From 9d6f5dc22d8385a701d4c016f87c43c5730ce356 Mon Sep 17 00:00:00 2001
From: Vitali sharp8n <jetframes@gmail.com>
Date: Mon, 14 Sep 2026 17:27:38 +0300
Subject: [PATCH] enchancements

---
 .../Routing/RoutingService.InvokeAgent.cs        |  4 ++++
 .../Routing/RoutingService.InvokeFunction.cs     |  3 +++
 .../Providers/Chat/ChatCompletionProvider.cs     | 16 +++++++++++++++-
 3 files changed, 22 insertions(+), 1 deletion(-)

diff --git a/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeAgent.cs b/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeAgent.cs
index e4465f27..adb169a4 100644
--- a/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeAgent.cs
+++ b/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeAgent.cs
@@ -124,6 +124,10 @@ public partial class RoutingService
                 var msg = RoleDialogModel.From(message,
                     role: AgentRole.Function,
                     content: message.Content);
+                // Preserve images produced by the tool (e.g. `browser_act screenshot`) so the next
+                // LLM round in this turn can attach them to the tool message for a multimodal model.
+                // They are not persisted (ConversationStorage ignores dialog.Files).
+                msg.Files = message.Files;
 
                 dialogs.Add(msg);
                 Context.AddDialogs([msg]);
diff --git a/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeFunction.cs b/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeFunction.cs
index 3850dcc1..bfec68c0 100644
--- a/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeFunction.cs
+++ b/src/Infrastructure/BotSharp.Core/Routing/RoutingService.InvokeFunction.cs
@@ -66,6 +66,9 @@ public partial class RoutingService
             message.RichContent = clonedMessage.RichContent;
             message.Data = clonedMessage.Data;
             message.MessageLabel = clonedMessage.MessageLabel;
+            // Tools may return images (e.g. a browser screenshot). Carry them back from the clone so
+            // the function dialog handed to the LLM can include the tool-produced image content.
+            message.Files = clonedMessage.Files;
         }
         catch (JsonException ex)
         {
diff --git a/src/Plugins/BotSharp.Plugin.DeepSeekAI/Providers/Chat/ChatCompletionProvider.cs b/src/Plugins/BotSharp.Plugin.DeepSeekAI/Providers/Chat/ChatCompletionProvider.cs
index 0bde4fac..eebc64d1 100644
--- a/src/Plugins/BotSharp.Plugin.DeepSeekAI/Providers/Chat/ChatCompletionProvider.cs
+++ b/src/Plugins/BotSharp.Plugin.DeepSeekAI/Providers/Chat/ChatCompletionProvider.cs
@@ -524,7 +524,21 @@ public class ChatCompletionProvider : IChatCompletion
 
                 messages.Add(assistantToolMsg);
 
-                messages.Add(new ToolChatMessage(message.ToolCallId.IfNullOrEmptyAs(message.FunctionName), message.LlmContent));
+                // A tool may produce images (e.g. `browser_act screenshot`): attach them to the tool
+                // message so a multimodal model can SEE the tool output instead of only reading the
+                // file name. Text-only models keep the plain text result. The gateway accepts image
+                // parts on a `tool` message; these files are transient (never persisted in history).
+                var toolContentParts = new List<ChatMessageContentPart>
+                {
+                    ChatMessageContentPart.CreateTextPart(message.LlmContent)
+                };
+                if (allowMultiModal && !message.Files.IsNullOrEmpty())
+                {
+                    CollectMessageContentParts(toolContentParts, message.Files, imageDetailLevel);
+                }
+
+                messages.Add(new ToolChatMessage(
+                    message.ToolCallId.IfNullOrEmptyAs(message.FunctionName), toolContentParts));
             }
             else if (message.Role == AgentRole.User)
             {