[Feature] 内置 TokenRhythm（基元律动）模型服务商 / Add TokenRhythm as a built-in provider #1376

提交前确认 / Before submitting

• 我已经搜索过现有 Issue，确认这不是重复建议。
• 我已经移除截图、配置和示例中的密钥及隐私信息。

当前版本 / Current version

dev @ f1bbcd8 (2026-10-06). This proposal is based on the source; I haven't run it on a device yet.

使用环境 / Environment

N/A. This is a model provider proposal, not tied to a device. The source was reviewed on Windows 11.

功能领域 / Feature area

Model or provider / 模型或服务商

要解决的问题 / Problem or motivation

TokenRhythm（基元律动） is an OpenAI-compatible model aggregator. It focuses on open-weight models from Chinese labs: DeepSeek V4, GLM 5.x, Qwen 3.7/3.8, Kimi K2.x, Doubao Seed 2.1, MiniMax, MiMo and LongCat. All 19 online chat models support tool calling.

Today Operit users have to add it under "其他供应商" and fill in the endpoint by hand. Then the thinking switch either does nothing or breaks requests:

• GLM-5.3, GLM-5.3-Flash, Kimi-K2.7-Code, LongCat-2.0 and MiniMax-M2.7 return HTTP 400 "当前模型必须开启深度思考" whenever any thinking-disable field is sent.
• deepseek-v4-pro-0813 and deepseek-v4-flash-0731 do not think unless thinking.type is set to enabled.

使用场景 / Use case

A user with a TokenRhythm key picks "TokenRhythm" in 模型聚合 and pastes the key. They chat with deepseek-v4-pro-0813 using tools, and switch thinking off for quick replies. The same user switches to glm-5.3, and requests keep working because the thinking switch is locked on for that model.

期望方案 / Proposed behavior

Add TOKENRHYTHM as a built-in provider, following the Novita (#425) and MiniMax (#987) pattern:

• endpoint https://tokenrhythm.studio/v1/chat/completions
• default model deepseek-v4-pro-0813
• routed through the existing OpenAI-compatible provider
• listed under 模型聚合

Thinking rules in ModelThinkingConfigDefaults based on live tests against the production API:

可接受结果 / Expected outcome

• "TokenRhythm" appears in 模型聚合, and the model list loads from /models.
• DeepSeek models offer an on/off switch plus low/high/max effort.
• Qwen and Seed models offer an on/off switch.
• Always-on models show thinking locked on, and never send a disable field (no HTTP 400).

截图、草图或参考图 / Screenshot, mockup or reference

• API docs: https://tokenrhythm.studio/docs/api-integration
• Public model catalog (no auth needed): https://tokenrhythm.studio/api/models

替代方案与补充信息 / Alternatives and context

The workaround today is "其他供应商" with a hand-written thinking config per model. I have an implementation ready (about 16 files, mostly one string per locale) and can open a PR to dev once this looks acceptable.

🤖 Generated with Claude Code