@yigitahmetsahin/captcha-solver 1.1.0 → 1.2.0

package/dist/index.cjs

@@ -30,25 +30,30 @@ var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: tru
 // src/index.ts
 var index_exports = {};
 __export(index_exports, {
+  Solver: () => Solver,
   imageToBase64: () => imageToBase64,
   preprocessCaptcha: () => preprocessCaptcha,
-  solveCaptchaImage: () => solveCaptchaImage
+  preprocessCaptchaToBuffer: () => preprocessCaptchaToBuffer
 });
 module.exports = __toCommonJS(index_exports);
 
 // src/solver.ts
-var import_openai = __toESM(require("openai"), 1);
+var import_ai = require("ai");
 
 // src/preprocess.ts
 var import_fs = __toESM(require("fs"), 1);
 var import_path = __toESM(require("path"), 1);
 var import_sharp = __toESM(require("sharp"), 1);
-async function preprocessCaptcha(imagePath) {
-  const absPath = import_path.default.resolve(imagePath);
-  const metadata = await (0, import_sharp.default)(absPath).metadata();
+async function preprocessCaptcha(input) {
+  const buf = await preprocessCaptchaToBuffer(input);
+  return buf.toString("base64");
+}
+async function preprocessCaptchaToBuffer(input) {
+  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)(absPath).blur(1.5).greyscale().toBuffer();
+  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;
@@ -61,14 +66,13 @@ async function preprocessCaptcha(imagePath) {
   const cropBottom = Math.floor(scaledH * 0.6);
   const cropW = cropRight - cropLeft;
   const cropH = cropBottom - cropTop;
-  const result = await (0, import_sharp.default)(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
+  return (0, import_sharp.default)(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
     top: 20,
     bottom: 20,
     left: 30,
     right: 30,
     background: { r: 255, g: 255, b: 255 }
   }).png().toBuffer();
-  return result.toString("base64");
 }
 function imageToBase64(imagePath) {
   const buffer = import_fs.default.readFileSync(imagePath);
@@ -81,60 +85,40 @@ 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.`;
-async function singleAttempt(client, base64Image, model, maxRetries) {
-  for (let retry = 0; retry <= maxRetries; retry++) {
-    try {
-      const isReasoningModel = model.startsWith("o");
-      const tokenParam = isReasoningModel ? { max_completion_tokens: 2e3 } : { max_tokens: 256 };
-      const response = await client.chat.completions.create({
-        model,
-        messages: [
-          {
-            role: "user",
-            content: [
-              { type: "text", text: PROMPT },
-              {
-                type: "image_url",
-                image_url: {
-                  url: `data:image/png;base64,${base64Image}`
-                }
-              }
-            ]
-          }
-        ],
-        temperature: 1,
-        ...tokenParam
-      });
-      const raw = response.choices[0]?.message?.content?.trim() ?? "";
-      const lower = raw.toLowerCase();
-      if (lower.includes("sorry") || lower.includes("can't help") || lower.includes("cannot help") || lower.includes("unable to") || lower.includes("i can't") || raw.length > 20) {
-        return null;
-      }
-      const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, "");
-      return cleaned || null;
-    } catch (_err) {
-      if (retry < maxRetries) {
-        await new Promise((r) => setTimeout(r, 1e3 * (retry + 1)));
-        continue;
-      }
-      return null;
+var DEFAULT_MODELS = {
+  openai: "gpt-4o",
+  anthropic: "claude-sonnet-4-20250514",
+  google: "gemini-2.0-flash"
+};
+async function resolveModel(apiKey, provider, modelId) {
+  switch (provider) {
+    case "openai": {
+      const { createOpenAI } = await import("@ai-sdk/openai");
+      return createOpenAI({ apiKey })(modelId);
     }
+    case "anthropic": {
+      const { createAnthropic } = await import("@ai-sdk/anthropic");
+      return createAnthropic({ apiKey })(modelId);
+    }
+    case "google": {
+      const { createGoogleGenerativeAI } = await import("@ai-sdk/google");
+      return createGoogleGenerativeAI({ apiKey })(modelId);
+    }
+    default:
+      throw new Error(
+        `Unknown provider "${provider}". Install the matching @ai-sdk/* package and pass the model directly.`
+      );
   }
-  return null;
 }
 var CONFUSION_GROUPS = {
-  // Thin vertical strokes → digit 1
   "1": "1",
   I: "1",
   L: "1",
-  // Round shapes → letter O
   O: "O",
   D: "O",
   "0": "O",
-  // Similar curves
   S: "S",
   "5": "S",
-  // Straight edges
   Z: "Z",
   "2": "Z"
 };
@@ -182,34 +166,115 @@ function majorityVote(attempts, expectedLength) {
   }
   return result.join("");
 }
-async function solveCaptchaImage(imagePath, options = {}) {
-  const { model = "o3", numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
-  const client = new import_openai.default({ apiKey: process.env.OPENAI_API_KEY });
-  const base64Processed = await preprocessCaptcha(imagePath);
-  const attempts = [];
-  const maxTotalCalls = numAttempts + 4;
-  let callCount = 0;
-  while (attempts.length < numAttempts && callCount < maxTotalCalls) {
-    callCount++;
-    const result = await singleAttempt(client, base64Processed, model, maxRetries);
-    if (result) {
-      attempts.push(result);
-      if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);
+var Solver = class {
+  _model = null;
+  _pendingModel = null;
+  /**
+   * Create a captcha solver.
+   *
+   * @example
+   * // Simple — defaults to OpenAI gpt-4o
+   * const solver = new Solver('sk-...');
+   *
+   * @example
+   * // Specify provider and model
+   * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });
+   *
+   * @example
+   * // Pass an AI SDK model directly
+   * import { createOpenAI } from '@ai-sdk/openai';
+   * const openai = createOpenAI({ apiKey: 'sk-...' });
+   * const solver = new Solver(openai('gpt-4o'));
+   */
+  constructor(keyOrModel, options) {
+    if (typeof keyOrModel === "string") {
+      const provider = options?.provider ?? "openai";
+      const modelId = options?.model ?? DEFAULT_MODELS[provider];
+      this._pendingModel = resolveModel(keyOrModel, provider, modelId);
     } else {
-      if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);
+      this._model = keyOrModel;
     }
   }
-  if (attempts.length === 0) {
-    if (verbose) console.log("  All attempts failed!");
-    return "";
+  async getModel() {
+    if (this._model) return this._model;
+    this._model = await this._pendingModel;
+    this._pendingModel = null;
+    return this._model;
   }
-  const answer = majorityVote(attempts, expectedLength);
-  return answer;
-}
+  /**
+   * Solve a captcha image.
+   *
+   * @param input - File path (string) or raw image Buffer
+   * @param options - Solve options (attempts, expected length, etc.)
+   * @returns The captcha text
+   */
+  async solve(input, options = {}) {
+    const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
+    const model = await this.getModel();
+    const imageBuffer = await preprocessCaptchaToBuffer(input);
+    const attempts = [];
+    const maxTotalCalls = numAttempts + 4;
+    let callCount = 0;
+    while (attempts.length < numAttempts && callCount < maxTotalCalls) {
+      callCount++;
+      const result = await this.singleAttempt(model, imageBuffer, maxRetries);
+      if (result) {
+        attempts.push(result);
+        if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);
+      } else {
+        if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);
+      }
+    }
+    if (attempts.length === 0) {
+      if (verbose) console.log("  All attempts failed!");
+      return "";
+    }
+    return majorityVote(attempts, expectedLength);
+  }
+  /**
+   * Make a single API call to read the captcha.
+   * Retries up to `maxRetries` times on failure.
+   */
+  async singleAttempt(model, imageBuffer, maxRetries) {
+    for (let retry = 0; retry <= maxRetries; retry++) {
+      try {
+        const { text } = await (0, import_ai.generateText)({
+          model,
+          messages: [
+            {
+              role: "user",
+              content: [
+                { type: "text", text: PROMPT },
+                { type: "image", image: imageBuffer }
+              ]
+            }
+          ],
+          temperature: 1,
+          maxOutputTokens: 256
+        });
+        const raw = text.trim();
+        const lower = raw.toLowerCase();
+        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, "");
+        return cleaned || null;
+      } catch (_err) {
+        if (retry < maxRetries) {
+          await new Promise((r) => setTimeout(r, 1e3 * (retry + 1)));
+          continue;
+        }
+        return null;
+      }
+    }
+    return null;
+  }
+};
 // Annotate the CommonJS export names for ESM import in node:
 0 && (module.exports = {
+  Solver,
   imageToBase64,
   preprocessCaptcha,
-  solveCaptchaImage
+  preprocessCaptchaToBuffer
 });
 //# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/index.ts","../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["export { solveCaptchaImage } from './solver.js';\nexport { preprocessCaptcha, imageToBase64 } from './preprocess.js';\n","import OpenAI from 'openai';\nimport { preprocessCaptcha } 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\ninterface SolverOptions {\n  /** OpenAI model to use (default: \"o3\") */\n  model?: string;\n  /** Number of voting attempts (default: 5) */\n  numAttempts?: number;\n  /** Expected captcha length — results of other lengths are discarded (default: undefined = no filter) */\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\n/**\n * Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n */\nasync function singleAttempt(\n  client: OpenAI,\n  base64Image: string,\n  model: string,\n  maxRetries: number\n): Promise<string | null> {\n  for (let retry = 0; retry <= maxRetries; retry++) {\n    try {\n      // Reasoning models (o3, o4-mini) use max_completion_tokens;\n      // Standard models (gpt-4o, gpt-4.1, gpt-5.4-mini) use max_tokens.\n      const isReasoningModel = model.startsWith('o');\n      const tokenParam = isReasoningModel ? { max_completion_tokens: 2000 } : { max_tokens: 256 };\n\n      const response = await client.chat.completions.create({\n        model,\n        messages: [\n          {\n            role: 'user',\n            content: [\n              { type: 'text', text: PROMPT },\n              {\n                type: 'image_url',\n                image_url: {\n                  url: `data:image/png;base64,${base64Image}`,\n                },\n              },\n            ],\n          },\n        ],\n        temperature: 1,\n        ...tokenParam,\n      });\n\n      const raw = response.choices[0]?.message?.content?.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; // Model refused — don't count as an attempt\n      }\n\n      // Clean: keep only uppercase letters and digits\n      const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, '');\n      return cleaned || null;\n    } catch (_err) {\n      if (retry < maxRetries) {\n        // Wait briefly before retry\n        await new Promise((r) => setTimeout(r, 1000 * (retry + 1)));\n        continue;\n      }\n      return null;\n    }\n  }\n  return null;\n}\n\n/**\n * Confusion groups: 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  // Thin vertical strokes → digit 1\n  '1': '1',\n  I: '1',\n  L: '1',\n  // Round shapes → letter O\n  O: 'O',\n  D: 'O',\n  '0': 'O',\n  // Similar curves\n  S: 'S',\n  '5': 'S',\n  // Straight edges\n  Z: 'Z',\n  '2': 'Z',\n};\n\n/**\n * Character-level majority vote across multiple attempts.\n *\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 * The canonical character for the winning group is used.\n */\nfunction majorityVote(attempts: string[], expectedLength?: number): string {\n  // Filter to expected length if specified\n  let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n  // If length filter removed everything, fall back to most common length\n  if (filtered.length === 0) {\n    filtered = attempts;\n  }\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    // Count raw characters\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    // Group by canonical form and sum counts\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    // Pick the group with the highest combined count\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/**\n * Solve a captcha image using OpenAI vision + preprocessing + majority voting.\n */\nexport async function solveCaptchaImage(\n  imagePath: string,\n  options: SolverOptions = {}\n): Promise<string> {\n  const { model = 'o3', numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;\n\n  const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n\n  // Preprocess the image\n  const base64Processed = await preprocessCaptcha(imagePath);\n\n  // Run attempts — retry refusals/failures to guarantee numAttempts valid results\n  const attempts: string[] = [];\n  const maxTotalCalls = numAttempts + 4; // allow up to 4 extra calls for refusals\n  let callCount = 0;\n  while (attempts.length < numAttempts && callCount < maxTotalCalls) {\n    callCount++;\n    const result = await singleAttempt(client, base64Processed, model, maxRetries);\n    if (result) {\n      attempts.push(result);\n      if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);\n    } else {\n      if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);\n    }\n  }\n\n  if (attempts.length === 0) {\n    if (verbose) console.log('  All attempts failed!');\n    return '';\n  }\n\n  // Majority vote\n  const answer = majorityVote(attempts, expectedLength);\n  return answer;\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 * Returns a base64-encoded PNG string.\n */\nexport async function preprocessCaptcha(imagePath: string): Promise<string> {\n  const absPath = path.resolve(imagePath);\n\n  // Read original dimensions for crop/resize calculations\n  const metadata = await sharp(absPath).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(absPath).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 = factor*input + 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  const result = await 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  return result.toString('base64');\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;;;ACAA,oBAAmB;;;ACAnB,gBAAe;AACf,kBAAiB;AACjB,mBAAkB;AAelB,eAAsB,kBAAkB,WAAoC;AAC1E,QAAM,UAAU,YAAAA,QAAK,QAAQ,SAAS;AAGtC,QAAM,WAAW,UAAM,aAAAC,SAAM,OAAO,EAAE,SAAS;AAC/C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAIvB,QAAM,WAAW,UAAM,aAAAA,SAAM,OAAO,EAAE,KAAK,GAAG,EAAE,UAAU,EAAE,SAAS;AAGrE,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,QAAM,SAAS,UAAM,aAAAA,SAAM,QAAQ,EAChC,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;AAEZ,SAAO,OAAO,SAAS,QAAQ;AACjC;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,UAAAC,QAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;AD1EA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAuBf,eAAe,cACb,QACA,aACA,OACA,YACwB;AACxB,WAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,QAAI;AAGF,YAAM,mBAAmB,MAAM,WAAW,GAAG;AAC7C,YAAM,aAAa,mBAAmB,EAAE,uBAAuB,IAAK,IAAI,EAAE,YAAY,IAAI;AAE1F,YAAM,WAAW,MAAM,OAAO,KAAK,YAAY,OAAO;AAAA,QACpD;AAAA,QACA,UAAU;AAAA,UACR;AAAA,YACE,MAAM;AAAA,YACN,SAAS;AAAA,cACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,cAC7B;AAAA,gBACE,MAAM;AAAA,gBACN,WAAW;AAAA,kBACT,KAAK,yBAAyB,WAAW;AAAA,gBAC3C;AAAA,cACF;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,QACA,aAAa;AAAA,QACb,GAAG;AAAA,MACL,CAAC;AAED,YAAM,MAAM,SAAS,QAAQ,CAAC,GAAG,SAAS,SAAS,KAAK,KAAK;AAG7D,YAAM,QAAQ,IAAI,YAAY;AAC9B,UACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,eAAO;AAAA,MACT;AAGA,YAAM,UAAU,IAAI,YAAY,EAAE,QAAQ,cAAc,EAAE;AAC1D,aAAO,WAAW;AAAA,IACpB,SAAS,MAAM;AACb,UAAI,QAAQ,YAAY;AAEtB,cAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AAMA,IAAM,mBAA2C;AAAA;AAAA,EAE/C,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA;AAAA,EAEH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA;AAAA,EAEL,GAAG;AAAA,EACH,KAAK;AAAA;AAAA,EAEL,GAAG;AAAA,EACH,KAAK;AACP;AASA,SAAS,aAAa,UAAoB,gBAAiC;AAEzE,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAGtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AAEA,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;AAEtC,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;AAGA,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;AAGA,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;AAKA,eAAsB,kBACpB,WACA,UAAyB,CAAC,GACT;AACjB,QAAM,EAAE,QAAQ,MAAM,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE1F,QAAM,SAAS,IAAI,cAAAC,QAAO,EAAE,QAAQ,QAAQ,IAAI,eAAe,CAAC;AAGhE,QAAM,kBAAkB,MAAM,kBAAkB,SAAS;AAGzD,QAAM,WAAqB,CAAC;AAC5B,QAAM,gBAAgB,cAAc;AACpC,MAAI,YAAY;AAChB,SAAO,SAAS,SAAS,eAAe,YAAY,eAAe;AACjE;AACA,UAAM,SAAS,MAAM,cAAc,QAAQ,iBAAiB,OAAO,UAAU;AAC7E,QAAI,QAAQ;AACV,eAAS,KAAK,MAAM;AACpB,UAAI,QAAS,SAAQ,IAAI,aAAa,SAAS,MAAM,KAAK,MAAM,EAAE;AAAA,IACpE,OAAO;AACL,UAAI,QAAS,SAAQ,IAAI,UAAU,SAAS,iCAAiC;AAAA,IAC/E;AAAA,EACF;AAEA,MAAI,SAAS,WAAW,GAAG;AACzB,QAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,WAAO;AAAA,EACT;AAGA,QAAM,SAAS,aAAa,UAAU,cAAc;AACpD,SAAO;AACT;","names":["path","sharp","fs","OpenAI"]}
+{"version":3,"sources":["../src/index.ts","../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["export { Solver } from './solver.js';\nexport type { SolverOptions, SolveOptions, Provider } from './solver.js';\nexport { preprocessCaptcha, preprocessCaptchaToBuffer, imageToBase64 } from './preprocess.js';\n","import type { LanguageModel } 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\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// ── 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 The captcha text\n   */\n  async solve(input: string | Buffer, options: SolveOptions = {}): Promise<string> {\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    // Run attempts — retry refusals/failures to guarantee numAttempts valid results\n    const attempts: string[] = [];\n    const maxTotalCalls = numAttempts + 4;\n    let callCount = 0;\n\n    while (attempts.length < numAttempts && callCount < maxTotalCalls) {\n      callCount++;\n      const result = await this.singleAttempt(model, imageBuffer, maxRetries);\n      if (result) {\n        attempts.push(result);\n        if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);\n      } else {\n        if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);\n      }\n    }\n\n    if (attempts.length === 0) {\n      if (verbose) console.log('  All attempts failed!');\n      return '';\n    }\n\n    return majorityVote(attempts, expectedLength);\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<string | null> {\n    for (let retry = 0; retry <= maxRetries; retry++) {\n      try {\n        const { text } = 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 || 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 = factor*input + 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;AA8Bf,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;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,GAAoB;AAC/E,UAAM,EAAE,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE5E,UAAM,QAAQ,MAAM,KAAK,SAAS;AAClC,UAAM,cAAc,MAAM,0BAA0B,KAAK;AAGzD,UAAM,WAAqB,CAAC;AAC5B,UAAM,gBAAgB,cAAc;AACpC,QAAI,YAAY;AAEhB,WAAO,SAAS,SAAS,eAAe,YAAY,eAAe;AACjE;AACA,YAAM,SAAS,MAAM,KAAK,cAAc,OAAO,aAAa,UAAU;AACtE,UAAI,QAAQ;AACV,iBAAS,KAAK,MAAM;AACpB,YAAI,QAAS,SAAQ,IAAI,aAAa,SAAS,MAAM,KAAK,MAAM,EAAE;AAAA,MACpE,OAAO;AACL,YAAI,QAAS,SAAQ,IAAI,UAAU,SAAS,iCAAiC;AAAA,MAC/E;AAAA,IACF;AAEA,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO;AAAA,IACT;AAEA,WAAO,aAAa,UAAU,cAAc;AAAA,EAC9C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,aACA,YACwB;AACxB,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,KAAK,IAAI,UAAM,wBAAa;AAAA,UAClC;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,WAAW;AAAA,MACpB,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

@@ -1,19 +1,58 @@
+import { LanguageModel } from 'ai';
+
+type Provider = 'openai' | 'anthropic' | 'google';
 interface SolverOptions {
-    /** OpenAI model to use (default: "o3") */
+    /** AI provider to use when constructing the model from an API key (default: "openai") */
+    provider?: Provider;
+    /** Model ID passed to the provider (default: "gpt-4o") */
     model?: string;
+}
+interface SolveOptions {
     /** Number of voting attempts (default: 5) */
     numAttempts?: number;
-    /** Expected captcha length — results of other lengths are discarded (default: undefined = no filter) */
+    /** Expected captcha length — results of other lengths are discarded */
     expectedLength?: number;
     /** Max retries per attempt on API failure (default: 2) */
     maxRetries?: number;
     /** Whether to log attempt details (default: true) */
     verbose?: boolean;
 }
-/**
- * Solve a captcha image using OpenAI vision + preprocessing + majority voting.
- */
-declare function solveCaptchaImage(imagePath: string, options?: SolverOptions): Promise<string>;
+declare class Solver {
+    private _model;
+    private _pendingModel;
+    /**
+     * Create a captcha solver.
+     *
+     * @example
+     * // Simple — defaults to OpenAI gpt-4o
+     * const solver = new Solver('sk-...');
+     *
+     * @example
+     * // Specify provider and model
+     * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });
+     *
+     * @example
+     * // Pass an AI SDK model directly
+     * import { createOpenAI } from '@ai-sdk/openai';
+     * const openai = createOpenAI({ apiKey: 'sk-...' });
+     * const solver = new Solver(openai('gpt-4o'));
+     */
+    constructor(keyOrModel: string | LanguageModel, options?: SolverOptions);
+    private getModel;
+    /**
+     * Solve a captcha image.
+     *
+     * @param input - File path (string) or raw image Buffer
+     * @param options - Solve options (attempts, expected length, etc.)
+     * @returns The captcha text
+     */
+    solve(input: string | Buffer, options?: SolveOptions): Promise<string>;
+    /**
+     * Make a single API call to read the captcha.
+     * Retries up to `maxRetries` times on failure.
+     */
+    private singleAttempt;
+}
 
 /**
  * Preprocess a captcha image using sharp (libvips).
@@ -26,12 +65,18 @@ declare function solveCaptchaImage(imagePath: string, options?: SolverOptions):
  *   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(imagePath: string): Promise<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 { imageToBase64, preprocessCaptcha, solveCaptchaImage };
+export { type Provider, type SolveOptions, Solver, type SolverOptions, imageToBase64, preprocessCaptcha, preprocessCaptchaToBuffer };

package/dist/index.d.ts

@@ -1,19 +1,58 @@
+import { LanguageModel } from 'ai';
+
+type Provider = 'openai' | 'anthropic' | 'google';
 interface SolverOptions {
-    /** OpenAI model to use (default: "o3") */
+    /** AI provider to use when constructing the model from an API key (default: "openai") */
+    provider?: Provider;
+    /** Model ID passed to the provider (default: "gpt-4o") */
     model?: string;
+}
+interface SolveOptions {
     /** Number of voting attempts (default: 5) */
     numAttempts?: number;
-    /** Expected captcha length — results of other lengths are discarded (default: undefined = no filter) */
+    /** Expected captcha length — results of other lengths are discarded */
     expectedLength?: number;
     /** Max retries per attempt on API failure (default: 2) */
     maxRetries?: number;
     /** Whether to log attempt details (default: true) */
     verbose?: boolean;
 }
-/**
- * Solve a captcha image using OpenAI vision + preprocessing + majority voting.
- */
-declare function solveCaptchaImage(imagePath: string, options?: SolverOptions): Promise<string>;
+declare class Solver {
+    private _model;
+    private _pendingModel;
+    /**
+     * Create a captcha solver.
+     *
+     * @example
+     * // Simple — defaults to OpenAI gpt-4o
+     * const solver = new Solver('sk-...');
+     *
+     * @example
+     * // Specify provider and model
+     * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });
+     *
+     * @example
+     * // Pass an AI SDK model directly
+     * import { createOpenAI } from '@ai-sdk/openai';
+     * const openai = createOpenAI({ apiKey: 'sk-...' });
+     * const solver = new Solver(openai('gpt-4o'));
+     */
+    constructor(keyOrModel: string | LanguageModel, options?: SolverOptions);
+    private getModel;
+    /**
+     * Solve a captcha image.
+     *
+     * @param input - File path (string) or raw image Buffer
+     * @param options - Solve options (attempts, expected length, etc.)
+     * @returns The captcha text
+     */
+    solve(input: string | Buffer, options?: SolveOptions): Promise<string>;
+    /**
+     * Make a single API call to read the captcha.
+     * Retries up to `maxRetries` times on failure.
+     */
+    private singleAttempt;
+}
 
 /**
  * Preprocess a captcha image using sharp (libvips).
@@ -26,12 +65,18 @@ declare function solveCaptchaImage(imagePath: string, options?: SolverOptions):
  *   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(imagePath: string): Promise<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 { imageToBase64, preprocessCaptcha, solveCaptchaImage };
+export { type Provider, type SolveOptions, Solver, type SolverOptions, imageToBase64, preprocessCaptcha, preprocessCaptchaToBuffer };

package/dist/index.js

@@ -1,16 +1,20 @@
 // src/solver.ts
-import OpenAI from "openai";
+import { generateText } from "ai";
 
 // src/preprocess.ts
 import fs from "fs";
 import path from "path";
 import sharp from "sharp";
-async function preprocessCaptcha(imagePath) {
-  const absPath = path.resolve(imagePath);
-  const metadata = await sharp(absPath).metadata();
+async function preprocessCaptcha(input) {
+  const buf = await preprocessCaptchaToBuffer(input);
+  return buf.toString("base64");
+}
+async function preprocessCaptchaToBuffer(input) {
+  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(absPath).blur(1.5).greyscale().toBuffer();
+  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;
@@ -23,14 +27,13 @@ async function preprocessCaptcha(imagePath) {
   const cropBottom = Math.floor(scaledH * 0.6);
   const cropW = cropRight - cropLeft;
   const cropH = cropBottom - cropTop;
-  const result = await sharp(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
+  return sharp(enhanced).extract({ left: cropLeft, top: cropTop, width: cropW, height: cropH }).extend({
     top: 20,
     bottom: 20,
     left: 30,
     right: 30,
     background: { r: 255, g: 255, b: 255 }
   }).png().toBuffer();
-  return result.toString("base64");
 }
 function imageToBase64(imagePath) {
   const buffer = fs.readFileSync(imagePath);
@@ -43,60 +46,40 @@ 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.`;
-async function singleAttempt(client, base64Image, model, maxRetries) {
-  for (let retry = 0; retry <= maxRetries; retry++) {
-    try {
-      const isReasoningModel = model.startsWith("o");
-      const tokenParam = isReasoningModel ? { max_completion_tokens: 2e3 } : { max_tokens: 256 };
-      const response = await client.chat.completions.create({
-        model,
-        messages: [
-          {
-            role: "user",
-            content: [
-              { type: "text", text: PROMPT },
-              {
-                type: "image_url",
-                image_url: {
-                  url: `data:image/png;base64,${base64Image}`
-                }
-              }
-            ]
-          }
-        ],
-        temperature: 1,
-        ...tokenParam
-      });
-      const raw = response.choices[0]?.message?.content?.trim() ?? "";
-      const lower = raw.toLowerCase();
-      if (lower.includes("sorry") || lower.includes("can't help") || lower.includes("cannot help") || lower.includes("unable to") || lower.includes("i can't") || raw.length > 20) {
-        return null;
-      }
-      const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, "");
-      return cleaned || null;
-    } catch (_err) {
-      if (retry < maxRetries) {
-        await new Promise((r) => setTimeout(r, 1e3 * (retry + 1)));
-        continue;
-      }
-      return null;
+var DEFAULT_MODELS = {
+  openai: "gpt-4o",
+  anthropic: "claude-sonnet-4-20250514",
+  google: "gemini-2.0-flash"
+};
+async function resolveModel(apiKey, provider, modelId) {
+  switch (provider) {
+    case "openai": {
+      const { createOpenAI } = await import("@ai-sdk/openai");
+      return createOpenAI({ apiKey })(modelId);
     }
+    case "anthropic": {
+      const { createAnthropic } = await import("@ai-sdk/anthropic");
+      return createAnthropic({ apiKey })(modelId);
+    }
+    case "google": {
+      const { createGoogleGenerativeAI } = await import("@ai-sdk/google");
+      return createGoogleGenerativeAI({ apiKey })(modelId);
+    }
+    default:
+      throw new Error(
+        `Unknown provider "${provider}". Install the matching @ai-sdk/* package and pass the model directly.`
+      );
   }
-  return null;
 }
 var CONFUSION_GROUPS = {
-  // Thin vertical strokes → digit 1
   "1": "1",
   I: "1",
   L: "1",
-  // Round shapes → letter O
   O: "O",
   D: "O",
   "0": "O",
-  // Similar curves
   S: "S",
   "5": "S",
-  // Straight edges
   Z: "Z",
   "2": "Z"
 };
@@ -144,33 +127,114 @@ function majorityVote(attempts, expectedLength) {
   }
   return result.join("");
 }
-async function solveCaptchaImage(imagePath, options = {}) {
-  const { model = "o3", numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
-  const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });
-  const base64Processed = await preprocessCaptcha(imagePath);
-  const attempts = [];
-  const maxTotalCalls = numAttempts + 4;
-  let callCount = 0;
-  while (attempts.length < numAttempts && callCount < maxTotalCalls) {
-    callCount++;
-    const result = await singleAttempt(client, base64Processed, model, maxRetries);
-    if (result) {
-      attempts.push(result);
-      if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);
+var Solver = class {
+  _model = null;
+  _pendingModel = null;
+  /**
+   * Create a captcha solver.
+   *
+   * @example
+   * // Simple — defaults to OpenAI gpt-4o
+   * const solver = new Solver('sk-...');
+   *
+   * @example
+   * // Specify provider and model
+   * const solver = new Solver('sk-ant-...', { provider: 'anthropic', model: 'claude-sonnet-4-20250514' });
+   *
+   * @example
+   * // Pass an AI SDK model directly
+   * import { createOpenAI } from '@ai-sdk/openai';
+   * const openai = createOpenAI({ apiKey: 'sk-...' });
+   * const solver = new Solver(openai('gpt-4o'));
+   */
+  constructor(keyOrModel, options) {
+    if (typeof keyOrModel === "string") {
+      const provider = options?.provider ?? "openai";
+      const modelId = options?.model ?? DEFAULT_MODELS[provider];
+      this._pendingModel = resolveModel(keyOrModel, provider, modelId);
     } else {
-      if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);
+      this._model = keyOrModel;
     }
   }
-  if (attempts.length === 0) {
-    if (verbose) console.log("  All attempts failed!");
-    return "";
+  async getModel() {
+    if (this._model) return this._model;
+    this._model = await this._pendingModel;
+    this._pendingModel = null;
+    return this._model;
   }
-  const answer = majorityVote(attempts, expectedLength);
-  return answer;
-}
+  /**
+   * Solve a captcha image.
+   *
+   * @param input - File path (string) or raw image Buffer
+   * @param options - Solve options (attempts, expected length, etc.)
+   * @returns The captcha text
+   */
+  async solve(input, options = {}) {
+    const { numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;
+    const model = await this.getModel();
+    const imageBuffer = await preprocessCaptchaToBuffer(input);
+    const attempts = [];
+    const maxTotalCalls = numAttempts + 4;
+    let callCount = 0;
+    while (attempts.length < numAttempts && callCount < maxTotalCalls) {
+      callCount++;
+      const result = await this.singleAttempt(model, imageBuffer, maxRetries);
+      if (result) {
+        attempts.push(result);
+        if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);
+      } else {
+        if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);
+      }
+    }
+    if (attempts.length === 0) {
+      if (verbose) console.log("  All attempts failed!");
+      return "";
+    }
+    return majorityVote(attempts, expectedLength);
+  }
+  /**
+   * Make a single API call to read the captcha.
+   * Retries up to `maxRetries` times on failure.
+   */
+  async singleAttempt(model, imageBuffer, maxRetries) {
+    for (let retry = 0; retry <= maxRetries; retry++) {
+      try {
+        const { text } = await generateText({
+          model,
+          messages: [
+            {
+              role: "user",
+              content: [
+                { type: "text", text: PROMPT },
+                { type: "image", image: imageBuffer }
+              ]
+            }
+          ],
+          temperature: 1,
+          maxOutputTokens: 256
+        });
+        const raw = text.trim();
+        const lower = raw.toLowerCase();
+        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, "");
+        return cleaned || null;
+      } catch (_err) {
+        if (retry < maxRetries) {
+          await new Promise((r) => setTimeout(r, 1e3 * (retry + 1)));
+          continue;
+        }
+        return null;
+      }
+    }
+    return null;
+  }
+};
 export {
+  Solver,
   imageToBase64,
   preprocessCaptcha,
-  solveCaptchaImage
+  preprocessCaptchaToBuffer
 };
 //# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -1 +1 @@
-{"version":3,"sources":["../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["import OpenAI from 'openai';\nimport { preprocessCaptcha } 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\ninterface SolverOptions {\n  /** OpenAI model to use (default: \"o3\") */\n  model?: string;\n  /** Number of voting attempts (default: 5) */\n  numAttempts?: number;\n  /** Expected captcha length — results of other lengths are discarded (default: undefined = no filter) */\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\n/**\n * Make a single API call to read the captcha.\n * Retries up to `maxRetries` times on failure.\n */\nasync function singleAttempt(\n  client: OpenAI,\n  base64Image: string,\n  model: string,\n  maxRetries: number\n): Promise<string | null> {\n  for (let retry = 0; retry <= maxRetries; retry++) {\n    try {\n      // Reasoning models (o3, o4-mini) use max_completion_tokens;\n      // Standard models (gpt-4o, gpt-4.1, gpt-5.4-mini) use max_tokens.\n      const isReasoningModel = model.startsWith('o');\n      const tokenParam = isReasoningModel ? { max_completion_tokens: 2000 } : { max_tokens: 256 };\n\n      const response = await client.chat.completions.create({\n        model,\n        messages: [\n          {\n            role: 'user',\n            content: [\n              { type: 'text', text: PROMPT },\n              {\n                type: 'image_url',\n                image_url: {\n                  url: `data:image/png;base64,${base64Image}`,\n                },\n              },\n            ],\n          },\n        ],\n        temperature: 1,\n        ...tokenParam,\n      });\n\n      const raw = response.choices[0]?.message?.content?.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; // Model refused — don't count as an attempt\n      }\n\n      // Clean: keep only uppercase letters and digits\n      const cleaned = raw.toUpperCase().replace(/[^A-Z0-9]/g, '');\n      return cleaned || null;\n    } catch (_err) {\n      if (retry < maxRetries) {\n        // Wait briefly before retry\n        await new Promise((r) => setTimeout(r, 1000 * (retry + 1)));\n        continue;\n      }\n      return null;\n    }\n  }\n  return null;\n}\n\n/**\n * Confusion groups: 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  // Thin vertical strokes → digit 1\n  '1': '1',\n  I: '1',\n  L: '1',\n  // Round shapes → letter O\n  O: 'O',\n  D: 'O',\n  '0': 'O',\n  // Similar curves\n  S: 'S',\n  '5': 'S',\n  // Straight edges\n  Z: 'Z',\n  '2': 'Z',\n};\n\n/**\n * Character-level majority vote across multiple attempts.\n *\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 * The canonical character for the winning group is used.\n */\nfunction majorityVote(attempts: string[], expectedLength?: number): string {\n  // Filter to expected length if specified\n  let filtered = expectedLength ? attempts.filter((a) => a.length === expectedLength) : attempts;\n\n  // If length filter removed everything, fall back to most common length\n  if (filtered.length === 0) {\n    filtered = attempts;\n  }\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    // Count raw characters\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    // Group by canonical form and sum counts\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    // Pick the group with the highest combined count\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/**\n * Solve a captcha image using OpenAI vision + preprocessing + majority voting.\n */\nexport async function solveCaptchaImage(\n  imagePath: string,\n  options: SolverOptions = {}\n): Promise<string> {\n  const { model = 'o3', numAttempts = 5, expectedLength, maxRetries = 2, verbose = true } = options;\n\n  const client = new OpenAI({ apiKey: process.env.OPENAI_API_KEY });\n\n  // Preprocess the image\n  const base64Processed = await preprocessCaptcha(imagePath);\n\n  // Run attempts — retry refusals/failures to guarantee numAttempts valid results\n  const attempts: string[] = [];\n  const maxTotalCalls = numAttempts + 4; // allow up to 4 extra calls for refusals\n  let callCount = 0;\n  while (attempts.length < numAttempts && callCount < maxTotalCalls) {\n    callCount++;\n    const result = await singleAttempt(client, base64Processed, model, maxRetries);\n    if (result) {\n      attempts.push(result);\n      if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);\n    } else {\n      if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);\n    }\n  }\n\n  if (attempts.length === 0) {\n    if (verbose) console.log('  All attempts failed!');\n    return '';\n  }\n\n  // Majority vote\n  const answer = majorityVote(attempts, expectedLength);\n  return answer;\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 * Returns a base64-encoded PNG string.\n */\nexport async function preprocessCaptcha(imagePath: string): Promise<string> {\n  const absPath = path.resolve(imagePath);\n\n  // Read original dimensions for crop/resize calculations\n  const metadata = await sharp(absPath).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(absPath).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 = factor*input + 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  const result = await 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  return result.toString('base64');\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,OAAO,YAAY;;;ACAnB,OAAO,QAAQ;AACf,OAAO,UAAU;AACjB,OAAO,WAAW;AAelB,eAAsB,kBAAkB,WAAoC;AAC1E,QAAM,UAAU,KAAK,QAAQ,SAAS;AAGtC,QAAM,WAAW,MAAM,MAAM,OAAO,EAAE,SAAS;AAC/C,QAAM,QAAQ,SAAS;AACvB,QAAM,QAAQ,SAAS;AAIvB,QAAM,WAAW,MAAM,MAAM,OAAO,EAAE,KAAK,GAAG,EAAE,UAAU,EAAE,SAAS;AAGrE,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,QAAM,SAAS,MAAM,MAAM,QAAQ,EAChC,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;AAEZ,SAAO,OAAO,SAAS,QAAQ;AACjC;AAKO,SAAS,cAAc,WAA2B;AACvD,QAAM,SAAS,GAAG,aAAa,SAAS;AACxC,SAAO,OAAO,SAAS,QAAQ;AACjC;;;AD1EA,IAAM,SAAS;AAAA;AAAA;AAAA;AAAA;AAuBf,eAAe,cACb,QACA,aACA,OACA,YACwB;AACxB,WAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,QAAI;AAGF,YAAM,mBAAmB,MAAM,WAAW,GAAG;AAC7C,YAAM,aAAa,mBAAmB,EAAE,uBAAuB,IAAK,IAAI,EAAE,YAAY,IAAI;AAE1F,YAAM,WAAW,MAAM,OAAO,KAAK,YAAY,OAAO;AAAA,QACpD;AAAA,QACA,UAAU;AAAA,UACR;AAAA,YACE,MAAM;AAAA,YACN,SAAS;AAAA,cACP,EAAE,MAAM,QAAQ,MAAM,OAAO;AAAA,cAC7B;AAAA,gBACE,MAAM;AAAA,gBACN,WAAW;AAAA,kBACT,KAAK,yBAAyB,WAAW;AAAA,gBAC3C;AAAA,cACF;AAAA,YACF;AAAA,UACF;AAAA,QACF;AAAA,QACA,aAAa;AAAA,QACb,GAAG;AAAA,MACL,CAAC;AAED,YAAM,MAAM,SAAS,QAAQ,CAAC,GAAG,SAAS,SAAS,KAAK,KAAK;AAG7D,YAAM,QAAQ,IAAI,YAAY;AAC9B,UACE,MAAM,SAAS,OAAO,KACtB,MAAM,SAAS,YAAY,KAC3B,MAAM,SAAS,aAAa,KAC5B,MAAM,SAAS,WAAW,KAC1B,MAAM,SAAS,SAAS,KACxB,IAAI,SAAS,IACb;AACA,eAAO;AAAA,MACT;AAGA,YAAM,UAAU,IAAI,YAAY,EAAE,QAAQ,cAAc,EAAE;AAC1D,aAAO,WAAW;AAAA,IACpB,SAAS,MAAM;AACb,UAAI,QAAQ,YAAY;AAEtB,cAAM,IAAI,QAAQ,CAAC,MAAM,WAAW,GAAG,OAAQ,QAAQ,EAAE,CAAC;AAC1D;AAAA,MACF;AACA,aAAO;AAAA,IACT;AAAA,EACF;AACA,SAAO;AACT;AAMA,IAAM,mBAA2C;AAAA;AAAA,EAE/C,KAAK;AAAA,EACL,GAAG;AAAA,EACH,GAAG;AAAA;AAAA,EAEH,GAAG;AAAA,EACH,GAAG;AAAA,EACH,KAAK;AAAA;AAAA,EAEL,GAAG;AAAA,EACH,KAAK;AAAA;AAAA,EAEL,GAAG;AAAA,EACH,KAAK;AACP;AASA,SAAS,aAAa,UAAoB,gBAAiC;AAEzE,MAAI,WAAW,iBAAiB,SAAS,OAAO,CAAC,MAAM,EAAE,WAAW,cAAc,IAAI;AAGtF,MAAI,SAAS,WAAW,GAAG;AACzB,eAAW;AAAA,EACb;AAEA,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;AAEtC,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;AAGA,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;AAGA,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;AAKA,eAAsB,kBACpB,WACA,UAAyB,CAAC,GACT;AACjB,QAAM,EAAE,QAAQ,MAAM,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE1F,QAAM,SAAS,IAAI,OAAO,EAAE,QAAQ,QAAQ,IAAI,eAAe,CAAC;AAGhE,QAAM,kBAAkB,MAAM,kBAAkB,SAAS;AAGzD,QAAM,WAAqB,CAAC;AAC5B,QAAM,gBAAgB,cAAc;AACpC,MAAI,YAAY;AAChB,SAAO,SAAS,SAAS,eAAe,YAAY,eAAe;AACjE;AACA,UAAM,SAAS,MAAM,cAAc,QAAQ,iBAAiB,OAAO,UAAU;AAC7E,QAAI,QAAQ;AACV,eAAS,KAAK,MAAM;AACpB,UAAI,QAAS,SAAQ,IAAI,aAAa,SAAS,MAAM,KAAK,MAAM,EAAE;AAAA,IACpE,OAAO;AACL,UAAI,QAAS,SAAQ,IAAI,UAAU,SAAS,iCAAiC;AAAA,IAC/E;AAAA,EACF;AAEA,MAAI,SAAS,WAAW,GAAG;AACzB,QAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,WAAO;AAAA,EACT;AAGA,QAAM,SAAS,aAAa,UAAU,cAAc;AACpD,SAAO;AACT;","names":[]}
+{"version":3,"sources":["../src/solver.ts","../src/preprocess.ts"],"sourcesContent":["import type { LanguageModel } 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\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// ── 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 The captcha text\n   */\n  async solve(input: string | Buffer, options: SolveOptions = {}): Promise<string> {\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    // Run attempts — retry refusals/failures to guarantee numAttempts valid results\n    const attempts: string[] = [];\n    const maxTotalCalls = numAttempts + 4;\n    let callCount = 0;\n\n    while (attempts.length < numAttempts && callCount < maxTotalCalls) {\n      callCount++;\n      const result = await this.singleAttempt(model, imageBuffer, maxRetries);\n      if (result) {\n        attempts.push(result);\n        if (verbose) console.log(`  Attempt ${attempts.length}: ${result}`);\n      } else {\n        if (verbose) console.log(`  Call ${callCount}: (refused/failed, retrying...)`);\n      }\n    }\n\n    if (attempts.length === 0) {\n      if (verbose) console.log('  All attempts failed!');\n      return '';\n    }\n\n    return majorityVote(attempts, expectedLength);\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<string | null> {\n    for (let retry = 0; retry <= maxRetries; retry++) {\n      try {\n        const { text } = 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 || 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 = factor*input + 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;AA8Bf,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;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,GAAoB;AAC/E,UAAM,EAAE,cAAc,GAAG,gBAAgB,aAAa,GAAG,UAAU,KAAK,IAAI;AAE5E,UAAM,QAAQ,MAAM,KAAK,SAAS;AAClC,UAAM,cAAc,MAAM,0BAA0B,KAAK;AAGzD,UAAM,WAAqB,CAAC;AAC5B,UAAM,gBAAgB,cAAc;AACpC,QAAI,YAAY;AAEhB,WAAO,SAAS,SAAS,eAAe,YAAY,eAAe;AACjE;AACA,YAAM,SAAS,MAAM,KAAK,cAAc,OAAO,aAAa,UAAU;AACtE,UAAI,QAAQ;AACV,iBAAS,KAAK,MAAM;AACpB,YAAI,QAAS,SAAQ,IAAI,aAAa,SAAS,MAAM,KAAK,MAAM,EAAE;AAAA,MACpE,OAAO;AACL,YAAI,QAAS,SAAQ,IAAI,UAAU,SAAS,iCAAiC;AAAA,MAC/E;AAAA,IACF;AAEA,QAAI,SAAS,WAAW,GAAG;AACzB,UAAI,QAAS,SAAQ,IAAI,wBAAwB;AACjD,aAAO;AAAA,IACT;AAEA,WAAO,aAAa,UAAU,cAAc;AAAA,EAC9C;AAAA;AAAA;AAAA;AAAA;AAAA,EAMA,MAAc,cACZ,OACA,aACA,YACwB;AACxB,aAAS,QAAQ,GAAG,SAAS,YAAY,SAAS;AAChD,UAAI;AACF,cAAM,EAAE,KAAK,IAAI,MAAM,aAAa;AAAA,UAClC;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,WAAW;AAAA,MACpB,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

@@ -1,7 +1,7 @@
 {
   "name": "@yigitahmetsahin/captcha-solver",
-  "version": "1.1.0",
-  "description": "AI-powered captcha solver using image preprocessing and OpenAI vision models",
+  "version": "1.2.0",
+  "description": "AI-powered captcha solver using image preprocessing and multi-provider vision models (Vercel AI SDK)",
   "main": "dist/index.js",
   "module": "dist/index.mjs",
   "types": "dist/index.d.ts",
@@ -32,7 +32,10 @@
     "captcha",
     "solver",
     "ocr",
+    "ai-sdk",
     "openai",
+    "anthropic",
+    "google",
     "vision",
     "image-processing",
     "typescript"
@@ -48,11 +51,12 @@
   },
   "homepage": "https://github.com/yigitahmetsahin/captcha-solver#readme",
   "dependencies": {
+    "ai": "^6.0.146",
     "dotenv": "^16.4.7",
-    "openai": "^4.77.0",
     "sharp": "^0.33.5"
   },
   "devDependencies": {
+    "@ai-sdk/openai": "^3.0.50",
     "@eslint/js": "^9.39.2",
     "@types/node": "^22.10.0",
     "@vitest/coverage-v8": "^4.0.18",
@@ -65,6 +69,22 @@
     "typescript-eslint": "^8.53.1",
     "vitest": "^4.0.17"
   },
+  "peerDependencies": {
+    "@ai-sdk/openai": ">=1.0.0",
+    "@ai-sdk/anthropic": ">=1.0.0",
+    "@ai-sdk/google": ">=1.0.0"
+  },
+  "peerDependenciesMeta": {
+    "@ai-sdk/openai": {
+      "optional": true
+    },
+    "@ai-sdk/anthropic": {
+      "optional": true
+    },
+    "@ai-sdk/google": {
+      "optional": true
+    }
+  },
   "engines": {
     "node": ">=24"
   },