[PSA-Team-source]

Summary

Adds uberllm_video, a hosted video_generation provider (provider uberllm, env UBERLLM_API_KEY). One key reaches seven video model families: Seedance 2.0, Wan 2.7, Kling 3 Standard, Veo 3.1 Fast, LTX 2.3 Fast, Hailuo 02 Standard and Vidu Q3 Turbo, for text-to-video, image-to-video (first frame) and reference-to-video. The same key and prepaid balance also serve UberLLM's image, music and speech models, which is the gap #569 describes; this PR covers video only, and those could follow as separate tools if you want them.

Mirrors #905 (same file set and lifecycle: pinned contract, submit-then-poll, resume_job, settled cost, dry_run quote, Layer 3 skill).

Related issue

Refs #569

Changes

• tools/video/uberllm_video.py: UberLLMVideo(BaseTool), auto-discovered and ranked by video_selector. model takes a family (seedance-2.0 default, the only one with all three routes, first in models for routing) or an exact UberLLM model id.
  • Submit POST /v1/video/generations with an Idempotency-Key, poll GET /v1/video/generations/{id}, download with the same key (the key is only sent to the UberLLM host). retry_policy.max_retries=0.
  • failed, canceled and expired end without charge (expired is mapped to terminal, since provider_jobs.poll would otherwise keep waiting on it).
  • Reports the job's usage.cost_usd as cost_usd with data.cost_status settled | pending; while the charge is pending it re-reads the job and otherwise reports the listed price.
  • Every payload is checked against the model's pinned JSON Schema with unknown fields refused, so a length, resolution or setting the model lacks fails locally instead of being ignored.
  • estimate_cost from pinned per-second rates, PriceQuoteRequired for an unpinned resolution, no network. dry_run: true returns UberLLM's exact free quote (validate: true). max_cost_usd has the job refused at no charge above that.
  • Local frames and references go to UberLLM's file store (POST /v1/files). The schema does not advertise image_url, so video_selector does not send local frames to fal first.
• tools/provider_contracts/uberllm_video.json: 16 model ids in 7 families, each with the request schema from GET /v1/models/{id} and per-resolution rates from free validate quotes (highest rate across audio settings, so estimates do not undercount). Verified 2026-11-10.
• tests/contracts/test_uberllm_video.py: 45 tests, all HTTP faked.
• .agents/skills/uberllm-video/SKILL.md (Layer 3) plus a row in ai-video-gen; docs/PROVIDERS.md env line, provider section, pricing and coverage rows; .env.example.

Testing

• pytest tests/contracts/test_uberllm_video.py tests/contracts/test_pika_video.py tests/tools/test_video_selector_routing.py -q: 164 passed.
• pytest tests/ -q (Python 3.11, macOS): 2954 passed, 20 skipped, 8 failed. The 8 failures are in test_audio_enhance_default_container, test_caption_workspace_isolation and test_showcase_card_security, and fail the same way on main here because the local ffmpeg build has no drawtext filter or libvorbis.
• Live dry_run through UberLLMVideo().execute(...): estimates matched the quotes exactly (Seedance 2.0 text 5 s $0.76; Seedance 2.0 reference 5 s $0.2684 with aspect_ratio 16:9 applied by default).

Tested with fal-ai/vidu/q3/text-to-video/turbo: 1 s 360p clip through execute(), ffprobe 640x352 h264 + aac, 1.04 s, cost_status settled at $0.02975, same as the preflight estimate.

Prices per model: https://uberllm.dev/models

I work on UberLLM.

Checklist

• The change is focused on a single logical concern.
• I ran the relevant tests locally (make test-contracts / make test) where applicable.
• I updated docs/README if behavior or usage changed.
• No unrelated files (build artifacts, local config) are included in the diff.