media/media.go

@@ -5,10 +5,16 @@
 // already satisfies the target's llm.Capabilities. Images that do not fit
 // are decoded, downscaled (never upscaled), and re-encoded into an allowed
 // format and byte budget. Anything that cannot honestly be made to fit —
-// undecodable formats, impossible byte budgets, too many images, images for
-// a text-only target — fails with an error wrapping llm.ErrUnsupported so a
-// failover chain can advance to a more capable target without a health
-// penalty.
+// undecodable formats, impossible byte budgets, images for a text-only
+// target — fails with an error wrapping llm.ErrUnsupported so a failover
+// chain can advance to a more capable target without a health penalty.
+//
+// Over-count is the exception: a request carrying more images than
+// MaxImagesPerReq does NOT fail — the oldest images are replaced with a short
+// text placeholder and the most-recent MaxImagesPerReq are kept, because a hard
+// refuse exhausts a chain whose targets share the same cap (e.g. an agent loop
+// accumulating a preview image per iteration). MaxImagesPerReq remains the
+// per-model knob (0 = no image support).
 //
 // Why a separate package: every provider would otherwise duplicate the same
 // decode/scale/encode pipeline. Providers keep only a cheap capability
@@ -52,11 +58,17 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 	if !caps.SupportsImages() {
 		return llm.Request{}, fmt.Errorf("media: %w: target does not accept image input (request carries %d image(s))", llm.ErrUnsupported, total)
 	}
-	// Why error instead of dropping the overflow: silently removing an image
-	// changes the question the caller asked; the honest move is to refuse and
-	// let a chain try a roomier target.
+	// Overflow: keep the most-recent MaxImagesPerReq images and replace each
+	// older one with a short text placeholder, rather than refusing the whole
+	// request. A hard refuse exhausts a failover chain whose targets share the
+	// same cap — e.g. an agent loop that accumulates a preview image per
+	// iteration past the cap makes EVERY target reject and the run dies. The
+	// placeholder preserves each message's turn structure and tells the model an
+	// earlier image was elided; the most recent images (the relevant ones in an
+	// iterative run) are retained. The per-model threshold stays configurable via
+	// Capabilities.MaxImagesPerReq (0 still means "no image support").
 	if total > caps.MaxImagesPerReq {
-		return llm.Request{}, fmt.Errorf("media: %w: request carries %d images, target allows at most %d per request", llm.ErrUnsupported, total, caps.MaxImagesPerReq)
+		req = dropOldestImages(req, total-caps.MaxImagesPerReq)
 	}
 
 	out := req
@@ -92,6 +104,53 @@ func Normalize(req llm.Request, caps llm.Capabilities) (llm.Request, error) {
 	return out, nil
 }
 
+// imageOverflowPlaceholder replaces an image dropped to fit a target's
+// per-request image cap. It keeps the message turn intact and tells the model
+// an earlier image was elided rather than silently changing the conversation.
+const imageOverflowPlaceholder = "[earlier image omitted to fit this model's per-request image limit]"
+
+// dropOldestImages replaces the n oldest image parts (front-to-back across the
+// message history) with imageOverflowPlaceholder text, keeping the most-recent
+// images and preserving every message's turn structure. Copy-on-write: the
+// input request is never mutated. n <= 0 returns req unchanged.
+func dropOldestImages(req llm.Request, n int) llm.Request {
+	if n <= 0 {
+		return req
+	}
+	out := req
+	out.Messages = make([]llm.Message, len(req.Messages))
+	copy(out.Messages, req.Messages)
+	dropped := 0
+	for mi := range out.Messages {
+		if dropped >= n {
+			break
+		}
+		if !hasImagePart(out.Messages[mi].Parts) {
+			continue
+		}
+		parts := make([]llm.Part, 0, len(out.Messages[mi].Parts))
+		for _, p := range out.Messages[mi].Parts {
+			if _, ok := p.(llm.ImagePart); ok && dropped < n {
+				dropped++
+				parts = append(parts, llm.Text(imageOverflowPlaceholder))
+				continue
+			}
+			parts = append(parts, p)
+		}
+		out.Messages[mi].Parts = parts
+	}
+	return out
+}
+
+func hasImagePart(parts []llm.Part) bool {
+	for _, p := range parts {
+		if _, ok := p.(llm.ImagePart); ok {
+			return true
+		}
+	}
+	return false
+}
+
 // Info reports an image part's sniffed format ("jpeg", "png", "gif", or
 // "webp") and pixel dimensions. It is a cheap metadata read — the pixels are
 // never decoded. webp is recognized by signature but not decodable with the

media/media_test.go

@@ -149,18 +149,50 @@ func TestNormalizeImagesUnsupported(t *testing.T) {
 	}
 }
 
-func TestNormalizeTooManyImages(t *testing.T) {
-	img := llm.Image("image/png", encPNG(t, gradient(4, 4)))
+func TestNormalizeTooManyImages_DropsOldest(t *testing.T) {
+	// 3 distinguishable images across 2 messages; cap = 2. Overflow no longer
+	// errors — the OLDEST image is replaced with a placeholder and the most-recent
+	// two (the relevant ones in an iterative run) are kept.
+	a := llm.Image("image/png", encPNG(t, gradient(2, 2))).(llm.ImagePart)
+	b := llm.Image("image/png", encPNG(t, gradient(4, 4))).(llm.ImagePart)
+	c := llm.Image("image/png", encPNG(t, gradient(8, 8))).(llm.ImagePart)
 	req := llm.Request{Messages: []llm.Message{
-		llm.UserParts(img, img),
-		llm.UserParts(img),
+		llm.UserParts(a, b),
+		llm.UserParts(c),
 	}}
-	_, err := Normalize(req, llm.Capabilities{MaxImagesPerReq: 2})
-	if !errors.Is(err, llm.ErrUnsupported) {
-		t.Fatalf("err = %v, want ErrUnsupported", err)
+	caps := llm.Capabilities{MaxImagesPerReq: 2, MaxImageDimension: 64, MaxImageBytes: 1 << 20, AllowedImageMIME: []string{"image/png"}}
+	out, err := Normalize(req, caps)
+	if err != nil {
+		t.Fatalf("drop-oldest overflow should not error: %v", err)
 	}
-	if !strings.Contains(err.Error(), "3 images") || !strings.Contains(err.Error(), "at most 2") {
-		t.Errorf("err message %q lacks the counts", err)
+	var imgs []llm.ImagePart
+	placeholders := 0
+	for _, m := range out.Messages {
+		for _, p := range m.Parts {
+			switch v := p.(type) {
+			case llm.ImagePart:
+				imgs = append(imgs, v)
+			case llm.TextPart:
+				if strings.Contains(v.Text, "omitted") {
+					placeholders++
+				}
+			}
+		}
+	}
+	if len(imgs) != 2 {
+		t.Fatalf("kept %d images, want 2 (the cap)", len(imgs))
+	}
+	if placeholders != 1 {
+		t.Errorf("placeholders = %d, want 1 for the dropped oldest image", placeholders)
+	}
+	for _, im := range imgs {
+		if bytes.Equal(im.Data, a.Data) {
+			t.Errorf("oldest image was kept; the most-recent two should survive")
+		}
+	}
+	// The input request must be untouched (copy-on-write).
+	if len(req.Messages[0].Parts) != 2 {
+		t.Errorf("input request was mutated: %+v", req.Messages[0].Parts)
 	}
 }