [docs/DEPLOY.md](docs/DEPLOY.md#不使用模板手动部署)。
+
 说明：Zeabur 使用仓库内 `Dockerfile` 直接构建时，不需要额外传入 `BUILD_VERSION`；镜像会优先读取该构建参数，未提供时自动回退到仓库根目录的 `VERSION` 文件。
 
 ### 方式三：Vercel 部署
@@ -283,7 +289,7 @@ base64 < config.json | tr -d '\n'
 
 ### 方式四：本地源码运行
 
-**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）
+**前置要求**：Go 1.26+，Node.js `20.19+` 或 `22.12+`（仅在需要构建 WebUI 时）；同时确保 `npm` 可用，建议 `npm 10+`
 
 ```bash
 # 1. 克隆仓库
@@ -302,7 +308,7 @@ go run ./cmd/ds2api
 
 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`
+> **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 和 npm）。你也可以手动构建：`./scripts/build-webui.sh`
 
 ## 配置说明
 
@@ -316,7 +322,7 @@ go run ./cmd/ds2api
 - `runtime`：账号并发、队列与 token 刷新策略，可通过 Admin Settings 热更新。
 - `auto_delete.mode`：请求结束后的远端会话清理策略，支持 `none` / `single` / `all`。
 - `history_split`：旧轮次拆分字段，已废弃并忽略，仅保留兼容旧配置。
-- `current_input_file`：唯一生效的独立拆分策略；默认开启且阈值为 `0`，触发时将完整上下文合并上传为隐藏上下文文件。
+- `current_input_file`：唯一生效的独立拆分策略；默认开启且阈值为 `0`，触发时将完整上下文合并上传为 `DS2API_HISTORY.txt` 上下文文件。
 - 如果关闭 `current_input_file`，请求会直接透传，不上传拆分上下文文件。
 - `thinking_injection`：默认开启；在最新 user 消息末尾追加思考增强提示词，提高高强度推理与工具调用前的思考稳定性；`prompt` 留空时使用内置默认提示词。
 
@@ -332,6 +338,7 @@ go run ./cmd/ds2api
 | **直通 token 模式** | 传入 token 不在 `config.keys` 中时，直接作为 DeepSeek token 使用 |
 
 可选请求头 `X-Ds2-Target-Account`：指定使用某个托管账号（值为 email 或 mobile）。
+如果指定账号不存在，或者当前管理账号队列已满，请求会返回 `429`；当前 `429` 不附带 `Retry-After` 头。若账号存在但登录/刷新失败，则返回对应的鉴权错误。
 Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `?key=` / `?api_key=` 作为调用方凭据。
 
 ## 并发模型
@@ -344,7 +351,7 @@ Gemini 路由还可以使用 `x-goog-api-key`，或在没有认证头时使用 `
 ```
 
 - 当 in-flight 槽位满时，请求进入等待队列，**不会立即 429**
-- 超出总承载上限后才返回 `429 Too Many Requests`
+- 超出总承载上限后才返回 `429 Too Many Requests`，当前响应不附带 `Retry-After`
 - `GET /admin/queue/status` 返回实时并发状态
 
 ## Tool Call 适配
@@ -422,10 +429,10 @@ npm run build --prefix webui
 
 工作流文件：`.github/workflows/release-artifacts.yml`
 
-- **触发条件**：仅在 GitHub Release `published` 时触发（普通 push 不会触发）
-- **构建产物**：多平台二进制包（`linux/amd64`、`linux/arm64`、`linux/armv7`、`darwin/amd64`、`darwin/arm64`、`windows/amd64`、`windows/arm64`）+ `sha256sums.txt`
+- **触发条件**：默认仅在 GitHub Release `published` 时自动触发；也支持在 Actions 页面手动 `workflow_dispatch`，并填写 `release_tag` 复跑/补发
+- **构建产物**：多平台二进制包（`linux/amd64`、`linux/arm64`、`linux/armv7`、`darwin/amd64`、`darwin/arm64`、`windows/amd64`、`windows/arm64`）、Linux Docker 镜像导出包 + `sha256sums.txt`
 - **容器镜像发布**：仅推送到 GHCR（`ghcr.io/cjackhwang/ds2api`）
-- **每个压缩包包含**：`ds2api` 可执行文件、`static/admin`、WASM 文件（同时支持内置 fallback）、`config.example.json` 配置示例、README、LICENSE
+- **每个二进制压缩包包含**：`ds2api` 可执行文件、`static/admin`、`config.example.json`、`.env.example`、`README.MD`、`README.en.md`、`LICENSE`
 
 ## 免责声明
 

README.en.md

@@ -16,7 +16,7 @@
 
 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).
+DS2API converts DeepSeek Web chat capability into OpenAI-compatible, Claude-compatible, and Gemini-compatible APIs. The core backend is Go-based, with a small Node Runtime bridge used for Vercel streaming, and the React WebUI admin panel lives 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)
 
@@ -116,7 +116,7 @@ For the full module-by-module architecture and directory responsibilities, see [
 
 | Capability | Details |
 | --- | --- |
-| OpenAI compatible | `GET /v1/models`, `GET /v1/models/{id}`, `POST /v1/chat/completions`, `POST /v1/responses`, `GET /v1/responses/{response_id}`, `POST /v1/embeddings`, `POST /v1/files` |
+| OpenAI compatible | `GET /v1/models`, `GET /v1/models/{id}`, `POST /v1/chat/completions`, `POST /v1/responses`, `GET /v1/responses/{response_id}`, `POST /v1/embeddings`, `POST /v1/files`, `GET /v1/files/{file_id}` |
 | Claude compatible | `GET /anthropic/v1/models`, `POST /anthropic/v1/messages`, `POST /anthropic/v1/messages/count_tokens` (plus shortcut paths `/v1/messages`, `/messages`) |
 | Gemini compatible | `POST /v1beta/models/{model}:generateContent`, `POST /v1beta/models/{model}:streamGenerateContent` (plus `/v1/models/{model}:*` paths) |
 | Unified CORS compatibility | `/v1/*`, `/anthropic/*`, `/v1beta/models/*`, and `/admin/*` share one CORS policy; on Vercel, the Node Runtime for `/v1/chat/completions` mirrors the same relaxed preflight behavior for third-party clients |
@@ -128,6 +128,8 @@ For the full module-by-module architecture and directory responsibilities, see [
 | WebUI Admin Panel | SPA at `/admin` (bilingual Chinese/English, dark mode, with server-side conversation history) |
 | Health Probes | `GET /healthz` (liveness), `GET /readyz` (readiness) |
 
+OpenAI `/v1/*` routes remain canonical, and DS2API also accepts root shortcuts such as `/models`, `/chat/completions`, `/responses`, `/embeddings`, `/files`, and `/files/{file_id}` for clients configured with the bare service URL.
+
 ## Platform Compatibility Matrix
 
 | Tier | Platform | Status |
@@ -150,9 +152,9 @@ For the full module-by-module architecture and directory responsibilities, see [
 | default | `deepseek-v4-flash-search` | enabled by default, request-controlled | ✅ |
 | expert | `deepseek-v4-pro-search` | enabled by default, request-controlled | ✅ |
 | vision | `deepseek-v4-vision` | enabled by default, request-controlled | ❌ |
-| vision | `deepseek-v4-vision-search` | enabled by default, request-controlled | ✅ |
 
 Besides native IDs, DS2API also accepts common aliases as input (for example `gpt-4.1`, `gpt-5`, `gpt-5-codex`, `o3`, `claude-*`, `gemini-*`), but `/v1/models` returns normalized DeepSeek native model IDs. The complete alias behavior is documented in [API.en.md](API.en.md#model-alias-resolution) and `config.example.json`.
+Current upstream vision support exposes only the `vision` lane and does not provide a separate search-enabled vision variant.
 
 ### Claude Endpoint (`GET /anthropic/v1/models`)
 
@@ -243,6 +245,10 @@ Rebuild after updates: `docker-compose up -d --build`
 2. After deployment, open `/admin` and login with `DS2API_ADMIN_KEY` shown in Zeabur env/template instructions.
 3. Import / edit config in Admin UI (it will be written and persisted to `/data/config.json`).
 
+Fresh Zeabur volumes can start without `/data/config.json`; DS2API will boot with an empty file-backed config and create the file on the first Admin UI save.
+
+For manual deployment without the template, create a Zeabur GitHub service, keep Root Directory as `/`, build with the repo-root `Dockerfile`, mount a persistent volume at `/data`, set `PORT=5001`, `DS2API_ADMIN_KEY=your-strong-secret`, and `DS2API_CONFIG_PATH=/data/config.json`, then expose HTTP port `5001`. See [docs/DEPLOY.en.md](docs/DEPLOY.en.md#manual-deployment-without-the-template) for the full guide.
+
 Note: when Zeabur builds directly from the repo `Dockerfile`, you do not need to pass `BUILD_VERSION`. The image prefers that build arg when provided, and automatically falls back to the repo-root `VERSION` file when it is absent.
 
 ### Option 3: Vercel
@@ -304,7 +310,7 @@ Common fields:
 - `runtime`: account concurrency, queueing, and token refresh behavior, hot-reloadable via Admin Settings.
 - `auto_delete.mode`: remote session cleanup after each request, supporting `none` / `single` / `all`.
 - `history_split`: legacy multi-turn history split field, now ignored and kept only for backward-compatible config loading.
-- `current_input_file`: the only active split mode; it is enabled by default and uploads the full context as a hidden context file once the character threshold is reached.
+- `current_input_file`: the only active split mode; it is enabled by default and uploads the full context as a `DS2API_HISTORY.txt` context file once the character threshold is reached.
 - If you turn off `current_input_file`, requests pass through directly without uploading any split context file.
 
 For the full environment variable list, see [docs/DEPLOY.en.md](docs/DEPLOY.en.md). For auth behavior, see [API.en.md](API.en.md#authentication).
@@ -407,10 +413,10 @@ npm run build --prefix webui
 
 Workflow: `.github/workflows/release-artifacts.yml`
 
-- **Trigger**: only on GitHub Release `published` (normal pushes do not trigger builds)
-- **Outputs**: multi-platform archives (`linux/amd64`, `linux/arm64`, `linux/armv7`, `darwin/amd64`, `darwin/arm64`, `windows/amd64`, `windows/arm64`) + `sha256sums.txt`
+- **Trigger**: by default only on GitHub Release `published`; you can also run it manually via `workflow_dispatch` and pass `release_tag` to rerun / backfill
+- **Outputs**: multi-platform binary archives (`linux/amd64`, `linux/arm64`, `linux/armv7`, `darwin/amd64`, `darwin/arm64`, `windows/amd64`, `windows/arm64`), Linux Docker image export tarballs, and `sha256sums.txt`
 - **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
-- **Each archive includes**: `ds2api` executable, `static/admin`, WASM file (with embedded fallback support), `config.example.json`-based config template, README, LICENSE
+- **Each binary archive includes**: the `ds2api` executable, `static/admin`, `config.example.json`, `.env.example`, `README.MD`, `README.en.md`, and `LICENSE`
 
 ## Disclaimer
 

VERSION

@@ -1 +1 @@
-4.1.3
+4.4.0

docs/ARCHITECTURE.en.md

@@ -25,6 +25,8 @@ ds2api/
 │   ├── chathistory/                      # Server-side conversation history storage/query
 │   ├── claudeconv/                       # Claude message conversion helpers
 │   ├── compat/                           # Compatibility and regression helpers
+│   ├── assistantturn/                    # Upstream output to canonical assistant turn / stream event semantics
+│   ├── completionruntime/                # Shared Go DeepSeek completion startup, non-stream collection, and retry
 │   ├── config/                           # Config loading/validation/hot reload
 │   ├── deepseek/                         # DeepSeek upstream client/protocol/transport
 │   │   ├── client/                       # Login/session/completion/upload/delete calls
@@ -171,6 +173,8 @@ flowchart LR
 - `internal/httpapi/openai/*`: OpenAI HTTP surface split into chat, responses, files, embeddings, history, and shared packages; chat/responses share the promptcompat, stream, and toolcall semantics.
 - `internal/httpapi/{claude,gemini}`: protocol wrappers that normalize into the same prompt compatibility semantics without duplicating upstream execution.
 - `internal/promptcompat`: compatibility core for turning OpenAI/Claude/Gemini requests into DeepSeek web-chat plain-text context.
+- `internal/assistantturn`: Go output-side canonical semantics, converting DeepSeek SSE collection results and stream finalization state into assistant turns and centralizing thinking, tool call, citation, usage, stop/error behavior.
+- `internal/completionruntime`: shared Go completion execution helpers for DeepSeek session/PoW/call startup, non-stream collection, and empty-output retry; streaming paths use it to start upstream requests, continue to use `internal/stream` for real-time consumption, and use `assistantturn` during finalization.
 - `internal/translatorcliproxy`: structure translation between Claude/Gemini and OpenAI.
 - `internal/deepseek/{client,protocol,transport}`: upstream requests, sessions, PoW adaptation, protocol constants, and transport details.
 - `internal/js/chat-stream` + `api/chat-stream.js`: Vercel Node streaming bridge; Go prepare/release owns auth, account lease, and completion payload assembly, while Node relays real-time SSE with Go-aligned finalization and tool sieve semantics.

docs/ARCHITECTURE.md

@@ -25,6 +25,8 @@ ds2api/
 │   ├── chathistory/                      # 服务器端对话记录存储与查询
 │   ├── claudeconv/                       # Claude 消息格式转换工具
 │   ├── compat/                           # 兼容性辅助与回归支持
+│   ├── assistantturn/                    # 上游输出到统一 assistant turn / stream event 的语义层
+│   ├── completionruntime/                # Go 主路径共享 DeepSeek completion 启动、非流式收集与 retry
 │   ├── config/                           # 配置加载、校验、热更新
 │   ├── deepseek/                         # DeepSeek 上游 client/protocol/transport
 │   │   ├── client/                       # 登录、会话、completion、上传/删除等上游调用
@@ -171,6 +173,8 @@ flowchart LR
 - `internal/httpapi/openai/*`：OpenAI HTTP surface，按 chat、responses、files、embeddings、history、shared 拆分；chat/responses 共享 promptcompat、stream、toolcall 等核心语义。
 - `internal/httpapi/{claude,gemini}`：协议输入输出适配，归一到同一套 prompt compatibility 语义，不重复实现上游调用逻辑。
 - `internal/promptcompat`：OpenAI/Claude/Gemini 请求到 DeepSeek 网页纯文本上下文的兼容内核。
+- `internal/assistantturn`：Go 输出侧统一语义层，把 DeepSeek SSE 收集结果和流式收尾状态归一成 assistant turn，集中处理 thinking、tool call、citation、usage、stop/error 语义。
+- `internal/completionruntime`：Go surface 共享的 completion 执行辅助，负责 DeepSeek session/PoW/call 启动、非流式 collect 和 empty-output retry；流式路径复用它启动上游请求，继续用 `internal/stream` 做实时消费，并在最终收尾阶段接入 `assistantturn`。
 - `internal/translatorcliproxy`：Claude/Gemini 与 OpenAI 结构互转。
 - `internal/deepseek/{client,protocol,transport}`：上游请求、会话、PoW 适配、协议常量与传输层。
 - `internal/js/chat-stream` + `api/chat-stream.js`：Vercel Node 流式桥；Go prepare/release 管理鉴权、账号租约和 completion payload，Node 侧负责实时 SSE 转发并保持 Go 对齐的终结态和 tool sieve 语义。

docs/CONTRIBUTING.en.md

@@ -36,7 +36,7 @@ go run ./cmd/ds2api
 cd webui
 
 # 2. Install dependencies
-npm install
+npm ci
 
 # 3. Start dev server (hot reload)
 npm run dev

docs/CONTRIBUTING.md

@@ -36,7 +36,7 @@ go run ./cmd/ds2api
 cd webui
 
 # 2. 安装依赖
-npm install
+npm ci
 
 # 3. 启动开发服务器（热更新）
 npm run dev

docs/DEPLOY.en.md

@@ -64,8 +64,8 @@ Use `config.json` as the single source of truth:
 
 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`
+- **Trigger**: by default only on Release `published`; you can also run it manually via `workflow_dispatch` and pass `release_tag` to rerun / backfill
+- **Outputs**: multi-platform binary archives, Linux Docker image export tarballs, and `sha256sums.txt`
 - **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
 
 | Platform | Architecture | Format |
@@ -131,6 +131,7 @@ docker-compose logs -f
 
 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).
 The compose template also defaults to `DS2API_CONFIG_PATH=/data/config.json` with `./config.json:/data/config.json` mounted, so deployments avoid read-only `/app` persistence issues by default.
+The image pre-creates `/data` and grants it to the non-root `ds2api` user. If you bind-mount a single host file, make sure `config.json` is readable/writable by the container user, for example with `chmod 644 config.json`; otherwise Linux UID/GID mismatches can still cause `open /data/config.json: permission denied`.
 Compatibility note: when `DS2API_CONFIG_PATH` is unset and runtime base dir is `/app`, newer versions prefer `/data/config.json`; if that file is missing but legacy `/app/config.json` exists, DS2API automatically falls back to the legacy path to avoid post-upgrade config loss.
 
 If you want a pinned version instead of `latest`, you can also pull a specific tag directly:
@@ -196,7 +197,7 @@ This repo includes a `zeabur.yaml` template for one-click deployment on Zeabur:
 Notes:
 
 - **Port**: DS2API listens on `5001` by default; the template sets `PORT=5001`.
-- **Persistent config**: the template mounts `/data` and sets `DS2API_CONFIG_PATH=/data/config.json`. After importing config in Admin UI, it will be written and persisted to this path.
+- **Persistent config**: the template mounts `/data` and sets `DS2API_CONFIG_PATH=/data/config.json`. On a fresh volume, DS2API starts with an empty file-backed config; after importing config in Admin UI, it will be written and persisted to this path.
 - **`open /app/config.json: permission denied`**: this means the instance is trying to persist runtime tokens to a read-only path (commonly `/app` inside the image).  
   Recommended handling:
   1. Set a writable path explicitly: `DS2API_CONFIG_PATH=/data/config.json` (and mount a persistent volume at `/data`);
@@ -205,6 +206,37 @@ Notes:
 - **Build version**: Zeabur / regular `docker build` does not require `BUILD_VERSION` by default. The image prefers that build arg when provided, and automatically falls back to the repo-root `VERSION` file when it is absent.
 - **First login**: after deployment, open `/admin` and login with `DS2API_ADMIN_KEY` shown in Zeabur env/template instructions (recommended: rotate to a strong secret after first login).
 
+#### Manual Deployment Without The Template
+
+If you do not want to use the `zeabur.yaml` one-click template, deploy directly from the repo root with Zeabur's GitHub integration:
+
+1. Fork this repo, or push the code to your own GitHub repository.
+2. In Zeabur Dashboard, create a Project, add a Service, then choose a GitHub/Git repository source.
+3. Select the repository and branch. Keep Root Directory as `/`.
+4. Use the Dockerfile build path. Zeabur auto-detects the repo-root `Dockerfile`; do not set `ZBPACK_IGNORE_DOCKERFILE=true`. If the UI asks for a Dockerfile name, enter `Dockerfile`.
+5. Add a persistent volume in the Service settings and mount it at `/data`.
+6. Configure environment variables:
+
+| Variable | Recommended value | Description |
+| --- | --- | --- |
+| `PORT` | `5001` | Service listen port; keep it aligned with the exposed Zeabur HTTP port. |
+| `DS2API_ADMIN_KEY` | Strong random string | Required admin login key. |
+| `DS2API_CONFIG_PATH` | `/data/config.json` | Recommended persistent config path. |
+| `LOG_LEVEL` | `INFO` | Optional log level. |
+| `DS2API_CONFIG_JSON` | Raw JSON or Base64 JSON | Optional config bootstrap from env. |
+| `DS2API_ENV_WRITEBACK` | `1` | Optional; enable only when using `DS2API_CONFIG_JSON` and you want the initial config written to `/data/config.json`. |
+
+7. Expose HTTP port `5001`. The health check path can be `/healthz`.
+8. After deployment, open `/admin`, login with `DS2API_ADMIN_KEY`, then import or edit config in Admin UI. A fresh volume does not need `/data/config.json` up front; the service boots first and creates the file on the first save.
+
+Troubleshooting:
+
+- **Startup log says `open /data/config.json: no such file or directory`**: make sure you deployed a version that includes the fresh-volume bootstrap fix, then redeploy the latest code.
+- **`open /app/config.json: permission denied`**: the config path still points at the read-only image directory; mount `/data` and set `DS2API_CONFIG_PATH=/data/config.json`.
+- **Config disappears after restart**: check that the `/data` persistent volume is mounted on this service. If you use `DS2API_CONFIG_JSON` but want Admin UI saves persisted, enable `DS2API_ENV_WRITEBACK=1`.
+
+References: Zeabur's official [GitHub/Git integration](https://zeabur.com/docs/en-US/deploy/github), [Dockerfile deployment](https://zeabur.com/docs/en-US/deploy/dockerfile), and [Volumes](https://zeabur.com/docs/data-management/volumes) docs.
+
 ---
 
 ## 3. Vercel Deployment
@@ -270,6 +302,7 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # optional for personal accounts
 | `VERCEL_TOKEN` | Vercel sync token | — |
 | `VERCEL_PROJECT_ID` | Vercel project ID | — |
 | `VERCEL_TEAM_ID` | Vercel team ID | — |
+| `DS2API_CHAT_HISTORY_PATH` | Chat history storage path (must be set to `/tmp/chat_history.json` on Vercel, otherwise unavailable due to read-only filesystem) | `data/chat_history.json` |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | Deployment protection bypass for internal Node→Go calls | — |
 
 ### 3.4 Vercel Architecture
@@ -359,6 +392,22 @@ If API responses return Vercel HTML `Authentication Required`:
 - **Option B**: Add `x-vercel-protection-bypass` header to requests
 - **Option C**: Set `VERCEL_AUTOMATION_BYPASS_SECRET` (or `DS2API_VERCEL_PROTECTION_BYPASS`) for internal Node→Go calls
 
+#### Chat History Unavailable (read-only file system)
+
+```text
+create chat history dir: mkdir /var/task/data: read-only file system
+```
+
+**Cause**: Vercel Serverless functions have a read-only filesystem (`/var/task`). Chat history fails because it cannot create directories there.
+
+**Fix**: Add the following in Vercel Project Settings → Environment Variables:
+
+```text
+DS2API_CHAT_HISTORY_PATH=/tmp/chat_history.json
+```
+
+`/tmp` is the only writable directory in Vercel Serverless. Data is ephemeral (not persisted across cold starts), but the feature works within a single instance lifetime.
+
 ### 3.6 Build Artifacts Not Committed
 
 - `static/admin` directory is not in Git
@@ -401,7 +450,7 @@ Or step by step:
 
 ```bash
 cd webui
-npm install
+npm ci
 npm run build
 # Output goes to static/admin/
 ```

docs/DEPLOY.md

@@ -64,8 +64,8 @@ cp config.example.json config.json
 
 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`
+- **触发条件**：默认仅在 Release `published` 时自动触发；也支持在 Actions 页面手动 `workflow_dispatch`，并填写 `release_tag` 复跑/补发
+- **构建产物**：多平台二进制压缩包、Linux Docker 镜像导出包 + `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 |
@@ -131,6 +131,7 @@ docker-compose logs -f
 
 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）。
 Compose 模板还会默认设置 `DS2API_CONFIG_PATH=/data/config.json` 并挂载 `./config.json:/data/config.json`，优先避免 `/app` 只读带来的配置持久化问题。
+镜像内会预创建 `/data` 并授权给非 root 的 `ds2api` 用户；如果你使用 bind mount 单文件，请确保宿主机 `config.json` 至少可被容器用户读取/写入，例如 `chmod 644 config.json`，否则 Linux UID/GID 不一致时仍可能出现 `open /data/config.json: permission denied`。
 兼容说明：若未设置 `DS2API_CONFIG_PATH` 且运行目录是 `/app`，新版本会优先使用 `/data/config.json`；当该文件不存在但检测到历史 `/app/config.json` 时，会自动回退读取旧路径，避免升级后“配置丢失”。
 
 如需固定版本，也可以直接拉取指定 tag：
@@ -196,7 +197,7 @@ healthcheck:
 部署要点：
 
 - **端口**：服务默认监听 `5001`，模板会固定设置 `PORT=5001`。
-- **配置持久化**：模板挂载卷 `/data`，并设置 `DS2API_CONFIG_PATH=/data/config.json`；在管理台导入配置后，会写入并持久化到该路径。
+- **配置持久化**：模板挂载卷 `/data`，并设置 `DS2API_CONFIG_PATH=/data/config.json`；首次空卷启动时会先使用空的文件模式配置，在管理台导入配置后，会写入并持久化到该路径。
 - **`open /app/config.json: permission denied`**：说明当前实例在尝试把运行时 token 持久化到只读路径（常见于镜像内 `/app`）。  
   处理建议：
   1. 显式设置可写路径：`DS2API_CONFIG_PATH=/data/config.json`（并挂载持久卷到 `/data`）；  
@@ -205,6 +206,37 @@ healthcheck:
 - **构建版本号**：Zeabur / 普通 `docker build` 默认不需要传 `BUILD_VERSION`；镜像会优先使用该构建参数，未提供时自动回退到仓库根目录的 `VERSION` 文件。
 - **首次登录**：部署完成后访问 `/admin`，使用 Zeabur 环境变量/模板指引中的 `DS2API_ADMIN_KEY` 登录（建议首次登录后自行更换为强密码）。
 
+#### 不使用模板手动部署
+
+如果你不想使用 `zeabur.yaml` 一键模板，可以直接用 Zeabur 的 GitHub 集成从仓库根目录构建：
+
+1. Fork 本仓库，或把代码推送到你自己的 GitHub 仓库。
+2. 在 Zeabur Dashboard 中创建 Project，然后添加 Service，选择 GitHub/Git 仓库来源。
+3. 选择仓库与分支，Root Directory 保持 `/`。
+4. 构建方式使用 Dockerfile。Zeabur 会自动检测仓库根目录的 `Dockerfile`；不要设置 `ZBPACK_IGNORE_DOCKERFILE=true`。如果界面要求填写 Dockerfile 名称，填写 `Dockerfile`。
+5. 在 Service 配置中添加持久卷，挂载目录填写 `/data`。
+6. 配置环境变量：
+
+| 变量 | 推荐值 | 说明 |
+| --- | --- | --- |
+| `PORT` | `5001` | 服务监听端口，需要和 Zeabur 暴露的 HTTP 端口一致。 |
+| `DS2API_ADMIN_KEY` | 强随机字符串 | 管理台登录密钥，必填。 |
+| `DS2API_CONFIG_PATH` | `/data/config.json` | 配置持久化路径，建议必填。 |
+| `LOG_LEVEL` | `INFO` | 可选，日志级别。 |
+| `DS2API_CONFIG_JSON` | 原始 JSON 或 Base64 JSON | 可选，用于用环境变量初始化配置。 |
+| `DS2API_ENV_WRITEBACK` | `1` | 可选；当设置了 `DS2API_CONFIG_JSON` 且希望首次启动后写入 `/data/config.json` 时再启用。 |
+
+7. 暴露 HTTP 端口 `5001`，健康检查路径可填 `/healthz`。
+8. 部署完成后访问 `/admin`，用 `DS2API_ADMIN_KEY` 登录，然后在管理台导入或编辑配置。首次空卷可以没有 `/data/config.json`，服务会先启动，第一次保存时自动创建该文件。
+
+常见问题：
+
+- **启动日志出现 `open /data/config.json: no such file or directory`**：请确认已经部署包含“首次空卷启动”修复的版本，并重新部署最新代码。
+- **出现 `open /app/config.json: permission denied`**：说明配置路径仍指向镜像内只读目录；设置持久卷 `/data`，并确认 `DS2API_CONFIG_PATH=/data/config.json`。
+- **管理台保存后重启配置丢失**：检查 `/data` 持久卷是否已挂载到当前服务；如果使用了 `DS2API_CONFIG_JSON`，但想让管理台保存落盘，请启用 `DS2API_ENV_WRITEBACK=1`。
+
+参考：Zeabur 官方文档的 [GitHub/Git 集成](https://zeabur.com/docs/en-US/deploy/github)、[Dockerfile 部署](https://zeabur.com/docs/zh-CN/deploy/dockerfile) 与 [Volumes](https://zeabur.com/docs/data-management/volumes)。
+
 ---
 
 ## 三、Vercel 部署
@@ -270,6 +302,7 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # 个人账号可留空
 | `VERCEL_TOKEN` | Vercel 同步 token | — |
 | `VERCEL_PROJECT_ID` | Vercel 项目 ID | — |
 | `VERCEL_TEAM_ID` | Vercel 团队 ID | — |
+| `DS2API_CHAT_HISTORY_PATH` | Chat history 存储路径（Vercel 上必须设为 `/tmp/chat_history.json`，否则因文件系统只读而不可用） | `data/chat_history.json` |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | 部署保护绕过密钥（内部 Node→Go 调用） | — |
 
 ### 3.3 运行时行为配置（通过 Admin API 设置）
@@ -369,6 +402,22 @@ No Output Directory named "public" found after the Build completed.
 - **方案 B**：请求中添加 `x-vercel-protection-bypass` 头
 - **方案 C**：设置 `VERCEL_AUTOMATION_BYPASS_SECRET`（或 `DS2API_VERCEL_PROTECTION_BYPASS`），仅影响内部 Node→Go 调用
 
+#### Chat History 不可用（read-only file system）
+
+```text
+create chat history dir: mkdir /var/task/data: read-only file system
+```
+
+**原因**：Vercel Serverless 函数的文件系统（`/var/task`）为只读，chat history 尝试在该路径下创建目录失败。
+
+**解决**：在 Vercel Project Settings → Environment Variables 中添加：
+
+```text
+DS2API_CHAT_HISTORY_PATH=/tmp/chat_history.json
+```
+
+`/tmp` 是 Vercel Serverless 环境中唯一可写的目录。数据在函数冷启动之间不会持久化（ephemeral），但在单个实例生命周期内功能正常。
+
 ### 3.6 仓库不提交构建产物
 
 - `static/admin` 目录不在 Git 中
@@ -411,7 +460,7 @@ go run ./cmd/ds2api
 
 ```bash
 cd webui
-npm install
+npm ci
 npm run build
 # 产物输出到 static/admin/
 ```

docs/DeepSeekSSE行为结构说明-2026-04-05.md

@@ -309,7 +309,18 @@ parse SSE block
 - 新模型可能增加新的 `p` 路径。
 - 新版本可能增加新的 fragment.type。
 - `CONTENT_FILTER` 的终态模板内容可能变化。
-- 自动续写相关状态（如 `INCOMPLETE` / `AUTO_CONTINUE`）当前主要来自实测与实现兼容逻辑，后续字段形态仍可能变化。
+- 自动续写相关状态（如 `INCOMPLETE` / `AUTO_CONTINUE`）当前主要来自实测与实现兼容逻辑，后续字段形态仍可能变化。当前实现不会仅因早期 `WIP` 状态就自动继续；只有显式 `INCOMPLETE` 或 `auto_continue` 信号才会触发 continue。
 - 解析器应当对未知字段、未知路径、未知事件保持容忍。
 
 如果你要把这份说明用于实际开发，建议同时保留原始流样本、回放脚本和回归测试，不要只依赖本文。
+
+## 2026-04-29 最近线上样本增量观察
+
+基于 `longtext-deepseek-v4-flash-20260429` 与 `longtext-deepseek-v4-pro-20260429` 两个真实账号长文本样本，近期格式变化要点如下：
+
+1. `data:` 事件中仍大量出现 `{"v":"..."}` 的无路径增量（`p` 缺失），解析器必须把空路径视为可见正文候选，而不能只依赖 `response/content`。
+2. 对象形态 `v`（如 `{"text":"..."}` / `{"content":"..."}`）仍会出现，且可能与无路径 chunk 混用；仅按字符串处理会导致正文丢块。
+3. 多轮 continuation 场景下，后续 chunk 可能不再重复显式 `status`，状态机需要保留上一轮 `INCOMPLETE` 语义直到出现终态。
+4. 2026-04-29 起客户端头部版本基线上调到 `x-client-version: 2.0.3`，否则部分账号会出现上游行为不一致（包括空输出与补轮异常）。
+
+建议：新增样本默认回放应优先覆盖「长文本 + 多轮 + 无路径 chunk」组合，避免只用短样本导致回归漏检。

docs/README.md

@@ -22,7 +22,7 @@
 
 ### 文档维护约定
 
-- 文档更新必须以实际代码实现为依据：总路由装配看 `internal/server/router.go`，协议/resource 路由看 `internal/httpapi/*/**/routes.go` 与 `internal/httpapi/admin/handler.go`，配置默认值看 `internal/config/*`，模型/alias 看 `internal/config/models.go`，prompt 兼容链路看 `docs/prompt-compatibility.md` 列出的代码入口。
+- 文档更新必须以实际代码实现为依据：总路由装配看 `internal/server/router.go`，协议/resource 路由看 `internal/httpapi/**/handler*.go` 与 `internal/httpapi/admin/handler.go`，配置默认值看 `internal/config/*`，模型/alias 看 `internal/config/models.go`，prompt 兼容链路看 `docs/prompt-compatibility.md` 列出的代码入口。
 - `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。
@@ -53,7 +53,7 @@ Recommended reading order:
 
 ### Maintenance conventions
 
-- Documentation updates must be grounded in the actual implementation: root routing lives in `internal/server/router.go`, protocol/resource routes live in `internal/httpapi/*/**/routes.go` and `internal/httpapi/admin/handler.go`, config defaults in `internal/config/*`, models/aliases in `internal/config/models.go`, and the prompt compatibility pipeline in the code entrypoints listed by `docs/prompt-compatibility.md`.
+- Documentation updates must be grounded in the actual implementation: root routing lives in `internal/server/router.go`, protocol/resource routes live in `internal/httpapi/**/handler*.go` and `internal/httpapi/admin/handler.go`, config defaults in `internal/config/*`, models/aliases in `internal/config/models.go`, and the prompt compatibility pipeline in the code entrypoints listed by `docs/prompt-compatibility.md`.
 - `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.

docs/TESTING.md

@@ -60,11 +60,10 @@ npm run build --prefix webui
 ./tests/scripts/check-refactor-line-gate.sh
 ./tests/scripts/check-node-split-syntax.sh
 ./tests/scripts/check-cross-build.sh
-
-# 历史阶段门禁：阶段 6 手工烟测签字检查（默认读取 plans/stage6-manual-smoke.md）
-./tests/scripts/check-stage6-manual-smoke.sh
 ```
 
+说明：`plans/stage6-manual-smoke.md` 已移除，阶段 6 手工烟测不再作为当前 CI 或发布门禁。
+
 ### 端到端测试 | End-to-End Tests
 
 ```bash

docs/prompt-compatibility.md

@@ -48,6 +48,8 @@ DS2API 当前的核心思路，不是把客户端传来的 `messages`、`tools`
   -> 文件收集 / inline 上传 / current input file（OpenAI 链路）
   -> completion payload
   -> 下游网页对话接口
+  -> assistantturn 输出语义归一（Go 非流式 + 流式收尾）
+  -> 各协议 renderer（OpenAI / Responses / Claude / Gemini）
 ```
 
 对应的关键代码入口：
@@ -72,6 +74,10 @@ DS2API 当前的核心思路，不是把客户端传来的 `messages`、`tools`
   [internal/promptcompat/thinking_injection.go](../internal/promptcompat/thinking_injection.go)
 - completion payload：
   [internal/promptcompat/standard_request.go](../internal/promptcompat/standard_request.go)
+- Go 输出侧 assistant turn：
+  [internal/assistantturn/turn.go](../internal/assistantturn/turn.go)
+- Go completion runtime：
+  [internal/completionruntime/nonstream.go](../internal/completionruntime/nonstream.go)
 
 ## 4. 下游真正收到的东西
 
@@ -98,13 +104,17 @@ DS2API 当前的核心思路，不是把客户端传来的 `messages`、`tools`
 - `prompt` 才是对话上下文主载体。
 - `ref_file_ids` 只承载文件引用，不承载普通文本消息。
 - `tools` 不会作为“原生工具 schema”直接下发给下游，而是被改写进 `prompt`。
+- 对外返回给客户端的 `prompt_tokens` / `input_tokens` / `promptTokenCount` 不再按“最后一条消息”或字符粗估近似返回，而是基于**完整上下文 prompt**做 tokenizer 计数；为了避免上下文实际超限但客户端误以为还能塞下，请求侧上下文 token 会额外保守上浮一点，宁可略大也不低估。
 - 当前 `/v1/chat/completions` 业务路径仍是“每次请求新建一个远端 `chat_session_id`，并默认发送 `parent_message_id: null`”；因此 DS2API 对外默认表现为“新会话 + prompt 拼历史”，而不是复用 DeepSeek 原生会话树。
 - 但 DeepSeek 远端本身支持同一 `chat_session_id` 的跨轮次持续对话。2026-04-27 已用项目内现有 DeepSeek client 做过一次不改业务代码的双轮实测：同一 `chat_session_id` 下，第 1 轮返回 `request_message_id=1` / `response_message_id=2` / 文本 `SESSION_TEST_ONE`；第 2 轮重新获取一次 PoW，并发送 `parent_message_id=2` 后，成功返回 `request_message_id=3` / `response_message_id=4` / 文本 `SESSION_TEST_TWO`。这说明“同远端会话持续聊天”能力存在，且每轮需要携带正确的 parent/message 链接信息，同时重新获取对应轮次可用的 PoW。
-- OpenAI Chat / Responses 原生走统一 OpenAI 标准化与 DeepSeek payload 组装；Claude / Gemini 会尽量复用 OpenAI prompt/tool 语义，其中 Gemini 直接复用 `promptcompat.BuildOpenAIPromptForAdapter`，Claude 消息接口在可代理场景会转换为 OpenAI chat 形态再执行。
+- OpenAI Chat / Responses 原生走统一 OpenAI 标准化与 DeepSeek payload 组装；Claude / Gemini 会尽量复用 OpenAI prompt/tool 语义，其中 Gemini 直接复用 `promptcompat.BuildOpenAIPromptForAdapter`。Go 主服务新增 `completionruntime` 启动层，统一执行 DeepSeek session/PoW/call；输出侧新增 `assistantturn` 语义层：非流式 OpenAI Chat / Responses / Claude / Gemini 会把 DeepSeek SSE 收集结果先归一成同一份 assistant turn，再分别渲染成各协议原生外形；流式 OpenAI Chat / Responses / Claude / Gemini 继续保持各协议实时 SSE framing，但最终收尾的 tool fallback、schema 归一、usage、empty-output / content-filter 错误语义同样由 `assistantturn` 判定。Claude / Gemini 的常规 Go 主路径不再依赖内部 `httptest` 转发到 OpenAI handler；`translatorcliproxy` 仍保留用于 Vercel bridge、兼容工具和回归测试。
+- Vercel Node 流式路径本轮不迁移，仍使用现有 Node bridge / stream-tool-sieve 实现；后续若变更 Node 流式语义，需要按 `assistantturn` 的 Go canonical 输出语义同步对齐。
 - 客户端传入的 thinking / reasoning 开关会被归一到下游 `thinking_enabled`。Gemini `generationConfig.thinkingConfig.thinkingBudget` 会翻译成同一套 thinking 开关；关闭时即使上游返回 `response/thinking_content`，兼容层也不会把它当作可见正文输出。若最终解析出的模型名带 `-nothinking` 后缀，则会无条件强制关闭 thinking，优先级高于请求体中的 `thinking` / `reasoning` / `reasoning_effort`。Claude surface 在流式请求且未显式声明 `thinking` 时，仍按 Anthropic 语义默认关闭；但在非流式代理场景，兼容层会内部开启一次下游 thinking，用于捕获“正文为空、工具调用落在 thinking 里”的情况，随后在回包前剥离用户不可见的 thinking block。
-- 对 OpenAI Chat / Responses 的非流式收尾，如果最终可见正文为空，兼容层会优先尝试把思维链中的独立 DSML / XML 工具块当作真实工具调用解析出来。流式链路也会在收尾阶段做同样的 fallback 检测，但不会因为思维链内容去中途拦截或改写流式输出；thinking / reasoning 增量仍按原样先发，只有在结束收尾时才可能补发最终工具调用结果。补发结果会作为本轮 assistant 的结构化 `tool_calls` / `function_call` 输出返回，而不是塞进 `content` 文本；如果客户端没有开启 thinking / reasoning，思维链只用于检测，不会作为 `reasoning_content` 或可见正文暴露。只有正文为空且思维链里也没有可执行工具调用时，才继续按空回复错误处理。
+- 对 OpenAI Chat / Responses 的非流式收尾，如果最终可见正文为空，兼容层会优先尝试把思维链中的独立 DSML / XML 工具块当作真实工具调用解析出来。流式链路也会在收尾阶段做同样的 fallback 检测，但不会因为思维链内容去中途拦截或改写流式输出；真正的工具识别始终基于原始上游文本，而不是基于“已经做过可见输出清洗”的版本，因此即使最终可见层会剥离完整 leaked DSML / XML `tool_calls` wrapper、并抑制全空参数或无效 wrapper 块，也不会影响真实工具调用转成结构化 `tool_calls` / `function_call`。补发结果会作为本轮 assistant 的结构化 `tool_calls` / `function_call` 输出返回，而不是塞进 `content` 文本；如果客户端没有开启 thinking / reasoning，思维链只用于检测，不会作为 `reasoning_content` 或可见正文暴露。只有正文为空且思维链里也没有可执行工具调用时，才继续按空回复错误处理。
 - OpenAI Chat / Responses 的空回复错误处理之前会默认做一次内部补偿重试：第一次上游完整结束后，如果最终可见正文为空、没有解析到工具调用、也没有已经向客户端流式发出工具调用，并且终止原因不是 `content_filter`，兼容层会复用同一个 `chat_session_id`、账号、token 与工具策略，把原始 completion `prompt` 追加固定后缀 `Previous reply had no visible output. Please regenerate the visible final answer or tool call now.` 后重新提交一次。重试遵循 DeepSeek 多轮对话协议：从第一次上游 SSE 流中提取 `response_message_id`，并在重试 payload 中设置 `parent_message_id` 为该值，使重试成为同一会话的后续轮次而非断裂的根消息；同时重新获取一次 PoW（若 PoW 获取失败则回退到原始 PoW）。该重试不会重新标准化消息、不会新建 session、不会切换账号，也不会向流式客户端插入重试标记；第二次 thinking / reasoning 会按正常增量直接接到第一次之后，并继续使用 overlap trim 去重。若第二次仍为空，终端错误码仍保持现有 `upstream_empty_output`；若任一尝试触发空 `content_filter`，不做补偿重试并保持 `content_filter` 错误。JS Vercel 运行时同样设置 `parent_message_id`，但因无法直接调用 PoW API 而复用原始 PoW。
 
+- OpenAI Chat / Responses 在最终可见正文渲染阶段，会把 DeepSeek 搜索返回中的 `[citation:N]` / `[reference:N]` 标记替换成对应 Markdown 链接。`citation` 标记按一基序号解析；`reference` 标记只有在同一段正文中出现 `[reference:0]`（允许冒号后有空格）时才按零基序号映射，并且不会影响同段正文里的 `citation` 标记。
+
 ## 5. prompt 是怎么拼出来的
 
 OpenAI Chat / Responses 在标准化后、current input file 之前，会默认执行 `thinking_injection` 增强。它参考 DeepSeek V4 “把控制指令放在 user 消息末尾更稳定”的用法，在最新 user message 后追加思考增强提示词。当前内置默认提示词以 `Reasoning Effort: Absolute maximum with no shortcuts permitted.` 开头，并继续要求模型充分分解问题、覆盖潜在路径与边界条件、把完整推演过程显式写出。该开关默认启用，可通过 `thinking_injection.enabled=false` 关闭；也可以通过 `thinking_injection.prompt` 自定义提示词，留空时使用内置默认提示词。
@@ -114,6 +124,11 @@ OpenAI Chat / Responses 在标准化后、current input file 之前，会默认
 - 普通请求会直接出现在最终 `prompt` 的最新 user block 末尾。
 - 如果触发 current input file，它会进入完整上下文文件中。
 
+另外，`MessagesPrepareWithThinking` 还会在最终 prompt 的最前面预置一段固定的 system 级“输出完整性约束（Output integrity guard）”：
+
+- 如果上游上下文、工具输出或解析后的文本出现乱码、损坏、部分解析、重复或其他畸形片段，不要模仿、不要回显，只输出给用户的正确内容。
+- 这段约束位于普通 system / tool prompt 之前，因此是当前最终 prompt 里的最高优先级前置指令。
+
 ### 5.1 角色标记
 
 最终 prompt 使用 DeepSeek 风格角色标记：
@@ -152,11 +167,13 @@ OpenAI Chat / Responses 在标准化后、current input file 之前，会默认
 
 工具调用正例现在优先示范官方 DSML 风格：`<|DSML|tool_calls>` → `<|DSML|invoke name="...">` → `<|DSML|parameter name="...">`。
 兼容层仍接受旧式纯 `<tool_calls>` wrapper，但提示词会优先要求模型输出官方 DSML 标签，并强调不能只输出 closing wrapper 而漏掉 opening tag。需要注意：这是“兼容 DSML 外壳，内部仍以 XML 解析语义为准”，不是原生 DSML 全链路实现；DSML 标签会在解析入口归一化回现有 XML 标签后继续走同一套 parser。
-数组参数使用 `<item>...</item>` 子节点表示；当某个参数体只包含 item 子节点时，Go / Node 解析器会把它还原成数组，避免 `questions` / `options` 这类 schema 中要求 array 的参数被误解析成 `{ "item": ... }` 对象。若模型把完整结构化 XML fragment 误包进 CDATA，兼容层会在保护 `content` / `command` 等原文字段的前提下，尝试把非原文字段中的 CDATA XML fragment 还原成 object / array。不过，如果 CDATA 只是单个平面的 XML/HTML 标签，例如 `<b>urgent</b>` 这种行内标记，兼容层会保留原始字符串，不会强行升成 object / array；只有明显表示结构的 CDATA 片段，例如多兄弟节点、嵌套子节点或 `item` 列表，才会触发结构化恢复。
+数组参数使用 `<item>...</item>` 子节点表示；当某个参数体只包含 item 子节点时，Go / Node 解析器会把它还原成数组，避免 `questions` / `options` 这类 schema 中要求 array 的参数被误解析成 `{ "item": ... }` 对象。除此之外，解析器还会回收一些更松散的列表写法，例如 JSON array 字面量或逗号分隔的 JSON 项序列，只要它们足够明确；但 `<item>` 仍然是首选形态。若模型把完整结构化 XML fragment 误包进 CDATA，兼容层会在保护 `content` / `command` 等原文字段的前提下，尝试把非原文字段中的 CDATA XML fragment 还原成 object / array。不过，如果 CDATA 只是单个平面的 XML/HTML 标签，例如 `<b>urgent</b>` 这种行内标记，兼容层会保留原始字符串，不会强行升成 object / array；只有明显表示结构的 CDATA 片段，例如多兄弟节点、嵌套子节点或 `item` 列表，才会触发结构化恢复。
+Go 侧读取 DeepSeek SSE 时不再依赖 `bufio.Scanner` 的固定 2MiB 单行上限；当写文件类工具把很长的 `content` 放在单个 `data:` 行里返回时，非流式收集、流式解析和 auto-continue 透传都会保留完整行，再进入同一套工具解析与序列化流程。
 在 assistant 最终回包阶段，如果某个 tool 参数在声明 schema 中明确是 `string`，兼容层会在把解析后的 `tool_calls` / `function_call` 重新序列化成 OpenAI / Responses / Claude 可见参数前，递归把该路径上的 number / bool / object / array 统一转成字符串；其中 object / array 会压成紧凑 JSON 字符串。这个保护只对 schema 明确声明为 string 的路径生效，不会改写本来就是 `number` / `boolean` / `object` / `array` 的参数。这样可以兼容 DeepSeek 输出了结构化片段、但上游客户端工具 schema 又严格要求字符串参数的场景（例如 `content`、`prompt`、`path`、`taskId` 等）。
 工具 schema 的权威来源始终是**当前请求实际携带的 schema**，而不是同名工具在其他 runtime（Claude Code / OpenCode / Codex 等）里的默认印象。兼容层现在会同时兼容 OpenAI 风格 `function.parameters`、直接工具对象上的 `parameters` / `input_schema`、以及 camelCase 的 `inputSchema` / `schema`，并在最终输出阶段按这份请求内 schema 决定是保留 array/object，还是仅对明确声明为 `string` 的路径做字符串化。该规则同样适用于 Claude 的流式收尾和 Vercel Node 流式 tool-call formatter，避免不同 runtime 因 schema shape 差异而出现同名工具参数类型漂移。
 正例中的工具名只会来自当前请求实际声明的工具；如果当前请求没有足够的已知工具形态，就省略对应的单工具、多工具或嵌套示例，避免把不可用工具名写进 prompt。
 对执行类工具，脚本内容必须进入执行参数本身：`Bash` / `execute_command` 使用 `command`，`exec_command` 使用 `cmd`；不要把脚本示范成 `path` / `content` 文件写入参数。
+如果当前请求声明了 `Read` / `read_file` 这类读取工具，兼容层会额外注入一条 read-tool cache guard：当读取结果只表示“文件未变更 / 已在历史中 / 请引用先前上下文 / 没有正文内容”时，模型必须把它视为内容不可用，不能反复调用同一个无正文读取；应改为请求完整正文读取能力，或向用户说明需要重新提供文件内容。这个约束只缓解客户端缓存返回空内容导致的死循环，DS2API 不会也无法凭空恢复客户端本地文件正文。
 
 OpenAI 路径实现：
 [internal/promptcompat/tool_prompt.go](../internal/promptcompat/tool_prompt.go)
@@ -233,6 +250,14 @@ OpenAI 文件相关实现：
 - 文件 ID 收集：
   [internal/promptcompat/file_refs.go](../internal/promptcompat/file_refs.go)
 
+OpenAI 的文件上传现在不再是“只传文件本体”的通用路径，而是会先根据请求里的 `model` 解析出 DeepSeek 的上传类型，并把它透传到上传接口的 `x-model-type`。当前可见的上传类型就是 `default` / `expert` / `vision`，其中 vision 请求上传图片时必须带上 `vision`，否则下游容易退回到仅文本或 OCR 语义。这个模型类型会同时用于：
+
+- `/v1/files` 这类独立文件上传入口
+- Chat / Responses 的 inline 图片、附件上传
+- current input file 触发时生成的 `DS2API_HISTORY.txt` 上下文文件
+
+也就是说，文件上传和完成请求的 `model_type` 现在是一致的：完成 payload 里仍然是 `model_type`，上传文件则会在 DeepSeek 上传阶段携带同样的模型类型信息。
+
 结论：
 
 - “systemprompt 文字”在 prompt 里
@@ -244,9 +269,10 @@ OpenAI 文件相关实现：
 
 兼容层现在只保留 `current_input_file` 这一种拆分方式；旧的 `history_split` 已废弃，只保留为兼容旧配置的字段，不再参与请求处理。
 
-- `current_input_file` 默认开启；它用于把“完整上下文”合并进隐藏上下文文件。当最新 user turn 的纯文本长度达到 `current_input_file.min_chars`（默认 `0`）时，兼容层会上传一个文件名为 `IGNORE.txt` 的上下文文件，并在 live prompt 中只保留一个中性的 user 消息要求模型直接回答最新请求，不再暴露文件名或要求模型读取本地文件。
+- `current_input_file` 默认开启；它用于把“完整上下文”合并进 `DS2API_HISTORY.txt` 上下文文件。当最新 user turn 的纯文本长度达到 `current_input_file.min_chars`（默认 `0`）时，兼容层会上传一个文件名为 `DS2API_HISTORY.txt` 的上下文文件。文件内容会先做 OpenAI 消息标准化，再序列化成按轮次编号的 `DS2API_HISTORY.txt` 风格 transcript，带有 `# DS2API_HISTORY.txt` 标题和 `=== N. ROLE ===` 分段；live prompt 中则会给出一个 continuation 语气的 user 消息，引导模型从 `DS2API_HISTORY.txt` 的最新状态继续推进，并直接回答最新请求，避免把任务拉回起点。
 - 如果 `current_input_file.enabled=false`，请求会直接透传，不上传任何拆分上下文文件。
 - 旧的 `history_split.enabled` / `history_split.trigger_after_turns` 会被读取进配置对象以保持兼容，但不会触发拆分上传，也不会影响 `current_input_file` 的默认开启。
+- 即使触发 `current_input_file` 后 live prompt 被缩短，对客户端回包里的上下文 token 统计，仍会沿用**拆分前的完整 prompt 语义**做计数，而不是按缩短后的占位 prompt 计算；否则会把真实上下文显著算小。
 
 相关实现：
 
@@ -257,16 +283,24 @@ OpenAI 文件相关实现：
 - 旧历史拆分兼容壳：
   [internal/httpapi/openai/history/history_split.go](../internal/httpapi/openai/history/history_split.go)
 
-当前输入转文件启用并触发时，上传文件的真实文件名是 `IGNORE.txt`，文件内容是完整 `messages` 上下文；它仍会先用 OpenAI 消息标准化和 DeepSeek 角色标记序列化，再包进 `IGNORE` 文件边界里：
+当前输入转文件启用并触发时，上传文件的真实文件名是 `DS2API_HISTORY.txt`，文件内容是完整 `messages` 上下文；它仍会先用 OpenAI 消息标准化和 DeepSeek 角色标记序列化，再按轮次编号成 `DS2API_HISTORY.txt` 风格的 transcript（不再注入文件边界标签）：
 
 ```text
-[uploaded filename]: IGNORE.txt
-[file content end]
+[uploaded filename]: DS2API_HISTORY.txt
+# DS2API_HISTORY.txt
+Prior conversation history and tool progress.
 
-<｜begin▁of▁sentence｜><｜System｜>...<｜User｜>...<｜Assistant｜>...<｜Tool｜>...<｜User｜>...
+=== 1. SYSTEM ===
+...
 
-[file name]: IGNORE
-[file content begin]
+=== 2. USER ===
+...
+
+=== 3. ASSISTANT ===
+...
+
+=== 4. TOOL ===
+...
 ```
 
 开启后，请求的 live prompt 不再直接内联完整上下文，而是保留一个 user role 的短提示，提示模型基于已提供上下文直接回答最新请求；上传后的 `file_id` 会进入 `ref_file_ids`。
@@ -318,7 +352,7 @@ OpenAI 文件相关实现：
 
 ```json
 {
-  "prompt": "<｜begin▁of▁sentence｜><｜System｜>原 system / developer\n\nYou have access to these tools: ...<｜end▁of▁instructions｜><｜User｜>The current request and prior conversation context have already been provided. Answer the latest user request directly.<｜Assistant｜>",
+  "prompt": "<｜begin▁of▁sentence｜><｜System｜>原 system / developer\n\nYou have access to these tools: ...<｜end▁of▁instructions｜><｜User｜>Continue from the latest state in the attached DS2API_HISTORY.txt context. Treat it as the current working state and answer the latest user request directly.<｜Assistant｜>",
   "ref_file_ids": [
     "file-current-input-ignore",
     "file-systemprompt",
@@ -333,7 +367,7 @@ OpenAI 文件相关实现：
 
 - 大部分结构化语义被压进 `prompt`
 - 文件保持文件
-- 需要时把完整上下文拆进隐藏上下文文件
+- 需要时把完整上下文拆进 `DS2API_HISTORY.txt` 上下文文件，并按轮次编号成 transcript
 
 ## 12. 修改时必须同步本文档的场景
 
@@ -346,7 +380,7 @@ OpenAI 文件相关实现：
 - tool result 注入方式变更
 - tool prompt 模板或 tool_choice 约束变更
 - inline 文件上传 / 文件引用收集规则变更
-- current input file 触发条件、上传格式、`IGNORE` 包装格式变更
+- current input file 触发条件、上传格式、`DS2API_HISTORY.txt` transcript 结构变更
 - 旧 `history_split` 兼容逻辑的读取、忽略或退化行为变更
 - completion payload 字段语义变更
 - Claude / Gemini 对这套统一语义的复用关系变更

docs/toolcall-semantics.md

@@ -26,7 +26,7 @@
 </tool_calls>
 ```
 
-这不是原生 DSML 全链路实现。DSML 只作为 prompt 外壳和解析入口别名；进入 parser 前会被归一化成 `<tool_calls>` / `<invoke>` / `<parameter>`，内部仍以现有 XML 解析语义为准。
+这不是原生 DSML 全链路实现。DSML 主要用于让模型有意识地输出协议标识，隔离普通 XML 语义；进入 parser 前会按固定本地标签名归一化成 `<tool_calls>` / `<invoke>` / `<parameter>`，内部仍以现有 XML 解析语义为准。
 
 约束：
 
@@ -39,7 +39,8 @@
 兼容修复：
 
 - 如果模型漏掉 opening wrapper，但后面仍输出了一个或多个 invoke 并以 closing wrapper 收尾，Go 解析链路会在解析前补回缺失的 opening wrapper。
-- 如果模型把 DSML 标签里的分隔符 `|` 写漏成空格（例如 `<|DSML tool_calls>` / `<|DSML invoke>` / `<|DSML parameter>`，或无 leading pipe 的 `<DSML tool_calls>` 形态），或把 `DSML` 与工具标签名直接黏连（例如 `<DSMLtool_calls>` / `<DSMLinvoke>` / `<DSMLparameter>`），或把最前面的 pipe 误写成全宽竖线（例如 `<｜DSML|tool_calls>` / `<｜DSML|invoke>` / `<｜DSML|parameter>`），Go / Node 会在固定工具标签名范围内归一化；相似但非工具标签名（如 `tool_calls_extra`）仍按普通文本处理。
+- Go / Node 解析层不再枚举每一种 DSML typo。它会把工具标签名前的 `DSML`、管道符 `|` / `｜`、空白、重复 leading `<` 视为可容忍的协议噪声，然后只匹配固定本地标签名 `tool_calls` / `invoke` / `parameter`。例如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<<DSML|DSML|tool_calls>` 都会归一化；相似但非固定标签名（如 `tool_calls_extra`）仍按普通文本处理。
+- 如果模型在固定工具标签名后多输出一个尾部管道符，例如 `<|DSML|tool_calls|` / `<|DSML|invoke|` / `<|DSML|parameter|`，兼容层会把这个尾部 `|` 当作异常标签终止符并补齐缺失的 `>`；如果后面已经有 `>`，也会消费这个多余 `|` 后再归一化。
 - 这是一个针对常见模型失误的窄修复，不改变推荐输出格式；prompt 仍要求模型直接输出完整 DSML 外壳。
 - 裸 `<invoke ...>` / `<parameter ...>` 不会被当成“已支持的工具语法”；只有 `tool_calls` wrapper 或可修复的缺失 opening wrapper 才会进入工具调用路径。
 
@@ -53,7 +54,7 @@
 
 在流式链路中（Go / Node 一致）：
 
-- DSML `<|DSML|tool_calls>` wrapper、兼容变体（`<dsml|tool_calls>`、`<｜tool_calls>`、`<|tool_calls>`、`<｜DSML|tool_calls>`）、窄容错空格分隔形态（如 `<|DSML tool_calls>`）、黏连形态（如 `<DSMLtool_calls>`）和 canonical `<tool_calls>` wrapper 都会进入结构化捕获
+- DSML `<|DSML|tool_calls>` wrapper、基于固定本地标签名的 DSML 噪声容错形态、尾部管道符形态（如 `<|DSML|tool_calls|`）和 canonical `<tool_calls>` wrapper 都会进入结构化捕获
 - 如果流里直接从 invoke 开始，但后面补上了 closing wrapper，Go 流式筛分也会按缺失 opening wrapper 的修复路径尝试恢复
 - 已识别成功的工具调用不会再次回流到普通文本
 - 不符合新格式的块不会执行，并继续按原样文本透传
@@ -61,6 +62,7 @@
 - 支持嵌套围栏（如 4 反引号嵌套 3 反引号）和 CDATA 内围栏保护
 - 如果模型把 `<![CDATA[` 打开后却没有闭合，流式扫描阶段仍会保守地继续缓冲，不会误把 CDATA 里的示例 XML 当成真实工具调用；在最终 parse / flush 恢复阶段，会对这类 loose CDATA 做窄修复，尽量保住外层已完整包裹的真实工具调用
 - 当文本中 mention 了某种标签名（如 `<dsml|tool_calls>` 或 Markdown inline code 里的 `<|DSML|tool_calls>`）而后面紧跟真正工具调用时，sieve 会跳过不可解析的 mention 候选并继续匹配后续真实工具块，不会因 mention 导致工具调用丢失，也不会截断 mention 后的正文
+- Go 侧 SSE 读取不再使用 `bufio.Scanner` 的固定 token 上限；单个 `data:` 行中包含很长的写文件参数时，非流式收集、流式解析与 auto-continue 透传都应保留完整行，再交给 tool parser 处理
 
 另外，`<parameter>` 的值如果本身是合法 JSON 字面量，也会按结构化值解析，而不是一律保留为字符串。例如 `123`、`true`、`null`、`[1,2]`、`{"a":1}` 都会还原成对应的 number / boolean / null / array / object。
 结构化 XML 参数也会还原为 JSON 结构：如果参数体只包含一个或多个 `<item>...</item>` 子节点，会输出数组；嵌套对象里的 item-only 字段也同样按数组处理。例如 `<parameter name="questions"><item><question>...</question></item></parameter>` 会输出 `{"questions":[{"question":"..."}]}`，而不是 `{"questions":{"item":...}}`。
@@ -94,7 +96,7 @@ node --test tests/node/stream-tool-sieve.test.js
 
 - DSML `<|DSML|tool_calls>` wrapper 正常解析
 - legacy canonical `<tool_calls>` wrapper 正常解析
-- 别名变体（`<dsml|tool_calls>`、`<｜tool_calls>`、`<|tool_calls>`）、DSML 空格分隔 typo（如 `<|DSML tool_calls>`）和黏连 typo（如 `<DSMLtool_calls>`）正常解析
+- 固定本地标签名的 DSML 噪声容错形态（如 `<DSML|tool_calls>`、`<<|DSML|tool_calls>`、`<|DSML tool_calls>`、`<DSMLtool_calls>`、`<<DSML|DSML|tool_calls>`）正常解析
 - 混搭标签（DSML wrapper + canonical inner）归一化后正常解析
 - 波浪线围栏 `~~~` 内的示例不执行
 - 嵌套围栏（4 反引号嵌套 3 反引号）内的示例不执行

go.mod

@@ -6,10 +6,13 @@ require (
 	github.com/andybalholm/brotli v1.2.1
 	github.com/go-chi/chi/v5 v5.2.5
 	github.com/google/uuid v1.6.0
+	github.com/hupe1980/go-tiktoken v0.0.10
 	github.com/refraction-networking/utls v1.8.2
 	github.com/router-for-me/CLIProxyAPI/v6 v6.9.14
 )
 
+require github.com/dlclark/regexp2 v1.11.5 // indirect
+
 require (
 	github.com/klauspost/compress v1.18.5 // indirect
 	github.com/sirupsen/logrus v1.9.4 // indirect

go.sum

@@ -2,10 +2,14 @@ github.com/andybalholm/brotli v1.2.1 h1:R+f5xP285VArJDRgowrfb9DqL18yVK0gKAW/F+eT
 github.com/andybalholm/brotli v1.2.1/go.mod h1:rzTDkvFWvIrjDXZHkuS16NPggd91W3kUSvPlQ1pLaKY=
 github.com/davecgh/go-spew v1.1.1 h1:vj9j/u1bqnvCEfJOwUhtlOARqs3+rkHYY13jYWTU97c=
 github.com/davecgh/go-spew v1.1.1/go.mod h1:J7Y8YcW2NihsgmVo/mv3lAwl/skON4iLHjSsI+c5H38=
+github.com/dlclark/regexp2 v1.11.5 h1:Q/sSnsKerHeCkc/jSTNq1oCm7KiVgUMZRDUoRu0JQZQ=
+github.com/dlclark/regexp2 v1.11.5/go.mod h1:DHkYz0B9wPfa6wondMfaivmHpzrQ3v9q8cnmRbL6yW8=
 github.com/go-chi/chi/v5 v5.2.5 h1:Eg4myHZBjyvJmAFjFvWgrqDTXFyOzjj7YIm3L3mu6Ug=
 github.com/go-chi/chi/v5 v5.2.5/go.mod h1:X7Gx4mteadT3eDOMTsXzmI4/rwUpOwBHLpAfupzFJP0=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
+github.com/hupe1980/go-tiktoken v0.0.10 h1:m6phOJaGyctqWdGIgwn9X8AfJvaG74tnQoDL+ntOUEQ=
+github.com/hupe1980/go-tiktoken v0.0.10/go.mod h1:NME6d8hrE+Jo+kLUZHhXShYV8e40hYkm4BbSLQKtvAo=
 github.com/klauspost/compress v1.18.5 h1:/h1gH5Ce+VWNLSWqPzOVn6XBO+vJbCNGvjoaGBFW2IE=
 github.com/klauspost/compress v1.18.5/go.mod h1:cwPg85FWrGar70rWktvGQj8/hthj3wpl0PGDogxkrSQ=
 github.com/pmezard/go-difflib v1.0.0 h1:4DBwDE0NGyQoBHbLQYPwSUPoCMWR5BEzIk/f1lZbAQM=
@@ -37,6 +41,8 @@ golang.org/x/net v0.52.0 h1:He/TN1l0e4mmR3QqHMT2Xab3Aj3L9qjbhRm78/6jrW0=
 golang.org/x/net v0.52.0/go.mod h1:R1MAz7uMZxVMualyPXb+VaqGSa3LIaUqk0eEt3w36Sw=
 golang.org/x/sys v0.42.0 h1:omrd2nAlyT5ESRdCLYdm3+fMfNFE/+Rf4bDIQImRJeo=
 golang.org/x/sys v0.42.0/go.mod h1:4GL1E5IUh+htKOUEOaiffhrAeqysfVGipDYzABqnCmw=
+golang.org/x/text v0.35.0 h1:JOVx6vVDFokkpaq1AEptVzLTpDe9KGpj5tR4/X+ybL8=
+golang.org/x/text v0.35.0/go.mod h1:khi/HExzZJ2pGnjenulevKNX1W67CUy0AsXcNubPGCA=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405 h1:yhCVgyC4o1eVCa2tZl7eS0r+SDo693bJlVdllGtEeKM=
 gopkg.in/check.v1 v0.0.0-20161208181325-20d25e280405/go.mod h1:Co6ibVJAznAaIkqp8huTwlJQCZ016jof/cbN4VW5Yz0=
 gopkg.in/yaml.v3 v3.0.1 h1:fxVm/GzAzEWqLHuvctI91KS9hhNmmWOoWu0XTYJS7CA=

internal/assistantturn/stream.go

@@ -0,0 +1,64 @@
+package assistantturn
+
+import (
+	"ds2api/internal/httpapi/openai/shared"
+	"ds2api/internal/sse"
+)
+
+type StreamEventType string
+
+const (
+	StreamEventTextDelta     StreamEventType = "text_delta"
+	StreamEventThinkingDelta StreamEventType = "thinking_delta"
+	StreamEventToolCall      StreamEventType = "tool_call"
+	StreamEventDone          StreamEventType = "done"
+	StreamEventError         StreamEventType = "error"
+	StreamEventPing          StreamEventType = "ping"
+)
+
+type StreamEvent struct {
+	Type     StreamEventType
+	Text     string
+	Thinking string
+	ToolCall any
+	Error    *OutputError
+	Usage    *Usage
+}
+
+type Accumulator struct {
+	inner shared.StreamAccumulator
+}
+
+type AccumulatorOptions struct {
+	ThinkingEnabled       bool
+	SearchEnabled         bool
+	StripReferenceMarkers bool
+}
+
+func NewAccumulator(opts AccumulatorOptions) *Accumulator {
+	return &Accumulator{
+		inner: shared.StreamAccumulator{
+			ThinkingEnabled:       opts.ThinkingEnabled,
+			SearchEnabled:         opts.SearchEnabled,
+			StripReferenceMarkers: opts.StripReferenceMarkers,
+		},
+	}
+}
+
+func (a *Accumulator) Apply(parsed sse.LineResult) shared.StreamAccumulatorResult {
+	if a == nil {
+		return shared.StreamAccumulatorResult{}
+	}
+	return a.inner.Apply(parsed)
+}
+
+func (a *Accumulator) Snapshot() (rawText, text, rawThinking, thinking, detectionThinking string) {
+	if a == nil {
+		return "", "", "", "", ""
+	}
+	return a.inner.RawText.String(),
+		a.inner.Text.String(),
+		a.inner.RawThinking.String(),
+		a.inner.Thinking.String(),
+		a.inner.ToolDetectionThinking.String()
+}

internal/assistantturn/turn.go

@@ -0,0 +1,227 @@
+package assistantturn
+
+import (
+	"net/http"
+	"strings"
+
+	"ds2api/internal/httpapi/openai/shared"
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/sse"
+	"ds2api/internal/toolcall"
+	"ds2api/internal/util"
+)
+
+type StopReason string
+
+const (
+	StopReasonStop          StopReason = "stop"
+	StopReasonToolCalls     StopReason = "tool_calls"
+	StopReasonContentFilter StopReason = "content_filter"
+	StopReasonError         StopReason = "error"
+)
+
+type Usage struct {
+	InputTokens     int
+	OutputTokens    int
+	ReasoningTokens int
+	TotalTokens     int
+}
+
+type OutputError struct {
+	Status  int
+	Message string
+	Code    string
+}
+
+type Turn struct {
+	Model             string
+	Prompt            string
+	RawText           string
+	RawThinking       string
+	DetectionThinking string
+	Text              string
+	Thinking          string
+	ToolCalls         []toolcall.ParsedToolCall
+	ParsedToolCalls   toolcall.ToolCallParseResult
+	CitationLinks     map[int]string
+	ContentFilter     bool
+	ResponseMessageID int
+	StopReason        StopReason
+	Usage             Usage
+	Error             *OutputError
+}
+
+type BuildOptions struct {
+	Model                 string
+	Prompt                string
+	RefFileTokens         int
+	SearchEnabled         bool
+	StripReferenceMarkers bool
+	ToolNames             []string
+	ToolsRaw              any
+	ToolChoice            promptcompat.ToolChoicePolicy
+}
+
+type StreamSnapshot struct {
+	RawText               string
+	VisibleText           string
+	RawThinking           string
+	VisibleThinking       string
+	DetectionThinking     string
+	ContentFilter         bool
+	CitationLinks         map[int]string
+	ResponseMessageID     int
+	AlreadyEmittedCalls   bool
+	AdditionalToolCalls   []toolcall.ParsedToolCall
+	AlreadyEmittedToolRaw bool
+}
+
+func BuildTurnFromCollected(result sse.CollectResult, opts BuildOptions) Turn {
+	thinking := shared.CleanVisibleOutput(result.Thinking, opts.StripReferenceMarkers)
+	text := shared.CleanVisibleOutput(result.Text, opts.StripReferenceMarkers)
+	if opts.SearchEnabled {
+		text = shared.ReplaceCitationMarkersWithLinks(text, result.CitationLinks)
+	}
+
+	parsed := shared.DetectAssistantToolCalls(result.Text, text, result.Thinking, result.ToolDetectionThinking, opts.ToolNames)
+	calls := toolcall.NormalizeParsedToolCallsForSchemas(parsed.Calls, opts.ToolsRaw)
+	parsed.Calls = calls
+
+	stopReason := StopReasonStop
+	if result.ContentFilter {
+		stopReason = StopReasonContentFilter
+	}
+	if len(calls) > 0 {
+		stopReason = StopReasonToolCalls
+	}
+
+	turn := Turn{
+		Model:             opts.Model,
+		Prompt:            opts.Prompt,
+		RawText:           result.Text,
+		RawThinking:       result.Thinking,
+		DetectionThinking: result.ToolDetectionThinking,
+		Text:              text,
+		Thinking:          thinking,
+		ToolCalls:         calls,
+		ParsedToolCalls:   parsed,
+		CitationLinks:     result.CitationLinks,
+		ContentFilter:     result.ContentFilter,
+		ResponseMessageID: result.ResponseMessageID,
+		StopReason:        stopReason,
+	}
+	turn.Usage = BuildUsage(opts.Model, opts.Prompt, thinking, text, opts.RefFileTokens)
+	turn.Error = ValidateTurn(turn, opts.ToolChoice)
+	if turn.Error != nil {
+		turn.StopReason = StopReasonError
+	}
+	return turn
+}
+
+func BuildTurnFromStreamSnapshot(snapshot StreamSnapshot, opts BuildOptions) Turn {
+	thinking := shared.CleanVisibleOutput(snapshot.VisibleThinking, opts.StripReferenceMarkers)
+	text := shared.CleanVisibleOutput(snapshot.VisibleText, opts.StripReferenceMarkers)
+	if opts.SearchEnabled {
+		text = shared.ReplaceCitationMarkersWithLinks(text, snapshot.CitationLinks)
+	}
+
+	parsed := shared.DetectAssistantToolCalls(snapshot.RawText, text, snapshot.RawThinking, snapshot.DetectionThinking, opts.ToolNames)
+	calls := parsed.Calls
+	if len(calls) == 0 && len(snapshot.AdditionalToolCalls) > 0 {
+		calls = snapshot.AdditionalToolCalls
+	}
+	calls = toolcall.NormalizeParsedToolCallsForSchemas(calls, opts.ToolsRaw)
+	parsed.Calls = calls
+
+	stopReason := StopReasonStop
+	if snapshot.ContentFilter {
+		stopReason = StopReasonContentFilter
+	}
+	if len(calls) > 0 || snapshot.AlreadyEmittedCalls || snapshot.AlreadyEmittedToolRaw {
+		stopReason = StopReasonToolCalls
+	}
+
+	turn := Turn{
+		Model:             opts.Model,
+		Prompt:            opts.Prompt,
+		RawText:           snapshot.RawText,
+		RawThinking:       snapshot.RawThinking,
+		DetectionThinking: snapshot.DetectionThinking,
+		Text:              text,
+		Thinking:          thinking,
+		ToolCalls:         calls,
+		ParsedToolCalls:   parsed,
+		CitationLinks:     snapshot.CitationLinks,
+		ContentFilter:     snapshot.ContentFilter,
+		ResponseMessageID: snapshot.ResponseMessageID,
+		StopReason:        stopReason,
+	}
+	turn.Usage = BuildUsage(opts.Model, opts.Prompt, thinking, text, opts.RefFileTokens)
+	if !snapshot.AlreadyEmittedCalls && !snapshot.AlreadyEmittedToolRaw {
+		turn.Error = ValidateTurn(turn, opts.ToolChoice)
+	}
+	if turn.Error != nil && len(calls) == 0 {
+		turn.StopReason = StopReasonError
+	}
+	return turn
+}
+
+func BuildUsage(model, prompt, thinking, text string, refFileTokens int) Usage {
+	inputTokens := util.CountPromptTokens(prompt, model) + refFileTokens
+	reasoningTokens := util.CountOutputTokens(thinking, model)
+	outputTokens := reasoningTokens + util.CountOutputTokens(text, model)
+	return Usage{
+		InputTokens:     inputTokens,
+		OutputTokens:    outputTokens,
+		ReasoningTokens: reasoningTokens,
+		TotalTokens:     inputTokens + outputTokens,
+	}
+}
+
+func ValidateTurn(turn Turn, policy promptcompat.ToolChoicePolicy) *OutputError {
+	if policy.IsRequired() && len(turn.ToolCalls) == 0 {
+		return &OutputError{
+			Status:  http.StatusUnprocessableEntity,
+			Message: "tool_choice requires at least one valid tool call.",
+			Code:    "tool_choice_violation",
+		}
+	}
+	if len(turn.ToolCalls) > 0 {
+		return nil
+	}
+	if strings.TrimSpace(turn.Text) != "" {
+		return nil
+	}
+	status, message, code := UpstreamEmptyOutputDetail(turn.ContentFilter, turn.Text, turn.Thinking)
+	return &OutputError{Status: status, Message: message, Code: code}
+}
+
+func UpstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
+	_ = text
+	if contentFilter {
+		return http.StatusBadRequest, "Upstream content filtered the response and returned no output.", "content_filter"
+	}
+	if strings.TrimSpace(thinking) != "" {
+		return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned reasoning without visible output.", "upstream_empty_output"
+	}
+	return http.StatusTooManyRequests, "Upstream account hit a rate limit and returned empty output.", "upstream_empty_output"
+}
+
+func ShouldRetryEmptyOutput(turn Turn, attempts, maxAttempts int) bool {
+	return attempts < maxAttempts &&
+		!turn.ContentFilter &&
+		len(turn.ToolCalls) == 0 &&
+		strings.TrimSpace(turn.Text) == "" &&
+		strings.TrimSpace(turn.Thinking) == ""
+}
+
+func FinishReason(turn Turn) string {
+	switch turn.StopReason {
+	case StopReasonToolCalls:
+		return "tool_calls"
+	case StopReasonContentFilter:
+		return "content_filter"
+	default:
+		return "stop"
+	}
+}

internal/assistantturn/turn_test.go

@@ -0,0 +1,100 @@
+package assistantturn
+
+import (
+	"testing"
+
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/sse"
+)
+
+func TestBuildTurnFromCollectedTextCitation(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{
+		Text:          "See [citation:1]",
+		CitationLinks: map[int]string{1: "https://example.com"},
+	}, BuildOptions{Model: "deepseek-v4-flash", Prompt: "prompt", SearchEnabled: true, StripReferenceMarkers: true})
+	if turn.Text != "See [1](https://example.com)" {
+		t.Fatalf("text mismatch: %q", turn.Text)
+	}
+	if turn.StopReason != StopReasonStop {
+		t.Fatalf("stop reason mismatch: %q", turn.StopReason)
+	}
+	if turn.Error != nil {
+		t.Fatalf("unexpected error: %#v", turn.Error)
+	}
+}
+
+func TestBuildTurnFromCollectedToolCall(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{
+		Text: `<tool_calls><invoke name="Write"><parameter name="content">{"x":1}</parameter></invoke></tool_calls>`,
+	}, BuildOptions{
+		ToolNames: []string{"Write"},
+		ToolsRaw: []any{map[string]any{
+			"name": "Write",
+			"input_schema": map[string]any{
+				"type": "object",
+				"properties": map[string]any{
+					"content": map[string]any{"type": "string"},
+				},
+			},
+		}},
+	})
+	if len(turn.ToolCalls) != 1 {
+		t.Fatalf("expected one tool call, got %d", len(turn.ToolCalls))
+	}
+	if turn.StopReason != StopReasonToolCalls {
+		t.Fatalf("stop reason mismatch: %q", turn.StopReason)
+	}
+	if _, ok := turn.ToolCalls[0].Input["content"].(string); !ok {
+		t.Fatalf("expected content coerced to string, got %#v", turn.ToolCalls[0].Input["content"])
+	}
+}
+
+func TestBuildTurnFromCollectedThinkingOnlyIsEmptyOutput(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{Thinking: "hidden"}, BuildOptions{})
+	if turn.Error == nil || turn.Error.Code != "upstream_empty_output" {
+		t.Fatalf("expected empty output error, got %#v", turn.Error)
+	}
+}
+
+func TestBuildTurnFromCollectedToolChoiceRequired(t *testing.T) {
+	turn := BuildTurnFromCollected(sse.CollectResult{Text: "hello"}, BuildOptions{
+		ToolChoice: promptcompat.ToolChoicePolicy{Mode: promptcompat.ToolChoiceRequired},
+	})
+	if turn.Error == nil || turn.Error.Code != "tool_choice_violation" {
+		t.Fatalf("expected tool choice violation, got %#v", turn.Error)
+	}
+}
+
+func TestBuildTurnFromStreamSnapshotUsesVisibleTextAndRawToolDetection(t *testing.T) {
+	turn := BuildTurnFromStreamSnapshot(StreamSnapshot{
+		RawText:     `<tool_calls><invoke name="Write"><parameter name="content">{"x":1}</parameter></invoke></tool_calls>`,
+		VisibleText: "",
+	}, BuildOptions{
+		ToolNames: []string{"Write"},
+		ToolsRaw: []any{map[string]any{
+			"name": "Write",
+			"schema": map[string]any{
+				"type": "object",
+				"properties": map[string]any{
+					"content": map[string]any{"type": "string"},
+				},
+			},
+		}},
+	})
+	if len(turn.ToolCalls) != 1 {
+		t.Fatalf("expected stream snapshot tool call, got %d", len(turn.ToolCalls))
+	}
+	if _, ok := turn.ToolCalls[0].Input["content"].(string); !ok {
+		t.Fatalf("expected stream snapshot schema coercion, got %#v", turn.ToolCalls[0].Input["content"])
+	}
+}
+
+func TestBuildTurnFromStreamSnapshotAlreadyEmittedToolAvoidsEmptyError(t *testing.T) {
+	turn := BuildTurnFromStreamSnapshot(StreamSnapshot{AlreadyEmittedCalls: true}, BuildOptions{})
+	if turn.Error != nil {
+		t.Fatalf("unexpected empty-output error after emitted tool call: %#v", turn.Error)
+	}
+	if turn.StopReason != StopReasonToolCalls {
+		t.Fatalf("stop reason mismatch: %q", turn.StopReason)
+	}
+}

internal/chathistory/store.go

@@ -14,6 +14,7 @@ import (
 	"github.com/google/uuid"
 
 	"ds2api/internal/config"
+	"ds2api/internal/util"
 )
 
 const (
@@ -309,8 +310,12 @@ func (s *Store) Update(id string, params UpdateParams) (Entry, error) {
 	if params.Status != "" {
 		item.Status = params.Status
 	}
-	item.ReasoningContent = params.ReasoningContent
-	item.Content = params.Content
+	if params.ReasoningContent != "" || item.ReasoningContent == "" {
+		item.ReasoningContent = params.ReasoningContent
+	}
+	if params.Content != "" || item.Content == "" {
+		item.Content = params.Content
+	}
 	item.Error = strings.TrimSpace(params.Error)
 	item.StatusCode = params.StatusCode
 	item.ElapsedMs = params.ElapsedMs
@@ -610,8 +615,8 @@ func buildPreview(item Entry) string {
 	if candidate == "" {
 		candidate = strings.TrimSpace(item.UserInput)
 	}
-	if len(candidate) > defaultPreviewAt {
-		return candidate[:defaultPreviewAt] + "..."
+	if truncated, ok := util.TruncateRunes(candidate, defaultPreviewAt); ok {
+		return truncated + "..."
 	}
 	return candidate
 }

internal/chathistory/store_test.go

@@ -8,6 +8,7 @@ import (
 	"strings"
 	"sync"
 	"testing"
+	"unicode/utf8"
 )
 
 func blockDetailDir(t *testing.T, detailDir string) func() {
@@ -105,6 +106,17 @@ func TestStoreCreatesAndPersistsEntries(t *testing.T) {
 	}
 }
 
+func TestBuildPreviewPreservesUTF8MB4Characters(t *testing.T) {
+	long := strings.Repeat("😀", defaultPreviewAt+1)
+	preview := buildPreview(Entry{Content: long})
+	if !utf8.ValidString(preview) {
+		t.Fatalf("expected valid utf-8 preview, got %q", preview)
+	}
+	if preview != strings.Repeat("😀", defaultPreviewAt)+"..." {
+		t.Fatalf("unexpected preview: %q", preview)
+	}
+}
+
 func TestStoreTrimsToConfiguredLimit(t *testing.T) {
 	path := filepath.Join(t.TempDir(), "chat_history.json")
 	store := New(path)
@@ -481,3 +493,112 @@ func TestStoreWritesOnlyChangedDetailFiles(t *testing.T) {
 		t.Fatalf("expected untouched detail file to remain byte-identical")
 	}
 }
+
+func TestUpdatePreservesContentWhenNewContentIsEmpty(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "chat_history.json")
+	store := New(path)
+
+	started, err := store.Start(StartParams{
+		CallerID:  "caller:abc",
+		Model:     "deepseek-v4-flash",
+		Stream:    true,
+		UserInput: "hello",
+	})
+	if err != nil {
+		t.Fatalf("start entry failed: %v", err)
+	}
+
+	if _, err := store.Update(started.ID, UpdateParams{
+		Status:           "streaming",
+		ReasoningContent: "let me think",
+		Content:          "I'll help you with that.",
+	}); err != nil {
+		t.Fatalf("progress update failed: %v", err)
+	}
+
+	updated, err := store.Update(started.ID, UpdateParams{
+		Status:    "success",
+		Content:   "",
+		Completed: true,
+	})
+	if err != nil {
+		t.Fatalf("success update failed: %v", err)
+	}
+
+	if updated.Content != "I'll help you with that." {
+		t.Fatalf("expected content to be preserved, got %q", updated.Content)
+	}
+	if updated.ReasoningContent != "let me think" {
+		t.Fatalf("expected reasoning content to be preserved, got %q", updated.ReasoningContent)
+	}
+
+	full, err := store.Get(started.ID)
+	if err != nil {
+		t.Fatalf("get entry failed: %v", err)
+	}
+	if full.Content != "I'll help you with that." {
+		t.Fatalf("expected persisted content to be preserved, got %q", full.Content)
+	}
+	if full.ReasoningContent != "let me think" {
+		t.Fatalf("expected persisted reasoning content to be preserved, got %q", full.ReasoningContent)
+	}
+}
+
+func TestUpdateAllowsSettingContentFromEmpty(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "chat_history.json")
+	store := New(path)
+
+	started, err := store.Start(StartParams{
+		CallerID:  "caller:abc",
+		Model:     "deepseek-v4-flash",
+		Stream:    true,
+		UserInput: "hello",
+	})
+	if err != nil {
+		t.Fatalf("start entry failed: %v", err)
+	}
+
+	updated, err := store.Update(started.ID, UpdateParams{
+		Status:  "success",
+		Content: "final answer",
+	})
+	if err != nil {
+		t.Fatalf("update failed: %v", err)
+	}
+	if updated.Content != "final answer" {
+		t.Fatalf("expected content to be set, got %q", updated.Content)
+	}
+}
+
+func TestUpdateAllowsOverwritingContentWithNewValue(t *testing.T) {
+	path := filepath.Join(t.TempDir(), "chat_history.json")
+	store := New(path)
+
+	started, err := store.Start(StartParams{
+		CallerID:  "caller:abc",
+		Model:     "deepseek-v4-flash",
+		Stream:    true,
+		UserInput: "hello",
+	})
+	if err != nil {
+		t.Fatalf("start entry failed: %v", err)
+	}
+
+	if _, err := store.Update(started.ID, UpdateParams{
+		Status:  "streaming",
+		Content: "partial",
+	}); err != nil {
+		t.Fatalf("first update failed: %v", err)
+	}
+
+	updated, err := store.Update(started.ID, UpdateParams{
+		Status:  "success",
+		Content: "final answer",
+	})
+	if err != nil {
+		t.Fatalf("second update failed: %v", err)
+	}
+	if updated.Content != "final answer" {
+		t.Fatalf("expected content to be overwritten, got %q", updated.Content)
+	}
+}

internal/completionruntime/nonstream.go

@@ -0,0 +1,170 @@
+package completionruntime
+
+import (
+	"context"
+	"fmt"
+	"io"
+	"net/http"
+	"strings"
+
+	"ds2api/internal/assistantturn"
+	"ds2api/internal/auth"
+	"ds2api/internal/config"
+	"ds2api/internal/httpapi/openai/shared"
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/sse"
+)
+
+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)
+	CallCompletion(ctx context.Context, a *auth.RequestAuth, payload map[string]any, powResp string, maxAttempts int) (*http.Response, error)
+}
+
+type Options struct {
+	StripReferenceMarkers bool
+	MaxAttempts           int
+	RetryEnabled          bool
+	RetryMaxAttempts      int
+}
+
+type NonStreamResult struct {
+	SessionID string
+	Payload   map[string]any
+	Turn      assistantturn.Turn
+	Attempts  int
+}
+
+type StartResult struct {
+	SessionID string
+	Payload   map[string]any
+	Pow       string
+	Response  *http.Response
+}
+
+func StartCompletion(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, stdReq promptcompat.StandardRequest, opts Options) (StartResult, *assistantturn.OutputError) {
+	maxAttempts := opts.MaxAttempts
+	if maxAttempts <= 0 {
+		maxAttempts = 3
+	}
+	sessionID, err := ds.CreateSession(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{}, authOutputError(a)
+	}
+	pow, err := ds.GetPow(ctx, a, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID}, &assistantturn.OutputError{Status: http.StatusUnauthorized, Message: "Failed to get PoW (invalid token or unknown error).", Code: "error"}
+	}
+	payload := stdReq.CompletionPayload(sessionID)
+	resp, err := ds.CallCompletion(ctx, a, payload, pow, maxAttempts)
+	if err != nil {
+		return StartResult{SessionID: sessionID, Payload: payload, Pow: pow}, &assistantturn.OutputError{Status: http.StatusInternalServerError, Message: "Failed to get completion.", Code: "error"}
+	}
+	return StartResult{SessionID: sessionID, Payload: payload, Pow: pow, Response: resp}, nil
+}
+
+func ExecuteNonStreamWithRetry(ctx context.Context, ds DeepSeekCaller, a *auth.RequestAuth, stdReq promptcompat.StandardRequest, opts Options) (NonStreamResult, *assistantturn.OutputError) {
+	start, startErr := StartCompletion(ctx, ds, a, stdReq, opts)
+	if startErr != nil {
+		return NonStreamResult{SessionID: start.SessionID, Payload: start.Payload}, startErr
+	}
+	maxAttempts := opts.MaxAttempts
+	if maxAttempts <= 0 {
+		maxAttempts = 3
+	}
+	sessionID := start.SessionID
+	payload := start.Payload
+	pow := start.Pow
+
+	attempts := 0
+	currentResp := start.Response
+	usagePrompt := stdReq.PromptTokenText
+	accumulatedThinking := ""
+	accumulatedRawThinking := ""
+	accumulatedToolDetectionThinking := ""
+	for {
+		turn, outErr := collectAttempt(currentResp, stdReq, usagePrompt, opts)
+		if outErr != nil {
+			return NonStreamResult{SessionID: sessionID, Payload: payload, Attempts: attempts}, outErr
+		}
+		accumulatedThinking += sse.TrimContinuationOverlap(accumulatedThinking, turn.Thinking)
+		accumulatedRawThinking += sse.TrimContinuationOverlap(accumulatedRawThinking, turn.RawThinking)
+		accumulatedToolDetectionThinking += sse.TrimContinuationOverlap(accumulatedToolDetectionThinking, turn.DetectionThinking)
+		turn.Thinking = accumulatedThinking
+		turn.RawThinking = accumulatedRawThinking
+		turn.DetectionThinking = accumulatedToolDetectionThinking
+		turn = assistantturn.BuildTurnFromCollected(sse.CollectResult{
+			Text:                  turn.RawText,
+			Thinking:              turn.RawThinking,
+			ToolDetectionThinking: turn.DetectionThinking,
+			ContentFilter:         turn.ContentFilter,
+			CitationLinks:         turn.CitationLinks,
+			ResponseMessageID:     turn.ResponseMessageID,
+		}, buildOptions(stdReq, usagePrompt, opts))
+
+		retryMax := opts.RetryMaxAttempts
+		if retryMax <= 0 {
+			retryMax = shared.EmptyOutputRetryMaxAttempts()
+		}
+		if !opts.RetryEnabled || !assistantturn.ShouldRetryEmptyOutput(turn, attempts, retryMax) {
+			return NonStreamResult{SessionID: sessionID, Payload: payload, Turn: turn, Attempts: attempts}, turn.Error
+		}
+
+		attempts++
+		config.Logger.Info("[completion_runtime_empty_retry] attempting synthetic retry", "surface", stdReq.Surface, "stream", false, "retry_attempt", attempts, "parent_message_id", turn.ResponseMessageID)
+		retryPow, powErr := ds.GetPow(ctx, a, maxAttempts)
+		if powErr != nil {
+			config.Logger.Warn("[completion_runtime_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", stdReq.Surface, "retry_attempt", attempts, "error", powErr)
+			retryPow = pow
+		}
+		retryPayload := shared.ClonePayloadForEmptyOutputRetry(payload, turn.ResponseMessageID)
+		nextResp, err := ds.CallCompletion(ctx, a, retryPayload, retryPow, maxAttempts)
+		if err != nil {
+			return NonStreamResult{SessionID: sessionID, Payload: payload, Turn: turn, Attempts: attempts}, &assistantturn.OutputError{Status: http.StatusInternalServerError, Message: "Failed to get completion.", Code: "error"}
+		}
+		usagePrompt = shared.UsagePromptWithEmptyOutputRetry(usagePrompt, attempts)
+		currentResp = nextResp
+	}
+}
+
+func collectAttempt(resp *http.Response, stdReq promptcompat.StandardRequest, usagePrompt string, opts Options) (assistantturn.Turn, *assistantturn.OutputError) {
+	defer func() {
+		if err := resp.Body.Close(); err != nil {
+			config.Logger.Warn("[completion_runtime] response body close failed", "surface", stdReq.Surface, "error", err)
+		}
+	}()
+	if resp.StatusCode != http.StatusOK {
+		body, _ := io.ReadAll(resp.Body)
+		message := strings.TrimSpace(string(body))
+		if message == "" {
+			message = http.StatusText(resp.StatusCode)
+		}
+		return assistantturn.Turn{}, &assistantturn.OutputError{Status: resp.StatusCode, Message: message, Code: "error"}
+	}
+	result := sse.CollectStream(resp, stdReq.Thinking, false)
+	return assistantturn.BuildTurnFromCollected(result, buildOptions(stdReq, usagePrompt, opts)), nil
+}
+
+func buildOptions(stdReq promptcompat.StandardRequest, prompt string, opts Options) assistantturn.BuildOptions {
+	return assistantturn.BuildOptions{
+		Model:                 stdReq.ResponseModel,
+		Prompt:                prompt,
+		RefFileTokens:         stdReq.RefFileTokens,
+		SearchEnabled:         stdReq.Search,
+		StripReferenceMarkers: opts.StripReferenceMarkers,
+		ToolNames:             stdReq.ToolNames,
+		ToolsRaw:              stdReq.ToolsRaw,
+		ToolChoice:            stdReq.ToolChoice,
+	}
+}
+
+func authOutputError(a *auth.RequestAuth) *assistantturn.OutputError {
+	if a != nil && a.UseConfigToken {
+		return &assistantturn.OutputError{Status: http.StatusUnauthorized, Message: "Account token is invalid. Please re-login the account in admin.", Code: "error"}
+	}
+	return &assistantturn.OutputError{Status: http.StatusUnauthorized, Message: "Invalid token. If this should be a DS2API key, add it to config.keys first.", Code: "error"}
+}
+
+func Errorf(status int, format string, args ...any) *assistantturn.OutputError {
+	return &assistantturn.OutputError{Status: status, Message: fmt.Sprintf(format, args...), Code: "error"}
+}

internal/completionruntime/nonstream_test.go

@@ -0,0 +1,120 @@
+package completionruntime
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+
+	"ds2api/internal/auth"
+	"ds2api/internal/promptcompat"
+)
+
+type fakeDeepSeekCaller struct {
+	responses []*http.Response
+	payloads  []map[string]any
+}
+
+func (f *fakeDeepSeekCaller) CreateSession(context.Context, *auth.RequestAuth, int) (string, error) {
+	return "session-1", nil
+}
+
+func (f *fakeDeepSeekCaller) GetPow(context.Context, *auth.RequestAuth, int) (string, error) {
+	return "pow", nil
+}
+
+func (f *fakeDeepSeekCaller) CallCompletion(_ context.Context, _ *auth.RequestAuth, payload map[string]any, _ string, _ int) (*http.Response, error) {
+	f.payloads = append(f.payloads, payload)
+	if len(f.responses) == 0 {
+		return sseHTTPResponse(http.StatusOK, `data: {"p":"response/content","v":"fallback"}`), nil
+	}
+	resp := f.responses[0]
+	f.responses = f.responses[1:]
+	return resp, nil
+}
+
+func TestExecuteNonStreamWithRetryBuildsCanonicalTurn(t *testing.T) {
+	ds := &fakeDeepSeekCaller{responses: []*http.Response{sseHTTPResponse(
+		http.StatusOK,
+		`data: {"response_message_id":42,"p":"response/content","v":"<tool_calls><invoke name=\"Write\"><parameter name=\"content\">{\"x\":1}</parameter></invoke></tool_calls>"}`,
+	)}}
+	stdReq := promptcompat.StandardRequest{
+		Surface:         "test",
+		ResponseModel:   "deepseek-v4-flash",
+		PromptTokenText: "prompt",
+		FinalPrompt:     "final prompt",
+		ToolNames:       []string{"Write"},
+		ToolsRaw: []any{map[string]any{
+			"name": "Write",
+			"input_schema": map[string]any{
+				"type": "object",
+				"properties": map[string]any{
+					"content": map[string]any{"type": "string"},
+				},
+			},
+		}},
+	}
+
+	result, outErr := ExecuteNonStreamWithRetry(context.Background(), ds, &auth.RequestAuth{}, stdReq, Options{})
+	if outErr != nil {
+		t.Fatalf("unexpected output error: %#v", outErr)
+	}
+	if result.SessionID != "session-1" {
+		t.Fatalf("session mismatch: %q", result.SessionID)
+	}
+	if got := result.Turn.ResponseMessageID; got != 42 {
+		t.Fatalf("response message id mismatch: %d", got)
+	}
+	if len(result.Turn.ToolCalls) != 1 {
+		t.Fatalf("expected one tool call, got %d", len(result.Turn.ToolCalls))
+	}
+	if _, ok := result.Turn.ToolCalls[0].Input["content"].(string); !ok {
+		t.Fatalf("expected schema-normalized string argument, got %#v", result.Turn.ToolCalls[0].Input["content"])
+	}
+	if result.Turn.Usage.InputTokens == 0 || result.Turn.Usage.TotalTokens == 0 {
+		t.Fatalf("expected usage to be populated, got %#v", result.Turn.Usage)
+	}
+}
+
+func TestExecuteNonStreamWithRetryUsesParentMessageForEmptyRetry(t *testing.T) {
+	ds := &fakeDeepSeekCaller{responses: []*http.Response{
+		sseHTTPResponse(http.StatusOK, `data: {"response_message_id":77,"p":"response/status","v":"FINISHED"}`),
+		sseHTTPResponse(http.StatusOK, `data: {"response_message_id":78,"p":"response/content","v":"ok"}`),
+	}}
+	stdReq := promptcompat.StandardRequest{
+		Surface:         "test",
+		ResponseModel:   "deepseek-v4-flash",
+		PromptTokenText: "prompt",
+		FinalPrompt:     "final prompt",
+	}
+
+	result, outErr := ExecuteNonStreamWithRetry(context.Background(), ds, &auth.RequestAuth{}, stdReq, Options{RetryEnabled: true})
+	if outErr != nil {
+		t.Fatalf("unexpected output error: %#v", outErr)
+	}
+	if result.Attempts != 1 {
+		t.Fatalf("expected one retry, got %d", result.Attempts)
+	}
+	if len(ds.payloads) != 2 {
+		t.Fatalf("expected two completion calls, got %d", len(ds.payloads))
+	}
+	if got := ds.payloads[1]["parent_message_id"]; got != 77 {
+		t.Fatalf("retry parent_message_id mismatch: %#v", got)
+	}
+	if result.Turn.Text != "ok" {
+		t.Fatalf("retry text mismatch: %q", result.Turn.Text)
+	}
+}
+
+func sseHTTPResponse(status int, lines ...string) *http.Response {
+	body := strings.Join(lines, "\n")
+	if !strings.HasSuffix(body, "\n") {
+		body += "\n"
+	}
+	return &http.Response{
+		StatusCode: status,
+		Header:     make(http.Header),
+		Body:       io.NopCloser(strings.NewReader(body)),
+	}
+}

internal/config/config_edge_test.go

@@ -79,13 +79,20 @@ func TestGetModelConfigDeepSeekExpertReasonerSearch(t *testing.T) {
 	}
 }
 
-func TestGetModelConfigDeepSeekVisionReasonerSearch(t *testing.T) {
-	thinking, search, ok := GetModelConfig("deepseek-v4-vision-search")
+func TestGetModelConfigDeepSeekVision(t *testing.T) {
+	thinking, search, ok := GetModelConfig("deepseek-v4-vision")
 	if !ok {
-		t.Fatal("expected ok for deepseek-v4-vision-search")
+		t.Fatal("expected ok for deepseek-v4-vision")
 	}
-	if !thinking || !search {
-		t.Fatalf("expected both true, got thinking=%v search=%v", thinking, search)
+	if !thinking || search {
+		t.Fatalf("expected thinking=true search=false, got thinking=%v search=%v", thinking, search)
+	}
+}
+
+func TestGetModelConfigDeepSeekVisionSearchUnsupported(t *testing.T) {
+	_, _, ok := GetModelConfig("deepseek-v4-vision-search")
+	if ok {
+		t.Fatal("expected deepseek-v4-vision-search to be unsupported")
 	}
 }
 
@@ -748,18 +755,16 @@ func TestOpenAIModelsResponse(t *testing.T) {
 		t.Fatal("expected non-empty models list")
 	}
 	expected := map[string]bool{
-		"deepseek-v4-flash":                    false,
-		"deepseek-v4-flash-nothinking":         false,
-		"deepseek-v4-pro":                      false,
-		"deepseek-v4-pro-nothinking":           false,
-		"deepseek-v4-flash-search":             false,
-		"deepseek-v4-flash-search-nothinking":  false,
-		"deepseek-v4-pro-search":               false,
-		"deepseek-v4-pro-search-nothinking":    false,
-		"deepseek-v4-vision":                   false,
-		"deepseek-v4-vision-nothinking":        false,
-		"deepseek-v4-vision-search":            false,
-		"deepseek-v4-vision-search-nothinking": false,
+		"deepseek-v4-flash":                   false,
+		"deepseek-v4-flash-nothinking":        false,
+		"deepseek-v4-pro":                     false,
+		"deepseek-v4-pro-nothinking":          false,
+		"deepseek-v4-flash-search":            false,
+		"deepseek-v4-flash-search-nothinking": false,
+		"deepseek-v4-pro-search":              false,
+		"deepseek-v4-pro-search-nothinking":   false,
+		"deepseek-v4-vision":                  false,
+		"deepseek-v4-vision-nothinking":       false,
 	}
 	for _, model := range data {
 		if _, ok := expected[model.ID]; ok {

internal/config/config_test.go

@@ -144,6 +144,44 @@ func TestLoadStoreIgnoresLegacyConfigJSONEnv(t *testing.T) {
 	}
 }
 
+func TestExplicitMissingConfigPathBootstrapsEmptyFileBackedStore(t *testing.T) {
+	path := t.TempDir() + "/config.json"
+
+	t.Setenv("DS2API_CONFIG_JSON", "")
+	t.Setenv("DS2API_CONFIG_PATH", path)
+
+	store, err := LoadStoreWithError()
+	if err != nil {
+		t.Fatalf("expected missing explicit config path to bootstrap, got: %v", err)
+	}
+	if store.IsEnvBacked() {
+		t.Fatal("expected bootstrap store to be file-backed")
+	}
+	if store.ConfigPath() != path {
+		t.Fatalf("ConfigPath() = %q, want %q", store.ConfigPath(), path)
+	}
+	if len(store.Keys()) != 0 || len(store.Accounts()) != 0 {
+		t.Fatalf("expected empty bootstrap config, got keys=%d accounts=%d", len(store.Keys()), len(store.Accounts()))
+	}
+	if _, statErr := os.Stat(path); !errors.Is(statErr, os.ErrNotExist) {
+		t.Fatalf("expected bootstrap not to create config until first save, stat err=%v", statErr)
+	}
+
+	if err := store.Update(func(c *Config) error {
+		c.Keys = []string{"first-key"}
+		return nil
+	}); err != nil {
+		t.Fatalf("update should persist bootstrap config: %v", err)
+	}
+	content, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("expected first update to write config: %v", err)
+	}
+	if !strings.Contains(string(content), "first-key") {
+		t.Fatalf("expected saved config to contain first key, got: %s", content)
+	}
+}
+
 func TestEnvBackedStoreWritebackBootstrapsMissingConfigFile(t *testing.T) {
 	tmp, err := os.CreateTemp(t.TempDir(), "config-*.json")
 	if err != nil {

internal/config/model_alias_test.go

@@ -144,10 +144,17 @@ func TestResolveModelCustomAliasToExpert(t *testing.T) {
 
 func TestResolveModelCustomAliasToVision(t *testing.T) {
 	got, ok := ResolveModel(mockModelAliasReader{
-		"my-vision-model": "deepseek-v4-vision-search",
+		"my-vision-model": "deepseek-v4-vision",
 	}, "my-vision-model")
-	if !ok || got != "deepseek-v4-vision-search" {
-		t.Fatalf("expected alias -> deepseek-v4-vision-search, got ok=%v model=%q", ok, got)
+	if !ok || got != "deepseek-v4-vision" {
+		t.Fatalf("expected alias -> deepseek-v4-vision, got ok=%v model=%q", ok, got)
+	}
+}
+
+func TestResolveModelHeuristicVisionIgnoresSearchSuffix(t *testing.T) {
+	got, ok := ResolveModel(nil, "gemini-vision-search")
+	if !ok || got != "deepseek-v4-vision" {
+		t.Fatalf("expected heuristic vision alias to resolve without search variant, got ok=%v model=%q", ok, got)
 	}
 }
 

internal/config/models.go

@@ -22,7 +22,6 @@ var deepSeekBaseModels = []ModelInfo{
 	{ID: "deepseek-v4-flash-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 	{ID: "deepseek-v4-pro-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 	{ID: "deepseek-v4-vision", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
-	{ID: "deepseek-v4-vision-search", Object: "model", Created: 1677610602, OwnedBy: "deepseek", Permission: []any{}},
 }
 
 var DeepSeekModels = appendNoThinkingVariants(deepSeekBaseModels)
@@ -67,7 +66,7 @@ func GetModelConfig(model string) (thinking bool, search bool, ok bool) {
 	switch baseModel {
 	case "deepseek-v4-flash", "deepseek-v4-pro", "deepseek-v4-vision":
 		return !noThinking, false, true
-	case "deepseek-v4-flash-search", "deepseek-v4-pro-search", "deepseek-v4-vision-search":
+	case "deepseek-v4-flash-search", "deepseek-v4-pro-search":
 		return !noThinking, true, true
 	default:
 		return false, false, false
@@ -81,7 +80,7 @@ func GetModelType(model string) (modelType string, ok bool) {
 		return "default", true
 	case "deepseek-v4-pro", "deepseek-v4-pro-search":
 		return "expert", true
-	case "deepseek-v4-vision", "deepseek-v4-vision-search":
+	case "deepseek-v4-vision":
 		return "vision", true
 	default:
 		return "", false
@@ -359,8 +358,6 @@ func resolveCanonicalModel(aliases map[string]string, model string) (string, boo
 	useSearch := strings.Contains(model, "search")
 
 	switch {
-	case useVision && useSearch:
-		return "deepseek-v4-vision-search", true
 	case useVision:
 		return "deepseek-v4-vision", true
 	case useReasoner && useSearch:

internal/config/paths.go

@@ -31,13 +31,20 @@ func ResolvePath(envKey, defaultRel string) string {
 
 func ConfigPath() string {
 	if strings.TrimSpace(os.Getenv("DS2API_CONFIG_PATH")) == "" && BaseDir() == "/app" {
-		// Official container images commonly run from /app where filesystem may be read-only.
-		// Prefer /data default so deployments can persist config/token state by mounting a volume.
-		return "/data/config.json"
+		return containerDefaultConfigPath()
 	}
 	return ResolvePath("DS2API_CONFIG_PATH", "config.json")
 }
 
+func containerDefaultConfigPath() string {
+	// Container images run as non-root by default. Only use /data when mounted/provisioned.
+	// Otherwise keep /app/config.json so admin-side save does not fail on MkdirAll("/data").
+	if st, err := os.Stat("/data"); err == nil && st.IsDir() {
+		return "/data/config.json"
+	}
+	return "/app/config.json"
+}
+
 func legacyContainerConfigPath() string {
 	return "/app/config.json"
 }

internal/config/paths_test.go

@@ -0,0 +1,28 @@
+package config
+
+import (
+	"os"
+	"testing"
+)
+
+func TestContainerDefaultConfigPath(t *testing.T) {
+	t.Run("fallback to /app when /data is missing", func(t *testing.T) {
+		// This test environment does not guarantee a writable/mounted /data.
+		// If /data is absent we must keep /app fallback to avoid persistence failures.
+		if _, err := os.Stat("/data"); err == nil {
+			t.Skip("/data exists in this environment; cannot validate missing-/data fallback")
+		}
+		if got := containerDefaultConfigPath(); got != "/app/config.json" {
+			t.Fatalf("containerDefaultConfigPath() = %q, want %q", got, "/app/config.json")
+		}
+	})
+
+	t.Run("prefer /data when /data directory exists", func(t *testing.T) {
+		if _, err := os.Stat("/data"); err != nil {
+			t.Skip("/data does not exist in this environment")
+		}
+		if got := containerDefaultConfigPath(); got != "/data/config.json" {
+			t.Fatalf("containerDefaultConfigPath() = %q, want %q", got, "/data/config.json")
+		}
+	})
+}

internal/config/store.go

@@ -52,11 +52,12 @@ func loadStore() (*Store, error) {
 
 func loadConfig() (Config, bool, error) {
 	rawCfg := strings.TrimSpace(os.Getenv("DS2API_CONFIG_JSON"))
+	path := ConfigPath()
 	if rawCfg != "" {
 		cfg, err := parseConfigString(rawCfg)
 		if err != nil {
 			if !IsVercel() && envWritebackEnabled() {
-				if fileCfg, fileErr := loadConfigFromFile(ConfigPath()); fileErr == nil {
+				if fileCfg, fileErr := loadConfigFromFile(path); fileErr == nil {
 					return fileCfg, false, nil
 				}
 			}
@@ -67,7 +68,7 @@ func loadConfig() (Config, bool, error) {
 		if IsVercel() || !envWritebackEnabled() {
 			return cfg, true, err
 		}
-		content, fileErr := os.ReadFile(ConfigPath())
+		content, fileErr := os.ReadFile(path)
 		if fileErr == nil {
 			var fileCfg Config
 			if unmarshalErr := json.Unmarshal(content, &fileCfg); unmarshalErr == nil {
@@ -79,7 +80,7 @@ func loadConfig() (Config, bool, error) {
 			if validateErr := ValidateConfig(cfg); validateErr != nil {
 				return cfg, true, validateErr
 			}
-			if writeErr := writeConfigFile(ConfigPath(), cfg.Clone()); writeErr == nil {
+			if writeErr := writeConfigFile(path, cfg.Clone()); writeErr == nil {
 				return cfg, false, err
 			} else {
 				Logger.Warn("[config] env writeback bootstrap failed", "error", writeErr)
@@ -87,7 +88,7 @@ func loadConfig() (Config, bool, error) {
 		}
 		return cfg, true, err
 	}
-	cfg, err := loadConfigFromFile(ConfigPath())
+	cfg, err := loadConfigFromFile(path)
 	if err != nil {
 		if shouldTryLegacyContainerConfigPath() {
 			legacyPath := legacyContainerConfigPath()
@@ -100,6 +101,10 @@ func loadConfig() (Config, bool, error) {
 			// Vercel may start without writable/present config; keep in-memory bootstrap config.
 			return Config{}, true, nil
 		}
+		if shouldBootstrapMissingConfigFile(err) {
+			Logger.Warn("[config] config file missing; starting with empty file-backed config", "path", path)
+			return Config{}, false, nil
+		}
 		return Config{}, false, err
 	}
 	if IsVercel() {
@@ -109,6 +114,10 @@ func loadConfig() (Config, bool, error) {
 	return cfg, false, nil
 }
 
+func shouldBootstrapMissingConfigFile(err error) bool {
+	return errors.Is(err, os.ErrNotExist) && strings.TrimSpace(os.Getenv("DS2API_CONFIG_PATH")) != ""
+}
+
 func loadConfigFromFile(path string) (Config, error) {
 	content, err := os.ReadFile(path)
 	if err != nil {

internal/deepseek/client/client_continue.go

@@ -7,6 +7,7 @@ import (
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	"encoding/json"
 	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"strings"
@@ -27,7 +28,7 @@ type continueState struct {
 }
 
 // wrapCompletionWithAutoContinue wraps the completion response body so that
-// if the upstream indicates the response is incomplete (WIP / INCOMPLETE /
+// if the upstream indicates the response is incomplete (INCOMPLETE /
 // AUTO_CONTINUE), ds2api will automatically call the DeepSeek continue
 // endpoint and splice the continuation SSE stream onto the original.
 // The caller sees a single, seamless SSE stream.
@@ -132,33 +133,51 @@ func pumpAutoContinue(ctx context.Context, pw *io.PipeWriter, initial io.ReadClo
 // sentinels are consumed (not forwarded) so that the downstream only sees
 // one final [DONE] at the very end.
 func streamBodyWithContinueState(ctx context.Context, pw *io.PipeWriter, body io.Reader, state *continueState) (bool, error) {
-	scanner := bufio.NewScanner(body)
-	scanner.Buffer(make([]byte, 0, 64*1024), 2*1024*1024)
+	reader := bufio.NewReaderSize(body, 64*1024)
 	hadDone := false
-	for scanner.Scan() {
+	for {
 		select {
 		case <-ctx.Done():
 			return hadDone, ctx.Err()
 		default:
 		}
-		line := append([]byte{}, scanner.Bytes()...)
-		trimmed := strings.TrimSpace(string(line))
-		if trimmed == "" {
-			continue
-		}
-		if strings.HasPrefix(trimmed, "data:") {
-			data := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
-			if data == "[DONE]" {
-				hadDone = true
-				continue
+		line, err := reader.ReadBytes('\n')
+		if len(line) == 0 && err != nil {
+			if err == io.EOF {
+				return hadDone, nil
 			}
-			state.observe(data)
+			return hadDone, err
 		}
-		if _, err := io.Copy(pw, bytes.NewReader(append(line, '\n'))); err != nil {
+		trimmed := strings.TrimSpace(string(line))
+		if trimmed != "" {
+			if strings.HasPrefix(trimmed, "data:") {
+				data := strings.TrimSpace(strings.TrimPrefix(trimmed, "data:"))
+				if data == "[DONE]" {
+					hadDone = true
+					if err != nil && err != io.EOF {
+						return hadDone, err
+					}
+					if err == io.EOF {
+						return hadDone, nil
+					}
+					continue
+				}
+				state.observe(data)
+			}
+			if !strings.HasSuffix(string(line), "\n") {
+				line = append(line, '\n')
+			}
+			if _, copyErr := io.Copy(pw, bytes.NewReader(line)); copyErr != nil {
+				return hadDone, copyErr
+			}
+		}
+		if err != nil {
+			if err == io.EOF {
+				return hadDone, nil
+			}
 			return hadDone, err
 		}
 	}
-	return hadDone, scanner.Err()
 }
 
 // observe extracts continue-relevant signals from an SSE JSON chunk.
@@ -174,49 +193,100 @@ func (s *continueState) observe(data string) {
 	if id := intFrom(chunk["response_message_id"]); id > 0 {
 		s.responseMessageID = id
 	}
-	// Path-based status: {"p": "response/status", "v": "FINISHED"}
-	if p, _ := chunk["p"].(string); p == "response/status" {
-		if status, _ := chunk["v"].(string); status != "" {
-			s.lastStatus = strings.TrimSpace(status)
-			if strings.EqualFold(s.lastStatus, "FINISHED") {
-				s.finished = true
-			}
-		}
+	s.observeDirectPatch(asString(chunk["p"]), chunk["v"])
+	if p, _ := chunk["p"].(string); p == "response" {
+		s.observeBatchPatches("response", chunk["v"])
+	} else {
+		s.observeBatchPatches("", chunk["v"])
 	}
-	// Nested v.response
-	v, _ := chunk["v"].(map[string]any)
-	if response, _ := v["response"].(map[string]any); response != nil {
-		if id := intFrom(response["message_id"]); id > 0 {
-			s.responseMessageID = id
-		}
-		if status, _ := response["status"].(string); status != "" {
-			s.lastStatus = strings.TrimSpace(status)
-			if strings.EqualFold(s.lastStatus, "FINISHED") {
-				s.finished = true
-			}
-		}
-		if autoContinue, ok := response["auto_continue"].(bool); ok && autoContinue {
+	if v, _ := chunk["v"].(map[string]any); v != nil {
+		s.observeResponseObject(v["response"])
+	}
+	if message, _ := chunk["message"].(map[string]any); message != nil {
+		s.observeResponseObject(message["response"])
+	}
+}
+
+func (s *continueState) observeDirectPatch(path string, value any) {
+	if s == nil {
+		return
+	}
+	switch strings.Trim(strings.TrimSpace(path), "/") {
+	case "response/status", "status", "response/quasi_status", "quasi_status":
+		s.setStatus(asString(value))
+	case "response/auto_continue", "auto_continue":
+		if v, ok := value.(bool); ok && v {
 			s.lastStatus = "AUTO_CONTINUE"
 		}
 	}
-	// Nested message.response
-	if message, _ := chunk["message"].(map[string]any); message != nil {
-		if response, _ := message["response"].(map[string]any); response != nil {
-			if id := intFrom(response["message_id"]); id > 0 {
-				s.responseMessageID = id
-			}
-			if status, _ := response["status"].(string); status != "" {
-				s.lastStatus = strings.TrimSpace(status)
-				if strings.EqualFold(s.lastStatus, "FINISHED") {
-					s.finished = true
-				}
+}
+
+func (s *continueState) observeResponseObject(raw any) {
+	if s == nil {
+		return
+	}
+	response, _ := raw.(map[string]any)
+	if response == nil {
+		return
+	}
+	if id := intFrom(response["message_id"]); id > 0 {
+		s.responseMessageID = id
+	}
+	s.setStatus(asString(response["status"]))
+	if autoContinue, ok := response["auto_continue"].(bool); ok && autoContinue {
+		s.lastStatus = "AUTO_CONTINUE"
+	}
+}
+
+func (s *continueState) observeBatchPatches(parentPath string, raw any) {
+	if s == nil {
+		return
+	}
+	patches, ok := raw.([]any)
+	if !ok {
+		return
+	}
+	for _, patch := range patches {
+		m, ok := patch.(map[string]any)
+		if !ok {
+			continue
+		}
+		path := strings.TrimSpace(asString(m["p"]))
+		if path == "" {
+			continue
+		}
+		fullPath := path
+		if parent := strings.Trim(strings.TrimSpace(parentPath), "/"); parent != "" && !strings.Contains(path, "/") {
+			fullPath = parent + "/" + path
+		}
+		switch strings.Trim(strings.TrimSpace(fullPath), "/") {
+		case "response/status", "status", "response/quasi_status", "quasi_status":
+			s.setStatus(asString(m["v"]))
+		case "response/auto_continue", "auto_continue":
+			if v, ok := m["v"].(bool); ok && v {
+				s.lastStatus = "AUTO_CONTINUE"
 			}
 		}
 	}
 }
 
-// shouldContinue returns true when the upstream indicates the response is
-// not yet finished and we have enough information to issue a continue request.
+func (s *continueState) setStatus(status string) {
+	if s == nil {
+		return
+	}
+	normalized := strings.TrimSpace(status)
+	if normalized == "" {
+		return
+	}
+	s.lastStatus = normalized
+	if strings.EqualFold(normalized, "FINISHED") || strings.EqualFold(normalized, "CONTENT_FILTER") {
+		s.finished = true
+	}
+}
+
+// shouldContinue returns true when the upstream explicitly indicates the
+// response is incomplete and we have enough information to issue a continue
+// request. Plain WIP is not sufficient because normal streams begin in WIP.
 func (s *continueState) shouldContinue() bool {
 	if s == nil {
 		return false
@@ -225,7 +295,7 @@ func (s *continueState) shouldContinue() bool {
 		return false
 	}
 	switch strings.ToUpper(strings.TrimSpace(s.lastStatus)) {
-	case "WIP", "INCOMPLETE", "AUTO_CONTINUE":
+	case "INCOMPLETE", "AUTO_CONTINUE":
 		return true
 	default:
 		return false
@@ -241,3 +311,19 @@ func (s *continueState) prepareForNextRound() {
 	s.finished = false
 	s.lastStatus = ""
 }
+
+func asString(v any) string {
+	if v == nil {
+		return ""
+	}
+	switch x := v.(type) {
+	case string:
+		return x
+	default:
+		s := strings.TrimSpace(strings.ReplaceAll(strings.TrimSpace(fmt.Sprint(v)), "\u0000", ""))
+		if s == "<nil>" {
+			return ""
+		}
+		return s
+	}
+}

internal/deepseek/client/client_continue_test.go

@@ -8,6 +8,7 @@ import (
 	"io"
 	"net/http"
 	"strings"
+	"sync/atomic"
 	"testing"
 
 	"ds2api/internal/auth"
@@ -124,6 +125,146 @@ func TestCallCompletionAutoContinueThreadsPowHeader(t *testing.T) {
 	}
 }
 
+func TestAutoContinueDoesNotTriggerOnPlainWIPWithoutExplicitContinuationSignal(t *testing.T) {
+	initialBody := strings.Join([]string{
+		`data: {"response_message_id":321,"v":{"response":{"message_id":321,"status":"WIP","auto_continue":false}}}`,
+		`data: [DONE]`,
+	}, "\n") + "\n"
+
+	var continueCalls atomic.Int32
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		continueCalls.Add(1)
+		return nil, errors.New("continue should not have been called")
+	})
+	defer func() { _ = body.Close() }()
+
+	out, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if continueCalls.Load() != 0 {
+		t.Fatalf("expected no continue calls, got %d", continueCalls.Load())
+	}
+	if !bytes.Contains(out, []byte(`"status":"WIP"`)) || !bytes.Contains(out, []byte(`data: [DONE]`)) {
+		t.Fatalf("expected original body to pass through unchanged, got=%s", string(out))
+	}
+}
+
+func TestAutoContinuePassesThroughLongSingleSSELine(t *testing.T) {
+	payload := strings.Repeat("x", 2*1024*1024+4096)
+	initialBody := `data: {"p":"response/content","v":"` + payload + `"}` + "\n" +
+		`data: [DONE]` + "\n"
+
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		return nil, errors.New("continue should not have been called")
+	})
+	defer func() { _ = body.Close() }()
+
+	out, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if !bytes.Contains(out, []byte(payload)) {
+		t.Fatalf("expected long SSE payload to pass through, got len=%d want payload len=%d", len(out), len(payload))
+	}
+	if !bytes.Contains(out, []byte(`data: [DONE]`)) {
+		t.Fatalf("expected final DONE sentinel in body, got len=%d", len(out))
+	}
+}
+
+func TestAutoContinueTriggersOnDirectQuasiStatusIncomplete(t *testing.T) {
+	initialBody := strings.Join([]string{
+		`data: {"response_message_id":321,"p":"response/content","v":"<tool_calls><invoke name=\"write_file\"><parameter name=\"content\"><![CDATA[part-one"}`,
+		`data: {"p":"response/quasi_status","v":"INCOMPLETE"}`,
+		`data: [DONE]`,
+	}, "\n") + "\n"
+
+	var continueCalls atomic.Int32
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		continueCalls.Add(1)
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     make(http.Header),
+			Body: io.NopCloser(strings.NewReader(
+				`data: {"response_message_id":322,"p":"response/content","v":"-part-two]]></parameter></invoke></tool_calls>"}` + "\n" +
+					`data: {"p":"response/status","v":"FINISHED"}` + "\n" +
+					`data: [DONE]` + "\n",
+			)),
+		}, nil
+	})
+	defer func() { _ = body.Close() }()
+
+	out, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if continueCalls.Load() != 1 {
+		t.Fatalf("expected exactly one continue call, got %d", continueCalls.Load())
+	}
+	if !bytes.Contains(out, []byte("part-one")) || !bytes.Contains(out, []byte("-part-two")) {
+		t.Fatalf("expected continued tool content in body, got=%s", string(out))
+	}
+}
+
+func TestAutoContinueTriggersOnResponseBatchQuasiStatusIncomplete(t *testing.T) {
+	initialBody := strings.Join([]string{
+		`data: {"response_message_id":321,"v":{"response":{"message_id":321,"status":"WIP","auto_continue":false}}}`,
+		`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":2413},{"p":"quasi_status","v":"INCOMPLETE"}]}`,
+		`data: [DONE]`,
+	}, "\n") + "\n"
+
+	var continueCalls atomic.Int32
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		continueCalls.Add(1)
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     make(http.Header),
+			Body: io.NopCloser(strings.NewReader(
+				`data: {"response_message_id":322,"p":"response/status","v":"FINISHED"}` + "\n" +
+					`data: [DONE]` + "\n",
+			)),
+		}, nil
+	})
+	defer func() { _ = body.Close() }()
+
+	out, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if continueCalls.Load() != 1 {
+		t.Fatalf("expected exactly one continue call, got %d", continueCalls.Load())
+	}
+	if !bytes.Contains(out, []byte(`"quasi_status","v":"INCOMPLETE"`)) || !bytes.Contains(out, []byte(`"v":"FINISHED"`)) {
+		t.Fatalf("expected continued output to include initial and final rounds, got=%s", string(out))
+	}
+}
+
+func TestAutoContinueDoesNotTriggerWhenResponseBatchQuasiStatusFinished(t *testing.T) {
+	initialBody := strings.Join([]string{
+		`data: {"response_message_id":321,"v":{"response":{"message_id":321,"status":"WIP","auto_continue":false}}}`,
+		`data: {"p":"response","o":"BATCH","v":[{"p":"accumulated_token_usage","v":2413},{"p":"quasi_status","v":"FINISHED"}]}`,
+		`data: [DONE]`,
+	}, "\n") + "\n"
+
+	var continueCalls atomic.Int32
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		continueCalls.Add(1)
+		return nil, errors.New("continue should not have been called")
+	})
+	defer func() { _ = body.Close() }()
+
+	out, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if continueCalls.Load() != 0 {
+		t.Fatalf("expected no continue calls, got %d", continueCalls.Load())
+	}
+	if !bytes.Contains(out, []byte(`"quasi_status","v":"FINISHED"`)) || !bytes.Contains(out, []byte(`data: [DONE]`)) {
+		t.Fatalf("expected original finished body to pass through unchanged, got=%s", string(out))
+	}
+}
+
 type failingOrCompletionDoer struct {
 	completionResp *http.Response
 }
@@ -134,3 +275,33 @@ func (d failingOrCompletionDoer) Do(req *http.Request) (*http.Response, error) {
 	}
 	return nil, errors.New("forced stream failure")
 }
+
+func TestAutoContinuePreservesIncompleteStateWhenNextChunkOmitsStatus(t *testing.T) {
+	initialBody := strings.Join([]string{
+		`data: {"response_message_id":321,"v":{"response":{"message_id":321,"status":"INCOMPLETE"}}}`,
+		`data: {"p":"response/content","v":{"text":"continued"}}`,
+		`data: [DONE]`,
+	}, "\n") + "\n"
+
+	var continueCalls atomic.Int32
+	body := newAutoContinueBody(context.Background(), io.NopCloser(strings.NewReader(initialBody)), "session-123", 8, func(context.Context, string, int) (*http.Response, error) {
+		continueCalls.Add(1)
+		return &http.Response{
+			StatusCode: http.StatusOK,
+			Header:     make(http.Header),
+			Body: io.NopCloser(strings.NewReader(
+				`data: {"response_message_id":322,"p":"response/status","v":"FINISHED"}` + "\n" +
+					`data: [DONE]` + "\n",
+			)),
+		}, nil
+	})
+	defer func() { _ = body.Close() }()
+
+	_, err := io.ReadAll(body)
+	if err != nil {
+		t.Fatalf("read body failed: %v", err)
+	}
+	if continueCalls.Load() != 1 {
+		t.Fatalf("expected exactly one continue call, got %d", continueCalls.Load())
+	}
+}

internal/deepseek/client/client_file_status.go

@@ -22,6 +22,9 @@ const (
 
 var fileReadySleep = time.Sleep
 
+// ErrUploadFileNotFound indicates that DeepSeek returned no matching uploaded file.
+var ErrUploadFileNotFound = errors.New("uploaded file not found")
+
 func (c *Client) waitForUploadedFile(ctx context.Context, a *auth.RequestAuth, result *UploadFileResult) error {
 	if result == nil || strings.TrimSpace(result.ID) == "" {
 		return nil
@@ -42,7 +45,7 @@ func (c *Client) waitForUploadedFile(ctx context.Context, a *auth.RequestAuth, r
 			return fmt.Errorf("waiting for file %s to become ready: %w", result.ID, err)
 		}
 
-		fetched, err := c.fetchUploadedFile(pollCtx, a, result.ID)
+		fetched, err := c.FetchUploadedFile(pollCtx, a, result.ID)
 		if err == nil && fetched != nil {
 			mergeUploadFileResults(result, fetched)
 			if isReadyUploadFileStatus(result.Status) {
@@ -65,7 +68,8 @@ func (c *Client) waitForUploadedFile(ctx context.Context, a *auth.RequestAuth, r
 	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) {
+// FetchUploadedFile returns metadata for an uploaded DeepSeek file by ID.
+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")
@@ -92,7 +96,7 @@ func (c *Client) fetchUploadedFile(ctx context.Context, a *auth.RequestAuth, fil
 
 	result := extractFetchedUploadFileResult(resp, fileID)
 	if result == nil || strings.TrimSpace(result.ID) == "" {
-		return nil, errors.New("fetch files succeeded without matching file data")
+		return nil, ErrUploadFileNotFound
 	}
 	result.Raw = resp
 	return result, nil

internal/deepseek/client/client_upload.go

@@ -23,6 +23,7 @@ type UploadFileRequest struct {
 	Filename    string
 	ContentType string
 	Purpose     string
+	ModelType   string
 	Data        []byte
 }
 
@@ -54,6 +55,7 @@ func (c *Client) UploadFile(ctx context.Context, a *auth.RequestAuth, req Upload
 		contentType = "application/octet-stream"
 	}
 	purpose := strings.TrimSpace(req.Purpose)
+	modelType := strings.ToLower(strings.TrimSpace(req.ModelType))
 	body, contentTypeHeader, err := buildUploadMultipartBody(filename, contentType, req.Data)
 	if err != nil {
 		return nil, err
@@ -64,6 +66,9 @@ func (c *Client) UploadFile(ctx context.Context, a *auth.RequestAuth, req Upload
 		"purpose":      purpose,
 		"bytes":        len(req.Data),
 	}
+	if modelType != "" {
+		capturePayload["model_type"] = modelType
+	}
 	captureSession := c.capture.Start("deepseek_upload_file", dsprotocol.DeepSeekUploadFileURL, a.AccountID, capturePayload)
 	attempts := 0
 	refreshed := false
@@ -81,6 +86,9 @@ func (c *Client) UploadFile(ctx context.Context, a *auth.RequestAuth, req Upload
 		}
 		headers := c.authHeaders(a.DeepSeekToken)
 		headers["Content-Type"] = contentTypeHeader
+		if modelType != "" {
+			headers["x-model-type"] = modelType
+		}
 		headers["x-ds-pow-response"] = powHeader
 		headers["x-file-size"] = strconv.Itoa(len(req.Data))
 		headers["x-thinking-enabled"] = "1"

internal/deepseek/client/client_upload_test.go

@@ -82,6 +82,7 @@ func TestUploadFileUsesUploadTargetPowAndMultipartHeaders(t *testing.T) {
 	var seenTargetPath string
 	var seenContentType string
 	var seenFileSize string
+	var seenModelType string
 	var seenBody string
 	call := 0
 	client := &Client{
@@ -96,6 +97,7 @@ func TestUploadFileUsesUploadTargetPowAndMultipartHeaders(t *testing.T) {
 				seenPow = req.Header.Get("x-ds-pow-response")
 				seenContentType = req.Header.Get("Content-Type")
 				seenFileSize = req.Header.Get("x-file-size")
+				seenModelType = req.Header.Get("x-model-type")
 				seenBody = string(bodyBytes)
 				return &http.Response{StatusCode: http.StatusOK, Header: make(http.Header), Body: io.NopCloser(strings.NewReader(uploadResponse)), Request: req}, nil
 			default:
@@ -112,6 +114,7 @@ func TestUploadFileUsesUploadTargetPowAndMultipartHeaders(t *testing.T) {
 		Filename:    "demo.txt",
 		ContentType: "text/plain",
 		Purpose:     "assistants",
+		ModelType:   "vision",
 		Data:        []byte("hello"),
 	}, 1)
 	if err != nil {
@@ -140,6 +143,9 @@ func TestUploadFileUsesUploadTargetPowAndMultipartHeaders(t *testing.T) {
 	if seenFileSize != "5" {
 		t.Fatalf("expected x-file-size=5, got %q", seenFileSize)
 	}
+	if seenModelType != "vision" {
+		t.Fatalf("expected x-model-type=vision, got %q", seenModelType)
+	}
 	if !strings.HasPrefix(seenContentType, "multipart/form-data; boundary=") {
 		t.Fatalf("expected multipart content type, got %q", seenContentType)
 	}

internal/deepseek/protocol/constants.go

@@ -159,6 +159,6 @@ func toStringSet(in []string) map[string]struct{} {
 
 const (
 	KeepAliveTimeout  = 5
-	StreamIdleTimeout = 90
-	MaxKeepaliveCount = 10
+	StreamIdleTimeout = 300
+	MaxKeepaliveCount = 40
 )

internal/deepseek/protocol/constants_shared.json

@@ -2,7 +2,7 @@
   "client": {
     "name": "DeepSeek",
     "platform": "android",
-    "version": "2.0.1",
+    "version": "2.0.4",
     "android_api_level": "35",
     "locale": "zh_CN"
   },
@@ -24,4 +24,4 @@
   "skip_exact_paths": [
     "response/search_status"
   ]
-}
+}

internal/deepseek/protocol/sse.go

@@ -2,20 +2,24 @@ package protocol
 
 import (
 	"bufio"
+	"io"
 	"net/http"
 )
 
 func ScanSSELines(resp *http.Response, onLine func([]byte) bool) error {
-	scanner := bufio.NewScanner(resp.Body)
-	buf := make([]byte, 0, 64*1024)
-	scanner.Buffer(buf, 2*1024*1024)
-	for scanner.Scan() {
-		if !onLine(scanner.Bytes()) {
-			break
+	reader := bufio.NewReaderSize(resp.Body, 64*1024)
+	for {
+		line, err := reader.ReadBytes('\n')
+		if len(line) > 0 {
+			if !onLine(line) {
+				return nil
+			}
+		}
+		if err != nil {
+			if err == io.EOF {
+				return nil
+			}
+			return err
 		}
 	}
-	if err := scanner.Err(); err != nil {
-		return err
-	}
-	return nil
 }

internal/deepseek/protocol/sse_test.go

@@ -0,0 +1,26 @@
+package protocol
+
+import (
+	"io"
+	"net/http"
+	"strings"
+	"testing"
+)
+
+func TestScanSSELinesHandlesLongSingleLine(t *testing.T) {
+	payload := strings.Repeat("x", 2*1024*1024+4096)
+	body := "data: {\"p\":\"response/content\",\"v\":\"" + payload + "\"}\n"
+	resp := &http.Response{Body: io.NopCloser(strings.NewReader(body))}
+
+	var got string
+	err := ScanSSELines(resp, func(line []byte) bool {
+		got = string(line)
+		return true
+	})
+	if err != nil {
+		t.Fatalf("ScanSSELines returned error: %v", err)
+	}
+	if !strings.Contains(got, payload) {
+		t.Fatalf("long SSE line was not preserved: got len=%d want payload len=%d", len(got), len(payload))
+	}
+}

internal/devcapture/store.go

@@ -10,6 +10,8 @@ import (
 	"sync"
 	"time"
 
+	"ds2api/internal/util"
+
 	"github.com/google/uuid"
 )
 
@@ -194,7 +196,8 @@ func (c *captureBody) append(chunk string) {
 	}
 	remain := maxLen - current
 	if len(chunk) > remain {
-		c.buf.WriteString(chunk[:remain])
+		truncated, _ := util.TruncateUTF8Bytes(chunk, remain)
+		c.buf.WriteString(truncated)
 		c.truncated = true
 		return
 	}

internal/devcapture/store_test.go

@@ -4,6 +4,7 @@ import (
 	"io"
 	"strings"
 	"testing"
+	"unicode/utf8"
 )
 
 func TestNewFromEnvDefaults(t *testing.T) {
@@ -82,3 +83,28 @@ func TestWrapBodyTruncatesByLimit(t *testing.T) {
 		t.Fatalf("expected account id, got %q", items[0].AccountID)
 	}
 }
+
+func TestWrapBodyTruncatesUTF8WithoutBreakingRune(t *testing.T) {
+	s := &Store{enabled: true, limit: 5, maxBodyBytes: 5}
+	session := s.Start("test", "http://x", "acc1", map[string]any{"x": 1})
+	if session == nil {
+		t.Fatal("expected session")
+	}
+	rc := session.WrapBody(io.NopCloser(strings.NewReader("😀xy")), 200)
+	_, _ = io.ReadAll(rc)
+	_ = rc.Close()
+
+	items := s.Snapshot()
+	if len(items) != 1 {
+		t.Fatalf("expected 1 item, got %d", len(items))
+	}
+	if !utf8.ValidString(items[0].ResponseBody) {
+		t.Fatalf("expected valid utf-8 response body, got %q", items[0].ResponseBody)
+	}
+	if items[0].ResponseBody != "😀x" {
+		t.Fatalf("expected rune-safe truncation, got %q", items[0].ResponseBody)
+	}
+	if !items[0].ResponseTruncated {
+		t.Fatal("expected truncated flag true")
+	}
+}

internal/format/claude/render.go

@@ -1,13 +1,56 @@
 package claude
 
 import (
+	"ds2api/internal/assistantturn"
 	"ds2api/internal/toolcall"
 	"fmt"
 	"time"
 
+	"ds2api/internal/prompt"
 	"ds2api/internal/util"
 )
 
+func BuildMessageResponseFromTurn(messageID, model string, turn assistantturn.Turn, exposeThinking bool) map[string]any {
+	content := make([]map[string]any, 0, 4)
+	if exposeThinking && turn.Thinking != "" {
+		content = append(content, map[string]any{"type": "thinking", "thinking": turn.Thinking})
+	}
+	stopReason := "end_turn"
+	if len(turn.ToolCalls) > 0 {
+		stopReason = "tool_use"
+		for i, tc := range turn.ToolCalls {
+			content = append(content, map[string]any{
+				"type":  "tool_use",
+				"id":    fmt.Sprintf("toolu_%d_%d", time.Now().Unix(), i),
+				"name":  tc.Name,
+				"input": tc.Input,
+			})
+		}
+	} else {
+		text := turn.Text
+		if text == "" && exposeThinking {
+			text = turn.Thinking
+		}
+		if text == "" {
+			text = "抱歉，没有生成有效的响应内容。"
+		}
+		content = append(content, map[string]any{"type": "text", "text": text})
+	}
+	return map[string]any{
+		"id":            messageID,
+		"type":          "message",
+		"role":          "assistant",
+		"model":         model,
+		"content":       content,
+		"stop_reason":   stopReason,
+		"stop_sequence": nil,
+		"usage": map[string]any{
+			"input_tokens":  turn.Usage.InputTokens,
+			"output_tokens": turn.Usage.OutputTokens,
+		},
+	}
+}
+
 func BuildMessageResponse(messageID, model string, normalizedMessages []any, finalThinking, finalText string, toolNames []string) map[string]any {
 	detected := toolcall.ParseToolCalls(finalText, toolNames)
 	if len(detected) == 0 && finalText == "" && finalThinking != "" {
@@ -43,8 +86,23 @@ func BuildMessageResponse(messageID, model string, normalizedMessages []any, fin
 		"stop_reason":   stopReason,
 		"stop_sequence": nil,
 		"usage": map[string]any{
-			"input_tokens":  util.EstimateTokens(fmt.Sprintf("%v", normalizedMessages)),
-			"output_tokens": util.EstimateTokens(finalThinking) + util.EstimateTokens(finalText),
+			"input_tokens":  util.CountPromptTokens(prompt.MessagesPrepareWithThinking(claudeMessageMaps(normalizedMessages), false), model),
+			"output_tokens": util.CountOutputTokens(finalThinking, model) + util.CountOutputTokens(finalText, model),
 		},
 	}
 }
+
+func claudeMessageMaps(messages []any) []map[string]any {
+	if len(messages) == 0 {
+		return nil
+	}
+	out := make([]map[string]any, 0, len(messages))
+	for _, item := range messages {
+		msg, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		out = append(out, msg)
+	}
+	return out
+}

internal/format/openai/render_chat.go

@@ -29,7 +29,7 @@ func BuildChatCompletionWithToolCalls(completionID, model, finalPrompt, finalThi
 		"created": time.Now().Unix(),
 		"model":   model,
 		"choices": []map[string]any{{"index": 0, "message": messageObj, "finish_reason": finishReason}},
-		"usage":   BuildChatUsage(finalPrompt, finalThinking, finalText),
+		"usage":   BuildChatUsageForModel(model, finalPrompt, finalThinking, finalText, 0),
 	}
 }
 

internal/format/openai/render_responses.go

@@ -70,7 +70,7 @@ func BuildResponseObjectFromItems(responseID, model, finalPrompt, finalThinking,
 		"model":       model,
 		"output":      output,
 		"output_text": outputText,
-		"usage":       BuildResponsesUsage(finalPrompt, finalThinking, finalText),
+		"usage":       BuildResponsesUsageForModel(model, finalPrompt, finalThinking, finalText, 0),
 	}
 }
 

internal/format/openai/render_test.go

@@ -6,6 +6,7 @@ import (
 	"testing"
 
 	"ds2api/internal/toolcall"
+	"ds2api/internal/util"
 )
 
 func TestBuildResponseObjectKeepsFencedToolPayloadAsText(t *testing.T) {
@@ -177,3 +178,17 @@ func TestBuildResponseObjectWithToolCallsCoercesSchemaDeclaredStringArguments(t
 		t.Fatalf("expected response content stringified by schema, got %#v", args["content"])
 	}
 }
+
+func TestBuildChatUsageForModelUsesConservativePromptCount(t *testing.T) {
+	prompt := strings.Repeat("上下文token ", 40)
+	usage := BuildChatUsageForModel("deepseek-v4-flash", prompt, "", "ok", 0)
+	promptTokens, _ := usage["prompt_tokens"].(int)
+	if promptTokens <= util.EstimateTokens(prompt) {
+		t.Fatalf("expected conservative prompt token count > rough estimate, got=%d estimate=%d", promptTokens, util.EstimateTokens(prompt))
+	}
+	totalTokens, _ := usage["total_tokens"].(int)
+	completionTokens, _ := usage["completion_tokens"].(int)
+	if totalTokens != promptTokens+completionTokens {
+		t.Fatalf("expected total tokens to add up, got usage=%#v", usage)
+	}
+}

internal/format/openai/render_usage.go

@@ -2,10 +2,10 @@ package openai
 
 import "ds2api/internal/util"
 
-func BuildChatUsage(finalPrompt, finalThinking, finalText string) map[string]any {
-	promptTokens := util.EstimateTokens(finalPrompt)
-	reasoningTokens := util.EstimateTokens(finalThinking)
-	completionTokens := util.EstimateTokens(finalText)
+func BuildChatUsageForModel(model, finalPrompt, finalThinking, finalText string, refFileTokens int) map[string]any {
+	promptTokens := util.CountPromptTokens(finalPrompt, model) + refFileTokens
+	reasoningTokens := util.CountOutputTokens(finalThinking, model)
+	completionTokens := util.CountOutputTokens(finalText, model)
 	return map[string]any{
 		"prompt_tokens":     promptTokens,
 		"completion_tokens": reasoningTokens + completionTokens,
@@ -16,13 +16,21 @@ func BuildChatUsage(finalPrompt, finalThinking, finalText string) map[string]any
 	}
 }
 
-func BuildResponsesUsage(finalPrompt, finalThinking, finalText string) map[string]any {
-	promptTokens := util.EstimateTokens(finalPrompt)
-	reasoningTokens := util.EstimateTokens(finalThinking)
-	completionTokens := util.EstimateTokens(finalText)
+func BuildChatUsage(finalPrompt, finalThinking, finalText string) map[string]any {
+	return BuildChatUsageForModel("", finalPrompt, finalThinking, finalText, 0)
+}
+
+func BuildResponsesUsageForModel(model, finalPrompt, finalThinking, finalText string, refFileTokens int) map[string]any {
+	promptTokens := util.CountPromptTokens(finalPrompt, model) + refFileTokens
+	reasoningTokens := util.CountOutputTokens(finalThinking, model)
+	completionTokens := util.CountOutputTokens(finalText, model)
 	return map[string]any{
 		"input_tokens":  promptTokens,
 		"output_tokens": reasoningTokens + completionTokens,
 		"total_tokens":  promptTokens + reasoningTokens + completionTokens,
 	}
 }
+
+func BuildResponsesUsage(finalPrompt, finalThinking, finalText string) map[string]any {
+	return BuildResponsesUsageForModel("", finalPrompt, finalThinking, finalText, 0)
+}

internal/httpapi/admin/rawsamples/handler_raw_samples.go

@@ -15,6 +15,7 @@ import (
 	"ds2api/internal/devcapture"
 	adminshared "ds2api/internal/httpapi/admin/shared"
 	"ds2api/internal/rawsample"
+	"ds2api/internal/util"
 )
 
 type captureChain struct {
@@ -479,10 +480,13 @@ func previewCaptureChainResponse(chain captureChain) string {
 
 func previewText(text string, limit int) string {
 	text = strings.TrimSpace(text)
-	if limit <= 0 || len(text) <= limit {
+	if limit <= 0 {
 		return text
 	}
-	return text[:limit] + "..."
+	if truncated, ok := util.TruncateRunes(text, limit); ok {
+		return truncated + "..."
+	}
+	return text
 }
 
 func captureChainHasTruncatedResponse(chain captureChain) bool {

internal/httpapi/admin/rawsamples/handler_raw_samples_test.go

@@ -10,6 +10,7 @@ import (
 	"path/filepath"
 	"strings"
 	"testing"
+	"unicode/utf8"
 
 	"ds2api/internal/devcapture"
 )
@@ -231,6 +232,16 @@ func TestCombineCaptureBodiesPreservesOrderAndSeparators(t *testing.T) {
 	}
 }
 
+func TestPreviewTextPreservesUTF8MB4Characters(t *testing.T) {
+	preview := previewText(strings.Repeat("😀", 281), 280)
+	if !utf8.ValidString(preview) {
+		t.Fatalf("expected valid utf-8 preview, got %q", preview)
+	}
+	if preview != strings.Repeat("😀", 280)+"..." {
+		t.Fatalf("unexpected preview: %q", preview)
+	}
+}
+
 func TestQueryRawSampleCapturesGroupsBySessionAndMatchesQuestion(t *testing.T) {
 	devcapture.Global().Clear()
 	defer devcapture.Global().Clear()

internal/httpapi/claude/handler_messages.go

@@ -3,12 +3,20 @@ package claude
 import (
 	"bytes"
 	"encoding/json"
+	"errors"
+	"fmt"
 	"io"
 	"net/http"
 	"net/http/httptest"
 	"strings"
+	"time"
 
+	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
+	claudefmt "ds2api/internal/format/claude"
+	"ds2api/internal/httpapi/requestbody"
+	"ds2api/internal/promptcompat"
 	streamengine "ds2api/internal/stream"
 	"ds2api/internal/translatorcliproxy"
 	"ds2api/internal/util"
@@ -20,20 +28,100 @@ func (h *Handler) Messages(w http.ResponseWriter, r *http.Request) {
 	if strings.TrimSpace(r.Header.Get("anthropic-version")) == "" {
 		r.Header.Set("anthropic-version", "2023-06-01")
 	}
-	if h.OpenAI == nil {
-		writeClaudeError(w, http.StatusInternalServerError, "OpenAI proxy backend unavailable.")
+	if isClaudeVercelProxyRequest(r) && h.proxyViaOpenAI(w, r, h.Store) {
 		return
 	}
-	if h.proxyViaOpenAI(w, r, h.Store) {
+	if h.Auth == nil || h.DS == nil {
+		if h.OpenAI != nil && h.proxyViaOpenAI(w, r, h.Store) {
+			return
+		}
+		writeClaudeError(w, http.StatusInternalServerError, "Claude runtime backend unavailable.")
 		return
 	}
-	writeClaudeError(w, http.StatusBadGateway, "Failed to proxy Claude request.")
+	if h.handleClaudeDirect(w, r) {
+		return
+	}
+	writeClaudeError(w, http.StatusBadGateway, "Failed to handle Claude request.")
+}
+
+func isClaudeVercelProxyRequest(r *http.Request) bool {
+	if r == nil || r.URL == nil {
+		return false
+	}
+	return strings.TrimSpace(r.URL.Query().Get("__stream_prepare")) == "1" ||
+		strings.TrimSpace(r.URL.Query().Get("__stream_release")) == "1"
+}
+
+func (h *Handler) handleClaudeDirect(w http.ResponseWriter, r *http.Request) bool {
+	raw, err := io.ReadAll(r.Body)
+	if err != nil {
+		if errors.Is(err, requestbody.ErrInvalidUTF8Body) {
+			writeClaudeError(w, http.StatusBadRequest, "invalid json")
+		} else {
+			writeClaudeError(w, http.StatusBadRequest, "invalid body")
+		}
+		return true
+	}
+	var req map[string]any
+	if err := json.Unmarshal(raw, &req); err != nil {
+		writeClaudeError(w, http.StatusBadRequest, "invalid json")
+		return true
+	}
+	exposeThinking := false
+	if enabled, ok := util.ResolveThinkingOverride(req); ok && enabled {
+		exposeThinking = true
+	} else if _, ok := util.ResolveThinkingOverride(req); !ok && !util.ToBool(req["stream"]) {
+		req["thinking"] = map[string]any{"type": "enabled"}
+	}
+	norm, err := normalizeClaudeRequest(h.Store, req)
+	if err != nil {
+		writeClaudeError(w, http.StatusBadRequest, err.Error())
+		return true
+	}
+	a, err := h.Auth.Determine(r)
+	if err != nil {
+		writeClaudeError(w, http.StatusUnauthorized, err.Error())
+		return true
+	}
+	defer h.Auth.Release(a)
+	if norm.Standard.Stream {
+		h.handleClaudeDirectStream(w, r, a, norm.Standard)
+		return true
+	}
+	result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, norm.Standard, completionruntime.Options{
+		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		RetryEnabled:          true,
+	})
+	if outErr != nil {
+		writeClaudeError(w, outErr.Status, outErr.Message)
+		return true
+	}
+	writeJSON(w, http.StatusOK, claudefmt.BuildMessageResponseFromTurn(
+		fmt.Sprintf("msg_%d", time.Now().UnixNano()),
+		norm.Standard.ResponseModel,
+		result.Turn,
+		exposeThinking,
+	))
+	return true
+}
+
+func (h *Handler) handleClaudeDirectStream(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, stdReq promptcompat.StandardRequest) {
+	start, outErr := completionruntime.StartCompletion(r.Context(), h.DS, a, stdReq, completionruntime.Options{})
+	if outErr != nil {
+		writeClaudeError(w, outErr.Status, outErr.Message)
+		return
+	}
+	h.handleClaudeStreamRealtime(w, r, start.Response, stdReq.ResponseModel, stdReq.Messages, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw)
 }
 
 func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, store ConfigReader) bool {
 	raw, err := io.ReadAll(r.Body)
 	if err != nil {
-		writeClaudeError(w, http.StatusBadRequest, "invalid body")
+		if errors.Is(err, requestbody.ErrInvalidUTF8Body) {
+			writeClaudeError(w, http.StatusBadRequest, "invalid json")
+		} else {
+			writeClaudeError(w, http.StatusBadRequest, "invalid body")
+		}
 		return true
 	}
 	var req map[string]any
@@ -206,6 +294,7 @@ func (h *Handler) handleClaudeStreamRealtime(w http.ResponseWriter, r *http.Requ
 		h.compatStripReferenceMarkers(),
 		toolNames,
 		toolsRaw,
+		buildClaudePromptTokenText(messages, thinkingEnabled),
 	)
 	streamRuntime.sendMessageStart()
 

internal/httpapi/claude/handler_stream_test.go

@@ -28,6 +28,18 @@ func makeClaudeSSEHTTPResponse(lines ...string) *http.Response {
 	}
 }
 
+func makeClaudeContentLine(t *testing.T, text string) string {
+	t.Helper()
+	line, err := json.Marshal(map[string]any{
+		"p": "response/content",
+		"v": text,
+	})
+	if err != nil {
+		t.Fatalf("marshal content line failed: %v", err)
+	}
+	return "data: " + string(line)
+}
+
 func parseClaudeFrames(t *testing.T, body string) []claudeFrame {
 	t.Helper()
 	chunks := strings.Split(body, "\n\n")
@@ -71,6 +83,17 @@ func findClaudeFrames(frames []claudeFrame, event string) []claudeFrame {
 	return out
 }
 
+func collectClaudeTextDeltas(frames []claudeFrame) string {
+	var combined strings.Builder
+	for _, f := range findClaudeFrames(frames, "content_block_delta") {
+		delta, _ := f.Payload["delta"].(map[string]any)
+		if delta["type"] == "text_delta" {
+			combined.WriteString(asString(delta["text"]))
+		}
+	}
+	return combined.String()
+}
+
 func TestHandleClaudeStreamRealtimeTextIncrementsWithEventHeaders(t *testing.T) {
 	h := &Handler{}
 	resp := makeClaudeSSEHTTPResponse(
@@ -96,8 +119,8 @@ func TestHandleClaudeStreamRealtimeTextIncrementsWithEventHeaders(t *testing.T)
 
 	frames := parseClaudeFrames(t, body)
 	deltas := findClaudeFrames(frames, "content_block_delta")
-	if len(deltas) < 2 {
-		t.Fatalf("expected at least 2 text deltas, got=%d body=%s", len(deltas), body)
+	if len(deltas) < 1 {
+		t.Fatalf("expected at least 1 text delta, got=%d body=%s", len(deltas), body)
 	}
 	combined := strings.Builder{}
 	for _, f := range deltas {
@@ -111,6 +134,52 @@ func TestHandleClaudeStreamRealtimeTextIncrementsWithEventHeaders(t *testing.T)
 	}
 }
 
+func TestHandleClaudeStreamRealtimeToolBufferedPlainTextDoesNotRepeatFinalText(t *testing.T) {
+	h := &Handler{}
+	want := "明白\n\nBash\nIN\npwd\nOUT\nok"
+	resp := makeClaudeSSEHTTPResponse(
+		makeClaudeContentLine(t, "明"),
+		makeClaudeContentLine(t, "白\n\nBash\nIN\npwd\n"),
+		makeClaudeContentLine(t, "OUT\nok"),
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
+
+	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "use tool"}}, false, false, []string{"Bash"}, nil)
+
+	frames := parseClaudeFrames(t, rec.Body.String())
+	if got := collectClaudeTextDeltas(frames); got != want {
+		t.Fatalf("unexpected combined text: got %q want %q body=%s", got, want, rec.Body.String())
+	}
+}
+
+func TestHandleClaudeStreamRealtimeTrimsContinuationReplay(t *testing.T) {
+	h := &Handler{}
+	prefix := strings.Repeat("A", 40)
+	resp := makeClaudeSSEHTTPResponse(
+		`data: {"p":"response/content","v":"`+prefix+`"}`,
+		`data: {"p":"response/content","v":"`+prefix+` tail"}`,
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/anthropic/v1/messages", nil)
+
+	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "hi"}}, false, false, nil, nil)
+
+	frames := parseClaudeFrames(t, rec.Body.String())
+	combined := strings.Builder{}
+	for _, f := range findClaudeFrames(frames, "content_block_delta") {
+		delta, _ := f.Payload["delta"].(map[string]any)
+		if delta["type"] == "text_delta" {
+			combined.WriteString(asString(delta["text"]))
+		}
+	}
+	if got, want := combined.String(), prefix+" tail"; got != want {
+		t.Fatalf("unexpected combined text: got %q want %q body=%s", got, want, rec.Body.String())
+	}
+}
+
 func TestHandleClaudeStreamRealtimeThinkingDelta(t *testing.T) {
 	h := &Handler{}
 	resp := makeClaudeSSEHTTPResponse(

internal/httpapi/claude/handler_tokens.go

@@ -3,8 +3,6 @@ package claude
 import (
 	"encoding/json"
 	"net/http"
-
-	"ds2api/internal/util"
 )
 
 func (h *Handler) CountTokens(w http.ResponseWriter, r *http.Request) {
@@ -26,26 +24,11 @@ func (h *Handler) CountTokens(w http.ResponseWriter, r *http.Request) {
 		writeClaudeError(w, http.StatusBadRequest, "Request must include 'model' and 'messages'.")
 		return
 	}
-	inputTokens := 0
-	if sys, ok := req["system"].(string); ok {
-		inputTokens += util.EstimateTokens(sys)
-	}
-	for _, item := range messages {
-		msg, ok := item.(map[string]any)
-		if !ok {
-			continue
-		}
-		inputTokens += 2
-		inputTokens += util.EstimateTokens(extractMessageContent(msg["content"]))
-	}
-	if tools, ok := req["tools"].([]any); ok {
-		for _, t := range tools {
-			b, _ := json.Marshal(t)
-			inputTokens += util.EstimateTokens(string(b))
-		}
-	}
-	if inputTokens < 1 {
-		inputTokens = 1
+	normalized, err := normalizeClaudeRequest(h.Store, req)
+	if err != nil {
+		writeClaudeError(w, http.StatusBadRequest, err.Error())
+		return
 	}
+	inputTokens := countClaudeInputTokens(normalized.Standard)
 	writeJSON(w, http.StatusOK, map[string]any{"input_tokens": inputTokens})
 }

internal/httpapi/claude/prompt_token_text.go

@@ -0,0 +1,7 @@
+package claude
+
+import "ds2api/internal/prompt"
+
+func buildClaudePromptTokenText(messages []any, thinkingEnabled bool) string {
+	return prompt.MessagesPrepareWithThinking(toMessageMaps(messages), thinkingEnabled)
+}

internal/httpapi/claude/standard_request.go

@@ -48,17 +48,18 @@ func normalizeClaudeRequest(store ConfigReader, req map[string]any) (claudeNorma
 
 	return claudeNormalizedRequest{
 		Standard: promptcompat.StandardRequest{
-			Surface:        "anthropic_messages",
-			RequestedModel: strings.TrimSpace(model),
-			ResolvedModel:  dsModel,
-			ResponseModel:  strings.TrimSpace(model),
-			Messages:       payload["messages"].([]any),
-			ToolsRaw:       toolsRequested,
-			FinalPrompt:    finalPrompt,
-			ToolNames:      toolNames,
-			Stream:         util.ToBool(req["stream"]),
-			Thinking:       thinkingEnabled,
-			Search:         searchEnabled,
+			Surface:         "anthropic_messages",
+			RequestedModel:  strings.TrimSpace(model),
+			ResolvedModel:   dsModel,
+			ResponseModel:   strings.TrimSpace(model),
+			Messages:        payload["messages"].([]any),
+			PromptTokenText: finalPrompt,
+			ToolsRaw:        toolsRequested,
+			FinalPrompt:     finalPrompt,
+			ToolNames:       toolNames,
+			Stream:          util.ToBool(req["stream"]),
+			Thinking:        thinkingEnabled,
+			Search:          searchEnabled,
 		},
 		NormalizedMessages: normalizedMessages,
 	}, nil

internal/httpapi/claude/stream_runtime_core.go

@@ -8,6 +8,8 @@ import (
 
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
+	"ds2api/internal/toolcall"
+	"ds2api/internal/toolstream"
 )
 
 type claudeStreamRuntime struct {
@@ -15,10 +17,11 @@ type claudeStreamRuntime struct {
 	rc       *http.ResponseController
 	canFlush bool
 
-	model     string
-	toolNames []string
-	messages  []any
-	toolsRaw  any
+	model           string
+	toolNames       []string
+	messages        []any
+	toolsRaw        any
+	promptTokenText string
 
 	thinkingEnabled       bool
 	searchEnabled         bool
@@ -29,11 +32,18 @@ type claudeStreamRuntime struct {
 	thinking  strings.Builder
 	text      strings.Builder
 
+	sieve                 toolstream.State
+	rawText               strings.Builder
+	rawThinking           strings.Builder
+	toolDetectionThinking strings.Builder
+	toolCallsDetected     bool
+
 	nextBlockIndex     int
 	thinkingBlockOpen  bool
 	thinkingBlockIndex int
 	textBlockOpen      bool
 	textBlockIndex     int
+	textEmitted        bool
 	ended              bool
 	upstreamErr        string
 }
@@ -49,6 +59,7 @@ func newClaudeStreamRuntime(
 	stripReferenceMarkers bool,
 	toolNames []string,
 	toolsRaw any,
+	promptTokenText string,
 ) *claudeStreamRuntime {
 	return &claudeStreamRuntime{
 		w:                     w,
@@ -62,6 +73,7 @@ func newClaudeStreamRuntime(
 		stripReferenceMarkers: stripReferenceMarkers,
 		toolNames:             toolNames,
 		toolsRaw:              toolsRaw,
+		promptTokenText:       promptTokenText,
 		messageID:             fmt.Sprintf("msg_%d", time.Now().UnixNano()),
 		thinkingBlockIndex:    -1,
 		textBlockIndex:        -1,
@@ -81,8 +93,28 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	}
 
 	contentSeen := false
+	for _, p := range parsed.ToolDetectionThinkingParts {
+		trimmed := sse.TrimContinuationOverlapFromBuilder(&s.toolDetectionThinking, p.Text)
+		if trimmed != "" {
+			s.toolDetectionThinking.WriteString(trimmed)
+		}
+	}
 	for _, p := range parsed.Parts {
-		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
+		var rawTrimmed string
+		if p.Type == "thinking" {
+			rawTrimmed = sse.TrimContinuationOverlapFromBuilder(&s.rawThinking, p.Text)
+		} else {
+			rawTrimmed = sse.TrimContinuationOverlapFromBuilder(&s.rawText, p.Text)
+		}
+		if rawTrimmed == "" {
+			continue
+		}
+		if p.Type == "thinking" {
+			s.rawThinking.WriteString(rawTrimmed)
+		} else {
+			s.rawText.WriteString(rawTrimmed)
+		}
+		cleanedText := cleanVisibleOutput(rawTrimmed, s.stripReferenceMarkers)
 		if cleanedText == "" {
 			continue
 		}
@@ -95,7 +127,7 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 			if !s.thinkingEnabled {
 				continue
 			}
-			trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
+			trimmed := sse.TrimContinuationOverlapFromBuilder(&s.thinking, cleanedText)
 			if trimmed == "" {
 				continue
 			}
@@ -125,44 +157,80 @@ func (s *claudeStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 			continue
 		}
 
-		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
-		if trimmed == "" {
-			continue
-		}
-		s.text.WriteString(trimmed)
-		if s.bufferToolContent {
-			if hasUnclosedCodeFence(s.text.String()) {
-				continue
+		s.text.WriteString(cleanedText)
+
+		if !s.bufferToolContent {
+			s.closeThinkingBlock()
+			if !s.textBlockOpen {
+				s.textBlockIndex = s.nextBlockIndex
+				s.nextBlockIndex++
+				s.send("content_block_start", map[string]any{
+					"type":  "content_block_start",
+					"index": s.textBlockIndex,
+					"content_block": map[string]any{
+						"type": "text",
+						"text": "",
+					},
+				})
+				s.textBlockOpen = true
 			}
-			continue
-		}
-		s.closeThinkingBlock()
-		if !s.textBlockOpen {
-			s.textBlockIndex = s.nextBlockIndex
-			s.nextBlockIndex++
-			s.send("content_block_start", map[string]any{
-				"type":  "content_block_start",
+			s.send("content_block_delta", map[string]any{
+				"type":  "content_block_delta",
 				"index": s.textBlockIndex,
-				"content_block": map[string]any{
-					"type": "text",
-					"text": "",
+				"delta": map[string]any{
+					"type": "text_delta",
+					"text": cleanedText,
 				},
 			})
-			s.textBlockOpen = true
+			s.textEmitted = true
+			continue
+		}
+
+		events := toolstream.ProcessChunk(&s.sieve, rawTrimmed, s.toolNames)
+		for _, evt := range events {
+			if len(evt.ToolCalls) > 0 {
+				s.closeTextBlock()
+				s.toolCallsDetected = true
+				normalized := toolcall.NormalizeParsedToolCallsForSchemas(evt.ToolCalls, s.toolsRaw)
+				for _, tc := range normalized {
+					idx := s.nextBlockIndex
+					s.nextBlockIndex++
+					s.sendToolUseBlock(idx, tc)
+				}
+				continue
+			}
+			if evt.Content == "" {
+				continue
+			}
+			cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
+			if cleaned == "" || (s.searchEnabled && sse.IsCitation(cleaned)) {
+				continue
+			}
+			s.closeThinkingBlock()
+			if !s.textBlockOpen {
+				s.textBlockIndex = s.nextBlockIndex
+				s.nextBlockIndex++
+				s.send("content_block_start", map[string]any{
+					"type":  "content_block_start",
+					"index": s.textBlockIndex,
+					"content_block": map[string]any{
+						"type": "text",
+						"text": "",
+					},
+				})
+				s.textBlockOpen = true
+			}
+			s.send("content_block_delta", map[string]any{
+				"type":  "content_block_delta",
+				"index": s.textBlockIndex,
+				"delta": map[string]any{
+					"type": "text_delta",
+					"text": cleaned,
+				},
+			})
+			s.textEmitted = true
 		}
-		s.send("content_block_delta", map[string]any{
-			"type":  "content_block_delta",
-			"index": s.textBlockIndex,
-			"delta": map[string]any{
-				"type": "text_delta",
-				"text": trimmed,
-			},
-		})
 	}
 
 	return streamengine.ParsedDecision{ContentSeen: contentSeen}
 }
-
-func hasUnclosedCodeFence(text string) bool {
-	return strings.Count(text, "```")%2 == 1
-}

internal/httpapi/claude/stream_runtime_emit.go

@@ -42,7 +42,10 @@ func (s *claudeStreamRuntime) sendPing() {
 }
 
 func (s *claudeStreamRuntime) sendMessageStart() {
-	inputTokens := util.EstimateTokens(fmt.Sprintf("%v", s.messages))
+	inputTokens := countClaudeInputTokensFromText(s.promptTokenText, s.model)
+	if inputTokens == 0 {
+		inputTokens = util.CountPromptTokens(fmt.Sprintf("%v", s.messages), s.model)
+	}
 	s.send("message_start", map[string]any{
 		"type": "message_start",
 		"message": map[string]any{

internal/httpapi/claude/stream_runtime_finalize.go

@@ -1,13 +1,15 @@
 package claude
 
 import (
+	"ds2api/internal/assistantturn"
+	"ds2api/internal/sse"
 	"ds2api/internal/toolcall"
+	"ds2api/internal/toolstream"
 	"encoding/json"
 	"fmt"
 	"time"
 
 	streamengine "ds2api/internal/stream"
-	"ds2api/internal/util"
 )
 
 func (s *claudeStreamRuntime) closeThinkingBlock() {
@@ -34,6 +36,32 @@ func (s *claudeStreamRuntime) closeTextBlock() {
 	s.textBlockIndex = -1
 }
 
+func (s *claudeStreamRuntime) sendToolUseBlock(idx int, tc toolcall.ParsedToolCall) {
+	s.send("content_block_start", map[string]any{
+		"type":  "content_block_start",
+		"index": idx,
+		"content_block": map[string]any{
+			"type":  "tool_use",
+			"id":    fmt.Sprintf("toolu_%d_%d", time.Now().Unix(), idx),
+			"name":  tc.Name,
+			"input": map[string]any{},
+		},
+	})
+	inputBytes, _ := json.Marshal(tc.Input)
+	s.send("content_block_delta", map[string]any{
+		"type":  "content_block_delta",
+		"index": idx,
+		"delta": map[string]any{
+			"type":         "input_json_delta",
+			"partial_json": string(inputBytes),
+		},
+	})
+	s.send("content_block_stop", map[string]any{
+		"type":  "content_block_stop",
+		"index": idx,
+	})
+}
+
 func (s *claudeStreamRuntime) finalize(stopReason string) {
 	if s.ended {
 		return
@@ -41,49 +69,80 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 	s.ended = true
 
 	s.closeThinkingBlock()
-	s.closeTextBlock()
-
-	finalThinking := s.thinking.String()
-	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
 
 	if s.bufferToolContent {
-		detected := toolcall.ParseStandaloneToolCalls(finalText, s.toolNames)
-		if len(detected) == 0 && finalText == "" && finalThinking != "" {
-			detected = toolcall.ParseStandaloneToolCalls(finalThinking, s.toolNames)
-		}
-		if len(detected) > 0 {
-			detected = toolcall.NormalizeParsedToolCallsForSchemas(detected, s.toolsRaw)
-			stopReason = "tool_use"
-			for i, tc := range detected {
-				idx := s.nextBlockIndex + i
-				s.send("content_block_start", map[string]any{
-					"type":  "content_block_start",
-					"index": idx,
-					"content_block": map[string]any{
-						"type":  "tool_use",
-						"id":    fmt.Sprintf("toolu_%d_%d", time.Now().Unix(), idx),
-						"name":  tc.Name,
-						"input": map[string]any{},
-					},
-				})
-
-				inputBytes, _ := json.Marshal(tc.Input)
+		for _, evt := range toolstream.Flush(&s.sieve, s.toolNames) {
+			if len(evt.ToolCalls) > 0 {
+				s.closeTextBlock()
+				s.toolCallsDetected = true
+				normalized := toolcall.NormalizeParsedToolCallsForSchemas(evt.ToolCalls, s.toolsRaw)
+				for _, tc := range normalized {
+					idx := s.nextBlockIndex
+					s.nextBlockIndex++
+					s.sendToolUseBlock(idx, tc)
+				}
+				continue
+			}
+			if evt.Content != "" {
+				cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
+				if cleaned == "" || (s.searchEnabled && sse.IsCitation(cleaned)) {
+					continue
+				}
+				if !s.textBlockOpen {
+					s.textBlockIndex = s.nextBlockIndex
+					s.nextBlockIndex++
+					s.send("content_block_start", map[string]any{
+						"type":  "content_block_start",
+						"index": s.textBlockIndex,
+						"content_block": map[string]any{
+							"type": "text",
+							"text": "",
+						},
+					})
+					s.textBlockOpen = true
+				}
 				s.send("content_block_delta", map[string]any{
 					"type":  "content_block_delta",
-					"index": idx,
+					"index": s.textBlockIndex,
 					"delta": map[string]any{
-						"type":         "input_json_delta",
-						"partial_json": string(inputBytes),
+						"type": "text_delta",
+						"text": cleaned,
 					},
 				})
-
-				s.send("content_block_stop", map[string]any{
-					"type":  "content_block_stop",
-					"index": idx,
-				})
+				s.textEmitted = true
 			}
-			s.nextBlockIndex += len(detected)
-		} else if finalText != "" {
+		}
+	}
+
+	s.closeTextBlock()
+
+	turn := assistantturn.BuildTurnFromStreamSnapshot(assistantturn.StreamSnapshot{
+		RawText:               s.rawText.String(),
+		VisibleText:           s.text.String(),
+		RawThinking:           s.rawThinking.String(),
+		VisibleThinking:       s.thinking.String(),
+		DetectionThinking:     s.toolDetectionThinking.String(),
+		AlreadyEmittedCalls:   s.toolCallsDetected,
+		AlreadyEmittedToolRaw: s.toolCallsDetected,
+	}, assistantturn.BuildOptions{
+		Model:                 s.model,
+		Prompt:                s.promptTokenText,
+		SearchEnabled:         s.searchEnabled,
+		StripReferenceMarkers: s.stripReferenceMarkers,
+		ToolNames:             s.toolNames,
+		ToolsRaw:              s.toolsRaw,
+	})
+	finalText := turn.Text
+
+	if s.bufferToolContent && !s.toolCallsDetected {
+		if len(turn.ToolCalls) > 0 {
+			stopReason = "tool_use"
+			for _, tc := range turn.ToolCalls {
+				idx := s.nextBlockIndex
+				s.nextBlockIndex++
+				s.sendToolUseBlock(idx, tc)
+			}
+		} else if finalText != "" && !s.textEmitted {
 			idx := s.nextBlockIndex
 			s.nextBlockIndex++
 			s.send("content_block_start", map[string]any{
@@ -102,6 +161,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 					"text": finalText,
 				},
 			})
+			s.textEmitted = true
 			s.send("content_block_stop", map[string]any{
 				"type":  "content_block_stop",
 				"index": idx,
@@ -109,7 +169,10 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 		}
 	}
 
-	outputTokens := util.EstimateTokens(finalThinking) + util.EstimateTokens(finalText)
+	if s.toolCallsDetected {
+		stopReason = "tool_use"
+	}
+
 	s.send("message_delta", map[string]any{
 		"type": "message_delta",
 		"delta": map[string]any{
@@ -117,7 +180,7 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 			"stop_sequence": nil,
 		},
 		"usage": map[string]any{
-			"output_tokens": outputTokens,
+			"output_tokens": turn.Usage.OutputTokens,
 		},
 	})
 	s.send("message_stop", map[string]any{"type": "message_stop"})

internal/httpapi/claude/token_count.go

@@ -0,0 +1,20 @@
+package claude
+
+import (
+	"strings"
+
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/util"
+)
+
+func countClaudeInputTokens(stdReq promptcompat.StandardRequest) int {
+	promptText := stdReq.PromptTokenText
+	if strings.TrimSpace(promptText) == "" {
+		promptText = stdReq.FinalPrompt
+	}
+	return countClaudeInputTokensFromText(promptText, stdReq.ResolvedModel)
+}
+
+func countClaudeInputTokensFromText(promptText, model string) int {
+	return util.CountPromptTokens(promptText, model)
+}

internal/httpapi/gemini/convert_request.go

@@ -33,19 +33,24 @@ func normalizeGeminiRequest(store ConfigReader, routeModel string, req map[strin
 
 	toolsRaw := convertGeminiTools(req["tools"])
 	finalPrompt, toolNames := promptcompat.BuildOpenAIPromptForAdapter(messagesRaw, toolsRaw, "", thinkingEnabled)
+	if len(toolNames) == 0 && len(toolsRaw) > 0 {
+		toolNames = []string{"__any_tool__"}
+	}
 	passThrough := collectGeminiPassThrough(req)
 
 	return promptcompat.StandardRequest{
-		Surface:        "google_gemini",
-		RequestedModel: requestedModel,
-		ResolvedModel:  resolvedModel,
-		ResponseModel:  requestedModel,
-		Messages:       messagesRaw,
-		FinalPrompt:    finalPrompt,
-		ToolNames:      toolNames,
-		Stream:         stream,
-		Thinking:       thinkingEnabled,
-		Search:         searchEnabled,
-		PassThrough:    passThrough,
+		Surface:         "google_gemini",
+		RequestedModel:  requestedModel,
+		ResolvedModel:   resolvedModel,
+		ResponseModel:   requestedModel,
+		Messages:        messagesRaw,
+		PromptTokenText: finalPrompt,
+		ToolsRaw:        toolsRaw,
+		FinalPrompt:     finalPrompt,
+		ToolNames:       toolNames,
+		Stream:          stream,
+		Thinking:        thinkingEnabled,
+		Search:          searchEnabled,
+		PassThrough:     passThrough,
 	}, nil
 }

internal/httpapi/gemini/handler_generate.go

@@ -2,8 +2,8 @@ package gemini
 
 import (
 	"bytes"
-	"ds2api/internal/toolcall"
 	"encoding/json"
+	"errors"
 	"io"
 	"net/http"
 	"net/http/httptest"
@@ -11,7 +11,13 @@ import (
 
 	"github.com/go-chi/chi/v5"
 
+	"ds2api/internal/assistantturn"
+	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
+	"ds2api/internal/httpapi/requestbody"
+	"ds2api/internal/promptcompat"
 	"ds2api/internal/sse"
+	"ds2api/internal/toolcall"
 	"ds2api/internal/translatorcliproxy"
 	"ds2api/internal/util"
 
@@ -19,20 +25,90 @@ import (
 )
 
 func (h *Handler) handleGenerateContent(w http.ResponseWriter, r *http.Request, stream bool) {
-	if h.OpenAI == nil {
-		writeGeminiError(w, http.StatusInternalServerError, "OpenAI proxy backend unavailable.")
+	if isGeminiVercelProxyRequest(r) && h.proxyViaOpenAI(w, r, stream) {
 		return
 	}
-	if h.proxyViaOpenAI(w, r, stream) {
+	if h.Auth == nil || h.DS == nil {
+		if h.OpenAI != nil && h.proxyViaOpenAI(w, r, stream) {
+			return
+		}
+		writeGeminiError(w, http.StatusInternalServerError, "Gemini runtime backend unavailable.")
 		return
 	}
-	writeGeminiError(w, http.StatusBadGateway, "Failed to proxy Gemini request.")
+	if h.handleGeminiDirect(w, r, stream) {
+		return
+	}
+	writeGeminiError(w, http.StatusBadGateway, "Failed to handle Gemini request.")
+}
+
+func isGeminiVercelProxyRequest(r *http.Request) bool {
+	if r == nil || r.URL == nil {
+		return false
+	}
+	return strings.TrimSpace(r.URL.Query().Get("__stream_prepare")) == "1" ||
+		strings.TrimSpace(r.URL.Query().Get("__stream_release")) == "1"
+}
+
+func (h *Handler) handleGeminiDirect(w http.ResponseWriter, r *http.Request, stream bool) bool {
+	raw, err := io.ReadAll(r.Body)
+	if err != nil {
+		if errors.Is(err, requestbody.ErrInvalidUTF8Body) {
+			writeGeminiError(w, http.StatusBadRequest, "invalid json")
+		} else {
+			writeGeminiError(w, http.StatusBadRequest, "invalid body")
+		}
+		return true
+	}
+	routeModel := strings.TrimSpace(chi.URLParam(r, "model"))
+	var req map[string]any
+	if err := json.Unmarshal(raw, &req); err != nil {
+		writeGeminiError(w, http.StatusBadRequest, "invalid json")
+		return true
+	}
+	stdReq, err := normalizeGeminiRequest(h.Store, routeModel, req, stream)
+	if err != nil {
+		writeGeminiError(w, http.StatusBadRequest, err.Error())
+		return true
+	}
+	a, err := h.Auth.Determine(r)
+	if err != nil {
+		writeGeminiError(w, http.StatusUnauthorized, err.Error())
+		return true
+	}
+	defer h.Auth.Release(a)
+	if stream {
+		h.handleGeminiDirectStream(w, r, a, stdReq)
+		return true
+	}
+	result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
+		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		RetryEnabled:          true,
+	})
+	if outErr != nil {
+		writeGeminiError(w, outErr.Status, outErr.Message)
+		return true
+	}
+	writeJSON(w, http.StatusOK, buildGeminiGenerateContentResponseFromTurn(result.Turn))
+	return true
+}
+
+func (h *Handler) handleGeminiDirectStream(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, stdReq promptcompat.StandardRequest) {
+	start, outErr := completionruntime.StartCompletion(r.Context(), h.DS, a, stdReq, completionruntime.Options{})
+	if outErr != nil {
+		writeGeminiError(w, outErr.Status, outErr.Message)
+		return
+	}
+	h.handleStreamGenerateContent(w, r, start.Response, stdReq.ResponseModel, stdReq.PromptTokenText, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw)
 }
 
 func (h *Handler) proxyViaOpenAI(w http.ResponseWriter, r *http.Request, stream bool) bool {
 	raw, err := io.ReadAll(r.Body)
 	if err != nil {
-		writeGeminiError(w, http.StatusBadRequest, "invalid body")
+		if errors.Is(err, requestbody.ErrInvalidUTF8Body) {
+			writeGeminiError(w, http.StatusBadRequest, "invalid json")
+		} else {
+			writeGeminiError(w, http.StatusBadRequest, "invalid body")
+		}
 		return true
 	}
 	routeModel := strings.TrimSpace(chi.URLParam(r, "model"))
@@ -227,7 +303,7 @@ func (h *Handler) handleNonStreamGenerateContent(w http.ResponseWriter, resp *ht
 //nolint:unused // retained for native Gemini non-stream handling path.
 func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, finalText string, toolNames []string) map[string]any {
 	parts := buildGeminiPartsFromFinal(finalText, finalThinking, toolNames)
-	usage := buildGeminiUsage(finalPrompt, finalThinking, finalText)
+	usage := buildGeminiUsage(model, finalPrompt, finalThinking, finalText)
 	return map[string]any{
 		"candidates": []map[string]any{
 			{
@@ -244,11 +320,53 @@ func buildGeminiGenerateContentResponse(model, finalPrompt, finalThinking, final
 	}
 }
 
+func buildGeminiGenerateContentResponseFromTurn(turn assistantturn.Turn) map[string]any {
+	parts := buildGeminiPartsFromTurn(turn)
+	return map[string]any{
+		"candidates": []map[string]any{
+			{
+				"index": 0,
+				"content": map[string]any{
+					"role":  "model",
+					"parts": parts,
+				},
+				"finishReason": "STOP",
+			},
+		},
+		"modelVersion": turn.Model,
+		"usageMetadata": map[string]any{
+			"promptTokenCount":     turn.Usage.InputTokens,
+			"candidatesTokenCount": turn.Usage.OutputTokens,
+			"totalTokenCount":      turn.Usage.TotalTokens,
+		},
+	}
+}
+
+func buildGeminiPartsFromTurn(turn assistantturn.Turn) []map[string]any {
+	if len(turn.ToolCalls) > 0 {
+		parts := make([]map[string]any, 0, len(turn.ToolCalls))
+		for _, tc := range turn.ToolCalls {
+			parts = append(parts, map[string]any{
+				"functionCall": map[string]any{
+					"name": tc.Name,
+					"args": tc.Input,
+				},
+			})
+		}
+		return parts
+	}
+	text := turn.Text
+	if text == "" {
+		text = turn.Thinking
+	}
+	return []map[string]any{{"text": text}}
+}
+
 //nolint:unused // retained for native Gemini non-stream handling path.
-func buildGeminiUsage(finalPrompt, finalThinking, finalText string) map[string]any {
-	promptTokens := util.EstimateTokens(finalPrompt)
-	reasoningTokens := util.EstimateTokens(finalThinking)
-	completionTokens := util.EstimateTokens(finalText)
+func buildGeminiUsage(model, finalPrompt, finalThinking, finalText string) map[string]any {
+	promptTokens := util.CountPromptTokens(finalPrompt, model)
+	reasoningTokens := util.CountOutputTokens(finalThinking, model)
+	completionTokens := util.CountOutputTokens(finalText, model)
 	return map[string]any{
 		"promptTokenCount":     promptTokens,
 		"candidatesTokenCount": reasoningTokens + completionTokens,

internal/httpapi/gemini/handler_stream_runtime.go

@@ -7,13 +7,14 @@ import (
 	"strings"
 	"time"
 
+	"ds2api/internal/assistantturn"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
 )
 
 //nolint:unused // retained for native Gemini stream handling path.
-func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Request, resp *http.Response, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string) {
+func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Request, resp *http.Response, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any) {
 	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
@@ -28,7 +29,7 @@ func (h *Handler) handleStreamGenerateContent(w http.ResponseWriter, r *http.Req
 
 	rc := http.NewResponseController(w)
 	_, canFlush := w.(http.Flusher)
-	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames)
+	runtime := newGeminiStreamRuntime(w, rc, canFlush, model, finalPrompt, thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames, toolsRaw)
 
 	initialType := "text"
 	if thinkingEnabled {
@@ -64,9 +65,11 @@ type geminiStreamRuntime struct {
 	bufferContent         bool
 	stripReferenceMarkers bool
 	toolNames             []string
+	toolsRaw              any
 
-	thinking strings.Builder
-	text     strings.Builder
+	accumulator       *assistantturn.Accumulator
+	contentFilter     bool
+	responseMessageID int
 }
 
 //nolint:unused // retained for native Gemini stream handling path.
@@ -80,6 +83,7 @@ func newGeminiStreamRuntime(
 	searchEnabled bool,
 	stripReferenceMarkers bool,
 	toolNames []string,
+	toolsRaw any,
 ) *geminiStreamRuntime {
 	return &geminiStreamRuntime{
 		w:                     w,
@@ -92,6 +96,12 @@ func newGeminiStreamRuntime(
 		bufferContent:         len(toolNames) > 0,
 		stripReferenceMarkers: stripReferenceMarkers,
 		toolNames:             toolNames,
+		toolsRaw:              toolsRaw,
+		accumulator: assistantturn.NewAccumulator(assistantturn.AccumulatorOptions{
+			ThinkingEnabled:       thinkingEnabled,
+			SearchEnabled:         searchEnabled,
+			StripReferenceMarkers: stripReferenceMarkers,
+		}),
 	}
 }
 
@@ -111,35 +121,24 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.ResponseMessageID > 0 {
+		s.responseMessageID = parsed.ResponseMessageID
+	}
 	if parsed.ContentFilter || parsed.ErrorMessage != "" || parsed.Stop {
+		if parsed.ContentFilter {
+			s.contentFilter = true
+		}
 		return streamengine.ParsedDecision{Stop: true}
 	}
 
-	contentSeen := false
-	for _, p := range parsed.Parts {
-		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
-		if cleanedText == "" {
-			continue
-		}
-		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(cleanedText) {
-			continue
-		}
-		contentSeen = true
+	accumulated := s.accumulator.Apply(parsed)
+	for _, p := range accumulated.Parts {
 		if p.Type == "thinking" {
-			if s.thinkingEnabled {
-				trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
-				if trimmed == "" {
-					continue
-				}
-				s.thinking.WriteString(trimmed)
-			}
 			continue
 		}
-		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
-		if trimmed == "" {
+		if p.RawText == "" || p.CitationOnly || p.VisibleText == "" {
 			continue
 		}
-		s.text.WriteString(trimmed)
 		if s.bufferContent {
 			continue
 		}
@@ -149,23 +148,38 @@ func (s *geminiStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Parse
 					"index": 0,
 					"content": map[string]any{
 						"role":  "model",
-						"parts": []map[string]any{{"text": trimmed}},
+						"parts": []map[string]any{{"text": p.VisibleText}},
 					},
 				},
 			},
 			"modelVersion": s.model,
 		})
 	}
-	return streamengine.ParsedDecision{ContentSeen: contentSeen}
+	return streamengine.ParsedDecision{ContentSeen: accumulated.ContentSeen}
 }
 
 //nolint:unused // retained for native Gemini stream handling path.
 func (s *geminiStreamRuntime) finalize() {
-	finalThinking := s.thinking.String()
-	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
+	rawText, text, rawThinking, thinking, detectionThinking := s.accumulator.Snapshot()
+	turn := assistantturn.BuildTurnFromStreamSnapshot(assistantturn.StreamSnapshot{
+		RawText:           rawText,
+		VisibleText:       text,
+		RawThinking:       rawThinking,
+		VisibleThinking:   thinking,
+		DetectionThinking: detectionThinking,
+		ContentFilter:     s.contentFilter,
+		ResponseMessageID: s.responseMessageID,
+	}, assistantturn.BuildOptions{
+		Model:                 s.model,
+		Prompt:                s.finalPrompt,
+		SearchEnabled:         s.searchEnabled,
+		StripReferenceMarkers: s.stripReferenceMarkers,
+		ToolNames:             s.toolNames,
+		ToolsRaw:              s.toolsRaw,
+	})
 
 	if s.bufferContent {
-		parts := buildGeminiPartsFromFinal(finalText, finalThinking, s.toolNames)
+		parts := buildGeminiPartsFromTurn(turn)
 		s.sendChunk(map[string]any{
 			"candidates": []map[string]any{
 				{
@@ -193,7 +207,11 @@ func (s *geminiStreamRuntime) finalize() {
 				"finishReason": "STOP",
 			},
 		},
-		"modelVersion":  s.model,
-		"usageMetadata": buildGeminiUsage(s.finalPrompt, finalThinking, finalText),
+		"modelVersion": s.model,
+		"usageMetadata": map[string]any{
+			"promptTokenCount":     turn.Usage.InputTokens,
+			"candidatesTokenCount": turn.Usage.OutputTokens,
+			"totalTokenCount":      turn.Usage.TotalTokens,
+		},
 	})
 }

internal/httpapi/openai/chat/chat_history_test.go

@@ -126,6 +126,7 @@ func TestStartChatHistoryRecoversFromTransientWriteFailure(t *testing.T) {
 	session := startChatHistory(historyStore, req, a, stdReq)
 	if session == nil {
 		t.Fatalf("expected session even when initial persistence fails")
+		return
 	}
 	if session.disabled {
 		t.Fatalf("expected session to remain active after transient start failure")
@@ -194,7 +195,7 @@ func TestHandleStreamContextCancelledMarksHistoryStopped(t *testing.T) {
 	rec := httptest.NewRecorder()
 	resp := makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"hello"}`, `data: [DONE]`)
 
-	h.handleStream(rec, req, resp, "cid-stop", "deepseek-v4-flash", "prompt", false, false, nil, nil, session)
+	h.handleStream(rec, req, resp, "cid-stop", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, session)
 
 	snapshot, err := historyStore.Snapshot()
 	if err != nil {
@@ -310,16 +311,16 @@ func TestChatCompletionsCurrentInputFilePersistsNeutralPrompt(t *testing.T) {
 	if len(ds.uploadCalls) != 1 {
 		t.Fatalf("expected current input upload to happen, got %d", len(ds.uploadCalls))
 	}
-	if ds.uploadCalls[0].Filename != "IGNORE.txt" {
-		t.Fatalf("expected IGNORE.txt upload, got %q", ds.uploadCalls[0].Filename)
+	if ds.uploadCalls[0].Filename != "DS2API_HISTORY.txt" {
+		t.Fatalf("expected DS2API_HISTORY.txt upload, got %q", ds.uploadCalls[0].Filename)
 	}
 	if full.HistoryText != string(ds.uploadCalls[0].Data) {
 		t.Fatalf("expected uploaded current input file to be persisted in history text")
 	}
 	if len(full.Messages) != 1 {
-		t.Fatalf("expected neutral prompt to be the only persisted message, got %#v", full.Messages)
+		t.Fatalf("expected continuation prompt to be the only persisted message, got %#v", full.Messages)
 	}
-	if !strings.Contains(full.Messages[0].Content, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral prompt to be persisted, got %#v", full.Messages[0])
+	if !strings.Contains(full.Messages[0].Content, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation prompt to be persisted, got %#v", full.Messages[0])
 	}
 }

internal/httpapi/openai/chat/chat_stream_runtime.go

@@ -5,7 +5,10 @@ import (
 	"net/http"
 	"strings"
 
+	"ds2api/internal/assistantturn"
 	openaifmt "ds2api/internal/format/openai"
+	"ds2api/internal/httpapi/openai/shared"
+	"ds2api/internal/promptcompat"
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
 	"ds2api/internal/toolstream"
@@ -16,12 +19,14 @@ type chatStreamRuntime struct {
 	rc       *http.ResponseController
 	canFlush bool
 
-	completionID string
-	created      int64
-	model        string
-	finalPrompt  string
-	toolNames    []string
-	toolsRaw     any
+	completionID  string
+	created       int64
+	model         string
+	finalPrompt   string
+	refFileTokens int
+	toolNames     []string
+	toolsRaw      any
+	toolChoice    promptcompat.ToolChoicePolicy
 
 	thinkingEnabled       bool
 	searchEnabled         bool
@@ -33,13 +38,11 @@ type chatStreamRuntime struct {
 	toolCallsEmitted     bool
 	toolCallsDoneEmitted bool
 
-	toolSieve             toolstream.State
-	streamToolCallIDs     map[int]string
-	streamToolNames       map[int]string
-	thinking              strings.Builder
-	toolDetectionThinking strings.Builder
-	text                  strings.Builder
-	responseMessageID     int
+	toolSieve         toolstream.State
+	streamToolCallIDs map[int]string
+	streamToolNames   map[int]string
+	accumulator       shared.StreamAccumulator
+	responseMessageID int
 
 	finalThinking     string
 	finalText         string
@@ -50,6 +53,32 @@ type chatStreamRuntime struct {
 	finalErrorCode    string
 }
 
+type chatDeltaBatch struct {
+	runtime *chatStreamRuntime
+	field   string
+	text    strings.Builder
+}
+
+func (b *chatDeltaBatch) append(field, text string) {
+	if text == "" {
+		return
+	}
+	if b.field != "" && b.field != field {
+		b.flush()
+	}
+	b.field = field
+	b.text.WriteString(text)
+}
+
+func (b *chatDeltaBatch) flush() {
+	if b.field == "" || b.text.Len() == 0 {
+		return
+	}
+	b.runtime.sendDelta(map[string]any{b.field: b.text.String()})
+	b.field = ""
+	b.text.Reset()
+}
+
 func newChatStreamRuntime(
 	w http.ResponseWriter,
 	rc *http.ResponseController,
@@ -63,6 +92,7 @@ func newChatStreamRuntime(
 	stripReferenceMarkers bool,
 	toolNames []string,
 	toolsRaw any,
+	toolChoice promptcompat.ToolChoicePolicy,
 	bufferToolContent bool,
 	emitEarlyToolDeltas bool,
 ) *chatStreamRuntime {
@@ -76,6 +106,7 @@ func newChatStreamRuntime(
 		finalPrompt:           finalPrompt,
 		toolNames:             toolNames,
 		toolsRaw:              toolsRaw,
+		toolChoice:            toolChoice,
 		thinkingEnabled:       thinkingEnabled,
 		searchEnabled:         searchEnabled,
 		stripReferenceMarkers: stripReferenceMarkers,
@@ -83,6 +114,11 @@ func newChatStreamRuntime(
 		emitEarlyToolDeltas:   emitEarlyToolDeltas,
 		streamToolCallIDs:     map[int]string{},
 		streamToolNames:       map[int]string{},
+		accumulator: shared.StreamAccumulator{
+			ThinkingEnabled:       thinkingEnabled,
+			SearchEnabled:         searchEnabled,
+			StripReferenceMarkers: stripReferenceMarkers,
+		},
 	}
 }
 
@@ -91,7 +127,13 @@ func (s *chatStreamRuntime) sendKeepAlive() {
 		return
 	}
 	_, _ = s.w.Write([]byte(": keep-alive\n\n"))
-	_ = s.rc.Flush()
+	s.sendChunk(openaifmt.BuildChatStreamChunk(
+		s.completionID,
+		s.created,
+		s.model,
+		[]map[string]any{},
+		nil,
+	))
 }
 
 func (s *chatStreamRuntime) sendChunk(v any) {
@@ -104,6 +146,23 @@ func (s *chatStreamRuntime) sendChunk(v any) {
 	}
 }
 
+func (s *chatStreamRuntime) sendDelta(delta map[string]any) {
+	if len(delta) == 0 {
+		return
+	}
+	if !s.firstChunkSent {
+		delta["role"] = "assistant"
+		s.firstChunkSent = true
+	}
+	s.sendChunk(openaifmt.BuildChatStreamChunk(
+		s.completionID,
+		s.created,
+		s.model,
+		[]map[string]any{openaifmt.BuildChatStreamDeltaChoice(0, delta)},
+		nil,
+	))
+}
+
 func (s *chatStreamRuntime) sendDone() {
 	_, _ = s.w.Write([]byte("data: [DONE]\n\n"))
 	if s.canFlush {
@@ -127,6 +186,15 @@ func (s *chatStreamRuntime) sendFailedChunk(status int, message, code string) {
 	s.sendDone()
 }
 
+func (s *chatStreamRuntime) markContextCancelled() {
+	s.finalErrorStatus = 499
+	s.finalErrorMessage = "request context cancelled"
+	s.finalErrorCode = string(streamengine.StopReasonContextCancelled)
+	s.finalThinking = s.accumulator.Thinking.String()
+	s.finalText = cleanVisibleOutput(s.accumulator.Text.String(), s.stripReferenceMarkers)
+	s.finalFinishReason = string(streamengine.StopReasonContextCancelled)
+}
+
 func (s *chatStreamRuntime) resetStreamToolCallState() {
 	s.streamToolCallIDs = map[int]string{}
 	s.streamToolNames = map[int]string{}
@@ -136,81 +204,71 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 	s.finalErrorStatus = 0
 	s.finalErrorMessage = ""
 	s.finalErrorCode = ""
-	finalThinking := s.thinking.String()
-	finalToolDetectionThinking := s.toolDetectionThinking.String()
-	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
-	s.finalThinking = finalThinking
-	s.finalText = finalText
-	detected := detectAssistantToolCalls(finalText, finalThinking, finalToolDetectionThinking, s.toolNames)
-	if len(detected.Calls) > 0 && !s.toolCallsDoneEmitted {
+	finalThinking := s.accumulator.Thinking.String()
+	finalToolDetectionThinking := s.accumulator.ToolDetectionThinking.String()
+	finalText := s.accumulator.Text.String()
+	turn := assistantturn.BuildTurnFromStreamSnapshot(assistantturn.StreamSnapshot{
+		RawText:               s.accumulator.RawText.String(),
+		VisibleText:           finalText,
+		RawThinking:           s.accumulator.RawThinking.String(),
+		VisibleThinking:       finalThinking,
+		DetectionThinking:     finalToolDetectionThinking,
+		ContentFilter:         finishReason == "content_filter",
+		ResponseMessageID:     s.responseMessageID,
+		AlreadyEmittedCalls:   s.toolCallsEmitted,
+		AlreadyEmittedToolRaw: s.toolCallsDoneEmitted,
+	}, assistantturn.BuildOptions{
+		Model:                 s.model,
+		Prompt:                s.finalPrompt,
+		RefFileTokens:         s.refFileTokens,
+		SearchEnabled:         s.searchEnabled,
+		StripReferenceMarkers: s.stripReferenceMarkers,
+		ToolNames:             s.toolNames,
+		ToolsRaw:              s.toolsRaw,
+		ToolChoice:            s.toolChoice,
+	})
+	s.finalThinking = turn.Thinking
+	s.finalText = turn.Text
+	if len(turn.ToolCalls) > 0 && !s.toolCallsDoneEmitted {
 		finishReason = "tool_calls"
-		delta := map[string]any{
-			"tool_calls": formatFinalStreamToolCallsWithStableIDs(detected.Calls, s.streamToolCallIDs, s.toolsRaw),
-		}
-		if !s.firstChunkSent {
-			delta["role"] = "assistant"
-			s.firstChunkSent = true
-		}
-		s.sendChunk(openaifmt.BuildChatStreamChunk(
-			s.completionID,
-			s.created,
-			s.model,
-			[]map[string]any{openaifmt.BuildChatStreamDeltaChoice(0, delta)},
-			nil,
-		))
+		s.sendDelta(map[string]any{
+			"tool_calls": formatFinalStreamToolCallsWithStableIDs(turn.ToolCalls, s.streamToolCallIDs, s.toolsRaw),
+		})
 		s.toolCallsEmitted = true
 		s.toolCallsDoneEmitted = true
 	} else if s.bufferToolContent {
+		batch := chatDeltaBatch{runtime: s}
 		for _, evt := range toolstream.Flush(&s.toolSieve, s.toolNames) {
 			if len(evt.ToolCalls) > 0 {
+				batch.flush()
 				finishReason = "tool_calls"
 				s.toolCallsEmitted = true
 				s.toolCallsDoneEmitted = true
-				tcDelta := map[string]any{
+				s.sendDelta(map[string]any{
 					"tool_calls": formatFinalStreamToolCallsWithStableIDs(evt.ToolCalls, s.streamToolCallIDs, s.toolsRaw),
-				}
-				if !s.firstChunkSent {
-					tcDelta["role"] = "assistant"
-					s.firstChunkSent = true
-				}
-				s.sendChunk(openaifmt.BuildChatStreamChunk(
-					s.completionID,
-					s.created,
-					s.model,
-					[]map[string]any{openaifmt.BuildChatStreamDeltaChoice(0, tcDelta)},
-					nil,
-				))
+				})
 				s.resetStreamToolCallState()
 			}
 			if evt.Content == "" {
 				continue
 			}
 			cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
-			if cleaned == "" {
+			if cleaned == "" || (s.searchEnabled && sse.IsCitation(cleaned)) {
 				continue
 			}
-			delta := map[string]any{
-				"content": cleaned,
-			}
-			if !s.firstChunkSent {
-				delta["role"] = "assistant"
-				s.firstChunkSent = true
-			}
-			s.sendChunk(openaifmt.BuildChatStreamChunk(
-				s.completionID,
-				s.created,
-				s.model,
-				[]map[string]any{openaifmt.BuildChatStreamDeltaChoice(0, delta)},
-				nil,
-			))
+			batch.append("content", cleaned)
 		}
+		batch.flush()
 	}
 
-	if len(detected.Calls) > 0 || s.toolCallsEmitted {
+	if len(turn.ToolCalls) > 0 || s.toolCallsEmitted {
 		finishReason = "tool_calls"
 	}
-	if len(detected.Calls) == 0 && !s.toolCallsEmitted && strings.TrimSpace(finalText) == "" {
-		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", finalText, finalThinking)
+	if len(turn.ToolCalls) == 0 && !s.toolCallsEmitted && strings.TrimSpace(turn.Text) == "" {
+		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", turn.Text, turn.Thinking)
+		if turn.Error != nil {
+			status, message, code = turn.Error.Status, turn.Error.Message, turn.Error.Code
+		}
 		if deferEmptyOutput {
 			s.finalErrorStatus = status
 			s.finalErrorMessage = message
@@ -220,7 +278,7 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 		s.sendFailedChunk(status, message, code)
 		return true
 	}
-	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
+	usage := chatUsageFromTurn(turn)
 	s.finalFinishReason = finishReason
 	s.finalUsage = usage
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
@@ -234,6 +292,17 @@ func (s *chatStreamRuntime) finalize(finishReason string, deferEmptyOutput bool)
 	return true
 }
 
+func chatUsageFromTurn(turn assistantturn.Turn) map[string]any {
+	return map[string]any{
+		"prompt_tokens":     turn.Usage.InputTokens,
+		"completion_tokens": turn.Usage.OutputTokens,
+		"total_tokens":      turn.Usage.TotalTokens,
+		"completion_tokens_details": map[string]any{
+			"reasoning_tokens": turn.Usage.ReasoningTokens,
+		},
+	}
+}
+
 func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedDecision {
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
@@ -242,7 +311,7 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 		s.responseMessageID = parsed.ResponseMessageID
 	}
 	if parsed.ContentFilter {
-		if strings.TrimSpace(s.text.String()) == "" {
+		if strings.TrimSpace(s.accumulator.Text.String()) == "" {
 			return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReason("content_filter")}
 		}
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReasonHandlerRequested}
@@ -254,109 +323,65 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 		return streamengine.ParsedDecision{Stop: true, StopReason: streamengine.StopReasonHandlerRequested}
 	}
 
-	newChoices := make([]map[string]any, 0, len(parsed.Parts))
-	contentSeen := false
-	for _, p := range parsed.ToolDetectionThinkingParts {
-		trimmed := sse.TrimContinuationOverlap(s.toolDetectionThinking.String(), p.Text)
-		if trimmed != "" {
-			s.toolDetectionThinking.WriteString(trimmed)
-		}
-	}
-	for _, p := range parsed.Parts {
-		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
-		if s.searchEnabled && sse.IsCitation(cleanedText) {
-			continue
-		}
-		if cleanedText == "" {
-			continue
-		}
-		contentSeen = true
-		delta := map[string]any{}
-		if !s.firstChunkSent {
-			delta["role"] = "assistant"
-			s.firstChunkSent = true
-		}
+	batch := chatDeltaBatch{runtime: s}
+	accumulated := s.accumulator.Apply(parsed)
+	for _, p := range accumulated.Parts {
 		if p.Type == "thinking" {
-			if s.thinkingEnabled {
-				trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
-				if trimmed == "" {
+			batch.append("reasoning_content", p.VisibleText)
+			continue
+		}
+		if p.RawText == "" {
+			continue
+		}
+		if p.CitationOnly {
+			continue
+		}
+		if !s.bufferToolContent {
+			batch.append("content", p.VisibleText)
+		} else {
+			events := toolstream.ProcessChunk(&s.toolSieve, p.RawText, s.toolNames)
+			for _, evt := range events {
+				if len(evt.ToolCallDeltas) > 0 {
+					if !s.emitEarlyToolDeltas {
+						continue
+					}
+					filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.streamToolNames)
+					if len(filtered) == 0 {
+						continue
+					}
+					formatted := formatIncrementalStreamToolCallDeltas(filtered, s.streamToolCallIDs)
+					if len(formatted) == 0 {
+						continue
+					}
+					batch.flush()
+					tcDelta := map[string]any{
+						"tool_calls": formatted,
+					}
+					s.toolCallsEmitted = true
+					s.sendDelta(tcDelta)
 					continue
 				}
-				s.thinking.WriteString(trimmed)
-				delta["reasoning_content"] = trimmed
-			}
-		} else {
-			trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
-			if trimmed == "" {
-				continue
-			}
-			s.text.WriteString(trimmed)
-			if !s.bufferToolContent {
-				delta["content"] = trimmed
-			} else {
-				events := toolstream.ProcessChunk(&s.toolSieve, trimmed, s.toolNames)
-				for _, evt := range events {
-					if len(evt.ToolCallDeltas) > 0 {
-						if !s.emitEarlyToolDeltas {
-							continue
-						}
-						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.streamToolNames)
-						if len(filtered) == 0 {
-							continue
-						}
-						formatted := formatIncrementalStreamToolCallDeltas(filtered, s.streamToolCallIDs)
-						if len(formatted) == 0 {
-							continue
-						}
-						tcDelta := map[string]any{
-							"tool_calls": formatted,
-						}
-						s.toolCallsEmitted = true
-						if !s.firstChunkSent {
-							tcDelta["role"] = "assistant"
-							s.firstChunkSent = true
-						}
-						newChoices = append(newChoices, openaifmt.BuildChatStreamDeltaChoice(0, tcDelta))
+				if len(evt.ToolCalls) > 0 {
+					batch.flush()
+					s.toolCallsEmitted = true
+					s.toolCallsDoneEmitted = true
+					tcDelta := map[string]any{
+						"tool_calls": formatFinalStreamToolCallsWithStableIDs(evt.ToolCalls, s.streamToolCallIDs, s.toolsRaw),
+					}
+					s.sendDelta(tcDelta)
+					s.resetStreamToolCallState()
+					continue
+				}
+				if evt.Content != "" {
+					cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
+					if cleaned == "" || (s.searchEnabled && sse.IsCitation(cleaned)) {
 						continue
 					}
-					if len(evt.ToolCalls) > 0 {
-						s.toolCallsEmitted = true
-						s.toolCallsDoneEmitted = true
-						tcDelta := map[string]any{
-							"tool_calls": formatFinalStreamToolCallsWithStableIDs(evt.ToolCalls, s.streamToolCallIDs, s.toolsRaw),
-						}
-						if !s.firstChunkSent {
-							tcDelta["role"] = "assistant"
-							s.firstChunkSent = true
-						}
-						newChoices = append(newChoices, openaifmt.BuildChatStreamDeltaChoice(0, tcDelta))
-						s.resetStreamToolCallState()
-						continue
-					}
-					if evt.Content != "" {
-						cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
-						if cleaned == "" {
-							continue
-						}
-						contentDelta := map[string]any{
-							"content": cleaned,
-						}
-						if !s.firstChunkSent {
-							contentDelta["role"] = "assistant"
-							s.firstChunkSent = true
-						}
-						newChoices = append(newChoices, openaifmt.BuildChatStreamDeltaChoice(0, contentDelta))
-					}
+					batch.append("content", cleaned)
 				}
 			}
 		}
-		if len(delta) > 0 {
-			newChoices = append(newChoices, openaifmt.BuildChatStreamDeltaChoice(0, delta))
-		}
 	}
-
-	if len(newChoices) > 0 {
-		s.sendChunk(openaifmt.BuildChatStreamChunk(s.completionID, s.created, s.model, newChoices, nil))
-	}
-	return streamengine.ParsedDecision{ContentSeen: contentSeen}
+	batch.flush()
+	return streamengine.ParsedDecision{ContentSeen: accumulated.ContentSeen}
 }

internal/httpapi/openai/chat/chat_stream_runtime_test.go

@@ -0,0 +1,87 @@
+package chat
+
+import (
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+	"time"
+
+	"ds2api/internal/promptcompat"
+)
+
+func TestChatStreamKeepAliveEmitsEmptyChoiceDataFrame(t *testing.T) {
+	rec := httptest.NewRecorder()
+	runtime := newChatStreamRuntime(
+		rec,
+		http.NewResponseController(rec),
+		true,
+		"chatcmpl-test",
+		time.Now().Unix(),
+		"deepseek-v4-flash",
+		"prompt",
+		false,
+		false,
+		true,
+		nil,
+		nil,
+		promptcompat.DefaultToolChoicePolicy(),
+		false,
+		false,
+	)
+
+	runtime.sendKeepAlive()
+
+	body := rec.Body.String()
+	if !strings.Contains(body, ": keep-alive\n\n") {
+		t.Fatalf("expected keep-alive comment, got %q", body)
+	}
+	frames, done := parseSSEDataFrames(t, body)
+	if done {
+		t.Fatalf("keep-alive must not emit [DONE], body=%q", body)
+	}
+	if len(frames) != 1 {
+		t.Fatalf("expected one data frame, got %d body=%q", len(frames), body)
+	}
+	if got := asString(frames[0]["id"]); got != "chatcmpl-test" {
+		t.Fatalf("expected completion id to be preserved, got %q", got)
+	}
+	if got := asString(frames[0]["object"]); got != "chat.completion.chunk" {
+		t.Fatalf("expected chat chunk object, got %q", got)
+	}
+	choices, _ := frames[0]["choices"].([]any)
+	if len(choices) != 0 {
+		t.Fatalf("expected empty choices heartbeat, got %#v", choices)
+	}
+}
+
+func TestChatStreamFinalizeEnforcesRequiredToolChoice(t *testing.T) {
+	rec := httptest.NewRecorder()
+	runtime := newChatStreamRuntime(
+		rec,
+		http.NewResponseController(rec),
+		true,
+		"chatcmpl-test",
+		time.Now().Unix(),
+		"deepseek-v4-flash",
+		"prompt",
+		false,
+		false,
+		true,
+		[]string{"Write"},
+		nil,
+		promptcompat.ToolChoicePolicy{Mode: promptcompat.ToolChoiceRequired},
+		true,
+		false,
+	)
+
+	if !runtime.finalize("stop", false) {
+		t.Fatalf("expected terminal error to be written")
+	}
+	if runtime.finalErrorCode != "tool_choice_violation" {
+		t.Fatalf("expected tool_choice_violation, got %q body=%s", runtime.finalErrorCode, rec.Body.String())
+	}
+	if !strings.Contains(rec.Body.String(), "tool_choice requires") {
+		t.Fatalf("expected tool choice error in stream body, got %s", rec.Body.String())
+	}
+}

internal/httpapi/openai/chat/empty_retry_runtime.go

@@ -7,15 +7,19 @@ import (
 	"strings"
 	"time"
 
+	"ds2api/internal/assistantturn"
 	"ds2api/internal/auth"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	openaifmt "ds2api/internal/format/openai"
+	"ds2api/internal/promptcompat"
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
 )
 
 type chatNonStreamResult struct {
+	rawThinking           string
+	rawText               string
 	thinking              string
 	toolDetectionThinking string
 	text                  string
@@ -24,13 +28,15 @@ type chatNonStreamResult struct {
 	body                  map[string]any
 	finishReason          string
 	responseMessageID     int
+	outputError           *assistantturn.OutputError
 }
 
-func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
+func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
 	attempts := 0
 	currentResp := resp
 	usagePrompt := finalPrompt
 	accumulatedThinking := ""
+	accumulatedRawThinking := ""
 	accumulatedToolDetectionThinking := ""
 	for {
 		result, ok := h.collectChatNonStreamAttempt(w, currentResp, completionID, model, usagePrompt, thinkingEnabled, searchEnabled, toolNames, toolsRaw)
@@ -38,15 +44,18 @@ func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Co
 			return
 		}
 		accumulatedThinking += sse.TrimContinuationOverlap(accumulatedThinking, result.thinking)
+		accumulatedRawThinking += sse.TrimContinuationOverlap(accumulatedRawThinking, result.rawThinking)
 		accumulatedToolDetectionThinking += sse.TrimContinuationOverlap(accumulatedToolDetectionThinking, result.toolDetectionThinking)
 		result.thinking = accumulatedThinking
+		result.rawThinking = accumulatedRawThinking
 		result.toolDetectionThinking = accumulatedToolDetectionThinking
-		detected := detectAssistantToolCalls(result.text, result.thinking, result.toolDetectionThinking, toolNames)
+		detected := detectAssistantToolCalls(result.rawText, result.text, result.rawThinking, result.toolDetectionThinking, toolNames)
 		result.detectedCalls = len(detected.Calls)
 		result.body = openaifmt.BuildChatCompletionWithToolCalls(completionID, model, usagePrompt, result.thinking, result.text, detected.Calls, toolsRaw)
+		addRefFileTokensToUsage(result.body, refFileTokens)
 		result.finishReason = chatFinishReason(result.body)
 		if !shouldRetryChatNonStream(result, attempts) {
-			h.finishChatNonStreamResult(w, result, attempts, usagePrompt, historySession)
+			h.finishChatNonStreamResult(w, result, attempts, usagePrompt, refFileTokens, historySession)
 			return
 		}
 
@@ -67,7 +76,7 @@ func (h *Handler) handleNonStreamWithRetry(w http.ResponseWriter, ctx context.Co
 			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "chat.completions", "stream", false, "retry_attempt", attempts, "error", err)
 			return
 		}
-		usagePrompt = usagePromptWithEmptyOutputRetry(finalPrompt, attempts)
+		usagePrompt = usagePromptWithEmptyOutputRetry(usagePrompt, attempts)
 		currentResp = nextResp
 	}
 }
@@ -80,39 +89,45 @@ func (h *Handler) collectChatNonStreamAttempt(w http.ResponseWriter, resp *http.
 		return chatNonStreamResult{}, false
 	}
 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
-	finalThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	finalToolDetectionThinking := cleanVisibleOutput(result.ToolDetectionThinking, stripReferenceMarkers)
-	finalText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if searchEnabled {
-		finalText = replaceCitationMarkersWithLinks(finalText, result.CitationLinks)
-	}
-	detected := detectAssistantToolCalls(finalText, finalThinking, finalToolDetectionThinking, toolNames)
-	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, usagePrompt, finalThinking, finalText, detected.Calls, toolsRaw)
+	turn := assistantturn.BuildTurnFromCollected(result, assistantturn.BuildOptions{
+		Model:                 model,
+		Prompt:                usagePrompt,
+		SearchEnabled:         searchEnabled,
+		StripReferenceMarkers: h.compatStripReferenceMarkers(),
+		ToolNames:             toolNames,
+		ToolsRaw:              toolsRaw,
+	})
+	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, usagePrompt, turn.Thinking, turn.Text, turn.ToolCalls, toolsRaw)
 	return chatNonStreamResult{
-		thinking:              finalThinking,
-		toolDetectionThinking: finalToolDetectionThinking,
-		text:                  finalText,
+		rawThinking:           result.Thinking,
+		rawText:               result.Text,
+		thinking:              turn.Thinking,
+		toolDetectionThinking: result.ToolDetectionThinking,
+		text:                  turn.Text,
 		contentFilter:         result.ContentFilter,
-		detectedCalls:         len(detected.Calls),
+		detectedCalls:         len(turn.ToolCalls),
 		body:                  respBody,
 		finishReason:          chatFinishReason(respBody),
 		responseMessageID:     result.ResponseMessageID,
+		outputError:           turn.Error,
 	}, true
 }
 
-func (h *Handler) finishChatNonStreamResult(w http.ResponseWriter, result chatNonStreamResult, attempts int, usagePrompt string, historySession *chatHistorySession) {
-	if result.detectedCalls == 0 && shouldWriteUpstreamEmptyOutputError(result.text) {
+func (h *Handler) finishChatNonStreamResult(w http.ResponseWriter, result chatNonStreamResult, attempts int, usagePrompt string, refFileTokens int, historySession *chatHistorySession) {
+	if result.detectedCalls == 0 && strings.TrimSpace(result.text) == "" {
 		status, message, code := upstreamEmptyOutputDetail(result.contentFilter, result.text, result.thinking)
+		if result.outputError != nil {
+			status, message, code = result.outputError.Status, result.outputError.Message, result.outputError.Code
+		}
 		if historySession != nil {
 			historySession.error(status, message, code, result.thinking, result.text)
 		}
-		writeUpstreamEmptyOutputError(w, result.text, result.thinking, result.contentFilter)
+		writeOpenAIErrorWithCode(w, status, message, code)
 		config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "chat.completions", "stream", false, "retry_attempts", attempts, "success_source", "none", "content_filter", result.contentFilter)
 		return
 	}
 	if historySession != nil {
-		historySession.success(http.StatusOK, result.thinking, result.text, result.finishReason, openaifmt.BuildChatUsage(usagePrompt, result.thinking, result.text))
+		historySession.success(http.StatusOK, result.thinking, result.text, result.finishReason, openaifmt.BuildChatUsageForModel("", usagePrompt, result.thinking, result.text, refFileTokens))
 	}
 	writeJSON(w, http.StatusOK, result.body)
 	source := "first_attempt"
@@ -136,11 +151,12 @@ func shouldRetryChatNonStream(result chatNonStreamResult, attempts int) bool {
 		attempts < emptyOutputRetryMaxAttempts() &&
 		!result.contentFilter &&
 		result.detectedCalls == 0 &&
-		strings.TrimSpace(result.text) == ""
+		strings.TrimSpace(result.text) == "" &&
+		strings.TrimSpace(result.thinking) == ""
 }
 
-func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
-	streamRuntime, initialType, ok := h.prepareChatStreamRuntime(w, resp, completionID, model, finalPrompt, thinkingEnabled, searchEnabled, toolNames, toolsRaw, historySession)
+func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, historySession *chatHistorySession) {
+	streamRuntime, initialType, ok := h.prepareChatStreamRuntime(w, resp, completionID, model, finalPrompt, refFileTokens, thinkingEnabled, searchEnabled, toolNames, toolsRaw, toolChoice, historySession)
 	if !ok {
 		return
 	}
@@ -182,7 +198,7 @@ func (h *Handler) handleStreamWithRetry(w http.ResponseWriter, r *http.Request,
 	}
 }
 
-func (h *Handler) prepareChatStreamRuntime(w http.ResponseWriter, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) (*chatStreamRuntime, string, bool) {
+func (h *Handler) prepareChatStreamRuntime(w http.ResponseWriter, resp *http.Response, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, historySession *chatHistorySession) (*chatStreamRuntime, string, bool) {
 	if resp.StatusCode != http.StatusOK {
 		defer func() { _ = resp.Body.Close() }()
 		body, _ := io.ReadAll(resp.Body)
@@ -208,8 +224,10 @@ func (h *Handler) prepareChatStreamRuntime(w http.ResponseWriter, resp *http.Res
 	streamRuntime := newChatStreamRuntime(
 		w, rc, canFlush, completionID, time.Now().Unix(), model, finalPrompt,
 		thinkingEnabled, searchEnabled, h.compatStripReferenceMarkers(), toolNames, toolsRaw,
+		toolChoice,
 		len(toolNames) > 0, h.toolcallFeatureMatchEnabled() && h.toolcallEarlyEmitHighConfidence(),
 	)
+	streamRuntime.refFileTokens = refFileTokens
 	return streamRuntime, initialType, true
 }
 
@@ -229,7 +247,7 @@ func (h *Handler) consumeChatStreamAttempt(r *http.Request, resp *http.Response,
 		OnParsed: func(parsed sse.LineResult) streamengine.ParsedDecision {
 			decision := streamRuntime.onParsed(parsed)
 			if historySession != nil {
-				historySession.progress(streamRuntime.thinking.String(), streamRuntime.text.String())
+				historySession.progress(streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String())
 			}
 			return decision
 		},
@@ -239,11 +257,15 @@ func (h *Handler) consumeChatStreamAttempt(r *http.Request, resp *http.Response,
 			}
 		},
 		OnContextDone: func() {
+			streamRuntime.markContextCancelled()
 			if historySession != nil {
-				historySession.stopped(streamRuntime.thinking.String(), streamRuntime.text.String(), string(streamengine.StopReasonContextCancelled))
+				historySession.stopped(streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String(), string(streamengine.StopReasonContextCancelled))
 			}
 		},
 	})
+	if streamRuntime.finalErrorCode == string(streamengine.StopReasonContextCancelled) {
+		return true, false
+	}
 	terminalWritten := streamRuntime.finalize(finalReason, allowDeferEmpty && finalReason != "content_filter")
 	if terminalWritten {
 		recordChatStreamHistory(streamRuntime, historySession)
@@ -257,7 +279,7 @@ func recordChatStreamHistory(streamRuntime *chatStreamRuntime, historySession *c
 		return
 	}
 	if streamRuntime.finalErrorMessage != "" {
-		historySession.error(streamRuntime.finalErrorStatus, streamRuntime.finalErrorMessage, streamRuntime.finalErrorCode, streamRuntime.thinking.String(), streamRuntime.text.String())
+		historySession.error(streamRuntime.finalErrorStatus, streamRuntime.finalErrorMessage, streamRuntime.finalErrorCode, streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String())
 		return
 	}
 	historySession.success(http.StatusOK, streamRuntime.finalThinking, streamRuntime.finalText, streamRuntime.finalFinishReason, streamRuntime.finalUsage)
@@ -266,7 +288,7 @@ func recordChatStreamHistory(streamRuntime *chatStreamRuntime, historySession *c
 func failChatStreamRetry(streamRuntime *chatStreamRuntime, historySession *chatHistorySession, status int, message, code string) {
 	streamRuntime.sendFailedChunk(status, message, code)
 	if historySession != nil {
-		historySession.error(status, message, code, streamRuntime.thinking.String(), streamRuntime.text.String())
+		historySession.error(status, message, code, streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String())
 	}
 }
 
@@ -275,6 +297,10 @@ func logChatStreamTerminal(streamRuntime *chatStreamRuntime, attempts int) {
 	if attempts > 0 {
 		source = "synthetic_retry"
 	}
+	if streamRuntime.finalErrorCode == string(streamengine.StopReasonContextCancelled) {
+		config.Logger.Info("[openai_empty_retry] terminal cancelled", "surface", "chat.completions", "stream", true, "retry_attempts", attempts, "error_code", streamRuntime.finalErrorCode)
+		return
+	}
 	if streamRuntime.finalErrorMessage != "" {
 		config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "chat.completions", "stream", true, "retry_attempts", attempts, "success_source", "none", "error_code", streamRuntime.finalErrorCode)
 		return

internal/httpapi/openai/chat/empty_retry_runtime_test.go

@@ -0,0 +1,87 @@
+package chat
+
+import (
+	"context"
+	"net/http"
+	"net/http/httptest"
+	"testing"
+	"time"
+
+	"ds2api/internal/chathistory"
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/stream"
+)
+
+func TestConsumeChatStreamAttemptMarksContextCancelledState(t *testing.T) {
+	historyStore := newTestChatHistoryStore(t)
+	entry, err := historyStore.Start(chathistory.StartParams{
+		CallerID:  "caller:test",
+		Model:     "deepseek-v4-flash",
+		Stream:    true,
+		UserInput: "hello",
+	})
+	if err != nil {
+		t.Fatalf("start history failed: %v", err)
+	}
+	session := &chatHistorySession{
+		store:       historyStore,
+		entryID:     entry.ID,
+		startedAt:   time.Now(),
+		lastPersist: time.Now(),
+		finalPrompt: "prompt",
+	}
+
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil).WithContext(ctx)
+	rec := httptest.NewRecorder()
+	streamRuntime := newChatStreamRuntime(
+		rec,
+		http.NewResponseController(rec),
+		true,
+		"cid-cancelled",
+		time.Now().Unix(),
+		"deepseek-v4-flash",
+		"prompt",
+		false,
+		false,
+		true,
+		nil,
+		nil,
+		promptcompat.DefaultToolChoicePolicy(),
+		false,
+		false,
+	)
+	resp := makeOpenAISSEHTTPResponse(
+		`data: {"p":"response/content","v":"hello"}`,
+		`data: [DONE]`,
+	)
+
+	h := &Handler{}
+	terminalWritten, retryable := h.consumeChatStreamAttempt(req, resp, streamRuntime, "text", false, session, true)
+	if !terminalWritten || retryable {
+		t.Fatalf("expected cancelled attempt to terminate without retry, got terminalWritten=%v retryable=%v", terminalWritten, retryable)
+	}
+	if got, want := streamRuntime.finalErrorCode, string(stream.StopReasonContextCancelled); got != want {
+		t.Fatalf("expected cancelled final error code %q, got %q", want, got)
+	}
+	if streamRuntime.finalErrorMessage == "" {
+		t.Fatalf("expected cancelled final error message to be preserved")
+	}
+
+	snapshot, err := historyStore.Snapshot()
+	if err != nil {
+		t.Fatalf("snapshot failed: %v", err)
+	}
+	if len(snapshot.Items) != 1 {
+		t.Fatalf("expected one history item, got %d", len(snapshot.Items))
+	}
+	full, err := historyStore.Get(snapshot.Items[0].ID)
+	if err != nil {
+		t.Fatalf("get detail failed: %v", err)
+	}
+	if full.Status != "stopped" {
+		t.Fatalf("expected stopped status, got %#v", full)
+	}
+}

internal/httpapi/openai/chat/handler.go

@@ -80,6 +80,10 @@ func writeOpenAIError(w http.ResponseWriter, status int, message string) {
 	shared.WriteOpenAIError(w, status, message)
 }
 
+func writeOpenAIErrorWithCode(w http.ResponseWriter, status int, message, code string) {
+	shared.WriteOpenAIErrorWithCode(w, status, message, code)
+}
+
 func openAIErrorType(status int) string {
 	return shared.OpenAIErrorType(status)
 }
@@ -108,8 +112,8 @@ func replaceCitationMarkersWithLinks(text string, links map[int]string) string {
 	return shared.ReplaceCitationMarkersWithLinks(text, links)
 }
 
-func shouldWriteUpstreamEmptyOutputError(text string) bool {
-	return shared.ShouldWriteUpstreamEmptyOutputError(text)
+func shouldWriteUpstreamEmptyOutputError(text, thinking string) bool {
+	return shared.ShouldWriteUpstreamEmptyOutputError(text, thinking)
 }
 
 func upstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
@@ -148,6 +152,6 @@ func formatFinalStreamToolCallsWithStableIDs(calls []toolcall.ParsedToolCall, id
 	return shared.FormatFinalStreamToolCallsWithStableIDs(calls, ids, toolsRaw)
 }
 
-func detectAssistantToolCalls(text, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
-	return shared.DetectAssistantToolCalls(text, exposedThinking, detectionThinking, toolNames)
+func detectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
+	return shared.DetectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking, toolNames)
 }

internal/httpapi/openai/chat/handler_chat.go

@@ -9,6 +9,7 @@ import (
 	"time"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	openaifmt "ds2api/internal/format/openai"
@@ -76,43 +77,40 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 	}
 	historySession := startChatHistory(h.ChatHistory, r, a, stdReq)
 
-	sessionID, err = h.DS.CreateSession(r.Context(), a, 3)
-	if err != nil {
-		if a.UseConfigToken {
+	if !stdReq.Stream {
+		result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
+			StripReferenceMarkers: h.compatStripReferenceMarkers(),
+			RetryEnabled:          true,
+		})
+		sessionID = result.SessionID
+		if outErr != nil {
 			if historySession != nil {
-				historySession.error(http.StatusUnauthorized, "Account token is invalid. Please re-login the account in admin.", "error", "", "")
+				historySession.error(outErr.Status, outErr.Message, outErr.Code, result.Turn.Thinking, result.Turn.Text)
 			}
-			writeOpenAIError(w, http.StatusUnauthorized, "Account token is invalid. Please re-login the account in admin.")
-		} else {
-			if historySession != nil {
-				historySession.error(http.StatusUnauthorized, "Invalid token. If this should be a DS2API key, add it to config.keys first.", "error", "", "")
-			}
-			writeOpenAIError(w, http.StatusUnauthorized, "Invalid token. If this should be a DS2API key, add it to config.keys first.")
+			writeOpenAIErrorWithCode(w, outErr.Status, outErr.Message, outErr.Code)
+			return
 		}
-		return
-	}
-	pow, err := h.DS.GetPow(r.Context(), a, 3)
-	if err != nil {
+		respBody := openaifmt.BuildChatCompletionWithToolCalls(result.SessionID, stdReq.ResponseModel, result.Turn.Prompt, result.Turn.Thinking, result.Turn.Text, result.Turn.ToolCalls, stdReq.ToolsRaw)
+		respBody["usage"] = chatUsageFromTurn(result.Turn)
+		finishReason := chatFinishReason(respBody)
 		if historySession != nil {
-			historySession.error(http.StatusUnauthorized, "Failed to get PoW (invalid token or unknown error).", "error", "", "")
+			historySession.success(http.StatusOK, result.Turn.Thinking, result.Turn.Text, finishReason, chatUsageFromTurn(result.Turn))
 		}
-		writeOpenAIError(w, http.StatusUnauthorized, "Failed to get PoW (invalid token or unknown error).")
+		writeJSON(w, http.StatusOK, respBody)
 		return
 	}
-	payload := stdReq.CompletionPayload(sessionID)
-	resp, err := h.DS.CallCompletion(r.Context(), a, payload, pow, 3)
-	if err != nil {
+
+	start, outErr := completionruntime.StartCompletion(r.Context(), h.DS, a, stdReq, completionruntime.Options{})
+	sessionID = start.SessionID
+	if outErr != nil {
 		if historySession != nil {
-			historySession.error(http.StatusInternalServerError, "Failed to get completion.", "error", "", "")
+			historySession.error(outErr.Status, outErr.Message, outErr.Code, "", "")
 		}
-		writeOpenAIError(w, http.StatusInternalServerError, "Failed to get completion.")
+		writeOpenAIErrorWithCode(w, outErr.Status, outErr.Message, outErr.Code)
 		return
 	}
-	if stdReq.Stream {
-		h.handleStreamWithRetry(w, r, a, resp, payload, pow, sessionID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, historySession)
-		return
-	}
-	h.handleNonStreamWithRetry(w, r.Context(), a, resp, payload, pow, sessionID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, historySession)
+	refFileTokens := stdReq.RefFileTokens
+	h.handleStreamWithRetry(w, r, a, start.Response, start.Payload, start.Pow, sessionID, stdReq.ResponseModel, stdReq.PromptTokenText, refFileTokens, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, stdReq.ToolChoice, historySession)
 }
 
 func (h *Handler) autoDeleteRemoteSession(ctx context.Context, a *auth.RequestAuth, sessionID string) {
@@ -148,7 +146,7 @@ func (h *Handler) autoDeleteRemoteSession(ctx context.Context, a *auth.RequestAu
 	}
 }
 
-func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
+func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
 	if resp.StatusCode != http.StatusOK {
 		defer func() { _ = resp.Body.Close() }()
 		body, _ := io.ReadAll(resp.Body)
@@ -162,13 +160,12 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, co
 
 	stripReferenceMarkers := h.compatStripReferenceMarkers()
 	finalThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	finalToolDetectionThinking := cleanVisibleOutput(result.ToolDetectionThinking, stripReferenceMarkers)
 	finalText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
 	if searchEnabled {
 		finalText = replaceCitationMarkersWithLinks(finalText, result.CitationLinks)
 	}
-	detected := detectAssistantToolCalls(finalText, finalThinking, finalToolDetectionThinking, toolNames)
-	if shouldWriteUpstreamEmptyOutputError(finalText) && len(detected.Calls) == 0 {
+	detected := detectAssistantToolCalls(result.Text, finalText, result.Thinking, result.ToolDetectionThinking, toolNames)
+	if shouldWriteUpstreamEmptyOutputError(finalText, finalThinking) && len(detected.Calls) == 0 {
 		status, message, code := upstreamEmptyOutputDetail(result.ContentFilter, finalText, finalThinking)
 		if historySession != nil {
 			historySession.error(status, message, code, finalThinking, finalText)
@@ -177,6 +174,9 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, co
 		return
 	}
 	respBody := openaifmt.BuildChatCompletionWithToolCalls(completionID, model, finalPrompt, finalThinking, finalText, detected.Calls, toolsRaw)
+	if refFileTokens > 0 {
+		addRefFileTokensToUsage(respBody, refFileTokens)
+	}
 	finishReason := "stop"
 	if choices, ok := respBody["choices"].([]map[string]any); ok && len(choices) > 0 {
 		if fr, _ := choices[0]["finish_reason"].(string); strings.TrimSpace(fr) != "" {
@@ -184,12 +184,12 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, resp *http.Response, co
 		}
 	}
 	if historySession != nil {
-		historySession.success(http.StatusOK, finalThinking, finalText, finishReason, openaifmt.BuildChatUsage(finalPrompt, finalThinking, finalText))
+		historySession.success(http.StatusOK, finalThinking, finalText, finishReason, openaifmt.BuildChatUsageForModel(model, finalPrompt, finalThinking, finalText, refFileTokens))
 	}
 	writeJSON(w, http.StatusOK, respBody)
 }
 
-func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *http.Response, completionID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
+func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *http.Response, completionID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, historySession *chatHistorySession) {
 	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
@@ -231,9 +231,11 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 		stripReferenceMarkers,
 		toolNames,
 		toolsRaw,
+		promptcompat.DefaultToolChoicePolicy(),
 		bufferToolContent,
 		emitEarlyToolDeltas,
 	)
+	streamRuntime.refFileTokens = refFileTokens
 
 	streamengine.ConsumeSSE(streamengine.ConsumeConfig{
 		Context:             r.Context(),
@@ -250,7 +252,7 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 		OnParsed: func(parsed sse.LineResult) streamengine.ParsedDecision {
 			decision := streamRuntime.onParsed(parsed)
 			if historySession != nil {
-				historySession.progress(streamRuntime.thinking.String(), streamRuntime.text.String())
+				historySession.progress(streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String())
 			}
 			return decision
 		},
@@ -264,14 +266,14 @@ func (h *Handler) handleStream(w http.ResponseWriter, r *http.Request, resp *htt
 				return
 			}
 			if streamRuntime.finalErrorMessage != "" {
-				historySession.error(streamRuntime.finalErrorStatus, streamRuntime.finalErrorMessage, streamRuntime.finalErrorCode, streamRuntime.thinking.String(), streamRuntime.text.String())
+				historySession.error(streamRuntime.finalErrorStatus, streamRuntime.finalErrorMessage, streamRuntime.finalErrorCode, streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String())
 				return
 			}
 			historySession.success(http.StatusOK, streamRuntime.finalThinking, streamRuntime.finalText, streamRuntime.finalFinishReason, streamRuntime.finalUsage)
 		},
 		OnContextDone: func() {
 			if historySession != nil {
-				historySession.stopped(streamRuntime.thinking.String(), streamRuntime.text.String(), string(streamengine.StopReasonContextCancelled))
+				historySession.stopped(streamRuntime.accumulator.Thinking.String(), streamRuntime.accumulator.Text.String(), string(streamengine.StopReasonContextCancelled))
 			}
 		},
 	})

internal/httpapi/openai/chat/handler_toolcall_test.go

@@ -1,6 +1,7 @@
 package chat
 
 import (
+	"context"
 	"encoding/json"
 	"io"
 	"net/http"
@@ -93,7 +94,7 @@ func TestHandleNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T) {
 	)
 	rec := httptest.NewRecorder()
 
-	h.handleNonStream(rec, resp, "cid-empty", "deepseek-v4-flash", "prompt", false, false, nil, nil, nil)
+	h.handleNonStream(rec, resp, "cid-empty", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, nil)
 	if rec.Code != http.StatusTooManyRequests {
 		t.Fatalf("expected status 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -112,7 +113,7 @@ func TestHandleNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWithoutOutp
 	)
 	rec := httptest.NewRecorder()
 
-	h.handleNonStream(rec, resp, "cid-empty-filtered", "deepseek-v4-flash", "prompt", false, false, nil, nil, nil)
+	h.handleNonStream(rec, resp, "cid-empty-filtered", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, nil)
 	if rec.Code != http.StatusBadRequest {
 		t.Fatalf("expected status 400 for filtered upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -131,14 +132,19 @@ func TestHandleNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testing.T) {
 	)
 	rec := httptest.NewRecorder()
 
-	h.handleNonStream(rec, resp, "cid-thinking-only", "deepseek-v4-pro", "prompt", true, false, nil, nil, 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())
+	h.handleNonStream(rec, resp, "cid-thinking-only", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, nil)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected status 200 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)
+	choices, _ := out["choices"].([]any)
+	if len(choices) == 0 {
+		t.Fatal("expected at least one choice")
+	}
+	first, _ := choices[0].(map[string]any)
+	msg, _ := first["message"].(map[string]any)
+	if asString(msg["reasoning_content"]) != "Only thinking" {
+		t.Fatalf("expected reasoning_content='Only thinking', got %#v", msg)
 	}
 }
 
@@ -150,7 +156,7 @@ func TestHandleNonStreamPromotesThinkingToolCallsWhenTextEmpty(t *testing.T) {
 	)
 	rec := httptest.NewRecorder()
 
-	h.handleNonStream(rec, resp, "cid-thinking-tool", "deepseek-v4-pro", "prompt", true, false, []string{"search"}, nil, nil)
+	h.handleNonStream(rec, resp, "cid-thinking-tool", "deepseek-v4-pro", "prompt", 0, true, false, []string{"search"}, nil, nil)
 	if rec.Code != http.StatusOK {
 		t.Fatalf("expected 200 for thinking tool calls, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -181,7 +187,7 @@ func TestHandleNonStreamPromotesHiddenThinkingDSMLToolCallsWhenTextEmpty(t *test
 	)
 	rec := httptest.NewRecorder()
 
-	h.handleNonStream(rec, resp, "cid-hidden-thinking-tool", "deepseek-v4-pro", "prompt", false, false, []string{"search"}, nil, nil)
+	h.handleNonStream(rec, resp, "cid-hidden-thinking-tool", "deepseek-v4-pro", "prompt", 0, false, false, []string{"search"}, nil, nil)
 	if rec.Code != http.StatusOK {
 		t.Fatalf("expected 200 for hidden thinking tool calls, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -211,7 +217,7 @@ func TestHandleStreamToolsPlainTextStreamsBeforeFinish(t *testing.T) {
 	rec := httptest.NewRecorder()
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid6", "deepseek-v4-flash", "prompt", false, false, []string{"search"}, nil, nil)
+	h.handleStream(rec, req, resp, "cid6", "deepseek-v4-flash", "prompt", 0, false, false, []string{"search"}, nil, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -239,6 +245,118 @@ func TestHandleStreamToolsPlainTextStreamsBeforeFinish(t *testing.T) {
 	}
 }
 
+func TestHandleStreamThinkingDisabledDoesNotLeakHiddenFragmentContinuations(t *testing.T) {
+	h := &Handler{}
+	resp := makeSSEHTTPResponse(
+		`data: {"p":"response/fragments","o":"APPEND","v":[{"type":"THINK","content":"我们"}]}`,
+		`data: {"p":"response/fragments/-1/content","v":"被"}`,
+		`data: {"v":"要求"}`,
+		`data: {"p":"response/fragments","o":"APPEND","v":[{"type":"RESPONSE","content":"答"}]}`,
+		`data: {"p":"response/fragments/-1/content","v":"案"}`,
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+
+	h.handleStream(rec, req, resp, "cid-hidden-fragment", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, nil)
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	content := strings.Builder{}
+	for _, frame := range frames {
+		choices, _ := frame["choices"].([]any)
+		for _, item := range choices {
+			choice, _ := item.(map[string]any)
+			delta, _ := choice["delta"].(map[string]any)
+			if c, ok := delta["content"].(string); ok {
+				content.WriteString(c)
+			}
+		}
+	}
+	if got := content.String(); got != "答案" {
+		t.Fatalf("expected only visible response text, got %q body=%s", got, rec.Body.String())
+	}
+}
+
+func TestHandleStreamEmitsSingleChoiceFramesForMultipleParsedParts(t *testing.T) {
+	h := &Handler{}
+	resp := makeSSEHTTPResponse(
+		`data: {"p":"response/fragments","o":"APPEND","v":[{"type":"THINK","content":"我们"},{"type":"THINK","content":"被"},{"type":"THINK","content":"要求"},{"type":"RESPONSE","content":"答"},{"type":"RESPONSE","content":"案"}]}`,
+		`data: [DONE]`,
+	)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+
+	h.handleStream(rec, req, resp, "cid-multi-parts", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, nil)
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	var reasoning, content strings.Builder
+	for _, frame := range frames {
+		choices, _ := frame["choices"].([]any)
+		if len(choices) != 1 {
+			t.Fatalf("expected exactly one choice per stream frame, got %d frame=%#v body=%s", len(choices), frame, rec.Body.String())
+		}
+		choice, _ := choices[0].(map[string]any)
+		delta, _ := choice["delta"].(map[string]any)
+		reasoning.WriteString(asString(delta["reasoning_content"]))
+		content.WriteString(asString(delta["content"]))
+	}
+	if got := reasoning.String(); got != "我们被要求" {
+		t.Fatalf("first-choice-only client would miss reasoning tokens: got %q body=%s", got, rec.Body.String())
+	}
+	if got := content.String(); got != "答案" {
+		t.Fatalf("first-choice-only client would miss content tokens: got %q body=%s", got, rec.Body.String())
+	}
+}
+
+func TestHandleStreamCoalescesSmallContentDeltas(t *testing.T) {
+	h := &Handler{}
+	lines := make([]string, 0, 101)
+	for i := 0; i < 100; i++ {
+		b, _ := json.Marshal(map[string]any{
+			"p": "response/content",
+			"v": "字",
+		})
+		lines = append(lines, "data: "+string(b))
+	}
+	lines = append(lines, "data: [DONE]")
+	resp := makeSSEHTTPResponse(lines...)
+	rec := httptest.NewRecorder()
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+
+	h.handleStream(rec, req, resp, "cid-coalesce", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, nil)
+
+	frames, done := parseSSEDataFrames(t, rec.Body.String())
+	if !done {
+		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
+	}
+	var content strings.Builder
+	contentDeltaFrames := 0
+	for _, frame := range frames {
+		choices, _ := frame["choices"].([]any)
+		if len(choices) != 1 {
+			t.Fatalf("expected exactly one choice per stream frame, got %d frame=%#v body=%s", len(choices), frame, rec.Body.String())
+		}
+		choice, _ := choices[0].(map[string]any)
+		delta, _ := choice["delta"].(map[string]any)
+		if c, ok := delta["content"].(string); ok {
+			contentDeltaFrames++
+			content.WriteString(c)
+		}
+	}
+	if got, want := content.String(), strings.Repeat("字", 100); got != want {
+		t.Fatalf("coalesced stream content mismatch: got %q want %q body=%s", got, want, rec.Body.String())
+	}
+	if contentDeltaFrames >= 100 {
+		t.Fatalf("expected coalescing to reduce 100 tiny content frames, got %d body=%s", contentDeltaFrames, rec.Body.String())
+	}
+}
+
 func TestHandleStreamIncompleteCapturedToolJSONFlushesAsTextOnFinalize(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
@@ -248,7 +366,7 @@ func TestHandleStreamIncompleteCapturedToolJSONFlushesAsTextOnFinalize(t *testin
 	rec := httptest.NewRecorder()
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid10", "deepseek-v4-flash", "prompt", false, false, []string{"search"}, nil, nil)
+	h.handleStream(rec, req, resp, "cid10", "deepseek-v4-flash", "prompt", 0, false, false, []string{"search"}, nil, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -282,7 +400,7 @@ func TestHandleStreamPromotesThinkingToolCallsOnFinalizeWithoutMidstreamIntercep
 	rec := httptest.NewRecorder()
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid-thinking-stream", "deepseek-v4-pro", "prompt", true, false, []string{"search"}, nil, nil)
+	h.handleStream(rec, req, resp, "cid-thinking-stream", "deepseek-v4-pro", "prompt", 0, true, false, []string{"search"}, nil, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -291,20 +409,16 @@ func TestHandleStreamPromotesThinkingToolCallsOnFinalizeWithoutMidstreamIntercep
 	if !streamHasToolCallsDelta(frames) {
 		t.Fatalf("expected tool_calls delta from finalize fallback, body=%s", rec.Body.String())
 	}
-	reasoningSeen := false
 	for _, frame := range frames {
 		choices, _ := frame["choices"].([]any)
 		for _, item := range choices {
 			choice, _ := item.(map[string]any)
 			delta, _ := choice["delta"].(map[string]any)
 			if asString(delta["reasoning_content"]) != "" {
-				reasoningSeen = true
+				t.Fatalf("did not expect leaked reasoning_content markup, body=%s", rec.Body.String())
 			}
 		}
 	}
-	if !reasoningSeen {
-		t.Fatalf("expected reasoning_content to stream before finalize fallback, body=%s", rec.Body.String())
-	}
 	if streamFinishReason(frames) != "tool_calls" {
 		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
 	}
@@ -319,7 +433,7 @@ func TestHandleStreamPromotesHiddenThinkingDSMLToolCallsOnFinalize(t *testing.T)
 	rec := httptest.NewRecorder()
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid-hidden-thinking-stream", "deepseek-v4-pro", "prompt", false, false, []string{"search"}, nil, nil)
+	h.handleStream(rec, req, resp, "cid-hidden-thinking-stream", "deepseek-v4-pro", "prompt", 0, false, false, []string{"search"}, nil, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -353,7 +467,7 @@ func TestHandleStreamEmitsDistinctToolCallIDsAcrossSeparateToolBlocks(t *testing
 	rec := httptest.NewRecorder()
 	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
 
-	h.handleStream(rec, req, resp, "cid-multi", "deepseek-v4-flash", "prompt", false, false, []string{"read_file", "search"}, nil, nil)
+	h.handleStream(rec, req, resp, "cid-multi", "deepseek-v4-flash", "prompt", 0, false, false, []string{"read_file", "search"}, nil, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -419,7 +533,7 @@ func TestHandleStreamCoercesSchemaDeclaredStringArgumentsOnFinalize(t *testing.T
 		},
 	}
 
-	h.handleStream(rec, req, resp, "cid-string-protect", "deepseek-v4-flash", "prompt", false, false, []string{"Write"}, toolsRaw, nil)
+	h.handleStream(rec, req, resp, "cid-string-protect", "deepseek-v4-flash", "prompt", 0, false, false, []string{"Write"}, toolsRaw, nil)
 
 	frames, done := parseSSEDataFrames(t, rec.Body.String())
 	if !done {
@@ -451,3 +565,45 @@ func TestHandleStreamCoercesSchemaDeclaredStringArgumentsOnFinalize(t *testing.T
 	}
 	t.Fatalf("expected at least one streamed tool call delta, body=%s", rec.Body.String())
 }
+
+func TestHandleNonStreamWithRetryIncludesRefFileTokensInUsage(t *testing.T) {
+	h := &Handler{}
+
+	run := func(refFileTokens int) map[string]any {
+		resp := makeSSEHTTPResponse(
+			`data: {"p":"response/content","v":"hello world"}`,
+			`data: [DONE]`,
+		)
+		rec := httptest.NewRecorder()
+		h.handleNonStreamWithRetry(rec, context.Background(), nil, resp, nil, "", "cid-ref", "deepseek-v4-flash", "prompt", refFileTokens, false, false, nil, nil, nil)
+		if rec.Code != http.StatusOK {
+			t.Fatalf("expected 200, got %d body=%s", rec.Code, rec.Body.String())
+		}
+		return decodeJSONBody(t, rec.Body.String())
+	}
+
+	base := run(0)
+	withRef := run(7)
+
+	baseUsage, _ := base["usage"].(map[string]any)
+	refUsage, _ := withRef["usage"].(map[string]any)
+	if baseUsage == nil || refUsage == nil {
+		t.Fatalf("expected usage objects, base=%#v ref=%#v", base["usage"], withRef["usage"])
+	}
+
+	getInt := func(m map[string]any, key string) int {
+		t.Helper()
+		v, ok := m[key].(float64)
+		if !ok {
+			t.Fatalf("expected numeric %s, got %#v", key, m[key])
+		}
+		return int(v)
+	}
+
+	if got := getInt(refUsage, "prompt_tokens") - getInt(baseUsage, "prompt_tokens"); got != 7 {
+		t.Fatalf("expected prompt_tokens delta 7, got %d", got)
+	}
+	if got := getInt(refUsage, "total_tokens") - getInt(baseUsage, "total_tokens"); got != 7 {
+		t.Fatalf("expected total_tokens delta 7, got %d", got)
+	}
+}

internal/httpapi/openai/chat/ref_file_tokens.go

@@ -0,0 +1,26 @@
+package chat
+
+// addRefFileTokensToUsage adds inline-uploaded file token estimates to an existing
+// usage map inside a response object. This keeps the token accounting aware of file
+// content that the upstream model processes but that is not part of the prompt text.
+func addRefFileTokensToUsage(obj map[string]any, refFileTokens int) {
+	if refFileTokens <= 0 || obj == nil {
+		return
+	}
+	usage, ok := obj["usage"].(map[string]any)
+	if !ok || usage == nil {
+		return
+	}
+	for _, key := range []string{"input_tokens", "prompt_tokens"} {
+		if v, ok := usage[key]; ok {
+			if n, ok := v.(int); ok {
+				usage[key] = n + refFileTokens
+			}
+		}
+	}
+	if v, ok := usage["total_tokens"]; ok {
+		if n, ok := v.(int); ok {
+			usage["total_tokens"] = n + refFileTokens
+		}
+	}
+}

internal/httpapi/openai/chat/vercel_prepare_test.go

@@ -130,8 +130,8 @@ func TestHandleVercelStreamPrepareAppliesCurrentInputFile(t *testing.T) {
 		t.Fatalf("expected payload object, got %#v", body["payload"])
 	}
 	promptText, _ := payload["prompt"].(string)
-	if !strings.Contains(promptText, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral prompt, got %s", promptText)
+	if !strings.Contains(promptText, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation prompt, got %s", promptText)
 	}
 	if strings.Contains(promptText, "first user turn") || strings.Contains(promptText, "latest user turn") {
 		t.Fatalf("expected original turns hidden from prompt, got %s", promptText)

internal/httpapi/openai/citation_links_test.go

@@ -54,3 +54,31 @@ func TestReplaceCitationMarkersWithLinksSupportsReferenceZeroBased(t *testing.T)
 		t.Fatalf("expected %q, got %q", want, got)
 	}
 }
+
+func TestReplaceCitationMarkersWithLinksKeepsCitationOneBasedWithZeroBasedReference(t *testing.T) {
+	raw := "引用[citation:1]，来源[reference:0]，后续[reference:1]。"
+	links := map[int]string{
+		1: "https://example.com/first",
+		2: "https://example.com/second",
+	}
+
+	got := replaceCitationMarkersWithLinks(raw, links)
+	want := "引用[1](https://example.com/first)，来源[0](https://example.com/first)，后续[1](https://example.com/second)。"
+	if got != want {
+		t.Fatalf("expected %q, got %q", want, got)
+	}
+}
+
+func TestReplaceCitationMarkersWithLinksDetectsSpacedReferenceZeroBased(t *testing.T) {
+	raw := "来源[reference: 0] 与 [reference: 1]。"
+	links := map[int]string{
+		1: "https://example.com/first",
+		2: "https://example.com/second",
+	}
+
+	got := replaceCitationMarkersWithLinks(raw, links)
+	want := "来源[0](https://example.com/first) 与 [1](https://example.com/second)。"
+	if got != want {
+		t.Fatalf("expected %q, got %q", want, got)
+	}
+}

internal/httpapi/openai/file_inline_upload_test.go

@@ -94,6 +94,9 @@ func TestPreprocessInlineFileInputsReplacesDataURLAndCollectsRefFileIDs(t *testi
 	if len(ds.uploadCalls) != 1 {
 		t.Fatalf("expected 1 upload, got %d", len(ds.uploadCalls))
 	}
+	if ds.uploadCalls[0].ModelType != "default" {
+		t.Fatalf("expected default model type when request omits model, got %q", ds.uploadCalls[0].ModelType)
+	}
 	if ds.lastCtx != ctx {
 		t.Fatalf("expected upload to use request context")
 	}
@@ -149,7 +152,7 @@ func TestPreprocessInlineFileInputsDeduplicatesIdenticalPayloads(t *testing.T) {
 func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
-	reqBody := `{"model":"deepseek-v4-flash","messages":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"image_url","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	reqBody := `{"model":"deepseek-v4-vision","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")
@@ -163,6 +166,9 @@ func TestChatCompletionsUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	if len(ds.uploadCalls) != 1 {
 		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
 	}
+	if ds.uploadCalls[0].ModelType != "vision" {
+		t.Fatalf("expected vision model type for vision request, got %q", ds.uploadCalls[0].ModelType)
+	}
 	if ds.completionReq == nil {
 		t.Fatal("expected completion payload to be captured")
 	}
@@ -177,7 +183,7 @@ func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
-	reqBody := `{"model":"deepseek-v4-flash","input":[{"role":"user","content":[{"type":"input_text","text":"hi"},{"type":"input_image","image_url":{"url":"data:image/png;base64,QUJDRA=="}}]}],"stream":false}`
+	reqBody := `{"model":"deepseek-v4-pro","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")
@@ -191,6 +197,9 @@ func TestResponsesUploadsInlineFilesBeforeCompletion(t *testing.T) {
 	if len(ds.uploadCalls) != 1 {
 		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
 	}
+	if ds.uploadCalls[0].ModelType != "expert" {
+		t.Fatalf("expected expert model type for pro request, got %q", ds.uploadCalls[0].ModelType)
+	}
 	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"])
@@ -216,6 +225,45 @@ func TestChatCompletionsInlineUploadFailureReturnsBadRequest(t *testing.T) {
 	}
 }
 
+func TestChatCompletionsInlineUploadLimitReturnsBadRequest(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	content := []any{map[string]any{"type": "input_text", "text": "hi"}}
+	for i := 0; i < 51; i++ {
+		content = append(content, map[string]any{
+			"type":      "image_url",
+			"image_url": map[string]any{"url": "data:image/png;base64,QUJDRA=="},
+		})
+	}
+	body, err := json.Marshal(map[string]any{
+		"model": "deepseek-v4-flash",
+		"messages": []any{map[string]any{
+			"role":    "user",
+			"content": content,
+		}},
+		"stream": false,
+	})
+	if err != nil {
+		t.Fatalf("marshal request: %v", err)
+	}
+	req := httptest.NewRequest(http.MethodPost, "/v1/chat/completions", strings.NewReader(string(body)))
+	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 !strings.Contains(rec.Body.String(), "exceeded maximum of 50 inline files per request") {
+		t.Fatalf("expected inline file limit error, got body=%s", rec.Body.String())
+	}
+	if ds.completionReq != nil {
+		t.Fatalf("did not expect completion call after inline file limit error")
+	}
+}
+
 func TestResponsesInlineUploadFailureReturnsInternalServerError(t *testing.T) {
 	ds := &inlineUploadDSStub{uploadErr: errors.New("boom")}
 	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}

internal/httpapi/openai/files/file_inline_upload.go

@@ -12,6 +12,7 @@ import (
 	"strings"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/config"
 	dsclient "ds2api/internal/deepseek/client"
 	"ds2api/internal/httpapi/openai/shared"
 	"ds2api/internal/promptcompat"
@@ -39,11 +40,13 @@ func (e *inlineFileUploadError) Error() string {
 }
 
 type inlineUploadState struct {
-	ctx          context.Context
-	handler      *Handler
-	auth         *auth.RequestAuth
-	uploadedByID map[string]string
-	uploadCount  int
+	ctx             context.Context
+	handler         *Handler
+	auth            *auth.RequestAuth
+	modelType       string
+	uploadedByID    map[string]string
+	uploadCount     int
+	inlineFileBytes int
 }
 
 type inlineDecodedFile struct {
@@ -57,10 +60,19 @@ func (h *Handler) PreprocessInlineFileInputs(ctx context.Context, a *auth.Reques
 	if h == nil || h.DS == nil || len(req) == 0 {
 		return nil
 	}
+	modelType := "default"
+	if requestedModel, ok := req["model"].(string); ok {
+		if resolvedModel, ok := config.ResolveModel(h.Store, requestedModel); ok {
+			if resolvedType, ok := config.GetModelType(resolvedModel); ok {
+				modelType = resolvedType
+			}
+		}
+	}
 	state := &inlineUploadState{
 		ctx:          ctx,
 		handler:      h,
 		auth:         a,
+		modelType:    modelType,
 		uploadedByID: map[string]string{},
 	}
 	for _, key := range []string{"messages", "input", "attachments"} {
@@ -75,6 +87,9 @@ func (h *Handler) PreprocessInlineFileInputs(ctx context.Context, a *auth.Reques
 	if refIDs := promptcompat.CollectOpenAIRefFileIDs(req); len(refIDs) > 0 {
 		req["ref_file_ids"] = stringsToAnySlice(refIDs)
 	}
+	if state.inlineFileBytes > 0 {
+		req["_inline_file_bytes"] = state.inlineFileBytes
+	}
 	return nil
 }
 
@@ -135,13 +150,15 @@ func (s *inlineUploadState) tryUploadBlock(block map[string]any) (map[string]any
 		return nil, false, nil
 	}
 	if s.uploadCount >= maxInlineFilesPerRequest {
-		return nil, true, fmt.Errorf("exceeded maximum of %d inline files per request", maxInlineFilesPerRequest)
+		err := fmt.Errorf("exceeded maximum of %d inline files per request", maxInlineFilesPerRequest)
+		return nil, true, &inlineFileUploadError{status: http.StatusBadRequest, message: err.Error(), err: err}
 	}
 	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++
+	s.inlineFileBytes += len(decoded.Data)
 	replacement := map[string]any{
 		"type":    decoded.ReplacementType,
 		"file_id": fileID,
@@ -168,6 +185,7 @@ func (s *inlineUploadState) uploadInlineFile(file inlineDecodedFile) (string, er
 	result, err := s.handler.DS.UploadFile(s.ctx, s.auth, dsclient.UploadFileRequest{
 		Filename:    file.Filename,
 		ContentType: contentType,
+		ModelType:   s.modelType,
 		Data:        file.Data,
 	}, 3)
 	if err != nil {

internal/httpapi/openai/files/handler_files.go

@@ -1,13 +1,18 @@
 package files
 
 import (
+	"context"
+	"errors"
 	"io"
 	"net/http"
 	"strings"
 	"time"
 
+	"github.com/go-chi/chi/v5"
+
 	"ds2api/internal/auth"
 	"ds2api/internal/chathistory"
+	"ds2api/internal/config"
 	dsclient "ds2api/internal/deepseek/client"
 	"ds2api/internal/httpapi/openai/shared"
 )
@@ -21,6 +26,10 @@ type Handler struct {
 	ChatHistory *chathistory.Store
 }
 
+type fileFetcher interface {
+	FetchUploadedFile(ctx context.Context, a *auth.RequestAuth, fileID string) (*dsclient.UploadFileResult, error)
+}
+
 func (h *Handler) UploadFile(w http.ResponseWriter, r *http.Request) {
 	a, err := h.Auth.Determine(r)
 	if err != nil {
@@ -66,10 +75,12 @@ func (h *Handler) UploadFile(w http.ResponseWriter, r *http.Request) {
 	if contentType == "" && len(data) > 0 {
 		contentType = http.DetectContentType(data)
 	}
+	modelType := resolveUploadModelType(h.Store, r)
 	result, err := h.DS.UploadFile(r.Context(), a, dsclient.UploadFileRequest{
 		Filename:    header.Filename,
 		ContentType: contentType,
 		Purpose:     strings.TrimSpace(r.FormValue("purpose")),
+		ModelType:   modelType,
 		Data:        data,
 	}, 3)
 	if err != nil {
@@ -82,6 +93,70 @@ func (h *Handler) UploadFile(w http.ResponseWriter, r *http.Request) {
 	shared.WriteJSON(w, http.StatusOK, buildOpenAIFileObject(result))
 }
 
+func (h *Handler) RetrieveFile(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
+		}
+		shared.WriteOpenAIError(w, status, detail)
+		return
+	}
+	defer h.Auth.Release(a)
+
+	fileID := strings.TrimSpace(chi.URLParam(r, "file_id"))
+	if fileID == "" {
+		shared.WriteOpenAIError(w, http.StatusBadRequest, "file_id is required")
+		return
+	}
+	fetcher, ok := h.DS.(fileFetcher)
+	if !ok {
+		shared.WriteOpenAIError(w, http.StatusNotImplemented, "file retrieval is not available")
+		return
+	}
+	result, err := fetcher.FetchUploadedFile(r.Context(), a, fileID)
+	if err != nil {
+		if errors.Is(err, dsclient.ErrUploadFileNotFound) {
+			shared.WriteOpenAIError(w, http.StatusNotFound, "file not found")
+			return
+		}
+		shared.WriteOpenAIError(w, http.StatusInternalServerError, "Failed to retrieve file.")
+		return
+	}
+	if result != nil && result.AccountID == "" {
+		result.AccountID = a.AccountID
+	}
+	shared.WriteJSON(w, http.StatusOK, buildOpenAIFileObject(result))
+}
+
+func resolveUploadModelType(store shared.ConfigReader, r *http.Request) string {
+	for _, candidate := range []string{r.FormValue("model_type"), r.Header.Get("X-Model-Type")} {
+		if modelType := normalizeUploadModelType(candidate); modelType != "" {
+			return modelType
+		}
+	}
+	requestedModel := strings.TrimSpace(r.FormValue("model"))
+	if requestedModel != "" {
+		if resolvedModel, ok := config.ResolveModel(store, requestedModel); ok {
+			if modelType, ok := config.GetModelType(resolvedModel); ok {
+				return modelType
+			}
+		}
+	}
+	return "default"
+}
+
+func normalizeUploadModelType(raw string) string {
+	switch strings.ToLower(strings.TrimSpace(raw)) {
+	case "default", "expert", "vision":
+		return strings.ToLower(strings.TrimSpace(raw))
+	default:
+		return ""
+	}
+}
+
 func buildOpenAIFileObject(result *dsclient.UploadFileResult) map[string]any {
 	if result == nil {
 		obj := map[string]any{

internal/httpapi/openai/files_route_test.go

@@ -43,6 +43,7 @@ func (managedFilesAuthStub) Release(_ *auth.RequestAuth) {}
 type filesRouteDSStub struct {
 	lastReq dsclient.UploadFileRequest
 	upload  *dsclient.UploadFileResult
+	fetched *dsclient.UploadFileResult
 	err     error
 }
 
@@ -65,6 +66,16 @@ func (m *filesRouteDSStub) UploadFile(_ context.Context, _ *auth.RequestAuth, re
 	return &dsclient.UploadFileResult{ID: "file-123", Filename: req.Filename, Bytes: int64(len(req.Data)), Purpose: req.Purpose, Status: "uploaded"}, nil
 }
 
+func (m *filesRouteDSStub) FetchUploadedFile(_ context.Context, _ *auth.RequestAuth, fileID string) (*dsclient.UploadFileResult, error) {
+	if m.err != nil {
+		return nil, m.err
+	}
+	if m.fetched != nil {
+		return m.fetched, nil
+	}
+	return &dsclient.UploadFileResult{ID: fileID, Filename: "notes.txt", Bytes: 11, Purpose: "assistants", Status: "processed"}, nil
+}
+
 func (m *filesRouteDSStub) CallCompletion(_ context.Context, _ *auth.RequestAuth, _ map[string]any, _ string, _ int) (*http.Response, error) {
 	return nil, errors.New("not implemented")
 }
@@ -77,7 +88,7 @@ func (m *filesRouteDSStub) DeleteAllSessionsForToken(_ context.Context, _ string
 	return nil
 }
 
-func newMultipartUploadRequest(t *testing.T, purpose string, filename string, data []byte) *http.Request {
+func newMultipartUploadRequest(t *testing.T, purpose string, filename string, data []byte, model string) *http.Request {
 	t.Helper()
 	var body bytes.Buffer
 	writer := multipart.NewWriter(&body)
@@ -86,6 +97,11 @@ func newMultipartUploadRequest(t *testing.T, purpose string, filename string, da
 			t.Fatalf("write purpose failed: %v", err)
 		}
 	}
+	if model != "" {
+		if err := writer.WriteField("model", model); err != nil {
+			t.Fatalf("write model failed: %v", err)
+		}
+	}
 	part, err := writer.CreateFormFile("file", filename)
 	if err != nil {
 		t.Fatalf("create form file failed: %v", err)
@@ -108,7 +124,7 @@ func TestFilesRouteUploadSuccess(t *testing.T) {
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
 
-	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"), "deepseek-v4-vision")
 	rec := httptest.NewRecorder()
 	r.ServeHTTP(rec, req)
 
@@ -121,6 +137,9 @@ func TestFilesRouteUploadSuccess(t *testing.T) {
 	if ds.lastReq.Purpose != "assistants" {
 		t.Fatalf("expected purpose assistants, got %q", ds.lastReq.Purpose)
 	}
+	if ds.lastReq.ModelType != "vision" {
+		t.Fatalf("expected vision model type, got %q", ds.lastReq.ModelType)
+	}
 	if string(ds.lastReq.Data) != "hello world" {
 		t.Fatalf("unexpected uploaded data: %q", string(ds.lastReq.Data))
 	}
@@ -145,7 +164,7 @@ func TestFilesRouteUploadIncludesAccountIDForManagedAccount(t *testing.T) {
 	r := chi.NewRouter()
 	registerOpenAITestRoutes(r, h)
 
-	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"))
+	req := newMultipartUploadRequest(t, "assistants", "notes.txt", []byte("hello world"), "deepseek-v4-vision")
 	rec := httptest.NewRecorder()
 	r.ServeHTTP(rec, req)
 
@@ -161,6 +180,54 @@ func TestFilesRouteUploadIncludesAccountIDForManagedAccount(t *testing.T) {
 	}
 }
 
+func TestFilesRouteRetrieveSuccess(t *testing.T) {
+	ds := &filesRouteDSStub{fetched: &dsclient.UploadFileResult{
+		ID:       "file-123",
+		Filename: "notes.txt",
+		Bytes:    11,
+		Purpose:  "assistants",
+		Status:   "processed",
+	}}
+	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: managedFilesAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	registerOpenAITestRoutes(r, h)
+
+	req := httptest.NewRequest(http.MethodGet, "/v1/files/file-123", nil)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	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["id"] != "file-123" || out["filename"] != "notes.txt" || out["status"] != "processed" {
+		t.Fatalf("unexpected file object: %#v", out)
+	}
+	if out["account_id"] != "acct-123" {
+		t.Fatalf("expected account_id acct-123, got %#v", out["account_id"])
+	}
+}
+
+func TestFilesRouteRetrieveNotFound(t *testing.T) {
+	ds := &filesRouteDSStub{err: dsclient.ErrUploadFileNotFound}
+	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: ds}
+	r := chi.NewRouter()
+	registerOpenAITestRoutes(r, h)
+
+	req := httptest.NewRequest(http.MethodGet, "/v1/files/missing-file", nil)
+	req.Header.Set("Authorization", "Bearer direct-token")
+	rec := httptest.NewRecorder()
+	r.ServeHTTP(rec, req)
+
+	if rec.Code != http.StatusNotFound {
+		t.Fatalf("expected 404, got %d body=%s", rec.Code, rec.Body.String())
+	}
+}
+
 func TestFilesRouteRejectsNonMultipart(t *testing.T) {
 	h := &openAITestSurface{Store: mockOpenAIConfig{wideInput: true}, Auth: streamStatusAuthStub{}, DS: &filesRouteDSStub{}}
 	r := chi.NewRouter()

internal/httpapi/openai/history/current_input_file.go

@@ -7,13 +7,14 @@ import (
 	"strings"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/config"
 	dsclient "ds2api/internal/deepseek/client"
 	"ds2api/internal/httpapi/openai/shared"
 	"ds2api/internal/promptcompat"
 )
 
 const (
-	currentInputFilename    = "IGNORE.txt"
+	currentInputFilename    = promptcompat.CurrentInputContextFilename
 	currentInputContentType = "text/plain; charset=utf-8"
 	currentInputPurpose     = "assistants"
 )
@@ -35,11 +36,15 @@ func (s Service) ApplyCurrentInputFile(ctx context.Context, a *auth.RequestAuth,
 	if strings.TrimSpace(fileText) == "" {
 		return stdReq, errors.New("current user input file produced empty transcript")
 	}
-
+	modelType := "default"
+	if resolvedType, ok := config.GetModelType(stdReq.ResolvedModel); ok {
+		modelType = resolvedType
+	}
 	result, err := s.DS.UploadFile(ctx, a, dsclient.UploadFileRequest{
 		Filename:    currentInputFilename,
 		ContentType: currentInputContentType,
 		Purpose:     currentInputPurpose,
+		ModelType:   modelType,
 		Data:        []byte(fileText),
 	}, 3)
 	if err != nil {
@@ -62,6 +67,9 @@ func (s Service) ApplyCurrentInputFile(ctx context.Context, a *auth.RequestAuth,
 	stdReq.CurrentInputFileApplied = true
 	stdReq.RefFileIDs = prependUniqueRefFileID(stdReq.RefFileIDs, fileID)
 	stdReq.FinalPrompt, stdReq.ToolNames = promptcompat.BuildOpenAIPrompt(messages, stdReq.ToolsRaw, "", stdReq.ToolChoice, stdReq.Thinking)
+	// Token accounting must reflect the actual downstream context:
+	// the uploaded DS2API_HISTORY.txt file content + the continuation live prompt.
+	stdReq.PromptTokenText = fileText + "\n" + stdReq.FinalPrompt
 	return stdReq, nil
 }
 
@@ -85,5 +93,5 @@ func latestUserInputForFile(messages []any) (int, string) {
 }
 
 func currentInputFilePrompt() string {
-	return "The current request and prior conversation context have already been provided. Answer the latest user request directly."
+	return "Continue from the latest state in the attached DS2API_HISTORY.txt context. Treat it as the current working state and answer the latest user request directly."
 }

internal/httpapi/openai/history_split_test.go

@@ -14,6 +14,7 @@ import (
 	"ds2api/internal/auth"
 	dsclient "ds2api/internal/deepseek/client"
 	"ds2api/internal/promptcompat"
+	"ds2api/internal/util"
 )
 
 func historySplitTestMessages() []any {
@@ -60,27 +61,32 @@ func (streamStatusManagedAuthStub) DetermineCaller(_ *http.Request) (*auth.Reque
 
 func (streamStatusManagedAuthStub) Release(_ *auth.RequestAuth) {}
 
-func TestBuildOpenAICurrentInputContextTranscriptUsesInjectedFileWrapper(t *testing.T) {
+func TestBuildOpenAICurrentInputContextTranscriptUsesNumberedHistorySections(t *testing.T) {
 	_, historyMessages := splitOpenAIHistoryMessages(historySplitTestMessages(), 1)
 	transcript := buildOpenAICurrentInputContextTranscript(historyMessages)
 
-	if !strings.HasPrefix(transcript, "[file content end]\n\n") {
-		t.Fatalf("expected injected file wrapper prefix, got %q", transcript)
+	if strings.Contains(transcript, "[file content end]") || strings.Contains(transcript, "[file content begin]") || strings.Contains(transcript, "[file name]:") {
+		t.Fatalf("expected transcript without file wrapper tags, got %q", transcript)
 	}
-	if !strings.Contains(transcript, "<｜begin▁of▁sentence｜>") {
-		t.Fatalf("expected serialized conversation markers, got %q", transcript)
+	if !strings.Contains(transcript, "# DS2API_HISTORY.txt") {
+		t.Fatalf("expected history transcript header, got %q", transcript)
 	}
-	if !strings.Contains(transcript, "first user turn") || !strings.Contains(transcript, "tool result") {
-		t.Fatalf("expected historical turns preserved, got %q", transcript)
+	if !strings.Contains(transcript, "Prior conversation history and tool progress.") {
+		t.Fatalf("expected history transcript description, got %q", transcript)
 	}
-	if !strings.Contains(transcript, "[reasoning_content]") || !strings.Contains(transcript, "hidden reasoning") {
-		t.Fatalf("expected reasoning block preserved, got %q", transcript)
-	}
-	if !strings.Contains(transcript, "<|DSML|tool_calls>") {
-		t.Fatalf("expected tool calls preserved, got %q", transcript)
-	}
-	if !strings.HasSuffix(transcript, "\n[file name]: IGNORE\n[file content begin]\n") {
-		t.Fatalf("expected injected file wrapper suffix, got %q", transcript)
+	for _, want := range []string{
+		"=== 1. USER ===",
+		"=== 2. ASSISTANT ===",
+		"=== 3. TOOL ===",
+		"first user turn",
+		"tool result",
+		"[reasoning_content]",
+		"hidden reasoning",
+		"<|DSML|tool_calls>",
+	} {
+		if !strings.Contains(transcript, want) {
+			t.Fatalf("expected transcript to contain %q, got %q", want, transcript)
+		}
 	}
 }
 
@@ -221,7 +227,7 @@ func TestApplyCurrentInputFileDisabledPassThrough(t *testing.T) {
 		DS: ds,
 	}
 	req := map[string]any{
-		"model":    "deepseek-v4-flash",
+		"model":    "deepseek-v4-vision",
 		"messages": historySplitTestMessages(),
 	}
 	stdReq, err := promptcompat.NormalizeOpenAIChatRequest(h.Store, req, "")
@@ -244,7 +250,7 @@ func TestApplyCurrentInputFileDisabledPassThrough(t *testing.T) {
 	}
 }
 
-func TestApplyCurrentInputFileUploadsFirstTurnWithInjectedWrapper(t *testing.T) {
+func TestApplyCurrentInputFileUploadsFirstTurnWithNumberedHistoryTranscript(t *testing.T) {
 	ds := &inlineUploadDSStub{}
 	h := &openAITestSurface{
 		Store: mockOpenAIConfig{
@@ -274,34 +280,90 @@ func TestApplyCurrentInputFileUploadsFirstTurnWithInjectedWrapper(t *testing.T)
 		t.Fatalf("expected 1 current input upload, got %d", len(ds.uploadCalls))
 	}
 	upload := ds.uploadCalls[0]
-	if upload.Filename != "IGNORE.txt" {
+	if upload.Filename != "DS2API_HISTORY.txt" {
 		t.Fatalf("unexpected upload filename: %q", upload.Filename)
 	}
 	uploadedText := string(upload.Data)
-	if !strings.HasPrefix(uploadedText, "[file content end]\n\n") {
-		t.Fatalf("expected injected file wrapper prefix, got %q", uploadedText)
+	if strings.Contains(uploadedText, "[file content end]") || strings.Contains(uploadedText, "[file content begin]") || strings.Contains(uploadedText, "[file name]:") {
+		t.Fatalf("expected uploaded transcript without file wrapper tags, got %q", uploadedText)
 	}
-	if !strings.Contains(uploadedText, "<｜begin▁of▁sentence｜><｜User｜>first turn content that is long enough") {
-		t.Fatalf("expected serialized current user turn markers, got %q", uploadedText)
+	for _, want := range []string{
+		"# DS2API_HISTORY.txt",
+		"=== 1. USER ===",
+		"first turn content that is long enough",
+	} {
+		if !strings.Contains(uploadedText, want) {
+			t.Fatalf("expected uploaded transcript to contain %q, got %q", want, uploadedText)
+		}
 	}
 	if !strings.Contains(uploadedText, promptcompat.ThinkingInjectionMarker) {
 		t.Fatalf("expected thinking injection in current input file, got %q", uploadedText)
 	}
-	if !strings.HasSuffix(uploadedText, "\n[file name]: IGNORE\n[file content begin]\n") {
-		t.Fatalf("expected injected file wrapper suffix, got %q", uploadedText)
-	}
+
 	if strings.Contains(out.FinalPrompt, "first turn content that is long enough") {
 		t.Fatalf("expected current input text to be replaced in live prompt, got %s", out.FinalPrompt)
 	}
-	if strings.Contains(out.FinalPrompt, "CURRENT_USER_INPUT.txt") || strings.Contains(out.FinalPrompt, "IGNORE.txt") || strings.Contains(out.FinalPrompt, "Read that file") {
+	if strings.Contains(out.FinalPrompt, "CURRENT_USER_INPUT.txt") || strings.Contains(out.FinalPrompt, "Read that file") {
 		t.Fatalf("expected live prompt not to instruct file reads, got %s", out.FinalPrompt)
 	}
-	if !strings.Contains(out.FinalPrompt, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral continuation instruction in live prompt, got %s", out.FinalPrompt)
+	if !strings.Contains(out.FinalPrompt, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation-oriented prompt in live prompt, got %s", out.FinalPrompt)
 	}
 	if len(out.RefFileIDs) != 1 || out.RefFileIDs[0] != "file-inline-1" {
 		t.Fatalf("expected current input file id in ref_file_ids, got %#v", out.RefFileIDs)
 	}
+	if !strings.Contains(out.PromptTokenText, "first turn content that is long enough") {
+		t.Fatalf("expected prompt token text to preserve original full context, got %q", out.PromptTokenText)
+	}
+	if !strings.Contains(out.PromptTokenText, "# DS2API_HISTORY.txt") || !strings.Contains(out.PromptTokenText, "=== 1. USER ===") {
+		t.Fatalf("expected prompt token text to include numbered history transcript, got %q", out.PromptTokenText)
+	}
+}
+
+func TestApplyCurrentInputFilePreservesFullContextPromptForTokenCounting(t *testing.T) {
+	ds := &inlineUploadDSStub{}
+	h := &openAITestSurface{
+		Store: mockOpenAIConfig{
+			wideInput:           true,
+			currentInputEnabled: true,
+			currentInputMin:     0,
+			thinkingInjection:   boolPtr(true),
+		},
+		DS: ds,
+	}
+	req := map[string]any{
+		"model":    "deepseek-v4-vision",
+		"messages": historySplitTestMessages(),
+	}
+	stdReq, err := promptcompat.NormalizeOpenAIChatRequest(h.Store, req, "")
+	if err != nil {
+		t.Fatalf("normalize failed: %v", err)
+	}
+
+	out, err := h.applyCurrentInputFile(context.Background(), &auth.RequestAuth{DeepSeekToken: "token"}, stdReq)
+	if err != nil {
+		t.Fatalf("apply current input file failed: %v", err)
+	}
+	if out.FinalPrompt == stdReq.FinalPrompt {
+		t.Fatalf("expected live prompt to be rewritten after current input file")
+	}
+	// PromptTokenText must include the uploaded file content (which contains the full context)
+	// plus the neutral live prompt — reflecting the actual downstream token cost.
+	if !strings.Contains(out.PromptTokenText, "first user turn") || !strings.Contains(out.PromptTokenText, "latest user turn") {
+		t.Fatalf("expected prompt token text to contain file context with full conversation, got %q", out.PromptTokenText)
+	}
+	if strings.Contains(out.PromptTokenText, "[file content end]") || strings.Contains(out.PromptTokenText, "[file name]:") {
+		t.Fatalf("expected prompt token text to omit file wrapper tags, got %q", out.PromptTokenText)
+	}
+	if !strings.Contains(out.PromptTokenText, "# DS2API_HISTORY.txt") || !strings.Contains(out.PromptTokenText, "=== 1. SYSTEM ===") {
+		t.Fatalf("expected prompt token text to include numbered history transcript, got %q", out.PromptTokenText)
+	}
+	if !strings.Contains(out.PromptTokenText, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected prompt token text to also include continuation prompt, got %q", out.PromptTokenText)
+	}
+	if strings.Contains(out.FinalPrompt, "first user turn") || strings.Contains(out.FinalPrompt, "latest user turn") {
+		t.Fatalf("expected live prompt to hide original turns, got %q", out.FinalPrompt)
+	}
 }
 
 func TestApplyCurrentInputFileUploadsFullContextFile(t *testing.T) {
@@ -316,7 +378,7 @@ func TestApplyCurrentInputFileUploadsFullContextFile(t *testing.T) {
 		DS: ds,
 	}
 	req := map[string]any{
-		"model":    "deepseek-v4-flash",
+		"model":    "deepseek-v4-vision",
 		"messages": historySplitTestMessages(),
 	}
 	stdReq, err := promptcompat.NormalizeOpenAIChatRequest(h.Store, req, "")
@@ -335,20 +397,23 @@ func TestApplyCurrentInputFileUploadsFullContextFile(t *testing.T) {
 		t.Fatalf("expected one current input upload, got %d", len(ds.uploadCalls))
 	}
 	upload := ds.uploadCalls[0]
-	if upload.Filename != "IGNORE.txt" {
-		t.Fatalf("expected IGNORE.txt upload, got %q", upload.Filename)
+	if upload.Filename != "DS2API_HISTORY.txt" {
+		t.Fatalf("expected DS2API_HISTORY.txt upload, got %q", upload.Filename)
+	}
+	if upload.ModelType != "vision" {
+		t.Fatalf("expected vision model type for vision request, got %q", upload.ModelType)
 	}
 	uploadedText := string(upload.Data)
-	for _, want := range []string{"system instructions", "first user turn", "hidden reasoning", "tool result", "latest user turn", promptcompat.ThinkingInjectionMarker} {
+	for _, want := range []string{"# DS2API_HISTORY.txt", "=== 1. SYSTEM ===", "=== 2. USER ===", "=== 3. ASSISTANT ===", "=== 4. TOOL ===", "=== 5. USER ===", "system instructions", "first user turn", "hidden reasoning", "tool result", "latest user turn", promptcompat.ThinkingInjectionMarker} {
 		if !strings.Contains(uploadedText, want) {
 			t.Fatalf("expected full context file to contain %q, got %q", want, uploadedText)
 		}
 	}
-	if strings.Contains(out.FinalPrompt, "first user turn") || strings.Contains(out.FinalPrompt, "latest user turn") || strings.Contains(out.FinalPrompt, "CURRENT_USER_INPUT.txt") || strings.Contains(out.FinalPrompt, "IGNORE.txt") || strings.Contains(out.FinalPrompt, "Read that file") {
-		t.Fatalf("expected live prompt to use only a neutral continuation instruction, got %s", out.FinalPrompt)
+	if strings.Contains(out.FinalPrompt, "first user turn") || strings.Contains(out.FinalPrompt, "latest user turn") || strings.Contains(out.FinalPrompt, "CURRENT_USER_INPUT.txt") || strings.Contains(out.FinalPrompt, "Read that file") {
+		t.Fatalf("expected live prompt to use only a continuation instruction, got %s", out.FinalPrompt)
 	}
-	if !strings.Contains(out.FinalPrompt, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral continuation instruction in live prompt, got %s", out.FinalPrompt)
+	if !strings.Contains(out.FinalPrompt, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation-oriented prompt in live prompt, got %s", out.FinalPrompt)
 	}
 }
 
@@ -380,6 +445,9 @@ func TestApplyCurrentInputFileCarriesHistoryText(t *testing.T) {
 	if out.HistoryText != string(ds.uploadCalls[0].Data) {
 		t.Fatalf("expected current input file flow to preserve uploaded text in history, got %q", out.HistoryText)
 	}
+	if !strings.Contains(out.HistoryText, "# DS2API_HISTORY.txt") || !strings.Contains(out.HistoryText, "=== 1. SYSTEM ===") {
+		t.Fatalf("expected history text to use numbered transcript format, got %q", out.HistoryText)
+	}
 }
 
 func TestChatCompletionsCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *testing.T) {
@@ -411,15 +479,18 @@ func TestChatCompletionsCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *t
 		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
 	}
 	upload := ds.uploadCalls[0]
-	if upload.Filename != "IGNORE.txt" {
+	if upload.Filename != "DS2API_HISTORY.txt" {
 		t.Fatalf("unexpected upload filename: %q", upload.Filename)
 	}
 	if upload.Purpose != "assistants" {
 		t.Fatalf("unexpected purpose: %q", upload.Purpose)
 	}
 	historyText := string(upload.Data)
-	if !strings.Contains(historyText, "[file content end]") || !strings.Contains(historyText, "[file name]: IGNORE") {
-		t.Fatalf("expected injected IGNORE wrapper, got %s", historyText)
+	if strings.Contains(historyText, "[file content end]") || strings.Contains(historyText, "[file content begin]") || strings.Contains(historyText, "[file name]:") {
+		t.Fatalf("expected history transcript without file wrapper tags, got %s", historyText)
+	}
+	if !strings.Contains(historyText, "# DS2API_HISTORY.txt") || !strings.Contains(historyText, "=== 1. SYSTEM ===") {
+		t.Fatalf("expected history transcript to use numbered sections, got %s", historyText)
 	}
 	if !strings.Contains(historyText, "latest user turn") {
 		t.Fatalf("expected full context to include latest turn, got %s", historyText)
@@ -428,8 +499,8 @@ func TestChatCompletionsCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *t
 		t.Fatal("expected completion payload to be captured")
 	}
 	promptText, _ := ds.completionReq["prompt"].(string)
-	if !strings.Contains(promptText, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral completion prompt, got %s", promptText)
+	if !strings.Contains(promptText, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation-oriented prompt, got %s", promptText)
 	}
 	if strings.Contains(promptText, "first user turn") || strings.Contains(promptText, "latest user turn") {
 		t.Fatalf("expected prompt to hide original turns, got %s", promptText)
@@ -438,6 +509,16 @@ func TestChatCompletionsCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *t
 	if len(refIDs) == 0 || refIDs[0] != "file-inline-1" {
 		t.Fatalf("expected uploaded current input file to be first ref_file_id, got %#v", ds.completionReq["ref_file_ids"])
 	}
+	var body map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &body); err != nil {
+		t.Fatalf("decode response failed: %v", err)
+	}
+	usage, _ := body["usage"].(map[string]any)
+	promptTokens := int(usage["prompt_tokens"].(float64))
+	neutralCount := util.CountPromptTokens(promptText, "deepseek-v4-flash")
+	if promptTokens <= neutralCount {
+		t.Fatalf("expected prompt_tokens to exceed neutral live prompt count (includes file context), got=%d neutral=%d", promptTokens, neutralCount)
+	}
 }
 
 func TestResponsesCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *testing.T) {
@@ -470,16 +551,30 @@ func TestResponsesCurrentInputFileUploadsContextAndKeepsNeutralPrompt(t *testing
 	if len(ds.uploadCalls) != 1 {
 		t.Fatalf("expected 1 upload call, got %d", len(ds.uploadCalls))
 	}
+	historyText := string(ds.uploadCalls[0].Data)
+	if !strings.Contains(historyText, "# DS2API_HISTORY.txt") || !strings.Contains(historyText, "=== 1. SYSTEM ===") {
+		t.Fatalf("expected uploaded history text to use numbered transcript format, got %s", historyText)
+	}
 	if ds.completionReq == nil {
 		t.Fatal("expected completion payload to be captured")
 	}
 	promptText, _ := ds.completionReq["prompt"].(string)
-	if !strings.Contains(promptText, "Answer the latest user request directly.") {
-		t.Fatalf("expected neutral completion prompt, got %s", promptText)
+	if !strings.Contains(promptText, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") {
+		t.Fatalf("expected continuation-oriented prompt, got %s", promptText)
 	}
 	if strings.Contains(promptText, "first user turn") || strings.Contains(promptText, "latest user turn") {
 		t.Fatalf("expected prompt to hide original turns, got %s", promptText)
 	}
+	var body map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &body); err != nil {
+		t.Fatalf("decode response failed: %v", err)
+	}
+	usage, _ := body["usage"].(map[string]any)
+	inputTokens := int(usage["input_tokens"].(float64))
+	neutralCount := util.CountPromptTokens(promptText, "deepseek-v4-flash")
+	if inputTokens <= neutralCount {
+		t.Fatalf("expected input_tokens to exceed neutral live prompt count (includes file context), got=%d neutral=%d", inputTokens, neutralCount)
+	}
 }
 
 func TestChatCompletionsCurrentInputFileMapsManagedAuthFailureTo401(t *testing.T) {
@@ -606,11 +701,15 @@ func TestCurrentInputFileWorksAcrossAutoDeleteModes(t *testing.T) {
 			if len(ds.uploadCalls) != 1 {
 				t.Fatalf("expected current input upload for mode=%s, got %d", mode, len(ds.uploadCalls))
 			}
+			historyText := string(ds.uploadCalls[0].Data)
+			if !strings.Contains(historyText, "# DS2API_HISTORY.txt") || !strings.Contains(historyText, "=== 1. SYSTEM ===") {
+				t.Fatalf("expected uploaded history text to use numbered transcript format, got %s", historyText)
+			}
 			if ds.completionReq == nil {
 				t.Fatalf("expected completion payload for mode=%s", mode)
 			}
 			promptText, _ := ds.completionReq["prompt"].(string)
-			if !strings.Contains(promptText, "Answer the latest user request directly.") || strings.Contains(promptText, "first user turn") || strings.Contains(promptText, "latest user turn") {
+			if !strings.Contains(promptText, "Continue from the latest state in the attached DS2API_HISTORY.txt context.") || strings.Contains(promptText, "first user turn") || strings.Contains(promptText, "latest user turn") {
 				t.Fatalf("unexpected prompt for mode=%s: %s", mode, promptText)
 			}
 		})

internal/httpapi/openai/leaked_output_sanitize_test.go

@@ -42,6 +42,14 @@ func TestSanitizeLeakedOutputRemovesDanglingThinkBlock(t *testing.T) {
 	}
 }
 
+func TestSanitizeLeakedOutputRemovesCompleteDSMLToolCallWrapper(t *testing.T) {
+	raw := "前置文本\n<｜DSML｜tool_calls>\n<｜DSML｜invoke name=\"Bash\">\n<｜DSML｜parameter name=\"command\"></｜DSML｜parameter>\n</｜DSML｜invoke>\n</｜DSML｜tool_calls>\n后置文本"
+	got := sanitizeLeakedOutput(raw)
+	if got != "前置文本\n\n后置文本" {
+		t.Fatalf("unexpected sanitize result for leaked dsml wrapper: %q", got)
+	}
+}
+
 func TestSanitizeLeakedOutputRemovesAgentXMLLeaks(t *testing.T) {
 	raw := "Done.<attempt_completion><result>Some final answer</result></attempt_completion>"
 	got := sanitizeLeakedOutput(raw)

internal/httpapi/openai/responses/empty_retry_runtime.go

@@ -1,7 +1,6 @@
 package responses
 
 import (
-	"context"
 	"io"
 	"net/http"
 	"strings"
@@ -10,121 +9,12 @@ import (
 	"ds2api/internal/auth"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
-	openaifmt "ds2api/internal/format/openai"
 	"ds2api/internal/promptcompat"
-	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
-	"ds2api/internal/toolcall"
 )
 
-type responsesNonStreamResult struct {
-	thinking              string
-	toolDetectionThinking string
-	text                  string
-	contentFilter         bool
-	parsed                toolcall.ToolCallParseResult
-	body                  map[string]any
-	responseMessageID     int
-}
-
-func (h *Handler) handleResponsesNonStreamWithRetry(w http.ResponseWriter, ctx context.Context, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
-	attempts := 0
-	currentResp := resp
-	usagePrompt := finalPrompt
-	accumulatedThinking := ""
-	accumulatedToolDetectionThinking := ""
-	for {
-		result, ok := h.collectResponsesNonStreamAttempt(w, currentResp, responseID, model, usagePrompt, thinkingEnabled, searchEnabled, toolNames, toolsRaw)
-		if !ok {
-			return
-		}
-		accumulatedThinking += sse.TrimContinuationOverlap(accumulatedThinking, result.thinking)
-		accumulatedToolDetectionThinking += sse.TrimContinuationOverlap(accumulatedToolDetectionThinking, result.toolDetectionThinking)
-		result.thinking = accumulatedThinking
-		result.toolDetectionThinking = accumulatedToolDetectionThinking
-		result.parsed = detectAssistantToolCalls(result.text, result.thinking, result.toolDetectionThinking, toolNames)
-		result.body = openaifmt.BuildResponseObjectWithToolCalls(responseID, model, usagePrompt, result.thinking, result.text, result.parsed.Calls, toolsRaw)
-
-		if !shouldRetryResponsesNonStream(result, attempts) {
-			h.finishResponsesNonStreamResult(w, result, attempts, owner, responseID, toolChoice, traceID)
-			return
-		}
-
-		attempts++
-		config.Logger.Info("[openai_empty_retry] attempting synthetic retry", "surface", "responses", "stream", false, "retry_attempt", attempts, "parent_message_id", result.responseMessageID)
-		retryPow, powErr := h.DS.GetPow(ctx, a, 3)
-		if powErr != nil {
-			config.Logger.Warn("[openai_empty_retry] retry PoW fetch failed, falling back to original PoW", "surface", "responses", "stream", false, "retry_attempt", attempts, "error", powErr)
-			retryPow = pow
-		}
-		nextResp, err := h.DS.CallCompletion(ctx, a, clonePayloadForEmptyOutputRetry(payload, result.responseMessageID), retryPow, 3)
-		if err != nil {
-			writeOpenAIError(w, http.StatusInternalServerError, "Failed to get completion.")
-			config.Logger.Warn("[openai_empty_retry] retry request failed", "surface", "responses", "stream", false, "retry_attempt", attempts, "error", err)
-			return
-		}
-		usagePrompt = usagePromptWithEmptyOutputRetry(finalPrompt, attempts)
-		currentResp = nextResp
-	}
-}
-
-func (h *Handler) collectResponsesNonStreamAttempt(w http.ResponseWriter, resp *http.Response, responseID, model, usagePrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any) (responsesNonStreamResult, bool) {
-	defer func() { _ = resp.Body.Close() }()
-	if resp.StatusCode != http.StatusOK {
-		body, _ := io.ReadAll(resp.Body)
-		writeOpenAIError(w, resp.StatusCode, strings.TrimSpace(string(body)))
-		return responsesNonStreamResult{}, false
-	}
-	result := sse.CollectStream(resp, thinkingEnabled, false)
-	stripReferenceMarkers := h.compatStripReferenceMarkers()
-	sanitizedThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	toolDetectionThinking := cleanVisibleOutput(result.ToolDetectionThinking, stripReferenceMarkers)
-	sanitizedText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
-	if searchEnabled {
-		sanitizedText = replaceCitationMarkersWithLinks(sanitizedText, result.CitationLinks)
-	}
-	textParsed := detectAssistantToolCalls(sanitizedText, sanitizedThinking, toolDetectionThinking, toolNames)
-	responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, model, usagePrompt, sanitizedThinking, sanitizedText, textParsed.Calls, toolsRaw)
-	return responsesNonStreamResult{
-		thinking:              sanitizedThinking,
-		toolDetectionThinking: toolDetectionThinking,
-		text:                  sanitizedText,
-		contentFilter:         result.ContentFilter,
-		parsed:                textParsed,
-		body:                  responseObj,
-		responseMessageID:     result.ResponseMessageID,
-	}, true
-}
-
-func (h *Handler) finishResponsesNonStreamResult(w http.ResponseWriter, result responsesNonStreamResult, attempts int, owner, responseID string, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
-	if len(result.parsed.Calls) == 0 && writeUpstreamEmptyOutputError(w, result.text, result.thinking, result.contentFilter) {
-		config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "responses", "stream", false, "retry_attempts", attempts, "success_source", "none", "content_filter", result.contentFilter)
-		return
-	}
-	logResponsesToolPolicyRejection(traceID, toolChoice, result.parsed, "text")
-	if toolChoice.IsRequired() && len(result.parsed.Calls) == 0 {
-		writeOpenAIErrorWithCode(w, http.StatusUnprocessableEntity, "tool_choice requires at least one valid tool call.", "tool_choice_violation")
-		return
-	}
-	h.getResponseStore().put(owner, responseID, result.body)
-	writeJSON(w, http.StatusOK, result.body)
-	source := "first_attempt"
-	if attempts > 0 {
-		source = "synthetic_retry"
-	}
-	config.Logger.Info("[openai_empty_retry] completed", "surface", "responses", "stream", false, "retry_attempts", attempts, "success_source", source)
-}
-
-func shouldRetryResponsesNonStream(result responsesNonStreamResult, attempts int) bool {
-	return emptyOutputRetryEnabled() &&
-		attempts < emptyOutputRetryMaxAttempts() &&
-		!result.contentFilter &&
-		len(result.parsed.Calls) == 0 &&
-		strings.TrimSpace(result.text) == ""
-}
-
-func (h *Handler) handleResponsesStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
-	streamRuntime, initialType, ok := h.prepareResponsesStreamRuntime(w, resp, owner, responseID, model, finalPrompt, thinkingEnabled, searchEnabled, toolNames, toolsRaw, toolChoice, traceID)
+func (h *Handler) handleResponsesStreamWithRetry(w http.ResponseWriter, r *http.Request, a *auth.RequestAuth, resp *http.Response, payload map[string]any, pow, owner, responseID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
+	streamRuntime, initialType, ok := h.prepareResponsesStreamRuntime(w, resp, owner, responseID, model, finalPrompt, refFileTokens, thinkingEnabled, searchEnabled, toolNames, toolsRaw, toolChoice, traceID)
 	if !ok {
 		return
 	}
@@ -165,7 +55,7 @@ func (h *Handler) handleResponsesStreamWithRetry(w http.ResponseWriter, r *http.
 	}
 }
 
-func (h *Handler) prepareResponsesStreamRuntime(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) (*responsesStreamRuntime, string, bool) {
+func (h *Handler) prepareResponsesStreamRuntime(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) (*responsesStreamRuntime, string, bool) {
 	if resp.StatusCode != http.StatusOK {
 		defer func() { _ = resp.Body.Close() }()
 		body, _ := io.ReadAll(resp.Body)
@@ -190,6 +80,7 @@ func (h *Handler) prepareResponsesStreamRuntime(w http.ResponseWriter, resp *htt
 			h.getResponseStore().put(owner, responseID, obj)
 		},
 	)
+	streamRuntime.refFileTokens = refFileTokens
 	streamRuntime.sendCreated()
 	return streamRuntime, initialType, true
 }
@@ -212,7 +103,13 @@ func (h *Handler) consumeResponsesStreamAttempt(r *http.Request, resp *http.Resp
 				finalReason = "content_filter"
 			}
 		},
+		OnContextDone: func() {
+			streamRuntime.markContextCancelled()
+		},
 	})
+	if streamRuntime.finalErrorCode == string(streamengine.StopReasonContextCancelled) {
+		return true, false
+	}
 	terminalWritten := streamRuntime.finalize(finalReason, allowDeferEmpty && finalReason != "content_filter")
 	if terminalWritten {
 		return true, false
@@ -225,6 +122,10 @@ func logResponsesStreamTerminal(streamRuntime *responsesStreamRuntime, attempts
 	if attempts > 0 {
 		source = "synthetic_retry"
 	}
+	if streamRuntime.finalErrorCode == string(streamengine.StopReasonContextCancelled) {
+		config.Logger.Info("[openai_empty_retry] terminal cancelled", "surface", "responses", "stream", true, "retry_attempts", attempts, "error_code", streamRuntime.finalErrorCode)
+		return
+	}
 	if streamRuntime.failed {
 		config.Logger.Info("[openai_empty_retry] terminal empty output", "surface", "responses", "stream", true, "retry_attempts", attempts, "success_source", "none", "error_code", streamRuntime.finalErrorCode)
 		return

internal/httpapi/openai/responses/empty_retry_runtime_test.go

@@ -0,0 +1,70 @@
+package responses
+
+import (
+	"context"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"testing"
+
+	"ds2api/internal/promptcompat"
+	"ds2api/internal/stream"
+)
+
+func makeResponsesOpenAISSEHTTPResponse(lines ...string) *http.Response {
+	body := strings.Join(lines, "\n")
+	if !strings.HasSuffix(body, "\n") {
+		body += "\n"
+	}
+	return &http.Response{
+		StatusCode: http.StatusOK,
+		Header:     make(http.Header),
+		Body:       io.NopCloser(strings.NewReader(body)),
+	}
+}
+
+func TestConsumeResponsesStreamAttemptMarksContextCancelledState(t *testing.T) {
+	ctx, cancel := context.WithCancel(context.Background())
+	cancel()
+
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil).WithContext(ctx)
+	rec := httptest.NewRecorder()
+	streamRuntime := newResponsesStreamRuntime(
+		rec,
+		http.NewResponseController(rec),
+		true,
+		"resp-cancelled",
+		"deepseek-v4-flash",
+		"prompt",
+		false,
+		false,
+		true,
+		nil,
+		nil,
+		false,
+		false,
+		promptcompat.DefaultToolChoicePolicy(),
+		"",
+		nil,
+	)
+	resp := makeResponsesOpenAISSEHTTPResponse(
+		`data: {"p":"response/content","v":"hello"}`,
+		`data: [DONE]`,
+	)
+
+	h := &Handler{}
+	terminalWritten, retryable := h.consumeResponsesStreamAttempt(req, resp, streamRuntime, "text", false, true)
+	if !terminalWritten || retryable {
+		t.Fatalf("expected cancelled attempt to terminate without retry, got terminalWritten=%v retryable=%v", terminalWritten, retryable)
+	}
+	if !streamRuntime.failed {
+		t.Fatalf("expected cancelled response stream to be marked failed")
+	}
+	if got, want := streamRuntime.finalErrorCode, string(stream.StopReasonContextCancelled); got != want {
+		t.Fatalf("expected cancelled final error code %q, got %q", want, got)
+	}
+	if streamRuntime.finalErrorMessage == "" {
+		t.Fatalf("expected cancelled final error message to be preserved")
+	}
+}

internal/httpapi/openai/responses/handler.go

@@ -130,6 +130,6 @@ func filterIncrementalToolCallDeltasByAllowed(deltas []toolstream.ToolCallDelta,
 	return shared.FilterIncrementalToolCallDeltasByAllowed(deltas, seenNames)
 }
 
-func detectAssistantToolCalls(text, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
-	return shared.DetectAssistantToolCalls(text, exposedThinking, detectionThinking, toolNames)
+func detectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
+	return shared.DetectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking, toolNames)
 }

internal/httpapi/openai/responses/ref_file_tokens.go

@@ -0,0 +1,26 @@
+package responses
+
+// addRefFileTokensToUsage adds inline-uploaded file token estimates to an existing
+// usage map inside a response object. This keeps the token accounting aware of file
+// content that the upstream model processes but that is not part of the prompt text.
+func addRefFileTokensToUsage(obj map[string]any, refFileTokens int) {
+	if refFileTokens <= 0 || obj == nil {
+		return
+	}
+	usage, ok := obj["usage"].(map[string]any)
+	if !ok || usage == nil {
+		return
+	}
+	for _, key := range []string{"input_tokens", "prompt_tokens"} {
+		if v, ok := usage[key]; ok {
+			if n, ok := v.(int); ok {
+				usage[key] = n + refFileTokens
+			}
+		}
+	}
+	if v, ok := usage["total_tokens"]; ok {
+		if n, ok := v.(int); ok {
+			usage["total_tokens"] = n + refFileTokens
+		}
+	}
+}

internal/httpapi/openai/responses/responses_handler.go

@@ -12,6 +12,7 @@ import (
 	"github.com/google/uuid"
 
 	"ds2api/internal/auth"
+	"ds2api/internal/completionruntime"
 	"ds2api/internal/config"
 	dsprotocol "ds2api/internal/deepseek/protocol"
 	openaifmt "ds2api/internal/format/openai"
@@ -92,36 +93,34 @@ func (h *Handler) Responses(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	sessionID, err := h.DS.CreateSession(r.Context(), a, 3)
-	if err != nil {
-		if a.UseConfigToken {
-			writeOpenAIError(w, http.StatusUnauthorized, "Account token is invalid. Please re-login the account in admin.")
-		} else {
-			writeOpenAIError(w, http.StatusUnauthorized, "Invalid token. If this should be a DS2API key, add it to config.keys first.")
+	responseID := "resp_" + strings.ReplaceAll(uuid.NewString(), "-", "")
+	if !stdReq.Stream {
+		result, outErr := completionruntime.ExecuteNonStreamWithRetry(r.Context(), h.DS, a, stdReq, completionruntime.Options{
+			StripReferenceMarkers: h.compatStripReferenceMarkers(),
+			RetryEnabled:          true,
+		})
+		if outErr != nil {
+			writeOpenAIErrorWithCode(w, outErr.Status, outErr.Message, outErr.Code)
+			return
 		}
-		return
-	}
-	pow, err := h.DS.GetPow(r.Context(), a, 3)
-	if err != nil {
-		writeOpenAIError(w, http.StatusUnauthorized, "Failed to get PoW (invalid token or unknown error).")
-		return
-	}
-	payload := stdReq.CompletionPayload(sessionID)
-	resp, err := h.DS.CallCompletion(r.Context(), a, payload, pow, 3)
-	if err != nil {
-		writeOpenAIError(w, http.StatusInternalServerError, "Failed to get completion.")
+		responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, stdReq.ResponseModel, result.Turn.Prompt, result.Turn.Thinking, result.Turn.Text, result.Turn.ToolCalls, stdReq.ToolsRaw)
+		responseObj["usage"] = responsesUsageFromTurn(result.Turn)
+		h.getResponseStore().put(owner, responseID, responseObj)
+		writeJSON(w, http.StatusOK, responseObj)
 		return
 	}
 
-	responseID := "resp_" + strings.ReplaceAll(uuid.NewString(), "-", "")
-	if stdReq.Stream {
-		h.handleResponsesStreamWithRetry(w, r, a, resp, payload, pow, owner, responseID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, stdReq.ToolChoice, traceID)
+	start, outErr := completionruntime.StartCompletion(r.Context(), h.DS, a, stdReq, completionruntime.Options{})
+	if outErr != nil {
+		writeOpenAIErrorWithCode(w, outErr.Status, outErr.Message, outErr.Code)
 		return
 	}
-	h.handleResponsesNonStreamWithRetry(w, r.Context(), a, resp, payload, pow, owner, responseID, stdReq.ResponseModel, stdReq.FinalPrompt, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, stdReq.ToolChoice, traceID)
+
+	refFileTokens := stdReq.RefFileTokens
+	h.handleResponsesStreamWithRetry(w, r, a, start.Response, start.Payload, start.Pow, owner, responseID, stdReq.ResponseModel, stdReq.PromptTokenText, refFileTokens, stdReq.Thinking, stdReq.Search, stdReq.ToolNames, stdReq.ToolsRaw, stdReq.ToolChoice, traceID)
 }
 
-func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
+func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Response, owner, responseID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
 	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
@@ -131,12 +130,11 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 	result := sse.CollectStream(resp, thinkingEnabled, true)
 	stripReferenceMarkers := h.compatStripReferenceMarkers()
 	sanitizedThinking := cleanVisibleOutput(result.Thinking, stripReferenceMarkers)
-	toolDetectionThinking := cleanVisibleOutput(result.ToolDetectionThinking, stripReferenceMarkers)
 	sanitizedText := cleanVisibleOutput(result.Text, stripReferenceMarkers)
 	if searchEnabled {
 		sanitizedText = replaceCitationMarkersWithLinks(sanitizedText, result.CitationLinks)
 	}
-	textParsed := detectAssistantToolCalls(sanitizedText, sanitizedThinking, toolDetectionThinking, toolNames)
+	textParsed := detectAssistantToolCalls(result.Text, sanitizedText, result.Thinking, result.ToolDetectionThinking, toolNames)
 	if len(textParsed.Calls) == 0 && writeUpstreamEmptyOutputError(w, sanitizedText, sanitizedThinking, result.ContentFilter) {
 		return
 	}
@@ -149,11 +147,14 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 	}
 
 	responseObj := openaifmt.BuildResponseObjectWithToolCalls(responseID, model, finalPrompt, sanitizedThinking, sanitizedText, textParsed.Calls, toolsRaw)
+	if refFileTokens > 0 {
+		addRefFileTokensToUsage(responseObj, refFileTokens)
+	}
 	h.getResponseStore().put(owner, responseID, responseObj)
 	writeJSON(w, http.StatusOK, responseObj)
 }
 
-func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request, resp *http.Response, owner, responseID, model, finalPrompt string, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
+func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request, resp *http.Response, owner, responseID, model, finalPrompt string, refFileTokens int, thinkingEnabled, searchEnabled bool, toolNames []string, toolsRaw any, toolChoice promptcompat.ToolChoicePolicy, traceID string) {
 	defer func() { _ = resp.Body.Close() }()
 	if resp.StatusCode != http.StatusOK {
 		body, _ := io.ReadAll(resp.Body)
@@ -195,6 +196,7 @@ func (h *Handler) handleResponsesStream(w http.ResponseWriter, r *http.Request,
 			h.getResponseStore().put(owner, responseID, obj)
 		},
 	)
+	streamRuntime.refFileTokens = refFileTokens
 	streamRuntime.sendCreated()
 
 	streamengine.ConsumeSSE(streamengine.ConsumeConfig{

internal/httpapi/openai/responses/responses_stream_delta_batch.go

@@ -0,0 +1,39 @@
+package responses
+
+import (
+	"strings"
+
+	openaifmt "ds2api/internal/format/openai"
+)
+
+type responsesDeltaBatch struct {
+	runtime *responsesStreamRuntime
+	kind    string
+	text    strings.Builder
+}
+
+func (b *responsesDeltaBatch) append(kind, text string) {
+	if text == "" {
+		return
+	}
+	if b.kind != "" && b.kind != kind {
+		b.flush()
+	}
+	b.kind = kind
+	b.text.WriteString(text)
+}
+
+func (b *responsesDeltaBatch) flush() {
+	if b.kind == "" || b.text.Len() == 0 {
+		return
+	}
+	text := b.text.String()
+	switch b.kind {
+	case "reasoning":
+		b.runtime.sendEvent("response.reasoning.delta", openaifmt.BuildResponsesReasoningDeltaPayload(b.runtime.responseID, text))
+	case "text":
+		b.runtime.emitTextDelta(text)
+	}
+	b.kind = ""
+	b.text.Reset()
+}

internal/httpapi/openai/responses/responses_stream_runtime_core.go

@@ -1,12 +1,14 @@
 package responses
 
 import (
+	"ds2api/internal/assistantturn"
 	"ds2api/internal/toolcall"
 	"net/http"
 	"strings"
 
 	"ds2api/internal/config"
 	openaifmt "ds2api/internal/format/openai"
+	"ds2api/internal/httpapi/openai/shared"
 	"ds2api/internal/promptcompat"
 	"ds2api/internal/sse"
 	streamengine "ds2api/internal/stream"
@@ -18,13 +20,14 @@ type responsesStreamRuntime struct {
 	rc       *http.ResponseController
 	canFlush bool
 
-	responseID  string
-	model       string
-	finalPrompt string
-	toolNames   []string
-	toolsRaw    any
-	traceID     string
-	toolChoice  promptcompat.ToolChoicePolicy
+	responseID    string
+	model         string
+	finalPrompt   string
+	refFileTokens int
+	toolNames     []string
+	toolsRaw      any
+	traceID       string
+	toolChoice    promptcompat.ToolChoicePolicy
 
 	thinkingEnabled       bool
 	searchEnabled         bool
@@ -35,29 +38,27 @@ type responsesStreamRuntime struct {
 	toolCallsEmitted     bool
 	toolCallsDoneEmitted bool
 
-	sieve                 toolstream.State
-	thinking              strings.Builder
-	toolDetectionThinking strings.Builder
-	text                  strings.Builder
-	visibleText           strings.Builder
-	responseMessageID     int
-	streamToolCallIDs     map[int]string
-	functionItemIDs       map[int]string
-	functionOutputIDs     map[int]int
-	functionArgs          map[int]string
-	functionDone          map[int]bool
-	functionAdded         map[int]bool
-	functionNames         map[int]string
-	messageItemID         string
-	messageOutputID       int
-	nextOutputID          int
-	messageAdded          bool
-	messagePartAdded      bool
-	sequence              int
-	failed                bool
-	finalErrorStatus      int
-	finalErrorMessage     string
-	finalErrorCode        string
+	sieve             toolstream.State
+	accumulator       shared.StreamAccumulator
+	visibleText       strings.Builder
+	responseMessageID int
+	streamToolCallIDs map[int]string
+	functionItemIDs   map[int]string
+	functionOutputIDs map[int]int
+	functionArgs      map[int]string
+	functionDone      map[int]bool
+	functionAdded     map[int]bool
+	functionNames     map[int]string
+	messageItemID     string
+	messageOutputID   int
+	nextOutputID      int
+	messageAdded      bool
+	messagePartAdded  bool
+	sequence          int
+	failed            bool
+	finalErrorStatus  int
+	finalErrorMessage string
+	finalErrorCode    string
 
 	persistResponse func(obj map[string]any)
 }
@@ -105,6 +106,11 @@ func newResponsesStreamRuntime(
 		toolChoice:            toolChoice,
 		traceID:               traceID,
 		persistResponse:       persistResponse,
+		accumulator: shared.StreamAccumulator{
+			ThinkingEnabled:       thinkingEnabled,
+			SearchEnabled:         searchEnabled,
+			StripReferenceMarkers: stripReferenceMarkers,
+		},
 	}
 }
 
@@ -136,21 +142,47 @@ func (s *responsesStreamRuntime) failResponse(status int, message, code string)
 	s.sendDone()
 }
 
+func (s *responsesStreamRuntime) markContextCancelled() {
+	s.failed = true
+	s.finalErrorStatus = 499
+	s.finalErrorMessage = "request context cancelled"
+	s.finalErrorCode = string(streamengine.StopReasonContextCancelled)
+}
+
 func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput bool) bool {
 	s.failed = false
 	s.finalErrorStatus = 0
 	s.finalErrorMessage = ""
 	s.finalErrorCode = ""
-	finalThinking := s.thinking.String()
-	finalToolDetectionThinking := s.toolDetectionThinking.String()
-	finalText := cleanVisibleOutput(s.text.String(), s.stripReferenceMarkers)
-
 	if s.bufferToolContent {
 		s.processToolStreamEvents(toolstream.Flush(&s.sieve, s.toolNames), true, true)
 	}
 
-	textParsed := detectAssistantToolCalls(finalText, finalThinking, finalToolDetectionThinking, s.toolNames)
-	detected := textParsed.Calls
+	finalThinking := s.accumulator.Thinking.String()
+	finalToolDetectionThinking := s.accumulator.ToolDetectionThinking.String()
+	finalText := s.accumulator.Text.String()
+	turn := assistantturn.BuildTurnFromStreamSnapshot(assistantturn.StreamSnapshot{
+		RawText:               s.accumulator.RawText.String(),
+		VisibleText:           finalText,
+		RawThinking:           s.accumulator.RawThinking.String(),
+		VisibleThinking:       finalThinking,
+		DetectionThinking:     finalToolDetectionThinking,
+		ContentFilter:         finishReason == "content_filter",
+		ResponseMessageID:     s.responseMessageID,
+		AlreadyEmittedCalls:   s.toolCallsEmitted,
+		AlreadyEmittedToolRaw: s.toolCallsDoneEmitted,
+	}, assistantturn.BuildOptions{
+		Model:                 s.model,
+		Prompt:                s.finalPrompt,
+		RefFileTokens:         s.refFileTokens,
+		SearchEnabled:         s.searchEnabled,
+		StripReferenceMarkers: s.stripReferenceMarkers,
+		ToolNames:             s.toolNames,
+		ToolsRaw:              s.toolsRaw,
+		ToolChoice:            s.toolChoice,
+	})
+	textParsed := turn.ParsedToolCalls
+	detected := turn.ToolCalls
 	s.logToolPolicyRejections(textParsed)
 
 	if len(detected) > 0 {
@@ -162,12 +194,15 @@ func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput
 
 	s.closeMessageItem()
 
-	if s.toolChoice.IsRequired() && len(detected) == 0 {
-		s.failResponse(http.StatusUnprocessableEntity, "tool_choice requires at least one valid tool call.", "tool_choice_violation")
+	if turn.Error != nil && turn.Error.Code == "tool_choice_violation" {
+		s.failResponse(turn.Error.Status, turn.Error.Message, turn.Error.Code)
 		return true
 	}
-	if len(detected) == 0 && strings.TrimSpace(finalText) == "" {
-		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", finalText, finalThinking)
+	if len(detected) == 0 && strings.TrimSpace(turn.Text) == "" {
+		status, message, code := upstreamEmptyOutputDetail(finishReason == "content_filter", turn.Text, turn.Thinking)
+		if turn.Error != nil {
+			status, message, code = turn.Error.Status, turn.Error.Message, turn.Error.Code
+		}
 		if deferEmptyOutput {
 			s.finalErrorStatus = status
 			s.finalErrorMessage = message
@@ -179,7 +214,7 @@ func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput
 	}
 	s.closeIncompleteFunctionItems()
 
-	obj := s.buildCompletedResponseObject(finalThinking, finalText, detected)
+	obj := s.buildCompletedResponseObject(turn.Thinking, turn.Text, detected)
 	if s.persistResponse != nil {
 		s.persistResponse(obj)
 	}
@@ -188,6 +223,14 @@ func (s *responsesStreamRuntime) finalize(finishReason string, deferEmptyOutput
 	return true
 }
 
+func responsesUsageFromTurn(turn assistantturn.Turn) map[string]any {
+	return map[string]any{
+		"input_tokens":  turn.Usage.InputTokens,
+		"output_tokens": turn.Usage.OutputTokens,
+		"total_tokens":  turn.Usage.TotalTokens,
+	}
+}
+
 func (s *responsesStreamRuntime) logToolPolicyRejections(textParsed toolcall.ToolCallParseResult) {
 	logRejected := func(parsed toolcall.ToolCallParseResult, channel string) {
 		rejected := filteredRejectedToolNamesForLog(parsed.RejectedToolNames)
@@ -219,46 +262,27 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 		return streamengine.ParsedDecision{Stop: true}
 	}
 
-	contentSeen := false
-	for _, p := range parsed.ToolDetectionThinkingParts {
-		trimmed := sse.TrimContinuationOverlap(s.toolDetectionThinking.String(), p.Text)
-		if trimmed != "" {
-			s.toolDetectionThinking.WriteString(trimmed)
-		}
-	}
-	for _, p := range parsed.Parts {
-		cleanedText := cleanVisibleOutput(p.Text, s.stripReferenceMarkers)
-		if cleanedText == "" {
-			continue
-		}
-		if p.Type != "thinking" && s.searchEnabled && sse.IsCitation(cleanedText) {
-			continue
-		}
-		contentSeen = true
+	batch := responsesDeltaBatch{runtime: s}
+	accumulated := s.accumulator.Apply(parsed)
+	for _, p := range accumulated.Parts {
 		if p.Type == "thinking" {
-			if !s.thinkingEnabled {
-				continue
-			}
-			trimmed := sse.TrimContinuationOverlap(s.thinking.String(), cleanedText)
-			if trimmed == "" {
-				continue
-			}
-			s.thinking.WriteString(trimmed)
-			s.sendEvent("response.reasoning.delta", openaifmt.BuildResponsesReasoningDeltaPayload(s.responseID, trimmed))
+			batch.append("reasoning", p.VisibleText)
 			continue
 		}
-
-		trimmed := sse.TrimContinuationOverlap(s.text.String(), cleanedText)
-		if trimmed == "" {
+		if p.RawText == "" {
+			continue
+		}
+		if p.CitationOnly {
 			continue
 		}
-		s.text.WriteString(trimmed)
 		if !s.bufferToolContent {
-			s.emitTextDelta(trimmed)
+			batch.append("text", p.VisibleText)
 			continue
 		}
-		s.processToolStreamEvents(toolstream.ProcessChunk(&s.sieve, trimmed, s.toolNames), true, true)
+		batch.flush()
+		s.processToolStreamEvents(toolstream.ProcessChunk(&s.sieve, p.RawText, s.toolNames), true, true)
 	}
 
-	return streamengine.ParsedDecision{ContentSeen: contentSeen}
+	batch.flush()
+	return streamengine.ParsedDecision{ContentSeen: accumulated.ContentSeen}
 }

internal/httpapi/openai/responses/responses_stream_runtime_events.go

@@ -4,6 +4,7 @@ import (
 	"encoding/json"
 
 	openaifmt "ds2api/internal/format/openai"
+	"ds2api/internal/sse"
 	"ds2api/internal/toolstream"
 )
 
@@ -43,7 +44,10 @@ func (s *responsesStreamRuntime) sendDone() {
 func (s *responsesStreamRuntime) processToolStreamEvents(events []toolstream.Event, emitContent bool, resetAfterToolCalls bool) {
 	for _, evt := range events {
 		if emitContent && evt.Content != "" {
-			s.emitTextDelta(evt.Content)
+			cleaned := cleanVisibleOutput(evt.Content, s.stripReferenceMarkers)
+			if cleaned != "" && (!s.searchEnabled || !sse.IsCitation(cleaned)) {
+				s.emitTextDelta(cleaned)
+			}
 		}
 		if len(evt.ToolCallDeltas) > 0 {
 			if !s.emitEarlyToolDeltas {

internal/httpapi/openai/responses/responses_stream_runtime_toolcalls_finalize.go

@@ -145,7 +145,7 @@ func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, fin
 		}
 	}
 
-	return openaifmt.BuildResponseObjectFromItems(
+	obj := openaifmt.BuildResponseObjectFromItems(
 		s.responseID,
 		s.model,
 		s.finalPrompt,
@@ -154,4 +154,8 @@ func (s *responsesStreamRuntime) buildCompletedResponseObject(finalThinking, fin
 		output,
 		outputText,
 	)
+	if s.refFileTokens > 0 {
+		addRefFileTokensToUsage(obj, s.refFileTokens)
+	}
+	return obj
 }

internal/httpapi/openai/responses/responses_stream_test.go

@@ -27,7 +27,7 @@ func TestHandleResponsesStreamDoesNotEmitReasoningTextCompatEvents(t *testing.T)
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 
 	body := rec.Body.String()
 	if !strings.Contains(body, "event: response.reasoning.delta") {
@@ -57,7 +57,7 @@ func TestHandleResponsesStreamEmitsOutputTextDoneBeforeContentPartDone(t *testin
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 	body := rec.Body.String()
 	if !strings.Contains(body, "event: response.output_text.done") {
 		t.Fatalf("expected response.output_text.done payload, body=%s", body)
@@ -91,7 +91,7 @@ func TestHandleResponsesStreamOutputTextDeltaCarriesItemIndexes(t *testing.T) {
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 	body := rec.Body.String()
 
 	deltaPayload, ok := extractSSEEventPayload(body, "response.output_text.delta")
@@ -109,6 +109,48 @@ func TestHandleResponsesStreamOutputTextDeltaCarriesItemIndexes(t *testing.T) {
 	}
 }
 
+func TestHandleResponsesStreamCoalescesSmallOutputTextDeltas(t *testing.T) {
+	h := &Handler{}
+	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
+	rec := httptest.NewRecorder()
+
+	var streamBody strings.Builder
+	for i := 0; i < 100; i++ {
+		b, _ := json.Marshal(map[string]any{
+			"p": "response/content",
+			"v": "字",
+		})
+		streamBody.WriteString("data: ")
+		streamBody.WriteString(string(b))
+		streamBody.WriteString("\n")
+	}
+	streamBody.WriteString("data: [DONE]\n")
+	resp := &http.Response{
+		StatusCode: http.StatusOK,
+		Body:       io.NopCloser(strings.NewReader(streamBody.String())),
+	}
+
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_coalesce", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+
+	payloads := extractSSEEventPayloads(rec.Body.String(), "response.output_text.delta")
+	if len(payloads) == 0 {
+		t.Fatalf("expected response.output_text.delta payloads, body=%s", rec.Body.String())
+	}
+	var content strings.Builder
+	for _, payload := range payloads {
+		content.WriteString(asString(payload["delta"]))
+	}
+	if got, want := content.String(), strings.Repeat("字", 100); got != want {
+		t.Fatalf("coalesced response content mismatch: got %q want %q body=%s", got, want, rec.Body.String())
+	}
+	if len(payloads) >= 100 {
+		t.Fatalf("expected coalescing to reduce 100 tiny text deltas, got %d body=%s", len(payloads), rec.Body.String())
+	}
+	if !strings.Contains(rec.Body.String(), "event: response.completed") {
+		t.Fatalf("expected completed event, body=%s", rec.Body.String())
+	}
+}
+
 func TestHandleResponsesStreamEmitsDistinctToolCallIDsAcrossSeparateToolBlocks(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
@@ -130,7 +172,7 @@ func TestHandleResponsesStreamEmitsDistinctToolCallIDsAcrossSeparateToolBlocks(t
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, []string{"read_file", "search"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, []string{"read_file", "search"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
 
 	body := rec.Body.String()
 	doneEvents := extractSSEEventPayloads(body, "response.function_call_arguments.done")
@@ -183,7 +225,7 @@ func TestHandleResponsesStreamRequiredToolChoiceFailure(t *testing.T) {
 		Mode:    promptcompat.ToolChoiceRequired,
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, []string{"read_file"}, nil, policy, "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, []string{"read_file"}, nil, policy, "")
 
 	body := rec.Body.String()
 	if !strings.Contains(body, "event: response.failed") {
@@ -213,7 +255,7 @@ func TestHandleResponsesStreamFailsWhenUpstreamHasOnlyThinking(t *testing.T) {
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 
 	body := rec.Body.String()
 	if !strings.Contains(body, "event: response.failed") {
@@ -251,11 +293,11 @@ func TestHandleResponsesStreamPromotesThinkingToolCallsOnFinalizeWithoutMidstrea
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", true, false, []string{"read_file"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, []string{"read_file"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
 
 	body := rec.Body.String()
-	if !strings.Contains(body, "event: response.reasoning.delta") {
-		t.Fatalf("expected reasoning delta in stream body, got %s", body)
+	if strings.Contains(body, "event: response.reasoning.delta") {
+		t.Fatalf("did not expect leaked reasoning delta in stream body, got %s", body)
 	}
 	if !strings.Contains(body, "event: response.function_call_arguments.done") {
 		t.Fatalf("expected finalize fallback function call event, got %s", body)
@@ -288,7 +330,7 @@ func TestHandleResponsesStreamPromotesHiddenThinkingDSMLToolCallsOnFinalize(t *t
 		Mode:    promptcompat.ToolChoiceRequired,
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_hidden", "deepseek-v4-pro", "prompt", false, false, []string{"read_file"}, nil, policy, "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_hidden", "deepseek-v4-pro", "prompt", 0, false, false, []string{"read_file"}, nil, policy, "")
 
 	body := rec.Body.String()
 	if strings.Contains(body, "event: response.reasoning.delta") {
@@ -317,7 +359,7 @@ func TestHandleResponsesNonStreamRequiredToolChoiceViolation(t *testing.T) {
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, []string{"read_file"}, nil, policy, "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, []string{"read_file"}, nil, policy, "")
 	if rec.Code != http.StatusUnprocessableEntity {
 		t.Fatalf("expected 422 for required tool_choice violation, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -344,7 +386,7 @@ func TestHandleResponsesNonStreamRequiredToolChoiceIgnoresThinkingToolPayloadWhe
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", true, false, []string{"read_file"}, nil, policy, "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, true, false, []string{"read_file"}, nil, policy, "")
 	if rec.Code != http.StatusUnprocessableEntity {
 		t.Fatalf("expected 422 for required tool_choice violation, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -366,7 +408,7 @@ func TestHandleResponsesNonStreamReturns429WhenUpstreamOutputEmpty(t *testing.T)
 		)),
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 	if rec.Code != http.StatusTooManyRequests {
 		t.Fatalf("expected 429 for empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -388,7 +430,7 @@ func TestHandleResponsesNonStreamReturnsContentFilterErrorWhenUpstreamFilteredWi
 		)),
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-flash", "prompt", 0, false, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
 	if rec.Code != http.StatusBadRequest {
 		t.Fatalf("expected 400 for filtered empty upstream output, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -410,14 +452,26 @@ func TestHandleResponsesNonStreamReturns429WhenUpstreamHasOnlyThinking(t *testin
 		)),
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
-	if rec.Code != http.StatusTooManyRequests {
-		t.Fatalf("expected 429 for thinking-only upstream output, got %d body=%s", rec.Code, rec.Body.String())
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, nil, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected 200 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)
+	output, _ := out["output"].([]any)
+	if len(output) == 0 {
+		t.Fatal("expected at least one output item")
+	}
+	first, _ := output[0].(map[string]any)
+	content, _ := first["content"].([]any)
+	if len(content) == 0 {
+		t.Fatal("expected at least one content item")
+	}
+	firstContent, _ := content[0].(map[string]any)
+	if asString(firstContent["type"]) != "reasoning" {
+		t.Fatalf("expected reasoning type, got %v", firstContent["type"])
+	}
+	if asString(firstContent["text"]) != "Only thinking" {
+		t.Fatalf("expected text='Only thinking', got %v", firstContent["text"])
 	}
 }
 
@@ -432,7 +486,7 @@ func TestHandleResponsesNonStreamPromotesThinkingToolCallsWhenTextEmpty(t *testi
 		)),
 	}
 
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", true, false, []string{"read_file"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_test", "deepseek-v4-pro", "prompt", 0, true, false, []string{"read_file"}, nil, promptcompat.DefaultToolChoicePolicy(), "")
 	if rec.Code != http.StatusOK {
 		t.Fatalf("expected 200 for thinking tool calls, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -462,7 +516,7 @@ func TestHandleResponsesNonStreamPromotesHiddenThinkingDSMLToolCallsWhenTextEmpt
 		Mode:    promptcompat.ToolChoiceRequired,
 		Allowed: map[string]struct{}{"read_file": {}},
 	}
-	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_hidden", "deepseek-v4-pro", "prompt", false, false, []string{"read_file"}, nil, policy, "")
+	h.handleResponsesNonStream(rec, resp, "owner-a", "resp_hidden", "deepseek-v4-pro", "prompt", 0, false, false, []string{"read_file"}, nil, policy, "")
 	if rec.Code != http.StatusOK {
 		t.Fatalf("expected 200 for hidden thinking tool calls, got %d body=%s", rec.Code, rec.Body.String())
 	}
@@ -509,7 +563,7 @@ func TestHandleResponsesStreamCoercesSchemaDeclaredStringArguments(t *testing.T)
 		Body:       io.NopCloser(strings.NewReader(streamBody)),
 	}
 
-	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_string_protect", "deepseek-v4-flash", "prompt", false, false, []string{"Write"}, toolsRaw, promptcompat.DefaultToolChoicePolicy(), "")
+	h.handleResponsesStream(rec, req, resp, "owner-a", "resp_string_protect", "deepseek-v4-flash", "prompt", 0, false, false, []string{"Write"}, toolsRaw, promptcompat.DefaultToolChoicePolicy(), "")
 
 	payload, ok := extractSSEEventPayload(rec.Body.String(), "response.function_call_arguments.done")
 	if !ok {

internal/httpapi/openai/shared/assistant_toolcalls.go

@@ -6,12 +6,12 @@ import (
 	"ds2api/internal/toolcall"
 )
 
-func DetectAssistantToolCalls(text, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
-	textParsed := toolcall.ParseStandaloneToolCallsDetailed(text, toolNames)
+func DetectAssistantToolCalls(rawText, visibleText, exposedThinking, detectionThinking string, toolNames []string) toolcall.ToolCallParseResult {
+	textParsed := toolcall.ParseStandaloneToolCallsDetailed(rawText, toolNames)
 	if len(textParsed.Calls) > 0 {
 		return textParsed
 	}
-	if strings.TrimSpace(text) != "" {
+	if strings.TrimSpace(visibleText) != "" {
 		return textParsed
 	}
 	thinking := detectionThinking

internal/httpapi/openai/shared/citation_links.go

@@ -13,7 +13,7 @@ func ReplaceCitationMarkersWithLinks(text string, links map[int]string) string {
 	if strings.TrimSpace(text) == "" || len(links) == 0 {
 		return text
 	}
-	zeroBased := strings.Contains(strings.ToLower(text), "[reference:0]")
+	zeroBasedReference := hasZeroBasedReferenceMarker(text)
 	return citationMarkerPattern.ReplaceAllStringFunc(text, func(match string) string {
 		sub := citationMarkerPattern.FindStringSubmatch(match)
 		if len(sub) < 3 {
@@ -24,7 +24,7 @@ func ReplaceCitationMarkersWithLinks(text string, links map[int]string) string {
 			return match
 		}
 		lookupIdx := idx
-		if zeroBased {
+		if strings.EqualFold(sub[1], "reference") && zeroBasedReference {
 			lookupIdx = idx + 1
 		}
 		url := strings.TrimSpace(links[lookupIdx])
@@ -34,3 +34,16 @@ func ReplaceCitationMarkersWithLinks(text string, links map[int]string) string {
 		return fmt.Sprintf("[%d](%s)", idx, url)
 	})
 }
+
+func hasZeroBasedReferenceMarker(text string) bool {
+	for _, sub := range citationMarkerPattern.FindAllStringSubmatch(text, -1) {
+		if len(sub) < 3 || !strings.EqualFold(sub[1], "reference") {
+			continue
+		}
+		idx, err := strconv.Atoi(strings.TrimSpace(sub[2]))
+		if err == nil && idx == 0 {
+			return true
+		}
+	}
+	return false
+}

internal/httpapi/openai/shared/leaked_output_sanitize.go

@@ -3,6 +3,8 @@ package shared
 import (
 	"regexp"
 	"strings"
+
+	"ds2api/internal/toolcall"
 )
 
 var emptyJSONFencePattern = regexp.MustCompile("(?is)```json\\s*```")
@@ -47,10 +49,42 @@ func sanitizeLeakedOutput(text string) string {
 	out = leakedThinkTagPattern.ReplaceAllString(out, "")
 	out = leakedBOSMarkerPattern.ReplaceAllString(out, "")
 	out = leakedMetaMarkerPattern.ReplaceAllString(out, "")
+	out = stripLeakedToolCallWrapperBlocks(out)
 	out = sanitizeLeakedAgentXMLBlocks(out)
 	return out
 }
 
+func stripLeakedToolCallWrapperBlocks(text string) string {
+	if text == "" {
+		return text
+	}
+	var b strings.Builder
+	pos := 0
+	for pos < len(text) {
+		tag, ok := toolcall.FindToolMarkupTagOutsideIgnored(text, pos)
+		if !ok {
+			b.WriteString(text[pos:])
+			break
+		}
+		if tag.Start > pos {
+			b.WriteString(text[pos:tag.Start])
+		}
+		if tag.Closing || tag.Name != "tool_calls" {
+			b.WriteString(text[tag.Start : tag.End+1])
+			pos = tag.End + 1
+			continue
+		}
+		closeTag, ok := toolcall.FindMatchingToolMarkupClose(text, tag)
+		if !ok {
+			b.WriteString(text[tag.Start : tag.End+1])
+			pos = tag.End + 1
+			continue
+		}
+		pos = closeTag.End + 1
+	}
+	return b.String()
+}
+
 func stripDanglingThinkSuffix(text string) string {
 	matches := leakedThinkTagPattern.FindAllStringIndex(text, -1)
 	if len(matches) == 0 {

internal/httpapi/openai/shared/stream_accumulator.go

@@ -0,0 +1,104 @@
+package shared
+
+import (
+	"strings"
+
+	"ds2api/internal/sse"
+)
+
+type StreamAccumulator struct {
+	ThinkingEnabled       bool
+	SearchEnabled         bool
+	StripReferenceMarkers bool
+
+	RawThinking           strings.Builder
+	Thinking              strings.Builder
+	ToolDetectionThinking strings.Builder
+	RawText               strings.Builder
+	Text                  strings.Builder
+}
+
+type StreamPartDelta struct {
+	Type         string
+	RawText      string
+	VisibleText  string
+	CitationOnly bool
+}
+
+type StreamAccumulatorResult struct {
+	ContentSeen bool
+	Parts       []StreamPartDelta
+}
+
+func (a *StreamAccumulator) Apply(parsed sse.LineResult) StreamAccumulatorResult {
+	out := StreamAccumulatorResult{}
+	for _, p := range parsed.ToolDetectionThinkingParts {
+		trimmed := sse.TrimContinuationOverlapFromBuilder(&a.ToolDetectionThinking, p.Text)
+		if trimmed != "" {
+			a.ToolDetectionThinking.WriteString(trimmed)
+		}
+	}
+	for _, p := range parsed.Parts {
+		if p.Type == "thinking" {
+			delta := a.applyThinkingPart(p.Text)
+			if delta.RawText != "" {
+				out.ContentSeen = true
+			}
+			if delta.RawText != "" || delta.VisibleText != "" {
+				out.Parts = append(out.Parts, delta)
+			}
+			continue
+		}
+		delta := a.applyTextPart(p.Text)
+		if delta.RawText != "" {
+			out.ContentSeen = true
+		}
+		if delta.RawText != "" || delta.VisibleText != "" || delta.CitationOnly {
+			out.Parts = append(out.Parts, delta)
+		}
+	}
+	return out
+}
+
+func (a *StreamAccumulator) applyThinkingPart(text string) StreamPartDelta {
+	rawTrimmed := sse.TrimContinuationOverlapFromBuilder(&a.RawThinking, text)
+	if rawTrimmed != "" {
+		a.RawThinking.WriteString(rawTrimmed)
+	}
+	delta := StreamPartDelta{Type: "thinking", RawText: rawTrimmed}
+	if !a.ThinkingEnabled || rawTrimmed == "" {
+		return delta
+	}
+	cleanedText := CleanVisibleOutput(rawTrimmed, a.StripReferenceMarkers)
+	if cleanedText == "" {
+		return delta
+	}
+	trimmed := sse.TrimContinuationOverlapFromBuilder(&a.Thinking, cleanedText)
+	if trimmed == "" {
+		return delta
+	}
+	a.Thinking.WriteString(trimmed)
+	delta.VisibleText = trimmed
+	return delta
+}
+
+func (a *StreamAccumulator) applyTextPart(text string) StreamPartDelta {
+	rawTrimmed := sse.TrimContinuationOverlapFromBuilder(&a.RawText, text)
+	if rawTrimmed == "" {
+		return StreamPartDelta{Type: "text"}
+	}
+	a.RawText.WriteString(rawTrimmed)
+	delta := StreamPartDelta{Type: "text", RawText: rawTrimmed}
+	cleanedText := CleanVisibleOutput(rawTrimmed, a.StripReferenceMarkers)
+	if a.SearchEnabled && sse.IsCitation(cleanedText) {
+		delta.CitationOnly = true
+		return delta
+	}
+	trimmed := sse.TrimContinuationOverlapFromBuilder(&a.Text, cleanedText)
+	if trimmed == "" {
+		return delta
+	}
+	a.Text.WriteString(trimmed)
+	delta.VisibleText = trimmed
+	return delta
+}

internal/httpapi/openai/shared/stream_accumulator_test.go

@@ -0,0 +1,97 @@
+package shared
+
+import (
+	"testing"
+
+	"ds2api/internal/sse"
+)
+
+func TestStreamAccumulatorAppliesThinkingAndTextDedupe(t *testing.T) {
+	acc := StreamAccumulator{ThinkingEnabled: true, StripReferenceMarkers: true}
+	thinkingPrefix := "this is a long thinking snapshot prefix used by DeepSeek continue replay"
+	textPrefix := "this is a long visible answer snapshot prefix used by DeepSeek continue replay"
+	first := acc.Apply(sse.LineResult{
+		Parsed: true,
+		Parts: []sse.ContentPart{
+			{Type: "thinking", Text: thinkingPrefix},
+			{Type: "text", Text: textPrefix},
+		},
+	})
+	second := acc.Apply(sse.LineResult{
+		Parsed: true,
+		Parts: []sse.ContentPart{
+			{Type: "thinking", Text: thinkingPrefix + " next"},
+			{Type: "text", Text: textPrefix + " world"},
+		},
+	})
+
+	if !first.ContentSeen || !second.ContentSeen {
+		t.Fatalf("expected both chunks to mark content seen")
+	}
+	if got := acc.RawThinking.String(); got != thinkingPrefix+" next" {
+		t.Fatalf("raw thinking = %q", got)
+	}
+	if got := acc.Thinking.String(); got != thinkingPrefix+" next" {
+		t.Fatalf("thinking = %q", got)
+	}
+	if got := acc.RawText.String(); got != textPrefix+" world" {
+		t.Fatalf("raw text = %q", got)
+	}
+	if got := acc.Text.String(); got != textPrefix+" world" {
+		t.Fatalf("text = %q", got)
+	}
+	if got := second.Parts[0].VisibleText; got != " next" {
+		t.Fatalf("thinking delta = %q", got)
+	}
+	if got := second.Parts[1].VisibleText; got != " world" {
+		t.Fatalf("text delta = %q", got)
+	}
+}
+
+func TestStreamAccumulatorKeepsHiddenThinkingForToolDetection(t *testing.T) {
+	acc := StreamAccumulator{ThinkingEnabled: false, StripReferenceMarkers: true}
+	result := acc.Apply(sse.LineResult{
+		Parsed: true,
+		Parts: []sse.ContentPart{
+			{Type: "thinking", Text: "<tool_calls></tool_calls>"},
+		},
+		ToolDetectionThinkingParts: []sse.ContentPart{
+			{Type: "thinking", Text: "detect"},
+			{Type: "thinking", Text: " tools"},
+		},
+	})
+
+	if !result.ContentSeen {
+		t.Fatalf("expected hidden thinking to count as upstream content")
+	}
+	if got := acc.RawThinking.String(); got != "<tool_calls></tool_calls>" {
+		t.Fatalf("raw thinking = %q", got)
+	}
+	if got := acc.Thinking.String(); got != "" {
+		t.Fatalf("visible thinking = %q", got)
+	}
+	if got := acc.ToolDetectionThinking.String(); got != "detect tools" {
+		t.Fatalf("tool detection thinking = %q", got)
+	}
+}
+
+func TestStreamAccumulatorSuppressesCitationTextWhenSearchEnabled(t *testing.T) {
+	acc := StreamAccumulator{SearchEnabled: true, StripReferenceMarkers: true}
+	result := acc.Apply(sse.LineResult{
+		Parsed: true,
+		Parts:  []sse.ContentPart{{Type: "text", Text: "[citation:1]"}},
+	})
+
+	if !result.ContentSeen {
+		t.Fatalf("expected citation chunk to mark upstream content")
+	}
+	if len(result.Parts) != 1 || !result.Parts[0].CitationOnly {
+		t.Fatalf("expected citation-only delta, got %#v", result.Parts)
+	}
+	if got := acc.RawText.String(); got != "[citation:1]" {
+		t.Fatalf("raw text = %q", got)
+	}
+	if got := acc.Text.String(); got != "" {
+		t.Fatalf("visible text = %q", got)
+	}
+}

internal/httpapi/openai/shared/upstream_empty.go

@@ -1,9 +1,12 @@
 package shared
 
-import "net/http"
+import (
+	"net/http"
+	"strings"
+)
 
-func ShouldWriteUpstreamEmptyOutputError(text string) bool {
-	return text == ""
+func ShouldWriteUpstreamEmptyOutputError(text, thinking string) bool {
+	return strings.TrimSpace(text) == "" && strings.TrimSpace(thinking) == ""
 }
 
 func UpstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int, string, string) {
@@ -18,7 +21,7 @@ func UpstreamEmptyOutputDetail(contentFilter bool, text, thinking string) (int,
 }
 
 func WriteUpstreamEmptyOutputError(w http.ResponseWriter, text, thinking string, contentFilter bool) bool {
-	if !ShouldWriteUpstreamEmptyOutputError(text) {
+	if !ShouldWriteUpstreamEmptyOutputError(text, thinking) {
 		return false
 	}
 	status, message, code := UpstreamEmptyOutputDetail(contentFilter, text, thinking)

internal/httpapi/openai/stream_status_test.go

@@ -345,7 +345,7 @@ func TestChatCompletionsStreamRetriesEmptyOutputOnSameSession(t *testing.T) {
 
 func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"response_message_id":99,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":99}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -380,9 +380,6 @@ func TestChatCompletionsNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if asString(message["content"]) != "visible" {
 		t.Fatalf("expected retry visible content, got %#v", message)
 	}
-	if !strings.Contains(asString(message["reasoning_content"]), "plan") {
-		t.Fatalf("expected first-attempt reasoning to be preserved, got %#v", message)
-	}
 }
 
 func TestChatCompletionsContentFilterDoesNotRetry(t *testing.T) {
@@ -499,7 +496,7 @@ func TestResponsesStreamRetriesThinkingOnlyOutput(t *testing.T) {
 
 func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	ds := &streamStatusDSSeqStub{resps: []*http.Response{
-		makeOpenAISSEHTTPResponse(`data: {"response_message_id":88,"p":"response/thinking_content","v":"plan"}`, "data: [DONE]"),
+		makeOpenAISSEHTTPResponse(`data: {"response_message_id":88}`, "data: [DONE]"),
 		makeOpenAISSEHTTPResponse(`data: {"p":"response/content","v":"visible"}`, "data: [DONE]"),
 	}}
 	h := &openAITestSurface{
@@ -540,9 +537,9 @@ func TestResponsesNonStreamRetriesThinkingOnlyOutput(t *testing.T) {
 	if len(content) == 0 {
 		t.Fatalf("expected content entries, got %#v", item)
 	}
-	reasoning, _ := content[0].(map[string]any)
-	if asString(reasoning["type"]) != "reasoning" || !strings.Contains(asString(reasoning["text"]), "plan") {
-		t.Fatalf("expected preserved reasoning entry, got %#v", content)
+	textEntry, _ := content[0].(map[string]any)
+	if asString(textEntry["type"]) != "output_text" || asString(textEntry["text"]) != "visible" {
+		t.Fatalf("expected visible text entry, got %#v", content)
 	}
 }