Replace temperature with reasoning effort and retry in-stream errors

- Providers get a default reasoning effort (low..max, empty = not sent);
  chat can override it per message. OpenAI sends reasoning_effort,
  Claude sends output_config.effort with adaptive thinking, and thinking
  blocks are replayed during the tool loop.
- Show a thinking indicator while waiting for model output.
- Surface in-stream upstream errors (e.g. overloaded) instead of
  breaking the SSE stream, and retry them up to 3 times when nothing
  has been emitted yet.

web/src/apis/settings.ts

@@ -2,6 +2,11 @@ import { http } from "@/http"
 
 export type Protocol = "openai" | "anthropic"
 
+/** 思考强度，与后端 llm.Effort 一致 */
+export const EFFORTS = ["low", "medium", "high", "xhigh", "max"] as const
+
+export type Effort = typeof EFFORTS[number]
+
 export interface Provider {
   id: number
   name: string
@@ -11,7 +16,8 @@ export interface Provider {
   model: string
   max_tokens: number
   context_window: number
-  temperature: number | null
+  /** 为空表示不传，由模型决定 */
+  reasoning_effort: Effort | null
   is_default: boolean
   updated_at: string
 }
@@ -25,7 +31,8 @@ export interface ProviderForm {
   model: string
   max_tokens: number
   context_window: number
-  temperature: number | null
+  /** 为空表示不传，由模型决定 */
+  reasoning_effort: Effort | null
   is_default: boolean
 }
 
@@ -53,6 +60,6 @@ export function deleteProvider(id: number): Promise<void> {
 }
 
 export function testProvider(id: number): Promise<TestResult> {
-  // 模型接口超时 120s 且 SDK 会重试，前端放宽等待时间
-  return http.post(`/settings/providers/${id}/test`, null, { timeout: 400000 })
+  // 模型接口读超时 600s 且 SDK 会重试 2 次，前端放宽等待时间
+  return http.post(`/settings/providers/${id}/test`, null, { timeout: 1900000 })
 }

web/src/pages/chat/components/MessageList.vue

@@ -18,6 +18,8 @@ export interface DisplayItem {
   error?: string
   /** 生成中：这次模型调用已结束，后续文字另起一条 */
   closed?: boolean
+  /** 生成中：模型正在思考，还没有输出文字 */
+  thinking?: boolean
 }
 
 defineProps<{ items: DisplayItem[] }>()
@@ -53,6 +55,9 @@ function prettyResult(result: string) {
       </div>
       <div v-else class="flex flex-col gap-2">
         <!-- markdown-it 已关闭 html，模型输出中的标签按文本显示 -->
+        <div v-if="item.thinking" class="text-xs text-[var(--el-text-color-secondary)]">
+          思考中…
+        </div>
         <div v-if="item.text" class="markdown-body" v-html="renderMarkdown(item.text)" />
         <el-collapse v-if="item.toolCalls.length" class="tool-cards">
           <el-collapse-item v-for="call in item.toolCalls" :key="call.id" :name="call.id">

web/src/pages/chat/index.vue

@@ -1,11 +1,11 @@
 <script lang="ts" setup>
 import type { ChatMessage } from "@/apis/chat"
-import type { Provider } from "@/apis/settings"
+import type { Effort, Provider } from "@/apis/settings"
 import type { SkillNode } from "@/apis/skills"
 import type { DisplayItem } from "@/pages/chat/components/MessageList.vue"
 import { ElNotification } from "element-plus"
 import { getChatSession, resetChat } from "@/apis/chat"
-import { listProviders } from "@/apis/settings"
+import { EFFORTS, listProviders } from "@/apis/settings"
 import { listSkillTree } from "@/apis/skills"
 import SkillTree from "@/components/SkillTree.vue"
 import SkillVersionDrawer from "@/components/SkillVersionDrawer.vue"
@@ -22,6 +22,8 @@ const contextTab = ref<InstanceType<typeof ContextTab>>()
 const providers = ref<Provider[]>([])
 const providerId = usePersisted<number | null>("chat.providerId", null)
 const provider = computed(() => providers.value.find(p => p.id === providerId.value))
+// 为空表示跟随模型配置里的默认值
+const effort = usePersisted<Effort | null>("chat.effort", null)
 
 const messages = ref<ChatMessage[]>([])
 const estimatedTokens = ref(0)
@@ -94,11 +96,19 @@ function currentAssistant(): DisplayItem {
 function onEvent(name: string, data: any) {
   if (name === "prompt") {
     estimatedTokens.value = data.estimated_tokens
+    // OpenAI 格式推理期间没有任何推送，先显示等待状态直到出字
+    currentAssistant().thinking = true
+  } else if (name === "thinking") {
+    currentAssistant().thinking = true
   } else if (name === "delta") {
-    currentAssistant().text += data.text
+    const item = currentAssistant()
+    item.thinking = false
+    item.text += data.text
   } else if (name === "usage") {
     // 一次模型调用结束；之后的工具调用挂在这条回复上，新的文字另起一条
-    currentAssistant().closed = true
+    const item = currentAssistant()
+    item.closed = true
+    item.thinking = false
   } else if (name === "tool_call") {
     const last = [...liveItems.value].reverse().find(i => i.kind === "assistant") ?? currentAssistant()
     last.toolCalls.push({ id: data.id, name: data.name, arguments: data.arguments, result: null })
@@ -108,6 +118,10 @@ function onEvent(name: string, data: any) {
       if (call)
         call.result = data.preview
     }
+    // 工具都返回后进入下一次模型调用
+    const last = liveItems.value.at(-1)
+    if (last?.closed && last.toolCalls.every(c => c.result !== null))
+      currentAssistant().thinking = true
   } else if (name === "skill_draft") {
     highlightIds.value = [...new Set([...highlightIds.value, data.skill_id])]
     loadSkills()
@@ -115,7 +129,9 @@ function onEvent(name: string, data: any) {
   } else if (name === "context_updated") {
     contextTab.value?.refreshFromChat()
   } else if (name === "error") {
-    currentAssistant().error = data.message
+    const item = currentAssistant()
+    item.thinking = false
+    item.error = data.message
   }
   scrollToBottom()
 }
@@ -131,11 +147,11 @@ async function submit() {
   input.value = ""
   liveItems.value = [{ key: `live-user${Date.now()}`, kind: "user", text, toolCalls: [] }]
   scrollToBottom()
-  await send({ text, provider_id: provider.value.id }, onEvent)
+  await send({ text, provider_id: provider.value.id, reasoning_effort: effort.value }, onEvent)
   // 无论完成、出错还是停止，都以服务端保存的会话为准；保留本地的错误提示
   const errors = liveItems.value.filter(i => i.error)
   await loadSession()
-  liveItems.value = errors.map(e => ({ ...e, text: "", toolCalls: [] }))
+  liveItems.value = errors.map(e => ({ ...e, text: "", toolCalls: [], thinking: false }))
 }
 
 function onKeydown(e: Event | KeyboardEvent) {
@@ -222,6 +238,9 @@ loadSkills()
               <el-select v-model="providerId" placeholder="选择模型配置" class="w-60" :disabled="streaming">
                 <el-option v-for="p in providers" :key="p.id" :value="p.id" :label="`${p.name}（${p.model}）`" />
               </el-select>
+              <el-select v-model="effort" class="w-44" :disabled="streaming" clearable :value-on-clear="null" :placeholder="`思考强度：跟随配置（${provider?.reasoning_effort ?? '不设置'}）`">
+                <el-option v-for="e in EFFORTS" :key="e" :value="e" :label="`思考强度：${e}`" />
+              </el-select>
               <el-tooltip content="下一轮发给模型的 system prompt（含已发布 skills、CONTEXT.md）与历史消息的估算大小" placement="bottom">
                 <span class="text-xs" :class="tokenRatio > 0.8 ? 'text-red-500 font-bold' : 'text-[var(--el-text-color-secondary)]'">
                   上下文约 {{ estimatedTokens.toLocaleString() }}{{ provider ? ` / ${provider.context_window.toLocaleString()}` : "" }} token

web/src/pages/settings/index.vue

@@ -1,6 +1,6 @@
 <script lang="ts" setup>
 import type { Provider, ProviderForm } from "@/apis/settings"
-import { createProvider, deleteProvider, listProviders, patchProvider, testProvider } from "@/apis/settings"
+import { createProvider, deleteProvider, EFFORTS, listProviders, patchProvider, testProvider } from "@/apis/settings"
 import { formatTime } from "@/utils/text"
 
 const PROTOCOL_LABELS = { openai: "OpenAI 格式（/v1/chat/completions）", anthropic: "Claude 格式（/v1/messages）" } as const
@@ -12,7 +12,7 @@ const saving = ref(false)
 const testing = ref<number | null>(null)
 
 function emptyForm(): ProviderForm {
-  return { name: "", protocol: "anthropic", base_url: "", api_key: "", model: "", max_tokens: 8192, context_window: 200000, temperature: null, is_default: false }
+  return { name: "", protocol: "anthropic", base_url: "", api_key: "", model: "", max_tokens: 8192, context_window: 200000, reasoning_effort: null, is_default: false }
 }
 const form = ref<ProviderForm>(emptyForm())
 
@@ -106,6 +106,11 @@ load()
       </el-table-column>
       <el-table-column prop="base_url" label="Base URL" min-width="200" />
       <el-table-column prop="model" label="模型" min-width="150" />
+      <el-table-column label="思考强度" width="90">
+        <template #default="{ row }">
+          {{ (row as Provider).reasoning_effort ?? "-" }}
+        </template>
+      </el-table-column>
       <el-table-column label="API Key" width="100">
         <template #default="{ row }">
           ****{{ row.api_key_hint }}
@@ -169,9 +174,11 @@ load()
           <el-input-number v-model="form.context_window" :min="4096" :max="2000000" :step="10000" controls-position="right" />
           <span class="ml-2 text-xs">token，超过时拒绝发送</span>
         </el-form-item>
-        <el-form-item label="Temperature">
-          <el-input-number v-model="form.temperature" :min="0" :max="2" :step="0.1" :value-on-clear="null" controls-position="right" placeholder="默认" />
-          <span class="ml-2 text-xs">留空使用模型默认值</span>
+        <el-form-item label="思考强度">
+          <el-select v-model="form.reasoning_effort" clearable placeholder="不设置" class="w-40" :value-on-clear="null">
+            <el-option v-for="e in EFFORTS" :key="e" :value="e" :label="e" />
+          </el-select>
+          <span class="ml-2 text-xs">Chat 默认值，可在 Chat 页按轮调整；留空则不传，由模型决定。越高越慢、越贵</span>
         </el-form-item>
         <el-form-item label="设为默认">
           <el-switch v-model="form.is_default" />