Merge pull request #174 from CJackHwang/dev

Merge pull request #171 from CJackHwang/codex/fix-issue-#170-in-ds2api

Enable env-backed config writeback and bootstrap missing config file

.env.example

@@ -10,6 +10,8 @@ DS2API_ADMIN_KEY=change-me
 DS2API_CONFIG_PATH=/app/config.json
 # 2) inline JSON or Base64 JSON
 # DS2API_CONFIG_JSON=
+# 3) legacy compatibility alias
+# CONFIG_JSON=
 
 # Optional: static admin assets path
 # DS2API_STATIC_ADMIN_DIR=/app/static/admin

.github/workflows/release-artifacts.yml

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

API.en.md

@@ -46,6 +46,7 @@ Use it per deployment mode:
 
 - Local run: read `config.json` directly
 - Docker / Vercel: generate Base64 from `config.json`, then set `DS2API_CONFIG_JSON`
+- Compatibility note: `DS2API_CONFIG_JSON` may also contain raw JSON directly; `CONFIG_JSON` is the legacy fallback variable
 
 ```bash
 DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
@@ -65,6 +66,7 @@ Two header formats accepted:
 | --- | --- |
 | Bearer Token | `Authorization: Bearer <token>` |
 | API Key Header | `x-api-key: <token>` (no `Bearer` prefix) |
+| Gemini-compatible | `x-goog-api-key: <token>` or `?key=<token>` / `?api_key=<token>` |
 
 **Auth behavior**:
 
@@ -72,6 +74,7 @@ Two header formats accepted:
 - 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.
+Gemini-compatible clients can also send `x-goog-api-key`, `?key=`, or `?api_key=` as the caller credential source.
 
 ### Admin Endpoints (`/admin/*`)
 
@@ -124,13 +127,16 @@ Two header formats accepted:
 | GET | `/admin/queue/status` | Admin | Account queue status |
 | POST | `/admin/accounts/test` | Admin | Test one account |
 | POST | `/admin/accounts/test-all` | Admin | Test all accounts |
+| POST | `/admin/accounts/sessions/delete-all` | Admin | Delete all sessions for one account |
 | POST | `/admin/import` | Admin | Batch import keys/accounts |
 | POST | `/admin/test` | Admin | Test API through service |
 | POST | `/admin/vercel/sync` | Admin | Sync config to Vercel |
 | GET | `/admin/vercel/status` | Admin | Vercel sync status |
+| POST | `/admin/vercel/status` | Admin | Vercel sync status / draft compare |
 | GET | `/admin/export` | Admin | Export config JSON/Base64 |
 | GET | `/admin/dev/captures` | Admin | Read local packet-capture entries |
 | DELETE | `/admin/dev/captures` | Admin | Clear local packet-capture entries |
+| GET | `/admin/version` | Admin | Check current version and latest Release |
 
 ---
 
@@ -580,6 +586,10 @@ Returns sanitized config.
 ```json
 {
   "keys": ["k1", "k2"],
+  "env_backed": false,
+  "env_source_present": true,
+  "env_writeback_enabled": true,
+  "config_path": "/data/config.json",
   "accounts": [
     {
       "identifier": "user@example.com",
@@ -599,7 +609,7 @@ Returns sanitized config.
 
 ### `POST /admin/config`
 
-Updatable fields: `keys`, `accounts`, `claude_mapping`.
+Only updates `keys`, `accounts`, and `claude_mapping`.
 
 **Request**:
 
@@ -620,25 +630,27 @@ Updatable fields: `keys`, `accounts`, `claude_mapping`.
 
 Reads runtime settings and status, including:
 
-- `admin` (JWT expiry, default-password warning, etc.)
-- `runtime` (`account_max_inflight`, `account_max_queue`, `global_max_inflight`)
-- `toolcall` / `responses` / `embeddings`
+- `success`
+- `admin` (`has_password_hash`, `jwt_expire_hours`, `jwt_valid_after_unix`, `default_password_warning`)
+- `runtime` (`account_max_inflight`, `account_max_queue`, `global_max_inflight`, `token_refresh_interval_hours`)
+- `responses` / `embeddings`
 - `auto_delete` (`sessions`)
 - `claude_mapping` / `model_aliases`
 - `env_backed`, `needs_vercel_sync`
+- `toolcall` policy is fixed to `feature_match + high` and is no longer returned or editable via settings
 
 ### `PUT /admin/settings`
 
 Hot-updates runtime settings. Supported fields:
 
 - `admin.jwt_expire_hours`
-- `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight`
-- `toolcall.mode` / `toolcall.early_emit_confidence`
+- `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.sessions`
 - `claude_mapping`
 - `model_aliases`
+- `toolcall` policy is fixed and is no longer writable through settings
 
 ### `POST /admin/settings/password`
 
@@ -650,6 +662,8 @@ Request example:
 {"new_password":"your-new-password"}
 ```
 
+It also accepts `{"password":"your-new-password"}`.
+
 ### `POST /admin/config/import`
 
 Imports full config with:
@@ -658,6 +672,8 @@ Imports full config with:
 - `mode=replace`
 
 The request can send config directly, or wrapped as `{"config": {...}, "mode":"merge"}`.
+Query params `?mode=merge` / `?mode=replace` are also supported.
+Import accepts `keys`, `accounts`, `claude_mapping` / `claude_model_mapping`, `model_aliases`, `admin`, `runtime`, `responses`, `embeddings`, and `auto_delete`; legacy `toolcall` fields are ignored.
 
 ### `GET /admin/config/export`
 
@@ -683,6 +699,7 @@ Exports full config in three forms: `config`, `json`, and `base64`.
 | --- | --- | --- |
 | `page` | `1` | ≥ 1 |
 | `page_size` | `10` | 1–100 |
+| `q` | empty | Filter by identifier / email / mobile |
 
 **Response**:
 
@@ -695,7 +712,8 @@ Exports full config in three forms: `config`, `json`, and `base64`.
       "mobile": "",
       "has_password": true,
       "has_token": true,
-      "token_preview": "abc..."
+      "token_preview": "abc...",
+      "test_status": "ok"
     }
   ],
   "total": 25,
@@ -705,6 +723,8 @@ Exports full config in three forms: `config`, `json`, and `base64`.
 }
 ```
 
+Returned items also include `test_status`, usually `ok` or `failed`.
+
 ### `POST /admin/accounts`
 
 ```json
@@ -757,10 +777,14 @@ Exports full config in three forms: `config`, `json`, and `base64`.
   "success": true,
   "response_time": 1240,
   "message": "API test successful (session creation only)",
-  "model": "deepseek-chat"
+  "model": "deepseek-chat",
+  "session_count": 0,
+  "config_writable": true
 }
 ```
 
+If a `message` is provided, `thinking` may also be included when the upstream response carries reasoning text.
+
 ### `POST /admin/accounts/test-all`
 
 Optional request field: `model`.
@@ -774,6 +798,25 @@ Optional request field: `model`.
 }
 ```
 
+The internal concurrency limit is currently fixed at 5.
+
+### `POST /admin/accounts/sessions/delete-all`
+
+Deletes all DeepSeek sessions for a specific account. Request example:
+
+```json
+{"identifier":"user@example.com"}
+```
+
+Response:
+
+```json
+{"success": true, "message": "删除成功"}
+```
+
+If the account is missing or deletion fails, `success` becomes `false` and `message` contains the error.
+The current handler returns the Chinese literal `删除成功` on success.
+
 ### `POST /admin/import`
 
 Batch import keys and accounts.
@@ -851,16 +894,25 @@ Or manual deploy required:
 }
 ```
 
+Failed account checks are returned in `failed_accounts`, and any saved Vercel credentials are returned in `saved_credentials`.
+
 ### `GET /admin/vercel/status`
 
 ```json
 {
   "synced": true,
   "last_sync_time": 1738400000,
-  "has_synced_before": true
+  "has_synced_before": true,
+  "env_backed": false,
+  "config_hash": "....",
+  "last_synced_hash": "....",
+  "draft_hash": "....",
+  "draft_differs": false
 }
 ```
 
+`POST /admin/vercel/status` can also accept `config_override` to compare a draft config against the current synced config.
+
 ### `GET /admin/export`
 
 ```json
@@ -870,6 +922,29 @@ Or manual deploy required:
 }
 ```
 
+This is the same payload as `GET /admin/config/export`, just with a shorter path.
+
+### `GET /admin/version`
+
+Checks the current build version and the latest GitHub Release:
+
+```json
+{
+  "success": true,
+  "current_version": "2.3.5",
+  "current_tag": "v2.3.5",
+  "source": "file:VERSION",
+  "checked_at": "2026-03-29T00:00:00Z",
+  "latest_tag": "v2.3.6",
+  "latest_version": "2.3.6",
+  "release_url": "https://github.com/CJackHwang/ds2api/releases/tag/v2.3.6",
+  "published_at": "2026-03-28T12:00:00Z",
+  "has_update": true
+}
+```
+
+If GitHub API access fails, the response includes `check_error` while still returning HTTP 200.
+
 ### `GET /admin/dev/captures`
 
 Reads local packet-capture status and recent entries (Admin auth required):

API.md

@@ -46,6 +46,7 @@ cp config.example.json config.json
 
 - 本地运行：直接读取 `config.json`
 - Docker / Vercel：从 `config.json` 生成 Base64，填入 `DS2API_CONFIG_JSON`
+- 兼容写法：`DS2API_CONFIG_JSON` 也可直接填原始 JSON；`CONFIG_JSON` 是旧版兼容回退变量
 
 ```bash
 DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
@@ -65,6 +66,7 @@ Vercel 一键部署可先只填 `DS2API_ADMIN_KEY`，部署后在 `/admin` 导
 | --- | --- |
 | Bearer Token | `Authorization: Bearer <token>` |
 | API Key Header | `x-api-key: <token>`（无 `Bearer` 前缀） |
+| Gemini 兼容 | `x-goog-api-key: <token>` 或 `?key=<token>` / `?api_key=<token>` |
 
 **鉴权行为**：
 
@@ -72,6 +74,7 @@ Vercel 一键部署可先只填 `DS2API_ADMIN_KEY`，部署后在 `/admin` 导
 - 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。
+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/*`）
 
@@ -124,13 +127,16 @@ Vercel 一键部署可先只填 `DS2API_ADMIN_KEY`，部署后在 `/admin` 导
 | GET | `/admin/queue/status` | Admin | 账号队列状态 |
 | POST | `/admin/accounts/test` | Admin | 测试单个账号 |
 | POST | `/admin/accounts/test-all` | Admin | 测试全部账号 |
+| POST | `/admin/accounts/sessions/delete-all` | Admin | 删除某账号的全部会话 |
 | POST | `/admin/import` | Admin | 批量导入 keys/accounts |
 | POST | `/admin/test` | Admin | 测试当前 API 可用性 |
 | POST | `/admin/vercel/sync` | Admin | 同步配置到 Vercel |
 | GET | `/admin/vercel/status` | Admin | Vercel 同步状态 |
+| POST | `/admin/vercel/status` | Admin | Vercel 同步状态 / 草稿对比 |
 | GET | `/admin/export` | Admin | 导出配置 JSON/Base64 |
 | GET | `/admin/dev/captures` | Admin | 查看本地抓包记录 |
 | DELETE | `/admin/dev/captures` | Admin | 清空本地抓包记录 |
+| GET | `/admin/version` | Admin | 查询当前版本与最新 Release |
 
 ---
 
@@ -286,7 +292,8 @@ data: [DONE]
 
 补充说明：
 
-- **非代码块上下文**下，工具 JSON 即使与普通文本混合，也会按特征识别并产出可执行 tool call（前后普通文本仍可透传）。
+- **非代码块上下文**下，工具负载即使与普通文本混合，也会按特征识别并产出可执行 tool call（前后普通文本仍可透传）。
+- 解析器以 XML/Markup 为最高优先级，并兼容 JSON、ANTML、text-kv 等格式输入；最终按客户端协议转译为对应 tool call 结构（OpenAI/Claude/Gemini）。
 - Markdown fenced code block（例如 ```json ... ```）中的 `tool_calls` 仅视为示例文本，不会被执行。
 
 ---
@@ -346,7 +353,8 @@ data: [DONE]
 ```
 
 流式场景下若 `tool_choice=required` 违规，会返回 `response.failed` 后结束（不再发送 `response.completed`）。
-未在 `tools` 声明中的工具名会被严格拒绝，不会作为有效 tool call 下发。
+
+> 当前版本说明：解析层默认“尽量提取结构化 tool call”，未启用基于 `tools` allow-list 的硬拒绝；是否执行仍应由你的工具执行器做白名单校验。
 
 ### `GET /v1/responses/{response_id}`
 
@@ -492,6 +500,8 @@ data: {"type":"message_stop"}
 }
 ```
 
+返回项还会包含 `test_status`，当前值通常为 `ok` 或 `failed`。
+
 ---
 
 ## Gemini 兼容接口
@@ -585,6 +595,10 @@ data: {"type":"message_stop"}
 ```json
 {
   "keys": ["k1", "k2"],
+  "env_backed": false,
+  "env_source_present": true,
+  "env_writeback_enabled": true,
+  "config_path": "/data/config.json",
   "accounts": [
     {
       "identifier": "user@example.com",
@@ -604,7 +618,7 @@ data: {"type":"message_stop"}
 
 ### `POST /admin/config`
 
-可更新 `keys`、`accounts`、`claude_mapping`。
+只更新 `keys`、`accounts`、`claude_mapping`。
 
 **请求**：
 
@@ -625,25 +639,27 @@ data: {"type":"message_stop"}
 
 读取运行时设置与状态，返回：
 
-- `admin`（JWT 过期、默认密码告警等）
-- `runtime`（`account_max_inflight`、`account_max_queue`、`global_max_inflight`）
-- `toolcall` / `responses` / `embeddings`
+- `success`
+- `admin`（`has_password_hash`、`jwt_expire_hours`、`jwt_valid_after_unix`、`default_password_warning`）
+- `runtime`（`account_max_inflight`、`account_max_queue`、`global_max_inflight`、`token_refresh_interval_hours`）
+- `responses` / `embeddings`
 - `auto_delete`（`sessions`）
 - `claude_mapping` / `model_aliases`
 - `env_backed`、`needs_vercel_sync`
+- `toolcall` 策略已固定为 `feature_match + high`，不再通过 settings 返回或修改
 
 ### `PUT /admin/settings`
 
 热更新运行时设置。支持更新：
 
 - `admin.jwt_expire_hours`
-- `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight`
-- `toolcall.mode` / `toolcall.early_emit_confidence`
+- `runtime.account_max_inflight` / `runtime.account_max_queue` / `runtime.global_max_inflight` / `runtime.token_refresh_interval_hours`
 - `responses.store_ttl_seconds`
 - `embeddings.provider`
 - `auto_delete.sessions`
 - `claude_mapping`
 - `model_aliases`
+- `toolcall` 策略已固定，不再作为可写入字段
 
 ### `POST /admin/settings/password`
 
@@ -655,6 +671,8 @@ data: {"type":"message_stop"}
 {"new_password":"your-new-password"}
 ```
 
+也兼容 `{"password":"your-new-password"}`。
+
 ### `POST /admin/config/import`
 
 导入完整配置，支持：
@@ -663,6 +681,8 @@ data: {"type":"message_stop"}
 - `mode=replace`
 
 请求可直接传配置对象，或使用 `{"config": {...}, "mode":"merge"}` 包裹格式。
+也支持在查询参数里传 `?mode=merge` / `?mode=replace`。
+导入时会接受 `keys`、`accounts`、`claude_mapping` / `claude_model_mapping`、`model_aliases`、`admin`、`runtime`、`responses`、`embeddings`、`auto_delete` 等字段；`toolcall` 相关字段会被忽略。
 
 ### `GET /admin/config/export`
 
@@ -688,6 +708,7 @@ data: {"type":"message_stop"}
 | --- | --- | --- |
 | `page` | `1` | ≥ 1 |
 | `page_size` | `10` | 1–100 |
+| `q` | 空 | 按 identifier / email / mobile 过滤 |
 
 **响应**：
 
@@ -700,7 +721,8 @@ data: {"type":"message_stop"}
       "mobile": "",
       "has_password": true,
       "has_token": true,
-      "token_preview": "abc..."
+      "token_preview": "abc...",
+      "test_status": "ok"
     }
   ],
   "total": 25,
@@ -762,10 +784,14 @@ data: {"type":"message_stop"}
   "success": true,
   "response_time": 1240,
   "message": "API 测试成功（仅会话创建）",
-  "model": "deepseek-chat"
+  "model": "deepseek-chat",
+  "session_count": 0,
+  "config_writable": true
 }
 ```
 
+如果传入 `message`，还会附带 `thinking`（当上游返回思考内容时）。
+
 ### `POST /admin/accounts/test-all`
 
 可选请求字段：`model`
@@ -779,6 +805,24 @@ data: {"type":"message_stop"}
 }
 ```
 
+内部并发上限当前固定为 5。
+
+### `POST /admin/accounts/sessions/delete-all`
+
+清空指定账号的所有 DeepSeek 会话。请求体示例：
+
+```json
+{"identifier":"user@example.com"}
+```
+
+响应：
+
+```json
+{"success": true, "message": "删除成功"}
+```
+
+如果账号不存在或删除失败，`success` 会是 `false`，`message` 会返回错误原因。
+
 ### `POST /admin/import`
 
 批量导入 keys 与 accounts。
@@ -856,16 +900,25 @@ data: {"type":"message_stop"}
 }
 ```
 
+失败校验的账号会通过 `failed_accounts` 返回；成功保存到 Vercel 的凭据会通过 `saved_credentials` 返回。
+
 ### `GET /admin/vercel/status`
 
 ```json
 {
   "synced": true,
   "last_sync_time": 1738400000,
-  "has_synced_before": true
+  "has_synced_before": true,
+  "env_backed": false,
+  "config_hash": "....",
+  "last_synced_hash": "....",
+  "draft_hash": "....",
+  "draft_differs": false
 }
 ```
 
+`POST /admin/vercel/status` 还可以携带 `config_override`，用于对比“草稿配置”和当前已同步配置。
+
 ### `GET /admin/export`
 
 ```json
@@ -875,6 +928,29 @@ data: {"type":"message_stop"}
 }
 ```
 
+该接口与 `GET /admin/config/export` 返回相同内容，只是路径更短。
+
+### `GET /admin/version`
+
+查询当前构建版本与 GitHub 最新 Release：
+
+```json
+{
+  "success": true,
+  "current_version": "2.3.5",
+  "current_tag": "v2.3.5",
+  "source": "file:VERSION",
+  "checked_at": "2026-03-29T00:00:00Z",
+  "latest_tag": "v2.3.6",
+  "latest_version": "2.3.6",
+  "release_url": "https://github.com/CJackHwang/ds2api/releases/tag/v2.3.6",
+  "published_at": "2026-03-28T12:00:00Z",
+  "has_update": true
+}
+```
+
+如果 GitHub API 不可用，响应里会额外包含 `check_error`，但 HTTP 状态仍为 200。
+
 ### `GET /admin/dev/captures`
 
 查看本地抓包状态与最近记录（需 Admin 鉴权）：

Dockerfile

@@ -34,7 +34,7 @@ CMD ["/usr/local/bin/ds2api"]
 
 FROM runtime-base AS runtime-from-source
 COPY --from=go-builder /out/ds2api /usr/local/bin/ds2api
-COPY --from=go-builder /app/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
+COPY --from=go-builder /app/internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm /app/sha3_wasm_bg.7b9ca65ddd.wasm
 COPY --from=go-builder /app/config.example.json /app/config.example.json
 COPY --from=webui-builder /app/static/admin /app/static/admin
 

README.MD

@@ -8,7 +8,7 @@
 ![Stars](https://img.shields.io/github/stars/CJackHwang/ds2api.svg)
 ![Forks](https://img.shields.io/github/forks/CJackHwang/ds2api.svg)
 [![Release](https://img.shields.io/github/v/release/CJackHwang/ds2api?display_name=tag)](https://github.com/CJackHwang/ds2api/releases)
-[![Docker](https://img.shields.io/badge/docker-ready-blue.svg)](DEPLOY.md)
+[![Docker](https://img.shields.io/badge/docker-ready-blue.svg)](docs/DEPLOY.md)
 [![Deploy on Zeabur](https://zeabur.com/button.svg)](https://zeabur.com/templates/L4CFHP)
 [![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/CJackHwang/ds2api)
 
@@ -16,6 +16,14 @@
 
 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`）。
 
+> **重要免责声明**
+>
+> 本仓库仅供学习、研究、个人实验和内部验证使用，不提供任何形式的商业授权、适用性保证或结果保证。
+>
+> 作者及仓库维护者不对因使用、修改、分发、部署或依赖本项目而产生的任何直接或间接损失、账号封禁、数据丢失、法律风险或第三方索赔负责。
+>
+> 请勿将本项目用于违反服务条款、协议、法律法规或平台规则的场景。商业使用前请自行确认 `LICENSE`、相关协议以及你是否获得了作者的书面许可。
+
 ## 架构概览
 
 ```mermaid
@@ -68,7 +76,7 @@ flowchart LR
 | 并发队列控制 | 每账号 in-flight 上限 + 等待队列，动态计算建议并发值 |
 | DeepSeek PoW | WASM 计算（`wazero`），无需外部 Node.js 依赖 |
 | Tool Calling | 防泄漏处理：非代码块高置信特征识别、`delta.tool_calls` 早发、结构化增量输出 |
-| Admin API | 配置管理、运行时设置热更新、账号测试 / 批量测试、导入导出、Vercel 同步 |
+| Admin API | 配置管理、运行时设置热更新、账号测试 / 批量测试、会话清理、导入导出、Vercel 同步、版本检查 |
 | WebUI 管理台 | `/admin` 单页应用（中英文双语、深色模式） |
 | 运维探针 | `GET /healthz`（存活）、`GET /readyz`（就绪） |
 
@@ -132,6 +140,7 @@ cp config.example.json config.json
 后续部署建议：
 - 本地运行：直接读取 `config.json`
 - Docker / Vercel：由 `config.json` 生成 `DS2API_CONFIG_JSON`（Base64）注入环境变量
+- 兼容写法：`DS2API_CONFIG_JSON` 也可以直接写原始 JSON；`CONFIG_JSON` 是旧版回退变量
 
 ### 方式一：本地运行
 
@@ -152,7 +161,7 @@ go run ./cmd/ds2api
 
 默认监听地址：`http://localhost:5001`
 
-> **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 install && npm run build`（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）。你也可以手动构建：`./scripts/build-webui.sh`
 
 ### 方式二：Docker 运行
 
@@ -170,6 +179,8 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
+默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请调整 `ports` 配置。
+
 更新镜像：`docker-compose up -d --build`
 
 #### Zeabur 一键部署（Dockerfile）
@@ -202,7 +213,7 @@ base64 < config.json | tr -d '\n'
 
 > **流式说明**：`/v1/chat/completions` 在 Vercel 上默认走 `api/chat-stream.js`（Node Runtime）以保证实时 SSE。鉴权、账号选择、会话/PoW 准备仍由 Go 内部 prepare 接口完成；流式响应（含 `tools`）在 Node 侧执行与 Go 对齐的输出组装与防泄漏处理。
 
-详细部署说明请参阅 [部署指南](DEPLOY.md)。
+详细部署说明请参阅 [部署指南](docs/DEPLOY.md)。
 
 ### 方式四：下载 Release 构建包
 
@@ -259,10 +270,6 @@ cp opencode.json.example opencode.json
   "compat": {
     "wide_input_strict_output": true
   },
-  "toolcall": {
-    "mode": "feature_match",
-    "early_emit_confidence": "high"
-  },
   "responses": {
     "store_ttl_seconds": 900
   },
@@ -279,7 +286,8 @@ cp opencode.json.example opencode.json
   "runtime": {
     "account_max_inflight": 2,
     "account_max_queue": 0,
-    "global_max_inflight": 0
+    "global_max_inflight": 0,
+    "token_refresh_interval_hours": 6
   },
   "auto_delete": {
     "sessions": false
@@ -292,12 +300,12 @@ cp opencode.json.example opencode.json
 - `token`：配置文件中即使填写也会在加载时被清空（不会从 `config.json` 读取 token）；实际 token 仅在运行时内存中维护并自动刷新
 - `model_aliases`：常见模型名（如 GPT/Codex/Claude）到 DeepSeek 模型的映射
 - `compat.wide_input_strict_output`：建议保持 `true`（当前实现默认宽进严出）
-- `toolcall`：固定采用特征匹配 + 高置信早发策略
+- `toolcall`：策略已固定为特征匹配 + 高置信早发，不再作为可配置项
 - `responses.store_ttl_seconds`：`/v1/responses/{id}` 的内存缓存 TTL
 - `embeddings.provider`：embedding 提供方（当前内置 `deterministic/mock/builtin`）
 - `claude_mapping`：字典中 `fast`/`slow` 后缀映射到对应 DeepSeek 模型（兼容读取 `claude_model_mapping`）
 - `admin`：管理后台设置（JWT 过期时间、密码哈希等），可通过 Admin Settings API 热更新
-- `runtime`：运行时参数（并发限制、队列大小），可通过 Admin Settings API 热更新；`account_max_queue=0`/`global_max_inflight=0` 表示按推荐值自动计算
+- `runtime`：运行时参数（并发限制、队列大小、托管账号 token 刷新间隔），可通过 Admin Settings API 热更新；`account_max_queue=0`/`global_max_inflight=0` 表示按推荐值自动计算，`token_refresh_interval_hours=6` 为默认强制重登间隔
 - `auto_delete.sessions`：是否在请求结束后自动清理 DeepSeek 会话（默认 `false`，可在 Settings 热更新）
 
 ### 环境变量
@@ -311,9 +319,14 @@ cp opencode.json.example opencode.json
 | `DS2API_JWT_EXPIRE_HOURS` | Admin JWT 过期小时数 | `24` |
 | `DS2API_CONFIG_PATH` | 配置文件路径 | `config.json` |
 | `DS2API_CONFIG_JSON` | 直接注入配置（JSON 或 Base64） | — |
+| `CONFIG_JSON` | 旧版兼容配置注入 | — |
+| `DS2API_ENV_WRITEBACK` | 环境变量模式下自动写回配置文件并切换文件模式（`1/true/yes/on`） | 关闭 |
 | `DS2API_WASM_PATH` | PoW WASM 文件路径 | 自动查找 |
 | `DS2API_STATIC_ADMIN_DIR` | 管理台静态文件目录 | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | 启动时自动构建 WebUI | 本地开启，Vercel 关闭 |
+| `DS2API_DEV_PACKET_CAPTURE` | 本地开发抓包开关（记录最近会话请求/响应体） | 本地非 Vercel 默认开启 |
+| `DS2API_DEV_PACKET_CAPTURE_LIMIT` | 本地抓包保留条数（超出自动淘汰） | `5` |
+| `DS2API_DEV_PACKET_CAPTURE_MAX_BODY_BYTES` | 单条响应体最大记录字节数 | `2097152` |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | 每账号最大并发 in-flight 请求数 | `2` |
 | `DS2API_ACCOUNT_CONCURRENCY` | 同上（兼容旧名） | — |
 | `DS2API_ACCOUNT_MAX_QUEUE` | 等待队列上限 | `recommended_concurrency` |
@@ -330,6 +343,8 @@ cp opencode.json.example opencode.json
 | `VERCEL_TEAM_ID` | Vercel 团队 ID | — |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | Vercel 部署保护绕过密钥（内部 Node→Go 调用） | — |
 
+> 提示：当检测到 `DS2API_CONFIG_JSON/CONFIG_JSON` 时，管理台会显示当前模式风险与自动持久化状态（含 `DS2API_CONFIG_PATH` 路径与模式切换说明）。
+
 ## 鉴权模式
 
 调用业务接口（`/v1/*`、`/anthropic/*`、Gemini 路由）时支持两种模式：
@@ -340,6 +355,7 @@ cp opencode.json.example opencode.json
 | **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）。
+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
 
@@ -356,13 +372,17 @@ cp opencode.json.example opencode.json
 
 ## Tool Call 适配
 
-当请求中带 `tools` 时，DS2API 会做防泄漏处理：
+当请求中带 `tools` 时，DS2API 会做防泄漏处理与结构化转译：
 
-1. 只在**非代码块上下文**启用 toolcall 特征识别（代码块示例不会触发）
-2. `responses` 流式严格使用官方 item 生命周期事件（`response.output_item.*`、`response.content_part.*`、`response.function_call_arguments.*`）
-3. 未在 `tools` 声明中的工具名会被严格拒绝，不会下发为有效 tool call
+1. 只在**非代码块上下文**启用执行型 toolcall 识别（代码块示例默认不触发）
+2. 解析层以 XML/Markup 为最高优先级，同时兼容 JSON / ANTML / invoke / text-kv，并统一归一到内部工具调用结构
+3. `responses` 流式严格使用官方 item 生命周期事件（`response.output_item.*`、`response.content_part.*`、`response.function_call_arguments.*`）
 4. `responses` 支持并执行 `tool_choice`（`auto`/`none`/`required`/强制函数）；`required` 违规时非流式返回 `422`，流式返回 `response.failed`
-5. 仅在通过策略校验后才会发出有效工具调用事件，避免错误工具名进入客户端执行链
+5. 客户端请求哪种协议，就按该协议返回工具调用（OpenAI/Claude/Gemini 各自原生结构）；模型侧优先约束输出规范 XML，再由兼容层转译
+
+> 说明：当前版本在 parser 层仍以“尽量解析成功”为优先，未启用基于 allow-list 的工具名硬拒绝。
+>
+> 想评估“把工具调用封装成 XML 再输入模型”的方案，可参考：`docs/toolcall-semantics.md`。
 
 ## 本地开发抓包工具
 
@@ -430,6 +450,7 @@ ds2api/
 ├── tests/
 │   ├── compat/              # 兼容性测试夹具与期望输出
 │   └── scripts/             # 统一测试脚本入口（unit/e2e）
+├── docs/                    # 部署 / 贡献 / 测试等辅助文档
 ├── static/admin/            # WebUI 构建产物（不提交到 Git）
 ├── .github/
 │   ├── workflows/           # GitHub Actions（质量门禁 + Release 自动构建）
@@ -449,9 +470,9 @@ ds2api/
 | 文档 | 说明 |
 | --- | --- |
 | [API.md](API.md) / [API.en.md](API.en.md) | API 接口文档（含请求/响应示例） |
-| [DEPLOY.md](DEPLOY.md) / [DEPLOY.en.md](DEPLOY.en.md) | 部署指南（本地/Docker/Vercel/systemd） |
-| [CONTRIBUTING.md](CONTRIBUTING.md) / [CONTRIBUTING.en.md](CONTRIBUTING.en.md) | 贡献指南 |
-| [TESTING.md](TESTING.md) | 测试集使用指南 |
+| [DEPLOY.md](docs/DEPLOY.md) / [DEPLOY.en.md](docs/DEPLOY.en.md) | 部署指南（本地/Docker/Vercel/systemd） |
+| [CONTRIBUTING.md](docs/CONTRIBUTING.md) / [CONTRIBUTING.en.md](docs/CONTRIBUTING.en.md) | 贡献指南 |
+| [TESTING.md](docs/TESTING.md) | 测试集使用指南 |
 
 ## 测试
 
@@ -481,7 +502,7 @@ npm ci --prefix webui && npm run build --prefix webui
 
 ## 测试
 
-详细测试指南请参阅 [TESTING.md](TESTING.md)。
+详细测试指南请参阅 [docs/TESTING.md](docs/TESTING.md)。
 
 ### 快速测试命令
 
@@ -507,4 +528,7 @@ go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
 
 ## 免责声明
 
-本项目基于逆向方式实现，仅供学习与研究使用。稳定性和可用性不作保证，请勿用于违反服务条款或法律法规的场景。
+本项目基于逆向方式实现，仅供学习、研究、个人实验和内部验证使用，不提供任何商业授权、稳定性保证或可用性保证。
+作者及仓库维护者不对因使用、修改、分发、部署或依赖本项目而产生的任何直接或间接损失、账号封禁、数据丢失、法律风险或第三方索赔负责。
+
+请勿将本项目用于违反服务条款、协议、法律法规或平台规则的场景。商业使用前请自行确认 `LICENSE`、相关协议以及你是否获得了作者的书面许可。

README.en.md

@@ -8,7 +8,7 @@
 ![Stars](https://img.shields.io/github/stars/CJackHwang/ds2api.svg)
 ![Forks](https://img.shields.io/github/forks/CJackHwang/ds2api.svg)
 [![Release](https://img.shields.io/github/v/release/CJackHwang/ds2api?display_name=tag)](https://github.com/CJackHwang/ds2api/releases)
-[![Docker](https://img.shields.io/badge/docker-ready-blue.svg)](DEPLOY.en.md)
+[![Docker](https://img.shields.io/badge/docker-ready-blue.svg)](docs/DEPLOY.en.md)
 [![Deploy on Zeabur](https://zeabur.com/button.svg)](https://zeabur.com/templates/L4CFHP)
 [![Deploy with Vercel](https://vercel.com/button)](https://vercel.com/new/clone?repository-url=https://github.com/CJackHwang/ds2api)
 
@@ -16,6 +16,14 @@ 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).
 
+> **Important Disclaimer**
+>
+> This repository is provided for learning, research, personal experimentation, and internal validation only. It does not grant any commercial authorization and comes with no warranty of fitness, stability, or results.
+>
+> The author and repository maintainers are not responsible for any direct or indirect loss, account suspension, data loss, legal risk, or third-party claims arising from use, modification, distribution, deployment, or reliance on this project.
+>
+> Do not use this project in ways that violate service terms, agreements, laws, or platform rules. Before any commercial use, review the `LICENSE`, the relevant terms, and confirm that you have the author's written permission.
+
 ## Architecture Overview
 
 ```mermaid
@@ -68,7 +76,7 @@ flowchart LR
 | Concurrency control | Per-account in-flight limit + waiting queue, dynamic recommended concurrency |
 | DeepSeek PoW | WASM solving via `wazero`, no external Node.js dependency |
 | Tool Calling | Anti-leak handling: non-code-block feature match, early `delta.tool_calls`, structured incremental output |
-| Admin API | Config management, runtime settings hot-reload, account testing/batch test, import/export, Vercel sync |
+| Admin API | Config management, runtime settings hot-reload, account testing/batch test, session cleanup, import/export, Vercel sync, version check |
 | WebUI Admin Panel | SPA at `/admin` (bilingual Chinese/English, dark mode) |
 | Health Probes | `GET /healthz` (liveness), `GET /readyz` (readiness) |
 
@@ -132,6 +140,7 @@ cp config.example.json config.json
 Recommended per deployment mode:
 - Local run: read `config.json` directly
 - Docker / Vercel: generate Base64 from `config.json` and inject as `DS2API_CONFIG_JSON`
+- Compatibility note: `DS2API_CONFIG_JSON` may also contain raw JSON directly; `CONFIG_JSON` is the legacy fallback variable
 
 ### Option 1: Local Run
 
@@ -152,7 +161,7 @@ go run ./cmd/ds2api
 
 Default URL: `http://localhost:5001`
 
-> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm install && npm run build` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
+> **WebUI auto-build**: On first local startup, if `static/admin` is missing, DS2API will auto-run `npm ci` (only when dependencies are missing) and `npm run build -- --outDir static/admin --emptyOutDir` (requires Node.js). You can also build manually: `./scripts/build-webui.sh`
 
 ### Option 2: Docker
 
@@ -170,6 +179,8 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
+The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, adjust the `ports` mapping.
+
 Rebuild after updates: `docker-compose up -d --build`
 
 #### Zeabur One-Click (Dockerfile)
@@ -202,7 +213,7 @@ base64 < config.json | tr -d '\n'
 
 > **Streaming note**: `/v1/chat/completions` on Vercel is routed to `api/chat-stream.js` (Node Runtime) for real-time SSE. Auth, account selection, and session/PoW preparation are still handled by the Go internal prepare endpoint; streaming output (including `tools`) is assembled on Node with Go-aligned anti-leak handling.
 
-For detailed deployment instructions, see the [Deployment Guide](DEPLOY.en.md).
+For detailed deployment instructions, see the [Deployment Guide](docs/DEPLOY.en.md).
 
 ### Option 4: Download Release Binaries
 
@@ -259,10 +270,6 @@ cp opencode.json.example opencode.json
   "compat": {
     "wide_input_strict_output": true
   },
-  "toolcall": {
-    "mode": "feature_match",
-    "early_emit_confidence": "high"
-  },
   "responses": {
     "store_ttl_seconds": 900
   },
@@ -279,7 +286,8 @@ cp opencode.json.example opencode.json
   "runtime": {
     "account_max_inflight": 2,
     "account_max_queue": 0,
-    "global_max_inflight": 0
+    "global_max_inflight": 0,
+    "token_refresh_interval_hours": 6
   },
   "auto_delete": {
     "sessions": false
@@ -292,12 +300,12 @@ cp opencode.json.example opencode.json
 - `token`: Even if set in `config.json`, it is cleared during load (DS2API does not read persisted tokens from config); runtime tokens are maintained/refreshed in memory only
 - `model_aliases`: Map common model names (GPT/Codex/Claude) to DeepSeek models
 - `compat.wide_input_strict_output`: Keep `true` (current default policy)
-- `toolcall`: Fixed to feature matching + high-confidence early emit
+- `toolcall`: Fixed to feature matching + high-confidence early emit, no longer configurable
 - `responses.store_ttl_seconds`: In-memory TTL for `/v1/responses/{id}`
 - `embeddings.provider`: Embeddings provider (`deterministic/mock/builtin` built-in)
 - `claude_mapping`: Maps `fast`/`slow` suffixes to corresponding DeepSeek models (still compatible with `claude_model_mapping`)
 - `admin`: Admin panel settings (JWT expiry, password hash, etc.), hot-reloadable via Admin Settings API
-- `runtime`: Runtime parameters (concurrency limits, queue sizes), hot-reloadable via Admin Settings API; `account_max_queue=0`/`global_max_inflight=0` means auto-calculate from recommended values
+- `runtime`: Runtime parameters (concurrency limits, queue sizes, managed token refresh interval), hot-reloadable via Admin Settings API; `account_max_queue=0`/`global_max_inflight=0` means auto-calculate from recommended values, `token_refresh_interval_hours=6` is the default forced re-login interval
 - `auto_delete.sessions`: Whether to auto-delete DeepSeek sessions after request completion (default `false`, hot-reloadable via Settings)
 
 ### Environment Variables
@@ -311,6 +319,8 @@ cp opencode.json.example opencode.json
 | `DS2API_JWT_EXPIRE_HOURS` | Admin JWT TTL in hours | `24` |
 | `DS2API_CONFIG_PATH` | Config file path | `config.json` |
 | `DS2API_CONFIG_JSON` | Inline config (JSON or Base64) | — |
+| `CONFIG_JSON` | Legacy compatibility config input | — |
+| `DS2API_ENV_WRITEBACK` | Auto-write env-backed config to file and transition to file mode (`1/true/yes/on`) | Disabled |
 | `DS2API_WASM_PATH` | PoW WASM file path | Auto-detect |
 | `DS2API_STATIC_ADMIN_DIR` | Admin static assets dir | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | Auto-build WebUI on startup | Enabled locally, disabled on Vercel |
@@ -330,6 +340,8 @@ cp opencode.json.example opencode.json
 | `VERCEL_TEAM_ID` | Vercel team ID | — |
 | `DS2API_VERCEL_PROTECTION_BYPASS` | Vercel deployment protection bypass for internal Node→Go calls | — |
 
+> Note: when `DS2API_CONFIG_JSON/CONFIG_JSON` is detected, the Admin UI shows mode risk and auto-persistence status (including `DS2API_CONFIG_PATH` and mode-transition hints).
+
 ## Authentication Modes
 
 For business endpoints (`/v1/*`, `/anthropic/*`, Gemini routes), DS2API supports two modes:
@@ -340,6 +352,7 @@ For business endpoints (`/v1/*`, `/anthropic/*`, Gemini routes), DS2API supports
 | **Direct token** | If the token is not in `config.keys`, DS2API treats it as a DeepSeek token directly |
 
 Optional header `X-Ds2-Target-Account`: Pin a specific managed account (value is email or mobile).
+Gemini routes also accept `x-goog-api-key`, or `?key=` / `?api_key=` when no auth header is present.
 
 ## Concurrency Model
 
@@ -431,6 +444,7 @@ ds2api/
 ├── tests/
 │   ├── compat/              # Compatibility fixtures and expected outputs
 │   └── scripts/             # Unified test script entrypoints (unit/e2e)
+├── docs/                    # Deployment / contributing / testing docs
 ├── static/admin/            # WebUI build output (not committed to Git)
 ├── .github/
 │   ├── workflows/           # GitHub Actions (quality gates + release automation)
@@ -450,9 +464,9 @@ ds2api/
 | Document | Description |
 | --- | --- |
 | [API.md](API.md) / [API.en.md](API.en.md) | API reference with request/response examples |
-| [DEPLOY.md](DEPLOY.md) / [DEPLOY.en.md](DEPLOY.en.md) | Deployment guide (local/Docker/Vercel/systemd) |
-| [CONTRIBUTING.md](CONTRIBUTING.md) / [CONTRIBUTING.en.md](CONTRIBUTING.en.md) | Contributing guide |
-| [TESTING.md](TESTING.md) | Testsuite guide |
+| [DEPLOY.md](docs/DEPLOY.md) / [DEPLOY.en.md](docs/DEPLOY.en.md) | Deployment guide (local/Docker/Vercel/systemd) |
+| [CONTRIBUTING.md](docs/CONTRIBUTING.md) / [CONTRIBUTING.en.md](docs/CONTRIBUTING.en.md) | Contributing guide |
+| [TESTING.md](docs/TESTING.md) | Testsuite guide |
 
 ## Testing
 
@@ -491,4 +505,7 @@ Workflow: `.github/workflows/release-artifacts.yml`
 
 ## Disclaimer
 
-This project is built through reverse engineering and is provided for learning and research only. Stability is not guaranteed. Do not use it in scenarios that violate terms of service or laws.
+This project is built through reverse engineering and is provided for learning, research, personal experimentation, and internal validation only. No commercial authorization is granted, and no warranty of stability, fitness, or results is provided.
+The author and repository maintainers are not responsible for any direct or indirect loss, account suspension, data loss, legal risk, or third-party claims arising from use, modification, distribution, deployment, or reliance on this project.
+
+Do not use this project in ways that violate service terms, agreements, laws, or platform rules. Before any commercial use, review the `LICENSE`, the relevant terms, and confirm that you have the author's written permission.

VERSION

@@ -1 +1 @@
-2.4.1
+2.5.1

docs/CONTRIBUTING.en.md

@@ -70,6 +70,7 @@ docker-compose -f docker-compose.dev.yml up
 5. Open a Pull Request
 
 > 💡 If you modify files under `webui/`, no manual build is needed — CI handles it automatically.
+> If you want to verify the generated `static/admin/` assets locally, you can still run `./scripts/build-webui.sh`.
 
 ## Build WebUI
 
@@ -129,6 +130,8 @@ ds2api/
 │       ├── components/      # Shared components
 │       └── locales/         # Language packs
 ├── scripts/                 # Build and test scripts
+├── tests/                   # Unit tests, Node tests, and end-to-end tests
+├── plans/                   # Plans, gates, and manual smoke-test records
 ├── static/admin/            # WebUI build output (not committed)
 ├── Dockerfile               # Multi-stage build
 ├── docker-compose.yml       # Production

docs/CONTRIBUTING.md

@@ -70,6 +70,7 @@ docker-compose -f docker-compose.dev.yml up
 5. 发起 Pull Request
 
 > 💡 如果修改了 `webui/` 目录下的文件，无需手动构建——CI 会自动处理。
+> 但如果你本地想验证 `static/admin/` 产物，还是可以手动运行 `./scripts/build-webui.sh`。
 
 ## WebUI 构建
 
@@ -129,6 +130,8 @@ ds2api/
 │       ├── components/      # 通用组件
 │       └── locales/         # 语言包
 ├── scripts/                 # 构建与测试脚本
+├── tests/                   # 单元测试、Node 测试与端到端测试
+├── plans/                   # 计划、门禁和手工烟测记录
 ├── static/admin/            # WebUI 构建产物（不提交）
 ├── Dockerfile               # 多阶段构建
 ├── docker-compose.yml       # 生产环境

docs/DEPLOY.en.md

@@ -32,6 +32,7 @@ Config source (choose one):
 
 - **File**: `config.json` (recommended for local/Docker)
 - **Environment variable**: `DS2API_CONFIG_JSON` (recommended for Vercel; supports raw JSON or Base64)
+- Compatibility note: `CONFIG_JSON` is the legacy fallback variable; `DS2API_CONFIG_JSON` may also contain raw JSON directly
 
 Unified recommendation (best practice):
 
@@ -69,7 +70,7 @@ Default address: `http://0.0.0.0:5001` (override with `PORT`).
 
 ### 1.2 WebUI Build
 
-On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm).
+On first local startup, if `static/admin/` is missing, DS2API will automatically attempt to build the WebUI (requires Node.js/npm; when dependencies are missing it runs `npm ci` first, then `npm run build -- --outDir static/admin --emptyOutDir`).
 
 Manual build:
 
@@ -123,6 +124,8 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
+The default `docker-compose.yml` maps host port `6011` to container port `5001`. If you want `5001` exposed directly, adjust the `ports` mapping.
+
 ### 2.2 Update
 
 ```bash
@@ -245,6 +248,7 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # optional for personal accounts
 | `DS2API_ACCOUNT_QUEUE_SIZE` | Alias (legacy compat) | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | Global inflight limit | `recommended_concurrency` |
 | `DS2API_MAX_INFLIGHT` | Alias (legacy compat) | — |
+| `DS2API_ENV_WRITEBACK` | When `DS2API_CONFIG_JSON` is present, auto-write to `DS2API_CONFIG_PATH` and switch to file-backed mode after success (`1/true/yes/on`) | Disabled |
 | `DS2API_VERCEL_INTERNAL_SECRET` | Hybrid streaming internal auth | Falls back to `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | Stream lease TTL | `900` |
 | `VERCEL_TOKEN` | Vercel sync token | — |
@@ -453,8 +457,8 @@ server {
 # Copy compiled binary and related files to target directory
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# Optional: if you want to use an external WASM file (override embedded one)
-# sudo cp sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
+# Optional: if you want to use an external WASM file (override the embedded one, from a release package or build output)
+# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DEPLOY.md

@@ -32,6 +32,7 @@
 
 - **文件方式**：`config.json`（推荐本地/Docker 使用）
 - **环境变量方式**：`DS2API_CONFIG_JSON`（推荐 Vercel 使用，支持 JSON 字符串或 Base64 编码）
+- 兼容写法：`CONFIG_JSON` 是旧版回退变量；`DS2API_CONFIG_JSON` 也可以直接写原始 JSON
 
 统一建议（最优实践）：
 
@@ -69,7 +70,7 @@ go run ./cmd/ds2api
 
 ### 1.2 WebUI 构建
 
-本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm）。
+本地首次启动时，若 `static/admin/` 不存在，服务会自动尝试构建 WebUI（需要 Node.js/npm；缺依赖时会先执行 `npm ci`，再执行 `npm run build -- --outDir static/admin --emptyOutDir`）。
 
 你也可以手动构建：
 
@@ -123,6 +124,8 @@ docker-compose up -d
 docker-compose logs -f
 ```
 
+默认 `docker-compose.yml` 会把宿主机 `6011` 映射到容器内的 `5001`。如果你希望直接对外暴露 `5001`，请调整 `ports` 配置。
+
 ### 2.2 更新
 
 ```bash
@@ -245,6 +248,7 @@ VERCEL_TEAM_ID=team_xxxxxxxxxxxx   # 个人账号可留空
 | `DS2API_ACCOUNT_QUEUE_SIZE` | 同上（兼容别名） | — |
 | `DS2API_GLOBAL_MAX_INFLIGHT` | 全局并发上限 | `recommended_concurrency` |
 | `DS2API_MAX_INFLIGHT` | 同上（兼容别名） | — |
+| `DS2API_ENV_WRITEBACK` | 检测到 `DS2API_CONFIG_JSON` 时自动写入 `DS2API_CONFIG_PATH`，并在成功后转为文件模式（`1/true/yes/on`） | 关闭 |
 | `DS2API_VERCEL_INTERNAL_SECRET` | 混合流式内部鉴权 | 回退用 `DS2API_ADMIN_KEY` |
 | `DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS` | 流式 lease TTL | `900` |
 | `VERCEL_TOKEN` | Vercel 同步 token | — |
@@ -453,8 +457,8 @@ server {
 # 将编译好的二进制文件和相关文件复制到目标目录
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# 可选：若你希望使用外置 WASM 文件（覆盖内置版本）
-# sudo cp sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
+# 可选：若你希望使用外置 WASM 文件（覆盖内置版本，来自 release 包或构建产物）
+# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/TESTING.md

@@ -1,6 +1,6 @@
 # DS2API 测试指南
 
-语言 / Language: [中文 + English](TESTING.md)
+语言 / Language: 中文 + English（同页）
 
 ## 概述 | Overview
 
@@ -14,6 +14,7 @@ DS2API 提供两个层级的测试：
 | 端到端测试 | `./tests/scripts/run-live.sh` | 使用真实账号执行全链路测试 |
 
 端到端测试集会录制完整的请求/响应日志，用于故障排查。
+Node 单元测试脚本会先做 `node --check` 语法门禁，再以 `--test-concurrency=1` 串行执行测试文件，减少模块级共享状态带来的干扰。
 
 ---
 
@@ -66,6 +67,8 @@ DS2API 提供两个层级的测试：
 
 4. **结果收集**：继续执行所有用例（不中断），写入最终汇总
 
+如果你只想跳过这些 preflight 检查，可以直接运行 `go run ./cmd/ds2api-tests --no-preflight`。
+
 ---
 
 ## CLI 参数 | CLI Flags

docs/toolcall-semantics.md

@@ -1,41 +1,72 @@
-# Tool call parsing semantics (Go canonical spec)
+# Tool call parsing semantics（Go/Node 统一语义）
 
-This document defines the cross-runtime contract for `ParseToolCallsDetailed` / `parseToolCallsDetailed`.
+本文档描述当前代码中 `ParseToolCallsDetailed` / `parseToolCallsDetailed` 的**实际行为**，用于对齐 Go 与 Node Runtime。
 
-## Output contract
+## 1) 输出结构（当前实现）
 
-- `calls`: accepted tool calls with normalized tool names.
-- `sawToolCallSyntax`: true when tool-call-like syntax is detected (`tool_calls`, `<tool_call>`, `<function_call>`, `<invoke>`) or a valid call is parsed.
-- `rejectedByPolicy`: true when parser extracted call syntax but all calls are rejected by allow-list policy.
-- `rejectedToolNames`: de-duplicated rejected tool names in first-seen order.
+- `calls`：解析得到的工具调用列表（`name` + `input`）。
+- `sawToolCallSyntax`：检测到工具调用语法特征时为 `true`（例如 `tool_calls`、`<tool_call>`、`<function_call>`、`<invoke>`、`function.name:`）。
+- `rejectedByPolicy`：当前实现固定为 `false`（预留字段，尚未启用 allow-list 拒绝）。
+- `rejectedToolNames`：当前实现固定为空数组（预留字段）。
 
-## Parse pipeline
+> 说明：`filterToolCallsDetailed` 当前仅做结构清洗，不做工具名策略拒绝。
 
-1. Strip fenced code blocks for non-standalone parsing.
-2. Build candidates from:
-   - full text,
-   - fenced JSON snippets,
-   - extracted JSON objects around `tool_calls`,
-   - first `{` to last `}` object slice.
-3. Parse each candidate in order:
-   - JSON payload parser (`tool_calls`, list, single call object),
-   - XML/Markup parser (`<tool_call>`, `<function_call>`, `<invoke>`; supports attributes + nested fields),
-   - Text KV fallback parser (`function.name: <name>` ... `function.arguments: {json}`).
-4. Stop at first candidate that yields at least one call.
+## 2) 解析管线
 
-## Name normalization policy
+1. **示例保护**：若判定为 fenced code block 示例上下文，则跳过执行型解析。
+2. **候选片段构建**：从完整文本中构建候选（原文、围绕 `tool_calls` 的 JSON 片段、首尾大括号切片等）。
+3. **按序尝试解析（命中即停）**：
+   - 对“明显 JSON 工具载荷候选”（以 `{`/`[` 开头且包含 `tool_calls`/`\"function\"`）先走 JSON 解析，避免 JSON 字符串内偶发 XML 片段误命中；
+   - 其余候选优先 XML 解析（`<tool_call>` / `<function_call>` / `<invoke>` / `tool_use` / `antml:function_call` 等）；
+   - JSON 解析（`{"tool_calls": [...]}`、列表、单对象）；
+   - Markup 解析；
+   - Text-KV 回退（如 `function.name:` + `function.arguments:`）。
+4. **兜底**：候选全部失败后，再对全文做 XML / Text-KV 回退。
 
-When matching parsed names against configured tools:
+## 3) XML 能力边界（当前）
 
-1. exact match,
-2. case-insensitive match,
-3. namespace tail match (`a.b.c` => `c`),
-4. loose alnum match (remove non `[a-z0-9]`, compare).
+当前已支持输入端的“多 XML/标记风格”解析，包括但不限于：
 
-## Standalone mode
+- `<tool_call><tool_name>...</tool_name><parameters>...</parameters></tool_call>`
+- `<function_call>tool</function_call><function parameter name="x">...</function parameter>`
+- `<invoke name="tool"><parameter name="x">...</parameter></invoke>`
+- `antml:function_call` / `antml:argument` / `antml:parameters`
+- `tool_use` 家族标签
 
-Standalone mode (`ParseStandaloneToolCallsDetailed`) parses the whole input directly (no candidate slicing), while still applying:
+但**输出端仍统一转换为 OpenAI 兼容 JSON 事件/对象**（`message.tool_calls`、`delta.tool_calls`、`response.function_call_arguments.*`）。
 
-- example-context guard,
-- JSON then markup fallback,
-- the same allow-list normalization policy.
+## 4) 关于“是否可以封装成 XML 再喂给模型”
+
+结论：**可以做，而且当前解析器已经能兼容 XML 作为输入格式之一**，但代码里并没有 `toolcall.prefer_xml_output` 这个开关。现有可调配置只有：
+
+- `toolcall.mode`：`feature_match` / `off`
+- `toolcall.early_emit_confidence`：`high` / `low` / `off`
+
+推荐思路仍然是“输入兼容层 + 输出按客户端协议渲染”：
+
+1. **Prompt 约束层**：如果你要尝试 XML-first，可以在系统提示词里约束模型输出规范 XML tool block（例如 `<tool_calls><tool_call>...</tool_call></tool_calls>`）。
+2. **解析兼容层**：继续在 parser 中同时接受 JSON / XML / ANTML / invoke / text-kv。
+3. **协议归一层**：无论模型输出什么格式，统一落到内部 `ParsedToolCall`。
+4. **对外渲染层**：根据客户端请求协议渲染（OpenAI / Claude / Gemini 各自格式）。
+
+这样可以同时获得：
+
+- 减少模型端 JSON 转义/引号错误；
+- 不破坏现有 SDK / 客户端生态；
+- 逐步灰度（按模型、按租户、按请求开关）。
+
+## 5) 落地建议（低风险迭代）
+
+- 继续使用现有的 `toolcall.mode=feature_match` 和 `toolcall.early_emit_confidence=high` 作为默认策略。
+- 如果要试 XML-first，把它放在 prompt 层或上游模板层，不要假设代码里已有专门的 XML 输出开关。
+- 增加观测指标：
+  - `toolcall_parse_source`（json/xml/markup/textkv）；
+  - `toolcall_parse_success_rate`；
+  - `toolcall_malformed_rate`；
+  - `toolcall_repair_rate`。
+- 先在 `responses` 链路灰度，再扩展 `chat.completions`。
+
+## 6) 兼容性提醒
+
+- 上游模型若输出混合文本 + XML，仍可能出现“半结构化”噪声，需要依赖现有 sieve 增量消费策略。
+- XML 不等于安全：仍需做 tool 名、参数 schema、执行权限的服务端校验。

internal/adapter/claude/handler_util_test.go

@@ -93,8 +93,11 @@ func TestNormalizeClaudeMessagesToolUseToAssistantToolCalls(t *testing.T) {
 		t.Fatalf("expected call id preserved, got %#v", call)
 	}
 	content, _ := m["content"].(string)
-	if !containsStr(content, "search_web") || !containsStr(content, `"arguments":"{\"query\":\"latest\"}"`) {
-		t.Fatalf("expected assistant content to include serialized tool call for prompt roundtrip, got %q", content)
+	if !containsStr(content, "<tool_calls>") || !containsStr(content, "<tool_name>search_web</tool_name>") {
+		t.Fatalf("expected assistant content to include XML tool call history, got %q", content)
+	}
+	if !containsStr(content, `<parameters>{"query":"latest"}</parameters>`) {
+		t.Fatalf("expected assistant content to include serialized parameters, got %q", content)
 	}
 }
 
@@ -248,14 +251,11 @@ func TestBuildClaudeToolPromptSingleTool(t *testing.T) {
 	if !containsStr(prompt, "Search the web") {
 		t.Fatalf("expected description in prompt")
 	}
-	if !containsStr(prompt, "tool_use") {
-		t.Fatalf("expected tool_use instruction in prompt")
+	if !containsStr(prompt, "<tool_calls>") {
+		t.Fatalf("expected XML tool_calls format in prompt")
 	}
-	if containsStr(prompt, "TOOL_CALL_HISTORY") || containsStr(prompt, "TOOL_RESULT_HISTORY") {
-		t.Fatalf("expected legacy tool history markers removed from prompt")
-	}
-	if !containsStr(prompt, "Do not print tool-call JSON in text") {
-		t.Fatalf("expected prompt to keep no tool-call-json instruction")
+	if !containsStr(prompt, "TOOL CALL FORMAT") {
+		t.Fatalf("expected tool call format header in prompt")
 	}
 }
 
@@ -301,12 +301,9 @@ func TestBuildClaudeToolPromptSupportsOpenAIStyleFunctionTool(t *testing.T) {
 func TestBuildClaudeToolPromptSkipsNonMap(t *testing.T) {
 	tools := []any{"not a map"}
 	prompt := buildClaudeToolPrompt(tools)
-	if prompt == "" {
-		t.Fatal("expected non-empty prompt even with invalid tools")
-	}
-	// Should still contain the intro and instruction
-	if !containsStr(prompt, "You are Claude") {
-		t.Fatalf("expected intro in prompt")
+	// No valid tools → empty prompt
+	if prompt != "" {
+		t.Fatalf("expected empty prompt for non-map tools, got: %q", prompt)
 	}
 }
 

internal/adapter/claude/handler_utils.go

@@ -4,6 +4,9 @@ import (
 	"encoding/json"
 	"fmt"
 	"strings"
+
+	"ds2api/internal/prompt"
+	"ds2api/internal/util"
 )
 
 func normalizeClaudeMessages(messages []any) []any {
@@ -70,22 +73,27 @@ func normalizeClaudeMessages(messages []any) []any {
 }
 
 func buildClaudeToolPrompt(tools []any) string {
-	parts := []string{"You are Claude, a helpful AI assistant. You have access to these tools:"}
+	toolSchemas := make([]string, 0, len(tools))
+	names := make([]string, 0, len(tools))
 	for _, t := range tools {
 		m, ok := t.(map[string]any)
 		if !ok {
 			continue
 		}
 		name, desc, schemaObj := extractClaudeToolMeta(m)
+		if name == "" {
+			continue
+		}
+		names = append(names, name)
 		schema, _ := json.Marshal(schemaObj)
-		parts = append(parts, fmt.Sprintf("Tool: %s\nDescription: %s\nParameters: %s", name, desc, schema))
+		toolSchemas = append(toolSchemas, fmt.Sprintf("Tool: %s\nDescription: %s\nParameters: %s", name, desc, schema))
 	}
-	parts = append(parts,
-		"When you need a tool, respond with Claude-native tool use (tool_use) using the provided tool schema. Do not print tool-call JSON in text.",
-		"Tool roundtrip context is included directly in the conversation messages (assistant tool_use/tool_calls and tool results).",
-		"After receiving a valid tool result, continue with final answer instead of repeating the same call unless required fields are still missing.",
-	)
-	return strings.Join(parts, "\n\n")
+	if len(toolSchemas) == 0 {
+		return ""
+	}
+	return "You have access to these tools:\n\n" +
+		strings.Join(toolSchemas, "\n\n") + "\n\n" +
+		util.BuildToolCallInstructions(names)
 }
 
 func formatClaudeToolResultForPrompt(block map[string]any) string {
@@ -146,7 +154,7 @@ func normalizeClaudeToolUseToAssistant(block map[string]any) map[string]any {
 	}
 	return map[string]any{
 		"role":       "assistant",
-		"content":    marshalCompactJSON(toolCalls),
+		"content":    prompt.FormatToolCallsForPrompt(toolCalls),
 		"tool_calls": toolCalls,
 	}
 }

internal/adapter/claude/stream_runtime_finalize.go

@@ -1,6 +1,7 @@
 package claude
 
 import (
+	"encoding/json"
 	"fmt"
 	"time"
 
@@ -60,9 +61,20 @@ func (s *claudeStreamRuntime) finalize(stopReason string) {
 						"type":  "tool_use",
 						"id":    fmt.Sprintf("toolu_%d_%d", time.Now().Unix(), idx),
 						"name":  tc.Name,
-						"input": tc.Input,
+						"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,

internal/adapter/openai/chat_stream_runtime.go

@@ -97,7 +97,7 @@ func (s *chatStreamRuntime) sendDone() {
 
 func (s *chatStreamRuntime) finalize(finishReason string) {
 	finalThinking := s.thinking.String()
-	finalText := sanitizeLeakedToolHistory(s.text.String())
+	finalText := sanitizeLeakedOutput(s.text.String())
 	detected := util.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
 	if len(detected.Calls) > 0 && !s.toolCallsDoneEmitted {
 		finishReason = "tool_calls"
@@ -141,7 +141,7 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 			if evt.Content == "" {
 				continue
 			}
-			cleaned := sanitizeLeakedToolHistory(evt.Content)
+			cleaned := sanitizeLeakedOutput(evt.Content)
 			if cleaned == "" {
 				continue
 			}
@@ -250,7 +250,7 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 						continue
 					}
 					if evt.Content != "" {
-						cleaned := sanitizeLeakedToolHistory(evt.Content)
+						cleaned := sanitizeLeakedOutput(evt.Content)
 						if cleaned == "" {
 							continue
 						}

internal/adapter/openai/handler_chat.go

@@ -105,7 +105,7 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, re
 	result := sse.CollectStream(resp, thinkingEnabled, true)
 
 	finalThinking := result.Thinking
-	finalText := sanitizeLeakedToolHistory(result.Text)
+	finalText := sanitizeLeakedOutput(result.Text)
 	respBody := openaifmt.BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText, toolNames)
 	writeJSON(w, http.StatusOK, respBody)
 }

internal/adapter/openai/handler_toolcall_format.go

@@ -53,7 +53,7 @@ func injectToolPrompt(messages []map[string]any, tools []any, policy util.ToolCh
 	if len(toolSchemas) == 0 {
 		return messages, names
 	}
-	toolPrompt := "You have access to these tools:\n\n" + strings.Join(toolSchemas, "\n\n") + "\n\nWhen you need to use tools, output ONLY this JSON object format:\n{\"tool_calls\": [{\"name\": \"tool_name\", \"input\": {\"param\": \"value\"}}]}\n\n【EXAMPLE】\nUser: Please check the weather in Beijing and Shanghai, and update my todo list.\nAssistant:\n{\"tool_calls\": [\n  {\"name\": \"get_weather\", \"input\": {\"city\": \"Beijing\"}},\n  {\"name\": \"get_weather\", \"input\": {\"city\": \"Shanghai\"}},\n  {\"name\": \"update_todo\", \"input\": {\"todos\": [{\"content\": \"Buy milk\"}, {\"content\": \"Write report\"}]}}\n]}\n\nIMPORTANT:\n1) If calling tools, output ONLY the JSON object above. Do NOT include any extra text.\n2) Do NOT wrap tool-call JSON in markdown/code fences (for example, do not use triple backticks).\n3) After receiving a tool result, you MUST use it to produce the final answer.\n4) Only call another tool when the previous result is missing required data or returned an error.\n5) JSON SYNTAX STRICTLY REQUIRED: All property names MUST be enclosed in double quotes (e.g., \"name\", not name).\n6) ARRAY FORMAT: If providing a list of items, you MUST enclose them in square brackets `[]` (e.g., \"todos\": [{\"item\": \"a\"}, {\"item\": \"b\"}]). DO NOT output comma-separated objects without brackets."
+	toolPrompt := "You have access to these tools:\n\n" + strings.Join(toolSchemas, "\n\n") + "\n\n" + buildToolCallInstructions(names)
 	if policy.Mode == util.ToolChoiceRequired {
 		toolPrompt += "\n7) For this response, you MUST call at least one tool from the allowed list."
 	}
@@ -73,6 +73,11 @@ func injectToolPrompt(messages []map[string]any, tools []any, policy util.ToolCh
 	return messages, names
 }
 
+// buildToolCallInstructions delegates to the shared util implementation.
+func buildToolCallInstructions(toolNames []string) string {
+	return util.BuildToolCallInstructions(toolNames)
+}
+
 func formatIncrementalStreamToolCallDeltas(deltas []toolCallDelta, ids map[int]string) []map[string]any {
 	if len(deltas) == 0 {
 		return nil

internal/adapter/openai/handler_toolcall_policy.go

@@ -1,19 +1,9 @@
 package openai
 
-import "strings"
-
 func (h *Handler) toolcallFeatureMatchEnabled() bool {
-	if h == nil || h.Store == nil {
-		return true
-	}
-	mode := strings.TrimSpace(strings.ToLower(h.Store.ToolcallMode()))
-	return mode == "" || mode == "feature_match"
+	return true
 }
 
 func (h *Handler) toolcallEarlyEmitHighConfidence() bool {
-	if h == nil || h.Store == nil {
-		return true
-	}
-	level := strings.TrimSpace(strings.ToLower(h.Store.ToolcallEarlyEmitConfidence()))
-	return level == "" || level == "high"
+	return true
 }

internal/adapter/openai/leaked_output_sanitize.go

@@ -0,0 +1,70 @@
+package openai
+
+import (
+	"regexp"
+)
+
+var emptyJSONFencePattern = regexp.MustCompile("(?is)```json\\s*```")
+var leakedToolCallArrayPattern = regexp.MustCompile(`(?is)\[\{\s*"function"\s*:\s*\{[\s\S]*?\}\s*,\s*"id"\s*:\s*"call[^"]*"\s*,\s*"type"\s*:\s*"function"\s*}\]`)
+var leakedToolResultBlobPattern = regexp.MustCompile(`(?is)<\s*\|\s*tool\s*\|\s*>\s*\{[\s\S]*?"tool_call_id"\s*:\s*"call[^"]*"\s*}`)
+
+// leakedMetaMarkerPattern matches DeepSeek special tokens in BOTH forms:
+//   - ASCII underscore: <｜end_of_sentence｜>
+//   - U+2581 variant:   <｜end▁of▁sentence｜>  (used in some DeepSeek outputs)
+var leakedMetaMarkerPattern = regexp.MustCompile(`(?i)<[｜\|]\s*(?:assistant|tool|end[_▁]of[_▁]sentence|end[_▁]of[_▁]thinking)\s*[｜\|]>`)
+
+// leakedAgentXMLBlockPatterns catch agent-style XML blocks that leak through
+// when the sieve fails to capture them. These are applied only to complete
+// wrapper blocks so standalone "<result>" examples in normal output remain
+// untouched.
+var leakedAgentXMLBlockPatterns = []*regexp.Regexp{
+	regexp.MustCompile(`(?is)<attempt_completion\b[^>]*>(.*?)</attempt_completion>`),
+	regexp.MustCompile(`(?is)<ask_followup_question\b[^>]*>(.*?)</ask_followup_question>`),
+	regexp.MustCompile(`(?is)<new_task\b[^>]*>(.*?)</new_task>`),
+}
+
+var leakedAgentWrapperTagPattern = regexp.MustCompile(`(?is)</?(?:attempt_completion|ask_followup_question|new_task)\b[^>]*>`)
+var leakedAgentWrapperPlusResultOpenPattern = regexp.MustCompile(`(?is)<(?:attempt_completion|ask_followup_question|new_task)\b[^>]*>\s*<result>`)
+var leakedAgentResultPlusWrapperClosePattern = regexp.MustCompile(`(?is)</result>\s*</(?:attempt_completion|ask_followup_question|new_task)\b[^>]*>`)
+var leakedAgentResultTagPattern = regexp.MustCompile(`(?is)</?result>`)
+
+func sanitizeLeakedOutput(text string) string {
+	if text == "" {
+		return text
+	}
+	out := emptyJSONFencePattern.ReplaceAllString(text, "")
+	out = leakedToolCallArrayPattern.ReplaceAllString(out, "")
+	out = leakedToolResultBlobPattern.ReplaceAllString(out, "")
+	out = leakedMetaMarkerPattern.ReplaceAllString(out, "")
+	out = sanitizeLeakedAgentXMLBlocks(out)
+	return out
+}
+
+func sanitizeLeakedAgentXMLBlocks(text string) string {
+	out := text
+	for _, pattern := range leakedAgentXMLBlockPatterns {
+		out = pattern.ReplaceAllStringFunc(out, func(match string) string {
+			submatches := pattern.FindStringSubmatch(match)
+			if len(submatches) < 2 {
+				return match
+			}
+			// Preserve the inner text so leaked agent instructions do not erase
+			// the actual answer, but strip the wrapper/result markup itself.
+			return leakedAgentResultTagPattern.ReplaceAllString(submatches[1], "")
+		})
+	}
+	// Fallback for truncated output streams: strip any dangling wrapper tags
+	// that were not part of a complete block replacement. If we detect leaked
+	// wrapper tags, strip only adjacent <result> tags to avoid exposing agent
+	// markup without altering unrelated user-visible <result> examples.
+	if leakedAgentWrapperTagPattern.MatchString(out) {
+		out = leakedAgentWrapperPlusResultOpenPattern.ReplaceAllStringFunc(out, func(match string) string {
+			return leakedAgentResultTagPattern.ReplaceAllString(match, "")
+		})
+		out = leakedAgentResultPlusWrapperClosePattern.ReplaceAllStringFunc(out, func(match string) string {
+			return leakedAgentResultTagPattern.ReplaceAllString(match, "")
+		})
+		out = leakedAgentWrapperTagPattern.ReplaceAllString(out, "")
+	}
+	return out
+}

internal/adapter/openai/leaked_output_sanitize_test.go

@@ -0,0 +1,68 @@
+package openai
+
+import "testing"
+
+func TestSanitizeLeakedOutputRemovesEmptyJSONFence(t *testing.T) {
+	raw := "before\n```json\n```\nafter"
+	got := sanitizeLeakedOutput(raw)
+	if got != "before\n\nafter" {
+		t.Fatalf("unexpected sanitized empty json fence: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesLeakedWireToolCallAndResult(t *testing.T) {
+	raw := "开始\n[{\"function\":{\"arguments\":\"{\\\"command\\\":\\\"java -version\\\"}\",\"name\":\"exec\"},\"id\":\"callb9a321\",\"type\":\"function\"}]< | Tool | >{\"content\":\"openjdk version 21\",\"tool_call_id\":\"callb9a321\"}\n结束"
+	got := sanitizeLeakedOutput(raw)
+	if got != "开始\n\n结束" {
+		t.Fatalf("unexpected sanitize result for leaked wire format: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesStandaloneMetaMarkers(t *testing.T) {
+	raw := "A<| end_of_sentence |><| Assistant |>B<| end_of_thinking |>C<｜end▁of▁thinking｜>D<｜end▁of▁sentence｜>E"
+	got := sanitizeLeakedOutput(raw)
+	if got != "ABCDE" {
+		t.Fatalf("unexpected sanitize result for meta markers: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesAgentXMLLeaks(t *testing.T) {
+	raw := "Done.<attempt_completion><result>Some final answer</result></attempt_completion>"
+	got := sanitizeLeakedOutput(raw)
+	if got != "Done.Some final answer" {
+		t.Fatalf("unexpected sanitize result for agent XML leak: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputPreservesStandaloneResultTags(t *testing.T) {
+	raw := "Example XML: <result>value</result>"
+	got := sanitizeLeakedOutput(raw)
+	if got != raw {
+		t.Fatalf("unexpected sanitize result for standalone result tag: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesDanglingAgentXMLOpeningTags(t *testing.T) {
+	raw := "Done.<attempt_completion><result>Some final answer"
+	got := sanitizeLeakedOutput(raw)
+	if got != "Done.Some final answer" {
+		t.Fatalf("unexpected sanitize result for dangling opening tags: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputRemovesDanglingAgentXMLClosingTags(t *testing.T) {
+	raw := "Done.Some final answer</result></attempt_completion>"
+	got := sanitizeLeakedOutput(raw)
+	if got != "Done.Some final answer" {
+		t.Fatalf("unexpected sanitize result for dangling closing tags: %q", got)
+	}
+}
+
+func TestSanitizeLeakedOutputPreservesUnrelatedResultTagsWhenWrapperLeaks(t *testing.T) {
+	raw := "Done.<attempt_completion><result>Some final answer\nExample XML: <result>value</result>"
+	got := sanitizeLeakedOutput(raw)
+	want := "Done.Some final answer\nExample XML: <result>value</result>"
+	if got != want {
+		t.Fatalf("unexpected sanitize result for mixed leaked wrapper + xml example: %q", got)
+	}
+}

internal/adapter/openai/message_normalize.go

@@ -1,7 +1,6 @@
 package openai
 
 import (
-	"encoding/json"
 	"strings"
 
 	"ds2api/internal/prompt"
@@ -55,7 +54,18 @@ func normalizeOpenAIMessagesForPrompt(raw []any, traceID string) []map[string]an
 }
 
 func buildAssistantContentForPrompt(msg map[string]any) string {
-	return strings.TrimSpace(normalizeOpenAIContentForPrompt(msg["content"]))
+	content := strings.TrimSpace(normalizeOpenAIContentForPrompt(msg["content"]))
+	toolHistory := prompt.FormatToolCallsForPrompt(msg["tool_calls"])
+	switch {
+	case content == "" && toolHistory == "":
+		return ""
+	case content == "":
+		return toolHistory
+	case toolHistory == "":
+		return content
+	default:
+		return content + "\n\n" + toolHistory
+	}
 }
 
 func buildToolContentForPrompt(msg map[string]any) string {
@@ -70,18 +80,6 @@ func normalizeOpenAIContentForPrompt(v any) string {
 	return prompt.NormalizeContent(v)
 }
 
-func normalizeToolArgumentString(raw string) string {
-	trimmed := strings.TrimSpace(raw)
-	if trimmed == "" {
-		return ""
-	}
-	if looksLikeConcatenatedJSON(trimmed) {
-		// Keep original payload to avoid silent argument rewrites.
-		return raw
-	}
-	return trimmed
-}
-
 func normalizeOpenAIRoleForPrompt(role string) string {
 	role = strings.ToLower(strings.TrimSpace(role))
 	if role == "developer" {
@@ -96,20 +94,3 @@ func asString(v any) string {
 	}
 	return ""
 }
-
-func looksLikeConcatenatedJSON(raw string) bool {
-	trimmed := strings.TrimSpace(raw)
-	if trimmed == "" {
-		return false
-	}
-	if strings.Contains(trimmed, "}{") || strings.Contains(trimmed, "][") {
-		return true
-	}
-	dec := json.NewDecoder(strings.NewReader(trimmed))
-	var first any
-	if err := dec.Decode(&first); err != nil {
-		return false
-	}
-	var second any
-	return dec.Decode(&second) == nil
-}

internal/adapter/openai/message_normalize_test.go

@@ -34,20 +34,23 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsAndToolResult(t *tes
 	}
 
 	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
-	if len(normalized) != 3 {
-		t.Fatalf("expected 3 normalized messages with tool-call-only assistant turn omitted, got %d", len(normalized))
+	if len(normalized) != 4 {
+		t.Fatalf("expected 4 normalized messages with assistant tool history preserved, got %d", len(normalized))
 	}
-	toolContent, _ := normalized[2]["content"].(string)
-	if !strings.Contains(toolContent, `"temp":18`) {
-		t.Fatalf("tool result should be transparently forwarded, got %q", toolContent)
+	assistantContent, _ := normalized[2]["content"].(string)
+	if !strings.Contains(assistantContent, "<tool_calls>") {
+		t.Fatalf("assistant tool history should be preserved in XML form, got %q", assistantContent)
 	}
-	if strings.Contains(toolContent, "[TOOL_RESULT_HISTORY]") {
-		t.Fatalf("tool history marker should not be injected: %q", toolContent)
+	if !strings.Contains(assistantContent, "<tool_name>get_weather</tool_name>") {
+		t.Fatalf("expected tool name in preserved history, got %q", assistantContent)
+	}
+	if !strings.Contains(normalized[3]["content"].(string), `"temp":18`) {
+		t.Fatalf("tool result should be transparently forwarded, got %#v", normalized[3]["content"])
 	}
 
 	prompt := util.MessagesPrepare(normalized)
-	if strings.Contains(prompt, "[TOOL_CALL_HISTORY]") || strings.Contains(prompt, "[TOOL_RESULT_HISTORY]") {
-		t.Fatalf("expected no synthetic history markers in prompt: %q", prompt)
+	if !strings.Contains(prompt, "<tool_calls>") {
+		t.Fatalf("expected preserved assistant tool history in prompt: %q", prompt)
 	}
 }
 
@@ -170,8 +173,15 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantMultipleToolCallsRemainSepara
 	}
 
 	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
-	if len(normalized) != 0 {
-		t.Fatalf("expected assistant tool_call-only message omitted, got %#v", normalized)
+	if len(normalized) != 1 {
+		t.Fatalf("expected assistant tool_call-only message preserved, got %#v", normalized)
+	}
+	content, _ := normalized[0]["content"].(string)
+	if strings.Count(content, "<tool_call>") != 2 {
+		t.Fatalf("expected two preserved tool call blocks, got %q", content)
+	}
+	if !strings.Contains(content, "<tool_name>search_web</tool_name>") || !strings.Contains(content, "<tool_name>eval_javascript</tool_name>") {
+		t.Fatalf("expected both tool names in preserved history, got %q", content)
 	}
 }
 
@@ -192,8 +202,12 @@ func TestNormalizeOpenAIMessagesForPrompt_PreservesConcatenatedToolArguments(t *
 	}
 
 	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
-	if len(normalized) != 0 {
-		t.Fatalf("expected assistant tool_call-only content omitted, got %#v", normalized)
+	if len(normalized) != 1 {
+		t.Fatalf("expected assistant tool_call-only content preserved, got %#v", normalized)
+	}
+	content, _ := normalized[0]["content"].(string)
+	if !strings.Contains(content, `{}{"query":"测试工具调用"}`) {
+		t.Fatalf("expected concatenated tool arguments preserved, got %q", content)
 	}
 }
 
@@ -215,7 +229,7 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsMissingNameAreDroppe
 
 	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
 	if len(normalized) != 0 {
-		t.Fatalf("expected assistant tool_calls without text omitted, got %#v", normalized)
+		t.Fatalf("expected assistant tool_calls without text to be dropped when name is missing, got %#v", normalized)
 	}
 }
 
@@ -237,8 +251,15 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantNilContentDoesNotInjectNullLi
 	}
 
 	normalized := normalizeOpenAIMessagesForPrompt(raw, "")
-	if len(normalized) != 0 {
-		t.Fatalf("expected nil-content assistant tool_call-only message omitted, got %#v", normalized)
+	if len(normalized) != 1 {
+		t.Fatalf("expected nil-content assistant tool_call-only message preserved, got %#v", normalized)
+	}
+	content, _ := normalized[0]["content"].(string)
+	if strings.Contains(content, "null") {
+		t.Fatalf("expected no null literal injection, got %q", content)
+	}
+	if !strings.Contains(content, "<tool_calls>") {
+		t.Fatalf("expected assistant tool history in normalized content, got %q", content)
 	}
 }
 

internal/adapter/openai/prompt_build_test.go

@@ -47,8 +47,11 @@ func TestBuildOpenAIFinalPrompt_HandlerPathIncludesToolRoundtripSemantics(t *tes
 	if !strings.Contains(finalPrompt, `"condition":"sunny"`) {
 		t.Fatalf("handler finalPrompt should preserve tool output content: %q", finalPrompt)
 	}
-	if strings.Contains(finalPrompt, "[TOOL_CALL_HISTORY]") || strings.Contains(finalPrompt, "[TOOL_RESULT_HISTORY]") {
-		t.Fatalf("handler finalPrompt should not include synthetic history markers: %q", finalPrompt)
+	if !strings.Contains(finalPrompt, "<tool_calls>") {
+		t.Fatalf("handler finalPrompt should preserve assistant tool history: %q", finalPrompt)
+	}
+	if !strings.Contains(finalPrompt, "<tool_name>get_weather</tool_name>") {
+		t.Fatalf("handler finalPrompt should include tool name history: %q", finalPrompt)
 	}
 }
 
@@ -71,16 +74,19 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 	}
 
 	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "")
-	if !strings.Contains(finalPrompt, "After receiving a tool result, you MUST use it to produce the final answer.") {
+	if !strings.Contains(finalPrompt, "After receiving a tool result, use it directly.") {
 		t.Fatalf("vercel prepare finalPrompt missing final-answer instruction: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "Only call another tool when the previous result is missing required data or returned an error.") {
+	if !strings.Contains(finalPrompt, "Only call another tool if the result is insufficient.") {
 		t.Fatalf("vercel prepare finalPrompt missing retry guard instruction: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "Do NOT wrap tool-call JSON in markdown/code fences") {
-		t.Fatalf("vercel prepare finalPrompt missing no-fence instruction: %q", finalPrompt)
+	if !strings.Contains(finalPrompt, "TOOL CALL FORMAT") {
+		t.Fatalf("vercel prepare finalPrompt missing xml format instruction: %q", finalPrompt)
+	}
+	if !strings.Contains(finalPrompt, "Do NOT wrap the XML in markdown code fences") {
+		t.Fatalf("vercel prepare finalPrompt missing no-fence xml instruction: %q", finalPrompt)
 	}
 	if strings.Contains(finalPrompt, "```json") {
-		t.Fatalf("vercel prepare finalPrompt should not require fenced json tool calls: %q", finalPrompt)
+		t.Fatalf("vercel prepare finalPrompt should not require fenced tool calls: %q", finalPrompt)
 	}
 }

internal/adapter/openai/responses_handler.go

@@ -113,7 +113,7 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 		return
 	}
 	result := sse.CollectStream(resp, thinkingEnabled, true)
-	sanitizedText := sanitizeLeakedToolHistory(result.Text)
+	sanitizedText := sanitizeLeakedOutput(result.Text)
 	textParsed := util.ParseStandaloneToolCallsDetailed(sanitizedText, toolNames)
 	logResponsesToolPolicyRejection(traceID, toolChoice, textParsed, "text")
 

internal/adapter/openai/responses_input_items.go

@@ -1,11 +1,11 @@
 package openai
 
 import (
-	"encoding/json"
 	"fmt"
 	"strings"
 
 	"ds2api/internal/config"
+	"ds2api/internal/prompt"
 )
 
 func normalizeResponsesInputItem(m map[string]any) map[string]any {
@@ -148,7 +148,7 @@ func normalizeResponsesInputItemWithState(m map[string]any, callNameByID map[str
 
 		functionPayload := map[string]any{
 			"name":      name,
-			"arguments": stringifyToolCallArguments(argsRaw),
+			"arguments": prompt.StringifyToolCallArguments(argsRaw),
 		}
 		call := map[string]any{
 			"type":     "function",
@@ -211,26 +211,3 @@ func normalizeResponsesFallbackPart(m map[string]any) string {
 	}
 	return strings.TrimSpace(fmt.Sprintf("%v", m))
 }
-
-func stringifyToolCallArguments(v any) string {
-	switch x := v.(type) {
-	case nil:
-		return "{}"
-	case string:
-		s := strings.TrimSpace(x)
-		if s == "" {
-			return "{}"
-		}
-		s = normalizeToolArgumentString(s)
-		if s == "" {
-			return "{}"
-		}
-		return s
-	default:
-		b, err := json.Marshal(x)
-		if err != nil || len(b) == 0 {
-			return "{}"
-		}
-		return string(b)
-	}
-}

internal/adapter/openai/responses_stream_runtime_core.go

@@ -97,7 +97,7 @@ func newResponsesStreamRuntime(
 
 func (s *responsesStreamRuntime) finalize() {
 	finalThinking := s.thinking.String()
-	finalText := sanitizeLeakedToolHistory(s.text.String())
+	finalText := sanitizeLeakedOutput(s.text.String())
 
 	if s.bufferToolContent {
 		s.processToolStreamEvents(flushToolSieve(&s.sieve, s.toolNames), true)
@@ -194,7 +194,7 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 			continue
 		}
 
-		cleanedText := sanitizeLeakedToolHistory(p.Text)
+		cleanedText := sanitizeLeakedOutput(p.Text)
 		if cleanedText == "" {
 			continue
 		}

internal/adapter/openai/tool_history_sanitize.go

@@ -1,23 +0,0 @@
-package openai
-
-import (
-	"regexp"
-)
-
-var leakedToolHistoryPattern = regexp.MustCompile(`(?is)\[TOOL_CALL_HISTORY\][\s\S]*?\[/TOOL_CALL_HISTORY\]|\[TOOL_RESULT_HISTORY\][\s\S]*?\[/TOOL_RESULT_HISTORY\]`)
-var emptyJSONFencePattern = regexp.MustCompile("(?is)```json\\s*```")
-var leakedToolCallArrayPattern = regexp.MustCompile(`(?is)\[\{\s*"function"\s*:\s*\{[\s\S]*?\}\s*,\s*"id"\s*:\s*"call[^"]*"\s*,\s*"type"\s*:\s*"function"\s*}\]`)
-var leakedToolResultBlobPattern = regexp.MustCompile(`(?is)<\s*\|\s*tool\s*\|\s*>\s*\{[\s\S]*?"tool_call_id"\s*:\s*"call[^"]*"\s*}`)
-var leakedMetaMarkerPattern = regexp.MustCompile(`(?is)<\s*\|\s*(?:assistant|tool|end_of_sentence|end_of_thinking)\s*\|\s*>`)
-
-func sanitizeLeakedToolHistory(text string) string {
-	if text == "" {
-		return text
-	}
-	out := leakedToolHistoryPattern.ReplaceAllString(text, "")
-	out = emptyJSONFencePattern.ReplaceAllString(out, "")
-	out = leakedToolCallArrayPattern.ReplaceAllString(out, "")
-	out = leakedToolResultBlobPattern.ReplaceAllString(out, "")
-	out = leakedMetaMarkerPattern.ReplaceAllString(out, "")
-	return out
-}

internal/adapter/openai/tool_history_sanitize_test.go

@@ -1,122 +0,0 @@
-package openai
-
-import "testing"
-
-func TestSanitizeLeakedToolHistoryRemovesMarkerBlocks(t *testing.T) {
-	raw := "前缀\n[TOOL_CALL_HISTORY]\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_CALL_HISTORY]\n后缀"
-	got := sanitizeLeakedToolHistory(raw)
-	if got != "前缀\n\n后缀" {
-		t.Fatalf("unexpected sanitized content: %q", got)
-	}
-}
-
-func TestSanitizeLeakedToolHistoryPreservesChunkWhitespace(t *testing.T) {
-	cases := []struct {
-		name string
-		raw  string
-		want string
-	}{
-		{
-			name: "trailing space kept",
-			raw:  "Hello ",
-			want: "Hello ",
-		},
-		{
-			name: "leading newline kept",
-			raw:  "\nworld",
-			want: "\nworld",
-		},
-		{
-			name: "surrounding whitespace around marker is preserved",
-			raw:  "A \n[TOOL_RESULT_HISTORY]\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_RESULT_HISTORY]\n B",
-			want: "A \n\n B",
-		},
-	}
-
-	for _, tc := range cases {
-		t.Run(tc.name, func(t *testing.T) {
-			got := sanitizeLeakedToolHistory(tc.raw)
-			if got != tc.want {
-				t.Fatalf("unexpected sanitize result, want %q got %q", tc.want, got)
-			}
-		})
-	}
-}
-
-func TestSanitizeLeakedToolHistoryRemovesEmptyJSONFence(t *testing.T) {
-	raw := "before\n```json\n```\nafter"
-	got := sanitizeLeakedToolHistory(raw)
-	if got != "before\n\nafter" {
-		t.Fatalf("unexpected sanitized empty json fence: %q", got)
-	}
-}
-
-func TestFlushToolSieveDropsToolHistoryLeak(t *testing.T) {
-	var state toolStreamSieveState
-	chunk := "[TOOL_CALL_HISTORY]\nstatus: already_called\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_CALL_HISTORY]"
-	evts := processToolSieveChunk(&state, chunk, []string{"exec"})
-	if len(evts) != 0 {
-		t.Fatalf("expected no immediate output before history block is complete, got %+v", evts)
-	}
-	flushed := flushToolSieve(&state, []string{"exec"})
-	if len(flushed) != 0 {
-		t.Fatalf("expected history block to be swallowed, got %+v", flushed)
-	}
-}
-
-func TestFlushToolSieveDropsToolResultHistoryLeak(t *testing.T) {
-	var state toolStreamSieveState
-	chunk := "[TOOL_RESULT_HISTORY]\nstatus: already_called\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_RESULT_HISTORY]"
-	evts := processToolSieveChunk(&state, chunk, []string{"exec"})
-	if len(evts) != 0 {
-		t.Fatalf("expected no immediate output before result history block is complete, got %+v", evts)
-	}
-	flushed := flushToolSieve(&state, []string{"exec"})
-	if len(flushed) != 0 {
-		t.Fatalf("expected result history block to be swallowed, got %+v", flushed)
-	}
-}
-
-func TestSanitizeLeakedToolHistoryRemovesLeakedWireToolCallAndResult(t *testing.T) {
-	raw := "开始\n[{\"function\":{\"arguments\":\"{\\\"command\\\":\\\"java -version\\\"}\",\"name\":\"exec\"},\"id\":\"callb9a321\",\"type\":\"function\"}]< | Tool | >{\"content\":\"openjdk version 21\",\"tool_call_id\":\"callb9a321\"}\n结束"
-	got := sanitizeLeakedToolHistory(raw)
-	if got != "开始\n\n结束" {
-		t.Fatalf("unexpected sanitize result for leaked wire format: %q", got)
-	}
-}
-
-func TestSanitizeLeakedToolHistoryRemovesStandaloneMetaMarkers(t *testing.T) {
-	raw := "A<| end_of_sentence |><| Assistant |>B<| end_of_thinking |>C"
-	got := sanitizeLeakedToolHistory(raw)
-	if got != "ABC" {
-		t.Fatalf("unexpected sanitize result for meta markers: %q", got)
-	}
-}
-
-func TestProcessToolSieveChunkSplitsResultHistoryBoundary(t *testing.T) {
-	var state toolStreamSieveState
-	parts := []string{
-		"Hello ",
-		"[TOOL_RESULT_HISTORY]\nstatus: already_called\n",
-		"function.name: exec\nfunction.arguments: {}\n[/TOOL_RESULT_HISTORY]",
-		"world",
-	}
-	var events []toolStreamEvent
-	for _, p := range parts {
-		events = append(events, processToolSieveChunk(&state, p, []string{"exec"})...)
-	}
-	events = append(events, flushToolSieve(&state, []string{"exec"})...)
-
-	var text string
-	for _, evt := range events {
-		if evt.Content != "" {
-			text += evt.Content
-		}
-		if len(evt.ToolCalls) > 0 {
-			t.Fatalf("did not expect parsed tool calls from history leak: %+v", evt.ToolCalls)
-		}
-	}
-	if text != "Hello world" {
-		t.Fatalf("expected clean text output preserving boundary spaces, got %q", text)
-	}
-}

internal/adapter/openai/tool_sieve_core.go

@@ -114,8 +114,14 @@ func flushToolSieve(state *toolStreamSieveState, toolNames []string) []toolStrea
 		} else {
 			content := state.capture.String()
 			if content != "" {
-				state.noteText(content)
-				events = append(events, toolStreamEvent{Content: content})
+				// If the captured text looks like an incomplete XML tool call block,
+				// swallow it to prevent leaking raw XML tags to the client.
+				if hasOpenXMLToolTag(content) {
+					// Drop it silently — incomplete tool call.
+				} else {
+					state.noteText(content)
+					events = append(events, toolStreamEvent{Content: content})
+				}
 			}
 		}
 		state.capture.Reset()
@@ -124,8 +130,14 @@ func flushToolSieve(state *toolStreamSieveState, toolNames []string) []toolStrea
 	}
 	if state.pending.Len() > 0 {
 		content := state.pending.String()
-		state.noteText(content)
-		events = append(events, toolStreamEvent{Content: content})
+		// Safety: if pending contains XML tool tag fragments (e.g. "tool_calls>"
+		// from a split closing tag), swallow them instead of leaking.
+		if hasOpenXMLToolTag(content) || looksLikeXMLToolTagFragment(content) {
+			// Drop it — likely an incomplete tool call fragment.
+		} else {
+			state.noteText(content)
+			events = append(events, toolStreamEvent{Content: content})
+		}
 		state.pending.Reset()
 	}
 	return events
@@ -159,6 +171,10 @@ func findSuspiciousPrefixStart(s string) int {
 			start = idx
 		}
 	}
+	// Also check for partial XML tool tag at end of string.
+	if xmlIdx := findPartialXMLToolTagStart(s); xmlIdx >= 0 && xmlIdx > start {
+		start = xmlIdx
+	}
 	return start
 }
 
@@ -167,7 +183,7 @@ func findToolSegmentStart(s string) int {
 		return -1
 	}
 	lower := strings.ToLower(s)
-	keywords := []string{"tool_calls", "\"function\"", "function.name:", "[tool_call_history]", "[tool_result_history]"}
+	keywords := []string{"tool_calls", "\"function\"", "function.name:"}
 	bestKeyIdx := -1
 	for _, kw := range keywords {
 		idx := strings.Index(lower, kw)
@@ -175,13 +191,32 @@ func findToolSegmentStart(s string) int {
 			bestKeyIdx = idx
 		}
 	}
+	// Also detect XML tool call tags.
+	for _, tag := range xmlToolTagsToDetect {
+		idx := strings.Index(lower, tag)
+		if idx >= 0 && (bestKeyIdx < 0 || idx < bestKeyIdx) {
+			bestKeyIdx = idx
+		}
+	}
 	if bestKeyIdx < 0 {
 		return -1
 	}
+	// For XML tags, the '<' is itself the segment start.
+	if bestKeyIdx < len(s) && s[bestKeyIdx] == '<' {
+		if fenceStart, ok := openFenceStartBefore(s, bestKeyIdx); ok {
+			return fenceStart
+		}
+		return bestKeyIdx
+	}
 	start := strings.LastIndex(s[:bestKeyIdx], "{")
 	if start < 0 {
 		start = bestKeyIdx
 	}
+	// If the keyword matched inside an XML tag (e.g. "tool_calls" in "<tool_calls>"),
+	// back up past the '<' to capture the full tag.
+	if start > 0 && s[start-1] == '<' {
+		start--
+	}
 	if fenceStart, ok := openFenceStartBefore(s, start); ok {
 		return fenceStart
 	}
@@ -193,9 +228,19 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	if captured == "" {
 		return "", nil, "", false
 	}
+
+	// Try XML tool call extraction first.
+	if xmlPrefix, xmlCalls, xmlSuffix, xmlReady := consumeXMLToolCapture(captured, toolNames); xmlReady {
+		return xmlPrefix, xmlCalls, xmlSuffix, true
+	}
+	// If XML tags are present but block is incomplete, keep buffering.
+	if hasOpenXMLToolTag(captured) {
+		return "", nil, "", false
+	}
+
 	lower := strings.ToLower(captured)
 	keyIdx := -1
-	keywords := []string{"tool_calls", "\"function\"", "function.name:", "[tool_call_history]", "[tool_result_history]"}
+	keywords := []string{"tool_calls", "\"function\"", "function.name:"}
 	for _, kw := range keywords {
 		idx := strings.Index(lower, kw)
 		if idx >= 0 && (keyIdx < 0 || idx < keyIdx) {
@@ -208,9 +253,6 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	}
 	start := strings.LastIndex(captured[:keyIdx], "{")
 	if start < 0 {
-		if blockStart, blockEnd, ok := extractToolHistoryBlock(captured, keyIdx); ok {
-			return captured[:blockStart], nil, captured[blockEnd:], true
-		}
 		start = keyIdx
 	}
 	obj, end, ok := extractJSONObjectFrom(captured, start)
@@ -234,67 +276,3 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	prefixPart, suffixPart = trimWrappingJSONFence(prefixPart, suffixPart)
 	return prefixPart, parsed.Calls, suffixPart, true
 }
-
-func extractToolHistoryBlock(captured string, keyIdx int) (start int, end int, ok bool) {
-	if keyIdx < 0 || keyIdx >= len(captured) {
-		return 0, 0, false
-	}
-	rest := strings.ToLower(captured[keyIdx:])
-	switch {
-	case strings.HasPrefix(rest, "[tool_call_history]"):
-		closeTag := "[/tool_call_history]"
-		closeIdx := strings.Index(rest, closeTag)
-		if closeIdx < 0 {
-			return 0, 0, false
-		}
-		return keyIdx, keyIdx + closeIdx + len(closeTag), true
-	case strings.HasPrefix(rest, "[tool_result_history]"):
-		closeTag := "[/tool_result_history]"
-		closeIdx := strings.Index(rest, closeTag)
-		if closeIdx < 0 {
-			return 0, 0, false
-		}
-		return keyIdx, keyIdx + closeIdx + len(closeTag), true
-	default:
-		return 0, 0, false
-	}
-}
-
-func trimWrappingJSONFence(prefix, suffix string) (string, string) {
-	trimmedPrefix := strings.TrimRight(prefix, " \t\r\n")
-	fenceIdx := strings.LastIndex(trimmedPrefix, "```")
-	if fenceIdx < 0 {
-		return prefix, suffix
-	}
-	// Only strip when the trailing fence in prefix behaves like an opening fence.
-	// A legitimate closing fence before a standalone tool JSON must be preserved.
-	if strings.Count(trimmedPrefix[:fenceIdx+3], "```")%2 == 0 {
-		return prefix, suffix
-	}
-	fenceHeader := strings.TrimSpace(trimmedPrefix[fenceIdx+3:])
-	if fenceHeader != "" && !strings.EqualFold(fenceHeader, "json") {
-		return prefix, suffix
-	}
-
-	trimmedSuffix := strings.TrimLeft(suffix, " \t\r\n")
-	if !strings.HasPrefix(trimmedSuffix, "```") {
-		return prefix, suffix
-	}
-	consumedLeading := len(suffix) - len(trimmedSuffix)
-	return trimmedPrefix[:fenceIdx], suffix[consumedLeading+3:]
-}
-
-func openFenceStartBefore(s string, pos int) (int, bool) {
-	if pos <= 0 || pos > len(s) {
-		return -1, false
-	}
-	segment := s[:pos]
-	lastFence := strings.LastIndex(segment, "```")
-	if lastFence < 0 {
-		return -1, false
-	}
-	if strings.Count(segment, "```")%2 == 1 {
-		return lastFence, true
-	}
-	return -1, false
-}

internal/adapter/openai/tool_sieve_jsonscan.go

@@ -1,5 +1,7 @@
 package openai
 
+import "strings"
+
 func extractJSONObjectFrom(text string, start int) (string, int, bool) {
 	if start < 0 || start >= len(text) || text[start] != '{' {
 		return "", 0, false
@@ -41,3 +43,42 @@ func extractJSONObjectFrom(text string, start int) (string, int, bool) {
 	}
 	return "", 0, false
 }
+
+func trimWrappingJSONFence(prefix, suffix string) (string, string) {
+	trimmedPrefix := strings.TrimRight(prefix, " \t\r\n")
+	fenceIdx := strings.LastIndex(trimmedPrefix, "```")
+	if fenceIdx < 0 {
+		return prefix, suffix
+	}
+	// Only strip when the trailing fence in prefix behaves like an opening fence.
+	// A legitimate closing fence before a standalone tool JSON must be preserved.
+	if strings.Count(trimmedPrefix[:fenceIdx+3], "```")%2 == 0 {
+		return prefix, suffix
+	}
+	fenceHeader := strings.TrimSpace(trimmedPrefix[fenceIdx+3:])
+	if fenceHeader != "" && !strings.EqualFold(fenceHeader, "json") {
+		return prefix, suffix
+	}
+
+	trimmedSuffix := strings.TrimLeft(suffix, " \t\r\n")
+	if !strings.HasPrefix(trimmedSuffix, "```") {
+		return prefix, suffix
+	}
+	consumedLeading := len(suffix) - len(trimmedSuffix)
+	return trimmedPrefix[:fenceIdx], suffix[consumedLeading+3:]
+}
+
+func openFenceStartBefore(s string, pos int) (int, bool) {
+	if pos <= 0 || pos > len(s) {
+		return -1, false
+	}
+	segment := s[:pos]
+	lastFence := strings.LastIndex(segment, "```")
+	if lastFence < 0 {
+		return -1, false
+	}
+	if strings.Count(segment, "```")%2 == 1 {
+		return lastFence, true
+	}
+	return -1, false
+}

internal/adapter/openai/tool_sieve_state.go

@@ -34,7 +34,8 @@ type toolCallDelta struct {
 	Arguments string
 }
 
-const toolSieveContextTailLimit = 256
+// Keep in sync with JS TOOL_SIEVE_CONTEXT_TAIL_LIMIT.
+const toolSieveContextTailLimit = 2048
 
 func (s *toolStreamSieveState) resetIncrementalToolState() {
 	s.disableDeltas = false

internal/adapter/openai/tool_sieve_xml.go

@@ -0,0 +1,147 @@
+package openai
+
+import (
+	"regexp"
+	"strings"
+
+	"ds2api/internal/util"
+)
+
+// --- XML tool call support for the streaming sieve ---
+
+var xmlToolCallClosingTags = []string{"</tool_calls>", "</tool_call>", "</invoke>", "</function_call>", "</function_calls>", "</tool_use>",
+	// Agent-style XML tags (Roo Code, Cline, etc.)
+	"</attempt_completion>", "</ask_followup_question>", "</new_task>", "</result>"}
+var xmlToolCallOpeningTags = []string{"<tool_calls", "<tool_call", "<invoke", "<function_call", "<function_calls", "<tool_use",
+	// Agent-style XML tags
+	"<attempt_completion", "<ask_followup_question", "<new_task", "<result"}
+
+// xmlToolCallTagPairs maps each opening tag to its expected closing tag.
+// Order matters: longer/wrapper tags must be checked first.
+var xmlToolCallTagPairs = []struct{ open, close string }{
+	{"<tool_calls", "</tool_calls>"},
+	{"<tool_call", "</tool_call>"},
+	{"<function_calls", "</function_calls>"},
+	{"<function_call", "</function_call>"},
+	{"<invoke", "</invoke>"},
+	{"<tool_use", "</tool_use>"},
+	// Agent-style: these are XML "tool call" patterns from coding agents.
+	// They get captured → parsed. If parsing fails, the block is consumed
+	// (swallowed) to prevent raw XML from leaking to the client.
+	{"<attempt_completion", "</attempt_completion>"},
+	{"<ask_followup_question", "</ask_followup_question>"},
+	{"<new_task", "</new_task>"},
+}
+
+// xmlToolCallBlockPattern matches a complete XML tool call block (wrapper or standalone).
+var xmlToolCallBlockPattern = regexp.MustCompile(`(?is)(<tool_calls>\s*(?:.*?)\s*</tool_calls>|<tool_call>\s*(?:.*?)\s*</tool_call>|<invoke\b[^>]*>(?:.*?)</invoke>|<function_calls?\b[^>]*>(?:.*?)</function_calls?>|<tool_use>(?:.*?)</tool_use>|<attempt_completion>(?:.*?)</attempt_completion>|<ask_followup_question>(?:.*?)</ask_followup_question>|<new_task>(?:.*?)</new_task>)`)
+
+// xmlToolTagsToDetect is the set of XML tag prefixes used by findToolSegmentStart.
+var xmlToolTagsToDetect = []string{"<tool_calls>", "<tool_calls\n", "<tool_call>", "<tool_call\n",
+	"<invoke ", "<invoke>", "<function_call", "<function_calls", "<tool_use>",
+	// Agent-style tags
+	"<attempt_completion>", "<ask_followup_question>", "<new_task>"}
+
+// consumeXMLToolCapture tries to extract complete XML tool call blocks from captured text.
+func consumeXMLToolCapture(captured string, toolNames []string) (prefix string, calls []util.ParsedToolCall, suffix string, ready bool) {
+	lower := strings.ToLower(captured)
+	// Find the FIRST matching open/close pair, preferring wrapper tags.
+	// Tag pairs are ordered longest-first (e.g. <tool_calls before <tool_call)
+	// so wrapper tags are checked before inner tags.
+	for _, pair := range xmlToolCallTagPairs {
+		openIdx := strings.Index(lower, pair.open)
+		if openIdx < 0 {
+			continue
+		}
+		// Find the LAST occurrence of the specific closing tag to get the outermost block.
+		closeIdx := strings.LastIndex(lower, pair.close)
+		if closeIdx < openIdx {
+			// Opening tag is present but its specific closing tag hasn't arrived.
+			// Return not-ready so we keep buffering — do NOT fall through to
+			// try inner pairs (e.g. <tool_call inside <tool_calls).
+			return "", nil, "", false
+		}
+		closeEnd := closeIdx + len(pair.close)
+
+		xmlBlock := captured[openIdx:closeEnd]
+		prefixPart := captured[:openIdx]
+		suffixPart := captured[closeEnd:]
+		parsed := util.ParseToolCalls(xmlBlock, toolNames)
+		if len(parsed) > 0 {
+			prefixPart, suffixPart = trimWrappingJSONFence(prefixPart, suffixPart)
+			return prefixPart, parsed, suffixPart, true
+		}
+		// Looks like XML tool syntax but failed to parse — consume it to avoid leak.
+		return prefixPart, nil, suffixPart, true
+	}
+	return "", nil, "", false
+}
+
+// hasOpenXMLToolTag returns true if captured text contains an XML tool opening tag
+// whose SPECIFIC closing tag has not appeared yet.
+func hasOpenXMLToolTag(captured string) bool {
+	lower := strings.ToLower(captured)
+	for _, pair := range xmlToolCallTagPairs {
+		if strings.Contains(lower, pair.open) {
+			if !strings.Contains(lower, pair.close) {
+				return true
+			}
+		}
+	}
+	return false
+}
+
+// findPartialXMLToolTagStart checks if the string ends with a partial XML tool tag
+// (e.g., "<tool_ca" or "<inv") and returns the position of the '<'.
+func findPartialXMLToolTagStart(s string) int {
+	lastLT := strings.LastIndex(s, "<")
+	if lastLT < 0 {
+		return -1
+	}
+	tail := s[lastLT:]
+	// If there's a '>' in the tail, the tag is closed — not partial.
+	if strings.Contains(tail, ">") {
+		return -1
+	}
+	lowerTail := strings.ToLower(tail)
+	// Check if the tail is a prefix of any known XML tool tag.
+	for _, tag := range xmlToolCallOpeningTags {
+		tagWithLT := tag
+		if !strings.HasPrefix(tagWithLT, "<") {
+			tagWithLT = "<" + tagWithLT
+		}
+		if strings.HasPrefix(tagWithLT, lowerTail) {
+			return lastLT
+		}
+	}
+	return -1
+}
+
+// looksLikeXMLToolTagFragment returns true if s looks like a fragment from a
+// split XML tool call tag — for example "tool_calls>" or "/tool_call>\n".
+// These fragments arise when '<' was consumed separately and the tail remains.
+func looksLikeXMLToolTagFragment(s string) bool {
+	trimmed := strings.TrimSpace(s)
+	if trimmed == "" {
+		return false
+	}
+	lower := strings.ToLower(trimmed)
+	// Check for closing tag tails like "tool_calls>" or "/tool_calls>"
+	fragments := []string{
+		"tool_calls>", "tool_call>", "/tool_calls>", "/tool_call>",
+		"function_calls>", "function_call>", "/function_calls>", "/function_call>",
+		"invoke>", "/invoke>", "tool_use>", "/tool_use>",
+		"tool_name>", "/tool_name>", "parameters>", "/parameters>",
+		// Agent-style tag fragments
+		"attempt_completion>", "/attempt_completion>",
+		"ask_followup_question>", "/ask_followup_question>",
+		"new_task>", "/new_task>",
+		"result>", "/result>",
+	}
+	for _, f := range fragments {
+		if strings.Contains(lower, f) {
+			return true
+		}
+	}
+	return false
+}

internal/adapter/openai/tool_sieve_xml_test.go

@@ -0,0 +1,319 @@
+package openai
+
+import (
+	"strings"
+	"testing"
+)
+
+func TestProcessToolSieveInterceptsXMLToolCallWithoutLeak(t *testing.T) {
+	var state toolStreamSieveState
+	// Simulate a model producing XML tool call output chunk by chunk.
+	chunks := []string{
+		"<tool_calls>\n",
+		"  <tool_call>\n",
+		"    <tool_name>read_file</tool_name>\n",
+		`    <parameters>{"path":"README.MD"}</parameters>` + "\n",
+		"  </tool_call>\n",
+		"</tool_calls>",
+	}
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent string
+	var toolCalls int
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if strings.Contains(textContent, "<tool_call") {
+		t.Fatalf("XML tool call content leaked to text: %q", textContent)
+	}
+	if strings.Contains(textContent, "read_file") {
+		t.Fatalf("tool name leaked to text: %q", textContent)
+	}
+	if toolCalls == 0 {
+		t.Fatal("expected tool calls to be extracted, got none")
+	}
+}
+
+func TestProcessToolSieveXMLWithLeadingText(t *testing.T) {
+	var state toolStreamSieveState
+	// Model outputs some prose then an XML tool call.
+	chunks := []string{
+		"Let me check the file.\n",
+		"<tool_calls>\n  <tool_call>\n    <tool_name>read_file</tool_name>\n",
+		`    <parameters>{"path":"go.mod"}</parameters>` + "\n  </tool_call>\n</tool_calls>",
+	}
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent string
+	var toolCalls int
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	// Leading text should be emitted.
+	if !strings.Contains(textContent, "Let me check the file.") {
+		t.Fatalf("expected leading text to be emitted, got %q", textContent)
+	}
+	// The XML itself should NOT leak.
+	if strings.Contains(textContent, "<tool_call") {
+		t.Fatalf("XML tool call content leaked to text: %q", textContent)
+	}
+	if toolCalls == 0 {
+		t.Fatal("expected tool calls to be extracted, got none")
+	}
+}
+
+func TestProcessToolSievePartialXMLTagHeldBack(t *testing.T) {
+	var state toolStreamSieveState
+	// Chunk ends with a partial XML tool tag.
+	events := processToolSieveChunk(&state, "Hello <tool_ca", []string{"read_file"})
+
+	var textContent string
+	for _, evt := range events {
+		textContent += evt.Content
+	}
+
+	// "Hello " should be emitted, but "<tool_ca" should be held back.
+	if strings.Contains(textContent, "<tool_ca") {
+		t.Fatalf("partial XML tag should not be emitted, got %q", textContent)
+	}
+	if !strings.Contains(textContent, "Hello") {
+		t.Fatalf("expected 'Hello' text to be emitted, got %q", textContent)
+	}
+}
+
+func TestFindToolSegmentStartDetectsXMLToolCalls(t *testing.T) {
+	cases := []struct {
+		name  string
+		input string
+		want  int
+	}{
+		{"tool_calls_tag", "some text <tool_calls>\n", 10},
+		{"tool_call_tag", "prefix <tool_call>\n", 7},
+		{"invoke_tag", "text <invoke name=\"foo\">body</invoke>", 5},
+		{"function_call_tag", "<function_call name=\"foo\">body</function_call>", 0},
+		{"no_xml", "just plain text", -1},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := findToolSegmentStart(tc.input)
+			if got != tc.want {
+				t.Fatalf("findToolSegmentStart(%q) = %d, want %d", tc.input, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestFindPartialXMLToolTagStart(t *testing.T) {
+	cases := []struct {
+		name  string
+		input string
+		want  int
+	}{
+		{"partial_tool_call", "Hello <tool_ca", 6},
+		{"partial_invoke", "Prefix <inv", 7},
+		{"partial_lt_only", "Text <", 5},
+		{"complete_tag", "Text <tool_call>done", -1},
+		{"no_lt", "plain text", -1},
+		{"closed_lt", "a < b > c", -1},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := findPartialXMLToolTagStart(tc.input)
+			if got != tc.want {
+				t.Fatalf("findPartialXMLToolTagStart(%q) = %d, want %d", tc.input, got, tc.want)
+			}
+		})
+	}
+}
+
+func TestHasOpenXMLToolTag(t *testing.T) {
+	if !hasOpenXMLToolTag("<tool_call>\n<tool_name>foo</tool_name>") {
+		t.Fatal("should detect open XML tool tag without closing tag")
+	}
+	if hasOpenXMLToolTag("<tool_call>\n<tool_name>foo</tool_name></tool_call>") {
+		t.Fatal("should return false when closing tag is present")
+	}
+	if hasOpenXMLToolTag("plain text without any XML") {
+		t.Fatal("should return false for plain text")
+	}
+}
+
+// Test the EXACT scenario the user reports: token-by-token streaming where
+// <tool_calls> tag arrives in small pieces.
+func TestProcessToolSieveTokenByTokenXMLNoLeak(t *testing.T) {
+	var state toolStreamSieveState
+	// Simulate DeepSeek model generating tokens one at a time.
+	chunks := []string{
+		"<",
+		"tool",
+		"_calls",
+		">\n",
+		"  <",
+		"tool",
+		"_call",
+		">\n",
+		"    <",
+		"tool",
+		"_name",
+		">",
+		"read",
+		"_file",
+		"</",
+		"tool",
+		"_name",
+		">\n",
+		"    <",
+		"parameters",
+		">",
+		`{"path"`,
+		`: "README.MD"`,
+		`}`,
+		"</",
+		"parameters",
+		">\n",
+		"  </",
+		"tool",
+		"_call",
+		">\n",
+		"</",
+		"tool",
+		"_calls",
+		">",
+	}
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent string
+	var toolCalls int
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if strings.Contains(textContent, "<tool_call") {
+		t.Fatalf("XML tool call content leaked to text in token-by-token mode: %q", textContent)
+	}
+	if strings.Contains(textContent, "tool_calls>") {
+		t.Fatalf("closing tag fragment leaked to text: %q", textContent)
+	}
+	if strings.Contains(textContent, "read_file") {
+		t.Fatalf("tool name leaked to text: %q", textContent)
+	}
+	if toolCalls == 0 {
+		t.Fatal("expected tool calls to be extracted, got none")
+	}
+}
+
+// Test that flushToolSieve on incomplete XML does NOT leak the raw XML content.
+func TestFlushToolSieveIncompleteXMLDoesNotLeak(t *testing.T) {
+	var state toolStreamSieveState
+	// XML block starts but stream ends before completion.
+	chunks := []string{
+		"<tool_calls>\n",
+		"  <tool_call>\n",
+		"    <tool_name>read_file</tool_name>\n",
+	}
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"read_file"})...)
+	}
+	// Stream ends abruptly - flush should NOT dump raw XML.
+	events = append(events, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent string
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+	}
+
+	if strings.Contains(textContent, "<tool_call") {
+		t.Fatalf("incomplete XML leaked on flush: %q", textContent)
+	}
+}
+
+// Test that the opening tag "<tool_calls>\n  " is NOT emitted as text content.
+func TestOpeningXMLTagNotLeakedAsContent(t *testing.T) {
+	var state toolStreamSieveState
+	// First chunk is the opening tag - should be held, not emitted.
+	evts1 := processToolSieveChunk(&state, "<tool_calls>\n  ", []string{"read_file"})
+	for _, evt := range evts1 {
+		if strings.Contains(evt.Content, "<tool_calls>") {
+			t.Fatalf("opening tag leaked on first chunk: %q", evt.Content)
+		}
+	}
+
+	// Remaining content arrives.
+	evts2 := processToolSieveChunk(&state, "<tool_call>\n    <tool_name>read_file</tool_name>\n    <parameters>{\"path\":\"README.MD\"}</parameters>\n  </tool_call>\n</tool_calls>", []string{"read_file"})
+	evts2 = append(evts2, flushToolSieve(&state, []string{"read_file"})...)
+
+	var textContent string
+	var toolCalls int
+	allEvents := append(evts1, evts2...)
+	for _, evt := range allEvents {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+		toolCalls += len(evt.ToolCalls)
+	}
+
+	if strings.Contains(textContent, "<tool_call") {
+		t.Fatalf("XML content leaked: %q", textContent)
+	}
+	if toolCalls == 0 {
+		t.Fatal("expected tool calls to be extracted")
+	}
+}
+
+func TestProcessToolSieveInterceptsAttemptCompletionLeak(t *testing.T) {
+	var state toolStreamSieveState
+	// Simulate an agent outputting attempt_completion XML tag 
+	// which shouldn't leak to text output, even if it fails to parse as a valid tool.
+	chunks := []string{
+		"Done with task.\n",
+		"<attempt_completion>\n",
+		"  <result>Here is the answer</result>\n",
+		"</attempt_completion>",
+	}
+	var events []toolStreamEvent
+	for _, c := range chunks {
+		events = append(events, processToolSieveChunk(&state, c, []string{"attempt_completion"})...)
+	}
+	events = append(events, flushToolSieve(&state, []string{"attempt_completion"})...)
+
+	var textContent string
+	for _, evt := range events {
+		if evt.Content != "" {
+			textContent += evt.Content
+		}
+	}
+
+	if !strings.Contains(textContent, "Done with task.\n") {
+		t.Fatalf("expected leading text to be emitted, got %q", textContent)
+	}
+
+	if strings.Contains(textContent, "<attempt_completion>") || strings.Contains(textContent, "result>") {
+		t.Fatalf("agent XML tag content leaked to text: %q", textContent)
+	}
+}

internal/adapter/openai/vercel_stream.go

@@ -93,18 +93,16 @@ func (h *Handler) handleVercelStreamPrepare(w http.ResponseWriter, r *http.Reque
 	}
 	leased = true
 	writeJSON(w, http.StatusOK, map[string]any{
-		"session_id":               sessionID,
-		"lease_id":                 leaseID,
-		"model":                    stdReq.ResponseModel,
-		"final_prompt":             stdReq.FinalPrompt,
-		"thinking_enabled":         stdReq.Thinking,
-		"search_enabled":           stdReq.Search,
-		"tool_names":               stdReq.ToolNames,
-		"toolcall_feature_match":   h.toolcallFeatureMatchEnabled(),
-		"toolcall_early_emit_high": h.toolcallEarlyEmitHighConfidence(),
-		"deepseek_token":           a.DeepSeekToken,
-		"pow_header":               powHeader,
-		"payload":                  payload,
+		"session_id":       sessionID,
+		"lease_id":         leaseID,
+		"model":            stdReq.ResponseModel,
+		"final_prompt":     stdReq.FinalPrompt,
+		"thinking_enabled": stdReq.Thinking,
+		"search_enabled":   stdReq.Search,
+		"tool_names":       stdReq.ToolNames,
+		"deepseek_token":   a.DeepSeekToken,
+		"pow_header":       powHeader,
+		"payload":          payload,
 	})
 }
 

internal/admin/deps.go

@@ -21,6 +21,9 @@ type ConfigStore interface {
 	Update(mutator func(*config.Config) error) error
 	ExportJSONAndBase64() (string, string, error)
 	IsEnvBacked() bool
+	IsEnvWritebackEnabled() bool
+	HasEnvConfigSource() bool
+	ConfigPath() string
 	SetVercelSync(hash string, ts int64) error
 	AdminPasswordHash() string
 	AdminJWTExpireHours() int
@@ -28,6 +31,7 @@ type ConfigStore interface {
 	RuntimeAccountMaxInflight() int
 	RuntimeAccountMaxQueue(defaultSize int) int
 	RuntimeGlobalMaxInflight(defaultSize int) int
+	RuntimeTokenRefreshIntervalHours() int
 	AutoDeleteSessions() bool
 }
 

internal/admin/handler_config_import.go

@@ -120,12 +120,6 @@ func (h *Handler) configImport(w http.ResponseWriter, r *http.Request) {
 					next.ModelAliases[k] = v
 				}
 			}
-			if strings.TrimSpace(incoming.Toolcall.Mode) != "" {
-				next.Toolcall.Mode = incoming.Toolcall.Mode
-			}
-			if strings.TrimSpace(incoming.Toolcall.EarlyEmitConfidence) != "" {
-				next.Toolcall.EarlyEmitConfidence = incoming.Toolcall.EarlyEmitConfidence
-			}
 			if incoming.Responses.StoreTTLSeconds > 0 {
 				next.Responses.StoreTTLSeconds = incoming.Responses.StoreTTLSeconds
 			}
@@ -150,6 +144,9 @@ func (h *Handler) configImport(w http.ResponseWriter, r *http.Request) {
 			if incoming.Runtime.GlobalMaxInflight > 0 {
 				next.Runtime.GlobalMaxInflight = incoming.Runtime.GlobalMaxInflight
 			}
+			if incoming.Runtime.TokenRefreshIntervalHours > 0 {
+				next.Runtime.TokenRefreshIntervalHours = incoming.Runtime.TokenRefreshIntervalHours
+			}
 		}
 
 		normalizeSettingsConfig(&next)

internal/admin/handler_config_read.go

@@ -8,9 +8,12 @@ import (
 func (h *Handler) getConfig(w http.ResponseWriter, _ *http.Request) {
 	snap := h.Store.Snapshot()
 	safe := map[string]any{
-		"keys":       snap.Keys,
-		"accounts":   []map[string]any{},
-		"env_backed": h.Store.IsEnvBacked(),
+		"keys":                  snap.Keys,
+		"accounts":              []map[string]any{},
+		"env_backed":            h.Store.IsEnvBacked(),
+		"env_source_present":    h.Store.HasEnvConfigSource(),
+		"env_writeback_enabled": h.Store.IsEnvWritebackEnabled(),
+		"config_path":           h.Store.ConfigPath(),
 		"claude_mapping": func() map[string]string {
 			if len(snap.ClaudeMapping) > 0 {
 				return snap.ClaudeMapping

internal/admin/handler_settings_parse.go

@@ -21,16 +21,15 @@ func boolFrom(v any) bool {
 	}
 }
 
-func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *config.RuntimeConfig, *config.ToolcallConfig, *config.ResponsesConfig, *config.EmbeddingsConfig, *config.AutoDeleteConfig, map[string]string, map[string]string, error) {
+func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *config.RuntimeConfig, *config.ResponsesConfig, *config.EmbeddingsConfig, *config.AutoDeleteConfig, map[string]string, map[string]string, error) {
 	var (
-		adminCfg       *config.AdminConfig
-		runtimeCfg     *config.RuntimeConfig
-		toolcallCfg    *config.ToolcallConfig
-		respCfg        *config.ResponsesConfig
-		embCfg         *config.EmbeddingsConfig
-		autoDeleteCfg  *config.AutoDeleteConfig
-		claudeMap      map[string]string
-		aliasMap       map[string]string
+		adminCfg      *config.AdminConfig
+		runtimeCfg    *config.RuntimeConfig
+		respCfg       *config.ResponsesConfig
+		embCfg        *config.EmbeddingsConfig
+		autoDeleteCfg *config.AutoDeleteConfig
+		claudeMap     map[string]string
+		aliasMap      map[string]string
 	)
 
 	if raw, ok := req["admin"].(map[string]any); ok {
@@ -38,7 +37,7 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["jwt_expire_hours"]; exists {
 			n := intFrom(v)
 			if n < 1 || n > 720 {
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("admin.jwt_expire_hours must be between 1 and 720")
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("admin.jwt_expire_hours must be between 1 and 720")
 			}
 			cfg.JWTExpireHours = n
 		}
@@ -50,59 +49,43 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		if v, exists := raw["account_max_inflight"]; exists {
 			n := intFrom(v)
 			if n < 1 || n > 256 {
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.account_max_inflight must be between 1 and 256")
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.account_max_inflight must be between 1 and 256")
 			}
 			cfg.AccountMaxInflight = n
 		}
 		if v, exists := raw["account_max_queue"]; exists {
 			n := intFrom(v)
 			if n < 1 || n > 200000 {
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.account_max_queue must be between 1 and 200000")
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.account_max_queue must be between 1 and 200000")
 			}
 			cfg.AccountMaxQueue = n
 		}
 		if v, exists := raw["global_max_inflight"]; exists {
 			n := intFrom(v)
 			if n < 1 || n > 200000 {
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be between 1 and 200000")
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be between 1 and 200000")
 			}
 			cfg.GlobalMaxInflight = n
 		}
+		if v, exists := raw["token_refresh_interval_hours"]; exists {
+			n := intFrom(v)
+			if n < 1 || n > 720 {
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.token_refresh_interval_hours must be between 1 and 720")
+			}
+			cfg.TokenRefreshIntervalHours = n
+		}
 		if cfg.AccountMaxInflight > 0 && cfg.GlobalMaxInflight > 0 && cfg.GlobalMaxInflight < cfg.AccountMaxInflight {
-			return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be >= runtime.account_max_inflight")
+			return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("runtime.global_max_inflight must be >= runtime.account_max_inflight")
 		}
 		runtimeCfg = cfg
 	}
 
-	if raw, ok := req["toolcall"].(map[string]any); ok {
-		cfg := &config.ToolcallConfig{}
-		if v, exists := raw["mode"]; exists {
-			mode := strings.ToLower(strings.TrimSpace(fmt.Sprintf("%v", v)))
-			switch mode {
-			case "feature_match", "off":
-				cfg.Mode = mode
-			default:
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("toolcall.mode must be feature_match or off")
-			}
-		}
-		if v, exists := raw["early_emit_confidence"]; exists {
-			level := strings.ToLower(strings.TrimSpace(fmt.Sprintf("%v", v)))
-			switch level {
-			case "high", "low", "off":
-				cfg.EarlyEmitConfidence = level
-			default:
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("toolcall.early_emit_confidence must be high, low or off")
-			}
-		}
-		toolcallCfg = cfg
-	}
-
 	if raw, ok := req["responses"].(map[string]any); ok {
 		cfg := &config.ResponsesConfig{}
 		if v, exists := raw["store_ttl_seconds"]; exists {
 			n := intFrom(v)
 			if n < 30 || n > 86400 {
-				return nil, nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("responses.store_ttl_seconds must be between 30 and 86400")
+				return nil, nil, nil, nil, nil, nil, nil, fmt.Errorf("responses.store_ttl_seconds must be between 30 and 86400")
 			}
 			cfg.StoreTTLSeconds = n
 		}
@@ -150,5 +133,5 @@ func parseSettingsUpdateRequest(req map[string]any) (*config.AdminConfig, *confi
 		autoDeleteCfg = cfg
 	}
 
-	return adminCfg, runtimeCfg, toolcallCfg, respCfg, embCfg, autoDeleteCfg, claudeMap, aliasMap, nil
+	return adminCfg, runtimeCfg, respCfg, embCfg, autoDeleteCfg, claudeMap, aliasMap, nil
 }

internal/admin/handler_settings_read.go

@@ -21,11 +21,11 @@ func (h *Handler) getSettings(w http.ResponseWriter, _ *http.Request) {
 			"default_password_warning": authn.UsingDefaultAdminKey(h.Store),
 		},
 		"runtime": map[string]any{
-			"account_max_inflight": h.Store.RuntimeAccountMaxInflight(),
-			"account_max_queue":    h.Store.RuntimeAccountMaxQueue(recommended),
-			"global_max_inflight":  h.Store.RuntimeGlobalMaxInflight(recommended),
+			"account_max_inflight":         h.Store.RuntimeAccountMaxInflight(),
+			"account_max_queue":            h.Store.RuntimeAccountMaxQueue(recommended),
+			"global_max_inflight":          h.Store.RuntimeGlobalMaxInflight(recommended),
+			"token_refresh_interval_hours": h.Store.RuntimeTokenRefreshIntervalHours(),
 		},
-		"toolcall":          snap.Toolcall,
 		"responses":         snap.Responses,
 		"embeddings":        snap.Embeddings,
 		"auto_delete":       snap.AutoDelete,

internal/admin/handler_settings_runtime.go

@@ -14,6 +14,9 @@ func validateMergedRuntimeSettings(current config.RuntimeConfig, incoming *confi
 		if incoming.GlobalMaxInflight > 0 {
 			merged.GlobalMaxInflight = incoming.GlobalMaxInflight
 		}
+		if incoming.TokenRefreshIntervalHours > 0 {
+			merged.TokenRefreshIntervalHours = incoming.TokenRefreshIntervalHours
+		}
 	}
 	return validateRuntimeSettings(merged)
 }

internal/admin/handler_settings_test.go

@@ -28,6 +28,25 @@ func TestGetSettingsDefaultPasswordWarning(t *testing.T) {
 	}
 }
 
+func TestGetSettingsIncludesTokenRefreshInterval(t *testing.T) {
+	h := newAdminTestHandler(t, `{
+		"keys":["k1"],
+		"runtime":{"token_refresh_interval_hours":9}
+	}`)
+	req := httptest.NewRequest(http.MethodGet, "/admin/settings", nil)
+	rec := httptest.NewRecorder()
+	h.getSettings(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	var body map[string]any
+	_ = json.Unmarshal(rec.Body.Bytes(), &body)
+	runtime, _ := body["runtime"].(map[string]any)
+	if got := intFrom(runtime["token_refresh_interval_hours"]); got != 9 {
+		t.Fatalf("expected token_refresh_interval_hours=9, got %d body=%v", got, body)
+	}
+}
+
 func TestUpdateSettingsValidation(t *testing.T) {
 	h := newAdminTestHandler(t, `{"keys":["k1"]}`)
 	payload := map[string]any{
@@ -44,6 +63,25 @@ func TestUpdateSettingsValidation(t *testing.T) {
 	}
 }
 
+func TestUpdateSettingsValidationRejectsTokenRefreshInterval(t *testing.T) {
+	h := newAdminTestHandler(t, `{"keys":["k1"]}`)
+	payload := map[string]any{
+		"runtime": map[string]any{
+			"token_refresh_interval_hours": 0,
+		},
+	}
+	b, _ := json.Marshal(payload)
+	req := httptest.NewRequest(http.MethodPut, "/admin/settings", bytes.NewReader(b))
+	rec := httptest.NewRecorder()
+	h.updateSettings(rec, req)
+	if rec.Code != http.StatusBadRequest {
+		t.Fatalf("expected 400, got %d body=%s", rec.Code, rec.Body.String())
+	}
+	if !bytes.Contains(rec.Body.Bytes(), []byte("runtime.token_refresh_interval_hours")) {
+		t.Fatalf("expected token refresh validation detail, got %s", rec.Body.String())
+	}
+}
+
 func TestUpdateSettingsValidationWithMergedRuntimeSnapshot(t *testing.T) {
 	h := newAdminTestHandler(t, `{
 		"keys":["k1"],
@@ -126,6 +164,29 @@ func TestUpdateSettingsHotReloadRuntime(t *testing.T) {
 	}
 }
 
+func TestUpdateSettingsHotReloadTokenRefreshInterval(t *testing.T) {
+	h := newAdminTestHandler(t, `{
+		"keys":["k1"],
+		"runtime":{"token_refresh_interval_hours":6}
+	}`)
+
+	payload := map[string]any{
+		"runtime": map[string]any{
+			"token_refresh_interval_hours": 12,
+		},
+	}
+	b, _ := json.Marshal(payload)
+	req := httptest.NewRequest(http.MethodPut, "/admin/settings", bytes.NewReader(b))
+	rec := httptest.NewRecorder()
+	h.updateSettings(rec, req)
+	if rec.Code != http.StatusOK {
+		t.Fatalf("status=%d body=%s", rec.Code, rec.Body.String())
+	}
+	if got := h.Store.RuntimeTokenRefreshIntervalHours(); got != 12 {
+		t.Fatalf("token_refresh_interval_hours=%d want=12", got)
+	}
+}
+
 func TestUpdateSettingsPasswordInvalidatesOldJWT(t *testing.T) {
 	hash := authn.HashAdminPassword("old-password")
 	h := newAdminTestHandler(t, `{"admin":{"password_hash":"`+hash+`"}}`)
@@ -207,6 +268,30 @@ func TestConfigImportMergeAndReplace(t *testing.T) {
 	}
 }
 
+func TestConfigImportAppliesTokenRefreshInterval(t *testing.T) {
+	h := newAdminTestHandler(t, `{"keys":["k1"]}`)
+
+	replace := map[string]any{
+		"mode": "replace",
+		"config": map[string]any{
+			"keys": []any{"k9"},
+			"runtime": map[string]any{
+				"token_refresh_interval_hours": 11,
+			},
+		},
+	}
+	replaceBytes, _ := json.Marshal(replace)
+	replaceReq := httptest.NewRequest(http.MethodPost, "/admin/config/import?mode=replace", bytes.NewReader(replaceBytes))
+	replaceRec := httptest.NewRecorder()
+	h.configImport(replaceRec, replaceReq)
+	if replaceRec.Code != http.StatusOK {
+		t.Fatalf("replace status=%d body=%s", replaceRec.Code, replaceRec.Body.String())
+	}
+	if got := h.Store.RuntimeTokenRefreshIntervalHours(); got != 11 {
+		t.Fatalf("token_refresh_interval_hours=%d want=11", got)
+	}
+}
+
 func TestConfigImportRejectsInvalidRuntimeBounds(t *testing.T) {
 	h := newAdminTestHandler(t, `{"keys":["k1"]}`)
 	payload := map[string]any{

internal/admin/handler_settings_write.go

@@ -17,7 +17,7 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 		return
 	}
 
-	adminCfg, runtimeCfg, toolcallCfg, responsesCfg, embeddingsCfg, autoDeleteCfg, claudeMap, aliasMap, err := parseSettingsUpdateRequest(req)
+	adminCfg, runtimeCfg, responsesCfg, embeddingsCfg, autoDeleteCfg, claudeMap, aliasMap, err := parseSettingsUpdateRequest(req)
 	if err != nil {
 		writeJSON(w, http.StatusBadRequest, map[string]any{"detail": err.Error()})
 		return
@@ -45,13 +45,8 @@ func (h *Handler) updateSettings(w http.ResponseWriter, r *http.Request) {
 			if runtimeCfg.GlobalMaxInflight > 0 {
 				c.Runtime.GlobalMaxInflight = runtimeCfg.GlobalMaxInflight
 			}
-		}
-		if toolcallCfg != nil {
-			if strings.TrimSpace(toolcallCfg.Mode) != "" {
-				c.Toolcall.Mode = strings.TrimSpace(toolcallCfg.Mode)
-			}
-			if strings.TrimSpace(toolcallCfg.EarlyEmitConfidence) != "" {
-				c.Toolcall.EarlyEmitConfidence = strings.TrimSpace(toolcallCfg.EarlyEmitConfidence)
+			if runtimeCfg.TokenRefreshIntervalHours > 0 {
+				c.Runtime.TokenRefreshIntervalHours = runtimeCfg.TokenRefreshIntervalHours
 			}
 		}
 		if responsesCfg != nil && responsesCfg.StoreTTLSeconds > 0 {

internal/admin/settings_validation.go

@@ -12,8 +12,6 @@ func normalizeSettingsConfig(c *config.Config) {
 		return
 	}
 	c.Admin.PasswordHash = strings.TrimSpace(c.Admin.PasswordHash)
-	c.Toolcall.Mode = strings.ToLower(strings.TrimSpace(c.Toolcall.Mode))
-	c.Toolcall.EarlyEmitConfidence = strings.ToLower(strings.TrimSpace(c.Toolcall.EarlyEmitConfidence))
 	c.Embeddings.Provider = strings.TrimSpace(c.Embeddings.Provider)
 }
 
@@ -27,20 +25,6 @@ func validateSettingsConfig(c config.Config) error {
 	if c.Responses.StoreTTLSeconds != 0 && (c.Responses.StoreTTLSeconds < 30 || c.Responses.StoreTTLSeconds > 86400) {
 		return fmt.Errorf("responses.store_ttl_seconds must be between 30 and 86400")
 	}
-	if mode := strings.TrimSpace(c.Toolcall.Mode); mode != "" {
-		switch mode {
-		case "feature_match", "off":
-		default:
-			return fmt.Errorf("toolcall.mode must be feature_match or off")
-		}
-	}
-	if level := strings.TrimSpace(c.Toolcall.EarlyEmitConfidence); level != "" {
-		switch level {
-		case "high", "low", "off":
-		default:
-			return fmt.Errorf("toolcall.early_emit_confidence must be high, low or off")
-		}
-	}
 	if c.Embeddings.Provider != "" && strings.TrimSpace(c.Embeddings.Provider) == "" {
 		return fmt.Errorf("embeddings.provider cannot be empty")
 	}
@@ -57,6 +41,9 @@ func validateRuntimeSettings(runtime config.RuntimeConfig) error {
 	if runtime.GlobalMaxInflight != 0 && (runtime.GlobalMaxInflight < 1 || runtime.GlobalMaxInflight > 200000) {
 		return fmt.Errorf("runtime.global_max_inflight must be between 1 and 200000")
 	}
+	if runtime.TokenRefreshIntervalHours != 0 && (runtime.TokenRefreshIntervalHours < 1 || runtime.TokenRefreshIntervalHours > 720) {
+		return fmt.Errorf("runtime.token_refresh_interval_hours must be between 1 and 720")
+	}
 	if runtime.AccountMaxInflight > 0 && runtime.GlobalMaxInflight > 0 && runtime.GlobalMaxInflight < runtime.AccountMaxInflight {
 		return fmt.Errorf("runtime.global_max_inflight must be >= runtime.account_max_inflight")
 	}

internal/auth/request.go

@@ -40,18 +40,16 @@ type Resolver struct {
 	Pool  *account.Pool
 	Login LoginFunc
 
-	mu                   sync.Mutex
-	tokenRefreshedAt     map[string]time.Time
-	tokenRefreshInterval time.Duration
+	mu               sync.Mutex
+	tokenRefreshedAt map[string]time.Time
 }
 
 func NewResolver(store *config.Store, pool *account.Pool, login LoginFunc) *Resolver {
 	return &Resolver{
-		Store:                store,
-		Pool:                 pool,
-		Login:                login,
-		tokenRefreshedAt:     map[string]time.Time{},
-		tokenRefreshInterval: 6 * time.Hour,
+		Store:            store,
+		Pool:             pool,
+		Login:            login,
+		tokenRefreshedAt: map[string]time.Time{},
 	}
 }
 
@@ -232,10 +230,14 @@ func (r *Resolver) ensureManagedToken(ctx context.Context, a *RequestAuth) error
 }
 
 func (r *Resolver) shouldForceRefresh(accountID string) bool {
+	if r == nil || r.Store == nil {
+		return false
+	}
 	if strings.TrimSpace(accountID) == "" {
 		return false
 	}
-	if r.tokenRefreshInterval <= 0 {
+	intervalHours := r.Store.RuntimeTokenRefreshIntervalHours()
+	if intervalHours <= 0 {
 		return false
 	}
 	now := time.Now()
@@ -246,7 +248,7 @@ func (r *Resolver) shouldForceRefresh(accountID string) bool {
 		r.tokenRefreshedAt[accountID] = now
 		return false
 	}
-	return now.Sub(last) >= r.tokenRefreshInterval
+	return now.Sub(last) >= time.Duration(intervalHours)*time.Hour
 }
 
 func (r *Resolver) markTokenRefreshedNow(accountID string) {

internal/auth/request_test.go

@@ -244,3 +244,60 @@ func TestDetermineManagedAccountForcesRefreshEverySixHours(t *testing.T) {
 		t.Fatalf("expected exactly one forced refresh login, got %d", got)
 	}
 }
+
+func TestDetermineManagedAccountUsesUpdatedRefreshInterval(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["managed-key"],
+		"accounts":[{"email":"acc@example.com","password":"pwd","token":"seed-token"}],
+		"runtime":{"token_refresh_interval_hours":6}
+	}`)
+	store := config.LoadStore()
+	if err := store.UpdateAccountToken("acc@example.com", "seed-token"); err != nil {
+		t.Fatalf("update token failed: %v", err)
+	}
+	pool := account.NewPool(store)
+
+	var loginCount int32
+	resolver := NewResolver(store, pool, func(_ context.Context, _ config.Account) (string, error) {
+		n := atomic.AddInt32(&loginCount, 1)
+		return "fresh-token-" + string(rune('0'+n)), nil
+	})
+
+	req, _ := http.NewRequest(http.MethodPost, "/v1/chat/completions", nil)
+	req.Header.Set("x-api-key", "managed-key")
+
+	a1, err := resolver.Determine(req)
+	if err != nil {
+		t.Fatalf("determine failed: %v", err)
+	}
+	if a1.DeepSeekToken != "seed-token" {
+		t.Fatalf("expected initial token without forced refresh, got %q", a1.DeepSeekToken)
+	}
+	resolver.Release(a1)
+	if got := atomic.LoadInt32(&loginCount); got != 0 {
+		t.Fatalf("expected no login before runtime update, got %d", got)
+	}
+
+	if err := store.Update(func(c *config.Config) error {
+		c.Runtime.TokenRefreshIntervalHours = 1
+		return nil
+	}); err != nil {
+		t.Fatalf("update runtime failed: %v", err)
+	}
+
+	resolver.mu.Lock()
+	resolver.tokenRefreshedAt["acc@example.com"] = time.Now().Add(-2 * time.Hour)
+	resolver.mu.Unlock()
+
+	a2, err := resolver.Determine(req)
+	if err != nil {
+		t.Fatalf("determine after runtime update failed: %v", err)
+	}
+	defer resolver.Release(a2)
+	if a2.DeepSeekToken != "fresh-token-1" {
+		t.Fatalf("expected refreshed token after runtime update, got %q", a2.DeepSeekToken)
+	}
+	if got := atomic.LoadInt32(&loginCount); got != 1 {
+		t.Fatalf("expected exactly one login after runtime update, got %d", got)
+	}
+}

internal/config/codec.go

@@ -32,15 +32,12 @@ func (c Config) MarshalJSON() ([]byte, error) {
 	if strings.TrimSpace(c.Admin.PasswordHash) != "" || c.Admin.JWTExpireHours > 0 || c.Admin.JWTValidAfterUnix > 0 {
 		m["admin"] = c.Admin
 	}
-	if c.Runtime.AccountMaxInflight > 0 || c.Runtime.AccountMaxQueue > 0 || c.Runtime.GlobalMaxInflight > 0 {
+	if c.Runtime.AccountMaxInflight > 0 || c.Runtime.AccountMaxQueue > 0 || c.Runtime.GlobalMaxInflight > 0 || c.Runtime.TokenRefreshIntervalHours > 0 {
 		m["runtime"] = c.Runtime
 	}
 	if c.Compat.WideInputStrictOutput != nil {
 		m["compat"] = c.Compat
 	}
-	if strings.TrimSpace(c.Toolcall.Mode) != "" || strings.TrimSpace(c.Toolcall.EarlyEmitConfidence) != "" {
-		m["toolcall"] = c.Toolcall
-	}
 	if c.Responses.StoreTTLSeconds > 0 {
 		m["responses"] = c.Responses
 	}
@@ -98,9 +95,7 @@ func (c *Config) UnmarshalJSON(b []byte) error {
 				return fmt.Errorf("invalid field %q: %w", k, err)
 			}
 		case "toolcall":
-			if err := json.Unmarshal(v, &c.Toolcall); err != nil {
-				return fmt.Errorf("invalid field %q: %w", k, err)
-			}
+			// Legacy field ignored. Toolcall policy is fixed and no longer configurable.
 		case "responses":
 			if err := json.Unmarshal(v, &c.Responses); err != nil {
 				return fmt.Errorf("invalid field %q: %w", k, err)
@@ -143,7 +138,6 @@ func (c Config) Clone() Config {
 		Compat: CompatConfig{
 			WideInputStrictOutput: cloneBoolPtr(c.Compat.WideInputStrictOutput),
 		},
-		Toolcall:         c.Toolcall,
 		Responses:        c.Responses,
 		Embeddings:       c.Embeddings,
 		AutoDelete:       c.AutoDelete,

internal/config/config.go

@@ -9,7 +9,6 @@ type Config struct {
 	Admin            AdminConfig       `json:"admin,omitempty"`
 	Runtime          RuntimeConfig     `json:"runtime,omitempty"`
 	Compat           CompatConfig      `json:"compat,omitempty"`
-	Toolcall         ToolcallConfig    `json:"toolcall,omitempty"`
 	Responses        ResponsesConfig   `json:"responses,omitempty"`
 	Embeddings       EmbeddingsConfig  `json:"embeddings,omitempty"`
 	AutoDelete       AutoDeleteConfig  `json:"auto_delete"`
@@ -62,14 +61,10 @@ type AdminConfig struct {
 }
 
 type RuntimeConfig struct {
-	AccountMaxInflight int `json:"account_max_inflight,omitempty"`
-	AccountMaxQueue    int `json:"account_max_queue,omitempty"`
-	GlobalMaxInflight  int `json:"global_max_inflight,omitempty"`
-}
-
-type ToolcallConfig struct {
-	Mode                string `json:"mode,omitempty"`
-	EarlyEmitConfidence string `json:"early_emit_confidence,omitempty"`
+	AccountMaxInflight        int `json:"account_max_inflight,omitempty"`
+	AccountMaxQueue           int `json:"account_max_queue,omitempty"`
+	GlobalMaxInflight         int `json:"global_max_inflight,omitempty"`
+	TokenRefreshIntervalHours int `json:"token_refresh_interval_hours,omitempty"`
 }
 
 type ResponsesConfig struct {

internal/config/config_edge_test.go

@@ -104,6 +104,9 @@ func TestConfigJSONRoundtrip(t *testing.T) {
 			"fast": "deepseek-chat",
 			"slow": "deepseek-reasoner",
 		},
+		Runtime: RuntimeConfig{
+			TokenRefreshIntervalHours: 12,
+		},
 		VercelSyncHash: "hash123",
 		VercelSyncTime: 1234567890,
 		AdditionalFields: map[string]any{
@@ -130,6 +133,9 @@ func TestConfigJSONRoundtrip(t *testing.T) {
 	if decoded.ClaudeMapping["fast"] != "deepseek-chat" {
 		t.Fatalf("unexpected claude mapping: %#v", decoded.ClaudeMapping)
 	}
+	if decoded.Runtime.TokenRefreshIntervalHours != 12 {
+		t.Fatalf("unexpected runtime refresh interval: %#v", decoded.Runtime.TokenRefreshIntervalHours)
+	}
 	if decoded.VercelSyncHash != "hash123" {
 		t.Fatalf("unexpected vercel sync hash: %q", decoded.VercelSyncHash)
 	}

internal/config/config_test.go

@@ -2,6 +2,7 @@ package config
 
 import (
 	"encoding/base64"
+	"errors"
 	"os"
 	"strings"
 	"testing"
@@ -79,6 +80,136 @@ func TestLoadStorePreservesFileBackedTokensForRuntime(t *testing.T) {
 	}
 }
 
+func TestEnvBackedStoreWritebackBootstrapsMissingConfigFile(t *testing.T) {
+	tmp, err := os.CreateTemp(t.TempDir(), "config-*.json")
+	if err != nil {
+		t.Fatalf("create temp config: %v", err)
+	}
+	path := tmp.Name()
+	_ = tmp.Close()
+	_ = os.Remove(path)
+
+	t.Setenv("DS2API_CONFIG_JSON", `{"keys":["k1"],"accounts":[{"email":"seed@example.com","password":"p"}]}`)
+	t.Setenv("CONFIG_JSON", "")
+	t.Setenv("DS2API_CONFIG_PATH", path)
+	t.Setenv("DS2API_ENV_WRITEBACK", "1")
+
+	store := LoadStore()
+	if store.IsEnvBacked() {
+		t.Fatalf("expected writeback bootstrap to become file-backed immediately")
+	}
+	if err := store.Update(func(c *Config) error {
+		c.Accounts = append(c.Accounts, Account{Email: "new@example.com", Password: "p2"})
+		return nil
+	}); err != nil {
+		t.Fatalf("update failed: %v", err)
+	}
+	content, err := os.ReadFile(path)
+	if err != nil {
+		t.Fatalf("read written config: %v", err)
+	}
+	if !strings.Contains(string(content), "seed@example.com") {
+		t.Fatalf("expected bootstrapped config to contain seed account, got: %s", content)
+	}
+	if !strings.Contains(string(content), "new@example.com") {
+		t.Fatalf("expected persisted config to contain added account, got: %s", content)
+	}
+
+	reloaded := LoadStore()
+	if reloaded.IsEnvBacked() {
+		t.Fatalf("expected reloaded store to prefer persisted config file")
+	}
+	accounts := reloaded.Accounts()
+	if len(accounts) != 2 {
+		t.Fatalf("expected 2 accounts after reload, got %d", len(accounts))
+	}
+}
+
+func TestEnvBackedStoreWritebackDoesNotBootstrapOnInvalidEnvJSON(t *testing.T) {
+	tmp, err := os.CreateTemp(t.TempDir(), "config-*.json")
+	if err != nil {
+		t.Fatalf("create temp config: %v", err)
+	}
+	path := tmp.Name()
+	_ = tmp.Close()
+	_ = os.Remove(path)
+
+	t.Setenv("DS2API_CONFIG_JSON", "{invalid-json")
+	t.Setenv("CONFIG_JSON", "")
+	t.Setenv("DS2API_CONFIG_PATH", path)
+	t.Setenv("DS2API_ENV_WRITEBACK", "1")
+
+	cfg, fromEnv, loadErr := loadConfig()
+	if loadErr == nil {
+		t.Fatalf("expected loadConfig error for invalid env json")
+	}
+	if !fromEnv {
+		t.Fatalf("expected fromEnv=true when parsing env config fails")
+	}
+	if len(cfg.Keys) != 0 || len(cfg.Accounts) != 0 {
+		t.Fatalf("expected empty config on parse failure, got keys=%d accounts=%d", len(cfg.Keys), len(cfg.Accounts))
+	}
+	if _, statErr := os.Stat(path); !errors.Is(statErr, os.ErrNotExist) {
+		t.Fatalf("expected no bootstrapped config file, stat err=%v", statErr)
+	}
+}
+
+func TestEnvBackedStoreWritebackFallsBackToPersistedFileOnInvalidEnvJSON(t *testing.T) {
+	tmp, err := os.CreateTemp(t.TempDir(), "config-*.json")
+	if err != nil {
+		t.Fatalf("create temp config: %v", err)
+	}
+	path := tmp.Name()
+	if _, err := tmp.WriteString(`{"keys":["file-key"],"accounts":[{"email":"persisted@example.com","password":"p"}]}`); err != nil {
+		t.Fatalf("write temp config: %v", err)
+	}
+	_ = tmp.Close()
+
+	t.Setenv("DS2API_CONFIG_JSON", "{invalid-json")
+	t.Setenv("CONFIG_JSON", "")
+	t.Setenv("DS2API_CONFIG_PATH", path)
+	t.Setenv("DS2API_ENV_WRITEBACK", "1")
+
+	cfg, fromEnv, loadErr := loadConfig()
+	if loadErr != nil {
+		t.Fatalf("expected fallback to persisted file, got error: %v", loadErr)
+	}
+	if fromEnv {
+		t.Fatalf("expected fallback to file-backed mode")
+	}
+	if len(cfg.Keys) != 1 || cfg.Keys[0] != "file-key" {
+		t.Fatalf("unexpected keys after fallback: %#v", cfg.Keys)
+	}
+	if len(cfg.Accounts) != 1 || cfg.Accounts[0].Email != "persisted@example.com" {
+		t.Fatalf("unexpected accounts after fallback: %#v", cfg.Accounts)
+	}
+}
+
+func TestRuntimeTokenRefreshIntervalHoursDefaultsToSix(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["k1"],
+		"accounts":[{"email":"u@example.com","password":"p"}]
+	}`)
+
+	store := LoadStore()
+	if got := store.RuntimeTokenRefreshIntervalHours(); got != 6 {
+		t.Fatalf("expected default refresh interval 6, got %d", got)
+	}
+}
+
+func TestRuntimeTokenRefreshIntervalHoursUsesConfigValue(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{
+		"keys":["k1"],
+		"accounts":[{"email":"u@example.com","password":"p"}],
+		"runtime":{"token_refresh_interval_hours":9}
+	}`)
+
+	store := LoadStore()
+	if got := store.RuntimeTokenRefreshIntervalHours(); got != 9 {
+		t.Fatalf("expected configured refresh interval 9, got %d", got)
+	}
+}
+
 func TestStoreUpdateAccountTokenKeepsIdentifierResolvable(t *testing.T) {
 	t.Setenv("DS2API_CONFIG_JSON", `{
 		"accounts":[{"email":"user@example.com","password":"p"}]

internal/config/store.go

@@ -40,12 +40,38 @@ func loadConfig() (Config, bool, error) {
 	}
 	if rawCfg != "" {
 		cfg, err := parseConfigString(rawCfg)
+		if err != nil {
+			if !IsVercel() && envWritebackEnabled() {
+				if fileCfg, fileErr := loadConfigFromFile(ConfigPath()); fileErr == nil {
+					return fileCfg, false, nil
+				}
+			}
+			return cfg, true, err
+		}
 		cfg.ClearAccountTokens()
 		cfg.DropInvalidAccounts()
+		if IsVercel() || !envWritebackEnabled() {
+			return cfg, true, err
+		}
+		content, fileErr := os.ReadFile(ConfigPath())
+		if fileErr == nil {
+			var fileCfg Config
+			if unmarshalErr := json.Unmarshal(content, &fileCfg); unmarshalErr == nil {
+				fileCfg.DropInvalidAccounts()
+				return fileCfg, false, err
+			}
+		}
+		if errors.Is(fileErr, os.ErrNotExist) {
+			if writeErr := writeConfigFile(ConfigPath(), cfg.Clone()); writeErr == nil {
+				return cfg, false, err
+			} else {
+				Logger.Warn("[config] env writeback bootstrap failed", "error", writeErr)
+			}
+		}
 		return cfg, true, err
 	}
 
-	content, err := os.ReadFile(ConfigPath())
+	cfg, err := loadConfigFromFile(ConfigPath())
 	if err != nil {
 		if IsVercel() {
 			// Vercel one-click deploy may start without a writable/present config file.
@@ -54,16 +80,6 @@ func loadConfig() (Config, bool, error) {
 		}
 		return Config{}, false, err
 	}
-	var cfg Config
-	if err := json.Unmarshal(content, &cfg); err != nil {
-		return Config{}, false, err
-	}
-	cfg.DropInvalidAccounts()
-	if strings.Contains(string(content), `"test_status"`) && !IsVercel() {
-		if b, err := json.MarshalIndent(cfg, "", "  "); err == nil {
-			_ = os.WriteFile(ConfigPath(), b, 0o644)
-		}
-	}
 	if IsVercel() {
 		// Vercel filesystem is ephemeral/read-only for runtime writes; avoid save errors.
 		return cfg, true, nil
@@ -71,6 +87,24 @@ func loadConfig() (Config, bool, error) {
 	return cfg, false, nil
 }
 
+func loadConfigFromFile(path string) (Config, error) {
+	content, err := os.ReadFile(path)
+	if err != nil {
+		return Config{}, err
+	}
+	var cfg Config
+	if err := json.Unmarshal(content, &cfg); err != nil {
+		return Config{}, err
+	}
+	cfg.DropInvalidAccounts()
+	if strings.Contains(string(content), `"test_status"`) && !IsVercel() {
+		if b, err := json.MarshalIndent(cfg, "", "  "); err == nil {
+			_ = os.WriteFile(path, b, 0o644)
+		}
+	}
+	return cfg, nil
+}
+
 func (s *Store) Snapshot() Config {
 	s.mu.RLock()
 	defer s.mu.RUnlock()
@@ -177,7 +211,7 @@ func (s *Store) Update(mutator func(*Config) error) error {
 func (s *Store) Save() error {
 	s.mu.Lock()
 	defer s.mu.Unlock()
-	if s.fromEnv {
+	if s.fromEnv && (IsVercel() || !envWritebackEnabled()) {
 		Logger.Info("[save_config] source from env, skip write")
 		return nil
 	}
@@ -187,11 +221,15 @@ func (s *Store) Save() error {
 	if err != nil {
 		return err
 	}
-	return os.WriteFile(s.path, b, 0o644)
+	if err := writeConfigBytes(s.path, b); err != nil {
+		return err
+	}
+	s.fromEnv = false
+	return nil
 }
 
 func (s *Store) saveLocked() error {
-	if s.fromEnv {
+	if s.fromEnv && (IsVercel() || !envWritebackEnabled()) {
 		Logger.Info("[save_config] source from env, skip write")
 		return nil
 	}
@@ -201,7 +239,11 @@ func (s *Store) saveLocked() error {
 	if err != nil {
 		return err
 	}
-	return os.WriteFile(s.path, b, 0o644)
+	if err := writeConfigBytes(s.path, b); err != nil {
+		return err
+	}
+	s.fromEnv = false
+	return nil
 }
 
 func (s *Store) IsEnvBacked() bool {

internal/config/store_accessors.go

@@ -43,23 +43,11 @@ func (s *Store) CompatWideInputStrictOutput() bool {
 }
 
 func (s *Store) ToolcallMode() string {
-	s.mu.RLock()
-	defer s.mu.RUnlock()
-	mode := strings.TrimSpace(strings.ToLower(s.cfg.Toolcall.Mode))
-	if mode == "" {
-		return "feature_match"
-	}
-	return mode
+	return "feature_match"
 }
 
 func (s *Store) ToolcallEarlyEmitConfidence() string {
-	s.mu.RLock()
-	defer s.mu.RUnlock()
-	level := strings.TrimSpace(strings.ToLower(s.cfg.Toolcall.EarlyEmitConfidence))
-	if level == "" {
-		return "high"
-	}
-	return level
+	return "high"
 }
 
 func (s *Store) ResponsesStoreTTLSeconds() int {
@@ -166,6 +154,15 @@ func (s *Store) RuntimeGlobalMaxInflight(defaultSize int) int {
 	return defaultSize
 }
 
+func (s *Store) RuntimeTokenRefreshIntervalHours() int {
+	s.mu.RLock()
+	defer s.mu.RUnlock()
+	if s.cfg.Runtime.TokenRefreshIntervalHours > 0 {
+		return s.cfg.Runtime.TokenRefreshIntervalHours
+	}
+	return 6
+}
+
 func (s *Store) AutoDeleteSessions() bool {
 	s.mu.RLock()
 	defer s.mu.RUnlock()

internal/config/store_env_writeback.go

@@ -0,0 +1,51 @@
+package config
+
+import (
+	"encoding/json"
+	"fmt"
+	"os"
+	"path/filepath"
+	"strings"
+)
+
+func envWritebackEnabled() bool {
+	v := strings.ToLower(strings.TrimSpace(os.Getenv("DS2API_ENV_WRITEBACK")))
+	return v == "1" || v == "true" || v == "yes" || v == "on"
+}
+
+func (s *Store) IsEnvWritebackEnabled() bool {
+	return envWritebackEnabled()
+}
+
+func (s *Store) HasEnvConfigSource() bool {
+	rawCfg := strings.TrimSpace(os.Getenv("DS2API_CONFIG_JSON"))
+	if rawCfg == "" {
+		rawCfg = strings.TrimSpace(os.Getenv("CONFIG_JSON"))
+	}
+	return rawCfg != ""
+}
+
+func (s *Store) ConfigPath() string {
+	return s.path
+}
+
+func writeConfigFile(path string, cfg Config) error {
+	persistCfg := cfg.Clone()
+	persistCfg.ClearAccountTokens()
+	b, err := json.MarshalIndent(persistCfg, "", "  ")
+	if err != nil {
+		return err
+	}
+	return writeConfigBytes(path, b)
+}
+
+func writeConfigBytes(path string, b []byte) error {
+	dir := filepath.Dir(path)
+	if dir == "." || dir == "" {
+		return os.WriteFile(path, b, 0o644)
+	}
+	if err := os.MkdirAll(dir, 0o755); err != nil {
+		return fmt.Errorf("mkdir config dir: %w", err)
+	}
+	return os.WriteFile(path, b, 0o644)
+}

internal/js/chat-stream/toolcall_policy.js

@@ -12,12 +12,10 @@ function resolveToolcallPolicy(prepBody, payloadTools) {
   if (toolNames.length === 0 && Array.isArray(payloadTools) && payloadTools.length > 0) {
     toolNames = ['__any_tool__'];
   }
-  const featureMatchEnabled = boolDefaultTrue(prepBody && prepBody.toolcall_feature_match);
-  const emitEarlyToolDeltas = featureMatchEnabled && boolDefaultTrue(prepBody && prepBody.toolcall_early_emit_high);
   return {
     toolNames,
     toolSieveEnabled: toolNames.length > 0,
-    emitEarlyToolDeltas,
+    emitEarlyToolDeltas: true,
   };
 }
 

internal/js/helpers/stream-tool-sieve/jsonscan.js

@@ -140,9 +140,33 @@ function extractJSONObjectFrom(text, start) {
   return { ok: false, end: 0 };
 }
 
+function trimWrappingJSONFence(prefix, suffix) {
+  const rightTrimmedPrefix = (prefix || '').replace(/[ \t\r\n]+$/g, '');
+  const fenceIdx = rightTrimmedPrefix.lastIndexOf('```');
+  if (fenceIdx < 0) return { prefix, suffix };
+  const fenceCount = (rightTrimmedPrefix.slice(0, fenceIdx + 3).match(/```/g) || []).length;
+  if (fenceCount % 2 === 0) {
+    return { prefix, suffix };
+  }
+  const header = rightTrimmedPrefix.slice(fenceIdx + 3).trim().toLowerCase();
+  if (header && header !== 'json') {
+    return { prefix, suffix };
+  }
+  const leftTrimmedSuffix = (suffix || '').replace(/^[ \t\r\n]+/g, '');
+  if (!leftTrimmedSuffix.startsWith('```')) {
+    return { prefix, suffix };
+  }
+  const consumed = (suffix || '').length - leftTrimmedSuffix.length;
+  return {
+    prefix: rightTrimmedPrefix.slice(0, fenceIdx),
+    suffix: (suffix || '').slice(consumed + 3),
+  };
+}
+
 module.exports = {
   findObjectFieldValueStart,
   parseJSONStringLiteral,
   skipSpaces,
   extractJSONObjectFrom,
+  trimWrappingJSONFence,
 };

internal/js/helpers/stream-tool-sieve/parse.js

@@ -52,11 +52,26 @@ function parseToolCallsDetailed(text, toolNames) {
   }
 
   const candidates = buildToolCallCandidates(normalized);
+  for (const c of candidates) {
+    if (!isLikelyJSONToolPayloadCandidate(c)) {
+      continue;
+    }
+    const jsonParsed = parseToolCallsPayload(c);
+    if (jsonParsed.length === 0) {
+      continue;
+    }
+    result.sawToolCallSyntax = true;
+    const filteredJSON = filterToolCallsDetailed(jsonParsed, toolNames);
+    result.calls = filteredJSON.calls;
+    result.rejectedToolNames = filteredJSON.rejectedToolNames;
+    result.rejectedByPolicy = filteredJSON.rejectedToolNames.length > 0 && filteredJSON.calls.length === 0;
+    return result;
+  }
   let parsed = [];
   for (const c of candidates) {
-    parsed = parseToolCallsPayload(c);
+    parsed = parseMarkupToolCalls(c);
     if (parsed.length === 0) {
-      parsed = parseMarkupToolCalls(c);
+      parsed = parseToolCallsPayload(c);
     }
     if (parsed.length === 0) {
       parsed = parseTextKVToolCalls(c);
@@ -101,9 +116,24 @@ function parseStandaloneToolCallsDetailed(text, toolNames) {
   const candidates = buildToolCallCandidates(trimmed);
   let parsed = [];
   for (const c of candidates) {
+    if (!isLikelyJSONToolPayloadCandidate(c)) {
+      continue;
+    }
     parsed = parseToolCallsPayload(c);
     if (parsed.length === 0) {
-      parsed = parseMarkupToolCalls(c);
+      continue;
+    }
+    result.sawToolCallSyntax = true;
+    const filteredJSON = filterToolCallsDetailed(parsed, toolNames);
+    result.calls = filteredJSON.calls;
+    result.rejectedToolNames = filteredJSON.rejectedToolNames;
+    result.rejectedByPolicy = filteredJSON.rejectedToolNames.length > 0 && filteredJSON.calls.length === 0;
+    return result;
+  }
+  for (const c of candidates) {
+    parsed = parseMarkupToolCalls(c);
+    if (parsed.length === 0) {
+      parsed = parseToolCallsPayload(c);
     }
     if (parsed.length === 0) {
       parsed = parseTextKVToolCalls(c);
@@ -198,6 +228,18 @@ function shouldSkipToolCallParsingForCodeFenceExample(text) {
   return !looksLikeToolCallSyntax(stripped);
 }
 
+function isLikelyJSONToolPayloadCandidate(text) {
+  const trimmed = toStringSafe(text).trim();
+  if (!trimmed) {
+    return false;
+  }
+  if (!(trimmed.startsWith('{') || trimmed.startsWith('['))) {
+    return false;
+  }
+  const lower = trimmed.toLowerCase();
+  return lower.includes('tool_calls') || lower.includes('"function"');
+}
+
 module.exports = {
   extractToolNames,
   parseToolCalls,

internal/js/helpers/stream-tool-sieve/parse_payload.js

@@ -6,6 +6,8 @@ const TOOL_CALL_MARKUP_SELFCLOSE_PATTERN = /<(?:[a-z0-9_:-]+:)?invoke\b([^>]*)\/
 const TOOL_CALL_MARKUP_KV_PATTERN = /<(?:[a-z0-9_:-]+:)?([a-z0-9_.-]+)\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?\1>/gi;
 const TOOL_CALL_MARKUP_ATTR_PATTERN = /(name|function|tool)\s*=\s*"([^"]+)"/i;
 const TOOL_CALL_MARKUP_NAME_PATTERNS = [
+  /<(?:[a-z0-9_:-]+:)?tool_name\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?tool_name>/i,
+  /<(?:[a-z0-9_:-]+:)?function_name\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?function_name>/i,
   /<(?:[a-z0-9_:-]+:)?name\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?name>/i,
   /<(?:[a-z0-9_:-]+:)?function\b[^>]*>([\s\S]*?)<\/(?:[a-z0-9_:-]+:)?function>/i,
 ];
@@ -100,7 +102,10 @@ function extractToolCallObjects(text) {
       const obj = extractJSONObjectFrom(raw, start);
       if (obj.ok) {
         out.push(raw.slice(start, obj.end).trim());
-        offset = obj.end;
+        // Ensure forward progress even when the matched keyword is outside
+        // the extracted JSON object (e.g. closing XML wrapper tags containing
+        // "tool_calls" after an earlier JSON arguments object).
+        offset = Math.max(obj.end, idx + matched.length);
         idx = -1;
         break;
       }

internal/js/helpers/stream-tool-sieve/sieve-xml.js

@@ -0,0 +1,100 @@
+'use strict';
+const { parseToolCalls } = require('./parse');
+
+// Tag pairs ordered longest-first: wrapper tags checked before inner tags.
+const XML_TOOL_TAG_PAIRS = [
+  { open: '<tool_calls', close: '</tool_calls>' },
+  { open: '<tool_call', close: '</tool_call>' },
+  { open: '<function_calls', close: '</function_calls>' },
+  { open: '<function_call', close: '</function_call>' },
+  { open: '<invoke', close: '</invoke>' },
+  { open: '<tool_use', close: '</tool_use>' },
+];
+
+const XML_TOOL_OPENING_TAGS = XML_TOOL_TAG_PAIRS.map(p => p.open);
+
+function consumeXMLToolCapture(captured, toolNames, trimWrappingJSONFence) {
+  const lower = captured.toLowerCase();
+  // Find the FIRST matching open/close pair, preferring wrapper tags.
+  for (const pair of XML_TOOL_TAG_PAIRS) {
+    const openIdx = lower.indexOf(pair.open);
+    if (openIdx < 0) {
+      continue;
+    }
+    // Find the LAST occurrence of the specific closing tag.
+    const closeIdx = lower.lastIndexOf(pair.close);
+    if (closeIdx < openIdx) {
+      // Opening tag present but specific closing tag hasn't arrived.
+      // Return not-ready — do NOT fall through to inner pairs.
+      return { ready: false, prefix: '', calls: [], suffix: '' };
+    }
+    const closeEnd = closeIdx + pair.close.length;
+    const xmlBlock = captured.slice(openIdx, closeEnd);
+    let prefixPart = captured.slice(0, openIdx);
+    let suffixPart = captured.slice(closeEnd);
+    const parsed = parseToolCalls(xmlBlock, toolNames);
+    if (Array.isArray(parsed) && parsed.length > 0) {
+      const trimmedFence = trimWrappingJSONFence(prefixPart, suffixPart);
+      return {
+        ready: true,
+        prefix: trimmedFence.prefix,
+        calls: parsed,
+        suffix: trimmedFence.suffix,
+      };
+    }
+    // XML tool syntax but failed to parse — consume to avoid leak.
+    return { ready: true, prefix: prefixPart, calls: [], suffix: suffixPart };
+  }
+  return { ready: false, prefix: '', calls: [], suffix: '' };
+}
+
+function hasOpenXMLToolTag(captured) {
+  const lower = captured.toLowerCase();
+  for (const pair of XML_TOOL_TAG_PAIRS) {
+    if (lower.includes(pair.open)) {
+      if (!lower.includes(pair.close)) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+function findPartialXMLToolTagStart(s) {
+  const lastLT = s.lastIndexOf('<');
+  if (lastLT < 0) {
+    return -1;
+  }
+  const tail = s.slice(lastLT);
+  if (tail.includes('>')) {
+    return -1;
+  }
+  const lowerTail = tail.toLowerCase();
+  for (const tag of XML_TOOL_OPENING_TAGS) {
+    const tagWithLT = tag.startsWith('<') ? tag : '<' + tag;
+    if (tagWithLT.startsWith(lowerTail)) {
+      return lastLT;
+    }
+  }
+  return -1;
+}
+
+function looksLikeXMLToolTagFragment(s) {
+  const trimmed = (s || '').trim();
+  if (!trimmed) return false;
+  const lower = trimmed.toLowerCase();
+  const fragments = [
+    'tool_calls>', 'tool_call>', '/tool_calls>', '/tool_call>',
+    'function_calls>', 'function_call>', '/function_calls>', '/function_call>',
+    'invoke>', '/invoke>', 'tool_use>', '/tool_use>',
+    'tool_name>', '/tool_name>', 'parameters>', '/parameters>',
+  ];
+  return fragments.some(f => lower.includes(f));
+}
+
+module.exports = {
+  consumeXMLToolCapture,
+  hasOpenXMLToolTag,
+  findPartialXMLToolTagStart,
+  looksLikeXMLToolTagFragment,
+};

internal/js/helpers/stream-tool-sieve/sieve.js

@@ -5,8 +5,18 @@ const {
   insideCodeFenceWithState,
 } = require('./state');
 const { parseStandaloneToolCallsDetailed } = require('./parse');
-const { extractJSONObjectFrom } = require('./jsonscan');
-const { TOOL_SEGMENT_KEYWORDS, earliestKeywordIndex } = require('./tool-keywords');
+const { extractJSONObjectFrom, trimWrappingJSONFence } = require('./jsonscan');
+const {
+  TOOL_SEGMENT_KEYWORDS,
+  XML_TOOL_SEGMENT_TAGS,
+  earliestKeywordIndex,
+} = require('./tool-keywords');
+const {
+  consumeXMLToolCapture: consumeXMLToolCaptureImpl,
+  hasOpenXMLToolTag,
+  findPartialXMLToolTagStart,
+  looksLikeXMLToolTagFragment,
+} = require('./sieve-xml');
 function processToolSieveChunk(state, chunk, toolNames) {
   if (!state) {
     return [];
@@ -106,16 +116,21 @@ function flushToolSieve(state, toolNames) {
         events.push({ type: 'text', text: consumed.suffix });
       }
     } else if (state.capture) {
-      noteText(state, state.capture);
-      events.push({ type: 'text', text: state.capture });
+      const content = state.capture;
+      if (!hasOpenXMLToolTag(content) && !looksLikeXMLToolTagFragment(content)) {
+        noteText(state, content);
+        events.push({ type: 'text', text: content });
+      }
     }
     state.capture = '';
     state.capturing = false;
     resetIncrementalToolState(state);
   }
   if (state.pending) {
-    noteText(state, state.pending);
-    events.push({ type: 'text', text: state.pending });
+    if (!hasOpenXMLToolTag(state.pending) && !looksLikeXMLToolTagFragment(state.pending)) {
+      noteText(state, state.pending);
+      events.push({ type: 'text', text: state.pending });
+    }
     state.pending = '';
   }
   return events;
@@ -144,6 +159,11 @@ function findSuspiciousPrefixStart(s) {
       start = idx;
     }
   }
+  // Also check for partial XML tool tag at end of string.
+  const xmlIdx = findPartialXMLToolTagStart(s);
+  if (xmlIdx >= 0 && xmlIdx > start) {
+    start = xmlIdx;
+  }
   return start;
 }
 
@@ -154,13 +174,35 @@ function findToolSegmentStart(state, s) {
   const lower = s.toLowerCase();
   let offset = 0;
   while (true) {
-    const { index: bestKeyIdx, keyword: matchedKeyword } = earliestKeywordIndex(lower, TOOL_SEGMENT_KEYWORDS, offset);
+    // Check JSON keywords.
+    let { index: bestKeyIdx, keyword: matchedKeyword } = earliestKeywordIndex(lower, TOOL_SEGMENT_KEYWORDS, offset);
+    // Also check XML tool tags.
+    for (const tag of XML_TOOL_SEGMENT_TAGS) {
+      const idx = lower.indexOf(tag, offset);
+      if (idx >= 0 && (bestKeyIdx < 0 || idx < bestKeyIdx)) {
+        bestKeyIdx = idx;
+        matchedKeyword = tag;
+      }
+    }
     if (bestKeyIdx < 0) {
       return -1;
     }
+    // For XML tags, the '<' is itself the segment start.
+    if (s[bestKeyIdx] === '<') {
+      if (!insideCodeFenceWithState(state, s.slice(0, bestKeyIdx))) {
+        return bestKeyIdx;
+      }
+      offset = bestKeyIdx + matchedKeyword.length;
+      continue;
+    }
     const keyIdx = bestKeyIdx;
     const start = s.slice(0, keyIdx).lastIndexOf('{');
-    const candidateStart = start >= 0 ? start : keyIdx;
+    let candidateStart = start >= 0 ? start : keyIdx;
+    // If the keyword matched inside an XML tag (e.g. "tool_calls" in "<tool_calls>"),
+    // back up past the '<' to capture the full tag.
+    if (candidateStart > 0 && s[candidateStart - 1] === '<') {
+      candidateStart--;
+    }
     if (!insideCodeFenceWithState(state, s.slice(0, candidateStart))) {
       return candidateStart;
     }
@@ -173,6 +215,17 @@ function consumeToolCapture(state, toolNames) {
   if (!captured) {
     return { ready: false, prefix: '', calls: [], suffix: '' };
   }
+
+  // Try XML tool call extraction first.
+  const xmlResult = consumeXMLToolCaptureImpl(captured, toolNames, trimWrappingJSONFence);
+  if (xmlResult.ready) {
+    return xmlResult;
+  }
+  // If XML tags are present but block is incomplete, keep buffering.
+  if (hasOpenXMLToolTag(captured)) {
+    return { ready: false, prefix: '', calls: [], suffix: '' };
+  }
+
   const lower = captured.toLowerCase();
   const { index: keyIdx } = earliestKeywordIndex(lower, TOOL_SEGMENT_KEYWORDS);
   if (keyIdx < 0) {
@@ -180,17 +233,6 @@ function consumeToolCapture(state, toolNames) {
   }
   const start = captured.slice(0, keyIdx).lastIndexOf('{');
   const actualStart = start >= 0 ? start : keyIdx;
-  if (start < 0) {
-    const history = extractToolHistoryBlock(captured, keyIdx);
-    if (history.ok) {
-      return {
-        ready: true,
-        prefix: captured.slice(0, history.start),
-        calls: [],
-        suffix: captured.slice(history.end),
-      };
-    }
-  }
   const obj = extractJSONObjectFrom(captured, actualStart);
   if (!obj.ok) {
     return { ready: false, prefix: '', calls: [], suffix: '' };
@@ -231,52 +273,6 @@ function consumeToolCapture(state, toolNames) {
   };
 }
 
-function extractToolHistoryBlock(captured, keyIdx) {
-  if (typeof captured !== 'string' || keyIdx < 0 || keyIdx >= captured.length) {
-    return { ok: false, start: 0, end: 0 };
-  }
-  const rest = captured.slice(keyIdx).toLowerCase();
-  if (rest.startsWith('[tool_call_history]')) {
-    const closeTag = '[/tool_call_history]';
-    const closeIdx = rest.indexOf(closeTag);
-    if (closeIdx < 0) {
-      return { ok: false, start: 0, end: 0 };
-    }
-    return { ok: true, start: keyIdx, end: keyIdx + closeIdx + closeTag.length };
-  }
-  if (rest.startsWith('[tool_result_history]')) {
-    const closeTag = '[/tool_result_history]';
-    const closeIdx = rest.indexOf(closeTag);
-    if (closeIdx < 0) {
-      return { ok: false, start: 0, end: 0 };
-    }
-    return { ok: true, start: keyIdx, end: keyIdx + closeIdx + closeTag.length };
-  }
-  return { ok: false, start: 0, end: 0 };
-}
-
-function trimWrappingJSONFence(prefix, suffix) {
-  const rightTrimmedPrefix = (prefix || '').replace(/[ \t\r\n]+$/g, '');
-  const fenceIdx = rightTrimmedPrefix.lastIndexOf('```');
-  if (fenceIdx < 0) return { prefix, suffix };
-  const fenceCount = (rightTrimmedPrefix.slice(0, fenceIdx + 3).match(/```/g) || []).length;
-  if (fenceCount % 2 === 0) {
-    return { prefix, suffix };
-  }
-  const header = rightTrimmedPrefix.slice(fenceIdx + 3).trim().toLowerCase();
-  if (header && header !== 'json') {
-    return { prefix, suffix };
-  }
-  const leftTrimmedSuffix = (suffix || '').replace(/^[ \t\r\n]+/g, '');
-  if (!leftTrimmedSuffix.startsWith('```')) {
-    return { prefix, suffix };
-  }
-  const consumed = (suffix || '').length - leftTrimmedSuffix.length;
-  return {
-    prefix: rightTrimmedPrefix.slice(0, fenceIdx),
-    suffix: (suffix || '').slice(consumed + 3),
-  };
-}
 module.exports = {
   processToolSieveChunk,
   flushToolSieve,

internal/js/helpers/stream-tool-sieve/state.js

@@ -1,6 +1,7 @@
 'use strict';
 
-const TOOL_SIEVE_CONTEXT_TAIL_LIMIT = 4096;
+// Keep in sync with Go toolSieveContextTailLimit.
+const TOOL_SIEVE_CONTEXT_TAIL_LIMIT = 2048;
 
 function createToolSieveState() {
   return {

internal/js/helpers/stream-tool-sieve/tool-keywords.js

@@ -4,8 +4,19 @@ const TOOL_SEGMENT_KEYWORDS = [
   'tool_calls',
   '"function"',
   'function.name:',
-  '[tool_call_history]',
-  '[tool_result_history]',
+];
+
+const XML_TOOL_SEGMENT_TAGS = [
+  '<tool_calls>', '<tool_calls\n', '<tool_call>', '<tool_call\n',
+  '<invoke ', '<invoke>', '<function_call', '<function_calls', '<tool_use>',
+];
+
+const XML_TOOL_OPENING_TAGS = [
+  '<tool_calls', '<tool_call', '<invoke', '<function_call', '<function_calls', '<tool_use',
+];
+
+const XML_TOOL_CLOSING_TAGS = [
+  '</tool_calls>', '</tool_call>', '</invoke>', '</function_call>', '</function_calls>', '</tool_use>',
 ];
 
 function earliestKeywordIndex(text, keywords = TOOL_SEGMENT_KEYWORDS, offset = 0) {
@@ -26,5 +37,8 @@ function earliestKeywordIndex(text, keywords = TOOL_SEGMENT_KEYWORDS, offset = 0
 
 module.exports = {
   TOOL_SEGMENT_KEYWORDS,
+  XML_TOOL_SEGMENT_TAGS,
+  XML_TOOL_OPENING_TAGS,
+  XML_TOOL_CLOSING_TAGS,
   earliestKeywordIndex,
 };

internal/prompt/messages.go

@@ -42,12 +42,15 @@ func MessagesPrepare(messages []map[string]any) string {
 			} else {
 				parts = append(parts, m.Text)
 			}
-		case "user", "system":
-			if i > 0 {
-				parts = append(parts, "<｜User｜>"+m.Text)
-			} else {
-				parts = append(parts, m.Text)
+		case "system":
+			// Clear system boundary improves R1 and V3 context understanding significantly
+			if strings.TrimSpace(m.Text) != "" {
+				parts = append(parts, "<system_instructions>\n"+strings.TrimSpace(m.Text)+"\n</system_instructions>\n\n")
 			}
+		case "user":
+			// Always prepend <｜User｜> to user messages. DeepSeek R1 reasoning triggers best
+			// and aligns context perfectly when the user turn is explicitly marked.
+			parts = append(parts, "<｜User｜>"+m.Text)
 		default:
 			parts = append(parts, m.Text)
 		}

internal/prompt/tool_calls.go

@@ -0,0 +1,137 @@
+package prompt
+
+import (
+	"encoding/json"
+	"strings"
+)
+
+var promptXMLTextEscaper = strings.NewReplacer(
+	"&", "&",
+	"<", "&lt;",
+	">", "&gt;",
+)
+
+// FormatToolCallsForPrompt renders a tool_calls slice into the canonical
+// prompt-visible history block used across adapters.
+func FormatToolCallsForPrompt(raw any) string {
+	calls, ok := raw.([]any)
+	if !ok || len(calls) == 0 {
+		return ""
+	}
+
+	blocks := make([]string, 0, len(calls))
+	for _, item := range calls {
+		call, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		block := formatToolCallForPrompt(call)
+		if block != "" {
+			blocks = append(blocks, block)
+		}
+	}
+	if len(blocks) == 0 {
+		return ""
+	}
+	return "<tool_calls>\n" + strings.Join(blocks, "\n") + "\n</tool_calls>"
+}
+
+// StringifyToolCallArguments normalizes tool arguments into a compact string
+// while preserving raw concatenated payloads when they already look like model
+// output rather than a single JSON object.
+func StringifyToolCallArguments(v any) string {
+	switch x := v.(type) {
+	case nil:
+		return "{}"
+	case string:
+		s := strings.TrimSpace(x)
+		if s == "" {
+			return "{}"
+		}
+		s = normalizeToolArgumentString(s)
+		if s == "" {
+			return "{}"
+		}
+		return s
+	default:
+		b, err := json.Marshal(x)
+		if err != nil || len(b) == 0 {
+			return "{}"
+		}
+		return string(b)
+	}
+}
+
+func formatToolCallForPrompt(call map[string]any) string {
+	if call == nil {
+		return ""
+	}
+
+	name := strings.TrimSpace(asString(call["name"]))
+	fn, _ := call["function"].(map[string]any)
+	if name == "" && fn != nil {
+		name = strings.TrimSpace(asString(fn["name"]))
+	}
+	if name == "" {
+		return ""
+	}
+
+	argsRaw := call["arguments"]
+	if argsRaw == nil {
+		argsRaw = call["input"]
+	}
+	if argsRaw == nil && fn != nil {
+		argsRaw = fn["arguments"]
+		if argsRaw == nil {
+			argsRaw = fn["input"]
+		}
+	}
+
+	return "  <tool_call>\n" +
+		"    <tool_name>" + escapeXMLText(name) + "</tool_name>\n" +
+		"    <parameters>" + escapeXMLText(StringifyToolCallArguments(argsRaw)) + "</parameters>\n" +
+		"  </tool_call>"
+}
+
+func normalizeToolArgumentString(raw string) string {
+	trimmed := strings.TrimSpace(raw)
+	if trimmed == "" {
+		return ""
+	}
+	if looksLikeConcatenatedJSON(trimmed) {
+		// Keep the original payload to avoid silently rewriting model output.
+		return raw
+	}
+	return trimmed
+}
+
+func looksLikeConcatenatedJSON(raw string) bool {
+	trimmed := strings.TrimSpace(raw)
+	if trimmed == "" {
+		return false
+	}
+	if strings.Contains(trimmed, "}{") || strings.Contains(trimmed, "][") {
+		return true
+	}
+	dec := json.NewDecoder(strings.NewReader(trimmed))
+	var first any
+	if err := dec.Decode(&first); err != nil {
+		return false
+	}
+	var second any
+	return dec.Decode(&second) == nil
+}
+
+func asString(v any) string {
+	if s, ok := v.(string); ok {
+		return s
+	}
+	return ""
+}
+
+func escapeXMLText(v string) string {
+	if v == "" {
+		return ""
+	}
+	return promptXMLTextEscaper.Replace(v)
+}

internal/prompt/tool_calls_test.go

@@ -0,0 +1,41 @@
+package prompt
+
+import "testing"
+
+func TestStringifyToolCallArgumentsPreservesConcatenatedJSON(t *testing.T) {
+	got := StringifyToolCallArguments(`{}{"query":"测试工具调用"}`)
+	if got != `{}{"query":"测试工具调用"}` {
+		t.Fatalf("expected raw concatenated JSON to be preserved, got %q", got)
+	}
+}
+
+func TestFormatToolCallsForPromptXML(t *testing.T) {
+	got := FormatToolCallsForPrompt([]any{
+		map[string]any{
+			"id": "call_1",
+			"function": map[string]any{
+				"name":      "search_web",
+				"arguments": map[string]any{"query": "latest"},
+			},
+		},
+	})
+	if got == "" {
+		t.Fatal("expected non-empty formatted tool calls")
+	}
+	if got != "<tool_calls>\n  <tool_call>\n    <tool_name>search_web</tool_name>\n    <parameters>{\"query\":\"latest\"}</parameters>\n  </tool_call>\n</tool_calls>" {
+		t.Fatalf("unexpected formatted tool call XML: %q", got)
+	}
+}
+
+func TestFormatToolCallsForPromptEscapesXMLEntities(t *testing.T) {
+	got := FormatToolCallsForPrompt([]any{
+		map[string]any{
+			"name":      "search<&>",
+			"arguments": `{"q":"a < b && c > d"}`,
+		},
+	})
+	want := "<tool_calls>\n  <tool_call>\n    <tool_name>search&lt;&amp;&gt;</tool_name>\n    <parameters>{\"q\":\"a &lt; b &amp;&amp; c &gt; d\"}</parameters>\n  </tool_call>\n</tool_calls>"
+	if got != want {
+		t.Fatalf("unexpected escaped tool call XML: %q", got)
+	}
+}

internal/sse/content_filter_leak.go

@@ -0,0 +1,31 @@
+package sse
+
+import "strings"
+
+func filterLeakedContentFilterParts(parts []ContentPart) []ContentPart {
+	if len(parts) == 0 {
+		return parts
+	}
+	out := make([]ContentPart, 0, len(parts))
+	for _, p := range parts {
+		cleaned := stripLeakedContentFilterSuffix(p.Text)
+		if strings.TrimSpace(cleaned) == "" {
+			continue
+		}
+		p.Text = cleaned
+		out = append(out, p)
+	}
+	return out
+}
+
+func stripLeakedContentFilterSuffix(text string) string {
+	if text == "" {
+		return text
+	}
+	upperText := strings.ToUpper(text)
+	idx := strings.Index(upperText, "CONTENT_FILTER")
+	if idx < 0 {
+		return text
+	}
+	return strings.TrimRight(text[:idx], " \t\r\n")
+}

internal/sse/line.go

@@ -40,6 +40,7 @@ func ParseDeepSeekContentLine(raw []byte, thinkingEnabled bool, currentType stri
 		}
 	}
 	parts, finished, nextType := ParseSSEChunkForContent(chunk, thinkingEnabled, currentType)
+	parts = filterLeakedContentFilterParts(parts)
 	return LineResult{
 		Parsed:   true,
 		Stop:     finished,

internal/sse/line_test.go

@@ -35,3 +35,33 @@ func TestParseDeepSeekContentLineContent(t *testing.T) {
 		t.Fatalf("unexpected parts: %#v", res.Parts)
 	}
 }
+
+func TestParseDeepSeekContentLineStripsLeakedContentFilterSuffix(t *testing.T) {
+	res := ParseDeepSeekContentLine([]byte(`data: {"p":"response/content","v":"正常输出CONTENT_FILTER你好，这个问题我暂时无法回答"}`), false, "text")
+	if !res.Parsed || res.Stop {
+		t.Fatalf("expected parsed non-stop result: %#v", res)
+	}
+	if len(res.Parts) != 1 || res.Parts[0].Text != "正常输出" {
+		t.Fatalf("unexpected parts after filter: %#v", res.Parts)
+	}
+}
+
+func TestParseDeepSeekContentLineDropsPureLeakedContentFilterChunk(t *testing.T) {
+	res := ParseDeepSeekContentLine([]byte(`data: {"p":"response/content","v":"CONTENT_FILTER你好，这个问题我暂时无法回答"}`), false, "text")
+	if !res.Parsed || res.Stop {
+		t.Fatalf("expected parsed non-stop result: %#v", res)
+	}
+	if len(res.Parts) != 0 {
+		t.Fatalf("expected empty parts, got %#v", res.Parts)
+	}
+}
+
+func TestParseDeepSeekContentLineTrimsFromContentFilterKeyword(t *testing.T) {
+	res := ParseDeepSeekContentLine([]byte(`data: {"p":"response/content","v":"模型会在命中 CONTENT_FILTER 时返回拒绝原因。"}`), false, "text")
+	if !res.Parsed || res.Stop {
+		t.Fatalf("expected parsed non-stop result: %#v", res)
+	}
+	if len(res.Parts) != 1 || res.Parts[0].Text != "模型会在命中" {
+		t.Fatalf("unexpected parts after filter: %#v", res.Parts)
+	}
+}

internal/util/messages_test.go

@@ -12,7 +12,7 @@ func TestMessagesPrepareBasic(t *testing.T) {
 	if got == "" {
 		t.Fatal("expected non-empty prompt")
 	}
-	if got != "Hello" {
+	if got != "<｜User｜>Hello" {
 		t.Fatalf("unexpected prompt: %q", got)
 	}
 }
@@ -55,7 +55,7 @@ func TestMessagesPrepareArrayTextVariants(t *testing.T) {
 		},
 	}
 	got := MessagesPrepare(messages)
-	if got != "line1\nline2" {
+	if got != "<｜User｜>line1\nline2" {
 		t.Fatalf("unexpected content from text variants: %q", got)
 	}
 }

internal/util/tool_prompt.go

@@ -0,0 +1,99 @@
+package util
+
+// BuildToolCallInstructions generates the unified tool-calling instruction block
+// used by all adapters (OpenAI, Claude, Gemini). It uses attention-optimized
+// structure: rules → negative examples → positive examples → anchor.
+//
+// The toolNames slice should contain the actual tool names available in the
+// current request; the function picks real names for examples.
+func BuildToolCallInstructions(toolNames []string) string {
+	// Pick real tool names for examples; fall back to generic names.
+	ex1 := "read_file"
+	ex2 := "write_to_file"
+	ex3 := "ask_followup_question"
+	used := map[string]bool{}
+	for _, n := range toolNames {
+		switch {
+		// Read/query-type tools
+		case !used["ex1"] && matchAny(n, "read_file", "list_files", "search_files", "Read", "Glob"):
+			ex1 = n
+			used["ex1"] = true
+		// Write/execute-type tools
+		case !used["ex2"] && matchAny(n, "write_to_file", "apply_diff", "execute_command", "Write", "Edit", "MultiEdit", "Bash"):
+			ex2 = n
+			used["ex2"] = true
+		// Interactive/meta tools
+		case !used["ex3"] && matchAny(n, "ask_followup_question", "attempt_completion", "update_todo_list", "Task"):
+			ex3 = n
+			used["ex3"] = true
+		}
+	}
+
+	return `TOOL CALL FORMAT — FOLLOW EXACTLY:
+
+When calling tools, emit ONLY raw XML. No text before, no text after, no markdown fences.
+
+<tool_calls>
+  <tool_call>
+    <tool_name>TOOL_NAME_HERE</tool_name>
+    <parameters>{"key":"value"}</parameters>
+  </tool_call>
+</tool_calls>
+
+RULES:
+1) Output ONLY the XML above when calling tools. Do NOT mix tool XML with regular text.
+2) <parameters> MUST contain a strict JSON object. All JSON keys and strings use double quotes.
+3) Multiple tools → multiple <tool_call> blocks inside ONE <tool_calls> root.
+4) Do NOT wrap the XML in markdown code fences (no triple backticks).
+5) After receiving a tool result, use it directly. Only call another tool if the result is insufficient.
+6) If you want to say something AND call a tool, output text first, then the XML block on its own.
+
+❌ WRONG — Do NOT do these:
+Wrong 1 — mixed text and XML:
+  I'll read the file for you. <tool_calls><tool_call>...
+Wrong 2 — describing tool calls in text:
+  [调用 Bash] {"command": "ls"}
+Wrong 3 — missing <tool_calls> wrapper:
+  <tool_call><tool_name>` + ex1 + `</tool_name><parameters>{}</parameters></tool_call>
+
+✅ CORRECT EXAMPLES:
+
+Example A — Single tool:
+<tool_calls>
+  <tool_call>
+    <tool_name>` + ex1 + `</tool_name>
+    <parameters>{"path":"src/main.go"}</parameters>
+  </tool_call>
+</tool_calls>
+
+Example B — Two tools in parallel:
+<tool_calls>
+  <tool_call>
+    <tool_name>` + ex1 + `</tool_name>
+    <parameters>{"path":"config.json"}</parameters>
+  </tool_call>
+  <tool_call>
+    <tool_name>` + ex2 + `</tool_name>
+    <parameters>{"path":"output.txt","content":"Hello world"}</parameters>
+  </tool_call>
+</tool_calls>
+
+Example C — Tool with complex nested JSON parameters:
+<tool_calls>
+  <tool_call>
+    <tool_name>` + ex3 + `</tool_name>
+    <parameters>{"question":"Which approach do you prefer?","follow_up":[{"text":"Option A"},{"text":"Option B"}]}</parameters>
+  </tool_call>
+</tool_calls>
+
+Remember: Output ONLY the <tool_calls>...</tool_calls> XML block when calling tools.`
+}
+
+func matchAny(name string, candidates ...string) bool {
+	for _, c := range candidates {
+		if name == c {
+			return true
+		}
+	}
+	return false
+}

internal/util/toolcalls_candidates.go

@@ -64,7 +64,7 @@ func extractToolCallObjects(text string) []string {
 	lower := strings.ToLower(text)
 	out := []string{}
 	offset := 0
-	keywords := []string{"tool_calls", "\"function\"", "function.name:", "[tool_call_history]"}
+	keywords := []string{"tool_calls", "\"function\"", "function.name:"}
 	for {
 		bestIdx := -1
 		matchedKeyword := ""

internal/util/toolcalls_parse.go

@@ -32,15 +32,31 @@ func ParseToolCallsDetailed(text string, availableToolNames []string) ToolCallPa
 	}
 
 	candidates := buildToolCallCandidates(text)
-	var parsed []ParsedToolCall
 	for _, candidate := range candidates {
+		if !isLikelyJSONToolPayloadCandidate(candidate) {
+			continue
+		}
 		tc := parseToolCallsPayload(candidate)
 		if len(tc) == 0 {
-			tc = parseXMLToolCalls(candidate)
+			continue
 		}
+		parsed := tc
+		calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
+		result.Calls = calls
+		result.RejectedToolNames = rejectedNames
+		result.RejectedByPolicy = len(rejectedNames) > 0 && len(calls) == 0
+		result.SawToolCallSyntax = true
+		return result
+	}
+	var parsed []ParsedToolCall
+	for _, candidate := range candidates {
+		tc := parseXMLToolCalls(candidate)
 		if len(tc) == 0 {
 			tc = parseMarkupToolCalls(candidate)
 		}
+		if len(tc) == 0 {
+			tc = parseToolCallsPayload(candidate)
+		}
 		if len(tc) == 0 {
 			tc = parseTextKVToolCalls(candidate)
 		}
@@ -84,17 +100,32 @@ func ParseStandaloneToolCallsDetailed(text string, availableToolNames []string)
 	candidates := buildToolCallCandidates(trimmed)
 	var parsed []ParsedToolCall
 	for _, candidate := range candidates {
-		candidate = strings.TrimSpace(candidate)
-		if candidate == "" {
+		if !isLikelyJSONToolPayloadCandidate(candidate) {
 			continue
 		}
 		parsed = parseToolCallsPayload(candidate)
 		if len(parsed) == 0 {
-			parsed = parseXMLToolCalls(candidate)
+			continue
 		}
+		result.SawToolCallSyntax = true
+		calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
+		result.Calls = calls
+		result.RejectedToolNames = rejectedNames
+		result.RejectedByPolicy = len(rejectedNames) > 0 && len(calls) == 0
+		return result
+	}
+	for _, candidate := range candidates {
+		candidate = strings.TrimSpace(candidate)
+		if candidate == "" {
+			continue
+		}
+		parsed = parseXMLToolCalls(candidate)
 		if len(parsed) == 0 {
 			parsed = parseMarkupToolCalls(candidate)
 		}
+		if len(parsed) == 0 {
+			parsed = parseToolCallsPayload(candidate)
+		}
 		if len(parsed) == 0 {
 			parsed = parseTextKVToolCalls(candidate)
 		}
@@ -165,6 +196,18 @@ func parseToolCallsPayload(payload string) []ParsedToolCall {
 	return nil
 }
 
+func isLikelyJSONToolPayloadCandidate(candidate string) bool {
+	trimmed := strings.TrimSpace(candidate)
+	if trimmed == "" {
+		return false
+	}
+	if !(strings.HasPrefix(trimmed, "{") || strings.HasPrefix(trimmed, "[")) {
+		return false
+	}
+	lower := strings.ToLower(trimmed)
+	return strings.Contains(lower, "tool_calls") || strings.Contains(lower, "\"function\"")
+}
+
 func isLikelyChatMessageEnvelope(v map[string]any) bool {
 	if v == nil {
 		return false

internal/util/toolcalls_parse_markup.go

@@ -104,6 +104,34 @@ func parseSingleXMLToolCall(block string) (ParsedToolCall, bool) {
 				}
 			case "parameters":
 				inParams = true
+				var node struct {
+					Inner string `xml:",innerxml"`
+				}
+				if err := dec.DecodeElement(&node, &t); err == nil {
+					inner := strings.TrimSpace(node.Inner)
+					if inner != "" {
+						if parsed := parseToolCallInput(inner); len(parsed) > 0 {
+							if len(parsed) == 1 {
+								if _, onlyRaw := parsed["_raw"]; onlyRaw {
+									if kv := parseMarkupKVObject(inner); len(kv) > 0 {
+										for k, vv := range kv {
+											params[k] = vv
+										}
+										break
+									}
+								}
+							}
+							for k, vv := range parsed {
+								params[k] = vv
+							}
+						} else if kv := parseMarkupKVObject(inner); len(kv) > 0 {
+							for k, vv := range kv {
+								params[k] = vv
+							}
+						}
+					}
+				}
+				inParams = false
 			case "tool_name", "name":
 				var v string
 				if err := dec.DecodeElement(&v, &t); err == nil && strings.TrimSpace(v) != "" {

internal/util/toolcalls_test.go

@@ -162,6 +162,34 @@ func TestParseToolCallsSupportsClaudeXMLToolCall(t *testing.T) {
 	}
 }
 
+func TestParseToolCallsSupportsCanonicalXMLParametersJSON(t *testing.T) {
+	text := `<tool_call><tool_name>get_weather</tool_name><parameters>{"city":"beijing","unit":"c"}</parameters></tool_call>`
+	calls := ParseToolCalls(text, []string{"get_weather"})
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "get_weather" {
+		t.Fatalf("expected tool name get_weather, got %q", calls[0].Name)
+	}
+	if calls[0].Input["city"] != "beijing" || calls[0].Input["unit"] != "c" {
+		t.Fatalf("expected parsed json parameters, got %#v", calls[0].Input)
+	}
+}
+
+func TestParseToolCallsPrefersJSONPayloadOverIncidentalXMLInString(t *testing.T) {
+	text := `{"tool_calls":[{"name":"search","input":{"q":"latest <tool_call><tool_name>wrong</tool_name><parameters>{\"x\":1}</parameters></tool_call>"}}]}`
+	calls := ParseToolCallsDetailed(text, []string{"search"}).Calls
+	if len(calls) != 1 {
+		t.Fatalf("expected 1 call, got %#v", calls)
+	}
+	if calls[0].Name != "search" {
+		t.Fatalf("expected tool name search, got %q", calls[0].Name)
+	}
+	if calls[0].Input["q"] == nil {
+		t.Fatalf("expected q argument from json payload, got %#v", calls[0].Input)
+	}
+}
+
 func TestParseToolCallsDetailedMarksXMLToolCallSyntax(t *testing.T) {
 	text := `<tool_call><tool_name>Bash</tool_name><parameters><command>pwd</command></parameters></tool_call>`
 	res := ParseToolCallsDetailed(text, []string{"bash"})

internal/util/toolcalls_textkv_test.go

@@ -6,14 +6,12 @@ import (
 
 func TestParseTextKVToolCalls_Basic(t *testing.T) {
 	text := `
-[TOOL_CALL_HISTORY]
 status: already_called
 origin: assistant
 not_user_input: true
 tool_call_id: call_3fcd15235eb94f7eae3a8de5a9cfa36b
 function.name: execute_command
 function.arguments: {"command":"cd scripts && python check_syntax.py example.py","cwd":null,"timeout":30}
-[/TOOL_CALL_HISTORY]
 
 Some other text thinking...
 `

internal/util/util_edge_test.go

@@ -162,13 +162,16 @@ func TestMessagesPrepareMergesConsecutiveSameRole(t *testing.T) {
 		{"role": "user", "content": "World"},
 	}
 	got := MessagesPrepare(messages)
+	if !strings.HasPrefix(got, "<｜User｜>") {
+		t.Fatalf("expected user marker at the start, got %q", got)
+	}
 	if !strings.Contains(got, "Hello") || !strings.Contains(got, "World") {
 		t.Fatalf("expected both messages, got %q", got)
 	}
-	// Should be merged without <｜User｜> between them
+	// Should be merged into a single user turn with one marker at the start.
 	count := strings.Count(got, "<｜User｜>")
-	if count != 0 {
-		t.Fatalf("expected no User marker for first message pair, got %d occurrences", count)
+	if count != 1 {
+		t.Fatalf("expected one User marker for the merged pair, got %d occurrences", count)
 	}
 }
 

plans/node-syntax-gate-targets.txt

@@ -16,6 +16,8 @@ internal/js/helpers/stream-tool-sieve.js
 internal/js/helpers/stream-tool-sieve/index.js
 internal/js/helpers/stream-tool-sieve/state.js
 internal/js/helpers/stream-tool-sieve/sieve.js
+internal/js/helpers/stream-tool-sieve/sieve-xml.js
 internal/js/helpers/stream-tool-sieve/jsonscan.js
 internal/js/helpers/stream-tool-sieve/parse.js
 internal/js/helpers/stream-tool-sieve/format.js
+internal/js/helpers/stream-tool-sieve/tool-keywords.js

plans/refactor-line-gate-targets.txt

@@ -53,6 +53,7 @@ internal/adapter/openai/responses_stream_runtime_events.go
 internal/adapter/openai/responses_stream_runtime_toolcalls.go
 internal/adapter/openai/tool_sieve_state.go
 internal/adapter/openai/tool_sieve_core.go
+internal/adapter/openai/tool_sieve_xml.go
 internal/adapter/openai/tool_sieve_jsonscan.go
 
 internal/util/toolcalls_parse.go
@@ -106,6 +107,7 @@ internal/js/helpers/stream-tool-sieve.js
 internal/js/helpers/stream-tool-sieve/index.js
 internal/js/helpers/stream-tool-sieve/state.js
 internal/js/helpers/stream-tool-sieve/sieve.js
+internal/js/helpers/stream-tool-sieve/sieve-xml.js
 internal/js/helpers/stream-tool-sieve/jsonscan.js
 internal/js/helpers/stream-tool-sieve/parse.js
 internal/js/helpers/stream-tool-sieve/format.js

tests/compat/expected/toolcalls_json_payload_with_incidental_xml_text.json

@@ -0,0 +1,13 @@
+{
+  "calls": [
+    {
+      "name": "search",
+      "input": {
+        "q": "latest <tool_call><tool_name>wrong</tool_name><parameters>{\"x\":1}</parameters></tool_call>"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/expected/toolcalls_xml_tool_name_parameters_json.json

@@ -0,0 +1,14 @@
+{
+  "calls": [
+    {
+      "name": "get_weather",
+      "input": {
+        "city": "beijing",
+        "unit": "c"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
+  "rejectedByPolicy": false,
+  "rejectedToolNames": []
+}

tests/compat/fixtures/toolcalls/json_payload_with_incidental_xml_text.json

@@ -0,0 +1,6 @@
+{
+  "text": "{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"latest <tool_call><tool_name>wrong</tool_name><parameters>{\\\"x\\\":1}</parameters></tool_call>\"}}]}",
+  "tool_names": [
+    "search"
+  ]
+}

tests/compat/fixtures/toolcalls/xml_tool_name_parameters_json.json

@@ -0,0 +1,6 @@
+{
+  "text": "<tool_call><tool_name>get_weather</tool_name><parameters>{\"city\":\"beijing\",\"unit\":\"c\"}</parameters></tool_call>",
+  "tool_names": [
+    "get_weather"
+  ]
+}

tests/node/chat-stream.test.js

@@ -34,7 +34,7 @@ test('resolveToolcallPolicy defaults to feature-match + early emit when prepare
   assert.equal(policy.emitEarlyToolDeltas, true);
 });
 
-test('resolveToolcallPolicy respects prepare flags and prepared tool names', () => {
+test('resolveToolcallPolicy ignores prepare flags and keeps early emit enabled', () => {
   const policy = resolveToolcallPolicy(
     {
       tool_names: [' prepped_tool ', '', null],
@@ -45,7 +45,7 @@ test('resolveToolcallPolicy respects prepare flags and prepared tool names', ()
   );
   assert.deepEqual(policy.toolNames, ['prepped_tool']);
   assert.equal(policy.toolSieveEnabled, true);
-  assert.equal(policy.emitEarlyToolDeltas, false);
+  assert.equal(policy.emitEarlyToolDeltas, true);
 });
 
 test('normalizePreparedToolNames filters empty values', () => {

tests/node/stream-tool-sieve.test.js

@@ -98,10 +98,8 @@ test('parseToolCalls ignores tool_call payloads that exist only inside fenced co
 
 test('parseToolCalls parses text-kv fallback payload', () => {
   const text = [
-    '[TOOL_CALL_HISTORY]',
     'function.name: execute_command',
     'function.arguments: {"command":"cd scripts && python check_syntax.py example.py","cwd":null,"timeout":30}',
-    '[/TOOL_CALL_HISTORY]',
     'Some other text thinking...',
   ].join('\n');
   const calls = parseToolCalls(text, ['execute_command']);
@@ -213,6 +211,40 @@ test('sieve flushes incomplete captured tool json as text on stream finalize', (
   assert.equal(leakedText.includes('{'), true);
 });
 
+test('sieve flushes incomplete captured XML tool blocks without leaking raw tags', () => {
+  const events = runSieve(
+    [
+      '前置正文G。',
+      '<tool_calls>\n',
+      '  <tool_call>\n',
+      '    <tool_name>read_file</tool_name>\n',
+    ],
+    ['read_file'],
+  );
+  const leakedText = collectText(events);
+  assert.equal(leakedText.includes('前置正文G。'), true);
+  assert.equal(leakedText.toLowerCase().includes('tool_calls'), false);
+  assert.equal(leakedText.includes('<tool_call'), false);
+});
+
+test('sieve captures XML wrapper tags with attributes without leaking wrapper text', () => {
+  const events = runSieve(
+    [
+      '前置正文H。',
+      '<tool_calls id="x"><tool_call><tool_name>read_file</tool_name><parameters>{"path":"README.MD"}</parameters></tool_call></tool_calls>',
+      '后置正文I。',
+    ],
+    ['read_file'],
+  );
+  const leakedText = collectText(events);
+  const hasToolCall = events.some((evt) => evt.type === 'tool_calls' && evt.calls?.length > 0);
+  assert.equal(hasToolCall, true);
+  assert.equal(leakedText.includes('前置正文H。'), true);
+  assert.equal(leakedText.includes('后置正文I。'), true);
+  assert.equal(leakedText.includes('<tool_calls id=\"x\">'), false);
+  assert.equal(leakedText.includes('</tool_calls>'), false);
+});
+
 test('sieve still intercepts large tool json payloads over previous capture limit', () => {
   const large = 'a'.repeat(9000);
   const payload = `{"tool_calls":[{"name":"read_file","input":{"path":"${large}"}}]}`;
@@ -238,54 +270,44 @@ test('sieve keeps plain text intact in tool mode when no tool call appears', ()
   assert.equal(leakedText, '你好，这是普通文本回复。请继续。');
 });
 
-test('sieve swallows leaked TOOL_CALL_HISTORY marker blocks', () => {
+test('sieve keeps plain "tool_calls" prose as text when no valid payload follows', () => {
   const events = runSieve(
-    [
-      '前置文本。',
-      '[TOOL_CALL_HISTORY]\nstatus: already_called\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_CALL_HISTORY]',
-      '后置文本。',
-    ],
-    ['exec'],
-  );
-  const leakedText = collectText(events);
-  const hasToolCall = events.some((evt) => evt.type === 'tool_calls');
-  assert.equal(hasToolCall, false);
-  assert.equal(leakedText.includes('前置文本。'), true);
-  assert.equal(leakedText.includes('后置文本。'), true);
-  assert.equal(leakedText.includes('[TOOL_CALL_HISTORY]'), false);
-});
-
-test('sieve swallows leaked TOOL_RESULT_HISTORY marker blocks', () => {
-  const events = runSieve(
-    [
-      '前置文本。',
-      '[TOOL_RESULT_HISTORY]\nstatus: already_called\nfunction.name: exec\nfunction.arguments: {}\n[/TOOL_RESULT_HISTORY]',
-      '后置文本。',
-    ],
-    ['exec'],
-  );
-  const leakedText = collectText(events);
-  const hasToolCall = events.some((evt) => evt.type === 'tool_calls');
-  assert.equal(hasToolCall, false);
-  assert.equal(leakedText.includes('前置文本。'), true);
-  assert.equal(leakedText.includes('后置文本。'), true);
-  assert.equal(leakedText.includes('[TOOL_RESULT_HISTORY]'), false);
-});
-
-test('sieve preserves text spacing when TOOL_RESULT_HISTORY spans chunks', () => {
-  const events = runSieve(
-    [
-      'Hello ',
-      '[TOOL_RESULT_HISTORY]\nstatus: already_called\n',
-      'function.name: exec\nfunction.arguments: {}\n[/TOOL_RESULT_HISTORY]',
-      'world',
-    ],
-    ['exec'],
+    ['前置。', '这里提到 tool_calls 只是解释，不是调用。', '后置。'],
+    ['read_file'],
   );
   const leakedText = collectText(events);
   const hasToolCall = events.some((evt) => evt.type === 'tool_calls' && evt.calls?.length > 0);
   assert.equal(hasToolCall, false);
-  assert.equal(leakedText, 'Hello world');
+  assert.equal(leakedText.includes('tool_calls'), true);
+  assert.equal(leakedText, '前置。这里提到 tool_calls 只是解释，不是调用。后置。');
+});
+
+test('sieve keeps numbered planning prose before a real tool payload (mobile-chat style)', () => {
+  const events = runSieve(
+    [
+      '好的，我会依次测试每个工具，先把所有工具都调用一遍，然后汇总结果给你看。\n\n1. 获取当前时间\n',
+      '{"tool_calls":[{"name":"get_current_time","input":{}}]}',
+    ],
+    ['get_current_time'],
+  );
+  const leakedText = collectText(events);
+  const finalCalls = events.filter((evt) => evt.type === 'tool_calls').flatMap((evt) => evt.calls || []);
+  assert.equal(finalCalls.length, 1);
+  assert.equal(finalCalls[0].name, 'get_current_time');
+  assert.equal(leakedText.includes('先把所有工具都调用一遍'), true);
+  assert.equal(leakedText.includes('1. 获取当前时间'), true);
+  assert.equal(leakedText.toLowerCase().includes('tool_calls'), false);
+});
+
+test('sieve keeps numbered planning prose when no tool payload follows', () => {
+  const events = runSieve(
+    ['好的，我会依次测试每个工具。\n\n1. 获取当前时间'],
+    ['get_current_time'],
+  );
+  const leakedText = collectText(events);
+  const hasToolCall = events.some((evt) => evt.type === 'tool_calls' && evt.calls?.length > 0);
+  assert.equal(hasToolCall, false);
+  assert.equal(leakedText, '好的，我会依次测试每个工具。\n\n1. 获取当前时间');
 });
 
 test('sieve emits unknown tool payload (no args) as executable tool call', () => {

vercel.json

@@ -4,7 +4,7 @@
   "outputDirectory": "static",
   "functions": {
     "api/chat-stream.js": {
-      "includeFiles": "**/sha3_wasm_bg.7b9ca65ddd.wasm",
+      "includeFiles": "internal/deepseek/assets/sha3_wasm_bg.7b9ca65ddd.wasm",
       "maxDuration": 300
     },
     "api/index.go": {

webui/src/features/account/AccountManagerContainer.jsx

@@ -64,6 +64,27 @@ export default function AccountManagerContainer({ config, onRefresh, onMessage,
 
     return (
         <div className="space-y-6">
+            {Boolean(config?.env_source_present) && (
+                <div className={`rounded-xl border px-4 py-3 text-sm ${
+                    config?.env_writeback_enabled
+                        ? (config?.env_backed ? 'border-amber-500/30 bg-amber-500/10 text-amber-600' : 'border-emerald-500/30 bg-emerald-500/10 text-emerald-600')
+                        : 'border-amber-500/30 bg-amber-500/10 text-amber-600'
+                }`}>
+                    <p className="font-medium">
+                        {config?.env_writeback_enabled
+                            ? (config?.env_backed
+                                ? t('accountManager.envModeWritebackPendingTitle')
+                                : t('accountManager.envModeWritebackActiveTitle'))
+                            : t('accountManager.envModeRiskTitle')}
+                    </p>
+                    <p className="mt-1 text-xs opacity-90">
+                        {config?.env_writeback_enabled
+                            ? t('accountManager.envModeWritebackDesc', { path: config?.config_path || 'config.json' })
+                            : t('accountManager.envModeRiskDesc')}
+                    </p>
+                </div>
+            )}
+
             <QueueCards queueStatus={queueStatus} t={t} />
 
             <ApiKeysPanel

webui/src/features/settings/BehaviorSection.jsx

@@ -3,35 +3,6 @@ export default function BehaviorSection({ t, form, setForm }) {
         <div className="bg-card border border-border rounded-xl p-5 space-y-4">
             <h3 className="font-semibold">{t('settings.behaviorTitle')}</h3>
             <div className="grid grid-cols-1 md:grid-cols-2 gap-4">
-                <label className="text-sm space-y-2">
-                    <span className="text-muted-foreground">{t('settings.toolcallMode')}</span>
-                    <select
-                        value={form.toolcall.mode}
-                        onChange={(e) => setForm((prev) => ({
-                            ...prev,
-                            toolcall: { ...prev.toolcall, mode: e.target.value },
-                        }))}
-                        className="w-full bg-background border border-border rounded-lg px-3 py-2"
-                    >
-                        <option value="feature_match">feature_match</option>
-                        <option value="off">off</option>
-                    </select>
-                </label>
-                <label className="text-sm space-y-2">
-                    <span className="text-muted-foreground">{t('settings.earlyEmitConfidence')}</span>
-                    <select
-                        value={form.toolcall.early_emit_confidence}
-                        onChange={(e) => setForm((prev) => ({
-                            ...prev,
-                            toolcall: { ...prev.toolcall, early_emit_confidence: e.target.value },
-                        }))}
-                        className="w-full bg-background border border-border rounded-lg px-3 py-2"
-                    >
-                        <option value="high">high</option>
-                        <option value="low">low</option>
-                        <option value="off">off</option>
-                    </select>
-                </label>
                 <label className="text-sm space-y-2">
                     <span className="text-muted-foreground">{t('settings.responsesTTL')}</span>
                     <input

webui/src/features/settings/RuntimeSection.jsx

@@ -2,7 +2,7 @@ export default function RuntimeSection({ t, form, setForm }) {
     return (
         <div className="bg-card border border-border rounded-xl p-5 space-y-4">
             <h3 className="font-semibold">{t('settings.runtimeTitle')}</h3>
-            <div className="grid grid-cols-1 md:grid-cols-3 gap-4">
+            <div className="grid grid-cols-1 md:grid-cols-2 lg:grid-cols-4 gap-4">
                 <label className="text-sm space-y-2">
                     <span className="text-muted-foreground">{t('settings.accountMaxInflight')}</span>
                     <input
@@ -42,6 +42,21 @@ export default function RuntimeSection({ t, form, setForm }) {
                         className="w-full bg-background border border-border rounded-lg px-3 py-2"
                     />
                 </label>
+                <label className="text-sm space-y-2">
+                    <span className="text-muted-foreground">{t('settings.tokenRefreshIntervalHours')}</span>
+                    <input
+                        type="number"
+                        min={1}
+                        max={720}
+                        step={1}
+                        value={form.runtime.token_refresh_interval_hours}
+                        onChange={(e) => setForm((prev) => ({
+                            ...prev,
+                            runtime: { ...prev.runtime, token_refresh_interval_hours: Number(e.target.value || 1) },
+                        }))}
+                        className="w-full bg-background border border-border rounded-lg px-3 py-2"
+                    />
+                </label>
             </div>
         </div>
     )

webui/src/features/settings/useSettingsForm.js

@@ -12,8 +12,7 @@ const MAX_AUTO_FETCH_FAILURES = 3
 
 const DEFAULT_FORM = {
     admin: { jwt_expire_hours: 24 },
-    runtime: { account_max_inflight: 2, account_max_queue: 10, global_max_inflight: 10 },
-    toolcall: { mode: 'feature_match', early_emit_confidence: 'high' },
+    runtime: { account_max_inflight: 2, account_max_queue: 10, global_max_inflight: 10, token_refresh_interval_hours: 6 },
     responses: { store_ttl_seconds: 900 },
     embeddings: { provider: '' },
     auto_delete: { sessions: false },
@@ -45,10 +44,7 @@ function fromServerForm(data) {
             account_max_inflight: Number(data.runtime?.account_max_inflight || 2),
             account_max_queue: Number(data.runtime?.account_max_queue || 10),
             global_max_inflight: Number(data.runtime?.global_max_inflight || 10),
-        },
-        toolcall: {
-            mode: data.toolcall?.mode || 'feature_match',
-            early_emit_confidence: data.toolcall?.early_emit_confidence || 'high',
+            token_refresh_interval_hours: Number(data.runtime?.token_refresh_interval_hours || 6),
         },
         responses: {
             store_ttl_seconds: Number(data.responses?.store_ttl_seconds || 900),
@@ -71,10 +67,7 @@ function toServerPayload(form) {
             account_max_inflight: Number(form.runtime.account_max_inflight),
             account_max_queue: Number(form.runtime.account_max_queue),
             global_max_inflight: Number(form.runtime.global_max_inflight),
-        },
-        toolcall: {
-            mode: String(form.toolcall.mode || '').trim(),
-            early_emit_confidence: String(form.toolcall.early_emit_confidence || '').trim(),
+            token_refresh_interval_hours: Number(form.runtime.token_refresh_interval_hours),
         },
         responses: { store_ttl_seconds: Number(form.responses.store_ttl_seconds) },
         embeddings: { provider: String(form.embeddings.provider || '').trim() },

webui/src/locales/en.json

@@ -138,7 +138,12 @@
         "sessionCount": "Sessions: {count}",
         "deleteAllSessions": "Delete all sessions",
         "deleteAllSessionsConfirm": "Are you sure you want to delete all sessions for this account? This action cannot be undone.",
-        "deleteAllSessionsSuccess": "Successfully deleted all sessions"
+        "deleteAllSessionsSuccess": "Successfully deleted all sessions",
+        "envModeRiskTitle": "Environment-variable config mode detected (persistence risk)",
+        "envModeRiskDesc": "Detected DS2API_CONFIG_JSON/CONFIG_JSON. If DS2API_ENV_WRITEBACK is not enabled, Admin UI edits are in-memory only and may be lost after restart.",
+        "envModeWritebackPendingTitle": "Env mode + auto-persistence enabled (pending file handoff)",
+        "envModeWritebackActiveTitle": "Env mode + auto-persistence active",
+        "envModeWritebackDesc": "The app will auto-create/write the config file and transition to file-backed mode. Current persistence path: {path}"
     },
     "apiTester": {
         "defaultMessage": "Hello, please introduce yourself in one sentence.",
@@ -222,13 +227,12 @@
         "passwordTooShort": "Password must be at least 4 characters.",
         "passwordUpdated": "Password updated. Please sign in again.",
         "passwordUpdateFailed": "Failed to update password.",
-        "runtimeTitle": "Concurrency & Queue",
+        "runtimeTitle": "Runtime",
         "accountMaxInflight": "Per-account max inflight",
         "accountMaxQueue": "Account max queue size",
         "globalMaxInflight": "Global max inflight",
+        "tokenRefreshIntervalHours": "Managed token refresh interval (hours)",
         "behaviorTitle": "Behavior",
-        "toolcallMode": "Toolcall mode",
-        "earlyEmitConfidence": "Early emit confidence",
         "responsesTTL": "Responses store TTL (seconds)",
         "embeddingsProvider": "Embeddings provider",
         "modelTitle": "Model mapping",

webui/src/locales/zh.json

@@ -138,7 +138,12 @@
         "sessionCount": "会话: {count}",
         "deleteAllSessions": "删除所有会话",
         "deleteAllSessionsConfirm": "确定要删除该账号的所有会话吗？此操作不可恢复。",
-        "deleteAllSessionsSuccess": "删除成功"
+        "deleteAllSessionsSuccess": "删除成功",
+        "envModeRiskTitle": "当前为环境变量配置模式（有持久化风险）",
+        "envModeRiskDesc": "检测到 DS2API_CONFIG_JSON/CONFIG_JSON。若未开启 DS2API_ENV_WRITEBACK，管理台改动仅在内存生效，重启可能丢失。",
+        "envModeWritebackPendingTitle": "环境变量模式 + 自动持久化已开启（等待落盘）",
+        "envModeWritebackActiveTitle": "环境变量模式 + 自动持久化已生效",
+        "envModeWritebackDesc": "程序会自动创建/写入配置文件并在后续切换为文件模式。当前持久化路径：{path}"
     },
     "apiTester": {
         "defaultMessage": "你好，请用一句话介绍你自己。",
@@ -222,13 +227,12 @@
         "passwordTooShort": "新密码至少 4 位",
         "passwordUpdated": "密码已更新，需重新登录",
         "passwordUpdateFailed": "密码更新失败",
-        "runtimeTitle": "并发与队列",
+        "runtimeTitle": "运行时设置",
         "accountMaxInflight": "每账号并发上限",
         "accountMaxQueue": "账号等待队列上限",
         "globalMaxInflight": "全局并发上限",
+        "tokenRefreshIntervalHours": "托管账号 Token 刷新间隔（小时）",
         "behaviorTitle": "行为设置",
-        "toolcallMode": "Toolcall 模式",
-        "earlyEmitConfidence": "早发置信度",
         "responsesTTL": "Responses 缓存 TTL（秒）",
         "embeddingsProvider": "Embeddings Provider",
         "modelTitle": "模型映射",