refactor: improve XML tool parsing robustness, update system prompt constraints, and simplify tool filtering logic

API.en.md

@@ -267,6 +267,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` models emit `delta.reasoning_content`
 - Text emits `delta.content`
 - Last chunk includes `finish_reason` and `usage`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 #### Tool Calls
 
@@ -383,6 +384,7 @@ Business auth required. Returns OpenAI-compatible embeddings shape.
 ## Claude-Compatible API
 
 Besides `/anthropic/v1/*`, DS2API also supports shortcut paths: `/v1/messages`, `/messages`, `/v1/messages/count_tokens`, `/messages/count_tokens`.
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `GET /anthropic/v1/models`
 
@@ -517,6 +519,7 @@ Supported paths:
 - `/v1/models/{model}:streamGenerateContent` (compat path)
 
 Authentication is the same as other business routes (`Authorization: Bearer <token>` or `x-api-key`).
+Implementation-wise this path is unified on the OpenAI Chat Completions parse-and-translate pipeline to avoid maintaining divergent parsing chains.
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -535,6 +538,7 @@ Returns SSE (`text/event-stream`), each chunk as `data: <json>`:
 - regular text: incremental text chunks
 - `tools` mode: buffered and emitted as `functionCall` at finalize phase
 - final chunk: includes `finishReason: "STOP"` and `usageMetadata`
+- Token counting prefers pass-through from upstream DeepSeek SSE (`accumulated_token_usage` / `token_usage`), and only falls back to local estimation when upstream usage is absent
 
 ---
 

API.md

@@ -267,6 +267,7 @@ data: [DONE]
 - `deepseek-reasoner` / `deepseek-reasoner-search` 模型输出 `delta.reasoning_content`
 - 普通文本输出 `delta.content`
 - 最后一段包含 `finish_reason` 和 `usage`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 #### Tool Calls
 
@@ -389,6 +390,7 @@ data: [DONE]
 ## Claude 兼容接口
 
 除标准路径 `/anthropic/v1/*` 外，还支持快捷路径 `/v1/messages`、`/messages`、`/v1/messages/count_tokens`、`/messages/count_tokens`。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `GET /anthropic/v1/models`
 
@@ -523,6 +525,7 @@ data: {"type":"message_stop"}
 - `/v1/models/{model}:streamGenerateContent`（兼容路径）
 
 鉴权方式同业务接口（`Authorization: Bearer <token>` 或 `x-api-key`）。
+实现上统一走 OpenAI Chat Completions 解析与回译链路，避免多套解析逻辑分叉维护。
 
 ### `POST /v1beta/models/{model}:generateContent`
 
@@ -541,6 +544,7 @@ data: {"type":"message_stop"}
 - 常规文本：持续返回增量文本 chunk
 - `tools` 场景：会缓冲并在结束时输出 `functionCall` 结构
 - 结束 chunk：包含 `finishReason: "STOP"` 与 `usageMetadata`
+- token 计数优先透传上游 DeepSeek SSE（如 `accumulated_token_usage` / `token_usage`）；仅在上游缺失时回退本地估算
 
 ---
 

README.MD

@@ -48,7 +48,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(并发槽位 + 等待队列)"]
             DSClient["DeepSeek Client\n(Session / Auth / HTTP)"]
-            Pow["PoW WASM\n(wazero 预加载)"]
+            Pow["PoW 实现\n(纯 Go 毫秒级)"]
             Tool["Tool Sieve\n(Go/Node 语义对齐)"]
         end
     end
@@ -95,7 +95,7 @@ flowchart LR
 | Gemini 兼容 | `POST /v1beta/models/{model}:generateContent`、`POST /v1beta/models/{model}:streamGenerateContent`（及 `/v1/models/{model}:*` 路径） |
 | 多账号轮询 | 自动 token 刷新、邮箱/手机号双登录方式 |
 | 并发队列控制 | 每账号 in-flight 上限 + 等待队列，动态计算建议并发值 |
-| DeepSeek PoW | WASM 计算（`wazero`），无需外部 Node.js 依赖 |
+| DeepSeek PoW | 纯 Go 高性能实现（DeepSeekHashV1），毫秒级响应 |
 | Tool Calling | 防泄漏处理：非代码块高置信特征识别、`delta.tool_calls` 早发、结构化增量输出 |
 | Admin API | 配置管理、运行时设置热更新、账号测试 / 批量测试、会话清理、导入导出、Vercel 同步、版本检查 |
 | WebUI 管理台 | `/admin` 单页应用（中英文双语、深色模式） |
@@ -344,7 +344,6 @@ cp opencode.json.example opencode.json
 | `DS2API_CONFIG_PATH` | 配置文件路径 | `config.json` |
 | `DS2API_CONFIG_JSON` | 直接注入配置（JSON 或 Base64） | — |
 | `DS2API_ENV_WRITEBACK` | 环境变量模式下自动写回配置文件并切换文件模式（`1/true/yes/on`） | 关闭 |
-| `DS2API_WASM_PATH` | PoW WASM 文件路径 | 自动查找 |
 | `DS2API_STATIC_ADMIN_DIR` | 管理台静态文件目录 | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | 启动时自动构建 WebUI | 本地开启，Vercel 关闭 |
 | `DS2API_DEV_PACKET_CAPTURE` | 本地开发抓包开关（记录最近会话请求/响应体） | 本地非 Vercel 默认开启 |
@@ -455,7 +454,7 @@ ds2api/
 │   ├── claudeconv/          # Claude 消息格式转换
 │   ├── compat/              # Go 版本兼容与回归测试辅助
 │   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek API 客户端、PoW WASM
+│   ├── deepseek/            # DeepSeek API 客户端、PoW 逻辑
 │   ├── js/                  # Node 运行时流式处理与兼容逻辑
 │   ├── devcapture/          # 开发抓包模块
 │   ├── rawsample/           # 原始流样本可见文本提取与回放辅助

README.en.md

@@ -48,7 +48,7 @@ flowchart LR
             Auth["Auth Resolver\n(API key / bearer / x-goog-api-key)"]
             Pool["Account Pool + Queue\n(in-flight slots + wait queue)"]
             DSClient["DeepSeek Client\n(session / auth / HTTP)"]
-            Pow["PoW WASM\n(wazero preload)"]
+            Pow["PoW Solver\n(Pure Go ms-level)"]
             Tool["Tool Sieve\n(Go/Node semantic parity)"]
         end
     end
@@ -95,7 +95,7 @@ flowchart LR
 | Gemini compatible | `POST /v1beta/models/{model}:generateContent`, `POST /v1beta/models/{model}:streamGenerateContent` (plus `/v1/models/{model}:*` paths) |
 | Multi-account rotation | Auto token refresh, email/mobile dual login |
 | Concurrency control | Per-account in-flight limit + waiting queue, dynamic recommended concurrency |
-| DeepSeek PoW | WASM solving via `wazero`, no external Node.js dependency |
+| DeepSeek PoW | Pure Go high-performance solver (DeepSeekHashV1), ms-level response |
 | Tool Calling | Anti-leak handling: non-code-block feature match, early `delta.tool_calls`, structured incremental output |
 | Admin API | Config management, runtime settings hot-reload, account testing/batch test, session cleanup, import/export, Vercel sync, version check |
 | WebUI Admin Panel | SPA at `/admin` (bilingual Chinese/English, dark mode) |
@@ -344,7 +344,6 @@ cp opencode.json.example opencode.json
 | `DS2API_CONFIG_PATH` | Config file path | `config.json` |
 | `DS2API_CONFIG_JSON` | Inline config (JSON or Base64) | — |
 | `DS2API_ENV_WRITEBACK` | Auto-write env-backed config to file and transition to file mode (`1/true/yes/on`) | Disabled |
-| `DS2API_WASM_PATH` | PoW WASM file path | Auto-detect |
 | `DS2API_STATIC_ADMIN_DIR` | Admin static assets dir | `static/admin` |
 | `DS2API_AUTO_BUILD_WEBUI` | Auto-build WebUI on startup | Enabled locally, disabled on Vercel |
 | `DS2API_ACCOUNT_MAX_INFLIGHT` | Max in-flight requests per account | `2` |
@@ -453,7 +452,7 @@ ds2api/
 │   ├── claudeconv/          # Claude message format conversion
 │   ├── compat/              # Go-version compatibility and regression helpers
 │   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek API client, PoW WASM
+│   ├── deepseek/            # DeepSeek API client, PoW logic
 │   ├── js/                  # Node runtime stream/compat logic
 │   ├── devcapture/          # Dev packet capture module
 │   ├── rawsample/           # Visible-text extraction and replay helpers for raw stream samples

docs/CONTRIBUTING.en.md

@@ -115,7 +115,7 @@ ds2api/
 │   ├── claudeconv/          # Claude message conversion
 │   ├── compat/              # Go-version compatibility and regression helpers
 │   ├── config/              # Config loading, validation, and hot-reload
-│   ├── deepseek/            # DeepSeek client, PoW WASM
+│   ├── deepseek/            # DeepSeek client, PoW logic
 │   ├── js/                  # Node runtime stream/compat logic
 │   ├── devcapture/          # Dev packet capture
 │   ├── format/              # Output formatting

docs/CONTRIBUTING.md

@@ -115,7 +115,7 @@ ds2api/
 │   ├── claudeconv/          # Claude 消息格式转换
 │   ├── compat/              # Go 版本兼容与回归测试辅助
 │   ├── config/              # 配置加载、校验与热更新
-│   ├── deepseek/            # DeepSeek 客户端、PoW WASM
+│   ├── deepseek/            # DeepSeek 客户端、PoW 逻辑
 │   ├── js/                  # Node 运行时流式/兼容逻辑
 │   ├── devcapture/          # 开发抓包
 │   ├── format/              # 输出格式化

docs/DEPLOY.en.md

@@ -366,7 +366,6 @@ Each archive includes:
 
 - `ds2api` executable (`ds2api.exe` on Windows)
 - `static/admin/` (built WebUI assets)
-- `sha3_wasm_bg.7b9ca65ddd.wasm` (optional; binary has embedded fallback)
 - `config.example.json`, `.env.example`
 - `README.MD`, `README.en.md`, `LICENSE`
 
@@ -456,8 +455,6 @@ server {
 # Copy compiled binary and related files to target directory
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# Optional: if you want to use an external WASM file (override the embedded one, from a release package or build output)
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

docs/DEPLOY.md

@@ -366,7 +366,6 @@ 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`（可选；程序内置 embed fallback）
 - `config.example.json`、`.env.example`
 - `README.MD`、`README.en.md`、`LICENSE`
 
@@ -456,8 +455,6 @@ server {
 # 将编译好的二进制文件和相关文件复制到目标目录
 sudo mkdir -p /opt/ds2api
 sudo cp ds2api config.json /opt/ds2api/
-# 可选：若你希望使用外置 WASM 文件（覆盖内置版本，来自 release 包或构建产物）
-# sudo cp /path/to/sha3_wasm_bg.7b9ca65ddd.wasm /opt/ds2api/
 sudo cp -r static/admin /opt/ds2api/static/admin
 ```
 

internal/adapter/openai/chat_stream_runtime.go

@@ -37,6 +37,7 @@ type chatStreamRuntime struct {
 	streamToolNames   map[int]string
 	thinking          strings.Builder
 	text              strings.Builder
+	promptTokens      int
 	outputTokens      int
 }
 
@@ -170,11 +171,16 @@ func (s *chatStreamRuntime) finalize(finishReason string) {
 		finishReason = "tool_calls"
 	}
 	usage := openaifmt.BuildChatUsage(s.finalPrompt, finalThinking, finalText)
+	if s.promptTokens > 0 {
+		usage["prompt_tokens"] = s.promptTokens
+	}
 	if s.outputTokens > 0 {
 		usage["completion_tokens"] = s.outputTokens
-		if prompt, ok := usage["prompt_tokens"].(int); ok {
-			usage["total_tokens"] = prompt + s.outputTokens
-		}
+	}
+	if s.promptTokens > 0 || s.outputTokens > 0 {
+		p := usage["prompt_tokens"].(int)
+		c := usage["completion_tokens"].(int)
+		usage["total_tokens"] = p + c
 	}
 	s.sendChunk(openaifmt.BuildChatStreamChunk(
 		s.completionID,
@@ -190,6 +196,9 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.PromptTokens > 0 {
+		s.promptTokens = parsed.PromptTokens
+	}
 	if parsed.OutputTokens > 0 {
 		s.outputTokens = parsed.OutputTokens
 	}
@@ -243,7 +252,7 @@ func (s *chatStreamRuntime) onParsed(parsed sse.LineResult) streamengine.ParsedD
 						if !s.emitEarlyToolDeltas {
 							continue
 						}
-						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.toolNames, s.streamToolNames)
+						filtered := filterIncrementalToolCallDeltasByAllowed(evt.ToolCallDeltas, s.streamToolNames)
 						if len(filtered) == 0 {
 							continue
 						}

internal/adapter/openai/handler_chat.go

@@ -131,12 +131,17 @@ func (h *Handler) handleNonStream(w http.ResponseWriter, ctx context.Context, re
 		return
 	}
 	respBody := openaifmt.BuildChatCompletion(completionID, model, finalPrompt, finalThinking, finalText, toolNames)
-	if result.OutputTokens > 0 {
+	if result.PromptTokens > 0 || result.OutputTokens > 0 {
 		if usage, ok := respBody["usage"].(map[string]any); ok {
-			usage["completion_tokens"] = result.OutputTokens
-			if prompt, ok := usage["prompt_tokens"].(int); ok {
-				usage["total_tokens"] = prompt + result.OutputTokens
+			if result.PromptTokens > 0 {
+				usage["prompt_tokens"] = result.PromptTokens
 			}
+			if result.OutputTokens > 0 {
+				usage["completion_tokens"] = result.OutputTokens
+			}
+			p, _ := usage["prompt_tokens"].(int)
+			c, _ := usage["completion_tokens"].(int)
+			usage["total_tokens"] = p + c
 		}
 	}
 	writeJSON(w, http.StatusOK, respBody)

internal/adapter/openai/handler_toolcall_format.go

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

internal/adapter/openai/prompt_build_test.go

@@ -74,16 +74,13 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 	}
 
 	finalPrompt, _ := buildOpenAIFinalPrompt(messages, tools, "")
-	if !strings.Contains(finalPrompt, "After receiving a tool result, use it directly.") {
-		t.Fatalf("vercel prepare finalPrompt missing final-answer instruction: %q", finalPrompt)
-	}
-	if !strings.Contains(finalPrompt, "Only call another tool if the result is insufficient.") {
-		t.Fatalf("vercel prepare finalPrompt missing retry guard instruction: %q", finalPrompt)
+	if !strings.Contains(finalPrompt, "Remember: Output ONLY the <tool_calls>...</tool_calls> XML block when calling tools.") {
+		t.Fatalf("vercel prepare finalPrompt missing final tool-call anchor instruction: %q", finalPrompt)
 	}
 	if !strings.Contains(finalPrompt, "TOOL CALL FORMAT") {
 		t.Fatalf("vercel prepare finalPrompt missing xml format instruction: %q", finalPrompt)
 	}
-	if !strings.Contains(finalPrompt, "Do NOT wrap the XML in markdown code fences") {
+	if !strings.Contains(finalPrompt, "Do NOT wrap XML in markdown fences") {
 		t.Fatalf("vercel prepare finalPrompt missing no-fence xml instruction: %q", finalPrompt)
 	}
 	if strings.Contains(finalPrompt, "```json") {

internal/adapter/openai/responses_handler.go

@@ -130,12 +130,17 @@ func (h *Handler) handleResponsesNonStream(w http.ResponseWriter, resp *http.Res
 	}
 
 	responseObj := openaifmt.BuildResponseObject(responseID, model, finalPrompt, sanitizedThinking, sanitizedText, toolNames)
-	if result.OutputTokens > 0 {
+	if result.PromptTokens > 0 || result.OutputTokens > 0 {
 		if usage, ok := responseObj["usage"].(map[string]any); ok {
-			usage["output_tokens"] = result.OutputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + result.OutputTokens
+			if result.PromptTokens > 0 {
+				usage["input_tokens"] = result.PromptTokens
 			}
+			if result.OutputTokens > 0 {
+				usage["output_tokens"] = result.OutputTokens
+			}
+			input, _ := usage["input_tokens"].(int)
+			output, _ := usage["output_tokens"].(int)
+			usage["total_tokens"] = input + output
 		}
 	}
 	h.getResponseStore().put(owner, responseID, responseObj)

internal/adapter/openai/responses_stream_runtime_core.go

@@ -51,6 +51,7 @@ type responsesStreamRuntime struct {
 	messagePartAdded  bool
 	sequence          int
 	failed            bool
+	promptTokens      int
 	outputTokens      int
 
 	persistResponse func(obj map[string]any)
@@ -152,9 +153,19 @@ func (s *responsesStreamRuntime) finalize() {
 	if s.outputTokens > 0 {
 		if usage, ok := obj["usage"].(map[string]any); ok {
 			usage["output_tokens"] = s.outputTokens
-			if input, ok := usage["input_tokens"].(int); ok {
-				usage["total_tokens"] = input + s.outputTokens
+		}
+	}
+	if s.promptTokens > 0 || s.outputTokens > 0 {
+		if usage, ok := obj["usage"].(map[string]any); ok {
+			if s.promptTokens > 0 {
+				usage["input_tokens"] = s.promptTokens
 			}
+			if s.outputTokens > 0 {
+				usage["output_tokens"] = s.outputTokens
+			}
+			input, _ := usage["input_tokens"].(int)
+			output, _ := usage["output_tokens"].(int)
+			usage["total_tokens"] = input + output
 		}
 	}
 	if s.persistResponse != nil {
@@ -185,6 +196,9 @@ func (s *responsesStreamRuntime) onParsed(parsed sse.LineResult) streamengine.Pa
 	if !parsed.Parsed {
 		return streamengine.ParsedDecision{}
 	}
+	if parsed.PromptTokens > 0 {
+		s.promptTokens = parsed.PromptTokens
+	}
 	if parsed.OutputTokens > 0 {
 		s.outputTokens = parsed.OutputTokens
 	}

internal/adapter/openai/responses_stream_runtime_events.go

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