feat: add MiniMax provider support for prompt refinement

[octo-patch]

Summary

This PR adds MiniMax as an alternative LLM provider for Open-Sora's prompt refinement step, alongside the existing OpenAI/GPT-4o integration.

Changes

• opensora/utils/prompt_refine.py

  • Add _get_client(model) — routes to MiniMax API (OpenAI-compatible) when model starts with "MiniMax", otherwise uses OpenAI
  • Add has_minimax_key() — checks whether MINIMAX_API_KEY is set
  • Add _strip_think_tags(text) — strips <think>…</think> chain-of-thought blocks emitted by MiniMax reasoning models
  • Add _extra_tokens(model) — allocates 500 extra output tokens for MiniMax models to accommodate thinking tokens
  • Add refine_prompts_by_minimax() — convenience wrapper that defaults to MiniMax-M2.7
  • refine_prompt() and refine_prompts() now accept an optional model parameter (defaults to PROMPT_MODEL env var, falls back to "gpt-4o")

• tests/test_minimax_prompt_refine.py — 18 unit tests covering routing, API key validation, model selection, temperature constraints, think-tag stripping, and the refine_prompts_by_minimax() wrapper

Supported models

Model	Description
MiniMax-M2.7	Default — peak performance
MiniMax-M2.7-highspeed	Faster variant

Usage

# Use MiniMax for prompt refinement
export MINIMAX_API_KEY=your_key_here
export PROMPT_MODEL=MiniMax-M2.7

Or programmatically:

from opensora.utils.prompt_refine import refine_prompts_by_minimax

refined = refine_prompts_by_minimax(["a cat playing in snow"], type="t2v")

API references

• Chat (OpenAI Compatible): https://platform.minimax.io/docs/api-reference/text-openai-api