@formwright/ai 0.1.0

package/LICENSE

@@ -0,0 +1,21 @@
+MIT License
+
+Copyright (c) 2026 Formwright contributors
+
+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 +1,85 @@
+# @formwright/ai
+
+> Generate a **validated** [Formwright](https://github.com/aliarsalan177/formwright) schema from a natural-language description — with any LLM.
+
+Describe a form in English; get back a schema the runtime accepts. Provider-agnostic:
+the validate → repair loop is built in, and the model call is pluggable — Claude
+(default), OpenAI/GPT, or anything else, each through its own native SDK.
+
+```bash
+npm i @formwright/ai
+```
+
+## Claude (default)
+
+```ts
+import { generateSchema } from "@formwright/ai";
+import { Form } from "@formwright/core";
+import "@formwright/dom";
+
+// Uses ANTHROPIC_API_KEY; model defaults to claude-opus-4-8
+const { schema, attempts } = await generateSchema(
+  "a checkout form: email, a plan dropdown (Free/Pro), and a promo code shown only when Pro is selected",
+);
+
+new Form(schema).mount(document.getElementById("root")!);
+```
+
+## OpenAI (GPT)
+
+Bring your own `openai` client — no extra dependency is bundled:
+
+```ts
+import OpenAI from "openai";
+import { generateSchema, openaiProvider } from "@formwright/ai";
+
+const { schema } = await generateSchema("a contact form with name, email, and message", {
+  provider: openaiProvider({ client: new OpenAI(), model: "gpt-4o" }),
+});
+```
+
+## Any other model (Gemini, Mistral, local, …)
+
+Wrap any async function that returns a candidate schema object; the repair loop does the rest:
+
+```ts
+import { generateSchema, defineProvider, buildPrompt } from "@formwright/ai";
+
+const gemini = defineProvider(async (input) => {
+  const text = await callYourModel({ system: input.system, prompt: buildPrompt(input) });
+  return JSON.parse(text); // return the parsed schema object
+});
+
+const { schema } = await generateSchema("an event RSVP form", { provider: gemini });
+```
+
+`input.repair` is populated on retries with the previous (invalid) output and the exact
+validation issues, so your provider can pass them back to the model.
+
+## How it works
+
+1. The provider proposes a schema object from your description.
+2. It's checked with `@formwright/schema`'s validator.
+3. If invalid, the precise, path-addressed issues are fed back for repair (up to
+   `maxRepairAttempts`, default 2).
+4. You get a `FormSchema` that the runtime is guaranteed to accept — or a thrown
+   `SchemaGenerationError` carrying the remaining issues.
+
+## API
+
+```ts
+generateSchema(description: string, options?: {
+  provider?: SchemaProvider;   // default: Claude
+  apiKey?: string;             // for the default Claude provider
+  model?: string;              // override the default model
+  maxRepairAttempts?: number;  // default 2
+  system?: string;             // override the Formwright DSL prompt
+  guidelines?: string;         // append extra guidance
+}): Promise<{ schema: FormSchema; attempts: number }>
+```
+
+Providers: `claudeProvider(opts)`, `openaiProvider({ client, model? })`, `defineProvider(fn)`.
+
+## License
+
+MIT

package/dist/chunk-I3HZFUC7.js

@@ -0,0 +1,156 @@
+import Anthropic from '@anthropic-ai/sdk';
+import { validateSchema } from '@formwright/schema';
+
+// src/claude.ts
+
+// src/openai.ts
+function openaiProvider(options) {
+  const model = options.model ?? "gpt-4o";
+  return {
+    async propose(input) {
+      const response = await options.client.chat.completions.create({
+        model,
+        response_format: { type: "json_object" },
+        messages: [
+          { role: "system", content: input.system },
+          {
+            role: "user",
+            content: `${buildPrompt(input)}
+
+Respond with ONLY the JSON schema object.`
+          }
+        ]
+      });
+      const content = response.choices[0]?.message?.content;
+      if (!content) throw new Error("OpenAI returned no content.");
+      return JSON.parse(content);
+    }
+  };
+}
+
+// src/index.ts
+var SchemaGenerationError = class extends Error {
+  issues;
+  constructor(message, issues) {
+    super(message);
+    this.name = "SchemaGenerationError";
+    this.issues = issues;
+  }
+};
+function defineProvider(propose) {
+  return { propose };
+}
+function buildPrompt(input) {
+  if (!input.repair) {
+    return `Design a Formwright form for: ${input.description}`;
+  }
+  return `Design a Formwright form for: ${input.description}
+
+Your previous attempt was INVALID:
+${JSON.stringify(input.repair.previous, null, 2)}
+
+Validation issues to fix:
+` + input.repair.issues.map((i) => `- ${i.path}: ${i.message}`).join("\n") + `
+
+Return a corrected schema that resolves every issue.`;
+}
+function unwrap(candidate) {
+  if (candidate && typeof candidate === "object" && "schema" in candidate) {
+    return candidate.schema;
+  }
+  return candidate;
+}
+async function generateSchema(description, options = {}) {
+  const provider = options.provider ?? await defaultProvider(options);
+  const maxRepair = options.maxRepairAttempts ?? 2;
+  const system = (options.system ?? SYSTEM_PROMPT) + (options.guidelines ? `
+
+Additional guidelines:
+${options.guidelines}` : "");
+  let repair;
+  let lastIssues = [];
+  for (let attempt = 1; attempt <= maxRepair + 1; attempt++) {
+    const input = repair ? { description, system, repair } : { description, system };
+    const candidate = await provider.propose(input);
+    const result = validateSchema(unwrap(candidate));
+    if (result.ok) return { schema: result.value, attempts: attempt };
+    lastIssues = result.issues;
+    if (attempt > maxRepair) break;
+    repair = { previous: unwrap(candidate), issues: result.issues };
+  }
+  throw new SchemaGenerationError(
+    `Could not produce a valid schema after ${maxRepair + 1} attempts.`,
+    lastIssues
+  );
+}
+async function defaultProvider(options) {
+  const { claudeProvider: claudeProvider2 } = await import('./claude-BNP64G4N.js');
+  const opts = {};
+  if (options.apiKey !== void 0) opts.apiKey = options.apiKey;
+  if (options.model !== void 0) opts.model = options.model;
+  return claudeProvider2(opts);
+}
+var SYSTEM_PROMPT = `You design forms as Formwright schemas \u2014 plain JSON, no code.
+
+A FormSchema has:
+- "id" (string), "version" (string, e.g. "1.0"), optional "title", and "fields" (non-empty array).
+- optional "submit": { "endpoint": { "method": "POST"|"GET"|"PUT"|"PATCH"|"DELETE", "url": string }, "transform"?, "onSuccess"?, "onError"? }.
+
+Each field has "id" (unique within its scope) and "type", plus optional "label", "placeholder", "help", "description", "defaultValue".
+Field types:
+- "text" | "email" | "password" | "number" | "textarea"
+- "checkbox" | "toggle" (boolean; toggle renders as a switch)
+- "select" | "radio" \u2014 need "options": [{ "label": string, "value": string|number }]
+- "group" \u2014 a nested object; needs "fields": [ ...child fields ]. Produces an object in the payload.
+- "collection" \u2014 a repeatable list of objects; needs "fields", optional "itemLabel", "addLabel", "minItems", "maxItems". Produces an array of objects.
+
+Validation (optional): "validation": { "kind": "string"|"number", "required"?, "min"?, "max"?, "minLength"?, "maxLength"?, "pattern"?, "format"?: "email"|"url"|"uuid", "message"? }.
+
+Conditional logic \u2014 data, not code \u2014 via "visibleWhen" / "enabledWhen" / "requiredWhen", a JSONLogic-style expression:
+{ "==": [a, b] }, "!=", ">", ">=", "<", "<=", { "in": [needle, haystack] }, { "and": [...] }, { "or": [...] }, { "not": x }, and { "var": "fieldId" } to read another field's value.
+Names resolve to a sibling first, then outward to the form root \u2014 so a field inside a group or collection row can react to an outer toggle.
+
+Other per-field options: "omit": true (keep in the UI but exclude from the payload), "labelPosition": "start"|"end" (checkbox/toggle), "layout" ("accordion" for group; "cards"|"accordion" for collection).
+
+Produce sensible labels, validation, and conditions matching the request. Prefer "toggle" for yes/no switches and add helpful placeholders. Output ONLY the schema object.`;
+
+// src/claude.ts
+var DEFAULT_CLAUDE_MODEL = "claude-opus-4-8";
+var TOOL_NAME = "emit_form_schema";
+var TOOL = {
+  name: TOOL_NAME,
+  description: "Emit the complete Formwright form schema as a structured object.",
+  input_schema: {
+    type: "object",
+    properties: {
+      schema: { type: "object", description: "A complete, valid Formwright FormSchema." }
+    },
+    required: ["schema"]
+  }
+};
+function claudeProvider(options = {}) {
+  const client = options.client ?? new Anthropic(options.apiKey ? { apiKey: options.apiKey } : void 0);
+  const model = options.model ?? DEFAULT_CLAUDE_MODEL;
+  const maxTokens = options.maxTokens ?? 16e3;
+  return {
+    async propose(input) {
+      const response = await client.messages.create({
+        model,
+        max_tokens: maxTokens,
+        system: input.system,
+        tools: [TOOL],
+        tool_choice: { type: "tool", name: TOOL_NAME },
+        messages: [{ role: "user", content: buildPrompt(input) }]
+      });
+      const toolUse = response.content.find(
+        (block) => block.type === "tool_use"
+      );
+      if (!toolUse) throw new Error("Claude did not emit a schema via the tool.");
+      return toolUse.input;
+    }
+  };
+}
+
+export { DEFAULT_CLAUDE_MODEL, SYSTEM_PROMPT, SchemaGenerationError, buildPrompt, claudeProvider, defineProvider, generateSchema, openaiProvider };
+//# sourceMappingURL=chunk-I3HZFUC7.js.map
+//# sourceMappingURL=chunk-I3HZFUC7.js.map

package/dist/chunk-I3HZFUC7.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/openai.ts","../src/index.ts","../src/claude.ts"],"names":["claudeProvider"],"mappings":";;;;;;AAkCO,SAAS,eAAe,OAAA,EAAgD;AAC7E,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,QAAA;AAC/B,EAAA,OAAO;AAAA,IACL,MAAM,QAAQ,KAAA,EAAuC;AACnD,MAAA,MAAM,WAAW,MAAM,OAAA,CAAQ,MAAA,CAAO,IAAA,CAAK,YAAY,MAAA,CAAO;AAAA,QAC5D,KAAA;AAAA,QACA,eAAA,EAAiB,EAAE,IAAA,EAAM,aAAA,EAAc;AAAA,QACvC,QAAA,EAAU;AAAA,UACR,EAAE,IAAA,EAAM,QAAA,EAAU,OAAA,EAAS,MAAM,MAAA,EAAO;AAAA,UACxC;AAAA,YACE,IAAA,EAAM,MAAA;AAAA,YACN,OAAA,EAAS,CAAA,EAAG,WAAA,CAAY,KAAK,CAAC;;AAAA,yCAAA;AAAA;AAChC;AACF,OACD,CAAA;AACD,MAAA,MAAM,OAAA,GAAU,QAAA,CAAS,OAAA,CAAQ,CAAC,GAAG,OAAA,EAAS,OAAA;AAC9C,MAAA,IAAI,CAAC,OAAA,EAAS,MAAM,IAAI,MAAM,6BAA6B,CAAA;AAC3D,MAAA,OAAO,IAAA,CAAK,MAAM,OAAO,CAAA;AAAA,IAC3B;AAAA,GACF;AACF;;;ACMO,IAAM,qBAAA,GAAN,cAAoC,KAAA,CAAM;AAAA,EACtC,MAAA;AAAA,EACT,WAAA,CAAY,SAAiB,MAAA,EAAoC;AAC/D,IAAA,KAAA,CAAM,OAAO,CAAA;AACb,IAAA,IAAA,CAAK,IAAA,GAAO,uBAAA;AACZ,IAAA,IAAA,CAAK,MAAA,GAAS,MAAA;AAAA,EAChB;AACF;AAGO,SAAS,eAAe,OAAA,EAAoE;AACjG,EAAA,OAAO,EAAE,OAAA,EAAQ;AACnB;AAGO,SAAS,YAAY,KAAA,EAA6B;AACvD,EAAA,IAAI,CAAC,MAAM,MAAA,EAAQ;AACjB,IAAA,OAAO,CAAA,8BAAA,EAAiC,MAAM,WAAW,CAAA,CAAA;AAAA,EAC3D;AACA,EAAA,OACE,CAAA,8BAAA,EAAiC,MAAM,WAAW;;AAAA;AAAA,EACX,KAAK,SAAA,CAAU,KAAA,CAAM,OAAO,QAAA,EAAU,IAAA,EAAM,CAAC,CAAC;;AAAA;AAAA,CAAA,GAErF,MAAM,MAAA,CAAO,MAAA,CAAO,GAAA,CAAI,CAAC,MAAM,CAAA,EAAA,EAAK,CAAA,CAAE,IAAI,CAAA,EAAA,EAAK,EAAE,OAAO,CAAA,CAAE,CAAA,CAAE,IAAA,CAAK,IAAI,CAAA,GACrE;;AAAA,oDAAA,CAAA;AAEJ;AAGA,SAAS,OAAO,SAAA,EAA6B;AAC3C,EAAA,IAAI,SAAA,IAAa,OAAO,SAAA,KAAc,QAAA,IAAY,YAAY,SAAA,EAAW;AACvE,IAAA,OAAQ,SAAA,CAAkC,MAAA;AAAA,EAC5C;AACA,EAAA,OAAO,SAAA;AACT;AAOA,eAAsB,cAAA,CACpB,WAAA,EACA,OAAA,GAA2B,EAAC,EACH;AACzB,EAAA,MAAM,QAAA,GAAW,OAAA,CAAQ,QAAA,IAAa,MAAM,gBAAgB,OAAO,CAAA;AACnE,EAAA,MAAM,SAAA,GAAY,QAAQ,iBAAA,IAAqB,CAAA;AAC/C,EAAA,MAAM,MAAA,GAAA,CACH,OAAA,CAAQ,MAAA,IAAU,aAAA,KAClB,QAAQ,UAAA,GAAa;;AAAA;AAAA,EAA+B,OAAA,CAAQ,UAAU,CAAA,CAAA,GAAK,EAAA,CAAA;AAE9E,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI,aAAyC,EAAC;AAE9C,EAAA,KAAA,IAAS,OAAA,GAAU,CAAA,EAAG,OAAA,IAAW,SAAA,GAAY,GAAG,OAAA,EAAA,EAAW;AACzD,IAAA,MAAM,KAAA,GAAsB,SAAS,EAAE,WAAA,EAAa,QAAQ,MAAA,EAAO,GAAI,EAAE,WAAA,EAAa,MAAA,EAAO;AAC7F,IAAA,MAAM,SAAA,GAAY,MAAM,QAAA,CAAS,OAAA,CAAQ,KAAK,CAAA;AAC9C,IAAA,MAAM,MAAA,GAAS,cAAA,CAAe,MAAA,CAAO,SAAS,CAAC,CAAA;AAC/C,IAAA,IAAI,MAAA,CAAO,IAAI,OAAO,EAAE,QAAQ,MAAA,CAAO,KAAA,EAAO,UAAU,OAAA,EAAQ;AAEhE,IAAA,UAAA,GAAa,MAAA,CAAO,MAAA;AACpB,IAAA,IAAI,UAAU,SAAA,EAAW;AACzB,IAAA,MAAA,GAAS,EAAE,QAAA,EAAU,MAAA,CAAO,SAAS,CAAA,EAAG,MAAA,EAAQ,OAAO,MAAA,EAAO;AAAA,EAChE;AAEA,EAAA,MAAM,IAAI,qBAAA;AAAA,IACR,CAAA,uCAAA,EAA0C,YAAY,CAAC,CAAA,UAAA,CAAA;AAAA,IACvD;AAAA,GACF;AACF;AAGA,eAAe,gBAAgB,OAAA,EAAmD;AAChF,EAAA,MAAM,EAAE,cAAA,EAAAA,eAAAA,EAAe,GAAI,MAAM,OAAO,sBAAa,CAAA;AACrD,EAAA,MAAM,OAA4C,EAAC;AACnD,EAAA,IAAI,OAAA,CAAQ,MAAA,KAAW,MAAA,EAAW,IAAA,CAAK,SAAS,OAAA,CAAQ,MAAA;AACxD,EAAA,IAAI,OAAA,CAAQ,KAAA,KAAU,MAAA,EAAW,IAAA,CAAK,QAAQ,OAAA,CAAQ,KAAA;AACtD,EAAA,OAAOA,gBAAe,IAAI,CAAA;AAC5B;AAEO,IAAM,aAAA,GAAgB,CAAA;;AAAA;AAAA;AAAA;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;AAAA;;AAAA;AAAA;AAAA;;AAAA;;AAAA,0KAAA;;;ACpItB,IAAM,oBAAA,GAAuB;AAUpC,IAAM,SAAA,GAAY,kBAAA;AAElB,IAAM,IAAA,GAAuB;AAAA,EAC3B,IAAA,EAAM,SAAA;AAAA,EACN,WAAA,EAAa,kEAAA;AAAA,EACb,YAAA,EAAc;AAAA,IACZ,IAAA,EAAM,QAAA;AAAA,IACN,UAAA,EAAY;AAAA,MACV,MAAA,EAAQ,EAAE,IAAA,EAAM,QAAA,EAAU,aAAa,0CAAA;AAA2C,KACpF;AAAA,IACA,QAAA,EAAU,CAAC,QAAQ;AAAA;AAEvB,CAAA;AAEO,SAAS,cAAA,CAAe,OAAA,GAAiC,EAAC,EAAmB;AAClF,EAAA,MAAM,MAAA,GACJ,OAAA,CAAQ,MAAA,IAAU,IAAI,SAAA,CAAU,OAAA,CAAQ,MAAA,GAAS,EAAE,MAAA,EAAQ,OAAA,CAAQ,MAAA,EAAO,GAAI,MAAS,CAAA;AACzF,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,oBAAA;AAC/B,EAAA,MAAM,SAAA,GAAY,QAAQ,SAAA,IAAa,IAAA;AAEvC,EAAA,OAAO;AAAA,IACL,MAAM,QAAQ,KAAA,EAAuC;AACnD,MAAA,MAAM,QAAA,GAAW,MAAM,MAAA,CAAO,QAAA,CAAS,MAAA,CAAO;AAAA,QAC5C,KAAA;AAAA,QACA,UAAA,EAAY,SAAA;AAAA,QACZ,QAAQ,KAAA,CAAM,MAAA;AAAA,QACd,KAAA,EAAO,CAAC,IAAI,CAAA;AAAA,QACZ,WAAA,EAAa,EAAE,IAAA,EAAM,MAAA,EAAQ,MAAM,SAAA,EAAU;AAAA,QAC7C,QAAA,EAAU,CAAC,EAAE,IAAA,EAAM,QAAQ,OAAA,EAAS,WAAA,CAAY,KAAK,CAAA,EAAG;AAAA,OACzD,CAAA;AACD,MAAA,MAAM,OAAA,GAAU,SAAS,OAAA,CAAQ,IAAA;AAAA,QAC/B,CAAC,KAAA,KAA2C,KAAA,CAAM,IAAA,KAAS;AAAA,OAC7D;AACA,MAAA,IAAI,CAAC,OAAA,EAAS,MAAM,IAAI,MAAM,4CAA4C,CAAA;AAC1E,MAAA,OAAO,OAAA,CAAQ,KAAA;AAAA,IACjB;AAAA,GACF;AACF","file":"chunk-I3HZFUC7.js","sourcesContent":["/**\n * OpenAI (GPT) provider — uses JSON mode on the Chat Completions API.\n *\n * To avoid a hard dependency on the `openai` package, pass your own client:\n *\n *   import OpenAI from \"openai\";\n *   import { generateSchema, openaiProvider } from \"@formwright/ai\";\n *   const { schema } = await generateSchema(\"a contact form\", {\n *     provider: openaiProvider({ client: new OpenAI() }),\n *   });\n *\n * Any client matching {@link OpenAILike} works (Azure OpenAI, compatible gateways).\n */\nimport { buildPrompt, type ProposeInput, type SchemaProvider } from \"./index.js\";\n\n/** Minimal structural shape of an OpenAI client's `chat.completions.create`. */\nexport interface OpenAILike {\n  chat: {\n    completions: {\n      create(body: {\n        model: string;\n        messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>;\n        response_format?: { type: \"json_object\" };\n      }): Promise<{ choices: Array<{ message: { content: string | null } }> }>;\n    };\n  };\n}\n\nexport interface OpenAIProviderOptions {\n  /** A configured OpenAI client (e.g. `new OpenAI()`). Required. */\n  readonly client: OpenAILike;\n  readonly model?: string;\n}\n\nexport function openaiProvider(options: OpenAIProviderOptions): SchemaProvider {\n  const model = options.model ?? \"gpt-4o\";\n  return {\n    async propose(input: ProposeInput): Promise<unknown> {\n      const response = await options.client.chat.completions.create({\n        model,\n        response_format: { type: \"json_object\" },\n        messages: [\n          { role: \"system\", content: input.system },\n          {\n            role: \"user\",\n            content: `${buildPrompt(input)}\\n\\nRespond with ONLY the JSON schema object.`,\n          },\n        ],\n      });\n      const content = response.choices[0]?.message?.content;\n      if (!content) throw new Error(\"OpenAI returned no content.\");\n      return JSON.parse(content) as unknown;\n    },\n  };\n}\n","/**\n * @formwright/ai — turn a natural-language form description into a *validated*\n * Formwright schema using an LLM.\n *\n *   import { generateSchema } from \"@formwright/ai\";\n *   const { schema } = await generateSchema(\"a signup form with a US-only state field\");\n *   new Form(schema).mount(el);\n *\n * Provider-agnostic: the validate → repair loop lives here, and the actual model\n * call is a pluggable {@link SchemaProvider}. A Claude provider ships built-in\n * (the default); an OpenAI provider and a custom-function provider let you use\n * GPT, Gemini, Mistral, a local model, or anything else — each through its own\n * native SDK, never a compatibility shim.\n *\n * Whatever the provider returns is checked with `@formwright/schema`'s validator;\n * if it's invalid, the precise issues are fed back for repair — so what you get\n * out always satisfies the runtime (or a thrown {@link SchemaGenerationError}).\n */\nimport { validateSchema, type FormSchema, type ValidationIssue } from \"@formwright/schema\";\n\nexport { claudeProvider, type ClaudeProviderOptions } from \"./claude.js\";\nexport { openaiProvider, type OpenAIProviderOptions, type OpenAILike } from \"./openai.js\";\n\n/** A pluggable model backend: produce a candidate schema object for a request. */\nexport interface SchemaProvider {\n  propose(input: ProposeInput): Promise<unknown>;\n}\n\nexport interface ProposeInput {\n  readonly description: string;\n  readonly system: string;\n  /** Present on a repair attempt: the previous (invalid) output and why it failed. */\n  readonly repair?: RepairContext;\n}\n\nexport interface RepairContext {\n  readonly previous: unknown;\n  readonly issues: readonly ValidationIssue[];\n}\n\nexport interface GenerateOptions {\n  /** Model backend. Defaults to Claude (needs `ANTHROPIC_API_KEY` or `apiKey`). */\n  readonly provider?: SchemaProvider;\n  /** Convenience for the default Claude provider when `provider` is omitted. */\n  readonly apiKey?: string;\n  readonly model?: string;\n  /** How many times to feed validation errors back for repair (default 2). */\n  readonly maxRepairAttempts?: number;\n  /** Override the Formwright DSL system prompt. */\n  readonly system?: string;\n  /** Extra guidance appended to the system prompt. */\n  readonly guidelines?: string;\n}\n\nexport interface GenerateResult {\n  readonly schema: FormSchema;\n  /** Number of model round-trips it took (1 = valid on the first try). */\n  readonly attempts: number;\n}\n\nexport class SchemaGenerationError extends Error {\n  readonly issues: readonly ValidationIssue[];\n  constructor(message: string, issues: readonly ValidationIssue[]) {\n    super(message);\n    this.name = \"SchemaGenerationError\";\n    this.issues = issues;\n  }\n}\n\n/** Wrap a plain async function as a provider (for Gemini, Mistral, local models, …). */\nexport function defineProvider(propose: (input: ProposeInput) => Promise<unknown>): SchemaProvider {\n  return { propose };\n}\n\n/** Build the user-facing instruction, including repair feedback when retrying. */\nexport function buildPrompt(input: ProposeInput): string {\n  if (!input.repair) {\n    return `Design a Formwright form for: ${input.description}`;\n  }\n  return (\n    `Design a Formwright form for: ${input.description}\\n\\n` +\n    `Your previous attempt was INVALID:\\n${JSON.stringify(input.repair.previous, null, 2)}\\n\\n` +\n    `Validation issues to fix:\\n` +\n    input.repair.issues.map((i) => `- ${i.path}: ${i.message}`).join(\"\\n\") +\n    `\\n\\nReturn a corrected schema that resolves every issue.`\n  );\n}\n\n/** Some providers wrap the schema under a `schema` key; accept either shape. */\nfunction unwrap(candidate: unknown): unknown {\n  if (candidate && typeof candidate === \"object\" && \"schema\" in candidate) {\n    return (candidate as { schema: unknown }).schema;\n  }\n  return candidate;\n}\n\n/**\n * Generate a validated {@link FormSchema} from a natural-language description.\n * Throws {@link SchemaGenerationError} if the model can't produce a valid schema\n * within `maxRepairAttempts`.\n */\nexport async function generateSchema(\n  description: string,\n  options: GenerateOptions = {},\n): Promise<GenerateResult> {\n  const provider = options.provider ?? (await defaultProvider(options));\n  const maxRepair = options.maxRepairAttempts ?? 2;\n  const system =\n    (options.system ?? SYSTEM_PROMPT) +\n    (options.guidelines ? `\\n\\nAdditional guidelines:\\n${options.guidelines}` : \"\");\n\n  let repair: RepairContext | undefined;\n  let lastIssues: readonly ValidationIssue[] = [];\n\n  for (let attempt = 1; attempt <= maxRepair + 1; attempt++) {\n    const input: ProposeInput = repair ? { description, system, repair } : { description, system };\n    const candidate = await provider.propose(input);\n    const result = validateSchema(unwrap(candidate));\n    if (result.ok) return { schema: result.value, attempts: attempt };\n\n    lastIssues = result.issues;\n    if (attempt > maxRepair) break;\n    repair = { previous: unwrap(candidate), issues: result.issues };\n  }\n\n  throw new SchemaGenerationError(\n    `Could not produce a valid schema after ${maxRepair + 1} attempts.`,\n    lastIssues,\n  );\n}\n\n/** Lazily construct the default (Claude) provider so OpenAI-only users don't need a key. */\nasync function defaultProvider(options: GenerateOptions): Promise<SchemaProvider> {\n  const { claudeProvider } = await import(\"./claude.js\");\n  const opts: { apiKey?: string; model?: string } = {};\n  if (options.apiKey !== undefined) opts.apiKey = options.apiKey;\n  if (options.model !== undefined) opts.model = options.model;\n  return claudeProvider(opts);\n}\n\nexport const SYSTEM_PROMPT = `You design forms as Formwright schemas — plain JSON, no code.\n\nA FormSchema has:\n- \"id\" (string), \"version\" (string, e.g. \"1.0\"), optional \"title\", and \"fields\" (non-empty array).\n- optional \"submit\": { \"endpoint\": { \"method\": \"POST\"|\"GET\"|\"PUT\"|\"PATCH\"|\"DELETE\", \"url\": string }, \"transform\"?, \"onSuccess\"?, \"onError\"? }.\n\nEach field has \"id\" (unique within its scope) and \"type\", plus optional \"label\", \"placeholder\", \"help\", \"description\", \"defaultValue\".\nField types:\n- \"text\" | \"email\" | \"password\" | \"number\" | \"textarea\"\n- \"checkbox\" | \"toggle\" (boolean; toggle renders as a switch)\n- \"select\" | \"radio\" — need \"options\": [{ \"label\": string, \"value\": string|number }]\n- \"group\" — a nested object; needs \"fields\": [ ...child fields ]. Produces an object in the payload.\n- \"collection\" — a repeatable list of objects; needs \"fields\", optional \"itemLabel\", \"addLabel\", \"minItems\", \"maxItems\". Produces an array of objects.\n\nValidation (optional): \"validation\": { \"kind\": \"string\"|\"number\", \"required\"?, \"min\"?, \"max\"?, \"minLength\"?, \"maxLength\"?, \"pattern\"?, \"format\"?: \"email\"|\"url\"|\"uuid\", \"message\"? }.\n\nConditional logic — data, not code — via \"visibleWhen\" / \"enabledWhen\" / \"requiredWhen\", a JSONLogic-style expression:\n{ \"==\": [a, b] }, \"!=\", \">\", \">=\", \"<\", \"<=\", { \"in\": [needle, haystack] }, { \"and\": [...] }, { \"or\": [...] }, { \"not\": x }, and { \"var\": \"fieldId\" } to read another field's value.\nNames resolve to a sibling first, then outward to the form root — so a field inside a group or collection row can react to an outer toggle.\n\nOther per-field options: \"omit\": true (keep in the UI but exclude from the payload), \"labelPosition\": \"start\"|\"end\" (checkbox/toggle), \"layout\" (\"accordion\" for group; \"cards\"|\"accordion\" for collection).\n\nProduce sensible labels, validation, and conditions matching the request. Prefer \"toggle\" for yes/no switches and add helpful placeholders. Output ONLY the schema object.`;\n","/**\n * Claude provider — emits the schema through a forced tool call on the Anthropic\n * Messages API. Default model is Claude Opus 4.8 (`claude-opus-4-8`).\n */\nimport Anthropic from \"@anthropic-ai/sdk\";\nimport { buildPrompt, type ProposeInput, type SchemaProvider } from \"./index.js\";\n\n/** Claude Opus 4.8 — the most capable model. */\nexport const DEFAULT_CLAUDE_MODEL = \"claude-opus-4-8\";\n\nexport interface ClaudeProviderOptions {\n  readonly apiKey?: string;\n  /** Bring your own configured client (takes precedence over `apiKey`). */\n  readonly client?: Anthropic;\n  readonly model?: string;\n  readonly maxTokens?: number;\n}\n\nconst TOOL_NAME = \"emit_form_schema\";\n\nconst TOOL: Anthropic.Tool = {\n  name: TOOL_NAME,\n  description: \"Emit the complete Formwright form schema as a structured object.\",\n  input_schema: {\n    type: \"object\",\n    properties: {\n      schema: { type: \"object\", description: \"A complete, valid Formwright FormSchema.\" },\n    },\n    required: [\"schema\"],\n  },\n};\n\nexport function claudeProvider(options: ClaudeProviderOptions = {}): SchemaProvider {\n  const client =\n    options.client ?? new Anthropic(options.apiKey ? { apiKey: options.apiKey } : undefined);\n  const model = options.model ?? DEFAULT_CLAUDE_MODEL;\n  const maxTokens = options.maxTokens ?? 16000;\n\n  return {\n    async propose(input: ProposeInput): Promise<unknown> {\n      const response = await client.messages.create({\n        model,\n        max_tokens: maxTokens,\n        system: input.system,\n        tools: [TOOL],\n        tool_choice: { type: \"tool\", name: TOOL_NAME },\n        messages: [{ role: \"user\", content: buildPrompt(input) }],\n      });\n      const toolUse = response.content.find(\n        (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n      );\n      if (!toolUse) throw new Error(\"Claude did not emit a schema via the tool.\");\n      return toolUse.input;\n    },\n  };\n}\n"]}

package/dist/claude-BNP64G4N.js

@@ -0,0 +1,3 @@
+export { DEFAULT_CLAUDE_MODEL, claudeProvider } from './chunk-I3HZFUC7.js';
+//# sourceMappingURL=claude-BNP64G4N.js.map
+//# sourceMappingURL=claude-BNP64G4N.js.map

package/dist/claude-BNP64G4N.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"names":[],"mappings":"","file":"claude-BNP64G4N.js"}

package/dist/index.cjs

@@ -0,0 +1,196 @@
+'use strict';
+
+var Anthropic = require('@anthropic-ai/sdk');
+var schema = require('@formwright/schema');
+
+function _interopDefault (e) { return e && e.__esModule ? e : { default: e }; }
+
+var Anthropic__default = /*#__PURE__*/_interopDefault(Anthropic);
+
+var __defProp = Object.defineProperty;
+var __getOwnPropNames = Object.getOwnPropertyNames;
+var __esm = (fn, res) => function __init() {
+  return fn && (res = (0, fn[__getOwnPropNames(fn)[0]])(fn = 0)), res;
+};
+var __export = (target, all) => {
+  for (var name in all)
+    __defProp(target, name, { get: all[name], enumerable: true });
+};
+
+// src/claude.ts
+var claude_exports = {};
+__export(claude_exports, {
+  DEFAULT_CLAUDE_MODEL: () => DEFAULT_CLAUDE_MODEL,
+  claudeProvider: () => claudeProvider
+});
+function claudeProvider(options = {}) {
+  const client = options.client ?? new Anthropic__default.default(options.apiKey ? { apiKey: options.apiKey } : void 0);
+  const model = options.model ?? DEFAULT_CLAUDE_MODEL;
+  const maxTokens = options.maxTokens ?? 16e3;
+  return {
+    async propose(input) {
+      const response = await client.messages.create({
+        model,
+        max_tokens: maxTokens,
+        system: input.system,
+        tools: [TOOL],
+        tool_choice: { type: "tool", name: TOOL_NAME },
+        messages: [{ role: "user", content: buildPrompt(input) }]
+      });
+      const toolUse = response.content.find(
+        (block) => block.type === "tool_use"
+      );
+      if (!toolUse) throw new Error("Claude did not emit a schema via the tool.");
+      return toolUse.input;
+    }
+  };
+}
+var DEFAULT_CLAUDE_MODEL, TOOL_NAME, TOOL;
+var init_claude = __esm({
+  "src/claude.ts"() {
+    init_index();
+    DEFAULT_CLAUDE_MODEL = "claude-opus-4-8";
+    TOOL_NAME = "emit_form_schema";
+    TOOL = {
+      name: TOOL_NAME,
+      description: "Emit the complete Formwright form schema as a structured object.",
+      input_schema: {
+        type: "object",
+        properties: {
+          schema: { type: "object", description: "A complete, valid Formwright FormSchema." }
+        },
+        required: ["schema"]
+      }
+    };
+  }
+});
+
+// src/openai.ts
+function openaiProvider(options) {
+  const model = options.model ?? "gpt-4o";
+  return {
+    async propose(input) {
+      const response = await options.client.chat.completions.create({
+        model,
+        response_format: { type: "json_object" },
+        messages: [
+          { role: "system", content: input.system },
+          {
+            role: "user",
+            content: `${buildPrompt(input)}
+
+Respond with ONLY the JSON schema object.`
+          }
+        ]
+      });
+      const content = response.choices[0]?.message?.content;
+      if (!content) throw new Error("OpenAI returned no content.");
+      return JSON.parse(content);
+    }
+  };
+}
+var init_openai = __esm({
+  "src/openai.ts"() {
+    init_index();
+  }
+});
+function defineProvider(propose) {
+  return { propose };
+}
+function buildPrompt(input) {
+  if (!input.repair) {
+    return `Design a Formwright form for: ${input.description}`;
+  }
+  return `Design a Formwright form for: ${input.description}
+
+Your previous attempt was INVALID:
+${JSON.stringify(input.repair.previous, null, 2)}
+
+Validation issues to fix:
+` + input.repair.issues.map((i) => `- ${i.path}: ${i.message}`).join("\n") + `
+
+Return a corrected schema that resolves every issue.`;
+}
+function unwrap(candidate) {
+  if (candidate && typeof candidate === "object" && "schema" in candidate) {
+    return candidate.schema;
+  }
+  return candidate;
+}
+async function generateSchema(description, options = {}) {
+  const provider = options.provider ?? await defaultProvider(options);
+  const maxRepair = options.maxRepairAttempts ?? 2;
+  const system = (options.system ?? exports.SYSTEM_PROMPT) + (options.guidelines ? `
+
+Additional guidelines:
+${options.guidelines}` : "");
+  let repair;
+  let lastIssues = [];
+  for (let attempt = 1; attempt <= maxRepair + 1; attempt++) {
+    const input = repair ? { description, system, repair } : { description, system };
+    const candidate = await provider.propose(input);
+    const result = schema.validateSchema(unwrap(candidate));
+    if (result.ok) return { schema: result.value, attempts: attempt };
+    lastIssues = result.issues;
+    if (attempt > maxRepair) break;
+    repair = { previous: unwrap(candidate), issues: result.issues };
+  }
+  throw new exports.SchemaGenerationError(
+    `Could not produce a valid schema after ${maxRepair + 1} attempts.`,
+    lastIssues
+  );
+}
+async function defaultProvider(options) {
+  const { claudeProvider: claudeProvider2 } = await Promise.resolve().then(() => (init_claude(), claude_exports));
+  const opts = {};
+  if (options.apiKey !== void 0) opts.apiKey = options.apiKey;
+  if (options.model !== void 0) opts.model = options.model;
+  return claudeProvider2(opts);
+}
+exports.SchemaGenerationError = void 0; exports.SYSTEM_PROMPT = void 0;
+var init_index = __esm({
+  "src/index.ts"() {
+    init_claude();
+    init_openai();
+    exports.SchemaGenerationError = class extends Error {
+      issues;
+      constructor(message, issues) {
+        super(message);
+        this.name = "SchemaGenerationError";
+        this.issues = issues;
+      }
+    };
+    exports.SYSTEM_PROMPT = `You design forms as Formwright schemas \u2014 plain JSON, no code.
+
+A FormSchema has:
+- "id" (string), "version" (string, e.g. "1.0"), optional "title", and "fields" (non-empty array).
+- optional "submit": { "endpoint": { "method": "POST"|"GET"|"PUT"|"PATCH"|"DELETE", "url": string }, "transform"?, "onSuccess"?, "onError"? }.
+
+Each field has "id" (unique within its scope) and "type", plus optional "label", "placeholder", "help", "description", "defaultValue".
+Field types:
+- "text" | "email" | "password" | "number" | "textarea"
+- "checkbox" | "toggle" (boolean; toggle renders as a switch)
+- "select" | "radio" \u2014 need "options": [{ "label": string, "value": string|number }]
+- "group" \u2014 a nested object; needs "fields": [ ...child fields ]. Produces an object in the payload.
+- "collection" \u2014 a repeatable list of objects; needs "fields", optional "itemLabel", "addLabel", "minItems", "maxItems". Produces an array of objects.
+
+Validation (optional): "validation": { "kind": "string"|"number", "required"?, "min"?, "max"?, "minLength"?, "maxLength"?, "pattern"?, "format"?: "email"|"url"|"uuid", "message"? }.
+
+Conditional logic \u2014 data, not code \u2014 via "visibleWhen" / "enabledWhen" / "requiredWhen", a JSONLogic-style expression:
+{ "==": [a, b] }, "!=", ">", ">=", "<", "<=", { "in": [needle, haystack] }, { "and": [...] }, { "or": [...] }, { "not": x }, and { "var": "fieldId" } to read another field's value.
+Names resolve to a sibling first, then outward to the form root \u2014 so a field inside a group or collection row can react to an outer toggle.
+
+Other per-field options: "omit": true (keep in the UI but exclude from the payload), "labelPosition": "start"|"end" (checkbox/toggle), "layout" ("accordion" for group; "cards"|"accordion" for collection).
+
+Produce sensible labels, validation, and conditions matching the request. Prefer "toggle" for yes/no switches and add helpful placeholders. Output ONLY the schema object.`;
+  }
+});
+init_index();
+
+exports.buildPrompt = buildPrompt;
+exports.claudeProvider = claudeProvider;
+exports.defineProvider = defineProvider;
+exports.generateSchema = generateSchema;
+exports.openaiProvider = openaiProvider;
+//# sourceMappingURL=index.cjs.map
+//# sourceMappingURL=index.cjs.map

package/dist/index.cjs.map

@@ -0,0 +1 @@
+{"version":3,"sources":["../src/claude.ts","../src/openai.ts","../src/index.ts"],"names":["Anthropic","SYSTEM_PROMPT","validateSchema","SchemaGenerationError","claudeProvider"],"mappings":";;;;;;;;;;;;;;;;;;;;AAAA,IAAA,cAAA,GAAA,EAAA;AAAA,QAAA,CAAA,cAAA,EAAA;AAAA,EAAA,oBAAA,EAAA,MAAA,oBAAA;AAAA,EAAA,cAAA,EAAA,MAAA;AAAA,CAAA,CAAA;AAgCO,SAAS,cAAA,CAAe,OAAA,GAAiC,EAAC,EAAmB;AAClF,EAAA,MAAM,MAAA,GACJ,OAAA,CAAQ,MAAA,IAAU,IAAIA,0BAAA,CAAU,OAAA,CAAQ,MAAA,GAAS,EAAE,MAAA,EAAQ,OAAA,CAAQ,MAAA,EAAO,GAAI,MAAS,CAAA;AACzF,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,oBAAA;AAC/B,EAAA,MAAM,SAAA,GAAY,QAAQ,SAAA,IAAa,IAAA;AAEvC,EAAA,OAAO;AAAA,IACL,MAAM,QAAQ,KAAA,EAAuC;AACnD,MAAA,MAAM,QAAA,GAAW,MAAM,MAAA,CAAO,QAAA,CAAS,MAAA,CAAO;AAAA,QAC5C,KAAA;AAAA,QACA,UAAA,EAAY,SAAA;AAAA,QACZ,QAAQ,KAAA,CAAM,MAAA;AAAA,QACd,KAAA,EAAO,CAAC,IAAI,CAAA;AAAA,QACZ,WAAA,EAAa,EAAE,IAAA,EAAM,MAAA,EAAQ,MAAM,SAAA,EAAU;AAAA,QAC7C,QAAA,EAAU,CAAC,EAAE,IAAA,EAAM,QAAQ,OAAA,EAAS,WAAA,CAAY,KAAK,CAAA,EAAG;AAAA,OACzD,CAAA;AACD,MAAA,MAAM,OAAA,GAAU,SAAS,OAAA,CAAQ,IAAA;AAAA,QAC/B,CAAC,KAAA,KAA2C,KAAA,CAAM,IAAA,KAAS;AAAA,OAC7D;AACA,MAAA,IAAI,CAAC,OAAA,EAAS,MAAM,IAAI,MAAM,4CAA4C,CAAA;AAC1E,MAAA,OAAO,OAAA,CAAQ,KAAA;AAAA,IACjB;AAAA,GACF;AACF;AAvDA,IAQa,sBAUP,SAAA,EAEA,IAAA;AApBN,IAAA,WAAA,GAAA,KAAA,CAAA;AAAA,EAAA,eAAA,GAAA;AAKA,IAAA,UAAA,EAAA;AAGO,IAAM,oBAAA,GAAuB,iBAAA;AAUpC,IAAM,SAAA,GAAY,kBAAA;AAElB,IAAM,IAAA,GAAuB;AAAA,MAC3B,IAAA,EAAM,SAAA;AAAA,MACN,WAAA,EAAa,kEAAA;AAAA,MACb,YAAA,EAAc;AAAA,QACZ,IAAA,EAAM,QAAA;AAAA,QACN,UAAA,EAAY;AAAA,UACV,MAAA,EAAQ,EAAE,IAAA,EAAM,QAAA,EAAU,aAAa,0CAAA;AAA2C,SACpF;AAAA,QACA,QAAA,EAAU,CAAC,QAAQ;AAAA;AACrB,KACF;AAAA,EAAA;AAAA,CAAA,CAAA;;;ACIO,SAAS,eAAe,OAAA,EAAgD;AAC7E,EAAA,MAAM,KAAA,GAAQ,QAAQ,KAAA,IAAS,QAAA;AAC/B,EAAA,OAAO;AAAA,IACL,MAAM,QAAQ,KAAA,EAAuC;AACnD,MAAA,MAAM,WAAW,MAAM,OAAA,CAAQ,MAAA,CAAO,IAAA,CAAK,YAAY,MAAA,CAAO;AAAA,QAC5D,KAAA;AAAA,QACA,eAAA,EAAiB,EAAE,IAAA,EAAM,aAAA,EAAc;AAAA,QACvC,QAAA,EAAU;AAAA,UACR,EAAE,IAAA,EAAM,QAAA,EAAU,OAAA,EAAS,MAAM,MAAA,EAAO;AAAA,UACxC;AAAA,YACE,IAAA,EAAM,MAAA;AAAA,YACN,OAAA,EAAS,CAAA,EAAG,WAAA,CAAY,KAAK,CAAC;;AAAA,yCAAA;AAAA;AAChC;AACF,OACD,CAAA;AACD,MAAA,MAAM,OAAA,GAAU,QAAA,CAAS,OAAA,CAAQ,CAAC,GAAG,OAAA,EAAS,OAAA;AAC9C,MAAA,IAAI,CAAC,OAAA,EAAS,MAAM,IAAI,MAAM,6BAA6B,CAAA;AAC3D,MAAA,OAAO,IAAA,CAAK,MAAM,OAAO,CAAA;AAAA,IAC3B;AAAA,GACF;AACF;AAtDA,IAAA,WAAA,GAAA,KAAA,CAAA;AAAA,EAAA,eAAA,GAAA;AAaA,IAAA,UAAA,EAAA;AAAA,EAAA;AAAA,CAAA,CAAA;ACyDO,SAAS,eAAe,OAAA,EAAoE;AACjG,EAAA,OAAO,EAAE,OAAA,EAAQ;AACnB;AAGO,SAAS,YAAY,KAAA,EAA6B;AACvD,EAAA,IAAI,CAAC,MAAM,MAAA,EAAQ;AACjB,IAAA,OAAO,CAAA,8BAAA,EAAiC,MAAM,WAAW,CAAA,CAAA;AAAA,EAC3D;AACA,EAAA,OACE,CAAA,8BAAA,EAAiC,MAAM,WAAW;;AAAA;AAAA,EACX,KAAK,SAAA,CAAU,KAAA,CAAM,OAAO,QAAA,EAAU,IAAA,EAAM,CAAC,CAAC;;AAAA;AAAA,CAAA,GAErF,MAAM,MAAA,CAAO,MAAA,CAAO,GAAA,CAAI,CAAC,MAAM,CAAA,EAAA,EAAK,CAAA,CAAE,IAAI,CAAA,EAAA,EAAK,EAAE,OAAO,CAAA,CAAE,CAAA,CAAE,IAAA,CAAK,IAAI,CAAA,GACrE;;AAAA,oDAAA,CAAA;AAEJ;AAGA,SAAS,OAAO,SAAA,EAA6B;AAC3C,EAAA,IAAI,SAAA,IAAa,OAAO,SAAA,KAAc,QAAA,IAAY,YAAY,SAAA,EAAW;AACvE,IAAA,OAAQ,SAAA,CAAkC,MAAA;AAAA,EAC5C;AACA,EAAA,OAAO,SAAA;AACT;AAOA,eAAsB,cAAA,CACpB,WAAA,EACA,OAAA,GAA2B,EAAC,EACH;AACzB,EAAA,MAAM,QAAA,GAAW,OAAA,CAAQ,QAAA,IAAa,MAAM,gBAAgB,OAAO,CAAA;AACnE,EAAA,MAAM,SAAA,GAAY,QAAQ,iBAAA,IAAqB,CAAA;AAC/C,EAAA,MAAM,MAAA,GAAA,CACH,OAAA,CAAQ,MAAA,IAAUC,qBAAA,KAClB,QAAQ,UAAA,GAAa;;AAAA;AAAA,EAA+B,OAAA,CAAQ,UAAU,CAAA,CAAA,GAAK,EAAA,CAAA;AAE9E,EAAA,IAAI,MAAA;AACJ,EAAA,IAAI,aAAyC,EAAC;AAE9C,EAAA,KAAA,IAAS,OAAA,GAAU,CAAA,EAAG,OAAA,IAAW,SAAA,GAAY,GAAG,OAAA,EAAA,EAAW;AACzD,IAAA,MAAM,KAAA,GAAsB,SAAS,EAAE,WAAA,EAAa,QAAQ,MAAA,EAAO,GAAI,EAAE,WAAA,EAAa,MAAA,EAAO;AAC7F,IAAA,MAAM,SAAA,GAAY,MAAM,QAAA,CAAS,OAAA,CAAQ,KAAK,CAAA;AAC9C,IAAA,MAAM,MAAA,GAASC,qBAAA,CAAe,MAAA,CAAO,SAAS,CAAC,CAAA;AAC/C,IAAA,IAAI,MAAA,CAAO,IAAI,OAAO,EAAE,QAAQ,MAAA,CAAO,KAAA,EAAO,UAAU,OAAA,EAAQ;AAEhE,IAAA,UAAA,GAAa,MAAA,CAAO,MAAA;AACpB,IAAA,IAAI,UAAU,SAAA,EAAW;AACzB,IAAA,MAAA,GAAS,EAAE,QAAA,EAAU,MAAA,CAAO,SAAS,CAAA,EAAG,MAAA,EAAQ,OAAO,MAAA,EAAO;AAAA,EAChE;AAEA,EAAA,MAAM,IAAIC,6BAAA;AAAA,IACR,CAAA,uCAAA,EAA0C,YAAY,CAAC,CAAA,UAAA,CAAA;AAAA,IACvD;AAAA,GACF;AACF;AAGA,eAAe,gBAAgB,OAAA,EAAmD;AAChF,EAAA,MAAM,EAAE,cAAA,EAAAC,eAAAA,EAAe,GAAI,MAAM,OAAA,CAAA,OAAA,EAAA,CAAA,IAAA,CAAA,OAAA,WAAA,EAAA,EAAA,cAAA,CAAA,CAAA;AACjC,EAAA,MAAM,OAA4C,EAAC;AACnD,EAAA,IAAI,OAAA,CAAQ,MAAA,KAAW,MAAA,EAAW,IAAA,CAAK,SAAS,OAAA,CAAQ,MAAA;AACxD,EAAA,IAAI,OAAA,CAAQ,KAAA,KAAU,MAAA,EAAW,IAAA,CAAK,QAAQ,OAAA,CAAQ,KAAA;AACtD,EAAA,OAAOA,gBAAe,IAAI,CAAA;AAC5B;AA9EaD,sCAAA,CAAA,CAgFAF;AA5Ib,IAAA,UAAA,GAAA,KAAA,CAAA;AAAA,EAAA,cAAA,GAAA;AAoBA,IAAA,WAAA,EAAA;AACA,IAAA,WAAA,EAAA;AAuCO,IAAME,6BAAA,GAAN,cAAoC,KAAA,CAAM;AAAA,MACtC,MAAA;AAAA,MACT,WAAA,CAAY,SAAiB,MAAA,EAAoC;AAC/D,QAAA,KAAA,CAAM,OAAO,CAAA;AACb,QAAA,IAAA,CAAK,IAAA,GAAO,uBAAA;AACZ,QAAA,IAAA,CAAK,MAAA,GAAS,MAAA;AAAA,MAChB;AAAA,KACF;AAyEO,IAAMF,qBAAA,GAAgB,CAAA;;AAAA;AAAA;AAAA;;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;AAAA;;AAAA;;AAAA;AAAA;AAAA;;AAAA;;AAAA,0KAAA,CAAA;AAAA,EAAA;AAAA,CAAA,CAAA","file":"index.cjs","sourcesContent":["/**\n * Claude provider — emits the schema through a forced tool call on the Anthropic\n * Messages API. Default model is Claude Opus 4.8 (`claude-opus-4-8`).\n */\nimport Anthropic from \"@anthropic-ai/sdk\";\nimport { buildPrompt, type ProposeInput, type SchemaProvider } from \"./index.js\";\n\n/** Claude Opus 4.8 — the most capable model. */\nexport const DEFAULT_CLAUDE_MODEL = \"claude-opus-4-8\";\n\nexport interface ClaudeProviderOptions {\n  readonly apiKey?: string;\n  /** Bring your own configured client (takes precedence over `apiKey`). */\n  readonly client?: Anthropic;\n  readonly model?: string;\n  readonly maxTokens?: number;\n}\n\nconst TOOL_NAME = \"emit_form_schema\";\n\nconst TOOL: Anthropic.Tool = {\n  name: TOOL_NAME,\n  description: \"Emit the complete Formwright form schema as a structured object.\",\n  input_schema: {\n    type: \"object\",\n    properties: {\n      schema: { type: \"object\", description: \"A complete, valid Formwright FormSchema.\" },\n    },\n    required: [\"schema\"],\n  },\n};\n\nexport function claudeProvider(options: ClaudeProviderOptions = {}): SchemaProvider {\n  const client =\n    options.client ?? new Anthropic(options.apiKey ? { apiKey: options.apiKey } : undefined);\n  const model = options.model ?? DEFAULT_CLAUDE_MODEL;\n  const maxTokens = options.maxTokens ?? 16000;\n\n  return {\n    async propose(input: ProposeInput): Promise<unknown> {\n      const response = await client.messages.create({\n        model,\n        max_tokens: maxTokens,\n        system: input.system,\n        tools: [TOOL],\n        tool_choice: { type: \"tool\", name: TOOL_NAME },\n        messages: [{ role: \"user\", content: buildPrompt(input) }],\n      });\n      const toolUse = response.content.find(\n        (block): block is Anthropic.ToolUseBlock => block.type === \"tool_use\",\n      );\n      if (!toolUse) throw new Error(\"Claude did not emit a schema via the tool.\");\n      return toolUse.input;\n    },\n  };\n}\n","/**\n * OpenAI (GPT) provider — uses JSON mode on the Chat Completions API.\n *\n * To avoid a hard dependency on the `openai` package, pass your own client:\n *\n *   import OpenAI from \"openai\";\n *   import { generateSchema, openaiProvider } from \"@formwright/ai\";\n *   const { schema } = await generateSchema(\"a contact form\", {\n *     provider: openaiProvider({ client: new OpenAI() }),\n *   });\n *\n * Any client matching {@link OpenAILike} works (Azure OpenAI, compatible gateways).\n */\nimport { buildPrompt, type ProposeInput, type SchemaProvider } from \"./index.js\";\n\n/** Minimal structural shape of an OpenAI client's `chat.completions.create`. */\nexport interface OpenAILike {\n  chat: {\n    completions: {\n      create(body: {\n        model: string;\n        messages: Array<{ role: \"system\" | \"user\" | \"assistant\"; content: string }>;\n        response_format?: { type: \"json_object\" };\n      }): Promise<{ choices: Array<{ message: { content: string | null } }> }>;\n    };\n  };\n}\n\nexport interface OpenAIProviderOptions {\n  /** A configured OpenAI client (e.g. `new OpenAI()`). Required. */\n  readonly client: OpenAILike;\n  readonly model?: string;\n}\n\nexport function openaiProvider(options: OpenAIProviderOptions): SchemaProvider {\n  const model = options.model ?? \"gpt-4o\";\n  return {\n    async propose(input: ProposeInput): Promise<unknown> {\n      const response = await options.client.chat.completions.create({\n        model,\n        response_format: { type: \"json_object\" },\n        messages: [\n          { role: \"system\", content: input.system },\n          {\n            role: \"user\",\n            content: `${buildPrompt(input)}\\n\\nRespond with ONLY the JSON schema object.`,\n          },\n        ],\n      });\n      const content = response.choices[0]?.message?.content;\n      if (!content) throw new Error(\"OpenAI returned no content.\");\n      return JSON.parse(content) as unknown;\n    },\n  };\n}\n","/**\n * @formwright/ai — turn a natural-language form description into a *validated*\n * Formwright schema using an LLM.\n *\n *   import { generateSchema } from \"@formwright/ai\";\n *   const { schema } = await generateSchema(\"a signup form with a US-only state field\");\n *   new Form(schema).mount(el);\n *\n * Provider-agnostic: the validate → repair loop lives here, and the actual model\n * call is a pluggable {@link SchemaProvider}. A Claude provider ships built-in\n * (the default); an OpenAI provider and a custom-function provider let you use\n * GPT, Gemini, Mistral, a local model, or anything else — each through its own\n * native SDK, never a compatibility shim.\n *\n * Whatever the provider returns is checked with `@formwright/schema`'s validator;\n * if it's invalid, the precise issues are fed back for repair — so what you get\n * out always satisfies the runtime (or a thrown {@link SchemaGenerationError}).\n */\nimport { validateSchema, type FormSchema, type ValidationIssue } from \"@formwright/schema\";\n\nexport { claudeProvider, type ClaudeProviderOptions } from \"./claude.js\";\nexport { openaiProvider, type OpenAIProviderOptions, type OpenAILike } from \"./openai.js\";\n\n/** A pluggable model backend: produce a candidate schema object for a request. */\nexport interface SchemaProvider {\n  propose(input: ProposeInput): Promise<unknown>;\n}\n\nexport interface ProposeInput {\n  readonly description: string;\n  readonly system: string;\n  /** Present on a repair attempt: the previous (invalid) output and why it failed. */\n  readonly repair?: RepairContext;\n}\n\nexport interface RepairContext {\n  readonly previous: unknown;\n  readonly issues: readonly ValidationIssue[];\n}\n\nexport interface GenerateOptions {\n  /** Model backend. Defaults to Claude (needs `ANTHROPIC_API_KEY` or `apiKey`). */\n  readonly provider?: SchemaProvider;\n  /** Convenience for the default Claude provider when `provider` is omitted. */\n  readonly apiKey?: string;\n  readonly model?: string;\n  /** How many times to feed validation errors back for repair (default 2). */\n  readonly maxRepairAttempts?: number;\n  /** Override the Formwright DSL system prompt. */\n  readonly system?: string;\n  /** Extra guidance appended to the system prompt. */\n  readonly guidelines?: string;\n}\n\nexport interface GenerateResult {\n  readonly schema: FormSchema;\n  /** Number of model round-trips it took (1 = valid on the first try). */\n  readonly attempts: number;\n}\n\nexport class SchemaGenerationError extends Error {\n  readonly issues: readonly ValidationIssue[];\n  constructor(message: string, issues: readonly ValidationIssue[]) {\n    super(message);\n    this.name = \"SchemaGenerationError\";\n    this.issues = issues;\n  }\n}\n\n/** Wrap a plain async function as a provider (for Gemini, Mistral, local models, …). */\nexport function defineProvider(propose: (input: ProposeInput) => Promise<unknown>): SchemaProvider {\n  return { propose };\n}\n\n/** Build the user-facing instruction, including repair feedback when retrying. */\nexport function buildPrompt(input: ProposeInput): string {\n  if (!input.repair) {\n    return `Design a Formwright form for: ${input.description}`;\n  }\n  return (\n    `Design a Formwright form for: ${input.description}\\n\\n` +\n    `Your previous attempt was INVALID:\\n${JSON.stringify(input.repair.previous, null, 2)}\\n\\n` +\n    `Validation issues to fix:\\n` +\n    input.repair.issues.map((i) => `- ${i.path}: ${i.message}`).join(\"\\n\") +\n    `\\n\\nReturn a corrected schema that resolves every issue.`\n  );\n}\n\n/** Some providers wrap the schema under a `schema` key; accept either shape. */\nfunction unwrap(candidate: unknown): unknown {\n  if (candidate && typeof candidate === \"object\" && \"schema\" in candidate) {\n    return (candidate as { schema: unknown }).schema;\n  }\n  return candidate;\n}\n\n/**\n * Generate a validated {@link FormSchema} from a natural-language description.\n * Throws {@link SchemaGenerationError} if the model can't produce a valid schema\n * within `maxRepairAttempts`.\n */\nexport async function generateSchema(\n  description: string,\n  options: GenerateOptions = {},\n): Promise<GenerateResult> {\n  const provider = options.provider ?? (await defaultProvider(options));\n  const maxRepair = options.maxRepairAttempts ?? 2;\n  const system =\n    (options.system ?? SYSTEM_PROMPT) +\n    (options.guidelines ? `\\n\\nAdditional guidelines:\\n${options.guidelines}` : \"\");\n\n  let repair: RepairContext | undefined;\n  let lastIssues: readonly ValidationIssue[] = [];\n\n  for (let attempt = 1; attempt <= maxRepair + 1; attempt++) {\n    const input: ProposeInput = repair ? { description, system, repair } : { description, system };\n    const candidate = await provider.propose(input);\n    const result = validateSchema(unwrap(candidate));\n    if (result.ok) return { schema: result.value, attempts: attempt };\n\n    lastIssues = result.issues;\n    if (attempt > maxRepair) break;\n    repair = { previous: unwrap(candidate), issues: result.issues };\n  }\n\n  throw new SchemaGenerationError(\n    `Could not produce a valid schema after ${maxRepair + 1} attempts.`,\n    lastIssues,\n  );\n}\n\n/** Lazily construct the default (Claude) provider so OpenAI-only users don't need a key. */\nasync function defaultProvider(options: GenerateOptions): Promise<SchemaProvider> {\n  const { claudeProvider } = await import(\"./claude.js\");\n  const opts: { apiKey?: string; model?: string } = {};\n  if (options.apiKey !== undefined) opts.apiKey = options.apiKey;\n  if (options.model !== undefined) opts.model = options.model;\n  return claudeProvider(opts);\n}\n\nexport const SYSTEM_PROMPT = `You design forms as Formwright schemas — plain JSON, no code.\n\nA FormSchema has:\n- \"id\" (string), \"version\" (string, e.g. \"1.0\"), optional \"title\", and \"fields\" (non-empty array).\n- optional \"submit\": { \"endpoint\": { \"method\": \"POST\"|\"GET\"|\"PUT\"|\"PATCH\"|\"DELETE\", \"url\": string }, \"transform\"?, \"onSuccess\"?, \"onError\"? }.\n\nEach field has \"id\" (unique within its scope) and \"type\", plus optional \"label\", \"placeholder\", \"help\", \"description\", \"defaultValue\".\nField types:\n- \"text\" | \"email\" | \"password\" | \"number\" | \"textarea\"\n- \"checkbox\" | \"toggle\" (boolean; toggle renders as a switch)\n- \"select\" | \"radio\" — need \"options\": [{ \"label\": string, \"value\": string|number }]\n- \"group\" — a nested object; needs \"fields\": [ ...child fields ]. Produces an object in the payload.\n- \"collection\" — a repeatable list of objects; needs \"fields\", optional \"itemLabel\", \"addLabel\", \"minItems\", \"maxItems\". Produces an array of objects.\n\nValidation (optional): \"validation\": { \"kind\": \"string\"|\"number\", \"required\"?, \"min\"?, \"max\"?, \"minLength\"?, \"maxLength\"?, \"pattern\"?, \"format\"?: \"email\"|\"url\"|\"uuid\", \"message\"? }.\n\nConditional logic — data, not code — via \"visibleWhen\" / \"enabledWhen\" / \"requiredWhen\", a JSONLogic-style expression:\n{ \"==\": [a, b] }, \"!=\", \">\", \">=\", \"<\", \"<=\", { \"in\": [needle, haystack] }, { \"and\": [...] }, { \"or\": [...] }, { \"not\": x }, and { \"var\": \"fieldId\" } to read another field's value.\nNames resolve to a sibling first, then outward to the form root — so a field inside a group or collection row can react to an outer toggle.\n\nOther per-field options: \"omit\": true (keep in the UI but exclude from the payload), \"labelPosition\": \"start\"|\"end\" (checkbox/toggle), \"layout\" (\"accordion\" for group; \"cards\"|\"accordion\" for collection).\n\nProduce sensible labels, validation, and conditions matching the request. Prefer \"toggle\" for yes/no switches and add helpful placeholders. Output ONLY the schema object.`;\n"]}

package/dist/index.d.cts

@@ -0,0 +1,129 @@
+import { ValidationIssue, FormSchema } from '@formwright/schema';
+import Anthropic from '@anthropic-ai/sdk';
+
+/**
+ * Claude provider — emits the schema through a forced tool call on the Anthropic
+ * Messages API. Default model is Claude Opus 4.8 (`claude-opus-4-8`).
+ */
+
+interface ClaudeProviderOptions {
+    readonly apiKey?: string;
+    /** Bring your own configured client (takes precedence over `apiKey`). */
+    readonly client?: Anthropic;
+    readonly model?: string;
+    readonly maxTokens?: number;
+}
+declare function claudeProvider(options?: ClaudeProviderOptions): SchemaProvider;
+
+/**
+ * OpenAI (GPT) provider — uses JSON mode on the Chat Completions API.
+ *
+ * To avoid a hard dependency on the `openai` package, pass your own client:
+ *
+ *   import OpenAI from "openai";
+ *   import { generateSchema, openaiProvider } from "@formwright/ai";
+ *   const { schema } = await generateSchema("a contact form", {
+ *     provider: openaiProvider({ client: new OpenAI() }),
+ *   });
+ *
+ * Any client matching {@link OpenAILike} works (Azure OpenAI, compatible gateways).
+ */
+
+/** Minimal structural shape of an OpenAI client's `chat.completions.create`. */
+interface OpenAILike {
+    chat: {
+        completions: {
+            create(body: {
+                model: string;
+                messages: Array<{
+                    role: "system" | "user" | "assistant";
+                    content: string;
+                }>;
+                response_format?: {
+                    type: "json_object";
+                };
+            }): Promise<{
+                choices: Array<{
+                    message: {
+                        content: string | null;
+                    };
+                }>;
+            }>;
+        };
+    };
+}
+interface OpenAIProviderOptions {
+    /** A configured OpenAI client (e.g. `new OpenAI()`). Required. */
+    readonly client: OpenAILike;
+    readonly model?: string;
+}
+declare function openaiProvider(options: OpenAIProviderOptions): SchemaProvider;
+
+/**
+ * @formwright/ai — turn a natural-language form description into a *validated*
+ * Formwright schema using an LLM.
+ *
+ *   import { generateSchema } from "@formwright/ai";
+ *   const { schema } = await generateSchema("a signup form with a US-only state field");
+ *   new Form(schema).mount(el);
+ *
+ * Provider-agnostic: the validate → repair loop lives here, and the actual model
+ * call is a pluggable {@link SchemaProvider}. A Claude provider ships built-in
+ * (the default); an OpenAI provider and a custom-function provider let you use
+ * GPT, Gemini, Mistral, a local model, or anything else — each through its own
+ * native SDK, never a compatibility shim.
+ *
+ * Whatever the provider returns is checked with `@formwright/schema`'s validator;
+ * if it's invalid, the precise issues are fed back for repair — so what you get
+ * out always satisfies the runtime (or a thrown {@link SchemaGenerationError}).
+ */
+
+/** A pluggable model backend: produce a candidate schema object for a request. */
+interface SchemaProvider {
+    propose(input: ProposeInput): Promise<unknown>;
+}
+interface ProposeInput {
+    readonly description: string;
+    readonly system: string;
+    /** Present on a repair attempt: the previous (invalid) output and why it failed. */
+    readonly repair?: RepairContext;
+}
+interface RepairContext {
+    readonly previous: unknown;
+    readonly issues: readonly ValidationIssue[];
+}
+interface GenerateOptions {
+    /** Model backend. Defaults to Claude (needs `ANTHROPIC_API_KEY` or `apiKey`). */
+    readonly provider?: SchemaProvider;
+    /** Convenience for the default Claude provider when `provider` is omitted. */
+    readonly apiKey?: string;
+    readonly model?: string;
+    /** How many times to feed validation errors back for repair (default 2). */
+    readonly maxRepairAttempts?: number;
+    /** Override the Formwright DSL system prompt. */
+    readonly system?: string;
+    /** Extra guidance appended to the system prompt. */
+    readonly guidelines?: string;
+}
+interface GenerateResult {
+    readonly schema: FormSchema;
+    /** Number of model round-trips it took (1 = valid on the first try). */
+    readonly attempts: number;
+}
+declare class SchemaGenerationError extends Error {
+    readonly issues: readonly ValidationIssue[];
+    constructor(message: string, issues: readonly ValidationIssue[]);
+}
+/** Wrap a plain async function as a provider (for Gemini, Mistral, local models, …). */
+declare function defineProvider(propose: (input: ProposeInput) => Promise<unknown>): SchemaProvider;
+/** Build the user-facing instruction, including repair feedback when retrying. */
+declare function buildPrompt(input: ProposeInput): string;
+/**
+ * Generate a validated {@link FormSchema} from a natural-language description.
+ * Throws {@link SchemaGenerationError} if the model can't produce a valid schema
+ * within `maxRepairAttempts`.
+ */
+declare function generateSchema(description: string, options?: GenerateOptions): Promise<GenerateResult>;
+declare const SYSTEM_PROMPT = "You design forms as Formwright schemas \u2014 plain JSON, no code.\n\nA FormSchema has:\n- \"id\" (string), \"version\" (string, e.g. \"1.0\"), optional \"title\", and \"fields\" (non-empty array).\n- optional \"submit\": { \"endpoint\": { \"method\": \"POST\"|\"GET\"|\"PUT\"|\"PATCH\"|\"DELETE\", \"url\": string }, \"transform\"?, \"onSuccess\"?, \"onError\"? }.\n\nEach field has \"id\" (unique within its scope) and \"type\", plus optional \"label\", \"placeholder\", \"help\", \"description\", \"defaultValue\".\nField types:\n- \"text\" | \"email\" | \"password\" | \"number\" | \"textarea\"\n- \"checkbox\" | \"toggle\" (boolean; toggle renders as a switch)\n- \"select\" | \"radio\" \u2014 need \"options\": [{ \"label\": string, \"value\": string|number }]\n- \"group\" \u2014 a nested object; needs \"fields\": [ ...child fields ]. Produces an object in the payload.\n- \"collection\" \u2014 a repeatable list of objects; needs \"fields\", optional \"itemLabel\", \"addLabel\", \"minItems\", \"maxItems\". Produces an array of objects.\n\nValidation (optional): \"validation\": { \"kind\": \"string\"|\"number\", \"required\"?, \"min\"?, \"max\"?, \"minLength\"?, \"maxLength\"?, \"pattern\"?, \"format\"?: \"email\"|\"url\"|\"uuid\", \"message\"? }.\n\nConditional logic \u2014 data, not code \u2014 via \"visibleWhen\" / \"enabledWhen\" / \"requiredWhen\", a JSONLogic-style expression:\n{ \"==\": [a, b] }, \"!=\", \">\", \">=\", \"<\", \"<=\", { \"in\": [needle, haystack] }, { \"and\": [...] }, { \"or\": [...] }, { \"not\": x }, and { \"var\": \"fieldId\" } to read another field's value.\nNames resolve to a sibling first, then outward to the form root \u2014 so a field inside a group or collection row can react to an outer toggle.\n\nOther per-field options: \"omit\": true (keep in the UI but exclude from the payload), \"labelPosition\": \"start\"|\"end\" (checkbox/toggle), \"layout\" (\"accordion\" for group; \"cards\"|\"accordion\" for collection).\n\nProduce sensible labels, validation, and conditions matching the request. Prefer \"toggle\" for yes/no switches and add helpful placeholders. Output ONLY the schema object.";
+
+export { type ClaudeProviderOptions, type GenerateOptions, type GenerateResult, type OpenAILike, type OpenAIProviderOptions, type ProposeInput, type RepairContext, SYSTEM_PROMPT, SchemaGenerationError, type SchemaProvider, buildPrompt, claudeProvider, defineProvider, generateSchema, openaiProvider };

package/dist/index.d.ts

@@ -0,0 +1,129 @@
+import { ValidationIssue, FormSchema } from '@formwright/schema';
+import Anthropic from '@anthropic-ai/sdk';
+
+/**
+ * Claude provider — emits the schema through a forced tool call on the Anthropic
+ * Messages API. Default model is Claude Opus 4.8 (`claude-opus-4-8`).
+ */
+
+interface ClaudeProviderOptions {
+    readonly apiKey?: string;
+    /** Bring your own configured client (takes precedence over `apiKey`). */
+    readonly client?: Anthropic;
+    readonly model?: string;
+    readonly maxTokens?: number;
+}
+declare function claudeProvider(options?: ClaudeProviderOptions): SchemaProvider;
+
+/**
+ * OpenAI (GPT) provider — uses JSON mode on the Chat Completions API.
+ *
+ * To avoid a hard dependency on the `openai` package, pass your own client:
+ *
+ *   import OpenAI from "openai";
+ *   import { generateSchema, openaiProvider } from "@formwright/ai";
+ *   const { schema } = await generateSchema("a contact form", {
+ *     provider: openaiProvider({ client: new OpenAI() }),
+ *   });
+ *
+ * Any client matching {@link OpenAILike} works (Azure OpenAI, compatible gateways).
+ */
+
+/** Minimal structural shape of an OpenAI client's `chat.completions.create`. */
+interface OpenAILike {
+    chat: {
+        completions: {
+            create(body: {
+                model: string;
+                messages: Array<{
+                    role: "system" | "user" | "assistant";
+                    content: string;
+                }>;
+                response_format?: {
+                    type: "json_object";
+                };
+            }): Promise<{
+                choices: Array<{
+                    message: {
+                        content: string | null;
+                    };
+                }>;
+            }>;
+        };
+    };
+}
+interface OpenAIProviderOptions {
+    /** A configured OpenAI client (e.g. `new OpenAI()`). Required. */
+    readonly client: OpenAILike;
+    readonly model?: string;
+}
+declare function openaiProvider(options: OpenAIProviderOptions): SchemaProvider;
+
+/**
+ * @formwright/ai — turn a natural-language form description into a *validated*
+ * Formwright schema using an LLM.
+ *
+ *   import { generateSchema } from "@formwright/ai";
+ *   const { schema } = await generateSchema("a signup form with a US-only state field");
+ *   new Form(schema).mount(el);
+ *
+ * Provider-agnostic: the validate → repair loop lives here, and the actual model
+ * call is a pluggable {@link SchemaProvider}. A Claude provider ships built-in
+ * (the default); an OpenAI provider and a custom-function provider let you use
+ * GPT, Gemini, Mistral, a local model, or anything else — each through its own
+ * native SDK, never a compatibility shim.
+ *
+ * Whatever the provider returns is checked with `@formwright/schema`'s validator;
+ * if it's invalid, the precise issues are fed back for repair — so what you get
+ * out always satisfies the runtime (or a thrown {@link SchemaGenerationError}).
+ */
+
+/** A pluggable model backend: produce a candidate schema object for a request. */
+interface SchemaProvider {
+    propose(input: ProposeInput): Promise<unknown>;
+}
+interface ProposeInput {
+    readonly description: string;
+    readonly system: string;
+    /** Present on a repair attempt: the previous (invalid) output and why it failed. */
+    readonly repair?: RepairContext;
+}
+interface RepairContext {
+    readonly previous: unknown;
+    readonly issues: readonly ValidationIssue[];
+}
+interface GenerateOptions {
+    /** Model backend. Defaults to Claude (needs `ANTHROPIC_API_KEY` or `apiKey`). */
+    readonly provider?: SchemaProvider;
+    /** Convenience for the default Claude provider when `provider` is omitted. */
+    readonly apiKey?: string;
+    readonly model?: string;
+    /** How many times to feed validation errors back for repair (default 2). */
+    readonly maxRepairAttempts?: number;
+    /** Override the Formwright DSL system prompt. */
+    readonly system?: string;
+    /** Extra guidance appended to the system prompt. */
+    readonly guidelines?: string;
+}
+interface GenerateResult {
+    readonly schema: FormSchema;
+    /** Number of model round-trips it took (1 = valid on the first try). */
+    readonly attempts: number;
+}
+declare class SchemaGenerationError extends Error {
+    readonly issues: readonly ValidationIssue[];
+    constructor(message: string, issues: readonly ValidationIssue[]);
+}
+/** Wrap a plain async function as a provider (for Gemini, Mistral, local models, …). */
+declare function defineProvider(propose: (input: ProposeInput) => Promise<unknown>): SchemaProvider;
+/** Build the user-facing instruction, including repair feedback when retrying. */
+declare function buildPrompt(input: ProposeInput): string;
+/**
+ * Generate a validated {@link FormSchema} from a natural-language description.
+ * Throws {@link SchemaGenerationError} if the model can't produce a valid schema
+ * within `maxRepairAttempts`.
+ */
+declare function generateSchema(description: string, options?: GenerateOptions): Promise<GenerateResult>;
+declare const SYSTEM_PROMPT = "You design forms as Formwright schemas \u2014 plain JSON, no code.\n\nA FormSchema has:\n- \"id\" (string), \"version\" (string, e.g. \"1.0\"), optional \"title\", and \"fields\" (non-empty array).\n- optional \"submit\": { \"endpoint\": { \"method\": \"POST\"|\"GET\"|\"PUT\"|\"PATCH\"|\"DELETE\", \"url\": string }, \"transform\"?, \"onSuccess\"?, \"onError\"? }.\n\nEach field has \"id\" (unique within its scope) and \"type\", plus optional \"label\", \"placeholder\", \"help\", \"description\", \"defaultValue\".\nField types:\n- \"text\" | \"email\" | \"password\" | \"number\" | \"textarea\"\n- \"checkbox\" | \"toggle\" (boolean; toggle renders as a switch)\n- \"select\" | \"radio\" \u2014 need \"options\": [{ \"label\": string, \"value\": string|number }]\n- \"group\" \u2014 a nested object; needs \"fields\": [ ...child fields ]. Produces an object in the payload.\n- \"collection\" \u2014 a repeatable list of objects; needs \"fields\", optional \"itemLabel\", \"addLabel\", \"minItems\", \"maxItems\". Produces an array of objects.\n\nValidation (optional): \"validation\": { \"kind\": \"string\"|\"number\", \"required\"?, \"min\"?, \"max\"?, \"minLength\"?, \"maxLength\"?, \"pattern\"?, \"format\"?: \"email\"|\"url\"|\"uuid\", \"message\"? }.\n\nConditional logic \u2014 data, not code \u2014 via \"visibleWhen\" / \"enabledWhen\" / \"requiredWhen\", a JSONLogic-style expression:\n{ \"==\": [a, b] }, \"!=\", \">\", \">=\", \"<\", \"<=\", { \"in\": [needle, haystack] }, { \"and\": [...] }, { \"or\": [...] }, { \"not\": x }, and { \"var\": \"fieldId\" } to read another field's value.\nNames resolve to a sibling first, then outward to the form root \u2014 so a field inside a group or collection row can react to an outer toggle.\n\nOther per-field options: \"omit\": true (keep in the UI but exclude from the payload), \"labelPosition\": \"start\"|\"end\" (checkbox/toggle), \"layout\" (\"accordion\" for group; \"cards\"|\"accordion\" for collection).\n\nProduce sensible labels, validation, and conditions matching the request. Prefer \"toggle\" for yes/no switches and add helpful placeholders. Output ONLY the schema object.";
+
+export { type ClaudeProviderOptions, type GenerateOptions, type GenerateResult, type OpenAILike, type OpenAIProviderOptions, type ProposeInput, type RepairContext, SYSTEM_PROMPT, SchemaGenerationError, type SchemaProvider, buildPrompt, claudeProvider, defineProvider, generateSchema, openaiProvider };

package/dist/index.js

@@ -0,0 +1,3 @@
+export { SYSTEM_PROMPT, SchemaGenerationError, buildPrompt, claudeProvider, defineProvider, generateSchema, openaiProvider } from './chunk-I3HZFUC7.js';
+//# sourceMappingURL=index.js.map
+//# sourceMappingURL=index.js.map

package/dist/index.js.map

@@ -0,0 +1 @@
+{"version":3,"sources":[],"names":[],"mappings":"","file":"index.js"}

package/package.json

@@ -0,0 +1,42 @@
+{
+  "name": "@formwright/ai",
+  "version": "0.1.0",
+  "description": "Generate a validated Formwright schema from a natural-language description using Claude.",
+  "license": "MIT",
+  "repository": {
+    "type": "git",
+    "url": "git+https://github.com/aliarsalan177/formwright.git",
+    "directory": "packages/ai"
+  },
+  "homepage": "https://github.com/aliarsalan177/formwright#readme",
+  "bugs": "https://github.com/aliarsalan177/formwright/issues",
+  "type": "module",
+  "sideEffects": false,
+  "files": [
+    "dist"
+  ],
+  "exports": {
+    ".": {
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js",
+      "require": "./dist/index.cjs"
+    }
+  },
+  "main": "./dist/index.cjs",
+  "module": "./dist/index.js",
+  "types": "./dist/index.d.ts",
+  "dependencies": {
+    "@anthropic-ai/sdk": "^0.68.0",
+    "@formwright/schema": "0.1.0"
+  },
+  "publishConfig": {
+    "access": "public"
+  },
+  "scripts": {
+    "build": "tsup",
+    "dev": "tsup --watch",
+    "typecheck": "tsc --noEmit",
+    "test": "vitest run",
+    "clean": "rm -rf dist .turbo"
+  }
+}