支持工具读取图片并注入多模态模型

AGENTS.md

@@ -86,7 +86,7 @@ Scheduler → SessionManager scheduled execution → AgentLoop → Scheduler del
 - **WebUI/TUI file transfer** streams bytes over authenticated HTTP and sends only short-lived upload IDs/attachment metadata over WebSocket; messages persist local media paths without guaranteeing later availability, and client responses must never expose those paths
 - **WebUI authentication** protects every management API and `/ws`; only static pairing assets, public health/status, and pairing submission may bypass device auth. Pair-code issuance requires both a real loopback peer and the filesystem-held admin token; never put bearer tokens in URLs or logs
 - **Providers** are pure HTTP clients; no bus/session/channel awareness
-- **Tools** are executed by `AgentLoop`; they receive raw arguments and return string results
+- **Tools** are executed by `AgentLoop`; they receive raw arguments and normally return text. Tools that produce model-consumable media use the structured `execute_with_media` side channel; model capability checks and provider content-block serialization stay outside tools
 
 ### Concurrency and Lifecycle Invariants
 

README.md

@@ -254,7 +254,7 @@ PicoBot 有两类记忆：
 | 工具 | 说明 |
 |------|------|
 | `calculator` | 数学表达式和统计计算 |
-| `file_read` / `file_write` / `file_edit` | 文件读写和编辑 |
+| `file_read` / `file_write` / `file_edit` | 文件读写和编辑；`file_read` 读取受支持图片时可将图片直接提供给多模态模型 |
 | `file_search` / `content_search` | 文件名和内容搜索 |
 | `bash` | 在 workspace 中执行 Shell 命令 |
 | `http_request` / `web_fetch` | HTTP 请求和网页文本抽取 |

docs/ARCHITECTURE.md

@@ -201,6 +201,8 @@ Gateway 在 `/` 提供随二进制编译的 HTML/CSS/JavaScript，不依赖外
 
 WebUI/TUI 文件字节通过受鉴权的 HTTP 接口流式传输，WebSocket 只携带短期 `upload_id` 和结构化附件描述。`UploadRegistry` 在内存中按 `cli_chat` chat scope 校验并消费待发送上传；消息继续以 `media_refs` 保存 Gateway 本地路径，不建立永久附件资产。下载接口必须通过 client、session、message 和附件序号反查路径，不能接受客户端路径。历史附件路径失效属于正常状态，不得影响历史文本读取。待发送但未进入消息的上传由 `TaskSupervisor` 所有的限时清理任务回收。Agent 上下文会为所有媒体注入内部路径清单，客户端响应不得暴露该路径。
 
+工具默认通过 `ToolResult` 返回文本；需要把图片等产物交给模型时，通过 `Tool::execute_with_media` 返回文本和结构化 `MediaRef`。工具只负责经过自身路径策略校验后声明媒体，不感知当前模型或 Provider。`AgentLoop` 仅将最新连续工具结果批次的媒体交给 `MediaHandlerRegistry`，旧工具媒体只回放文本和路径，避免历史 Base64 膨胀。OpenAI-compatible Provider 保持 `tool` 结果为文本，并在完整工具批次后构造仅存在于请求内的临时多模态 `user` 消息；Anthropic Provider 将媒体放入对应 `tool_result.content`。媒体加载、格式或能力检查失败必须降级成文本，不得使历史记录不可读取。
+
 `AuthManager` 默认保护所有管理 API 和 `/ws`。静态资源、`/health`、`/api/auth/status` 与 `/api/auth/pair` 保持公开，使未配对浏览器只能加载配对界面。`picobot pair` 使用权限为 `0600` 的本机管理密钥调用仅接受真实回环连接的 `/api/auth/code`；反向代理即使从回环连接也无法在没有该密钥时签发代码。配对码为 8 位、5 分钟有效、单次消费，并按来源实施失败锁定。浏览器收到 HttpOnly、SameSite=Strict Cookie，CLI 使用 Bearer token；服务端仅持久化 SHA-256 哈希。`--revoke-all` 的持久化成功后才清空内存令牌，活动 WebSocket 每 5 秒复核身份并回收已撤销连接。
 
 同源 `/api/*` 管理接口只提供显式白名单能力：

docs/FILE_TRANSFER_DESIGN.md

@@ -259,21 +259,27 @@ Agent 继续使用 `send_message(files=...)`。文件路径由服务端内部产
 
 ### 7.3 Agent 上下文中的路径
 
-构造发给模型的用户消息时，无论附件类型是否被模型原生支持，都先增加一个结构化附件清单。清单包含安全文件名、媒体类型和 Gateway 内部路径，例如：
+构造发给模型的用户消息时，无论附件类型是否被模型原生支持，都把结构化附件清单和用户正文放在同一个文本内容块中。清单包含安全文件名、扩展名、媒体类型、MIME、当前可取得的大小、Gateway 内部路径和内容交付状态，例如：
 
 ```text
-[附件清单：path 是 Gateway 内部存储路径，可供文件工具读取]
+[随本条用户消息同时提交的附件。path 是 Gateway 内部存储路径，可供文件工具读取；content_delivery 说明附件内容是否另以模型原生内容块提供。]
 [
   {
     "name": "report.pdf",
+    "extension": "pdf",
     "media_type": "file",
-    "path": "/gateway/media/report.pdf"
+    "mime_type": "application/pdf",
+    "size_bytes": 12345,
+    "path": "/gateway/media/report.pdf",
+    "content_delivery": "content is not embedded in this model request; path remains available to file tools"
   }
 ]
 ```
 
 随后再追加图片等原生多模态 content block。这样支持视觉输入的模型既能看到图片内容，也知道其文件路径；普通文件同样可以由 LLM 使用 `file_read`、Bash 等工具读取。附件路径只进入服务端到 Provider 的模型上下文，不进入 WebSocket/HTTP 客户端响应。
 
+`file_read` 读取 PNG、JPEG、GIF 或 WebP 时不把 Base64 当作工具文本返回，而是通过结构化工具媒体侧通道返回规范化路径。`AgentLoop` 根据当前模型能力构造原生图片块；只有最新连续工具结果批次携带图片内容，旧结果仅保留文本路径。OpenAI-compatible 请求把工具图片汇总为工具批次后的临时 `user` 多模态消息，Anthropic 请求把图片放入对应 `tool_result`。这些 Provider 请求视图不写入历史，消息仍只持久化路径引用。
+
 历史路径已经失效时，清单仍反映消息所记录的原路径；工具读取失败应作为普通、可解释的“文件已移动或删除”结果返回，不能导致 Agent loop panic。
 
 ## 8. WebUI 交互

src/agent/agent_loop.rs

@@ -19,12 +19,31 @@ const MAX_TOOL_RESULT_CHARS: usize = 16_000;
 /// Minimum characters to keep when truncating
 const TRUNCATION_SUFFIX_LEN: usize = 200;
 
+enum MediaOrigin<'a> {
+    User,
+    Tool(&'a str),
+    Message,
+}
+
+fn should_include_message_media(messages: &[ChatMessage], index: usize) -> bool {
+    let message = &messages[index];
+    if message.role != "tool" {
+        return true;
+    }
+    let active_tool_start = messages
+        .iter()
+        .rposition(|candidate| candidate.role != "tool")
+        .map_or(0, |last_non_tool| last_non_tool + 1);
+    index >= active_tool_start
+}
+
 /// Build content blocks from text and media, respecting model input capabilities
 fn build_content_blocks(
     text: &str,
     media_refs: &[MediaRef],
     input_types: &[String],
     registry: &MediaHandlerRegistry,
+    origin: MediaOrigin<'_>,
 ) -> Vec<ContentBlock> {
     let mut blocks = Vec::new();
 
@@ -56,20 +75,30 @@ fn build_content_blocks(
                     "content_delivery": if native_input {
                         "also included as a model-native content block"
                     } else {
-                        "content is not embedded; use a file tool with path to inspect it"
+                        "content is not embedded in this model request; path remains available to file tools"
                     },
                 })
             })
             .collect::<Vec<_>>();
         let manifest = serde_json::Value::Array(attachments).to_string();
+        let note = match origin {
+            MediaOrigin::User if text.is_empty() => {
+                "用户发送了以下附件。path 是 Gateway 内部存储路径，可供文件工具读取；附件内容未必已嵌入模型输入。".to_string()
+            }
+            MediaOrigin::User => {
+                "随本条用户消息同时提交的附件。path 是 Gateway 内部存储路径，可供文件工具读取；content_delivery 说明附件内容是否另以模型原生内容块提供。".to_string()
+            }
+            MediaOrigin::Tool(tool_name) => format!(
+                "工具 {tool_name} 返回了以下附件。content_delivery 说明附件内容是否另以模型原生内容块提供。"
+            ),
+            MediaOrigin::Message => {
+                "本条消息包含以下附件。content_delivery 说明附件内容是否另以模型原生内容块提供。".to_string()
+            }
+        };
         let message_text = if text.is_empty() {
-            format!(
+            format!("[{note}]\n{manifest}")
-                "[用户发送了以下附件。path 是 Gateway 内部存储路径，可供文件工具读取；附件内容未必已嵌入模型输入。]\n{manifest}"
-            )
         } else {
-            format!(
+            format!("{text}\n\n[{note}]\n{manifest}")
-                "{text}\n\n[随本条用户消息同时提交的附件。path 是 Gateway 内部存储路径，可供文件工具读取；content_delivery 说明附件内容是否另以模型原生内容块提供。]\n{manifest}"
-            )
         };
         blocks.push(ContentBlock::text(message_text));
 
@@ -422,15 +451,21 @@ impl AgentLoop {
         &self.tools
     }
 
-    fn chat_message_to_llm_message(&self, m: &ChatMessage) -> Message {
+    fn chat_message_to_llm_message(&self, m: &ChatMessage, include_media: bool) -> Message {
-        let content = if m.media_refs.is_empty() {
+        let content = if m.media_refs.is_empty() || !include_media {
             vec![ContentBlock::text(&m.content)]
         } else {
+            let origin = match m.role.as_str() {
+                "user" => MediaOrigin::User,
+                "tool" => MediaOrigin::Tool(m.tool_name.as_deref().unwrap_or("unknown")),
+                _ => MediaOrigin::Message,
+            };
             build_content_blocks(
                 &m.content,
                 &m.media_refs,
                 &self.input_types,
                 &self.media_registry,
+                origin,
             )
         };
 
@@ -444,6 +479,17 @@ impl AgentLoop {
         }
     }
 
+    fn messages_for_llm(&self, messages: &[ChatMessage]) -> Vec<Message> {
+        messages
+            .iter()
+            .enumerate()
+            .map(|(index, message)| {
+                let include_media = should_include_message_media(messages, index);
+                self.chat_message_to_llm_message(message, include_media)
+            })
+            .collect()
+    }
+
     /// Process a message using the provided conversation history.
     /// History management is handled externally by SessionManager.
     ///
@@ -501,10 +547,7 @@ impl AgentLoop {
             }
 
             // Convert messages to LLM format
-            let messages_for_llm: Vec<Message> = messages
+            let messages_for_llm = self.messages_for_llm(&messages);
-                .iter()
-                .map(|m| self.chat_message_to_llm_message(m))
-                .collect();
 
             // Build request
             let tools = if self.tools.has_tools() {
@@ -601,19 +644,21 @@ impl AgentLoop {
                             "Loop warning: {}",
                             msg
                         );
-                        let tool_message = ChatMessage::tool(
+                        let tool_message = ChatMessage::tool_with_media(
                             tool_call.id.clone(),
                             tool_call.name.clone(),
                             format!("{}\n\n[上一条结果]\n{}", msg, truncated_output),
+                            result.media_refs.clone(),
                         );
                         messages.push(tool_message.clone());
                         emitted_messages.push(tool_message);
                     }
                     LoopDetectionResult::Ok => {
-                        let tool_message = ChatMessage::tool(
+                        let tool_message = ChatMessage::tool_with_media(
                             tool_call.id.clone(),
                             tool_call.name.clone(),
                             truncated_output,
+                            result.media_refs.clone(),
                         );
                         messages.push(tool_message.clone());
                         emitted_messages.push(tool_message);
@@ -641,10 +686,7 @@ impl AgentLoop {
         messages.push(summary_request);
 
         // Convert messages to LLM format
-        let messages_for_llm: Vec<Message> = messages
+        let messages_for_llm = self.messages_for_llm(&messages);
-            .iter()
-            .map(|m| self.chat_message_to_llm_message(m))
-            .collect();
 
         let request = ChatCompletionRequest {
             messages: messages_for_llm,
@@ -784,10 +826,14 @@ impl AgentLoop {
             }
         };
 
-        match tool.execute(tool_call.arguments.clone()).await {
+        match tool.execute_with_media(tool_call.arguments.clone()).await {
-            Ok(result) => {
+            Ok(result_with_media) => {
+                let result = result_with_media.result;
                 if result.success {
-                    ToolExecutionOutcome::success(result.output)
+                    ToolExecutionOutcome::success_with_media(
+                        result.output,
+                        result_with_media.media_refs,
+                    )
                 } else {
                     let error = result.error.unwrap_or_default();
                     ToolExecutionOutcome::failure(format!("Error: {}", error), Some(error))
@@ -805,6 +851,8 @@ impl AgentLoop {
 mod tests {
     use super::*;
     use crate::observability::{MultiObserver, Observer};
+    use crate::providers::{ChatCompletionResponse, Usage};
+    use crate::tools::FileReadTool;
 
     struct TestObserver {
         events: std::sync::Mutex<Vec<ObserverEvent>>,
@@ -844,6 +892,113 @@ mod tests {
         assert_eq!(multi.len(), 1);
     }
 
+    struct ToolMediaProvider {
+        image_path: String,
+        requests: std::sync::Mutex<Vec<ChatCompletionRequest>>,
+    }
+
+    #[async_trait::async_trait]
+    impl LLMProvider for ToolMediaProvider {
+        async fn chat(
+            &self,
+            request: ChatCompletionRequest,
+        ) -> Result<ChatCompletionResponse, Box<dyn std::error::Error + Send + Sync>> {
+            let call_number = {
+                let mut requests = self.requests.lock().unwrap();
+                requests.push(request);
+                requests.len()
+            };
+            Ok(ChatCompletionResponse {
+                id: format!("response-{call_number}"),
+                model: "vision-test".to_string(),
+                content: if call_number == 1 {
+                    String::new()
+                } else {
+                    "image seen".to_string()
+                },
+                reasoning_content: None,
+                tool_calls: if call_number == 1 {
+                    vec![ToolCall {
+                        id: "call-image".to_string(),
+                        name: "file_read".to_string(),
+                        arguments: serde_json::json!({ "path": self.image_path }),
+                    }]
+                } else {
+                    Vec::new()
+                },
+                usage: Usage {
+                    prompt_tokens: 1,
+                    completion_tokens: 1,
+                    total_tokens: 2,
+                    cached_tokens: None,
+                    cache_read_input_tokens: None,
+                    cache_creation_input_tokens: None,
+                },
+            })
+        }
+
+        fn ptype(&self) -> &str {
+            "test"
+        }
+
+        fn name(&self) -> &str {
+            "tool-media-test"
+        }
+
+        fn model_id(&self) -> &str {
+            "vision-test"
+        }
+    }
+
+    #[tokio::test]
+    async fn file_read_media_reaches_the_next_model_iteration() {
+        use std::io::Write;
+
+        let mut image = tempfile::Builder::new().suffix(".png").tempfile().unwrap();
+        image.write_all(b"\x89PNG\r\n\x1a\nminimal").unwrap();
+        let provider = Arc::new(ToolMediaProvider {
+            image_path: image.path().to_string_lossy().into_owned(),
+            requests: std::sync::Mutex::new(Vec::new()),
+        });
+        let tools = Arc::new(ToolRegistry::new());
+        tools.register(FileReadTool::new());
+        let agent = AgentLoop::with_provider_and_tools(
+            provider.clone(),
+            tools,
+            2,
+            "vision-test".to_string(),
+            std::env::current_dir().unwrap(),
+            vec!["text".to_string(), "image".to_string()],
+        );
+
+        let result = agent
+            .process(vec![ChatMessage::user("inspect the image")])
+            .await
+            .unwrap();
+
+        assert_eq!(result.final_response.content, "image seen");
+        let requests = provider.requests.lock().unwrap();
+        assert_eq!(requests.len(), 2);
+        let tool_result = requests[1]
+            .messages
+            .iter()
+            .find(|message| message.role == "tool")
+            .unwrap();
+        assert!(
+            tool_result
+                .content
+                .iter()
+                .any(|block| matches!(block, ContentBlock::ImageUrl { .. }))
+        );
+        assert!(result.emitted_messages.iter().any(|message| {
+            message.role == "tool"
+                && message
+                    .media_refs
+                    .iter()
+                    .any(|media| media.media_type == "image")
+        }));
+    }
+
     #[test]
     fn test_should_execute_in_parallel_single_tool() {
         // Would need a proper setup with AgentLoop to test fully
@@ -904,6 +1059,7 @@ mod tests {
             }],
             &[],
             &registry,
+            MediaOrigin::User,
         );
 
         assert_eq!(blocks.len(), 1);
@@ -912,7 +1068,7 @@ mod tests {
                     && text.contains("随本条用户消息同时提交的附件")
                     && text.contains("missing.png")
                     && text.contains("\"media_type\":\"image\"")
-                    && text.contains("content is not embedded")));
+                    && text.contains("content is not embedded in this model request")));
     }
 
     #[test]
@@ -926,6 +1082,7 @@ mod tests {
             }],
             &[],
             &registry,
+            MediaOrigin::User,
         );
 
         assert_eq!(blocks.len(), 1);
@@ -935,7 +1092,7 @@ mod tests {
                     && text.contains("application/vnd.openxmlformats-officedocument.wordprocessingml.document")
                     && text.contains("\"extension\":\"docx\"")
                     && text.contains("\"size_bytes\":null")
-                    && text.contains("content is not embedded")));
+                    && text.contains("content is not embedded in this model request")));
     }
 
     #[test]
@@ -943,7 +1100,7 @@ mod tests {
         use std::io::Write;
 
         let mut image = tempfile::Builder::new().suffix(".png").tempfile().unwrap();
-        image.write_all(b"image bytes").unwrap();
+        image.write_all(b"\x89PNG\r\n\x1a\nminimal").unwrap();
         let path = image.path().to_string_lossy().into_owned();
         let registry = MediaHandlerRegistry::with_defaults();
 
@@ -955,6 +1112,7 @@ mod tests {
             }],
             &["image".to_string()],
             &registry,
+            MediaOrigin::User,
         );
 
         assert!(matches!(blocks.first(), Some(ContentBlock::Text { text })
@@ -964,6 +1122,79 @@ mod tests {
                     && text.contains("model-native content block")));
         assert!(matches!(blocks.get(1), Some(ContentBlock::ImageUrl { .. })));
     }
+
+    #[test]
+    fn test_build_content_blocks_labels_tool_media_without_user_wording() {
+        use std::io::Write;
+
+        let mut image = tempfile::Builder::new().suffix(".png").tempfile().unwrap();
+        image.write_all(b"\x89PNG\r\n\x1a\nminimal").unwrap();
+        let path = image.path().to_string_lossy().into_owned();
+        let registry = MediaHandlerRegistry::with_defaults();
+
+        let blocks = build_content_blocks(
+            "Image file ready for visual inspection.",
+            &[MediaRef {
+                path,
+                media_type: "image".to_string(),
+            }],
+            &["image".to_string()],
+            &registry,
+            MediaOrigin::Tool("file_read"),
+        );
+
+        assert!(matches!(blocks.first(), Some(ContentBlock::Text { text })
+                if text.contains("工具 file_read 返回了以下附件")
+                    && !text.contains("用户消息")));
+        assert!(matches!(blocks.get(1), Some(ContentBlock::ImageUrl { .. })));
+    }
+
+    #[test]
+    fn only_the_trailing_tool_batch_replays_tool_media() {
+        let mut messages = vec![
+            ChatMessage::assistant_with_tool_calls(
+                "",
+                vec![ToolCall {
+                    id: "old-call".to_string(),
+                    name: "file_read".to_string(),
+                    arguments: serde_json::json!({}),
+                }],
+            ),
+            ChatMessage::tool_with_media(
+                "old-call",
+                "file_read",
+                "old",
+                vec![MediaRef {
+                    path: "/tmp/old.png".to_string(),
+                    media_type: "image".to_string(),
+                }],
+            ),
+            ChatMessage::assistant("continue"),
+            ChatMessage::assistant_with_tool_calls(
+                "",
+                vec![ToolCall {
+                    id: "new-call".to_string(),
+                    name: "file_read".to_string(),
+                    arguments: serde_json::json!({}),
+                }],
+            ),
+            ChatMessage::tool_with_media(
+                "new-call",
+                "file_read",
+                "new",
+                vec![MediaRef {
+                    path: "/tmp/new.png".to_string(),
+                    media_type: "image".to_string(),
+                }],
+            ),
+        ];
+
+        assert!(!should_include_message_media(&messages, 1));
+        assert!(should_include_message_media(&messages, 4));
+
+        messages.push(ChatMessage::user("next turn"));
+        assert!(!should_include_message_media(&messages, 4));
+    }
 }
 
 #[derive(Debug)]

src/agent/media_handler.rs

@@ -4,6 +4,8 @@ use std::io::Read;
 
 use crate::bus::message::ContentBlock;
 
+const MAX_NATIVE_IMAGE_BYTES: u64 = 10 * 1024 * 1024;
+
 pub trait MediaHandler: Send + Sync {
     fn media_type(&self) -> &str;
     fn handle(&self, path: &str) -> Result<Vec<ContentBlock>, MediaHandlerError>;
@@ -41,28 +43,60 @@ impl MediaHandler for ImageHandler {
     }
 
     fn handle(&self, path: &str) -> Result<Vec<ContentBlock>, MediaHandlerError> {
-        let (mime_type, base64_data) =
+        let (mime_type, base64_data) = encode_image_to_base64(path)?;
-            encode_image_to_base64(path).map_err(MediaHandlerError::Io)?;
         let url = format!("data:{};base64,{}", mime_type, base64_data);
         Ok(vec![ContentBlock::image_url(url)])
     }
 }
 
-fn encode_image_to_base64(path: &str) -> Result<(String, String), std::io::Error> {
+fn encode_image_to_base64(path: &str) -> Result<(String, String), MediaHandlerError> {
     use base64::{Engine as _, engine::general_purpose::STANDARD};
 
-    let mut file = std::fs::File::open(path)?;
+    let metadata = std::fs::metadata(path).map_err(MediaHandlerError::Io)?;
-    let mut buffer = Vec::new();
+    if metadata.len() > MAX_NATIVE_IMAGE_BYTES {
-    file.read_to_end(&mut buffer)?;
+        return Err(MediaHandlerError::UnsupportedFormat(format!(
+            "image is too large: {} bytes (max {} bytes)",
+            metadata.len(),
+            MAX_NATIVE_IMAGE_BYTES
+        )));
+    }
 
     let mime = mime_guess::from_path(path)
         .first_or_octet_stream()
         .to_string();
+    if !matches!(
+        mime.as_str(),
+        "image/png" | "image/jpeg" | "image/gif" | "image/webp"
+    ) {
+        return Err(MediaHandlerError::UnsupportedFormat(format!(
+            "unsupported image MIME type: {mime}"
+        )));
+    }
+
+    let mut file = std::fs::File::open(path).map_err(MediaHandlerError::Io)?;
+    let mut buffer = Vec::new();
+    file.read_to_end(&mut buffer)
+        .map_err(MediaHandlerError::Io)?;
+    if !valid_image_signature(&buffer, &mime) {
+        return Err(MediaHandlerError::UnsupportedFormat(format!(
+            "file content does not match image MIME type: {mime}"
+        )));
+    }
 
     let encoded = STANDARD.encode(&buffer);
     Ok((mime, encoded))
 }
 
+fn valid_image_signature(bytes: &[u8], mime: &str) -> bool {
+    match mime {
+        "image/png" => bytes.starts_with(b"\x89PNG\r\n\x1a\n"),
+        "image/jpeg" => bytes.starts_with(&[0xff, 0xd8, 0xff]),
+        "image/gif" => bytes.starts_with(b"GIF87a") || bytes.starts_with(b"GIF89a"),
+        "image/webp" => bytes.len() >= 12 && bytes.starts_with(b"RIFF") && &bytes[8..12] == b"WEBP",
+        _ => false,
+    }
+}
+
 pub struct MediaHandlerRegistry {
     handlers: HashMap<String, Box<dyn MediaHandler>>,
 }

src/bus/message.rs

@@ -215,13 +215,22 @@ impl ChatMessage {
         tool_call_id: impl Into<String>,
         tool_name: impl Into<String>,
         content: impl Into<String>,
+    ) -> Self {
+        Self::tool_with_media(tool_call_id, tool_name, content, Vec::new())
+    }
+
+    pub fn tool_with_media(
+        tool_call_id: impl Into<String>,
+        tool_name: impl Into<String>,
+        content: impl Into<String>,
+        media_refs: Vec<MediaRef>,
     ) -> Self {
         Self {
             id: uuid::Uuid::new_v4().to_string(),
             role: "tool".to_string(),
             content: content.into(),
             reasoning_content: None,
-            media_refs: Vec::new(),
+            media_refs,
             timestamp: current_timestamp(),
             tool_call_id: Some(tool_call_id.into()),
             tool_name: Some(tool_name.into()),

src/observability/mod.rs

@@ -5,6 +5,8 @@
 
 use std::time::Duration;
 
+use crate::bus::MediaRef;
+
 /// Events emitted during agent and tool execution.
 #[derive(Debug, Clone)]
 pub enum ObserverEvent {
@@ -57,6 +59,8 @@ pub struct ToolExecutionOutcome {
     pub error_reason: Option<String>,
     /// How long the tool took to execute.
     pub duration: Duration,
+    /// Structured media returned by the tool for the next model iteration.
+    pub media_refs: Vec<MediaRef>,
 }
 
 impl ToolExecutionOutcome {
@@ -67,6 +71,18 @@ impl ToolExecutionOutcome {
             success: true,
             error_reason: None,
             duration: Duration::ZERO,
+            media_refs: Vec::new(),
+        }
+    }
+
+    /// Create a successful outcome carrying structured media artifacts.
+    pub fn success_with_media(output: String, media_refs: Vec<MediaRef>) -> Self {
+        Self {
+            output,
+            success: true,
+            error_reason: None,
+            duration: Duration::ZERO,
+            media_refs,
         }
     }
 
@@ -77,6 +93,7 @@ impl ToolExecutionOutcome {
             success: false,
             error_reason,
             duration: Duration::ZERO,
+            media_refs: Vec::new(),
         }
     }
 }

src/providers/anthropic.rs

@@ -5,7 +5,7 @@ use std::collections::HashMap;
 use std::time::Duration;
 
 use super::traits::Usage;
-use super::{ChatCompletionRequest, ChatCompletionResponse, LLMProvider, Tool, ToolCall};
+use super::{ChatCompletionRequest, ChatCompletionResponse, LLMProvider, Message, Tool, ToolCall};
 use crate::bus::message::ContentBlock;
 use crate::storage::Storage;
 use std::sync::Arc;
@@ -139,6 +139,44 @@ struct AnthropicMessage {
     content: Vec<serde_json::Value>,
 }
 
+fn convert_messages(messages: &[Message]) -> Vec<AnthropicMessage> {
+    messages
+        .iter()
+        .map(|message| {
+            let role = if message.role == "tool" {
+                "user".to_string()
+            } else {
+                message.role.clone()
+            };
+            let content = if let Some(ref tool_call_id) = message.tool_call_id {
+                vec![serde_json::json!({
+                    "type": "tool_result",
+                    "tool_use_id": tool_call_id,
+                    "content": convert_content_blocks(&message.content, false),
+                })]
+            } else {
+                let mut blocks = convert_content_blocks(&message.content, message.role == "system");
+                if let Some(tool_calls) = message
+                    .tool_calls
+                    .as_ref()
+                    .filter(|calls| !calls.is_empty())
+                {
+                    for tool_call in tool_calls {
+                        blocks.push(serde_json::json!({
+                            "type": "tool_use",
+                            "id": tool_call.id,
+                            "name": tool_call.name,
+                            "input": tool_call.arguments,
+                        }));
+                    }
+                }
+                blocks
+            };
+            AnthropicMessage { role, content }
+        })
+        .collect()
+}
+
 #[derive(Serialize)]
 struct AnthropicTool {
     name: String,
@@ -216,50 +254,7 @@ impl LLMProvider for AnthropicProvider {
 
         let body = AnthropicRequest {
             model: self.model_id.clone(),
-            messages: request
+            messages: convert_messages(&request.messages),
-                .messages
-                .iter()
-                .map(|m| {
-                    let role = if m.role == "tool" {
-                        // Anthropic uses "user" role for tool result messages
-                        "user".to_string()
-                    } else {
-                        m.role.clone()
-                    };
-                    let content = if let Some(ref tc_id) = m.tool_call_id {
-                        // Tool result: wrap as tool_result content block
-                        let output = m
-                            .content
-                            .iter()
-                            .filter_map(|b| match b {
-                                ContentBlock::Text { text } => Some(text.as_str()),
-                                _ => None,
-                            })
-                            .collect::<Vec<_>>()
-                            .join("");
-                        vec![serde_json::json!({
-                            "type": "tool_result",
-                            "tool_use_id": tc_id,
-                            "content": output,
-                        })]
-                    } else {
-                        let mut blocks = convert_content_blocks(&m.content, m.role == "system");
-                        // Append tool_use blocks from assistant messages with tool calls
-                        if let Some(tool_calls) = m.tool_calls.as_ref().filter(|c| !c.is_empty()) {
-                            for tc in tool_calls {
-                                blocks.push(serde_json::json!({
-                                    "type": "tool_use",
-                                    "id": tc.id,
-                                    "name": tc.name,
-                                    "input": tc.arguments,
-                                }));
-                            }
-                        }
-                        blocks
-                    };
-                    AnthropicMessage { role, content }
-                })
-                .collect(),
             max_tokens,
             temperature: request.temperature.or(self.temperature),
             tools,
@@ -482,4 +477,29 @@ mod tests {
         let value = serde_json::to_value(tool).unwrap();
         assert_eq!(value["cache_control"]["type"], "ephemeral");
     }
+
+    #[test]
+    fn tool_result_preserves_native_image_blocks() {
+        let messages = vec![Message {
+            role: "tool".to_string(),
+            content: vec![
+                ContentBlock::text("image ready"),
+                ContentBlock::image_url("data:image/png;base64,AAAA"),
+            ],
+            reasoning_content: None,
+            tool_call_id: Some("call_1".to_string()),
+            name: Some("file_read".to_string()),
+            tool_calls: None,
+        }];
+
+        let converted = convert_messages(&messages);
+        let result = &converted[0].content[0];
+
+        assert_eq!(converted[0].role, "user");
+        assert_eq!(result["type"], "tool_result");
+        assert_eq!(result["content"][0]["type"], "text");
+        assert_eq!(result["content"][1]["type"], "image");
+        assert_eq!(result["content"][1]["source"]["media_type"], "image/png");
+        assert_eq!(result["content"][1]["source"]["data"], "AAAA");
+    }
 }

src/providers/openai.rs

@@ -6,7 +6,7 @@ use std::collections::HashMap;
 use std::time::Duration;
 
 use super::traits::Usage;
-use super::{ChatCompletionRequest, ChatCompletionResponse, LLMProvider, ToolCall};
+use super::{ChatCompletionRequest, ChatCompletionResponse, LLMProvider, Message, ToolCall};
 use crate::bus::message::ContentBlock;
 use crate::storage::Storage;
 use std::sync::Arc;
@@ -32,6 +32,116 @@ fn convert_content_blocks(blocks: &[ContentBlock]) -> Value {
     )
 }
 
+fn text_content(blocks: &[ContentBlock]) -> String {
+    blocks
+        .iter()
+        .filter_map(|block| match block {
+            ContentBlock::Text { text } => Some(text.as_str()),
+            ContentBlock::ImageUrl { .. } => None,
+        })
+        .collect::<Vec<_>>()
+        .join("\n")
+}
+
+fn regular_message_json(message: &Message) -> Value {
+    if message.role == "tool" {
+        json!({
+            "role": message.role,
+            "content": text_content(&message.content),
+            "tool_call_id": message.tool_call_id,
+            "name": message.name,
+        })
+    } else if message.role == "assistant"
+        && message
+            .tool_calls
+            .as_ref()
+            .is_some_and(|calls| !calls.is_empty())
+    {
+        let mut value = json!({
+            "role": message.role,
+            "content": convert_content_blocks(&message.content),
+            "tool_calls": message.tool_calls.as_ref().map(|calls| {
+                calls.iter().map(|call| json!({
+                    "id": call.id,
+                    "type": "function",
+                    "function": {
+                        "name": call.name,
+                        "arguments": serde_json::to_string(&call.arguments).unwrap_or_else(|_| "null".to_string())
+                    }
+                })).collect::<Vec<_>>()
+            })
+        });
+        if let Some(ref reasoning_content) = message.reasoning_content {
+            value["reasoning_content"] = json!(reasoning_content);
+        }
+        value
+    } else {
+        let mut value = json!({
+            "role": message.role,
+            "content": convert_content_blocks(&message.content)
+        });
+        if message.role == "assistant"
+            && let Some(ref reasoning_content) = message.reasoning_content
+        {
+            value["reasoning_content"] = json!(reasoning_content);
+        }
+        value
+    }
+}
+
+/// OpenAI-compatible APIs do not consistently accept image parts on `tool`
+/// messages. Keep the tool-call contract textual and append one temporary user
+/// message containing media returned by the complete contiguous tool batch.
+fn convert_messages(messages: &[Message]) -> Vec<Value> {
+    let mut converted = Vec::with_capacity(messages.len());
+    let mut tool_media = Vec::new();
+    let mut tool_media_labels = Vec::new();
+
+    let flush_tool_media =
+        |converted: &mut Vec<Value>, tool_media: &mut Vec<Value>, labels: &mut Vec<String>| {
+            if tool_media.is_empty() {
+                return;
+            }
+            let mut content = vec![json!({
+                "type": "text",
+                "text": format!("[以下媒体由本轮工具调用返回]\n{}", labels.join("\n")),
+            })];
+            content.append(tool_media);
+            converted.push(json!({
+                "role": "user",
+                "content": content,
+            }));
+            labels.clear();
+        };
+
+    for message in messages {
+        if message.role != "tool" {
+            flush_tool_media(&mut converted, &mut tool_media, &mut tool_media_labels);
+        }
+        converted.push(regular_message_json(message));
+        if message.role == "tool" {
+            for block in &message.content {
+                if let ContentBlock::ImageUrl { image_url } = block {
+                    let label = format!(
+                        "- {} / {}",
+                        message.name.as_deref().unwrap_or("tool"),
+                        message.tool_call_id.as_deref().unwrap_or("unknown")
+                    );
+                    if tool_media_labels.last() != Some(&label) {
+                        tool_media_labels.push(label);
+                    }
+                    tool_media.push(json!({
+                        "type": "image_url",
+                        "image_url": { "url": image_url.url },
+                    }));
+                }
+            }
+        }
+    }
+    flush_tool_media(&mut converted, &mut tool_media, &mut tool_media_labels);
+    converted
+}
+
 pub struct OpenAIProvider {
     client: Client,
     name: String,
@@ -83,46 +193,7 @@ impl OpenAIProvider {
     fn build_request_body(&self, request: &ChatCompletionRequest) -> Value {
         let mut body = json!({
             "model": self.model_id,
-            "messages": request.messages.iter().map(|m| {
+            "messages": convert_messages(&request.messages),
-                if m.role == "tool" {
-                    json!({
-                        "role": m.role,
-                        "content": convert_content_blocks(&m.content),
-                        "tool_call_id": m.tool_call_id,
-                        "name": m.name,
-                    })
-                } else if m.role == "assistant" && m.tool_calls.as_ref().is_some_and(|c| !c.is_empty()) {
-                    let mut msg = json!({
-                        "role": m.role,
-                        "content": convert_content_blocks(&m.content),
-                        "tool_calls": m.tool_calls.as_ref().map(|calls| {
-                            calls.iter().map(|call| json!({
-                                "id": call.id,
-                                "type": "function",
-                                "function": {
-                                    "name": call.name,
-                                    "arguments": serde_json::to_string(&call.arguments).unwrap_or_else(|_| "null".to_string())
-                                }
-                            })).collect::<Vec<_>>()
-                        })
-                    });
-                    if let Some(ref rc) = m.reasoning_content {
-                        msg["reasoning_content"] = json!(rc);
-                    }
-                    msg
-                } else {
-                    let mut msg = json!({
-                        "role": m.role,
-                        "content": convert_content_blocks(&m.content)
-                    });
-                    if m.role == "assistant"
-                        && let Some(ref rc) = m.reasoning_content
-                    {
-                        msg["reasoning_content"] = json!(rc);
-                    }
-                    msg
-                }
-            }).collect::<Vec<_>>(),
             "temperature": request.temperature.or(self.temperature).unwrap_or(0.7),
             "max_tokens": request.max_tokens.or(self.max_tokens),
         });
@@ -450,6 +521,37 @@ mod tests {
         );
     }
 
+    #[test]
+    fn tool_images_are_appended_after_the_complete_tool_batch() {
+        let messages = vec![
+            Message::tool("call_1", "file_read", "first image"),
+            Message {
+                role: "tool".to_string(),
+                content: vec![
+                    ContentBlock::text("second image"),
+                    ContentBlock::image_url("data:image/png;base64,AAAA"),
+                ],
+                reasoning_content: None,
+                tool_call_id: Some("call_2".to_string()),
+                name: Some("file_read".to_string()),
+                tool_calls: None,
+            },
+        ];
+
+        let converted = convert_messages(&messages);
+
+        assert_eq!(converted.len(), 3);
+        assert_eq!(converted[0]["role"], "tool");
+        assert_eq!(converted[1]["role"], "tool");
+        assert_eq!(converted[2]["role"], "user");
+        assert_eq!(converted[2]["content"][1]["type"], "image_url");
+        assert_eq!(
+            converted[2]["content"][1]["image_url"]["url"],
+            "data:image/png;base64,AAAA"
+        );
+        assert_eq!(converted[1]["content"], "second image");
+    }
+
     #[test]
     fn test_decode_response_accepts_null_tool_calls() {
         let text = r#"{

src/tools/file_read.rs

@@ -1,9 +1,11 @@
 use async_trait::async_trait;
 use encoding_rs::*;
 use serde_json::json;
+use std::io::Read;
 
+use crate::bus::MediaRef;
 use crate::tools::path_utils;
-use crate::tools::traits::{Tool, ToolResult};
+use crate::tools::traits::{Tool, ToolResult, ToolResultWithMedia};
 
 const MAX_CHARS: usize = 128_000;
 const MAX_FILE_BYTES: u64 = 5 * 1024 * 1024;
@@ -39,7 +41,7 @@ impl Tool for FileReadTool {
     }
 
     fn description(&self) -> &str {
-        "Read the contents of a file. Returns numbered lines. Use offset and limit to paginate through large files."
+        "Read a file. Text returns numbered lines with offset/limit pagination; supported images are provided as visual media to vision-capable models."
     }
 
     fn parameters_schema(&self) -> serde_json::Value {
@@ -266,6 +268,111 @@ impl Tool for FileReadTool {
             }
         }
     }
+
+    async fn execute_with_media(
+        &self,
+        args: serde_json::Value,
+    ) -> anyhow::Result<ToolResultWithMedia> {
+        if let Some(image_result) = self.inspect_image_result(&args) {
+            return Ok(image_result);
+        }
+        self.execute(args).await.map(Into::into)
+    }
+}
+
+impl FileReadTool {
+    fn inspect_image_result(&self, args: &serde_json::Value) -> Option<ToolResultWithMedia> {
+        let path = args.get("path")?.as_str()?;
+        let resolved = path_utils::resolve_path(path, self.allowed_dir.as_deref()).ok()?;
+        let mime = mime_guess::from_path(&resolved)
+            .first_or_octet_stream()
+            .to_string();
+        if !is_supported_image_mime(&mime) {
+            return None;
+        }
+
+        let failure = |error: String| ToolResultWithMedia {
+            result: ToolResult {
+                success: false,
+                output: String::new(),
+                error: Some(error),
+            },
+            media_refs: Vec::new(),
+        };
+
+        if !resolved.exists() {
+            return Some(failure(format!("File not found: {path}")));
+        }
+        if !resolved.is_file() {
+            return Some(failure(format!("Not a file: {path}")));
+        }
+        let metadata = match std::fs::metadata(&resolved) {
+            Ok(metadata) => metadata,
+            Err(error) => {
+                return Some(failure(format!("Failed to inspect file: {error}")));
+            }
+        };
+        if metadata.len() > MAX_FILE_BYTES {
+            return Some(failure(format!(
+                "File too large to read safely: {} bytes (max {} bytes).",
+                metadata.len(),
+                MAX_FILE_BYTES
+            )));
+        }
+        if let Err(error) = validate_image_signature(&resolved, &mime) {
+            return Some(failure(error));
+        }
+
+        let canonical = std::fs::canonicalize(&resolved).unwrap_or(resolved);
+        let canonical_path = canonical.to_string_lossy().into_owned();
+        Some(ToolResultWithMedia {
+            result: ToolResult {
+                success: true,
+                output: format!(
+                    "Image file ready for visual inspection.\nPath: {canonical_path}\nMIME: {mime}\nSize: {} bytes",
+                    metadata.len()
+                ),
+                error: None,
+            },
+            media_refs: vec![MediaRef {
+                path: canonical_path,
+                media_type: "image".to_string(),
+            }],
+        })
+    }
+}
+
+fn is_supported_image_mime(mime: &str) -> bool {
+    matches!(
+        mime,
+        "image/png" | "image/jpeg" | "image/gif" | "image/webp"
+    )
+}
+
+fn validate_image_signature(path: &std::path::Path, mime: &str) -> Result<(), String> {
+    let mut file =
+        std::fs::File::open(path).map_err(|error| format!("Failed to open image file: {error}"))?;
+    let mut header = [0_u8; 12];
+    let read = file
+        .read(&mut header)
+        .map_err(|error| format!("Failed to inspect image file: {error}"))?;
+    let header = &header[..read];
+    let valid = match mime {
+        "image/png" => header.starts_with(b"\x89PNG\r\n\x1a\n"),
+        "image/jpeg" => header.starts_with(&[0xff, 0xd8, 0xff]),
+        "image/gif" => header.starts_with(b"GIF87a") || header.starts_with(b"GIF89a"),
+        "image/webp" => {
+            header.len() >= 12 && header.starts_with(b"RIFF") && &header[8..12] == b"WEBP"
+        }
+        _ => false,
+    };
+    if valid {
+        Ok(())
+    } else {
+        Err(format!(
+            "File content does not match the expected image format: {mime}"
+        ))
+    }
 }
 
 fn decode_text(bytes: &[u8]) -> (Option<String>, Option<&'static str>) {
@@ -414,4 +521,48 @@ mod tests {
         assert!(!result.success);
         assert!(result.error.unwrap().contains("Binary file too large"));
     }
+
+    #[tokio::test]
+    async fn image_read_returns_structured_media_without_base64_text() {
+        let mut file = tempfile::Builder::new().suffix(".png").tempfile().unwrap();
+        file.write_all(b"\x89PNG\r\n\x1a\nminimal").unwrap();
+
+        let result = FileReadTool::new()
+            .execute_with_media(json!({ "path": file.path() }))
+            .await
+            .unwrap();
+
+        assert!(result.result.success);
+        assert_eq!(result.media_refs.len(), 1);
+        assert_eq!(result.media_refs[0].media_type, "image");
+        assert_eq!(
+            result.media_refs[0].path,
+            std::fs::canonicalize(file.path())
+                .unwrap()
+                .to_string_lossy()
+        );
+        assert!(result.result.output.contains("MIME: image/png"));
+        assert!(!result.result.output.contains("base64"));
+    }
+
+    #[tokio::test]
+    async fn image_read_rejects_mismatched_file_signature() {
+        let mut file = tempfile::Builder::new().suffix(".png").tempfile().unwrap();
+        file.write_all(b"not a png").unwrap();
+
+        let result = FileReadTool::new()
+            .execute_with_media(json!({ "path": file.path() }))
+            .await
+            .unwrap();
+
+        assert!(!result.result.success);
+        assert!(result.media_refs.is_empty());
+        assert!(
+            result
+                .result
+                .error
+                .unwrap()
+                .contains("does not match the expected image format")
+        );
+    }
 }

src/tools/mod.rs

@@ -41,7 +41,7 @@ pub use pty::{PtyManager, PtyTool};
 pub use registry::ToolRegistry;
 pub use send_message::SendMessageTool;
 pub use todo::TodoTool;
-pub use traits::{OutboundMessenger, Tool, ToolResult};
+pub use traits::{OutboundMessenger, Tool, ToolResult, ToolResultWithMedia};
 pub use web_fetch::WebFetchTool;
 
 use crate::agent::SubAgentManager;

src/tools/traits.rs

@@ -1,4 +1,4 @@
-use crate::bus::{MediaItem, MessageSource};
+use crate::bus::{MediaItem, MediaRef, MessageSource};
 use async_trait::async_trait;
 
 #[derive(Debug, Clone)]
@@ -8,6 +8,23 @@ pub struct ToolResult {
     pub error: Option<String>,
 }
 
+/// A tool result plus media artifacts that should be made available to a
+/// capable model on the next agent iteration.
+#[derive(Debug, Clone)]
+pub struct ToolResultWithMedia {
+    pub result: ToolResult,
+    pub media_refs: Vec<MediaRef>,
+}
+
+impl From<ToolResult> for ToolResultWithMedia {
+    fn from(result: ToolResult) -> Self {
+        Self {
+            result,
+            media_refs: Vec::new(),
+        }
+    }
+}
+
 #[async_trait]
 pub trait Tool: Send + Sync + 'static {
     fn name(&self) -> &str;
@@ -15,6 +32,15 @@ pub trait Tool: Send + Sync + 'static {
     fn parameters_schema(&self) -> serde_json::Value;
     async fn execute(&self, args: serde_json::Value) -> anyhow::Result<ToolResult>;
 
+    /// Execute the tool and return structured media artifacts when applicable.
+    /// Most tools only return text and use this default implementation.
+    async fn execute_with_media(
+        &self,
+        args: serde_json::Value,
+    ) -> anyhow::Result<ToolResultWithMedia> {
+        self.execute(args).await.map(Into::into)
+    }
+
     /// Whether this tool is side-effect free and safe to parallelize.
     fn read_only(&self) -> bool {
         false