feat: add OpenAI/Anthropic tools support with tool emulation

- Parse tools/tool_choice from OpenAI and Anthropic requests - Inject tool definitions into system prompt via toolemulation - Parse action blocks (```json action) from model responses - Retry logic for forced tool_choice (any/required) - Return proper tool_calls / tool_use in responses - Support streaming tools via collect-and-replay pattern - Add tool history projection (assistant tool_calls + tool results) - Model ID normalization: use official names (Qwen3.6-Plus, etc.) - Fix resolveSessionMode to use Fresh mode when tools present
2026-04-25 13:37:58 +08:00
parent c49b4b63e7
commit 74bbd8e6d2
13 changed files with 648 additions and 115 deletions
--- a/README.md
+++ b/README.md
@@ -14,6 +14,8 @@ Current scope:
 - one request at a time
 - supports Windows named-pipe transport and local websocket transport
 - directly uses Lingma IPC, not DOM/CDP
+- OpenAI-compatible `tools` / `tool_choice` support (tool emulation via prompt engineering)
+- Anthropic-compatible `tools` / `tool_choice` support

 ## Run

@@ -69,6 +71,41 @@ Recommended layout:
 }
 ```

+## macOS / Linux
+
+This project also works on macOS and Linux via **WebSocket transport**. The Windows named-pipe transport is automatically skipped on non-Windows platforms.
+
+### Run on macOS
+
+```bash
+cd ~/OpenSources/lingma-ipc-proxy
+go run ./cmd/lingma-ipc-proxy --transport websocket --port 8095
+
+# Or use auto-detect (will discover websocket port from Lingma's shared client cache)
+go run ./cmd/lingma-ipc-proxy --port 8095
+```
+
+### Build on macOS / Linux
+
+```bash
+cd ~/OpenSources/lingma-ipc-proxy
+go build -o ./dist/lingma-ipc-proxy ./cmd/lingma-ipc-proxy
+```
+
+### macOS Config Example
+
+```json
+{
+  "host": "127.0.0.1",
+  "port": 8095,
+  "transport": "websocket",
+  "mode": "agent",
+  "shell_type": "zsh",
+  "session_mode": "auto",
+  "timeout": 120
+}
+```
+
 ## Build

 Build a Windows executable:
--- a/cmd/lingma-ipc-proxy/main.go
+++ b/cmd/lingma-ipc-proxy/main.go
@@ -10,6 +10,7 @@ import (
 	"os"
 	"os/signal"
 	"path/filepath"
+	"runtime"
 	"strconv"
 	"strings"
 	"syscall"
@@ -88,7 +89,7 @@ func loadConfig() (service.Config, string) {
 		Transport:   lingmaipc.TransportAuto,
 		Cwd:         currentDir(),
 		Mode:        "agent",
-		ShellType:   "powershell",
+		ShellType:   defaultShellType(),
 		SessionMode: service.SessionModeAuto,
 		Timeout:     120 * time.Second,
 	}
@@ -299,3 +300,10 @@ func valueOr(value string, fallback string) string {
 	}
 	return fallback
 }
+
+func defaultShellType() string {
+	if runtime.GOOS == "windows" {
+		return "powershell"
+	}
+	return "zsh"
+}
--- a/go.mod
+++ b/go.mod
@@ -1,6 +1,6 @@
 module lingma-ipc-proxy

-go 1.25.0
+go 1.21

 require (
 	github.com/Microsoft/go-winio v0.6.2
--- a/internal/httpapi/server.go
+++ b/internal/httpapi/server.go
@@ -9,6 +9,7 @@ import (
 	"time"

 	"lingma-ipc-proxy/internal/service"
+	"lingma-ipc-proxy/internal/toolemulation"
 )

 type Server struct {
@@ -23,6 +24,8 @@ type anthropicRequest struct {
 	System     any          `json:"system,omitempty"`
 	Messages   []rawMessage `json:"messages"`
 	Stream     bool         `json:"stream,omitempty"`
+	Tools      any          `json:"tools,omitempty"`
+	ToolChoice any          `json:"tool_choice,omitempty"`
 }

 type openAIChatRequest struct {
@@ -31,11 +34,15 @@ type openAIChatRequest struct {
 	Stream              bool         `json:"stream,omitempty"`
 	MaxTokens           int          `json:"max_tokens,omitempty"`
 	MaxCompletionTokens int          `json:"max_completion_tokens,omitempty"`
+	Tools               any          `json:"tools,omitempty"`
+	ToolChoice          any          `json:"tool_choice,omitempty"`
 }

 type rawMessage struct {
 	Role       string `json:"role"`
 	Content    any    `json:"content"`
+	ToolCalls  []any  `json:"tool_calls,omitempty"`
+	ToolCallID string `json:"tool_call_id,omitempty"`
 }

 type modelResponse struct {
@@ -170,13 +177,26 @@ func (s *Server) handleAnthropicMessages(w http.ResponseWriter, r *http.Request)
 		return
 	}

+	content := []map[string]any{{"type": "text", "text": result.Text}}
+	stopReason := "end_turn"
+	if len(result.ToolCalls) > 0 {
+		for _, tc := range result.ToolCalls {
+			content = append(content, map[string]any{
+				"type":  "tool_use",
+				"id":    tc.ID,
+				"name":  tc.Name,
+				"input": tc.Arguments,
+			})
+		}
+		stopReason = "tool_use"
+	}
 	writeJSON(w, http.StatusOK, map[string]any{
 		"id":            fmt.Sprintf("msg_%d", time.Now().UnixNano()),
 		"type":          "message",
 		"role":          "assistant",
-		"content":       []map[string]any{{"type": "text", "text": result.Text}},
+		"content":       content,
 		"model":         result.Model,
-		"stop_reason":   "end_turn",
+		"stop_reason":   stopReason,
 		"stop_sequence": nil,
 		"usage": map[string]any{
 			"input_tokens":  result.InputTokens,
@@ -224,6 +244,27 @@ func (s *Server) handleOpenAIChatCompletions(w http.ResponseWriter, r *http.Requ
 	}

 	created := time.Now().Unix()
+	message := map[string]any{
+		"role":    "assistant",
+		"content": result.Text,
+	}
+	finishReason := "stop"
+	if len(result.ToolCalls) > 0 {
+		toolCalls := make([]map[string]any, 0, len(result.ToolCalls))
+		for _, tc := range result.ToolCalls {
+			argsJSON, _ := json.Marshal(tc.Arguments)
+			toolCalls = append(toolCalls, map[string]any{
+				"id":   tc.ID,
+				"type": "function",
+				"function": map[string]any{
+					"name":      tc.Name,
+					"arguments": string(argsJSON),
+				},
+			})
+		}
+		message["tool_calls"] = toolCalls
+		finishReason = "tool_calls"
+	}
 	writeJSON(w, http.StatusOK, map[string]any{
 		"id":      fmt.Sprintf("chatcmpl-%d", time.Now().UnixNano()),
 		"object":  "chat.completion",
@@ -232,11 +273,8 @@ func (s *Server) handleOpenAIChatCompletions(w http.ResponseWriter, r *http.Requ
 		"choices": []map[string]any{
 			{
 				"index":         0,
-				"message": map[string]any{
-					"role":    "assistant",
-					"content": result.Text,
-				},
-				"finish_reason": "stop",
+				"message":       message,
+				"finish_reason": finishReason,
 			},
 		},
 		"usage": map[string]any{
@@ -254,17 +292,73 @@ func (s *Server) handleAnthropicStream(w http.ResponseWriter, r *http.Request, r
 		return
 	}

+	model := strings.TrimSpace(req.Model)
+	if model == "" {
+		model = "lingma"
+	}
+	msgID := fmt.Sprintf("msg_%d", time.Now().UnixNano())
+
+	if len(req.Tools) > 0 {
+		result, err := s.svc.Generate(r.Context(), req)
+		if err != nil {
+			writeAnthropicError(w, http.StatusInternalServerError, "api_error", err.Error())
+			return
+		}
+		streamingHeaders(w)
+		_ = writeSSEEvent(w, flusher, "message_start", map[string]any{
+			"type": "message_start",
+			"message": map[string]any{
+				"id": msgID, "type": "message", "role": "assistant", "content": []any{},
+				"model": model, "stop_reason": nil, "stop_sequence": nil,
+				"usage": map[string]any{"input_tokens": 0, "output_tokens": 0},
+			},
+		})
+		_ = writeSSEEvent(w, flusher, "content_block_start", map[string]any{
+			"type":  "content_block_start", "index": 0,
+			"content_block": map[string]any{"type": "text", "text": ""},
+		})
+		if result.Text != "" {
+			_ = writeSSEEvent(w, flusher, "content_block_delta", map[string]any{
+				"type":  "content_block_delta", "index": 0,
+				"delta": map[string]any{"type": "text_delta", "text": result.Text},
+			})
+		}
+		_ = writeSSEEvent(w, flusher, "content_block_stop", map[string]any{
+			"type": "content_block_stop", "index": 0,
+		})
+		for i, tc := range result.ToolCalls {
+			_ = writeSSEEvent(w, flusher, "content_block_start", map[string]any{
+				"type":  "content_block_start", "index": i + 1,
+				"content_block": map[string]any{"type": "tool_use", "id": tc.ID, "name": tc.Name, "input": map[string]any{}},
+			})
+			argsJSON, _ := json.Marshal(tc.Arguments)
+			_ = writeSSEEvent(w, flusher, "content_block_delta", map[string]any{
+				"type":  "content_block_delta", "index": i + 1,
+				"delta": map[string]any{"type": "input_json_delta", "partial_json": string(argsJSON)},
+			})
+			_ = writeSSEEvent(w, flusher, "content_block_stop", map[string]any{
+				"type": "content_block_stop", "index": i + 1,
+			})
+		}
+		stopReason := "end_turn"
+		if len(result.ToolCalls) > 0 {
+			stopReason = "tool_use"
+		}
+		_ = writeSSEEvent(w, flusher, "message_delta", map[string]any{
+			"type":  "message_delta",
+			"delta": map[string]any{"stop_reason": stopReason, "stop_sequence": nil},
+			"usage": map[string]any{"output_tokens": result.OutputTokens},
+		})
+		_ = writeSSEEvent(w, flusher, "message_stop", map[string]any{"type": "message_stop"})
+		return
+	}
+
 	events, done, err := s.svc.GenerateStream(r.Context(), req)
 	if err != nil {
 		writeAnthropicError(w, http.StatusInternalServerError, "api_error", err.Error())
 		return
 	}

-	model := strings.TrimSpace(req.Model)
-	if model == "" {
-		model = "lingma"
-	}
-	msgID := fmt.Sprintf("msg_%d", time.Now().UnixNano())
 	streamingHeaders(w)
 	if err := writeSSEEvent(w, flusher, "message_start", map[string]any{
 		"type": "message_start",
@@ -383,18 +477,65 @@ func (s *Server) handleOpenAIStream(w http.ResponseWriter, r *http.Request, req
 		return
 	}

-	events, done, err := s.svc.GenerateStream(r.Context(), req)
-	if err != nil {
-		writeOpenAIError(w, http.StatusInternalServerError, "api_error", err.Error())
-		return
-	}
-
 	model := strings.TrimSpace(req.Model)
 	if model == "" {
 		model = "lingma"
 	}
 	chatID := fmt.Sprintf("chatcmpl-%d", time.Now().UnixNano())
 	created := time.Now().Unix()
+
+	if len(req.Tools) > 0 {
+		result, err := s.svc.Generate(r.Context(), req)
+		if err != nil {
+			writeOpenAIError(w, http.StatusInternalServerError, "api_error", err.Error())
+			return
+		}
+		streamingHeaders(w)
+		_ = writeOpenAIChunk(w, flusher, map[string]any{
+			"id": chatID, "object": "chat.completion.chunk", "created": created, "model": model,
+			"choices": []map[string]any{{"index": 0, "delta": map[string]any{"role": "assistant"}, "finish_reason": nil}},
+		})
+		if result.Text != "" {
+			_ = writeOpenAIChunk(w, flusher, map[string]any{
+				"id": chatID, "object": "chat.completion.chunk", "created": created, "model": model,
+				"choices": []map[string]any{{"index": 0, "delta": map[string]any{"content": result.Text}, "finish_reason": nil}},
+			})
+		}
+		for i, tc := range result.ToolCalls {
+			argsJSON, _ := json.Marshal(tc.Arguments)
+			_ = writeOpenAIChunk(w, flusher, map[string]any{
+				"id": chatID, "object": "chat.completion.chunk", "created": created, "model": model,
+				"choices": []map[string]any{{
+					"index": 0,
+					"delta": map[string]any{
+						"tool_calls": []map[string]any{{
+							"index": i, "id": tc.ID, "type": "function",
+							"function": map[string]any{"name": tc.Name, "arguments": string(argsJSON)},
+						}},
+					},
+					"finish_reason": nil,
+				}},
+			})
+		}
+		finishReason := "stop"
+		if len(result.ToolCalls) > 0 {
+			finishReason = "tool_calls"
+		}
+		_ = writeOpenAIChunk(w, flusher, map[string]any{
+			"id": chatID, "object": "chat.completion.chunk", "created": created, "model": model,
+			"choices": []map[string]any{{"index": 0, "delta": map[string]any{}, "finish_reason": finishReason}},
+		})
+		_, _ = fmt.Fprint(w, "data: [DONE]\n\n")
+		flusher.Flush()
+		return
+	}
+
+	events, done, err := s.svc.GenerateStream(r.Context(), req)
+	if err != nil {
+		writeOpenAIError(w, http.StatusInternalServerError, "api_error", err.Error())
+		return
+	}
+
 	streamingHeaders(w)
 	if err := writeOpenAIChunk(w, flusher, map[string]any{
 		"id":      chatID,
@@ -493,22 +634,41 @@ func normalizeAnthropicRequest(req anthropicRequest) (service.ChatRequest, error
 	messages := make([]service.ChatMessage, 0, len(req.Messages))
 	for _, message := range req.Messages {
 		role := strings.ToLower(strings.TrimSpace(message.Role))
-		text := strings.TrimSpace(extractText(message.Content))
-		if role != "user" && role != "assistant" {
-			continue
+		switch role {
+		case "user":
+			text, toolResults := extractAnthropicUserContent(message.Content)
+			for _, tr := range toolResults {
+				prompt := toolemulation.ActionOutputPrompt(tr.ToolUseID, tr.Content)
+				if prompt != "" {
+					messages = append(messages, service.ChatMessage{Role: "user", Text: prompt})
 				}
-		if text == "" {
-			continue
 			}
+			if text != "" {
 				messages = append(messages, service.ChatMessage{Role: role, Text: text})
 			}
+		case "assistant":
+			text, calls := extractAnthropicAssistantContent(message.Content)
+			projected := toolemulation.AssistantToolCallsToText(text, calls)
+			if projected != "" {
+				messages = append(messages, service.ChatMessage{Role: role, Text: projected})
+			}
+		}
+	}
 	if len(messages) == 0 {
 		return service.ChatRequest{}, fmt.Errorf("no user or assistant messages found")
 	}
+
+	toolChoice := toolemulation.ToolChoice{Mode: "auto"}
+	if req.ToolChoice != nil {
+		toolChoice = toolemulation.ExtractToolChoice(req.ToolChoice)
+	}
+
 	return service.ChatRequest{
 		Model:      strings.TrimSpace(req.Model),
 		System:     strings.TrimSpace(extractText(req.System)),
 		Messages:   messages,
+		Tools:      toolemulation.ExtractAnthropicTools(req.Tools),
+		ToolChoice: toolChoice,
 	}, nil
 }

@@ -517,16 +677,31 @@ func normalizeOpenAIRequest(req openAIChatRequest) (service.ChatRequest, error)
 	systemParts := make([]string, 0, 2)
 	for _, message := range req.Messages {
 		role := strings.ToLower(strings.TrimSpace(message.Role))
-		text := strings.TrimSpace(extractText(message.Content))
-		if text == "" {
-			continue
-		}
 		switch role {
 		case "system":
+			text := strings.TrimSpace(extractText(message.Content))
+			if text != "" {
 				systemParts = append(systemParts, text)
-		case "user", "assistant":
+			}
+		case "user":
+			text := strings.TrimSpace(extractText(message.Content))
+			if text != "" {
 				messages = append(messages, service.ChatMessage{Role: role, Text: text})
 			}
+		case "assistant":
+			text := strings.TrimSpace(extractText(message.Content))
+			calls := extractOpenAIToolCalls(message.ToolCalls)
+			projected := toolemulation.AssistantToolCallsToText(text, calls)
+			if projected != "" {
+				messages = append(messages, service.ChatMessage{Role: role, Text: projected})
+			}
+		case "tool":
+			output := strings.TrimSpace(extractText(message.Content))
+			prompt := toolemulation.ActionOutputPrompt(message.ToolCallID, output)
+			if prompt != "" {
+				messages = append(messages, service.ChatMessage{Role: "user", Text: prompt})
+			}
+		}
 	}
 	if len(messages) == 0 {
 		return service.ChatRequest{}, fmt.Errorf("no user or assistant messages found")
@@ -535,6 +710,8 @@ func normalizeOpenAIRequest(req openAIChatRequest) (service.ChatRequest, error)
 		Model:      strings.TrimSpace(req.Model),
 		System:     strings.Join(systemParts, "\n\n"),
 		Messages:   messages,
+		Tools:      toolemulation.ExtractTools(req.Tools),
+		ToolChoice: toolemulation.ExtractToolChoice(req.ToolChoice),
 	}, nil
 }

@@ -681,3 +858,117 @@ func (s *Server) release() {
 	default:
 	}
 }
+
+func extractOpenAIToolCalls(raw []any) []toolemulation.ToolCall {
+	if len(raw) == 0 {
+		return nil
+	}
+	out := make([]toolemulation.ToolCall, 0, len(raw))
+	for _, item := range raw {
+		m, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		id := stringFromAny(m["id"])
+		fn, ok := m["function"].(map[string]any)
+		if !ok {
+			continue
+		}
+		name := stringFromAny(fn["name"])
+		if name == "" {
+			continue
+		}
+		argsRaw := stringFromAny(fn["arguments"])
+		var args map[string]any
+		if argsRaw != "" {
+			_ = json.Unmarshal([]byte(argsRaw), &args)
+		}
+		out = append(out, toolemulation.ToolCall{
+			ID:        id,
+			Name:      name,
+			Arguments: args,
+		})
+	}
+	return out
+}
+
+type anthropicToolResult struct {
+	ToolUseID string
+	Content   string
+}
+
+func extractAnthropicUserContent(content any) (string, []anthropicToolResult) {
+	text := extractText(content)
+	items, ok := content.([]any)
+	if !ok {
+		return text, nil
+	}
+	var results []anthropicToolResult
+	var textParts []string
+	for _, item := range items {
+		m, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		switch stringFromAny(m["type"]) {
+		case "text":
+			if t := stringFromAny(m["text"]); t != "" {
+				textParts = append(textParts, t)
+			}
+		case "tool_result":
+			toolUseID := stringFromAny(m["tool_use_id"])
+			resultText := extractText(m["content"])
+			if resultText != "" {
+				results = append(results, anthropicToolResult{
+					ToolUseID: toolUseID,
+					Content:   resultText,
+				})
+			}
+		}
+	}
+	if len(textParts) > 0 {
+		text = strings.Join(textParts, "\n")
+	}
+	return text, results
+}
+
+func extractAnthropicAssistantContent(content any) (string, []toolemulation.ToolCall) {
+	text := extractText(content)
+	items, ok := content.([]any)
+	if !ok {
+		return text, nil
+	}
+	calls := make([]toolemulation.ToolCall, 0, len(items))
+	var textParts []string
+	for _, item := range items {
+		m, ok := item.(map[string]any)
+		if !ok {
+			continue
+		}
+		switch stringFromAny(m["type"]) {
+		case "text":
+			if t := stringFromAny(m["text"]); t != "" {
+				textParts = append(textParts, t)
+			}
+		case "tool_use":
+			id := stringFromAny(m["id"])
+			name := stringFromAny(m["name"])
+			if name == "" {
+				continue
+			}
+			var args map[string]any
+			if rawInput, ok := m["input"].(map[string]any); ok {
+				args = rawInput
+			}
+			calls = append(calls, toolemulation.ToolCall{
+				ID:        id,
+				Name:      name,
+				Arguments: args,
+			})
+		}
+	}
+	if len(textParts) > 0 {
+		text = strings.Join(textParts, "\n")
+	}
+	return text, calls
+}
--- a/internal/lingmaipc/client.go
+++ b/internal/lingmaipc/client.go
@@ -17,9 +17,6 @@ import (
 )

 const (
-	PipeDir    = `\\.\pipe\`
-	PipePrefix = "lingma-"
-
 	MetaRequestID       = "ai-coding/request-id"
 	MetaMode            = "ai-coding/mode"
 	MetaModel           = "ai-coding/model"
--- a/internal/lingmaipc/pipe_const_other.go
+++ b/internal/lingmaipc/pipe_const_other.go
@@ -0,0 +1,8 @@
+//go:build !windows
+
+package lingmaipc
+
+const (
+	PipeDir    = ""
+	PipePrefix = ""
+)
--- a/internal/lingmaipc/pipe_const_windows.go
+++ b/internal/lingmaipc/pipe_const_windows.go
@@ -0,0 +1,8 @@
+//go:build windows
+
+package lingmaipc
+
+const (
+	PipeDir    = `\\.\pipe\`
+	PipePrefix = "lingma-"
+)
--- a/internal/lingmaipc/pipe_other.go
+++ b/internal/lingmaipc/pipe_other.go
@@ -0,0 +1,12 @@
+//go:build !windows
+
+package lingmaipc
+
+import (
+	"context"
+	"errors"
+)
+
+func connectPipeTransport(ctx context.Context, pipePath string) (framedTransport, error) {
+	return nil, errors.New("pipe transport is only supported on Windows")
+}
--- a/internal/lingmaipc/pipe_windows.go
+++ b/internal/lingmaipc/pipe_windows.go
@@ -0,0 +1,57 @@
+//go:build windows
+
+package lingmaipc
+
+import (
+	"context"
+	"fmt"
+	"net"
+	"sync"
+
+	winio "github.com/Microsoft/go-winio"
+)
+
+type pipeTransport struct {
+	path   string
+	conn   net.Conn
+	reader *framedReader
+	write  sync.Mutex
+}
+
+func connectPipeTransport(ctx context.Context, pipePath string) (framedTransport, error) {
+	conn, err := winio.DialPipeContext(ctx, pipePath)
+	if err != nil {
+		return nil, fmt.Errorf("connect Lingma IPC pipe %s: %w", pipePath, err)
+	}
+	return &pipeTransport{
+		path:   pipePath,
+		conn:   conn,
+		reader: newFramedReader(conn),
+	}, nil
+}
+
+func (t *pipeTransport) ReadFrame() ([]byte, error) {
+	return t.reader.ReadFrame()
+}
+
+func (t *pipeTransport) WriteFrame(body []byte) error {
+	t.write.Lock()
+	defer t.write.Unlock()
+
+	frame := []byte(fmt.Sprintf("Content-Length: %d\r\n\r\n", len(body)))
+	if _, err := t.conn.Write(frame); err != nil {
+		return fmt.Errorf("write frame header: %w", err)
+	}
+	if _, err := t.conn.Write(body); err != nil {
+		return fmt.Errorf("write frame body: %w", err)
+	}
+	return nil
+}
+
+func (t *pipeTransport) Close() error {
+	return t.conn.Close()
+}
+
+func (t *pipeTransport) Address() string {
+	return t.path
+}
--- a/internal/lingmaipc/transport.go
+++ b/internal/lingmaipc/transport.go
@@ -8,7 +8,6 @@ import (
 	"errors"
 	"fmt"
 	"io"
-	"net"
 	"net/url"
 	"os"
 	"path/filepath"
@@ -19,7 +18,6 @@ import (
 	"sync"
 	"time"

-	winio "github.com/Microsoft/go-winio"
 	"github.com/gorilla/websocket"
 )

@@ -74,16 +72,23 @@ func ResolveDialOptions(transport Transport, explicitPipe string, explicitWebSoc
 			return DialOptions{Transport: TransportWebSocket, WebSocketURL: wsURL}, nil
 		}

+		if runtime.GOOS == "windows" {
 			pipePath, pipeErr := ResolvePipePath(explicitPipe)
 			if pipeErr == nil {
 				return DialOptions{Transport: TransportPipe, PipePath: pipePath}, nil
 			}
+			wsURL, wsErr := ResolveWebSocketURL(explicitWebSocketURL)
+			if wsErr == nil {
+				return DialOptions{Transport: TransportWebSocket, WebSocketURL: wsURL}, nil
+			}
+			return DialOptions{}, fmt.Errorf("resolve Lingma transport automatically: pipe: %w; websocket: %v", pipeErr, wsErr)
+		}

 		wsURL, wsErr := ResolveWebSocketURL(explicitWebSocketURL)
 		if wsErr == nil {
 			return DialOptions{Transport: TransportWebSocket, WebSocketURL: wsURL}, nil
 		}
-		return DialOptions{}, fmt.Errorf("resolve Lingma transport automatically: pipe: %w; websocket: %v", pipeErr, wsErr)
+		return DialOptions{}, fmt.Errorf("resolve Lingma transport automatically on %s: websocket: %w", runtime.GOOS, wsErr)
 	case TransportPipe:
 		pipePath, err := ResolvePipePath(explicitPipe)
 		if err != nil {
@@ -307,51 +312,6 @@ func connectTransport(ctx context.Context, opts DialOptions) (framedTransport, e
 	}
 }

-type pipeTransport struct {
-	path   string
-	conn   net.Conn
-	reader *framedReader
-	write  sync.Mutex
-}
-
-func connectPipeTransport(ctx context.Context, pipePath string) (*pipeTransport, error) {
-	conn, err := winio.DialPipeContext(ctx, pipePath)
-	if err != nil {
-		return nil, fmt.Errorf("connect Lingma IPC pipe %s: %w", pipePath, err)
-	}
-	return &pipeTransport{
-		path:   pipePath,
-		conn:   conn,
-		reader: newFramedReader(conn),
-	}, nil
-}
-
-func (t *pipeTransport) ReadFrame() ([]byte, error) {
-	return t.reader.ReadFrame()
-}
-
-func (t *pipeTransport) WriteFrame(body []byte) error {
-	t.write.Lock()
-	defer t.write.Unlock()
-
-	frame := []byte(fmt.Sprintf("Content-Length: %d\r\n\r\n", len(body)))
-	if _, err := t.conn.Write(frame); err != nil {
-		return fmt.Errorf("write frame header: %w", err)
-	}
-	if _, err := t.conn.Write(body); err != nil {
-		return fmt.Errorf("write frame body: %w", err)
-	}
-	return nil
-}
-
-func (t *pipeTransport) Close() error {
-	return t.conn.Close()
-}
-
-func (t *pipeTransport) Address() string {
-	return t.path
-}
-
 type websocketTransport struct {
 	url     string
 	conn    *websocket.Conn
--- a/internal/service/service.go
+++ b/internal/service/service.go
@@ -12,6 +12,7 @@ import (
 	"time"

 	"lingma-ipc-proxy/internal/lingmaipc"
+	"lingma-ipc-proxy/internal/toolemulation"
 )

 type SessionMode string
@@ -45,6 +46,8 @@ type ChatRequest struct {
 	Model      string
 	System     string
 	Messages   []ChatMessage
+	Tools      []toolemulation.ToolDef
+	ToolChoice toolemulation.ToolChoice
 }

 type ChatResult struct {
@@ -62,6 +65,7 @@ type ChatResult struct {
 	Endpoint         string
 	Transport        string
 	EffectiveSession SessionMode
+	ToolCalls        []toolemulation.ToolCall
 }

 type StreamEvent struct {
@@ -77,6 +81,7 @@ type Model struct {
 	ID         string `json:"id"`
 	Name       string `json:"name"`
 	Scene      string `json:"scene,omitempty"`
+	InternalID string `json:"-"`
 }

 type State struct {
@@ -97,6 +102,7 @@ type Service struct {
 	transport       lingmaipc.Transport
 	stickySessionID string
 	stickyModelID   string
+	modelMap        map[string]string // official name -> internal id
 }

 type promptRunResult struct {
@@ -170,6 +176,16 @@ func (s *Service) ListModels(ctx context.Context) ([]Model, error) {
 	if len(models) == 0 {
 		models = []Model{{ID: "lingma", Name: "Lingma", Scene: "default"}}
 	}
+
+	s.mu.Lock()
+	s.modelMap = make(map[string]string, len(models))
+	for _, m := range models {
+		if m.InternalID != "" {
+			s.modelMap[m.ID] = m.InternalID
+		}
+	}
+	s.mu.Unlock()
+
 	return models, nil
 }

@@ -235,17 +251,19 @@ func (s *Service) generateLocked(
 		_ = s.deleteSessionLocked(cleanupCtx, ipcClient, sessionID)
 	}()

+	internalModelID := s.resolveInternalModelID(req.Model)
+
 	requestID := lingmaipc.CreateRequestID("serve")
 	meta := lingmaipc.CreateMeta(lingmaipc.MetaOptions{
 		RequestID:       requestID,
 		Mode:            s.cfg.Mode,
-		Model:           req.Model,
+		Model:           internalModelID,
 		ShellType:       s.cfg.ShellType,
 		CurrentFilePath: s.cfg.CurrentFilePath,
 		EnabledMCP:      []any{},
 	})

-	modelID := strings.TrimSpace(req.Model)
+	modelID := strings.TrimSpace(internalModelID)
 	if modelID != "" && s.shouldSetModel(sessionID, effectiveMode, modelID) {
 		if err := ipcClient.Request(requestCtx, "session/set_model", map[string]any{
 			"sessionId": sessionID,
@@ -284,6 +302,28 @@ func (s *Service) generateLocked(
 	}

 	result = s.buildChatResult(req, sessionID, requestID, prompt, runResult, effectiveMode)
+
+	if len(req.Tools) > 0 {
+		calls, remaining, parseErr := toolemulation.ParseActionBlocks(result.Text, toolemulation.Config{})
+		if parseErr == nil && len(calls) > 0 {
+			result.Text = remaining
+			result.ToolCalls = calls
+		} else if (req.ToolChoice.Mode == "any" || req.ToolChoice.Mode == "tool") && len(calls) == 0 {
+			if !toolemulation.LooksLikeRefusal(result.Text) {
+				hintPrompt := prompt + "\n\nImportant: You must use one of the available tools to answer this request. Output a \"```json action\" block."
+				retryResult, retryErr := s.runPromptLocked(requestCtx, ipcClient, sessionID, hintPrompt, requestID, meta, onDelta)
+				if retryErr == nil && retryResult != nil {
+					retryCalls, retryRemaining, retryParseErr := toolemulation.ParseActionBlocks(retryResult.AssistantText, toolemulation.Config{})
+					if retryParseErr == nil && len(retryCalls) > 0 {
+						result.Text = retryRemaining
+						result.ToolCalls = retryCalls
+						result.OutputTokens = estimateTokens(retryResult.AssistantText)
+					}
+				}
+			}
+		}
+	}
+
 	return result, nil
 }

@@ -546,7 +586,7 @@ func resolveSessionMode(req ChatRequest, configured SessionMode) SessionMode {
 	if configured != SessionModeAuto {
 		return configured
 	}
-	if strings.TrimSpace(req.System) != "" || len(filteredMessages(req.Messages)) > 1 {
+	if len(req.Tools) > 0 || strings.TrimSpace(req.System) != "" || len(filteredMessages(req.Messages)) > 1 {
 		return SessionModeFresh
 	}
 	return SessionModeReuse
@@ -567,13 +607,35 @@ func buildLingmaPrompt(req ChatRequest, mode SessionMode) (string, error) {
 	if mode == SessionModeReuse {
 		return lastUser, nil
 	}
-	if strings.TrimSpace(req.System) == "" && len(messages) == 1 {
+
+	system := strings.TrimSpace(req.System)
+	if len(req.Tools) > 0 {
+		system = toolemulation.InjectTooling(system, req.Tools, req.ToolChoice)
+	}
+
+	if system == "" && len(messages) == 1 {
 		return lastUser, nil
 	}

+	if len(req.Tools) > 0 {
+		parts := make([]string, 0, len(messages)+2)
+		if system != "" {
+			parts = append(parts, system)
+		}
+		for _, message := range messages {
+			role := "User"
+			if message.Role == "assistant" {
+				role = "Assistant"
+			}
+			parts = append(parts, fmt.Sprintf("%s: %s", role, message.Text))
+		}
+		parts = append(parts, "Assistant:")
+		return strings.Join(parts, "\n\n"), nil
+	}
+
 	parts := make([]string, 0, len(messages)+4)
-	if strings.TrimSpace(req.System) != "" {
-		parts = append(parts, "System instructions:", strings.TrimSpace(req.System))
+	if system != "" {
+		parts = append(parts, "System instructions:", system)
 	}
 	parts = append(parts, "Conversation transcript:")
 	for _, message := range messages {
@@ -627,7 +689,7 @@ func extractModels(raw any) []Model {
 				if name == "" {
 					name = id
 				}
-				seen[id] = Model{ID: id, Name: name, Scene: currentScene}
+				seen[name] = Model{ID: name, Name: name, Scene: currentScene, InternalID: id}
 			}
 			for key, child := range typed {
 				nextScene := currentScene
@@ -657,6 +719,15 @@ func likelyModelID(id string) bool {
 	return strings.Contains(lowered, "qwen") || strings.Contains(lowered, "model") || strings.Contains(lowered, "auto") || strings.Contains(lowered, "coder")
 }

+func (s *Service) resolveInternalModelID(officialName string) string {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	if internalID, ok := s.modelMap[officialName]; ok && internalID != "" {
+		return internalID
+	}
+	return officialName
+}
+
 func isSceneKey(key string) bool {
 	switch strings.ToLower(strings.TrimSpace(key)) {
 	case "assistant", "chat", "developer", "inline", "quest":
--- a/lingma-ipc-proxy.macos.json
+++ b/lingma-ipc-proxy.macos.json
@@ -0,0 +1,10 @@
+{
+  "host": "127.0.0.1",
+  "port": 8095,
+  "transport": "websocket",
+  "mode": "agent",
+  "shell_type": "zsh",
+  "session_mode": "auto",
+  "timeout": 120,
+  "cwd": "/Users/tiancheng"
+}
--- a/scripts/test-macos.sh
+++ b/scripts/test-macos.sh
@@ -0,0 +1,74 @@
+#!/bin/bash
+# lingma-ipc-proxy macOS 功能测试脚本
+# 用法: ./scripts/test-macos.sh [host:port]
+
+ENDPOINT="${1:-127.0.0.1:8095}"
+MODEL="dashscope_qwen3_coder"
+PASS=0
+FAIL=0
+
+assert_contains() {
+    local response="$1"
+    local expected="$2"
+    local test_name="$3"
+    if echo "$response" | grep -q "$expected"; then
+        echo "  ✅ $test_name"
+        PASS=$((PASS + 1))
+    else
+        echo "  ❌ $test_name"
+        echo "     期望包含: $expected"
+        echo "     实际响应: $(echo "$response" | head -c 200)"
+        FAIL=$((FAIL + 1))
+    fi
+}
+
+echo "========================================"
+echo "lingma-ipc-proxy macOS 功能测试"
+echo "端点: http://$ENDPOINT"
+echo "========================================"
+
+# 1. 测试 /v1/models
+echo ""
+echo "[1/4] 测试 /v1/models"
+RESPONSE=$(curl -s "http://$ENDPOINT/v1/models" 2>/dev/null || echo "ERROR")
+assert_contains "$RESPONSE" "dashscope_qwen3_coder" "模型列表包含 Qwen3-Coder"
+assert_contains "$RESPONSE" "kmodel" "模型列表包含 Kimi"
+assert_contains "$RESPONSE" '"object":"list"' "响应格式正确"
+
+# 2. 测试 /v1/chat/completions 非流式
+echo ""
+echo "[2/4] 测试 /v1/chat/completions (非流式)"
+RESPONSE=$(curl -s -X POST "http://$ENDPOINT/v1/chat/completions" \
+    -H 'Content-Type: application/json' \
+    -d "{\"model\":\"$MODEL\",\"messages\":[{\"role\":\"user\",\"content\":\"1+1=?\"}],\"stream\":false}" 2>/dev/null || echo "ERROR")
+assert_contains "$RESPONSE" "2" "回答包含正确答案"
+assert_contains "$RESPONSE" "chat.completion" "响应类型正确"
+assert_contains "$RESPONSE" "stop" "finish_reason 为 stop"
+
+# 3. 测试 /v1/chat/completions 流式
+echo ""
+echo "[3/4] 测试 /v1/chat/completions (流式 SSE)"
+RESPONSE=$(curl -s -N -X POST "http://$ENDPOINT/v1/chat/completions" \
+    -H 'Content-Type: application/json' \
+    -d "{\"model\":\"$MODEL\",\"messages\":[{\"role\":\"user\",\"content\":\"1+1=?\"}],\"stream\":true}" 2>/dev/null || echo "ERROR")
+assert_contains "$RESPONSE" "data:" "包含 SSE data: 前缀"
+assert_contains "$RESPONSE" "chat.completion.chunk" "chunk 类型正确"
+
+# 4. 测试 /v1/messages (Anthropic 格式)
+echo ""
+echo "[4/4] 测试 /v1/messages (Anthropic 格式)"
+RESPONSE=$(curl -s -X POST "http://$ENDPOINT/v1/messages" \
+    -H 'Content-Type: application/json' \
+    -d "{\"model\":\"$MODEL\",\"messages\":[{\"role\":\"user\",\"content\":\"2+2=?\"}],\"stream\":false}" 2>/dev/null || echo "ERROR")
+assert_contains "$RESPONSE" "4" "回答包含正确答案"
+assert_contains "$RESPONSE" "end_turn" "stop_reason 为 end_turn"
+
+# 汇总
+echo ""
+echo "========================================"
+echo "测试结果: $PASS 通过, $FAIL 失败"
+echo "========================================"
+
+if [ "$FAIL" -gt 0 ]; then
+    exit 1
+fi