Heygen Avatar

HeyGen Avatar Provider

Client

lib/marketing-agency/heygen/client.ts → createHeyGenClient(config?) returns:

• createAvatarVideo(request) → { id, provider:'heygen', status }
• getVideoStatus(videoId) → { status: queued|processing|completed|failed, videoUrl? }
• waitForCompletion(videoId, { timeoutMs, intervalMs, sleep? }) → final job (bounded poll)

config: { apiKey?, baseUrl?, fetchImpl?, allowMockFallback? }. apiKey defaults to process.env.HEYGEN_API_KEY. fetchImpl is injectable for tests.

Request shape

{
  avatarId: string,
  script: string,                 // used only for the text→TTS path
  consent: HeyGenConsentMetadata, // REQUIRED — { subjectName, sourceRef, confirmedAt }
  audioUrl?: string,              // PREFERRED — ElevenLabs audio for lip-sync
  voiceId?: string,               // HeyGen voice for the text path (no audioUrl)
  dimension?: { width, height },  // default 1280×720
  title?: string,
}

Voice pairing (preferred path)

ElevenLabs is the voice source (lib/services/ai/voice-generation.ts → generateSpeech returns audioBase64). The composition helper lib/marketing-agency/studio/avatar-video.ts (generateAvatarVideo) does: generateSpeech → upload to a public URL → createAvatarVideo({ audioUrl }). Prefer this (real cloned voice + lip-sync) over HeyGen's built-in text TTS.

Hard rules (NEVER)

• NEVER call createAvatarVideo without consent — it throws HeyGenConsentError by design.
• NEVER hard-code the API key — HEYGEN_API_KEY lives in Vercel env only.
• NEVER make a silent mock stand in for a live call. No key → HeyGenConfigurationError (matches the Artlist provider gate). The mock job is returned only when allowMockFallback: true is explicitly passed (dev/preview).
• NEVER publish a video still in queued/processing — poll to completed first.

Status mapping

HeyGen completed|success → completed; failed|error → failed; pending|waiting|processing → processing.

Cost awareness

HeyGen renders + ElevenLabs synthesis are billed per use. Track cost-of-generation per client (margin vs retainer) — see client-content-studio. Don't loop renders without a cap.

Testing

Inject fetchImpl (and sleep for waitForCompletion) — never hit the network in unit tests. See tests/unit/lib/heygen-client.test.ts.

Anti-patterns

• ❌ Throwing away the consent gate to "simplify" — it's a legal/likeness control.
• ❌ Using HeyGen text TTS when an ElevenLabs cloned voice is available (off-brand voice).
• ❌ Blocking a serverless function on a long synchronous poll — prefer queue + status check.

Relation to the grounded media system (REAL IMAGES ONLY)

The HeyGen avatar path is a consented-likeness exception — a different modality (talking heads) — that sits alongside, never replaces, the grounded pipeline. It is registered in the Real Images Only spec's sanctioned-exceptions ledger with an open item: consent/ allowlist enforcement is a separate slice, not yet folded into the reference-library grounding checks — don't assume createAvatarVideo inherits generateImage's BLOCKED semantics; its own consent gate (HeyGenConsentError) is the control that applies here.

Any accompanying imagery (thumbnails, backgrounds, b-roll, seed frames) MUST come from lib/services/ai/image-generation.ts generateImage() (or the generate_image MCP tool), grounded-by-default on public/reference-library/ + the private bucket reference-library-private — direct image-provider calls fail the CI guard test tests/unit/ai/no-direct-image-apis.test.ts. Non-avatar video goes through the grounded default in lib/services/ai/video/generation-service.ts (seeds from owned refs; GroundingBlockedError when no coverage). Completed HeyGen renders should be fed to the media worker (Railway ffmpeg) / frame-extraction scripts to grow the owned corpus.

Visual generation (binding): see .claude/rules/real-images-only.md + grounded-visuals. Direct provider calls fail CI.

Foundation & Gate Wiring (SYN-1050)

Adopted from the senior-skill standard so every artefact this skill produces is checked against the locked foundation before it lands.

Reads at every invocation (never cached — re-read each run):

• .claude/memory/ceo-foundation.md — consent records, Aid Rule (Q3.1.1) on RestoreAssist, cross-client boundary (Phase 3.4), no fabricated client metrics.
• .claude/memory/verification-gates.md — gate state for any claim referenced.

Output gate: every client-facing artefact this skill produces routes through brand-voice-enforce before the CEO batched-review queue. A REJECT blocks the artefact until the quoted offending string is fixed.

Evidence standard: every quantitative or factual claim carries exactly one tag — [VERIFIED] / [INFERENCE] / [UNCONFIRMED]. Untagged = defect (.claude/rules/fabel-evidence-standard.md). Never state a projected result as fact.

Spec: see spec.md in this skill directory.