llm-pi-ai: configurable reasoning-effort display names (e.g. On/Off for binary thinking) ｜ 思考等级支持配置显示名（如二值思考显示为 开/关）

[shin4]

背景

有些模型的思考只有「开 / 关」两档，没有强度分级。例如小米 MiMo V2.6（Pro / Flash）：官方文档中控制思考的唯一参数是 thinking.type（enabled / disabled，默认开启），没有 reasoning_effort。

用 llm-pi-ai 接入时，配置本身没有问题：

reasoningEfforts: { off: none, high: high }   # 配合 compat.thinkingFormat: deepseek、supportsReasoningEffort: false

请求层完全正确（Off → thinking.type: disabled，High → enabled），但界面上显示的是 Off / High，而用户实际面对的是「开 / 关」。

原因（0.1.7-rc.2）

• 等级 id 只能取 pi-ai 的固定集合（off / minimal / low / medium / high / xhigh / max，packages/llm/llm-pi-ai/src/catalog.ts:74），没有 on。
• 显示名由 id 首字母大写生成，不可配置（packages/llm/llm-pi-ai/src/adapter.ts:197）。
• LlmReasoningEffortInfo.name 是普通字符串，模型选择器直接把它当作标签渲染（packages/client/ui-model-selection/src/client/ModelSelect.tsx:102-116），中间没有按 id 本地化——这也是 [i18n] Request Chinese translation for reasoning effort options #6117 提到的中文缺失的根源。
• 插件无法补救：名称在适配器的 resolveModel 中生成，llm 服务校验后直接交给界面，中间没有可拦截的事件或插槽；locale 字典也不允许其他插件覆盖同一命名空间。

建议

A. 允许在配置里声明每档的显示名（部署方可控，向后兼容）

reasoningEfforts 的值除字符串外，也接受对象：

reasoningEfforts:
  off:  { wire: none, name: { zh: 关, en: Off } }
  high: { wire: high, name: { zh: 开, en: On } }

仍接受现有的字符串写法（high: high），未声明 name 时沿用现在的首字母大写规则。name 可以是字符串，也可以是按 locale 区分的对象。

B. 模型选择器按等级 id 本地化（同时解决 #6117）

• 在 ui-model-selection 的 locale 中加入 effort.off / low / medium / high / max … 以及 effort.providerDefault 的中文（目前中英文都是 "Default"），按 id 查找，未知 id 回退到适配器给的名称。
• 模型只提供 off + 一档时，自动显示为「关 / 开」（Off / On）。

A 让部署方为特定模型给出准确的名称；B 让所有模型默认就有本地化的显示。两者互不冲突，任选其一都能解决问题。

现状

我们在 Mio（基于 DeepSeek Harness 构建的 MiMo 桌面应用）中通过配置 reasoning: high 去掉了与 High 重复的 "Default" 项，菜单现在只剩 Off / High，但名称仍无法改成「开 / 关」。

相关讨论：#6117（等级名称本地化）、#6389（pi-ai 按模型声明思考等级）、#4627（llm-pi-ai 按模型设置默认思考等级）。

---

Background

Some models offer thinking as a plain on/off switch with no intensity levels. Xiaomi MiMo V2.6 (Pro / Flash) is one: per its official docs, the only thinking control is thinking.type (enabled / disabled, on by default) — there is no reasoning_effort.

Configuring it through llm-pi-ai works:

reasoningEfforts: { off: none, high: high }   # with compat.thinkingFormat: deepseek, supportsReasoningEffort: false

The wire is exactly right (Off → thinking.type: disabled, High → enabled), but the UI shows Off / High for what is really On / Off.

Why (0.1.7-rc.2)

• Level ids are limited to pi-ai's fixed set (off / minimal / low / medium / high / xhigh / max, packages/llm/llm-pi-ai/src/catalog.ts:74); there is no on.
• The display name is the capitalized id and cannot be configured (packages/llm/llm-pi-ai/src/adapter.ts:197).
• LlmReasoningEffortInfo.name is a plain string that the model picker renders as the label (packages/client/ui-model-selection/src/client/ModelSelect.tsx:102-116), with no per-id localization in between — also the root of the missing Chinese labels in [i18n] Request Chinese translation for reasoning effort options #6117.
• A plugin cannot fix it: names are produced in the adapter's resolveModel and passed through the llm service to the UI with no interceptable event or slot, and locale dictionaries reject a second registration for the same namespace.

Proposal

A. Let configuration declare a display name per level (deployment-controlled, backward compatible)

Accept an object as well as a string in reasoningEfforts:

reasoningEfforts:
  off:  { wire: none, name: { zh: 关, en: Off } }
  high: { wire: high, name: { zh: 开, en: On } }

The existing string form (high: high) keeps working, and a level without name keeps today's capitalized id. name may be a plain string or a per-locale object.

B. Localize effort labels by id in the model picker (also resolves #6117)

• Add effort.off / low / medium / high / max … keys (and a real Chinese effort.providerDefault, which is "Default" in both dictionaries today) to ui-model-selection, looked up by id with the adapter's name as the fallback for unknown ids.
• When a model offers off plus exactly one level, label them Off / On (关 / 开).

A gives deployments exact names for specific models; B gives every model localized labels by default. They are complementary; either one solves this.

Current state

In Mio, a MiMo desktop app built on DeepSeek Harness, we set reasoning: high to drop the "Default" entry that only repeated High, so the menu now shows Off / High — but the names still cannot say On / Off.

Related: #6117 (effort label localization), #6389 (per-model thinking levels for pi-ai routes), #4627 (per-model default thinking level in llm-pi-ai).