diff --git a/cmd/cyrene/vision.go b/cmd/cyrene/vision.go
index 397a3cd..42ec356 100644
--- a/cmd/cyrene/vision.go
+++ b/cmd/cyrene/vision.go
@@ -1,6 +1,7 @@
 package main
 
 import (
+	"path/filepath"
 	"strings"
 
 	"git.yeij.top/AskaEth/Cyrene_v2/internal/logger"
@@ -30,6 +31,10 @@ func (a *app) wireVision() {
 	}
 	adapter := vision.NewOpenAICompatible(a.cfg.LLM,
 		vision.WithModel(a.cfg.LLM.VisionModel),
+		// 同源 upload 产物（`/api/v1/files/<name>`）走**本地读取**：与自己取回相比更快，
+		// 且在「服务端路由不到自己的广告 host」（如安卓模拟器 10.0.2.2、反向代理改写）
+		// 的部署下才取得到图。目录必须与 upload 的落盘位置一致 = `<DataDir>/uploads`。
+		vision.WithUploadsDir(filepath.Join(resolveDataDir(a.cfg), "uploads")),
 		vision.WithLogger(a.log))
 	if err := vision.RegisterTools(a.caps,
 		vision.NewVisionTool(visionToolName, adapter),
diff --git a/internal/runtime/multimodal/vision/adapter.go b/internal/runtime/multimodal/vision/adapter.go
index 09d6bd7..74ba704 100644
--- a/internal/runtime/multimodal/vision/adapter.go
+++ b/internal/runtime/multimodal/vision/adapter.go
@@ -40,11 +40,12 @@ type OCRAdapter interface {
 // 模型 + OCR 模型」）：Describe 用 model（LLM_VISION_MODEL），Extract 用
 // ocrModel（LLM_OCR_MODEL），未配 OCR 模型时回落到同一个模型。
 type OpenAICompatibleAdapter struct {
-	cfg      config.LLMConfig
-	model    string
-	ocrModel string
-	client   *http.Client
-	log      *slog.Logger
+	cfg        config.LLMConfig
+	model      string
+	ocrModel   string
+	client     *http.Client
+	log        *slog.Logger
+	uploadsDir string
 }
 
 // OpenAIOption 用于定制 OpenAICompatibleAdapter。
@@ -87,6 +88,18 @@ func WithOCRModel(m string) OpenAIOption {
 	}
 }
 
+// WithUploadsDir 配置本机上传目录（<DataDir>/uploads，即网关静态路由
+// /api/v1/files/ 的根），使指向本服务的上传 URL 直接走本地文件读取，
+// 而不是绕一圈网络回自己的 HTTP 路由。留空时上传 URL 按普通 URL 取回，
+// 行为与配置前一致。
+func WithUploadsDir(dir string) OpenAIOption {
+	return func(a *OpenAICompatibleAdapter) {
+		if d := strings.TrimSpace(dir); d != "" {
+			a.uploadsDir = d
+		}
+	}
+}
+
 // NewOpenAICompatible 用 cfg 构造一个适配器，未覆盖时使用
 // cfg.Model。
 func NewOpenAICompatible(cfg config.LLMConfig, opts ...OpenAIOption) *OpenAICompatibleAdapter {
@@ -163,19 +176,40 @@ func (a *OpenAICompatibleAdapter) ask(ctx context.Context, model string, img Ima
 	return a.decode(resp.Body)
 }
 
-// imageURL 把 ImageRef 转换为线上格式期望的 image_url 取值：http(s)
-// URL 原样透传，其余一律转为 base64 data: URI。
+// imageURL 返回线上 image_url 的取值。无论来源是 http(s) URL、本地路径
+// 还是 base64，送上游的字节一律由服务端自己取回并内联为 data: URI：
+// 把同源（局域网）URL 原样透传给云 provider，对方根本取不到，
+// 自部署形态下视觉链路会整条不通（B1 上传大图的实际意义也就落空）。
+// URL 取回沿用 Resolve→fetchURL 的既有守卫（http/https 白名单 + 50MiB
+// 体积上限），并加取回超时（imageFetchTimeout）。
 func (a *OpenAICompatibleAdapter) imageURL(ctx context.Context, img ImageRef) (string, error) {
-	if img.Source == SourceURL {
-		return img.Data, nil
+	ref := a.localizeRef(img)
+	if ref.Source == SourceURL {
+		var cancel context.CancelFunc
+		ctx, cancel = context.WithTimeout(ctx, imageFetchTimeout)
+		defer cancel()
 	}
-	mime, data, err := Resolve(ctx, img, a.client)
+	mime, data, err := Resolve(ctx, ref, a.client)
 	if err != nil {
 		return "", err
 	}
 	return DataURI(mime, data), nil
 }
 
+// localizeRef 把指向本服务上传路由的 URL 改写成 uploads 目录下的本地路径。
+// uploadsDir 未配置、或 URL 不是本服务的上传形状时原样返回（此时仍会按
+// 普通 URL 取回，因此行为不退化）；文件缺失等读取错误由 Resolve 如实报告。
+func (a *OpenAICompatibleAdapter) localizeRef(img ImageRef) ImageRef {
+	if img.Source != SourceURL || a.uploadsDir == "" {
+		return img
+	}
+	path, ok := localUploadPath(img.Data, a.uploadsDir)
+	if !ok {
+		return img
+	}
+	return ImageRef{Source: SourcePath, Data: path}
+}
+
 // openAIVisionRequest 是发往 /chat/completions 的 JSON 请求体。
 type openAIVisionRequest struct {
 	Model     string          `json:"model"`
diff --git a/internal/runtime/multimodal/vision/errors.go b/internal/runtime/multimodal/vision/errors.go
index bb32e3d..9caa868 100644
--- a/internal/runtime/multimodal/vision/errors.go
+++ b/internal/runtime/multimodal/vision/errors.go
@@ -17,6 +17,9 @@ var (
 	ErrUpstream = errors.New("vision: upstream request failed")
 	// ErrBadImage 在图像无法读取或解码时返回。
 	ErrBadImage = errors.New("vision: cannot read image")
+	// ErrNotImage 在来源声明的类型与真实字节都判不出图像时返回
+	// （例如把 HTML/PROXY 错误页当成图片送进模型）。
+	ErrNotImage = errors.New("vision: source is not an image")
 	// ErrBadArgs 在工具参数格式错误时由该工具返回。
 	ErrBadArgs = errors.New("vision: bad tool arguments")
 )
diff --git a/internal/runtime/multimodal/vision/helpers_test.go b/internal/runtime/multimodal/vision/helpers_test.go
new file mode 100644
index 0000000..bac535b
--- /dev/null
+++ b/internal/runtime/multimodal/vision/helpers_test.go
@@ -0,0 +1,107 @@
+package vision
+
+import (
+	"encoding/json"
+	"fmt"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"os"
+	"path/filepath"
+	"sync"
+	"sync/atomic"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/config"
+)
+
+// writeFileBytes 写一个测试用的字节文件。
+func writeFileBytes(path string, data []byte) error {
+	if err := os.WriteFile(path, data, 0o600); err != nil {
+		return fmt.Errorf("write test file: %w", err)
+	}
+	return nil
+}
+
+// writeFileBytesAll 在 dir 下写 name（含父目录），用于摆放测试用的上传产物。
+func writeFileBytesAll(dir, name string, data []byte) error {
+	if err := os.MkdirAll(dir, 0o700); err != nil {
+		return fmt.Errorf("create test dir: %w", err)
+	}
+	return writeFileBytes(filepath.Join(dir, name), data)
+}
+
+// imageURLRecorder 是一个 /chat/completions endpoint，记录发往上游的
+// image_url 取值，使测试能断言「服务端真的把图取回并内联了」而不是
+// 原样透传 URL。
+type imageURLRecorder struct {
+	mu    sync.Mutex
+	calls int
+	urls  []string
+}
+
+func (r *imageURLRecorder) seen() (int, []string) {
+	r.mu.Lock()
+	defer r.mu.Unlock()
+	return r.calls, append([]string(nil), r.urls...)
+}
+
+// newImageURLRecorderServer 起一个上游 endpoint，把每次请求里的
+// image_url.url 记下来。
+func newImageURLRecorderServer(t *testing.T) (*httptest.Server, *imageURLRecorder) {
+	t.Helper()
+	rec := &imageURLRecorder{}
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		var body struct {
+			Messages []struct {
+				Content []struct {
+					ImageURL *struct {
+						URL string `json:"url"`
+					} `json:"image_url"`
+				} `json:"content"`
+			} `json:"messages"`
+		}
+		if err := json.NewDecoder(r.Body).Decode(&body); err != nil {
+			t.Errorf("decode upstream body: %v", err)
+		}
+		rec.mu.Lock()
+		rec.calls++
+		for _, m := range body.Messages {
+			for _, part := range m.Content {
+				if part.ImageURL != nil {
+					rec.urls = append(rec.urls, part.ImageURL.URL)
+				}
+			}
+		}
+		rec.mu.Unlock()
+		w.Header().Set("Content-Type", "application/json")
+		_, _ = io.WriteString(w, `{"choices":[{"message":{"content":"a small square"}}]}`)
+	}))
+	t.Cleanup(srv.Close)
+	return srv, rec
+}
+
+// newImageServer 起一个提供图像字节的来源服务，contentType 为 "" 时
+// 刻意不发 Content-Type（考验字节嗅探）。
+func newImageServer(t *testing.T, data []byte, contentType string, hits *atomic.Int32) *httptest.Server {
+	t.Helper()
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		if hits != nil {
+			hits.Add(1)
+		}
+		if contentType == "" {
+			w.Header()["Content-Type"] = nil
+			w.Header().Set("Content-Type", "")
+		} else {
+			w.Header().Set("Content-Type", contentType)
+		}
+		_, _ = w.Write(data)
+	}))
+	t.Cleanup(srv.Close)
+	return srv
+}
+
+// newAdapter 构造一个指向 up 的适配器，opts 追加在必需项之后。
+func newAdapter(up string, opts ...OpenAIOption) *OpenAICompatibleAdapter {
+	return NewOpenAICompatible(config.LLMConfig{BaseURL: up, Model: "vision-test"}, opts...)
+}
diff --git a/internal/runtime/multimodal/vision/image.go b/internal/runtime/multimodal/vision/image.go
index 906932c..c94b95d 100644
--- a/internal/runtime/multimodal/vision/image.go
+++ b/internal/runtime/multimodal/vision/image.go
@@ -8,6 +8,7 @@ import (
 	"net/http"
 	"os"
 	"strings"
+	"time"
 )
 
 // 读取图像字节与响应体的体积上限，使恶意或
@@ -18,6 +19,11 @@ const (
 	maxFetchedBody = 50 << 20
 )
 
+// imageFetchTimeout 限制一次图像 URL 取回的时长。适配器与消息解析器会
+// 为模型代取 URL，没有上限时一个卡住或恶意的地址会一直占着这一轮
+// （fetchURL 复用默认 client，其本身不设整体超时）。
+const imageFetchTimeout = 30 * time.Second
+
 // DataURI 由 mime 与原始字节构造一个 base64 data: 图像 URI。
 func DataURI(mime string, data []byte) string {
 	return "data:" + mime + ";base64," + base64.StdEncoding.EncodeToString(data)
@@ -40,7 +46,8 @@ func Resolve(ctx context.Context, ref ImageRef, client *http.Client) (mime strin
 	}
 }
 
-// readFile 读取一个图像文件，并从扩展名推断其 MIME 类型。
+// readFile 读取一个图像文件。扩展名给出声明的 MIME，真实字节嗅探是否决权
+// （见 imageMIMEFor）：改名的 HTML 不会被当成图片读进模型。
 func readFile(path string) (string, []byte, error) {
 	f, err := os.Open(path)
 	if err != nil {
@@ -51,7 +58,11 @@ func readFile(path string) (string, []byte, error) {
 	if err != nil {
 		return "", nil, fmt.Errorf("vision: %w: %v", ErrBadImage, err)
 	}
-	return mimeByExt(path), data, nil
+	mime, err := imageMIMEFor(mimeByExt(path), data)
+	if err != nil {
+		return "", nil, err
+	}
+	return mime, data, nil
 }
 
 // decodeBase64 返回 base64/data: 载荷解码后的字节及其 MIME。
@@ -72,6 +83,8 @@ func decodeBase64(src string) (string, []byte, error) {
 }
 
 // fetchURL 通过 http(s) 下载一张图像；只允许 http/https 协议。
+// 声明的 Content-Type 只作参考，真实字节嗅探说了算（见 imageMIMEFor）：
+// 网关/PROXY 返回的 HTML 错误页绝不会被当成 png 送进模型。
 func fetchURL(ctx context.Context, rawURL string, client *http.Client) (string, []byte, error) {
 	if !strings.HasPrefix(rawURL, "http://") && !strings.HasPrefix(rawURL, "https://") {
 		return "", nil, fmt.Errorf("vision: %w: %q", ErrUnsupportedSource, rawURL)
@@ -95,7 +108,35 @@ func fetchURL(ctx context.Context, rawURL string, client *http.Client) (string,
 	if err != nil {
 		return "", nil, fmt.Errorf("vision: %w: %v", ErrFetch, err)
 	}
-	return mimeByContentType(resp.Header.Get("Content-Type")), data, nil
+	mime, err := imageMIMEFor(resp.Header.Get("Content-Type"), data)
+	if err != nil {
+		return "", nil, err
+	}
+	return mime, data, nil
+}
+
+// inlineImageSrc 把图像来源字符串取回并内联为 data: URI，供主模型
+// 直接消费。已是 data: URI 的原样返回（客户端现有的 base64 路径因此
+// 逐字节不变）。URL 取回复用 Resolve→fetchURL 的既有守卫
+// （协议白名单 + 体积上限），并加取回超时。
+func inlineImageSrc(ctx context.Context, src string) (string, error) {
+	if strings.HasPrefix(src, "data:") {
+		return src, nil
+	}
+	ref, err := ParseImageRef(src)
+	if err != nil {
+		return "", err
+	}
+	if ref.Source == SourceURL {
+		var cancel context.CancelFunc
+		ctx, cancel = context.WithTimeout(ctx, imageFetchTimeout)
+		defer cancel()
+	}
+	mime, data, err := Resolve(ctx, ref, nil)
+	if err != nil {
+		return "", err
+	}
+	return DataURI(mime, data), nil
 }
 
 // splitDataURI 解析 "data:<mime>;base64,<payload>" 前缀。src 不是
@@ -124,46 +165,6 @@ func splitDataURI(src string) (mime, payload string, ok bool) {
 	return mime, payload, true
 }
 
-// mimeByExt 从文件路径扩展名推断 MIME 类型。
-func mimeByExt(path string) string {
-	switch strings.ToLower(pathExtension(path)) {
-	case ".png":
-		return "image/png"
-	case ".jpg", ".jpeg":
-		return "image/jpeg"
-	case ".gif":
-		return "image/gif"
-	case ".webp":
-		return "image/webp"
-	default:
-		return "image/png"
-	}
-}
-
-// mimeByContentType 从 HTTP Content-Type 头返回 MIME 类型，
-// 缺失或非图像时默认为 image/png。
-func mimeByContentType(ct string) string {
-	if ct == "" || !strings.HasPrefix(ct, "image/") {
-		return "image/png"
-	}
-	return strings.TrimSpace(strings.Split(ct, ";")[0])
-}
-
-// pathExtension 返回带点号的文件扩展名，没有则返回空串。
-func pathExtension(path string) string {
-	base := path
-	if idx := strings.LastIndexByte(path, '/'); idx >= 0 {
-		base = path[idx+1:]
-	} else if idx := strings.LastIndexByte(path, '\\'); idx >= 0 {
-		base = path[idx+1:]
-	}
-	dot := strings.LastIndexByte(base, '.')
-	if dot < 0 {
-		return ""
-	}
-	return base[dot:]
-}
-
 // fileExists 报告 path 是否存在于磁盘且为常规文件。
 func fileExists(path string) bool {
 	info, err := os.Stat(path)
diff --git a/internal/runtime/multimodal/vision/mime.go b/internal/runtime/multimodal/vision/mime.go
new file mode 100644
index 0000000..f41151d
--- /dev/null
+++ b/internal/runtime/multimodal/vision/mime.go
@@ -0,0 +1,148 @@
+package vision
+
+import (
+	"fmt"
+	"net/http"
+	"strings"
+)
+
+// sniffBytes 是字节嗅探读取的前缀长度：http.DetectContentType 只看前
+// 512 字节。
+const sniffBytes = 512
+
+// knownRasterMIME 把一个 MIME 归一化到允许内联进视觉模型的光栅图
+// 白名单。白名单与网关上传同口径（internal/services/gateway/upload.go
+// 的 uploadExtForMIME）：SVG/HTML 是可执行内容，绝不进模型。
+// 第二个返回值为 false 时该 MIME 不是白名单内的图像类型。
+func knownRasterMIME(mime string) (string, bool) {
+	switch mediaType(mime) {
+	case "image/png":
+		return "image/png", true
+	case "image/jpeg", "image/jpg":
+		// image/jpg 不是标准写法，但确有其事，归一化到 image/jpeg。
+		return "image/jpeg", true
+	case "image/gif":
+		return "image/gif", true
+	case "image/webp":
+		return "image/webp", true
+	case "image/bmp":
+		return "image/bmp", true
+	default:
+		return "", false
+	}
+}
+
+// imageMediaType 返回 Content-Type 里声明的 image/* 媒体类型（去掉参数、
+// 转小写）。缺失、非 image/*、或载体可执行的 image/svg+xml 都返回 false。
+func imageMediaType(ct string) (string, bool) {
+	m := mediaType(ct)
+	if !strings.HasPrefix(m, "image/") || m == "image/svg+xml" {
+		return "", false
+	}
+	return m, true
+}
+
+// mediaType 归一化一个 MIME 字符串：截掉 ";charset=..." 参数、去空白、转小写。
+func mediaType(mime string) string {
+	return strings.ToLower(strings.TrimSpace(strings.Split(mime, ";")[0]))
+}
+
+// sniffMIME 按真实字节嗅探类型。它是对声明类型（HTTP Content-Type 或文件
+// 扩展名）的否决权：两者都可能骗人，字节不会。
+func sniffMIME(data []byte) string {
+	if len(data) > sniffBytes {
+		data = data[:sniffBytes]
+	}
+	return http.DetectContentType(data)
+}
+
+// isTextualMIME 报告 mime 是否是可读文本 —— HTML/PROXY 错误页、SVG、
+// JSON 等「看起来像图但其实是文本」的载荷都落在这里。
+func isTextualMIME(mime string) bool {
+	m := mediaType(mime)
+	if strings.HasPrefix(m, "text/") {
+		return true
+	}
+	switch m {
+	case "application/json", "application/xml", "application/xhtml+xml", "application/javascript":
+		return true
+	default:
+		return strings.HasSuffix(m, "+xml")
+	}
+}
+
+// imageMIMEFor 决定内联给视觉模型的 MIME。declared 是来源声明的类型
+// （HTTP Content-Type 或文件扩展名），data 是真实字节：
+//
+//   - 字节能判出白名单光栅图 → 以字节为准（声明的类型只作参考，
+//     谎报或错标的类型骗不了模型）；
+//   - 字节是文本（HTML/PROXY 错误页、SVG）→ 报错，绝不喂给模型；
+//   - 字节判不定（application/octet-stream 等）且来源明确声明 image/* →
+//     采信声明（否则 avif 这类嗅探不出的格式会被无谓拒绝）；
+//   - 其余（含类型缺失且字节判不定）→ 报错。
+//
+// 绝不静默回退成 image/png：把一张 HTML 错误页当 PNG 交给视觉模型，
+// 会让模型对着乱码编描述。
+func imageMIMEFor(declared string, data []byte) (string, error) {
+	sniffed := sniffMIME(data)
+	if sniffedMIME, ok := knownRasterMIME(sniffed); ok {
+		return sniffedMIME, nil
+	}
+	declaredMIME, declaredImage := imageMediaType(declared)
+	if isTextualMIME(sniffed) {
+		return "", fmt.Errorf("vision: %w: declared %s but the payload is %s",
+			ErrNotImage, declaredLabel(declared), mediaType(sniffed))
+	}
+	if declaredImage {
+		if known, ok := knownRasterMIME(declaredMIME); ok {
+			return known, nil
+		}
+		return declaredMIME, nil
+	}
+	return "", fmt.Errorf("vision: %w: declared %s, payload is %s",
+		ErrNotImage, declaredLabel(declared), mediaType(sniffed))
+}
+
+// declaredLabel 把声明的类型渲染成错误信息里的可读标签；缺失时给出
+// 明确措辞，避免错误信息里出现空的引号。错误信息只含类型名，
+// 绝不含图像字节或来源 URL（后者可能带凭据）。
+func declaredLabel(declared string) string {
+	if m := mediaType(declared); m != "" {
+		return fmt.Sprintf("content type %q", m)
+	}
+	return "no content type"
+}
+
+// mimeByExt 给出文件路径扩展名声明的 MIME；未知扩展名返回 ""，交给字节
+// 嗅探裁决（不再静默假定 png）。
+func mimeByExt(path string) string {
+	switch strings.ToLower(pathExtension(path)) {
+	case ".png":
+		return "image/png"
+	case ".jpg", ".jpeg":
+		return "image/jpeg"
+	case ".gif":
+		return "image/gif"
+	case ".webp":
+		return "image/webp"
+	case ".bmp":
+		return "image/bmp"
+	default:
+		return ""
+	}
+}
+
+// pathExtension 返回带点号的文件扩展名，没有则返回空串。
+func pathExtension(path string) string {
+	base := path
+	if idx := strings.LastIndexByte(path, '/'); idx >= 0 {
+		base = path[idx+1:]
+	} else if idx := strings.LastIndexByte(path, '\\'); idx >= 0 {
+		base = path[idx+1:]
+	}
+	dot := strings.LastIndexByte(base, '.')
+	if dot < 0 {
+		return ""
+	}
+	return base[dot:]
+}
diff --git a/internal/runtime/multimodal/vision/mime_test.go b/internal/runtime/multimodal/vision/mime_test.go
new file mode 100644
index 0000000..c4f0053
--- /dev/null
+++ b/internal/runtime/multimodal/vision/mime_test.go
@@ -0,0 +1,152 @@
+package vision
+
+import (
+	"bytes"
+	"encoding/base64"
+	"errors"
+	"image"
+	"image/color"
+	"image/png"
+	"testing"
+)
+
+// pngBytes 返回一张真实的最小 PNG（1x1），使「字节嗅探」在测试里
+// 有真实依据，而不是靠手写常量。
+func pngBytes(t *testing.T) []byte {
+	t.Helper()
+	img := image.NewRGBA(image.Rect(0, 0, 1, 1))
+	img.Set(0, 0, color.RGBA{R: 0x11, G: 0x22, B: 0x33, A: 0xff})
+	var buf bytes.Buffer
+	if err := png.Encode(&buf, img); err != nil {
+		t.Fatalf("encode png: %v", err)
+	}
+	return buf.Bytes()
+}
+
+// htmlBytes 模拟网关/PROXY 的错误页：非图像响应体的典型形态。
+func htmlBytes() []byte {
+	return []byte("<!DOCTYPE html><html><body>502 Bad Gateway</body></html>")
+}
+
+// webpBytes 构造一个满足 RIFF/WEBPVP8 嗅探签名的 WebP 头。
+func webpBytes() []byte {
+	return append([]byte("RIFF\x00\x00\x00\x00WEBPVP8 "), make([]byte, 16)...)
+}
+
+// bmpBytes 构造一个满足 "BM" 嗅探签名的 BMP 头。
+func bmpBytes() []byte {
+	return append([]byte("BM"), make([]byte, 32)...)
+}
+
+// TestImageMIMEFor 覆盖「内容类型不能瞎猜」的裁决表：声明的类型只作参考，
+// 真实字节嗅探是否决权，判不出图像一律报错（绝不静默回退 image/png）。
+func TestImageMIMEFor(t *testing.T) {
+	t.Parallel()
+	tests := []struct {
+		name     string
+		declared string
+		data     []byte
+		wantMIME string
+		wantErr  bool
+	}{
+		{name: "png with correct type", declared: "image/png", data: pngBytes(t), wantMIME: "image/png"},
+		{name: "png with charset param", declared: "image/png; charset=binary", data: pngBytes(t), wantMIME: "image/png"},
+		{name: "png with missing type sniffs", declared: "", data: pngBytes(t), wantMIME: "image/png"},
+		{name: "bytes win over a wrong declaration", declared: "image/jpg", data: pngBytes(t), wantMIME: "image/png"},
+		{name: "webp sniffs", declared: "", data: webpBytes(), wantMIME: "image/webp"},
+		{name: "bmp sniffs", declared: "", data: bmpBytes(), wantMIME: "image/bmp"},
+		{name: "html declared as html", declared: "text/html; charset=utf-8", data: htmlBytes(), wantErr: true},
+		{name: "html with missing type", declared: "", data: htmlBytes(), wantErr: true},
+		{name: "html falsely declared png", declared: "image/png", data: htmlBytes(), wantErr: true},
+		{name: "svg rejected", declared: "image/svg+xml", data: []byte(`<svg xmlns="http://www.w3.org/2000/svg"/>`), wantErr: true},
+		{name: "empty payload", declared: "", data: nil, wantErr: true},
+		{name: "unknown binary without type", declared: "", data: []byte{0x00, 0x01, 0x02, 0x03}, wantErr: true},
+		{name: "declared image with unsniffable bytes trusts the declaration",
+			declared: "image/avif", data: []byte{0x00, 0x00, 0x00, 0x20}, wantMIME: "image/avif"},
+		{name: "declared jpg alias normalises in the fallback",
+			declared: "image/jpg", data: []byte{0x00, 0x01, 0x02, 0x03}, wantMIME: "image/jpeg"},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got, err := imageMIMEFor(tt.declared, tt.data)
+			if tt.wantErr {
+				if err == nil {
+					t.Fatalf("imageMIMEFor(%q) = %q, want error", tt.declared, got)
+				}
+				if !errors.Is(err, ErrNotImage) {
+					t.Fatalf("error = %v, want ErrNotImage", err)
+				}
+				return
+			}
+			if err != nil {
+				t.Fatalf("imageMIMEFor(%q) error: %v", tt.declared, err)
+			}
+			if got != tt.wantMIME {
+				t.Fatalf("mime = %q, want %q", got, tt.wantMIME)
+			}
+		})
+	}
+}
+
+// TestMimeByExtLeavesUnknownEmpty 钉住未知扩展名不再被猜成 png —— 交给
+// 字节嗅探裁决。
+func TestMimeByExtLeavesUnknownEmpty(t *testing.T) {
+	t.Parallel()
+	if got := mimeByExt("/tmp/notes.txt"); got != "" {
+		t.Fatalf("mimeByExt(.txt) = %q, want empty (let the bytes decide)", got)
+	}
+	if got := mimeByExt("/tmp/cat.JPEG"); got != "image/jpeg" {
+		t.Fatalf("mimeByExt(.JPEG) = %q, want image/jpeg", got)
+	}
+	if got := mimeByExt("/tmp/shot.bmp"); got != "image/bmp" {
+		t.Fatalf("mimeByExt(.bmp) = %q, want image/bmp", got)
+	}
+}
+
+// TestReadFileRejectsRenamedHTML 钉住本地路径也不喂错类型：把 HTML 改名成
+// .png 不能再被当成 PNG 读进模型。
+func TestReadFileRejectsRenamedHTML(t *testing.T) {
+	t.Parallel()
+	path := t.TempDir() + "/liar.png"
+	if err := writeFileBytes(path, htmlBytes()); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	mime, data, err := readFile(path)
+	if err == nil {
+		t.Fatalf("readFile = (%q, %d bytes), want error", mime, len(data))
+	}
+	if !errors.Is(err, ErrNotImage) {
+		t.Fatalf("error = %v, want ErrNotImage", err)
+	}
+}
+
+// TestReadFileKeepsKnownImage 钉住正常本地图片不退化（含新增的 bmp）。
+func TestReadFileKeepsKnownImage(t *testing.T) {
+	t.Parallel()
+	path := t.TempDir() + "/cat.png"
+	want := pngBytes(t)
+	if err := writeFileBytes(path, want); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	mime, data, err := readFile(path)
+	if err != nil {
+		t.Fatalf("readFile: %v", err)
+	}
+	if mime != "image/png" {
+		t.Fatalf("mime = %q, want image/png", mime)
+	}
+	if !bytes.Equal(data, want) {
+		t.Fatal("bytes changed on read")
+	}
+}
+
+// TestDataURIStable 是 base64 路径逐字节不变的底线。
+func TestDataURIStable(t *testing.T) {
+	t.Parallel()
+	data := pngBytes(t)
+	want := "data:image/png;base64," + base64.StdEncoding.EncodeToString(data)
+	if got := DataURI("image/png", data); got != want {
+		t.Fatalf("DataURI = %q, want %q", got, want)
+	}
+}
diff --git a/internal/runtime/multimodal/vision/resolver.go b/internal/runtime/multimodal/vision/resolver.go
index df55f03..ffe147c 100644
--- a/internal/runtime/multimodal/vision/resolver.go
+++ b/internal/runtime/multimodal/vision/resolver.go
@@ -73,8 +73,16 @@ func (r *MessageResolver) resolveImage(ctx context.Context, imageURL string) (mo
 	decision := r.router.Route(target, RouteRequest{})
 	switch decision.Kind {
 	case KindMain:
-		// 支持视觉的主模型直接看到这张图。
-		return model.Content{Type: model.ContentTypeImage, ImageURL: imageURL}, nil
+		// 支持视觉的主模型直接看到这张图；但 URL 与本地路径必须先由服务端
+		// 取回并内联为 data: URI——云 provider 取不到自部署的局域网地址
+		// （如 http://192.168.x.x:10942/api/v1/files/…），Anthropic 协议更是
+		// 只接受 base64 data: URI（model/anthropic_encode.go 会拒绝裸 URL）。
+		// 取回失败时返回错误，由 loop 保留原消息（该轮不因此中断）。
+		src, err := inlineImageSrc(ctx, imageURL)
+		if err != nil {
+			return model.Content{}, err
+		}
+		return model.Content{Type: model.ContentTypeImage, ImageURL: src}, nil
 	case KindVisionTool:
 		return r.invokeTool(ctx, decision.Tool, map[string]any{
 			"image":  imageURL,
diff --git a/internal/runtime/multimodal/vision/upload.go b/internal/runtime/multimodal/vision/upload.go
new file mode 100644
index 0000000..043b8f4
--- /dev/null
+++ b/internal/runtime/multimodal/vision/upload.go
@@ -0,0 +1,89 @@
+package vision
+
+import (
+	"net/url"
+	"path/filepath"
+	"strings"
+)
+
+// uploadFilePrefix 是本服务提供已存上传的静态路由前缀，与网关
+// internal/services/gateway/upload.go 的 uploadFilePrefix 一致。
+// 它位于 /api/v1/ 之下，因此属于版本化客户端契约。
+const uploadFilePrefix = "/api/v1/files/"
+
+// uploadNameBytes 是上传文件主名的字节数（16 字节 = 32 位十六进制），
+// 与网关 randomUploadName 一致。
+const uploadNameBytes = 16
+
+// localUploadPath 判断 rawURL 是否指向本服务的已存上传，若是则返回它在
+// uploadsDir 下的本地路径。命中时直接读盘：比绕一圈 HTTP 回自己更快，
+// 也不需要网络可达（自部署时同源 URL 用的是局域网地址，服务端自己
+// 未必路由得到，云 provider 则必然取不到）。
+//
+// 认可的形状严格为 <http|https>://<host>/api/v1/files/<name>，且 <name>
+// 必须通过 validUploadName。查询串、片段、多余路径段、任何分隔符或
+// 点号段都在触碰文件系统之前被拒绝；最后的包含性检查是
+// 第二道防线，即使名字校验被改弱，路径也不得逃出 uploadsDir。
+//
+// 注意 host 不参与判断：按网关设计，128 位随机文件名本身就是能力
+// （静态上传路由刻意不认证），因此指向别处的同名 URL 也只解析到
+// 本机同名上传，不会替调用方去取第三方内容。
+func localUploadPath(rawURL, uploadsDir string) (string, bool) {
+	if strings.TrimSpace(uploadsDir) == "" {
+		return "", false
+	}
+	u, err := url.Parse(rawURL)
+	if err != nil {
+		return "", false
+	}
+	if u.Scheme != "http" && u.Scheme != "https" {
+		return "", false
+	}
+	if u.Host == "" || u.RawQuery != "" || u.Fragment != "" {
+		return "", false
+	}
+	name, ok := strings.CutPrefix(u.Path, uploadFilePrefix)
+	if !ok || !validUploadName(name) {
+		return "", false
+	}
+	path := filepath.Join(uploadsDir, name)
+	if filepath.Dir(path) != filepath.Clean(uploadsDir) {
+		return "", false
+	}
+	return path, true
+}
+
+// validUploadName 报告 name 是否恰好是 32 位小写十六进制主名加上白名单
+// 内的图片扩展名。它与网关静态路由的路径穿越防护同口径
+// （upload.go 的 validUploadName）：分隔符、点、双点及任何其他形态都在
+// 触碰文件系统之前被拒绝。
+//
+// 这里不复用网关那份实现是为了不让本包依赖 HTTP 服务层
+// （runtime → services 是反向依赖）；两处规则必须一起改。
+func validUploadName(name string) bool {
+	if name == "" || name != filepath.Base(name) || strings.ContainsAny(name, "/\\") {
+		return false
+	}
+	stem, ext, ok := strings.Cut(name, ".")
+	if !ok || len(stem) != uploadNameBytes*2 || ext == "" {
+		return false
+	}
+	for _, c := range stem {
+		if !isLowerHex(c) {
+			return false
+		}
+	}
+	return knownUploadExt(ext)
+}
+
+// knownUploadExt 报告 ext（不含点）是否是白名单内的图片扩展名，与网关
+// uploadExtForMIME 的可写侧一致：只含光栅图。
+func knownUploadExt(ext string) bool {
+	_, ok := knownRasterMIME(mimeByExt("upload." + strings.ToLower(ext)))
+	return ok
+}
+
+// isLowerHex 报告 c 是否是 ASCII 小写十六进制数字。
+func isLowerHex(c rune) bool {
+	return (c >= '0' && c <= '9') || (c >= 'a' && c <= 'f')
+}
diff --git a/internal/runtime/multimodal/vision/upload_test.go b/internal/runtime/multimodal/vision/upload_test.go
new file mode 100644
index 0000000..594c0ef
--- /dev/null
+++ b/internal/runtime/multimodal/vision/upload_test.go
@@ -0,0 +1,177 @@
+package vision
+
+import (
+	"context"
+	"errors"
+	"path/filepath"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/config"
+)
+
+// uploadTestName 是一个形状合法的上传文件名（32 位小写十六进制 + .png）。
+const uploadTestName = "0123456789abcdef0123456789abcdef.png"
+
+// TestLocalUploadPath 是上传 URL → 本地路径的路径穿越防护用例表：
+// 只有严格形状才被接受，任何分隔符/点号段/编码变体都在触碰文件系统前被拒。
+func TestLocalUploadPath(t *testing.T) {
+	t.Parallel()
+	dir := filepath.Join("tmp", "uploads")
+	tests := []struct {
+		name    string
+		rawURL  string
+		dir     string
+		want    string
+		wantHit bool
+	}{
+		{name: "http upload url", rawURL: "http://192.168.1.5:10942/api/v1/files/" + uploadTestName,
+			dir: dir, want: filepath.Join(dir, uploadTestName), wantHit: true},
+		{name: "https upload url", rawURL: "https://box.local/api/v1/files/" + uploadTestName,
+			dir: dir, want: filepath.Join(dir, uploadTestName), wantHit: true},
+		{name: "traversal", rawURL: "http://h/api/v1/files/../../../data/cyrene.db", dir: dir, wantHit: false},
+		{name: "encoded traversal", rawURL: "http://h/api/v1/files/%2e%2e%2fcyrene.db", dir: dir, wantHit: false},
+		{name: "backslash", rawURL: `http://h/api/v1/files/..\cyrene.db`, dir: dir, wantHit: false},
+		{name: "absolute path name", rawURL: "http://h/api/v1/files//etc/passwd", dir: dir, wantHit: false},
+		{name: "uppercase hex", rawURL: "http://h/api/v1/files/" + strings.ToUpper(uploadTestName), dir: dir, wantHit: false},
+		{name: "short stem", rawURL: "http://h/api/v1/files/abcdef.png", dir: dir, wantHit: false},
+		{name: "non hex stem", rawURL: "http://h/api/v1/files/0123456789abcdef0123456789abcdeg.png", dir: dir, wantHit: false},
+		{name: "html extension", rawURL: "http://h/api/v1/files/" + strings.TrimSuffix(uploadTestName, ".png") + ".html", dir: dir, wantHit: false},
+		{name: "svg extension", rawURL: "http://h/api/v1/files/" + strings.TrimSuffix(uploadTestName, ".png") + ".svg", dir: dir, wantHit: false},
+		{name: "extra segment", rawURL: "http://h/api/v1/files/sub/" + uploadTestName, dir: dir, wantHit: false},
+		{name: "query string", rawURL: "http://h/api/v1/files/" + uploadTestName + "?x=1", dir: dir, wantHit: false},
+		{name: "fragment", rawURL: "http://h/api/v1/files/" + uploadTestName + "#f", dir: dir, wantHit: false},
+		{name: "other route", rawURL: "http://h/api/v1/other/" + uploadTestName, dir: dir, wantHit: false},
+		{name: "ftp scheme", rawURL: "ftp://h/api/v1/files/" + uploadTestName, dir: dir, wantHit: false},
+		{name: "relative url", rawURL: "/api/v1/files/" + uploadTestName, dir: dir, wantHit: false},
+		{name: "remote without uploads dir", rawURL: "http://h/api/v1/files/" + uploadTestName, dir: "", wantHit: false},
+		{name: "empty name", rawURL: "http://h/api/v1/files/", dir: dir, wantHit: false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			t.Parallel()
+			got, ok := localUploadPath(tt.rawURL, tt.dir)
+			if ok != tt.wantHit {
+				t.Fatalf("localUploadPath(%q, %q) = (%q, %v), want hit=%v", tt.rawURL, tt.dir, got, ok, tt.wantHit)
+			}
+			if tt.wantHit && got != tt.want {
+				t.Fatalf("path = %q, want %q", got, tt.want)
+			}
+		})
+	}
+}
+
+// TestOpenAICompatibleAdapter_ReadsOwnUploadLocally 覆盖同源 upload 产物的
+// 快路径：URL 指向本服务的 /api/v1/files/<name> 时直接读盘，不发起网络
+// 请求（URL 的 host 在此不可达，若走了 HTTP 必然失败/超时）。
+func TestOpenAICompatibleAdapter_ReadsOwnUploadLocally(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	dataDir := t.TempDir()
+	uploads := filepath.Join(dataDir, "uploads")
+	if err := writeFileBytesAll(uploads, uploadTestName, png); err != nil {
+		t.Fatalf("seed upload: %v", err)
+	}
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL, WithUploadsDir(uploads))
+
+	// 127.0.0.1:1 上是没有服务的：一旦退回网络取回就会立刻失败。
+	srcURL := "http://127.0.0.1:1/api/v1/files/" + uploadTestName
+	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: srcURL}, ""); err != nil {
+		t.Fatalf("Describe via the local upload path: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 || urls[0] != DataURI("image/png", png) {
+		t.Fatalf("image_url = %q, want the local file inlined", urls)
+	}
+}
+
+// TestOpenAICompatibleAdapter_MissingUploadFailsLoudly 钉住「认得的上传 URL
+// 但文件不在」不会偷偷退回网络：错误如实上报（不静默换成别的内容）。
+func TestOpenAICompatibleAdapter_MissingUploadFailsLoudly(t *testing.T) {
+	t.Parallel()
+	uploads := filepath.Join(t.TempDir(), "uploads")
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL, WithUploadsDir(uploads))
+
+	srcURL := "http://127.0.0.1:1/api/v1/files/" + uploadTestName
+	_, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: srcURL}, "")
+	if err == nil {
+		t.Fatal("Describe succeeded without the upload on disk, want an error")
+	}
+	if !errors.Is(err, ErrBadImage) {
+		t.Fatalf("error = %v, want ErrBadImage", err)
+	}
+	if calls, _ := rec.seen(); calls != 0 {
+		t.Fatalf("upstream was called %d times, want 0", calls)
+	}
+}
+
+// TestOpenAICompatibleAdapter_TraversalNotReadLocally 是越界文件名的反证
+// 用例：本地 secret.png 与远端返回的图内容不同，`../` 名称必须落到远端
+// 取回（而不是被读成本地越界文件）。
+func TestOpenAICompatibleAdapter_TraversalNotReadLocally(t *testing.T) {
+	t.Parallel()
+	local := pngBytes(t)
+	remote := append([]byte{}, pngBytes(t)...)
+	remote = append(remote, 0x00) // 内容不同即可，仍是合法 PNG 前缀
+	var hits atomic.Int32
+
+	dataDir := t.TempDir()
+	uploads := filepath.Join(dataDir, "uploads")
+	if err := writeFileBytesAll(uploads, uploadTestName, local); err != nil {
+		t.Fatalf("seed uploads dir: %v", err)
+	}
+	if err := writeFileBytes(filepath.Join(dataDir, "secret.png"), local); err != nil {
+		t.Fatalf("seed secret: %v", err)
+	}
+
+	imgSrv := newImageServer(t, remote, "image/png", &hits)
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL, WithUploadsDir(uploads))
+
+	srcURL := imgSrv.URL + "/api/v1/files/../secret.png"
+	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: srcURL}, ""); err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 {
+		t.Fatalf("image_url = %v, want one part", urls)
+	}
+	if urls[0] == DataURI("image/png", local) {
+		t.Fatal("the ../ name was resolved to a local file outside uploads")
+	}
+	if urls[0] != DataURI("image/png", remote) {
+		t.Fatalf("image_url = %q, want the remote payload", urls[0])
+	}
+	if hits.Load() != 1 {
+		t.Fatalf("source fetched %d times, want 1", hits.Load())
+	}
+}
+
+// TestWithUploadsDirIsOptional 钉住未配置上传目录时行为不变：上传形状的
+// URL 仍按普通 URL 取回。
+func TestWithUploadsDirIsOptional(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	var hits atomic.Int32
+	imgSrv := newImageServer(t, png, "image/png", &hits)
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := NewOpenAICompatible(config.LLMConfig{BaseURL: upSrv.URL, Model: "vision-test"})
+
+	_, err := ad.Describe(context.Background(), ImageRef{
+		Source: SourceURL,
+		Data:   imgSrv.URL + "/api/v1/files/" + uploadTestName,
+	}, "")
+	if err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 || urls[0] != DataURI("image/png", png) {
+		t.Fatalf("image_url = %q, want the fetched payload", urls)
+	}
+	if hits.Load() != 1 {
+		t.Fatalf("source fetched %d times, want 1", hits.Load())
+	}
+}
diff --git a/internal/runtime/multimodal/vision/url_test.go b/internal/runtime/multimodal/vision/url_test.go
new file mode 100644
index 0000000..6cf021b
--- /dev/null
+++ b/internal/runtime/multimodal/vision/url_test.go
@@ -0,0 +1,192 @@
+package vision
+
+import (
+	"context"
+	"errors"
+	"strings"
+	"sync/atomic"
+	"testing"
+
+	"git.yeij.top/AskaEth/Cyrene_v2/internal/runtime/model"
+)
+
+// TestOpenAICompatibleAdapter_InlinesFetchedURL 是本缺陷的主用例：来源是
+// http(s) URL 时，服务端必须自己把图取回并内联成 data: URI 再交给
+// provider —— 原样透传同源（局域网）URL 时云 provider 根本取不到。
+func TestOpenAICompatibleAdapter_InlinesFetchedURL(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	var hits atomic.Int32
+	imgSrv := newImageServer(t, png, "image/png", &hits)
+	srcURL := imgSrv.URL + "/cat.png"
+
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	text, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: srcURL}, "")
+	if err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	if text != "a small square" {
+		t.Fatalf("text = %q", text)
+	}
+	calls, urls := rec.seen()
+	if calls != 1 || len(urls) != 1 {
+		t.Fatalf("upstream calls = %d urls = %v, want one image_url", calls, urls)
+	}
+	if want := DataURI("image/png", png); urls[0] != want {
+		t.Fatalf("image_url = %q, want the inlined %q", urls[0], want)
+	}
+	if strings.Contains(urls[0], srcURL) || strings.HasPrefix(urls[0], "http") {
+		t.Fatalf("image_url still points at the source URL behaviour: %q", urls[0])
+	}
+	if hits.Load() != 1 {
+		t.Fatalf("source fetched %d times, want 1", hits.Load())
+	}
+}
+
+// TestOpenAICompatibleAdapter_RejectsHTMLResponse 是非图像响应体的用例：
+// 网关/PROXY 的 text/html 错误页必须报错，而不是被当成 image/png
+// 喂给视觉模型；错误信息里不得出现页面内容（不打印图像内容）。
+func TestOpenAICompatibleAdapter_RejectsHTMLResponse(t *testing.T) {
+	t.Parallel()
+	imgSrv := newImageServer(t, htmlBytes(), "text/html; charset=utf-8", nil)
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	_, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: imgSrv.URL + "/err"}, "")
+	if err == nil {
+		t.Fatal("Describe succeeded on a text/html payload, want an error")
+	}
+	if !errors.Is(err, ErrNotImage) {
+		t.Fatalf("error = %v, want ErrNotImage", err)
+	}
+	if strings.Contains(err.Error(), "502 Bad Gateway") {
+		t.Fatalf("error leaks the response body: %v", err)
+	}
+	if calls, _ := rec.seen(); calls != 0 {
+		t.Fatalf("upstream was called %d times, want 0 (fail before the model)", calls)
+	}
+}
+
+// TestOpenAICompatibleAdapter_RejectsHTMLMislabelledAsPNG 覆盖缺陷第 2 条：
+// 服务端谎报 image/png 时，字节嗅探仍必须否决它。
+func TestOpenAICompatibleAdapter_RejectsHTMLMislabelledAsPNG(t *testing.T) {
+	t.Parallel()
+	imgSrv := newImageServer(t, htmlBytes(), "image/png", nil)
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	_, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: imgSrv.URL + "/liar.png"}, "")
+	if !errors.Is(err, ErrNotImage) {
+		t.Fatalf("error = %v, want ErrNotImage", err)
+	}
+	if calls, _ := rec.seen(); calls != 0 {
+		t.Fatalf("upstream was called %d times, want 0", calls)
+	}
+}
+
+// TestOpenAICompatibleAdapter_MissingContentTypeSniffs 覆盖「类型缺失」：
+// 真实 PNG 字节仍应被字节嗅探救回，且内联的 MIME 正确。
+func TestOpenAICompatibleAdapter_MissingContentTypeSniffs(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	imgSrv := newImageServer(t, png, "", nil)
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourceURL, Data: imgSrv.URL + "/n"}, ""); err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 || urls[0] != DataURI("image/png", png) {
+		t.Fatalf("image_url = %v, want the sniffed png inlined", urls)
+	}
+}
+
+// TestOpenAICompatibleAdapter_Base64Unchanged 是行为不退化的底线：客户端
+// 现在用的 data: URI 路径必须逐字节不变地继续工作。
+func TestOpenAICompatibleAdapter_Base64Unchanged(t *testing.T) {
+	t.Parallel()
+	dataURI := DataURI("image/png", pngBytes(t))
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourceBase64, Data: dataURI}, ""); err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 || urls[0] != dataURI {
+		t.Fatalf("image_url = %q, want the input data: URI byte for byte", urls[0])
+	}
+}
+
+// TestOpenAICompatibleAdapter_LocalPathUnchanged 钉住本地路径来源同样被
+// 内联且字节不变。
+func TestOpenAICompatibleAdapter_LocalPathUnchanged(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	path := t.TempDir() + "/cat.png"
+	if err := writeFileBytes(path, png); err != nil {
+		t.Fatalf("write: %v", err)
+	}
+	upSrv, rec := newImageURLRecorderServer(t)
+	ad := newAdapter(upSrv.URL)
+
+	if _, err := ad.Describe(context.Background(), ImageRef{Source: SourcePath, Data: path}, ""); err != nil {
+		t.Fatalf("Describe: %v", err)
+	}
+	_, urls := rec.seen()
+	if len(urls) != 1 || urls[0] != DataURI("image/png", png) {
+		t.Fatalf("image_url = %q, want the file inlined", urls)
+	}
+}
+
+// TestMessageResolver_InlinesURLForMainModel 覆盖主模型直连那条支路：
+// 支持视觉的主模型拿到的也必须是内联后的 data: URI（Anthropic 协议更是
+// 只接受 data: URI），而不是原 URL。
+func TestMessageResolver_InlinesURLForMainModel(t *testing.T) {
+	t.Parallel()
+	png := pngBytes(t)
+	imgSrv := newImageServer(t, png, "image/png", nil)
+	res := NewMessageResolver(NewRouter(WithVisionTool("vision")), func() ModelInfo {
+		return ModelInfo{Name: "vision-model", SupportsVision: true}
+	}, func(_ context.Context, _ string, _ map[string]any) (any, error) {
+		t.Fatal("tool must not be invoked for a sighted main model")
+		return nil, nil
+	})
+
+	msg, err := res.Resolve(context.Background(), model.ImageMessage(model.RoleUser, "x", imgSrv.URL+"/cat.png"))
+	if err != nil {
+		t.Fatalf("Resolve: %v", err)
+	}
+	var got string
+	for _, p := range msg.Content {
+		if p.Type == model.ContentTypeImage {
+			got = p.ImageURL
+		}
+	}
+	if want := DataURI("image/png", png); got != want {
+		t.Fatalf("image part = %q, want %q", got, want)
+	}
+}
+
+// TestMessageResolver_KeepsDataURIByteForByte 钉住解析器对 data: URI 的
+// 短路：客户端送来的 base64 图不被重编码。
+func TestMessageResolver_KeepsDataURIByteForByte(t *testing.T) {
+	t.Parallel()
+	in := DataURI("image/png", pngBytes(t))
+	res := NewMessageResolver(NewRouter(WithVisionTool("vision")), func() ModelInfo {
+		return ModelInfo{Name: "vision-model", SupportsVision: true}
+	}, nil)
+
+	msg, err := res.Resolve(context.Background(), model.ImageMessage(model.RoleUser, "x", in))
+	if err != nil {
+		t.Fatalf("Resolve: %v", err)
+	}
+	for _, p := range msg.Content {
+		if p.Type == model.ContentTypeImage && p.ImageURL != in {
+			t.Fatalf("image part = %q, want the untouched input", p.ImageURL)
+		}
+	}
+}