fix: keep tool call replay when assistant text interleaves call and result

trimReplayDanglingAssistantToolCalls only collected tool results that
immediately followed the assistant tool-call message. Models such as
gpt-5.3-codex-spark may emit the function_call item before the
explanation text within one response, so history replay order becomes
assistant[tool_call] -> assistant[text] -> tool[result]. The call was
misjudged as dangling and stripped while the tool result survived,
producing a function_call_output without a matching function_call that
the Responses API rejects with 400.

- widen the response collection window to skip interleaved plain
  assistant text messages, and drop orphan tool results in the same pass
- synthesize a placeholder function_call (or drop the output when the
  tool name is unknown) in normalizeOpenAIResponsesInput so conversations
  already persisted with corrupted history can resume

internal/backend/agent/model/openai.go

@@ -1956,6 +1956,7 @@ func normalizeOpenAIResponsesInput(messages []Message) (string, []map[string]any
 	instructionParts := make([]string, 0, 2)
 	items := make([]map[string]any, 0, len(messages))
 	responsesCallIDs := make(map[string]string)
+	emittedCallIDs := make(map[string]struct{})
 	activeAssistantReasoningKey := ""
 	for _, message := range messages {
 		role := strings.TrimSpace(message.Role)
@@ -1968,6 +1969,26 @@ func normalizeOpenAIResponsesInput(messages []Message) (string, []map[string]any
 		}
 		if role == "tool" && strings.TrimSpace(message.ToolCallID) != "" {
 			callID := openAIResponsesToolMessageCallID(message, responsesCallIDs)
+			if strings.TrimSpace(callID) != "" {
+				if _, ok := emittedCallIDs[callID]; !ok {
+					// 历史损坏时可能出现没有配对 function_call 的工具结果
+					// （例如旧版本回放逻辑剥离了 assistant 调用但保留了结果）。
+					// Responses API 会直接拒绝这种 input，这里补一个占位 function_call
+					// 让旧会话可以继续；无法补齐时丢弃该结果。
+					if name := strings.TrimSpace(message.Name); name != "" {
+						items = append(items, map[string]any{
+							"type":      "function_call",
+							"call_id":   callID,
+							"name":      name,
+							"arguments": "{}",
+							"status":    "completed",
+						})
+						emittedCallIDs[callID] = struct{}{}
+					} else {
+						continue
+					}
+				}
+			}
 			var output any = openAIResponsesMessageText(message)
 			if hasImageContentParts(message.ContentParts) {
 				content, err := openAIResponsesMessageContent(message, false)
@@ -2034,6 +2055,9 @@ func normalizeOpenAIResponsesInput(messages []Message) (string, []map[string]any
 					toolItem["status"] = "completed"
 				}
 				items = append(items, toolItem)
+				if strings.TrimSpace(callID) != "" {
+					emittedCallIDs[strings.TrimSpace(callID)] = struct{}{}
+				}
 			}
 		}
 	}

internal/backend/agent/model/openai_responses_orphan_output_test.go

@@ -0,0 +1,96 @@
+package modeladapter
+
+import "testing"
+
+// 历史损坏时（旧版本回放逻辑剥离了 assistant 调用但保留了结果），
+// function_call_output 会缺少配对的 function_call，Responses API 会拒绝。
+// 这里验证 adapter 会为孤儿结果补一个占位 function_call，让旧会话可以继续。
+func TestNormalizeOpenAIResponsesInputSynthesizesCallForOrphanToolOutput(t *testing.T) {
+	messages := []Message{
+		{Role: "user", Content: "query"},
+		{Role: "assistant", Content: "我先快速定位上下文"},
+		{Role: "tool", Name: "Grep", ToolCallID: "call_xrN6", Content: "grep result"},
+		{Role: "user", Content: "next"},
+	}
+
+	_, items, err := normalizeOpenAIResponsesInput(messages)
+	if err != nil {
+		t.Fatalf("normalizeOpenAIResponsesInput failed: %v", err)
+	}
+
+	var callIndexes []int
+	var outputIndexes []int
+	for index, item := range items {
+		if item["type"] == "function_call" && item["call_id"] == "call_xrN6" {
+			callIndexes = append(callIndexes, index)
+		}
+		if item["type"] == "function_call_output" && item["call_id"] == "call_xrN6" {
+			outputIndexes = append(outputIndexes, index)
+		}
+	}
+	if len(callIndexes) != 1 {
+		t.Fatalf("expected 1 synthesized function_call, got %d: %+v", len(callIndexes), items)
+	}
+	if len(outputIndexes) != 1 || outputIndexes[0] != callIndexes[0]+1 {
+		t.Fatalf("expected function_call_output right after synthesized function_call, calls=%v outputs=%v", callIndexes, outputIndexes)
+	}
+	if got := items[callIndexes[0]]["name"]; got != "Grep" {
+		t.Fatalf("expected synthesized function_call name Grep, got %v", got)
+	}
+}
+
+// 连工具名都没有的孤儿结果只能丢弃，避免上游 400。
+func TestNormalizeOpenAIResponsesInputDropsNamelessOrphanToolOutput(t *testing.T) {
+	messages := []Message{
+		{Role: "user", Content: "query"},
+		{Role: "tool", ToolCallID: "call_unknown", Content: "orphan result"},
+		{Role: "user", Content: "next"},
+	}
+
+	_, items, err := normalizeOpenAIResponsesInput(messages)
+	if err != nil {
+		t.Fatalf("normalizeOpenAIResponsesInput failed: %v", err)
+	}
+	for _, item := range items {
+		if item["type"] == "function_call_output" {
+			t.Fatalf("expected orphan function_call_output to be dropped, got %+v", item)
+		}
+	}
+}
+
+// 正常配对的调用与结果不应受防御逻辑影响。
+func TestNormalizeOpenAIResponsesInputKeepsPairedCallAndOutput(t *testing.T) {
+	messages := []Message{
+		{Role: "user", Content: "query"},
+		{
+			Role: "assistant",
+			ToolCalls: []ToolCallDescriptor{{
+				ID:   "call_xrN6",
+				Type: "function",
+				Function: ToolCallFunctionShape{
+					Name:      "Grep",
+					Arguments: `{"pattern":"ref"}`,
+				},
+			}},
+		},
+		{Role: "tool", Name: "Grep", ToolCallID: "call_xrN6", Content: "grep result"},
+		{Role: "user", Content: "next"},
+	}
+
+	_, items, err := normalizeOpenAIResponsesInput(messages)
+	if err != nil {
+		t.Fatalf("normalizeOpenAIResponsesInput failed: %v", err)
+	}
+	var calls, outputs int
+	for _, item := range items {
+		if item["type"] == "function_call" && item["call_id"] == "call_xrN6" {
+			calls++
+		}
+		if item["type"] == "function_call_output" && item["call_id"] == "call_xrN6" {
+			outputs++
+		}
+	}
+	if calls != 1 || outputs != 1 {
+		t.Fatalf("expected exactly one paired function_call/output, got calls=%d outputs=%d", calls, outputs)
+	}
+}

internal/backend/agent/model/tool_image_test.go

@@ -31,12 +31,14 @@ func TestToolImageProviderEncodings(t *testing.T) {
 		if err != nil {
 			t.Fatalf("normalizeOpenAIResponsesInput() error = %v", err)
 		}
-		if len(items) != 1 || items[0]["type"] != "function_call_output" {
+		// 孤儿 tool 结果（无前置 assistant 调用）会补一个占位 function_call，
+		// 保证每个 function_call_output 都有配对调用。
+		if len(items) != 2 || items[0]["type"] != "function_call" || items[0]["call_id"] != "call-1" || items[1]["type"] != "function_call_output" {
 			t.Fatalf("openai responses items = %#v", items)
 		}
-		content, ok := items[0]["output"].([]map[string]any)
+		content, ok := items[1]["output"].([]map[string]any)
 		if !ok || len(content) != 2 {
-			t.Fatalf("openai responses output = %#v", items[0]["output"])
+			t.Fatalf("openai responses output = %#v", items[1]["output"])
 		}
 		if content[0]["type"] != "input_text" || content[1]["type"] != "input_image" {
 			t.Fatalf("openai responses content = %#v", content)