mcp-duojie-image 0.6.0 → 0.8.0

package/README.md

@@ -5,7 +5,9 @@ MCP server for image generation via duojie.games GPT image API.
 ## Features
 
 - Text-to-image generation with `gpt-image-2`
-- Reference image editing via local image paths, with optional image roles
+- 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)
 - 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
@@ -61,6 +63,8 @@ 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.
+
 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.
 
 Response:

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. Return after tool call; don't wait/check.",
+    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.",
     inputSchema: z.object({
         prompt: z.string().min(1).describe("Prompt"),
         output_dir: z.string().describe("Absolute output dir"),
@@ -105,6 +105,8 @@ 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 {
@@ -112,6 +114,7 @@ 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/prompt.js

@@ -83,5 +83,6 @@ export function preparePrompt(opts) {
     return {
         prompt,
         imagePaths: inputImages?.map((image) => image.path) ?? [],
+        imageInputs: inputImages,
     };
 }

package/dist/provider.js

@@ -10,11 +10,28 @@ async function fetchWithRetry(url, init, timeoutMs, retries = 0) {
         try {
             const res = await fetch(url, { ...init, signal: controller.signal });
             clearTimeout(timer);
+            // 5xx: 上游随机故障(EOF等)，立刻 abort body 连接并重试，不等 chunked 终止
+            if (res.status >= 500 && attempt < retries) {
+                try {
+                    controller.abort();
+                }
+                catch { }
+                const delayMs = 1000 * Math.pow(2, attempt) + Math.random() * 500;
+                await new Promise((r) => setTimeout(r, delayMs));
+                continue;
+            }
             return res;
         }
         catch (err) {
             clearTimeout(timer);
             const isAbort = err instanceof Error && err.name === "AbortError";
+            // AbortError from our manual abort above is not a real timeout
+            if (isAbort && attempt < retries) {
+                lastErr = err;
+                const delayMs = 1000 * Math.pow(2, attempt) + Math.random() * 500;
+                await new Promise((r) => setTimeout(r, delayMs));
+                continue;
+            }
             if (isAbort) {
                 throw new AppError("E_PROVIDER_TIMEOUT", "Request timed out");
             }
@@ -77,6 +94,10 @@ async function handleProviderError(res) {
     if (res.status === 429) {
         throw new AppError("E_PROVIDER_RATE_LIMIT", "Rate limit exceeded");
     }
+    // 5xx: 上游随机故障，不读 body（chunked 连接挂 120s），直接秒回
+    if (res.status >= 500) {
+        throw new AppError("E_PROVIDER_BAD_REQUEST", `Upstream error ${res.status} (body skipped to avoid chunked hang)`);
+    }
     const text = await res.text().catch(() => "");
     throw new AppError("E_PROVIDER_BAD_REQUEST", `Upstream error ${res.status}: ${text.slice(0, 200)}`);
 }
@@ -84,12 +105,16 @@ async function generateOpenAIImage(config, opts) {
     const prompt = opts.prompt;
     const size = mapOpenAISize(opts.size, opts.aspectRatio);
     const hasReferenceImages = Boolean(opts.referenceImages?.length);
-    const url = `${config.baseUrl}${hasReferenceImages ? "/v1/images/edits" : "/v1/images/generations"}`;
+    // edit_target → /v1/images/edits (multipart form)
+    // reference/style/composition → /v1/images/generations (JSON body + base64 data URIs)
+    const useEdits = hasReferenceImages && opts.useEditsEndpoint === true;
+    const url = `${config.baseUrl}${useEdits ? "/v1/images/edits" : "/v1/images/generations"}`;
     let body;
     let headers = {
         Authorization: `Bearer ${config.apiKey}`,
     };
-    if (hasReferenceImages) {
+    if (useEdits) {
+        // edit_target: multipart form（修改已有图片）
         const form = new FormData();
         form.set("model", GPT_IMAGE_MODEL);
         form.set("prompt", prompt);
@@ -104,7 +129,24 @@ async function generateOpenAIImage(config, opts) {
         }
         body = form;
     }
+    else if (hasReferenceImages) {
+        // reference/style/composition: generations + base64 data URIs in JSON body
+        const imageUrls = [];
+        for (const imgPath of opts.referenceImages) {
+            const data = await resizeIfNeeded(imgPath, opts.referenceMaxSide);
+            imageUrls.push(`data:image/webp;base64,${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" };
         body = JSON.stringify({
             model: GPT_IMAGE_MODEL,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "mcp-duojie-image",
-  "version": "0.6.0",
+  "version": "0.8.0",
   "description": "MCP server for image generation via duojie.games",
   "type": "module",
   "main": "dist/index.js",