anonymize

import { tool, text, type ToolsProviderController } from "@lmstudio/sdk";
import { z } from "zod";
import { configSchematics } from "./configSchematics";
import { anonymize } from "./anonymizer";

export async function toolsProvider(ctl: ToolsProviderController) {
  const config = ctl.getPluginConfig(configSchematics);
  const configCustomTerms = config.get("alwaysRedact");

  const anonymizeText = tool({
    name: "anonymize_text",
    description: text`
      Redact personally identifiable information (PII) from a block of text.

      The tool DETECTS AUTOMATICALLY (with proper validation, not just regex):
        • Credit / bank card numbers (Luhn-validated)
        • French NIR / numéro de sécurité sociale (mod-97 checksum)
        • IBAN (mod-97 checksum)
        • French phone numbers (0X XX XX XX XX, +33 X XX XX XX XX, 0033…)
        • Email addresses

      The tool DOES NOT detect names. You (the model) must identify any personal
      names in the text and pass them in \`names\`. Pass full names like
      "Jean Dupont", not just first names, to avoid false matches against common
      words. You can also pass arbitrary strings to redact via \`custom_terms\`
      (e.g. company names, addresses, project codenames).

      Replacement uses STABLE TYPED PSEUDONYMS:
        \`[NOM_1]\`, \`[CB_1]\`, \`[NIR_1]\`, \`[IBAN_1]\`, \`[TEL_1]\`,
        \`[EMAIL_1]\`, \`[CUSTOM_1]\`, …
      Same value → same pseudonym throughout the text, so coreference is
      preserved (e.g. "Jean … Jean" both become "[NOM_1]"). The response
      includes a \`mapping\` from pseudonym to original value.

      Typical workflow: read a file with the filesystem plugin, call
      anonymize_text on its content, then write the redacted version back.
    `,
    parameters: {
      text: z.string().min(1).describe("The text to anonymize."),
      names: z.array(z.string().min(1)).optional().describe(
        "Personal names to redact. Pass full names ('Jean Dupont'), not just first names.",
      ),
      custom_terms: z.array(z.string().min(1)).optional().describe(
        "Other strings to redact: company names, addresses, project codenames, etc.",
      ),
      include_mapping: z.boolean().optional().describe(
        "Include the pseudonym → original value mapping in the response. Default true.",
      ),
    },
    implementation: (args) => {
      const result = anonymize(args.text, {
        names: args.names,
        customTerms: args.custom_terms,
        configCustomTerms,
      });
      return {
        anonymized: result.anonymized,
        counts: result.counts,
        ...(args.include_mapping === false ? {} : { mapping: result.mapping }),
      };
    },
  });

  return [anonymizeText];
}