Skip to main content

21 · Adding a New Modality (Avatar, Music, etc.) — Exhaustive Guide

This guide covers introducing an entirely new modality category to NeuroLink — one that doesn't fit into existing slots (LLM chat, TTS, STT, Realtime, video, image-gen).

Concrete examples this guide enables:

  • Avatar / Lip-sync (D-ID, Synthesia, MuseTalk via Replicate, HeyGen)
  • Music generation (Suno, Udio, Beatoven, ElevenLabs Music, Lyria)
  • 3D generation (Tripo, Meshy, Rodin) — speculative
  • Sound effects (ElevenLabs SFX, Stable Audio) — speculative

The pattern follows what TTS / STT / Realtime did in commit 27a31c32 and what §A of 19-adding-video-provider.md extracts.

When to use this guide: when no existing modality / processor is a good home for the new capability. If you're tempted to put a music generator into TTSProcessor or a 3D model into the image-gen pathway, stop and use this guide instead.

The 11-step pattern​

Each new modality requires:

  1. Type file — src/lib/types/<modality>.ts (the <Modality>Handler interface, <Modality>Options, <Modality>Result)
  2. Processor utility — src/lib/utils/<modality>Processor.ts (registry + dispatch)
  3. Module directory — src/lib/<modality>/providers/ (handler classes)
  4. output.mode extension — add the new mode value to the union in src/lib/types/generate.ts
  5. output.<modality> config block — options shape under the output block
  6. Result block — result.<modality> field for output payloads
  7. Dispatcher in baseProvider.ts — route output.mode === "<modality>" to a new handler method
  8. Registration in providerRegistry.ts — register first-party handlers
  9. CLI surface — extend output-mode choice + new flags
  10. Test suite — test/continuous-test-suite-<modality>.ts + test:<modality> script
  11. Documentation — feature page, getting-started directory, provider-integration journal

Worked example: Avatar / Lip-sync​

This walkthrough adds the Avatar modality (D-ID as the first handler). Substitute "Avatar" → "Music" / "Audio" / "ThreeD" as needed.

Step 1 — Type file​

File: src/lib/types/avatar.ts — NEW.

/**
 * Avatar / Lip-sync Type Definitions
 *
 * Types for generating talking-head videos by combining a portrait image
 * with narration audio (or text + TTS provider).
 *
 * @module types/avatar
 */

import type { TTSAudioFormat } from "./tts.js";

/**
 * Output formats for avatar videos.
 */
export type AvatarVideoFormat = "mp4" | "webm" | "mov";

/**
 * Quality presets for avatar generation. Provider-specific mappings:
 * - D-ID: "standard" → 720p, "hd" → 1080p
 * - HeyGen: "standard" → 720p, "hd" → 1080p with enhancement
 * - MuseTalk (Replicate): single quality only; "hd" is no-op
 */
export type AvatarQuality = "standard" | "hd";

/**
 * Options for avatar video generation.
 */
export type AvatarOptions = {
  /** Source portrait image (Buffer or absolute file path) */
  image: Buffer | string;

  /**
   * Audio source. Either:
   * - Pre-recorded narration (Buffer or path) — direct lip-sync
   * - Text — handler must produce TTS first using `ttsProvider`
   */
  audio?: Buffer | string;
  text?: string;

  /** TTS provider when `text` is provided (default: "openai-tts") */
  ttsProvider?: string;
  /** Voice id passed through to the TTS provider */
  voice?: string;

  /** Avatar provider override */
  provider?: string;
  /** Output quality */
  quality?: AvatarQuality;
  /** Output format (default: "mp4") */
  format?: AvatarVideoFormat;
  /** Output file path (optional — if omitted, buffer is returned in result) */
  output?: string;
  /** Provider-specific options (cast to <Provider>AvatarOptions inside handlers) */
  [k: string]: unknown;
};

/**
 * Result of avatar generation.
 */
export type AvatarResult = {
  /** Generated video buffer */
  buffer: Buffer;
  /** Output format */
  format: AvatarVideoFormat;
  /** File size in bytes */
  size: number;
  /** Duration in seconds (if known) */
  duration?: number;
  /** Provider used */
  provider?: string;
  /** Performance metadata */
  metadata?: {
    latency: number;
    provider?: string;
    model?: string;
    [k: string]: unknown;
  };
};

/**
 * Handler contract for avatar generation providers.
 */
export type AvatarHandler = {
  /**
   * Generate a talking-head video.
   * Implementations enforce their own timeouts (recommended: 5 minutes).
   */
  generate(options: AvatarOptions): Promise<AvatarResult>;

  /** Validate the provider is configured (auth, base URL, etc.). */
  isConfigured(): boolean;

  /** Maximum supported audio length in seconds (provider-specific). */
  readonly maxAudioDurationSeconds?: number;

  /** Supported output formats. */
  readonly supportedFormats?: AvatarVideoFormat[];
};

CLAUDE.md compliance:

  • Rule 7 — uses type not interface ✓
  • Rule 8 — file is avatar.ts not avatarTypes.ts ✓
  • Rule 9 — types prefixed Avatar* (globally unique) ✓
  • Rule 11 — lives in src/lib/types/, not a local types directory ✓

Add to src/lib/types/index.ts: export * from "./avatar.js"; (rule 10 — barrel-only).

Step 2 — Processor utility​

File: src/lib/utils/avatarProcessor.ts — NEW.

Mirror src/lib/utils/ttsProcessor.ts:75-352. The shape is identical; substitute names:

import { logger } from "./logger.js";
import { ErrorCategory, ErrorSeverity } from "../constants/enums.js";
import { NeuroLinkError } from "./errorHandling.js";
import {
  SpanSerializer,
  SpanType,
  SpanStatus,
  getMetricsAggregator,
} from "../observability/index.js";
import type {
  AvatarHandler,
  AvatarOptions,
  AvatarResult,
} from "../types/index.js";

export const AVATAR_ERROR_CODES = {
  PROVIDER_NOT_SUPPORTED: "AVATAR_PROVIDER_NOT_SUPPORTED",
  PROVIDER_NOT_CONFIGURED: "AVATAR_PROVIDER_NOT_CONFIGURED",
  GENERATION_FAILED: "AVATAR_GENERATION_FAILED",
  POLL_TIMEOUT: "AVATAR_POLL_TIMEOUT",
  INVALID_INPUT: "AVATAR_INVALID_INPUT",
  AUDIO_TOO_LONG: "AVATAR_AUDIO_TOO_LONG",
} as const;

export class AvatarError extends NeuroLinkError {
  constructor(opts: {
    code: string;
    message: string;
    category?: ErrorCategory;
    severity?: ErrorSeverity;
    retriable?: boolean;
    context?: Record<string, unknown>;
    originalError?: Error;
  }) {
    super({
      code: opts.code,
      message: opts.message,
      category: opts.category ?? ErrorCategory.EXECUTION,
      severity: opts.severity ?? ErrorSeverity.HIGH,
      retriable: opts.retriable ?? false,
      context: opts.context,
      originalError: opts.originalError,
    });
    this.name = "AvatarError";
  }
}

export class AvatarProcessor {
  private static readonly handlers = new Map<string, AvatarHandler>();

  static registerHandler(providerName: string, handler: AvatarHandler): void {
    if (!providerName) throw new Error("Provider name required");
    if (!handler) throw new Error("Handler required");
    const key = providerName.toLowerCase();
    if (this.handlers.has(key)) {
      logger.warn(`[AvatarProcessor] Overwriting handler for: ${key}`);
    }
    this.handlers.set(key, handler);
    logger.debug(`[AvatarProcessor] Registered: ${key}`);
  }

  static supports(providerName: string): boolean {
    return providerName ? this.handlers.has(providerName.toLowerCase()) : false;
  }

  private static getHandler(providerName: string): AvatarHandler | undefined {
    return this.handlers.get(providerName.toLowerCase());
  }

  static async generate(
    provider: string,
    options: AvatarOptions,
  ): Promise<AvatarResult> {
    const span = SpanSerializer.createSpan(SpanType.AVATAR, "avatar.generate", {
      "avatar.provider": provider,
      "avatar.quality": options.quality,
      "avatar.format": options.format,
    });
    try {
      const handler = this.getHandler(provider);
      if (!handler) {
        throw new AvatarError({
          code: AVATAR_ERROR_CODES.PROVIDER_NOT_SUPPORTED,
          message: `Avatar provider "${provider}" is not registered. Available: ${Array.from(this.handlers.keys()).join(", ")}`,
          category: ErrorCategory.CONFIGURATION,
          severity: ErrorSeverity.HIGH,
          retriable: false,
        });
      }
      if (!handler.isConfigured()) {
        throw new AvatarError({
          code: AVATAR_ERROR_CODES.PROVIDER_NOT_CONFIGURED,
          message: `Avatar provider "${provider}" is not configured`,
          category: ErrorCategory.CONFIGURATION,
          severity: ErrorSeverity.HIGH,
          retriable: false,
        });
      }
      const result = await handler.generate(options);
      const ended = SpanSerializer.endSpan(span, SpanStatus.OK);
      getMetricsAggregator().recordSpan(ended);
      return result;
    } catch (err) {
      const ended = SpanSerializer.endSpan(
        span,
        SpanStatus.ERROR,
        err instanceof Error ? err.message : String(err),
      );
      getMetricsAggregator().recordSpan(ended);
      throw err;
    }
  }
}

Add SpanType.AVATAR to src/lib/types/span.ts so observability surfaces avatar operations as a distinct span category (mirrors what 27a31c32 did for SpanType.STT).

Step 3 — Module directory + first handler​

Directory: src/lib/avatar/ — NEW.

src/lib/avatar/
├── index.ts                      # Barrel re-exports
└── providers/
    ├── DIDAvatar.ts              # First handler implementation
    └── (future: HeyGenAvatar.ts, ReplicateAvatar.ts, etc.)

File: src/lib/avatar/providers/DIDAvatar.ts — NEW.

Skeleton:

import { ErrorCategory, ErrorSeverity } from "../../constants/enums.js";
import {
  AVATAR_ERROR_CODES,
  AvatarError,
} from "../../utils/avatarProcessor.js";
import type {
  AvatarHandler,
  AvatarOptions,
  AvatarResult,
  AvatarVideoFormat,
} from "../../types/index.js";
import { logger } from "../../utils/logger.js";

const DEFAULT_BASE_URL = "https://api.d-id.com";
const POLL_INTERVAL_MS = 3_000;
const TOTAL_TIMEOUT_MS = 5 * 60_000;

export class DIDAvatar implements AvatarHandler {
  public readonly maxAudioDurationSeconds = 60;
  public readonly supportedFormats: AvatarVideoFormat[] = ["mp4"];

  private readonly apiKey: string | null;
  private readonly baseUrl: string;

  constructor(apiKey?: string) {
    this.apiKey = (apiKey ?? process.env.D_ID_API_KEY ?? "").trim() || null;
    this.baseUrl = process.env.D_ID_BASE_URL ?? DEFAULT_BASE_URL;
  }

  isConfigured(): boolean {
    return this.apiKey !== null;
  }

  async generate(options: AvatarOptions): Promise<AvatarResult> {
    if (!this.apiKey) {
      throw new AvatarError({
        code: AVATAR_ERROR_CODES.PROVIDER_NOT_CONFIGURED,
        message: "D_ID_API_KEY not set",
        category: ErrorCategory.CONFIGURATION,
        severity: ErrorSeverity.HIGH,
        retriable: false,
      });
    }

    const startTime = Date.now();
    const sourceUrl = await this.uploadImage(options.image);
    const audioUrl = options.audio
      ? await this.uploadAudio(options.audio)
      : undefined;

    // POST /talks
    const submitResp = await fetch(`${this.baseUrl}/talks`, {
      method: "POST",
      headers: {
        Authorization: `Basic ${this.apiKey}`,
        "Content-Type": "application/json",
      },
      body: JSON.stringify({
        source_url: sourceUrl,
        script: audioUrl
          ? { type: "audio", audio_url: audioUrl }
          : {
              type: "text",
              input: options.text,
              provider: {
                type: "microsoft",
                voice_id: options.voice ?? "en-US-JennyNeural",
              },
            },
      }),
    });

    if (!submitResp.ok) {
      throw new AvatarError({
        code: AVATAR_ERROR_CODES.GENERATION_FAILED,
        message: `D-ID submit failed: ${submitResp.status} ${await submitResp.text()}`,
        category: ErrorCategory.EXECUTION,
        severity: ErrorSeverity.HIGH,
        retriable: submitResp.status >= 500,
      });
    }

    const { id } = await submitResp.json();
    const videoBuffer = await this.pollUntilComplete(id, startTime);

    return {
      buffer: videoBuffer,
      format: "mp4",
      size: videoBuffer.length,
      provider: "d-id",
      metadata: {
        latency: Date.now() - startTime,
        provider: "d-id",
        talkId: id,
      },
    };
  }

  private async pollUntilComplete(
    id: string,
    startTime: number,
  ): Promise<Buffer> {
    while (Date.now() - startTime < TOTAL_TIMEOUT_MS) {
      const statusResp = await fetch(`${this.baseUrl}/talks/${id}`, {
        headers: { Authorization: `Basic ${this.apiKey}` },
      });
      const data = await statusResp.json();
      if (data.status === "done") {
        const dl = await fetch(data.result_url);
        return Buffer.from(await dl.arrayBuffer());
      }
      if (data.status === "error") {
        throw new AvatarError({
          code: AVATAR_ERROR_CODES.GENERATION_FAILED,
          message: `D-ID generation failed: ${data.error?.description ?? "unknown"}`,
          category: ErrorCategory.EXECUTION,
          severity: ErrorSeverity.HIGH,
          retriable: false,
        });
      }
      await new Promise((r) => setTimeout(r, POLL_INTERVAL_MS));
    }
    throw new AvatarError({
      code: AVATAR_ERROR_CODES.POLL_TIMEOUT,
      message: `D-ID talk ${id} did not complete within ${TOTAL_TIMEOUT_MS / 1000}s`,
      category: ErrorCategory.TIMEOUT,
      severity: ErrorSeverity.MEDIUM,
      retriable: true,
    });
  }

  private async uploadImage(image: Buffer | string): Promise<string> {
    // POST /images, returns { url }
    // Accept Buffer or path
    // ... (see real impl pattern)
    return "https://...";
  }

  private async uploadAudio(audio: Buffer | string): Promise<string> {
    // POST /audios, returns { url }
    return "https://...";
  }
}

File: src/lib/avatar/index.ts — NEW.

export {
  DIDAvatar,
  DIDAvatar as DIDAvatarHandler,
} from "./providers/DIDAvatar.js";
// Re-export the processor for caller convenience
export {
  AvatarProcessor,
  AvatarError,
  AVATAR_ERROR_CODES,
} from "../utils/avatarProcessor.js";

Step 4 — Extend output.mode​

File: src/lib/types/generate.ts. Two locations (rules 100, 855 — both GenerateOptions shapes):

-    mode?: "text" | "video" | "ppt";
+    mode?: "text" | "video" | "ppt" | "avatar";

The output.mode union appears twice in this file (one in input options, one in result types). Update both. CLAUDE.md rule 5 mandates this be additive — never remove existing values.

Step 5 — output.avatar config block​

In the same file, add the per-mode config:

export type GenerateOptions = {
  // ... existing ...
  output?: {
    mode?: "text" | "video" | "ppt" | "avatar";
    format?: "json" | "structured" | "text" | "binary";
    video?: VideoOutputOptions;
    ppt?: PptOutputOptions;
+   avatar?: AvatarOptions; // imported from "./avatar.js"
  };
};

Step 6 — result.avatar field​

export type GenerateResult = {
  // ... existing ...
+ avatar?: AvatarResult;
  video?: VideoGenerationResult;
  ppt?: PptResult;
};

Step 7 — Dispatcher in baseProvider.ts​

File: src/lib/core/baseProvider.ts:815. Add an if branch alongside the existing video / ppt routing:

 try {
   if (options.output?.mode === "video") {
     return await this.handleVideoGeneration(options, startTime);
   }
+  if (options.output?.mode === "avatar") {
+    return await this.handleAvatarGeneration(options, startTime);
+  }
+  if (options.output?.mode === "ppt") {
+    return await this.handlePptGeneration(options, startTime);
+  }
   ...

Add the method (mirror handleVideoGeneration at line 1750):

private async handleAvatarGeneration(
  options: TextGenerationOptions,
  startTime: number,
): Promise<EnhancedGenerateResult> {
  const { AvatarProcessor, AvatarError, AVATAR_ERROR_CODES } =
    await import("../utils/avatarProcessor.js");

  const provider = options.output?.avatar?.provider ?? "d-id";
  const avatarOpts = options.output?.avatar;

  if (!avatarOpts) {
    throw new AvatarError({
      code: AVATAR_ERROR_CODES.INVALID_INPUT,
      message: "output.avatar config is required when output.mode is 'avatar'",
      retriable: false,
    });
  }

  // Resolve image (Buffer | path | URL handling, same as handleVideoGeneration)
  const imageBuffer = await this.resolveImage(avatarOpts.image);

  // Resolve audio: either provided directly or generated via TTS-then-pass
  let audioBuffer: Buffer | undefined;
  if (avatarOpts.audio) {
    audioBuffer = await this.resolveAudio(avatarOpts.audio);
  } else if (avatarOpts.text) {
    const { TTSProcessor } = await import("../utils/ttsProcessor.js");
    const ttsResult = await TTSProcessor.synthesize(
      avatarOpts.text,
      avatarOpts.ttsProvider ?? "openai-tts",
      { format: "mp3", voice: avatarOpts.voice },
    );
    audioBuffer = ttsResult.buffer;
  }

  const result = await AvatarProcessor.generate(provider, {
    ...avatarOpts,
    image: imageBuffer,
    audio: audioBuffer,
  });

  const baseResult: EnhancedGenerateResult = {
    content: avatarOpts.text ?? "",
    provider,
    model: provider, // avatar providers don't have separate model names
    usage: { input: 0, output: 0, total: 0 },
    avatar: result,
  };

  return await this.enhanceResult(baseResult, options, startTime);
}

Step 8 — Registration in providerRegistry.ts​

After the existing voice / video registration blocks (~line 670):

// ===== AVATAR HANDLER REGISTRATION =====
try {
  const { AvatarProcessor } = await import("../utils/avatarProcessor.js");
  const { DIDAvatar } = await import("../avatar/providers/DIDAvatar.js");
  AvatarProcessor.registerHandler("d-id", new DIDAvatar());
  logger.debug("Avatar handlers registered: d-id");
} catch (err) {
  logger.warn(
    `[ProviderRegistry] d-id avatar registration failed: ${err instanceof Error ? err.message : String(err)}`,
  );
}

The block is wrapped in its own try/catch — same fault-tolerance contract as voice. Future avatar handlers (HeyGen, MuseTalk via Replicate) add another try { ... } catch {} inside this block.

Step 9 — CLI surface​

File: src/cli/factories/commandFactory.ts. Two edits:

 // Output mode help text (line ~393)
- "Output mode: 'text' for standard generation, 'video' for video, 'ppt' for presentation",
+ "Output mode: 'text' for standard generation, 'video' for video, 'ppt' for presentation, 'avatar' for talking-head video",

 // Output mode dispatch (line ~2450)
+ if (argv.outputMode === "avatar") {
+   options.output = {
+     mode: "avatar" as const,
+     avatar: {
+       provider: argv.avatarProvider,
+       image: argv.avatarImage,
+       audio: argv.avatarAudio,
+       text: argv.avatarText,
+       voice: argv.avatarVoice,
+       quality: argv.avatarQuality,
+       format: argv.avatarFormat,
+     },
+   };
+ }

Add CLI flags to the option schema (around the existing --output-mode section):

"avatar-provider": {
  type: "string",
  description: "Avatar provider (d-id, replicate, heygen)",
},
"avatar-image": {
  type: "string",
  description: "Path to source portrait image",
},
"avatar-audio": {
  type: "string",
  description: "Path to narration audio (alternative to --avatar-text)",
},
"avatar-text": {
  type: "string",
  description: "Text for the avatar to speak (uses TTS internally)",
},
"avatar-voice": {
  type: "string",
  description: "TTS voice id when --avatar-text is used",
},
"avatar-quality": {
  type: "string",
  choices: ["standard", "hd"],
  description: "Avatar video quality",
},

Step 10 — Test suite​

File: test/continuous-test-suite-avatar.ts — NEW. Mirror test/continuous-test-suite-voice.ts shape:

import { NeuroLink } from "../src/lib/index.js";
import { logger } from "../src/lib/utils/logger.js";
import * as fs from "node:fs/promises";

const tests = [
  {
    category: "Avatar — D-ID",
    name: "generate() produces talking-head video from image + audio",
    fn: async () => {
      if (!process.env.D_ID_API_KEY) return true;
      const image = await fs.readFile("test/fixtures/portrait.jpg");
      const audio = await fs.readFile("test/fixtures/narration.mp3");
      const nl = new NeuroLink();
      const result = await nl.generate({
        provider: "vertex", // any LLM, not used for avatar mode
        output: {
          mode: "avatar",
          avatar: { provider: "d-id", image, audio, format: "mp4" },
        },
      });
      if (!result.avatar?.buffer) return false;
      // Validate MP4 magic bytes (ftyp at offset 4)
      return result.avatar.buffer.subarray(4, 8).toString() === "ftyp";
    },
  },
  // ... text-driven, registration check, error case, etc.
];

// Boilerplate runner — copy from continuous-test-suite-voice.ts

File: package.json — add test:avatar script:

 "test:voice": "pnpm exec tsx test/continuous-test-suite-voice.ts",
+"test:avatar": "pnpm exec tsx test/continuous-test-suite-avatar.ts",

Step 11 — Documentation​

docs/features/avatar.md — NEW​

Use docs/features/tts.md as the template. Sections:

  1. Overview — what avatar generation is, when to use it
  2. Quick Start — D-ID minimal example
  3. Supported Providers — table (D-ID, future HeyGen, MuseTalk)
  4. Input Options — image source (path/URL/Buffer), audio sources (direct vs TTS)
  5. Output Formats — mp4, webm, mov per provider
  6. Quality Tiers — standard / hd mappings per provider
  7. Streaming — note that avatar is async-only (no streaming)
  8. Pricing reference

docs/getting-started/providers/d-id.md — NEW​

Per-provider guide. Same template as elevenlabs.md.

docs/provider-integration/<NN>-avatar-integration.md — NEW (optional implementation journal)​

For non-trivial work, document the architectural decisions, the wire format, edge cases. Use 14-voice-speech-integration.md as the template.

Cross-reference updates​

FileUpdate
docs/features/index.mdAdd an "Avatar" link
docs/getting-started/providers/index.mdAdd the new providers
docs/reference/provider-comparison.mdAdd an Avatar section
README.mdMention the new modality
docs-site/sidebars.tsAdd the new pages

Generic shape — Music modality (for reference)​

Substitute "Avatar" → "Music" / "Audio" / "Music3D" in every step.

The Music modality differs from Avatar in two ways:

  1. No image input — MusicOptions takes prompt (text) and optional referenceAudio (Buffer).
  2. Variable output length — MusicResult.duration is the practical maximum, providers can return 30s–5min depending on subscription.

Concrete handlers to add (in priority order):

HandlerAPINotes
Beatoven (src/lib/music/providers/BeatovenMusic.ts)https://public-api.beatoven.aiAsync track gen + composition; complex auth
ElevenLabs Music (src/lib/music/providers/ElevenLabsMusic.ts)https://api.elevenlabs.io/v1/sound-generationDistinct from ElevenLabs TTS — different endpoint, different account billing.
Lyria 3 Pro (src/lib/music/providers/LyriaMusic.ts)https://generativelanguage.googleapis.com/v1beta/models/lyria-3-pro:generateContentGoogle Generative AI — auth via API key
Suno (src/lib/music/providers/SunoMusic.ts)(no public API yet — speculative)
Udio (src/lib/music/providers/UdioMusic.ts)(no public API yet — speculative)

The ElevenLabs Music endpoint is distinct from the ElevenLabs TTS endpoint. Naming: ElevenLabsMusic (in src/lib/music/providers/) vs ElevenLabsTTS (in src/lib/voice/providers/). The two share an ELEVENLABS_API_KEY env var (one ElevenLabs account); the handlers are independent.

Type-naming conflicts to avoid (CLAUDE.md rule 9)​

Globally unique type names with domain prefixes are enforced by the unique-type-names ESLint rule. For Avatar:

Don't useUse insteadReason
OptionsAvatarOptionsBare Options collides everywhere
HandlerAvatarHandlerConflicts with TTSHandler, STTHandler, VideoHandler
ResultAvatarResultConflicts with TTSResult, STTResult, VideoGenerationResult
VideoFormatAvatarVideoFormatConflicts with potential video-only VideoFormat
QualityAvatarQualityConflicts with TTSQuality

For Music: MusicOptions, MusicHandler, MusicResult, MusicFormat, MusicGenre, etc.

Validation gates​

pnpm run check
pnpm run lint
pnpm run build
pnpm run test:avatar    # the new modality test suite
pnpm run test:providers # cross-modality sanity
# Real API smoke test:
export D_ID_API_KEY=...
pnpm run cli generate --output-mode avatar \
  --avatar-provider d-id --avatar-image portrait.jpg --avatar-text "Hello world"

Common pitfalls​

PitfallFix
Tried to put Avatar in TTSProcessor because it produces audio-driven outputTTS is text → audio. Avatar is image + audio → video. Separate processor.
Tried to put Music in TTSProcessor because it produces audioTTS is voiced speech with prosody. Music is melodic / harmonic content. Separate processor.
Forgot to add SpanType.<MODALITY>Observability dashboards lose the new modality category
Bare Options, Handler, Result type namesESLint unique-type-names rule fails the build
Created src/lib/avatar/types/ instead of src/lib/types/avatar.tsESLint no-local-types-folder rule fails
Removed an existing output.mode valuePublic API break (CLAUDE.md rule 5). Always additive.
Forgot the second output.mode location in generate.tsType checking passes; runtime dispatch silently falls through to text mode
Did not implement TTS-pass-through for text → audio → avatarCaller must always provide pre-recorded audio; TTS-driven avatar generation requires the chain (TTS → audio → avatar)
Did not document in docs/features/index.mdModality is invisible to discovery; users won't find it
Did not add to provider-comparison.mdDecision-making docs stale immediately

When to NOT add a new modality​

If the new capability:

  • Maps to an existing modality with a different transport (e.g., a new TTS provider) → use the existing modality guide (16 etc.)
  • Is a tool, not a modality (e.g., a search-knowledge-base tool, a code-execution tool) → use custom tools (docs/sdk-custom-tools.md)
  • Is a transformation of existing output (e.g., subtitle burning on video) → ffmpeg pipeline / utility module under src/lib/adapters/<modality>/
  • Is one-off and unlikely to have multiple providers → custom tool or service module, not a full modality category

A new modality is justified when:

  • ≥2 providers in the space (so the registry pays for itself)
  • Distinct input/output shape from existing modalities
  • Caller-facing config that doesn't fit an existing output.mode

See also​