neuroverse

Instructions

Implementation Reference

• npm/src/index.ts:76-126 (handler)

  The handler implementation for the `neuroverse_process` tool, which orchestrates language detection, intent extraction, safety checks, and optional execution.

  server.registerTool(
    "neuroverse_process",
    {
      title: "Process Multilingual Input",
      description: `Process mixed-language input through the full NeuroVerse pipeline.
  
  Pipeline: Language Detect → Normalise → Intent Extract → Safety Check → (optional) Execute
  
  Supported languages: Tamil, Hindi, Telugu, Kannada + English (code-switched).
  
  Args:
    - text (string): Raw user input, possibly code-switched
    - user_id (string): User / agent identifier (default: "anonymous")
    - execute (boolean): Whether to also execute the intent (default: true)
  
  Returns:
    JSON with keys: language, intent, safety, execution (if execute=true)
  
  Examples:
    - "anna indha file ah csv convert pannu" → detects Tamil+English, extracts convert_format
    - "report banao sales ka" → detects Hindi+English, extracts generate_report
    - "drop database production" → BLOCKED by safety layer`,
      inputSchema: ProcessInputSchema,
      annotations: {
        readOnlyHint: false,
        destructiveHint: false,
        idempotentHint: true,
        openWorldHint: true,
      },
    },
    async (params) => {
      const lang = detectLanguage(params.text);
      const intent = await extractIntent(lang.normalizedText);
      const safety = await checkSafety(intent, params.text);
  
      const result: Record<string, unknown> = {
        language: lang,
        intent,
        safety,
      };
  
      if (params.execute) {
        const exec = await executeIntent(intent, safety);
        result["execution"] = exec;
      }
  
      return {
        content: [{ type: "text" as const, text: JSON.stringify(result, null, 2) }],
      };
    }
  );

• npm/src/index.ts:58-74 (schema)

  Input validation schema (using Zod) for the `neuroverse_process` tool.

  const ProcessInputSchema = z
    .object({
      text: z
        .string()
        .min(1, "Text is required")
        .max(5000, "Text must not exceed 5000 characters")
        .describe("Raw user input (may be code-switched, e.g. Tamil+English)"),
      user_id: z
        .string()
        .default("anonymous")
        .describe("Identifier for the user / agent"),
      execute: z
        .boolean()
        .default(true)
        .describe("If true, also execute the extracted intent after safety check"),
    })
    .strict();