Merge pull request #111 from CJackHwang/dev

Merge pull request #110 from CJackHwang/codex/align-js-runtime-with-go-runtime-logic Align Vercel JS stream tool-call delta handling with Go runtime
Merge pull request #110 from CJackHwang/codex/align-js-runtime-with-go-runtime-logic
2026-05-04 00:15:28 +08:00 · 2026-03-20 10:05:25 +08:00 · 2026-03-20 09:49:08 +08:00 · 2026-03-20 09:47:22 +08:00 · 2026-03-20 09:36:45 +08:00 · 2026-03-20 03:12:25 +08:00
42 changed files with 1052 additions and 555 deletions
--- a/.env.example
+++ b/.env.example
@@ -1,93 +1,15 @@
-# DS2API environment template (Go runtime)
-# Copy this file to .env and adjust values.
-# Updated: 2026-02
-
-# ---------------------------------------------------------------
-# Runtime
-# ---------------------------------------------------------------
-# HTTP listen port (default: 5001)
+# DS2API runtime
 PORT=5001
-
-# Log level: DEBUG | INFO | WARN | ERROR
 LOG_LEVEL=INFO

-# Max concurrent inflight requests per account in managed-key mode.
-# Default: 2
-# Recommended client concurrency is calculated dynamically as:
-#   account_count * DS2API_ACCOUNT_MAX_INFLIGHT
-# So by default it is account_count * 2.
-# Requests beyond inflight slots enter a waiting queue first.
-# Default queue size equals recommended concurrency, so 429 starts after:
-#   account_count * DS2API_ACCOUNT_MAX_INFLIGHT * 2
-# Alias: DS2API_ACCOUNT_CONCURRENCY
-# DS2API_ACCOUNT_MAX_INFLIGHT=2
+# Admin authentication
+DS2API_ADMIN_KEY=change-me

-# Optional waiting queue size override for managed-key mode.
-# Default: recommended_concurrency (same as account_count * inflight_limit)
-# Alias: DS2API_ACCOUNT_QUEUE_SIZE
-# DS2API_ACCOUNT_MAX_QUEUE=10
+# Config loading (choose one)
+# 1) file-based config
+DS2API_CONFIG_PATH=/app/config.json
+# 2) inline JSON or Base64 JSON
+# DS2API_CONFIG_JSON=

-# ---------------------------------------------------------------
-# Admin auth
-# ---------------------------------------------------------------
-# Admin key for /admin login and protected admin APIs.
-# Default is "admin" when unset, but setting it explicitly is recommended.
-DS2API_ADMIN_KEY=admin
-
-# Optional JWT signing secret for admin token.
-# Defaults to DS2API_ADMIN_KEY when unset.
-# DS2API_JWT_SECRET=change-me
-
-# Optional admin JWT validity in hours (default: 24)
-# DS2API_JWT_EXPIRE_HOURS=24
-
-# ---------------------------------------------------------------
-# Config source (choose one)
-# ---------------------------------------------------------------
-# Option A: config file path (local/dev recommended)
-# DS2API_CONFIG_PATH=config.json
-
-# Option B: JSON string
-# DS2API_CONFIG_JSON={"keys":["your-api-key"],"accounts":[{"email":"user@example.com","password":"xxx","token":""}]}
-
-# Option C: Base64 encoded JSON (recommended for Vercel env var)
-# DS2API_CONFIG_JSON=eyJrZXlzIjpbInlvdXItYXBpLWtleSJdLCJhY2NvdW50cyI6W3siZW1haWwiOiJ1c2VyQGV4YW1wbGUuY29tIiwicGFzc3dvcmQiOiJ4eHgiLCJ0b2tlbiI6IiJ9XX0=
-#
-# Generate from local config.json:
-#   DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
-
-# ---------------------------------------------------------------
-# Paths (optional)
-# ---------------------------------------------------------------
-# WASM file used for PoW solving
-# DS2API_WASM_PATH=sha3_wasm_bg.7b9ca65ddd.wasm
-
-# Built admin static assets directory
-# DS2API_STATIC_ADMIN_DIR=static/admin
-
-# Auto-build WebUI on startup when static/admin is missing.
-# Default: enabled on local/Docker, disabled on Vercel.
-# DS2API_AUTO_BUILD_WEBUI=true
-
-# Internal auth secret used by the Vercel hybrid streaming path
-# (Go prepare endpoint <-> Node stream function).
-# Optional: falls back to DS2API_ADMIN_KEY when unset.
-# DS2API_VERCEL_INTERNAL_SECRET=change-me
-
-# Stream lease TTL seconds for Vercel hybrid streaming.
-# During this window, the managed account stays occupied until Node calls release.
-# Default: 900 (15 minutes)
-# DS2API_VERCEL_STREAM_LEASE_TTL_SECONDS=900
-
-# ---------------------------------------------------------------
-# Vercel sync integration (optional)
-# ---------------------------------------------------------------
-# VERCEL_TOKEN=your-vercel-token
-# VERCEL_PROJECT_ID=prj_xxxxxxxxxxxx
-# VERCEL_TEAM_ID=team_xxxxxxxxxxxx
-
-# Optional: Vercel deployment protection bypass secret.
-# If deployment protection is enabled, DS2API will use this value as
-# x-vercel-protection-bypass for internal Node->Go calls on Vercel.
-# You can also use VERCEL_AUTOMATION_BYPASS_SECRET directly.
-# DS2API_VERCEL_PROTECTION_BYPASS=your-bypass-secret
+# Optional: static admin assets path
+# DS2API_STATIC_ADMIN_DIR=/app/static/admin
--- a/CONTRIBUTING.en.md
+++ b/CONTRIBUTING.en.md
@@ -99,7 +99,7 @@ ds2api/
 ├── api/
 │   ├── index.go             # Vercel Serverless Go entry
 │   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── helpers/             # Node.js helper modules
+│   └── (rewrite targets in vercel.json)
 ├── internal/
 │   ├── account/             # Account pool and concurrency queue
 │   ├── adapter/
@@ -112,6 +112,7 @@ ds2api/
 │   ├── compat/              # Compatibility helpers
 │   ├── config/              # Config loading and hot-reload
 │   ├── deepseek/            # DeepSeek client, PoW WASM
+│   ├── js/                  # Node runtime stream/compat logic
 │   ├── devcapture/          # Dev packet capture
 │   ├── format/              # Output formatting
 │   ├── prompt/              # Prompt building
@@ -123,7 +124,9 @@ ds2api/
 │   └── webui/               # WebUI static hosting
 ├── webui/                   # React WebUI source
 │   └── src/
-│       ├── components/      # Components
+│       ├── app/             # Routing, auth, config state
+│       ├── features/        # Feature modules
+│       ├── components/      # Shared components
 │       └── locales/         # Language packs
 ├── scripts/                 # Build and test scripts
 ├── static/admin/            # WebUI build output (not committed)
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -99,7 +99,7 @@ ds2api/
 ├── api/
 │   ├── index.go             # Vercel Serverless Go 入口
 │   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── helpers/             # Node.js 辅助模块
+│   └── (rewrite targets in vercel.json)
 ├── internal/
 │   ├── account/             # 账号池与并发队列
 │   ├── adapter/
@@ -112,6 +112,7 @@ ds2api/
 │   ├── compat/              # 兼容性辅助
 │   ├── config/              # 配置加载与热更新
 │   ├── deepseek/            # DeepSeek 客户端、PoW WASM
+│   ├── js/                  # Node 运行时流式/兼容逻辑
 │   ├── devcapture/          # 开发抓包
 │   ├── format/              # 输出格式化
 │   ├── prompt/              # Prompt 构建
@@ -123,7 +124,9 @@ ds2api/
 │   └── webui/               # WebUI 静态托管
 ├── webui/                   # React WebUI 源码
 │   └── src/
-│       ├── components/      # 组件
+│       ├── app/             # 路由、鉴权、配置状态
+│       ├── features/        # 业务功能模块
+│       ├── components/      # 通用组件
 │       └── locales/         # 语言包
 ├── scripts/                 # 构建与测试脚本
 ├── static/admin/            # WebUI 构建产物（不提交）
--- a/DEPLOY.en.md
+++ b/DEPLOY.en.md
@@ -113,12 +113,8 @@ go build -o ds2api ./cmd/ds2api
 # Copy env template
 cp .env.example .env

-# Generate single-line Base64 from config.json
-DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
-
-# Edit .env and set:
+# Edit .env and set at least:
 #   DS2API_ADMIN_KEY=your-admin-key
-#   DS2API_CONFIG_JSON=${DS2API_CONFIG_JSON}

 # Start
 docker-compose up -d
@@ -366,7 +362,7 @@ Each archive includes:

 - `ds2api` executable (`ds2api.exe` on Windows)
 - `static/admin/` (built WebUI assets)
- `sha3_wasm_bg.7b9ca65ddd.wasm`
+- `sha3_wasm_bg.7b9ca65ddd.wasm` (optional; binary has embedded fallback)
 - `config.example.json`, `.env.example`
 - `README.MD`, `README.en.md`, `LICENSE`

@@ -455,7 +451,9 @@ server {
 ```bash
 # Copy compiled binary and related files to target directory
 sudo mkdir -p /opt/ds2api
-sudo cp ds2api config.json sha3_wasm_bg.7b9ca65ddd.wasm /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/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```

--- a/DEPLOY.md
+++ b/DEPLOY.md
@@ -113,12 +113,8 @@ go build -o ds2api ./cmd/ds2api
 # 复制环境变量模板
 cp .env.example .env

-# 从 config.json 生成单行 Base64
-DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
-
-# 编辑 .env（请改成你的强密码），设置：
+# 编辑 .env（请改成你的强密码），至少设置：
 #   DS2API_ADMIN_KEY=your-admin-key
-#   DS2API_CONFIG_JSON=${DS2API_CONFIG_JSON}

 # 启动
 docker-compose up -d
@@ -366,7 +362,7 @@ No Output Directory named "public" found after the Build completed.

 - `ds2api` 可执行文件（Windows 为 `ds2api.exe`）
 - `static/admin/`（WebUI 构建产物）
- `sha3_wasm_bg.7b9ca65ddd.wasm`
+- `sha3_wasm_bg.7b9ca65ddd.wasm`（可选；程序内置 embed fallback）
 - `config.example.json`、`.env.example`
 - `README.MD`、`README.en.md`、`LICENSE`

@@ -455,7 +451,9 @@ server {
 ```bash
 # 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 sha3_wasm_bg.7b9ca65ddd.wasm /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/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```

--- a/README.MD
+++ b/README.MD
@@ -160,17 +160,13 @@ go run ./cmd/ds2api
 # 1. 准备环境变量文件
 cp .env.example .env

-# 2. 从 config.json 生成 DS2API_CONFIG_JSON（单行 Base64）
-DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
-
-# 3. 编辑 .env，设置：
+# 2. 编辑 .env（至少设置 DS2API_ADMIN_KEY）
 #    DS2API_ADMIN_KEY=请替换为强密码
-#    DS2API_CONFIG_JSON=${DS2API_CONFIG_JSON}

-# 4. 启动
+# 3. 启动
 docker-compose up -d

-# 5. 查看日志
+# 4. 查看日志
 docker-compose logs -f
 ```

@@ -397,7 +393,7 @@ ds2api/
 ├── api/
 │   ├── index.go             # Vercel Serverless Go 入口
 │   ├── chat-stream.js       # Vercel Node.js 流式转发
-│   └── helpers/             # Node.js 辅助模块
+│   └── (rewrite targets in vercel.json)
 ├── internal/
 │   ├── account/             # 账号池与并发队列
 │   ├── adapter/
@@ -410,6 +406,7 @@ ds2api/
 │   ├── compat/              # 兼容性辅助
 │   ├── config/              # 配置加载与热更新
 │   ├── deepseek/            # DeepSeek API 客户端、PoW WASM
+│   ├── js/                  # Node 运行时流式处理与兼容逻辑
 │   ├── devcapture/          # 开发抓包模块
 │   ├── format/              # 输出格式化
 │   ├── prompt/              # Prompt 构建
@@ -420,7 +417,9 @@ ds2api/
 │   └── webui/               # WebUI 静态文件托管与自动构建
 ├── webui/                   # React WebUI 源码（Vite + Tailwind）
 │   └── src/
-│       ├── components/      # AccountManager / ApiTester / BatchImport / VercelSync / Login / LandingPage
+│       ├── app/             # 路由、鉴权、配置状态管理
+│       ├── features/        # 业务功能模块（account/settings/vercel/apiTester）
+│       ├── components/      # 登录/落地页等通用组件
 │       └── locales/         # 中英文语言包（zh.json / en.json）
 ├── scripts/
 │   └── build-webui.sh       # WebUI 手动构建脚本
@@ -476,6 +475,23 @@ go run ./cmd/ds2api-tests \
 npm ci --prefix webui && npm run build --prefix webui
 ```

+## 测试
+
+详细测试指南请参阅 [TESTING.md](TESTING.md)。
+
+### 快速测试命令
+
+```bash
+# 运行所有单元测试
+go test ./...
+
+# 运行 tool calls 相关测试（调试工具调用问题）
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+
+# 运行端到端测试
+./tests/scripts/run-live.sh
+```
+
 ## Release 自动构建（GitHub Actions）

 工作流文件：`.github/workflows/release-artifacts.yml`
@@ -483,7 +499,7 @@ npm ci --prefix webui && npm run build --prefix webui
 - **触发条件**：仅在 GitHub Release `published` 时触发（普通 push 不会触发）
 - **构建产物**：多平台二进制包（`linux/amd64`、`linux/arm64`、`darwin/amd64`、`darwin/arm64`、`windows/amd64`）+ `sha256sums.txt`
 - **容器镜像发布**：仅推送到 GHCR（`ghcr.io/cjackhwang/ds2api`）
- **每个压缩包包含**：`ds2api` 可执行文件、`static/admin`、WASM 文件、配置示例、README、LICENSE
+- **每个压缩包包含**：`ds2api` 可执行文件、`static/admin`、WASM 文件（同时支持内置 fallback）、配置示例、README、LICENSE

 ## 免责声明

--- a/README.en.md
+++ b/README.en.md
@@ -160,17 +160,13 @@ Default URL: `http://localhost:5001`
 # 1. Prepare env file
 cp .env.example .env

-# 2. Generate DS2API_CONFIG_JSON from config.json (single-line Base64)
-DS2API_CONFIG_JSON="$(base64 < config.json | tr -d '\n')"
-
-# 3. Edit .env and set:
+# 2. Edit .env (at least set DS2API_ADMIN_KEY)
 #    DS2API_ADMIN_KEY=replace-with-a-strong-secret
-#    DS2API_CONFIG_JSON=${DS2API_CONFIG_JSON}

-# 4. Start
+# 3. Start
 docker-compose up -d

-# 5. View logs
+# 4. View logs
 docker-compose logs -f
 ```

@@ -398,7 +394,7 @@ ds2api/
 ├── api/
 │   ├── index.go             # Vercel Serverless Go entry
 │   ├── chat-stream.js       # Vercel Node.js stream relay
-│   └── helpers/             # Node.js helper modules
+│   └── (rewrite targets in vercel.json)
 ├── internal/
 │   ├── account/             # Account pool and concurrency queue
 │   ├── adapter/
@@ -411,6 +407,7 @@ ds2api/
 │   ├── compat/              # Compatibility helpers
 │   ├── config/              # Config loading and hot-reload
 │   ├── deepseek/            # DeepSeek API client, PoW WASM
+│   ├── js/                  # Node runtime stream/compat logic
 │   ├── devcapture/          # Dev packet capture module
 │   ├── format/              # Output formatting
 │   ├── prompt/              # Prompt construction
@@ -421,7 +418,9 @@ ds2api/
 │   └── webui/               # WebUI static file serving and auto-build
 ├── webui/                   # React WebUI source (Vite + Tailwind)
 │   └── src/
-│       ├── components/      # AccountManager / ApiTester / BatchImport / VercelSync / Login / LandingPage
+│       ├── app/             # Routing, auth, config state
+│       ├── features/        # Feature modules (account/settings/vercel/apiTester)
+│       ├── components/      # Shared UI pieces (login/landing, etc.)
 │       └── locales/         # Language packs (zh.json / en.json)
 ├── scripts/
 │   └── build-webui.sh       # Manual WebUI build script
@@ -484,7 +483,7 @@ Workflow: `.github/workflows/release-artifacts.yml`
 - **Trigger**: only on GitHub Release `published` (normal pushes do not trigger builds)
 - **Outputs**: multi-platform archives (`linux/amd64`, `linux/arm64`, `darwin/amd64`, `darwin/arm64`, `windows/amd64`) + `sha256sums.txt`
 - **Container publishing**: GHCR only (`ghcr.io/cjackhwang/ds2api`)
- **Each archive includes**: `ds2api` executable, `static/admin`, WASM file, config template, README, LICENSE
+- **Each archive includes**: `ds2api` executable, `static/admin`, WASM file (with embedded fallback support), config template, README, LICENSE

 ## Disclaimer

--- a/TESTING.md
+++ b/TESTING.md
@@ -51,7 +51,7 @@ DS2API 提供两个层级的测试：
 1. **Preflight 检查**：
   - `go test ./... -count=1`（单元测试）
   - `./tests/scripts/check-node-split-syntax.sh`（Node 拆分模块语法门禁）
-   - `node --test`（如仓库存在 Node 单测文件时执行；当前默认以 Go 测试 + Node 语法门禁为主）
+   - `node --test tests/node/stream-tool-sieve.test.js tests/node/chat-stream.test.js tests/node/js_compat_test.js`
   - `npm run build --prefix webui`（WebUI 构建检查）

 2. **隔离启动**：复制 `config.json` 到临时目录，启动独立服务进程
@@ -173,6 +173,50 @@ rg "<trace_id>" artifacts/testsuite/<run_id>/server.log
 go test ./...
 ```

+### 运行特定模块的单元测试
+
+```bash
+# 运行 tool calls 相关测试（推荐用于调试 tool call 解析问题）
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+
+# 运行单个测试用例
+go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/util/
+
+# 运行 format 相关测试
+go test -v ./internal/format/...
+
+# 运行 adapter 相关测试
+go test -v ./internal/adapter/openai/...
+```
+
+### 调试 Tool Call 问题 | Debugging Tool Call Issues
+
+当遇到 DeepSeek 工具调用解析问题时，可以使用以下方法：
+
+```bash
+# 1. 运行 tool calls 相关的所有测试
+go test -v -run 'TestParseToolCalls|TestRepair' ./internal/util/
+
+# 2. 查看测试输出中的详细调试信息
+go test -v -run TestParseToolCallsWithDeepSeekHallucination ./internal/util/ 2>&1
+
+# 3. 检查具体测试用例的修复效果
+# 测试用例位于 internal/util/toolcalls_test.go，包含：
+# - TestParseToolCallsWithDeepSeekHallucination: DeepSeek 典型幻觉输出
+# - TestRepairLooseJSONWithNestedObjects: 嵌套对象的方括号修复
+# - TestParseToolCallsWithMixedWindowsPaths: Windows 路径处理
+```
+
+### 运行 Node.js 测试
+
+```bash
+# 运行 Node 测试
+node --test tests/node/stream-tool-sieve.test.js
+
+# 或使用脚本
+./tests/scripts/run-unit-node.sh
+```
+
 ### 跑端到端测试（跳过 preflight）

 ```bash
--- a/internal/adapter/claude/handler_stream_test.go
+++ b/internal/adapter/claude/handler_stream_test.go
@@ -358,7 +358,7 @@ func TestHandleClaudeStreamRealtimeToolSafetyAcrossStructuredFormats(t *testing.
 	}
 }

-func TestHandleClaudeStreamRealtimeDoesNotStopOnUnclosedFencedToolExample(t *testing.T) {
+func TestHandleClaudeStreamRealtimePromotesUnclosedFencedToolExample(t *testing.T) {
 	h := &Handler{}
 	resp := makeClaudeSSEHTTPResponse(
 		"data: {\"p\":\"response/content\",\"v\":\"Here is an example:\\n```json\\n{\\\"tool_calls\\\":[{\\\"name\\\":\\\"Bash\\\",\\\"input\\\":{\\\"command\\\":\\\"pwd\\\"}}]}\"}",
@@ -371,22 +371,27 @@ func TestHandleClaudeStreamRealtimeDoesNotStopOnUnclosedFencedToolExample(t *tes
 	h.handleClaudeStreamRealtime(rec, req, resp, "claude-sonnet-4-5", []any{map[string]any{"role": "user", "content": "show example only"}}, false, false, []string{"Bash"})

 	frames := parseClaudeFrames(t, rec.Body.String())
+	foundToolUse := false
 	for _, f := range findClaudeFrames(frames, "content_block_start") {
 		contentBlock, _ := f.Payload["content_block"].(map[string]any)
 		if contentBlock["type"] == "tool_use" {
-			t.Fatalf("unexpected tool_use for fenced example, body=%s", rec.Body.String())
-		}
-	}
-
-	foundEndTurn := false
-	for _, f := range findClaudeFrames(frames, "message_delta") {
-		delta, _ := f.Payload["delta"].(map[string]any)
-		if delta["stop_reason"] == "end_turn" {
-			foundEndTurn = true
+			foundToolUse = true
 			break
 		}
 	}
-	if !foundEndTurn {
-		t.Fatalf("expected stop_reason=end_turn, body=%s", rec.Body.String())
+	if !foundToolUse {
+		t.Fatalf("expected tool_use for fenced example, body=%s", rec.Body.String())
+	}
+
+	foundToolStop := false
+	for _, f := range findClaudeFrames(frames, "message_delta") {
+		delta, _ := f.Payload["delta"].(map[string]any)
+		if delta["stop_reason"] == "tool_use" {
+			foundToolStop = true
+			break
+		}
+	}
+	if !foundToolStop {
+		t.Fatalf("expected stop_reason=tool_use, body=%s", rec.Body.String())
 	}
 }
--- a/internal/adapter/openai/chat_stream_runtime.go
+++ b/internal/adapter/openai/chat_stream_runtime.go
@@ -98,11 +98,11 @@ func (s *chatStreamRuntime) sendDone() {
 func (s *chatStreamRuntime) finalize(finishReason string) {
 	finalThinking := s.thinking.String()
 	finalText := s.text.String()
-	detected := util.ParseStandaloneToolCalls(finalText, s.toolNames)
-	if len(detected) > 0 && !s.toolCallsDoneEmitted {
+	detected := util.ParseStandaloneToolCallsDetailed(finalText, s.toolNames)
+	if len(detected.Calls) > 0 && !s.toolCallsDoneEmitted {
 		finishReason = "tool_calls"
 		delta := map[string]any{
-			"tool_calls": formatFinalStreamToolCallsWithStableIDs(detected, s.streamToolCallIDs),
+			"tool_calls": formatFinalStreamToolCallsWithStableIDs(detected.Calls, s.streamToolCallIDs),
 		}
 		if !s.firstChunkSent {
 			delta["role"] = "assistant"
@@ -158,7 +158,7 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 		}
 	}

-	if len(detected) > 0 || s.toolCallsEmitted {
+	if len(detected.Calls) > 0 || s.toolCallsEmitted {
 		finishReason = "tool_calls"
 	}
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
--- a/internal/adapter/openai/handler_chat.go
+++ b/internal/adapter/openai/handler_chat.go
@@ -42,7 +42,9 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 		// 2. 新请求可能获取到同一账号并开始使用
 		// 3. 异步删除仍在进行，会截断新请求正在使用的会话
 		if h.Store.AutoDeleteSessions() && a.DeepSeekToken != "" {
-			err := h.DS.DeleteAllSessionsForToken(context.Background(), a.DeepSeekToken)
+			deleteCtx, cancel := context.WithTimeout(r.Context(), 10*time.Second)
+			defer cancel()
+			err := h.DS.DeleteAllSessionsForToken(deleteCtx, a.DeepSeekToken)
 			if err != nil {
 				config.Logger.Warn("[auto_delete_sessions] failed", "account", a.AccountID, "error", err)
 			} else {
@@ -51,7 +53,7 @@ func (h *Handler) ChatCompletions(w http.ResponseWriter, r *http.Request) {
 		}
 		h.Auth.Release(a)
 	}()
-	
+
 	r = r.WithContext(auth.WithAuth(r.Context(), a))

 	var req map[string]any
--- a/internal/adapter/openai/handler_toolcall_format.go
+++ b/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 format (no other text):\n{\"tool_calls\": [{\"name\": \"tool_name\", \"input\": {\"param\": \"value\"}}]}\n\nHistory markers in conversation:\n- [TOOL_CALL_HISTORY]...[/TOOL_CALL_HISTORY] means a tool call you already made earlier.\n- [TOOL_RESULT_HISTORY]...[/TOOL_RESULT_HISTORY] means the runtime returned a tool result (not user input).\n\nIMPORTANT:\n1) If calling tools, output ONLY the JSON. The response must start with { and end with }.\n2) After receiving a tool result, you MUST use it to produce the final answer.\n3) Only call another tool when the previous result is missing required data or returned an error.\n4) Do not repeat a tool call that is already satisfied by an existing [TOOL_RESULT_HISTORY] block."
+	toolPrompt := "You have access to these tools:\n\n" + strings.Join(toolSchemas, "\n\n") + "\n\nWhen you need to use tools, output ONLY a JSON code block like this:\n```json\n{\"tool_calls\": [{\"name\": \"tool_name\", \"input\": {\"param\": \"value\"}}]}\n```\n\n【EXAMPLE】\nUser: Please check the weather in Beijing and Shanghai, and update my todo list.\nAssistant:\n```json\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```\n\nHistory markers in conversation:\n- [TOOL_CALL_HISTORY]...[/TOOL_CALL_HISTORY] means a tool call you already made earlier.\n- [TOOL_RESULT_HISTORY]...[/TOOL_RESULT_HISTORY] means the runtime returned a tool result (not user input).\n\nIMPORTANT:\n1) If calling tools, output ONLY the JSON code block. The response must start with ```json and end with ```.\n2) After receiving a tool result, you MUST use it to produce the final answer.\n3) Only call another tool when the previous result is missing required data or returned an error.\n4) Do not repeat a tool call that is already satisfied by an existing [TOOL_RESULT_HISTORY] block.\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."
 	if policy.Mode == util.ToolChoiceRequired {
 		toolPrompt += "\n5) For this response, you MUST call at least one tool from the allowed list."
 	}
--- a/internal/adapter/openai/handler_toolcall_test.go
+++ b/internal/adapter/openai/handler_toolcall_test.go
@@ -211,7 +211,7 @@ func TestHandleNonStreamUnknownToolNotIntercepted(t *testing.T) {
 	}
 }

-func TestHandleNonStreamEmbeddedToolCallExampleRemainsText(t *testing.T) {
+func TestHandleNonStreamEmbeddedToolCallExamplePromotesToolCall(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		`data: {"p":"response/content","v":"下面是示例："}`,
@@ -229,20 +229,21 @@ func TestHandleNonStreamEmbeddedToolCallExampleRemainsText(t *testing.T) {
 	out := decodeJSONBody(t, rec.Body.String())
 	choices, _ := out["choices"].([]any)
 	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] != "stop" {
-		t.Fatalf("expected finish_reason=stop, got %#v", choice["finish_reason"])
+	if choice["finish_reason"] != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
 	}
 	msg, _ := choice["message"].(map[string]any)
-	if _, ok := msg["tool_calls"]; ok {
-		t.Fatalf("did not expect tool_calls field for embedded example: %#v", msg["tool_calls"])
+	toolCalls, _ := msg["tool_calls"].([]any)
+	if len(toolCalls) != 1 {
+		t.Fatalf("expected one tool_call field for embedded example: %#v", msg["tool_calls"])
 	}
 	content, _ := msg["content"].(string)
-	if !strings.Contains(content, "下面是示例：") || !strings.Contains(content, "请勿执行。") || !strings.Contains(content, `"tool_calls"`) {
-		t.Fatalf("expected embedded example to remain plain text, got %#v", content)
+	if strings.Contains(content, `"tool_calls"`) {
+		t.Fatalf("expected raw tool_calls json stripped from content, got %#v", content)
 	}
 }

-func TestHandleNonStreamFencedToolCallExampleNotIntercepted(t *testing.T) {
+func TestHandleNonStreamFencedToolCallExamplePromotesToolCall(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		"data: {\"p\":\"response/content\",\"v\":\"```json\\n{\\\"tool_calls\\\":[{\\\"name\\\":\\\"search\\\",\\\"input\\\":{\\\"q\\\":\\\"go\\\"}}]}\\n```\"}",
@@ -258,16 +259,17 @@ func TestHandleNonStreamFencedToolCallExampleNotIntercepted(t *testing.T) {
 	out := decodeJSONBody(t, rec.Body.String())
 	choices, _ := out["choices"].([]any)
 	choice, _ := choices[0].(map[string]any)
-	if choice["finish_reason"] != "stop" {
-		t.Fatalf("expected finish_reason=stop, got %#v", choice["finish_reason"])
+	if choice["finish_reason"] != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, got %#v", choice["finish_reason"])
 	}
 	msg, _ := choice["message"].(map[string]any)
-	if _, ok := msg["tool_calls"]; ok {
-		t.Fatalf("did not expect tool_calls field for fenced example: %#v", msg["tool_calls"])
+	toolCalls, _ := msg["tool_calls"].([]any)
+	if len(toolCalls) != 1 {
+		t.Fatalf("expected one tool_call field for fenced example: %#v", msg["tool_calls"])
 	}
 	content, _ := msg["content"].(string)
-	if !strings.Contains(content, "```json") || !strings.Contains(content, `"tool_calls"`) {
-		t.Fatalf("expected fenced tool example to pass through as text, got %q", content)
+	if strings.Contains(content, `"tool_calls"`) {
+		t.Fatalf("expected raw tool_calls json stripped from content, got %q", content)
 	}
 }

@@ -615,7 +617,7 @@ func TestHandleStreamToolCallWithSameChunkTrailingTextRemainsText(t *testing.T)
 	}
 }

-func TestHandleStreamFencedToolCallSnippetRemainsText(t *testing.T) {
+func TestHandleStreamFencedToolCallSnippetPromotesToolCall(t *testing.T) {
 	h := &Handler{}
 	resp := makeSSEHTTPResponse(
 		fmt.Sprintf(`data: {"p":"response/content","v":%q}`, "下面是调用示例：\n```json\n"),
@@ -631,8 +633,8 @@ func TestHandleStreamFencedToolCallSnippetRemainsText(t *testing.T) {
 	if !done {
 		t.Fatalf("expected [DONE], body=%s", rec.Body.String())
 	}
-	if streamHasToolCallsDelta(frames) {
-		t.Fatalf("did not expect tool_calls delta for fenced snippet, body=%s", rec.Body.String())
+	if !streamHasToolCallsDelta(frames) {
+		t.Fatalf("expected tool_calls delta for fenced snippet, body=%s", rec.Body.String())
 	}
 	content := strings.Builder{}
 	for _, frame := range frames {
@@ -646,11 +648,11 @@ func TestHandleStreamFencedToolCallSnippetRemainsText(t *testing.T) {
 		}
 	}
 	got := content.String()
-	if !strings.Contains(got, "```json") || !strings.Contains(strings.ToLower(got), "tool_calls") {
-		t.Fatalf("expected fenced tool snippet in content, got=%q", got)
+	if strings.Contains(strings.ToLower(got), "tool_calls") {
+		t.Fatalf("expected raw fenced tool_calls snippet stripped from content, got=%q", got)
 	}
-	if streamFinishReason(frames) != "stop" {
-		t.Fatalf("expected finish_reason=stop, body=%s", rec.Body.String())
+	if streamFinishReason(frames) != "tool_calls" {
+		t.Fatalf("expected finish_reason=tool_calls, body=%s", rec.Body.String())
 	}
 }

--- a/internal/adapter/openai/responses_stream_test.go
+++ b/internal/adapter/openai/responses_stream_test.go
@@ -297,7 +297,7 @@ func TestHandleResponsesStreamOutputTextDeltaCarriesItemIndexes(t *testing.T) {
 	}
 }

-func TestHandleResponsesStreamThinkingAndMixedToolExampleRemainMessageOnly(t *testing.T) {
+func TestHandleResponsesStreamThinkingAndMixedToolExampleEmitsFunctionCall(t *testing.T) {
 	h := &Handler{}
 	req := httptest.NewRequest(http.MethodPost, "/v1/responses", nil)
 	rec := httptest.NewRecorder()
@@ -333,6 +333,7 @@ func TestHandleResponsesStreamThinkingAndMixedToolExampleRemainMessageOnly(t *te
 	responseObj, _ := completedPayload["response"].(map[string]any)
 	output, _ := responseObj["output"].([]any)
 	hasMessage := false
+	hasFunctionCall := false
 	for _, item := range output {
 		m, _ := item.(map[string]any)
 		if m == nil {
@@ -342,12 +343,15 @@ func TestHandleResponsesStreamThinkingAndMixedToolExampleRemainMessageOnly(t *te
 			hasMessage = true
 		}
 		if asString(m["type"]) == "function_call" {
-			t.Fatalf("did not expect function_call output for mixed prose tool example, output=%#v", output)
+			hasFunctionCall = true
 		}
 	}
 	if !hasMessage {
 		t.Fatalf("expected message output for mixed prose tool example, output=%#v", output)
 	}
+	if !hasFunctionCall {
+		t.Fatalf("expected function_call output for mixed prose tool example, output=%#v", output)
+	}
 }

 func TestHandleResponsesStreamToolChoiceNoneRejectsFunctionCall(t *testing.T) {
--- a/internal/adapter/openai/stream_status_test.go
+++ b/internal/adapter/openai/stream_status_test.go
@@ -171,15 +171,15 @@ func TestResponsesNonStreamMixedProseToolPayloadHandlerPath(t *testing.T) {
 		t.Fatalf("decode response failed: %v body=%s", err, rec.Body.String())
 	}
 	outputText, _ := out["output_text"].(string)
-	if outputText == "" {
-		t.Fatalf("expected output_text preserved for mixed prose payload")
+	if outputText != "" {
+		t.Fatalf("expected output_text hidden for mixed prose tool payload, got %q", outputText)
 	}
 	output, _ := out["output"].([]any)
 	if len(output) != 1 {
 		t.Fatalf("expected one output item, got %#v", output)
 	}
 	first, _ := output[0].(map[string]any)
-	if first["type"] != "message" {
-		t.Fatalf("expected message output item, got %#v", output)
+	if first["type"] != "function_call" {
+		t.Fatalf("expected function_call output item, got %#v", output)
 	}
 }
--- a/internal/adapter/openai/tool_sieve_core.go
+++ b/internal/adapter/openai/tool_sieve_core.go
@@ -167,22 +167,22 @@ func findToolSegmentStart(s string) int {
 		return -1
 	}
 	lower := strings.ToLower(s)
-	offset := 0
-	for {
-		keyRel := strings.Index(lower[offset:], "tool_calls")
-		if keyRel < 0 {
-			return -1
+	keywords := []string{"tool_calls", "function.name:", "[tool_call_history]"}
+	bestKeyIdx := -1
+	for _, kw := range keywords {
+		idx := strings.Index(lower, kw)
+		if idx >= 0 && (bestKeyIdx < 0 || idx < bestKeyIdx) {
+			bestKeyIdx = idx
 		}
-		keyIdx := offset + keyRel
-		start := strings.LastIndex(s[:keyIdx], "{")
-		if start < 0 {
-			start = keyIdx
-		}
-		if !insideCodeFence(s[:start]) {
-			return start
-		}
-		offset = keyIdx + len("tool_calls")
 	}
+	if bestKeyIdx < 0 {
+		return -1
+	}
+	start := strings.LastIndex(s[:bestKeyIdx], "{")
+	if start < 0 {
+		start = bestKeyIdx
+	}
+	return start
 }

 func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix string, calls []util.ParsedToolCall, suffix string, ready bool) {
@@ -191,13 +191,22 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 		return "", nil, "", false
 	}
 	lower := strings.ToLower(captured)
-	keyIdx := strings.Index(lower, "tool_calls")
+	
+	keyIdx := -1
+	keywords := []string{"tool_calls", "function.name:", "[tool_call_history]"}
+	for _, kw := range keywords {
+		idx := strings.Index(lower, kw)
+		if idx >= 0 && (keyIdx < 0 || idx < keyIdx) {
+			keyIdx = idx
+		}
+	}
+	
 	if keyIdx < 0 {
 		return "", nil, "", false
 	}
 	start := strings.LastIndex(captured[:keyIdx], "{")
 	if start < 0 {
-		return "", nil, "", false
+		start = keyIdx
 	}
 	obj, end, ok := extractJSONObjectFrom(captured, start)
 	if !ok {
@@ -205,9 +214,6 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 	}
 	prefixPart := captured[:start]
 	suffixPart := captured[end:]
-	if insideCodeFence(state.recentTextTail + prefixPart) {
-		return captured, nil, "", true
-	}
 	parsed := util.ParseStandaloneToolCallsDetailed(obj, toolNames)
 	if len(parsed.Calls) == 0 {
 		if parsed.SawToolCallSyntax && parsed.RejectedByPolicy {
@@ -215,6 +221,9 @@ func consumeToolCapture(state *toolStreamSieveState, toolNames []string) (prefix
 			// consume it to avoid leaking raw tool_calls JSON to user content.
 			return prefixPart, nil, suffixPart, true
 		}
+		// If it has obvious keywords but failed to parse even after loose repair,
+		// we still might want to intercept it if it looks like an attempt at tool call.
+		// For now, keep the original logic but rely on loose JSON repair.
 		return captured, nil, "", true
 	}
 	return prefixPart, parsed.Calls, suffixPart, true
--- a/internal/adapter/openai/tool_sieve_incremental.go
+++ b/internal/adapter/openai/tool_sieve_incremental.go
@@ -19,9 +19,6 @@ func buildIncrementalToolDeltas(state *toolStreamSieveState) []toolCallDelta {
 	if start < 0 {
 		return nil
 	}
-	if insideCodeFence(state.recentTextTail + captured[:start]) {
-		return nil
-	}
 	certainSingle, hasMultiple := classifyToolCallsIncrementalSafety(captured, keyIdx)
 	if hasMultiple {
 		state.disableDeltas = true
--- a/internal/admin/handler_accounts_testing.go
+++ b/internal/admin/handler_accounts_testing.go
@@ -247,8 +247,18 @@ func (h *Handler) deleteAllSessions(w http.ResponseWriter, r *http.Request) {
 	// 删除所有会话
 	err := h.DS.DeleteAllSessionsForToken(r.Context(), token)
 	if err != nil {
-		writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "删除失败: " + err.Error()})
-		return
+		// token 可能过期，尝试重新登录并重试一次
+		newToken, loginErr := h.DS.Login(r.Context(), acc)
+		if loginErr != nil {
+			writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "删除失败: " + err.Error()})
+			return
+		}
+		token = newToken
+		_ = h.Store.UpdateAccountToken(acc.Identifier(), token)
+		if retryErr := h.DS.DeleteAllSessionsForToken(r.Context(), token); retryErr != nil {
+			writeJSON(w, http.StatusOK, map[string]any{"success": false, "message": "删除失败: " + retryErr.Error()})
+			return
+		}
 	}

 	writeJSON(w, http.StatusOK, map[string]any{"success": true, "message": "删除成功"})
--- a/internal/admin/handler_accounts_testing_test.go
+++ b/internal/admin/handler_accounts_testing_test.go
@@ -1,9 +1,12 @@
 package admin

 import (
+	"bytes"
 	"context"
+	"encoding/json"
 	"errors"
 	"net/http"
+	"net/http/httptest"
 	"strings"
 	"testing"

@@ -13,10 +16,13 @@ import (
 )

 type testingDSMock struct {
-	loginCalls          int
-	createSessionCalls  int
-	getPowCalls         int
-	callCompletionCalls int
+	loginCalls                 int
+	createSessionCalls         int
+	getPowCalls                int
+	callCompletionCalls        int
+	deleteAllSessionsCalls     int
+	deleteAllSessionsError     error
+	deleteAllSessionsErrorOnce bool
 }

 func (m *testingDSMock) Login(_ context.Context, _ config.Account) (string, error) {
@@ -40,6 +46,14 @@ func (m *testingDSMock) CallCompletion(_ context.Context, _ *auth.RequestAuth, _
 }

 func (m *testingDSMock) DeleteAllSessionsForToken(_ context.Context, _ string) error {
+	m.deleteAllSessionsCalls++
+	if m.deleteAllSessionsError != nil {
+		err := m.deleteAllSessionsError
+		if m.deleteAllSessionsErrorOnce {
+			m.deleteAllSessionsError = nil
+		}
+		return err
+	}
 	return nil
 }

@@ -83,3 +97,38 @@ func TestTestAccount_BatchModeOnlyCreatesSession(t *testing.T) {
 		t.Fatalf("expected test status ok, got %q", updated.TestStatus)
 	}
 }
+
+func TestDeleteAllSessions_RetryWithReloginOnDeleteFailure(t *testing.T) {
+	t.Setenv("DS2API_CONFIG_JSON", `{"accounts":[{"email":"batch@example.com","password":"pwd","token":"expired-token"}]}`)
+	store := config.LoadStore()
+	ds := &testingDSMock{deleteAllSessionsError: errors.New("token expired"), deleteAllSessionsErrorOnce: true}
+	h := &Handler{Store: store, DS: ds}
+
+	req := httptest.NewRequest(http.MethodPost, "/delete-all", bytes.NewBufferString(`{"identifier":"batch@example.com"}`))
+	rec := httptest.NewRecorder()
+	h.deleteAllSessions(rec, req)
+
+	if rec.Code != http.StatusOK {
+		t.Fatalf("expected status 200, got %d", rec.Code)
+	}
+	var resp map[string]any
+	if err := json.Unmarshal(rec.Body.Bytes(), &resp); err != nil {
+		t.Fatalf("unmarshal response: %v", err)
+	}
+	if ok, _ := resp["success"].(bool); !ok {
+		t.Fatalf("expected success response, got %#v", resp)
+	}
+	if ds.loginCalls != 1 {
+		t.Fatalf("expected relogin once, got %d", ds.loginCalls)
+	}
+	if ds.deleteAllSessionsCalls != 2 {
+		t.Fatalf("expected delete called twice, got %d", ds.deleteAllSessionsCalls)
+	}
+	updated, ok := store.FindAccount("batch@example.com")
+	if !ok {
+		t.Fatal("expected account")
+	}
+	if updated.Token != "new-token" {
+		t.Fatalf("expected refreshed token persisted, got %q", updated.Token)
+	}
+}
--- a/internal/deepseek/client_auth.go
+++ b/internal/deepseek/client_auth.go
@@ -62,8 +62,8 @@ func (c *Client) CreateSession(ctx context.Context, a *auth.RequestAuth, maxAtte
 			attempts++
 			continue
 		}
-		code := intFrom(resp["code"])
-		if status == http.StatusOK && code == 0 {
+		code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+		if status == http.StatusOK && code == 0 && bizCode == 0 {
 			data, _ := resp["data"].(map[string]any)
 			bizData, _ := data["biz_data"].(map[string]any)
 			sessionID, _ := bizData["id"].(string)
@@ -71,10 +71,9 @@ func (c *Client) CreateSession(ctx context.Context, a *auth.RequestAuth, maxAtte
 				return sessionID, nil
 			}
 		}
-		msg, _ := resp["msg"].(string)
-		config.Logger.Warn("[create_session] failed", "status", status, "code", code, "msg", msg, "use_config_token", a.UseConfigToken, "account", a.AccountID)
+		config.Logger.Warn("[create_session] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "use_config_token", a.UseConfigToken, "account", a.AccountID)
 		if a.UseConfigToken {
-			if isTokenInvalid(status, code, msg) && !refreshed {
+			if isTokenInvalid(status, code, bizCode, msg, bizMsg) && !refreshed {
 				if c.Auth.RefreshToken(ctx, a) {
 					refreshed = true
 					continue
@@ -96,6 +95,7 @@ func (c *Client) GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts in
 		maxAttempts = c.maxRetries
 	}
 	attempts := 0
+	refreshed := false
 	for attempts < maxAttempts {
 		headers := c.authHeaders(a.DeepSeekToken)
 		resp, status, err := c.postJSONWithStatus(ctx, c.regular, DeepSeekCreatePowURL, headers, map[string]any{"target_path": "/api/v0/chat/completion"})
@@ -104,8 +104,8 @@ func (c *Client) GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts in
 			attempts++
 			continue
 		}
-		code := intFrom(resp["code"])
-		if status == http.StatusOK && code == 0 {
+		code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+		if status == http.StatusOK && code == 0 && bizCode == 0 {
 			data, _ := resp["data"].(map[string]any)
 			bizData, _ := data["biz_data"].(map[string]any)
 			challenge, _ := bizData["challenge"].(map[string]any)
@@ -116,15 +116,16 @@ func (c *Client) GetPow(ctx context.Context, a *auth.RequestAuth, maxAttempts in
 			}
 			return BuildPowHeader(challenge, answer)
 		}
-		msg, _ := resp["msg"].(string)
-		config.Logger.Warn("[get_pow] failed", "status", status, "code", code, "msg", msg, "use_config_token", a.UseConfigToken, "account", a.AccountID)
+		config.Logger.Warn("[get_pow] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "use_config_token", a.UseConfigToken, "account", a.AccountID)
 		if a.UseConfigToken {
-			if isTokenInvalid(status, code, msg) {
+			if isTokenInvalid(status, code, bizCode, msg, bizMsg) && !refreshed {
 				if c.Auth.RefreshToken(ctx, a) {
+					refreshed = true
 					continue
 				}
 			}
 			if c.Auth.SwitchAccount(ctx, a) {
+				refreshed = false
 				attempts++
 				continue
 			}
@@ -143,15 +144,34 @@ func (c *Client) authHeaders(token string) map[string]string {
 	return headers
 }

-func isTokenInvalid(status int, code int, msg string) bool {
-	msg = strings.ToLower(msg)
+func isTokenInvalid(status int, code int, bizCode int, msg string, bizMsg string) bool {
+	msg = strings.ToLower(strings.TrimSpace(msg) + " " + strings.TrimSpace(bizMsg))
 	if status == http.StatusUnauthorized || status == http.StatusForbidden {
 		return true
 	}
-	if code == 40001 || code == 40002 || code == 40003 {
+	if code == 40001 || code == 40002 || code == 40003 || bizCode == 40001 || bizCode == 40002 || bizCode == 40003 {
 		return true
 	}
-	return strings.Contains(msg, "token") || strings.Contains(msg, "unauthorized")
+	return strings.Contains(msg, "token") ||
+		strings.Contains(msg, "unauthorized") ||
+		strings.Contains(msg, "expired") ||
+		strings.Contains(msg, "not login") ||
+		strings.Contains(msg, "login required") ||
+		strings.Contains(msg, "invalid jwt")
+}
+
+func extractResponseStatus(resp map[string]any) (code int, bizCode int, msg string, bizMsg string) {
+	code = intFrom(resp["code"])
+	msg, _ = resp["msg"].(string)
+	data, _ := resp["data"].(map[string]any)
+	bizCode = intFrom(data["biz_code"])
+	bizMsg, _ = data["biz_msg"].(string)
+	if strings.TrimSpace(bizMsg) == "" {
+		if bizData, ok := data["biz_data"].(map[string]any); ok {
+			bizMsg, _ = bizData["msg"].(string)
+		}
+	}
+	return code, bizCode, msg, bizMsg
 }

 func normalizeMobileForLogin(raw string) (mobile string, areaCode any) {
--- a/internal/deepseek/client_session.go
+++ b/internal/deepseek/client_session.go
@@ -6,6 +6,7 @@ import (
 	"fmt"
 	"net/http"
 	"net/url"
+	"strings"

 	"ds2api/internal/auth"
 	"ds2api/internal/config"
@@ -22,12 +23,12 @@ type SessionInfo struct {

 // SessionStats 会话统计结果
 type SessionStats struct {
-	AccountID      string  // 账号标识 (email 或 mobile)
-	FirstPageCount int     // 第一页会话数量（当 HasMore 为 true 时，真实总数可能更大）
-	PinnedCount    int     // 置顶会话数量
-	HasMore        bool    // 是否还有更多页
-	Success        bool    // 请求是否成功
-	ErrorMessage   string  // 错误信息
+	AccountID      string // 账号标识 (email 或 mobile)
+	FirstPageCount int    // 第一页会话数量（当 HasMore 为 true 时，真实总数可能更大）
+	PinnedCount    int    // 置顶会话数量
+	HasMore        bool   // 是否还有更多页
+	Success        bool   // 请求是否成功
+	ErrorMessage   string // 错误信息
 }

 // GetSessionCount 获取单个账号的会话数量
@@ -56,8 +57,8 @@ func (c *Client) GetSessionCount(ctx context.Context, a *auth.RequestAuth, maxAt
 			continue
 		}

-		code := intFrom(resp["code"])
-		if status == http.StatusOK && code == 0 {
+		code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+		if status == http.StatusOK && code == 0 && bizCode == 0 {
 			data, _ := resp["data"].(map[string]any)
 			bizData, _ := data["biz_data"].(map[string]any)
 			chatSessions, _ := bizData["chat_sessions"].([]any)
@@ -79,12 +80,11 @@ func (c *Client) GetSessionCount(ctx context.Context, a *auth.RequestAuth, maxAt
 			return stats, nil
 		}

-		msg, _ := resp["msg"].(string)
 		stats.ErrorMessage = fmt.Sprintf("status=%d, code=%d, msg=%s", status, code, msg)
-		config.Logger.Warn("[get_session_count] failed", "status", status, "code", code, "msg", msg, "account", a.AccountID)
+		config.Logger.Warn("[get_session_count] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "account", a.AccountID)

 		if a.UseConfigToken {
-			if isTokenInvalid(status, code, msg) && !refreshed {
+			if isTokenInvalid(status, code, bizCode, msg, bizMsg) && !refreshed {
 				if c.Auth.RefreshToken(ctx, a) {
 					refreshed = true
 					continue
@@ -114,9 +114,11 @@ func (c *Client) GetSessionCountForToken(ctx context.Context, token string) (*Se
 		return nil, err
 	}

-	code := intFrom(resp["code"])
-	if status != http.StatusOK || code != 0 {
-		msg, _ := resp["msg"].(string)
+	code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+	if status != http.StatusOK || code != 0 || bizCode != 0 {
+		if strings.TrimSpace(bizMsg) != "" {
+			msg = bizMsg
+		}
 		return nil, fmt.Errorf("request failed: status=%d, code=%d, msg=%s", status, code, msg)
 	}

--- a/internal/deepseek/client_session_delete.go
+++ b/internal/deepseek/client_session_delete.go
@@ -49,18 +49,17 @@ func (c *Client) DeleteSession(ctx context.Context, a *auth.RequestAuth, session
 			continue
 		}

-		code := intFrom(resp["code"])
-		if status == http.StatusOK && code == 0 {
+		code, bizCode, msg, bizMsg := extractResponseStatus(resp)
+		if status == http.StatusOK && code == 0 && bizCode == 0 {
 			result.Success = true
 			return result, nil
 		}

-		msg, _ := resp["msg"].(string)
 		result.ErrorMessage = fmt.Sprintf("status=%d, code=%d, msg=%s", status, code, msg)
-		config.Logger.Warn("[delete_session] failed", "status", status, "code", code, "msg", msg, "session_id", sessionID)
+		config.Logger.Warn("[delete_session] failed", "status", status, "code", code, "biz_code", bizCode, "msg", msg, "biz_msg", bizMsg, "session_id", sessionID)

 		if a.UseConfigToken {
-			if isTokenInvalid(status, code, msg) && !refreshed {
+			if isTokenInvalid(status, code, bizCode, msg, bizMsg) && !refreshed {
 				if c.Auth.RefreshToken(ctx, a) {
 					refreshed = true
 					continue
--- a/internal/format/openai/render_chat.go
+++ b/internal/format/openai/render_chat.go
@@ -8,15 +8,15 @@ import (
 )

 func BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText string, toolNames []string) map[string]any {
-	detected := util.ParseStandaloneToolCalls(finalText, toolNames)
+	detected := util.ParseStandaloneToolCallsDetailed(finalText, toolNames)
 	finishReason := "stop"
 	messageObj := map[string]any{"role": "assistant", "content": finalText}
 	if strings.TrimSpace(finalThinking) != "" {
 		messageObj["reasoning_content"] = finalThinking
 	}
-	if len(detected) > 0 {
+	if len(detected.Calls) > 0 {
 		finishReason = "tool_calls"
-		messageObj["tool_calls"] = util.FormatOpenAIToolCalls(detected)
+		messageObj["tool_calls"] = util.FormatOpenAIToolCalls(detected.Calls)
 		messageObj["content"] = nil
 	}

--- a/internal/format/openai/render_responses.go
+++ b/internal/format/openai/render_responses.go
@@ -13,12 +13,12 @@ import (
 func BuildResponseObject(responseID, model, finalPrompt, finalThinking, finalText string, toolNames []string) map[string]any {
 	// Strict mode: only standalone, structured tool-call payloads are treated
 	// as executable tool calls.
-	detected := util.ParseStandaloneToolCalls(finalText, toolNames)
+	detected := util.ParseStandaloneToolCallsDetailed(finalText, toolNames)
 	exposedOutputText := finalText
 	output := make([]any, 0, 2)
-	if len(detected) > 0 {
+	if len(detected.Calls) > 0 {
 		exposedOutputText = ""
-		output = append(output, toResponsesFunctionCallItems(detected)...)
+		output = append(output, toResponsesFunctionCallItems(detected.Calls)...)
 	} else {
 		content := make([]any, 0, 2)
 		if finalThinking != "" {
--- a/internal/format/openai/render_test.go
+++ b/internal/format/openai/render_test.go
@@ -45,7 +45,7 @@ func TestBuildResponseObjectToolCallsFollowChatShape(t *testing.T) {
 	}
 }

-func TestBuildResponseObjectTreatsMixedProseToolPayloadAsText(t *testing.T) {
+func TestBuildResponseObjectPromotesMixedProseToolPayloadToFunctionCall(t *testing.T) {
 	obj := BuildResponseObject(
 		"resp_test",
 		"gpt-4o",
@@ -56,20 +56,20 @@ func TestBuildResponseObjectTreatsMixedProseToolPayloadAsText(t *testing.T) {
 	)

 	outputText, _ := obj["output_text"].(string)
-	if outputText == "" {
-		t.Fatalf("expected output_text preserved for mixed prose payload")
+	if outputText != "" {
+		t.Fatalf("expected output_text hidden for mixed prose tool payload, got %q", outputText)
 	}
 	output, _ := obj["output"].([]any)
 	if len(output) != 1 {
-		t.Fatalf("expected one message output item, got %#v", obj["output"])
+		t.Fatalf("expected one function_call output item, got %#v", obj["output"])
 	}
 	first, _ := output[0].(map[string]any)
-	if first["type"] != "message" {
-		t.Fatalf("expected message output type, got %#v", first["type"])
+	if first["type"] != "function_call" {
+		t.Fatalf("expected function_call output type, got %#v", first["type"])
 	}
 }

-func TestBuildResponseObjectFencedToolPayloadRemainsText(t *testing.T) {
+func TestBuildResponseObjectPromotesFencedToolPayloadToFunctionCall(t *testing.T) {
 	obj := BuildResponseObject(
 		"resp_test",
 		"gpt-4o",
@@ -80,16 +80,16 @@ func TestBuildResponseObjectFencedToolPayloadRemainsText(t *testing.T) {
 	)

 	outputText, _ := obj["output_text"].(string)
-	if outputText == "" {
-		t.Fatalf("expected output_text preserved for fenced example")
+	if outputText != "" {
+		t.Fatalf("expected output_text hidden for fenced tool payload, got %q", outputText)
 	}
 	output, _ := obj["output"].([]any)
 	if len(output) != 1 {
-		t.Fatalf("expected one message output item, got %#v", obj["output"])
+		t.Fatalf("expected one function_call output item, got %#v", obj["output"])
 	}
 	first, _ := output[0].(map[string]any)
-	if first["type"] != "message" {
-		t.Fatalf("expected message output type, got %#v", first["type"])
+	if first["type"] != "function_call" {
+		t.Fatalf("expected function_call output type, got %#v", first["type"])
 	}
 }

--- a/internal/js/chat-stream/vercel_stream.js
+++ b/internal/js/chat-stream/vercel_stream.js
@@ -1,33 +1,22 @@
 'use strict';

 const {
-  extractToolNames,
  createToolSieveState,
  processToolSieveChunk,
  flushToolSieve,
  parseStandaloneToolCalls,
  formatOpenAIStreamToolCalls,
 } = require('../helpers/stream-tool-sieve');
-const {
-  BASE_HEADERS,
-} = require('../shared/deepseek-constants');
-
-const {
-  writeOpenAIError,
-} = require('./error_shape');
-const {
-  parseChunkForContent,
-  isCitation,
-} = require('./sse_parse');
-const {
-  buildUsage,
-} = require('./token_usage');
+const { BASE_HEADERS } = require('../shared/deepseek-constants');
+const { writeOpenAIError } = require('./error_shape');
+const { parseChunkForContent, isCitation } = require('./sse_parse');
+const { buildUsage } = require('./token_usage');
 const {
  resolveToolcallPolicy,
+  formatIncrementalToolCallDeltas,
+  filterIncrementalToolCallDeltasByAllowed,
 } = require('./toolcall_policy');
-const {
-  createChatCompletionEmitter,
-} = require('./stream_emitter');
+const { createChatCompletionEmitter } = require('./stream_emitter');
 const {
  asString,
  isAbortError,
@@ -57,6 +46,7 @@ async function handleVercelStream(req, res, rawBody, payload) {
  const searchEnabled = toBool(prep.body.search_enabled);
  const toolPolicy = resolveToolcallPolicy(prep.body, payload.tools);
  const toolNames = toolPolicy.toolNames;
+  const emitEarlyToolDeltas = toolPolicy.emitEarlyToolDeltas;

  if (!model || !leaseID || !deepseekToken || !powHeader || !completionPayload) {
    writeOpenAIError(res, 500, 'invalid vercel prepare response');
@@ -132,6 +122,7 @@ async function handleVercelStream(req, res, rawBody, payload) {
    const toolSieveState = createToolSieveState();
    let toolCallsEmitted = false;
    const streamToolCallIDs = new Map();
+    const streamToolNames = new Map();
    const decoder = new TextDecoder();
    reader = completionRes.body.getReader();
    let buffered = '';
@@ -255,6 +246,18 @@ async function handleVercelStream(req, res, rawBody, payload) {
              }
              const events = processToolSieveChunk(toolSieveState, p.text, toolNames);
              for (const evt of events) {
+                if (evt.type === 'tool_call_deltas') {
+                  if (!emitEarlyToolDeltas) {
+                    continue;
+                  }
+                  const filtered = filterIncrementalToolCallDeltasByAllowed(evt.deltas, toolNames, streamToolNames);
+                  const formatted = formatIncrementalToolCallDeltas(filtered, streamToolCallIDs);
+                  if (formatted.length > 0) {
+                    toolCallsEmitted = true;
+                    sendDeltaFrame({ tool_calls: formatted });
+                  }
+                  continue;
+                }
                if (evt.type === 'tool_calls') {
                  toolCallsEmitted = true;
                  sendDeltaFrame({ tool_calls: formatOpenAIStreamToolCalls(evt.calls, streamToolCallIDs) });
--- a/internal/js/helpers/stream-tool-sieve/parse.js
+++ b/internal/js/helpers/stream-tool-sieve/parse.js
@@ -2,10 +2,8 @@

 const {
  toStringSafe,
-  looksLikeToolExampleContext,
 } = require('./state');
 const {
-  stripFencedCodeBlocks,
  buildToolCallCandidates,
  parseToolCallsPayload,
  parseMarkupToolCalls,
@@ -38,16 +36,13 @@ function parseToolCalls(text, toolNames) {

 function parseToolCallsDetailed(text, toolNames) {
  const result = emptyParseResult();
-  if (!toStringSafe(text)) {
+  const normalized = toStringSafe(text);
+  if (!normalized) {
    return result;
  }
-  const sanitized = stripFencedCodeBlocks(text);
-  if (!toStringSafe(sanitized)) {
-    return result;
-  }
-  result.sawToolCallSyntax = looksLikeToolCallSyntax(sanitized);
+  result.sawToolCallSyntax = looksLikeToolCallSyntax(normalized);

-  const candidates = buildToolCallCandidates(sanitized);
+  const candidates = buildToolCallCandidates(normalized);
  let parsed = [];
  for (const c of candidates) {
    parsed = parseToolCallsPayload(c);
@@ -63,9 +58,9 @@ function parseToolCallsDetailed(text, toolNames) {
    }
  }
  if (parsed.length === 0) {
-    parsed = parseMarkupToolCalls(sanitized);
+    parsed = parseMarkupToolCalls(normalized);
    if (parsed.length === 0) {
-      parsed = parseTextKVToolCalls(sanitized);
+      parsed = parseTextKVToolCalls(normalized);
      if (parsed.length === 0) {
        return result;
      }
@@ -90,22 +85,29 @@ function parseStandaloneToolCallsDetailed(text, toolNames) {
  if (!trimmed) {
    return result;
  }
-  if (trimmed.includes('```')) {
-    return result;
-  }
-  if (looksLikeToolExampleContext(trimmed)) {
-    return result;
-  }
  result.sawToolCallSyntax = looksLikeToolCallSyntax(trimmed);
-  let parsed = parseToolCallsPayload(trimmed);
+  const candidates = buildToolCallCandidates(trimmed);
+  let parsed = [];
+  for (const c of candidates) {
+    parsed = parseToolCallsPayload(c);
+    if (parsed.length === 0) {
+      parsed = parseMarkupToolCalls(c);
+    }
+    if (parsed.length === 0) {
+      parsed = parseTextKVToolCalls(c);
+    }
+    if (parsed.length > 0) {
+      break;
+    }
+  }
  if (parsed.length === 0) {
    parsed = parseMarkupToolCalls(trimmed);
-  }
-  if (parsed.length === 0) {
-    parsed = parseTextKVToolCalls(trimmed);
-  }
-  if (parsed.length === 0) {
-    return result;
+    if (parsed.length === 0) {
+      parsed = parseTextKVToolCalls(trimmed);
+      if (parsed.length === 0) {
+        return result;
+      }
+    }
  }

  result.sawToolCallSyntax = true;
--- a/internal/js/helpers/stream-tool-sieve/sieve.js
+++ b/internal/js/helpers/stream-tool-sieve/sieve.js
@@ -46,6 +46,9 @@ function processToolSieveChunk(state, chunk, toolNames) {
      if (Array.isArray(consumed.calls) && consumed.calls.length > 0) {
        state.pendingToolRaw = captured;
        state.pendingToolCalls = consumed.calls;
+        if (consumed.suffix) {
+          state.pending = consumed.suffix + state.pending;
+        }
        continue;
      }
      if (consumed.prefix) {
@@ -165,19 +168,34 @@ function findToolSegmentStart(s) {
    return -1;
  }
  const lower = s.toLowerCase();
+  const keywords = ['tool_calls', 'function.name:', '[tool_call_history]'];
  let offset = 0;
  // eslint-disable-next-line no-constant-condition
  while (true) {
-    const keyIdx = lower.indexOf('tool_calls', offset);
-    if (keyIdx < 0) {
+    let bestKeyIdx = -1;
+    let matchedKeyword = '';
+
+    for (const kw of keywords) {
+      const idx = lower.indexOf(kw, offset);
+      if (idx >= 0) {
+        if (bestKeyIdx < 0 || idx < bestKeyIdx) {
+          bestKeyIdx = idx;
+          matchedKeyword = kw;
+        }
+      }
+    }
+
+    if (bestKeyIdx < 0) {
      return -1;
    }
+
+    const keyIdx = bestKeyIdx;
    const start = s.slice(0, keyIdx).lastIndexOf('{');
    const candidateStart = start >= 0 ? start : keyIdx;
    if (!insideCodeFence(s.slice(0, candidateStart))) {
      return candidateStart;
    }
-    offset = keyIdx + 'tool_calls'.length;
+    offset = keyIdx + matchedKeyword.length;
  }
 }

@@ -187,20 +205,28 @@ function consumeToolCapture(state, toolNames) {
    return { ready: false, prefix: '', calls: [], suffix: '' };
  }
  const lower = captured.toLowerCase();
-  const keyIdx = lower.indexOf('tool_calls');
+  
+  let keyIdx = -1;
+  const keywords = ['tool_calls', 'function.name:', '[tool_call_history]'];
+  for (const kw of keywords) {
+    const idx = lower.indexOf(kw);
+    if (idx >= 0 && (keyIdx < 0 || idx < keyIdx)) {
+      keyIdx = idx;
+    }
+  }
+  
  if (keyIdx < 0) {
    return { ready: false, prefix: '', calls: [], suffix: '' };
  }
  const start = captured.slice(0, keyIdx).lastIndexOf('{');
-  if (start < 0) {
-    return { ready: false, prefix: '', calls: [], suffix: '' };
-  }
-  const obj = extractJSONObjectFrom(captured, start);
+  const actualStart = start >= 0 ? start : keyIdx;
+  
+  const obj = extractJSONObjectFrom(captured, actualStart);
  if (!obj.ok) {
    return { ready: false, prefix: '', calls: [], suffix: '' };
  }

-  const prefixPart = captured.slice(0, start);
+  const prefixPart = captured.slice(0, actualStart);
  const suffixPart = captured.slice(obj.end);

  if (insideCodeFence((state.recentTextTail || '') + prefixPart)) {
@@ -212,16 +238,7 @@ function consumeToolCapture(state, toolNames) {
    };
  }

-  if ((state.recentTextTail || '').trim() !== '' || prefixPart.trim() !== '' || suffixPart.trim() !== '') {
-    return {
-      ready: true,
-      prefix: captured,
-      calls: [],
-      suffix: '',
-    };
-  }
-
-  const parsed = parseStandaloneToolCallsDetailed(captured.slice(start, obj.end), toolNames);
+  const parsed = parseStandaloneToolCallsDetailed(captured.slice(actualStart, obj.end), toolNames);
  if (!Array.isArray(parsed.calls) || parsed.calls.length === 0) {
    if (parsed.sawToolCallSyntax && parsed.rejectedByPolicy) {
      return {
--- a/internal/util/toolcalls_candidates.go
+++ b/internal/util/toolcalls_candidates.go
@@ -20,7 +20,7 @@ func buildToolCallCandidates(text string) []string {
 		}
 	}

-	// best-effort extraction around "tool_calls" key in mixed text payloads.
+	// best-effort extraction around tool call keywords in mixed text payloads.
 	candidates = append(candidates, extractToolCallObjects(trimmed)...)

 	// best-effort object slice: from first '{' to last '}'
@@ -57,25 +57,65 @@ func extractToolCallObjects(text string) []string {
 	lower := strings.ToLower(text)
 	out := []string{}
 	offset := 0
+	keywords := []string{"tool_calls", "function.name:", "[tool_call_history]"}
 	for {
-		idx := strings.Index(lower[offset:], "tool_calls")
-		if idx < 0 {
+		bestIdx := -1
+		matchedKeyword := ""
+		for _, kw := range keywords {
+			idx := strings.Index(lower[offset:], kw)
+			if idx >= 0 {
+				absIdx := offset + idx
+				if bestIdx < 0 || absIdx < bestIdx {
+					bestIdx = absIdx
+					matchedKeyword = kw
+				}
+			}
+		}
+
+		if bestIdx < 0 {
 			break
 		}
-		idx += offset
-		start := strings.LastIndex(text[:idx], "{")
-		for start >= 0 {
+
+		idx := bestIdx
+		// Avoid backtracking too far to prevent OOM on malicious or very long strings
+		searchLimit := idx - 2000
+		if searchLimit < offset {
+			searchLimit = offset
+		}
+		
+		start := strings.LastIndex(text[searchLimit:idx], "{")
+		if start >= 0 {
+			start += searchLimit
+		}
+		
+		if start < 0 {
+			offset = idx + len(matchedKeyword)
+			continue
+		}
+
+		foundObj := false
+		for start >= searchLimit {
 			candidate, end, ok := extractJSONObject(text, start)
 			if ok {
 				// Move forward to avoid repeatedly matching the same object.
 				offset = end
 				out = append(out, strings.TrimSpace(candidate))
+				foundObj = true
 				break
 			}
-			start = strings.LastIndex(text[:start], "{")
+			// Try previous '{'
+			if start > searchLimit {
+				prevStart := strings.LastIndex(text[searchLimit:start], "{")
+				if prevStart >= 0 {
+					start = searchLimit + prevStart
+					continue
+				}
+			}
+			break
 		}
-		if start < 0 {
-			offset = idx + len("tool_calls")
+		
+		if !foundObj {
+			offset = idx + len(matchedKeyword)
 		}
 	}
 	return out
@@ -88,7 +128,12 @@ func extractJSONObject(text string, start int) (string, int, bool) {
 	depth := 0
 	quote := byte(0)
 	escaped := false
-	for i := start; i < len(text); i++ {
+	// Limit scan length to avoid OOM on unclosed objects
+	maxLen := start + 50000
+	if maxLen > len(text) {
+		maxLen = len(text)
+	}
+	for i := start; i < maxLen; i++ {
 		ch := text[i]
 		if quote != 0 {
 			if escaped {
--- a/internal/util/toolcalls_input_parse.go
+++ b/internal/util/toolcalls_input_parse.go
@@ -0,0 +1,108 @@
+package util
+
+import (
+	"encoding/json"
+	"strings"
+	"unicode"
+)
+
+func parseToolCallInput(v any) map[string]any {
+	switch x := v.(type) {
+	case nil:
+		return map[string]any{}
+	case map[string]any:
+		return x
+	case string:
+		raw := strings.TrimSpace(x)
+		if raw == "" {
+			return map[string]any{}
+		}
+		var parsed map[string]any
+		if err := json.Unmarshal([]byte(raw), &parsed); err == nil && parsed != nil {
+			repairPathLikeControlChars(parsed)
+			return parsed
+		}
+		// Try to repair invalid backslashes (common in Windows paths output by models)
+		repaired := repairInvalidJSONBackslashes(raw)
+		if repaired != raw {
+			if err := json.Unmarshal([]byte(repaired), &parsed); err == nil && parsed != nil {
+				repairPathLikeControlChars(parsed)
+				return parsed
+			}
+		}
+		// Try to repair loose JSON in string argument as well
+		repairedLoose := RepairLooseJSON(raw)
+		if repairedLoose != raw {
+			if err := json.Unmarshal([]byte(repairedLoose), &parsed); err == nil && parsed != nil {
+				repairPathLikeControlChars(parsed)
+				return parsed
+			}
+		}
+		return map[string]any{"_raw": raw}
+	default:
+		b, err := json.Marshal(x)
+		if err != nil {
+			return map[string]any{}
+		}
+		var parsed map[string]any
+		if err := json.Unmarshal(b, &parsed); err == nil && parsed != nil {
+			return parsed
+		}
+		return map[string]any{}
+	}
+}
+
+func repairPathLikeControlChars(m map[string]any) {
+	for k, v := range m {
+		switch vv := v.(type) {
+		case map[string]any:
+			repairPathLikeControlChars(vv)
+		case []any:
+			for _, item := range vv {
+				if child, ok := item.(map[string]any); ok {
+					repairPathLikeControlChars(child)
+				}
+			}
+		case string:
+			if isPathLikeKey(k) && containsControlRune(vv) {
+				m[k] = escapeControlRunes(vv)
+			}
+		}
+	}
+}
+
+func isPathLikeKey(key string) bool {
+	k := strings.ToLower(strings.TrimSpace(key))
+	return strings.Contains(k, "path") || strings.Contains(k, "file")
+}
+
+func containsControlRune(s string) bool {
+	for _, r := range s {
+		if unicode.IsControl(r) {
+			return true
+		}
+	}
+	return false
+}
+
+func escapeControlRunes(s string) string {
+	var b strings.Builder
+	b.Grow(len(s) + 8)
+	for _, r := range s {
+		switch r {
+		case '\b':
+			b.WriteString(`\b`)
+		case '\f':
+			b.WriteString(`\f`)
+		case '\n':
+			b.WriteString(`\n`)
+		case '\r':
+			b.WriteString(`\r`)
+		case '\t':
+			b.WriteString(`\t`)
+		default:
+			b.WriteRune(r)
+		}
+	}
+	return b.String()
+}
--- a/internal/util/toolcalls_json_repair.go
+++ b/internal/util/toolcalls_json_repair.go
@@ -0,0 +1,79 @@
+package util
+
+import (
+	"regexp"
+	"strings"
+)
+
+func repairInvalidJSONBackslashes(s string) string {
+	if !strings.Contains(s, "\\") {
+		return s
+	}
+	var out strings.Builder
+	out.Grow(len(s) + 10)
+	runes := []rune(s)
+	for i := 0; i < len(runes); i++ {
+		if runes[i] == '\\' {
+			if i+1 < len(runes) {
+				next := runes[i+1]
+				switch next {
+				case '"', '\\', '/', 'b', 'f', 'n', 'r', 't':
+					out.WriteRune('\\')
+					out.WriteRune(next)
+					i++
+					continue
+				case 'u':
+					if i+5 < len(runes) {
+						isHex := true
+						for j := 1; j <= 4; j++ {
+							r := runes[i+1+j]
+							if !((r >= '0' && r <= '9') || (r >= 'a' && r <= 'f') || (r >= 'A' && r <= 'F')) {
+								isHex = false
+								break
+							}
+						}
+						if isHex {
+							out.WriteRune('\\')
+							out.WriteRune('u')
+							for j := 1; j <= 4; j++ {
+								out.WriteRune(runes[i+1+j])
+							}
+							i += 5
+							continue
+						}
+					}
+				}
+			}
+			// Not a valid escape sequence, double it
+			out.WriteString("\\\\")
+		} else {
+			out.WriteRune(runes[i])
+		}
+	}
+	return out.String()
+}
+
+var unquotedKeyPattern = regexp.MustCompile(`([{,]\s*)([a-zA-Z_][a-zA-Z0-9_]*)\s*:`)
+
+// missingArrayBracketsPattern identifies a sequence of two or more JSON objects separated by commas
+// that immediately follow a colon, which indicates a missing array bracket `[` `]`.
+// E.g., "key": {"a": 1}, {"b": 2} -> "key": [{"a": 1}, {"b": 2}]
+// NOTE: The pattern uses (?:[^{}]|\{[^{}]*\})* to support single-level nested {} objects,
+// which handles cases like {"content": "x", "input": {"q": "y"}}
+var missingArrayBracketsPattern = regexp.MustCompile(`(:\s*)(\{(?:[^{}]|\{[^{}]*\})*\}(?:\s*,\s*\{(?:[^{}]|\{[^{}]*\})*\})+)`)
+
+func RepairLooseJSON(s string) string {
+	s = strings.TrimSpace(s)
+	if s == "" {
+		return s
+	}
+	// 1. Replace unquoted keys: {key: -> {"key":
+	s = unquotedKeyPattern.ReplaceAllString(s, `$1"$2":`)
+
+	// 2. Heuristic: Fix missing array brackets for list of objects
+	// e.g., : {obj1}, {obj2} -> : [{obj1}, {obj2}]
+	// This specifically addresses DeepSeek's "list hallucination"
+	s = missingArrayBracketsPattern.ReplaceAllString(s, `$1[$2]`)
+
+	return s
+}
--- a/internal/util/toolcalls_parse.go
+++ b/internal/util/toolcalls_parse.go
@@ -16,7 +16,6 @@ type ToolCallParseResult struct {
 	RejectedByPolicy  bool
 	RejectedToolNames []string
 }
-
 func ParseToolCalls(text string, availableToolNames []string) []ParsedToolCall {
 	return ParseToolCallsDetailed(text, availableToolNames).Calls
 }
@@ -26,10 +25,6 @@ func ParseToolCallsDetailed(text string, availableToolNames []string) ToolCallPa
 	if strings.TrimSpace(text) == "" {
 		return result
 	}
-	text = stripFencedCodeBlocks(text)
-	if strings.TrimSpace(text) == "" {
-		return result
-	}
 	result.SawToolCallSyntax = looksLikeToolCallSyntax(text)

 	candidates := buildToolCallCandidates(text)
@@ -68,7 +63,6 @@ func ParseToolCallsDetailed(text string, availableToolNames []string) ToolCallPa
 	result.RejectedByPolicy = len(rejectedNames) > 0 && len(calls) == 0
 	return result
 }
-
 func ParseStandaloneToolCalls(text string, availableToolNames []string) []ParsedToolCall {
 	return ParseStandaloneToolCallsDetailed(text, availableToolNames).Calls
 }
@@ -79,17 +73,15 @@ func ParseStandaloneToolCallsDetailed(text string, availableToolNames []string)
 	if trimmed == "" {
 		return result
 	}
-	if looksLikeToolExampleContext(trimmed) {
-		return result
-	}
 	result.SawToolCallSyntax = looksLikeToolCallSyntax(trimmed)
-	candidates := []string{trimmed}
+	candidates := buildToolCallCandidates(trimmed)
+	var parsed []ParsedToolCall
 	for _, candidate := range candidates {
 		candidate = strings.TrimSpace(candidate)
 		if candidate == "" {
 			continue
 		}
-		parsed := parseToolCallsPayload(candidate)
+		parsed = parseToolCallsPayload(candidate)
 		if len(parsed) == 0 {
 			parsed = parseXMLToolCalls(candidate)
 		}
@@ -100,14 +92,23 @@ func ParseStandaloneToolCallsDetailed(text string, availableToolNames []string)
 			parsed = parseTextKVToolCalls(candidate)
 		}
 		if len(parsed) > 0 {
-			result.SawToolCallSyntax = true
-			calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
-			result.Calls = calls
-			result.RejectedToolNames = rejectedNames
-			result.RejectedByPolicy = len(rejectedNames) > 0 && len(calls) == 0
-			return result
+			break
 		}
 	}
+	if len(parsed) == 0 {
+		parsed = parseXMLToolCalls(trimmed)
+		if len(parsed) == 0 {
+			parsed = parseTextKVToolCalls(trimmed)
+			if len(parsed) == 0 {
+				return result
+			}
+		}
+	}
+	result.SawToolCallSyntax = true
+	calls, rejectedNames := filterToolCallsDetailed(parsed, availableToolNames)
+	result.Calls = calls
+	result.RejectedToolNames = rejectedNames
+	result.RejectedByPolicy = len(rejectedNames) > 0 && len(calls) == 0
 	return result
 }

@@ -171,7 +172,13 @@ func resolveAllowedToolName(name string, allowed map[string]struct{}, allowedCan
 func parseToolCallsPayload(payload string) []ParsedToolCall {
 	var decoded any
 	if err := json.Unmarshal([]byte(payload), &decoded); err != nil {
-		return nil
+		// Try to repair backslashes first! Because LLMs often mix these two problems.
+		repaired := repairInvalidJSONBackslashes(payload)
+		// Try loose repair on top of that
+		repaired = RepairLooseJSON(repaired)
+		if err := json.Unmarshal([]byte(repaired), &decoded); err != nil {
+			return nil
+		}
 	}
 	switch v := decoded.(type) {
 	case map[string]any:
@@ -248,32 +255,3 @@ func parseToolCallItem(m map[string]any) (ParsedToolCall, bool) {
 		Input: parseToolCallInput(inputRaw),
 	}, true
 }
-
-func parseToolCallInput(v any) map[string]any {
-	switch x := v.(type) {
-	case nil:
-		return map[string]any{}
-	case map[string]any:
-		return x
-	case string:
-		raw := strings.TrimSpace(x)
-		if raw == "" {
-			return map[string]any{}
-		}
-		var parsed map[string]any
-		if err := json.Unmarshal([]byte(raw), &parsed); err == nil && parsed != nil {
-			return parsed
-		}
-		return map[string]any{"_raw": raw}
-	default:
-		b, err := json.Marshal(x)
-		if err != nil {
-			return map[string]any{}
-		}
-		var parsed map[string]any
-		if err := json.Unmarshal(b, &parsed); err == nil && parsed != nil {
-			return parsed
-		}
-		return map[string]any{}
-	}
-}
--- a/internal/util/toolcalls_test.go
+++ b/internal/util/toolcalls_test.go
@@ -1,6 +1,9 @@
 package util

-import "testing"
+import (
+	"strings"
+	"testing"
+)

 func TestParseToolCalls(t *testing.T) {
 	text := `prefix {"tool_calls":[{"name":"search","input":{"q":"golang"}}]} suffix`
@@ -19,8 +22,8 @@ func TestParseToolCalls(t *testing.T) {
 func TestParseToolCallsFromFencedJSON(t *testing.T) {
 	text := "I will call tools now\n```json\n{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"news\"}}]}\n```"
 	calls := ParseToolCalls(text, []string{"search"})
-	if len(calls) != 0 {
-		t.Fatalf("expected fenced tool_call example to be ignored, got %#v", calls)
+	if len(calls) != 1 {
+		t.Fatalf("expected fenced tool_call payload to be parsed, got %#v", calls)
 	}
 }

@@ -96,10 +99,10 @@ func TestFormatOpenAIToolCalls(t *testing.T) {
 	}
 }

-func TestParseStandaloneToolCallsOnlyMatchesStandalonePayload(t *testing.T) {
+func TestParseStandaloneToolCallsSupportsMixedProsePayload(t *testing.T) {
 	mixed := `这里是示例：{"tool_calls":[{"name":"search","input":{"q":"go"}}]}`
-	if calls := ParseStandaloneToolCalls(mixed, []string{"search"}); len(calls) != 0 {
-		t.Fatalf("expected standalone parser to ignore mixed prose, got %#v", calls)
+	if calls := ParseStandaloneToolCalls(mixed, []string{"search"}); len(calls) != 1 {
+		t.Fatalf("expected standalone parser to parse mixed prose payload, got %#v", calls)
 	}

 	standalone := `{"tool_calls":[{"name":"search","input":{"q":"go"}}]}`
@@ -109,10 +112,10 @@ func TestParseStandaloneToolCallsOnlyMatchesStandalonePayload(t *testing.T) {
 	}
 }

-func TestParseStandaloneToolCallsIgnoresFencedCodeBlock(t *testing.T) {
+func TestParseStandaloneToolCallsParsesFencedCodeBlock(t *testing.T) {
 	fenced := "```json\n{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}\n```"
-	if calls := ParseStandaloneToolCalls(fenced, []string{"search"}); len(calls) != 0 {
-		t.Fatalf("expected fenced tool_call example to be ignored, got %#v", calls)
+	if calls := ParseStandaloneToolCalls(fenced, []string{"search"}); len(calls) != 1 {
+		t.Fatalf("expected fenced tool_call payload to be parsed, got %#v", calls)
 	}
 }

@@ -279,3 +282,238 @@ func TestParseToolCallsDoesNotAcceptMismatchedMarkupTags(t *testing.T) {
 		t.Fatalf("expected mismatched tags to be rejected, got %#v", calls)
 	}
 }
+
+func TestRepairInvalidJSONBackslashes(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{`{"path": "C:\Users\name"}`, `{"path": "C:\\Users\name"}`},
+		{`{"cmd": "cd D:\git_codes"}`, `{"cmd": "cd D:\\git_codes"}`},
+		{`{"text": "line1\nline2"}`, `{"text": "line1\nline2"}`},
+		{`{"path": "D:\\back\\slash"}`, `{"path": "D:\\back\\slash"}`},
+		{`{"unicode": "\u2705"}`, `{"unicode": "\u2705"}`},
+		{`{"invalid_u": "\u123"}`, `{"invalid_u": "\\u123"}`},
+	}
+
+	for _, tt := range tests {
+		got := repairInvalidJSONBackslashes(tt.input)
+		if got != tt.expected {
+			t.Errorf("repairInvalidJSONBackslashes(%s) = %s; want %s", tt.input, got, tt.expected)
+		}
+	}
+}
+
+func TestRepairLooseJSON(t *testing.T) {
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{`{tool_calls: [{"name": "search", "input": {"q": "go"}}]}`, `{"tool_calls": [{"name": "search", "input": {"q": "go"}}]}`},
+		{`{name: "search", input: {q: "go"}}`, `{"name": "search", "input": {"q": "go"}}`},
+	}
+
+	for _, tt := range tests {
+		got := RepairLooseJSON(tt.input)
+		if got != tt.expected {
+			t.Errorf("RepairLooseJSON(%s) = %s; want %s", tt.input, got, tt.expected)
+		}
+	}
+}
+
+func TestParseToolCallsWithUnquotedKeys(t *testing.T) {
+	text := `这里是列表：{tool_calls: [{"name": "todowrite", "input": {"todos": "test"}}]}`
+	availableTools := []string{"todowrite"}
+
+	parsed := ParseToolCalls(text, availableTools)
+	if len(parsed) != 1 {
+		t.Fatalf("expected 1 tool call, got %d", len(parsed))
+	}
+	if parsed[0].Name != "todowrite" {
+		t.Errorf("expected tool todowrite, got %s", parsed[0].Name)
+	}
+}
+
+func TestParseToolCallsWithInvalidBackslashes(t *testing.T) {
+	// DeepSeek sometimes outputs Windows paths with single backslashes in JSON strings
+	// Note: using raw string to simulate what AI actually sends in the stream
+	text := `好的，执行以下命令：{"name": "execute_command", "input": "{\"command\": \"cd D:\git_codes && dir\"}"}`
+	availableTools := []string{"execute_command"}
+
+	parsed := ParseToolCalls(text, availableTools)
+	// If standard JSON fails, buildToolCallCandidates should still extract the object,
+	// and parseToolCallsPayload should repair it.
+	if len(parsed) != 1 {
+		// If it still fails, let's see why
+		candidates := buildToolCallCandidates(text)
+		t.Logf("Candidates: %v", candidates)
+		t.Fatalf("expected 1 tool call, got %d", len(parsed))
+	}
+
+	cmd, ok := parsed[0].Input["command"].(string)
+	if !ok {
+		t.Fatalf("expected command string in input, got %v", parsed[0].Input)
+	}
+
+	expected := "cd D:\\git_codes && dir"
+	if cmd != expected {
+		t.Errorf("expected command %q, got %q", expected, cmd)
+	}
+}
+
+func TestParseToolCallsWithDeepSeekHallucination(t *testing.T) {
+	// 模拟 DeepSeek 典型的幻觉输出：未加引号的键名 + 包含 Windows 路径的嵌套 JSON 字符串 + 漏掉列表的方括号
+	text := `检测到实施意图——实现经典算法。需在misc/目录创建Python文件。
+关键约束:
+1. Windows UTF-8编码处理
+2. 必须用绝对路径导入
+3. 禁止write覆盖已有文件（misc/目录允许创建新文件）
+将任务分解并委托：
+- 研究8皇后算法模式（并行探索）
+- 实现带可视化输出的解决方案（unspecified-high）
+先创建todo列表追踪步骤。
+{tool_calls: [{"name": "todowrite", "input": {"todos": {"content": "研究8皇后问题算法模式（回溯法）和输出格式", "status": "pending", "priority": "high"}, {"content": "在misc/目录创建8皇后Python脚本，包含完整解决方案和可视化输出", "status": "pending", "priority": "high"}, {"content": "验证脚本正确性（运行测试）", "status": "pending", "priority": "medium"}}}]}`
+
+	availableTools := []string{"todowrite"}
+	parsed := ParseToolCalls(text, availableTools)
+
+	if len(parsed) != 1 {
+		cands := buildToolCallCandidates(text)
+		for i, c := range cands {
+			t.Logf("CAND %d: %s", i, c)
+			repaired := RepairLooseJSON(c)
+			t.Logf("  REPAIRED: %s", repaired)
+		}
+		t.Fatalf("expected 1 tool call, got %d. Candidates: %v", len(parsed), buildToolCallCandidates(text))
+	}
+
+	if parsed[0].Name != "todowrite" {
+		t.Errorf("expected tool name 'todowrite', got %q", parsed[0].Name)
+	}
+
+	todos, ok := parsed[0].Input["todos"].([]any)
+	if !ok {
+		t.Fatalf("expected 'todos' to be parsed as a list, got %T: %#v", parsed[0].Input["todos"], parsed[0].Input["todos"])
+	}
+	if len(todos) != 3 {
+		t.Errorf("expected 3 todo items, got %d", len(todos))
+	}
+}
+
+func TestParseToolCallsWithMixedWindowsPaths(t *testing.T) {
+	// 更复杂的案例：嵌套 JSON 字符串中的反斜杠未转义
+	text := `关键约束: 1. Windows UTF-8编码处理 2. 必须用绝对路径导入 D:\git_codes\ds2api\misc
+{tool_calls: [{"name": "write_file", "input": "{\"path\": \"D:\\git_codes\\ds2api\\misc\\queens.py\", \"content\": \"print('hello')\"}"}]}`
+
+	availableTools := []string{"write_file"}
+	parsed := ParseToolCalls(text, availableTools)
+
+	if len(parsed) != 1 {
+		t.Fatalf("expected 1 tool call from mixed text with paths, got %d", len(parsed))
+	}
+
+	path, _ := parsed[0].Input["path"].(string)
+	// 在解析后的 Go map 中，反斜杠应该被还原
+	if !strings.Contains(path, "D:\\git_codes") && !strings.Contains(path, "D:/git_codes") {
+		t.Errorf("expected path to contain Windows style separators, got %q", path)
+	}
+}
+
+func TestParseToolCallInputRepairsControlCharsInPath(t *testing.T) {
+	in := `{"path":"D:\tmp\new\readme.txt","content":"line1\nline2"}`
+	parsed := parseToolCallInput(in)
+
+	path, ok := parsed["path"].(string)
+	if !ok {
+		t.Fatalf("expected path string in parsed input, got %#v", parsed["path"])
+	}
+	if path != `D:\tmp\new\readme.txt` {
+		t.Fatalf("expected repaired windows path, got %q", path)
+	}
+
+	content, ok := parsed["content"].(string)
+	if !ok {
+		t.Fatalf("expected content string in parsed input, got %#v", parsed["content"])
+	}
+	if content != "line1\nline2" {
+		t.Fatalf("expected non-path field to keep decoded escapes, got %q", content)
+	}
+}
+
+func TestRepairLooseJSONWithNestedObjects(t *testing.T) {
+	// 测试嵌套对象的修复：DeepSeek 幻觉输出，每个元素内部包含嵌套 {}
+	// 注意：正则只支持单层嵌套，不支持更深层次的嵌套
+	tests := []struct {
+		name     string
+		input    string
+		expected string
+	}{
+		// 1. 单层嵌套对象（核心修复目标）
+		{
+			name:     "单层嵌套 - 2个元素",
+			input:    `"todos": {"content": "研究算法", "input": {"q": "8 queens"}}, {"content": "实现", "input": {"path": "queens.py"}}`,
+			expected: `"todos": [{"content": "研究算法", "input": {"q": "8 queens"}}, {"content": "实现", "input": {"path": "queens.py"}}]`,
+		},
+		// 2. 3个单层嵌套对象
+		{
+			name:     "3个单层嵌套对象",
+			input:    `"items": {"a": {"x":1}}, {"b": {"y":2}}, {"c": {"z":3}}`,
+			expected: `"items": [{"a": {"x":1}}, {"b": {"y":2}}, {"c": {"z":3}}]`,
+		},
+		// 3. 混合嵌套：有些字段是对象，有些是原始值
+		{
+			name:     "混合嵌套 - 对象和原始值混合",
+			input:    `"items": {"name": "test", "config": {"timeout": 30}}, {"name": "test2", "config": {"timeout": 60}}`,
+			expected: `"items": [{"name": "test", "config": {"timeout": 30}}, {"name": "test2", "config": {"timeout": 60}}]`,
+		},
+		// 4. 4个嵌套对象（边界测试）
+		{
+			name:     "4个嵌套对象",
+			input:    `"todos": {"id": 1}, {"id": 2}, {"id": 3}, {"id": 4}`,
+			expected: `"todos": [{"id": 1}, {"id": 2}, {"id": 3}, {"id": 4}]`,
+		},
+		// 5. DeepSeek 典型幻觉：无空格逗号分隔
+		{
+			name:     "无空格逗号分隔",
+			input:    `"results": {"name": "a"}, {"name": "b"}, {"name": "c"}`,
+			expected: `"results": [{"name": "a"}, {"name": "b"}, {"name": "c"}]`,
+		},
+		// 6. 嵌套数组（数组在对象内，不是深层嵌套）
+		{
+			name:     "对象内包含数组",
+			input:    `"data": {"items": [1,2,3]}, {"items": [4,5,6]}`,
+			expected: `"data": [{"items": [1,2,3]}, {"items": [4,5,6]}]`,
+		},
+		// 7. 真实的 DeepSeek 8皇后问题输出
+		{
+			name:     "DeepSeek 8皇后真实输出",
+			input:    `"todos": {"content": "研究8皇后算法", "status": "pending"}, {"content": "实现Python脚本", "status": "pending"}, {"content": "验证结果", "status": "pending"}`,
+			expected: `"todos": [{"content": "研究8皇后算法", "status": "pending"}, {"content": "实现Python脚本", "status": "pending"}, {"content": "验证结果", "status": "pending"}]`,
+		},
+		// 8. 简单无嵌套对象（回归测试）
+		{
+			name:     "简单无嵌套对象",
+			input:    `"items": {"a": 1}, {"b": 2}`,
+			expected: `"items": [{"a": 1}, {"b": 2}]`,
+		},
+		// 9. 更复杂的单层嵌套
+		{
+			name:     "复杂单层嵌套",
+			input:    `"functions": {"name": "execute", "input": {"command": "ls"}}, {"name": "read", "input": {"file": "a.txt"}}`,
+			expected: `"functions": [{"name": "execute", "input": {"command": "ls"}}, {"name": "read", "input": {"file": "a.txt"}}]`,
+		},
+		// 10. 5个嵌套对象
+		{
+			name:     "5个嵌套对象",
+			input:    `"tasks": {"id":1}, {"id":2}, {"id":3}, {"id":4}, {"id":5}`,
+			expected: `"tasks": [{"id":1}, {"id":2}, {"id":3}, {"id":4}, {"id":5}]`,
+		},
+	}
+
+	for _, tt := range tests {
+		got := RepairLooseJSON(tt.input)
+		if got != tt.expected {
+			t.Errorf("[%s] RepairLooseJSON with nested objects:\n  input:    %s\n  got:      %s\n  expected: %s", tt.name, tt.input, got, tt.expected)
+		}
+	}
+}
--- a/internal/util/util_edge_test.go
+++ b/internal/util/util_edge_test.go
@@ -409,8 +409,8 @@ func TestParseToolCallsWithFunctionWrapper(t *testing.T) {
 func TestParseStandaloneToolCallsFencedCodeBlock(t *testing.T) {
 	fenced := "Here's an example:\n```json\n{\"tool_calls\":[{\"name\":\"search\",\"input\":{\"q\":\"go\"}}]}\n```\nDon't execute this."
 	calls := ParseStandaloneToolCalls(fenced, []string{"search"})
-	if len(calls) != 0 {
-		t.Fatalf("expected fenced code block ignored, got %d calls", len(calls))
+	if len(calls) != 1 {
+		t.Fatalf("expected fenced code block to be parsed, got %d calls", len(calls))
 	}
 }

--- a/misc/deepseek_functioncalling_bug/report.md
+++ b/misc/deepseek_functioncalling_bug/report.md
@@ -1,101 +0,0 @@
-# DeepSeek Function Calling 缺陷分析与 ds2api 的增强修复策略
-
-> **相关 PR**: #74 (代码核心实现) 与 #75 (Merge to dev)
-> **问题背景**: 解决因包括 DeepSeek 在内的部分模型在函数调用（Function Calling/Tool Call）表现不够“规范”，从而导致工具调用失败的问题。
-
-## 一、底层架构对比：为什么会产生 Function Calling 缺陷？
-
-在探讨缺陷前，我们需要理解两种 Function Calling 的底层结构差异：
-
-### 1. OpenAI 的原生结构化返回 (API 级分离)
-在 OpenAI 的规范中，**聊天文字与工具调用是在底层的 JSON 结构中被硬性拆分的**：
-* 聊天废话存放在 `response.choices[0].message.content` 里。
-* 工具请求存放在单独的数组 `response.choices[0].message.tool_calls` 里。
-
-**优势：** 这种设计对客户端极其友好。客户端只需判断 `tool_calls` 是否为空，就能决定是执行代码还是渲染文字。它支持同时并发多个工具请求，且底层的生成殷勤被严格训练和约束，极少抛出语法错误的 JSON。
-
-### 2. DeepSeek 等模型的“单文本流”机制
-相比之下，部分未经深度专门微调的模型（或者在特定的通信适配层中），它们依然倾向于把一切内容打包成一个纯文本流吐出。这就是为什么它们的输出往往不仅包含了本该属于 `tool_calls` 结构里的 JSON，还会像个“老实人”一样夹杂了属于 `content` 里的散文。
-
---
-
-## 二、DeepSeek 在 Function Calling 上的特定缺陷表现
-
-相比于 OpenAI 严格遵循 API 约定的原生结构，DeepSeek 等开源/国产推理模型在工具调用时，经常会暴露出以下三种典型的“不守规矩”的输出行为：
-
-### 1. 混合输出：散文文本与工具 JSON 混杂 (Mixed Prose Streams)
-当应用要求模型直接返回工具请求时，DeepSeek 有时候会**“忍不住想和用户搭话”**。
-它常常前置一段解释性废话，中间插入工具调用的 JSON 参数，并在末尾再补上一句总结：
-```text
-好的，我这就帮你读取 README.md 的内容：
-{"tool_calls":[{"name":"read_file","input":{"path":"README.md"}}]}
-请稍等片刻，我马上把它读出来。
-```
-**旧版系统痛点：**
-原有的代码存在**严格模式（Strict Mode）**校验：
-```go
-// 如果解析到的 JSON 块前后存在任何非空字符串，就放弃当作工具调用！
-if strings.TrimSpace(state.recentTextTail) != "" || strings.TrimSpace(prefixPart) != "" ... {
-    return captured, nil, "", true
-}
-```
-这直接导致上述结构被网关认定是一段“普通聊天”，直接原封不动地返回给用户，这直接干挂了后续的工具自动执行流程。
-
-### 2. 工具名格式幻觉：擅自修改或前缀化工具名称
-由于 DeepSeek 的预训练数据中有大量的代码和不同的平台结构，它在回复工具名称时，常常无法忠实于 System Prompt 中提供的纯命名（也就是 `name: "read_file"`），而是加上前缀或者拼写变形，例如：
-* `{"name": "mcp.search_web"}` （自带命名空间）
-* `{"name": "tools.read_file"}`
-* `{"name": "search-web"}` （下划线变成了中划线）
-
-**旧版系统痛点：**
-旧版系统对于工具名的匹配几乎只有“绝对相等”的字典级比对，只要差了一个字符或加了前缀，就会由于找不到合法工具而直接失败。
-
-### 3. Role 角色的非标准返回
-在部分工具通信流的响应中，返回的内容其所属的 `role` 没有被标准化处理，可能携带意料之外的属性，或是与下游严格比对出现冲突。
-
---
-
-## 二、PR #74 的代码增强修复方案
-
-为了解决大模型这种自身的不规范行为，PR #74 在系统的中间层网关联入了一个**极其包容的容错引擎**。它并不强制要求模型“改过自新”，而是主动做了以下三块增强：
-
-### 1. 从流中分离混合内容（废除 Strict Mode）
-修改了 `internal/adapter/openai/tool_sieve_core.go`。
-取消了前后包裹文本的拦截逻辑。当系统扫描到流式结构中有完整的 `{"tool_calls":...}` 时，它会将废话和 JSON 分发到不同的事件流中：
-```go
-if prefix != "" {
-    // 将前面的“好的，帮你读文件”剥离出来作为常规文本输出
-    state.noteText(prefix)
-    events = append(events, toolStreamEvent{Content: prefix})
-}
-// 捕获并拦截中间的工具请求，进行背后执行
-state.pendingToolCalls = calls
-```
-**效果：** 用户的屏幕上只能看到正常的文字交流，而后端的工具也会立刻挂载。
-
-### 2. 多级宽容匹配引擎 (Resolve Allowed Tool Name)
-在 `internal/util/toolcalls_parse.go` 中，新增了一个由严到松降级匹配的强大漏斗策略函数 `resolveAllowedToolName`：
-
-1. **绝对匹配**：和以前一样，`read_file` == `read_file`。
-2. **忽略大小写**：`Read_File` 算作合法。
-3. **命名空间抹除**：通过寻找最后一个 `.` 来剥离前缀，强制将 `mcp.search_web` 还原出真实的 `search_web`。
-4. **终极正则清洗**：
-   引入 `var toolNameLoosePattern = regexp.MustCompile(`[^a-z0-9]+`)`。
-   这个正则剥离了字符串里所有的符号、空格、格式符。
-   将传入的 `read-file` 洗除符号成为 `readfile`，并去和系统中所有合法工具同样清洗后的版本进行比较。只要核心字母一致，即算作匹配成功。
-
-### 3. Role 归一化 (Normalize OpenAIRoleForPrompt)
-在 `internal/adapter/openai/responses_input_items.go` 等处，引入了特定的 `normalizeOpenAIRoleForPrompt(role)` 清洗，保证输入和传递给上游的 Role 枚举始终受控，消除了因为意外的身份字段传参崩溃。
-
---
-
-## 报告总结与 tool_sieve 的本质作用
-
-PR #74 / #75 并没有从模型本身开刀，而是基于**网关应足够健壮**的设计哲学。
-
-**其实整个增强实现，本质上实现了一个名为 `tool_sieve` (工具筛子) 的中间层网关。**
-面对 DeepSeek 这种吐出一团混合了聊天文字与 JSON 面团的“不标准”数据流，`tool_sieve` 就像一个勤劳的高精度筛子，不仅人工揉开了面团：
-1. 它把散文分拣出来，塞回标准结构的 `content` 字段去展示；
-2. 剥离并清洗出有瑕疵的 JSON 块，按照 OpenAI 的标准格式小心翼翼地放进 `tool_calls` 结构里去等待执行。
-
-这意味着，即便 AI 被配置了奇怪的回复设定、加粗了强调语言，甚至是犯了标点符号拼写小失误，**只要它输出了可以拼凑成工具指令的 JSON 核心单元，整个中继层就能将其挽救，并把正确的工具结果呈现给模型和用户**。 这不仅修复了缺陷，更极大地增强了工具网关的通用性和鲁棒性。
--- a/plans/refactor-baseline.md
+++ b/plans/refactor-baseline.md
@@ -1,32 +0,0 @@
-# DS2API Refactor Baseline (Historical Snapshot)
-
- Snapshot time: `2026-02-22T08:53:54Z`
- Snapshot branch: `dev`
- Snapshot HEAD: `5d3989a`
- Scope: backend + node api + webui large-file decoupling (no behavior change)
-
-## Gate Commands
-
-1. `./tests/scripts/run-unit-all.sh`
-   - Result: PASS
-   - Includes:
-     - `go test ./...`
-     - `node --test api/helpers/stream-tool-sieve.test.js api/chat-stream.test.js api/compat/js_compat_test.js`
-2. `npm --prefix webui run build`
-   - Result: PASS
-3. `./tests/scripts/check-refactor-line-gate.sh`
-   - Result: PASS (`checked=131 missing=0 over_limit=0`)
-4. Stage gates (1-5) replay:
-   - `go test ./internal/config ./internal/admin ./internal/account ./internal/deepseek ./internal/format/openai` -> PASS
-   - `go test ./internal/adapter/openai ./internal/util ./internal/sse ./internal/compat` -> PASS
-   - `go test ./internal/adapter/claude ./internal/adapter/gemini ./internal/config` -> PASS
-   - `go test ./internal/testsuite ./cmd/ds2api-tests` -> PASS
-   - `node --test api/helpers/stream-tool-sieve.test.js api/chat-stream.test.js api/compat/js_compat_test.js` -> PASS
-5. Final full regression:
-   - `go test ./... -count=1` -> PASS
-
-## Notes
-
- This file records a historical baseline for refactor process tracking.
- It is not intended to represent the current repository HEAD.
- Frontend manual smoke for phase 6 still requires human execution and sign-off.
--- a/plans/refactor-line-gate.md
+++ b/plans/refactor-line-gate.md
@@ -1,22 +0,0 @@
-# Refactor Line Gate
-
-## Rules
-
-1. Backend production files upper bound: `<= 300` lines.
-2. Frontend (`webui/`) production files upper bound: `<= 500` lines.
-3. Entry/facade files upper bound: `<= 120` lines.
-4. Scope is limited to target files in `plans/refactor-line-gate-targets.txt`.
-5. Test files are out of scope for this gate.
-
-## Command
-
-```bash
-./tests/scripts/check-refactor-line-gate.sh
-```
-
-## Naming Note
-
- Original split plan used `internal/admin/handler_accounts_test.go` for account probing logic.
- In Go, `*_test.go` files are test-only compilation units and cannot host production handlers.
- The production file is implemented as `internal/admin/handler_accounts_testing.go`.
-
--- a/tests/compat/expected/toolcalls_fenced_json.json
+++ b/tests/compat/expected/toolcalls_fenced_json.json
@@ -1,6 +1,13 @@
 {
-  "calls": [],
-  "sawToolCallSyntax": false,
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
  "rejectedByPolicy": false,
  "rejectedToolNames": []
-}
+}
--- a/tests/compat/expected/toolcalls_standalone_fenced_example.json
+++ b/tests/compat/expected/toolcalls_standalone_fenced_example.json
@@ -1,6 +1,13 @@
 {
-  "calls": [],
-  "sawToolCallSyntax": false,
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
+  "sawToolCallSyntax": true,
  "rejectedByPolicy": false,
  "rejectedToolNames": []
-}
+}
--- a/tests/compat/expected/toolcalls_standalone_mixed_prose.json
+++ b/tests/compat/expected/toolcalls_standalone_mixed_prose.json
@@ -1,6 +1,13 @@
 {
-  "calls": [],
+  "calls": [
+    {
+      "name": "read_file",
+      "input": {
+        "path": "README.MD"
+      }
+    }
+  ],
  "sawToolCallSyntax": true,
  "rejectedByPolicy": false,
  "rejectedToolNames": []
-}
+}
--- a/tests/node/stream-tool-sieve.test.js
+++ b/tests/node/stream-tool-sieve.test.js
@@ -91,7 +91,9 @@ test('parseToolCalls supports fenced json and function.arguments string payload'
    '```',
  ].join('\n');
  const calls = parseToolCalls(text, ['read_file']);
-  assert.equal(calls.length, 0);
+  assert.equal(calls.length, 1);
+  assert.equal(calls[0].name, 'read_file');
+  assert.equal(calls[0].input.path, 'README.md');
 });

 test('parseToolCalls parses text-kv fallback payload', () => {
@@ -122,19 +124,19 @@ test('parseToolCalls parses multiple text-kv fallback payloads', () => {
  assert.equal(calls[1].name, 'bash');
 });

-test('parseStandaloneToolCalls only matches standalone payload and ignores mixed prose', () => {
+test('parseStandaloneToolCalls parses mixed prose payload', () => {
  const mixed = '这里是示例：{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}，请勿执行。';
  const standalone = '{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}';
  const mixedCalls = parseStandaloneToolCalls(mixed, ['read_file']);
  const standaloneCalls = parseStandaloneToolCalls(standalone, ['read_file']);
-  assert.equal(mixedCalls.length, 0);
+  assert.equal(mixedCalls.length, 1);
  assert.equal(standaloneCalls.length, 1);
 });

-test('parseStandaloneToolCalls ignores fenced code block tool_call examples', () => {
+test('parseStandaloneToolCalls parses fenced code block tool_call payload', () => {
  const fenced = ['```json', '{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}', '```'].join('\n');
  const calls = parseStandaloneToolCalls(fenced, ['read_file']);
-  assert.equal(calls.length, 0);
+  assert.equal(calls.length, 1);
 });


@@ -259,28 +261,28 @@ test('sieve emits final tool_calls for split arguments payload without increment
  assert.deepEqual(finalCalls[0].input, { path: 'README.MD', mode: 'head' });
 });

-test('sieve keeps tool json as text when leading prose exists (strict mode)', () => {
+test('sieve still emits tool_calls when leading prose exists before tool json', () => {
  const events = runSieve(
    ['我将调用工具。', '{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}'],
    ['read_file'],
  );
  const hasTool = events.some((evt) => (evt.type === 'tool_calls' && evt.calls?.length > 0) || (evt.type === 'tool_call_deltas' && evt.deltas?.length > 0));
  const leakedText = collectText(events);
-  assert.equal(hasTool, false);
+  assert.equal(hasTool, true);
  assert.equal(leakedText.includes('我将调用工具。'), true);
-  assert.equal(leakedText.toLowerCase().includes('tool_calls'), true);
+  assert.equal(leakedText.toLowerCase().includes('tool_calls'), false);
 });

-test('sieve keeps same-chunk trailing prose payload as text in strict mode', () => {
+test('sieve emits tool_calls and keeps trailing prose when payload and prose share a chunk', () => {
  const events = runSieve(
    ['{"tool_calls":[{"name":"read_file","input":{"path":"README.MD"}}]}然后继续解释。'],
    ['read_file'],
  );
  const hasTool = events.some((evt) => (evt.type === 'tool_calls' && evt.calls?.length > 0) || (evt.type === 'tool_call_deltas' && evt.deltas?.length > 0));
  const leakedText = collectText(events);
-  assert.equal(hasTool, false);
+  assert.equal(hasTool, true);
  assert.equal(leakedText.includes('然后继续解释。'), true);
-  assert.equal(leakedText.toLowerCase().includes('tool_calls'), true);
+  assert.equal(leakedText.toLowerCase().includes('tool_calls'), false);
 });

 test('formatOpenAIStreamToolCalls reuses ids with the same idStore', () => {
--- a/tests/repair_json_tool.go
+++ b/tests/repair_json_tool.go
@@ -0,0 +1,77 @@
+package main
+
+import (
+	"fmt"
+	"strings"
+)
+
+func repairInvalidJSONBackslashes(s string) string {
+	if !strings.Contains(s, "\\") {
+		return s
+	}
+	var out strings.Builder
+	out.Grow(len(s) + 10)
+	runes := []rune(s)
+	for i := 0; i < len(runes); i++ {
+		if runes[i] == '\\' {
+			if i+1 < len(runes) {
+				next := runes[i+1]
+				switch next {
+				case '"', '\\', '/', 'b', 'f', 'n', 'r', 't':
+					out.WriteRune('\\')
+					out.WriteRune(next)
+					i++
+					continue
+				case 'u':
+					if i+5 < len(runes) {
+						isHex := true
+						for j := 1; j <= 4; j++ {
+							r := runes[i+1+j]
+							if !((r >= '0' && r <= '9') || (r >= 'a' && r <= 'f') || (r >= 'A' && r <= 'F')) {
+								isHex = false
+								break
+							}
+						}
+						if isHex {
+							out.WriteRune('\\')
+							out.WriteRune('u')
+							for j := 1; j <= 4; j++ {
+								out.WriteRune(runes[i+1+j])
+							}
+							i += 5
+							continue
+						}
+					}
+				}
+			}
+			// Not a valid escape sequence, double it
+			out.WriteString("\\\\")
+		} else {
+			out.WriteRune(runes[i])
+		}
+	}
+	return out.String()
+}
+
+func main() {
+	tests := []struct {
+		input    string
+		expected string
+	}{
+		{`{"path": "C:\Users\name"}`, `{"path": "C:\\Users\\name"}`},
+		{`{"cmd": "cd D:\git_codes"}`, `{"cmd": "cd D:\\git_codes"}`},
+		{`{"text": "line1\nline2"}`, `{"text": "line1\nline2"}`},
+		{`{"path": "D:\\back\\slash"}`, `{"path": "D:\\back\\slash"}`},
+		{`{"unicode": "\u2705"}`, `{"unicode": "\u2705"}`},
+		{`{"invalid_u": "\u123"}`, `{"invalid_u": "\\u123"}`},
+	}
+
+	for _, tt := range tests {
+		got := repairInvalidJSONBackslashes(tt.input)
+		if got != tt.expected {
+			fmt.Printf("FAIL: input=%s\n  got=%s\n  exp=%s\n", tt.input, got, tt.expected)
+		} else {
+			fmt.Printf("PASS: input=%s\n", tt.input)
+		}
+	}
+}
Author	SHA1	Message	Date
CJACK.	b3eae22cef	Merge pull request #111 from CJackHwang/dev Merge pull request #110 from CJackHwang/codex/align-js-runtime-with-go-runtime-logic Align Vercel JS stream tool-call delta handling with Go runtime	2026-03-20 10:05:25 +08:00
CJACK.	7af0098d1b	Merge pull request #110 from CJackHwang/codex/align-js-runtime-with-go-runtime-logic Align Vercel JS stream tool-call delta handling with Go runtime	2026-03-20 09:49:08 +08:00
CJACK.	17405be300	shrink vercel stream module under line gate limit	2026-03-20 09:47:22 +08:00
CJACK.	5bc03e5de6	align vercel js stream toolcall delta behavior with go runtime	2026-03-20 09:36:45 +08:00
CJACK.	5a5f93148d	Merge pull request #109 from CJackHwang/dev Merge pull request #108 from CJackHwang/codex/clean-up-unused-files-and-update-documentation-uiip50 docs: refresh deployment/testing guides and remove stale investigation report	2026-03-20 03:12:25 +08:00
CJACK.	32dc5b6099	Merge pull request #108 from CJackHwang/codex/clean-up-unused-files-and-update-documentation-uiip50 docs: refresh deployment/testing guides and remove stale investigation report	2026-03-20 03:08:09 +08:00
CJACK.	7936d4675f	Merge pull request #107 from CJackHwang/codex/clean-up-unused-files-and-update-documentation docs: prune stale files and refresh docs, add .env.example, align READMEs/DEPLOY/CONTRIBUTING	2026-03-20 03:07:21 +08:00
CJACK.	808eafa7c6	docs: refresh deployment/testing guides and prune stale report	2026-03-20 03:05:36 +08:00
CJACK.	bcb8ed6df2	docs: prune stale docs and refresh project documentation	2026-03-20 03:05:22 +08:00
CJACK.	8ec5dcc0cc	Merge pull request #106 from CJackHwang/dev Merge pull request #105 from CJackHwang/codex/fix-issues-found-in-review Merge pull request #104 from CJackHwang/codex/revert-to-commit-efb484b Restore tool-call parsing and repair logic; remove accidental split files	2026-03-20 02:53:30 +08:00
CJACK.	88a79f212d	Fix path control-char repair on JSON fallback parses	2026-03-20 02:52:27 +08:00
CJACK.	b1f8d6192f	Merge pull request #105 from CJackHwang/codex/fix-issues-found-in-review Merge pull request #104 from CJackHwang/codex/revert-to-commit-efb484b Restore tool-call parsing and repair logic; remove accidental split files	2026-03-20 02:38:35 +08:00
CJACK.	acfb3b225d	Split toolcall input parsing to satisfy line gate	2026-03-20 02:37:23 +08:00
CJACK.	99a6164000	Fix path corruption when parsing tool call JSON strings	2026-03-20 02:31:37 +08:00
CJACK.	e49d9d33e2	Merge pull request #104 from CJackHwang/codex/revert-to-commit-efb484b Restore tool-call parsing and repair logic; remove accidental split files	2026-03-20 02:17:52 +08:00
CJACK.	184a3d1e4e	Sync Node tool-call parsing with aggressive fenced/mixed policy	2026-03-20 02:16:37 +08:00
CJACK.	c4ec14f49a	Fix refactor line gate for toolcalls_parse	2026-03-20 02:12:34 +08:00
CJACK.	fb5fc0e885	Default to aggressive tool-call interception in mixed/fenced text	2026-03-20 02:03:46 +08:00
CJACK.	20b603666d	Allow standalone parser to detect mixed prose tool JSON	2026-03-20 02:03:32 +08:00
CJACK.	4d549b7102	Revert "Merge branch 'dev' into codex/fix-issues-found-in-review" This reverts commit `33b0d1d144`, reversing changes made to `efb484ba4f`.	2026-03-20 01:38:11 +08:00
CJACK.	33b0d1d144	Merge branch 'dev' into codex/fix-issues-found-in-review	2026-03-20 01:23:00 +08:00
CJACK.	41c0f7ce28	Merge pull request #102 from CJackHwang/dev Merge pull request #99 from CJackHwang/codex/refactor-toolcalls_parse.go-for-line-limits Codex-generated pull request	2026-03-20 01:18:05 +08:00
CJACK.	efb484ba4f	Merge pull request #103 from CJackHwang/codex/fix-threshold-issue-and-audit-pr fix: unblock PR #101 line gate and improve PoW/token retry handling	2026-03-20 01:16:46 +08:00
CJACK.	145501d4a5	fix(tool-sieve): allow mixed prose + tool json interception	2026-03-20 01:15:32 +08:00
CJACK.	2d5103997b	fix(tool-sieve): keep mixed prose tool json in strict text mode	2026-03-20 01:15:15 +08:00
CJACK.	52e7e7aae8	fix: unblock line gate and harden pow token recovery	2026-03-20 00:50:05 +08:00
CJACK.	5b5a4000d7	Merge pull request #99 from CJackHwang/codex/refactor-toolcalls_parse.go-for-line-limits Codex-generated pull request	2026-03-19 21:06:45 +08:00
CJACK.	2bbf603148	fix: address PR #97 review findings	2026-03-18 00:52:24 +08:00
CJACK.	d14b8a0664	Stabilize tool-call parsing and pass refactor gate	2026-03-18 00:45:28 +08:00
CJACK.	f16e0b579e	Merge pull request #92 from valkryhx/main fix(toolcall): fix deepseek function calling bug and add json repair	2026-03-18 00:15:47 +08:00
CJACK.	43cbc4aac0	Merge pull request #97 from CJackHwang/dev Merge pull request #96 from CJackHwang/codex/update-ci-line-count-limits-cihke3 ci: ignore test files in line gate and raise frontend limit to 500	2026-03-18 00:15:03 +08:00
huangxun	cf569f4749	docs: add testing documentation for tool call debugging - Add targeted test commands to TESTING.md for debugging tool call issues - Add quick test commands reference in README.md - Document specific test cases for DeepSeek tool call parsing	2026-03-17 16:41:16 +08:00
huangxun	c9c59f2490	refactor(toolcall): enhance tool call extraction with multiple keywords and safety limits - Add support for multiple keywords: tool_calls, function.name:, [tool_call_history] - Add OOM protection with search limits in extractToolCallObjects - Add max scan length limit in extractJSONObject to prevent OOM on unclosed objects - Update tool_sieve to handle more tool call patterns - Add loose JSON repair in parseToolCallPayload for better error recovery This improves DeepSeek tool call parsing robustness.	2026-03-17 16:28:27 +08:00
huangxun	16216cc2ca	fix(toolcalls): support nested objects in missing array brackets repair - Upgrade missingArrayBracketsPattern regex to support single-level nested {} objects - This fixes DeepSeek's list hallucination where tool call JSON objects contain nested fields like {"input": {"q": "value"}} - Add comprehensive test cases covering 2-5 nested objects, mixed nested/primitive fields, and real DeepSeek 8-queen output patterns - Add RepairLooseJSON function to repair unquoted keys and missing array brackets Fixes: DeepSeek tool call parsing with nested JSON objects	2026-03-17 16:24:16 +08:00
huangxun	7318d1f4a8	fix(toolcall): fix deepseek function calling bug and add json repair - Fix: Expand stream sieve keywords to support function.name: and [TOOL_CALL_HISTORY] - Fix: Add repairInvalidJSONBackslashes to handle unescaped backslashes in Windows paths - Sync: Update JS stream sieve to match Go implementation - Test: Add unit tests for backslash repair and deepseek format parsing - Tool: Move repair json test tool to tests/repair_json_tool.go	2026-03-13 13:47:40 +08:00