npm - @yigitahmetsahin/captcha-solver - Versions diffs - 2.0.1 → 3.0.0 - Mend

@yigitahmetsahin/captcha-solver 2.0.1 → 3.0.0

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (7) hide show

package/dist/index.cjs CHANGED Viewed

@@ -30,8 +30,10 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 // src/index.ts
 var index_exports = {};
 __export(index_exports, {
+  LEGACY_CONFUSION_GROUPS: () => LEGACY_CONFUSION_GROUPS,
   Solver: () => Solver,
   imageToBase64: () => imageToBase64,
+  majorityVote: () => majorityVote,
   preprocessCaptcha: () => preprocessCaptcha,
   preprocessCaptchaToBuffer: () => preprocessCaptchaToBuffer
 });
@@ -44,47 +46,98 @@ var import_ai = require("ai");
 var import_fs = __toESM(require("fs"), 1);
 var import_path = __toESM(require("path"), 1);
 var import_sharp = __toESM(require("sharp"), 1);
-async function preprocessCaptcha(input) {
-  const buf = await preprocessCaptchaToBuffer(input);
+var LEGACY_CROP = { left: 0.1, top: 0.02, right: 0.9, bottom: 0.6 };
+async function preprocessCaptcha(input, options) {
+  const buf = await preprocessCaptchaToBuffer(input, options);
   return buf.toString("base64");
 }
-async function preprocessCaptchaToBuffer(input) {
+async function preprocessCaptchaToBuffer(input, options) {
+  const {
+    blur = 1.5,
+    scale = 4,
+    contrast = 3,
+    sharpen = true,
+    crop = "auto",
+    padding = true,
+    negate = false,
+    greyscale = true
+  } = options ?? {};
   const source = typeof input === "string" ? import_path.default.resolve(input) : input;
   const metadata = await (0, import_sharp.default)(source).metadata();
   const origW = metadata.width;
   const origH = metadata.height;
-  const smoothed = await (0, import_sharp.default)(source).blur(1.5).greyscale().toBuffer();
-  const upscaled = await (0, import_sharp.default)(smoothed).resize(origW * 4, origH * 4, { kernel: "lanczos3" }).toBuffer();
-  const stats = await (0, import_sharp.default)(upscaled).stats();
-  const mean = stats.channels[0].mean;
-  const enhanced = await (0, import_sharp.default)(upscaled).linear(3, mean * (1 - 3)).sharpen({ sigma: 1, m1: 2, m2: 1 }).toBuffer();
-  const scaledW = origW * 4;
-  const scaledH = origH * 4;
-  const cropLeft = Math.floor(scaledW * 0.1);
-  const cropTop = Math.floor(scaledH * 0.02);
-  const cropRight = Math.floor(scaledW * 0.9);
-  const cropBottom = Math.floor(scaledH * 0.6);
-  const cropW = cropRight - cropLeft;
-  const cropH = cropBottom - cropTop;
-  return (0, import_sharp.default)(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
-    top: 20,
-    bottom: 20,
-    left: 30,
-    right: 30,
+  let pipeline = (0, import_sharp.default)(source);
+  if (blur > 0) pipeline = pipeline.blur(blur);
+  if (greyscale) pipeline = pipeline.greyscale();
+  const smoothed = await pipeline.toBuffer();
+  const upscaled = await (0, import_sharp.default)(smoothed).resize(origW * scale, origH * scale, { kernel: "lanczos3" }).toBuffer();
+  let enhanced;
+  if (contrast !== 1) {
+    const stats = await (0, import_sharp.default)(upscaled).stats();
+    const mean = stats.channels[0].mean;
+    let pipe = (0, import_sharp.default)(upscaled).linear(contrast, mean * (1 - contrast));
+    if (sharpen) pipe = pipe.sharpen({ sigma: 1, m1: 2, m2: 1 });
+    enhanced = await pipe.toBuffer();
+  } else {
+    enhanced = sharpen ? await (0, import_sharp.default)(upscaled).sharpen({ sigma: 1, m1: 2, m2: 1 }).toBuffer() : upscaled;
+  }
+  let cropped;
+  if (crop === "none") {
+    cropped = enhanced;
+  } else if (crop === "auto") {
+    cropped = await autoCrop(enhanced);
+  } else {
+    const fractions = crop === "legacy" ? LEGACY_CROP : crop;
+    const scaledW = origW * scale;
+    const scaledH = origH * scale;
+    const cropLeft = Math.floor(scaledW * fractions.left);
+    const cropTop = Math.floor(scaledH * fractions.top);
+    const cropRight = Math.floor(scaledW * fractions.right);
+    const cropBottom = Math.floor(scaledH * fractions.bottom);
+    const cropW = cropRight - cropLeft;
+    const cropH = cropBottom - cropTop;
+    cropped = await (0, import_sharp.default)(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).toBuffer();
+  }
+  const final = negate ? await (0, import_sharp.default)(cropped).negate().toBuffer() : cropped;
+  if (padding === false) {
+    return (0, import_sharp.default)(final).png().toBuffer();
+  }
+  const pad = typeof padding === "number" ? padding : void 0;
+  const vPad = pad ?? 20;
+  const hPad = pad ?? 30;
+  return (0, import_sharp.default)(final).extend({
+    top: vPad,
+    bottom: vPad,
+    left: hPad,
+    right: hPad,
     background: { r: 255, g: 255, b: 255 }
   }).png().toBuffer();
 }
+async function autoCrop(enhanced) {
+  try {
+    const trimmed = (0, import_sharp.default)(enhanced).trim({ threshold: 30 });
+    const trimmedBuf = await trimmed.toBuffer({ resolveWithObject: true });
+    const { width, height } = trimmedBuf.info;
+    if (width > 2 && height > 2) {
+      return trimmedBuf.data;
+    }
+  } catch {
+  }
+  return enhanced;
+}
 function imageToBase64(imagePath) {
   const buffer = import_fs.default.readFileSync(imagePath);
   return buffer.toString("base64");
 }
 // src/solver.ts
-var PROMPT = `You are an assistant helping a visually impaired person read distorted text from an image.
-The text contains uppercase letters A-Z and/or digits 0-9.
-A thin vertical stroke is the digit 1. Never read it as the letter I or L.
-A round closed shape is the letter O, not the letter D.
-Output ONLY the exact characters you read, nothing else.`;
+var PROMPT = `You are an expert OCR assistant reading distorted text from a CAPTCHA image.
+Two versions of the same captcha are provided. Cross-reference both to determine the correct text.
+The text may contain uppercase letters (A-Z), lowercase letters (a-z), and/or digits (0-9).
+Pay close attention to:
+- Letter case: lowercase "e" has a horizontal bar inside, digit "0" does not. Lowercase "r" has a short descender, uppercase "T" has a flat top.
+- Similar shapes: "5" has a flat top + curved bottom, "S" is fully curved. "4" has an angled stroke, "A" has a pointed top. "6" has a closed bottom loop, "8" has two loops. "2" has a curved top + flat bottom, "Z" has all straight lines.
+Output ONLY the exact characters you read, preserving case. Nothing else.`;
 var DEFAULT_MODELS = {
   openai: "gpt-4o",
   anthropic: "claude-sonnet-4-20250514",
@@ -110,7 +163,7 @@ async function resolveModel(apiKey, provider, modelId) {
       );
   }
 }
-var CONFUSION_GROUPS = {
+var LEGACY_CONFUSION_GROUPS = {
   "1": "1",
   I: "1",
   L: "1",
@@ -122,7 +175,7 @@ var CONFUSION_GROUPS = {
   Z: "Z",
   "2": "Z"
 };
-function majorityVote(attempts, expectedLength) {
+function majorityVote(attempts, expectedLength, groups) {
   let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;
   if (filtered.length === 0) {
     filtered = attempts;
@@ -142,6 +195,7 @@ function majorityVote(attempts, expectedLength) {
   }
   const sameLenAttempts = filtered.filter((a) => a.length === bestLen);
   if (sameLenAttempts.length === 0) return filtered[0];
+  const useGroups = groups && typeof groups === "object" ? groups : void 0;
   const result = [];
   for (let pos = 0; pos < bestLen; pos++) {
     const charCounts = /* @__PURE__ */ new Map();
@@ -149,20 +203,32 @@ function majorityVote(attempts, expectedLength) {
       const ch = a[pos];
       charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);
     }
-    const groupCounts = /* @__PURE__ */ new Map();
-    for (const [ch, count] of charCounts) {
-      const canonical = CONFUSION_GROUPS[ch] ?? ch;
-      groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);
-    }
-    let bestGroup = "";
-    let bestGroupCount = 0;
-    for (const [canonical, count] of groupCounts) {
-      if (count > bestGroupCount) {
-        bestGroup = canonical;
-        bestGroupCount = count;
+    if (useGroups) {
+      const groupCounts = /* @__PURE__ */ new Map();
+      for (const [ch, count] of charCounts) {
+        const canonical = useGroups[ch] ?? ch;
+        groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);
+      }
+      let bestGroup = "";
+      let bestGroupCount = 0;
+      for (const [canonical, count] of groupCounts) {
+        if (count > bestGroupCount) {
+          bestGroup = canonical;
+          bestGroupCount = count;
+        }
+      }
+      result.push(bestGroup);
+    } else {
+      let bestChar = "";
+      let bestCharCount = 0;
+      for (const [ch, count] of charCounts) {
+        if (count > bestCharCount) {
+          bestChar = ch;
+          bestCharCount = count;
+        }
       }
+      result.push(bestChar);
     }
-    result.push(bestGroup);
   }
   return result.join("");
 }
@@ -258,11 +324,32 @@ var Solver = class {
    * @returns Solved text, per-attempt answers, and token usage
    */
   async solve(input, options = {}) {
-    const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
+    const {
+      numAttempts = 7,
+      expectedLength,
+      maxRetries = 2,
+      verbose = true,
+      confusionGroups = false,
+      preprocess
+    } = options;
     const model = await this.getModel();
-    const imageBuffer = await preprocessCaptchaToBuffer(input);
+    const [enhancedBuffer, colorBuffer] = await Promise.all([
+      preprocessCaptchaToBuffer(input, preprocess),
+      preprocessCaptchaToBuffer(input, {
+        blur: 0,
+        scale: 4,
+        contrast: 1,
+        sharpen: false,
+        crop: "none",
+        padding: 40,
+        greyscale: false
+      })
+    ]);
     const results = await Promise.all(
-      Array.from({ length: numAttempts }, () => this.singleAttempt(model, imageBuffer, maxRetries))
+      Array.from(
+        { length: numAttempts },
+        () => this.singleAttempt(model, enhancedBuffer, colorBuffer, maxRetries)
+      )
     );
     const valid = results.filter((r) => r !== null);
     if (verbose) {
@@ -275,13 +362,18 @@ var Solver = class {
       if (verbose) console.log("  All attempts failed!");
       return { text: "", attempts, usage, attemptUsages };
     }
-    return { text: majorityVote(attempts, expectedLength), attempts, usage, attemptUsages };
+    return {
+      text: majorityVote(attempts, expectedLength, confusionGroups),
+      attempts,
+      usage,
+      attemptUsages
+    };
   }
   /**
    * Make a single API call to read the captcha.
    * Retries up to `maxRetries` times on failure.
    */
-  async singleAttempt(model, imageBuffer, maxRetries) {
+  async singleAttempt(model, primaryBuffer, secondaryBuffer, maxRetries) {
     for (let retry = 0; retry <= maxRetries; retry++) {
       try {
         const { text, usage } = await (0, import_ai.generateText)({
@@ -291,7 +383,8 @@ var Solver = class {
               role: "user",
               content: [
                 { type: "text", text: PROMPT },
-                { type: "image", image: imageBuffer }
+                { type: "image", image: primaryBuffer },
+                { type: "image", image: secondaryBuffer }
               ]
             }
           ],
@@ -303,7 +396,7 @@ var Solver = class {
         if (lower.includes("sorry") || lower.includes("can't help") || lower.includes("cannot help") || lower.includes("unable to") || lower.includes("i can't") || raw.length > 20) {
           return null;
         }
-        const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, "");
+        const cleaned = raw.replace(/[^A-Za-z0-9]/g, "");
         return cleaned ? { text: cleaned, usage } : null;
       } catch (_err) {
         if (retry < maxRetries) {
@@ -318,8 +411,10 @@ var Solver = class {
 };
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
+  LEGACY_CONFUSION_GROUPS,
   Solver,
   imageToBase64,
+  majorityVote,
   preprocessCaptcha,
   preprocessCaptchaToBuffer
 });

package/dist/index.cjs.map CHANGED Viewed

	@@ -1 +1 @@
1	- {"version":3,"sources":["../src/index.ts","../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["export { Solver } from './solver.js';\nexport type { SolverOptions, SolveOptions, SolveResult, Provider } from './solver.js';\nexport type { LanguageModelUsage } from 'ai';\nexport { preprocessCaptcha, preprocessCaptchaToBuffer, imageToBase64 } from './preprocess.js';\n","import type { LanguageModel, LanguageModelUsage } from 'ai';\nimport { generateText } from 'ai';\nimport { preprocessCaptchaToBuffer } from './preprocess.js';\n\nconst PROMPT = `You are an assistant helping a visually impaired person read distorted text from an image.\nThe text contains uppercase letters A-Z and/or digits 0-9.\nA thin vertical stroke is the digit 1. Never read it as the letter I or L.\nA round closed shape is the letter O, not the letter D.\nOutput ONLY the exact characters you read, nothing else.`;\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport type Provider = 'openai' \| 'anthropic' \| 'google';\n\nexport interface SolverOptions {\n /** AI provider to use when constructing the model from an API key (default: \"openai\") /\n provider?: Provider;\n /* Model ID passed to the provider (default: \"gpt-4o\") /\n model?: string;\n}\n\nexport interface SolveOptions {\n /* Number of voting attempts (default: 5) /\n numAttempts?: number;\n /* Expected captcha length — results of other lengths are discarded /\n expectedLength?: number;\n /* Max retries per attempt on API failure (default: 2) /\n maxRetries?: number;\n /* Whether to log attempt details (default: true) /\n verbose?: boolean;\n}\n\nexport interface SolveResult {\n /* The solved captcha text (majority-voted) /\n text: string;\n /* Per-attempt raw answers (before voting) /\n attempts: string[];\n /* Aggregated token usage across all parallel attempts /\n usage: LanguageModelUsage;\n /* Per-attempt usage breakdown /\n attemptUsages: LanguageModelUsage[];\n}\n\ninterface AttemptResult {\n text: string;\n usage: LanguageModelUsage;\n}\n\n// ── Provider resolution ──────────────────────────────────────────────\n\nconst DEFAULT_MODELS: Record<Provider, string> = {\n openai: 'gpt-4o',\n anthropic: 'claude-sonnet-4-20250514',\n google: 'gemini-2.0-flash',\n};\n\nasync function resolveModel(\n apiKey: string,\n provider: Provider,\n modelId: string\n): Promise<LanguageModel> {\n switch (provider) {\n case 'openai': {\n const { createOpenAI } = await import('@ai-sdk/openai');\n return createOpenAI({ apiKey })(modelId);\n }\n case 'anthropic': {\n // @ts-expect-error — optional peer dependency\n const { createAnthropic } = await import('@ai-sdk/anthropic');\n return createAnthropic({ apiKey })(modelId);\n }\n case 'google': {\n // @ts-expect-error — optional peer dependency\n const { createGoogleGenerativeAI } = await import('@ai-sdk/google');\n return createGoogleGenerativeAI({ apiKey })(modelId);\n }\n default:\n throw new Error(\n `Unknown provider \"${provider}\". Install the matching @ai-sdk/ package and pass the model directly.`\n );\n }\n}\n\n// ── Confusion groups ─────────────────────────────────────────────────\n\n/*\n Characters the model commonly misreads as each other.\n * Each group maps to its canonical (most likely correct) character.\n /\nconst CONFUSION_GROUPS: Record<string, string> = {\n '1': '1',\n I: '1',\n L: '1',\n O: 'O',\n D: 'O',\n '0': 'O',\n S: 'S',\n '5': 'S',\n Z: 'Z',\n '2': 'Z',\n};\n\n// ── Majority voting ──────────────────────────────────────────────────\n\n/\n Character-level majority vote across multiple attempts.\n * Uses confusion-aware voting: characters that the model commonly\n * confuses (e.g. 1/I/L, O/D/0) are grouped together during counting.\n /\nfunction majorityVote(attempts: string[], expectedLength?: number): string {\n let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n if (filtered.length === 0) {\n filtered = attempts;\n }\n if (filtered.length === 0) return '';\n\n // Find most common length\n const lenCounts = new Map<number, number>();\n for (const a of filtered) {\n lenCounts.set(a.length, (lenCounts.get(a.length) ?? 0) + 1);\n }\n let bestLen = 0;\n let bestCount = 0;\n for (const [len, count] of lenCounts) {\n if (count > bestCount) {\n bestLen = len;\n bestCount = count;\n }\n }\n\n const sameLenAttempts = filtered.filter((a) => a.length === bestLen);\n if (sameLenAttempts.length === 0) return filtered[0];\n\n // Vote per character position with confusion-aware grouping\n const result: string[] = [];\n for (let pos = 0; pos < bestLen; pos++) {\n const charCounts = new Map<string, number>();\n for (const a of sameLenAttempts) {\n const ch = a[pos];\n charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);\n }\n\n const groupCounts = new Map<string, number>();\n for (const [ch, count] of charCounts) {\n const canonical = CONFUSION_GROUPS[ch] ?? ch;\n groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);\n }\n\n let bestGroup = '';\n let bestGroupCount = 0;\n for (const [canonical, count] of groupCounts) {\n if (count > bestGroupCount) {\n bestGroup = canonical;\n bestGroupCount = count;\n }\n }\n\n result.push(bestGroup);\n }\n\n return result.join('');\n}\n\n// ── Usage aggregation ────────────────────────────────────────────────\n\nfunction sumOptional(a: number \| undefined, b: number \| undefined): number \| undefined {\n if (a === undefined && b === undefined) return undefined;\n return (a ?? 0) + (b ?? 0);\n}\n\nfunction aggregateUsage(usages: LanguageModelUsage[]): LanguageModelUsage {\n const zero: LanguageModelUsage = {\n inputTokens: undefined,\n inputTokenDetails: {\n noCacheTokens: undefined,\n cacheReadTokens: undefined,\n cacheWriteTokens: undefined,\n },\n outputTokens: undefined,\n outputTokenDetails: {\n textTokens: undefined,\n reasoningTokens: undefined,\n },\n totalTokens: undefined,\n };\n return usages.reduce<LanguageModelUsage>(\n (acc, u) => ({\n inputTokens: sumOptional(acc.inputTokens, u.inputTokens),\n inputTokenDetails: {\n noCacheTokens: sumOptional(\n acc.inputTokenDetails.noCacheTokens,\n u.inputTokenDetails.noCacheTokens\n ),\n cacheReadTokens: sumOptional(\n acc.inputTokenDetails.cacheReadTokens,\n u.inputTokenDetails.cacheReadTokens\n ),\n cacheWriteTokens: sumOptional(\n acc.inputTokenDetails.cacheWriteTokens,\n u.inputTokenDetails.cacheWriteTokens\n ),\n },\n outputTokens: sumOptional(acc.outputTokens, u.outputTokens),\n outputTokenDetails: {\n textTokens: sumOptional(acc.outputTokenDetails.textTokens, u.outputTokenDetails.textTokens),\n reasoningTokens: sumOptional(\n acc.outputTokenDetails.reasoningTokens,\n u.outputTokenDetails.reasoningTokens\n ),\n },\n totalTokens: sumOptional(acc.totalTokens, u.totalTokens),\n }),\n zero\n );\n}\n\n// ── Solver class ─────────────────────────────────────────────────────\n\nexport class Solver {\n private _model: LanguageModel \| null = null;\n private _pendingModel: Promise<LanguageModel> \| null = null;\n\n /\n Create a captcha solver.\n \n @example\n * // Simple — defaults to OpenAI gpt-4o\n * const solver = new Solver('sk-...');\n \n @example\n * // Specify provider and model\n * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });\n \n @example\n * // Pass an AI SDK model directly\n * import { createOpenAI } from '@ai-sdk/openai';\n * const openai = createOpenAI({ apiKey: 'sk-...' });\n * const solver = new Solver(openai('gpt-4o'));\n /\n constructor(keyOrModel: string \| LanguageModel, options?: SolverOptions) {\n if (typeof keyOrModel === 'string') {\n const provider = options?.provider ?? 'openai';\n const modelId = options?.model ?? DEFAULT_MODELS[provider];\n // Lazily resolve the model on first use\n this._pendingModel = resolveModel(keyOrModel, provider, modelId);\n } else {\n this._model = keyOrModel;\n }\n }\n\n private async getModel(): Promise<LanguageModel> {\n if (this._model) return this._model;\n this._model = await this._pendingModel!;\n this._pendingModel = null;\n return this._model;\n }\n\n /\n Solve a captcha image.\n \n @param input - File path (string) or raw image Buffer\n * @param options - Solve options (attempts, expected length, etc.)\n * @returns Solved text, per-attempt answers, and token usage\n /\n async solve(input: string \| Buffer, options: SolveOptions = {}): Promise<SolveResult> {\n const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;\n\n const model = await this.getModel();\n const imageBuffer = await preprocessCaptchaToBuffer(input);\n\n // Fire all attempts in parallel for speed\n const results = await Promise.all(\n Array.from({ length: numAttempts }, () => this.singleAttempt(model, imageBuffer, maxRetries))\n );\n const valid = results.filter((r): r is AttemptResult => r !== null);\n if (verbose) {\n valid.forEach((r, i) => console.log(` Attempt ${i + 1}: ${r.text}`));\n }\n\n const attempts = valid.map((r) => r.text);\n const attemptUsages = valid.map((r) => r.usage);\n const usage = aggregateUsage(attemptUsages);\n\n if (attempts.length === 0) {\n if (verbose) console.log(' All attempts failed!');\n return { text: '', attempts, usage, attemptUsages };\n }\n\n return { text: majorityVote(attempts, expectedLength), attempts, usage, attemptUsages };\n }\n\n /\n Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n /\n private async singleAttempt(\n model: LanguageModel,\n imageBuffer: Buffer,\n maxRetries: number\n ): Promise<AttemptResult \| null> {\n for (let retry = 0; retry <= maxRetries; retry++) {\n try {\n const { text, usage } = await generateText({\n model,\n messages: [\n {\n role: 'user',\n content: [\n { type: 'text', text: PROMPT },\n { type: 'image', image: imageBuffer },\n ],\n },\n ],\n temperature: 1,\n maxOutputTokens: 256,\n });\n\n const raw = text.trim();\n\n // Detect refusals\n const lower = raw.toLowerCase();\n if (\n lower.includes('sorry') \|\|\n lower.includes(\"can't help\") \|\|\n lower.includes('cannot help') \|\|\n lower.includes('unable to') \|\|\n lower.includes(\"i can't\") \|\|\n raw.length > 20\n ) {\n return null;\n }\n\n // Clean: keep only uppercase letters and digits\n const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, '');\n return cleaned ? { text: cleaned, usage } : null;\n } catch (_err) {\n if (retry < maxRetries) {\n await new Promise((r) => setTimeout(r, 1000 (retry + 1)));\n continue;\n }\n return null;\n }\n }\n return null;\n }\n}\n","import fs from 'fs';\nimport path from 'path';\nimport sharp from 'sharp';\n\n/*\n Preprocess a captcha image using sharp (libvips).\n \n Pipeline:\n * 1. Gaussian blur in color space (smooths dither pattern)\n * 2. Grayscale conversion\n * 3. Upscale 4× with Lanczos\n * 4. Contrast boost (3× around image mean) + sharpen\n * 5. Crop decorative borders\n * 6. Add white padding\n \n Accepts a file path or a raw image Buffer.\n * Returns a base64-encoded PNG string.\n /\nexport async function preprocessCaptcha(input: string \| Buffer): Promise<string> {\n const buf = await preprocessCaptchaToBuffer(input);\n return buf.toString('base64');\n}\n\n/\n Same preprocessing pipeline as `preprocessCaptcha`, but returns the\n * resulting PNG as a raw Buffer (useful for AI SDK image content parts).\n /\nexport async function preprocessCaptchaToBuffer(input: string \| Buffer): Promise<Buffer> {\n const source = typeof input === 'string' ? path.resolve(input) : input;\n\n // Read original dimensions for crop/resize calculations\n const metadata = await sharp(source).metadata();\n const origW = metadata.width!;\n const origH = metadata.height!;\n\n // Step 1-2: Blur in color space (smooths dither pattern) → greyscale\n // Separate from resize to prevent pipeline reordering\n const smoothed = await sharp(source).blur(1.5).greyscale().toBuffer();\n\n // Step 3: Upscale 4× with Lanczos\n const upscaled = await sharp(smoothed)\n .resize(origW 4, origH * 4, { kernel: 'lanczos3' })\n .toBuffer();\n\n // Step 4: Contrast 3× around actual image mean + sharpen\n // Matches PIL's ImageEnhance.Contrast: output = factorinput + mean(1-factor)\n const stats = await sharp(upscaled).stats();\n const mean = stats.channels[0].mean;\n const enhanced = await sharp(upscaled)\n .linear(3.0, mean * (1 - 3.0))\n .sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 })\n .toBuffer();\n\n // Step 5: Crop decorative borders\n // Remove 10% left/right, 2% top, 40% bottom (keep top 60%)\n // Math.floor matches Python's int() truncation\n const scaledW = origW * 4;\n const scaledH = origH * 4;\n const cropLeft = Math.floor(scaledW * 0.1);\n const cropTop = Math.floor(scaledH * 0.02);\n const cropRight = Math.floor(scaledW * 0.9);\n const cropBottom = Math.floor(scaledH * 0.6);\n const cropW = cropRight - cropLeft;\n const cropH = cropBottom - cropTop;\n\n // Step 5-6: Crop → add white padding → output PNG\n return sharp(enhanced)\n .extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH })\n .extend({\n top: 20,\n bottom: 20,\n left: 30,\n right: 30,\n background: { r: 255, g: 255, b: 255 },\n })\n .png()\n .toBuffer();\n}\n\n/*\n Read an image file and return its base64-encoded content.\n */\nexport function imageToBase64(imagePath: string): string {\n const buffer = fs.readFileSync(imagePath);\n return buffer.toString('base64');\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACCA,gBAA6B;;;ACD7B,gBAAe;AACf,kBAAiB;AACjB,mBAAkB;AAgBlB,eAAsB,kBAAkB,OAAyC;AAC/E,QAAM,MAAM,MAAM,0BAA0B,KAAK;AACjD,SAAO,IAAI,SAAS,QAAQ;AAC9B;AAMA,eAAsB,0BAA0B,OAAyC;AACvF,QAAM,SAAS,OAAO,UAAU,WAAW,YAAAA,QAAK,QAAQ,KAAK,IAAI;AAGjE,QAAM,WAAW,UAAM,aAAAC,SAAM,MAAM,EAAE,SAAS;AAC9C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAIvB,QAAM,WAAW,UAAM,aAAAA,SAAM,MAAM,EAAE,KAAK,GAAG,EAAE,UAAU,EAAE,SAAS;AAGpE,QAAM,WAAW,UAAM,aAAAA,SAAM,QAAQ,EAClC,OAAO,QAAQ,GAAG,QAAQ,GAAG,EAAE,QAAQ,WAAW,CAAC,EACnD,SAAS;AAIZ,QAAM,QAAQ,UAAM,aAAAA,SAAM,QAAQ,EAAE,MAAM;AAC1C,QAAM,OAAO,MAAM,SAAS,CAAC,EAAE;AAC/B,QAAM,WAAW,UAAM,aAAAA,SAAM,QAAQ,EAClC,OAAO,GAAK,QAAQ,IAAI,EAAI,EAC5B,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC,EACxC,SAAS;AAKZ,QAAM,UAAU,QAAQ;AACxB,QAAM,UAAU,QAAQ;AACxB,QAAM,WAAW,KAAK,MAAM,UAAU,GAAG;AACzC,QAAM,UAAU,KAAK,MAAM,UAAU,IAAI;AACzC,QAAM,YAAY,KAAK,MAAM,UAAU,GAAG;AAC1C,QAAM,aAAa,KAAK,MAAM,UAAU,GAAG;AAC3C,QAAM,QAAQ,YAAY;AAC1B,QAAM,QAAQ,aAAa;AAG3B,aAAO,aAAAA,SAAM,QAAQ,EAClB,QAAQ,EAAE,MAAM,UAAU,KAAK,SAAS,OAAO,OAAO,QAAQ,MAAM,CAAC,EACrE,OAAO;AAAA,IACN,KAAK;AAAA,IACL,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,OAAO;AAAA,IACP,YAAY,EAAE,GAAG,KAAK,GAAG,KAAK,GAAG,IAAI;AAAA,EACvC,CAAC,EACA,IAAI,EACJ,SAAS;AACd;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,UAAAC,QAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;ADjFA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AA8Cf,IAAM,iBAA2C;AAAA,EAC/C,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,QAAQ;AACV;AAEA,eAAe,aACb,QACA,UACA,SACwB;AACxB,UAAQ,UAAU;AAAA,IAChB,KAAK,UAAU;AACb,YAAM,EAAE,aAAa,IAAI,MAAM,OAAO,gBAAgB;AACtD,aAAO,aAAa,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACzC;AAAA,IACA,KAAK,aAAa;AAEhB,YAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,mBAAmB;AAC5D,aAAO,gBAAgB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IAC5C;AAAA,IACA,KAAK,UAAU;AAEb,YAAM,EAAE,yBAAyB,IAAI,MAAM,OAAO,gBAAgB;AAClE,aAAO,yBAAyB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACrD;AAAA,IACA;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,QAAQ;AAAA,MAC/B;AAAA,EACJ;AACF;AAQA,IAAM,mBAA2C;AAAA,EAC/C,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AACP;AASA,SAAS,aAAa,UAAoB,gBAAiC;AACzE,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAEtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAGlC,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,KAAK,UAAU;AACxB,cAAU,IAAI,EAAE,SAAS,UAAU,IAAI,EAAE,MAAM,KAAK,KAAK,CAAC;AAAA,EAC5D;AACA,MAAI,UAAU;AACd,MAAI,YAAY;AAChB,aAAW,CAAC,KAAK,KAAK,KAAK,WAAW;AACpC,QAAI,QAAQ,WAAW;AACrB,gBAAU;AACV,kBAAY;AAAA,IACd;AAAA,EACF;AAEA,QAAM,kBAAkB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnE,MAAI,gBAAgB,WAAW,EAAG,QAAO,SAAS,CAAC;AAGnD,QAAM,SAAmB,CAAC;AAC1B,WAAS,MAAM,GAAG,MAAM,SAAS,OAAO;AACtC,UAAM,aAAa,oBAAI,IAAoB;AAC3C,eAAW,KAAK,iBAAiB;AAC/B,YAAM,KAAK,EAAE,GAAG;AAChB,iBAAW,IAAI,KAAK,WAAW,IAAI,EAAE,KAAK,KAAK,CAAC;AAAA,IAClD;AAEA,UAAM,cAAc,oBAAI,IAAoB;AAC5C,eAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,YAAM,YAAY,iBAAiB,EAAE,KAAK;AAC1C,kBAAY,IAAI,YAAY,YAAY,IAAI,SAAS,KAAK,KAAK,KAAK;AAAA,IACtE;AAEA,QAAI,YAAY;AAChB,QAAI,iBAAiB;AACrB,eAAW,CAAC,WAAW,KAAK,KAAK,aAAa;AAC5C,UAAI,QAAQ,gBAAgB;AAC1B,oBAAY;AACZ,yBAAiB;AAAA,MACnB;AAAA,IACF;AAEA,WAAO,KAAK,SAAS;AAAA,EACvB;AAEA,SAAO,OAAO,KAAK,EAAE;AACvB;AAIA,SAAS,YAAY,GAAuB,GAA2C;AACrF,MAAI,MAAM,UAAa,MAAM,OAAW,QAAO;AAC/C,UAAQ,KAAK,MAAM,KAAK;AAC1B;AAEA,SAAS,eAAe,QAAkD;AACxE,QAAM,OAA2B;AAAA,IAC/B,aAAa;AAAA,IACb,mBAAmB;AAAA,MACjB,eAAe;AAAA,MACf,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,IACpB;AAAA,IACA,cAAc;AAAA,IACd,oBAAoB;AAAA,MAClB,YAAY;AAAA,MACZ,iBAAiB;AAAA,IACnB;AAAA,IACA,aAAa;AAAA,EACf;AACA,SAAO,OAAO;AAAA,IACZ,CAAC,KAAK,OAAO;AAAA,MACX,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,MACvD,mBAAmB;AAAA,QACjB,eAAe;AAAA,UACb,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,iBAAiB;AAAA,UACf,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,kBAAkB;AAAA,UAChB,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,MACF;AAAA,MACA,cAAc,YAAY,IAAI,cAAc,EAAE,YAAY;AAAA,MAC1D,oBAAoB;AAAA,QAClB,YAAY,YAAY,IAAI,mBAAmB,YAAY,EAAE,mBAAmB,UAAU;AAAA,QAC1F,iBAAiB;AAAA,UACf,IAAI,mBAAmB;AAAA,UACvB,EAAE,mBAAmB;AAAA,QACvB;AAAA,MACF;AAAA,MACA,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,IACzD;AAAA,IACA;AAAA,EACF;AACF;AAIO,IAAM,SAAN,MAAa;AAAA,EACV,SAA+B;AAAA,EAC/B,gBAA+C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAmBvD,YAAY,YAAoC,SAAyB;AACvE,QAAI,OAAO,eAAe,UAAU;AAClC,YAAM,WAAW,SAAS,YAAY;AACtC,YAAM,UAAU,SAAS,SAAS,eAAe,QAAQ;AAEzD,WAAK,gBAAgB,aAAa,YAAY,UAAU,OAAO;AAAA,IACjE,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAc,WAAmC;AAC/C,QAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,SAAK,SAAS,MAAM,KAAK;AACzB,SAAK,gBAAgB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,MAAM,OAAwB,UAAwB,CAAC,GAAyB;AACpF,UAAM,EAAE,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE5E,UAAM,QAAQ,MAAM,KAAK,SAAS;AAClC,UAAM,cAAc,MAAM,0BAA0B,KAAK;AAGzD,UAAM,UAAU,MAAM,QAAQ;AAAA,MAC5B,MAAM,KAAK,EAAE,QAAQ,YAAY,GAAG,MAAM,KAAK,cAAc,OAAO,aAAa,UAAU,CAAC;AAAA,IAC9F;AACA,UAAM,QAAQ,QAAQ,OAAO,CAAC,MAA0B,MAAM,IAAI;AAClE,QAAI,SAAS;AACX,YAAM,QAAQ,CAAC,GAAG,MAAM,QAAQ,IAAI,aAAa,IAAI,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC;AAAA,IACtE;AAEA,UAAM,WAAW,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI;AACxC,UAAM,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,eAAe,aAAa;AAE1C,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO,EAAE,MAAM,IAAI,UAAU,OAAO,cAAc;AAAA,IACpD;AAEA,WAAO,EAAE,MAAM,aAAa,UAAU,cAAc,GAAG,UAAU,OAAO,cAAc;AAAA,EACxF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,aACA,YAC+B;AAC/B,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,MAAM,MAAM,IAAI,UAAM,wBAAa;AAAA,UACzC;AAAA,UACA,UAAU;AAAA,YACR;AAAA,cACE,MAAM;AAAA,cACN,SAAS;AAAA,gBACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,gBAC7B,EAAE,MAAM,SAAS,OAAO,YAAY;AAAA,cACtC;AAAA,YACF;AAAA,UACF;AAAA,UACA,aAAa;AAAA,UACb,iBAAiB;AAAA,QACnB,CAAC;AAED,cAAM,MAAM,KAAK,KAAK;AAGtB,cAAM,QAAQ,IAAI,YAAY;AAC9B,YACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,iBAAO;AAAA,QACT;AAGA,cAAM,UAAU,IAAI,YAAY,EAAE,QAAQ,cAAc,EAAE;AAC1D,eAAO,UAAU,EAAE,MAAM,SAAS,MAAM,IAAI;AAAA,MAC9C,SAAS,MAAM;AACb,YAAI,QAAQ,YAAY;AACtB,gBAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;","names":["path","sharp","fs"]}
1	+ {"version":3,"sources":["../src/index.ts","../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["export { Solver } from './solver.js';\nexport { majorityVote, LEGACY_CONFUSION_GROUPS } from './solver.js';\nexport type { SolverOptions, SolveOptions, SolveResult, Provider } from './solver.js';\nexport type { LanguageModelUsage } from 'ai';\nexport { preprocessCaptcha, preprocessCaptchaToBuffer, imageToBase64 } from './preprocess.js';\nexport type { PreprocessOptions, CropFractions } from './preprocess.js';\n","import type { LanguageModel, LanguageModelUsage } from 'ai';\nimport { generateText } from 'ai';\nimport type { PreprocessOptions } from './preprocess.js';\nimport { preprocessCaptchaToBuffer } from './preprocess.js';\n\nconst PROMPT = `You are an expert OCR assistant reading distorted text from a CAPTCHA image.\nTwo versions of the same captcha are provided. Cross-reference both to determine the correct text.\nThe text may contain uppercase letters (A-Z), lowercase letters (a-z), and/or digits (0-9).\nPay close attention to:\n- Letter case: lowercase \"e\" has a horizontal bar inside, digit \"0\" does not. Lowercase \"r\" has a short descender, uppercase \"T\" has a flat top.\n- Similar shapes: \"5\" has a flat top + curved bottom, \"S\" is fully curved. \"4\" has an angled stroke, \"A\" has a pointed top. \"6\" has a closed bottom loop, \"8\" has two loops. \"2\" has a curved top + flat bottom, \"Z\" has all straight lines.\nOutput ONLY the exact characters you read, preserving case. Nothing else.`;\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport type Provider = 'openai' \| 'anthropic' \| 'google';\n\nexport interface SolverOptions {\n /** AI provider to use when constructing the model from an API key (default: \"openai\") /\n provider?: Provider;\n /* Model ID passed to the provider (default: \"gpt-4o\") /\n model?: string;\n}\n\nexport interface SolveOptions {\n /* Number of voting attempts (default: 7) /\n numAttempts?: number;\n /* Expected captcha length — results of other lengths are discarded /\n expectedLength?: number;\n /* Max retries per attempt on API failure (default: 2) /\n maxRetries?: number;\n /* Whether to log attempt details (default: true) /\n verbose?: boolean;\n /\n Confusion groups for majority voting.\n * Pass a Record<string, string> to merge visually similar characters,\n * or `false` to disable (default: false).\n * Use LEGACY_CONFUSION_GROUPS to restore pre-3.0 behavior.\n /\n confusionGroups?: Record<string, string> \| false;\n /* Preprocessing options passed to the image pipeline /\n preprocess?: PreprocessOptions;\n}\n\nexport interface SolveResult {\n /* The solved captcha text (majority-voted) /\n text: string;\n /* Per-attempt raw answers (before voting) /\n attempts: string[];\n /* Aggregated token usage across all parallel attempts /\n usage: LanguageModelUsage;\n /* Per-attempt usage breakdown /\n attemptUsages: LanguageModelUsage[];\n}\n\ninterface AttemptResult {\n text: string;\n usage: LanguageModelUsage;\n}\n\n// ── Provider resolution ──────────────────────────────────────────────\n\nconst DEFAULT_MODELS: Record<Provider, string> = {\n openai: 'gpt-4o',\n anthropic: 'claude-sonnet-4-20250514',\n google: 'gemini-2.0-flash',\n};\n\nasync function resolveModel(\n apiKey: string,\n provider: Provider,\n modelId: string\n): Promise<LanguageModel> {\n switch (provider) {\n case 'openai': {\n const { createOpenAI } = await import('@ai-sdk/openai');\n return createOpenAI({ apiKey })(modelId);\n }\n case 'anthropic': {\n // @ts-expect-error — optional peer dependency\n const { createAnthropic } = await import('@ai-sdk/anthropic');\n return createAnthropic({ apiKey })(modelId);\n }\n case 'google': {\n // @ts-expect-error — optional peer dependency\n const { createGoogleGenerativeAI } = await import('@ai-sdk/google');\n return createGoogleGenerativeAI({ apiKey })(modelId);\n }\n default:\n throw new Error(\n `Unknown provider \"${provider}\". Install the matching @ai-sdk/ package and pass the model directly.`\n );\n }\n}\n\n// ── Confusion groups ─────────────────────────────────────────────────\n\n/*\n Pre-3.0 confusion groups that merge visually similar characters.\n * Opt-in via `{ confusionGroups: LEGACY_CONFUSION_GROUPS }`.\n \n Maps: 1/I/L → '1', O/D/0 → 'O', S/5 → 'S', Z/2 → 'Z'\n /\nexport const LEGACY_CONFUSION_GROUPS: Record<string, string> = {\n '1': '1',\n I: '1',\n L: '1',\n O: 'O',\n D: 'O',\n '0': 'O',\n S: 'S',\n '5': 'S',\n Z: 'Z',\n '2': 'Z',\n};\n\n// ── Majority voting ──────────────────────────────────────────────────\n\n/\n Character-level majority vote across multiple attempts.\n * When `groups` is provided, visually similar characters are merged\n * during counting (e.g. 1/I/L all count toward '1').\n /\nexport function majorityVote(\n attempts: string[],\n expectedLength?: number,\n groups?: Record<string, string> \| false\n): string {\n let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n if (filtered.length === 0) {\n filtered = attempts;\n }\n if (filtered.length === 0) return '';\n\n // Find most common length\n const lenCounts = new Map<number, number>();\n for (const a of filtered) {\n lenCounts.set(a.length, (lenCounts.get(a.length) ?? 0) + 1);\n }\n let bestLen = 0;\n let bestCount = 0;\n for (const [len, count] of lenCounts) {\n if (count > bestCount) {\n bestLen = len;\n bestCount = count;\n }\n }\n\n const sameLenAttempts = filtered.filter((a) => a.length === bestLen);\n if (sameLenAttempts.length === 0) return filtered[0];\n\n const useGroups = groups && typeof groups === 'object' ? groups : undefined;\n\n // Vote per character position\n const result: string[] = [];\n for (let pos = 0; pos < bestLen; pos++) {\n const charCounts = new Map<string, number>();\n for (const a of sameLenAttempts) {\n const ch = a[pos];\n charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);\n }\n\n if (useGroups) {\n // Confusion-aware voting\n const groupCounts = new Map<string, number>();\n for (const [ch, count] of charCounts) {\n const canonical = useGroups[ch] ?? ch;\n groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);\n }\n\n let bestGroup = '';\n let bestGroupCount = 0;\n for (const [canonical, count] of groupCounts) {\n if (count > bestGroupCount) {\n bestGroup = canonical;\n bestGroupCount = count;\n }\n }\n result.push(bestGroup);\n } else {\n // Simple majority — pick the most frequent raw character\n let bestChar = '';\n let bestCharCount = 0;\n for (const [ch, count] of charCounts) {\n if (count > bestCharCount) {\n bestChar = ch;\n bestCharCount = count;\n }\n }\n result.push(bestChar);\n }\n }\n\n return result.join('');\n}\n\n// ── Usage aggregation ────────────────────────────────────────────────\n\nfunction sumOptional(a: number \| undefined, b: number \| undefined): number \| undefined {\n if (a === undefined && b === undefined) return undefined;\n return (a ?? 0) + (b ?? 0);\n}\n\nfunction aggregateUsage(usages: LanguageModelUsage[]): LanguageModelUsage {\n const zero: LanguageModelUsage = {\n inputTokens: undefined,\n inputTokenDetails: {\n noCacheTokens: undefined,\n cacheReadTokens: undefined,\n cacheWriteTokens: undefined,\n },\n outputTokens: undefined,\n outputTokenDetails: {\n textTokens: undefined,\n reasoningTokens: undefined,\n },\n totalTokens: undefined,\n };\n return usages.reduce<LanguageModelUsage>(\n (acc, u) => ({\n inputTokens: sumOptional(acc.inputTokens, u.inputTokens),\n inputTokenDetails: {\n noCacheTokens: sumOptional(\n acc.inputTokenDetails.noCacheTokens,\n u.inputTokenDetails.noCacheTokens\n ),\n cacheReadTokens: sumOptional(\n acc.inputTokenDetails.cacheReadTokens,\n u.inputTokenDetails.cacheReadTokens\n ),\n cacheWriteTokens: sumOptional(\n acc.inputTokenDetails.cacheWriteTokens,\n u.inputTokenDetails.cacheWriteTokens\n ),\n },\n outputTokens: sumOptional(acc.outputTokens, u.outputTokens),\n outputTokenDetails: {\n textTokens: sumOptional(acc.outputTokenDetails.textTokens, u.outputTokenDetails.textTokens),\n reasoningTokens: sumOptional(\n acc.outputTokenDetails.reasoningTokens,\n u.outputTokenDetails.reasoningTokens\n ),\n },\n totalTokens: sumOptional(acc.totalTokens, u.totalTokens),\n }),\n zero\n );\n}\n\n// ── Solver class ─────────────────────────────────────────────────────\n\nexport class Solver {\n private _model: LanguageModel \| null = null;\n private _pendingModel: Promise<LanguageModel> \| null = null;\n\n /\n Create a captcha solver.\n \n @example\n * // Simple — defaults to OpenAI gpt-4o\n * const solver = new Solver('sk-...');\n \n @example\n * // Specify provider and model\n * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });\n \n @example\n * // Pass an AI SDK model directly\n * import { createOpenAI } from '@ai-sdk/openai';\n * const openai = createOpenAI({ apiKey: 'sk-...' });\n * const solver = new Solver(openai('gpt-4o'));\n /\n constructor(keyOrModel: string \| LanguageModel, options?: SolverOptions) {\n if (typeof keyOrModel === 'string') {\n const provider = options?.provider ?? 'openai';\n const modelId = options?.model ?? DEFAULT_MODELS[provider];\n // Lazily resolve the model on first use\n this._pendingModel = resolveModel(keyOrModel, provider, modelId);\n } else {\n this._model = keyOrModel;\n }\n }\n\n private async getModel(): Promise<LanguageModel> {\n if (this._model) return this._model;\n this._model = await this._pendingModel!;\n this._pendingModel = null;\n return this._model;\n }\n\n /\n Solve a captcha image.\n \n @param input - File path (string) or raw image Buffer\n * @param options - Solve options (attempts, expected length, etc.)\n * @returns Solved text, per-attempt answers, and token usage\n /\n async solve(input: string \| Buffer, options: SolveOptions = {}): Promise<SolveResult> {\n const {\n numAttempts = 7,\n expectedLength,\n maxRetries = 2,\n verbose = true,\n confusionGroups = false,\n preprocess,\n } = options;\n\n const model = await this.getModel();\n\n // Two complementary views:\n // 1. Enhanced grayscale (high contrast + auto-crop) — great for clear text\n // 2. Color original (upscaled, no greyscale, no contrast) — preserves subtle features\n const [enhancedBuffer, colorBuffer] = await Promise.all([\n preprocessCaptchaToBuffer(input, preprocess),\n preprocessCaptchaToBuffer(input, {\n blur: 0,\n scale: 4,\n contrast: 1.0,\n sharpen: false,\n crop: 'none',\n padding: 40,\n greyscale: false,\n }),\n ]);\n\n // Fire all attempts in parallel for speed\n const results = await Promise.all(\n Array.from({ length: numAttempts }, () =>\n this.singleAttempt(model, enhancedBuffer, colorBuffer, maxRetries)\n )\n );\n const valid = results.filter((r): r is AttemptResult => r !== null);\n if (verbose) {\n valid.forEach((r, i) => console.log(` Attempt ${i + 1}: ${r.text}`));\n }\n\n const attempts = valid.map((r) => r.text);\n const attemptUsages = valid.map((r) => r.usage);\n const usage = aggregateUsage(attemptUsages);\n\n if (attempts.length === 0) {\n if (verbose) console.log(' All attempts failed!');\n return { text: '', attempts, usage, attemptUsages };\n }\n\n return {\n text: majorityVote(attempts, expectedLength, confusionGroups),\n attempts,\n usage,\n attemptUsages,\n };\n }\n\n /\n Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n /\n private async singleAttempt(\n model: LanguageModel,\n primaryBuffer: Buffer,\n secondaryBuffer: Buffer,\n maxRetries: number\n ): Promise<AttemptResult \| null> {\n for (let retry = 0; retry <= maxRetries; retry++) {\n try {\n const { text, usage } = await generateText({\n model,\n messages: [\n {\n role: 'user',\n content: [\n { type: 'text', text: PROMPT },\n { type: 'image', image: primaryBuffer },\n { type: 'image', image: secondaryBuffer },\n ],\n },\n ],\n temperature: 1,\n maxOutputTokens: 256,\n });\n\n const raw = text.trim();\n\n // Detect refusals\n const lower = raw.toLowerCase();\n if (\n lower.includes('sorry') \|\|\n lower.includes(\"can't help\") \|\|\n lower.includes('cannot help') \|\|\n lower.includes('unable to') \|\|\n lower.includes(\"i can't\") \|\|\n raw.length > 20\n ) {\n return null;\n }\n\n // Clean: keep only uppercase letters and digits\n const cleaned = raw.replace(/[^A-Za-z0-9]/g, '');\n return cleaned ? { text: cleaned, usage } : null;\n } catch (_err) {\n if (retry < maxRetries) {\n await new Promise((r) => setTimeout(r, 1000 (retry + 1)));\n continue;\n }\n return null;\n }\n }\n return null;\n }\n}\n","import fs from 'fs';\nimport path from 'path';\nimport sharp from 'sharp';\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport interface CropFractions {\n /** Fraction from left edge (0–1, default: 0.1) /\n left: number;\n /* Fraction from top edge (0–1, default: 0.02) /\n top: number;\n /* Fraction from left to keep (0–1, default: 0.9) /\n right: number;\n /* Fraction from top to keep (0–1, default: 0.6) /\n bottom: number;\n}\n\nexport interface PreprocessOptions {\n /* Gaussian blur radius (default: 1.5). Set to 0 to skip. /\n blur?: number;\n /* Upscale factor (default: 4) /\n scale?: number;\n /* Contrast multiplier around image mean (default: 3.0). Set to 1 to skip. /\n contrast?: number;\n /* Enable unsharp-mask sharpening (default: true) /\n sharpen?: boolean;\n /\n Crop mode (default: 'auto'):\n * - 'auto' – trim whitespace after contrast enhancement, with margin\n * - 'legacy' – fixed-percentage crop (original behavior)\n * - 'none' – skip cropping\n * - CropFractions – custom crop percentages\n /\n crop?: 'auto' \| 'legacy' \| 'none' \| CropFractions;\n /* Add white padding around the result (default: true). Pass false to skip, or a number for custom px. /\n padding?: boolean \| number;\n /* Invert colors (negate) after processing (default: false) /\n negate?: boolean;\n /* Convert to greyscale (default: true) /\n greyscale?: boolean;\n}\n\nconst LEGACY_CROP: CropFractions = { left: 0.1, top: 0.02, right: 0.9, bottom: 0.6 };\n\n// ── Public API ───────────────────────────────────────────────────────\n\n/\n Preprocess a captcha image and return a base64-encoded PNG string.\n /\nexport async function preprocessCaptcha(\n input: string \| Buffer,\n options?: PreprocessOptions\n): Promise<string> {\n const buf = await preprocessCaptchaToBuffer(input, options);\n return buf.toString('base64');\n}\n\n/\n Preprocess a captcha image and return the resulting PNG as a raw Buffer.\n \n Pipeline:\n * 1. Gaussian blur in color space (smooths dither pattern)\n * 2. Grayscale conversion\n * 3. Upscale with Lanczos\n * 4. Contrast boost around image mean + sharpen\n * 5. Crop (auto-detect, legacy fixed, none, or custom)\n * 6. Add white padding\n /\nexport async function preprocessCaptchaToBuffer(\n input: string \| Buffer,\n options?: PreprocessOptions\n): Promise<Buffer> {\n const {\n blur = 1.5,\n scale = 4,\n contrast = 3.0,\n sharpen = true,\n crop = 'auto',\n padding = true,\n negate = false,\n greyscale = true,\n } = options ?? {};\n\n const source = typeof input === 'string' ? path.resolve(input) : input;\n\n // Read original dimensions\n const metadata = await sharp(source).metadata();\n const origW = metadata.width!;\n const origH = metadata.height!;\n\n // Step 1-2: Blur (optional) + greyscale (optional)\n let pipeline = sharp(source);\n if (blur > 0) pipeline = pipeline.blur(blur);\n if (greyscale) pipeline = pipeline.greyscale();\n const smoothed = await pipeline.toBuffer();\n\n // Step 3: Upscale with Lanczos\n const upscaled = await sharp(smoothed)\n .resize(origW scale, origH * scale, { kernel: 'lanczos3' })\n .toBuffer();\n\n // Step 4: Contrast boost + sharpen\n let enhanced: Buffer;\n if (contrast !== 1.0) {\n const stats = await sharp(upscaled).stats();\n const mean = stats.channels[0].mean;\n let pipe = sharp(upscaled).linear(contrast, mean * (1 - contrast));\n if (sharpen) pipe = pipe.sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 });\n enhanced = await pipe.toBuffer();\n } else {\n enhanced = sharpen\n ? await sharp(upscaled).sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 }).toBuffer()\n : upscaled;\n }\n\n // Step 5: Crop\n let cropped: Buffer;\n if (crop === 'none') {\n cropped = enhanced;\n } else if (crop === 'auto') {\n cropped = await autoCrop(enhanced);\n } else {\n const fractions = crop === 'legacy' ? LEGACY_CROP : crop;\n const scaledW = origW * scale;\n const scaledH = origH * scale;\n const cropLeft = Math.floor(scaledW * fractions.left);\n const cropTop = Math.floor(scaledH * fractions.top);\n const cropRight = Math.floor(scaledW * fractions.right);\n const cropBottom = Math.floor(scaledH * fractions.bottom);\n const cropW = cropRight - cropLeft;\n const cropH = cropBottom - cropTop;\n cropped = await sharp(enhanced)\n .extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH })\n .toBuffer();\n }\n\n // Step 6: Negate (optional)\n const final = negate ? await sharp(cropped).negate().toBuffer() : cropped;\n\n // Step 7: Padding\n if (padding === false) {\n return sharp(final).png().toBuffer();\n }\n const pad = typeof padding === 'number' ? padding : undefined;\n const vPad = pad ?? 20;\n const hPad = pad ?? 30;\n return sharp(final)\n .extend({\n top: vPad,\n bottom: vPad,\n left: hPad,\n right: hPad,\n background: { r: 255, g: 255, b: 255 },\n })\n .png()\n .toBuffer();\n}\n\n/*\n Auto-crop: use sharp.trim() to detect the content bounding box after\n * contrast enhancement, then add a small margin. Falls back to the\n * untrimmed image if trim removes everything.\n /\nasync function autoCrop(enhanced: Buffer): Promise<Buffer> {\n try {\n const trimmed = sharp(enhanced).trim({ threshold: 30 });\n const trimmedBuf = await trimmed.toBuffer({ resolveWithObject: true });\n\n // If trim left a reasonable image, add a margin\n const { width, height } = trimmedBuf.info;\n if (width > 2 && height > 2) {\n return trimmedBuf.data;\n }\n } catch {\n // trim() can throw if image is uniform — fall through\n }\n\n // Fallback: return untrimmed\n return enhanced;\n}\n\n/\n Read an image file and return its base64-encoded content.\n */\nexport function imageToBase64(imagePath: string): string {\n const buffer = fs.readFileSync(imagePath);\n return buffer.toString('base64');\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACCA,gBAA6B;;;ACD7B,gBAAe;AACf,kBAAiB;AACjB,mBAAkB;AAwClB,IAAM,cAA6B,EAAE,MAAM,KAAK,KAAK,MAAM,OAAO,KAAK,QAAQ,IAAI;AAOnF,eAAsB,kBACpB,OACA,SACiB;AACjB,QAAM,MAAM,MAAM,0BAA0B,OAAO,OAAO;AAC1D,SAAO,IAAI,SAAS,QAAQ;AAC9B;AAaA,eAAsB,0BACpB,OACA,SACiB;AACjB,QAAM;AAAA,IACJ,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,WAAW;AAAA,IACX,UAAU;AAAA,IACV,OAAO;AAAA,IACP,UAAU;AAAA,IACV,SAAS;AAAA,IACT,YAAY;AAAA,EACd,IAAI,WAAW,CAAC;AAEhB,QAAM,SAAS,OAAO,UAAU,WAAW,YAAAA,QAAK,QAAQ,KAAK,IAAI;AAGjE,QAAM,WAAW,UAAM,aAAAC,SAAM,MAAM,EAAE,SAAS;AAC9C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAGvB,MAAI,eAAW,aAAAA,SAAM,MAAM;AAC3B,MAAI,OAAO,EAAG,YAAW,SAAS,KAAK,IAAI;AAC3C,MAAI,UAAW,YAAW,SAAS,UAAU;AAC7C,QAAM,WAAW,MAAM,SAAS,SAAS;AAGzC,QAAM,WAAW,UAAM,aAAAA,SAAM,QAAQ,EAClC,OAAO,QAAQ,OAAO,QAAQ,OAAO,EAAE,QAAQ,WAAW,CAAC,EAC3D,SAAS;AAGZ,MAAI;AACJ,MAAI,aAAa,GAAK;AACpB,UAAM,QAAQ,UAAM,aAAAA,SAAM,QAAQ,EAAE,MAAM;AAC1C,UAAM,OAAO,MAAM,SAAS,CAAC,EAAE;AAC/B,QAAI,WAAO,aAAAA,SAAM,QAAQ,EAAE,OAAO,UAAU,QAAQ,IAAI,SAAS;AACjE,QAAI,QAAS,QAAO,KAAK,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC;AACjE,eAAW,MAAM,KAAK,SAAS;AAAA,EACjC,OAAO;AACL,eAAW,UACP,UAAM,aAAAA,SAAM,QAAQ,EAAE,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC,EAAE,SAAS,IACzE;AAAA,EACN;AAGA,MAAI;AACJ,MAAI,SAAS,QAAQ;AACnB,cAAU;AAAA,EACZ,WAAW,SAAS,QAAQ;AAC1B,cAAU,MAAM,SAAS,QAAQ;AAAA,EACnC,OAAO;AACL,UAAM,YAAY,SAAS,WAAW,cAAc;AACpD,UAAM,UAAU,QAAQ;AACxB,UAAM,UAAU,QAAQ;AACxB,UAAM,WAAW,KAAK,MAAM,UAAU,UAAU,IAAI;AACpD,UAAM,UAAU,KAAK,MAAM,UAAU,UAAU,GAAG;AAClD,UAAM,YAAY,KAAK,MAAM,UAAU,UAAU,KAAK;AACtD,UAAM,aAAa,KAAK,MAAM,UAAU,UAAU,MAAM;AACxD,UAAM,QAAQ,YAAY;AAC1B,UAAM,QAAQ,aAAa;AAC3B,cAAU,UAAM,aAAAA,SAAM,QAAQ,EAC3B,QAAQ,EAAE,MAAM,UAAU,KAAK,SAAS,OAAO,OAAO,QAAQ,MAAM,CAAC,EACrE,SAAS;AAAA,EACd;AAGA,QAAM,QAAQ,SAAS,UAAM,aAAAA,SAAM,OAAO,EAAE,OAAO,EAAE,SAAS,IAAI;AAGlE,MAAI,YAAY,OAAO;AACrB,eAAO,aAAAA,SAAM,KAAK,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC;AACA,QAAM,MAAM,OAAO,YAAY,WAAW,UAAU;AACpD,QAAM,OAAO,OAAO;AACpB,QAAM,OAAO,OAAO;AACpB,aAAO,aAAAA,SAAM,KAAK,EACf,OAAO;AAAA,IACN,KAAK;AAAA,IACL,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,OAAO;AAAA,IACP,YAAY,EAAE,GAAG,KAAK,GAAG,KAAK,GAAG,IAAI;AAAA,EACvC,CAAC,EACA,IAAI,EACJ,SAAS;AACd;AAOA,eAAe,SAAS,UAAmC;AACzD,MAAI;AACF,UAAM,cAAU,aAAAA,SAAM,QAAQ,EAAE,KAAK,EAAE,WAAW,GAAG,CAAC;AACtD,UAAM,aAAa,MAAM,QAAQ,SAAS,EAAE,mBAAmB,KAAK,CAAC;AAGrE,UAAM,EAAE,OAAO,OAAO,IAAI,WAAW;AACrC,QAAI,QAAQ,KAAK,SAAS,GAAG;AAC3B,aAAO,WAAW;AAAA,IACpB;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,SAAO;AACT;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,UAAAC,QAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;ADtLA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAyDf,IAAM,iBAA2C;AAAA,EAC/C,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,QAAQ;AACV;AAEA,eAAe,aACb,QACA,UACA,SACwB;AACxB,UAAQ,UAAU;AAAA,IAChB,KAAK,UAAU;AACb,YAAM,EAAE,aAAa,IAAI,MAAM,OAAO,gBAAgB;AACtD,aAAO,aAAa,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACzC;AAAA,IACA,KAAK,aAAa;AAEhB,YAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,mBAAmB;AAC5D,aAAO,gBAAgB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IAC5C;AAAA,IACA,KAAK,UAAU;AAEb,YAAM,EAAE,yBAAyB,IAAI,MAAM,OAAO,gBAAgB;AAClE,aAAO,yBAAyB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACrD;AAAA,IACA;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,QAAQ;AAAA,MAC/B;AAAA,EACJ;AACF;AAUO,IAAM,0BAAkD;AAAA,EAC7D,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AACP;AASO,SAAS,aACd,UACA,gBACA,QACQ;AACR,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAEtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAGlC,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,KAAK,UAAU;AACxB,cAAU,IAAI,EAAE,SAAS,UAAU,IAAI,EAAE,MAAM,KAAK,KAAK,CAAC;AAAA,EAC5D;AACA,MAAI,UAAU;AACd,MAAI,YAAY;AAChB,aAAW,CAAC,KAAK,KAAK,KAAK,WAAW;AACpC,QAAI,QAAQ,WAAW;AACrB,gBAAU;AACV,kBAAY;AAAA,IACd;AAAA,EACF;AAEA,QAAM,kBAAkB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnE,MAAI,gBAAgB,WAAW,EAAG,QAAO,SAAS,CAAC;AAEnD,QAAM,YAAY,UAAU,OAAO,WAAW,WAAW,SAAS;AAGlE,QAAM,SAAmB,CAAC;AAC1B,WAAS,MAAM,GAAG,MAAM,SAAS,OAAO;AACtC,UAAM,aAAa,oBAAI,IAAoB;AAC3C,eAAW,KAAK,iBAAiB;AAC/B,YAAM,KAAK,EAAE,GAAG;AAChB,iBAAW,IAAI,KAAK,WAAW,IAAI,EAAE,KAAK,KAAK,CAAC;AAAA,IAClD;AAEA,QAAI,WAAW;AAEb,YAAM,cAAc,oBAAI,IAAoB;AAC5C,iBAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,cAAM,YAAY,UAAU,EAAE,KAAK;AACnC,oBAAY,IAAI,YAAY,YAAY,IAAI,SAAS,KAAK,KAAK,KAAK;AAAA,MACtE;AAEA,UAAI,YAAY;AAChB,UAAI,iBAAiB;AACrB,iBAAW,CAAC,WAAW,KAAK,KAAK,aAAa;AAC5C,YAAI,QAAQ,gBAAgB;AAC1B,sBAAY;AACZ,2BAAiB;AAAA,QACnB;AAAA,MACF;AACA,aAAO,KAAK,SAAS;AAAA,IACvB,OAAO;AAEL,UAAI,WAAW;AACf,UAAI,gBAAgB;AACpB,iBAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,YAAI,QAAQ,eAAe;AACzB,qBAAW;AACX,0BAAgB;AAAA,QAClB;AAAA,MACF;AACA,aAAO,KAAK,QAAQ;AAAA,IACtB;AAAA,EACF;AAEA,SAAO,OAAO,KAAK,EAAE;AACvB;AAIA,SAAS,YAAY,GAAuB,GAA2C;AACrF,MAAI,MAAM,UAAa,MAAM,OAAW,QAAO;AAC/C,UAAQ,KAAK,MAAM,KAAK;AAC1B;AAEA,SAAS,eAAe,QAAkD;AACxE,QAAM,OAA2B;AAAA,IAC/B,aAAa;AAAA,IACb,mBAAmB;AAAA,MACjB,eAAe;AAAA,MACf,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,IACpB;AAAA,IACA,cAAc;AAAA,IACd,oBAAoB;AAAA,MAClB,YAAY;AAAA,MACZ,iBAAiB;AAAA,IACnB;AAAA,IACA,aAAa;AAAA,EACf;AACA,SAAO,OAAO;AAAA,IACZ,CAAC,KAAK,OAAO;AAAA,MACX,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,MACvD,mBAAmB;AAAA,QACjB,eAAe;AAAA,UACb,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,iBAAiB;AAAA,UACf,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,kBAAkB;AAAA,UAChB,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,MACF;AAAA,MACA,cAAc,YAAY,IAAI,cAAc,EAAE,YAAY;AAAA,MAC1D,oBAAoB;AAAA,QAClB,YAAY,YAAY,IAAI,mBAAmB,YAAY,EAAE,mBAAmB,UAAU;AAAA,QAC1F,iBAAiB;AAAA,UACf,IAAI,mBAAmB;AAAA,UACvB,EAAE,mBAAmB;AAAA,QACvB;AAAA,MACF;AAAA,MACA,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,IACzD;AAAA,IACA;AAAA,EACF;AACF;AAIO,IAAM,SAAN,MAAa;AAAA,EACV,SAA+B;AAAA,EAC/B,gBAA+C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAmBvD,YAAY,YAAoC,SAAyB;AACvE,QAAI,OAAO,eAAe,UAAU;AAClC,YAAM,WAAW,SAAS,YAAY;AACtC,YAAM,UAAU,SAAS,SAAS,eAAe,QAAQ;AAEzD,WAAK,gBAAgB,aAAa,YAAY,UAAU,OAAO;AAAA,IACjE,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAc,WAAmC;AAC/C,QAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,SAAK,SAAS,MAAM,KAAK;AACzB,SAAK,gBAAgB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,MAAM,OAAwB,UAAwB,CAAC,GAAyB;AACpF,UAAM;AAAA,MACJ,cAAc;AAAA,MACd;AAAA,MACA,aAAa;AAAA,MACb,UAAU;AAAA,MACV,kBAAkB;AAAA,MAClB;AAAA,IACF,IAAI;AAEJ,UAAM,QAAQ,MAAM,KAAK,SAAS;AAKlC,UAAM,CAAC,gBAAgB,WAAW,IAAI,MAAM,QAAQ,IAAI;AAAA,MACtD,0BAA0B,OAAO,UAAU;AAAA,MAC3C,0BAA0B,OAAO;AAAA,QAC/B,MAAM;AAAA,QACN,OAAO;AAAA,QACP,UAAU;AAAA,QACV,SAAS;AAAA,QACT,MAAM;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,MACb,CAAC;AAAA,IACH,CAAC;AAGD,UAAM,UAAU,MAAM,QAAQ;AAAA,MAC5B,MAAM;AAAA,QAAK,EAAE,QAAQ,YAAY;AAAA,QAAG,MAClC,KAAK,cAAc,OAAO,gBAAgB,aAAa,UAAU;AAAA,MACnE;AAAA,IACF;AACA,UAAM,QAAQ,QAAQ,OAAO,CAAC,MAA0B,MAAM,IAAI;AAClE,QAAI,SAAS;AACX,YAAM,QAAQ,CAAC,GAAG,MAAM,QAAQ,IAAI,aAAa,IAAI,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC;AAAA,IACtE;AAEA,UAAM,WAAW,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI;AACxC,UAAM,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,eAAe,aAAa;AAE1C,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO,EAAE,MAAM,IAAI,UAAU,OAAO,cAAc;AAAA,IACpD;AAEA,WAAO;AAAA,MACL,MAAM,aAAa,UAAU,gBAAgB,eAAe;AAAA,MAC5D;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,eACA,iBACA,YAC+B;AAC/B,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,MAAM,MAAM,IAAI,UAAM,wBAAa;AAAA,UACzC;AAAA,UACA,UAAU;AAAA,YACR;AAAA,cACE,MAAM;AAAA,cACN,SAAS;AAAA,gBACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,gBAC7B,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,gBACtC,EAAE,MAAM,SAAS,OAAO,gBAAgB;AAAA,cAC1C;AAAA,YACF;AAAA,UACF;AAAA,UACA,aAAa;AAAA,UACb,iBAAiB;AAAA,QACnB,CAAC;AAED,cAAM,MAAM,KAAK,KAAK;AAGtB,cAAM,QAAQ,IAAI,YAAY;AAC9B,YACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,iBAAO;AAAA,QACT;AAGA,cAAM,UAAU,IAAI,QAAQ,iBAAiB,EAAE;AAC/C,eAAO,UAAU,EAAE,MAAM,SAAS,MAAM,IAAI;AAAA,MAC9C,SAAS,MAAM;AACb,YAAI,QAAQ,YAAY;AACtB,gBAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;","names":["path","sharp","fs"]}

package/dist/index.d.cts CHANGED Viewed

@@ -1,6 +1,61 @@
 import { LanguageModelUsage, LanguageModel } from 'ai';
 export { LanguageModelUsage } from 'ai';
+interface CropFractions {
+    /** Fraction from left edge (0–1, default: 0.1) */
+    left: number;
+    /** Fraction from top edge (0–1, default: 0.02) */
+    top: number;
+    /** Fraction from left to keep (0–1, default: 0.9) */
+    right: number;
+    /** Fraction from top to keep (0–1, default: 0.6) */
+    bottom: number;
+}
+interface PreprocessOptions {
+    /** Gaussian blur radius (default: 1.5). Set to 0 to skip. */
+    blur?: number;
+    /** Upscale factor (default: 4) */
+    scale?: number;
+    /** Contrast multiplier around image mean (default: 3.0). Set to 1 to skip. */
+    contrast?: number;
+    /** Enable unsharp-mask sharpening (default: true) */
+    sharpen?: boolean;
+    /**
+     * Crop mode (default: 'auto'):
+     *  - 'auto'   – trim whitespace after contrast enhancement, with margin
+     *  - 'legacy' – fixed-percentage crop (original behavior)
+     *  - 'none'   – skip cropping
+     *  - CropFractions – custom crop percentages
+     */
+    crop?: 'auto' | 'legacy' | 'none' | CropFractions;
+    /** Add white padding around the result (default: true). Pass false to skip, or a number for custom px. */
+    padding?: boolean | number;
+    /** Invert colors (negate) after processing (default: false) */
+    negate?: boolean;
+    /** Convert to greyscale (default: true) */
+    greyscale?: boolean;
+}
+/**
+ * Preprocess a captcha image and return a base64-encoded PNG string.
+ */
+declare function preprocessCaptcha(input: string | Buffer, options?: PreprocessOptions): Promise<string>;
+/**
+ * Preprocess a captcha image and return the resulting PNG as a raw Buffer.
+ *
+ * Pipeline:
+ *   1. Gaussian blur in color space (smooths dither pattern)
+ *   2. Grayscale conversion
+ *   3. Upscale with Lanczos
+ *   4. Contrast boost around image mean + sharpen
+ *   5. Crop (auto-detect, legacy fixed, none, or custom)
+ *   6. Add white padding
+ */
+declare function preprocessCaptchaToBuffer(input: string | Buffer, options?: PreprocessOptions): Promise<Buffer>;
+/**
+ * Read an image file and return its base64-encoded content.
+ */
+declare function imageToBase64(imagePath: string): string;
 type Provider = 'openai' | 'anthropic' | 'google';
 interface SolverOptions {
     /** AI provider to use when constructing the model from an API key (default: "openai") */
@@ -9,7 +64,7 @@ interface SolverOptions {
     model?: string;
 }
 interface SolveOptions {
-    /** Number of voting attempts (default: 5) */
+    /** Number of voting attempts (default: 7) */
     numAttempts?: number;
     /** Expected captcha length — results of other lengths are discarded */
     expectedLength?: number;
@@ -17,6 +72,15 @@ interface SolveOptions {
     maxRetries?: number;
     /** Whether to log attempt details (default: true) */
     verbose?: boolean;
+    /**
+     * Confusion groups for majority voting.
+     * Pass a Record<string, string> to merge visually similar characters,
+     * or `false` to disable (default: false).
+     * Use LEGACY_CONFUSION_GROUPS to restore pre-3.0 behavior.
+     */
+    confusionGroups?: Record<string, string> | false;
+    /** Preprocessing options passed to the image pipeline */
+    preprocess?: PreprocessOptions;
 }
 interface SolveResult {
     /** The solved captcha text (majority-voted) */
@@ -28,6 +92,19 @@ interface SolveResult {
     /** Per-attempt usage breakdown */
     attemptUsages: LanguageModelUsage[];
 }
+/**
+ * Pre-3.0 confusion groups that merge visually similar characters.
+ * Opt-in via `{ confusionGroups: LEGACY_CONFUSION_GROUPS }`.
+ *
+ * Maps: 1/I/L → '1', O/D/0 → 'O', S/5 → 'S', Z/2 → 'Z'
+ */
+declare const LEGACY_CONFUSION_GROUPS: Record<string, string>;
+/**
+ * Character-level majority vote across multiple attempts.
+ * When `groups` is provided, visually similar characters are merged
+ * during counting (e.g. 1/I/L all count toward '1').
+ */
+declare function majorityVote(attempts: string[], expectedLength?: number, groups?: Record<string, string> | false): string;
 declare class Solver {
     private _model;
     private _pendingModel;
@@ -65,29 +142,4 @@ declare class Solver {
     private singleAttempt;
 }
-/**
- * Preprocess a captcha image using sharp (libvips).
- *
- * Pipeline:
- *   1. Gaussian blur in color space (smooths dither pattern)
- *   2. Grayscale conversion
- *   3. Upscale 4× with Lanczos
- *   4. Contrast boost (3× around image mean) + sharpen
- *   5. Crop decorative borders
- *   6. Add white padding
- *
- * Accepts a file path or a raw image Buffer.
- * Returns a base64-encoded PNG string.
- */
-declare function preprocessCaptcha(input: string | Buffer): Promise<string>;
-/**
- * Same preprocessing pipeline as `preprocessCaptcha`, but returns the
- * resulting PNG as a raw Buffer (useful for AI SDK image content parts).
- */
-declare function preprocessCaptchaToBuffer(input: string | Buffer): Promise<Buffer>;
-/**
- * Read an image file and return its base64-encoded content.
- */
-declare function imageToBase64(imagePath: string): string;
-export { type Provider, type SolveOptions, type SolveResult, Solver, type SolverOptions, imageToBase64, preprocessCaptcha, preprocessCaptchaToBuffer };
+export { type CropFractions, LEGACY_CONFUSION_GROUPS, type PreprocessOptions, type Provider, type SolveOptions, type SolveResult, Solver, type SolverOptions, imageToBase64, majorityVote, preprocessCaptcha, preprocessCaptchaToBuffer };

package/dist/index.d.ts CHANGED Viewed

@@ -1,6 +1,61 @@
 import { LanguageModelUsage, LanguageModel } from 'ai';
 export { LanguageModelUsage } from 'ai';
+interface CropFractions {
+    /** Fraction from left edge (0–1, default: 0.1) */
+    left: number;
+    /** Fraction from top edge (0–1, default: 0.02) */
+    top: number;
+    /** Fraction from left to keep (0–1, default: 0.9) */
+    right: number;
+    /** Fraction from top to keep (0–1, default: 0.6) */
+    bottom: number;
+}
+interface PreprocessOptions {
+    /** Gaussian blur radius (default: 1.5). Set to 0 to skip. */
+    blur?: number;
+    /** Upscale factor (default: 4) */
+    scale?: number;
+    /** Contrast multiplier around image mean (default: 3.0). Set to 1 to skip. */
+    contrast?: number;
+    /** Enable unsharp-mask sharpening (default: true) */
+    sharpen?: boolean;
+    /**
+     * Crop mode (default: 'auto'):
+     *  - 'auto'   – trim whitespace after contrast enhancement, with margin
+     *  - 'legacy' – fixed-percentage crop (original behavior)
+     *  - 'none'   – skip cropping
+     *  - CropFractions – custom crop percentages
+     */
+    crop?: 'auto' | 'legacy' | 'none' | CropFractions;
+    /** Add white padding around the result (default: true). Pass false to skip, or a number for custom px. */
+    padding?: boolean | number;
+    /** Invert colors (negate) after processing (default: false) */
+    negate?: boolean;
+    /** Convert to greyscale (default: true) */
+    greyscale?: boolean;
+}
+/**
+ * Preprocess a captcha image and return a base64-encoded PNG string.
+ */
+declare function preprocessCaptcha(input: string | Buffer, options?: PreprocessOptions): Promise<string>;
+/**
+ * Preprocess a captcha image and return the resulting PNG as a raw Buffer.
+ *
+ * Pipeline:
+ *   1. Gaussian blur in color space (smooths dither pattern)
+ *   2. Grayscale conversion
+ *   3. Upscale with Lanczos
+ *   4. Contrast boost around image mean + sharpen
+ *   5. Crop (auto-detect, legacy fixed, none, or custom)
+ *   6. Add white padding
+ */
+declare function preprocessCaptchaToBuffer(input: string | Buffer, options?: PreprocessOptions): Promise<Buffer>;
+/**
+ * Read an image file and return its base64-encoded content.
+ */
+declare function imageToBase64(imagePath: string): string;
 type Provider = 'openai' | 'anthropic' | 'google';
 interface SolverOptions {
     /** AI provider to use when constructing the model from an API key (default: "openai") */
@@ -9,7 +64,7 @@ interface SolverOptions {
     model?: string;
 }
 interface SolveOptions {
-    /** Number of voting attempts (default: 5) */
+    /** Number of voting attempts (default: 7) */
     numAttempts?: number;
     /** Expected captcha length — results of other lengths are discarded */
     expectedLength?: number;
@@ -17,6 +72,15 @@ interface SolveOptions {
     maxRetries?: number;
     /** Whether to log attempt details (default: true) */
     verbose?: boolean;
+    /**
+     * Confusion groups for majority voting.
+     * Pass a Record<string, string> to merge visually similar characters,
+     * or `false` to disable (default: false).
+     * Use LEGACY_CONFUSION_GROUPS to restore pre-3.0 behavior.
+     */
+    confusionGroups?: Record<string, string> | false;
+    /** Preprocessing options passed to the image pipeline */
+    preprocess?: PreprocessOptions;
 }
 interface SolveResult {
     /** The solved captcha text (majority-voted) */
@@ -28,6 +92,19 @@ interface SolveResult {
     /** Per-attempt usage breakdown */
     attemptUsages: LanguageModelUsage[];
 }
+/**
+ * Pre-3.0 confusion groups that merge visually similar characters.
+ * Opt-in via `{ confusionGroups: LEGACY_CONFUSION_GROUPS }`.
+ *
+ * Maps: 1/I/L → '1', O/D/0 → 'O', S/5 → 'S', Z/2 → 'Z'
+ */
+declare const LEGACY_CONFUSION_GROUPS: Record<string, string>;
+/**
+ * Character-level majority vote across multiple attempts.
+ * When `groups` is provided, visually similar characters are merged
+ * during counting (e.g. 1/I/L all count toward '1').
+ */
+declare function majorityVote(attempts: string[], expectedLength?: number, groups?: Record<string, string> | false): string;
 declare class Solver {
     private _model;
     private _pendingModel;
@@ -65,29 +142,4 @@ declare class Solver {
     private singleAttempt;
 }
-/**
- * Preprocess a captcha image using sharp (libvips).
- *
- * Pipeline:
- *   1. Gaussian blur in color space (smooths dither pattern)
- *   2. Grayscale conversion
- *   3. Upscale 4× with Lanczos
- *   4. Contrast boost (3× around image mean) + sharpen
- *   5. Crop decorative borders
- *   6. Add white padding
- *
- * Accepts a file path or a raw image Buffer.
- * Returns a base64-encoded PNG string.
- */
-declare function preprocessCaptcha(input: string | Buffer): Promise<string>;
-/**
- * Same preprocessing pipeline as `preprocessCaptcha`, but returns the
- * resulting PNG as a raw Buffer (useful for AI SDK image content parts).
- */
-declare function preprocessCaptchaToBuffer(input: string | Buffer): Promise<Buffer>;
-/**
- * Read an image file and return its base64-encoded content.
- */
-declare function imageToBase64(imagePath: string): string;
-export { type Provider, type SolveOptions, type SolveResult, Solver, type SolverOptions, imageToBase64, preprocessCaptcha, preprocessCaptchaToBuffer };
+export { type CropFractions, LEGACY_CONFUSION_GROUPS, type PreprocessOptions, type Provider, type SolveOptions, type SolveResult, Solver, type SolverOptions, imageToBase64, majorityVote, preprocessCaptcha, preprocessCaptchaToBuffer };

package/dist/index.js CHANGED Viewed

@@ -5,47 +5,98 @@ import { generateText } from "ai";
 import fs from "fs";
 import path from "path";
 import sharp from "sharp";
-async function preprocessCaptcha(input) {
-  const buf = await preprocessCaptchaToBuffer(input);
+var LEGACY_CROP = { left: 0.1, top: 0.02, right: 0.9, bottom: 0.6 };
+async function preprocessCaptcha(input, options) {
+  const buf = await preprocessCaptchaToBuffer(input, options);
   return buf.toString("base64");
 }
-async function preprocessCaptchaToBuffer(input) {
+async function preprocessCaptchaToBuffer(input, options) {
+  const {
+    blur = 1.5,
+    scale = 4,
+    contrast = 3,
+    sharpen = true,
+    crop = "auto",
+    padding = true,
+    negate = false,
+    greyscale = true
+  } = options ?? {};
   const source = typeof input === "string" ? path.resolve(input) : input;
   const metadata = await sharp(source).metadata();
   const origW = metadata.width;
   const origH = metadata.height;
-  const smoothed = await sharp(source).blur(1.5).greyscale().toBuffer();
-  const upscaled = await sharp(smoothed).resize(origW * 4, origH * 4, { kernel: "lanczos3" }).toBuffer();
-  const stats = await sharp(upscaled).stats();
-  const mean = stats.channels[0].mean;
-  const enhanced = await sharp(upscaled).linear(3, mean * (1 - 3)).sharpen({ sigma: 1, m1: 2, m2: 1 }).toBuffer();
-  const scaledW = origW * 4;
-  const scaledH = origH * 4;
-  const cropLeft = Math.floor(scaledW * 0.1);
-  const cropTop = Math.floor(scaledH * 0.02);
-  const cropRight = Math.floor(scaledW * 0.9);
-  const cropBottom = Math.floor(scaledH * 0.6);
-  const cropW = cropRight - cropLeft;
-  const cropH = cropBottom - cropTop;
-  return sharp(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
-    top: 20,
-    bottom: 20,
-    left: 30,
-    right: 30,
+  let pipeline = sharp(source);
+  if (blur > 0) pipeline = pipeline.blur(blur);
+  if (greyscale) pipeline = pipeline.greyscale();
+  const smoothed = await pipeline.toBuffer();
+  const upscaled = await sharp(smoothed).resize(origW * scale, origH * scale, { kernel: "lanczos3" }).toBuffer();
+  let enhanced;
+  if (contrast !== 1) {
+    const stats = await sharp(upscaled).stats();
+    const mean = stats.channels[0].mean;
+    let pipe = sharp(upscaled).linear(contrast, mean * (1 - contrast));
+    if (sharpen) pipe = pipe.sharpen({ sigma: 1, m1: 2, m2: 1 });
+    enhanced = await pipe.toBuffer();
+  } else {
+    enhanced = sharpen ? await sharp(upscaled).sharpen({ sigma: 1, m1: 2, m2: 1 }).toBuffer() : upscaled;
+  }
+  let cropped;
+  if (crop === "none") {
+    cropped = enhanced;
+  } else if (crop === "auto") {
+    cropped = await autoCrop(enhanced);
+  } else {
+    const fractions = crop === "legacy" ? LEGACY_CROP : crop;
+    const scaledW = origW * scale;
+    const scaledH = origH * scale;
+    const cropLeft = Math.floor(scaledW * fractions.left);
+    const cropTop = Math.floor(scaledH * fractions.top);
+    const cropRight = Math.floor(scaledW * fractions.right);
+    const cropBottom = Math.floor(scaledH * fractions.bottom);
+    const cropW = cropRight - cropLeft;
+    const cropH = cropBottom - cropTop;
+    cropped = await sharp(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).toBuffer();
+  }
+  const final = negate ? await sharp(cropped).negate().toBuffer() : cropped;
+  if (padding === false) {
+    return sharp(final).png().toBuffer();
+  }
+  const pad = typeof padding === "number" ? padding : void 0;
+  const vPad = pad ?? 20;
+  const hPad = pad ?? 30;
+  return sharp(final).extend({
+    top: vPad,
+    bottom: vPad,
+    left: hPad,
+    right: hPad,
     background: { r: 255, g: 255, b: 255 }
   }).png().toBuffer();
 }
+async function autoCrop(enhanced) {
+  try {
+    const trimmed = sharp(enhanced).trim({ threshold: 30 });
+    const trimmedBuf = await trimmed.toBuffer({ resolveWithObject: true });
+    const { width, height } = trimmedBuf.info;
+    if (width > 2 && height > 2) {
+      return trimmedBuf.data;
+    }
+  } catch {
+  }
+  return enhanced;
+}
 function imageToBase64(imagePath) {
   const buffer = fs.readFileSync(imagePath);
   return buffer.toString("base64");
 }
 // src/solver.ts
-var PROMPT = `You are an assistant helping a visually impaired person read distorted text from an image.
-The text contains uppercase letters A-Z and/or digits 0-9.
-A thin vertical stroke is the digit 1. Never read it as the letter I or L.
-A round closed shape is the letter O, not the letter D.
-Output ONLY the exact characters you read, nothing else.`;
+var PROMPT = `You are an expert OCR assistant reading distorted text from a CAPTCHA image.
+Two versions of the same captcha are provided. Cross-reference both to determine the correct text.
+The text may contain uppercase letters (A-Z), lowercase letters (a-z), and/or digits (0-9).
+Pay close attention to:
+- Letter case: lowercase "e" has a horizontal bar inside, digit "0" does not. Lowercase "r" has a short descender, uppercase "T" has a flat top.
+- Similar shapes: "5" has a flat top + curved bottom, "S" is fully curved. "4" has an angled stroke, "A" has a pointed top. "6" has a closed bottom loop, "8" has two loops. "2" has a curved top + flat bottom, "Z" has all straight lines.
+Output ONLY the exact characters you read, preserving case. Nothing else.`;
 var DEFAULT_MODELS = {
   openai: "gpt-4o",
   anthropic: "claude-sonnet-4-20250514",
@@ -71,7 +122,7 @@ async function resolveModel(apiKey, provider, modelId) {
       );
   }
 }
-var CONFUSION_GROUPS = {
+var LEGACY_CONFUSION_GROUPS = {
   "1": "1",
   I: "1",
   L: "1",
@@ -83,7 +134,7 @@ var CONFUSION_GROUPS = {
   Z: "Z",
   "2": "Z"
 };
-function majorityVote(attempts, expectedLength) {
+function majorityVote(attempts, expectedLength, groups) {
   let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;
   if (filtered.length === 0) {
     filtered = attempts;
@@ -103,6 +154,7 @@ function majorityVote(attempts, expectedLength) {
   }
   const sameLenAttempts = filtered.filter((a) => a.length === bestLen);
   if (sameLenAttempts.length === 0) return filtered[0];
+  const useGroups = groups && typeof groups === "object" ? groups : void 0;
   const result = [];
   for (let pos = 0; pos < bestLen; pos++) {
     const charCounts = /* @__PURE__ */ new Map();
@@ -110,20 +162,32 @@ function majorityVote(attempts, expectedLength) {
       const ch = a[pos];
       charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);
     }
-    const groupCounts = /* @__PURE__ */ new Map();
-    for (const [ch, count] of charCounts) {
-      const canonical = CONFUSION_GROUPS[ch] ?? ch;
-      groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);
-    }
-    let bestGroup = "";
-    let bestGroupCount = 0;
-    for (const [canonical, count] of groupCounts) {
-      if (count > bestGroupCount) {
-        bestGroup = canonical;
-        bestGroupCount = count;
+    if (useGroups) {
+      const groupCounts = /* @__PURE__ */ new Map();
+      for (const [ch, count] of charCounts) {
+        const canonical = useGroups[ch] ?? ch;
+        groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);
+      }
+      let bestGroup = "";
+      let bestGroupCount = 0;
+      for (const [canonical, count] of groupCounts) {
+        if (count > bestGroupCount) {
+          bestGroup = canonical;
+          bestGroupCount = count;
+        }
+      }
+      result.push(bestGroup);
+    } else {
+      let bestChar = "";
+      let bestCharCount = 0;
+      for (const [ch, count] of charCounts) {
+        if (count > bestCharCount) {
+          bestChar = ch;
+          bestCharCount = count;
+        }
       }
+      result.push(bestChar);
     }
-    result.push(bestGroup);
   }
   return result.join("");
 }
@@ -219,11 +283,32 @@ var Solver = class {
    * @returns Solved text, per-attempt answers, and token usage
    */
   async solve(input, options = {}) {
-    const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
+    const {
+      numAttempts = 7,
+      expectedLength,
+      maxRetries = 2,
+      verbose = true,
+      confusionGroups = false,
+      preprocess
+    } = options;
     const model = await this.getModel();
-    const imageBuffer = await preprocessCaptchaToBuffer(input);
+    const [enhancedBuffer, colorBuffer] = await Promise.all([
+      preprocessCaptchaToBuffer(input, preprocess),
+      preprocessCaptchaToBuffer(input, {
+        blur: 0,
+        scale: 4,
+        contrast: 1,
+        sharpen: false,
+        crop: "none",
+        padding: 40,
+        greyscale: false
+      })
+    ]);
     const results = await Promise.all(
-      Array.from({ length: numAttempts }, () => this.singleAttempt(model, imageBuffer, maxRetries))
+      Array.from(
+        { length: numAttempts },
+        () => this.singleAttempt(model, enhancedBuffer, colorBuffer, maxRetries)
+      )
     );
     const valid = results.filter((r) => r !== null);
     if (verbose) {
@@ -236,13 +321,18 @@ var Solver = class {
       if (verbose) console.log("  All attempts failed!");
       return { text: "", attempts, usage, attemptUsages };
     }
-    return { text: majorityVote(attempts, expectedLength), attempts, usage, attemptUsages };
+    return {
+      text: majorityVote(attempts, expectedLength, confusionGroups),
+      attempts,
+      usage,
+      attemptUsages
+    };
   }
   /**
    * Make a single API call to read the captcha.
    * Retries up to `maxRetries` times on failure.
    */
-  async singleAttempt(model, imageBuffer, maxRetries) {
+  async singleAttempt(model, primaryBuffer, secondaryBuffer, maxRetries) {
     for (let retry = 0; retry <= maxRetries; retry++) {
       try {
         const { text, usage } = await generateText({
@@ -252,7 +342,8 @@ var Solver = class {
               role: "user",
               content: [
                 { type: "text", text: PROMPT },
-                { type: "image", image: imageBuffer }
+                { type: "image", image: primaryBuffer },
+                { type: "image", image: secondaryBuffer }
               ]
             }
           ],
@@ -264,7 +355,7 @@ var Solver = class {
         if (lower.includes("sorry") || lower.includes("can't help") || lower.includes("cannot help") || lower.includes("unable to") || lower.includes("i can't") || raw.length > 20) {
           return null;
         }
-        const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, "");
+        const cleaned = raw.replace(/[^A-Za-z0-9]/g, "");
         return cleaned ? { text: cleaned, usage } : null;
       } catch (_err) {
         if (retry < maxRetries) {
@@ -278,8 +369,10 @@ var Solver = class {
   }
 };
 export {
+  LEGACY_CONFUSION_GROUPS,
   Solver,
   imageToBase64,
+  majorityVote,
   preprocessCaptcha,
   preprocessCaptchaToBuffer
 };

package/dist/index.js.map CHANGED Viewed

	@@ -1 +1 @@
1	- {"version":3,"sources":["../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["import type { LanguageModel, LanguageModelUsage } from 'ai';\nimport { generateText } from 'ai';\nimport { preprocessCaptchaToBuffer } from './preprocess.js';\n\nconst PROMPT = `You are an assistant helping a visually impaired person read distorted text from an image.\nThe text contains uppercase letters A-Z and/or digits 0-9.\nA thin vertical stroke is the digit 1. Never read it as the letter I or L.\nA round closed shape is the letter O, not the letter D.\nOutput ONLY the exact characters you read, nothing else.`;\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport type Provider = 'openai' \| 'anthropic' \| 'google';\n\nexport interface SolverOptions {\n /** AI provider to use when constructing the model from an API key (default: \"openai\") /\n provider?: Provider;\n /* Model ID passed to the provider (default: \"gpt-4o\") /\n model?: string;\n}\n\nexport interface SolveOptions {\n /* Number of voting attempts (default: 5) /\n numAttempts?: number;\n /* Expected captcha length — results of other lengths are discarded /\n expectedLength?: number;\n /* Max retries per attempt on API failure (default: 2) /\n maxRetries?: number;\n /* Whether to log attempt details (default: true) /\n verbose?: boolean;\n}\n\nexport interface SolveResult {\n /* The solved captcha text (majority-voted) /\n text: string;\n /* Per-attempt raw answers (before voting) /\n attempts: string[];\n /* Aggregated token usage across all parallel attempts /\n usage: LanguageModelUsage;\n /* Per-attempt usage breakdown /\n attemptUsages: LanguageModelUsage[];\n}\n\ninterface AttemptResult {\n text: string;\n usage: LanguageModelUsage;\n}\n\n// ── Provider resolution ──────────────────────────────────────────────\n\nconst DEFAULT_MODELS: Record<Provider, string> = {\n openai: 'gpt-4o',\n anthropic: 'claude-sonnet-4-20250514',\n google: 'gemini-2.0-flash',\n};\n\nasync function resolveModel(\n apiKey: string,\n provider: Provider,\n modelId: string\n): Promise<LanguageModel> {\n switch (provider) {\n case 'openai': {\n const { createOpenAI } = await import('@ai-sdk/openai');\n return createOpenAI({ apiKey })(modelId);\n }\n case 'anthropic': {\n // @ts-expect-error — optional peer dependency\n const { createAnthropic } = await import('@ai-sdk/anthropic');\n return createAnthropic({ apiKey })(modelId);\n }\n case 'google': {\n // @ts-expect-error — optional peer dependency\n const { createGoogleGenerativeAI } = await import('@ai-sdk/google');\n return createGoogleGenerativeAI({ apiKey })(modelId);\n }\n default:\n throw new Error(\n `Unknown provider \"${provider}\". Install the matching @ai-sdk/ package and pass the model directly.`\n );\n }\n}\n\n// ── Confusion groups ─────────────────────────────────────────────────\n\n/*\n Characters the model commonly misreads as each other.\n * Each group maps to its canonical (most likely correct) character.\n /\nconst CONFUSION_GROUPS: Record<string, string> = {\n '1': '1',\n I: '1',\n L: '1',\n O: 'O',\n D: 'O',\n '0': 'O',\n S: 'S',\n '5': 'S',\n Z: 'Z',\n '2': 'Z',\n};\n\n// ── Majority voting ──────────────────────────────────────────────────\n\n/\n Character-level majority vote across multiple attempts.\n * Uses confusion-aware voting: characters that the model commonly\n * confuses (e.g. 1/I/L, O/D/0) are grouped together during counting.\n /\nfunction majorityVote(attempts: string[], expectedLength?: number): string {\n let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n if (filtered.length === 0) {\n filtered = attempts;\n }\n if (filtered.length === 0) return '';\n\n // Find most common length\n const lenCounts = new Map<number, number>();\n for (const a of filtered) {\n lenCounts.set(a.length, (lenCounts.get(a.length) ?? 0) + 1);\n }\n let bestLen = 0;\n let bestCount = 0;\n for (const [len, count] of lenCounts) {\n if (count > bestCount) {\n bestLen = len;\n bestCount = count;\n }\n }\n\n const sameLenAttempts = filtered.filter((a) => a.length === bestLen);\n if (sameLenAttempts.length === 0) return filtered[0];\n\n // Vote per character position with confusion-aware grouping\n const result: string[] = [];\n for (let pos = 0; pos < bestLen; pos++) {\n const charCounts = new Map<string, number>();\n for (const a of sameLenAttempts) {\n const ch = a[pos];\n charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);\n }\n\n const groupCounts = new Map<string, number>();\n for (const [ch, count] of charCounts) {\n const canonical = CONFUSION_GROUPS[ch] ?? ch;\n groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);\n }\n\n let bestGroup = '';\n let bestGroupCount = 0;\n for (const [canonical, count] of groupCounts) {\n if (count > bestGroupCount) {\n bestGroup = canonical;\n bestGroupCount = count;\n }\n }\n\n result.push(bestGroup);\n }\n\n return result.join('');\n}\n\n// ── Usage aggregation ────────────────────────────────────────────────\n\nfunction sumOptional(a: number \| undefined, b: number \| undefined): number \| undefined {\n if (a === undefined && b === undefined) return undefined;\n return (a ?? 0) + (b ?? 0);\n}\n\nfunction aggregateUsage(usages: LanguageModelUsage[]): LanguageModelUsage {\n const zero: LanguageModelUsage = {\n inputTokens: undefined,\n inputTokenDetails: {\n noCacheTokens: undefined,\n cacheReadTokens: undefined,\n cacheWriteTokens: undefined,\n },\n outputTokens: undefined,\n outputTokenDetails: {\n textTokens: undefined,\n reasoningTokens: undefined,\n },\n totalTokens: undefined,\n };\n return usages.reduce<LanguageModelUsage>(\n (acc, u) => ({\n inputTokens: sumOptional(acc.inputTokens, u.inputTokens),\n inputTokenDetails: {\n noCacheTokens: sumOptional(\n acc.inputTokenDetails.noCacheTokens,\n u.inputTokenDetails.noCacheTokens\n ),\n cacheReadTokens: sumOptional(\n acc.inputTokenDetails.cacheReadTokens,\n u.inputTokenDetails.cacheReadTokens\n ),\n cacheWriteTokens: sumOptional(\n acc.inputTokenDetails.cacheWriteTokens,\n u.inputTokenDetails.cacheWriteTokens\n ),\n },\n outputTokens: sumOptional(acc.outputTokens, u.outputTokens),\n outputTokenDetails: {\n textTokens: sumOptional(acc.outputTokenDetails.textTokens, u.outputTokenDetails.textTokens),\n reasoningTokens: sumOptional(\n acc.outputTokenDetails.reasoningTokens,\n u.outputTokenDetails.reasoningTokens\n ),\n },\n totalTokens: sumOptional(acc.totalTokens, u.totalTokens),\n }),\n zero\n );\n}\n\n// ── Solver class ─────────────────────────────────────────────────────\n\nexport class Solver {\n private _model: LanguageModel \| null = null;\n private _pendingModel: Promise<LanguageModel> \| null = null;\n\n /\n Create a captcha solver.\n \n @example\n * // Simple — defaults to OpenAI gpt-4o\n * const solver = new Solver('sk-...');\n \n @example\n * // Specify provider and model\n * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });\n \n @example\n * // Pass an AI SDK model directly\n * import { createOpenAI } from '@ai-sdk/openai';\n * const openai = createOpenAI({ apiKey: 'sk-...' });\n * const solver = new Solver(openai('gpt-4o'));\n /\n constructor(keyOrModel: string \| LanguageModel, options?: SolverOptions) {\n if (typeof keyOrModel === 'string') {\n const provider = options?.provider ?? 'openai';\n const modelId = options?.model ?? DEFAULT_MODELS[provider];\n // Lazily resolve the model on first use\n this._pendingModel = resolveModel(keyOrModel, provider, modelId);\n } else {\n this._model = keyOrModel;\n }\n }\n\n private async getModel(): Promise<LanguageModel> {\n if (this._model) return this._model;\n this._model = await this._pendingModel!;\n this._pendingModel = null;\n return this._model;\n }\n\n /\n Solve a captcha image.\n \n @param input - File path (string) or raw image Buffer\n * @param options - Solve options (attempts, expected length, etc.)\n * @returns Solved text, per-attempt answers, and token usage\n /\n async solve(input: string \| Buffer, options: SolveOptions = {}): Promise<SolveResult> {\n const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;\n\n const model = await this.getModel();\n const imageBuffer = await preprocessCaptchaToBuffer(input);\n\n // Fire all attempts in parallel for speed\n const results = await Promise.all(\n Array.from({ length: numAttempts }, () => this.singleAttempt(model, imageBuffer, maxRetries))\n );\n const valid = results.filter((r): r is AttemptResult => r !== null);\n if (verbose) {\n valid.forEach((r, i) => console.log(` Attempt ${i + 1}: ${r.text}`));\n }\n\n const attempts = valid.map((r) => r.text);\n const attemptUsages = valid.map((r) => r.usage);\n const usage = aggregateUsage(attemptUsages);\n\n if (attempts.length === 0) {\n if (verbose) console.log(' All attempts failed!');\n return { text: '', attempts, usage, attemptUsages };\n }\n\n return { text: majorityVote(attempts, expectedLength), attempts, usage, attemptUsages };\n }\n\n /\n Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n /\n private async singleAttempt(\n model: LanguageModel,\n imageBuffer: Buffer,\n maxRetries: number\n ): Promise<AttemptResult \| null> {\n for (let retry = 0; retry <= maxRetries; retry++) {\n try {\n const { text, usage } = await generateText({\n model,\n messages: [\n {\n role: 'user',\n content: [\n { type: 'text', text: PROMPT },\n { type: 'image', image: imageBuffer },\n ],\n },\n ],\n temperature: 1,\n maxOutputTokens: 256,\n });\n\n const raw = text.trim();\n\n // Detect refusals\n const lower = raw.toLowerCase();\n if (\n lower.includes('sorry') \|\|\n lower.includes(\"can't help\") \|\|\n lower.includes('cannot help') \|\|\n lower.includes('unable to') \|\|\n lower.includes(\"i can't\") \|\|\n raw.length > 20\n ) {\n return null;\n }\n\n // Clean: keep only uppercase letters and digits\n const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, '');\n return cleaned ? { text: cleaned, usage } : null;\n } catch (_err) {\n if (retry < maxRetries) {\n await new Promise((r) => setTimeout(r, 1000 (retry + 1)));\n continue;\n }\n return null;\n }\n }\n return null;\n }\n}\n","import fs from 'fs';\nimport path from 'path';\nimport sharp from 'sharp';\n\n/*\n Preprocess a captcha image using sharp (libvips).\n \n Pipeline:\n * 1. Gaussian blur in color space (smooths dither pattern)\n * 2. Grayscale conversion\n * 3. Upscale 4× with Lanczos\n * 4. Contrast boost (3× around image mean) + sharpen\n * 5. Crop decorative borders\n * 6. Add white padding\n \n Accepts a file path or a raw image Buffer.\n * Returns a base64-encoded PNG string.\n /\nexport async function preprocessCaptcha(input: string \| Buffer): Promise<string> {\n const buf = await preprocessCaptchaToBuffer(input);\n return buf.toString('base64');\n}\n\n/\n Same preprocessing pipeline as `preprocessCaptcha`, but returns the\n * resulting PNG as a raw Buffer (useful for AI SDK image content parts).\n /\nexport async function preprocessCaptchaToBuffer(input: string \| Buffer): Promise<Buffer> {\n const source = typeof input === 'string' ? path.resolve(input) : input;\n\n // Read original dimensions for crop/resize calculations\n const metadata = await sharp(source).metadata();\n const origW = metadata.width!;\n const origH = metadata.height!;\n\n // Step 1-2: Blur in color space (smooths dither pattern) → greyscale\n // Separate from resize to prevent pipeline reordering\n const smoothed = await sharp(source).blur(1.5).greyscale().toBuffer();\n\n // Step 3: Upscale 4× with Lanczos\n const upscaled = await sharp(smoothed)\n .resize(origW 4, origH * 4, { kernel: 'lanczos3' })\n .toBuffer();\n\n // Step 4: Contrast 3× around actual image mean + sharpen\n // Matches PIL's ImageEnhance.Contrast: output = factorinput + mean(1-factor)\n const stats = await sharp(upscaled).stats();\n const mean = stats.channels[0].mean;\n const enhanced = await sharp(upscaled)\n .linear(3.0, mean * (1 - 3.0))\n .sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 })\n .toBuffer();\n\n // Step 5: Crop decorative borders\n // Remove 10% left/right, 2% top, 40% bottom (keep top 60%)\n // Math.floor matches Python's int() truncation\n const scaledW = origW * 4;\n const scaledH = origH * 4;\n const cropLeft = Math.floor(scaledW * 0.1);\n const cropTop = Math.floor(scaledH * 0.02);\n const cropRight = Math.floor(scaledW * 0.9);\n const cropBottom = Math.floor(scaledH * 0.6);\n const cropW = cropRight - cropLeft;\n const cropH = cropBottom - cropTop;\n\n // Step 5-6: Crop → add white padding → output PNG\n return sharp(enhanced)\n .extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH })\n .extend({\n top: 20,\n bottom: 20,\n left: 30,\n right: 30,\n background: { r: 255, g: 255, b: 255 },\n })\n .png()\n .toBuffer();\n}\n\n/*\n Read an image file and return its base64-encoded content.\n */\nexport function imageToBase64(imagePath: string): string {\n const buffer = fs.readFileSync(imagePath);\n return buffer.toString('base64');\n}\n"],"mappings":";AACA,SAAS,oBAAoB;;;ACD7B,OAAO,QAAQ;AACf,OAAO,UAAU;AACjB,OAAO,WAAW;AAgBlB,eAAsB,kBAAkB,OAAyC;AAC/E,QAAM,MAAM,MAAM,0BAA0B,KAAK;AACjD,SAAO,IAAI,SAAS,QAAQ;AAC9B;AAMA,eAAsB,0BAA0B,OAAyC;AACvF,QAAM,SAAS,OAAO,UAAU,WAAW,KAAK,QAAQ,KAAK,IAAI;AAGjE,QAAM,WAAW,MAAM,MAAM,MAAM,EAAE,SAAS;AAC9C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAIvB,QAAM,WAAW,MAAM,MAAM,MAAM,EAAE,KAAK,GAAG,EAAE,UAAU,EAAE,SAAS;AAGpE,QAAM,WAAW,MAAM,MAAM,QAAQ,EAClC,OAAO,QAAQ,GAAG,QAAQ,GAAG,EAAE,QAAQ,WAAW,CAAC,EACnD,SAAS;AAIZ,QAAM,QAAQ,MAAM,MAAM,QAAQ,EAAE,MAAM;AAC1C,QAAM,OAAO,MAAM,SAAS,CAAC,EAAE;AAC/B,QAAM,WAAW,MAAM,MAAM,QAAQ,EAClC,OAAO,GAAK,QAAQ,IAAI,EAAI,EAC5B,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC,EACxC,SAAS;AAKZ,QAAM,UAAU,QAAQ;AACxB,QAAM,UAAU,QAAQ;AACxB,QAAM,WAAW,KAAK,MAAM,UAAU,GAAG;AACzC,QAAM,UAAU,KAAK,MAAM,UAAU,IAAI;AACzC,QAAM,YAAY,KAAK,MAAM,UAAU,GAAG;AAC1C,QAAM,aAAa,KAAK,MAAM,UAAU,GAAG;AAC3C,QAAM,QAAQ,YAAY;AAC1B,QAAM,QAAQ,aAAa;AAG3B,SAAO,MAAM,QAAQ,EAClB,QAAQ,EAAE,MAAM,UAAU,KAAK,SAAS,OAAO,OAAO,QAAQ,MAAM,CAAC,EACrE,OAAO;AAAA,IACN,KAAK;AAAA,IACL,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,OAAO;AAAA,IACP,YAAY,EAAE,GAAG,KAAK,GAAG,KAAK,GAAG,IAAI;AAAA,EACvC,CAAC,EACA,IAAI,EACJ,SAAS;AACd;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,GAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;ADjFA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AA8Cf,IAAM,iBAA2C;AAAA,EAC/C,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,QAAQ;AACV;AAEA,eAAe,aACb,QACA,UACA,SACwB;AACxB,UAAQ,UAAU;AAAA,IAChB,KAAK,UAAU;AACb,YAAM,EAAE,aAAa,IAAI,MAAM,OAAO,gBAAgB;AACtD,aAAO,aAAa,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACzC;AAAA,IACA,KAAK,aAAa;AAEhB,YAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,mBAAmB;AAC5D,aAAO,gBAAgB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IAC5C;AAAA,IACA,KAAK,UAAU;AAEb,YAAM,EAAE,yBAAyB,IAAI,MAAM,OAAO,gBAAgB;AAClE,aAAO,yBAAyB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACrD;AAAA,IACA;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,QAAQ;AAAA,MAC/B;AAAA,EACJ;AACF;AAQA,IAAM,mBAA2C;AAAA,EAC/C,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AACP;AASA,SAAS,aAAa,UAAoB,gBAAiC;AACzE,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAEtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAGlC,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,KAAK,UAAU;AACxB,cAAU,IAAI,EAAE,SAAS,UAAU,IAAI,EAAE,MAAM,KAAK,KAAK,CAAC;AAAA,EAC5D;AACA,MAAI,UAAU;AACd,MAAI,YAAY;AAChB,aAAW,CAAC,KAAK,KAAK,KAAK,WAAW;AACpC,QAAI,QAAQ,WAAW;AACrB,gBAAU;AACV,kBAAY;AAAA,IACd;AAAA,EACF;AAEA,QAAM,kBAAkB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnE,MAAI,gBAAgB,WAAW,EAAG,QAAO,SAAS,CAAC;AAGnD,QAAM,SAAmB,CAAC;AAC1B,WAAS,MAAM,GAAG,MAAM,SAAS,OAAO;AACtC,UAAM,aAAa,oBAAI,IAAoB;AAC3C,eAAW,KAAK,iBAAiB;AAC/B,YAAM,KAAK,EAAE,GAAG;AAChB,iBAAW,IAAI,KAAK,WAAW,IAAI,EAAE,KAAK,KAAK,CAAC;AAAA,IAClD;AAEA,UAAM,cAAc,oBAAI,IAAoB;AAC5C,eAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,YAAM,YAAY,iBAAiB,EAAE,KAAK;AAC1C,kBAAY,IAAI,YAAY,YAAY,IAAI,SAAS,KAAK,KAAK,KAAK;AAAA,IACtE;AAEA,QAAI,YAAY;AAChB,QAAI,iBAAiB;AACrB,eAAW,CAAC,WAAW,KAAK,KAAK,aAAa;AAC5C,UAAI,QAAQ,gBAAgB;AAC1B,oBAAY;AACZ,yBAAiB;AAAA,MACnB;AAAA,IACF;AAEA,WAAO,KAAK,SAAS;AAAA,EACvB;AAEA,SAAO,OAAO,KAAK,EAAE;AACvB;AAIA,SAAS,YAAY,GAAuB,GAA2C;AACrF,MAAI,MAAM,UAAa,MAAM,OAAW,QAAO;AAC/C,UAAQ,KAAK,MAAM,KAAK;AAC1B;AAEA,SAAS,eAAe,QAAkD;AACxE,QAAM,OAA2B;AAAA,IAC/B,aAAa;AAAA,IACb,mBAAmB;AAAA,MACjB,eAAe;AAAA,MACf,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,IACpB;AAAA,IACA,cAAc;AAAA,IACd,oBAAoB;AAAA,MAClB,YAAY;AAAA,MACZ,iBAAiB;AAAA,IACnB;AAAA,IACA,aAAa;AAAA,EACf;AACA,SAAO,OAAO;AAAA,IACZ,CAAC,KAAK,OAAO;AAAA,MACX,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,MACvD,mBAAmB;AAAA,QACjB,eAAe;AAAA,UACb,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,iBAAiB;AAAA,UACf,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,kBAAkB;AAAA,UAChB,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,MACF;AAAA,MACA,cAAc,YAAY,IAAI,cAAc,EAAE,YAAY;AAAA,MAC1D,oBAAoB;AAAA,QAClB,YAAY,YAAY,IAAI,mBAAmB,YAAY,EAAE,mBAAmB,UAAU;AAAA,QAC1F,iBAAiB;AAAA,UACf,IAAI,mBAAmB;AAAA,UACvB,EAAE,mBAAmB;AAAA,QACvB;AAAA,MACF;AAAA,MACA,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,IACzD;AAAA,IACA;AAAA,EACF;AACF;AAIO,IAAM,SAAN,MAAa;AAAA,EACV,SAA+B;AAAA,EAC/B,gBAA+C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAmBvD,YAAY,YAAoC,SAAyB;AACvE,QAAI,OAAO,eAAe,UAAU;AAClC,YAAM,WAAW,SAAS,YAAY;AACtC,YAAM,UAAU,SAAS,SAAS,eAAe,QAAQ;AAEzD,WAAK,gBAAgB,aAAa,YAAY,UAAU,OAAO;AAAA,IACjE,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAc,WAAmC;AAC/C,QAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,SAAK,SAAS,MAAM,KAAK;AACzB,SAAK,gBAAgB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,MAAM,OAAwB,UAAwB,CAAC,GAAyB;AACpF,UAAM,EAAE,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE5E,UAAM,QAAQ,MAAM,KAAK,SAAS;AAClC,UAAM,cAAc,MAAM,0BAA0B,KAAK;AAGzD,UAAM,UAAU,MAAM,QAAQ;AAAA,MAC5B,MAAM,KAAK,EAAE,QAAQ,YAAY,GAAG,MAAM,KAAK,cAAc,OAAO,aAAa,UAAU,CAAC;AAAA,IAC9F;AACA,UAAM,QAAQ,QAAQ,OAAO,CAAC,MAA0B,MAAM,IAAI;AAClE,QAAI,SAAS;AACX,YAAM,QAAQ,CAAC,GAAG,MAAM,QAAQ,IAAI,aAAa,IAAI,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC;AAAA,IACtE;AAEA,UAAM,WAAW,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI;AACxC,UAAM,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,eAAe,aAAa;AAE1C,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO,EAAE,MAAM,IAAI,UAAU,OAAO,cAAc;AAAA,IACpD;AAEA,WAAO,EAAE,MAAM,aAAa,UAAU,cAAc,GAAG,UAAU,OAAO,cAAc;AAAA,EACxF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,aACA,YAC+B;AAC/B,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,MAAM,MAAM,IAAI,MAAM,aAAa;AAAA,UACzC;AAAA,UACA,UAAU;AAAA,YACR;AAAA,cACE,MAAM;AAAA,cACN,SAAS;AAAA,gBACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,gBAC7B,EAAE,MAAM,SAAS,OAAO,YAAY;AAAA,cACtC;AAAA,YACF;AAAA,UACF;AAAA,UACA,aAAa;AAAA,UACb,iBAAiB;AAAA,QACnB,CAAC;AAED,cAAM,MAAM,KAAK,KAAK;AAGtB,cAAM,QAAQ,IAAI,YAAY;AAC9B,YACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,iBAAO;AAAA,QACT;AAGA,cAAM,UAAU,IAAI,YAAY,EAAE,QAAQ,cAAc,EAAE;AAC1D,eAAO,UAAU,EAAE,MAAM,SAAS,MAAM,IAAI;AAAA,MAC9C,SAAS,MAAM;AACb,YAAI,QAAQ,YAAY;AACtB,gBAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;","names":[]}
1	+ {"version":3,"sources":["../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["import type { LanguageModel, LanguageModelUsage } from 'ai';\nimport { generateText } from 'ai';\nimport type { PreprocessOptions } from './preprocess.js';\nimport { preprocessCaptchaToBuffer } from './preprocess.js';\n\nconst PROMPT = `You are an expert OCR assistant reading distorted text from a CAPTCHA image.\nTwo versions of the same captcha are provided. Cross-reference both to determine the correct text.\nThe text may contain uppercase letters (A-Z), lowercase letters (a-z), and/or digits (0-9).\nPay close attention to:\n- Letter case: lowercase \"e\" has a horizontal bar inside, digit \"0\" does not. Lowercase \"r\" has a short descender, uppercase \"T\" has a flat top.\n- Similar shapes: \"5\" has a flat top + curved bottom, \"S\" is fully curved. \"4\" has an angled stroke, \"A\" has a pointed top. \"6\" has a closed bottom loop, \"8\" has two loops. \"2\" has a curved top + flat bottom, \"Z\" has all straight lines.\nOutput ONLY the exact characters you read, preserving case. Nothing else.`;\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport type Provider = 'openai' \| 'anthropic' \| 'google';\n\nexport interface SolverOptions {\n /** AI provider to use when constructing the model from an API key (default: \"openai\") /\n provider?: Provider;\n /* Model ID passed to the provider (default: \"gpt-4o\") /\n model?: string;\n}\n\nexport interface SolveOptions {\n /* Number of voting attempts (default: 7) /\n numAttempts?: number;\n /* Expected captcha length — results of other lengths are discarded /\n expectedLength?: number;\n /* Max retries per attempt on API failure (default: 2) /\n maxRetries?: number;\n /* Whether to log attempt details (default: true) /\n verbose?: boolean;\n /\n Confusion groups for majority voting.\n * Pass a Record<string, string> to merge visually similar characters,\n * or `false` to disable (default: false).\n * Use LEGACY_CONFUSION_GROUPS to restore pre-3.0 behavior.\n /\n confusionGroups?: Record<string, string> \| false;\n /* Preprocessing options passed to the image pipeline /\n preprocess?: PreprocessOptions;\n}\n\nexport interface SolveResult {\n /* The solved captcha text (majority-voted) /\n text: string;\n /* Per-attempt raw answers (before voting) /\n attempts: string[];\n /* Aggregated token usage across all parallel attempts /\n usage: LanguageModelUsage;\n /* Per-attempt usage breakdown /\n attemptUsages: LanguageModelUsage[];\n}\n\ninterface AttemptResult {\n text: string;\n usage: LanguageModelUsage;\n}\n\n// ── Provider resolution ──────────────────────────────────────────────\n\nconst DEFAULT_MODELS: Record<Provider, string> = {\n openai: 'gpt-4o',\n anthropic: 'claude-sonnet-4-20250514',\n google: 'gemini-2.0-flash',\n};\n\nasync function resolveModel(\n apiKey: string,\n provider: Provider,\n modelId: string\n): Promise<LanguageModel> {\n switch (provider) {\n case 'openai': {\n const { createOpenAI } = await import('@ai-sdk/openai');\n return createOpenAI({ apiKey })(modelId);\n }\n case 'anthropic': {\n // @ts-expect-error — optional peer dependency\n const { createAnthropic } = await import('@ai-sdk/anthropic');\n return createAnthropic({ apiKey })(modelId);\n }\n case 'google': {\n // @ts-expect-error — optional peer dependency\n const { createGoogleGenerativeAI } = await import('@ai-sdk/google');\n return createGoogleGenerativeAI({ apiKey })(modelId);\n }\n default:\n throw new Error(\n `Unknown provider \"${provider}\". Install the matching @ai-sdk/ package and pass the model directly.`\n );\n }\n}\n\n// ── Confusion groups ─────────────────────────────────────────────────\n\n/*\n Pre-3.0 confusion groups that merge visually similar characters.\n * Opt-in via `{ confusionGroups: LEGACY_CONFUSION_GROUPS }`.\n \n Maps: 1/I/L → '1', O/D/0 → 'O', S/5 → 'S', Z/2 → 'Z'\n /\nexport const LEGACY_CONFUSION_GROUPS: Record<string, string> = {\n '1': '1',\n I: '1',\n L: '1',\n O: 'O',\n D: 'O',\n '0': 'O',\n S: 'S',\n '5': 'S',\n Z: 'Z',\n '2': 'Z',\n};\n\n// ── Majority voting ──────────────────────────────────────────────────\n\n/\n Character-level majority vote across multiple attempts.\n * When `groups` is provided, visually similar characters are merged\n * during counting (e.g. 1/I/L all count toward '1').\n /\nexport function majorityVote(\n attempts: string[],\n expectedLength?: number,\n groups?: Record<string, string> \| false\n): string {\n let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n if (filtered.length === 0) {\n filtered = attempts;\n }\n if (filtered.length === 0) return '';\n\n // Find most common length\n const lenCounts = new Map<number, number>();\n for (const a of filtered) {\n lenCounts.set(a.length, (lenCounts.get(a.length) ?? 0) + 1);\n }\n let bestLen = 0;\n let bestCount = 0;\n for (const [len, count] of lenCounts) {\n if (count > bestCount) {\n bestLen = len;\n bestCount = count;\n }\n }\n\n const sameLenAttempts = filtered.filter((a) => a.length === bestLen);\n if (sameLenAttempts.length === 0) return filtered[0];\n\n const useGroups = groups && typeof groups === 'object' ? groups : undefined;\n\n // Vote per character position\n const result: string[] = [];\n for (let pos = 0; pos < bestLen; pos++) {\n const charCounts = new Map<string, number>();\n for (const a of sameLenAttempts) {\n const ch = a[pos];\n charCounts.set(ch, (charCounts.get(ch) ?? 0) + 1);\n }\n\n if (useGroups) {\n // Confusion-aware voting\n const groupCounts = new Map<string, number>();\n for (const [ch, count] of charCounts) {\n const canonical = useGroups[ch] ?? ch;\n groupCounts.set(canonical, (groupCounts.get(canonical) ?? 0) + count);\n }\n\n let bestGroup = '';\n let bestGroupCount = 0;\n for (const [canonical, count] of groupCounts) {\n if (count > bestGroupCount) {\n bestGroup = canonical;\n bestGroupCount = count;\n }\n }\n result.push(bestGroup);\n } else {\n // Simple majority — pick the most frequent raw character\n let bestChar = '';\n let bestCharCount = 0;\n for (const [ch, count] of charCounts) {\n if (count > bestCharCount) {\n bestChar = ch;\n bestCharCount = count;\n }\n }\n result.push(bestChar);\n }\n }\n\n return result.join('');\n}\n\n// ── Usage aggregation ────────────────────────────────────────────────\n\nfunction sumOptional(a: number \| undefined, b: number \| undefined): number \| undefined {\n if (a === undefined && b === undefined) return undefined;\n return (a ?? 0) + (b ?? 0);\n}\n\nfunction aggregateUsage(usages: LanguageModelUsage[]): LanguageModelUsage {\n const zero: LanguageModelUsage = {\n inputTokens: undefined,\n inputTokenDetails: {\n noCacheTokens: undefined,\n cacheReadTokens: undefined,\n cacheWriteTokens: undefined,\n },\n outputTokens: undefined,\n outputTokenDetails: {\n textTokens: undefined,\n reasoningTokens: undefined,\n },\n totalTokens: undefined,\n };\n return usages.reduce<LanguageModelUsage>(\n (acc, u) => ({\n inputTokens: sumOptional(acc.inputTokens, u.inputTokens),\n inputTokenDetails: {\n noCacheTokens: sumOptional(\n acc.inputTokenDetails.noCacheTokens,\n u.inputTokenDetails.noCacheTokens\n ),\n cacheReadTokens: sumOptional(\n acc.inputTokenDetails.cacheReadTokens,\n u.inputTokenDetails.cacheReadTokens\n ),\n cacheWriteTokens: sumOptional(\n acc.inputTokenDetails.cacheWriteTokens,\n u.inputTokenDetails.cacheWriteTokens\n ),\n },\n outputTokens: sumOptional(acc.outputTokens, u.outputTokens),\n outputTokenDetails: {\n textTokens: sumOptional(acc.outputTokenDetails.textTokens, u.outputTokenDetails.textTokens),\n reasoningTokens: sumOptional(\n acc.outputTokenDetails.reasoningTokens,\n u.outputTokenDetails.reasoningTokens\n ),\n },\n totalTokens: sumOptional(acc.totalTokens, u.totalTokens),\n }),\n zero\n );\n}\n\n// ── Solver class ─────────────────────────────────────────────────────\n\nexport class Solver {\n private _model: LanguageModel \| null = null;\n private _pendingModel: Promise<LanguageModel> \| null = null;\n\n /\n Create a captcha solver.\n \n @example\n * // Simple — defaults to OpenAI gpt-4o\n * const solver = new Solver('sk-...');\n \n @example\n * // Specify provider and model\n * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });\n \n @example\n * // Pass an AI SDK model directly\n * import { createOpenAI } from '@ai-sdk/openai';\n * const openai = createOpenAI({ apiKey: 'sk-...' });\n * const solver = new Solver(openai('gpt-4o'));\n /\n constructor(keyOrModel: string \| LanguageModel, options?: SolverOptions) {\n if (typeof keyOrModel === 'string') {\n const provider = options?.provider ?? 'openai';\n const modelId = options?.model ?? DEFAULT_MODELS[provider];\n // Lazily resolve the model on first use\n this._pendingModel = resolveModel(keyOrModel, provider, modelId);\n } else {\n this._model = keyOrModel;\n }\n }\n\n private async getModel(): Promise<LanguageModel> {\n if (this._model) return this._model;\n this._model = await this._pendingModel!;\n this._pendingModel = null;\n return this._model;\n }\n\n /\n Solve a captcha image.\n \n @param input - File path (string) or raw image Buffer\n * @param options - Solve options (attempts, expected length, etc.)\n * @returns Solved text, per-attempt answers, and token usage\n /\n async solve(input: string \| Buffer, options: SolveOptions = {}): Promise<SolveResult> {\n const {\n numAttempts = 7,\n expectedLength,\n maxRetries = 2,\n verbose = true,\n confusionGroups = false,\n preprocess,\n } = options;\n\n const model = await this.getModel();\n\n // Two complementary views:\n // 1. Enhanced grayscale (high contrast + auto-crop) — great for clear text\n // 2. Color original (upscaled, no greyscale, no contrast) — preserves subtle features\n const [enhancedBuffer, colorBuffer] = await Promise.all([\n preprocessCaptchaToBuffer(input, preprocess),\n preprocessCaptchaToBuffer(input, {\n blur: 0,\n scale: 4,\n contrast: 1.0,\n sharpen: false,\n crop: 'none',\n padding: 40,\n greyscale: false,\n }),\n ]);\n\n // Fire all attempts in parallel for speed\n const results = await Promise.all(\n Array.from({ length: numAttempts }, () =>\n this.singleAttempt(model, enhancedBuffer, colorBuffer, maxRetries)\n )\n );\n const valid = results.filter((r): r is AttemptResult => r !== null);\n if (verbose) {\n valid.forEach((r, i) => console.log(` Attempt ${i + 1}: ${r.text}`));\n }\n\n const attempts = valid.map((r) => r.text);\n const attemptUsages = valid.map((r) => r.usage);\n const usage = aggregateUsage(attemptUsages);\n\n if (attempts.length === 0) {\n if (verbose) console.log(' All attempts failed!');\n return { text: '', attempts, usage, attemptUsages };\n }\n\n return {\n text: majorityVote(attempts, expectedLength, confusionGroups),\n attempts,\n usage,\n attemptUsages,\n };\n }\n\n /\n Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n /\n private async singleAttempt(\n model: LanguageModel,\n primaryBuffer: Buffer,\n secondaryBuffer: Buffer,\n maxRetries: number\n ): Promise<AttemptResult \| null> {\n for (let retry = 0; retry <= maxRetries; retry++) {\n try {\n const { text, usage } = await generateText({\n model,\n messages: [\n {\n role: 'user',\n content: [\n { type: 'text', text: PROMPT },\n { type: 'image', image: primaryBuffer },\n { type: 'image', image: secondaryBuffer },\n ],\n },\n ],\n temperature: 1,\n maxOutputTokens: 256,\n });\n\n const raw = text.trim();\n\n // Detect refusals\n const lower = raw.toLowerCase();\n if (\n lower.includes('sorry') \|\|\n lower.includes(\"can't help\") \|\|\n lower.includes('cannot help') \|\|\n lower.includes('unable to') \|\|\n lower.includes(\"i can't\") \|\|\n raw.length > 20\n ) {\n return null;\n }\n\n // Clean: keep only uppercase letters and digits\n const cleaned = raw.replace(/[^A-Za-z0-9]/g, '');\n return cleaned ? { text: cleaned, usage } : null;\n } catch (_err) {\n if (retry < maxRetries) {\n await new Promise((r) => setTimeout(r, 1000 (retry + 1)));\n continue;\n }\n return null;\n }\n }\n return null;\n }\n}\n","import fs from 'fs';\nimport path from 'path';\nimport sharp from 'sharp';\n\n// ── Types ────────────────────────────────────────────────────────────\n\nexport interface CropFractions {\n /** Fraction from left edge (0–1, default: 0.1) /\n left: number;\n /* Fraction from top edge (0–1, default: 0.02) /\n top: number;\n /* Fraction from left to keep (0–1, default: 0.9) /\n right: number;\n /* Fraction from top to keep (0–1, default: 0.6) /\n bottom: number;\n}\n\nexport interface PreprocessOptions {\n /* Gaussian blur radius (default: 1.5). Set to 0 to skip. /\n blur?: number;\n /* Upscale factor (default: 4) /\n scale?: number;\n /* Contrast multiplier around image mean (default: 3.0). Set to 1 to skip. /\n contrast?: number;\n /* Enable unsharp-mask sharpening (default: true) /\n sharpen?: boolean;\n /\n Crop mode (default: 'auto'):\n * - 'auto' – trim whitespace after contrast enhancement, with margin\n * - 'legacy' – fixed-percentage crop (original behavior)\n * - 'none' – skip cropping\n * - CropFractions – custom crop percentages\n /\n crop?: 'auto' \| 'legacy' \| 'none' \| CropFractions;\n /* Add white padding around the result (default: true). Pass false to skip, or a number for custom px. /\n padding?: boolean \| number;\n /* Invert colors (negate) after processing (default: false) /\n negate?: boolean;\n /* Convert to greyscale (default: true) /\n greyscale?: boolean;\n}\n\nconst LEGACY_CROP: CropFractions = { left: 0.1, top: 0.02, right: 0.9, bottom: 0.6 };\n\n// ── Public API ───────────────────────────────────────────────────────\n\n/\n Preprocess a captcha image and return a base64-encoded PNG string.\n /\nexport async function preprocessCaptcha(\n input: string \| Buffer,\n options?: PreprocessOptions\n): Promise<string> {\n const buf = await preprocessCaptchaToBuffer(input, options);\n return buf.toString('base64');\n}\n\n/\n Preprocess a captcha image and return the resulting PNG as a raw Buffer.\n \n Pipeline:\n * 1. Gaussian blur in color space (smooths dither pattern)\n * 2. Grayscale conversion\n * 3. Upscale with Lanczos\n * 4. Contrast boost around image mean + sharpen\n * 5. Crop (auto-detect, legacy fixed, none, or custom)\n * 6. Add white padding\n /\nexport async function preprocessCaptchaToBuffer(\n input: string \| Buffer,\n options?: PreprocessOptions\n): Promise<Buffer> {\n const {\n blur = 1.5,\n scale = 4,\n contrast = 3.0,\n sharpen = true,\n crop = 'auto',\n padding = true,\n negate = false,\n greyscale = true,\n } = options ?? {};\n\n const source = typeof input === 'string' ? path.resolve(input) : input;\n\n // Read original dimensions\n const metadata = await sharp(source).metadata();\n const origW = metadata.width!;\n const origH = metadata.height!;\n\n // Step 1-2: Blur (optional) + greyscale (optional)\n let pipeline = sharp(source);\n if (blur > 0) pipeline = pipeline.blur(blur);\n if (greyscale) pipeline = pipeline.greyscale();\n const smoothed = await pipeline.toBuffer();\n\n // Step 3: Upscale with Lanczos\n const upscaled = await sharp(smoothed)\n .resize(origW scale, origH * scale, { kernel: 'lanczos3' })\n .toBuffer();\n\n // Step 4: Contrast boost + sharpen\n let enhanced: Buffer;\n if (contrast !== 1.0) {\n const stats = await sharp(upscaled).stats();\n const mean = stats.channels[0].mean;\n let pipe = sharp(upscaled).linear(contrast, mean * (1 - contrast));\n if (sharpen) pipe = pipe.sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 });\n enhanced = await pipe.toBuffer();\n } else {\n enhanced = sharpen\n ? await sharp(upscaled).sharpen({ sigma: 1.0, m1: 2.0, m2: 1.0 }).toBuffer()\n : upscaled;\n }\n\n // Step 5: Crop\n let cropped: Buffer;\n if (crop === 'none') {\n cropped = enhanced;\n } else if (crop === 'auto') {\n cropped = await autoCrop(enhanced);\n } else {\n const fractions = crop === 'legacy' ? LEGACY_CROP : crop;\n const scaledW = origW * scale;\n const scaledH = origH * scale;\n const cropLeft = Math.floor(scaledW * fractions.left);\n const cropTop = Math.floor(scaledH * fractions.top);\n const cropRight = Math.floor(scaledW * fractions.right);\n const cropBottom = Math.floor(scaledH * fractions.bottom);\n const cropW = cropRight - cropLeft;\n const cropH = cropBottom - cropTop;\n cropped = await sharp(enhanced)\n .extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH })\n .toBuffer();\n }\n\n // Step 6: Negate (optional)\n const final = negate ? await sharp(cropped).negate().toBuffer() : cropped;\n\n // Step 7: Padding\n if (padding === false) {\n return sharp(final).png().toBuffer();\n }\n const pad = typeof padding === 'number' ? padding : undefined;\n const vPad = pad ?? 20;\n const hPad = pad ?? 30;\n return sharp(final)\n .extend({\n top: vPad,\n bottom: vPad,\n left: hPad,\n right: hPad,\n background: { r: 255, g: 255, b: 255 },\n })\n .png()\n .toBuffer();\n}\n\n/*\n Auto-crop: use sharp.trim() to detect the content bounding box after\n * contrast enhancement, then add a small margin. Falls back to the\n * untrimmed image if trim removes everything.\n /\nasync function autoCrop(enhanced: Buffer): Promise<Buffer> {\n try {\n const trimmed = sharp(enhanced).trim({ threshold: 30 });\n const trimmedBuf = await trimmed.toBuffer({ resolveWithObject: true });\n\n // If trim left a reasonable image, add a margin\n const { width, height } = trimmedBuf.info;\n if (width > 2 && height > 2) {\n return trimmedBuf.data;\n }\n } catch {\n // trim() can throw if image is uniform — fall through\n }\n\n // Fallback: return untrimmed\n return enhanced;\n}\n\n/\n Read an image file and return its base64-encoded content.\n */\nexport function imageToBase64(imagePath: string): string {\n const buffer = fs.readFileSync(imagePath);\n return buffer.toString('base64');\n}\n"],"mappings":";AACA,SAAS,oBAAoB;;;ACD7B,OAAO,QAAQ;AACf,OAAO,UAAU;AACjB,OAAO,WAAW;AAwClB,IAAM,cAA6B,EAAE,MAAM,KAAK,KAAK,MAAM,OAAO,KAAK,QAAQ,IAAI;AAOnF,eAAsB,kBACpB,OACA,SACiB;AACjB,QAAM,MAAM,MAAM,0BAA0B,OAAO,OAAO;AAC1D,SAAO,IAAI,SAAS,QAAQ;AAC9B;AAaA,eAAsB,0BACpB,OACA,SACiB;AACjB,QAAM;AAAA,IACJ,OAAO;AAAA,IACP,QAAQ;AAAA,IACR,WAAW;AAAA,IACX,UAAU;AAAA,IACV,OAAO;AAAA,IACP,UAAU;AAAA,IACV,SAAS;AAAA,IACT,YAAY;AAAA,EACd,IAAI,WAAW,CAAC;AAEhB,QAAM,SAAS,OAAO,UAAU,WAAW,KAAK,QAAQ,KAAK,IAAI;AAGjE,QAAM,WAAW,MAAM,MAAM,MAAM,EAAE,SAAS;AAC9C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAGvB,MAAI,WAAW,MAAM,MAAM;AAC3B,MAAI,OAAO,EAAG,YAAW,SAAS,KAAK,IAAI;AAC3C,MAAI,UAAW,YAAW,SAAS,UAAU;AAC7C,QAAM,WAAW,MAAM,SAAS,SAAS;AAGzC,QAAM,WAAW,MAAM,MAAM,QAAQ,EAClC,OAAO,QAAQ,OAAO,QAAQ,OAAO,EAAE,QAAQ,WAAW,CAAC,EAC3D,SAAS;AAGZ,MAAI;AACJ,MAAI,aAAa,GAAK;AACpB,UAAM,QAAQ,MAAM,MAAM,QAAQ,EAAE,MAAM;AAC1C,UAAM,OAAO,MAAM,SAAS,CAAC,EAAE;AAC/B,QAAI,OAAO,MAAM,QAAQ,EAAE,OAAO,UAAU,QAAQ,IAAI,SAAS;AACjE,QAAI,QAAS,QAAO,KAAK,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC;AACjE,eAAW,MAAM,KAAK,SAAS;AAAA,EACjC,OAAO;AACL,eAAW,UACP,MAAM,MAAM,QAAQ,EAAE,QAAQ,EAAE,OAAO,GAAK,IAAI,GAAK,IAAI,EAAI,CAAC,EAAE,SAAS,IACzE;AAAA,EACN;AAGA,MAAI;AACJ,MAAI,SAAS,QAAQ;AACnB,cAAU;AAAA,EACZ,WAAW,SAAS,QAAQ;AAC1B,cAAU,MAAM,SAAS,QAAQ;AAAA,EACnC,OAAO;AACL,UAAM,YAAY,SAAS,WAAW,cAAc;AACpD,UAAM,UAAU,QAAQ;AACxB,UAAM,UAAU,QAAQ;AACxB,UAAM,WAAW,KAAK,MAAM,UAAU,UAAU,IAAI;AACpD,UAAM,UAAU,KAAK,MAAM,UAAU,UAAU,GAAG;AAClD,UAAM,YAAY,KAAK,MAAM,UAAU,UAAU,KAAK;AACtD,UAAM,aAAa,KAAK,MAAM,UAAU,UAAU,MAAM;AACxD,UAAM,QAAQ,YAAY;AAC1B,UAAM,QAAQ,aAAa;AAC3B,cAAU,MAAM,MAAM,QAAQ,EAC3B,QAAQ,EAAE,MAAM,UAAU,KAAK,SAAS,OAAO,OAAO,QAAQ,MAAM,CAAC,EACrE,SAAS;AAAA,EACd;AAGA,QAAM,QAAQ,SAAS,MAAM,MAAM,OAAO,EAAE,OAAO,EAAE,SAAS,IAAI;AAGlE,MAAI,YAAY,OAAO;AACrB,WAAO,MAAM,KAAK,EAAE,IAAI,EAAE,SAAS;AAAA,EACrC;AACA,QAAM,MAAM,OAAO,YAAY,WAAW,UAAU;AACpD,QAAM,OAAO,OAAO;AACpB,QAAM,OAAO,OAAO;AACpB,SAAO,MAAM,KAAK,EACf,OAAO;AAAA,IACN,KAAK;AAAA,IACL,QAAQ;AAAA,IACR,MAAM;AAAA,IACN,OAAO;AAAA,IACP,YAAY,EAAE,GAAG,KAAK,GAAG,KAAK,GAAG,IAAI;AAAA,EACvC,CAAC,EACA,IAAI,EACJ,SAAS;AACd;AAOA,eAAe,SAAS,UAAmC;AACzD,MAAI;AACF,UAAM,UAAU,MAAM,QAAQ,EAAE,KAAK,EAAE,WAAW,GAAG,CAAC;AACtD,UAAM,aAAa,MAAM,QAAQ,SAAS,EAAE,mBAAmB,KAAK,CAAC;AAGrE,UAAM,EAAE,OAAO,OAAO,IAAI,WAAW;AACrC,QAAI,QAAQ,KAAK,SAAS,GAAG;AAC3B,aAAO,WAAW;AAAA,IACpB;AAAA,EACF,QAAQ;AAAA,EAER;AAGA,SAAO;AACT;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,GAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;ADtLA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAyDf,IAAM,iBAA2C;AAAA,EAC/C,QAAQ;AAAA,EACR,WAAW;AAAA,EACX,QAAQ;AACV;AAEA,eAAe,aACb,QACA,UACA,SACwB;AACxB,UAAQ,UAAU;AAAA,IAChB,KAAK,UAAU;AACb,YAAM,EAAE,aAAa,IAAI,MAAM,OAAO,gBAAgB;AACtD,aAAO,aAAa,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACzC;AAAA,IACA,KAAK,aAAa;AAEhB,YAAM,EAAE,gBAAgB,IAAI,MAAM,OAAO,mBAAmB;AAC5D,aAAO,gBAAgB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IAC5C;AAAA,IACA,KAAK,UAAU;AAEb,YAAM,EAAE,yBAAyB,IAAI,MAAM,OAAO,gBAAgB;AAClE,aAAO,yBAAyB,EAAE,OAAO,CAAC,EAAE,OAAO;AAAA,IACrD;AAAA,IACA;AACE,YAAM,IAAI;AAAA,QACR,qBAAqB,QAAQ;AAAA,MAC/B;AAAA,EACJ;AACF;AAUO,IAAM,0BAAkD;AAAA,EAC7D,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AAAA,EACL,GAAG;AAAA,EACH,KAAK;AACP;AASO,SAAS,aACd,UACA,gBACA,QACQ;AACR,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAEtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AACA,MAAI,SAAS,WAAW,EAAG,QAAO;AAGlC,QAAM,YAAY,oBAAI,IAAoB;AAC1C,aAAW,KAAK,UAAU;AACxB,cAAU,IAAI,EAAE,SAAS,UAAU,IAAI,EAAE,MAAM,KAAK,KAAK,CAAC;AAAA,EAC5D;AACA,MAAI,UAAU;AACd,MAAI,YAAY;AAChB,aAAW,CAAC,KAAK,KAAK,KAAK,WAAW;AACpC,QAAI,QAAQ,WAAW;AACrB,gBAAU;AACV,kBAAY;AAAA,IACd;AAAA,EACF;AAEA,QAAM,kBAAkB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,OAAO;AACnE,MAAI,gBAAgB,WAAW,EAAG,QAAO,SAAS,CAAC;AAEnD,QAAM,YAAY,UAAU,OAAO,WAAW,WAAW,SAAS;AAGlE,QAAM,SAAmB,CAAC;AAC1B,WAAS,MAAM,GAAG,MAAM,SAAS,OAAO;AACtC,UAAM,aAAa,oBAAI,IAAoB;AAC3C,eAAW,KAAK,iBAAiB;AAC/B,YAAM,KAAK,EAAE,GAAG;AAChB,iBAAW,IAAI,KAAK,WAAW,IAAI,EAAE,KAAK,KAAK,CAAC;AAAA,IAClD;AAEA,QAAI,WAAW;AAEb,YAAM,cAAc,oBAAI,IAAoB;AAC5C,iBAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,cAAM,YAAY,UAAU,EAAE,KAAK;AACnC,oBAAY,IAAI,YAAY,YAAY,IAAI,SAAS,KAAK,KAAK,KAAK;AAAA,MACtE;AAEA,UAAI,YAAY;AAChB,UAAI,iBAAiB;AACrB,iBAAW,CAAC,WAAW,KAAK,KAAK,aAAa;AAC5C,YAAI,QAAQ,gBAAgB;AAC1B,sBAAY;AACZ,2BAAiB;AAAA,QACnB;AAAA,MACF;AACA,aAAO,KAAK,SAAS;AAAA,IACvB,OAAO;AAEL,UAAI,WAAW;AACf,UAAI,gBAAgB;AACpB,iBAAW,CAAC,IAAI,KAAK,KAAK,YAAY;AACpC,YAAI,QAAQ,eAAe;AACzB,qBAAW;AACX,0BAAgB;AAAA,QAClB;AAAA,MACF;AACA,aAAO,KAAK,QAAQ;AAAA,IACtB;AAAA,EACF;AAEA,SAAO,OAAO,KAAK,EAAE;AACvB;AAIA,SAAS,YAAY,GAAuB,GAA2C;AACrF,MAAI,MAAM,UAAa,MAAM,OAAW,QAAO;AAC/C,UAAQ,KAAK,MAAM,KAAK;AAC1B;AAEA,SAAS,eAAe,QAAkD;AACxE,QAAM,OAA2B;AAAA,IAC/B,aAAa;AAAA,IACb,mBAAmB;AAAA,MACjB,eAAe;AAAA,MACf,iBAAiB;AAAA,MACjB,kBAAkB;AAAA,IACpB;AAAA,IACA,cAAc;AAAA,IACd,oBAAoB;AAAA,MAClB,YAAY;AAAA,MACZ,iBAAiB;AAAA,IACnB;AAAA,IACA,aAAa;AAAA,EACf;AACA,SAAO,OAAO;AAAA,IACZ,CAAC,KAAK,OAAO;AAAA,MACX,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,MACvD,mBAAmB;AAAA,QACjB,eAAe;AAAA,UACb,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,iBAAiB;AAAA,UACf,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,QACA,kBAAkB;AAAA,UAChB,IAAI,kBAAkB;AAAA,UACtB,EAAE,kBAAkB;AAAA,QACtB;AAAA,MACF;AAAA,MACA,cAAc,YAAY,IAAI,cAAc,EAAE,YAAY;AAAA,MAC1D,oBAAoB;AAAA,QAClB,YAAY,YAAY,IAAI,mBAAmB,YAAY,EAAE,mBAAmB,UAAU;AAAA,QAC1F,iBAAiB;AAAA,UACf,IAAI,mBAAmB;AAAA,UACvB,EAAE,mBAAmB;AAAA,QACvB;AAAA,MACF;AAAA,MACA,aAAa,YAAY,IAAI,aAAa,EAAE,WAAW;AAAA,IACzD;AAAA,IACA;AAAA,EACF;AACF;AAIO,IAAM,SAAN,MAAa;AAAA,EACV,SAA+B;AAAA,EAC/B,gBAA+C;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EAmBvD,YAAY,YAAoC,SAAyB;AACvE,QAAI,OAAO,eAAe,UAAU;AAClC,YAAM,WAAW,SAAS,YAAY;AACtC,YAAM,UAAU,SAAS,SAAS,eAAe,QAAQ;AAEzD,WAAK,gBAAgB,aAAa,YAAY,UAAU,OAAO;AAAA,IACjE,OAAO;AACL,WAAK,SAAS;AAAA,IAChB;AAAA,EACF;AAAA,EAEA,MAAc,WAAmC;AAC/C,QAAI,KAAK,OAAQ,QAAO,KAAK;AAC7B,SAAK,SAAS,MAAM,KAAK;AACzB,SAAK,gBAAgB;AACrB,WAAO,KAAK;AAAA,EACd;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA,EASA,MAAM,MAAM,OAAwB,UAAwB,CAAC,GAAyB;AACpF,UAAM;AAAA,MACJ,cAAc;AAAA,MACd;AAAA,MACA,aAAa;AAAA,MACb,UAAU;AAAA,MACV,kBAAkB;AAAA,MAClB;AAAA,IACF,IAAI;AAEJ,UAAM,QAAQ,MAAM,KAAK,SAAS;AAKlC,UAAM,CAAC,gBAAgB,WAAW,IAAI,MAAM,QAAQ,IAAI;AAAA,MACtD,0BAA0B,OAAO,UAAU;AAAA,MAC3C,0BAA0B,OAAO;AAAA,QAC/B,MAAM;AAAA,QACN,OAAO;AAAA,QACP,UAAU;AAAA,QACV,SAAS;AAAA,QACT,MAAM;AAAA,QACN,SAAS;AAAA,QACT,WAAW;AAAA,MACb,CAAC;AAAA,IACH,CAAC;AAGD,UAAM,UAAU,MAAM,QAAQ;AAAA,MAC5B,MAAM;AAAA,QAAK,EAAE,QAAQ,YAAY;AAAA,QAAG,MAClC,KAAK,cAAc,OAAO,gBAAgB,aAAa,UAAU;AAAA,MACnE;AAAA,IACF;AACA,UAAM,QAAQ,QAAQ,OAAO,CAAC,MAA0B,MAAM,IAAI;AAClE,QAAI,SAAS;AACX,YAAM,QAAQ,CAAC,GAAG,MAAM,QAAQ,IAAI,aAAa,IAAI,CAAC,KAAK,EAAE,IAAI,EAAE,CAAC;AAAA,IACtE;AAEA,UAAM,WAAW,MAAM,IAAI,CAAC,MAAM,EAAE,IAAI;AACxC,UAAM,gBAAgB,MAAM,IAAI,CAAC,MAAM,EAAE,KAAK;AAC9C,UAAM,QAAQ,eAAe,aAAa;AAE1C,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO,EAAE,MAAM,IAAI,UAAU,OAAO,cAAc;AAAA,IACpD;AAEA,WAAO;AAAA,MACL,MAAM,aAAa,UAAU,gBAAgB,eAAe;AAAA,MAC5D;AAAA,MACA;AAAA,MACA;AAAA,IACF;AAAA,EACF;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,eACA,iBACA,YAC+B;AAC/B,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,MAAM,MAAM,IAAI,MAAM,aAAa;AAAA,UACzC;AAAA,UACA,UAAU;AAAA,YACR;AAAA,cACE,MAAM;AAAA,cACN,SAAS;AAAA,gBACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,gBAC7B,EAAE,MAAM,SAAS,OAAO,cAAc;AAAA,gBACtC,EAAE,MAAM,SAAS,OAAO,gBAAgB;AAAA,cAC1C;AAAA,YACF;AAAA,UACF;AAAA,UACA,aAAa;AAAA,UACb,iBAAiB;AAAA,QACnB,CAAC;AAED,cAAM,MAAM,KAAK,KAAK;AAGtB,cAAM,QAAQ,IAAI,YAAY;AAC9B,YACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,iBAAO;AAAA,QACT;AAGA,cAAM,UAAU,IAAI,QAAQ,iBAAiB,EAAE;AAC/C,eAAO,UAAU,EAAE,MAAM,SAAS,MAAM,IAAI;AAAA,MAC9C,SAAS,MAAM;AACb,YAAI,QAAQ,YAAY;AACtB,gBAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,QACF;AACA,eAAO;AAAA,MACT;AAAA,IACF;AACA,WAAO;AAAA,EACT;AACF;","names":[]}

package/package.json CHANGED Viewed

@@ -1,6 +1,6 @@
 {
   "name": "@yigitahmetsahin/captcha-solver",
-  "version": "2.0.1",
+  "version": "3.0.0",
   "description": "AI-powered captcha solver using image preprocessing and multi-provider vision models (Vercel AI SDK)",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
@@ -26,7 +26,8 @@
     "format": "prettier --write .",
     "lint": "npm run format && eslint src --fix && tsc --noEmit -p tsconfig.check.json",
     "lint:check": "prettier --check . && eslint src && tsc --noEmit -p tsconfig.check.json",
-    "prepublishOnly": "npm run build"
+    "prepublishOnly": "npm run build",
+    "prepare": "husky"
   },
   "keywords": [
     "captcha",
@@ -62,6 +63,7 @@
     "@vitest/coverage-v8": "^4.0.18",
     "eslint": "^9.39.2",
     "eslint-config-prettier": "^10.1.8",
+    "husky": "^9.1.7",
     "prettier": "^3.8.1",
     "tsup": "^8.5.1",
     "tsx": "^4.19.0",
@@ -70,9 +72,9 @@
     "vitest": "^4.0.17"
   },
   "peerDependencies": {
-    "@ai-sdk/openai": ">=1.0.0",
     "@ai-sdk/anthropic": ">=1.0.0",
-    "@ai-sdk/google": ">=1.0.0"
+    "@ai-sdk/google": ">=1.0.0",
+    "@ai-sdk/openai": ">=1.0.0"
   },
   "peerDependenciesMeta": {
     "@ai-sdk/openai": {