Treat MiniMax M2 thinking as always on - #2238

Suggest stripping rather than raising here.

The MiniMax docs say the opposite of this message: "For M2.x models, thinking cannot be disabled; thinking: {"type": "disabled"} is accepted but thinking remains on." The API takes the parameter, it just ignores it.

Practical cost: this turns a currently-working call into a hard failure at LLM.__init__ (llm.py:250), which runs on every completion and on Test Connection — not only at save time. And the form schema has no thinking property, so it can only ever fire for SDK/API callers, who are the least served by a message that contradicts the provider.

The enable_thinking: false raise above is the one worth keeping — that's a user explicitly asking for something impossible.

This only ever deletes a None.

By the time it runs, thinking is guaranteed absent from adapter_metadata for M2 — the helper raises if it was passed, and the enable_thinking branch skips M2 — so Pydantic emits the declared field as None:

MiniMaxLLMParameters(model='minimax/MiniMax-M2.7', api_key='k').model_dump()
  → thinking key present: True | value: None

That's the same None the M3 path keeps and has shipped with since #2166. Popping it only for M2 leaves two shapes for "no thinking" (M2: key absent, M3: key None), which is an easy thing to trip over later.

Suggest deleting both lines. Re-validation stays safe — the helper short-circuits on thinking is None, confirmed with validate({'model': 'MiniMax-M2.7', 'api_key': 'k', 'thinking': None}).

The assertion at line 192 then becomes assert validated["thinking"] is None.

Suggest dropping this block.

It encodes "what is an M2 model" a third time — _is_minimax_m2_model (base1.py:557) and the prefix-stripping in validate_model are the other two — in a different language, with hand-expanded case-insensitivity because JSON Schema has no /i flag. Nothing fails if the two drift, so the next model family silently loses the hint while tests stay green.

One thing worth knowing before merging either way: I ran this schema through @rjsf/utils@5 + validator-ajv8@5, which is what RjsfFormLayout.jsx uses. The condition does resolve correctly — M2 gets const: true and the checkbox defaults to checked — but unchecking yields two errors, and the second is raw ajv output:

[".enable_thinking must be equal to constant", "must match \"then\" schema"]

transformErrors (RjsfFormLayout.jsx:177) has no if case, so must match "then" schema reaches the user. No other adapter schema in the repo uses allOf/if, so this would become the pattern others copy.

The reworded description just above already conveys always-on, and the backend still rejects enable_thinking: false. If we want form-time feedback later, ui:widget: hidden for M2 reads better than a checkbox that accepts only one value — but it needs the same duplication solved, so it belongs in its own change.

allOf[0] couples this to schema layout — add a second conditional and it silently asserts the wrong branch. The Draft202012Validator loop just below already covers the same rule through behaviour, which is the version worth keeping.

Also from jsonschema import Draft202012Validator at line 329 should sit at the top of the file per our import convention.

Both go away if the allOf block is dropped.