mcp-duojie-image 0.8.0 → 0.9.0

package/README.md

@@ -5,9 +5,7 @@ MCP server for image generation via duojie.games GPT image API.
 ## Features
 
 - Text-to-image generation with `gpt-image-2`
-- Reference image support with role-based endpoint routing:
-  - `reference` / `style_reference` / `composition_reference` / `insert_source` → `/v1/images/generations` (JSON body + base64 data URIs; stable endpoint)
-  - `edit_target` → `/v1/images/edits` (multipart form; for modifying an existing image)
+- Reference image support: requests with reference images route through `/v1/images/edits` (multipart form); text-only requests route through `/v1/images/generations` (JSON body)
 - Prompt structuring helpers for use case, asset type, constraints, and effective prompt echo
 - Optional chroma-key transparent background post-processing
 - Async submission with task status lookup
@@ -63,7 +61,7 @@ Supported `input_images.role` values:
 - `composition_reference`
 - `insert_source`
 
-**Endpoint routing:** Images with the `edit_target` role are sent to `/v1/images/edits` (multipart form). All other roles (`reference`, `style_reference`, `composition_reference`, `insert_source`) are sent to `/v1/images/generations` as base64 data URIs in a JSON body. This mirrors Codex's imagegen skill, which treats style/composition/mood references as "generate" rather than "edit". If a request mixes `edit_target` with other roles, the edits endpoint is used.
+**Endpoint routing:** Any request with reference images is sent to `/v1/images/edits` as a multipart form, which correctly handles both reference images and the `size`/`aspect_ratio` parameters. Text-only requests (no reference images) are sent to `/v1/images/generations` as a JSON body. The `role` field on `input_images` is used only for prompt structuring, not for endpoint selection.
 
 When both `input_images` and `reference_images` are provided, `input_images` takes precedence. `generate_image` returns `effective_prompt` so callers can diagnose the exact prompt sent upstream.
 

package/dist/index.js

@@ -34,7 +34,7 @@ const server = new McpServer({
     version: "0.1.0",
 });
 server.registerTool("generate_image", {
-    description: "Generate image async with gpt-image-2. Reference/style/composition images route through /v1/images/generations (stable); edit_target role routes through /v1/images/edits. Return after tool call; don't wait/check.",
+    description: "Generate image async with gpt-image-2. Reference images route through /v1/images/edits (multipart form); text-only requests route through /v1/images/generations. Return after tool call; don't wait/check.",
     inputSchema: z.object({
         prompt: z.string().min(1).describe("Prompt"),
         output_dir: z.string().describe("Absolute output dir"),
@@ -105,8 +105,6 @@ server.registerTool("generate_image", {
     }
     const task = createTask();
     updateTask(task.id, { effectivePrompt: preparedPrompt.prompt });
-    // 只有 edit_target 角色才走 edits 端点；reference/style/composition 走 generations
-    const hasEditTarget = preparedPrompt.imageInputs?.some((img) => img.role === "edit_target") ?? false;
     // Start background generation (don't await)
     (async () => {
         try {
@@ -114,7 +112,6 @@ server.registerTool("generate_image", {
             let imageBuffer = await generateImage(config, {
                 prompt: preparedPrompt.prompt,
                 referenceImages: preparedPrompt.imagePaths,
-                useEditsEndpoint: hasEditTarget,
                 size: args.size ?? "1K",
                 aspectRatio: args.aspect_ratio,
                 referenceMaxSide: args.preserve_reference_detail ? false : args.reference_max_side,

package/dist/provider.js

@@ -105,16 +105,16 @@ async function generateOpenAIImage(config, opts) {
     const prompt = opts.prompt;
     const size = mapOpenAISize(opts.size, opts.aspectRatio);
     const hasReferenceImages = Boolean(opts.referenceImages?.length);
-    // edit_target → /v1/images/edits (multipart form)
-    // reference/style/composition → /v1/images/generations (JSON body + base64 data URIs)
-    const useEdits = hasReferenceImages && opts.useEditsEndpoint === true;
+    // 有参考图 → /v1/images/edits（multipart form）
+    // 无参考图 → /v1/images/generations（JSON body，纯文本）
+    const useEdits = hasReferenceImages;
     const url = `${config.baseUrl}${useEdits ? "/v1/images/edits" : "/v1/images/generations"}`;
     let body;
     let headers = {
         Authorization: `Bearer ${config.apiKey}`,
     };
     if (useEdits) {
-        // edit_target: multipart form（修改已有图片）
+        // 有参考图：multipart form
         const form = new FormData();
         form.set("model", GPT_IMAGE_MODEL);
         form.set("prompt", prompt);
@@ -129,22 +129,6 @@ async function generateOpenAIImage(config, opts) {
         }
         body = form;
     }
-    else if (hasReferenceImages) {
-        // reference/style/composition: generations + base64 data URIs in JSON body
-        const imageUrls = [];
-        for (const imgPath of opts.referenceImages) {
-            const data = await resizeIfNeeded(imgPath, opts.referenceMaxSide);
-            imageUrls.push(`data:image/webp;base64,${data.toString("base64")}`);
-        }
-        headers = { ...headers, "Content-Type": "application/json" };
-        body = JSON.stringify({
-            model: GPT_IMAGE_MODEL,
-            prompt,
-            size,
-            image: imageUrls,
-            ...(opts.aspectRatio ? { aspect_ratio: opts.aspectRatio } : {}),
-        });
-    }
     else {
         // 纯文本生成
         headers = { ...headers, "Content-Type": "application/json" };

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mcp-duojie-image",
-  "version": "0.8.0",
+  "version": "0.9.0",
   "description": "MCP server for image generation via duojie.games",
   "type": "module",
   "main": "dist/index.js",