fix(openai): infer previous response for item references

.gitignore

@@ -122,7 +122,7 @@ scripts
 .code-review-state
 #openspec/
 code-reviews/
-#AGENTS.md
+AGENTS.md
 backend/cmd/server/server
 deploy/docker-compose.override.yml
 .gocache/

backend/internal/service/openai_ws_forwarder.go

@@ -1379,10 +1379,12 @@ func shouldInferIngressFunctionCallOutputPreviousResponseID(
 	if signals.HasFunctionCallOutputMissingCallID {
 		return false
 	}
-	// If the client already sent tool-call context or item_reference anchors,
-	// treat this as a full replay / self-contained continuation payload rather
-	// than downgrading it into an inferred delta continuation.
-	if signals.HasToolCallContext || signals.HasItemReferenceForAllCallIDs {
+	// If the client already sent the actual tool-call context, treat this as
+	// a full replay / self-contained continuation payload rather than
+	// downgrading it into an inferred delta continuation. item_reference alone
+	// is not enough on the store=false WS path: it still needs a valid prior
+	// response anchor so upstream can resolve the referenced function_call.
+	if signals.HasToolCallContext {
 		return false
 	}
 	return strings.TrimSpace(expectedPreviousResponseID) != ""

backend/internal/service/openai_ws_forwarder_ingress_session_test.go

@@ -1488,7 +1488,7 @@ func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_StoreDisabledFun
 	require.False(t, gjson.Get(requestToJSONString(captureConn.writes[1]), "previous_response_id").Exists(), "请求已包含 function_call 上下文时不应自动补齐 previous_response_id")
 }
 
-func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_StoreDisabledFunctionCallOutputSkipsAutoAttachWhenItemReferencesPresent(t *testing.T) {
+func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_StoreDisabledFunctionCallOutputAutoAttachWhenOnlyItemReferencesPresent(t *testing.T) {
 	gin.SetMode(gin.TestMode)
 
 	cfg := &config.Config{}
@@ -1619,7 +1619,7 @@ func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_StoreDisabledFun
 
 	require.Equal(t, 1, captureDialer.DialCount())
 	require.Len(t, captureConn.writes, 2)
-	require.False(t, gjson.Get(requestToJSONString(captureConn.writes[1]), "previous_response_id").Exists(), "请求已包含 item_reference 锚点时不应自动补齐 previous_response_id")
+	require.Equal(t, "resp_auto_prev_ref_1", gjson.Get(requestToJSONString(captureConn.writes[1]), "previous_response_id").String(), "仅有 item_reference 不足以自包含 function_call_output，应回填上一轮响应 ID")
 }
 
 func TestOpenAIGatewayService_ProxyResponsesWebSocketFromClient_PreflightPingFailReconnectsBeforeTurn(t *testing.T) {

backend/internal/service/openai_ws_forwarder_ingress_test.go

@@ -303,12 +303,12 @@ func TestShouldInferIngressFunctionCallOutputPreviousResponseID(t *testing.T) {
 			want:             false,
 		},
 		{
-			name:             "skip_when_item_reference_already_covers_all_call_ids",
+			name:             "infer_when_only_item_reference_covers_call_ids",
 			storeDisabled:    true,
 			turn:             2,
 			signals:          ToolContinuationSignals{HasFunctionCallOutput: true, HasItemReferenceForAllCallIDs: true},
 			expectedPrevious: "resp_2",
-			want:             false,
+			want:             true,
 		},
 		{
 			name:             "skip_when_function_call_output_missing_call_id",