fix: skip thinking param for MiniMax Anthropic-compatible endpoints

[nintendomustdie]

MiniMax's Anthropic-compatible API does not support the thinking parameter. This change adds a MiniMax check to skip thinking when using minimax-cn provider.

[teknium1]

Thanks for the PR, but closing — this reverts a deliberate change from commit d442f25 ("fix: align MiniMax provider with official API docs", Apr 10) which removed exactly this guard after reading MiniMax's official docs.

Per https://platform.minimax.io/docs/api-reference/text-anthropic-api, MiniMax's Anthropic-compat endpoint does support the thinking parameter in manual mode (type=enabled + budget_tokens). The adapter already sends manual thinking (not adaptive) to MiniMax — see _supports_adaptive_thinking(model) branch in build_anthropic_kwargs. Tests test_minimax_m27_gets_manual_thinking and test_minimax_m25_gets_manual_thinking pin this behavior, and this PR would break them.

The only recent report of Messages.create() got an unexpected keyword argument 'thinking' was issue #5777, which turned out to be an old local anthropic Python SDK that didn't know the kwarg — not MiniMax's server rejecting it. That was fixed in #5782 by skipping thinking kwargs when the installed SDK doesn't support them. If you hit that error on hermes update, upgrading the anthropic package resolves it.

If you're seeing a real rejection from MiniMax's server (a 4xx HTTP response body, not a local TypeError), please open an issue with:

• the full HTTP response body
• which endpoint (api.minimax.io/anthropic vs api.minimaxi.com/anthropic)
• the installed anthropic SDK version (pip show anthropic)
• an updated link to the MiniMax docs if they've changed

A targeted fix would key on _requires_bearer_auth(base_url) rather than a model-name substring, so it handles all MiniMax Anthropic-compat deployments consistently.