feat(ai-models): add opt-in TwelveLabs Pegasus caption provider

[mohit-twelvelabs]

Hi! I'm Mohit, I work at TwelveLabs (@mohit-twelvelabs).

What this adds

An opt-in TwelveLabsAdapter in @twick/ai-models that generates time-aligned captions directly from a video URL using TwelveLabs Pegasus video understanding — no separate transcription step. It implements the existing ProviderAdapter contract (the package README already invites provider implementations like Sora/HeyGen/ElevenLabs), so it slots into ProviderRegistry / GenerationOrchestrator exactly like any other provider and emits the standard TimelineCaptionPatch.

import { ProviderRegistry, GenerationOrchestrator, TwelveLabsAdapter } from "@twick/ai-models";

const registry = new ProviderRegistry();
registry.registerAdapter(new TwelveLabsAdapter());
registry.setProviderConfig({ provider: "twelvelabs", apiKey: process.env.TWELVELABS_API_KEY });

const job = await new GenerationOrchestrator(registry).createJob({
  type: "caption",
  provider: "twelvelabs",
  input: { videoUrl: "https://example.com/clip.mp4", language: "en" },
});

Under the hood it creates a Pegasus analyze task (POST /analyze/tasks, time_based_metadata mode) on startJob, and the orchestrator polls getJobStatus (GET /analyze/tasks/{id}) until the task is ready, then maps each segment to a TimedTextSegment.

Why it helps this project

Captioning is a core editor workflow. Pegasus produces semantically-aware, time-aligned captions straight from a video URL, giving Twick apps an AI caption option behind the same provider-agnostic seam — and the orchestrator's built-in fallback means it can sit alongside other providers without lock-in.

Opt-in / non-breaking

• Adds "twelvelabs" to the AIModelProvider union and a new src/providers/twelvelabs-adapter.ts; touches no existing provider, default, or runtime path.
• No new runtime dependencies — uses native fetch (Node 20+), consistent with the rest of the repo.

How it was tested

• No-network unit test (twelvelabs-adapter.test.ts, node:test, zero new deps) covering caption parsing, metadata fallback, and malformed input — passing (3/3).
• pnpm --filter @twick/ai-models build (tsc typecheck + vite lib build + dts rollup) — passing.
• Live end-to-end against the real API: created a real Pegasus analyze task and polled it to ready. The returned result.data ({"caption":[{"start_time":0,"end_time":10,"metadata":{"text":"..."}}]}) parses correctly into { text, startMs: 0, endMs: 10000 }. Auth, both routes, request payload, and status transitions (queued → processing → ready) all verified.

A changeset (minor for @twick/ai-models) is included.

Per CONTRIBUTING, significant changes are encouraged to start with an issue — happy to open one or adjust the approach (segmentation prompt, sync vs async, etc.) to match maintainer preferences. You can grab a free API key at https://twelvelabs.io — there's a generous free tier.