feat: add TwelveLabs Pegasus as a video-understanding provider

[mohit-twelvelabs]

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

This PR adds TwelveLabs Pegasus as an opt-in video-understanding provider for camera and event analysis.

What it adds

• A new TwelveLabs provider that calls the Pegasus /analyze endpoint, wired into the ProviderFactory, config flow (new setup step + dropdown entry), constants, and English strings/translations — following the exact same pattern as the existing providers (it's API-key only, like Anthropic/Mistral).
• Default model pegasus1.5, endpoint, and error constants.
• Unit tests for the provider (no network) plus one opt-in live smoke test gated on TWELVELABS_API_KEY.

Why it helps this project

Pegasus is purpose-built for video understanding rather than independent stills. LLM Vision already extracts keyframes from videos, camera snapshots, and Frigate events, so this provider re-encodes those frames back into a short MP4 (using the ffmpeg binary the integration already requires) and sends that clip to Pegasus. That lets Pegasus reason about motion across the frames — a good complement to the existing frame-by-frame VLM providers for event summaries.

Opt-in / non-breaking

Nothing changes for existing users: no defaults are altered and no existing provider behavior is touched. Pegasus only runs if a user explicitly adds a TwelveLabs entry. No new Python dependency is introduced — the provider uses raw aiohttp exactly like the other HTTP providers.

How it was tested

• pytest tests/test_providers.py tests/test_config_flow.py — all pass (the live test is skipped without a key).
• Verified the full path end-to-end against the real API outside the test harness: extracted JPEG frames → ffmpeg MP4 → Pegasus /analyze returns HTTP 200 with a correct motion-aware description. Confirmed both the url and base64_string video-context paths work, and that Pegasus enforces max_tokens >= 512 (the provider clamps to this).
• Ran black on the changed files.

You can grab a free API key at https://twelvelabs.io — there's a generous free tier.

Notes

Per CONTRIBUTING this targets the latest v1.8.0-beta branch. I used AI assistance for boilerplate and to verify API behavior, but I understand and stand behind every line; happy to adjust anything to fit your conventions.

[valentinfrlch]

Hi @mohit-twelvelabs, thanks for creating this PR! Sounds very interesting, I'll take a look at it as soon as I have a little more time. Perhaps we can integrate it even more deeply, if that's something that would benefit accuracy.

[mohit-twelvelabs]

Thanks @valentinfrlch! No rush at all — and I'd love to integrate it more deeply if it helps accuracy. Pegasus can do targeted prompted analysis (e.g. structured event descriptions, scene Q&A), and Marengo embeddings could back semantic search over camera events if that's ever useful for the project — happy to follow your lead on the right shape.

In the meantime I fixed the failing validate-hassfest check in 27a9043: hassfest's TRANSLATIONS validator doesn't allow URLs in strings.json/translations/en.json, so I removed the https://twelvelabs.io links from the TwelveLabs step description and api_key description (now neutral "from the TwelveLabs dashboard" wording). CI should be green now.