fix(media): failed to store the media to /data/media and add image part

eslint.config.mjs

@@ -6,7 +6,7 @@ import vue from 'eslint-plugin-vue'
 export default [
   ...tseslint.configs.recommended,
   ...vue.configs['flat/recommended'],
-  { ignores: ['**/node_modules/**', '**/dist/**', '**/cache/**', 'packages/sdk/src/**'] },
+  { ignores: ['**/node_modules/**', '**/dist/**', '**/cache/**', '**/target/**', 'packages/sdk/src/**'] },
   {
     files: ['packages/**/*.{js,jsx,ts,tsx}', 'apps/**/*.{js,jsx,ts,tsx}'],
     languageOptions: {
@@ -56,6 +56,8 @@ export default [
     files: [
       'apps/web/src/pages/chat/components/tool-call-edit.vue',
       'apps/web/src/pages/chat/components/tool-call-write.vue',
+      'apps/web/src/pages/home/components/tool-call-edit.vue',
+      'apps/web/src/pages/home/components/tool-call-write.vue',
     ],
     rules: {
       'vue/no-v-html': 'off',

internal/channel/inbound/channel.go

@@ -417,7 +417,10 @@ func (p *ChannelInboundProcessor) HandleInbound(ctx context.Context, cfg channel
 		if _, loaded := p.pipeline.GetIC(sessionID); !loaded {
 			p.replayPipelineSession(ctx, sessionID)
 		}
-		event := pipelinepkg.AdaptInbound(msg, sessionID, identity.ChannelIdentityID, identity.DisplayName)
+		pipelineMsg := msg
+		pipelineMsg.Message = msg.Message
+		pipelineMsg.Message.Attachments = resolvedAttachments
+		event := pipelinepkg.AdaptInbound(pipelineMsg, sessionID, identity.ChannelIdentityID, identity.DisplayName)
 		if p.eventStore != nil {
 			eid, persistErr := p.eventStore.PersistEvent(ctx, identity.BotID, sessionID, event)
 			if persistErr != nil {

internal/channel/inbound/channel_test.go

@@ -24,6 +24,7 @@ import (
 	dbsqlc "github.com/memohai/memoh/internal/db/sqlc"
 	"github.com/memohai/memoh/internal/media"
 	messagepkg "github.com/memohai/memoh/internal/message"
+	pipelinepkg "github.com/memohai/memoh/internal/pipeline"
 	"github.com/memohai/memoh/internal/schedule"
 )
 
@@ -1223,6 +1224,82 @@ func TestChannelInboundProcessorIngestsQQFileAttachmentKeepsOriginalExtWhenMimeG
 	}
 }
 
+func TestChannelInboundProcessorPipelineUsesResolvedAttachments(t *testing.T) {
+	server := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, _ *http.Request) {
+		w.Header().Set("Content-Type", "image/jpeg")
+		_, _ = w.Write([]byte("fake-telegram-photo"))
+	}))
+	defer server.Close()
+
+	channelIdentitySvc := &fakeChannelIdentityService{channelIdentity: identities.ChannelIdentity{ID: "channelIdentity-pipeline-asset"}}
+	policySvc := &fakePolicyService{}
+	chatSvc := &fakeChatService{resolveResult: route.ResolveConversationResult{ChatID: "chat-pipeline-asset", RouteID: "route-pipeline-asset"}}
+	gateway := &fakeChatGateway{
+		resp: conversation.ChatResponse{
+			Messages: []conversation.ModelMessage{
+				{Role: "assistant", Content: conversation.NewTextContent("ok")},
+			},
+		},
+	}
+	processor := NewChannelInboundProcessor(slog.Default(), nil, chatSvc, chatSvc, gateway, channelIdentitySvc, policySvc, nil, "", 0)
+	mediaSvc := &fakeMediaIngestor{nextID: "asset-pipeline-photo", nextMime: "image/jpeg"}
+	processor.SetMediaService(mediaSvc)
+	processor.SetSessionEnsurer(&fakeSessionEnsurer{activeSession: SessionResult{ID: "session-pipeline-asset", Type: "chat"}})
+	pipeline := pipelinepkg.NewPipeline(pipelinepkg.RenderParams{})
+	processor.SetPipeline(pipeline, nil, nil)
+	sender := &fakeReplySender{}
+
+	cfg := channel.ChannelConfig{ID: "cfg-pipeline-asset", BotID: "bot-1", ChannelType: channel.ChannelTypeTelegram}
+	msg := channel.InboundMessage{
+		BotID:   "bot-1",
+		Channel: channel.ChannelTypeTelegram,
+		Message: channel.Message{
+			ID:   "msg-pipeline-asset-1",
+			Text: "photo test",
+			Attachments: []channel.Attachment{
+				{
+					Type:        channel.AttachmentImage,
+					URL:         server.URL + "/file/bot123/photo.jpg",
+					PlatformKey: "tg-photo-1",
+					Name:        "photo.jpg",
+					Mime:        "image/jpeg",
+				},
+			},
+		},
+		ReplyTarget: "12345",
+		Sender:      channel.Identity{SubjectID: "telegram-user"},
+		Conversation: channel.Conversation{
+			ID:   "12345",
+			Type: channel.ConversationTypePrivate,
+		},
+	}
+
+	if err := processor.HandleInbound(context.Background(), cfg, msg, sender); err != nil {
+		t.Fatalf("unexpected error: %v", err)
+	}
+	if mediaSvc.calls != 1 {
+		t.Fatalf("expected media ingest to be called once, got %d", mediaSvc.calls)
+	}
+
+	ic, ok := pipeline.GetIC("session-pipeline-asset")
+	if !ok {
+		t.Fatal("expected pipeline session to be created")
+	}
+	if len(ic.Nodes) == 0 || ic.Nodes[0].Message == nil {
+		t.Fatal("expected first pipeline node to be a message")
+	}
+	atts := ic.Nodes[0].Message.Attachments
+	if len(atts) != 1 {
+		t.Fatalf("expected one pipeline attachment, got %d", len(atts))
+	}
+	if got := atts[0].FilePath; got != "/data/media/test/asset-pipeline-photo" {
+		t.Fatalf("expected pipeline attachment path to use media store, got %q", got)
+	}
+	if strings.Contains(atts[0].FilePath, "api.telegram.org") {
+		t.Fatalf("expected pipeline attachment path to avoid telegram url, got %q", atts[0].FilePath)
+	}
+}
+
 func TestChannelInboundProcessorPersonalGroupNonOwnerIgnored(t *testing.T) {
 	channelIdentitySvc := &fakeChannelIdentityService{channelIdentity: identities.ChannelIdentity{ID: "channelIdentity-member"}}
 	policySvc := &fakePolicyService{ownerUserID: "channelIdentity-owner"}

internal/conversation/flow/resolver.go

@@ -145,6 +145,39 @@ func (r *Resolver) Pipeline() *pipelinepkg.Pipeline {
 	return r.pipeline
 }
 
+// InlineImageAttachments resolves image content hashes to sdk.ImagePart values
+// using the configured asset loader. Intended for the discuss driver to inline
+// images from new RC segments before calling the LLM.
+func (r *Resolver) InlineImageAttachments(ctx context.Context, botID string, refs []pipelinepkg.ImageAttachmentRef) []sdk.ImagePart {
+	if r == nil || r.assetLoader == nil || len(refs) == 0 {
+		return nil
+	}
+	var parts []sdk.ImagePart
+	for _, ref := range refs {
+		contentHash := strings.TrimSpace(ref.ContentHash)
+		if contentHash == "" {
+			continue
+		}
+		dataURL, mime, err := r.inlineAssetAsDataURL(ctx, botID, contentHash, "image", strings.TrimSpace(ref.Mime))
+		if err != nil {
+			if r.logger != nil {
+				r.logger.Warn(
+					"inline discuss image attachment failed",
+					slog.Any("error", err),
+					slog.String("bot_id", botID),
+					slog.String("content_hash", contentHash),
+				)
+			}
+			continue
+		}
+		parts = append(parts, sdk.ImagePart{
+			Image:     dataURL,
+			MediaType: mime,
+		})
+	}
+	return parts
+}
+
 type usageInfo struct {
 	InputTokens  *int `json:"inputTokens"`
 	OutputTokens *int `json:"outputTokens"`

internal/pipeline/adapt.go

@@ -90,11 +90,12 @@ func adaptAttachments(atts []channel.Attachment) []Attachment {
 	result := make([]Attachment, 0, len(atts))
 	for _, a := range atts {
 		att := Attachment{
-			Type:     string(a.Type),
-			MimeType: strings.TrimSpace(a.Mime),
-			FileName: strings.TrimSpace(a.Name),
-			Width:    a.Width,
-			Height:   a.Height,
+			Type:        string(a.Type),
+			MimeType:    strings.TrimSpace(a.Mime),
+			FileName:    strings.TrimSpace(a.Name),
+			ContentHash: strings.TrimSpace(a.ContentHash),
+			Width:       a.Width,
+			Height:      a.Height,
 		}
 		if a.DurationMs > 0 {
 			att.Duration = int(a.DurationMs / 1000)

internal/pipeline/driver.go

@@ -27,6 +27,7 @@ type ResolveRunConfigResult struct {
 // rounds. Implemented by flow.Resolver.
 type RunConfigResolver interface {
 	ResolveRunConfig(ctx context.Context, botID, sessionID, channelIdentityID, currentPlatform, replyTarget, conversationType, chatToken string) (ResolveRunConfigResult, error)
+	InlineImageAttachments(ctx context.Context, botID string, refs []ImageAttachmentRef) []sdk.ImagePart
 	StoreRound(ctx context.Context, botID, sessionID, channelIdentityID, currentPlatform string, messages []sdk.Message, modelID string) error
 }
 
@@ -256,6 +257,17 @@ func (d *DiscussDriver) handleReplyWithAgent(ctx context.Context, sess *discussS
 	runConfig.SessionType = sessionpkg.TypeDiscuss
 	runConfig.Query = ""
 
+	// Inline image attachments from new RC segments so the model receives
+	// them as native vision input (ImagePart) on the first encounter.
+	// Subsequent turns only see the file path in the XML rendering.
+	if runConfig.SupportsImageInput && d.deps.Resolver != nil {
+		imageRefs := extractNewImageRefs(rc, sess.lastProcessedMs)
+		if len(imageRefs) > 0 {
+			imageParts := d.deps.Resolver.InlineImageAttachments(ctx, cfg.BotID, imageRefs)
+			injectImagePartsIntoLastUserMessage(runConfig.Messages, imageParts)
+		}
+	}
+
 	isMentioned := wasRecentlyMentioned(rc, sess.lastProcessedMs)
 	lateBinding := buildLateBindingPrompt(isMentioned)
 	runConfig.Messages = append(runConfig.Messages, sdk.UserMessage(lateBinding))
@@ -375,6 +387,41 @@ func (d *DiscussDriver) loadTurnResponses(ctx context.Context, sessionID string)
 	return trs
 }
 
+// extractNewImageRefs collects ImageAttachmentRef entries from RC segments
+// that arrived after afterMs (i.e. new since the last LLM call).
+func extractNewImageRefs(rc RenderedContext, afterMs int64) []ImageAttachmentRef {
+	var refs []ImageAttachmentRef
+	for _, seg := range rc {
+		if seg.ReceivedAtMs > afterMs && !seg.IsMyself {
+			refs = append(refs, seg.ImageRefs...)
+		}
+	}
+	return refs
+}
+
+// injectImagePartsIntoLastUserMessage appends ImageParts to the last user
+// message in msgs so the model receives inline vision input.
+func injectImagePartsIntoLastUserMessage(msgs []sdk.Message, parts []sdk.ImagePart) {
+	if len(parts) == 0 {
+		return
+	}
+	extra := make([]sdk.MessagePart, 0, len(parts))
+	for _, p := range parts {
+		if strings.TrimSpace(p.Image) != "" {
+			extra = append(extra, p)
+		}
+	}
+	if len(extra) == 0 {
+		return
+	}
+	for i := len(msgs) - 1; i >= 0; i-- {
+		if msgs[i].Role == sdk.MessageRoleUser {
+			msgs[i].Content = append(msgs[i].Content, extra...)
+			return
+		}
+	}
+}
+
 func wasRecentlyMentioned(rc RenderedContext, afterMs int64) bool {
 	for _, seg := range rc {
 		if seg.ReceivedAtMs > afterMs && (seg.MentionsMe || seg.RepliesToMe) {

internal/pipeline/driver_test.go

@@ -0,0 +1,243 @@
+package pipeline
+
+import (
+	"context"
+	"strings"
+	"testing"
+
+	sdk "github.com/memohai/twilight-ai/sdk"
+
+	agentpkg "github.com/memohai/memoh/internal/agent"
+)
+
+func TestExtractNewImageRefs(t *testing.T) {
+	rc := RenderedContext{
+		{ReceivedAtMs: 100, ImageRefs: []ImageAttachmentRef{{ContentHash: "old-hash", Mime: "image/png"}}},
+		{ReceivedAtMs: 200, IsMyself: true, ImageRefs: []ImageAttachmentRef{{ContentHash: "self-hash"}}},
+		{ReceivedAtMs: 300, ImageRefs: []ImageAttachmentRef{{ContentHash: "new-hash", Mime: "image/jpeg"}}},
+		{ReceivedAtMs: 400, ImageRefs: nil},
+	}
+
+	refs := extractNewImageRefs(rc, 150)
+	if len(refs) != 1 {
+		t.Fatalf("expected 1 ref, got %d", len(refs))
+	}
+	if refs[0].ContentHash != "new-hash" {
+		t.Fatalf("expected new-hash, got %q", refs[0].ContentHash)
+	}
+	if refs[0].Mime != "image/jpeg" {
+		t.Fatalf("expected image/jpeg, got %q", refs[0].Mime)
+	}
+}
+
+func TestExtractNewImageRefs_IncludesMultiple(t *testing.T) {
+	rc := RenderedContext{
+		{ReceivedAtMs: 100},
+		{ReceivedAtMs: 200, ImageRefs: []ImageAttachmentRef{
+			{ContentHash: "a"},
+			{ContentHash: "b"},
+		}},
+		{ReceivedAtMs: 300, ImageRefs: []ImageAttachmentRef{{ContentHash: "c"}}},
+	}
+	refs := extractNewImageRefs(rc, 50)
+	if len(refs) != 3 {
+		t.Fatalf("expected 3 refs, got %d", len(refs))
+	}
+}
+
+func TestInjectImagePartsIntoLastUserMessage(t *testing.T) {
+	msgs := []sdk.Message{
+		sdk.UserMessage("hello"),
+		sdk.AssistantMessage("hi"),
+		sdk.UserMessage("look at this"),
+	}
+	parts := []sdk.ImagePart{
+		{Image: "data:image/png;base64,abc", MediaType: "image/png"},
+	}
+
+	injectImagePartsIntoLastUserMessage(msgs, parts)
+
+	lastUser := msgs[2]
+	if len(lastUser.Content) != 2 {
+		t.Fatalf("expected 2 content parts, got %d", len(lastUser.Content))
+	}
+	imgPart, ok := lastUser.Content[1].(sdk.ImagePart)
+	if !ok {
+		t.Fatalf("expected ImagePart, got %T", lastUser.Content[1])
+	}
+	if imgPart.Image != "data:image/png;base64,abc" {
+		t.Fatalf("unexpected image: %q", imgPart.Image)
+	}
+}
+
+func TestInjectImagePartsIntoLastUserMessage_Empty(t *testing.T) {
+	msgs := []sdk.Message{sdk.UserMessage("hello")}
+	injectImagePartsIntoLastUserMessage(msgs, nil)
+	if len(msgs[0].Content) != 1 {
+		t.Fatalf("expected no change, got %d parts", len(msgs[0].Content))
+	}
+}
+
+func TestInjectImagePartsIntoLastUserMessage_SkipsEmptyImage(t *testing.T) {
+	msgs := []sdk.Message{sdk.UserMessage("hello")}
+	parts := []sdk.ImagePart{{Image: "", MediaType: "image/png"}}
+	injectImagePartsIntoLastUserMessage(msgs, parts)
+	if len(msgs[0].Content) != 1 {
+		t.Fatalf("expected no change, got %d parts", len(msgs[0].Content))
+	}
+}
+
+func TestHandleReplyWithAgent_InlinesImages(t *testing.T) {
+	rc := RenderedContext{
+		{
+			ReceivedAtMs: 200,
+			Content:      []RenderedContentPiece{{Type: "text", Text: `<message id="1">photo</message>`}},
+			ImageRefs:    []ImageAttachmentRef{{ContentHash: "img-hash", Mime: "image/jpeg"}},
+		},
+	}
+
+	fakeAgent := &fakeDiscussStreamer{}
+
+	resolver := &fakeRunConfigResolver{
+		resolveResult: ResolveRunConfigResult{
+			RunConfig: agentpkg.RunConfig{
+				SupportsImageInput: true,
+			},
+			ModelID: "model-1",
+		},
+		inlineFn: func(_ context.Context, _ string, refs []ImageAttachmentRef) []sdk.ImagePart {
+			if len(refs) != 1 || refs[0].ContentHash != "img-hash" {
+				t.Fatalf("unexpected refs: %v", refs)
+			}
+			return []sdk.ImagePart{{Image: "data:image/jpeg;base64,FAKE", MediaType: "image/jpeg"}}
+		},
+	}
+
+	driver := NewDiscussDriver(DiscussDriverDeps{
+		Pipeline: NewPipeline(RenderParams{}),
+		Resolver: resolver,
+	})
+
+	sess := &discussSession{
+		config: DiscussSessionConfig{
+			BotID:     "bot-1",
+			SessionID: "sess-1",
+		},
+		lastProcessedMs: 0,
+	}
+
+	driver.handleReplyWithAgent(context.Background(), sess, rc, driver.logger, fakeAgent)
+
+	if fakeAgent.lastConfig == nil {
+		t.Fatal("expected agent to be called")
+	}
+
+	msgs := fakeAgent.lastConfig.Messages
+	var userMsgs []sdk.Message
+	for _, m := range msgs {
+		if m.Role == sdk.MessageRoleUser {
+			userMsgs = append(userMsgs, m)
+		}
+	}
+	if len(userMsgs) < 2 {
+		t.Fatalf("expected at least 2 user messages (rc + late binding), got %d", len(userMsgs))
+	}
+	rcMsg := userMsgs[0]
+	hasImage := false
+	for _, part := range rcMsg.Content {
+		if imgPart, ok := part.(sdk.ImagePart); ok {
+			hasImage = true
+			if !strings.HasPrefix(imgPart.Image, "data:image/jpeg;base64,") {
+				t.Fatalf("unexpected image data: %q", imgPart.Image)
+			}
+		}
+	}
+	if !hasImage {
+		t.Fatal("expected image part in RC user message")
+	}
+}
+
+func TestHandleReplyWithAgent_NoInlineWhenNoVision(t *testing.T) {
+	rc := RenderedContext{
+		{
+			ReceivedAtMs: 200,
+			Content:      []RenderedContentPiece{{Type: "text", Text: `<message id="1">photo</message>`}},
+			ImageRefs:    []ImageAttachmentRef{{ContentHash: "img-hash", Mime: "image/jpeg"}},
+		},
+	}
+
+	fakeAgent := &fakeDiscussStreamer{}
+
+	resolver := &fakeRunConfigResolver{
+		resolveResult: ResolveRunConfigResult{
+			RunConfig: agentpkg.RunConfig{
+				SupportsImageInput: false,
+			},
+			ModelID: "model-1",
+		},
+		inlineFn: func(_ context.Context, _ string, _ []ImageAttachmentRef) []sdk.ImagePart {
+			t.Fatal("should not be called when model doesn't support vision")
+			return nil
+		},
+	}
+
+	driver := NewDiscussDriver(DiscussDriverDeps{
+		Pipeline: NewPipeline(RenderParams{}),
+		Resolver: resolver,
+	})
+
+	sess := &discussSession{
+		config: DiscussSessionConfig{
+			BotID:     "bot-1",
+			SessionID: "sess-1",
+		},
+		lastProcessedMs: 0,
+	}
+
+	driver.handleReplyWithAgent(context.Background(), sess, rc, driver.logger, fakeAgent)
+
+	if fakeAgent.lastConfig == nil {
+		t.Fatal("expected agent to be called")
+	}
+	for _, m := range fakeAgent.lastConfig.Messages {
+		for _, part := range m.Content {
+			if _, ok := part.(sdk.ImagePart); ok {
+				t.Fatal("should not have image parts when vision is not supported")
+			}
+		}
+	}
+}
+
+// --- Test helpers ---
+
+type fakeDiscussStreamer struct {
+	lastConfig *agentpkg.RunConfig
+}
+
+func (f *fakeDiscussStreamer) Stream(_ context.Context, cfg agentpkg.RunConfig) <-chan agentpkg.StreamEvent {
+	f.lastConfig = &cfg
+	ch := make(chan agentpkg.StreamEvent, 1)
+	ch <- agentpkg.StreamEvent{Type: agentpkg.EventAgentEnd}
+	close(ch)
+	return ch
+}
+
+type fakeRunConfigResolver struct {
+	resolveResult ResolveRunConfigResult
+	inlineFn      func(ctx context.Context, botID string, refs []ImageAttachmentRef) []sdk.ImagePart
+}
+
+func (f *fakeRunConfigResolver) ResolveRunConfig(_ context.Context, _, _, _, _, _, _, _ string) (ResolveRunConfigResult, error) {
+	return f.resolveResult, nil
+}
+
+func (f *fakeRunConfigResolver) InlineImageAttachments(ctx context.Context, botID string, refs []ImageAttachmentRef) []sdk.ImagePart {
+	if f.inlineFn != nil {
+		return f.inlineFn(ctx, botID, refs)
+	}
+	return nil
+}
+
+func (*fakeRunConfigResolver) StoreRound(_ context.Context, _, _, _, _ string, _ []sdk.Message, _ string) error {
+	return nil
+}

internal/pipeline/rendering.go

@@ -14,6 +14,13 @@ type RenderedContentPiece struct {
 	URL  string `json:"url,omitempty"`
 }
 
+// ImageAttachmentRef holds the content hash and MIME type of an image
+// attachment that can be inlined as a vision input via the media store.
+type ImageAttachmentRef struct {
+	ContentHash string `json:"content_hash"`
+	Mime        string `json:"mime,omitempty"`
+}
+
 // RenderedSegment is a single segment of rendered context, one per IC node.
 type RenderedSegment struct {
 	ReceivedAtMs int64                  `json:"received_at_ms"`
@@ -22,6 +29,7 @@ type RenderedSegment struct {
 	IsSelfSent   bool                   `json:"is_self_sent,omitempty"`
 	MentionsMe   bool                   `json:"mentions_me,omitempty"`
 	RepliesToMe  bool                   `json:"replies_to_me,omitempty"`
+	ImageRefs    []ImageAttachmentRef   `json:"image_refs,omitempty"`
 }
 
 // RenderedContext is the output of the Rendering layer — a slice of segments.
@@ -140,6 +148,16 @@ func renderMessage(msg *ICMessage, params RenderParams) RenderedSegment {
 
 	pieces := []RenderedContentPiece{{Type: "text", Text: text}}
 
+	var imageRefs []ImageAttachmentRef
+	for _, att := range msg.Attachments {
+		if strings.EqualFold(att.Type, "image") && att.ContentHash != "" {
+			imageRefs = append(imageRefs, ImageAttachmentRef{
+				ContentHash: att.ContentHash,
+				Mime:        att.MimeType,
+			})
+		}
+	}
+
 	return RenderedSegment{
 		ReceivedAtMs: msg.ReceivedAtMs,
 		Content:      pieces,
@@ -147,6 +165,7 @@ func renderMessage(msg *ICMessage, params RenderParams) RenderedSegment {
 		IsSelfSent:   msg.IsSelfSent,
 		MentionsMe:   mentionsMe,
 		RepliesToMe:  repliesToMe,
+		ImageRefs:    imageRefs,
 	}
 }
 

internal/pipeline/rendering_test.go

@@ -0,0 +1,73 @@
+package pipeline
+
+import (
+	"testing"
+
+	"github.com/memohai/memoh/internal/channel"
+)
+
+func TestRenderMessage_ImageRefsPopulated(t *testing.T) {
+	msg := &ICMessage{
+		MessageID:    "msg-1",
+		ReceivedAtMs: 100,
+		TimestampSec: 100,
+		Content:      []ContentNode{{Type: "text", Text: "photo"}},
+		Attachments: []Attachment{
+			{Type: "image", ContentHash: "hash-1", MimeType: "image/jpeg", FilePath: "/data/media/bot/ab/hash-1.jpg"},
+			{Type: "file", ContentHash: "hash-2", MimeType: "application/pdf", FilePath: "/data/media/bot/cd/hash-2.pdf"},
+			{Type: "image", MimeType: "image/png"},
+		},
+		Conversation: ConversationMeta{Channel: "telegram", ConversationType: "private"},
+	}
+
+	seg := renderMessage(msg, RenderParams{})
+
+	if len(seg.ImageRefs) != 1 {
+		t.Fatalf("expected 1 image ref (only images with ContentHash), got %d", len(seg.ImageRefs))
+	}
+	if seg.ImageRefs[0].ContentHash != "hash-1" {
+		t.Fatalf("expected hash-1, got %q", seg.ImageRefs[0].ContentHash)
+	}
+	if seg.ImageRefs[0].Mime != "image/jpeg" {
+		t.Fatalf("expected image/jpeg, got %q", seg.ImageRefs[0].Mime)
+	}
+}
+
+func TestRenderMessage_NoImageRefs(t *testing.T) {
+	msg := &ICMessage{
+		MessageID:    "msg-2",
+		ReceivedAtMs: 200,
+		TimestampSec: 200,
+		Content:      []ContentNode{{Type: "text", Text: "text only"}},
+		Conversation: ConversationMeta{Channel: "telegram", ConversationType: "private"},
+	}
+
+	seg := renderMessage(msg, RenderParams{})
+
+	if len(seg.ImageRefs) != 0 {
+		t.Fatalf("expected 0 image refs, got %d", len(seg.ImageRefs))
+	}
+}
+
+func TestAdaptAttachments_ContentHash(t *testing.T) {
+	atts := []channel.Attachment{
+		{Type: channel.AttachmentImage, ContentHash: "abc123", URL: "/data/media/bot/ab/abc123.jpg", Mime: "image/jpeg"},
+		{Type: channel.AttachmentFile, URL: "https://example.com/doc.pdf", Mime: "application/pdf"},
+	}
+	got := adaptAttachments(atts)
+	if len(got) != 2 {
+		t.Fatalf("expected 2 attachments, got %d", len(got))
+	}
+	if got[0].ContentHash != "abc123" || got[0].MimeType != "image/jpeg" {
+		t.Fatalf("unexpected first attachment: %+v", got[0])
+	}
+	if got[0].FilePath != "/data/media/bot/ab/abc123.jpg" {
+		t.Fatalf("expected FilePath from URL, got %q", got[0].FilePath)
+	}
+	if got[1].Type != "file" || got[1].MimeType != "application/pdf" {
+		t.Fatalf("unexpected second attachment: %+v", got[1])
+	}
+	if got[1].FilePath != "https://example.com/doc.pdf" {
+		t.Fatalf("expected FilePath from URL, got %q", got[1].FilePath)
+	}
+}

internal/pipeline/types.go

@@ -45,6 +45,8 @@ type Attachment struct {
 	AltText      string `json:"alt_text,omitempty"`
 	// FilePath is the workspace path where the attachment is stored.
 	FilePath string `json:"file_path,omitempty"`
+	// ContentHash is the media-store content hash for persisted attachments.
+	ContentHash string `json:"content_hash,omitempty"`
 }
 
 // ForwardInfo describes a forwarded message origin.