@mindstudio-ai/remy 0.1.197 → 0.1.199

package/dist/headless.js

@@ -4232,7 +4232,15 @@ var SYSTEM_PROMPT = readAsset(
   "subagents/designExpert/tools/images/enhance-image-prompt.md"
 );
 async function enhanceImagePrompt(params) {
-  const { brief, width, height, transparentBackground, onLog, model } = params;
+  const {
+    brief,
+    width,
+    height,
+    transparentBackground,
+    hasReferenceImage,
+    onLog,
+    model
+  } = params;
   const contextParts = [
     `Dimensions: ${width}x${height}${width > height ? " (landscape)" : width < height ? " (portrait)" : " (square)"}`
   ];
@@ -4241,6 +4249,11 @@ async function enhanceImagePrompt(params) {
       "Transparent background: yes \u2014 the background will be removed. Focus on the subject as an isolated element."
     );
   }
+  if (hasReferenceImage) {
+    contextParts.push(
+      "Reference image: yes \u2014 a reference image is provided to the generation model alongside this prompt to guide style, subject, or composition. Complement it; don't re-describe what it already carries."
+    );
+  }
   const context = `<context>
 ${contextParts.join("\n")}
 </context>`;
@@ -4271,6 +4284,7 @@ async function generateImageAssets(opts) {
     prompts,
     sourceImages,
     transparentBackground,
+    enhancePrompts,
     onLog,
     imageGenerationModel: genModel,
     imageAnalysisModel,
@@ -4280,21 +4294,29 @@ async function generateImageAssets(opts) {
   const height = opts.height || 2048;
   const config = { width, height };
   if (sourceImages?.length) {
+    const [firstImage] = sourceImages;
     config.images = sourceImages;
-  }
-  const isEdit = !!sourceImages?.length;
-  const enhancedPrompts = isEdit ? prompts : await Promise.all(
+    config.source_images = sourceImages;
+    config.image_ref = sourceImages;
+    config.image = firstImage;
+    config.image_url = firstImage;
+    config.source_image = firstImage;
+    config.source = firstImage;
+  }
+  const hasReference = !!sourceImages?.length;
+  const enhancedPrompts = enhancePrompts ? await Promise.all(
     prompts.map(
       (brief) => enhanceImagePrompt({
         brief,
         width,
         height,
         transparentBackground,
+        hasReferenceImage: hasReference,
         onLog,
         model: imagePromptEnhancerModel
       })
     )
-  );
+  ) : prompts;
   let imageUrls;
   if (enhancedPrompts.length === 1) {
     const step = JSON.stringify({
@@ -4364,7 +4386,7 @@ async function generateImageAssets(opts) {
       if (url.startsWith("Error")) {
         return {
           prompt: prompts[i],
-          ...!isEdit && { enhancedPrompt: enhancedPrompts[i] },
+          ...enhancePrompts && { enhancedPrompt: enhancedPrompts[i] },
           error: url
         };
       }
@@ -4377,7 +4399,7 @@ async function generateImageAssets(opts) {
       return {
         url,
         prompt: prompts[i],
-        ...!isEdit && { enhancedPrompt: enhancedPrompts[i] },
+        ...enhancePrompts && { enhancedPrompt: enhancedPrompts[i] },
         analysis,
         width,
         height
@@ -4402,6 +4424,10 @@ var definition6 = {
         },
         description: "One or more image briefs describing what you want. Focus on subject, mood, style, and intended use \u2014 the tool optimizes your brief into a model-ready prompt automatically. Multiple briefs run in parallel."
       },
+      referenceImage: {
+        type: "string",
+        description: "Optional URL of a single reference image to guide the generation \u2014 for style, subject, character consistency, or composition. Your prompt still describes the desired result; the reference conditions it. Applies to every prompt in the batch."
+      },
       width: {
         type: "number",
         description: "Image width in pixels. Default 2048. Range: 2048-4096."
@@ -4424,6 +4450,8 @@ async function execute6(input, onLog, context) {
     width: input.width,
     height: input.height,
     transparentBackground: input.transparentBackground,
+    sourceImages: input.referenceImage ? [input.referenceImage] : void 0,
+    enhancePrompts: true,
     onLog,
     imageGenerationModel: resolveModel(
       "imageGeneration",
@@ -4493,6 +4521,7 @@ async function execute7(input, onLog, context) {
     width: input.width,
     height: input.height,
     transparentBackground: input.transparentBackground,
+    enhancePrompts: false,
     onLog,
     imageGenerationModel: resolveModel(
       "imageGeneration",

package/dist/index.js

@@ -4957,7 +4957,15 @@ var init_screenshot3 = __esm({
 
 // src/subagents/designExpert/tools/images/enhancePrompt.ts
 async function enhanceImagePrompt(params) {
-  const { brief, width, height, transparentBackground, onLog, model } = params;
+  const {
+    brief,
+    width,
+    height,
+    transparentBackground,
+    hasReferenceImage,
+    onLog,
+    model
+  } = params;
   const contextParts = [
     `Dimensions: ${width}x${height}${width > height ? " (landscape)" : width < height ? " (portrait)" : " (square)"}`
   ];
@@ -4966,6 +4974,11 @@ async function enhanceImagePrompt(params) {
       "Transparent background: yes \u2014 the background will be removed. Focus on the subject as an isolated element."
     );
   }
+  if (hasReferenceImage) {
+    contextParts.push(
+      "Reference image: yes \u2014 a reference image is provided to the generation model alongside this prompt to guide style, subject, or composition. Complement it; don't re-describe what it already carries."
+    );
+  }
   const context = `<context>
 ${contextParts.join("\n")}
 </context>`;
@@ -5006,6 +5019,7 @@ async function generateImageAssets(opts) {
     prompts,
     sourceImages,
     transparentBackground,
+    enhancePrompts,
     onLog,
     imageGenerationModel: genModel,
     imageAnalysisModel,
@@ -5015,21 +5029,29 @@ async function generateImageAssets(opts) {
   const height = opts.height || 2048;
   const config = { width, height };
   if (sourceImages?.length) {
+    const [firstImage] = sourceImages;
     config.images = sourceImages;
-  }
-  const isEdit = !!sourceImages?.length;
-  const enhancedPrompts = isEdit ? prompts : await Promise.all(
+    config.source_images = sourceImages;
+    config.image_ref = sourceImages;
+    config.image = firstImage;
+    config.image_url = firstImage;
+    config.source_image = firstImage;
+    config.source = firstImage;
+  }
+  const hasReference = !!sourceImages?.length;
+  const enhancedPrompts = enhancePrompts ? await Promise.all(
     prompts.map(
       (brief) => enhanceImagePrompt({
         brief,
         width,
         height,
         transparentBackground,
+        hasReferenceImage: hasReference,
         onLog,
         model: imagePromptEnhancerModel
       })
     )
-  );
+  ) : prompts;
   let imageUrls;
   if (enhancedPrompts.length === 1) {
     const step = JSON.stringify({
@@ -5099,7 +5121,7 @@ async function generateImageAssets(opts) {
       if (url.startsWith("Error")) {
         return {
           prompt: prompts[i],
-          ...!isEdit && { enhancedPrompt: enhancedPrompts[i] },
+          ...enhancePrompts && { enhancedPrompt: enhancedPrompts[i] },
           error: url
         };
       }
@@ -5112,7 +5134,7 @@ async function generateImageAssets(opts) {
       return {
         url,
         prompt: prompts[i],
-        ...!isEdit && { enhancedPrompt: enhancedPrompts[i] },
+        ...enhancePrompts && { enhancedPrompt: enhancedPrompts[i] },
         analysis,
         width,
         height
@@ -5144,6 +5166,8 @@ async function execute6(input, onLog, context) {
     width: input.width,
     height: input.height,
     transparentBackground: input.transparentBackground,
+    sourceImages: input.referenceImage ? [input.referenceImage] : void 0,
+    enhancePrompts: true,
     onLog,
     imageGenerationModel: resolveModel(
       "imageGeneration",
@@ -5182,6 +5206,10 @@ var init_generateImages = __esm({
             },
             description: "One or more image briefs describing what you want. Focus on subject, mood, style, and intended use \u2014 the tool optimizes your brief into a model-ready prompt automatically. Multiple briefs run in parallel."
           },
+          referenceImage: {
+            type: "string",
+            description: "Optional URL of a single reference image to guide the generation \u2014 for style, subject, character consistency, or composition. Your prompt still describes the desired result; the reference conditions it. Applies to every prompt in the batch."
+          },
           width: {
             type: "number",
             description: "Image width in pixels. Default 2048. Range: 2048-4096."
@@ -5214,6 +5242,7 @@ async function execute7(input, onLog, context) {
     width: input.width,
     height: input.height,
     transparentBackground: input.transparentBackground,
+    enhancePrompts: false,
     onLog,
     imageGenerationModel: resolveModel(
       "imageGeneration",

package/dist/subagents/designExpert/tools/images/enhance-image-prompt.md

@@ -15,12 +15,19 @@ Examples of good density:
 These are non-negotiable. Violating them produces bad output.
 
 - **No hex codes.** The model renders hex codes as visible text in the image. Describe colors by name and relationship: "deep emerald green with a smooth satin finish" or "warm sand beige fading into pale desaturated blue" — never "#7C3AED".
-- **No quoted strings.** Any single or double quoted string gets rendered as literal text in the image.
 - **No physical object framing.** Words like "artwork", "painting", "canvas", "print", "app icon", "square digital artwork" produce photorealistic mockups of a painting in a frame or an icon inset on a background. Describe the visual content directly.
-- **No text triggers.** Words like "poster", "magazine cover", "editorial spread", "sign", or brand names risk rendering literal text, mastheads, or mockup layouts. If you want an editorial photography *style*, describe the photographic qualities — not the format.
 - **Describe what you want, not what you don't want.** Negation doesn't work — "street with no cars" activates "cars." Say "empty street" instead.
 - **No body part positioning.** Don't describe specific arrangements of arms, legs, or limbs.
-- **No brand names.** Things like "Apple style" or "Nintendo style" will generate literal logos in the output.
+- **No other brands as a style shortcut.** Don't borrow another company's identity as shorthand — "Apple style", "Nintendo style" — it renders that company's literal logo. (A brand's *own* name or wordmark on its *own* asset is intended text, not this — see Text & wordmarks below.)
+
+## Text & wordmarks
+
+The model renders text well — but only the text you tell it to, so quotation marks mean "render this literally." Use them deliberately.
+
+- **Reproduce intended text exactly.** When the brief names a wordmark, brand name, label, sign, headline, or UI copy, carry the exact string through in quotes — e.g. a wordmark reading "Solid Credit". Never drop it, paraphrase it, or genericize it to "a wordmark": the literal text is usually the whole point of the asset, and if you omit it the model fills the space with an invented placeholder.
+- **Direct the typography and placement.** Specify weight, case, color, and position so the text lands where the designer wants it — "a near-black grotesque sans-serif wordmark reading 'Solid Credit', centered directly below the mark."
+- **Keep it short.** Wordmarks, labels, and short taglines render reliably; full sentences and paragraphs degrade into garbled glyphs. Trim long copy to the few words that matter, or leave it out.
+- **Don't summon text you don't want.** When the image should have no text, don't quote stray descriptive phrases, and avoid format words that imply copy — "poster", "magazine cover", "sign", "billboard" — which can produce spurious text or mastheads. For an editorial *style*, describe the photographic qualities, not the format.
 
 ## Composition
 
@@ -34,6 +41,7 @@ You'll receive context about the generation parameters. Use them:
 
 - **Dimensions**: If the image is wide (landscape), compose horizontally. If tall (portrait), compose vertically. If square, center the subject.
 - **Transparent background**: The background will be removed after generation and the image will be trimmed to the subject bounds (no extra padding). Don't describe elaborate backgrounds — focus on the subject. Describe it as an isolated element.
+- **Reference image**: When a reference image is provided, the generation model receives it alongside your prompt to guide style, subject, or composition. Write the prompt to *complement* the reference, not duplicate it: describe the scene, action, and anything new or changed, and lean on the reference for what it already establishes (a specific face, product, logo, or art style). Don't exhaustively re-describe those — over-specifying competes with the reference image and can distort it.
 
 ## Photography prompts
 

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@mindstudio-ai/remy",
-  "version": "0.1.197",
+  "version": "0.1.199",
   "description": "MindStudio coding agent",
   "repository": {
     "type": "git",