fix: complete Telegram image pipeline for Claude Code agents

crates/openfang-api/src/routes.rs

@@ -6852,6 +6852,7 @@ pub async fn set_provider_key(
                     model: model_id,
                     api_key_env: env_var.clone(),
                     base_url: None,
+                    vision_model: None,
                 };
                 let mut guard = state
                     .kernel

crates/openfang-channels/src/bridge.rs

@@ -528,25 +528,34 @@ async fn dispatch_message(
         return;
     }
 
-    // For images: download, base64 encode, and send as multimodal content blocks
+    // For images: build content blocks with the image URL for vision models.
+    // We pass the original URL rather than downloading + base64-encoding because
+    // many providers (DashScope/Qwen, OpenAI) prefer or require direct URLs.
     if let ChannelContent::Image { ref url, ref caption } = message.content {
-        let blocks = download_image_to_blocks(url, caption.as_deref()).await;
-        if blocks.iter().any(|b| matches!(b, ContentBlock::Image { .. })) {
-            // We have actual image data — send as structured blocks for vision
-            dispatch_with_blocks(
-                blocks,
-                message,
-                handle,
-                router,
-                adapter,
-                ct_str,
-                thread_id,
-                output_format,
-            )
-            .await;
-            return;
+        let mut blocks = Vec::new();
+        if let Some(cap) = caption {
+            if !cap.is_empty() {
+                blocks.push(ContentBlock::Text {
+                    text: cap.clone(),
+                    provider_metadata: None,
+                });
+            }
         }
-        // Image download failed — fall through to text description below
+        blocks.push(ContentBlock::ImageUrl {
+            url: url.clone(),
+        });
+        dispatch_with_blocks(
+            blocks,
+            message,
+            handle,
+            router,
+            adapter,
+            ct_str,
+            thread_id,
+            output_format,
+        )
+        .await;
+        return;
     }
 
     let text = match &message.content {

crates/openfang-kernel/src/kernel.rs

@@ -2262,6 +2262,55 @@ impl OpenFangKernel {
             }
         }
 
+        // Vision model selection for image content.
+        // Priority: 1) explicit vision_model from config (forced override)
+        //           2) current agent model if it supports vision (no swap needed)
+        //           3) error — no vision capability available
+        if let Some(ref blocks) = content_blocks {
+            let has_images = blocks.iter().any(|b| {
+                matches!(
+                    b,
+                    openfang_types::message::ContentBlock::Image { .. }
+                        | openfang_types::message::ContentBlock::ImageUrl { .. }
+                )
+            });
+            if has_images {
+                if let Some(ref vision_model) = self.config.default_model.vision_model {
+                    // Explicit vision_model configured — always use it
+                    info!(
+                        agent = %manifest.name,
+                        current_model = %manifest.model.model,
+                        vision_model = %vision_model,
+                        "Swapping to configured vision model for image content"
+                    );
+                    manifest.model.model = vision_model.clone();
+                    manifest.model.provider = self.config.default_model.provider.clone();
+                } else {
+                    // No vision_model forced — check if current model handles vision
+                    let current_supports_vision = self
+                        .model_catalog
+                        .read()
+                        .ok()
+                        .and_then(|cat| cat.find_model(&manifest.model.model).map(|m| m.supports_vision))
+                        .unwrap_or(false);
+
+                    if current_supports_vision {
+                        info!(
+                            agent = %manifest.name,
+                            model = %manifest.model.model,
+                            "Current model supports vision — no swap needed"
+                        );
+                    } else {
+                        warn!(
+                            agent = %manifest.name,
+                            model = %manifest.model.model,
+                            "Image received but no vision_model configured and current model lacks vision support"
+                        );
+                    }
+                }
+            }
+        }
+
         let driver = self.resolve_driver(&manifest)?;
 
         // Look up model's actual context window from the catalog

crates/openfang-memory/src/session.rs

@@ -584,6 +584,9 @@ impl SessionStore {
                             ContentBlock::Image { media_type, .. } => {
                                 text_parts.push(format!("[image: {media_type}]"));
                             }
+                            ContentBlock::ImageUrl { ref url } => {
+                                text_parts.push(format!("[image: {url}]"));
+                            }
                             ContentBlock::Thinking { thinking } => {
                                 text_parts.push(format!(
                                     "[thinking: {}]",

crates/openfang-runtime/src/compactor.rs

@@ -399,6 +399,9 @@ fn build_conversation_text(messages: &[Message], config: &CompactionConfig) -> S
                         ContentBlock::Image { media_type, .. } => {
                             conversation_text.push_str(&format!("[Image: {media_type}]\n\n"));
                         }
+                        ContentBlock::ImageUrl { url } => {
+                            conversation_text.push_str(&format!("[Image: {url}]\n\n"));
+                        }
                         ContentBlock::Thinking { .. } => {}
                         ContentBlock::Unknown => {}
                     }

crates/openfang-runtime/src/drivers/anthropic.rs

@@ -573,6 +573,12 @@ fn convert_message(msg: &Message) -> ApiMessage {
                             data: data.clone(),
                         },
                     }),
+                    ContentBlock::ImageUrl { url } => {
+                        // Anthropic requires base64; pass as text description for now.
+                        Some(ApiContentBlock::Text {
+                            text: format!("[Image: {url}]"),
+                        })
+                    }
                     ContentBlock::ToolUse { id, name, input, .. } => Some(ApiContentBlock::ToolUse {
                         id: id.clone(),
                         name: name.clone(),