feat(agent): retain image context across turns for follow-ups (#303)

mateclaw-server/src/main/java/vip/mate/agent/BaseAgent.java

@@ -955,6 +955,11 @@ public abstract class BaseAgent {
         if (decision.strategy() == MultimodalRoutingDecision.Strategy.SIDECAR
                 && mediaCaptionService != null
                 && decision.sidecarModel() != null) {
+            // The user's actual question (text parts only, excluding media markers)
+            // so the vision model tailors its description to what was asked rather
+            // than emitting a generic caption.
+            String userQuestion = extractUserQuestion(parts);
+            boolean captionPersisted = false;
             for (MessageContentPart part : parts) {
                 if (part == null) continue;
                 String contentType = part.getContentType();
@@ -963,13 +968,23 @@ public abstract class BaseAgent {
                         && !contentType.contains("svg");
                 if (!isImage) continue;
                 MediaCaptionService.CaptionResult result = mediaCaptionService.caption(
-                        decision.sidecarModel(), part, userLocale);
+                        decision.sidecarModel(), part, userLocale, userQuestion);
                 if (result.isFailure()) {
                     log.warn("[{}] Sidecar caption failed for {}: {}",
                             agentName, part.getFileName(), result.failure().getMessage());
-                    textBuilder.append("\n\n[系统提示] 视觉模型未能解析附件 ")
-                            .append(part.getFileName())
-                            .append("，请稍后重试或在「设置 → 模型」检查视觉模型配置。");
+                    if (isRemoteOnlyAttachment(part)) {
+                        // The image was never downloaded locally (only a remote
+                        // channel URL survives) — for WeCom/aibot that URL points
+                        // at short-lived AES-encrypted bytes, so captioning can
+                        // never succeed until media download is enabled.
+                        textBuilder.append("\n\n[系统提示] 图片 ")
+                                .append(part.getFileName())
+                                .append(" 未下载到本地，无法识别；请在「设置 → 渠道」开启该渠道的媒体下载。");
+                    } else {
+                        textBuilder.append("\n\n[系统提示] 视觉模型未能解析附件 ")
+                                .append(part.getFileName())
+                                .append("，请稍后重试或在「设置 → 模型」检查视觉模型配置。");
+                    }
                     continue;
                 }
                 textBuilder.append("\n\n[图片附件描述: ")
@@ -977,9 +992,17 @@ public abstract class BaseAgent {
                         .append("]\n")
                         .append(result.description())
                         .append("\n[/图片附件描述]");
+                // Persist the caption onto the part so later turns retain the image
+                // content: history user messages replay as text only, and without a
+                // stored caption every follow-up question loses the attachment.
+                part.setCaption(result.description());
+                captionPersisted = true;
                 String identifier = identifyPart(part);
                 if (identifier != null) sidecarHandledIdentifiers.add(identifier);
             }
+            if (captionPersisted && conversationService != null) {
+                conversationService.updateMessageParts(message, parts);
+            }
         }
 
         List<Media> mediaList = new ArrayList<>();
@@ -1048,7 +1071,7 @@ public abstract class BaseAgent {
             if (mediaPath == null) {
                 log.warn("[{}] {} file not found for attachment: {}, path: {}, mediaId: {}",
                         agentName, isVideo ? "Video" : "Image", part.getFileName(), part.getPath(), part.getMediaId());
-                skippedAttachments.add(part.getFileName() + "(文件未找到)");
+                skippedAttachments.add(part.getFileName() + unresolvedAttachmentReason(part));
                 continue;
             }
             try {
@@ -1083,6 +1106,52 @@ public abstract class BaseAgent {
         return new CurrentTurnUserMessage(built, decision);
     }
 
+    /**
+     * Concatenate the text parts of a message into the user's question, dropping
+     * image/file/media parts. Returns {@code null} when there is no usable text
+     * (e.g. an image-only IM message), which makes the caption fall back to the
+     * generic full-description prompt.
+     */
+    private static String extractUserQuestion(List<MessageContentPart> parts) {
+        if (parts == null || parts.isEmpty()) return null;
+        StringBuilder sb = new StringBuilder();
+        for (MessageContentPart part : parts) {
+            if (part == null || !"text".equals(part.getType())) continue;
+            String text = part.getText();
+            if (text == null || text.isBlank()) continue;
+            if (sb.length() > 0) sb.append('\n');
+            sb.append(text.trim());
+        }
+        String question = sb.toString().trim();
+        return question.isEmpty() ? null : question;
+    }
+
+    /**
+     * True when an attachment has no resolvable local file and its only locator
+     * is a remote http(s) URL — i.e. the IM channel never downloaded it locally.
+     * For WeCom/aibot images that URL points at short-lived AES-encrypted bytes,
+     * so it is unusable as-is. Lets callers turn a generic "file not found" into
+     * an actionable hint instead of a dead end.
+     */
+    private static boolean isRemoteOnlyAttachment(MessageContentPart part) {
+        if (part == null) return false;
+        if (part.getPath() != null && !part.getPath().isBlank()) return false;
+        String locator = part.getMediaId();
+        if (locator == null || locator.isBlank()) locator = part.getFileUrl();
+        return locator != null && (locator.startsWith("http://") || locator.startsWith("https://"));
+    }
+
+    /**
+     * Reason string appended to a skipped attachment whose local file could not
+     * be resolved — distinguishes "never downloaded" (channel media download
+     * off) from a genuine missing-file so the user gets an actionable message.
+     */
+    private static String unresolvedAttachmentReason(MessageContentPart part) {
+        return isRemoteOnlyAttachment(part)
+                ? "(图片未下载到本地，无法识别；请在「设置 → 渠道」开启该渠道的媒体下载)"
+                : "(文件未找到)";
+    }
+
     /**
      * Stable identifier for de-duplicating parts already handled by the sidecar
      * pass. Falls back across {@code path → mediaId → fileName} since not every
@@ -1194,7 +1263,14 @@ public abstract class BaseAgent {
                 if ("user".equals(msg.getRole())) {
                     // 用 DB 中的实际内容（可能包含 contentParts），不用传入的 text
                     String content = conversationService.renderMessageContent(msg);
-                    return buildUserMessageForCurrentTurn(msg, content != null && !content.isBlank() ? content : userMessageText);
+                    CurrentTurnUserMessage built = buildUserMessageForCurrentTurn(
+                            msg, content != null && !content.isBlank() ? content : userMessageText);
+                    // Vision-capable models replay history as text only, so a
+                    // follow-up question about an earlier image would otherwise be
+                    // answered blind. Re-attach the most recent image to this turn
+                    // so the model actually re-sees it. (Text-only models instead
+                    // rely on the persisted sidecar caption — see buildUserMessageInternal.)
+                    return maybeCarryRecentImage(history, i, msg, built);
                 }
             }
         } catch (Exception e) {
@@ -1204,6 +1280,85 @@ public abstract class BaseAgent {
         return new CurrentTurnUserMessage(new UserMessage(userMessageText), null);
     }
 
+    /** How far back (in messages) to look for an image to carry into a follow-up turn. */
+    private static final int CARRY_IMAGE_LOOKBACK = 8;
+
+    /**
+     * For a vision-capable model, re-attach the most recent image from a recent
+     * earlier turn to the current user message when the current turn carries no
+     * image of its own. History is replayed as text only (see {@link #toSpringMessage}),
+     * so without this a follow-up like "what's in the top-left of that photo?" is
+     * answered blind. Bounded to a single image within {@link #CARRY_IMAGE_LOOKBACK}
+     * messages so a long conversation doesn't re-send pixels on every turn.
+     *
+     * <p>No-op (returns {@code built} unchanged) when: the model can't see images,
+     * the current turn already has an image, no recent image exists, or the recent
+     * image has no resolvable local file (e.g. an undownloaded channel URL).
+     */
+    private CurrentTurnUserMessage maybeCarryRecentImage(List<MessageEntity> history, int currentIdx,
+                                                         MessageEntity currentMsg, CurrentTurnUserMessage built) {
+        try {
+            if (built == null || !modelSupportsVision()) return built;
+            if (messageHasImagePart(currentMsg)) return built; // current turn already carries an image
+
+            int from = Math.max(0, currentIdx - CARRY_IMAGE_LOOKBACK);
+            for (int j = currentIdx - 1; j >= from; j--) {
+                MessageEntity m = history.get(j);
+                if (m == null || !"user".equals(m.getRole())) continue;
+                List<MessageContentPart> parts = conversationService.parseMessageParts(m);
+                for (int k = parts.size() - 1; k >= 0; k--) {
+                    MessageContentPart part = parts.get(k);
+                    if (!isResolvableImagePart(part)) continue;
+                    Path imgPath = resolveImagePath(part.getPath());
+                    if (imgPath == null && part.getMediaId() != null) imgPath = resolveImagePath(part.getMediaId());
+                    if (imgPath == null) continue;
+                    String contentType = part.getContentType();
+                    if (contentType == null || "image/*".equals(contentType)) contentType = "image/jpeg";
+                    try {
+                        Media carried = new Media(MimeType.valueOf(contentType), new FileSystemResource(imgPath));
+                        UserMessage orig = built.userMessage();
+                        String name = part.getFileName() == null ? "image" : part.getFileName();
+                        String text = (orig.getText() == null ? "" : orig.getText())
+                                + "\n\n[系统提示] 以下图片是用户本次对话中较早发送的「" + name
+                                + "」，当前问题很可能与它相关。请直接查看该图片作答，不要凭记忆猜测。";
+                        List<Media> media = new ArrayList<>();
+                        if (orig.getMedia() != null) media.addAll(orig.getMedia());
+                        media.add(carried);
+                        log.debug("[{}] Carried recent image {} into follow-up turn for vision model",
+                                agentName, name);
+                        return new CurrentTurnUserMessage(
+                                UserMessage.builder().text(text).media(media).build(),
+                                built.routingDecision());
+                    } catch (Exception e) {
+                        log.debug("[{}] Failed to carry recent image {}: {}",
+                                agentName, part.getFileName(), e.getMessage());
+                        return built;
+                    }
+                }
+            }
+        } catch (Exception e) {
+            log.debug("[{}] maybeCarryRecentImage failed: {}", agentName, e.getMessage());
+        }
+        return built;
+    }
+
+    private boolean messageHasImagePart(MessageEntity message) {
+        for (MessageContentPart part : conversationService.parseMessageParts(message)) {
+            if (isResolvableImagePart(part)) return true;
+        }
+        return false;
+    }
+
+    /** An image part (not SVG) — the raster kind a multimodal API can ingest. */
+    private static boolean isResolvableImagePart(MessageContentPart part) {
+        if (part == null) return false;
+        String type = part.getType();
+        String contentType = part.getContentType();
+        boolean isImage = ("image".equals(type) || "file".equals(type))
+                && contentType != null && contentType.startsWith("image/");
+        return isImage && !contentType.contains("svg");
+    }
+
     protected Path resolveImagePath(String relativePath) {
         if (relativePath == null || relativePath.isBlank()) {
             return null;

mateclaw-server/src/main/java/vip/mate/llm/routing/MediaCaptionService.java

@@ -42,6 +42,18 @@ public class MediaCaptionService {
     private final RetryTemplate retryTemplate;
 
     public CaptionResult caption(ModelConfigEntity visionModel, MessageContentPart imagePart, Locale locale) {
+        return caption(visionModel, imagePart, locale, null);
+    }
+
+    /**
+     * Caption an image, optionally tailored to a user question. When
+     * {@code userQuestion} is non-blank the vision model is asked to answer it
+     * directly (in addition to describing the image), so multi-turn follow-ups
+     * get an answer rather than a generic description. When blank, falls back to
+     * the factual full-description prompt.
+     */
+    public CaptionResult caption(ModelConfigEntity visionModel, MessageContentPart imagePart, Locale locale,
+                                 String userQuestion) {
         if (visionModel == null || imagePart == null) {
             return CaptionResult.failure(0, new IllegalArgumentException("vision model or image part is null"));
         }
@@ -59,7 +71,7 @@ public class MediaCaptionService {
             ChatModel chatModel = chatModelFactory.buildFor(visionModel, retryTemplate);
             ChatClient client = ChatClient.create(chatModel);
             UserMessage userMessage = UserMessage.builder()
-                    .text(buildPrompt(locale, imagePart.getFileName()))
+                    .text(buildPrompt(locale, imagePart.getFileName(), userQuestion))
                     .media(List.of(new Media(MimeType.valueOf(contentType), new FileSystemResource(mediaPath))))
                     .build();
             String description = client.prompt()
@@ -87,9 +99,27 @@ public class MediaCaptionService {
      * (matches the primary user base) but switches to English so vision-model output
      * matches the chat language and avoids polluting English-only contexts.
      */
-    private String buildPrompt(Locale locale, String fileName) {
+    private String buildPrompt(Locale locale, String fileName, String userQuestion) {
         boolean english = locale != null && Locale.ENGLISH.getLanguage().equalsIgnoreCase(locale.getLanguage());
         String fileHint = (fileName == null || fileName.isBlank()) ? "" : " (" + fileName + ")";
+        boolean hasQuestion = userQuestion != null && !userQuestion.isBlank();
+
+        if (hasQuestion) {
+            String question = userQuestion.trim();
+            // Question-aware: extract everything relevant to the user's ask, then
+            // answer it. Answering in the question's own language keeps the caption
+            // consistent with the chat regardless of the configured locale.
+            if (english) {
+                return "Look at this image" + fileHint + " and answer the user's question. "
+                        + "First note any details relevant to the question (objects, scene, visible text/OCR, "
+                        + "numbers, layout), then answer directly. Reply in the same language as the question. "
+                        + "Question: " + question;
+            }
+            return "请仔细查看这张图片" + fileHint + "，并回答用户的问题。"
+                    + "先指出与问题相关的细节（物体、场景、画面文字/OCR、数字、排版等），再直接作答。"
+                    + "用与问题相同的语言回复。问题：" + question;
+        }
+
         if (english) {
             return "Describe this image" + fileHint
                     + " concisely: list the main objects, scene, any visible text (OCR), "

mateclaw-server/src/main/java/vip/mate/workspace/conversation/ConversationService.java

@@ -1033,10 +1033,38 @@ public class ConversationService {
             name = "未命名";
         }
         String path = safe(part.getPath());
-        if (path.isBlank()) {
-            return label + " " + name;
+        StringBuilder rendered = new StringBuilder(label).append(' ').append(name);
+        if (!path.isBlank()) {
+            rendered.append("（路径: ").append(path).append("）");
+        }
+        // A persisted caption (vision sidecar output) carries the image content
+        // into later turns: history user messages replay as text only, so without
+        // this the model would lose all knowledge of the attachment after turn 1.
+        String caption = safe(part.getCaption());
+        if (!caption.isBlank()) {
+            rendered.append("\n[图片内容] ").append(caption.trim());
+        }
+        return rendered.toString();
+    }
+
+    /**
+     * Overwrite a message's {@code content_parts} with an updated list — used by
+     * the vision sidecar to persist generated captions back onto image parts so
+     * later turns retain the image description (history replay is text-only).
+     * Best-effort: a serialization or DB failure is logged, not propagated, so
+     * the in-flight chat turn is never broken by a caption write.
+     */
+    public void updateMessageParts(MessageEntity message, List<MessageContentPart> parts) {
+        if (message == null || message.getId() == null || parts == null || parts.isEmpty()) {
+            return;
+        }
+        try {
+            message.setContentParts(serializeParts(parts));
+            messageMapper.updateById(message);
+        } catch (Exception e) {
+            log.warn("Failed to persist updated content_parts for message {}: {}",
+                    message.getId(), e.getMessage());
         }
-        return label + " " + name + "（路径: " + path + "）";
     }
 
     private void appendSegment(StringBuilder builder, String text) {

mateclaw-server/src/main/java/vip/mate/workspace/conversation/model/MessageContentPart.java

@@ -49,6 +49,15 @@ public class MessageContentPart {
      */
     private String mediaId;
 
+    /**
+     * Vision-model description of an image/video part, produced by the sidecar
+     * captioning path when the primary model is text-only. Persisted so the
+     * description survives into later turns: history replay sends user messages
+     * as text, and without a stored caption the image content would be lost on
+     * every follow-up question. Null for non-media parts or when no captioning ran.
+     */
+    private String caption;
+
     // ==================== 工厂方法 ====================
 
     public static MessageContentPart text(String text) {

mateclaw-server/src/test/java/vip/mate/agent/BaseAgentCarryRecentImageTest.java

@@ -0,0 +1,176 @@
+package vip.mate.agent;
+
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+import org.springframework.ai.chat.messages.UserMessage;
+import vip.mate.llm.service.ModelCapabilityService;
+import vip.mate.workspace.conversation.ConversationService;
+import vip.mate.workspace.conversation.model.MessageContentPart;
+import vip.mate.workspace.conversation.model.MessageEntity;
+
+import java.nio.file.Files;
+import java.nio.file.Path;
+import java.util.EnumSet;
+import java.util.List;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.mockito.ArgumentMatchers.any;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.when;
+
+/**
+ * Issue #303 follow-up: a vision-capable model replays history as text only, so a
+ * follow-up question about an earlier image was answered blind. The current turn
+ * must re-attach the most recent image so the model actually re-sees it.
+ */
+class BaseAgentCarryRecentImageTest {
+
+    @Test
+    @DisplayName("Vision model + follow-up with no image → most recent image is carried into the turn")
+    void followUp_carriesRecentImage() throws Exception {
+        Path img = Files.createTempFile("carry-test", ".jpg");
+        Files.write(img, new byte[]{(byte) 0xFF, (byte) 0xD8, (byte) 0xFF, 0x00});
+        try {
+            TestAgent agent = visionAgent();
+
+            MessageEntity imgTurn = userMsg("看看这张图");
+            MessageContentPart imagePart = imagePart(img.toAbsolutePath().toString());
+            MessageEntity asst = assistantMsg("图里是手写的字");
+            MessageEntity followUp = userMsg("左上角有没有小字");
+
+            List<MessageEntity> history = List.of(imgTurn, asst, followUp);
+            when(agent.conversationService.listMessages("c1")).thenReturn(history);
+            when(agent.conversationService.renderMessageContent(followUp)).thenReturn("左上角有没有小字");
+            when(agent.conversationService.parseMessageParts(imgTurn)).thenReturn(List.of(imagePart));
+            when(agent.conversationService.parseMessageParts(followUp)).thenReturn(List.of());
+            when(agent.conversationService.parseMessageParts(asst)).thenReturn(List.of());
+
+            UserMessage result = agent.callBuildCurrent("c1", "左上角有没有小字");
+
+            assertTrue(result.getMedia() != null && result.getMedia().size() == 1,
+                    "the recent image must be re-attached to the follow-up turn");
+            assertTrue(result.getText().contains("较早发送的"),
+                    "a note must explain the carried image to the model");
+        } finally {
+            Files.deleteIfExists(img);
+        }
+    }
+
+    @Test
+    @DisplayName("Text-only model → no image carried (relies on persisted caption instead)")
+    void textOnlyModel_doesNotCarry() throws Exception {
+        Path img = Files.createTempFile("carry-test", ".jpg");
+        Files.write(img, new byte[]{(byte) 0xFF, (byte) 0xD8, (byte) 0xFF, 0x00});
+        try {
+            TestAgent agent = newAgent(EnumSet.of(ModelCapabilityService.Modality.TEXT));
+            MessageEntity imgTurn = userMsg("看看这张图");
+            MessageEntity followUp = userMsg("左上角有没有小字");
+            List<MessageEntity> history = List.of(imgTurn, followUp);
+            when(agent.conversationService.listMessages("c1")).thenReturn(history);
+            when(agent.conversationService.renderMessageContent(followUp)).thenReturn("左上角有没有小字");
+            when(agent.conversationService.parseMessageParts(any())).thenReturn(List.of());
+
+            UserMessage result = agent.callBuildCurrent("c1", "左上角有没有小字");
+
+            assertTrue(result.getMedia() == null || result.getMedia().isEmpty(),
+                    "text-only model must not get raw image bytes carried over");
+        } finally {
+            Files.deleteIfExists(img);
+        }
+    }
+
+    @Test
+    @DisplayName("Current turn already has an image → nothing extra carried")
+    void currentTurnHasImage_noCarry() throws Exception {
+        Path older = Files.createTempFile("carry-old", ".jpg");
+        Path now = Files.createTempFile("carry-now", ".jpg");
+        Files.write(older, new byte[]{(byte) 0xFF, (byte) 0xD8, (byte) 0xFF, 0x00});
+        Files.write(now, new byte[]{(byte) 0xFF, (byte) 0xD8, (byte) 0xFF, 0x00});
+        try {
+            TestAgent agent = visionAgent();
+            MessageEntity oldTurn = userMsg("第一张");
+            MessageEntity curTurn = userMsg("第二张");
+            List<MessageEntity> history = List.of(oldTurn, curTurn);
+            when(agent.conversationService.listMessages("c1")).thenReturn(history);
+            when(agent.conversationService.renderMessageContent(curTurn)).thenReturn("第二张");
+            when(agent.conversationService.parseMessageParts(oldTurn))
+                    .thenReturn(List.of(imagePart(older.toAbsolutePath().toString())));
+            when(agent.conversationService.parseMessageParts(curTurn))
+                    .thenReturn(List.of(imagePart(now.toAbsolutePath().toString())));
+
+            UserMessage result = agent.callBuildCurrent("c1", "第二张");
+
+            assertTrue(result.getMedia() != null && result.getMedia().size() == 1,
+                    "only the current turn's own image should be present — no extra carry");
+            assertFalse(result.getText().contains("较早发送的"),
+                    "no carry note when the current turn already has an image");
+        } finally {
+            Files.deleteIfExists(older);
+            Files.deleteIfExists(now);
+        }
+    }
+
+    // ---------- scaffold ----------
+
+    private static MessageContentPart imagePart(String path) {
+        MessageContentPart p = new MessageContentPart();
+        p.setType("image");
+        p.setContentType("image/jpeg");
+        p.setFileName("image.jpg");
+        p.setPath(path);
+        return p;
+    }
+
+    private static MessageEntity userMsg(String content) {
+        MessageEntity m = new MessageEntity();
+        m.setRole("user");
+        m.setContent(content);
+        return m;
+    }
+
+    private static MessageEntity assistantMsg(String content) {
+        MessageEntity m = new MessageEntity();
+        m.setRole("assistant");
+        m.setContent(content);
+        return m;
+    }
+
+    private static TestAgent visionAgent() {
+        return newAgent(EnumSet.of(ModelCapabilityService.Modality.VISION, ModelCapabilityService.Modality.TEXT));
+    }
+
+    private static TestAgent newAgent(EnumSet<ModelCapabilityService.Modality> caps) {
+        ConversationService conv = mock(ConversationService.class);
+        TestAgent agent = new TestAgent(conv);
+        agent.modelCapabilities = caps;
+        agent.modelName = "test-model";
+        agent.agentName = "test-agent";
+        return agent;
+    }
+
+    static class TestAgent extends BaseAgent {
+        TestAgent(ConversationService conv) {
+            super(null, conv);
+        }
+
+        UserMessage callBuildCurrent(String conversationId, String text) {
+            return buildCurrentUserMessageWithRouting(conversationId, text).userMessage();
+        }
+
+        @Override
+        public String chat(String userMessage, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+
+        @Override
+        public reactor.core.publisher.Flux<String> chatStream(String userMessage, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+
+        @Override
+        public String execute(String goal, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+    }
+}

mateclaw-server/src/test/java/vip/mate/agent/BaseAgentImageCaptionPersistTest.java

@@ -0,0 +1,163 @@
+package vip.mate.agent;
+
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+import org.mockito.ArgumentCaptor;
+import org.springframework.ai.chat.messages.UserMessage;
+import vip.mate.llm.model.ModelConfigEntity;
+import vip.mate.llm.routing.MediaCaptionService;
+import vip.mate.llm.routing.MultimodalRouter;
+import vip.mate.llm.routing.model.MultimodalRoutingDecision;
+import vip.mate.llm.service.ModelCapabilityService;
+import vip.mate.workspace.conversation.ConversationService;
+import vip.mate.workspace.conversation.model.MessageContentPart;
+import vip.mate.workspace.conversation.model.MessageEntity;
+
+import java.util.EnumSet;
+import java.util.List;
+
+import static org.junit.jupiter.api.Assertions.assertEquals;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+import static org.mockito.ArgumentMatchers.any;
+import static org.mockito.ArgumentMatchers.eq;
+import static org.mockito.Mockito.mock;
+import static org.mockito.Mockito.verify;
+import static org.mockito.Mockito.when;
+
+/**
+ * Issue #303: when a text-only primary model captions an uploaded image via the
+ * vision sidecar, the caption must (1) be tailored to the user's actual question
+ * and (2) be persisted back onto the message part, so later turns — which replay
+ * user messages as text only — retain the image content instead of losing it.
+ */
+class BaseAgentImageCaptionPersistTest {
+
+    private static final String DESCRIPTION = "图中是一段 NullPointerException 堆栈，发生在 UserService.login 第 42 行。";
+
+    @Test
+    @DisplayName("Sidecar caption is persisted onto the image part and folded into the prompt")
+    void sidecarCaption_persistedAndInjected() {
+        TestHarness h = newHarness();
+        MessageContentPart image = imagePart();
+        MessageEntity msg = userMessage();
+        when(h.agent.conversationService.parseMessageParts(msg))
+                .thenReturn(List.of(MessageContentPart.text("图里的报错是什么"), image));
+
+        UserMessage result = h.agent.callBuildCurrentTurn(msg, "图里的报错是什么");
+
+        // (1) caption stored on the part → survives into later turns
+        assertEquals(DESCRIPTION, image.getCaption(), "caption must be written onto the image part");
+        verify(h.agent.conversationService).updateMessageParts(eq(msg), any());
+        // (2) caption folded into the current-turn prompt text
+        assertTrue(result.getText().contains(DESCRIPTION), "caption must be injected into the prompt");
+        assertTrue(result.getText().contains("[图片附件描述"), "caption must be wrapped in the attachment marker");
+    }
+
+    @Test
+    @DisplayName("The user's text question is passed to the caption service")
+    void userQuestion_passedToCaption() {
+        TestHarness h = newHarness();
+        MessageEntity msg = userMessage();
+        when(h.agent.conversationService.parseMessageParts(msg))
+                .thenReturn(List.of(MessageContentPart.text("报错的行号是多少"), imagePart()));
+
+        h.agent.callBuildCurrentTurn(msg, "报错的行号是多少");
+
+        ArgumentCaptor<String> question = ArgumentCaptor.forClass(String.class);
+        verify(h.caption).caption(any(), any(), any(), question.capture());
+        assertEquals("报错的行号是多少", question.getValue(),
+                "the user's question (text part) must drive a context-aware caption");
+    }
+
+    @Test
+    @DisplayName("Image-only message (no text part) → caption called with null question")
+    void imageOnly_nullQuestion() {
+        TestHarness h = newHarness();
+        MessageEntity msg = userMessage();
+        // WeChat Work image upload: only an image part, content placeholder "[图片]".
+        when(h.agent.conversationService.parseMessageParts(msg))
+                .thenReturn(List.of(imagePart()));
+
+        h.agent.callBuildCurrentTurn(msg, "[图片]");
+
+        ArgumentCaptor<String> question = ArgumentCaptor.forClass(String.class);
+        verify(h.caption).caption(any(), any(), any(), question.capture());
+        assertEquals(null, question.getValue(),
+                "no text part → null question → caption falls back to generic description");
+    }
+
+    // ---------- scaffold ----------
+
+    private static MessageContentPart imagePart() {
+        MessageContentPart p = new MessageContentPart();
+        p.setType("image");
+        p.setContentType("image/png");
+        p.setFileName("err.png");
+        p.setMediaId("media-1");
+        return p;
+    }
+
+    private static MessageEntity userMessage() {
+        MessageEntity m = new MessageEntity();
+        m.setId(1001L);
+        m.setRole("user");
+        m.setContent("[图片]");
+        return m;
+    }
+
+    private TestHarness newHarness() {
+        ConversationService conv = mock(ConversationService.class);
+        MultimodalRouter router = mock(MultimodalRouter.class);
+        MediaCaptionService caption = mock(MediaCaptionService.class);
+        ModelConfigEntity sidecar = mock(ModelConfigEntity.class);
+
+        when(router.route(any(), any())).thenReturn(
+                MultimodalRoutingDecision.sidecar(sidecar,
+                        EnumSet.of(ModelCapabilityService.Modality.VISION),
+                        EnumSet.of(ModelCapabilityService.Modality.VISION)));
+        when(caption.caption(any(), any(), any(), any()))
+                .thenReturn(MediaCaptionService.CaptionResult.success(DESCRIPTION, 12L, false));
+
+        TestAgent agent = new TestAgent(conv);
+        agent.multimodalRouter = router;
+        agent.mediaCaptionService = caption;
+        agent.modelCapabilities = EnumSet.noneOf(ModelCapabilityService.Modality.class);
+        agent.modelName = "text-only-model";
+        agent.agentName = "test-agent";
+
+        TestHarness h = new TestHarness();
+        h.agent = agent;
+        h.caption = caption;
+        return h;
+    }
+
+    static class TestHarness {
+        TestAgent agent;
+        MediaCaptionService caption;
+    }
+
+    static class TestAgent extends BaseAgent {
+        TestAgent(ConversationService conv) {
+            super(null, conv);
+        }
+
+        UserMessage callBuildCurrentTurn(MessageEntity msg, String renderedContent) {
+            return buildUserMessageForCurrentTurn(msg, renderedContent).userMessage();
+        }
+
+        @Override
+        public String chat(String userMessage, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+
+        @Override
+        public reactor.core.publisher.Flux<String> chatStream(String userMessage, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+
+        @Override
+        public String execute(String goal, String conversationId) {
+            throw new UnsupportedOperationException();
+        }
+    }
+}

mateclaw-server/src/test/java/vip/mate/agent/BaseAgentMultimodalSkipNoticeTest.java

@@ -86,6 +86,35 @@ class BaseAgentMultimodalSkipNoticeTest {
                 "image must NOT be injected when model has no VISION capability");
     }
 
+    @Test
+    @DisplayName("Vision-capable model + image with only a remote URL (never downloaded) → actionable 'enable media download' hint")
+    void remoteOnlyImage_emitsDownloadHint() {
+        // Issue #303 follow-up: WeCom/aibot delivers images as short-lived,
+        // AES-encrypted COS URLs. With channel media download off, the part is
+        // stored URL-only (path=null, mediaId=https://...). The model supports
+        // vision, so we reach the file-resolution branch — which must surface an
+        // actionable hint instead of a dead-end "文件未找到".
+        TestAgent agent = newAgentWithCaps(
+                EnumSet.of(ModelCapabilityService.Modality.VISION, ModelCapabilityService.Modality.TEXT));
+        MessageEntity msg = userMessage("看看这张图");
+        MessageContentPart remote = new MessageContentPart();
+        remote.setType("image");
+        remote.setContentType("image/jpeg");
+        remote.setFileName("image.jpg");
+        remote.setMediaId("https://ww-aibot-img.cos.example.com/x?sign=abc");
+        when(agent.conversationService.parseMessageParts(msg)).thenReturn(List.of(remote));
+
+        UserMessage result = agent.callBuildUserMessage(msg, "看看这张图");
+
+        String text = result.getText();
+        assertTrue(text.contains("未下载到本地"),
+                "remote-only attachment must hint that it was never downloaded locally");
+        assertTrue(text.contains("开启") && text.contains("媒体下载"),
+                "hint must tell the user to enable channel media download");
+        assertTrue(result.getMedia() == null || result.getMedia().isEmpty(),
+                "a remote URL must not be injected as Media");
+    }
+
     @Test
     @DisplayName("No attachments → no system notice, prompt text unchanged")
     void noAttachments_noNoticeAdded() {

mateclaw-server/src/test/java/vip/mate/llm/routing/MediaCaptionServiceBuildPromptTest.java

@@ -0,0 +1,70 @@
+package vip.mate.llm.routing;
+
+import org.junit.jupiter.api.DisplayName;
+import org.junit.jupiter.api.Test;
+
+import java.lang.reflect.Method;
+import java.util.Locale;
+
+import static org.junit.jupiter.api.Assertions.assertFalse;
+import static org.junit.jupiter.api.Assertions.assertTrue;
+
+/**
+ * Pins the two-stage prompt contract of {@link MediaCaptionService}: when a user
+ * question is supplied the vision model must be asked to answer it (so multi-turn
+ * follow-ups get a tailored answer instead of a generic caption), and when no
+ * question is supplied it must fall back to the factual full-description prompt.
+ *
+ * <p>{@code buildPrompt} is private — it is the smallest unit that captures the
+ * branching, so it is exercised via reflection rather than driving a real
+ * (networked) vision call.
+ */
+class MediaCaptionServiceBuildPromptTest {
+
+    private static String buildPrompt(Locale locale, String fileName, String userQuestion) throws Exception {
+        // Dependencies are unused by buildPrompt; null is fine for this unit.
+        MediaCaptionService service = new MediaCaptionService(null, null);
+        Method m = MediaCaptionService.class.getDeclaredMethod(
+                "buildPrompt", Locale.class, String.class, String.class);
+        m.setAccessible(true);
+        return (String) m.invoke(service, locale, fileName, userQuestion);
+    }
+
+    @Test
+    @DisplayName("Question + Chinese locale → question-aware prompt embedding the question")
+    void chineseWithQuestion_isQuestionAware() throws Exception {
+        String prompt = buildPrompt(Locale.SIMPLIFIED_CHINESE, "err.png", "图里的报错是什么");
+
+        assertTrue(prompt.contains("图里的报错是什么"), "the user's question must be embedded verbatim");
+        assertTrue(prompt.contains("回答用户的问题"), "must instruct the model to answer, not just describe");
+        assertFalse(prompt.contains("不超过 300 字"),
+                "question-aware prompt must not reuse the generic description template");
+    }
+
+    @Test
+    @DisplayName("Question + English locale → English question-aware prompt")
+    void englishWithQuestion_isQuestionAware() throws Exception {
+        String prompt = buildPrompt(Locale.ENGLISH, "err.png", "What is the error message?");
+
+        assertTrue(prompt.contains("What is the error message?"));
+        assertTrue(prompt.contains("answer the user's question"));
+    }
+
+    @Test
+    @DisplayName("Blank question → generic Chinese description prompt")
+    void chineseNoQuestion_isGeneric() throws Exception {
+        String prompt = buildPrompt(Locale.SIMPLIFIED_CHINESE, "photo.jpg", "  ");
+
+        assertTrue(prompt.contains("请用一段简洁的中文描述这张图片"));
+        assertFalse(prompt.contains("回答用户的问题"));
+    }
+
+    @Test
+    @DisplayName("Null question + English → generic English description prompt")
+    void englishNoQuestion_isGeneric() throws Exception {
+        String prompt = buildPrompt(Locale.ENGLISH, null, null);
+
+        assertTrue(prompt.contains("Describe this image"));
+        assertFalse(prompt.contains("answer the user's question"));
+    }
+}