[octo-patch]

Reason: NadirClaw's ordinary MiniMax choices do not include MiniMax-M3.1-Flash, and older LiteLLM catalogs reject its supported reasoning parameters before sending a request.

• Make minimax resolve to MiniMax-M3.1-Flash and expose Flash followed by MiniMax-M3 in the built-in MiniMax catalog.
• Add MiniMax to the setup wizard with current choices; retain legacy/custom metadata and explicit model requests without advertising them as ordinary choices.
• Preserve MiniMax reasoning history alongside tool calls and images, and allow caller-supplied thinking/effort parameters per request in streaming and non-streaming LiteLLM calls. Do not inject absent reasoning options.
• Add real SDK HTTP-boundary regressions for international, China, and custom endpoints, saved/legacy IDs, reasoning, tools, images, streaming, and the public chat route. Include LiteLLM's MCP tool adapter in dev dependencies.
• Tighten the existing Python 3.10 LiteLLM cap to <1.97.0: 1.97.0 fails response-model construction on that interpreter with unresolved reasoning types; 1.96.0 passes the transport suite.

Validation:

• Full GitHub CI passes on exact head 1f00a5cbd89d729355b0ec2b595d9c3dff988e7a: 966 passed, 1 skipped on each of Python 3.10, 3.11, and 3.12 (run).
• 274 relevant routing, setup, credentials, thinking, tools, streaming, and MiniMax tests pass on each of LiteLLM 1.82.0 and 1.106.0 (Python 3.11.15; pytest 9.1.1), and LiteLLM 1.96.0 on Python 3.10.12.
• These include 28 new regressions: one catalog test and 27 real HTTP/public-route cases. Only HTTP is mocked in the transport tests.
• Earlier local full-suite attempt in the lightweight environment: 890 passed, 45 failed, 5 errors, 2 skipped because sentence-transformers/its ML stack was unavailable. The complete CI environments now validate the full suite successfully.
• git diff --check passes. No paid inference or real API credentials were used.

References:

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

[doramirdor]

Reviewed. Verified against MiniMax docs: MiniMax-M3.1-Flash is listed with a 1M context, and the $0.15/$0.60 per-M pricing matches. M3's $0.60/$2.40 is the >512k tier, which is a conservative estimate as the comment says. The reasoning_content passthrough is limited to minimax, allowed_openai_params lists only the params that are present, M2.7 stays resolvable but is hidden from the wizard, and the 3.10 LiteLLM cap is tightened with a reason given. CI is green on 3.10–3.12 at head 1f00a5c. Merging, thanks.