refactor(agent): replace XML api_call with native function calling

Migrate the Telegram bot agent from an XML tag hack (<api_call>) to OpenAI-native function calling via CallWithRequestFull. Key changes: - mcp/interface.go: add parseMCPResponseFull to clientHooks interface - mcp/client.go: route callWithRequestFull through hooks for overridability - mcp/claude_client.go: override parseMCPResponseFull for Claude response format (tool_use blocks instead of choices[].message.tool_calls) - telegram/agent/agent.go: rewrite Run() to use CallWithRequestFull; define api_request tool with JSON Schema; implement tool-call loop with role="tool" result messages; remove XML parsing entirely - telegram/agent/apicall.go: remove parseAPICall (dead code) - telegram/agent/prompt.go: simplify — remove XML format instructions, replace with concise api_request tool usage instructions - telegram/agent/agent_test.go: rebuild all tests using LLMResponse objects; add TestNarrationStructurallyImpossible, TestOnChunkCalledWithFinalReply, TestToolCallIDPropagated; remove XML-specific tests Architecture advantage: with native function calling, the LLM returns EITHER ToolCalls OR Content — never both. Narration is now structurally impossible at the protocol level, not just enforced by prompt rules. All 11 agent tests pass. mcp package tests pass.
2026-07-10 06:20:58 +08:00 · 2026-03-08 17:10:07 +08:00
parent 5f47dd13db
commit 9fcf44af65
8 changed files with 472 additions and 320 deletions
--- a/mcp/claude_client.go
+++ b/mcp/claude_client.go
@@ -92,6 +92,68 @@ func (c *ClaudeClient) buildMCPRequestBody(systemPrompt, userPrompt string) map[
 	return requestBody
 }

+// parseMCPResponseFull Claude response format — handles both text and tool_use blocks.
+func (c *ClaudeClient) parseMCPResponseFull(body []byte) (*LLMResponse, error) {
+	var response struct {
+		Content []struct {
+			Type  string          `json:"type"`
+			Text  string          `json:"text,omitempty"`
+			ID    string          `json:"id,omitempty"`
+			Name  string          `json:"name,omitempty"`
+			Input json.RawMessage `json:"input,omitempty"`
+		} `json:"content"`
+		Usage struct {
+			InputTokens  int `json:"input_tokens"`
+			OutputTokens int `json:"output_tokens"`
+		} `json:"usage"`
+		Error *struct {
+			Type    string `json:"type"`
+			Message string `json:"message"`
+		} `json:"error"`
+	}
+
+	if err := json.Unmarshal(body, &response); err != nil {
+		return nil, fmt.Errorf("failed to parse Claude response: %w, body: %s", err, string(body))
+	}
+	if response.Error != nil {
+		return nil, fmt.Errorf("Claude API error: %s - %s", response.Error.Type, response.Error.Message)
+	}
+
+	totalTokens := response.Usage.InputTokens + response.Usage.OutputTokens
+	if TokenUsageCallback != nil && totalTokens > 0 {
+		TokenUsageCallback(TokenUsage{
+			Provider:         c.Provider,
+			Model:            c.Model,
+			PromptTokens:     response.Usage.InputTokens,
+			CompletionTokens: response.Usage.OutputTokens,
+			TotalTokens:      totalTokens,
+		})
+	}
+
+	result := &LLMResponse{}
+	for _, block := range response.Content {
+		switch block.Type {
+		case "text":
+			result.Content = block.Text
+		case "tool_use":
+			// Claude returns arguments as a JSON object (Input field); convert to string.
+			argsJSON, err := json.Marshal(block.Input)
+			if err != nil {
+				argsJSON = []byte("{}")
+			}
+			result.ToolCalls = append(result.ToolCalls, ToolCall{
+				ID:   block.ID,
+				Type: "function",
+				Function: ToolCallFunction{
+					Name:      block.Name,
+					Arguments: string(argsJSON),
+				},
+			})
+		}
+	}
+	return result, nil
+}
+
 // parseMCPResponse Claude has different response format
 func (c *ClaudeClient) parseMCPResponse(body []byte) (string, error) {
 	var response struct {
--- a/mcp/client.go
+++ b/mcp/client.go
@@ -234,10 +234,21 @@ func (client *Client) marshalRequestBody(requestBody map[string]any) ([]byte, er
 }

 func (client *Client) parseMCPResponse(body []byte) (string, error) {
+	r, err := client.parseMCPResponseFull(body)
+	if err != nil {
+		return "", err
+	}
+	return r.Content, nil
+}
+
+// parseMCPResponseFull parses the OpenAI-format response body and returns both
+// the text content and any tool calls.
+func (client *Client) parseMCPResponseFull(body []byte) (*LLMResponse, error) {
 	var result struct {
 		Choices []struct {
 			Message struct {
-				Content string `json:"content"`
+				Content   string     `json:"content"`
+				ToolCalls []ToolCall `json:"tool_calls"`
 			} `json:"message"`
 		} `json:"choices"`
 		Usage struct {
@@ -248,11 +259,11 @@ func (client *Client) parseMCPResponse(body []byte) (string, error) {
 	}

 	if err := json.Unmarshal(body, &result); err != nil {
-		return "", fmt.Errorf("failed to parse response: %w", err)
+		return nil, fmt.Errorf("failed to parse response: %w", err)
 	}

 	if len(result.Choices) == 0 {
-		return "", fmt.Errorf("API returned empty response")
+		return nil, fmt.Errorf("API returned empty response")
 	}

 	// Report token usage if callback is set
@@ -266,7 +277,11 @@ func (client *Client) parseMCPResponse(body []byte) (string, error) {
 		})
 	}

-	return result.Choices[0].Message.Content, nil
+	msg := result.Choices[0].Message
+	return &LLMResponse{
+		Content:   msg.Content,
+		ToolCalls: msg.ToolCalls,
+	}, nil
 }

 func (client *Client) buildUrl() string {
@@ -427,6 +442,70 @@ func (client *Client) CallWithRequest(req *Request) (string, error) {
 	return "", fmt.Errorf("still failed after %d retries: %w", maxRetries, lastErr)
 }

+// CallWithRequestFull calls the AI API and returns both text content and tool calls.
+func (client *Client) CallWithRequestFull(req *Request) (*LLMResponse, error) {
+	if client.APIKey == "" {
+		return nil, fmt.Errorf("AI API key not set, please call SetAPIKey first")
+	}
+	if req.Model == "" {
+		req.Model = client.Model
+	}
+
+	var lastErr error
+	maxRetries := client.config.MaxRetries
+	for attempt := 1; attempt <= maxRetries; attempt++ {
+		if attempt > 1 {
+			client.logger.Warnf("⚠️  AI API call failed, retrying (%d/%d)...", attempt, maxRetries)
+		}
+		result, err := client.callWithRequestFull(req)
+		if err == nil {
+			return result, nil
+		}
+		lastErr = err
+		if !client.hooks.isRetryableError(err) {
+			return nil, err
+		}
+		if attempt < maxRetries {
+			waitTime := client.config.RetryWaitBase * time.Duration(attempt)
+			time.Sleep(waitTime)
+		}
+	}
+	return nil, fmt.Errorf("still failed after %d retries: %w", maxRetries, lastErr)
+}
+
+// callWithRequestFull single call that returns LLMResponse (content + tool calls).
+func (client *Client) callWithRequestFull(req *Request) (*LLMResponse, error) {
+	client.logger.Infof("📡 [%s] Request AI Server (full): BaseURL: %s", client.String(), client.BaseURL)
+
+	requestBody := client.buildRequestBodyFromRequest(req)
+	jsonData, err := client.hooks.marshalRequestBody(requestBody)
+	if err != nil {
+		return nil, err
+	}
+
+	url := client.hooks.buildUrl()
+	httpReq, err := client.hooks.buildRequest(url, jsonData)
+	if err != nil {
+		return nil, fmt.Errorf("failed to create request: %w", err)
+	}
+
+	resp, err := client.httpClient.Do(httpReq)
+	if err != nil {
+		return nil, fmt.Errorf("failed to send request: %w", err)
+	}
+	defer resp.Body.Close()
+
+	body, err := io.ReadAll(resp.Body)
+	if err != nil {
+		return nil, fmt.Errorf("failed to read response: %w", err)
+	}
+	if resp.StatusCode != http.StatusOK {
+		return nil, fmt.Errorf("API returned error (status %d): %s", resp.StatusCode, string(body))
+	}
+
+	return client.hooks.parseMCPResponseFull(body)
+}
+
 // callWithRequest single AI API call (using Request object)
 func (client *Client) callWithRequest(req *Request) (string, error) {
 	// Print current AI configuration
@@ -481,13 +560,23 @@ func (client *Client) callWithRequest(req *Request) (string, error) {

 // buildRequestBodyFromRequest builds request body from Request object
 func (client *Client) buildRequestBodyFromRequest(req *Request) map[string]any {
-	// Convert Message to API format
-	messages := make([]map[string]string, 0, len(req.Messages))
+	// Convert Message to API format — must use map[string]any to support
+	// tool-call messages (tool_calls, tool_call_id fields).
+	messages := make([]map[string]any, 0, len(req.Messages))
 	for _, msg := range req.Messages {
-		messages = append(messages, map[string]string{
-			"role":    msg.Role,
-			"content": msg.Content,
-		})
+		m := map[string]any{"role": msg.Role}
+		if len(msg.ToolCalls) > 0 {
+			// Assistant message that contains tool invocations.
+			// content must be null/omitted for OpenAI compatibility.
+			m["tool_calls"] = msg.ToolCalls
+		} else if msg.ToolCallID != "" {
+			// Tool result message (role="tool").
+			m["tool_call_id"] = msg.ToolCallID
+			m["content"] = msg.Content
+		} else {
+			m["content"] = msg.Content
+		}
+		messages = append(messages, m)
 	}

 	// Build basic request body
--- a/mcp/interface.go
+++ b/mcp/interface.go
@@ -15,6 +15,11 @@ type AIClient interface {
 	// onChunk is called with the full accumulated text so far (not raw deltas).
 	// Returns the complete final text when done.
 	CallWithRequestStream(req *Request, onChunk func(string)) (string, error)
+	// CallWithRequestFull returns both text content and tool calls.
+	// Use this when the request includes Tools — the LLM may respond with
+	// either a plain text reply (LLMResponse.Content) or tool invocations
+	// (LLMResponse.ToolCalls), but not both.
+	CallWithRequestFull(req *Request) (*LLMResponse, error)
 }

 // clientHooks internal hook interface (for subclass to override specific steps)
@@ -30,5 +35,6 @@ type clientHooks interface {
 	setAuthHeader(reqHeaders http.Header)
 	marshalRequestBody(requestBody map[string]any) ([]byte, error)
 	parseMCPResponse(body []byte) (string, error)
+	parseMCPResponseFull(body []byte) (*LLMResponse, error)
 	isRetryableError(err error) bool
 }
--- a/mcp/request.go
+++ b/mcp/request.go
@@ -1,9 +1,34 @@
 package mcp

-// Message represents a conversation message
+// Message represents a conversation message.
+// Supports plain messages (Role+Content), assistant tool-call messages (ToolCalls),
+// and tool result messages (Role="tool", ToolCallID, Content).
 type Message struct {
-	Role    string `json:"role"`    // "system", "user", "assistant"
-	Content string `json:"content"` // Message content
+	Role       string     `json:"role"`                  // "system", "user", "assistant", "tool"
+	Content    string     `json:"content,omitempty"`     // Text content (omitted when ToolCalls present)
+	ToolCalls  []ToolCall `json:"tool_calls,omitempty"`  // Set by assistant when calling tools
+	ToolCallID string     `json:"tool_call_id,omitempty"` // Set on role="tool" result messages
+}
+
+// ToolCall is a single function call requested by the LLM.
+type ToolCall struct {
+	ID       string           `json:"id"`       // Unique call ID (e.g. "call_abc123")
+	Type     string           `json:"type"`     // Always "function"
+	Function ToolCallFunction `json:"function"` // Function name and JSON-serialised arguments
+}
+
+// ToolCallFunction holds the function name and raw JSON arguments string.
+type ToolCallFunction struct {
+	Name      string `json:"name"`      // Function name
+	Arguments string `json:"arguments"` // JSON-encoded argument object
+}
+
+// LLMResponse is returned by CallWithRequestFull and carries both the assistant
+// text reply (Content) and any structured tool calls (ToolCalls).
+// Exactly one of the two fields will be non-empty for a well-formed response.
+type LLMResponse struct {
+	Content   string     // Plain-text reply (final answer)
+	ToolCalls []ToolCall // Structured tool invocations
 }

 // Tool represents a tool/function that AI can call
--- a/telegram/agent/agent.go
+++ b/telegram/agent/agent.go
@@ -12,8 +12,39 @@ import (

 const maxIterations = 10

+// apiRequestTool is the single tool exposed to the LLM.
+// Native function calling means the LLM returns EITHER ToolCalls OR Content — never both.
+// This makes narration structurally impossible: text cannot appear alongside a tool call.
+var apiRequestTool = mcp.Tool{
+	Type: "function",
+	Function: mcp.FunctionDef{
+		Name:        "api_request",
+		Description: "Call the NOFX trading system REST API",
+		Parameters: map[string]any{
+			"type": "object",
+			"properties": map[string]any{
+				"method": map[string]any{
+					"type":        "string",
+					"enum":        []string{"GET", "POST", "PUT", "DELETE"},
+					"description": "HTTP method",
+				},
+				"path": map[string]any{
+					"type":        "string",
+					"description": "API path; include query params in path: /api/positions?trader_id=xxx",
+				},
+				"body": map[string]any{
+					"type":        "object",
+					"description": "Request body; use {} for GET requests",
+				},
+			},
+			"required": []string{"method", "path", "body"},
+		},
+	},
+}
+
 // Agent is a stateful AI agent for one Telegram chat.
-// It has a single tool (api_call) and an unbounded decision loop.
+// It exposes a single "api_request" tool and runs a loop until the LLM
+// returns a plain-text reply (no tool calls).
 type Agent struct {
 	apiTool      *apiCallTool
 	getLLM       func() mcp.AIClient
@@ -34,17 +65,15 @@ func New(apiPort int, botToken, userID string, getLLM func() mcp.AIClient, syste
 }

 // GenerateBotToken creates a long-lived JWT for the bot's internal API calls.
-// userID must match the actual registered user's ID so that bot-made changes
-// are visible in the frontend (they share the same user namespace).
+// userID must match the actual registered user's ID so bot-made changes
+// are visible in the frontend (shared user namespace).
 func GenerateBotToken(userID string) (string, error) {
 	return auth.GenerateJWT(userID, "bot@internal")
 }

 // buildAccountContext fetches the live account state (models, exchanges, strategies, traders,
-// and per-trader account summary + statistics) via the local API and returns it as a formatted
-// string for injection into the LLM context. This gives the LLM immediate awareness of what
-// is already configured and the current financial state, so it never asks the user for
-// information that already exists.
+// and per-trader account summary + statistics) and returns it as a formatted string for
+// injection into the LLM context at the start of each conversation.
 func (a *Agent) buildAccountContext() string {
 	type q struct {
 		label string
@@ -91,137 +120,104 @@ func (a *Agent) buildAccountContext() string {
 	return sb.String()
 }

-// Run processes one user message through the agent loop.
-// Loop: LLM decides -> if <api_call>: execute, append result, loop -> if no tag: return reply.
+// Run processes one user message through the native function-calling agent loop.
 //
-// On the first message of a conversation, the current account state (models, exchanges,
-// strategies, traders) is automatically fetched and injected so the LLM knows what is
-// already configured without asking the user to repeat themselves.
+// Architecture:
+//   - LLM receives the api_request tool definition alongside conversation history.
+//   - LLM response is EITHER ToolCalls (execute API) OR Content (final reply) — never both.
+//     This is enforced by the protocol: narration is structurally impossible.
+//   - Loop continues until the LLM returns a plain-text reply (no tool calls).
 //
-// onChunk is optional. When non-nil, each LLM call is streamed:
-//   - Chunks are forwarded to onChunk until an <api_call> tag appears in the accumulated text.
-//   - After an api_call iteration completes, onChunk("⏳") resets the display to a thinking indicator.
-//   - The final reply is streamed progressively via onChunk.
+// On the first message of a conversation the live account state is fetched and injected.
+// onChunk is optional; when set it is called once with the complete final reply text.
 func (a *Agent) Run(userMessage string, onChunk func(string)) string {
 	llm := a.getLLM()
 	if llm == nil {
 		return "AI assistant unavailable. Please configure an AI model in the Web UI."
 	}

-	// Build turn messages: history context prefix + current user message.
-	// On the very first message (no history), prepend a live account state snapshot so the
-	// LLM immediately knows what models, exchanges, strategies, and traders are configured.
+	// Build initial user message: prepend account state on first turn, history on subsequent turns.
 	histCtx := a.memory.BuildContext()
-	var firstMsg string
+	var firstUserContent string
 	if histCtx == "" {
-		// First message in this conversation — fetch and inject account state.
 		accountCtx := a.buildAccountContext()
-		firstMsg = accountCtx + "\n[User Message]\n" + userMessage
+		firstUserContent = accountCtx + "\n[User Message]\n" + userMessage
 	} else {
-		firstMsg = histCtx + "\n---\nUser: " + userMessage
+		firstUserContent = histCtx + "\n---\nUser: " + userMessage
 	}
-	turnMsgs := []mcp.Message{mcp.NewUserMessage(firstMsg)}

-	var lastResp string
+	turnMsgs := []mcp.Message{mcp.NewUserMessage(firstUserContent)}

 	for i := 0; i < maxIterations; i++ {
 		req, err := mcp.NewRequestBuilder().
 			WithSystemPrompt(a.systemPrompt).
 			AddConversationHistory(turnMsgs).
+			AddTool(apiRequestTool).
+			WithToolChoice("auto").
 			Build()
 		if err != nil {
 			logger.Errorf("Agent: failed to build request: %v", err)
 			break
 		}

-		var resp string
-		if onChunk != nil {
-			// Stream this call; suppress chunks once an <api_call> tag appears.
-			// Also hold back the last (len("<api_call>")-1) chars of accumulated text to
-			// avoid showing partial opening tags (e.g. "<", "<ap") before we can detect them.
-			const tagLen = len("<api_call>") // 10
-			const safeOffset = tagLen - 1    // 9: max prefix of tag we might have received
-
-			var apiTagSeen bool
-			resp, err = llm.CallWithRequestStream(req, func(accumulated string) {
-				if apiTagSeen {
-					return
-				}
-				if idx := strings.Index(accumulated, "<api_call>"); idx >= 0 {
-					apiTagSeen = true
-					// Forward only the text that appeared before the tag.
-					if display := strings.TrimSpace(accumulated[:idx]); display != "" {
-						onChunk(display)
-					}
-					return
-				}
-				// Forward only the "safe" prefix — hold back the last safeOffset chars
-				// in case they are the beginning of an <api_call> tag.
-				if safe := len(accumulated) - safeOffset; safe > 0 {
-					onChunk(accumulated[:safe])
-				}
-			})
-		} else {
-			resp, err = llm.CallWithRequest(req)
-		}
+		resp, err := llm.CallWithRequestFull(req)
 		if err != nil {
 			logger.Errorf("Agent: LLM call failed (iteration %d): %v", i+1, err)
 			return "AI assistant temporarily unavailable. Please try again."
 		}
-		lastResp = resp

-		apiReq, textBefore := parseAPICall(resp)
-		if apiReq == nil {
-			// No api_call tag — LLM gave a final answer (already streamed if onChunk set).
-			reply := stripAPICallTag(strings.TrimSpace(resp))
+		// No tool calls → LLM returned a final text reply.
+		if len(resp.ToolCalls) == 0 {
+			reply := strings.TrimSpace(resp.Content)
+			if onChunk != nil {
+				onChunk(reply)
+			}
 			a.memory.Add("user", userMessage)
 			a.memory.Add("assistant", reply)
 			return reply
 		}

-		// api_call iteration — reset display to thinking indicator before executing.
+		// Tool call iteration — show thinking indicator.
 		if onChunk != nil {
 			onChunk("⏳")
 		}

-		logger.Infof("Agent: iter=%d %s %s", i+1, apiReq.Method, apiReq.Path)
-		result := a.apiTool.execute(apiReq)
+		// Append assistant message carrying the tool calls (no content field).
+		turnMsgs = append(turnMsgs, mcp.Message{
+			Role:      "assistant",
+			ToolCalls: resp.ToolCalls,
+		})

-		if textBefore != "" {
-			turnMsgs = append(turnMsgs, mcp.NewAssistantMessage(textBefore))
-		}
-		turnMsgs = append(turnMsgs, mcp.NewUserMessage(
-			fmt.Sprintf("[API result: %s %s]\n%s", apiReq.Method, apiReq.Path, result),
-		))
-	}
-
-	// Safety: max iterations reached — ask LLM for a final summary (non-streaming).
-	logger.Warnf("Agent: max iterations (%d) reached", maxIterations)
-	turnMsgs = append(turnMsgs, mcp.NewUserMessage("Please summarize the results and give the user a final reply."))
-	if finalReq, err := mcp.NewRequestBuilder().
-		WithSystemPrompt(a.systemPrompt).
-		AddConversationHistory(turnMsgs).
-		Build(); err == nil {
-		if finalResp, err := llm.CallWithRequest(finalReq); err == nil {
-			lastResp = finalResp
+		// Execute each tool call and append the results as tool messages.
+		for _, tc := range resp.ToolCalls {
+			var apiReq apiRequest
+			if err := json.Unmarshal([]byte(tc.Function.Arguments), &apiReq); err != nil {
+				logger.Errorf("Agent: invalid tool args for call %s: %v", tc.ID, err)
+				turnMsgs = append(turnMsgs, mcp.Message{
+					Role:       "tool",
+					ToolCallID: tc.ID,
+					Content:    fmt.Sprintf(`{"error":"invalid arguments: %s"}`, err.Error()),
+				})
+				continue
+			}
+			logger.Infof("Agent: iter=%d tool=%s %s %s", i+1, tc.ID, apiReq.Method, apiReq.Path)
+			result := a.apiTool.execute(&apiReq)
+			turnMsgs = append(turnMsgs, mcp.Message{
+				Role:       "tool",
+				ToolCallID: tc.ID,
+				Content:    result,
+			})
 		}
 	}

-	reply := stripAPICallTag(strings.TrimSpace(lastResp))
+	// Safety: max iterations reached.
+	logger.Warnf("Agent: max iterations (%d) reached for message: %q", maxIterations, userMessage)
+	reply := "操作已完成，请检查您的账户查看最新状态。"
 	a.memory.Add("user", userMessage)
 	a.memory.Add("assistant", reply)
 	return reply
 }

-// stripAPICallTag removes any <api_call>...</api_call> fragment from s.
-// Used as a defensive layer to ensure tags never leak to the user.
-func stripAPICallTag(s string) string {
-	if idx := strings.Index(s, "<api_call>"); idx >= 0 {
-		return strings.TrimSpace(s[:idx])
-	}
-	return s
-}
-
 // ResetMemory clears conversation history (called on /start).
 func (a *Agent) ResetMemory() {
 	a.memory.ResetFull()
--- a/telegram/agent/agent_test.go
+++ b/telegram/agent/agent_test.go
@@ -11,34 +11,73 @@ import (
 	"nofx/mcp"
 )

+// mockLLM implements mcp.AIClient using pre-programmed LLMResponse objects.
+// Native function calling: CallWithRequestFull is the primary method;
+// CallWithRequest and CallWithRequestStream are stubs kept for interface compliance.
 type mockLLM struct {
-	responses []string
+	responses []*mcp.LLMResponse
 	calls     int
 	lastMsgs  []mcp.Message
 }

-func (m *mockLLM) SetAPIKey(_, _, _ string)                    {}
-func (m *mockLLM) SetTimeout(_ time.Duration)                  {}
-func (m *mockLLM) CallWithMessages(_, _ string) (string, error) { return m.next() }
+func (m *mockLLM) SetAPIKey(_, _, _ string)   {}
+func (m *mockLLM) SetTimeout(_ time.Duration) {}
+
+func (m *mockLLM) CallWithMessages(_, _ string) (string, error) { return "", nil }
+
 func (m *mockLLM) CallWithRequest(req *mcp.Request) (string, error) {
+	r, err := m.next()
+	if err != nil {
+		return "", err
+	}
+	return r.Content, nil
+}
+
+func (m *mockLLM) CallWithRequestStream(req *mcp.Request, onChunk func(string)) (string, error) {
+	r, err := m.next()
+	if err != nil {
+		return "", err
+	}
+	if onChunk != nil {
+		onChunk(r.Content)
+	}
+	return r.Content, nil
+}
+
+func (m *mockLLM) CallWithRequestFull(req *mcp.Request) (*mcp.LLMResponse, error) {
 	m.lastMsgs = req.Messages
 	return m.next()
 }
-func (m *mockLLM) CallWithRequestStream(req *mcp.Request, onChunk func(string)) (string, error) {
-	m.lastMsgs = req.Messages
-	r, err := m.next()
-	if onChunk != nil {
-		onChunk(r)
-	}
-	return r, err
-}
-func (m *mockLLM) next() (string, error) {
+
+func (m *mockLLM) next() (*mcp.LLMResponse, error) {
 	if m.calls < len(m.responses) {
 		r := m.responses[m.calls]
 		m.calls++
 		return r, nil
 	}
-	return "OK", nil
+	return &mcp.LLMResponse{Content: "OK"}, nil
+}
+
+// toolCall builds a mock LLM response that contains a single tool invocation.
+func toolCall(id, method, path string, body string) *mcp.LLMResponse {
+	if body == "" {
+		body = "{}"
+	}
+	return &mcp.LLMResponse{
+		ToolCalls: []mcp.ToolCall{{
+			ID:   id,
+			Type: "function",
+			Function: mcp.ToolCallFunction{
+				Name:      "api_request",
+				Arguments: fmt.Sprintf(`{"method":%q,"path":%q,"body":%s}`, method, path, body),
+			},
+		}},
+	}
+}
+
+// textReply builds a mock LLM response with a plain-text final answer.
+func textReply(content string) *mcp.LLMResponse {
+	return &mcp.LLMResponse{Content: content}
 }

 func mockGetLLM(llm *mockLLM) func() mcp.AIClient {
@@ -70,9 +109,9 @@ func mockAPIServer(handlers map[string]string) (*httptest.Server, int) {

 // ── Basic agent behaviour ──────────────────────────────────────────────────

-// TestAgentDirectReply: LLM replies without api_call — one call, direct reply.
+// TestAgentDirectReply: LLM replies with text (no tool calls) — one LLM call.
 func TestAgentDirectReply(t *testing.T) {
-	llm := &mockLLM{responses: []string{"Hello! How can I help you?"}}
+	llm := &mockLLM{responses: []*mcp.LLMResponse{textReply("Hello! How can I help you?")}}
 	a := New(8080, "tok", "test-user", mockGetLLM(llm), testPrompt)

 	reply := a.Run("hello", nil)
@@ -85,16 +124,16 @@ func TestAgentDirectReply(t *testing.T) {
 	}
 }

-// TestAgentAPICall: LLM calls API, gets result, gives final reply — two LLM calls.
+// TestAgentAPICall: LLM makes one tool call, gets result, gives final reply — two LLM calls.
 func TestAgentAPICall(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
 		"/api/my-traders": `[{"trader_id":"t1","trader_name":"BTC Trader","is_running":false}]`,
 	})
 	defer srv.Close()

-	llm := &mockLLM{responses: []string{
-		`<api_call>{"method":"GET","path":"/api/my-traders","body":{}}</api_call>`,
-		"You have one trader: BTC Trader.",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "GET", "/api/my-traders", "{}"),
+		textReply("You have one trader: BTC Trader."),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)

@@ -108,7 +147,7 @@ func TestAgentAPICall(t *testing.T) {
 	}
 }

-// TestAgentMultiStep: LLM chains two API calls before final reply — three LLM calls.
+// TestAgentMultiStep: LLM chains two tool calls before final reply — three LLM calls.
 func TestAgentMultiStep(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
 		"/api/account":   `{"total_equity":1000}`,
@@ -116,105 +155,118 @@ func TestAgentMultiStep(t *testing.T) {
 	})
 	defer srv.Close()

-	llm := &mockLLM{responses: []string{
-		`<api_call>{"method":"GET","path":"/api/account","body":{}}</api_call>`,
-		`<api_call>{"method":"GET","path":"/api/positions","body":{}}</api_call>`,
-		"Account looks healthy and no open positions.",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "GET", "/api/account", "{}"),
+		toolCall("c2", "GET", "/api/positions", "{}"),
+		textReply("Account looks healthy and no open positions."),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)

 	reply := a.Run("show me account status", nil)

 	if llm.calls != 3 {
-		t.Fatalf("expected 3 LLM calls (2 api + 1 final), got %d", llm.calls)
+		t.Fatalf("expected 3 LLM calls (2 tool + 1 final), got %d", llm.calls)
 	}
 	if reply != "Account looks healthy and no open positions." {
 		t.Fatalf("unexpected final reply: %q", reply)
 	}
 }

-// TestAgentAPIResultInContext: API result must appear in next LLM message.
+// TestAgentAPIResultInContext: tool result must appear as a tool message in the next LLM call.
 func TestAgentAPIResultInContext(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
 		"/api/account": `{"balance":1234.56}`,
 	})
 	defer srv.Close()

-	llm := &mockLLM{responses: []string{
-		`<api_call>{"method":"GET","path":"/api/account","body":{}}</api_call>`,
-		"Balance is 1234.56 USDT.",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "GET", "/api/account", "{}"),
+		textReply("Balance is 1234.56 USDT."),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
 	a.Run("show balance", nil)

+	// The last request must contain a tool-result message with the balance data.
 	found := false
 	for _, msg := range llm.lastMsgs {
-		if strings.Contains(msg.Content, "API result") || strings.Contains(msg.Content, "balance") {
+		if msg.Role == "tool" && strings.Contains(msg.Content, "balance") {
 			found = true
 			break
 		}
 	}
 	if !found {
-		t.Fatalf("API result not found in subsequent LLM context")
+		t.Fatalf("tool result message not found in subsequent LLM context; messages: %+v", llm.lastMsgs)
 	}
 }

-// ── NO NARRATION tests ─────────────────────────────────────────────────────
+// ── Narration-free architecture tests ─────────────────────────────────────

-// TestNoNarrationBeforeAPICall: any text before <api_call> must NOT reach the user.
-// The agent strips text-before-tag and only forwards it as assistant context.
-func TestNoNarrationBeforeAPICall(t *testing.T) {
+// TestNarrationStructurallyImpossible: when ToolCalls are present in the response,
+// any Content field is ignored and never surfaced to the user.
+// In real LLM APIs, Content is always empty alongside ToolCalls, but we verify
+// our agent handles a malformed response defensively.
+func TestNarrationStructurallyImpossible(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
 		"/api/strategies": `[{"id":"s1","name":"BTC Trend"}]`,
 	})
 	defer srv.Close()

-	narrations := []string{
-		"现在我将为您创建策略。\n",
-		"好的，我来帮你查询。",
-		"Let me check this for you. ",
-		"正在处理...",
-		"I will call the API now. ",
+	// Simulate a (malformed) response that has both Content and ToolCalls.
+	malformed := &mcp.LLMResponse{
+		Content: "现在我将为您查询策略。", // narration — must NOT reach user
+		ToolCalls: []mcp.ToolCall{{
+			ID:   "c1",
+			Type: "function",
+			Function: mcp.ToolCallFunction{
+				Name:      "api_request",
+				Arguments: `{"method":"GET","path":"/api/strategies","body":{}}`,
+			},
+		}},
 	}

-	for _, narration := range narrations {
-		llm := &mockLLM{responses: []string{
-			// LLM outputs narration before the api_call tag (bad behaviour we must handle)
-			narration + `<api_call>{"method":"GET","path":"/api/strategies","body":{}}</api_call>`,
-			"你有1个策略：BTC Trend。",
-		}}
-		a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
-		reply := a.Run("查询我的策略", nil)
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		malformed,
+		textReply("你有1个策略：BTC Trend。"),
+	}}
+	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
+	reply := a.Run("查询我的策略", nil)

-		// Final reply must not contain narration fragments
-		if strings.Contains(reply, "现在我将") || strings.Contains(reply, "Let me") ||
-			strings.Contains(reply, "正在处理") || strings.Contains(reply, "好的，我来") ||
-			strings.Contains(reply, "I will call") {
-			t.Fatalf("narration leaked into reply for input %q: got %q", narration, reply)
-		}
-		// api_call tag must not appear in reply
-		if strings.Contains(reply, "<api_call>") {
-			t.Fatalf("api_call tag leaked into reply: %q", reply)
-		}
+	if strings.Contains(reply, "现在我将") {
+		t.Fatalf("narration leaked into final reply: %q", reply)
+	}
+	if reply != "你有1个策略：BTC Trend。" {
+		t.Fatalf("unexpected reply: %q", reply)
 	}
 }

-// TestAPICallTagNotLeakedToUser: <api_call> tag must never appear in returned reply.
-func TestAPICallTagNotLeakedToUser(t *testing.T) {
+// TestOnChunkCalledWithFinalReply: onChunk receives the complete final reply.
+func TestOnChunkCalledWithFinalReply(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
-		"/api/account": `{"total_equity":500}`,
+		"/api/account": `{"equity":500}`,
 	})
 	defer srv.Close()

-	llm := &mockLLM{responses: []string{
-		`<api_call>{"method":"GET","path":"/api/account","body":{}}</api_call>`,
-		`账户余额 500 USDT。<api_call>{"method":"GET","path":"/api/account","body":{}}</api_call>`,
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "GET", "/api/account", "{}"),
+		textReply("Equity: 500 USDT."),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
-	reply := a.Run("show balance", nil)

-	if strings.Contains(reply, "<api_call>") {
-		t.Fatalf("api_call tag leaked to user: %q", reply)
+	var chunks []string
+	reply := a.Run("show equity", func(chunk string) {
+		chunks = append(chunks, chunk)
+	})
+
+	if reply != "Equity: 500 USDT." {
+		t.Fatalf("unexpected reply: %q", reply)
+	}
+	// Should have received ⏳ for the tool call, then the final reply.
+	if len(chunks) < 2 {
+		t.Fatalf("expected at least 2 chunks (⏳ + final), got: %v", chunks)
+	}
+	lastChunk := chunks[len(chunks)-1]
+	if lastChunk != "Equity: 500 USDT." {
+		t.Fatalf("last chunk should be final reply, got: %q", lastChunk)
 	}
 }

@@ -224,18 +276,15 @@ func TestAPICallTagNotLeakedToUser(t *testing.T) {
 // Verifies: POST strategy → GET verify → final reply shows strategy info.
 func TestCreateStrategyWorkflow(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
-		"POST /api/strategies":    `{"id":"s1","name":"BTC趋势"}`,
-		"GET /api/strategies/s1":  `{"id":"s1","name":"BTC趋势","config":{"coin_source":{"source_type":"static","static_coins":["BTC/USDT"]},"leverage":5}}`,
+		"POST /api/strategies":   `{"id":"s1","name":"BTC趋势"}`,
+		"GET /api/strategies/s1": `{"id":"s1","name":"BTC趋势","config":{"coin_source":{"source_type":"static","static_coins":["BTC/USDT"]},"leverage":5}}`,
 	})
 	defer srv.Close()

-	llm := &mockLLM{responses: []string{
-		// Step 1: create strategy
-		`<api_call>{"method":"POST","path":"/api/strategies","body":{"name":"BTC趋势","config":{}}}</api_call>`,
-		// Step 2: verify strategy
-		`<api_call>{"method":"GET","path":"/api/strategies/s1","body":{}}</api_call>`,
-		// Step 3: final reply
-		"策略已创建：BTC趋势，币种 BTC/USDT，杠杆 5x。",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "POST", "/api/strategies", `{"name":"BTC趋势","config":{}}`),
+		toolCall("c2", "GET", "/api/strategies/s1", "{}"),
+		textReply("策略已创建：BTC趋势，币种 BTC/USDT，杠杆 5x。"),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
 	reply := a.Run("帮我配置个btc趋势交易的策略", nil)
@@ -243,12 +292,12 @@ func TestCreateStrategyWorkflow(t *testing.T) {
 	if llm.calls != 3 {
 		t.Fatalf("expected 3 LLM calls, got %d", llm.calls)
 	}
-	if reply == "" || strings.Contains(reply, "<api_call>") {
-		t.Fatalf("bad final reply: %q", reply)
+	if reply == "" {
+		t.Fatalf("empty final reply")
 	}
 }

-// TestFullSetupWorkflow: create strategy → create trader → start trader.
+// TestFullSetupWorkflow: create strategy → verify → create trader → start trader.
 // This is the "帮我配置策略并跑起来" workflow.
 func TestFullSetupWorkflow(t *testing.T) {
 	calls := map[string]int{}
@@ -272,17 +321,12 @@ func TestFullSetupWorkflow(t *testing.T) {
 	var port int
 	fmt.Sscanf(srv.Listener.Addr().String(), "127.0.0.1:%d", &port)

-	llm := &mockLLM{responses: []string{
-		// 1. create strategy
-		`<api_call>{"method":"POST","path":"/api/strategies","body":{"name":"BTC趋势"}}</api_call>`,
-		// 2. verify strategy
-		`<api_call>{"method":"GET","path":"/api/strategies/s1","body":{}}</api_call>`,
-		// 3. create trader
-		`<api_call>{"method":"POST","path":"/api/traders","body":{"name":"BTC趋势交易员","strategy_id":"s1"}}</api_call>`,
-		// 4. start trader
-		`<api_call>{"method":"POST","path":"/api/traders/tr1/start","body":{}}</api_call>`,
-		// 5. final reply
-		"策略和交易员已创建并启动！BTC趋势交易员正在运行。",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "POST", "/api/strategies", `{"name":"BTC趋势"}`),
+		toolCall("c2", "GET", "/api/strategies/s1", "{}"),
+		toolCall("c3", "POST", "/api/traders", `{"name":"BTC趋势交易员","strategy_id":"s1"}`),
+		toolCall("c4", "POST", "/api/traders/tr1/start", "{}"),
+		textReply("策略和交易员已创建并启动！BTC趋势交易员正在运行。"),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
 	reply := a.Run("帮我配置个btc趋势交易的策略交易 跑起来", nil)
@@ -290,7 +334,6 @@ func TestFullSetupWorkflow(t *testing.T) {
 	if llm.calls != 5 {
 		t.Fatalf("expected 5 LLM calls, got %d", llm.calls)
 	}
-	// Verify each API was called
 	if calls["POST /api/strategies"] != 1 {
 		t.Errorf("expected 1 POST /api/strategies, got %d", calls["POST /api/strategies"])
 	}
@@ -300,8 +343,8 @@ func TestFullSetupWorkflow(t *testing.T) {
 	if calls["POST /api/traders/tr1/start"] != 1 {
 		t.Errorf("expected 1 POST /api/traders/tr1/start, got %d", calls["POST /api/traders/tr1/start"])
 	}
-	if strings.Contains(reply, "<api_call>") {
-		t.Fatalf("api_call tag in final reply: %q", reply)
+	if reply == "" {
+		t.Fatalf("empty final reply")
 	}
 }

@@ -324,10 +367,10 @@ func TestStartExistingTrader(t *testing.T) {
 	var port int
 	fmt.Sscanf(srv.Listener.Addr().String(), "127.0.0.1:%d", &port)

-	llm := &mockLLM{responses: []string{
-		`<api_call>{"method":"GET","path":"/api/my-traders","body":{}}</api_call>`,
-		`<api_call>{"method":"POST","path":"/api/traders/tr1/start","body":{}}</api_call>`,
-		"交易员 BTC Trader 已启动。",
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("c1", "GET", "/api/my-traders", "{}"),
+		toolCall("c2", "POST", "/api/traders/tr1/start", "{}"),
+		textReply("交易员 BTC Trader 已启动。"),
 	}}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
 	reply := a.Run("启动交易员", nil)
@@ -335,96 +378,62 @@ func TestStartExistingTrader(t *testing.T) {
 	if calls["POST /api/traders/tr1/start"] != 1 {
 		t.Errorf("expected trader to be started, got %d start calls", calls["POST /api/traders/tr1/start"])
 	}
-	if strings.Contains(reply, "<api_call>") {
-		t.Fatalf("api_call tag in reply: %q", reply)
+	if reply != "交易员 BTC Trader 已启动。" {
+		t.Fatalf("unexpected reply: %q", reply)
 	}
 }

-// ── Parser tests ───────────────────────────────────────────────────────────
+// ── Safety limit ───────────────────────────────────────────────────────────

-// TestParseAPICall: unit tests for the XML tag parser.
-func TestParseAPICall(t *testing.T) {
-	t.Run("valid call no text before", func(t *testing.T) {
-		resp := `<api_call>{"method":"POST","path":"/api/traders/t1/stop","body":{}}</api_call>`
-		req, text := parseAPICall(resp)
-		if req == nil {
-			t.Fatal("expected api_call, got nil")
-		}
-		if req.Method != "POST" || req.Path != "/api/traders/t1/stop" {
-			t.Fatalf("unexpected req: %+v", req)
-		}
-		if text != "" {
-			t.Fatalf("expected empty text before tag, got: %q", text)
-		}
-	})
-
-	t.Run("text before tag is captured", func(t *testing.T) {
-		resp := `Stopping trader.<api_call>{"method":"POST","path":"/api/traders/t1/stop","body":{}}</api_call>`
-		req, text := parseAPICall(resp)
-		if req == nil {
-			t.Fatal("expected api_call, got nil")
-		}
-		if text != "Stopping trader." {
-			t.Fatalf("unexpected text before tag: %q", text)
-		}
-	})
-
-	t.Run("no call tag", func(t *testing.T) {
-		req, text := parseAPICall("Just a reply.")
-		if req != nil {
-			t.Fatal("expected nil api_call")
-		}
-		if text != "Just a reply." {
-			t.Fatalf("expected original text, got %q", text)
-		}
-	})
-
-	t.Run("malformed JSON", func(t *testing.T) {
-		req, _ := parseAPICall(`<api_call>NOT JSON</api_call>`)
-		if req != nil {
-			t.Fatal("expected nil for malformed JSON")
-		}
-	})
-}
-
-// TestStripAPICallTag: defensive cleanup of stray tags in final reply.
-func TestStripAPICallTag(t *testing.T) {
-	cases := []struct {
-		input string
-		want  string
-	}{
-		{`正常回复`, `正常回复`},
-		{`回复<api_call>{"method":"GET","path":"/x"}</api_call>`, `回复`},
-		{`<api_call>{"method":"GET","path":"/x"}</api_call>`, ``},
-	}
-	for _, c := range cases {
-		got := stripAPICallTag(c.input)
-		if strings.TrimSpace(got) != c.want {
-			t.Errorf("stripAPICallTag(%q) = %q, want %q", c.input, got, c.want)
-		}
-	}
-}
-
-// TestMaxIterations: agent stops after maxIterations and returns a summary.
+// TestMaxIterations: agent terminates after maxIterations and returns fallback message.
 func TestMaxIterations(t *testing.T) {
 	srv, port := mockAPIServer(map[string]string{
 		"/api/account": `{"ok":true}`,
 	})
 	defer srv.Close()

-	// Always returns another api_call — should hit max iterations
-	responses := make([]string, maxIterations+2)
+	// Always returns another tool call — should hit max iterations.
+	responses := make([]*mcp.LLMResponse, maxIterations+2)
 	for i := range responses {
-		responses[i] = `<api_call>{"method":"GET","path":"/api/account","body":{}}</api_call>`
+		responses[i] = toolCall(fmt.Sprintf("c%d", i), "GET", "/api/account", "{}")
 	}
-	responses[maxIterations] = "Final summary after max iterations."

 	llm := &mockLLM{responses: responses}
 	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
 	reply := a.Run("loop forever", nil)

-	if strings.Contains(reply, "<api_call>") {
-		t.Fatalf("api_call tag in reply after max iterations: %q", reply)
+	if reply == "" {
+		t.Fatalf("expected a fallback reply, got empty string")
+	}
+	// Agent should have made exactly maxIterations tool-call LLM calls.
+	if llm.calls != maxIterations {
+		t.Fatalf("expected %d LLM calls (max iterations), got %d", maxIterations, llm.calls)
+	}
+}
+
+// TestToolCallIDPropagated: tool result messages carry the correct ToolCallID.
+func TestToolCallIDPropagated(t *testing.T) {
+	srv, port := mockAPIServer(map[string]string{
+		"/api/account": `{"balance":999}`,
+	})
+	defer srv.Close()
+
+	llm := &mockLLM{responses: []*mcp.LLMResponse{
+		toolCall("call-xyz-123", "GET", "/api/account", "{}"),
+		textReply("Balance is 999."),
+	}}
+	a := New(port, "tok", "test-user", mockGetLLM(llm), testPrompt)
+	a.Run("check balance", nil)
+
+	// Find the tool result message and verify ToolCallID matches.
+	found := false
+	for _, msg := range llm.lastMsgs {
+		if msg.Role == "tool" && msg.ToolCallID == "call-xyz-123" {
+			found = true
+			break
+		}
+	}
+	if !found {
+		t.Fatalf("tool result with ToolCallID='call-xyz-123' not found in messages: %+v", llm.lastMsgs)
 	}
-	_ = reply // just confirm it terminates
 }
--- a/telegram/agent/apicall.go
+++ b/telegram/agent/apicall.go
@@ -19,7 +19,7 @@ type apiCallTool struct {
 	client  *http.Client
 }

-// apiRequest is the parsed structure from the LLM's <api_call> tag.
+// apiRequest holds the arguments decoded from the LLM's api_request tool call.
 type apiRequest struct {
 	Method string         `json:"method"`
 	Path   string         `json:"path"`
@@ -86,24 +86,3 @@ func (t *apiCallTool) execute(req *apiRequest) string {
 	return string(body)
 }

-// parseAPICall extracts <api_call>...</api_call> from LLM response.
-// Returns (nil, original) if not found or malformed JSON.
-func parseAPICall(resp string) (*apiRequest, string) {
-	const openTag = "<api_call>"
-	const closeTag = "</api_call>"
-
-	start := strings.Index(resp, openTag)
-	end := strings.Index(resp, closeTag)
-	if start < 0 || end < 0 || end <= start {
-		return nil, resp
-	}
-
-	jsonStr := strings.TrimSpace(resp[start+len(openTag) : end])
-	var req apiRequest
-	if err := json.Unmarshal([]byte(jsonStr), &req); err != nil {
-		logger.Warnf("Agent: failed to parse api_call JSON %q: %v", jsonStr, err)
-		return nil, resp
-	}
-
-	return &req, strings.TrimSpace(resp[:start])
-}
--- a/telegram/agent/prompt.go
+++ b/telegram/agent/prompt.go
@@ -13,35 +13,22 @@ func BuildAgentPrompt(apiDocs, userID string) string {
 - All API calls are made on behalf of this user
 - When asked "which user / username / email" — answer with this user ID directly, no API call needed

-## Tool: api_call
-
-When you need to call the API, your ENTIRE response must be ONLY the tag — nothing else:
-<api_call>{"method":"GET","path":"/api/xxx","body":{}}</api_call>
-
-When you have a final answer (no more API calls needed), reply with plain text — NO tag at all.
-
-ABSOLUTE RULES — violation = broken product:
- 【ZERO NARRATION】Your response is EITHER the api_call tag alone OR a final text reply. NEVER both except api_call at the very end.
- NEVER output ANY text before an api_call tag. No "好的", no "现在", no "我将", no "Let me", no "I will", no "正在", no "Creating...", no ellipsis, NOTHING.
- NEVER more than one <api_call> tag per response
+## Tool: api_request
+Use the api_request tool to call the NOFX REST API:
 - method: "GET" | "POST" | "PUT" | "DELETE"
+- path: API path; query params go in the path: /api/positions?trader_id=xxx
 - body: JSON object (use {} for GET requests)
- query parameters go in the path: /api/positions?trader_id=xxx

 ## NOFX API Documentation

 %s

 ## Behavior Rules
-1. 【SILENT ACTION】When you need to call an API: output ONLY the <api_call> tag. Zero words before it.
-2. Only ONE <api_call> tag per response, always alone with nothing else
-3. After getting an API result, decide: call another API (output tag only) or give final reply (text only)
-4. If the API returns success (2xx), the operation succeeded — do not retry
-5. Reply in the same language the user used (中文→中文, English→English)
-6. Keep final replies concise — show results, not process
-7. Ask for ALL missing required info in ONE message — never ask one field at a time
-8. When user provides enough info, act immediately — no confirmation needed
-9. Be decisive — infer intent from context, use schema to fill in smart defaults
+1. Reply in the same language the user used (中文→中文, English→English)
+2. Keep final replies concise — show results, not process
+3. Ask for ALL missing required info in ONE message — never ask one field at a time
+4. When user provides enough info, act immediately — no confirmation needed
+5. Be decisive — infer intent from context, use schema to fill in smart defaults

 ## Verification Rule (CRITICAL)
 After ANY PUT or POST that creates or modifies a resource:
@@ -56,7 +43,6 @@ After ANY PUT or POST that creates or modifies a resource:
 - "AI model not enabled": tell user to enable the model first via PUT /api/models
 - "Exchange not enabled": tell user to enable the exchange first
 - 5xx: server error, ask user to try again
- stream interrupted / unavailable: apologize briefly and ask user to retry

 ## Account State (injected at conversation start)
 At the start of each new conversation, a [Current Account State] block is provided with: