[octo-patch]

Reason: Add current MiniMax presets with MiniMax-M3.1-Flash selected by default and MiniMax-M3 as the other ordinary choice.

The settings preset fills the endpoint and model only when Apply is clicked. Existing saved model IDs, legacy endpoints, and manual/custom configuration remain usable without automatic migration. The global preset uses https://api.minimax.io/v1; China uses the currently documented https://api.minimax.cn/v1.

Both rewrite paths continue to use Chat Completions. For Flash on those two direct endpoints, requests use low reasoning effort and a 16,384-token completion budget so thinking has room before the short answer. They omit the ignored temperature parameter. Other models and custom routes keep their existing request parameters. Model metadata reflects Flash's adaptive thinking, 1M context, 524,288 maximum output, and current global pricing; M3 retains its two standard input-length price tiers.

Validation:

• npm test in web/frontend: 2 catalog regressions passed without installing frontend dependencies.
• Changed Python files parse with Python 3.8 syntax; git diff --check and staged secret scan passed.
• Added HTTPX mocked-transport tests that inspect actual serialized requests from both rewrite callers, including regions, saved/custom IDs, and preserved multimodal/tool/reasoning history, plus settings persistence tests.
• Exact-head fork CI, commit 0bb477a7fb3cf3278698e335f8177d30f72ae315: Python 3.14 passed 193 tests with 2 skips and the offline prompt-bank check. Python 3.8 passed 191 tests with 2 skips, including all 29 added regressions, and failed two existing session-history tests using Path.write_text(newline=...).
• Unchanged upstream CI, base cf7deab2d5b28c1f70f5b5498461848b69024a15, reproduces the same two Python 3.8 failures. The build job was skipped because its Python dependency failed.
• Upstream PR CI requires maintainer approval. Local frontend/distribution builds remain paused by workspace capacity policy (latest shared snapshot about 19.8 GiB available versus the required 110 GiB). Local Python setup timed out on the shared install lock; its queued retry was canceled after remote validation completed. No paid inference was run.

References:

• https://platform.minimax.io/docs/api-reference/text-openai-api.md
• https://platform.minimaxi.com/docs/api-reference/text-openai-api.md
• https://platform.minimax.io/docs/guides/models-intro.md
• https://platform.minimax.io/docs/pricing/overview.md

[octo-patch]

I updated the China OpenAI-compatible preset and request handling to use https://api.minimaxi.com/v1 and pushed a new commit. I ran 29 focused Python tests and 2 frontend provider-catalog tests.