fix: keep tool call in provider sanitize when text interleaves call and result

trimDanglingAssistantToolCalls in the provider router held the same
consecutive-tool-message assumption as the replay projector: when a
model emits function_call items before explanation text within one
response (e.g. gpt-5.3-codex-spark), the sanitized sequence becomes
assistant[tool_call] -> assistant[text] -> tool[result] and responded
calls were stripped at the provider boundary even though the projector
already replayed them correctly. The orphan tool result then forced the
Responses adapter to synthesize a placeholder function_call with empty
arguments, degrading the arguments the model sees on in-loop requests.

Widen the response collection window to skip interleaved plain
assistant text messages and drop orphan tool results, mirroring the
projector fix.

internal/backend/agent/model/router.go

@@ -339,57 +339,82 @@ func mergeProviderReasoningMetadata(last *Message, current Message) {
 	}
 }
 
+// providerToolResponseWindowEnd 返回 assistant tool-call 消息的响应收集窗口右边界（不含）。
+// 窗口内除 tool 结果消息外，还允许出现同轮穿插的纯文本 assistant 消息：
+// 部分模型（如 gpt-5.3-codex-spark）会在同一条响应里先输出 function_call 再输出说明文本，
+// 回放顺序为 assistant[tool_call] -> assistant[text] -> tool[result]，若只收集紧邻的
+// tool 消息，会把有结果回放的调用误判为悬空。
+func providerToolResponseWindowEnd(messages []Message, index int) int {
+	end := index + 1
+	for end < len(messages) {
+		candidate := messages[end]
+		switch {
+		case strings.TrimSpace(candidate.Role) == "tool":
+			end++
+		case strings.TrimSpace(candidate.Role) == "assistant" && len(candidate.ToolCalls) == 0:
+			end++
+		default:
+			return end
+		}
+	}
+	return end
+}
+
 func trimDanglingAssistantToolCalls(input []Message) []Message {
 	if len(input) == 0 {
 		return nil
 	}
-	trimmed := make([]Message, 0, len(input))
+	survivingToolCallIDs := make(map[string]struct{})
-	for index := 0; index < len(input); index++ {
+	for index, message := range input {
-		message := cloneProviderMessage(input[index])
 		if strings.TrimSpace(message.Role) != "assistant" || len(message.ToolCalls) == 0 {
-			trimmed = append(trimmed, message)
 			continue
 		}
-
-		end := index + 1
 		responded := make(map[string]struct{}, len(message.ToolCalls))
-		for end < len(input) && strings.TrimSpace(input[end].Role) == "tool" {
+		for scan := index + 1; scan < providerToolResponseWindowEnd(input, index); scan++ {
-			toolCallID := strings.TrimSpace(input[end].ToolCallID)
+			if strings.TrimSpace(input[scan].Role) != "tool" {
-			if toolCallID != "" {
+				continue
+			}
+			if toolCallID := strings.TrimSpace(input[scan].ToolCallID); toolCallID != "" {
 				responded[toolCallID] = struct{}{}
 			}
-			end++
 		}
-
-		nextToolCalls := make([]ToolCallDescriptor, 0, len(message.ToolCalls))
-		allowedToolCallIDs := make(map[string]struct{}, len(message.ToolCalls))
 		for _, toolCall := range message.ToolCalls {
-			toolCallID := strings.TrimSpace(toolCall.ID)
+			if toolCallID := strings.TrimSpace(toolCall.ID); toolCallID != "" {
-			if _, ok := responded[toolCallID]; !ok {
+				if _, ok := responded[toolCallID]; ok {
+					survivingToolCallIDs[toolCallID] = struct{}{}
+				}
+			}
+		}
+	}
+	trimmed := make([]Message, 0, len(input))
+	for _, item := range input {
+		message := cloneProviderMessage(item)
+		if strings.TrimSpace(message.Role) == "assistant" && len(message.ToolCalls) > 0 {
+			nextToolCalls := make([]ToolCallDescriptor, 0, len(message.ToolCalls))
+			for _, toolCall := range message.ToolCalls {
+				if _, ok := survivingToolCallIDs[strings.TrimSpace(toolCall.ID)]; !ok {
 					continue
 				}
-			item := toolCall
+				toolCall.Index = len(nextToolCalls)
-			item.Index = len(nextToolCalls)
+				nextToolCalls = append(nextToolCalls, toolCall)
-			nextToolCalls = append(nextToolCalls, item)
-			allowedToolCallIDs[toolCallID] = struct{}{}
 			}
-
+			if len(nextToolCalls) == 0 {
-		if len(nextToolCalls) > 0 {
+				if strings.TrimSpace(message.Content) == "" && len(message.ContentParts) == 0 && strings.TrimSpace(message.ReasoningContent) == "" {
-			message.ToolCalls = nextToolCalls
-			trimmed = append(trimmed, message)
-			for toolIndex := index + 1; toolIndex < end; toolIndex++ {
-				toolMessage := cloneProviderMessage(input[toolIndex])
-				if _, ok := allowedToolCallIDs[strings.TrimSpace(toolMessage.ToolCallID)]; !ok {
 					continue
 				}
-				trimmed = append(trimmed, toolMessage)
-			}
-		} else if strings.TrimSpace(message.Content) != "" || len(message.ContentParts) > 0 || strings.TrimSpace(message.ReasoningContent) != "" {
 				message.ToolCalls = nil
-			trimmed = append(trimmed, message)
+			} else {
+				message.ToolCalls = nextToolCalls
 			}
-
+			trimmed = append(trimmed, message)
-		index = end - 1
+			continue
+		}
+		if strings.TrimSpace(message.Role) == "tool" && strings.TrimSpace(message.ToolCallID) != "" {
+			if _, ok := survivingToolCallIDs[strings.TrimSpace(message.ToolCallID)]; !ok {
+				continue
+			}
+		}
+		trimmed = append(trimmed, message)
 	}
 	return trimmed
 }

internal/backend/agent/model/router_dangling_tool_call_test.go

@@ -0,0 +1,77 @@
+package modeladapter
+
+import "testing"
+
+func routerToolCall(id string, name string) ToolCallDescriptor {
+	return ToolCallDescriptor{
+		ID:   id,
+		Type: "function",
+		Function: ToolCallFunctionShape{
+			Name:      name,
+			Arguments: `{"pattern":"ref"}`,
+		},
+	}
+}
+
+// 模型在同一条响应里先输出 function_call 再输出说明文本时，
+// sanitize 后的消息序列为 assistant[tool_call] -> assistant[text] -> tool[result]。
+// trimDanglingAssistantToolCalls 需要越过中间的文本消息收集工具结果，
+// 否则会产生孤儿 function_call_output（Responses API 400）。
+func TestTrimDanglingAssistantToolCallsKeepsInterleavedTextResponses(t *testing.T) {
+	input := []Message{
+		{Role: "user", Content: "query"},
+		{
+			Role: "assistant",
+			ToolCalls: []ToolCallDescriptor{
+				routerToolCall("call_1", "Grep"),
+				routerToolCall("call_2", "Read"),
+			},
+		},
+		{Role: "tool", Name: "Grep", ToolCallID: "call_1", Content: "grep result"},
+		{Role: "assistant", Content: "我先快速定位上下文"},
+		{Role: "tool", Name: "Read", ToolCallID: "call_2", Content: "read result"},
+		{Role: "user", Content: "next"},
+	}
+
+	trimmed := sanitizeProviderMessages(input)
+	if len(trimmed) != 6 {
+		t.Fatalf("expected 6 messages, got %d: %+v", len(trimmed), trimmed)
+	}
+	if len(trimmed[1].ToolCalls) != 2 {
+		t.Fatalf("expected both tool calls to survive, got %+v", trimmed[1].ToolCalls)
+	}
+	if trimmed[3].Role != "assistant" || trimmed[3].Content == "" {
+		t.Fatalf("expected interleaved assistant text to survive, got %+v", trimmed[3])
+	}
+	if trimmed[4].ToolCallID != "call_2" {
+		t.Fatalf("expected call_2 result to survive, got %+v", trimmed[4])
+	}
+}
+
+// 完全没有结果回放的调用仍应被剥离，且孤儿 tool 结果不得保留。
+func TestTrimDanglingAssistantToolCallsDropsUnrespondedCallsAndOrphanResults(t *testing.T) {
+	input := []Message{
+		{Role: "user", Content: "query"},
+		{
+			Role: "assistant",
+			ToolCalls: []ToolCallDescriptor{
+				routerToolCall("call_1", "Grep"),
+				routerToolCall("call_2", "Read"),
+			},
+		},
+		{Role: "tool", Name: "Grep", ToolCallID: "call_1", Content: "grep result"},
+		{Role: "tool", Name: "Read", ToolCallID: "call_3", Content: "orphan result"},
+		{Role: "user", Content: "next"},
+	}
+
+	trimmed := sanitizeProviderMessages(input)
+	if len(trimmed) != 4 {
+		t.Fatalf("expected 4 messages, got %d: %+v", len(trimmed), trimmed)
+	}
+	if len(trimmed[1].ToolCalls) != 1 || trimmed[1].ToolCalls[0].ID != "call_1" {
+		t.Fatalf("expected only call_1 to survive, got %+v", trimmed[1].ToolCalls)
+	}
+	if trimmed[2].ToolCallID != "call_1" {
+		t.Fatalf("expected only call_1 result to survive, got %+v", trimmed[2])
+	}
+}