feat(provider): add pipeline reasoning controls

docs/review/model-reasoning-control-design.md

@@ -0,0 +1,476 @@
+# 模型思考控制设计方案
+
+> 日期：2026-07-31
+> 状态：Phase 1 已审核并实现
+> 范围：LangBot 主仓库的模型配置、LiteLLM 请求层、Local Agent、Web 管理面板、监控与测试
+
+## 1. 结论
+
+建议为 LangBot 增加一套与厂商参数解耦的“思考策略”模型，并明确区分三个概念：
+
+1. **思考能力**：模型是否支持思考，以及支持开关、档位还是 token 预算。
+2. **思考策略**：一次请求选择厂商默认、关闭、开启或指定思考档位。
+3. **思考展示**：是否把模型返回的思考内容展示给最终用户。
+
+现有 `remove-think` 只属于第 3 类。它会过滤输出，但不会阻止模型思考，也不会降低思考 token、费用或延迟。新能力不应复用或改写这个字段。
+
+推荐实现原则：
+
+- 默认值为 `provider_default`，不向上游增加任何新参数，现有模型行为完全不变。
+- 用户显式选择的策略必须被准确执行；无法准确执行时返回明确错误，不静默降级。
+- LangBot 内部只保存统一策略，Provider 请求层负责翻译成各厂商参数。
+- `extra_args` 保留为高级逃生口，但不能成为主 UI 的思考配置方式。
+- 模型页只管理并展示能力；可写策略归属于 Local Agent 流水线，同一模型可在不同业务中使用不同思考量。
+- 原始 reasoning 数据与展示文本分开保存，保证多轮对话、工具调用和签名字段不丢失。
+
+## 2. 调研结论
+
+### 2.1 可验证资料
+
+本次结论基于以下可验证来源：
+
+- OpenAI 官方 Reasoning Guide：`reasoning.effort` 的可选值由模型决定，可包括 `none`、`minimal`、`low`、`medium`、`high`、`xhigh`、`max`；低档位偏向低延迟和低 token，高档位偏向质量。
+  - https://developers.openai.com/api/docs/guides/reasoning#reasoning-effort
+- LangBot 锁定的 LiteLLM `1.88.1` 实现。`uv.lock` 已锁定该版本，本地缓存中的适配代码可以确认 LangBot 实际依赖所支持的翻译行为。
+- LangBot 当前实现：模型级 `extra_args` 会在 `LiteLLMRequester._build_completion_args()` 中直接合并到 `acompletion()` 参数。
+
+Anthropic、Google 和 LiteLLM 的官方文档域名在本次环境中被浏览器策略禁止访问，因此下表中这些厂商的结论以 LiteLLM `1.88.1` 实际适配代码为准。实施前应再用对应厂商官方文档做一次参数范围核验，尤其是模型代际和允许值。
+
+### 2.2 厂商差异矩阵
+
+| Provider / 生态 | 可控制能力 | LiteLLM 1.88.1 统一入口 | 关键限制 | 建议支持级别 |
+| --- | --- | --- | --- | --- |
+| OpenAI | 思考档位，部分模型支持 `none` | `reasoning_effort` | 每个模型支持的档位不同，不能把 `none` 当成通用能力 | 首批完整支持 |
+| Anthropic | 旧模型使用 extended thinking + token budget；新模型可用 adaptive thinking + effort | `reasoning_effort` 或 `thinking` | `none` 表示不发送 thinking；新旧模型的映射不同 | 首批完整支持 |
+| Gemini | 2.x 主要映射为 `thinkingBudget`；3.x 主要映射为 `thinkingLevel` | `reasoning_effort` 或 `thinking` | Gemini 3 的 `none` 可能只能降到最低档，不能保证真正关闭 | 首批支持，但严格限制关闭语义 |
+| DeepSeek | 开启/关闭；当前适配不支持预算档位 | `thinking={type: enabled}`；非 `none` effort 会映射成开启 | 多轮思考模式要求回传 `reasoning_content` | 首批开关支持 |
+| xAI | 思考档位 | `reasoning_effort` | 仅 reasoning-capable 模型接受 | 首批完整支持 |
+| Ollama | `think` 布尔值；部分模型接受 low/medium/high | `reasoning_effort` | 非 gpt-oss 模型的档位可能退化为布尔开关 | 首批支持，按模型能力裁剪 UI |
+| OpenRouter | 聚合多厂商的 reasoning 参数 | `reasoning_effort`、`thinking` | 实际能力由路由后的模型决定 | 首批支持，能力未知时要求测试 |
+| Volcengine / Doubao | `thinking.type` 支持 enabled/disabled/auto | LiteLLM `volcengine` 适配器支持 `thinking` | LangBot 当前 manifest 使用 `openai`，不会进入该适配器 | 第二批，先修正路由并回归 |
+| Bailian / Qwen | 厂商兼容接口有独立思考开关/预算 | LiteLLM `dashscope` 适配器目前未提供统一 reasoning 映射 | LangBot 当前 manifest 使用 `openai`，只能通过高级参数透传 | 第二批，实施前核对官方字段 |
+| 其他 OpenAI-compatible 网关 | 取决于网关 | 尝试标准 `reasoning_effort` | 不能仅凭模型名推断完整能力 | 保守支持，默认不自动开启 |
+
+### 2.3 对 LangBot 的直接含义
+
+不能把这个功能实现成单一 `enable_thinking: bool`，原因如下：
+
+- 有的模型只有开关，有的模型只有档位，有的模型允许精确 token 预算。
+- 有的模型本身始终推理，只能降低思考量，无法真正关闭。
+- 同一个通用档位在不同厂商会映射成不同的实际预算。
+- 聚合网关和自定义 OpenAI-compatible 服务无法可靠地通过模型名识别能力。
+- “不展示思考内容”不等于“关闭思考”。
+
+## 3. 当前项目现状
+
+### 3.1 已有能力
+
+- `LLMModel.extra_args` 是 JSON 字段，Web 端已有通用高级参数编辑器。
+- `LiteLLMRequester` 会按“模型级 `extra_args`，再调用级 `extra_args`”的顺序合并参数。
+- LiteLLM 已统一处理多个 Provider 的 `reasoning_effort`、`thinking` 和返回的 `reasoning_content`。
+- `LocalAgentRunner` 的非流式、流式、工具调用和 fallback 路径都经过 `RuntimeProvider.invoke_llm*()`。
+- `remove-think` 已能控制 `<think>` 或独立 reasoning 内容是否进入展示文本。
+- Gemini 工具调用所需的 `provider_specific_fields` / thought signature 已有保留逻辑和单元测试。
+
+### 3.2 现有缺口
+
+- 管理员只能手写 `extra_args`，没有统一语义、能力提示和校验。
+- `remove-think` 名称容易被误解为关闭模型思考。
+- 模型扫描只识别 `vision` 和 `func_call`，没有 reasoning 能力。
+- 当前返回处理会把 `reasoning_content` 拼进 `<think>` 文本后删除原字段，可能损失多轮思考所需的结构化数据。
+- DeepSeek 思考模式需要在后续轮次回传 `reasoning_content`，当前链路不能保证完整保留。
+- Pipeline 只能选择模型，不能针对业务覆盖模型的思考策略。
+- 监控只记录总输入/输出 token，没有单独展示 reasoning token。
+- 部分 Provider manifest 仍声明为通用 `openai`，导致 LiteLLM 的厂商专用翻译器不会生效。
+
+### 3.3 预计改动地图
+
+| 层 | 主要文件 | 责任 |
+| --- | --- | --- |
+| 持久化 | `src/langbot/pkg/entity/persistence/model.py`、`src/langbot/pkg/persistence/alembic/versions/` | 新增 `reasoning_config` JSON 列和 Alembic 迁移 |
+| 模型服务 | `src/langbot/pkg/api/http/service/model.py` | CRUD 校验、冲突检测、测试模型时使用统一策略 |
+| HTTP 控制器 | `src/langbot/pkg/api/http/controller/groups/provider/models.py` | 继续复用现有模型路由，不新增平行 API |
+| 模型管理 | `src/langbot/pkg/provider/modelmgr/modelmgr.py` | 临时模型、数据库模型与扫描结果加载新字段 |
+| 请求抽象 | `src/langbot/pkg/provider/modelmgr/requester.py` | 定义能力查询和 reasoning 参数构建接口 |
+| LiteLLM 适配 | `src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py` | 能力识别、策略翻译、参数合并、reasoning 返回保留 |
+| Provider manifest | `src/langbot/pkg/provider/modelmgr/requesters/*.yaml` | 必要时修正 Provider 路由；相关变更放到独立阶段 |
+| Agent 调用 | `src/langbot/pkg/provider/runners/localagent.py` | 所有非流式、流式、工具调用、fallback 路径传递统一策略 |
+| Pipeline 元数据 | `src/langbot/templates/metadata/pipeline/ai.yaml` | 第二阶段加入 Pipeline 级覆盖 |
+| 输出配置 | `src/langbot/templates/metadata/pipeline/output.yaml` | 保留键名，澄清 `remove-think` 只控制展示 |
+| Web 类型/API | `web/src/app/infra/entities/api/index.ts`、`web/src/app/infra/http/BackendClient.ts` | 增加配置与能力响应类型 |
+| 模型 UI | `web/src/app/home/components/models-dialog/` | 能力标记、策略控件、校验、模型测试 |
+| i18n | `web/src/i18n/locales/` | 至少补齐英文、简体中文及项目已有覆盖语言 |
+| 测试 | `tests/unit_tests/provider/`、`web/tests/` | 翻译、服务、流式 round-trip、前端状态测试 |
+
+Phase 1 不修改 `langbot-plugin-sdk` 的公共实体或运行时协议。现有 `provider_message.Message.provider_specific_fields` 已可承载 Provider 原始 reasoning 数据；只有后续要把 reasoning 升级为跨插件公开实体时，才需要跨仓库 SDK 变更。
+
+## 4. 领域模型
+
+### 4.1 统一策略
+
+新增 `ReasoningConfig`，保存于 LLM 模型，Pipeline 可提供同结构覆盖。产品层只暴露一个离散档位：
+
+```json
+{
+  "level": "provider_default"
+}
+```
+
+字段定义：
+
+| 字段 | 类型 | 含义 |
+| --- | --- | --- |
+| `level` | `provider_default \| disabled \| enabled \| minimal \| low \| medium \| high \| xhigh \| max` | 同时表达开关和思考强度 |
+
+校验规则：
+
+- `provider_default`：不发送任何 reasoning 参数，保持厂商和模型默认行为。
+- `disabled`：明确关闭；仅当模型可真正关闭时允许保存/运行。
+- `enabled`：明确开启，但由 Provider 决定具体强度，适用于只有开关的模型。
+- `minimal` 到 `max`：明确开启，并指定强度；仅允许选择模型实际支持的档位。
+- 厂商的 `auto` 统一映射为 `provider_default`，不再增加一个重复状态。
+- 精确 token 预算不进入主数据结构。少数需要预算的场景继续通过高级参数配置，并由模型测试接口校验。
+
+### 4.2 能力描述
+
+沿用现有 `LLMModel.abilities`，新增 `reasoning` 能力标记。同时由后端在 API 返回中计算只读的 `reasoning_capabilities`：
+
+```json
+{
+  "supported": true,
+  "controls": ["toggle", "effort"],
+  "efforts": ["none", "low", "medium", "high"],
+  "can_disable": true,
+  "source": "litellm"
+}
+```
+
+设计约束：
+
+- `abilities` 仍是用户可编辑的粗粒度能力，符合现有 `vision`、`func_call` 模式。
+- `reasoning_capabilities` 不持久化，优先从 LiteLLM 模型元数据计算，避免模型升级后数据库残留过期能力。
+- 无法识别的自定义模型返回 `supported: null`、`source: unknown`，不猜测。
+- 用户可手动添加 `reasoning` ability，但未知能力模型必须先通过“测试模型”验证显式策略。
+- UI 只展示后端声明可用的控件；未知模型保留 Provider Default 和高级参数入口。
+
+### 4.3 持久化
+
+在 `llm_models` 表新增 JSON 列：
+
+```text
+reasoning_config JSON NOT NULL DEFAULT {"level":"provider_default"}
+```
+
+使用 Alembic 新迁移，不修改冻结的 legacy migration。
+
+该列作为已实现版本的兼容字段保留；新的模型页不再提供写入口，Local Agent 请求以流水线中按模型 UUID 保存的策略为准。
+
+不建议把内部策略塞进 `extra_args`，原因是当前 `extra_args` 会原样发送给 LiteLLM；使用保留键会让内部元数据泄漏到上游，并使高级参数与产品配置难以区分。
+
+## 5. 配置优先级与请求流程
+
+### 5.1 优先级
+
+```text
+Pipeline 当前候选模型策略
+        ↓ 缺少配置时固定为 provider_default
+Provider / 模型默认行为
+```
+
+请求参数合并顺序：
+
+```text
+基础参数
+  -> 模型 extra_args
+  -> 调用级 extra_args
+  -> 统一 reasoning 策略翻译结果（最后应用）
+```
+
+统一策略最后应用，可以确保流水线行为不受模型页历史设置影响。为了避免用户困惑，保存和测试时要检测 `extra_args` 中的冲突字段；当 `level != provider_default` 时，发现以下字段应直接报错：
+
+- `reasoning_effort`
+- `thinking`
+- `reasoning`
+- `extra_body` 内已知的 `thinking`、`enable_thinking`、`thinking_budget` 等字段
+
+当 `level == provider_default` 时继续允许这些高级参数，保证旧配置兼容。
+
+### 5.2 翻译层
+
+在 `pkg/provider/modelmgr/` 内新增独立的 reasoning 规范化模块，职责是：
+
+1. 读取当前流水线候选模型的请求级策略。
+2. 查询 `ProviderAPIRequester.get_reasoning_capabilities(model)`。
+3. 严格校验策略是否可以准确执行。
+4. 生成 LiteLLM 参数，不直接发 HTTP。
+5. 返回可观测的“最终生效策略”供日志和测试使用。
+
+建议接口：
+
+```python
+class ProviderAPIRequester:
+    def get_reasoning_capabilities(self, model: RuntimeLLMModel) -> ReasoningCapabilities: ...
+
+    def build_reasoning_args(
+        self,
+        model: RuntimeLLMModel,
+        config: ReasoningConfig,
+    ) -> dict[str, Any]: ...
+```
+
+LiteLLMRequester 默认优先生成统一参数：
+
+- 强度档位：`reasoning_effort=<level>`
+- 仅开启：`thinking={"type":"enabled"}` 或 Provider 等价参数
+- 关闭：优先 `reasoning_effort="none"`
+- 高级参数中的精确预算：`thinking={"type":"enabled","budget_tokens":N}`
+
+Provider 特例只放在 requester 翻译层，不进入 Pipeline 或平台适配器。
+
+### 5.3 Provider 特例
+
+- **Gemini 3**：如果 LiteLLM 能力表不能确认真正关闭，`disabled` 必须报“不支持关闭，可选择 Provider Default 或最低档”，不能把 `none` 静默映射成 low/minimal。
+- **DeepSeek**：所有非 `none` 档位最终都只是开启。能力 API 只返回 `toggle`，UI 不显示档位；多轮必须保存并回传 `reasoning_content`。
+- **Ollama**：仅对明确支持等级的模型展示 effort；其他模型只展示开关。
+- **OpenRouter**：以路由后的模型能力为准。模型未知时允许 Provider Default，显式策略必须通过测试接口。
+- **Volcengine**：使用 `thinking.type=enabled/disabled/auto`。应先让该 requester 进入 LiteLLM `volcengine` 适配器，或增加等价的明确翻译，不能依赖模型名。
+- **Bailian/Qwen**：作为第二批 Provider 专用翻译。实施前核对官方字段、模型范围、预算上下限和流式返回结构，不凭经验写接口。
+
+## 6. 返回数据与思考展示
+
+### 6.1 保留原始 reasoning
+
+当前 `LiteLLMRequester` 会读取 `reasoning_content`，将其拼接成 `<think>` 文本，再删除原字段。建议改为：
+
+```text
+上游 reasoning_content
+  ├─ 原样保存在 Message.provider_specific_fields.reasoning_content
+  └─ 根据 remove-think 决定是否渲染为 <think>...</think>
+```
+
+流式路径需要在 accumulator 中分别累计 `content` 与 `reasoning_content`，最终消息必须携带结构化 reasoning。不能只依赖已经渲染的 `<think>` 文本反向解析。
+
+这样可以同时满足：
+
+- `remove-think=true` 时用户看不到思考内容，但多轮协议仍能回传必要数据。
+- `remove-think=false` 时保持当前用户体验。
+- DeepSeek 多轮 thinking 不丢上下文。
+- Gemini thought signature、Anthropic thinking block 等 Provider 字段可以继续按结构化方式 round-trip。
+
+### 6.2 现有字段处理
+
+保留数据库和 Pipeline 配置键 `remove-think`，避免破坏兼容。Web 文案改为更准确的：
+
+- 中文：`向用户展示思考过程`
+- 英文：`Show reasoning process`
+
+UI 使用正向开关，保存时转换回 `remove-think = !showReasoning`。文案必须强调它只影响展示，不影响模型是否思考、token 或费用。
+
+## 7. Web 管理面板
+
+### 7.1 模型编辑
+
+模型页只承担能力管理和只读展示：
+
+1. `Reasoning` ability 复选框与 Vision、Function Calling 并列，供无法自动识别的自定义模型手动声明能力。
+2. 模型卡片使用简短图标或 badge 标识 reasoning 能力。
+3. 模型页不提供可写思考挡位，避免模型默认值与流水线策略形成两个控制源。
+
+### 7.2 Local Agent 流水线策略
+
+在 Local Agent 的主模型和每一个 fallback 模型下分别显示紧凑离散滑杆：
+
+1. `Provider 默认` 始终为首个选项；选择它时不向上游增加任何思考参数。
+2. 完整档位顺序为：`Provider 默认 / 关闭 / 开启 / 最低 / 低 / 中 / 高 / 极高 / 最大`。
+3. 前端只渲染后端为该模型返回的可用档位；仅开关模型显示 `Provider 默认 / 关闭 / 开启`。
+4. 模型不能真正关闭时不提供 `关闭`；能力未知时只显示不可调的 `Provider 默认`。
+5. 主模型和 fallback 分别保存策略，切换候选模型时不会把一个模型的挡位错误应用到另一个模型。
+6. Dify、Coze、Langflow、n8n 等外部 Runner 不显示该控件，因为 LangBot 不直接发起其内部模型请求。
+
+流水线配置保持旧格式兼容，并在模型选择对象中增加按 UUID 保存的映射：
+
+```json
+{
+  "model": {
+    "primary": "primary-model-uuid",
+    "fallbacks": ["fallback-model-uuid"],
+    "reasoning": {
+      "primary-model-uuid": "high"
+    }
+  }
+}
+```
+
+`provider_default` 不写入映射；缺少 `reasoning` 的旧流水线天然等价于全部使用 Provider 默认。
+
+滑杆交互要求：轨道使用现有主色和中性灰，不使用渐变；当前档位同时显示文字；支持键盘方向键和正确的 ARIA value text；窄屏下不溢出。
+
+### 7.3 i18n
+
+新增文案至少覆盖 `en_US`、`zh_Hans`；`ja_JP` 在模型面板现有同类字段已覆盖时同步补齐。不要把厂商参数名直接作为用户文案。
+
+## 8. API、MCP 与 Skill
+
+### 8.1 HTTP API
+
+模型 CRUD 增加：
+
+- 请求字段：`reasoning_config`
+- 响应字段：`reasoning_config`
+- 只读字段：`reasoning_capabilities`
+
+模型测试接口必须使用与真实请求完全相同的规范化和翻译逻辑，并在失败时返回可操作错误，例如：
+
+```text
+Model gemini-3-... cannot disable reasoning.
+Supported controls: effort=[low, medium, high].
+```
+
+可选增加只读调试信息，仅在测试接口返回：
+
+```json
+{
+  "effective_reasoning": {
+    "level": "low",
+    "translated_keys": ["reasoning_effort"]
+  }
+}
+```
+
+不得返回 API key、完整请求正文或原始思考内容。
+
+### 8.2 MCP 与技能
+
+当前 MCP 仅列出模型 Provider，没有完整模型 CRUD 工具。如果本次不新增 agent-accessible HTTP 操作，则无需强行新增 MCP 工具。
+
+如果后续让 Agent 修改模型思考策略，则必须同一提交更新：
+
+- `src/langbot/pkg/api/mcp/server.py`
+- 对应的 `skills/` 文档
+- 参数 schema 和安全说明
+
+## 9. 监控与可观测性
+
+控制思考量后，管理员需要判断质量、延迟和成本是否值得。建议第二阶段增加：
+
+- `reasoning_tokens`：从 `completion_tokens_details.reasoning_tokens` 或 Provider 等价字段提取。
+- `effective_reasoning_level`：记录规范化后的生效档位，不记录原始思考内容。
+- 模型监控页展示输入 token、可见输出 token、reasoning token、总延迟。
+- Provider 不返回细分 token 时显示未知，不推算。
+
+安全要求：日志、监控、debug API 默认都不得记录 reasoning 原文。思考内容可能包含敏感信息或系统提示，不应因为新增配置而扩大持久化范围。
+
+## 10. 兼容与迁移
+
+### 10.1 数据迁移
+
+- 所有现有 LLM 记录迁移为 `{"level":"provider_default"}`。
+- 不自动解析或迁移现有 `extra_args` 中的 reasoning 参数，避免误判嵌套结构和 Provider 语义。
+- UI 检测到旧 `extra_args` reasoning 字段时显示“由高级参数控制”，统一策略保持 Provider Default。
+- 用户主动改成统一策略时，要求先移除冲突高级参数。
+
+### 10.2 运行时兼容
+
+- `provider_default` 不产生任何新增请求参数。
+- 不改变现有 `remove-think` 的存储键和默认值。
+- 不改变已有 Provider 的 `litellm_provider`，除非该 Provider 在专项回归后单独切换。
+- `drop_params` 不能用于掩盖显式 reasoning 配置错误；显式策略被丢弃应视为失败。
+- 自托管和 toB 环境中的自定义兼容接口保持可用，未知能力不阻止 Provider Default 请求。
+
+## 11. 实施拆分
+
+### Phase 1：统一基础设施与主流 Provider
+
+- Alembic 增加 `llm_models.reasoning_config`。
+- Backend 模型实体、CRUD、测试接口支持统一配置。
+- LiteLLMRequester 增加能力查询、严格校验和参数翻译。
+- 支持 OpenAI、Anthropic、Gemini、DeepSeek、xAI、Ollama、OpenRouter 的已验证 LiteLLM 路径。
+- 修复结构化 reasoning 的非流式/流式保留。
+- 模型面板增加 reasoning ability 与只读能力标识。
+- Local Agent 主模型和每个 fallback 增加独立的请求级策略。
+
+### Phase 2：国内 Provider
+
+- 专项核对并支持 Volcengine/Doubao、Bailian/Qwen。
+- 对相关 requester 的 `litellm_provider` 变更做独立回归，避免把 reasoning 功能和通用请求行为回归混在一起。
+- 补齐扫描结果中的 reasoning capability。
+
+### Phase 3：监控与评估
+
+- 持久化 reasoning token 和生效策略。
+- 监控页增加 reasoning 成本/延迟指标。
+- 建立不同 effort 的离线质量、首 token 延迟、总耗时和 token 对比基线。
+
+## 12. 测试方案
+
+### 12.1 单元测试
+
+- `ReasoningConfig` 所有合法/非法组合。
+- `provider_default` 不产生任何新增参数。
+- 显式配置覆盖模型/调用 `extra_args` 的顺序。
+- reasoning 配置与高级参数冲突时拒绝。
+- OpenAI 档位原样映射。
+- Anthropic 档位映射，以及高级参数预算兼容。
+- Gemini 2 budget、Gemini 3 level，以及不支持真正关闭时拒绝。
+- DeepSeek 只显示/接受 toggle，非 `none` effort 不伪装成不同档位。
+- Ollama 布尔与分级模型差异。
+- Volcengine enabled/disabled/auto 翻译。
+- 未知 Provider 只允许 Provider Default，或在显式测试后使用标准参数。
+- 非流式 `reasoning_content` 保存到 `provider_specific_fields`。
+- 流式 reasoning 分片累计后仍能 round-trip。
+- Gemini thought signature 和工具调用现有测试不能回归。
+
+### 12.2 服务与持久化测试
+
+- 新建、读取、更新模型的 `reasoning_config`。
+- Alembic 从当前 head 升级后默认值正确。
+- 模型测试接口与真实 Local Agent 使用同一翻译函数。
+- 旧模型、旧 `extra_args` 和 `remove-think` 行为不变。
+
+### 12.3 前端测试
+
+- 能力不同的模型显示正确控件。
+- 离散滑杆只能停在后端返回的可用档位。
+- 当前档位文字、键盘操作和 ARIA value text 正确。
+- 仅开关模型、不可关闭模型、完整档位模型分别显示正确刻度。
+- fallback 能力不兼容时阻止保存并给出明确提示。
+- 中英文文案完整，移动端 Popover 不溢出。
+
+### 12.4 Provider 冒烟测试
+
+至少选取以下真实或可控 mock：
+
+- 一个支持 `none` 的 OpenAI reasoning 模型。
+- 一个不支持 `none` 的 reasoning 模型。
+- 一个 Anthropic adaptive thinking 模型。
+- 一个 Gemini 2.x 与一个 Gemini 3.x 模型。
+- 一个 DeepSeek hybrid thinking 模型，执行两轮含工具调用对话。
+- 一个 Ollama 本地 reasoning 模型。
+- 一个 OpenAI-compatible 自定义网关，验证 Provider Default 完全不变。
+
+每个模型比较 Provider Default、最低档、中档、高档或关闭，记录成功率、首 token 延迟、总耗时、总 token 和 reasoning token（若可用）。
+
+## 13. 风险与控制
+
+| 风险 | 影响 | 控制措施 |
+| --- | --- | --- |
+| 将“最低思考”误当成“关闭” | 用户以为节省了成本，实际仍在推理 | `can_disable` 严格校验，不静默降级 |
+| 模型能力表过期 | 新模型无法配置或旧模型报错 | 能力未知时保守；允许测试；升级 LiteLLM 时回归 |
+| 高 effort 导致延迟/费用陡增 | 用户体验和预算风险 | 默认 Provider Default；UI 提示；后续监控 reasoning token |
+| `extra_args` 与统一配置冲突 | 实际生效值不可预测 | 保存/测试时拒绝冲突；统一策略最后应用 |
+| reasoning 原文进入日志 | 敏感信息泄露 | 不记录原文，只记录策略和 token |
+| 多轮 reasoning 丢失 | 工具调用或后续轮次失败/降质 | 结构化保存并 round-trip；流式专项测试 |
+| 修改 Provider 路由造成通用回归 | 非 reasoning 请求也受影响 | 国内 Provider 路由放第二阶段，独立提交和回归 |
+
+## 14. 需要审核确认的决策
+
+1. **是否同意三层分离**：能力、策略、展示互不替代，保留 `remove-think` 仅控制展示。
+2. **是否同意严格语义**：显式关闭无法准确执行时直接报错，不自动降为最低思考。
+3. **是否同意请求级配置**：流水线按模型 UUID 保存挡位，不把产品配置塞进 `extra_args`。
+4. **是否同意 Runner 边界**：仅 Local Agent 展示控制项，外部 Runner 由其外部系统管理模型策略。
+5. **是否同意保守默认**：所有现有模型迁移为 Provider Default，不自动开启、关闭或迁移旧高级参数。
+6. **是否把结构化 reasoning 保留纳入第一阶段**：这是 DeepSeek 多轮和工具调用正确性的必要条件，建议必须纳入。
+
+## 15. 推荐审核结果
+
+建议按以上 6 项全部通过，并将 Phase 1 作为一个完整功能单元实施。不要只增加前端开关或只在 `extra_args` 中写 `reasoning_effort`；那样虽然改动小，但会继续混淆展示与推理、无法处理 Provider 差异，也无法保证多轮对话正确性。

src/langbot/pkg/api/http/service/model.py

@@ -9,6 +9,7 @@ from ....core import app
 from ....entity.persistence import model as persistence_model
 from ....entity.persistence import pipeline as persistence_pipeline
 from ....provider.modelmgr import requester as model_requester
+from ....provider.modelmgr import reasoning as model_reasoning
 from ....workspace.errors import WorkspaceNotFoundError
 from .secrets import mask_secret_value, redact_secrets, restore_secret_placeholders
 from .tenant import TenantContext, require_workspace_uuid, scope_statement
@@ -54,6 +55,53 @@ def _redact_model_secrets(model_data: dict) -> dict:
     return redacted
 
 
+def _normalize_llm_reasoning(model_data: dict) -> None:
+    model_data['reasoning_config'] = model_reasoning.validate_reasoning_config(
+        model_data.get('reasoning_config'),
+        model_data.get('abilities'),
+        model_data.get('extra_args'),
+    )
+
+
+def _validate_llm_reasoning_capability(
+    model_entity: persistence_model.LLMModel,
+    runtime_provider: model_requester.RuntimeProvider,
+) -> None:
+    config = model_reasoning.normalize_reasoning_config(model_entity.reasoning_config)
+    if config['level'] == 'provider_default':
+        return
+
+    runtime_model = model_requester.RuntimeLLMModel(
+        execution_context=runtime_provider.execution_context,
+        model_entity=model_entity,
+        provider=runtime_provider,
+    )
+    capabilities = runtime_provider.requester.get_reasoning_capabilities(runtime_model)
+    model_reasoning.validate_reasoning_capabilities(config, capabilities, model_entity.name)
+
+
+def _reasoning_capabilities(ap: app.Application, model: persistence_model.LLMModel) -> dict:
+    model_mgr = getattr(ap, 'model_mgr', None)
+    runtime_models = getattr(model_mgr, 'llm_model_dict', {}) if model_mgr is not None else {}
+    for runtime_model in runtime_models.values():
+        if (
+            runtime_model.model_entity.uuid == model.uuid
+            and runtime_model.model_entity.workspace_uuid == model.workspace_uuid
+        ):
+            return runtime_model.provider.requester.get_reasoning_capabilities(runtime_model)
+    return model_reasoning.default_reasoning_capabilities(
+        supported='reasoning' in (model.abilities or []),
+        source='manual' if 'reasoning' in (model.abilities or []) else 'unknown',
+    )
+
+
+def _serialize_llm_model(ap: app.Application, model: persistence_model.LLMModel) -> dict:
+    model_dict = ap.persistence_mgr.serialize_model(persistence_model.LLMModel, model)
+    model_dict['reasoning_config'] = model_reasoning.normalize_reasoning_config(model_dict.get('reasoning_config'))
+    model_dict['reasoning_capabilities'] = _reasoning_capabilities(ap, model)
+    return model_dict
+
+
 async def _validate_provider_supports(
     ap: app.Application,
     context: TenantContext,
@@ -147,7 +195,7 @@ class LLMModelsService:
 
         models_list = []
         for model in models:
-            model_dict = self.ap.persistence_mgr.serialize_model(persistence_model.LLMModel, model)
+            model_dict = _serialize_llm_model(self.ap, model)
             provider = providers.get(model.provider_uuid)
             if provider:
                 provider_dict = self.ap.persistence_mgr.serialize_model(persistence_model.ModelProvider, provider)
@@ -178,7 +226,7 @@ class LLMModelsService:
             )
         )
         models = result.all()
-        serialized = [self.ap.persistence_mgr.serialize_model(persistence_model.LLMModel, m) for m in models]
+        serialized = [_serialize_llm_model(self.ap, model) for model in models]
         return serialized if include_secret else [_redact_model_secrets(model) for model in serialized]
 
     async def create_llm_model(
@@ -214,13 +262,17 @@ class LLMModelsService:
 
         await _require_workspace_provider(self.ap, context, model_data['provider_uuid'])
         await _validate_provider_supports(self.ap, context, model_data['provider_uuid'], 'llm')
+        _normalize_llm_reasoning(model_data)
+
+        runtime_provider = await _require_runtime_provider(self.ap, context, model_data['provider_uuid'])
+        model_entity = persistence_model.LLMModel(**model_data)
+        _validate_llm_reasoning_capability(model_entity, runtime_provider)
 
         await self.ap.persistence_mgr.execute_async(sqlalchemy.insert(persistence_model.LLMModel).values(**model_data))
 
-        runtime_provider = await _require_runtime_provider(self.ap, context, model_data['provider_uuid'])
         runtime_llm_model = await self.ap.model_mgr.load_llm_model_with_provider(
             context,
-            persistence_model.LLMModel(**model_data),
+            model_entity,
             runtime_provider,
         )
         await self.ap.model_mgr.cache_llm_model(context, runtime_llm_model)
@@ -268,7 +320,7 @@ class LLMModelsService:
         if model is None:
             return None
 
-        model_dict = self.ap.persistence_mgr.serialize_model(persistence_model.LLMModel, model)
+        model_dict = _serialize_llm_model(self.ap, model)
 
         # Get provider
         provider_result = await self.ap.persistence_mgr.execute_async(
@@ -323,6 +375,18 @@ class LLMModelsService:
         await _require_workspace_provider(self.ap, context, provider_uuid)
         await _validate_provider_supports(self.ap, context, provider_uuid, 'llm')
 
+        merged_model_data = {
+            key: value
+            for key, value in {**existing_model, **model_data, 'provider_uuid': provider_uuid}.items()
+            if key not in {'provider', 'created_at', 'updated_at', 'reasoning_capabilities'}
+        }
+        _normalize_llm_reasoning(merged_model_data)
+        model_data['reasoning_config'] = merged_model_data['reasoning_config']
+
+        runtime_provider = await _require_runtime_provider(self.ap, context, provider_uuid)
+        model_entity = persistence_model.LLMModel(**_runtime_model_data(model_uuid, merged_model_data))
+        _validate_llm_reasoning_capability(model_entity, runtime_provider)
+
         result = await self.ap.persistence_mgr.execute_async(
             scope_statement(
                 sqlalchemy.update(persistence_model.LLMModel)
@@ -336,19 +400,9 @@ class LLMModelsService:
             raise WorkspaceNotFoundError('Model not found')
 
         await self.ap.model_mgr.remove_llm_model(context, model_uuid)
-        runtime_provider = await _require_runtime_provider(self.ap, context, provider_uuid)
         runtime_llm_model = await self.ap.model_mgr.load_llm_model_with_provider(
             context,
-            persistence_model.LLMModel(
-                **_runtime_model_data(
-                    model_uuid,
-                    {
-                        key: value
-                        for key, value in {**existing_model, **model_data, 'provider_uuid': provider_uuid}.items()
-                        if key not in {'provider', 'created_at', 'updated_at'}
-                    },
-                )
-            ),
+            model_entity,
             runtime_provider,
         )
         await self.ap.model_mgr.cache_llm_model(context, runtime_llm_model)
@@ -376,6 +430,7 @@ class LLMModelsService:
                 raise WorkspaceNotFoundError('Model not found')
             runtime_llm_model = await self.ap.model_mgr.get_model_by_uuid(context, model_uuid)
         else:
+            _normalize_llm_reasoning(model_data)
             runtime_llm_model = await self.ap.model_mgr.init_temporary_runtime_llm_model(context, model_data)
 
         extra_args = model_data.get('extra_args', {})

src/langbot/pkg/entity/persistence/model.py

@@ -48,6 +48,12 @@ class LLMModel(Base):
     provider_uuid = sqlalchemy.Column(sqlalchemy.String(255), nullable=False)
     abilities = sqlalchemy.Column(sqlalchemy.JSON, nullable=False, default=[])
     context_length = sqlalchemy.Column(sqlalchemy.Integer, nullable=True)
+    reasoning_config = sqlalchemy.Column(
+        sqlalchemy.JSON,
+        nullable=False,
+        default=lambda: {'level': 'provider_default'},
+        server_default=sqlalchemy.text('\'{"level":"provider_default"}\''),
+    )
     extra_args = sqlalchemy.Column(sqlalchemy.JSON, nullable=False, default={})
     prefered_ranking = sqlalchemy.Column(sqlalchemy.Integer, nullable=False, default=0)
     created_at = sqlalchemy.Column(sqlalchemy.DateTime, nullable=False, server_default=sqlalchemy.func.now())

src/langbot/pkg/persistence/alembic/versions/0018_add_llm_reasoning_config.py

@@ -0,0 +1,57 @@
+"""add llm reasoning config
+
+Revision ID: 0018_llm_reasoning_config
+Revises: 0017_oss_workspace_identity
+Create Date: 2026-07-27
+"""
+
+from __future__ import annotations
+
+import sqlalchemy as sa
+from alembic import op
+
+revision = '0018_llm_reasoning_config'
+down_revision = '0017_oss_workspace_identity'
+branch_labels = None
+depends_on = None
+
+
+_LLM_MODELS = sa.table(
+    'llm_models',
+    sa.column('reasoning_config', sa.JSON()),
+)
+
+
+def upgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    if 'llm_models' not in inspector.get_table_names():
+        return
+
+    columns = {column['name'] for column in inspector.get_columns('llm_models')}
+    if 'reasoning_config' in columns:
+        return
+
+    op.add_column(
+        'llm_models',
+        sa.Column(
+            'reasoning_config',
+            sa.JSON(),
+            nullable=True,
+            server_default=sa.text('\'{"level":"provider_default"}\''),
+        ),
+    )
+    conn.execute(_LLM_MODELS.update().values(reasoning_config={'level': 'provider_default'}))
+    with op.batch_alter_table('llm_models') as batch_op:
+        batch_op.alter_column('reasoning_config', existing_type=sa.JSON(), nullable=False)
+
+
+def downgrade() -> None:
+    conn = op.get_bind()
+    inspector = sa.inspect(conn)
+    if 'llm_models' not in inspector.get_table_names():
+        return
+    columns = {column['name'] for column in inspector.get_columns('llm_models')}
+    if 'reasoning_config' in columns:
+        with op.batch_alter_table('llm_models') as batch_op:
+            batch_op.drop_column('reasoning_config')

src/langbot/pkg/provider/modelmgr/modelmgr.py

@@ -649,6 +649,7 @@ class ModelManager:
             provider_uuid=runtime_provider.provider_entity.uuid,
             abilities=model_info.get('abilities', []),
             context_length=model_info.get('context_length'),
+            reasoning_config=model_info.get('reasoning_config', {'level': 'provider_default'}),
             extra_args=model_info.get('extra_args', {}),
         )
         return self._build_llm_model(execution_context, model_entity, runtime_provider)

src/langbot/pkg/provider/modelmgr/reasoning.py

@@ -0,0 +1,110 @@
+from __future__ import annotations
+
+import typing
+
+
+ReasoningLevel = typing.Literal[
+    'provider_default',
+    'disabled',
+    'enabled',
+    'minimal',
+    'low',
+    'medium',
+    'high',
+    'xhigh',
+    'max',
+]
+
+REASONING_LEVELS: tuple[str, ...] = (
+    'provider_default',
+    'disabled',
+    'enabled',
+    'minimal',
+    'low',
+    'medium',
+    'high',
+    'xhigh',
+    'max',
+)
+DEFAULT_REASONING_CONFIG: dict[str, str] = {'level': 'provider_default'}
+
+_CONFLICTING_TOP_LEVEL_ARGS = {'reasoning_effort', 'thinking', 'reasoning'}
+_CONFLICTING_EXTRA_BODY_ARGS = {'thinking', 'enable_thinking', 'thinking_budget', 'reasoning'}
+
+
+def normalize_reasoning_config(value: typing.Any) -> dict[str, str]:
+    """Return the canonical model reasoning configuration."""
+    if value is None:
+        return dict(DEFAULT_REASONING_CONFIG)
+    if not isinstance(value, dict):
+        raise ValueError('reasoning_config must be an object')
+
+    unknown_fields = set(value) - {'level'}
+    if unknown_fields:
+        raise ValueError(f'Unsupported reasoning_config fields: {", ".join(sorted(unknown_fields))}')
+
+    level = value.get('level', 'provider_default')
+    if level not in REASONING_LEVELS:
+        raise ValueError(f'Unsupported reasoning level: {level}')
+    return {'level': typing.cast(str, level)}
+
+
+def validate_reasoning_config(
+    value: typing.Any,
+    abilities: typing.Iterable[str] | None,
+    extra_args: typing.Any,
+) -> dict[str, str]:
+    """Validate a model-facing reasoning config and conflicting raw arguments."""
+    config = normalize_reasoning_config(value)
+    if config['level'] == 'provider_default':
+        return config
+
+    if 'reasoning' not in set(abilities or []):
+        raise ValueError('The reasoning ability must be enabled before selecting a reasoning level')
+
+    conflicts = find_reasoning_arg_conflicts(extra_args)
+    if conflicts:
+        raise ValueError('reasoning_config conflicts with advanced parameters: ' + ', '.join(conflicts))
+    return config
+
+
+def find_reasoning_arg_conflicts(extra_args: typing.Any) -> list[str]:
+    if not isinstance(extra_args, dict):
+        return []
+
+    conflicts = [key for key in sorted(_CONFLICTING_TOP_LEVEL_ARGS) if key in extra_args]
+    extra_body = extra_args.get('extra_body')
+    if isinstance(extra_body, dict):
+        conflicts.extend(f'extra_body.{key}' for key in sorted(_CONFLICTING_EXTRA_BODY_ARGS) if key in extra_body)
+    return conflicts
+
+
+def validate_reasoning_capabilities(
+    config: typing.Any,
+    capabilities: typing.Mapping[str, typing.Any],
+    model_name: str,
+) -> None:
+    """Ensure an explicit reasoning level can be honored by the requester."""
+    level = normalize_reasoning_config(config)['level']
+    if level == 'provider_default':
+        return
+
+    available_levels = capabilities.get('levels')
+    if not isinstance(available_levels, list):
+        available_levels = []
+    if capabilities.get('supported') is not True or level not in available_levels:
+        available_text = ', '.join(str(item) for item in available_levels) or 'provider_default'
+        raise ValueError(
+            f'Reasoning level "{level}" is not supported by model {model_name}. Available levels: {available_text}'
+        )
+
+
+def default_reasoning_capabilities(
+    supported: bool = False,
+    source: str = 'unknown',
+) -> dict[str, typing.Any]:
+    return {
+        'supported': supported,
+        'levels': ['provider_default'],
+        'source': source,
+    }

src/langbot/pkg/provider/modelmgr/requester.py

@@ -10,6 +10,7 @@ from ...entity.persistence import model as persistence_model
 from ...workspace.errors import WorkspaceInvariantError
 import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
 from . import token
+from . import reasoning
 import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
 import langbot_plugin.api.entities.builtin.provider.message as provider_message
 
@@ -377,11 +378,15 @@ class RuntimeLLMModel:
     provider: RuntimeProvider
     """提供商实例"""
 
+    reasoning_config_override: dict[str, str] | None
+    """Request-scoped reasoning policy supplied by the active pipeline."""
+
     def __init__(
         self,
         execution_context: ExecutionContext,
         model_entity: persistence_model.LLMModel,
         provider: RuntimeProvider,
+        reasoning_config_override: dict[str, str] | None = None,
     ):
         _ensure_same_execution_scope(provider.execution_context, execution_context, resource='LLM model')
         if model_entity.workspace_uuid != execution_context.workspace_uuid:
@@ -391,6 +396,7 @@ class RuntimeLLMModel:
         self.execution_context = execution_context
         self.model_entity = model_entity
         self.provider = provider
+        self.reasoning_config_override = reasoning_config_override
 
 
 class RuntimeEmbeddingModel:
@@ -482,6 +488,13 @@ class ProviderAPIRequester(metaclass=abc.ABCMeta):
         """
         raise NotImplementedError('This provider does not support model scanning')
 
+    def get_reasoning_capabilities(self, model: RuntimeLLMModel) -> dict[str, typing.Any]:
+        """Return normalized reasoning controls supported by a model."""
+        return reasoning.default_reasoning_capabilities(
+            supported='reasoning' in (model.model_entity.abilities or []),
+            source='manual' if 'reasoning' in (model.model_entity.abilities or []) else 'unknown',
+        )
+
     @abc.abstractmethod
     async def invoke_llm(
         self,

src/langbot/pkg/provider/modelmgr/requesters/litellmchat.py

@@ -7,7 +7,7 @@ import typing
 import litellm
 from litellm import acompletion, aembedding, arerank
 
-from .. import errors, requester
+from .. import errors, reasoning, requester
 from ....utils import httpclient
 import langbot_plugin.api.entities.builtin.resource.tool as resource_tool
 import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
@@ -164,6 +164,19 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
     _EMBEDDING_MODEL_HINTS = ('embedding', 'embed', 'bge-', 'e5-', 'm3e', 'gte-', 'text-embedding')
     _RERANK_MODEL_HINTS = ('rerank', 're-rank', 're_rank')
+    _INFERRED_EFFORT_PROVIDERS = frozenset(
+        {
+            'anthropic',
+            'gemini',
+            'groq',
+            'mistral',
+            'openai',
+            'openrouter',
+            'together_ai',
+            'xai',
+        }
+    )
+    _INFERRED_TOGGLE_PROVIDERS = frozenset({'deepseek', 'ollama', 'volcengine'})
 
     default_config: dict[str, typing.Any] = {
         'base_url': '',
@@ -201,7 +214,10 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             return False
 
         provider = self._get_custom_llm_provider()
-        candidates: list[tuple[str, str | None]] = [(model_name, provider)]
+        candidates: list[tuple[str, str | None]] = [
+            (candidate, None) for candidate in self._metadata_model_candidates(model_name)
+        ]
+        candidates.append((model_name, provider))
         litellm_model_name = self._build_litellm_model_name(model_name)
         if litellm_model_name != model_name:
             candidates.append((litellm_model_name, None))
@@ -268,6 +284,14 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 deduped_candidates.append(candidate)
         return deduped_candidates
 
+    @staticmethod
+    def _metadata_model_candidates(model_name: str) -> list[str]:
+        """Return known equivalent model IDs used only for LiteLLM metadata lookup."""
+        normalized_model_name = (model_name or '').lower()
+        if normalized_model_name.startswith('mimo-v2.5'):
+            return [f'openrouter/xiaomi/{normalized_model_name}']
+        return []
+
     def _known_context_length_fallback(self, model_name: str) -> int | None:
         normalized_model_name = (model_name or '').lower()
         if normalized_model_name.startswith('deepseek-v4-'):
@@ -287,7 +311,8 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
         if not callable(helper):
             return self._known_context_length_fallback(model_name)
 
-        candidates = [model_name]
+        candidates = self._metadata_model_candidates(model_name)
+        candidates.append(model_name)
         litellm_model_name = self._build_litellm_model_name(model_name)
         if litellm_model_name != model_name:
             candidates.append(litellm_model_name)
@@ -314,6 +339,142 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
     def _supports_vision(self, model_name: str) -> bool:
         return self._safe_litellm_bool_helper('supports_vision', model_name)
 
+    def _supports_reasoning(self, model_name: str) -> bool:
+        return self._safe_litellm_bool_helper('supports_reasoning', model_name)
+
+    def _reasoning_provider(self, model_name: str) -> str:
+        provider = (self._get_custom_llm_provider() or '').lower()
+        if provider:
+            return provider
+
+        normalized_name = (model_name or '').lower()
+        if '/' in normalized_name:
+            prefix = normalized_name.split('/', 1)[0]
+            if prefix in {
+                'anthropic',
+                'deepseek',
+                'gemini',
+                'groq',
+                'mistral',
+                'ollama',
+                'openai',
+                'openrouter',
+                'together_ai',
+                'volcengine',
+                'xai',
+            }:
+                return prefix
+
+        candidates = self._metadata_provider_candidates(model_name)
+        return candidates[0] if candidates else ''
+
+    def _safe_model_info(self, model_name: str) -> dict[str, typing.Any]:
+        helper = getattr(litellm, 'get_model_info', None)
+        if not callable(helper):
+            return {}
+
+        candidates = [
+            *self._metadata_model_candidates(model_name),
+            model_name,
+            self._build_litellm_model_name(model_name),
+        ]
+        for candidate in candidates:
+            try:
+                info = helper(candidate)
+            except Exception:
+                continue
+            if isinstance(info, dict):
+                return info
+            model_dump = getattr(info, 'model_dump', None)
+            if callable(model_dump):
+                try:
+                    dumped = model_dump()
+                    if isinstance(dumped, dict):
+                        return dumped
+                except Exception:
+                    continue
+        return {}
+
+    def get_reasoning_capabilities(self, model: requester.RuntimeLLMModel) -> dict[str, typing.Any]:
+        model_name = model.model_entity.name
+        abilities = model.model_entity.abilities or []
+        detected = self._supports_reasoning(model_name)
+        declared = 'reasoning' in abilities
+        provider = self._reasoning_provider(model_name)
+        inferred = provider in self._INFERRED_EFFORT_PROVIDERS | self._INFERRED_TOGGLE_PROVIDERS
+        supported = detected or declared or inferred
+        if not supported:
+            return reasoning.default_reasoning_capabilities()
+
+        normalized_name = model_name.lower()
+        levels = ['provider_default']
+
+        if provider == 'deepseek':
+            if 'reasoner' not in normalized_name and '-r1' not in normalized_name:
+                levels.append('disabled')
+            levels.append('enabled')
+        elif provider == 'volcengine':
+            levels.extend(['disabled', 'enabled'])
+        elif provider == 'ollama':
+            levels.append('disabled')
+            if normalized_name.startswith('gpt-oss') or '/gpt-oss' in normalized_name:
+                levels.extend(['low', 'medium', 'high'])
+            else:
+                levels.append('enabled')
+        elif not detected:
+            levels.extend(['low', 'medium', 'high'])
+        else:
+            model_info = self._safe_model_info(model_name)
+            supports_none = model_info.get('supports_none_reasoning_effort') is True
+            if provider == 'anthropic':
+                supports_none = True
+            if provider == 'gemini' and 'gemini-3' in normalized_name:
+                supports_none = False
+            if supports_none:
+                levels.append('disabled')
+
+            for level in ('minimal', 'low', 'medium', 'high'):
+                flag = model_info.get(f'supports_{level}_reasoning_effort')
+                if flag is not False:
+                    levels.append(level)
+            for level in ('xhigh', 'max'):
+                if model_info.get(f'supports_{level}_reasoning_effort') is True:
+                    levels.append(level)
+
+        return {
+            'supported': True,
+            'levels': list(dict.fromkeys(levels)),
+            'source': 'litellm' if detected else ('provider' if inferred else 'manual'),
+        }
+
+    def _build_reasoning_args(self, model: requester.RuntimeLLMModel) -> dict[str, typing.Any]:
+        raw_config = getattr(model, 'reasoning_config_override', None)
+        if raw_config is None:
+            raw_config = getattr(model.model_entity, 'reasoning_config', None)
+        if not isinstance(raw_config, dict):
+            raw_config = None
+        config = reasoning.normalize_reasoning_config(raw_config)
+        level = config['level']
+        if level == 'provider_default':
+            return {}
+
+        capabilities = self.get_reasoning_capabilities(model)
+        try:
+            reasoning.validate_reasoning_capabilities(config, capabilities, model.model_entity.name)
+        except ValueError as exc:
+            raise errors.RequesterError(str(exc)) from exc
+
+        provider = self._reasoning_provider(model.model_entity.name)
+        if level == 'disabled':
+            if provider == 'volcengine':
+                return {'thinking': {'type': 'disabled'}}
+            return {'reasoning_effort': 'none'}
+        if level == 'enabled':
+            if provider in {'deepseek', 'volcengine'}:
+                return {'thinking': {'type': 'enabled'}}
+            return {'reasoning_effort': 'low'}
+        return {'reasoning_effort': level}
+
     def _infer_model_type(self, model_id: str) -> str:
         normalized_id = (model_id or '').lower()
         if any(kw in normalized_id for kw in self._RERANK_MODEL_HINTS):
@@ -344,6 +505,11 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             )
             if supports_provider_reported_vision or self._supports_vision(model_id):
                 abilities.append('vision')
+            supports_provider_reported_reasoning = bool(
+                model_payload and model_payload.get('supports_reasoning') is True
+            )
+            if supports_provider_reported_reasoning or self._supports_reasoning(model_id):
+                abilities.append('reasoning')
             scanned_model['abilities'] = abilities
 
             context_length = self._context_length_from_scan_payload(model_payload)
@@ -670,6 +836,21 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
             args.update(model.model_entity.extra_args)
         args.update(extra_args)
 
+        reasoning_args = self._build_reasoning_args(model)
+        if reasoning_args:
+            conflicts = reasoning.find_reasoning_arg_conflicts(model.model_entity.extra_args)
+            conflicts.extend(reasoning.find_reasoning_arg_conflicts(extra_args))
+            if conflicts:
+                raise errors.RequesterError(
+                    'reasoning_config conflicts with advanced parameters: ' + ', '.join(dict.fromkeys(conflicts))
+                )
+            args.update(reasoning_args)
+            if 'reasoning_effort' in reasoning_args and self._reasoning_provider(model.model_entity.name) == 'openai':
+                allowed_openai_params = args.get('allowed_openai_params') or []
+                if not isinstance(allowed_openai_params, (list, tuple, set)):
+                    raise errors.RequesterError('allowed_openai_params must be an array')
+                args['allowed_openai_params'] = list(dict.fromkeys([*allowed_openai_params, 'reasoning_effort']))
+
         if funcs:
             tools = await self.ap.tool_mgr.generate_tools_for_openai(funcs)
             if tools:
@@ -699,6 +880,10 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
             content = message_data.get('content', '')
             reasoning_content = message_data.get('reasoning_content', None)
+            if reasoning_content:
+                provider_fields = dict(message_data.get('provider_specific_fields') or {})
+                provider_fields['reasoning_content'] = reasoning_content
+                message_data['provider_specific_fields'] = provider_fields
             message_data['content'] = self._process_thinking_content(content, reasoning_content, remove_think)
 
             if 'reasoning_content' in message_data:
@@ -760,13 +945,13 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
                 delta_content = delta.get('content', '')
                 reasoning_content = delta.get('reasoning_content', '')
+                provider_fields = dict(delta.get('provider_specific_fields') or {})
 
                 # Handle reasoning_content based on remove_think flag
                 if reasoning_content:
+                    provider_fields['reasoning_content'] = reasoning_content
                     if remove_think:
-                        # Skip reasoning content when remove_think is True
-                        chunk_idx += 1
-                        continue
+                        delta_content = None
                     else:
                         # Use reasoning_content as the displayed content
                         delta_content = reasoning_content
@@ -779,7 +964,7 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
 
                 tool_calls = self._normalize_stream_tool_calls(delta.get('tool_calls'), tool_call_state)
 
-                if chunk_idx == 0 and not delta_content and not tool_calls:
+                if chunk_idx == 0 and not delta_content and not tool_calls and not provider_fields:
                     chunk_idx += 1
                     continue
 
@@ -791,8 +976,8 @@ class LiteLLMRequester(requester.ProviderAPIRequester):
                 }
 
                 # Preserve provider_specific_fields from delta (e.g., Gemini thought_signatures)
-                if delta.get('provider_specific_fields'):
-                    chunk_data['provider_specific_fields'] = delta['provider_specific_fields']
+                if provider_fields:
+                    chunk_data['provider_specific_fields'] = provider_fields
 
                 chunk_data = {k: v for k, v in chunk_data.items() if v is not None}
                 yield provider_message.MessageChunk(**chunk_data)

src/langbot/pkg/provider/runners/localagent.py

@@ -6,6 +6,7 @@ import typing
 from .. import runner
 from ...telemetry import features as telemetry_features
 from ..modelmgr import requester as modelmgr_requester
+from ..modelmgr import reasoning as modelmgr_reasoning
 from ..tools.loaders.native import EXEC_TOOL_NAME
 import langbot_plugin.api.entities.builtin.pipeline.query as pipeline_query
 import langbot_plugin.api.entities.builtin.provider.message as provider_message
@@ -60,6 +61,7 @@ class _StreamAccumulator:
         self.msg_idx = 0
         self.accumulated_content = initial_content or ''
         self.last_role = 'assistant'
+        self.provider_specific_fields: dict[str, typing.Any] = {}
         self.msg_sequence = msg_sequence
         self.remove_think = remove_think
         self._think_state = None
@@ -94,6 +96,14 @@ class _StreamAccumulator:
                 if tool_call.function and tool_call.function.arguments:
                     self.tool_calls_map[tool_call.id].function.arguments += tool_call.function.arguments
 
+        if msg.provider_specific_fields:
+            for key, value in msg.provider_specific_fields.items():
+                if key == 'reasoning_content' and isinstance(value, str):
+                    previous = self.provider_specific_fields.get(key, '')
+                    self.provider_specific_fields[key] = f'{previous}{value}'
+                else:
+                    self.provider_specific_fields[key] = value
+
         if msg.is_final:
             self._flush_think_state()
 
@@ -103,6 +113,7 @@ class _StreamAccumulator:
                 role=self.last_role,
                 content=self._maybe_strip_think(self.accumulated_content),
                 tool_calls=list(self.tool_calls_map.values()) if (self.tool_calls_map and msg.is_final) else None,
+                provider_specific_fields=(self.provider_specific_fields or None) if msg.is_final else None,
                 is_final=msg.is_final,
                 msg_sequence=self.msg_sequence,
             )
@@ -115,6 +126,7 @@ class _StreamAccumulator:
             role=self.last_role,
             content=self._maybe_strip_think(self.accumulated_content),
             tool_calls=list(self.tool_calls_map.values()) if self.tool_calls_map else None,
+            provider_specific_fields=self.provider_specific_fields or None,
             msg_sequence=self.msg_sequence,
         )
 
@@ -233,9 +245,10 @@ class LocalAgentRunner(runner.RequestRunner):
                     execution_context,
                     query.use_llm_model_uuid,
                 )
-                candidates.append(primary)
             except ValueError:
                 self.ap.logger.warning(f'Primary model {query.use_llm_model_uuid} not found')
+            else:
+                candidates.append(LocalAgentRunner._apply_pipeline_reasoning_config(query, primary))
 
         # Fallback models
         fallback_uuids = (query.variables or {}).get('_fallback_model_uuids', [])
@@ -245,12 +258,34 @@ class LocalAgentRunner(runner.RequestRunner):
                     execution_context,
                     fb_uuid,
                 )
-                candidates.append(fb_model)
             except ValueError:
                 self.ap.logger.warning(f'Fallback model {fb_uuid} not found, skipping')
+            else:
+                candidates.append(LocalAgentRunner._apply_pipeline_reasoning_config(query, fb_model))
 
         return candidates
 
+    @staticmethod
+    def _apply_pipeline_reasoning_config(
+        query: pipeline_query.Query,
+        model: modelmgr_requester.RuntimeLLMModel,
+    ) -> modelmgr_requester.RuntimeLLMModel:
+        local_agent_config = query.pipeline_config.get('ai', {}).get('local-agent', {})
+        model_config = local_agent_config.get('model', {})
+        reasoning_by_model = model_config.get('reasoning', {}) if isinstance(model_config, dict) else {}
+        level = (
+            reasoning_by_model.get(model.model_entity.uuid, 'provider_default')
+            if isinstance(reasoning_by_model, dict)
+            else 'provider_default'
+        )
+        reasoning_config = modelmgr_reasoning.normalize_reasoning_config({'level': level})
+        return modelmgr_requester.RuntimeLLMModel(
+            execution_context=model.execution_context,
+            model_entity=model.model_entity,
+            provider=model.provider,
+            reasoning_config_override=reasoning_config,
+        )
+
     async def _invoke_with_fallback(
         self,
         query: pipeline_query.Query,

src/langbot/templates/metadata/pipeline/ai.yaml

@@ -92,6 +92,7 @@ stages:
         default:
           primary: ''
           fallbacks: []
+          reasoning: {}
       - name: max-round
         label:
           en_US: Max Round

tests/integration/persistence/test_migrations.py

@@ -9,8 +9,11 @@ Run: uv run pytest tests/integration/persistence/test_migrations.py -q
 
 from __future__ import annotations
 
+import json
+
 import pytest
 import sqlalchemy
+from sqlalchemy import text
 from sqlalchemy.ext.asyncio import create_async_engine
 
 from langbot.pkg.entity.persistence.base import Base
@@ -190,6 +193,66 @@ class TestSQLiteMigrationUpgrade:
         await run_alembic_upgrade(sqlite_engine, 'head')
         assert await get_alembic_current(sqlite_engine) == _get_script_head()
 
+    @pytest.mark.asyncio
+    async def test_reasoning_config_migrates_existing_models(self, sqlite_engine):
+        """Upgrade from 0017 backfills reasoning config and keeps a database default."""
+        async with sqlite_engine.begin() as conn:
+            await conn.execute(
+                text(
+                    """
+                    CREATE TABLE llm_models (
+                        uuid VARCHAR(255) PRIMARY KEY,
+                        name VARCHAR(255) NOT NULL,
+                        provider_uuid VARCHAR(255) NOT NULL,
+                        abilities JSON NOT NULL,
+                        context_length INTEGER,
+                        extra_args JSON NOT NULL,
+                        prefered_ranking INTEGER NOT NULL DEFAULT 0
+                    )
+                    """
+                )
+            )
+            await conn.execute(
+                text(
+                    """
+                    INSERT INTO llm_models (
+                        uuid, name, provider_uuid, abilities, extra_args, prefered_ranking
+                    ) VALUES (
+                        'existing-model', 'Existing Model', 'provider', '[]', '{}', 0
+                    )
+                    """
+                )
+            )
+
+        await run_alembic_stamp(sqlite_engine, '0017_oss_workspace_identity')
+        await run_alembic_upgrade(sqlite_engine, 'head')
+
+        async with sqlite_engine.begin() as conn:
+            columns = await conn.run_sync(lambda sync_conn: sqlalchemy.inspect(sync_conn).get_columns('llm_models'))
+            reasoning_column = next(column for column in columns if column['name'] == 'reasoning_config')
+            assert reasoning_column['nullable'] is False
+
+            existing_value = (
+                await conn.execute(text("SELECT reasoning_config FROM llm_models WHERE uuid = 'existing-model'"))
+            ).scalar_one()
+            assert json.loads(existing_value) == {'level': 'provider_default'}
+
+            await conn.execute(
+                text(
+                    """
+                    INSERT INTO llm_models (
+                        uuid, name, provider_uuid, abilities, extra_args, prefered_ranking
+                    ) VALUES (
+                        'new-model', 'New Model', 'provider', '[]', '{}', 0
+                    )
+                    """
+                )
+            )
+            new_value = (
+                await conn.execute(text("SELECT reasoning_config FROM llm_models WHERE uuid = 'new-model'"))
+            ).scalar_one()
+            assert json.loads(new_value) == {'level': 'provider_default'}
+
 
 class TestSQLiteMigrationFreshDatabase:
     """Tests for fresh database workflow."""

tests/unit_tests/api/service/test_model_service.py

@@ -17,12 +17,14 @@ import pytest
 from unittest.mock import AsyncMock, Mock
 from types import SimpleNamespace
 
+from langbot.pkg.api.http.context import ExecutionContext
 from langbot.pkg.api.http.service.model import (
     LLMModelsService,
     EmbeddingModelsService,
     RerankModelsService,
     _parse_provider_api_keys,
     _runtime_model_data,
+    _serialize_llm_model,
     _validate_provider_supports,
 )
 from langbot.pkg.api.http.service import model as model_service_module
@@ -64,15 +66,19 @@ def _create_mock_llm_model(
     abilities: list = None,
     context_length: int | None = None,
     extra_args: dict = None,
+    reasoning_config: dict = None,
 ) -> Mock:
     """Helper to create mock LLMModel entity."""
     model = Mock(spec=LLMModel)
+    model.workspace_uuid = WORKSPACE_UUID
     model.uuid = model_uuid
     model.name = name
     model.provider_uuid = provider_uuid
     model.abilities = abilities or []
     model.context_length = context_length
     model.extra_args = extra_args or {}
+    model.reasoning_config = reasoning_config or {'level': 'provider_default'}
+    model.prefered_ranking = 0
     return model
 
 
@@ -156,6 +162,26 @@ def _create_runtime_model_mgr() -> SimpleNamespace:
     return manager
 
 
+def _create_reasoning_runtime_provider(capabilities: dict) -> SimpleNamespace:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid=WORKSPACE_UUID,
+        placement_generation=1,
+    )
+    return SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=ModelProvider(
+            workspace_uuid=WORKSPACE_UUID,
+            uuid='provider-uuid',
+            name='Reasoning Provider',
+            requester='openai',
+            base_url='https://api.openai.com',
+            api_keys=[],
+        ),
+        requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities)),
+    )
+
+
 class TestParseProviderApiKeys:
     """Tests for _parse_provider_api_keys helper function."""
 
@@ -209,6 +235,42 @@ class TestRuntimeModelData:
         assert result['extra_args'] == {'temp': 0.7}
 
 
+class TestSerializeLLMModel:
+    def test_includes_runtime_reasoning_capabilities(self):
+        model = _create_mock_llm_model(
+            abilities=['reasoning'],
+            reasoning_config={'level': 'high'},
+        )
+        capabilities = {
+            'supported': True,
+            'levels': ['provider_default', 'low', 'high'],
+            'source': 'litellm',
+        }
+        runtime_model = SimpleNamespace(
+            model_entity=model,
+            provider=SimpleNamespace(
+                requester=SimpleNamespace(get_reasoning_capabilities=Mock(return_value=capabilities))
+            ),
+        )
+        ap = SimpleNamespace(
+            persistence_mgr=SimpleNamespace(
+                serialize_model=Mock(
+                    return_value={
+                        'uuid': model.uuid,
+                        'name': model.name,
+                        'reasoning_config': {'level': 'high'},
+                    }
+                )
+            ),
+            model_mgr=SimpleNamespace(llm_model_dict={('workspace', model.uuid): runtime_model}),
+        )
+
+        serialized = _serialize_llm_model(ap, model)
+
+        assert serialized['reasoning_config'] == {'level': 'high'}
+        assert serialized['reasoning_capabilities'] == capabilities
+
+
 class TestLLMModelsServiceGetLLMModels:
     """Tests for LLMModelsService.get_llm_models method."""
 
@@ -580,6 +642,66 @@ class TestLLMModelsServiceCreateLLMModel:
         ap.provider_service.find_or_create_provider.assert_called_once()
         assert result_uuid is not None
 
+    async def test_create_llm_model_validates_explicit_reasoning_level(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock(return_value=_create_mock_result([])))
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default', 'low', 'high'],
+                'source': 'litellm',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        await service.create_llm_model(
+            WORKSPACE_UUID,
+            {
+                'uuid': 'reasoning-model',
+                'name': 'Reasoning Model',
+                'provider_uuid': 'provider-uuid',
+                'abilities': ['reasoning'],
+                'reasoning_config': {'level': 'high'},
+                'extra_args': {},
+            },
+            preserve_uuid=True,
+            auto_set_to_default_pipeline=False,
+        )
+
+        runtime_entity = ap.model_mgr.load_llm_model_with_provider.await_args.args[1]
+        assert runtime_entity.reasoning_config == {'level': 'high'}
+
+    async def test_create_llm_model_rejects_unsupported_reasoning_before_insert(self):
+        ap = SimpleNamespace()
+        ap.persistence_mgr = SimpleNamespace(execute_async=AsyncMock())
+        runtime_provider = _create_reasoning_runtime_provider(
+            {
+                'supported': True,
+                'levels': ['provider_default'],
+                'source': 'manual',
+            }
+        )
+        ap.model_mgr = _create_runtime_model_mgr()
+        ap.model_mgr.provider_dict = {'provider-uuid': runtime_provider}
+
+        service = LLMModelsService(ap)
+        with pytest.raises(ValueError, match='Available levels: provider_default'):
+            await service.create_llm_model(
+                WORKSPACE_UUID,
+                {
+                    'name': 'Unknown Reasoning Model',
+                    'provider_uuid': 'provider-uuid',
+                    'abilities': ['reasoning'],
+                    'reasoning_config': {'level': 'high'},
+                    'extra_args': {},
+                },
+                auto_set_to_default_pipeline=False,
+            )
+
+        ap.persistence_mgr.execute_async.assert_not_awaited()
+
 
 class TestLLMModelsServiceUpdateLLMModel:
     """Tests for LLMModelsService.update_llm_model method."""
@@ -595,7 +717,10 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.remove_llm_model = AsyncMock()
         ap.model_mgr.load_llm_model_with_provider = AsyncMock(return_value=Mock())
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(
+            side_effect=[_create_mock_result(first_item=existing_model), _create_mock_result()]
+        )
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data())
@@ -623,7 +748,8 @@ class TestLLMModelsServiceUpdateLLMModel:
         ap.model_mgr.provider_dict = {}  # Empty
         ap.model_mgr.remove_llm_model = AsyncMock()
 
-        ap.persistence_mgr.execute_async = AsyncMock()
+        existing_model = _create_mock_llm_model()
+        ap.persistence_mgr.execute_async = AsyncMock(return_value=_create_mock_result(first_item=existing_model))
 
         service = LLMModelsService(ap)
         service.get_llm_model = AsyncMock(return_value=_existing_llm_data('nonexistent-provider'))

tests/unit_tests/provider/test_litellmchat.py

@@ -1304,6 +1304,7 @@ class TestScanModels:
         )
         requester._supports_function_calling = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
         requester._supports_vision = Mock(side_effect=lambda model_id: model_id == 'gpt-4o')
+        requester._supports_reasoning = Mock(side_effect=lambda model_id: model_id == 'o3')
         requester._safe_context_length = Mock(side_effect=lambda model_id: 128000 if model_id == 'gpt-4o' else None)
 
         mock_response = Mock()
@@ -1311,6 +1312,7 @@ class TestScanModels:
             return_value={
                 'data': [
                     {'id': 'gpt-4o'},
+                    {'id': 'o3'},
                     {'id': 'text-embedding-3-small'},
                     {'id': 'bge-reranker-v2'},
                 ]
@@ -1327,6 +1329,7 @@ class TestScanModels:
         by_id = {model['id']: model for model in result['models']}
         assert by_id['gpt-4o']['abilities'] == ['func_call', 'vision']
         assert by_id['gpt-4o']['context_length'] == 128000
+        assert by_id['o3']['abilities'] == ['reasoning']
         assert by_id['text-embedding-3-small']['type'] == 'embedding'
         assert by_id['bge-reranker-v2']['type'] == 'rerank'
 
@@ -1374,8 +1377,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'get_model_info') as mock_get_model_info:
-            mock_get_model_info.side_effect = (
-                lambda model: {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
+            mock_get_model_info.side_effect = lambda model: (
+                {'max_input_tokens': 131072} if model == 'moonshot/moonshot-v1-128k' else {}
             )
 
             assert requester._safe_context_length('moonshot-v1-128k') == 131072
@@ -1404,8 +1407,8 @@ class TestScanModels:
         )
 
         with patch.object(litellmchat.litellm, 'supports_function_calling') as mock_supports_function_calling:
-            mock_supports_function_calling.side_effect = (
-                lambda model, custom_llm_provider=None: model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
+            mock_supports_function_calling.side_effect = lambda model, custom_llm_provider=None: (
+                model == 'moonshot/kimi-k2.6' and custom_llm_provider is None
             )
 
             assert requester._supports_function_calling('kimi-k2.6') is True

tests/unit_tests/provider/test_model_service.py

@@ -249,7 +249,11 @@ async def test_updated_llm_model_is_immediately_usable_by_local_agent_pipeline()
         'ai': {
             'runner': {'runner': 'local-agent'},
             'local-agent': {
-                'model': {'primary': model_uuid, 'fallbacks': []},
+                'model': {
+                    'primary': model_uuid,
+                    'fallbacks': [],
+                    'reasoning': {model_uuid: 'high'},
+                },
                 'prompt': [],
                 'knowledge-bases': [],
             },
@@ -293,3 +297,130 @@ async def test_updated_llm_model_is_immediately_usable_by_local_agent_pipeline()
     candidates = await LocalAgentRunner._get_model_candidates(runner, processed_query)
 
     assert [model.model_entity.uuid for model in candidates] == [model_uuid]
+    assert candidates[0].reasoning_config_override == {'level': 'high'}
+
+
+@pytest.mark.asyncio
+async def test_local_agent_applies_reasoning_per_fallback_model():
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    provider = Mock(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+    )
+    primary = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='primary-model',
+            name='primary',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    fallback = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='fallback-model',
+            name='fallback',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    models = {'primary-model': primary, 'fallback-model': fallback}
+    runner = SimpleNamespace(
+        ap=SimpleNamespace(
+            model_mgr=SimpleNamespace(
+                get_model_by_uuid=AsyncMock(side_effect=lambda _context, model_uuid: models[model_uuid]),
+            ),
+            logger=Mock(),
+        )
+    )
+    query = SimpleNamespace(
+        use_llm_model_uuid='primary-model',
+        variables={'_fallback_model_uuids': ['fallback-model']},
+        pipeline_config={
+            'ai': {
+                'local-agent': {
+                    'model': {
+                        'primary': 'primary-model',
+                        'fallbacks': ['fallback-model'],
+                        'reasoning': {
+                            'primary-model': 'low',
+                            'fallback-model': 'high',
+                        },
+                    }
+                }
+            }
+        },
+        _execution_context=execution_context,
+    )
+
+    candidates = await LocalAgentRunner._get_model_candidates(runner, query)
+
+    assert [candidate.reasoning_config_override for candidate in candidates] == [
+        {'level': 'low'},
+        {'level': 'high'},
+    ]
+
+
+def test_local_agent_rejects_invalid_pipeline_reasoning_level():
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    provider = Mock(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+    )
+    model = requester.RuntimeLLMModel(
+        execution_context,
+        persistence_model.LLMModel(
+            workspace_uuid='workspace-test',
+            uuid='primary-model',
+            name='primary',
+            provider_uuid='provider',
+            abilities=['reasoning'],
+            extra_args={},
+        ),
+        provider,
+    )
+    query = SimpleNamespace(
+        pipeline_config={
+            'ai': {
+                'local-agent': {
+                    'model': {
+                        'primary': 'primary-model',
+                        'fallbacks': [],
+                        'reasoning': {'primary-model': 'turbo'},
+                    }
+                }
+            }
+        }
+    )
+
+    with pytest.raises(ValueError, match='Unsupported reasoning level'):
+        LocalAgentRunner._apply_pipeline_reasoning_config(query, model)

tests/unit_tests/provider/test_reasoning_control.py

@@ -0,0 +1,408 @@
+from __future__ import annotations
+
+from types import SimpleNamespace
+from unittest.mock import AsyncMock
+
+import pytest
+
+import langbot_plugin.api.entities.builtin.provider.message as provider_message
+from langbot.pkg.api.http.context import ExecutionContext
+from langbot.pkg.entity.persistence import model as persistence_model
+from langbot.pkg.provider.modelmgr import errors, reasoning, requester
+from langbot.pkg.provider.modelmgr.requesters import litellmchat
+from langbot.pkg.provider.modelmgr.requesters.litellmchat import LiteLLMRequester
+from langbot.pkg.provider.runners.localagent import _StreamAccumulator
+
+
+def _runtime_model(
+    request: LiteLLMRequester,
+    level: str = 'provider_default',
+    name: str = 'reasoning-model',
+    abilities: list[str] | None = None,
+) -> requester.RuntimeLLMModel:
+    execution_context = ExecutionContext(
+        instance_uuid='instance-test',
+        workspace_uuid='workspace-test',
+        placement_generation=1,
+    )
+    entity = persistence_model.LLMModel(
+        workspace_uuid='workspace-test',
+        uuid='reasoning-model',
+        name=name,
+        provider_uuid='provider-test',
+        abilities=abilities if abilities is not None else ['reasoning'],
+        reasoning_config={'level': level},
+        extra_args={},
+    )
+    provider = SimpleNamespace(
+        execution_context=execution_context,
+        provider_entity=persistence_model.ModelProvider(
+            workspace_uuid='workspace-test',
+            uuid='provider-test',
+            name='provider',
+            requester='openai',
+            base_url='https://example.com',
+            api_keys=[],
+        ),
+        requester=request,
+        token_mgr=SimpleNamespace(),
+    )
+    return requester.RuntimeLLMModel(execution_context, entity, provider)
+
+
+def _requester(provider: str = '') -> LiteLLMRequester:
+    return LiteLLMRequester(SimpleNamespace(), {'custom_llm_provider': provider})
+
+
+def test_reasoning_config_normalization_and_conflicts():
+    assert reasoning.normalize_reasoning_config(None) == {'level': 'provider_default'}
+    assert reasoning.normalize_reasoning_config({}) == {'level': 'provider_default'}
+    assert reasoning.validate_reasoning_config(
+        {'level': 'high'},
+        ['reasoning'],
+        {},
+    ) == {'level': 'high'}
+
+    with pytest.raises(ValueError, match='Unsupported reasoning level'):
+        reasoning.normalize_reasoning_config({'level': 'turbo'})
+    with pytest.raises(ValueError, match='reasoning ability'):
+        reasoning.validate_reasoning_config({'level': 'low'}, [], {})
+    with pytest.raises(ValueError, match='extra_body.thinking_budget'):
+        reasoning.validate_reasoning_config(
+            {'level': 'low'},
+            ['reasoning'],
+            {'extra_body': {'thinking_budget': 1024}},
+        )
+
+
+def test_manual_reasoning_model_exposes_conservative_effort_levels(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request))
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'low', 'medium', 'high'],
+        'source': 'manual',
+    }
+
+
+def test_provider_protocol_exposes_reasoning_for_unknown_model(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: pytest.fail('metadata should not be queried'))
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='future-reasoning-model', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'low', 'medium', 'high'],
+        'source': 'provider',
+    }
+
+
+def test_unknown_unmarked_model_without_provider_stays_safe(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='unknown-model', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': False,
+        'levels': ['provider_default'],
+        'source': 'unknown',
+    }
+
+
+def test_mimo_native_model_uses_known_equivalent_litellm_metadata(monkeypatch):
+    request = _requester('openai')
+
+    def supports_reasoning(model: str, custom_llm_provider: str | None = None) -> bool:
+        return model == 'openrouter/xiaomi/mimo-v2.5'
+
+    def get_model_info(model: str) -> dict:
+        if model == 'openrouter/xiaomi/mimo-v2.5':
+            return {'supports_reasoning': True}
+        raise ValueError('unknown model')
+
+    monkeypatch.setattr(litellmchat.litellm, 'supports_reasoning', supports_reasoning)
+    monkeypatch.setattr(litellmchat.litellm, 'get_model_info', get_model_info)
+
+    capabilities = request.get_reasoning_capabilities(
+        _runtime_model(request, name='mimo-v2.5', abilities=[])
+    )
+
+    assert capabilities == {
+        'supported': True,
+        'levels': ['provider_default', 'minimal', 'low', 'medium', 'high'],
+        'source': 'litellm',
+    }
+
+
+def test_openai_reasoning_levels_follow_litellm_metadata(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        request,
+        '_safe_model_info',
+        lambda _: {
+            'supports_none_reasoning_effort': True,
+            'supports_minimal_reasoning_effort': False,
+            'supports_low_reasoning_effort': True,
+            'supports_xhigh_reasoning_effort': True,
+        },
+    )
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='gpt-5'))
+
+    assert capabilities['source'] == 'litellm'
+    assert capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+        'xhigh',
+    ]
+
+
+def test_reasoning_argument_translation(monkeypatch):
+    openai_request = _requester('openai')
+    monkeypatch.setattr(openai_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        openai_request,
+        '_safe_model_info',
+        lambda _: {'supports_none_reasoning_effort': True},
+    )
+
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'provider_default')) == {}
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'disabled')) == {
+        'reasoning_effort': 'none'
+    }
+    assert openai_request._build_reasoning_args(_runtime_model(openai_request, 'high')) == {'reasoning_effort': 'high'}
+
+    deepseek_request = _requester('deepseek')
+    monkeypatch.setattr(deepseek_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(deepseek_request, '_safe_model_info', lambda _: {})
+    assert deepseek_request._build_reasoning_args(
+        _runtime_model(deepseek_request, 'enabled', name='deepseek-chat')
+    ) == {'thinking': {'type': 'enabled'}}
+
+
+def test_pipeline_reasoning_override_takes_precedence(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='gpt-5')
+
+    model.reasoning_config_override = {'level': 'provider_default'}
+    assert request._build_reasoning_args(model) == {}
+
+    model.reasoning_config_override = {'level': 'low'}
+    assert request._build_reasoning_args(model) == {'reasoning_effort': 'low'}
+
+
+def test_deepseek_provider_is_inferred_from_model_name(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    capabilities = request.get_reasoning_capabilities(_runtime_model(request, name='deepseek-chat'))
+
+    assert capabilities['levels'] == ['provider_default', 'disabled', 'enabled']
+
+
+def test_always_on_reasoning_models_do_not_offer_disabled(monkeypatch):
+    deepseek_request = _requester('deepseek')
+    monkeypatch.setattr(deepseek_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(deepseek_request, '_safe_model_info', lambda _: {})
+    deepseek_capabilities = deepseek_request.get_reasoning_capabilities(
+        _runtime_model(deepseek_request, name='deepseek-r1')
+    )
+    assert deepseek_capabilities['levels'] == ['provider_default', 'enabled']
+
+    gemini_request = _requester('gemini')
+    monkeypatch.setattr(gemini_request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(
+        gemini_request,
+        '_safe_model_info',
+        lambda _: {'supports_none_reasoning_effort': True},
+    )
+    gemini_capabilities = gemini_request.get_reasoning_capabilities(_runtime_model(gemini_request, name='gemini-3-pro'))
+    assert 'disabled' not in gemini_capabilities['levels']
+    with pytest.raises(errors.RequesterError, match='not supported'):
+        gemini_request._build_reasoning_args(_runtime_model(gemini_request, 'disabled', name='gemini-3-pro'))
+
+
+def test_toggle_and_effort_provider_capabilities(monkeypatch):
+    ollama_request = _requester('ollama')
+    monkeypatch.setattr(ollama_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(ollama_request, '_safe_model_info', lambda _: {})
+
+    toggle_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='qwen3'))
+    assert toggle_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'enabled',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'enabled', name='qwen3')) == {
+        'reasoning_effort': 'low'
+    }
+
+    effort_capabilities = ollama_request.get_reasoning_capabilities(_runtime_model(ollama_request, name='gpt-oss:20b'))
+    assert effort_capabilities['levels'] == [
+        'provider_default',
+        'disabled',
+        'low',
+        'medium',
+        'high',
+    ]
+    assert ollama_request._build_reasoning_args(_runtime_model(ollama_request, 'high', name='gpt-oss:20b')) == {
+        'reasoning_effort': 'high'
+    }
+
+    volcengine_request = _requester('volcengine')
+    monkeypatch.setattr(volcengine_request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(volcengine_request, '_safe_model_info', lambda _: {})
+    assert volcengine_request._build_reasoning_args(
+        _runtime_model(volcengine_request, 'disabled', name='doubao-seed')
+    ) == {'thinking': {'type': 'disabled'}}
+
+
+def test_explicit_unsupported_level_raises(monkeypatch):
+    request = _requester()
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default'):
+        request._build_reasoning_args(_runtime_model(request, 'high', abilities=[]))
+
+
+def test_provider_inference_rejects_levels_outside_conservative_profile(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: False)
+
+    with pytest.raises(errors.RequesterError, match='Available levels: provider_default, low, medium, high'):
+        request._build_reasoning_args(_runtime_model(request, 'xhigh', abilities=[]))
+
+
+@pytest.mark.asyncio
+async def test_completion_args_reject_reasoning_extra_arg_conflicts(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high')
+    model.model_entity.extra_args = {'reasoning_effort': 'low'}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    with pytest.raises(errors.RequesterError, match='conflicts with advanced parameters'):
+        await request._build_completion_args(model, [])
+
+
+@pytest.mark.asyncio
+async def test_openai_compatible_reasoning_effort_is_explicitly_allowed(monkeypatch):
+    request = _requester('openai')
+    monkeypatch.setattr(request, '_supports_reasoning', lambda _: True)
+    monkeypatch.setattr(request, '_safe_model_info', lambda _: {})
+    model = _runtime_model(request, 'high', name='deepseek-v4-flash')
+    model.model_entity.extra_args = {'allowed_openai_params': ['custom_extension']}
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert args['reasoning_effort'] == 'high'
+    assert args['allowed_openai_params'] == ['custom_extension', 'reasoning_effort']
+
+
+@pytest.mark.asyncio
+async def test_provider_default_does_not_allow_or_send_reasoning_effort():
+    request = _requester('openai')
+    model = _runtime_model(request, 'provider_default', name='deepseek-v4-flash')
+    model.provider.token_mgr.get_token = lambda: 'test-token'
+
+    args = await request._build_completion_args(model, [])
+
+    assert 'reasoning_effort' not in args
+    assert 'allowed_openai_params' not in args
+
+
+class _Dumpable:
+    def __init__(self, data: dict):
+        self.data = data
+
+    def model_dump(self) -> dict:
+        return dict(self.data)
+
+
+@pytest.mark.asyncio
+async def test_non_stream_reasoning_content_is_preserved(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+    response = SimpleNamespace(
+        choices=[
+            SimpleNamespace(
+                message=_Dumpable(
+                    {
+                        'role': 'assistant',
+                        'content': 'answer',
+                        'reasoning_content': 'private reasoning',
+                    }
+                )
+            )
+        ],
+        usage=None,
+    )
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=response))
+
+    message, _ = await request.invoke_llm(None, _runtime_model(request), [], remove_think=True)
+
+    assert message.content == 'answer'
+    assert message.provider_specific_fields == {'reasoning_content': 'private reasoning'}
+
+
+@pytest.mark.asyncio
+async def test_stream_reasoning_round_trip_with_hidden_display(monkeypatch):
+    request = _requester('deepseek')
+    request._build_completion_args = AsyncMock(return_value={})
+
+    async def chunks():
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'role': 'assistant', 'reasoning_content': 'private '}),
+                    finish_reason=None,
+                )
+            ],
+            usage=None,
+        )
+        yield SimpleNamespace(
+            choices=[
+                SimpleNamespace(
+                    delta=_Dumpable({'content': 'answer'}),
+                    finish_reason='stop',
+                )
+            ],
+            usage=None,
+        )
+
+    monkeypatch.setattr(litellmchat, 'acompletion', AsyncMock(return_value=chunks()))
+    accumulator = _StreamAccumulator(remove_think=True)
+    emitted: provider_message.MessageChunk | None = None
+
+    async for chunk in request.invoke_llm_stream(
+        None,
+        _runtime_model(request),
+        [],
+        remove_think=True,
+    ):
+        emitted = accumulator.add(chunk) or emitted
+
+    assert emitted is not None
+    assert emitted.content == 'answer'
+    assert emitted.provider_specific_fields == {'reasoning_content': 'private '}

tests/unit_tests/provider/test_requester_base.py

@@ -400,6 +400,7 @@ def test_runtime_llm_model_initialization(runtime_llm_model, fake_persistence_da
     assert model.model_entity.abilities == model_entity.abilities
     assert model.model_entity.extra_args == model_entity.extra_args
     assert model.provider is not None
+    assert model.reasoning_config_override is None
 
 
 def test_runtime_llm_model_provider_ref(runtime_llm_model):

web/src/app/home/components/dynamic-form/DynamicFormComponent.tsx

@@ -144,6 +144,7 @@ function getValueSchema(spec: DynamicFormValueSpec) {
       return z.object({
         primary: z.string(),
         fallbacks: z.array(z.string()),
+        reasoning: z.record(z.string()),
       });
     case DynamicFormItemType.PROMPT_EDITOR:
       return z.array(
@@ -488,12 +489,24 @@ export default function DynamicFormComponent({
                 (v): v is string => typeof v === 'string',
               )
             : [],
+          reasoning:
+            obj.reasoning != null &&
+            typeof obj.reasoning === 'object' &&
+            !Array.isArray(obj.reasoning)
+              ? Object.fromEntries(
+                  Object.entries(obj.reasoning).filter(
+                    (entry): entry is [string, string] =>
+                      typeof entry[1] === 'string',
+                  ),
+                )
+              : {},
         };
       }
       // Legacy string format or any other unexpected type
       return {
         primary: typeof value === 'string' ? value : '',
         fallbacks: [],
+        reasoning: {},
       };
     }
     if (item.type === 'prompt-editor') {

web/src/app/home/components/dynamic-form/DynamicFormItemComponent.tsx

@@ -25,6 +25,7 @@ import {
   EmbeddingModel,
   RerankModel,
   PluginTool,
+  ReasoningLevel,
 } from '@/app/infra/entities/api';
 import { toast } from 'sonner';
 import { useTranslation } from 'react-i18next';
@@ -66,6 +67,9 @@ import SettingsDialog, {
 } from '@/app/home/components/settings-dialog/SettingsDialog';
 import ToolResourceSelectors from '@/app/home/components/dynamic-form/ToolResourceSelectors';
 import { LANGBOT_MODELS_PROVIDER_REQUESTER } from '@/app/home/components/models-dialog/types';
+import ReasoningLevelPicker, {
+  REASONING_LEVELS,
+} from '@/app/home/components/reasoning/ReasoningLevelPicker';
 
 function hasUsableUuid<T extends { uuid?: string | null }>(
   item: T,
@@ -874,7 +878,11 @@ export default function DynamicFormItemComponent({
       ];
 
       const rawModelValue = field.value;
-      const modelValue: { primary: string; fallbacks: string[] } =
+      const modelValue: {
+        primary: string;
+        fallbacks: string[];
+        reasoning: Record<string, ReasoningLevel>;
+      } =
         rawModelValue != null &&
         typeof rawModelValue === 'object' &&
         !Array.isArray(rawModelValue)
@@ -893,10 +901,31 @@ export default function DynamicFormItemComponent({
                       .fallbacks as unknown[]
                   ).filter((v): v is string => typeof v === 'string')
                 : [],
+              reasoning:
+                (rawModelValue as Record<string, unknown>).reasoning != null &&
+                typeof (rawModelValue as Record<string, unknown>).reasoning ===
+                  'object' &&
+                !Array.isArray(
+                  (rawModelValue as Record<string, unknown>).reasoning,
+                )
+                  ? (Object.fromEntries(
+                      Object.entries(
+                        (rawModelValue as Record<string, unknown>)
+                          .reasoning as Record<string, unknown>,
+                      ).filter(
+                        (entry): entry is [string, ReasoningLevel] =>
+                          typeof entry[1] === 'string' &&
+                          REASONING_LEVELS.includes(
+                            entry[1] as ReasoningLevel,
+                          ),
+                      ),
+                    ) as Record<string, ReasoningLevel>)
+                  : {},
             }
           : {
               primary: typeof rawModelValue === 'string' ? rawModelValue : '',
               fallbacks: [],
+              reasoning: {},
             };
 
       const renderModelSelect = (
@@ -1043,20 +1072,78 @@ export default function DynamicFormItemComponent({
         field.onChange({ ...modelValue, ...patch });
       };
 
+      const updateModelReasoning = (
+        modelUuid: string,
+        level: ReasoningLevel,
+      ) => {
+        if (!modelUuid) return;
+        const updated = { ...modelValue.reasoning };
+        if (level === 'provider_default') {
+          delete updated[modelUuid];
+        } else {
+          updated[modelUuid] = level;
+        }
+        updateValue({ reasoning: updated });
+      };
+
+      const replaceModel = (
+        currentUuid: string,
+        nextUuid: string,
+        patch: Partial<typeof modelValue>,
+      ) => {
+        const nextValue = { ...modelValue, ...patch };
+        const updatedReasoning = { ...modelValue.reasoning };
+        const currentModelStillSelected =
+          nextValue.primary === currentUuid ||
+          nextValue.fallbacks.includes(currentUuid);
+        if (
+          currentUuid &&
+          currentUuid !== nextUuid &&
+          !currentModelStillSelected
+        ) {
+          delete updatedReasoning[currentUuid];
+        }
+        updateValue({ ...nextValue, reasoning: updatedReasoning });
+      };
+
+      const renderReasoningPicker = (modelUuid: string) => {
+        if (!modelUuid) return null;
+        const model = llmModels.find((candidate) => candidate.uuid === modelUuid);
+        const currentLevel =
+          modelValue.reasoning[modelUuid] || 'provider_default';
+        const availableLevels = model?.reasoning_capabilities?.levels || [
+          'provider_default',
+        ];
+        const levels = REASONING_LEVELS.filter(
+          (level) =>
+            availableLevels.includes(level) || level === currentLevel,
+        );
+
+        return (
+          <ReasoningLevelPicker
+            value={currentLevel}
+            levels={levels}
+            onChange={(level) => updateModelReasoning(modelUuid, level)}
+          />
+        );
+      };
+
       const addFallbackModel = () => {
         updateValue({ fallbacks: [...modelValue.fallbacks, ''] });
       };
 
       const updateFallbackModel = (index: number, value: string) => {
         const updated = [...modelValue.fallbacks];
+        const currentUuid = updated[index];
         updated[index] = value;
-        updateValue({ fallbacks: updated });
+        replaceModel(currentUuid, value, { fallbacks: updated });
       };
 
       const removeFallbackModel = (index: number) => {
         const updated = [...modelValue.fallbacks];
+        const removedUuid = updated[index];
         updated.splice(index, 1);
-        updateValue({ fallbacks: updated });
+        replaceModel(removedUuid, '', { fallbacks: updated });
       };
 
       const moveFallbackModel = (index: number, direction: 'up' | 'down') => {
@@ -1081,10 +1168,12 @@ export default function DynamicFormItemComponent({
               <div className="min-w-0 flex-1">
                 {renderModelSelect(
                   modelValue.primary,
-                  (val) => updateValue({ primary: val }),
+                  (val) =>
+                    replaceModel(modelValue.primary, val, { primary: val }),
                   t('models.selectModel'),
                 )}
               </div>
+              {renderReasoningPicker(modelValue.primary)}
               <Tooltip>
                 <TooltipTrigger asChild>
                   <Button
@@ -1118,15 +1207,18 @@ export default function DynamicFormItemComponent({
               </p>
               {modelValue.fallbacks.map((fbUuid: string, index: number) => (
                 <div key={index} className="flex min-w-0 items-center gap-2">
-                  <span className="text-xs text-muted-foreground w-4 shrink-0">
+                  <span className="w-4 shrink-0 text-xs text-muted-foreground">
                     {index + 1}.
                   </span>
-                  <div className="min-w-0 flex-1">
-                    {renderModelSelect(
-                      fbUuid,
-                      (val) => updateFallbackModel(index, val),
-                      t('models.selectModel'),
-                    )}
+                  <div className="flex min-w-0 flex-1 items-center gap-1.5">
+                    <div className="min-w-0 flex-1">
+                      {renderModelSelect(
+                        fbUuid,
+                        (val) => updateFallbackModel(index, val),
+                        t('models.selectModel'),
+                      )}
+                    </div>
+                    {renderReasoningPicker(fbUuid)}
                   </div>
                   <div className="flex gap-1 shrink-0">
                     <Button

web/src/app/home/components/dynamic-form/DynamicFormSaveValues.ts

@@ -5,6 +5,56 @@ export type DynamicFormSaveValueSpec = Pick<
   'default' | 'name' | 'type'
 >;
 
+const reasoningLevels = new Set([
+  'disabled',
+  'enabled',
+  'minimal',
+  'low',
+  'medium',
+  'high',
+  'xhigh',
+  'max',
+]);
+
+function normalizeModelFallbackValue(value: unknown): {
+  primary: string;
+  fallbacks: string[];
+  reasoning: Record<string, string>;
+} {
+  const raw =
+    value != null && typeof value === 'object' && !Array.isArray(value)
+      ? (value as Record<string, unknown>)
+      : {};
+  const primary =
+    typeof raw.primary === 'string'
+      ? raw.primary
+      : typeof value === 'string'
+        ? value
+        : '';
+  const fallbacks = Array.isArray(raw.fallbacks)
+    ? raw.fallbacks.filter(
+        (fallback): fallback is string => typeof fallback === 'string',
+      )
+    : [];
+  const selectedModels = new Set([primary, ...fallbacks].filter(Boolean));
+  const rawReasoning =
+    raw.reasoning != null &&
+    typeof raw.reasoning === 'object' &&
+    !Array.isArray(raw.reasoning)
+      ? (raw.reasoning as Record<string, unknown>)
+      : {};
+  const reasoning = Object.fromEntries(
+    Object.entries(rawReasoning).filter(
+      ([modelUuid, level]) =>
+        selectedModels.has(modelUuid) &&
+        typeof level === 'string' &&
+        reasoningLevels.has(level),
+    ),
+  ) as Record<string, string>;
+
+  return { primary, fallbacks, reasoning };
+}
+
 /**
  * Build the value snapshot emitted to parent forms for persistence.
  * Only single-line string fields trim surrounding whitespace; multiline text
@@ -16,10 +66,14 @@ export function normalizeDynamicFormValuesForSave(
 ): Record<string, unknown> {
   return specs.reduce<Record<string, unknown>>((values, spec) => {
     const value = formValues[spec.name] ?? spec.default;
-    values[spec.name] =
-      spec.type === 'string' && typeof value === 'string'
-        ? value.trim()
-        : value;
+    if (spec.type === 'model-fallback-selector') {
+      values[spec.name] = normalizeModelFallbackValue(value);
+    } else {
+      values[spec.name] =
+        spec.type === 'string' && typeof value === 'string'
+          ? value.trim()
+          : value;
+    }
     return values;
   }, {});
 }

web/src/app/home/components/models-dialog/ModelsPanel.tsx

@@ -1,7 +1,7 @@
 import { useState, useEffect } from 'react';
 import { Plus, Boxes } from 'lucide-react';
 import { httpClient, systemInfo } from '@/app/infra/http/HttpClient';
-import { ModelProvider } from '@/app/infra/entities/api';
+import { ModelProvider, ReasoningConfig } from '@/app/infra/entities/api';
 import {
   Dialog,
   DialogContent,
@@ -15,6 +15,7 @@ import ProviderForm from './component/provider-form/ProviderForm';
 import { ProviderCard } from './components';
 import {
   ExtraArg,
+  DEFAULT_REASONING_CONFIG,
   ModelType,
   ScanModelsResult,
   SelectedScannedModel,
@@ -285,6 +286,7 @@ export default function ModelsPanel({
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) {
     if (!name.trim()) {
@@ -300,6 +302,7 @@ export default function ModelsPanel({
           name,
           provider_uuid: providerUuid,
           abilities,
+          reasoning_config: reasoningConfig,
           context_length: parseContextLength(
             contextLength,
             t('models.contextLengthInvalid'),
@@ -361,6 +364,7 @@ export default function ModelsPanel({
             name: item.model.name,
             provider_uuid: providerUuid,
             abilities: item.abilities,
+            reasoning_config: DEFAULT_REASONING_CONFIG,
             context_length: item.model.context_length ?? null,
             extra_args: {},
           } as never);
@@ -398,6 +402,7 @@ export default function ModelsPanel({
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) {
     if (!name.trim()) {
@@ -413,6 +418,7 @@ export default function ModelsPanel({
           name,
           provider_uuid: providerUuid,
           abilities,
+          reasoning_config: reasoningConfig,
           context_length: parseContextLength(
             contextLength,
             t('models.contextLengthInvalid'),
@@ -469,6 +475,7 @@ export default function ModelsPanel({
     modelType: ModelType,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) {
     setIsTesting(true);
     setTestResult(null);
@@ -491,6 +498,7 @@ export default function ModelsPanel({
           provider_uuid: '',
           provider: providerData,
           abilities,
+          reasoning_config: reasoningConfig,
           extra_args: extraArgsObj,
         } as never);
       } else if (modelType === 'embedding') {
@@ -554,13 +562,21 @@ export default function ModelsPanel({
         onSpaceLogin={handleSpaceLogin}
         onOpenAddModel={() => setAddModelPopoverOpen(provider.uuid)}
         onCloseAddModel={() => setAddModelPopoverOpen(null)}
-        onAddModel={(modelType, name, abilities, extraArgs, contextLength) =>
+        onAddModel={(
+          modelType,
+          name,
+          abilities,
+          extraArgs,
+          reasoningConfig,
+          contextLength,
+        ) =>
           handleAddModel(
             provider.uuid,
             modelType,
             name,
             abilities,
             extraArgs,
+            reasoningConfig,
             contextLength,
           )
         }
@@ -576,6 +592,7 @@ export default function ModelsPanel({
           name,
           abilities,
           extraArgs,
+          reasoningConfig,
           contextLength,
         ) =>
           handleUpdateModel(
@@ -585,6 +602,7 @@ export default function ModelsPanel({
             name,
             abilities,
             extraArgs,
+            reasoningConfig,
             contextLength,
           )
         }
@@ -593,8 +611,15 @@ export default function ModelsPanel({
         onDeleteModel={(modelId, modelType) =>
           handleDeleteModel(provider.uuid, modelId, modelType)
         }
-        onTestModel={(name, modelType, abilities, extraArgs) =>
-          handleTestModel(provider.uuid, name, modelType, abilities, extraArgs)
+        onTestModel={(name, modelType, abilities, extraArgs, reasoningConfig) =>
+          handleTestModel(
+            provider.uuid,
+            name,
+            modelType,
+            abilities,
+            extraArgs,
+            reasoningConfig,
+          )
         }
         isSubmitting={isSubmitting}
         isTesting={isTesting}

web/src/app/home/components/models-dialog/components/AddModelPopover.tsx

@@ -6,6 +6,7 @@ import {
   ArrowUpDown,
   Eye,
   Wrench,
+  BrainCircuit,
   Check,
   RefreshCw,
 } from 'lucide-react';
@@ -20,8 +21,12 @@ import {
 } from '@/components/ui/popover';
 import { Tabs, TabsList, TabsTrigger } from '@/components/ui/tabs';
 import { useTranslation } from 'react-i18next';
-import { ScannedProviderModel } from '@/app/infra/entities/api';
 import {
+  ReasoningConfig,
+  ScannedProviderModel,
+} from '@/app/infra/entities/api';
+import {
+  DEFAULT_REASONING_CONFIG,
   ExtraArg,
   ModelType,
   ScanModelsResult,
@@ -42,6 +47,7 @@ interface AddModelPopoverProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onScanModels: (modelType?: ModelType) => Promise<ScanModelsResult>;
@@ -54,6 +60,7 @@ interface AddModelPopoverProps {
     modelType: ModelType,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) => Promise<void>;
   isSubmitting: boolean;
   isTesting: boolean;
@@ -143,11 +150,24 @@ export default function AddModelPopover({
       tab === 'llm' && contextLength.trim()
         ? Number(contextLength.trim())
         : null;
-    await onAddModel(tab, name, abilities, extraArgs, parsedContextLength);
+    await onAddModel(
+      tab,
+      name,
+      abilities,
+      extraArgs,
+      DEFAULT_REASONING_CONFIG,
+      parsedContextLength,
+    );
   };
 
   const handleTest = async () => {
-    await onTestModel(name, tab, tab === 'llm' ? abilities : [], extraArgs);
+    await onTestModel(
+      name,
+      tab,
+      tab === 'llm' ? abilities : [],
+      extraArgs,
+      DEFAULT_REASONING_CONFIG,
+    );
   };
 
   const handleScan = async () => {
@@ -322,7 +342,7 @@ export default function AddModelPopover({
                   {tab === 'llm' && (
                     <div className="space-y-2">
                       <Label>{t('models.abilities')}</Label>
-                      <div className="flex gap-4">
+                      <div className="flex flex-wrap gap-4">
                         <div className="flex items-center gap-2">
                           <Checkbox
                             id="add-vision"
@@ -349,6 +369,19 @@ export default function AddModelPopover({
                             {t('models.functionCallAbility')}
                           </Label>
                         </div>
+                        <div className="flex items-center gap-2">
+                          <Checkbox
+                            id="add-reasoning"
+                            checked={abilities.includes('reasoning')}
+                            onCheckedChange={(checked) =>
+                              toggleAbility('reasoning', checked as boolean)
+                            }
+                          />
+                          <Label htmlFor="add-reasoning" className="text-sm">
+                            <BrainCircuit className="h-3 w-3 inline mr-1" />
+                            {t('models.reasoningAbility')}
+                          </Label>
+                        </div>
                       </div>
                     </div>
                   )}

web/src/app/home/components/models-dialog/components/ModelItem.tsx

@@ -1,5 +1,5 @@
 import { useState, useEffect } from 'react';
-import { Trash2, Eye, Wrench, Check } from 'lucide-react';
+import { Trash2, Eye, Wrench, Check, BrainCircuit } from 'lucide-react';
 import { Button } from '@/components/ui/button';
 import { Input } from '@/components/ui/input';
 import { Label } from '@/components/ui/label';
@@ -11,8 +11,17 @@ import {
   PopoverTrigger,
 } from '@/components/ui/popover';
 import { useTranslation } from 'react-i18next';
-import { LLMModel, EmbeddingModel } from '@/app/infra/entities/api';
-import { ExtraArg, ModelType, TestResult } from '../types';
+import {
+  LLMModel,
+  EmbeddingModel,
+  ReasoningConfig,
+} from '@/app/infra/entities/api';
+import {
+  DEFAULT_REASONING_CONFIG,
+  ExtraArg,
+  ModelType,
+  TestResult,
+} from '../types';
 import ExtraArgsEditor from './ExtraArgsEditor';
 import { userInfo } from '@/app/infra/http';
 
@@ -32,12 +41,14 @@ interface ModelItemProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onTestModel: (
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) => Promise<void>;
   isSubmitting: boolean;
   isTesting: boolean;
@@ -103,7 +114,6 @@ export default function ModelItem({
   const [editExtraArgs, setEditExtraArgs] = useState<ExtraArg[]>(
     convertExtraArgsToArray(model.extra_args),
   );
-
   const isEditOpen = editModelPopoverOpen === model.uuid;
   const isDeleteOpen = deleteConfirmOpen === model.uuid;
 
@@ -133,12 +143,20 @@ export default function ModelItem({
       editName,
       editAbilities,
       editExtraArgs,
+      modelType === 'llm'
+        ? (model as LLMModel).reasoning_config || DEFAULT_REASONING_CONFIG
+        : DEFAULT_REASONING_CONFIG,
       parsedContextLength,
     );
   };
 
   const handleTest = async () => {
-    await onTestModel(editName, editAbilities, editExtraArgs);
+    await onTestModel(
+      editName,
+      editAbilities,
+      editExtraArgs,
+      DEFAULT_REASONING_CONFIG,
+    );
   };
 
   const toggleAbility = (ability: string, checked: boolean) => {
@@ -149,6 +167,12 @@ export default function ModelItem({
     }
   };
 
+  const supportsReasoning =
+    modelType === 'llm' &&
+    ((model as LLMModel).reasoning_capabilities?.supported === true ||
+      (model as LLMModel).abilities?.includes('reasoning'));
+  const canSaveModel = !isLangBotModels;
+
   // Check if popover should be disabled (space models when not logged in)
   const isPopoverDisabled =
     !canManage || (isLangBotModels && userInfo?.account_type !== 'space');
@@ -194,6 +218,12 @@ export default function ModelItem({
                   <Wrench className="h-3 w-3" />
                 </Badge>
               )}
+            {supportsReasoning && (
+              <Badge variant="outline" className="text-xs gap-1">
+                <BrainCircuit className="h-3 w-3" />
+                {t('models.reasoningAbility')}
+              </Badge>
+            )}
           </div>
           {canManage && !isLangBotModels && (
             <Popover
@@ -270,7 +300,7 @@ export default function ModelItem({
           {modelType === 'llm' && (
             <div className="space-y-2">
               <Label>{t('models.abilities')}</Label>
-              <div className="flex gap-4">
+              <div className="flex flex-wrap gap-4">
                 <div className="flex items-center gap-2">
                   <Checkbox
                     id={`edit-vision-${model.uuid}`}
@@ -305,6 +335,23 @@ export default function ModelItem({
                     {t('models.functionCallAbility')}
                   </Label>
                 </div>
+                <div className="flex items-center gap-2">
+                  <Checkbox
+                    id={`edit-reasoning-${model.uuid}`}
+                    checked={editAbilities.includes('reasoning')}
+                    disabled={isLangBotModels}
+                    onCheckedChange={(checked) =>
+                      toggleAbility('reasoning', checked as boolean)
+                    }
+                  />
+                  <Label
+                    htmlFor={`edit-reasoning-${model.uuid}`}
+                    className="text-sm"
+                  >
+                    <BrainCircuit className="h-3 w-3 inline mr-1" />
+                    {t('models.reasoningAbility')}
+                  </Label>
+                </div>
               </div>
             </div>
           )}
@@ -336,7 +383,7 @@ export default function ModelItem({
           />
 
           <div className="flex gap-2">
-            {!isLangBotModels && (
+            {canSaveModel && (
               <Button
                 className="flex-1"
                 size="sm"
@@ -347,7 +394,7 @@ export default function ModelItem({
               </Button>
             )}
             <Button
-              className={isLangBotModels ? 'w-full' : 'flex-1'}
+              className={canSaveModel ? 'flex-1' : 'w-full'}
               size="sm"
               variant="outline"
               onClick={handleTest}

web/src/app/home/components/models-dialog/components/ProviderCard.tsx

@@ -9,7 +9,7 @@ import {
   Radar,
 } from 'lucide-react';
 import { httpClient, systemInfo } from '@/app/infra/http/HttpClient';
-import { ModelProvider } from '@/app/infra/entities/api';
+import { ModelProvider, ReasoningConfig } from '@/app/infra/entities/api';
 import { Button } from '@/components/ui/button';
 import {
   Collapsible,
@@ -63,6 +63,7 @@ interface ProviderCardProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onScanModels: (modelType?: ModelType) => Promise<ScanModelsResult>;
@@ -78,6 +79,7 @@ interface ProviderCardProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onOpenDeleteConfirm: (modelId: string) => void;
@@ -88,6 +90,7 @@ interface ProviderCardProps {
     modelType: ModelType,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) => Promise<void>;
   isSubmitting: boolean;
   isTesting: boolean;
@@ -430,6 +433,7 @@ export default function ProviderCard({
                       name,
                       abilities,
                       extraArgs,
+                      reasoningConfig,
                       contextLength,
                     ) =>
                       onUpdateModel(
@@ -438,11 +442,23 @@ export default function ProviderCard({
                         name,
                         abilities,
                         extraArgs,
+                        reasoningConfig,
                         contextLength,
                       )
                     }
-                    onTestModel={(name, abilities, extraArgs) =>
-                      onTestModel(name, 'llm', abilities, extraArgs)
+                    onTestModel={(
+                      name,
+                      abilities,
+                      extraArgs,
+                      reasoningConfig,
+                    ) =>
+                      onTestModel(
+                        name,
+                        'llm',
+                        abilities,
+                        extraArgs,
+                        reasoningConfig,
+                      )
                     }
                     isSubmitting={isSubmitting}
                     isTesting={isTesting}
@@ -464,17 +480,34 @@ export default function ProviderCard({
                     onOpenDeleteConfirm={onOpenDeleteConfirm}
                     onCloseDeleteConfirm={onCloseDeleteConfirm}
                     onDeleteModel={() => onDeleteModel(model.uuid, 'embedding')}
-                    onUpdateModel={(name, abilities, extraArgs) =>
+                    onUpdateModel={(
+                      name,
+                      abilities,
+                      extraArgs,
+                      reasoningConfig,
+                    ) =>
                       onUpdateModel(
                         model.uuid,
                         'embedding',
                         name,
                         abilities,
                         extraArgs,
+                        reasoningConfig,
                       )
                     }
-                    onTestModel={(name, abilities, extraArgs) =>
-                      onTestModel(name, 'embedding', abilities, extraArgs)
+                    onTestModel={(
+                      name,
+                      abilities,
+                      extraArgs,
+                      reasoningConfig,
+                    ) =>
+                      onTestModel(
+                        name,
+                        'embedding',
+                        abilities,
+                        extraArgs,
+                        reasoningConfig,
+                      )
                     }
                     isSubmitting={isSubmitting}
                     isTesting={isTesting}
@@ -496,17 +529,34 @@ export default function ProviderCard({
                     onOpenDeleteConfirm={onOpenDeleteConfirm}
                     onCloseDeleteConfirm={onCloseDeleteConfirm}
                     onDeleteModel={() => onDeleteModel(model.uuid, 'rerank')}
-                    onUpdateModel={(name, abilities, extraArgs) =>
+                    onUpdateModel={(
+                      name,
+                      abilities,
+                      extraArgs,
+                      reasoningConfig,
+                    ) =>
                       onUpdateModel(
                         model.uuid,
                         'rerank',
                         name,
                         abilities,
                         extraArgs,
+                        reasoningConfig,
                       )
                     }
-                    onTestModel={(name, abilities, extraArgs) =>
-                      onTestModel(name, 'rerank', abilities, extraArgs)
+                    onTestModel={(
+                      name,
+                      abilities,
+                      extraArgs,
+                      reasoningConfig,
+                    ) =>
+                      onTestModel(
+                        name,
+                        'rerank',
+                        abilities,
+                        extraArgs,
+                        reasoningConfig,
+                      )
                     }
                     isSubmitting={isSubmitting}
                     isTesting={isTesting}

web/src/app/home/components/models-dialog/types.ts

@@ -5,6 +5,7 @@ import {
   ModelProvider,
   ProviderScanDebugInfo,
   ScannedProviderModel,
+  ReasoningConfig,
 } from '@/app/infra/entities/api';
 
 export type ExtraArg = {
@@ -16,6 +17,10 @@ export type ExtraArg = {
 
 export type ModelType = 'llm' | 'embedding' | 'rerank';
 
+export const DEFAULT_REASONING_CONFIG: ReasoningConfig = {
+  level: 'provider_default',
+};
+
 export interface ProviderModels {
   llm: LLMModel[];
   embedding: EmbeddingModel[];
@@ -53,12 +58,14 @@ export interface ModelItemProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onTest: (
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) => Promise<void>;
   isSubmitting: boolean;
   isTesting: boolean;
@@ -90,6 +97,7 @@ export interface ProviderCardProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onScanModels: (modelType?: ModelType) => Promise<ScanModelsResult>;
@@ -105,6 +113,7 @@ export interface ProviderCardProps {
     name: string,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
     contextLength?: number | null,
   ) => Promise<void>;
   onOpenDeleteConfirm: (modelId: string) => void;
@@ -115,6 +124,7 @@ export interface ProviderCardProps {
     modelType: ModelType,
     abilities: string[],
     extraArgs: ExtraArg[],
+    reasoningConfig: ReasoningConfig,
   ) => Promise<void>;
   isSubmitting: boolean;
   isTesting: boolean;

web/src/app/home/components/reasoning/ReasoningLevelPicker.module.css

@@ -0,0 +1,86 @@
+.control {
+  position: relative;
+  height: 2.5rem;
+  margin-top: 0.75rem;
+}
+
+.track,
+.fill {
+  position: absolute;
+  top: 50%;
+  left: 0;
+  height: 2rem;
+  transform: translateY(-50%);
+  border-radius: 1rem;
+}
+
+.track {
+  width: 100%;
+  background: var(--muted);
+}
+
+.fill {
+  background: var(--primary);
+}
+
+.tick {
+  position: absolute;
+  top: 50%;
+  width: 0.3rem;
+  height: 0.3rem;
+  transform: translate(-50%, -50%);
+  border-radius: 50%;
+  background: color-mix(in oklch, var(--muted-foreground) 46%, transparent);
+  pointer-events: none;
+}
+
+.tickActive {
+  background: color-mix(in oklch, var(--primary-foreground) 42%, transparent);
+}
+
+.input {
+  position: absolute;
+  inset: 0;
+  width: 100%;
+  height: 100%;
+  margin: 0;
+  appearance: none;
+  background: transparent;
+  cursor: pointer;
+}
+
+.input::-webkit-slider-runnable-track {
+  height: 2rem;
+  background: transparent;
+}
+
+.input::-webkit-slider-thumb {
+  width: 2.125rem;
+  height: 2.125rem;
+  margin-top: -0.0625rem;
+  appearance: none;
+  border: 1px solid var(--border);
+  border-radius: 50%;
+  background: var(--background);
+  box-shadow: 0 1px 4px color-mix(in oklch, var(--foreground) 20%, transparent);
+}
+
+.input::-moz-range-track {
+  height: 2rem;
+  background: transparent;
+}
+
+.input::-moz-range-thumb {
+  width: 2.125rem;
+  height: 2.125rem;
+  border: 1px solid var(--border);
+  border-radius: 50%;
+  background: var(--background);
+  box-shadow: 0 1px 4px color-mix(in oklch, var(--foreground) 20%, transparent);
+}
+
+.input:focus-visible {
+  outline: 2px solid var(--ring);
+  outline-offset: 2px;
+  border-radius: 1rem;
+}

web/src/app/home/components/reasoning/ReasoningLevelPicker.tsx

@@ -0,0 +1,123 @@
+import { BrainCircuit, ChevronDown, ChevronRight } from 'lucide-react';
+import { useTranslation } from 'react-i18next';
+import { ReasoningLevel } from '@/app/infra/entities/api';
+import { Button } from '@/components/ui/button';
+import {
+  Popover,
+  PopoverContent,
+  PopoverTrigger,
+} from '@/components/ui/popover';
+import styles from './ReasoningLevelPicker.module.css';
+
+export const REASONING_LEVELS: ReasoningLevel[] = [
+  'provider_default',
+  'disabled',
+  'enabled',
+  'minimal',
+  'low',
+  'medium',
+  'high',
+  'xhigh',
+  'max',
+];
+
+export const REASONING_LEVEL_LABEL_KEYS: Record<ReasoningLevel, string> = {
+  provider_default: 'models.reasoningLevels.providerDefault',
+  disabled: 'models.reasoningLevels.disabled',
+  enabled: 'models.reasoningLevels.enabled',
+  minimal: 'models.reasoningLevels.minimal',
+  low: 'models.reasoningLevels.low',
+  medium: 'models.reasoningLevels.medium',
+  high: 'models.reasoningLevels.high',
+  xhigh: 'models.reasoningLevels.xhigh',
+  max: 'models.reasoningLevels.max',
+};
+
+interface ReasoningLevelPickerProps {
+  value: ReasoningLevel;
+  levels: ReasoningLevel[];
+  disabled?: boolean;
+  onChange: (value: ReasoningLevel) => void;
+}
+
+export default function ReasoningLevelPicker({
+  value,
+  levels,
+  disabled = false,
+  onChange,
+}: ReasoningLevelPickerProps) {
+  const { t } = useTranslation();
+  const safeLevels: ReasoningLevel[] =
+    levels.length > 0 ? levels : ['provider_default'];
+  const safeValue: ReasoningLevel = safeLevels.includes(value)
+    ? value
+    : safeLevels[0];
+  const currentLabel = t(REASONING_LEVEL_LABEL_KEYS[safeValue]);
+  const isExplicit = safeValue !== 'provider_default';
+  const currentIndex = Math.max(0, safeLevels.indexOf(safeValue));
+  const denominator = Math.max(1, safeLevels.length - 1);
+  const progress = currentIndex / denominator;
+
+  return (
+    <Popover>
+      <PopoverTrigger asChild>
+        <Button
+          type="button"
+          variant="outline"
+          size="sm"
+          disabled={disabled || safeLevels.length <= 1}
+          aria-label={`${t('models.reasoningLevel')}: ${currentLabel}`}
+          className="h-9 w-9 shrink-0 gap-1.5 px-2.5 text-xs font-normal sm:w-auto sm:max-w-36"
+        >
+          <BrainCircuit
+            className={`size-4 shrink-0 ${isExplicit ? 'text-primary' : 'text-muted-foreground'}`}
+          />
+          <span className="hidden min-w-0 truncate sm:block">
+            {currentLabel}
+          </span>
+          <ChevronDown className="hidden size-3.5 shrink-0 text-muted-foreground sm:block" />
+        </Button>
+      </PopoverTrigger>
+      <PopoverContent align="end" className="w-[272px] p-4">
+        <div className="flex h-5 items-center gap-0.5 text-sm text-muted-foreground">
+          <span>{currentLabel}</span>
+          <ChevronRight className="size-3.5" />
+        </div>
+        <div className={styles.control}>
+          <div className={styles.track} />
+          <div
+            className={styles.fill}
+            style={{
+              width: `calc(17px + (100% - 34px) * ${progress})`,
+            }}
+          />
+          {safeLevels.map((level, index) => {
+            const tickProgress = index / denominator;
+            return (
+              <span
+                key={level}
+                className={`${styles.tick} ${index <= currentIndex ? styles.tickActive : ''}`}
+                style={{
+                  left: `calc(17px + (100% - 34px) * ${tickProgress})`,
+                }}
+              />
+            );
+          })}
+          <input
+            className={styles.input}
+            type="range"
+            min={0}
+            max={Math.max(0, safeLevels.length - 1)}
+            step={1}
+            value={currentIndex}
+            aria-label={t('models.reasoningLevel')}
+            aria-valuetext={currentLabel}
+            onChange={(event) =>
+              onChange(safeLevels[Number(event.target.value)])
+            }
+          />
+        </div>
+      </PopoverContent>
+    </Popover>
+  );
+}

web/src/app/infra/entities/api/index.ts

@@ -99,9 +99,32 @@ export interface LLMModel {
   provider?: ModelProvider;
   abilities?: string[];
   context_length?: number | null;
+  reasoning_config?: ReasoningConfig;
+  reasoning_capabilities?: ReasoningCapabilities;
   extra_args?: object;
 }
 
+export type ReasoningLevel =
+  | 'provider_default'
+  | 'disabled'
+  | 'enabled'
+  | 'minimal'
+  | 'low'
+  | 'medium'
+  | 'high'
+  | 'xhigh'
+  | 'max';
+
+export interface ReasoningConfig {
+  level: ReasoningLevel;
+}
+
+export interface ReasoningCapabilities {
+  supported: boolean;
+  levels: ReasoningLevel[];
+  source: 'litellm' | 'provider' | 'manual' | 'unknown';
+}
+
 export interface ApiRespProviderEmbeddingModels {
   models: EmbeddingModel[];
 }

web/src/i18n/locales/en-US.ts

@@ -214,6 +214,19 @@ const enUS = {
     selectModelAbilities: 'Select model abilities',
     visionAbility: 'Vision Ability',
     functionCallAbility: 'Function Call',
+    reasoningAbility: 'Reasoning',
+    reasoningLevel: 'Reasoning level',
+    reasoningLevels: {
+      providerDefault: 'Provider default',
+      disabled: 'Off',
+      enabled: 'On',
+      minimal: 'Minimal',
+      low: 'Low',
+      medium: 'Medium',
+      high: 'High',
+      xhigh: 'Extra high',
+      max: 'Maximum',
+    },
     contextLength: 'Context Window',
     contextLengthPlaceholder: 'Unknown',
     contextLengthInvalid: 'Context window must be a positive integer',

web/src/i18n/locales/ja-JP.ts

@@ -217,6 +217,19 @@ const jaJP = {
     selectModelAbilities: 'モデル機能を選択',
     visionAbility: '視覚機能',
     functionCallAbility: '関数呼び出し',
+    reasoningAbility: '推論',
+    reasoningLevel: '推論レベル',
+    reasoningLevels: {
+      providerDefault: 'Provider デフォルト',
+      disabled: 'オフ',
+      enabled: 'オン',
+      minimal: '最小',
+      low: '低',
+      medium: '中',
+      high: '高',
+      xhigh: '最高',
+      max: '最大',
+    },
     contextLength: 'コンテキストウィンドウ',
     contextLengthPlaceholder: '不明',
     contextLengthInvalid:

web/src/i18n/locales/zh-Hans.ts

@@ -204,6 +204,19 @@ const zhHans = {
     selectModelAbilities: '选择模型能力',
     visionAbility: '视觉能力',
     functionCallAbility: '函数调用',
+    reasoningAbility: '思考能力',
+    reasoningLevel: '思考档位',
+    reasoningLevels: {
+      providerDefault: 'Provider 默认',
+      disabled: '关闭',
+      enabled: '开启',
+      minimal: '最低',
+      low: '低',
+      medium: '中',
+      high: '高',
+      xhigh: '极高',
+      max: '最大',
+    },
     contextLength: '上下文窗口',
     contextLengthPlaceholder: '未知',
     contextLengthInvalid: '上下文窗口必须是正整数',

web/tests/unit/dynamic-form-save-values.test.mjs

@@ -34,12 +34,26 @@ test('normalizes only single-line text fields in a dynamic form save snapshot',
     { name: 'multiline', type: 'text', default: '' },
     { name: 'string-list', type: 'array[string]', default: [] },
     { name: 'count', type: 'integer', default: 0 },
+    {
+      name: 'model',
+      type: 'model-fallback-selector',
+      default: { primary: '', fallbacks: [], reasoning: {} },
+    },
   ];
   const values = {
     'single-line': '\t  hello  world \n',
     multiline: '  keep multiline whitespace  \n',
     'string-list': ['  first  ', ' second '],
     count: 3,
+    model: {
+      primary: 'primary-model',
+      fallbacks: ['fallback-model'],
+      reasoning: {
+        'primary-model': 'high',
+        'fallback-model': 'provider_default',
+        'removed-model': 'medium',
+      },
+    },
   };
 
   assert.deepEqual(normalizeDynamicFormValuesForSave(specs, values), {
@@ -47,5 +61,12 @@ test('normalizes only single-line text fields in a dynamic form save snapshot',
     multiline: '  keep multiline whitespace  \n',
     'string-list': ['  first  ', ' second '],
     count: 3,
+    model: {
+      primary: 'primary-model',
+      fallbacks: ['fallback-model'],
+      reasoning: {
+        'primary-model': 'high',
+      },
+    },
   });
 });