[octo-patch]

Reason: New MiniMax connections and ordinary discovered text choices should prefer MiniMax-M3.1-Flash, with MiniMax-M3 retained, while explicit saved legacy/custom requests continue working.

Summary

• Default direct MiniMax connections to Flash then M3; translate direct global/China wire IDs and keep nontext and non-MiniMax entries unchanged. Filter older MiniMax text models from discovered routing and dashboard suggestions without rewriting saved provider lists or explicit requests.
• Apply Flash adaptive thinking and valid effort at the actual HTTP boundary, defaulting to max, removing unsupported service_tier and reasoning_split=false. Native Anthropic uses output_config.effort; OpenAI uses reasoning_effort. M3 retains adaptive/disabled thinking and drops effort tuning.
• Preserve tools, signed/native reasoning history, OpenRouter reasoning details, images/video, and local Responses follow-up history. Keep the existing Chat Completions backend; native MiniMax Responses support is not denied or newly implemented here.
• Scope direct context/output/pricing metadata to verified MiniMax hosts. OpenRouter's catalog verifies minimax/minimax-m3.1-flash:free, mandatory reasoning, zero rates, 1,000,000 context and 524,288 output. Gateway metadata remains gateway-sourced. Commonstack's unauthenticated catalog returns 401, so no Flash slug is guessed.
• Document endpoints, controls, Flash prices and conservative M3 upper-band estimates without inventing the exact 512k boundary. Existing M3/cache/tier support from this PR is retained.

Validation

• Full suite: 756 passed, 1 skipped (python -m pytest tests -q -p pytest_asyncio.plugin -p no:cacheprovider), including actual proxy HTTP serialization tests for direct global/China Chat, native Anthropic, OpenRouter stream/nonstream, local Responses image/tool/history round trips, saved/custom/legacy controls and discovery metadata.
• Focused provider/proxy/protocol/cache suite: 202 passed.
• Ruff 0.15.12 on changed Python modules/tests: passed. git diff --check and project secret scan: passed.
• Credential-free env -i, private venv, cached libraries read-only, CPU threads capped at two. HTTPX 0.28.1, Starlette 1.0.0, pytest 9.0.3, pytest-asyncio 1.3.0 and AnyIO 4.13.0 match uv.lock. Cached NumPy 2.4.6 and Uvicorn 0.47.0 differ from pinned 2.4.4/0.46.0. A full frozen dependency install was not completed under the shared disk resource gate. Initial cached-environment CLI/async failures were resolved by using the private interpreter and enabling the pinned async plugin; the final full run is green.
• No paid inference. No repository CI workflows/checks are configured; remote head was checked after push.

Sources: https://platform.minimax.io/docs/api-reference/text-openai-api.md ; https://platform.minimax.io/docs/guides/pricing-paygo.md ; https://openrouter.ai/api/v1/models (checked 2026-10-20).

[octo-patch]

I aligned MiniMax-M3 pricing with the provider defaults and updated the related metadata assertions. I ran 95 targeted tests, Ruff, and git diff --check.

[octo-patch]

Updated the MiniMax default model list to MiniMax-M3 and MiniMax-M2.7, added China endpoint provider detection, and pushed a new commit. I ran the full test suite (730 passed, 1 skipped), Ruff, and git diff --check.