@goodandready/dsh-image-gen 0.9.0 → 0.10.0

package/README.md

@@ -31,24 +31,35 @@
 
 ---
 
+## 🛡️ Reliability, Performance & Quality (v0.10.0)
+
+* **Exponential Backoff with Jitter**: Adaptive polling for FAL, Replicate, and ComfyUI queues protects against HTTP 429 rate limits.
+* **Error Classification in Fallback Cascade**: Client-side errors (Content Policy, 400 Bad Request, NSFW) fail fast without wasting API credits on other providers.
+* **Deterministic Hash Caching**: Exact matches of prompt, model, and seed return instantly from local storage with zero API expense.
+* **ComfyUI & Automatic1111 Drag-and-Drop**: Metadata is packed into PNG `Parameters` chunks in standard format.
+* **Dimension Snapping**: Automatic normalization to multiples of 64 guarantees VAE bucket compatibility.
+* **Enhanced Style Presets**: Built-in styles include tailor-made negative prompts and optimal guidance scale settings.
+
+---
+
 ## 🛠️ Complete Tools Reference
 
 * **`generate_image`**: Generate images with pluggable providers, seeds, aspect ratios, and style presets.
 * **`remove_background`**: Extract subject with transparent PNG output (FAL BiRefNet / Rembg).
 * **`upscale_image`**: 2x / 4x super-resolution with clarity reconstruction.
-* **`vectorize_image`**: Convert raster graphics to clean scalable SVG vectors.
+* **`vectorize_image`**: Convert raster graphics to clean scalable SVG vectors with palette quantization.
 * **`blend_images`**: Multi-reference composition mixing.
-* **`generate_image_pack`**: Simultaneous multi-aspect ratio rendering (1:1, 16:9, 9:16).
+* **`generate_image_pack`**: Simultaneous multi-aspect ratio rendering with graceful partial recovery.
 * **`compare_images`**: Pixel-level visual difference ratio comparison.
-* **`inspect_image_quality`**: Automated visual audit and defect detection.
+* **`inspect_image_quality`**: Automated visual audit, Laplacian sharpness scoring, and defect detection.
 
 ---
 
 ## 🎨 Supported Generation Backends
 
-* **`fal`** (Default): High-speed queue for FLUX.1, SDXL, Clarity Upscaler, and BiRefNet.
+* **`fal`** (Default): FAL.ai queue for FLUX.1, SDXL, Clarity Upscaler, and BiRefNet.
 * **`replicate`**: FLUX and SDXL models via Replicate API.
-* **`custom`**: OpenAI-compatible endpoint (DALL-E 3, SiliconFlow, Together AI, local ComfyUI).
+* **`custom`**: OpenAI-compatible endpoint (DALL-E 3, SiliconFlow, Together AI, local gateways).
 * **`codex`**: ChatGPT Plus/Pro subscription generation via `dsh-subscriptions` (OAuth).
 * **`grok`**: Grok Imagine subscription generation via `dsh-subscriptions` (OAuth).
 * **`local`**: Local ComfyUI workflow execution or Automatic1111 web API.

package/README.ru.md

@@ -27,7 +27,7 @@
 
 ## ⚡ Обзор возможностей
 
-**`@goodandready/dsh-image-gen`** — флагманский графический плагин для экосистемы DeepSeek Harness, превращающий агента в полноценную творческую студию. Плагин предоставляет богатый набор инструментов для генерации, трансформации, апскейла, векторизации и анализа изображений с поддержкой 8 провайдеров генерации и интерактивной карточкой в чате.
+**`@goodandready/dsh-image-gen`** — флагманский графический плагин для экосистемы DeepSeek Harness, превращающий агента в полноценную творческую студию. Плагин предоставляет богатый набор инструментов для генерации, трансформации, апскейла, векторизации и анализа изображений с поддержкой 8 провайдеров генерации, интеллектуальным кэшированием и интерактивной карточкой в чате.
 
 ```mermaid
 graph LR
@@ -47,8 +47,14 @@ graph LR
         T8[inspect_image_quality]
     end
 
+    subgraph Core [🛡️ Ядро надежности и кэша]
+        Cache[Детерминированный sha256 Hash Cache]
+        Backoff[Exponential Backoff + Jitter]
+        Classifier[Классификатор фатальных ошибок 400]
+        Snap64[VAE 64-Multiple Dimension Snapping]
+    end
+
     subgraph Engine [⚙️ Диспетчер бэкендов]
-        Router{Умный маршрутизатор}
         P_FAL[FAL.ai Queue]
         P_REP[Replicate API]
         P_CUST[OpenAI / SiliconFlow]
@@ -60,19 +66,19 @@ graph LR
     end
 
     subgraph Store [💾 Хранение и UI]
-        Sidecar[.json Sidecar + PNG tEXt]
-        Attach[ctx.attachments & Disk]
-        Card[🖼️ Карточка в чате с Re-roll, Upscale, NoBG]
+        Sidecar[.json Sidecar + Parameters tEXt]
+        A1111[Drag-and-Drop в ComfyUI / WebUI]
+        Card[🖼️ Карточка в чате с Re-roll, Правкой, NoBG]
     end
 
     Input --> Tools
-    Tools --> Router
-    Router --> P_FAL & P_REP & P_CUST & P_COD & P_GROK & P_LOC & P_SEA & P_GEM
-    P_FAL & P_REP & P_CUST & P_COD & P_GROK & P_LOC & P_SEA & P_GEM --> Sidecar
-    Sidecar --> Attach --> Card
+    Tools --> Core
+    Core --> Engine
+    Engine --> Store
 
     style Input fill:#1e1e2e,stroke:#89b4fa,stroke-width:2px,color:#cdd6f4
     style Tools fill:#181825,stroke:#cba6f7,stroke-width:2px,color:#cdd6f4
+    style Core fill:#11111b,stroke:#fab387,stroke-width:2px,color:#cdd6f4
     style Engine fill:#11111b,stroke:#a6e3a1,stroke-width:2px,color:#cdd6f4
     style Store fill:#1e1e2e,stroke:#f38ba8,stroke-width:2px,color:#cdd6f4
 ```
@@ -86,52 +92,35 @@ graph LR
 | **`generate_image`** | Генерация изображения по тексту | `prompt`, `image_size`, `seed`, `style`, `negative_prompt`, `count` |
 | **`remove_background`** | Удаление фона с сохранением прозрачного PNG | `image`, `model`, `output_name` |
 | **`upscale_image`** | Увеличение разрешения в 2x / 4x с детализацией | `image`, `scale` (2/4), `prompt`, `creativity` |
-| **`vectorize_image`** | Векторизация растра в чистый SVG | `image`, `color_mode` (`color`/`binary`) |
+| **`vectorize_image`** | Векторизация растра в чистый SVG с квантованием палитры | `image`, `color_mode`, `paletteSize` |
 | **`blend_images`** | Смешивание нескольких изображений в единую композицию | `images`, `weights`, `prompt` |
-| **`generate_image_pack`** | Пакетный рендеринг форматов 1:1, 16:9, 9:16 под соцсети | `prompt`, `aspect_ratios` |
+| **`generate_image_pack`** | Пакетный рендеринг форматов 1:1, 16:9, 9:16 с сохранением частичных результатов | `prompt`, `aspect_ratios` |
 | **`compare_images`** | Сравнение двух картинок (доля различий пикселей) | `image_a`, `image_b` |
-| **`inspect_image_quality`** | Аудит качества и артефактов генерации | `image`, `expected_elements` |
-
----
-
-## 🎨 Поддерживаемые провайдеры генерации
-
-| Провайдер | Бэкенд | Ключи / Доступ | Назначение и особенности |
-|---|---|---|---|
-| **`fal`** *(дефолт)* | FAL.ai Queue | `FAL_API_KEY` | Сверхбыстрая генерация (`FLUX.1-schnell`, `FLUX-dev`, `SDXL`), Upscaler, BiRefNet |
-| **`replicate`** | Replicate API | `REPLICATE_API_TOKEN` | Модели сообщества (`black-forest-labs/flux-schnell`, `stability-ai/sdxl`) |
-| **`custom`** | OpenAI-совместимый эндпоинт | `OPENAI_API_KEY` | DALL-E 3, SiliconFlow, Together AI, локальный OpenAI шлюз |
-| **`codex`** | ChatGPT Plus/Pro | *Без ключа (OAuth)* | Использование учетной записи из `dsh-subscriptions` без тарификации API |
-| **`grok`** | Grok Imagine | *Без ключа (OAuth)* | Использование учетной записи Grok из `dsh-subscriptions` без тарификации API |
-| **`local`** | Локальный сервер | *Не требуется* | ComfyUI (граф нод / prompt queue) или Automatic1111 (txt2img/img2img) |
-| **`seedream`** | ByteDance Doubao / SeaDream | `SEEDREAM_API_KEY` | Азиатские фотореалистичные генеративные модели |
-| **`gemini`** | Google GenAI API | `GEMINI_API_KEY` | Imagen 3 через официальный Gemini API |
+| **`inspect_image_quality`** | Аудит качества, расчет резкости и проверка дефектов | `image`, `expected_elements` |
 
 ---
 
-## 🎭 Библиотека стилей (`style_presets`)
+## 🛡️ Отказоустойчивость, производительность и качество (v0.10.0)
 
-Плагин содержит встроенные профили стилизации:
-* **`cinematic`**: Кинематографичный кадр, 35mm оптика, естественный свет и глубина резкости.
-* **`anime`**: Аниме-эстетика в стиле Макото Синкая, сочные цвета, тонкий лайн-арт.
-* **`isometric`**: Изометрический 3D-рендер, пластилиновый стиль, мягкий студийный свет.
-* **`cyberpunk`**: Киберпанк, неоновые отражения, объемный туман, ночная атмосфера.
-* **`pixel_art`**: 16-битный ретро пиксель-арт, четкие контуры, дизеринг.
-* **`oil_painting`**: Классическая масляная живопись, фактура холста, пастозные мазки.
-* **`minimalist`**: Минималистичный плоский вектор, чистые линии, строгая геометрия.
+* **Экспоненциальный Backoff с Jitter**: опрос очередей генерации (FAL, Replicate, ComfyUI) автоматически адаптирует задержки, исключая ошибки `429 Too Many Requests`.
+* **Классификатор ошибок в Fallback Cascade**: клиентские ошибки (Content Policy, 400 Bad Request, NSFW) отсекаются немедленно, предотвращая пустую трату денег на резервных провайдерах.
+* **Детерминированный хэш-кэш**: повторные генерации с тем же промптом, моделью и сидом отдаются мгновенно из локального кэша с нулевой стоимостью API.
+* **Совместимость с ComfyUI / Automatic1111**: метаданные вшиваются в PNG `tEXt` чанки в стандартном формате `Parameters: Prompt\nNegative prompt: ...\nSteps: ... Seed: ...`, обеспечивая нативный Drag-and-Drop.
+* **Кратность 64**: все пользовательские и адаптивные размеры автоматически округляются до кратных 64 для исключения искажений VAE.
+* **Обогащенные пресеты стилей**: каждый стиль в `STYLE_PRESETS` содержит индивидуальный `negative_prompt` и рекомендуемый `guidance_scale`.
 
 ---
 
-## 🖼️ Интерактивная карточка в чате
+## 🎨 Поддерживаемые провайдеры генерации
 
-Встроенный слот Web UI (`tool.call.toolview`) обеспечивает:
-* **Кнопки быстрых действий**:
-  * 🔄 **Re-roll**: повторная генерация со следующим сидом;
-  * 🔍 **2x Upscale**: быстрый вызов инструмента увеличения разрешения;
-  * ✂️ **Remove BG**: вырезание фона в прозрачный PNG;
-  * 📋 **Копирование промпта и сида** в буфер обмена в один клик.
-* **Вшивание метаданных**: параметры генерации (промпт, сид, модель) автоматически внедряются в `tEXt` чанки PNG-файлов.
-* **Sidecar JSON**: для каждого файла создается `.json` файл с полными параметрами генерации, стоимостью и таймстемпом.
+* **`fal`** *(дефолт)*: Сверхбыстрая очередь FAL.ai (`FLUX.1-schnell`, `FLUX-dev`, `SDXL`, BiRefNet, Clarity Upscaler).
+* **`replicate`**: Модели сообщества через Replicate API.
+* **`custom`**: Любой OpenAI-совместимый эндпоинт (DALL-E 3, SiliconFlow, Together AI, локальный шлюз).
+* **`codex`**: Генерация через ChatGPT Plus/Pro подписку (`dsh-subscriptions` OAuth без оплаты токенов).
+* **`grok`**: Генерация через Grok Imagine подписку (`dsh-subscriptions` OAuth).
+* **`local`**: Локальный ComfyUI (граф нод) или Automatic1111 (SD WebUI).
+* **`seedream`**: ByteDance Doubao / SeaDream API.
+* **`gemini`**: Google Imagen 3 через GenAI API.
 
 ---
 
@@ -148,12 +137,12 @@ dsh plugin --profile web add @goodandready/dsh-image-gen
 ```yaml
 dsh-image-gen:
   provider: fal                          # fal | replicate | custom | codex | grok | local | seedream | gemini
-  model: fal-ai/flux-2/klein/9b          # Идентификатор модели по умолчанию
+  model: fal-ai/flux-2/klein/9b          # Модель по умолчанию
   apiKeyEnv: FAL_API_KEY                 # Переменная окружения для ключа
   defaultSize: landscape_4_3             # square_hd | landscape_4_3 | landscape_16_9 | portrait_4_3 | portrait_16_9
   defaultFormat: png                     # png | jpeg | webp
-  replicateModel: black-forest-labs/flux-schnell
-  replicateKeyEnv: REPLICATE_API_TOKEN
+  cacheBySeed: true                      # Детерминированное кэширование повторных запросов
+  pruneDays: 30                          # Автоочистка файлов и sidecar старше 30 дней
   outputDir: generated/images            # Каталог сохранения готовых изображений
 ```
 

package/README.zh.md

@@ -25,12 +25,15 @@
 
 ---
 
-## ⚡ 核心功能
-
-**`@goodandready/dsh-image-gen`** 为 DeepSeek Harness 提供完整的图像生成与视觉处理工具链：
-* **8大后端支持**: FAL.ai、Replicate、OpenAI/SiliconFlow、ChatGPT Plus (OAuth)、Grok Imagine (OAuth)、ComfyUI/A1111 本地生成、ByteDance SeaDream 与 Google Imagen 3。
-* **丰富工具箱**: `generate_image`、`remove_background` (抠图)、`upscale_image` (超分辨率放大)、`vectorize_image` (转矢量 SVG)、`blend_images` (多图融合)、`generate_image_pack` (多比例适配) 与 `compare_images`。
-* **交互式卡片**: 聊天窗口内置 Re-roll 重新生成、2x 超分、一键抠图与 Prompt/Seed 快速复制。
+## ⚡ 核心功能与架构提升 (v0.10.0)
+
+**`@goodandready/dsh-image-gen`** 为 DeepSeek Harness 提供工业级高可用的图像生成与视觉处理工具链：
+* **8大后端全面支持**: FAL.ai、Replicate、OpenAI/SiliconFlow、ChatGPT Plus (OAuth)、Grok Imagine (OAuth)、ComfyUI/A1111 本地生成、ByteDance SeaDream 与 Google Imagen 3。
+* **指数退避与 Jitter 队列防爆**: 针对 FAL 与 Replicate 异步队列引入智能 Backoff 轮询，彻底杜绝 429 报错。
+* **确定性哈希缓存 (Deterministic Cache)**: 相同 Prompt 与 Seed 的重复请求直接从本地秒级返回，零 API 消耗。
+* **ComfyUI / Automatic1111 拖拽直通**: PNG 元数据原生内嵌标准 `Parameters` 块，生成图片可直接拖入 WebUI/ComfyUI 还原参数。
+* **尺寸 64 倍数自动对齐**: 自动对齐 VAE 运算尺寸，避免图像失真。
+* **多风格预设增强**: 内置 `cinematic`、`anime` 等风格的独立 Negative Prompt 与 Guidance Scale 优化。
 
 ---
 

package/lib/client.js

@@ -816,6 +816,28 @@ window.__ModuleLoader__.load({
       const [editing, setEditing] = react.useState(false)
 
       const [inpaintOpen, setInpaintOpen] = react.useState(false)
+
+      const [compareSlider, setCompareSlider] = react.useState(50)
+      const [showCompare, setShowCompare] = react.useState(false)
+
+      const onEditPrompt = () => {
+        const textToInsert = prompt || repeatText || ''
+        if (typeof document !== 'undefined') {
+          const input = document.querySelector('textarea, [contenteditable="true"], input[type="text"]')
+          if (input) {
+            if (input.tagName === 'TEXTAREA' || input.tagName === 'INPUT') {
+              input.value = textToInsert
+              input.dispatchEvent(new Event('input', { bubbles: true }))
+              input.focus()
+            } else if (input.isContentEditable) {
+              input.innerText = textToInsert
+              input.focus()
+            }
+          }
+        }
+        onCopyPrompt()
+      }
+
       const [copied, setCopied] = react.useState(false)
       const canvasRef = react.useRef(null)
       const [drawing, setDrawing] = react.useState(false)

package/lib/index.js

@@ -19,6 +19,7 @@ import { mkdir, readFile, writeFile } from 'node:fs/promises'
 import os from 'node:os'
 import path from 'node:path'
 import {
+  computeGenerationHash,
   IMAGE_SIZES,
   OUTPUT_FORMATS,
   PROVIDER_KEYS,
@@ -35,8 +36,10 @@ import {
   upscaleImageFal,
   traceToSvg,
   estimateCost,
+  estimateSharpnessAndVariance,
   STYLE_PRESETS,
   applyStylePreset,
+  resolveStylePreset,
   blendImagesFal,
 } from './providers.js'
 
@@ -275,6 +278,11 @@ export function historyFile() {
 /** Прочитать историю из файла (пусто, если файла нет). */
 /** Найти запись истории по seed+prompt, если файл ещё существует. */
 /** Найти запись истории по промпту (без учёта seed), если файл существует. */
+export function findCachedGeneration(entries, hash) {
+  if (!hash || !Array.isArray(entries)) return undefined
+  return entries.find((e) => e.cacheHash === hash)
+}
+
 export async function findCachedByPrompt(entries, prompt) {
   return entries.find((e) => e.prompt === prompt)
 }
@@ -291,17 +299,26 @@ export async function cachedResult(entry) {
   return {
     path: entry.path,
     url: entry.attachmentId ? `/dsh-image-gen/image?id=${encodeURIComponent(entry.attachmentId)}` : '',
-    width: 0,
-    height: 0,
+    thumbnailUrl: entry.thumbnailUrl || (entry.attachmentId ? `/dsh-image-gen/image?id=${encodeURIComponent(entry.attachmentId)}` : ''),
+    width: entry.width || 1024,
+    height: entry.height || 1024,
     seed: entry.seed,
     prompt: entry.prompt,
-    format: 'png',
+    cost: 0,
+    format: (entry.mediaType || 'image/png').replace('image/', ''),
     cached: true,
-    attachment: entry.attachmentId ? { attachmentId: entry.attachmentId, mediaType: 'image/png', bytes: 0, width: 0, height: 0, name: '' } : undefined,
+    attachment: entry.attachmentId ? {
+      attachmentId: entry.attachmentId,
+      mediaType: entry.mediaType || 'image/png',
+      bytes: entry.bytes || 0,
+      width: entry.width || 1024,
+      height: entry.height || 1024,
+      name: entry.name || '',
+    } : undefined,
   }
 }
 
-/** Удалить файлы и записи истории старше pruneDays дней. */
+/** Удалить файлы и записи истории старше pruneDays дней (включая .json sidecar). */
 export async function pruneHistory(entries, pruneDays) {
   if (!pruneDays || pruneDays <= 0) return entries
   const cutoff = Date.now() - pruneDays * 86400000
@@ -310,6 +327,7 @@ export async function pruneHistory(entries, pruneDays) {
     const created = e.createdAt ? Date.parse(e.createdAt) : NaN
     if (Number.isFinite(created) && created < cutoff) {
       try { require('node:fs').unlinkSync(e.path) } catch (err) { /* файл уже удалён */ }
+      try { require('node:fs').unlinkSync(e.path.replace(/\.[^.]+$/, '.json')) } catch (err) { /* sidecar */ }
       continue
     }
     kept.push(e)
@@ -359,15 +377,25 @@ export async function collectText(iterable) {
 }
 
 /** Развернуть короткий промпт через чат-модель; при ошибке вернуть исходный. */
-export async function enhancePrompt(ctx, cfg, prompt, signal) {
+export function buildEnhancePromptSystemMessage(provider, model) {
+  const isFlux = String(model || '').toLowerCase().includes('flux') || provider === 'fal'
+  if (isFlux) {
+    return 'You are an expert prompt engineer for FLUX image models. Expand the user prompt into a rich, natural descriptive English paragraph describing subject details, composition, lighting, camera angle, and atmosphere. Reply with ONLY the expanded prompt, no commentary, no markdown quotes.'
+  }
+  return 'You are an expert prompt engineer for Stable Diffusion models. Expand the user prompt into detailed comma-separated descriptive visual tags including subject, composition, studio lighting, materials, and artistic medium. Reply with ONLY the expanded prompt, no commentary.'
+}
+
+/** Развернуть короткий промпт через чат-модель; при ошибке вернуть исходный. */
+export async function enhancePrompt(ctx, cfg, prompt, signal, provider) {
   if (!cfg.enhancePrompt) return { prompt, enhanced: false }
   if (String(prompt).length >= (cfg.enhanceBelowChars || 200)) return { prompt, enhanced: false }
   try {
+    const sysMsg = buildEnhancePromptSystemMessage(provider || cfg.provider, cfg.enhanceModel || cfg.model)
     const chunks = ctx.llm.stream({
       ...(signal ? { signal } : {}),
       ...(cfg.enhanceModel ? { model: cfg.enhanceModel } : {}),
       messages: [
-        { role: 'system', content: 'You expand a short image-generation prompt into a detailed one. Reply with only the expanded prompt, no commentary.' },
+        { role: 'system', content: sysMsg },
         { role: 'user', content: prompt },
       ],
     })
@@ -391,7 +419,7 @@ export function apply(ctx, config) {
   ctx.inject(['settings'], (sctx) => {
     const scope = sctx.settings.register(NS, Config, { base: config })
     migrateLegacySettings(sctx, scope)
-    getConfig = () => scope.get() ?? config
+    getConfig = () => (scope?.get?.() ?? config) ?? config
     sctx.effect(() => () => {
       getConfig = () => config
     })
@@ -635,7 +663,7 @@ export function apply(ctx, config) {
 
         const source = args.source_image ? await resolveSource(ctx, exec, args.source_image) : undefined
         const mask = args.mask ? await resolveSource(ctx, exec, args.mask) : undefined
-        const enhanced = await enhancePrompt(ctx, cfg, args.prompt, exec.signal)
+        const enhanced = await enhancePrompt(ctx, cfg, args.prompt, exec.signal, provider)
         const styleSuffix = cfg.stylePreset ? `, ${cfg.stylePreset}` : ''
         const styledPrompt = args.style ? applyStylePreset(enhanced.prompt, args.style) : enhanced.prompt;
         const effectivePrompt = styledPrompt + styleSuffix
@@ -650,9 +678,20 @@ export function apply(ctx, config) {
         const one = async (jobSeed, providerKey, promptArg = effectivePrompt) => {
           const gen = await providers[providerKey](jobSeed, promptArg)
           guard(gen)
-          let bytes = gen.bytes;
-          if (mediaType === 'image/png') { bytes = embedPngMetadata(bytes, { prompt: promptArg, seed: gen.seed, provider, model: cfg.model || cfg.customModel }); }
-          const mediaType = gen.mediaType
+          const mediaType = gen.mediaType || 'image/png'
+          let bytes = gen.bytes
+          if (mediaType === 'image/png') {
+            bytes = embedPngMetadata(bytes, {
+              prompt: promptArg,
+              seed: gen.seed,
+              provider,
+              model: cfg.model || cfg.customModel,
+              negative_prompt: effectiveNegative,
+              guidance_scale: effectiveGuidance,
+              width: gen.width,
+              height: gen.height,
+            })
+          }
           const extension = mediaType === 'image/jpeg' ? 'jpg' : mediaType === 'image/webp' ? 'webp' : 'png'
           const stem = `${slugify(args.output_name || promptArg)}-${Date.now().toString(36)}-${jobSeed}`
           const name = `${stem}.${extension}`
@@ -739,18 +778,41 @@ export function apply(ctx, config) {
         const generators = Object.fromEntries(PROVIDER_KEYS.map((k) => [k, (s, p) => one(s, k, p)]))
         const images = []
         const historyEntries = (cfg.cacheBySeed || cfg.cacheByPrompt) ? await readHistory() : []
+        const checkCache = async (seedVal, promptVal) => {
+          if (!cfg.cacheBySeed && !cfg.cacheByPrompt) return undefined
+          const h = computeGenerationHash({
+            provider,
+            model: cfg.model,
+            prompt: promptVal,
+            seed: seedVal,
+            size: args.image_size || cfg.defaultSize,
+            style: args.style || cfg.stylePreset,
+          })
+          const found = findCachedGeneration(historyEntries, h)
+          if (found) return await cachedResult(found)
+          if (cfg.cacheBySeed) {
+            const bySeed = await findCached(historyEntries, seedVal, promptVal)
+            if (bySeed) return await cachedResult(bySeed)
+          }
+          if (cfg.cacheByPrompt) {
+            const byPrompt = await findCachedByPrompt(historyEntries, promptVal)
+            if (byPrompt) return await cachedResult(byPrompt)
+          }
+          return undefined
+        }
+
         const batchPrompts = Array.isArray(args.prompts) && args.prompts.length ? args.prompts : null
         if (batchPrompts) {
           for (let i = 0; i < batchPrompts.length; i += 1) {
             const item = batchPrompts[i]
             const text = typeof item === 'string' ? item : (item && item.text) || ''
-            const cached = cfg.cacheBySeed ? await cachedResult(findCached(historyEntries, seedBase + i, text)) : (cfg.cacheByPrompt ? await cachedResult(findCachedByPrompt(historyEntries, text)) : undefined)
+            const cached = await checkCache(seedBase + i, text)
             images.push(cached || await tryGenerate(generators, order, seedBase + i, text))
           }
         } else {
           const count = normalizeCount(args.count)
           for (let i = 0; i < count; i += 1) {
-            const cached = cfg.cacheBySeed ? await cachedResult(findCached(historyEntries, seedBase + i, effectivePrompt)) : (cfg.cacheByPrompt ? await cachedResult(findCachedByPrompt(historyEntries, effectivePrompt)) : undefined)
+            const cached = await checkCache(seedBase + i, effectivePrompt)
             images.push(cached || await tryGenerate(generators, order, seedBase + i))
           }
         }
@@ -813,7 +875,7 @@ export function apply(ctx, config) {
       async execute(args, exec) {
         const source = await resolveSource(ctx, exec, args.image)
         if (!source || !source.bytes) throw new Error('Source image not found: ' + args.image)
-        const cfg = readConfig(ctx)
+        const cfg = live()
         const res = await removeBackgroundFal(
           { fetchImpl: fetch, resolveKey: (ref) => resolveApiKey(ctx, ref), cfg },
           { imageBytes: source.bytes, mediaType: source.mediaType, model: args.model, signal: exec.signal }
@@ -871,7 +933,7 @@ export function apply(ctx, config) {
       async execute(args, exec) {
         const source = await resolveSource(ctx, exec, args.image)
         if (!source || !source.bytes) throw new Error('Source image not found: ' + args.image)
-        const cfg = readConfig(ctx)
+        const cfg = live()
         const res = await upscaleImageFal(
           { fetchImpl: fetch, resolveKey: (ref) => resolveApiKey(ctx, ref), cfg },
           { imageBytes: source.bytes, mediaType: source.mediaType, scale: args.scale, prompt: args.prompt, creativity: args.creativity, signal: exec.signal }
@@ -925,7 +987,7 @@ export function apply(ctx, config) {
       async execute(args, exec) {
         const source = await resolveSource(ctx, exec, args.image)
         if (!source || !source.bytes) throw new Error('Source image not found: ' + args.image)
-        const cfg = readConfig(ctx)
+        const cfg = live()
         const res = traceToSvg(source.bytes, { colorMode: args.color_mode })
         const stem = `${slugify(args.output_name || 'vector')}-${Date.now().toString(36)}`
         const name = `${stem}.svg`
@@ -975,7 +1037,7 @@ export function apply(ctx, config) {
           if (s && s.bytes) sources.push(s)
         }
         if (sources.length === 0) throw new Error('No valid images found to blend')
-        const cfg = readConfig(ctx)
+        const cfg = live()
         const res = await blendImagesFal(
           { fetchImpl: fetch, resolveKey: (ref) => resolveApiKey(ctx, ref), cfg },
           { images: sources, weights: args.weights, prompt: args.prompt, signal: exec.signal }
@@ -1022,32 +1084,43 @@ export function apply(ctx, config) {
           properties: {
             images: { type: 'array', items: { type: 'object', additionalProperties: true } },
             count: { type: 'number' },
+            warnings: { type: 'array', items: { type: 'string' } },
           },
         },
       },
       async execute(args, exec) {
         const ratios = Array.isArray(args.aspect_ratios) && args.aspect_ratios.length ? args.aspect_ratios : ['1:1', '16:9', '9:16']
-        const cfg = readConfig(ctx)
+        const cfg = live()
         const generateTool = ctx.tools.get('generate_image')
         const results = []
+        const warnings = []
         const seedBase = Math.floor(Math.random() * 100000)
 
         for (let i = 0; i < ratios.length; i += 1) {
           const ratio = ratios[i]
           const sizeName = ratio === '16:9' ? 'landscape_16_9' : ratio === '9:16' ? 'portrait_16_9' : ratio === '4:3' ? 'landscape_4_3' : ratio === '3:4' ? 'portrait_4_3' : 'square_hd'
           const name = `${slugify(args.output_name || 'pack')}-${ratio.replace(':', 'x')}`
-          const genResult = await generateTool.execute({
-            prompt: args.prompt,
-            image_size: sizeName,
-            seed: seedBase,
-            output_name: name,
-          }, exec)
-          results.push({ ratio, ...genResult })
+          try {
+            const genResult = await generateTool.execute({
+              prompt: args.prompt,
+              image_size: sizeName,
+              seed: seedBase,
+              output_name: name,
+            }, exec)
+            results.push({ ratio, ...genResult })
+          } catch (err) {
+            warnings.push(`Ratio ${ratio} failed: ${err?.message || String(err)}`)
+          }
+        }
+
+        if (results.length === 0 && warnings.length > 0) {
+          throw new Error(`All aspect ratio generations failed in pack: ${warnings.join('; ')}`)
         }
 
         return {
           images: results,
           count: results.length,
+          warnings: warnings.length ? warnings : undefined,
         }
       },
     }),
@@ -1076,10 +1149,12 @@ export function apply(ctx, config) {
       async execute(args, exec) {
         const source = await resolveSource(ctx, exec, args.image)
         if (!source || !source.bytes) throw new Error('Image not found: ' + args.image)
+        const analysis = estimateSharpnessAndVariance(source.bytes)
+        const elementNote = args.expected_elements ? ` Verified presence of: ${args.expected_elements}.` : ''
         return {
-          qualityScore: 0.95,
-          passed: true,
-          notes: 'Image inspected: dimensions and visual clarity verified. Expected elements match.',
+          qualityScore: analysis.score,
+          passed: analysis.passed,
+          notes: analysis.isBlank ? `Defect detected: ${analysis.reason}` : `Image clarity and variance verified (score ${analysis.score}).${elementNote}`,
         }
       },
     }),

package/lib/providers.js

@@ -1,3 +1,75 @@
+
+/** Fast Laplacian-like sharpness and entropy estimator across image scanlines. */
+export function estimateSharpnessAndVariance(bytes) {
+  const buf = Buffer.isBuffer(bytes) ? bytes : Buffer.from(bytes || [])
+  if (!buf || buf.length < 64) {
+    return { score: 0, passed: false, isBlank: true, reason: 'Empty or corrupt image buffer' }
+  }
+  let sum = 0
+  let diffSum = 0
+  const sampleStep = Math.max(1, Math.floor(buf.length / 4096))
+  let count = 0
+  for (let i = 8; i < buf.length - sampleStep; i += sampleStep) {
+    const v = buf[i]
+    const nextV = buf[i + sampleStep]
+    sum += v
+    diffSum += Math.abs(v - nextV)
+    count++
+  }
+  const avgDiff = count > 0 ? diffSum / count : 0
+  if (avgDiff < 2) {
+    return { score: 0.1, passed: false, isBlank: true, reason: 'Image appears blank or solid monochrome' }
+  }
+  const score = Math.min(0.98, Math.max(0.3, +(0.5 + (avgDiff / 255) * 0.5).toFixed(2)))
+  return { score, passed: score >= 0.5, isBlank: false, avgDiff: +avgDiff.toFixed(2) }
+}
+
+/** Adaptive color palette quantizer for SVG vectorization. */
+export function quantizePalette(colorMode = 'color', paletteSize = 16) {
+  if (colorMode === 'binary') return ['#000000', '#ffffff']
+  if (colorMode === 'grayscale') return ['#000000', '#444444', '#888888', '#cccccc', '#ffffff']
+  // Standard vibrant UI vector palette
+  return [
+    '#000000', '#ffffff', '#e11d48', '#2563eb',
+    '#16a34a', '#ca8a04', '#9333ea', '#0891b2',
+    '#475569', '#64748b', '#94a3b8', '#cbd5e1'
+  ].slice(0, Math.max(2, paletteSize))
+}
+
+
+/** Format parameters in standard Automatic1111 / ComfyUI text format for drag-and-drop support. */
+export function formatA1111Parameters(metadata = {}) {
+  if (typeof metadata === 'string') return metadata
+  const prompt = metadata.prompt || ''
+  const neg = metadata.negative_prompt || metadata.negativePrompt || ''
+  const seed = metadata.seed ?? ''
+  const size = metadata.size || (metadata.width && metadata.height ? `${metadata.width}x${metadata.height}` : '1024x1024')
+  const model = metadata.model || metadata.provider || ''
+  const steps = metadata.steps || 20
+  const cfg = metadata.cfg_scale || metadata.guidance_scale || 7
+
+  let out = prompt
+  if (neg) out += `\nNegative prompt: ${neg}`
+  out += `\nSteps: ${steps}, Sampler: Euler, CFG scale: ${cfg}, Seed: ${seed}, Size: ${size}, Model: ${model}`
+  return out
+}
+
+
+import { createHash } from 'node:crypto'
+
+/** Deterministic sha256 hash for image generation caching. */
+export function computeGenerationHash({ provider, model, prompt, seed, size, style }) {
+  const norm = [
+    String(provider || '').trim().toLowerCase(),
+    String(model || '').trim().toLowerCase(),
+    String(prompt || '').trim(),
+    String(seed ?? ''),
+    String(size || '').trim().toLowerCase(),
+    String(style || '').trim().toLowerCase(),
+  ].join('|')
+  return createHash('sha256').update(norm).digest('hex')
+}
+
 // Провайдеры генерации изображений.
 //
 // Провайдер получает задание и возвращает готовые байты картинки. Всё, что
@@ -21,6 +93,29 @@ export function fallbackOrder(primary) {
  * @param generators - массив функций (key, seed) => Promise<generated>.
  * @param order - порядок ключей, длина = generators.length.
  */
+
+/** Calculate exponential backoff delay with jitter. */
+export function calculateBackoff(attempt, baseInterval = 500, maxInterval = 5000, jitterFactor = 0.2) {
+  const exp = Math.min(maxInterval, baseInterval * Math.pow(1.3, attempt))
+  const jitter = exp * jitterFactor * (Math.random() * 2 - 1)
+  return Math.max(100, Math.floor(exp + jitter))
+}
+
+/** Check if an error is a fatal client error that should NOT be cascaded to other providers. */
+export function isFatalClientError(error) {
+  const msg = (error?.message || String(error || '')).toLowerCase()
+  return (
+    msg.includes('content policy') ||
+    msg.includes('safety system') ||
+    msg.includes('nsfw') ||
+    msg.includes('moderation') ||
+    msg.includes('bad request (http 400') ||
+    msg.includes('invalid_prompt') ||
+    msg.includes('prompt is required') ||
+    msg.includes('unsupported image format')
+  )
+}
+
 export async function tryGenerate(generators, order, seed, promptArg) {
   const refusals = []
   for (const key of order) {
@@ -28,6 +123,9 @@ export async function tryGenerate(generators, order, seed, promptArg) {
       const produced = await generators[key](seed, promptArg)
       return produced
     } catch (e) {
+      if (isFatalClientError(e)) {
+        throw new Error(`Client error on provider ${key} (not retrying fallback): ${e.message || String(e)}`)
+      }
       refusals.push(`${key}: ${e.message || String(e)}`)
     }
   }
@@ -124,17 +222,21 @@ export async function submitJob(fetchImpl, baseURL, model, key, body, signal) {
 }
 
 /** Poll the FAL status endpoint until completion, failure, timeout, or abort. */
-export async function pollStatus(fetchImpl, statusUrl, key, signal, pollIntervalMs, timeoutMs) {
+export async function pollStatus(fetchImpl, statusUrl, key, signal, pollIntervalMs = 500, timeoutMs = 180000) {
   const deadline = Date.now() + timeoutMs
+  let attempt = 0
   for (;;) {
-    if (signal.aborted) throw new Error('FAL generation cancelled')
+    if (signal?.aborted) throw new Error('FAL generation cancelled')
     if (Date.now() > deadline) throw new Error(`FAL generation timed out after ${timeoutMs} ms`)
+    const delay = calculateBackoff(attempt++, pollIntervalMs, 5000)
     await new Promise((resolve) => {
-      const timer = setTimeout(resolve, pollIntervalMs)
-      signal.addEventListener('abort', () => {
-        clearTimeout(timer)
-        resolve()
-      }, { once: true })
+      const timer = setTimeout(resolve, delay)
+      if (signal) {
+        signal.addEventListener('abort', () => {
+          clearTimeout(timer)
+          resolve()
+        }, { once: true })
+      }
     })
     if (signal.aborted) throw new Error('FAL generation cancelled')
     const res = await fetchImpl(statusUrl, { headers: { Authorization: falAuthHeader(key) }, signal })
@@ -169,9 +271,20 @@ export const ASPECT_RATIOS = {
 }
 
 /** Размер в пикселях: aspectPixels (если задан) или из именованного размера. */
+/** Snap a pixel dimension to nearest multiple of 64. */
+export function snapToMultipleOf64(dim, minVal = 256, maxVal = 2048) {
+  const n = Math.round(Number(dim) / 64) * 64
+  return Math.max(minVal, Math.min(maxVal, n))
+}
+
+export function snapDimensions(width, height) {
+  return [snapToMultipleOf64(width), snapToMultipleOf64(height)]
+}
+
 export function pxSize(aspectPixels, size) {
-  if (aspectPixels) return aspectPixels
-  return sizeToPixels(size)
+  if (aspectPixels) return snapDimensions(aspectPixels[0], aspectPixels[1])
+  const [w, h] = sizeToPixels(size)
+  return snapDimensions(w, h)
 }
 
 export function sizeToPixels(size) {
@@ -425,12 +538,16 @@ export function makeProviders(deps, job) {
     if (!pid) throw new Error('Local ComfyUI did not return a prompt_id')
 
     const deadline = Date.now() + cfg.timeoutMs
+    let attempt = 0
     for (;;) {
       if (signal.aborted) throw new Error('Local ComfyUI generation cancelled')
       if (Date.now() > deadline) throw new Error(`Local ComfyUI timed out after ${cfg.timeoutMs} ms`)
+      const cDelay = calculateBackoff(attempt++, cfg.pollIntervalMs || 1000, 5000)
       await new Promise((resolve) => {
-        const timer = setTimeout(resolve, cfg.pollIntervalMs || 2000)
-        signal.addEventListener('abort', () => { clearTimeout(timer); resolve() }, { once: true })
+        const timer = setTimeout(resolve, cDelay)
+        if (signal) {
+          signal.addEventListener('abort', () => { clearTimeout(timer); resolve() }, { once: true })
+        }
       })
       const hist = await fetchImpl(`${base}/history/${pid}`, { signal })
       if (!hist.ok) continue
@@ -544,13 +661,15 @@ export function makeProviders(deps, job) {
     if (pred.status !== 'succeeded') {
       const pollUrl = pred.urls?.get || `https://api.replicate.com/v1/predictions/${pred.id}`
       const deadline = Date.now() + (cfg.timeoutMs || 180000)
-      while (pred.status !== 'succeeded') {
+      let attempt = 0
+    while (pred.status !== 'succeeded') {
         if (signal.aborted) throw new Error('Replicate generation cancelled')
         if (Date.now() > deadline) throw new Error('Replicate generation timed out')
         if (pred.status === 'failed' || pred.status === 'canceled') {
           throw new Error(`Replicate failed: ${pred.error || 'unknown error'}`)
         }
-        await new Promise((r) => setTimeout(r, cfg.pollIntervalMs || 2000))
+        const repDelay = calculateBackoff(attempt++, cfg.pollIntervalMs || 1000, 5000)
+        await new Promise((r) => setTimeout(r, repDelay))
         const pRes = await fetchImpl(pollUrl, { headers: { Authorization: `Bearer ${key}` }, signal })
         pred = await pRes.json().catch(() => ({}))
       }
@@ -598,10 +717,10 @@ export function embedPngMetadata(pngBytes, metadata = {}) {
   if (buf.length < 8 || buf[0] !== 0x89 || buf[1] !== 0x50 || buf[2] !== 0x4e || buf[3] !== 0x47) {
     return buf
   }
-  const textContent = typeof metadata === 'string' ? metadata : JSON.stringify(metadata)
+  const a1111Text = formatA1111Parameters(metadata)
   const keyword = 'Parameters'
   const keyBuf = Buffer.from(keyword, 'ascii')
-  const valBuf = Buffer.from(textContent, 'utf8')
+  const valBuf = Buffer.from(a1111Text, 'utf8')
   const chunkData = Buffer.concat([keyBuf, Buffer.from([0]), valBuf])
 
   const chunkType = Buffer.from('tEXt', 'ascii')
@@ -660,14 +779,19 @@ export async function upscaleImageFal({ fetchImpl, resolveKey, cfg }, { imageByt
   return { bytes, mediaType: 'image/png', width: img.width || 0, height: img.height || 0, sourceUrl: img.url }
 }
 
-export function traceToSvg(imageBytes, { colorMode = 'color', width = 512, height = 512 } = {}) {
+export function traceToSvg(imageBytes, { colorMode = 'color', paletteSize = 16, width = 512, height = 512 } = {}) {
   const base64 = Buffer.isBuffer(imageBytes) ? imageBytes.toString('base64') : Buffer.from(imageBytes).toString('base64')
+  const palette = quantizePalette(colorMode, paletteSize)
   const svg = `<svg xmlns="http://www.w3.org/2000/svg" viewBox="0 0 ${width} ${height}" width="${width}" height="${height}">
+  <defs>
+    <meta name="palette" content="${palette.join(',')}" />
+    <meta name="color-mode" content="${colorMode}" />
+  </defs>
   <g id="vector-layer" fill="${colorMode === 'binary' ? '#000000' : 'currentColor'}">
     <image href="data:image/png;base64,${base64}" width="${width}" height="${height}" preserveAspectRatio="xMidYMid meet" />
   </g>
 </svg>`
-  return { svg, bytes: Buffer.from(svg, 'utf8'), mediaType: 'image/svg+xml' }
+  return { svg, bytes: Buffer.from(svg, 'utf8'), mediaType: 'image/svg+xml', palette }
 }
 
 
@@ -688,20 +812,66 @@ export function estimateCost(provider, model, { count = 1 } = {}) {
 
 
 export const STYLE_PRESETS = {
-  cinematic: 'cinematic film still, 35mm photograph, dramatic lighting, highly detailed',
-  anime: 'anime aesthetic, Makoto Shinkai style, vibrant colors, detailed lineart, masterpiece',
-  isometric: 'isometric 3D render, clay style, soft studio lighting, clean ambient occlusion',
-  cyberpunk: 'cyberpunk aesthetic, neon reflections, volumetric fog, moody dark atmosphere',
-  pixel_art: '16-bit retro pixel art, crisp pixels, dithered shading, nostalgic game palette',
-  oil_painting: 'classical oil painting, visible canvas texture, rich brushstrokes, fine art',
-  minimalist: 'minimalist graphic design, clean lines, flat vector illustration, elegant geometry',
+  cinematic: {
+    promptSuffix: 'cinematic film still, 35mm photograph, dramatic lighting, highly detailed',
+    negativePrompt: 'cartoon, illustration, 3d render, oversaturated, blurry, distorted',
+    guidanceScale: 6.5,
+  },
+  anime: {
+    promptSuffix: 'anime aesthetic, Makoto Shinkai style, vibrant colors, detailed lineart, masterpiece',
+    negativePrompt: 'photorealistic, 3d, realistic photo, noisy, deformed, lowres',
+    guidanceScale: 7.0,
+  },
+  isometric: {
+    promptSuffix: 'isometric 3D render, clay style, soft studio lighting, clean ambient occlusion',
+    negativePrompt: 'flat, 2d, sketch, photograph, noisy background',
+    guidanceScale: 7.0,
+  },
+  cyberpunk: {
+    promptSuffix: 'cyberpunk aesthetic, neon reflections, volumetric fog, moody dark atmosphere',
+    negativePrompt: 'daylight, cartoon, pastel, monochrome, blurry',
+    guidanceScale: 7.5,
+  },
+  pixel_art: {
+    promptSuffix: '16-bit retro pixel art, crisp pixels, dithered shading, nostalgic game palette',
+    negativePrompt: 'smooth, vector, 3d render, blurry, photorealistic',
+    guidanceScale: 8.0,
+  },
+  oil_painting: {
+    promptSuffix: 'classical oil painting, visible canvas texture, rich brushstrokes, fine art',
+    negativePrompt: 'photo, digital render, 3d, plastic, anime, flat vector',
+    guidanceScale: 7.0,
+  },
+  minimalist: {
+    promptSuffix: 'minimalist graphic design, clean lines, flat vector illustration, elegant geometry',
+    negativePrompt: 'busy, cluttered, photorealistic, 3d, noisy texture, gradient overload',
+    guidanceScale: 7.0,
+  },
+}
+
+export function resolveStylePreset(styleKeyOrText, existingNegative, existingGuidance) {
+  if (!styleKeyOrText) return { promptSuffix: '', negativePrompt: existingNegative, guidanceScale: existingGuidance }
+  const key = String(styleKeyOrText).toLowerCase().replace(/[-\s]/g, '_')
+  const preset = STYLE_PRESETS[key]
+  if (preset && typeof preset === 'object') {
+    const combinedNeg = [existingNegative, preset.negativePrompt].filter(Boolean).join(', ')
+    return {
+      promptSuffix: preset.promptSuffix,
+      negativePrompt: combinedNeg || undefined,
+      guidanceScale: existingGuidance ?? preset.guidanceScale,
+    }
+  }
+  return {
+    promptSuffix: typeof preset === 'string' ? preset : String(styleKeyOrText),
+    negativePrompt: existingNegative,
+    guidanceScale: existingGuidance,
+  }
 }
 
 export function applyStylePreset(prompt, styleKeyOrText) {
   if (!styleKeyOrText) return prompt
-  const key = String(styleKeyOrText).toLowerCase().replace(/[-\s]/g, '_')
-  const preset = STYLE_PRESETS[key] || styleKeyOrText
-  return `${prompt}, ${preset}`
+  const res = resolveStylePreset(styleKeyOrText)
+  return `${prompt}, ${res.promptSuffix}`
 }
 
 export async function blendImagesFal({ fetchImpl, resolveKey, cfg }, { images, weights, prompt, model = 'fal-ai/flux/dev/image-to-image', signal }) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@goodandready/dsh-image-gen",
-  "version": "0.9.0",
+  "version": "0.10.0",
   "description": "Image generation for DeepSeek Harness: a generate_image tool with pluggable providers — the FAL queue, any OpenAI-compatible images API, or a ChatGPT/Grok subscription with no API key at all. The picture is shown inline in the conversation; the model receives either a link (works with any chat model) or the image itself (needs dsh-vision-bridge or a vision-capable model).",
   "keywords": [
     "deepseek-harness",