json-from-llm 0.1.0

package/CHANGELOG.md

@@ -0,0 +1,22 @@
+# Changelog
+
+All notable changes to this project are documented here. The format follows
+[Keep a Changelog](https://keepachangelog.com/en/1.1.0/) and the project adheres
+to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+## [0.1.0] - 2026-06-04
+
+### Added
+
+- `extractJson(text, options?)` and `tryExtractJson(text, options?)` — recover a
+  JSON value from LLM output wrapped in reasoning/thinking tags, markdown fences
+  or prose.
+- Strips `<think>` / `<thinking>` / `<reasoning>` blocks before scanning.
+- String-aware balanced-value scanner and trailing-comma repair (never corrupt
+  string contents).
+- `expect: 'object' | 'array' | 'any'` option and `repair` toggle.
+- Low-level exports: `stripReasoning`, `fencedBlocks`, `balancedSpans`,
+  `removeTrailingCommas`.
+- Zero runtime dependencies; ESM + CJS builds with type declarations.
+
+[0.1.0]: https://github.com/slegarraga/json-from-llm/releases/tag/v0.1.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Sebastian Legarraga
+
+Permission is hereby granted, free of charge, to any person obtaining a copy
+of this software and associated documentation files (the "Software"), to deal
+in the Software without restriction, including without limitation the rights
+to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
+copies of the Software, and to permit persons to whom the Software is
+furnished to do so, subject to the following conditions:
+
+The above copyright notice and this permission notice shall be included in all
+copies or substantial portions of the Software.
+
+THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
+IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
+FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
+AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
+LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
+OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
+SOFTWARE.

package/README.md

@@ -0,0 +1,82 @@
+# json-from-llm
+
+> Extract valid JSON from an LLM response — even when it's wrapped in reasoning/thinking tags, markdown fences or prose. **Zero dependencies.**
+
+You asked for JSON. The model gave you:
+
+````text
+<think>
+Let me reason about this. The score should reflect... maybe {draft: 6}?
+</think>
+Sure! Here's the result:
+```json
+{"score": 8, "reason": "clear"}
+```
+Hope that helps!
+````
+
+`JSON.parse` throws on all of that. `json-from-llm` returns `{ score: 8, reason: "clear" }`.
+
+```ts
+import { extractJson } from 'json-from-llm';
+
+const data = extractJson<{ score: number }>(modelOutput);
+```
+
+## Why
+
+- **Reasoning-model aware.** Strips `<think>` / `<thinking>` blocks first, so brace-laden reasoning (a real cause of `No object generated` failures with DeepSeek R1, Gemini 2.5 thinking, prompted Claude) never gets mistaken for the payload.
+- **Handles the real wrappers.** Markdown fences (`json` and bare ```), conversational prose before/after, and the JSON sitting bare in the text.
+- **String-aware, never corrupts.** The scanner and the trailing-comma repair both respect string contents — a `}` or `,` inside `"a string value"` is left alone.
+- **Conservative repair.** Removes trailing commas (the most common malformation); it will never rewrite your data.
+- **Two entry points.** `extractJson` throws on failure; `tryExtractJson` returns `{ found }`.
+- **Zero dependencies**, ESM + CJS, fully typed.
+
+## Install
+
+```sh
+npm install json-from-llm
+```
+
+## API
+
+### `extractJson<T>(text, options?) => T`
+
+Returns the extracted JSON value, or throws `JsonExtractionError` if none can be recovered.
+
+### `tryExtractJson<T>(text, options?) => { found: true, value: T } | { found: false }`
+
+The non-throwing variant.
+
+### Options
+
+```ts
+interface ExtractOptions {
+  repair?: boolean; // remove trailing commas (default true)
+  expect?: 'object' | 'array' | 'any'; // restrict the top-level type (default 'any')
+}
+```
+
+`expect` is handy when prose contains a stray array but you want the object:
+
+```ts
+extractJson('[1,2] then the answer {"a":1}', { expect: 'object' }); // { a: 1 }
+```
+
+### Algorithm
+
+1. Strip `<think>` / `<thinking>` / `<reasoning>` blocks.
+2. Prefer the contents of fenced `json (or bare `) code blocks.
+3. Otherwise scan for the first balanced `{…}` / `[…]` that parses, string-aware.
+4. If parsing fails, apply conservative repair (trailing commas) and retry.
+
+The low-level pieces (`stripReasoning`, `fencedBlocks`, `balancedSpans`, `removeTrailingCommas`) are exported too.
+
+## Related
+
+- [`tool-schema`](https://www.npmjs.com/package/tool-schema) — turn a JSON Schema into a provider tool/function schema (define the shape you then extract).
+- [`llm-sse`](https://www.npmjs.com/package/llm-sse) · [`llm-messages`](https://www.npmjs.com/package/llm-messages) · [`llm-errors`](https://www.npmjs.com/package/llm-errors) — the provider-portability suite.
+
+## License
+
+MIT © Sebastian Legarraga

package/dist/index.cjs

@@ -0,0 +1,211 @@
+"use strict";
+var __defProp = Object.defineProperty;
+var __getOwnPropDesc = Object.getOwnPropertyDescriptor;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __hasOwnProp = Object.prototype.hasOwnProperty;
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+var __copyProps = (to, from, except, desc) => {
+  if (from && typeof from === "object" || typeof from === "function") {
+    for (let key of __getOwnPropNames(from))
+      if (!__hasOwnProp.call(to, key) && key !== except)
+        __defProp(to, key, { get: () => from[key], enumerable: !(desc = __getOwnPropDesc(from, key)) || desc.enumerable });
+  }
+  return to;
+};
+var __toCommonJS = (mod) => __copyProps(__defProp({}, "__esModule", { value: true }), mod);
+
+// src/index.ts
+var index_exports = {};
+__export(index_exports, {
+  JsonExtractionError: () => JsonExtractionError,
+  balancedSpans: () => balancedSpans,
+  extractJson: () => extractJson,
+  fencedBlocks: () => fencedBlocks,
+  removeTrailingCommas: () => removeTrailingCommas,
+  stripReasoning: () => stripReasoning,
+  tryExtractJson: () => tryExtractJson
+});
+module.exports = __toCommonJS(index_exports);
+
+// src/repair.ts
+function removeTrailingCommas(json) {
+  let out = "";
+  let inString = false;
+  let escaped = false;
+  for (let i = 0; i < json.length; i++) {
+    const ch = json[i];
+    if (inString) {
+      out += ch;
+      if (escaped) {
+        escaped = false;
+      } else if (ch === "\\") {
+        escaped = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (ch === '"') {
+      inString = true;
+      out += ch;
+      continue;
+    }
+    if (ch === ",") {
+      let j = i + 1;
+      while (j < json.length && (json[j] === " " || json[j] === "\n" || json[j] === "\r" || json[j] === "	")) {
+        j++;
+      }
+      if (json[j] === "}" || json[j] === "]") {
+        continue;
+      }
+    }
+    out += ch;
+  }
+  return out;
+}
+
+// src/scan.ts
+function balancedSpans(text) {
+  const spans = [];
+  let i = 0;
+  while (i < text.length) {
+    const ch = text[i];
+    if (ch === "{" || ch === "[") {
+      const end = matchBalanced(text, i);
+      if (end !== -1) {
+        spans.push(text.slice(i, end));
+        i = end;
+        continue;
+      }
+    }
+    i++;
+  }
+  return spans;
+}
+function matchBalanced(text, start) {
+  let depth = 0;
+  let inString = false;
+  let escaped = false;
+  for (let i = start; i < text.length; i++) {
+    const ch = text[i];
+    if (inString) {
+      if (escaped) {
+        escaped = false;
+      } else if (ch === "\\") {
+        escaped = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (ch === '"') {
+      inString = true;
+    } else if (ch === "{" || ch === "[") {
+      depth++;
+    } else if (ch === "}" || ch === "]") {
+      depth--;
+      if (depth === 0) {
+        return i + 1;
+      }
+    }
+  }
+  return -1;
+}
+
+// src/strip.ts
+var REASONING_TAGS = /<(think|thinking|reasoning|thought)>[\s\S]*?<\/\1>/gi;
+function stripReasoning(text) {
+  return text.replace(REASONING_TAGS, "");
+}
+var FENCE = /```[^\S\n]*([a-zA-Z0-9_+-]*)[^\S\n]*\n?([\s\S]*?)```/g;
+function fencedBlocks(text) {
+  const blocks = [];
+  FENCE.lastIndex = 0;
+  let match;
+  while ((match = FENCE.exec(text)) !== null) {
+    const lang = match[1].toLowerCase();
+    const content = match[2].trim();
+    if (content.length > 0 && (lang === "" || lang.includes("json"))) {
+      blocks.push(content);
+    }
+  }
+  return blocks;
+}
+
+// src/types.ts
+var JsonExtractionError = class extends Error {
+  constructor(message, text) {
+    super(message);
+    this.text = text;
+    this.name = "JsonExtractionError";
+  }
+  text;
+};
+
+// src/extract.ts
+function parseCandidate(candidate, repair) {
+  try {
+    return { ok: true, value: JSON.parse(candidate) };
+  } catch {
+  }
+  if (repair) {
+    try {
+      return { ok: true, value: JSON.parse(removeTrailingCommas(candidate)) };
+    } catch {
+    }
+  }
+  return { ok: false };
+}
+function matchesExpect(value, expect) {
+  if (expect === "any") {
+    return true;
+  }
+  if (expect === "array") {
+    return Array.isArray(value);
+  }
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function tryExtractJson(text, options = {}) {
+  if (typeof text !== "string" || text.length === 0) {
+    return { found: false };
+  }
+  const repair = options.repair ?? true;
+  const expect = options.expect ?? "any";
+  const cleaned = stripReasoning(text);
+  const candidates = [];
+  for (const block of fencedBlocks(cleaned)) {
+    candidates.push(block, ...balancedSpans(block));
+  }
+  candidates.push(...balancedSpans(cleaned));
+  for (const candidate of candidates) {
+    const parsed = parseCandidate(candidate, repair);
+    if (parsed.ok && matchesExpect(parsed.value, expect)) {
+      return { found: true, value: parsed.value };
+    }
+  }
+  return { found: false };
+}
+function extractJson(text, options = {}) {
+  const result = tryExtractJson(text, options);
+  if (!result.found) {
+    throw new JsonExtractionError(
+      "No JSON value could be extracted from the text.",
+      text
+    );
+  }
+  return result.value;
+}
+// Annotate the CommonJS export names for ESM import in node:
+0 && (module.exports = {
+  JsonExtractionError,
+  balancedSpans,
+  extractJson,
+  fencedBlocks,
+  removeTrailingCommas,
+  stripReasoning,
+  tryExtractJson
+});
+//# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/index.ts","../src/repair.ts","../src/scan.ts","../src/strip.ts","../src/types.ts","../src/extract.ts"],"sourcesContent":["export { extractJson, tryExtractJson } from './extract.ts';\nexport { stripReasoning, fencedBlocks } from './strip.ts';\nexport { balancedSpans } from './scan.ts';\nexport { removeTrailingCommas } from './repair.ts';\nexport { JsonExtractionError } from './types.ts';\nexport type { ExtractOptions, ExtractResult } from './types.ts';\n","/**\n * Remove trailing commas (`{\"a\":1,}` → `{\"a\":1}`, `[1,2,]` → `[1,2]`), which\n * models emit frequently. String-aware: a comma inside a string value is never\n * touched, so this can only ever fix structure, never corrupt content.\n */\nexport function removeTrailingCommas(json: string): string {\n  let out = '';\n  let inString = false;\n  let escaped = false;\n\n  for (let i = 0; i < json.length; i++) {\n    const ch = json[i];\n\n    if (inString) {\n      out += ch;\n      if (escaped) {\n        escaped = false;\n      } else if (ch === '\\\\') {\n        escaped = true;\n      } else if (ch === '\"') {\n        inString = false;\n      }\n      continue;\n    }\n\n    if (ch === '\"') {\n      inString = true;\n      out += ch;\n      continue;\n    }\n\n    if (ch === ',') {\n      let j = i + 1;\n      while (\n        j < json.length &&\n        (json[j] === ' ' ||\n          json[j] === '\\n' ||\n          json[j] === '\\r' ||\n          json[j] === '\\t')\n      ) {\n        j++;\n      }\n      if (json[j] === '}' || json[j] === ']') {\n        continue; // drop the trailing comma\n      }\n    }\n\n    out += ch;\n  }\n\n  return out;\n}\n","/**\n * Find the substrings of complete, balanced JSON objects/arrays in `text`,\n * in document order. String-aware: braces and brackets inside JSON strings do\n * not affect nesting, so prose like `\"the } char\"` won't break the scan.\n */\nexport function balancedSpans(text: string): string[] {\n  const spans: string[] = [];\n  let i = 0;\n  while (i < text.length) {\n    const ch = text[i];\n    if (ch === '{' || ch === '[') {\n      const end = matchBalanced(text, i);\n      if (end !== -1) {\n        spans.push(text.slice(i, end));\n        i = end;\n        continue;\n      }\n    }\n    i++;\n  }\n  return spans;\n}\n\n/** Return the index just past the balanced value starting at `start`, or -1. */\nfunction matchBalanced(text: string, start: number): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n\n  for (let i = start; i < text.length; i++) {\n    const ch = text[i];\n\n    if (inString) {\n      if (escaped) {\n        escaped = false;\n      } else if (ch === '\\\\') {\n        escaped = true;\n      } else if (ch === '\"') {\n        inString = false;\n      }\n      continue;\n    }\n\n    if (ch === '\"') {\n      inString = true;\n    } else if (ch === '{' || ch === '[') {\n      depth++;\n    } else if (ch === '}' || ch === ']') {\n      depth--;\n      if (depth === 0) {\n        return i + 1;\n      }\n    }\n  }\n\n  return -1;\n}\n","/**\n * Remove model \"thinking\" / reasoning blocks. Reasoning models (DeepSeek R1,\n * Qwen, and prompted Claude/Gemini setups) emit `<think>…</think>` or\n * `<thinking>…</thinking>` before the answer, and that text frequently contains\n * brace-laden prose that would otherwise be mistaken for the payload.\n */\nconst REASONING_TAGS = /<(think|thinking|reasoning|thought)>[\\s\\S]*?<\\/\\1>/gi;\n\nexport function stripReasoning(text: string): string {\n  return text.replace(REASONING_TAGS, '');\n}\n\n/**\n * Return the inner contents of fenced code blocks that could hold JSON: blocks\n * tagged ```json / ```jsonc / ```json5, or untagged ``` blocks. Other languages\n * (```python, ```ts) are skipped — they won't contain the answer JSON.\n */\nconst FENCE = /```[^\\S\\n]*([a-zA-Z0-9_+-]*)[^\\S\\n]*\\n?([\\s\\S]*?)```/g;\n\nexport function fencedBlocks(text: string): string[] {\n  const blocks: string[] = [];\n  FENCE.lastIndex = 0;\n  let match: RegExpExecArray | null;\n  while ((match = FENCE.exec(text)) !== null) {\n    const lang = match[1].toLowerCase();\n    const content = match[2].trim();\n    if (content.length > 0 && (lang === '' || lang.includes('json'))) {\n      blocks.push(content);\n    }\n  }\n  return blocks;\n}\n","/** Options for {@link extractJson} and {@link tryExtractJson}. */\nexport interface ExtractOptions {\n  /**\n   * Apply conservative, string-aware repairs before parsing — currently the\n   * removal of trailing commas, which models emit often. Never rewrites string\n   * contents. Default `true`.\n   */\n  repair?: boolean;\n  /**\n   * Restrict which top-level JSON value to accept: an `'object'`, an `'array'`,\n   * or `'any'` (the default).\n   */\n  expect?: 'object' | 'array' | 'any';\n}\n\n/** The result of {@link tryExtractJson}. */\nexport type ExtractResult<T> =\n  | { found: true; value: T }\n  | { found: false; value?: undefined };\n\n/** Thrown by {@link extractJson} when no JSON value can be recovered. */\nexport class JsonExtractionError extends Error {\n  constructor(\n    message: string,\n    /** The original text that no JSON could be extracted from. */\n    public readonly text: string,\n  ) {\n    super(message);\n    this.name = 'JsonExtractionError';\n  }\n}\n","import { removeTrailingCommas } from './repair.ts';\nimport { balancedSpans } from './scan.ts';\nimport { fencedBlocks, stripReasoning } from './strip.ts';\nimport { JsonExtractionError } from './types.ts';\nimport type { ExtractOptions, ExtractResult } from './types.ts';\n\nfunction parseCandidate(\n  candidate: string,\n  repair: boolean,\n): { ok: true; value: unknown } | { ok: false } {\n  try {\n    return { ok: true, value: JSON.parse(candidate) };\n  } catch {\n    // fall through to repair\n  }\n  if (repair) {\n    try {\n      return { ok: true, value: JSON.parse(removeTrailingCommas(candidate)) };\n    } catch {\n      // unrecoverable\n    }\n  }\n  return { ok: false };\n}\n\nfunction matchesExpect(\n  value: unknown,\n  expect: 'object' | 'array' | 'any',\n): boolean {\n  if (expect === 'any') {\n    return true;\n  }\n  if (expect === 'array') {\n    return Array.isArray(value);\n  }\n  return typeof value === 'object' && value !== null && !Array.isArray(value);\n}\n\n/**\n * Extract a JSON value from LLM output without throwing.\n *\n * Strips `<think>` / `<thinking>` reasoning blocks, prefers fenced ```json\n * code blocks, then scans for the first balanced object/array that parses\n * (applying conservative repair). Returns `{ found: false }` if nothing parses.\n *\n * @example\n * ```ts\n * const r = tryExtractJson<{ score: number }>('<think>...</think>\\n{\"score\":7}');\n * if (r.found) console.log(r.value.score); // 7\n * ```\n */\nexport function tryExtractJson<T = unknown>(\n  text: string,\n  options: ExtractOptions = {},\n): ExtractResult<T> {\n  if (typeof text !== 'string' || text.length === 0) {\n    return { found: false };\n  }\n\n  const repair = options.repair ?? true;\n  const expect = options.expect ?? 'any';\n  const cleaned = stripReasoning(text);\n\n  // Candidate substrings, highest confidence first: fenced blocks (and any\n  // balanced values inside them), then balanced values anywhere in the text.\n  const candidates: string[] = [];\n  for (const block of fencedBlocks(cleaned)) {\n    candidates.push(block, ...balancedSpans(block));\n  }\n  candidates.push(...balancedSpans(cleaned));\n\n  for (const candidate of candidates) {\n    const parsed = parseCandidate(candidate, repair);\n    if (parsed.ok && matchesExpect(parsed.value, expect)) {\n      return { found: true, value: parsed.value as T };\n    }\n  }\n  return { found: false };\n}\n\n/**\n * Extract a JSON value from LLM output, throwing {@link JsonExtractionError}\n * if none can be recovered. See {@link tryExtractJson} for the algorithm.\n */\nexport function extractJson<T = unknown>(\n  text: string,\n  options: ExtractOptions = {},\n): T {\n  const result = tryExtractJson<T>(text, options);\n  if (!result.found) {\n    throw new JsonExtractionError(\n      'No JSON value could be extracted from the text.',\n      text,\n    );\n  }\n  return result.value;\n}\n"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;;ACKO,SAAS,qBAAqB,MAAsB;AACzD,MAAI,MAAM;AACV,MAAI,WAAW;AACf,MAAI,UAAU;AAEd,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,KAAK,KAAK,CAAC;AAEjB,QAAI,UAAU;AACZ,aAAO;AACP,UAAI,SAAS;AACX,kBAAU;AAAA,MACZ,WAAW,OAAO,MAAM;AACtB,kBAAU;AAAA,MACZ,WAAW,OAAO,KAAK;AACrB,mBAAW;AAAA,MACb;AACA;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,iBAAW;AACX,aAAO;AACP;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,UAAI,IAAI,IAAI;AACZ,aACE,IAAI,KAAK,WACR,KAAK,CAAC,MAAM,OACX,KAAK,CAAC,MAAM,QACZ,KAAK,CAAC,MAAM,QACZ,KAAK,CAAC,MAAM,MACd;AACA;AAAA,MACF;AACA,UAAI,KAAK,CAAC,MAAM,OAAO,KAAK,CAAC,MAAM,KAAK;AACtC;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,SAAO;AACT;;;AC9CO,SAAS,cAAc,MAAwB;AACpD,QAAM,QAAkB,CAAC;AACzB,MAAI,IAAI;AACR,SAAO,IAAI,KAAK,QAAQ;AACtB,UAAM,KAAK,KAAK,CAAC;AACjB,QAAI,OAAO,OAAO,OAAO,KAAK;AAC5B,YAAM,MAAM,cAAc,MAAM,CAAC;AACjC,UAAI,QAAQ,IAAI;AACd,cAAM,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAC7B,YAAI;AACJ;AAAA,MACF;AAAA,IACF;AACA;AAAA,EACF;AACA,SAAO;AACT;AAGA,SAAS,cAAc,MAAc,OAAuB;AAC1D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AAEd,WAAS,IAAI,OAAO,IAAI,KAAK,QAAQ,KAAK;AACxC,UAAM,KAAK,KAAK,CAAC;AAEjB,QAAI,UAAU;AACZ,UAAI,SAAS;AACX,kBAAU;AAAA,MACZ,WAAW,OAAO,MAAM;AACtB,kBAAU;AAAA,MACZ,WAAW,OAAO,KAAK;AACrB,mBAAW;AAAA,MACb;AACA;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,iBAAW;AAAA,IACb,WAAW,OAAO,OAAO,OAAO,KAAK;AACnC;AAAA,IACF,WAAW,OAAO,OAAO,OAAO,KAAK;AACnC;AACA,UAAI,UAAU,GAAG;AACf,eAAO,IAAI;AAAA,MACb;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;;;AClDA,IAAM,iBAAiB;AAEhB,SAAS,eAAe,MAAsB;AACnD,SAAO,KAAK,QAAQ,gBAAgB,EAAE;AACxC;AAOA,IAAM,QAAQ;AAEP,SAAS,aAAa,MAAwB;AACnD,QAAM,SAAmB,CAAC;AAC1B,QAAM,YAAY;AAClB,MAAI;AACJ,UAAQ,QAAQ,MAAM,KAAK,IAAI,OAAO,MAAM;AAC1C,UAAM,OAAO,MAAM,CAAC,EAAE,YAAY;AAClC,UAAM,UAAU,MAAM,CAAC,EAAE,KAAK;AAC9B,QAAI,QAAQ,SAAS,MAAM,SAAS,MAAM,KAAK,SAAS,MAAM,IAAI;AAChE,aAAO,KAAK,OAAO;AAAA,IACrB;AAAA,EACF;AACA,SAAO;AACT;;;ACVO,IAAM,sBAAN,cAAkC,MAAM;AAAA,EAC7C,YACE,SAEgB,MAChB;AACA,UAAM,OAAO;AAFG;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EAJkB;AAKpB;;;ACxBA,SAAS,eACP,WACA,QAC8C;AAC9C,MAAI;AACF,WAAO,EAAE,IAAI,MAAM,OAAO,KAAK,MAAM,SAAS,EAAE;AAAA,EAClD,QAAQ;AAAA,EAER;AACA,MAAI,QAAQ;AACV,QAAI;AACF,aAAO,EAAE,IAAI,MAAM,OAAO,KAAK,MAAM,qBAAqB,SAAS,CAAC,EAAE;AAAA,IACxE,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO,EAAE,IAAI,MAAM;AACrB;AAEA,SAAS,cACP,OACA,QACS;AACT,MAAI,WAAW,OAAO;AACpB,WAAO;AAAA,EACT;AACA,MAAI,WAAW,SAAS;AACtB,WAAO,MAAM,QAAQ,KAAK;AAAA,EAC5B;AACA,SAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAeO,SAAS,eACd,MACA,UAA0B,CAAC,GACT;AAClB,MAAI,OAAO,SAAS,YAAY,KAAK,WAAW,GAAG;AACjD,WAAO,EAAE,OAAO,MAAM;AAAA,EACxB;AAEA,QAAM,SAAS,QAAQ,UAAU;AACjC,QAAM,SAAS,QAAQ,UAAU;AACjC,QAAM,UAAU,eAAe,IAAI;AAInC,QAAM,aAAuB,CAAC;AAC9B,aAAW,SAAS,aAAa,OAAO,GAAG;AACzC,eAAW,KAAK,OAAO,GAAG,cAAc,KAAK,CAAC;AAAA,EAChD;AACA,aAAW,KAAK,GAAG,cAAc,OAAO,CAAC;AAEzC,aAAW,aAAa,YAAY;AAClC,UAAM,SAAS,eAAe,WAAW,MAAM;AAC/C,QAAI,OAAO,MAAM,cAAc,OAAO,OAAO,MAAM,GAAG;AACpD,aAAO,EAAE,OAAO,MAAM,OAAO,OAAO,MAAW;AAAA,IACjD;AAAA,EACF;AACA,SAAO,EAAE,OAAO,MAAM;AACxB;AAMO,SAAS,YACd,MACA,UAA0B,CAAC,GACxB;AACH,QAAM,SAAS,eAAkB,MAAM,OAAO;AAC9C,MAAI,CAAC,OAAO,OAAO;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,SAAO,OAAO;AAChB;","names":[]}

package/dist/index.d.cts

@@ -0,0 +1,69 @@
+/** Options for {@link extractJson} and {@link tryExtractJson}. */
+interface ExtractOptions {
+    /**
+     * Apply conservative, string-aware repairs before parsing — currently the
+     * removal of trailing commas, which models emit often. Never rewrites string
+     * contents. Default `true`.
+     */
+    repair?: boolean;
+    /**
+     * Restrict which top-level JSON value to accept: an `'object'`, an `'array'`,
+     * or `'any'` (the default).
+     */
+    expect?: 'object' | 'array' | 'any';
+}
+/** The result of {@link tryExtractJson}. */
+type ExtractResult<T> = {
+    found: true;
+    value: T;
+} | {
+    found: false;
+    value?: undefined;
+};
+/** Thrown by {@link extractJson} when no JSON value can be recovered. */
+declare class JsonExtractionError extends Error {
+    /** The original text that no JSON could be extracted from. */
+    readonly text: string;
+    constructor(message: string, 
+    /** The original text that no JSON could be extracted from. */
+    text: string);
+}
+
+/**
+ * Extract a JSON value from LLM output without throwing.
+ *
+ * Strips `<think>` / `<thinking>` reasoning blocks, prefers fenced ```json
+ * code blocks, then scans for the first balanced object/array that parses
+ * (applying conservative repair). Returns `{ found: false }` if nothing parses.
+ *
+ * @example
+ * ```ts
+ * const r = tryExtractJson<{ score: number }>('<think>...</think>\n{"score":7}');
+ * if (r.found) console.log(r.value.score); // 7
+ * ```
+ */
+declare function tryExtractJson<T = unknown>(text: string, options?: ExtractOptions): ExtractResult<T>;
+/**
+ * Extract a JSON value from LLM output, throwing {@link JsonExtractionError}
+ * if none can be recovered. See {@link tryExtractJson} for the algorithm.
+ */
+declare function extractJson<T = unknown>(text: string, options?: ExtractOptions): T;
+
+declare function stripReasoning(text: string): string;
+declare function fencedBlocks(text: string): string[];
+
+/**
+ * Find the substrings of complete, balanced JSON objects/arrays in `text`,
+ * in document order. String-aware: braces and brackets inside JSON strings do
+ * not affect nesting, so prose like `"the } char"` won't break the scan.
+ */
+declare function balancedSpans(text: string): string[];
+
+/**
+ * Remove trailing commas (`{"a":1,}` → `{"a":1}`, `[1,2,]` → `[1,2]`), which
+ * models emit frequently. String-aware: a comma inside a string value is never
+ * touched, so this can only ever fix structure, never corrupt content.
+ */
+declare function removeTrailingCommas(json: string): string;
+
+export { type ExtractOptions, type ExtractResult, JsonExtractionError, balancedSpans, extractJson, fencedBlocks, removeTrailingCommas, stripReasoning, tryExtractJson };

package/dist/index.d.ts

@@ -0,0 +1,69 @@
+/** Options for {@link extractJson} and {@link tryExtractJson}. */
+interface ExtractOptions {
+    /**
+     * Apply conservative, string-aware repairs before parsing — currently the
+     * removal of trailing commas, which models emit often. Never rewrites string
+     * contents. Default `true`.
+     */
+    repair?: boolean;
+    /**
+     * Restrict which top-level JSON value to accept: an `'object'`, an `'array'`,
+     * or `'any'` (the default).
+     */
+    expect?: 'object' | 'array' | 'any';
+}
+/** The result of {@link tryExtractJson}. */
+type ExtractResult<T> = {
+    found: true;
+    value: T;
+} | {
+    found: false;
+    value?: undefined;
+};
+/** Thrown by {@link extractJson} when no JSON value can be recovered. */
+declare class JsonExtractionError extends Error {
+    /** The original text that no JSON could be extracted from. */
+    readonly text: string;
+    constructor(message: string, 
+    /** The original text that no JSON could be extracted from. */
+    text: string);
+}
+
+/**
+ * Extract a JSON value from LLM output without throwing.
+ *
+ * Strips `<think>` / `<thinking>` reasoning blocks, prefers fenced ```json
+ * code blocks, then scans for the first balanced object/array that parses
+ * (applying conservative repair). Returns `{ found: false }` if nothing parses.
+ *
+ * @example
+ * ```ts
+ * const r = tryExtractJson<{ score: number }>('<think>...</think>\n{"score":7}');
+ * if (r.found) console.log(r.value.score); // 7
+ * ```
+ */
+declare function tryExtractJson<T = unknown>(text: string, options?: ExtractOptions): ExtractResult<T>;
+/**
+ * Extract a JSON value from LLM output, throwing {@link JsonExtractionError}
+ * if none can be recovered. See {@link tryExtractJson} for the algorithm.
+ */
+declare function extractJson<T = unknown>(text: string, options?: ExtractOptions): T;
+
+declare function stripReasoning(text: string): string;
+declare function fencedBlocks(text: string): string[];
+
+/**
+ * Find the substrings of complete, balanced JSON objects/arrays in `text`,
+ * in document order. String-aware: braces and brackets inside JSON strings do
+ * not affect nesting, so prose like `"the } char"` won't break the scan.
+ */
+declare function balancedSpans(text: string): string[];
+
+/**
+ * Remove trailing commas (`{"a":1,}` → `{"a":1}`, `[1,2,]` → `[1,2]`), which
+ * models emit frequently. String-aware: a comma inside a string value is never
+ * touched, so this can only ever fix structure, never corrupt content.
+ */
+declare function removeTrailingCommas(json: string): string;
+
+export { type ExtractOptions, type ExtractResult, JsonExtractionError, balancedSpans, extractJson, fencedBlocks, removeTrailingCommas, stripReasoning, tryExtractJson };

package/dist/index.js

@@ -0,0 +1,178 @@
+// src/repair.ts
+function removeTrailingCommas(json) {
+  let out = "";
+  let inString = false;
+  let escaped = false;
+  for (let i = 0; i < json.length; i++) {
+    const ch = json[i];
+    if (inString) {
+      out += ch;
+      if (escaped) {
+        escaped = false;
+      } else if (ch === "\\") {
+        escaped = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (ch === '"') {
+      inString = true;
+      out += ch;
+      continue;
+    }
+    if (ch === ",") {
+      let j = i + 1;
+      while (j < json.length && (json[j] === " " || json[j] === "\n" || json[j] === "\r" || json[j] === "	")) {
+        j++;
+      }
+      if (json[j] === "}" || json[j] === "]") {
+        continue;
+      }
+    }
+    out += ch;
+  }
+  return out;
+}
+
+// src/scan.ts
+function balancedSpans(text) {
+  const spans = [];
+  let i = 0;
+  while (i < text.length) {
+    const ch = text[i];
+    if (ch === "{" || ch === "[") {
+      const end = matchBalanced(text, i);
+      if (end !== -1) {
+        spans.push(text.slice(i, end));
+        i = end;
+        continue;
+      }
+    }
+    i++;
+  }
+  return spans;
+}
+function matchBalanced(text, start) {
+  let depth = 0;
+  let inString = false;
+  let escaped = false;
+  for (let i = start; i < text.length; i++) {
+    const ch = text[i];
+    if (inString) {
+      if (escaped) {
+        escaped = false;
+      } else if (ch === "\\") {
+        escaped = true;
+      } else if (ch === '"') {
+        inString = false;
+      }
+      continue;
+    }
+    if (ch === '"') {
+      inString = true;
+    } else if (ch === "{" || ch === "[") {
+      depth++;
+    } else if (ch === "}" || ch === "]") {
+      depth--;
+      if (depth === 0) {
+        return i + 1;
+      }
+    }
+  }
+  return -1;
+}
+
+// src/strip.ts
+var REASONING_TAGS = /<(think|thinking|reasoning|thought)>[\s\S]*?<\/\1>/gi;
+function stripReasoning(text) {
+  return text.replace(REASONING_TAGS, "");
+}
+var FENCE = /```[^\S\n]*([a-zA-Z0-9_+-]*)[^\S\n]*\n?([\s\S]*?)```/g;
+function fencedBlocks(text) {
+  const blocks = [];
+  FENCE.lastIndex = 0;
+  let match;
+  while ((match = FENCE.exec(text)) !== null) {
+    const lang = match[1].toLowerCase();
+    const content = match[2].trim();
+    if (content.length > 0 && (lang === "" || lang.includes("json"))) {
+      blocks.push(content);
+    }
+  }
+  return blocks;
+}
+
+// src/types.ts
+var JsonExtractionError = class extends Error {
+  constructor(message, text) {
+    super(message);
+    this.text = text;
+    this.name = "JsonExtractionError";
+  }
+  text;
+};
+
+// src/extract.ts
+function parseCandidate(candidate, repair) {
+  try {
+    return { ok: true, value: JSON.parse(candidate) };
+  } catch {
+  }
+  if (repair) {
+    try {
+      return { ok: true, value: JSON.parse(removeTrailingCommas(candidate)) };
+    } catch {
+    }
+  }
+  return { ok: false };
+}
+function matchesExpect(value, expect) {
+  if (expect === "any") {
+    return true;
+  }
+  if (expect === "array") {
+    return Array.isArray(value);
+  }
+  return typeof value === "object" && value !== null && !Array.isArray(value);
+}
+function tryExtractJson(text, options = {}) {
+  if (typeof text !== "string" || text.length === 0) {
+    return { found: false };
+  }
+  const repair = options.repair ?? true;
+  const expect = options.expect ?? "any";
+  const cleaned = stripReasoning(text);
+  const candidates = [];
+  for (const block of fencedBlocks(cleaned)) {
+    candidates.push(block, ...balancedSpans(block));
+  }
+  candidates.push(...balancedSpans(cleaned));
+  for (const candidate of candidates) {
+    const parsed = parseCandidate(candidate, repair);
+    if (parsed.ok && matchesExpect(parsed.value, expect)) {
+      return { found: true, value: parsed.value };
+    }
+  }
+  return { found: false };
+}
+function extractJson(text, options = {}) {
+  const result = tryExtractJson(text, options);
+  if (!result.found) {
+    throw new JsonExtractionError(
+      "No JSON value could be extracted from the text.",
+      text
+    );
+  }
+  return result.value;
+}
+export {
+  JsonExtractionError,
+  balancedSpans,
+  extractJson,
+  fencedBlocks,
+  removeTrailingCommas,
+  stripReasoning,
+  tryExtractJson
+};
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/repair.ts","../src/scan.ts","../src/strip.ts","../src/types.ts","../src/extract.ts"],"sourcesContent":["/**\n * Remove trailing commas (`{\"a\":1,}` → `{\"a\":1}`, `[1,2,]` → `[1,2]`), which\n * models emit frequently. String-aware: a comma inside a string value is never\n * touched, so this can only ever fix structure, never corrupt content.\n */\nexport function removeTrailingCommas(json: string): string {\n  let out = '';\n  let inString = false;\n  let escaped = false;\n\n  for (let i = 0; i < json.length; i++) {\n    const ch = json[i];\n\n    if (inString) {\n      out += ch;\n      if (escaped) {\n        escaped = false;\n      } else if (ch === '\\\\') {\n        escaped = true;\n      } else if (ch === '\"') {\n        inString = false;\n      }\n      continue;\n    }\n\n    if (ch === '\"') {\n      inString = true;\n      out += ch;\n      continue;\n    }\n\n    if (ch === ',') {\n      let j = i + 1;\n      while (\n        j < json.length &&\n        (json[j] === ' ' ||\n          json[j] === '\\n' ||\n          json[j] === '\\r' ||\n          json[j] === '\\t')\n      ) {\n        j++;\n      }\n      if (json[j] === '}' || json[j] === ']') {\n        continue; // drop the trailing comma\n      }\n    }\n\n    out += ch;\n  }\n\n  return out;\n}\n","/**\n * Find the substrings of complete, balanced JSON objects/arrays in `text`,\n * in document order. String-aware: braces and brackets inside JSON strings do\n * not affect nesting, so prose like `\"the } char\"` won't break the scan.\n */\nexport function balancedSpans(text: string): string[] {\n  const spans: string[] = [];\n  let i = 0;\n  while (i < text.length) {\n    const ch = text[i];\n    if (ch === '{' || ch === '[') {\n      const end = matchBalanced(text, i);\n      if (end !== -1) {\n        spans.push(text.slice(i, end));\n        i = end;\n        continue;\n      }\n    }\n    i++;\n  }\n  return spans;\n}\n\n/** Return the index just past the balanced value starting at `start`, or -1. */\nfunction matchBalanced(text: string, start: number): number {\n  let depth = 0;\n  let inString = false;\n  let escaped = false;\n\n  for (let i = start; i < text.length; i++) {\n    const ch = text[i];\n\n    if (inString) {\n      if (escaped) {\n        escaped = false;\n      } else if (ch === '\\\\') {\n        escaped = true;\n      } else if (ch === '\"') {\n        inString = false;\n      }\n      continue;\n    }\n\n    if (ch === '\"') {\n      inString = true;\n    } else if (ch === '{' || ch === '[') {\n      depth++;\n    } else if (ch === '}' || ch === ']') {\n      depth--;\n      if (depth === 0) {\n        return i + 1;\n      }\n    }\n  }\n\n  return -1;\n}\n","/**\n * Remove model \"thinking\" / reasoning blocks. Reasoning models (DeepSeek R1,\n * Qwen, and prompted Claude/Gemini setups) emit `<think>…</think>` or\n * `<thinking>…</thinking>` before the answer, and that text frequently contains\n * brace-laden prose that would otherwise be mistaken for the payload.\n */\nconst REASONING_TAGS = /<(think|thinking|reasoning|thought)>[\\s\\S]*?<\\/\\1>/gi;\n\nexport function stripReasoning(text: string): string {\n  return text.replace(REASONING_TAGS, '');\n}\n\n/**\n * Return the inner contents of fenced code blocks that could hold JSON: blocks\n * tagged ```json / ```jsonc / ```json5, or untagged ``` blocks. Other languages\n * (```python, ```ts) are skipped — they won't contain the answer JSON.\n */\nconst FENCE = /```[^\\S\\n]*([a-zA-Z0-9_+-]*)[^\\S\\n]*\\n?([\\s\\S]*?)```/g;\n\nexport function fencedBlocks(text: string): string[] {\n  const blocks: string[] = [];\n  FENCE.lastIndex = 0;\n  let match: RegExpExecArray | null;\n  while ((match = FENCE.exec(text)) !== null) {\n    const lang = match[1].toLowerCase();\n    const content = match[2].trim();\n    if (content.length > 0 && (lang === '' || lang.includes('json'))) {\n      blocks.push(content);\n    }\n  }\n  return blocks;\n}\n","/** Options for {@link extractJson} and {@link tryExtractJson}. */\nexport interface ExtractOptions {\n  /**\n   * Apply conservative, string-aware repairs before parsing — currently the\n   * removal of trailing commas, which models emit often. Never rewrites string\n   * contents. Default `true`.\n   */\n  repair?: boolean;\n  /**\n   * Restrict which top-level JSON value to accept: an `'object'`, an `'array'`,\n   * or `'any'` (the default).\n   */\n  expect?: 'object' | 'array' | 'any';\n}\n\n/** The result of {@link tryExtractJson}. */\nexport type ExtractResult<T> =\n  | { found: true; value: T }\n  | { found: false; value?: undefined };\n\n/** Thrown by {@link extractJson} when no JSON value can be recovered. */\nexport class JsonExtractionError extends Error {\n  constructor(\n    message: string,\n    /** The original text that no JSON could be extracted from. */\n    public readonly text: string,\n  ) {\n    super(message);\n    this.name = 'JsonExtractionError';\n  }\n}\n","import { removeTrailingCommas } from './repair.ts';\nimport { balancedSpans } from './scan.ts';\nimport { fencedBlocks, stripReasoning } from './strip.ts';\nimport { JsonExtractionError } from './types.ts';\nimport type { ExtractOptions, ExtractResult } from './types.ts';\n\nfunction parseCandidate(\n  candidate: string,\n  repair: boolean,\n): { ok: true; value: unknown } | { ok: false } {\n  try {\n    return { ok: true, value: JSON.parse(candidate) };\n  } catch {\n    // fall through to repair\n  }\n  if (repair) {\n    try {\n      return { ok: true, value: JSON.parse(removeTrailingCommas(candidate)) };\n    } catch {\n      // unrecoverable\n    }\n  }\n  return { ok: false };\n}\n\nfunction matchesExpect(\n  value: unknown,\n  expect: 'object' | 'array' | 'any',\n): boolean {\n  if (expect === 'any') {\n    return true;\n  }\n  if (expect === 'array') {\n    return Array.isArray(value);\n  }\n  return typeof value === 'object' && value !== null && !Array.isArray(value);\n}\n\n/**\n * Extract a JSON value from LLM output without throwing.\n *\n * Strips `<think>` / `<thinking>` reasoning blocks, prefers fenced ```json\n * code blocks, then scans for the first balanced object/array that parses\n * (applying conservative repair). Returns `{ found: false }` if nothing parses.\n *\n * @example\n * ```ts\n * const r = tryExtractJson<{ score: number }>('<think>...</think>\\n{\"score\":7}');\n * if (r.found) console.log(r.value.score); // 7\n * ```\n */\nexport function tryExtractJson<T = unknown>(\n  text: string,\n  options: ExtractOptions = {},\n): ExtractResult<T> {\n  if (typeof text !== 'string' || text.length === 0) {\n    return { found: false };\n  }\n\n  const repair = options.repair ?? true;\n  const expect = options.expect ?? 'any';\n  const cleaned = stripReasoning(text);\n\n  // Candidate substrings, highest confidence first: fenced blocks (and any\n  // balanced values inside them), then balanced values anywhere in the text.\n  const candidates: string[] = [];\n  for (const block of fencedBlocks(cleaned)) {\n    candidates.push(block, ...balancedSpans(block));\n  }\n  candidates.push(...balancedSpans(cleaned));\n\n  for (const candidate of candidates) {\n    const parsed = parseCandidate(candidate, repair);\n    if (parsed.ok && matchesExpect(parsed.value, expect)) {\n      return { found: true, value: parsed.value as T };\n    }\n  }\n  return { found: false };\n}\n\n/**\n * Extract a JSON value from LLM output, throwing {@link JsonExtractionError}\n * if none can be recovered. See {@link tryExtractJson} for the algorithm.\n */\nexport function extractJson<T = unknown>(\n  text: string,\n  options: ExtractOptions = {},\n): T {\n  const result = tryExtractJson<T>(text, options);\n  if (!result.found) {\n    throw new JsonExtractionError(\n      'No JSON value could be extracted from the text.',\n      text,\n    );\n  }\n  return result.value;\n}\n"],"mappings":";AAKO,SAAS,qBAAqB,MAAsB;AACzD,MAAI,MAAM;AACV,MAAI,WAAW;AACf,MAAI,UAAU;AAEd,WAAS,IAAI,GAAG,IAAI,KAAK,QAAQ,KAAK;AACpC,UAAM,KAAK,KAAK,CAAC;AAEjB,QAAI,UAAU;AACZ,aAAO;AACP,UAAI,SAAS;AACX,kBAAU;AAAA,MACZ,WAAW,OAAO,MAAM;AACtB,kBAAU;AAAA,MACZ,WAAW,OAAO,KAAK;AACrB,mBAAW;AAAA,MACb;AACA;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,iBAAW;AACX,aAAO;AACP;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,UAAI,IAAI,IAAI;AACZ,aACE,IAAI,KAAK,WACR,KAAK,CAAC,MAAM,OACX,KAAK,CAAC,MAAM,QACZ,KAAK,CAAC,MAAM,QACZ,KAAK,CAAC,MAAM,MACd;AACA;AAAA,MACF;AACA,UAAI,KAAK,CAAC,MAAM,OAAO,KAAK,CAAC,MAAM,KAAK;AACtC;AAAA,MACF;AAAA,IACF;AAEA,WAAO;AAAA,EACT;AAEA,SAAO;AACT;;;AC9CO,SAAS,cAAc,MAAwB;AACpD,QAAM,QAAkB,CAAC;AACzB,MAAI,IAAI;AACR,SAAO,IAAI,KAAK,QAAQ;AACtB,UAAM,KAAK,KAAK,CAAC;AACjB,QAAI,OAAO,OAAO,OAAO,KAAK;AAC5B,YAAM,MAAM,cAAc,MAAM,CAAC;AACjC,UAAI,QAAQ,IAAI;AACd,cAAM,KAAK,KAAK,MAAM,GAAG,GAAG,CAAC;AAC7B,YAAI;AACJ;AAAA,MACF;AAAA,IACF;AACA;AAAA,EACF;AACA,SAAO;AACT;AAGA,SAAS,cAAc,MAAc,OAAuB;AAC1D,MAAI,QAAQ;AACZ,MAAI,WAAW;AACf,MAAI,UAAU;AAEd,WAAS,IAAI,OAAO,IAAI,KAAK,QAAQ,KAAK;AACxC,UAAM,KAAK,KAAK,CAAC;AAEjB,QAAI,UAAU;AACZ,UAAI,SAAS;AACX,kBAAU;AAAA,MACZ,WAAW,OAAO,MAAM;AACtB,kBAAU;AAAA,MACZ,WAAW,OAAO,KAAK;AACrB,mBAAW;AAAA,MACb;AACA;AAAA,IACF;AAEA,QAAI,OAAO,KAAK;AACd,iBAAW;AAAA,IACb,WAAW,OAAO,OAAO,OAAO,KAAK;AACnC;AAAA,IACF,WAAW,OAAO,OAAO,OAAO,KAAK;AACnC;AACA,UAAI,UAAU,GAAG;AACf,eAAO,IAAI;AAAA,MACb;AAAA,IACF;AAAA,EACF;AAEA,SAAO;AACT;;;AClDA,IAAM,iBAAiB;AAEhB,SAAS,eAAe,MAAsB;AACnD,SAAO,KAAK,QAAQ,gBAAgB,EAAE;AACxC;AAOA,IAAM,QAAQ;AAEP,SAAS,aAAa,MAAwB;AACnD,QAAM,SAAmB,CAAC;AAC1B,QAAM,YAAY;AAClB,MAAI;AACJ,UAAQ,QAAQ,MAAM,KAAK,IAAI,OAAO,MAAM;AAC1C,UAAM,OAAO,MAAM,CAAC,EAAE,YAAY;AAClC,UAAM,UAAU,MAAM,CAAC,EAAE,KAAK;AAC9B,QAAI,QAAQ,SAAS,MAAM,SAAS,MAAM,KAAK,SAAS,MAAM,IAAI;AAChE,aAAO,KAAK,OAAO;AAAA,IACrB;AAAA,EACF;AACA,SAAO;AACT;;;ACVO,IAAM,sBAAN,cAAkC,MAAM;AAAA,EAC7C,YACE,SAEgB,MAChB;AACA,UAAM,OAAO;AAFG;AAGhB,SAAK,OAAO;AAAA,EACd;AAAA,EAJkB;AAKpB;;;ACxBA,SAAS,eACP,WACA,QAC8C;AAC9C,MAAI;AACF,WAAO,EAAE,IAAI,MAAM,OAAO,KAAK,MAAM,SAAS,EAAE;AAAA,EAClD,QAAQ;AAAA,EAER;AACA,MAAI,QAAQ;AACV,QAAI;AACF,aAAO,EAAE,IAAI,MAAM,OAAO,KAAK,MAAM,qBAAqB,SAAS,CAAC,EAAE;AAAA,IACxE,QAAQ;AAAA,IAER;AAAA,EACF;AACA,SAAO,EAAE,IAAI,MAAM;AACrB;AAEA,SAAS,cACP,OACA,QACS;AACT,MAAI,WAAW,OAAO;AACpB,WAAO;AAAA,EACT;AACA,MAAI,WAAW,SAAS;AACtB,WAAO,MAAM,QAAQ,KAAK;AAAA,EAC5B;AACA,SAAO,OAAO,UAAU,YAAY,UAAU,QAAQ,CAAC,MAAM,QAAQ,KAAK;AAC5E;AAeO,SAAS,eACd,MACA,UAA0B,CAAC,GACT;AAClB,MAAI,OAAO,SAAS,YAAY,KAAK,WAAW,GAAG;AACjD,WAAO,EAAE,OAAO,MAAM;AAAA,EACxB;AAEA,QAAM,SAAS,QAAQ,UAAU;AACjC,QAAM,SAAS,QAAQ,UAAU;AACjC,QAAM,UAAU,eAAe,IAAI;AAInC,QAAM,aAAuB,CAAC;AAC9B,aAAW,SAAS,aAAa,OAAO,GAAG;AACzC,eAAW,KAAK,OAAO,GAAG,cAAc,KAAK,CAAC;AAAA,EAChD;AACA,aAAW,KAAK,GAAG,cAAc,OAAO,CAAC;AAEzC,aAAW,aAAa,YAAY;AAClC,UAAM,SAAS,eAAe,WAAW,MAAM;AAC/C,QAAI,OAAO,MAAM,cAAc,OAAO,OAAO,MAAM,GAAG;AACpD,aAAO,EAAE,OAAO,MAAM,OAAO,OAAO,MAAW;AAAA,IACjD;AAAA,EACF;AACA,SAAO,EAAE,OAAO,MAAM;AACxB;AAMO,SAAS,YACd,MACA,UAA0B,CAAC,GACxB;AACH,QAAM,SAAS,eAAkB,MAAM,OAAO;AAC9C,MAAI,CAAC,OAAO,OAAO;AACjB,UAAM,IAAI;AAAA,MACR;AAAA,MACA;AAAA,IACF;AAAA,EACF;AACA,SAAO,OAAO;AAChB;","names":[]}

package/package.json

@@ -0,0 +1,74 @@
+{
+  "name": "json-from-llm",
+  "version": "0.1.0",
+  "description": "Extract valid JSON from an LLM response, even when it is wrapped in reasoning/thinking tags, markdown fences or prose. Zero dependencies.",
+  "keywords": [
+    "llm",
+    "json",
+    "extract",
+    "parse",
+    "structured-output",
+    "openai",
+    "anthropic",
+    "gemini",
+    "deepseek",
+    "reasoning",
+    "thinking",
+    "repair",
+    "ai",
+    "agents"
+  ],
+  "license": "MIT",
+  "author": "Sebastian Legarraga",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/slegarraga/json-from-llm.git"
+  },
+  "homepage": "https://github.com/slegarraga/json-from-llm#readme",
+  "bugs": {
+    "url": "https://github.com/slegarraga/json-from-llm/issues"
+  },
+  "type": "module",
+  "main": "./dist/index.cjs",
+  "module": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js",
+      "require": "./dist/index.cjs"
+    },
+    "./package.json": "./package.json"
+  },
+  "files": [
+    "dist",
+    "README.md",
+    "LICENSE",
+    "CHANGELOG.md"
+  ],
+  "engines": {
+    "node": ">=18"
+  },
+  "sideEffects": false,
+  "scripts": {
+    "build": "tsup",
+    "typecheck": "tsc --noEmit",
+    "test": "vitest run",
+    "test:watch": "vitest",
+    "lint": "eslint .",
+    "format": "prettier --write .",
+    "format:check": "prettier --check .",
+    "prepublishOnly": "npm run build",
+    "prepare": "npm run build"
+  },
+  "devDependencies": {
+    "@eslint/js": "^10.0.1",
+    "@types/node": "^25.9.1",
+    "eslint": "^10.4.1",
+    "prettier": "^3.4.2",
+    "tsup": "^8.3.5",
+    "typescript": "^5.7.2",
+    "typescript-eslint": "^8.60.0",
+    "vitest": "^2.1.8"
+  }
+}