Refresh minimax_video to the official MiniMax API (direct global/CN endpoints, Hailuo 2.3 contract)

[octo-patch]

Reason: Refresh minimax_video to the first-party MiniMax video API with direct global and CN endpoints and the current Hailuo 2.3 request/response contract.

minimax_video previously routed through a third-party gateway with gateway credentials and gateway-specific model slugs, so it lacked the direct MiniMax endpoints and the current first-party request/response contract. This refreshes the tool to call the official MiniMax video API directly.

Changes

• tools/video/minimax_video.py
  • Auth via MINIMAX_API_KEY with Authorization: Bearer <key> (was gateway credentials).
  • Region selection: MINIMAX_REGION=global (default, https://api.minimax.io) or cn (https://api.minimaxi.com), with an optional MINIMAX_BASE_URL override.
  • Current model slugs (MiniMax-Hailuo-2.3, MiniMax-Hailuo-2.3-Fast, MiniMax-Hailuo-02, and the T2V-01/I2V-01 families), default MiniMax-Hailuo-2.3.
  • Direct request/response contract: POST /v1/video_generation → poll GET /v1/query/video_generation → GET /v1/files/retrieve → download the returned URL. Request fields: model, prompt, first_frame_image (image-to-video), prompt_optimizer, fast_pretreatment, duration, resolution, callback_url. Errors surface base_resp.status_code/status_msg, with API-key redaction in error messages.
  • Idempotency key now covers the output-affecting fields (model, operation, first_frame_image, prompt_optimizer, fast_pretreatment, duration, resolution).
• .env.example and docs/PROVIDERS.md: document MINIMAX_API_KEY / MINIMAX_REGION / MINIMAX_BASE_URL and move minimax_video from the gateway entries to a dedicated direct-API section.
• tests/tools/test_minimax_video.py: new contract test covering registry discovery, availability, region routing, the text-to-video direct-API request/response flow, the image-to-video first-frame requirement, and base_resp error surfacing. HTTP is mocked with monkeypatch.setattr on the live requests module so the shared requests import stays intact for the rest of the suite.

Checks

• python -m pytest tests/tools/test_minimax_video.py -v — 6 passed
• python -m pytest tests/contracts/test_jimeng_video.py tests/tools/test_provider_model_defaults.py tests/tools/test_video_selector_routing.py -q — 76 passed (no regressions)
• python -m py_compile tools/base_tool.py tools/tool_registry.py tools/video/minimax_video.py tests/tools/test_minimax_video.py

[0xDevNinja]

Thanks for moving minimax onto the first-party API — the get_status fix (declaring env:MINIMAX_API_KEY instead of gating on FAL_KEY) and the redacted-key error path are both real improvements, and the text_to_video flow looks good end to end.

One merge blocker on the image_to_video path when minimax is reached through video_selector:

The schema swap renames the reference input from image_url to first_frame_image and drops image_url entirely, but video_selector only knows how to bridge a shared reference image into image_url:

# tools/video/video_selector.py (unchanged by this PR)
if adapted.get("operation") == "image_to_video" and adapted.get("reference_image_path"):
    tool_props = getattr(tool, "input_schema", {}).get("properties", {})
    if "image_url" in tool_props and "image_url" not in adapted:   # <- never true for minimax now
        adapted["image_url"] = upload_image_fal(adapted["reference_image_path"])

minimax still advertises supports["image_to_video"] = True, so the candidate filter (supports.get("image_to_video") or "image_url" in props or "reference_image_url" in props) keeps selecting it for image_to_video. But nothing in video_selector.py or tools/video/_shared.py maps any shared input to first_frame_image — I grepped both, no references. So a selector-routed image_to_video request (e.g. preferred_provider=minimax with reference_image_path) reaches execute() with first_frame_image unset and hard-fails:

image_to_video requires 'first_frame_image'.

This worked before the PR: the old image_url prop was populated by the selector's upload_image_fal bridge. So it's a regression in the selector→provider contract for image_to_video, not just a rename. It's surfaced (not silent) and fallback tools are returned, but the provider can no longer serve image_to_video via the selector.

Two ways to close it:

• teach the selector to populate first_frame_image for providers whose schema exposes it (parallel to the existing image_url branch), or
• accept reference_image_url / image_url as an alias in minimax and map it to first_frame_image inside execute().

Either one plus a selector→minimax image_to_video test would make this mergeable. text_to_video itself is fine — this is scoped to the image_to_video route.