[octo-patch]

Reason: Add MiniMax as a first-class provider with MiniMax-M3.1-Flash as the default and MiniMax-M3 as the second ordinary choice, while preserving saved and manually entered model IDs.

Changes:

• Add the MiniMax Chat Completions preset with global and China-region endpoints and 1,000,000-token context metadata for both current models.
• Seed the dashboard picker from the current catalog and keep discovery from reintroducing older ordinary choices. Existing saved/custom IDs and external-adapter aliases are not migrated.
• Use adaptive thinking for the current models in the built-in SDK/terminal path. Flash defaults to max effort, accepts the documented five effort values, and rejects disabled thinking, unsupported effort, service_tier, and reasoning_split=false; M3 accepts adaptive/disabled thinking and omits effort.
• Preserve existing image and tool-history handling and the explicit Responses adapter. This change does not add video input or alter external-adapter behavior.

Validation:

• Node 24.21.0 (CI major) and Node 22.23.2: all 56 repository tests passed with test concurrency limited to 2.
• node tools/check.mjs, POSIX launcher syntax checks, and git diff --check passed.
• New tests drive the real local HTTP adapter with mocked upstream transport and the real AgentManager with an injected SDK query: both regional Chat Completions URLs, streaming/nonstreaming, thinking/effort, saved/legacy/custom IDs, images, tool history, and explicit Responses selection.
• Tests ran in a credential-free environment; no paid inference was performed. The official installed runtime/real executable smoke test was not run locally because heavy setup is paused by the workspace resource floor. The SDK query is mocked; cross-platform/real-runtime validation is left to the repository workflow.

References:

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