diff --git a/internal/adapter/claude/handler.go b/internal/adapter/claude/handler.go
index 282b569..593c1bc 100644
--- a/internal/adapter/claude/handler.go
+++ b/internal/adapter/claude/handler.go
@@ -250,7 +250,7 @@ func normalizeClaudeMessages(messages []any) []any {
 					}
 				}
 				if typeStr == "tool_result" {
-					parts = append(parts, fmt.Sprintf("%v", b["content"]))
+					parts = append(parts, formatClaudeToolResultForPrompt(b))
 				}
 			}
 			copied["content"] = strings.Join(parts, "\n")
@@ -272,10 +272,36 @@ func buildClaudeToolPrompt(tools []any) string {
 		schema, _ := json.Marshal(m["input_schema"])
 		parts = append(parts, fmt.Sprintf("Tool: %s\nDescription: %s\nParameters: %s", name, desc, schema))
 	}
-	parts = append(parts, "When you need to use tools, you can call multiple tools in one response. Output ONLY JSON like {\"tool_calls\":[{\"name\":\"tool\",\"input\":{}}]}")
+	parts = append(parts,
+		"When you need to use tools, you can call multiple tools in one response. Output ONLY JSON like {\"tool_calls\":[{\"name\":\"tool\",\"input\":{}}]}",
+		"History markers in conversation: [TOOL_CALL_HISTORY]...[/TOOL_CALL_HISTORY] are your previous tool calls; [TOOL_RESULT_HISTORY]...[/TOOL_RESULT_HISTORY] are runtime tool outputs, not user input.",
+		"After a valid [TOOL_RESULT_HISTORY], continue with final answer instead of repeating the same call unless required fields are still missing.",
+	)
 	return strings.Join(parts, "\n\n")
 }
 
+func formatClaudeToolResultForPrompt(block map[string]any) string {
+	if block == nil {
+		return ""
+	}
+	toolCallID := strings.TrimSpace(fmt.Sprintf("%v", block["tool_use_id"]))
+	if toolCallID == "" {
+		toolCallID = strings.TrimSpace(fmt.Sprintf("%v", block["tool_call_id"]))
+	}
+	if toolCallID == "" {
+		toolCallID = "unknown"
+	}
+	name := strings.TrimSpace(fmt.Sprintf("%v", block["name"]))
+	if name == "" {
+		name = "unknown"
+	}
+	content := strings.TrimSpace(fmt.Sprintf("%v", block["content"]))
+	if content == "" {
+		content = "null"
+	}
+	return fmt.Sprintf("[TOOL_RESULT_HISTORY]\nstatus: already_returned\norigin: tool_runtime\nnot_user_input: true\ntool_call_id: %s\nname: %s\ncontent: %s\n[/TOOL_RESULT_HISTORY]", toolCallID, name, content)
+}
+
 func hasSystemMessage(messages []any) bool {
 	for _, m := range messages {
 		msg, ok := m.(map[string]any)
diff --git a/internal/adapter/claude/handler_util_test.go b/internal/adapter/claude/handler_util_test.go
index 73d2fab..ae75d8e 100644
--- a/internal/adapter/claude/handler_util_test.go
+++ b/internal/adapter/claude/handler_util_test.go
@@ -1,6 +1,7 @@
 package claude
 
 import (
+	"strings"
 	"testing"
 )
 
@@ -48,8 +49,9 @@ func TestNormalizeClaudeMessagesToolResult(t *testing.T) {
 	}
 	got := normalizeClaudeMessages(msgs)
 	m := got[0].(map[string]any)
-	if m["content"] != "tool output" {
-		t.Fatalf("expected 'tool output', got %q", m["content"])
+	content, _ := m["content"].(string)
+	if !strings.Contains(content, "[TOOL_RESULT_HISTORY]") || !strings.Contains(content, "content: tool output") {
+		t.Fatalf("expected serialized tool result marker, got %q", content)
 	}
 }
 
diff --git a/internal/adapter/openai/handler.go b/internal/adapter/openai/handler.go
index 28a451c..ac8666f 100644
--- a/internal/adapter/openai/handler.go
+++ b/internal/adapter/openai/handler.go
@@ -233,7 +233,7 @@ func injectToolPrompt(messages []map[string]any, tools []any) ([]map[string]any,
 	if len(toolSchemas) == 0 {
 		return messages, names
 	}
-	toolPrompt := "You have access to these tools:\n\n" + strings.Join(toolSchemas, "\n\n") + "\n\nWhen you need to use tools, output ONLY this JSON format (no other text):\n{\"tool_calls\": [{\"name\": \"tool_name\", \"input\": {\"param\": \"value\"}}]}\n\nIMPORTANT:\n1) If calling tools, output ONLY the JSON. The response must start with { and end with }.\n2) After receiving a tool result, you MUST use it to produce the final answer.\n3) Only call another tool when the previous result is missing required data or returned an error."
+	toolPrompt := "You have access to these tools:\n\n" + strings.Join(toolSchemas, "\n\n") + "\n\nWhen you need to use tools, output ONLY this JSON format (no other text):\n{\"tool_calls\": [{\"name\": \"tool_name\", \"input\": {\"param\": \"value\"}}]}\n\nHistory markers in conversation:\n- [TOOL_CALL_HISTORY]...[/TOOL_CALL_HISTORY] means a tool call you already made earlier.\n- [TOOL_RESULT_HISTORY]...[/TOOL_RESULT_HISTORY] means the runtime returned a tool result (not user input).\n\nIMPORTANT:\n1) If calling tools, output ONLY the JSON. The response must start with { and end with }.\n2) After receiving a tool result, you MUST use it to produce the final answer.\n3) Only call another tool when the previous result is missing required data or returned an error.\n4) Do not repeat a tool call that is already satisfied by an existing [TOOL_RESULT_HISTORY] block."
 
 	for i := range messages {
 		if messages[i]["role"] == "system" {
diff --git a/internal/adapter/openai/message_normalize.go b/internal/adapter/openai/message_normalize.go
index 3ebd1e7..c0ab7d2 100644
--- a/internal/adapter/openai/message_normalize.go
+++ b/internal/adapter/openai/message_normalize.go
@@ -86,7 +86,7 @@ func formatAssistantToolCallsForPrompt(msg map[string]any) string {
 			if args == "" {
 				args = "{}"
 			}
-			entries = append(entries, fmt.Sprintf("Tool call:\n- tool_call_id: %s\n- function.name: %s\n- function.arguments: %s", id, name, args))
+			entries = append(entries, fmt.Sprintf("[TOOL_CALL_HISTORY]\nstatus: already_called\norigin: assistant\nnot_user_input: true\ntool_call_id: %s\nfunction.name: %s\nfunction.arguments: %s\n[/TOOL_CALL_HISTORY]", id, name, args))
 		}
 	}
 
@@ -99,7 +99,7 @@ func formatAssistantToolCallsForPrompt(msg map[string]any) string {
 		if args == "" {
 			args = "{}"
 		}
-		entries = append(entries, fmt.Sprintf("Tool call:\n- tool_call_id: call_legacy\n- function.name: %s\n- function.arguments: %s", name, args))
+		entries = append(entries, fmt.Sprintf("[TOOL_CALL_HISTORY]\nstatus: already_called\norigin: assistant\nnot_user_input: true\ntool_call_id: call_legacy\nfunction.name: %s\nfunction.arguments: %s\n[/TOOL_CALL_HISTORY]", name, args))
 	}
 
 	return strings.Join(entries, "\n\n")
@@ -124,7 +124,7 @@ func formatToolResultForPrompt(msg map[string]any) string {
 		content = "null"
 	}
 
-	return fmt.Sprintf("Tool result:\n- tool_call_id: %s\n- name: %s\n- content: %s", toolCallID, name, content)
+	return fmt.Sprintf("[TOOL_RESULT_HISTORY]\nstatus: already_returned\norigin: tool_runtime\nnot_user_input: true\ntool_call_id: %s\nname: %s\ncontent: %s\n[/TOOL_RESULT_HISTORY]", toolCallID, name, content)
 }
 
 func normalizeOpenAIContentForPrompt(v any) string {
diff --git a/internal/adapter/openai/message_normalize_test.go b/internal/adapter/openai/message_normalize_test.go
index bb648d3..27849d7 100644
--- a/internal/adapter/openai/message_normalize_test.go
+++ b/internal/adapter/openai/message_normalize_test.go
@@ -38,18 +38,19 @@ func TestNormalizeOpenAIMessagesForPrompt_AssistantToolCallsAndToolResult(t *tes
 		t.Fatalf("expected 4 normalized messages, got %d", len(normalized))
 	}
 	assistantContent, _ := normalized[2]["content"].(string)
-	if !strings.Contains(assistantContent, "tool_call_id: call_1") ||
+	if !strings.Contains(assistantContent, "[TOOL_CALL_HISTORY]") ||
+		!strings.Contains(assistantContent, "tool_call_id: call_1") ||
 		!strings.Contains(assistantContent, "function.name: get_weather") ||
 		!strings.Contains(assistantContent, "function.arguments: {\"city\":\"beijing\"}") {
 		t.Fatalf("assistant tool call not serialized correctly: %q", assistantContent)
 	}
 	toolContent, _ := normalized[3]["content"].(string)
-	if !strings.Contains(toolContent, "Tool result:") || !strings.Contains(toolContent, "name: get_weather") {
+	if !strings.Contains(toolContent, "[TOOL_RESULT_HISTORY]") || !strings.Contains(toolContent, "name: get_weather") {
 		t.Fatalf("tool result not serialized correctly: %q", toolContent)
 	}
 
 	prompt := util.MessagesPrepare(normalized)
-	if !strings.Contains(prompt, "tool_call_id: call_1") || !strings.Contains(prompt, "Tool result:") {
+	if !strings.Contains(prompt, "tool_call_id: call_1") || !strings.Contains(prompt, "[TOOL_RESULT_HISTORY]") {
 		t.Fatalf("expected prompt to include tool call + result semantics: %q", prompt)
 	}
 }
diff --git a/internal/adapter/openai/prompt_build_test.go b/internal/adapter/openai/prompt_build_test.go
index 1833860..878af73 100644
--- a/internal/adapter/openai/prompt_build_test.go
+++ b/internal/adapter/openai/prompt_build_test.go
@@ -46,7 +46,7 @@ func TestBuildOpenAIFinalPrompt_HandlerPathIncludesToolRoundtripSemantics(t *tes
 	}
 	if !strings.Contains(finalPrompt, "tool_call_id: call_1") ||
 		!strings.Contains(finalPrompt, "function.name: get_weather") ||
-		!strings.Contains(finalPrompt, "Tool result:") ||
+		!strings.Contains(finalPrompt, "[TOOL_RESULT_HISTORY]") ||
 		!strings.Contains(finalPrompt, `"condition":"sunny"`) {
 		t.Fatalf("handler finalPrompt missing tool roundtrip semantics: %q", finalPrompt)
 	}
@@ -77,4 +77,7 @@ func TestBuildOpenAIFinalPrompt_VercelPreparePathKeepsFinalAnswerInstruction(t *
 	if !strings.Contains(finalPrompt, "Only call another tool when the previous result is missing required data or returned an error.") {
 		t.Fatalf("vercel prepare finalPrompt missing retry guard instruction: %q", finalPrompt)
 	}
+	if !strings.Contains(finalPrompt, "[TOOL_RESULT_HISTORY]") {
+		t.Fatalf("vercel prepare finalPrompt missing history marker instruction: %q", finalPrompt)
+	}
 }