clicktrack/lib/analysis/ai-ctp.ts

/**
 * AI-assisted CTP document generation
 *
 * Takes the results of BPM detection (and optional song metadata) and uses
 * Claude to produce a plausible, well-structured CTP document.
 *
 * Claude is asked to:
 *   - Divide the song into typical sections (Intro, Verse, Chorus, Bridge…)
 *   - Assign realistic start bars for each section
 *   - Note any tempo changes it would expect for the song/genre
 *   - Return a fully valid CTP 1.0 JSON document
 *
 * The caller should treat the result as a *draft* — the generated sections
 * are educated guesses and should be verified against the recording.
 */

import Anthropic from "@anthropic-ai/sdk";
import type { CTPDocument } from "@/lib/ctp/schema";

const client = new Anthropic();

// ─── Input / output types ─────────────────────────────────────────────────────

export interface AnalysisInput {
  bpm: number;
  duration: number;      // seconds
  title?: string;
  artist?: string;
  mbid?: string | null;
  contributedBy?: string;
}

// ─── JSON Schema for structured output ───────────────────────────────────────
// Must be strict (no additionalProperties, all required fields present).

const CTP_SCHEMA = {
  type: "object",
  additionalProperties: false,
  required: ["version", "metadata", "count_in", "sections"],
  properties: {
    version: { type: "string", enum: ["1.0"] },
    metadata: {
      type: "object",
      additionalProperties: false,
      required: [
        "title", "artist", "mbid", "duration_seconds",
        "contributed_by", "verified", "created_at",
      ],
      properties: {
        title:            { type: "string" },
        artist:           { type: "string" },
        mbid:             { type: ["string", "null"] },
        duration_seconds: { type: "number" },
        contributed_by:   { type: "string" },
        verified:         { type: "boolean" },
        created_at:       { type: "string" },
      },
    },
    count_in: {
      type: "object",
      additionalProperties: false,
      required: ["enabled", "bars", "use_first_section_tempo"],
      properties: {
        enabled:                 { type: "boolean" },
        bars:                    { type: "integer", minimum: 1, maximum: 8 },
        use_first_section_tempo: { type: "boolean" },
      },
    },
    sections: {
      type: "array",
      minItems: 1,
      items: {
        type: "object",
        additionalProperties: false,
        required: ["label", "start_bar", "time_signature", "transition"],
        // bpm is required for step, bpm_start/bpm_end for ramp — handled via oneOf
        // but we keep this schema simple (strict mode) and validate downstream with Zod.
        properties: {
          label:      { type: "string" },
          start_bar:  { type: "integer", minimum: 1 },
          bpm:        { type: "number" },
          bpm_start:  { type: "number" },
          bpm_end:    { type: "number" },
          transition: { type: "string", enum: ["step", "ramp"] },
          time_signature: {
            type: "object",
            additionalProperties: false,
            required: ["numerator", "denominator"],
            properties: {
              numerator:   { type: "integer", minimum: 1, maximum: 32 },
              denominator: { type: "integer", enum: [1, 2, 4, 8, 16, 32] },
            },
          },
        },
      },
    },
  },
};

// ─── System prompt ────────────────────────────────────────────────────────────

const SYSTEM_PROMPT = `\
You are an expert music producer and session musician assisting cover bands with click tracks.

You will receive automated BPM detection results for a song and must generate a CTP (Click Track Protocol) document describing the song's full tempo map.

CTP rules:
- "version" must be "1.0"
- sections[0].start_bar must be 1
- sections must be sorted by start_bar ascending, with no gaps
- Step sections have a single "bpm" field; ramp sections have "bpm_start" and "bpm_end" (no "bpm" field)
- All BPM values must be between 20 and 400
- time_signature.denominator must be a power of 2 (1, 2, 4, 8, 16, or 32)
- metadata.verified must be false (this is AI-generated, not human-verified)
- metadata.created_at must be an ISO 8601 datetime string

Guidelines for section layout:
- Use typical pop/rock section names: Intro, Verse, Pre-Chorus, Chorus, Bridge, Outro
- Estimate bar counts based on song duration and BPM (bars = duration_seconds × BPM / 60 / beats_per_bar)
- Most songs are 4/4; note any unusual meters if you know the song
- If you know the song has a tempo change (ritardando, double-time feel, key change with tempo shift), model it with a ramp or step section
- If unsure about sections, use a single constant-tempo section covering the whole song
- Use the detected BPM as the primary tempo — do not invent a different BPM unless the song is well-known to have a different tempo

The output is a draft for human review. Add reasonable section structure based on the song's typical arrangement.`;

// ─── Main function ────────────────────────────────────────────────────────────

export async function generateCTPWithAI(input: AnalysisInput): Promise<CTPDocument> {
  const { bpm, duration, title, artist, mbid, contributedBy } = input;

  const approxBars = Math.round((duration * bpm) / 60 / 4); // assuming 4/4

  const userMessage = `\
Generate a CTP document for the following song:

Title: ${title ?? "Unknown Title"}
Artist: ${artist ?? "Unknown Artist"}
MusicBrainz ID: ${mbid ?? "unknown"}
Detected BPM: ${bpm}
Duration: ${duration.toFixed(1)} seconds (~${approxBars} bars at 4/4)
Contributed by: ${contributedBy ?? "anonymous"}

Create a plausible section layout for this song. If this is a well-known song, use your knowledge of its actual arrangement. If not, use a sensible generic structure.`;

  const response = await client.messages.create({
    model: "claude-opus-4-6",
    max_tokens: 2048,
    thinking: { type: "adaptive" },
    system: SYSTEM_PROMPT,
    messages: [{ role: "user", content: userMessage }],
    output_config: {
      format: {
        type: "json_schema",
        schema: CTP_SCHEMA,
      },
    },
  });

  const textBlock = response.content.find((b) => b.type === "text");
  if (!textBlock || textBlock.type !== "text") {
    throw new Error("Claude did not return a text block");
  }

  let parsed: unknown;
  try {
    parsed = JSON.parse(textBlock.text);
  } catch {
    throw new Error(`Claude returned invalid JSON: ${textBlock.text.slice(0, 200)}`);
  }

  // Stamp the current timestamp if Claude left a placeholder
  const doc = parsed as CTPDocument;
  if (!doc.metadata.created_at || doc.metadata.created_at.includes("placeholder")) {
    doc.metadata.created_at = new Date().toISOString();
  }

  return doc;
}