Route the MiniMax OpenAI-compatible endpoint (global api.minimax.io and
China api.minimaxi.com hosts, plus the MiniMax-M3/MiniMax-M2.7 model ids)
to the thinking_type disable branch so that disabling thinking writes
thinking:{type:"disabled"} and drops reasoning_effort, matching the
existing MiMo/deepseek/glm behavior.
AnthropicAdapter.Stream built thinking config (buildAnthropicThinkingConfig)
and wrote it into body only inside the `if len(body)==0` normal-construction
block. The RequestBodyOverride branch skipped it entirely — a disabled effort
on the override path left whatever thinking config the override body carried,
violating user intent and diverging from openai.go, where
applyOpenAIThinkingDisable runs unconditionally after both branches.
Also, on the normal path, disabled only wrote thinking:{type:disabled} but
left a stale output_config (set by a prior adaptive turn or by
AnthropicExtraParams) in place — an explicitly disabled request could still
carry output_config.effort=high, a contradictory payload.
Fix:
- Extract applyAnthropicThinkingConfig(body, req), called unconditionally
after the override/normal block, mirroring openai.go:1860.
- disabled: force thinking:{type:"disabled"}, delete output_config, set
thinking_disabled_provider_param=thinking.type knob.
- adaptive (AnthropicThinkingEffort non-empty): write
thinking:{type:adaptive,display:summarized} + output_config.
- empty effort: no-op.
- buildAnthropicThinkingConfig retained for stable-message-count / message
normalization signals inside the normal block.
Tests: anthropic_thinking_disable_test.go covers all three branches, alias
normalization (off→disabled), disabled overriding existing adaptive config,
and the override-path symmetry scenario.
openAIThinkingDisableKind only matched dashscope/qwen/aliyun,
deepseek/bigmodel/z.ai/zhipu, and gpt-5.1+/gpt-6. Xiaomi MiMo
(api.xiaomimimo.com, model mimo-*) fell through to the default
branch, so when a user set reasoning=disabled the adapter wrote
no disable field and MiMo kept thinking enabled (extra latency +
reasoning token cost contrary to user intent).
Add xiaomimimo/mimo to the thinking_type case — MiMo uses the same
thinking:{type:"disabled"} shape as deepseek/glm, confirmed in
MiMo docs.
- openai.go: +2 conditions in openAIThinkingDisableKind
- openai_thinking_disable_test.go: cover MiMo (base/model/ultraspeed),
regress deepseek/glm/z.ai/qwen/gpt-5.1/gpt-6/unknown, and verify
applyOpenAIThinkingDisable writes thinking:{type:disabled} +
deletes reasoning_effort for MiMo
- Deleted outdated test files for OpenAI endpoint functionality.
- Improved the OpenAIEndpointURL function to better handle custom paths and version segments.
- Introduced a new helper function, stripEndpointVersionPrefix, to streamline endpoint processing.
- Introduced a new endpoint option for OpenAI integrations, allowing users to specify a custom path.
- Updated relevant components and validation logic to accommodate the new endpoint.
- Enhanced documentation and error messages to reflect the addition of the custom endpoint option.
上玄