diff --git a/dashboard/src/components/Pagination.vue b/dashboard/src/components/Pagination.vue
new file mode 100644
index 0000000..fe9877c
--- /dev/null
+++ b/dashboard/src/components/Pagination.vue
@@ -0,0 +1,48 @@
+<template>
+  <div class="pagination">
+    <button class="page-btn" :disabled="page <= 1" @click="$emit('page-change', page - 1)">‹</button>
+    <template v-for="p in visiblePages">
+      <span v-if="p === '...'" :key="p + '-ellipsis'" class="page-ellipsis">…</span>
+      <button v-else :key="p" class="page-btn" :class="{ active: p === page }" @click="$emit('page-change', p)">{{ p }}</button>
+    </template>
+    <button class="page-btn" :disabled="page >= totalPages" @click="$emit('page-change', page + 1)">›</button>
+    <span class="page-info">共 {{ total }} 条</span>
+  </div>
+</template>
+
+<script setup>
+import { computed } from 'vue'
+
+const props = defineProps({
+  page: { type: Number, required: true },
+  totalPages: { type: Number, required: true },
+  total: { type: Number, default: 0 }
+})
+
+defineEmits(['page-change'])
+
+const visiblePages = computed(() => {
+  const tp = props.totalPages
+  const cur = props.page
+  if (tp <= 7) return Array.from({ length: tp }, (_, i) => i + 1)
+  const pages = []
+  if (cur <= 4) {
+    pages.push(1, 2, 3, 4, 5, '...', tp)
+  } else if (cur >= tp - 3) {
+    pages.push(1, '...', tp - 4, tp - 3, tp - 2, tp - 1, tp)
+  } else {
+    pages.push(1, '...', cur - 1, cur, cur + 1, '...', tp)
+  }
+  return pages
+})
+</script>
+
+<style scoped>
+.pagination { display: flex; align-items: center; gap: 0.25rem; flex-wrap: wrap; }
+.page-btn { padding: 0.25rem 0.5rem; border: 1px solid #d1d5db; background: white; border-radius: 0.25rem; cursor: pointer; font-size: 0.875rem; }
+.page-btn:hover:not(:disabled) { background: #f3f4f6; }
+.page-btn:disabled { opacity: 0.4; cursor: not-allowed; }
+.page-btn.active { background: #3b82f6; color: white; border-color: #3b82f6; }
+.page-ellipsis { padding: 0.25rem 0.25rem; color: #6b7280; }
+.page-info { font-size: 0.875rem; color: #6b7280; margin-left: 0.5rem; }
+</style>
\ No newline at end of file
diff --git a/dashboard/src/components/ProcessBlock.vue b/dashboard/src/components/ProcessBlock.vue
index ab86c82..e9e7c2a 100644
--- a/dashboard/src/components/ProcessBlock.vue
+++ b/dashboard/src/components/ProcessBlock.vue
@@ -142,7 +142,8 @@ const allItems = computed(() => {
               displayContent = JSON.stringify(parsed, null, 2)
             }
           } catch (e) {
-            // 不是 JSON，保持原样
+            // 不是 JSON，保持原样显示原始内容
+            console.warn('Failed to parse tool result JSON:', e)
           }
           
           match.resultSummary = displayContent.slice(0, 200)
diff --git a/dashboard/src/utils/streamManager.js b/dashboard/src/utils/streamManager.js
index 6ea968b..90ae290 100644
--- a/dashboard/src/utils/streamManager.js
+++ b/dashboard/src/utils/streamManager.js
@@ -149,7 +149,8 @@ class StreamManager {
           const data = JSON.parse(line.slice(6))
           this._handleEvent(conversationId, currentEvent, data, streamStore, null)
         } catch (e) {
-          // 忽略解析错误
+          console.error('SSE parse error in _processLines:', e, 'line:', line)
+          streamStore.errorStream(conversationId, `Parse error: ${e.message}`)
         }
       }
     }
diff --git a/dashboard/src/utils/streamStore.js b/dashboard/src/utils/streamStore.js
index 639897c..910fe6b 100644
--- a/dashboard/src/utils/streamStore.js
+++ b/dashboard/src/utils/streamStore.js
@@ -48,7 +48,16 @@ export const useStreamStore = defineStore('stream', () => {
     
     const idx = state.process_steps.findIndex(s => s.id === step.id)
     if (idx >= 0) {
-      state.process_steps[idx] = step
+      // 对于 thinking/text 步骤，后端发送的是增量内容（基于 offset）
+      // 需要追加到已有内容上，而不是替换
+      if (step.type === 'thinking' || step.type === 'text') {
+        const existing = state.process_steps[idx]
+        existing.content = (existing.content || '') + (step.content || '')
+        // 触发响应式更新
+        state.process_steps[idx] = { ...existing }
+      } else {
+        state.process_steps[idx] = step
+      }
     } else {
       state.process_steps.push(step)
     }
diff --git a/dashboard/src/views/ConversationView.vue b/dashboard/src/views/ConversationView.vue
index b1e76cd..c67a964 100644
--- a/dashboard/src/views/ConversationView.vue
+++ b/dashboard/src/views/ConversationView.vue
@@ -48,11 +48,12 @@
           </div>
         </div>
 
-        <div v-if="totalPages > 1" class="sidebar-pagination">
-          <button @click="page--; fetchData()" :disabled="page === 1" class="btn-page">‹</button>
-          <span>{{ page }} / {{ totalPages }}</span>
-          <button @click="page++; fetchData()" :disabled="page >= totalPages" class="btn-page">›</button>
-        </div>
+        <Pagination
+          v-if="totalPages > 1"
+          :page="page"
+          :total-pages="totalPages"
+          @page-change="handlePageChange"
+        />
       </aside>
 
       <!-- 右侧内容区 - 对话界面 -->
@@ -170,6 +171,7 @@ import { formatDate } from '../utils/useFormatters.js'
 import ProcessBlock from '../components/ProcessBlock.vue'
 import MessageNav from '../components/MessageNav.vue'
 import MessageBubble from '../components/MessageBubble.vue'
+import Pagination from '../components/Pagination.vue'
 
 const {
   list,
@@ -277,6 +279,12 @@ const handleRegenerateMessage = async (msgId) => {
 }
 
 
+// 处理分页切换
+const handlePageChange = (newPage) => {
+  page.value = newPage
+  fetchData()
+}
+
 const onProviderChange = () => {
   const p = providers.value.find(p => p.id === form.value.provider_id)
   if (p) form.value.model = p.default_model || ''
diff --git a/dashboard/src/views/SettingsView.vue b/dashboard/src/views/SettingsView.vue
index 4db1338..bf9c80f 100644
--- a/dashboard/src/views/SettingsView.vue
+++ b/dashboard/src/views/SettingsView.vue
@@ -155,6 +155,7 @@
       <div class="section-title">
         <span class="section-icon">🔌</span>
         <span class="section-text">LLM Provider</span>
+        <span class="section-count">{{ providers.length }}</span>
       </div>
 
       <div v-if="loading" class="loading"><div class="spinner"></div>加载中...</div>
@@ -163,16 +164,17 @@
         <table class="settings-table">
           <thead>
             <tr>
-              <th>名称</th>
+              <th>名称 / 类型</th>
               <th>API / 模型</th>
               <th class="switch-col">启用</th>
               <th class="ops-col">操作</th>
             </tr>
           </thead>
           <tbody>
-            <tr v-for="p in providers" :key="p.id">
+            <tr v-for="p in paginatedProviders" :key="p.id">
               <td class="name-col">
                 <div class="provider-name">{{ p.name }}</div>
+                <div class="provider-type">{{ p.provider_type }}</div>
                 <div class="provider-badges">
                   <span v-if="p.is_default" class="badge badge-default">默认</span>
                   <span v-if="p.enabled" class="badge badge-enabled">启用</span>
@@ -180,9 +182,9 @@
                 </div>
               </td>
               <td class="info-col">
-                <div class="info-item">{{ p.base_url }}</div>
-                <div class="info-item sub">模型: {{ p.default_model }}</div>
-                <div class="info-item sub">最大Tokens: {{ p.max_tokens || 8192 }}</div>
+                <div class="info-item" :title="p.base_url">{{ p.base_url }}</div>
+                <div class="info-item sub">模型: <code>{{ p.default_model }}</code></div>
+                <div class="info-item sub">最大 Tokens: {{ p.max_tokens || 8192 }}</div>
               </td>
               <td class="switch-col">
                 <label class="switch" @click.prevent="toggleEnabled(p)">
@@ -208,7 +210,15 @@
           <tfoot>
             <tr>
               <td colspan="4" class="table-footer">
-                <button @click="showModal = true" class="btn-primary">
+                <Pagination
+                  v-if="providerTotalPages > 1"
+                  :page="providerPage"
+                  :total-pages="providerTotalPages"
+                  :total="providers.length"
+                  @page-change="handleProviderPageChange"
+                  class="pagination-inline"
+                />
+                <button @click="showModal = true; editing = null" class="btn-primary">
                   <svg width="16" height="16" style="margin-right: 6px;"><use href="#plus-icon"/></svg>
                   添加 Provider
                 </button>
@@ -240,7 +250,7 @@
             </tr>
           </thead>
           <tbody>
-            <tr v-for="u in users" :key="u.id">
+            <tr v-for="u in paginatedUsers" :key="u.id">
               <td class="name-col">{{ u.username }}</td>
               <td class="info-col">{{ u.email || '-' }}</td>
               <td>{{ u.role }}</td>
@@ -257,6 +267,19 @@
               </td>
             </tr>
           </tbody>
+          <tfoot v-if="userTotalPages > 1">
+            <tr>
+              <td colspan="5" class="table-footer">
+                <Pagination
+                  :page="userPage"
+                  :total-pages="userTotalPages"
+                  :total="users.length"
+                  @page-change="handleUserPageChange"
+                  class="pagination-inline"
+                />
+              </td>
+            </tr>
+          </tfoot>
         </table>
       </div>
       <div v-else class="empty-card">暂无用户</div>
@@ -308,23 +331,55 @@
 
     <!-- 添加/编辑 Provider 模态框 -->
     <div v-if="showModal" class="modal-overlay" @click.self="closeModal">
-      <div class="modal">
+      <div class="modal provider-modal">
         <h2>{{ editing ? '编辑 Provider' : '添加 Provider' }}</h2>
         
-        <div class="form-group"><label>名称</label><input v-model="form.name" placeholder="如: 我的 DeepSeek" /></div>
-        <div class="form-group"><label>Base URL</label><input v-model="form.base_url" placeholder="https://api.deepseek.com/chat/completions" /></div>
-        <div class="form-group"><label>API Key</label><input v-model="form.api_key" type="text" :placeholder="editing ? '留空则保持原密码' : 'sk-...'" /></div>
+        <div class="form-group">
+          <label>名称 <span class="required">*</span></label>
+          <input v-model="form.name" placeholder="如: 我的 DeepSeek" />
+        </div>
         
         <div class="form-group">
-          <label>模型名称</label>
+          <label>Provider 类型 <span class="required">*</span></label>
+          <select v-model="form.provider_type" class="form-select">
+            <option value="openai">OpenAI</option>
+            <option value="anthropic">Anthropic</option>
+            <option value="deepseek">DeepSeek</option>
+            <option value="glm">GLM (智谱)</option>
+            <option value="openai-compatible">OpenAI 兼容</option>
+          </select>
+          <span class="hint">不同类型的 API 格式和认证方式不同</span>
+        </div>
+        
+        <div class="form-group">
+          <label>Base URL <span class="required">*</span></label>
+          <input v-model="form.base_url" placeholder="https://api.deepseek.com/v1" />
+        </div>
+        
+        <div class="form-group">
+          <label>API Key</label>
+          <input v-model="form.api_key" type="text" :placeholder="editing ? '留空则保持原密码' : 'sk-...'" />
+          <span v-if="editing" class="hint">留空则不修改已有 Key</span>
+        </div>
+        
+        <div class="form-group">
+          <label>模型名称 <span class="required">*</span></label>
           <input v-model="form.default_model" placeholder="deepseek-chat / gpt-4" required />
         </div>
+        
         <div class="form-group">
           <label>最大 Tokens</label>
           <input v-model.number="form.max_tokens" type="number" placeholder="8192" min="1" />
           <span class="hint">单次回复最大 token 数，默认 8192</span>
         </div>
 
+        <div class="form-group checkbox-group">
+          <label class="checkbox-label">
+            <input v-model="form.is_default" type="checkbox" />
+            <span>设为默认 Provider</span>
+          </label>
+        </div>
+
         <div v-if="formError" class="error">{{ formError }}</div>
 
         <div class="modal-actions">
@@ -343,11 +398,12 @@
 </template>
 
 <script setup>
-import { ref, onMounted } from 'vue'
+import { ref, onMounted, computed } from 'vue'
 import { providersAPI } from '../utils/api.js'
 import { useAuth } from '../utils/useAuth.js'
 import { authAPI } from '../utils/api.js'
 import { useRouter } from 'vue-router'
+import Pagination from '../components/Pagination.vue'
 
 const router = useRouter()
 const { logout, user: currentUser } = useAuth()
@@ -465,8 +521,32 @@ const testing = ref(null)
 const testResult = ref(null)
 const formError = ref('')
 
+// Provider 分页
+const providerPage = ref(1)
+const providerPageSize = 10
+const providerTotalPages = computed(() => Math.max(1, Math.ceil(providers.value.length / providerPageSize)))
+const paginatedProviders = computed(() => {
+  const start = (providerPage.value - 1) * providerPageSize
+  return providers.value.slice(start, start + providerPageSize)
+})
+const handleProviderPageChange = (newPage) => {
+  providerPage.value = newPage
+}
+
+// 用户分页
+const userPage = ref(1)
+const userPageSize = 10
+const userTotalPages = computed(() => Math.max(1, Math.ceil(users.value.length / userPageSize)))
+const paginatedUsers = computed(() => {
+  const start = (userPage.value - 1) * userPageSize
+  return users.value.slice(start, start + userPageSize)
+})
+const handleUserPageChange = (newPage) => {
+  userPage.value = newPage
+}
+
 const form = ref({
-  name: '', base_url: '', api_key: '', default_model: '', max_tokens: 8192
+  name: '', provider_type: 'openai', base_url: '', api_key: '', default_model: '', max_tokens: 8192, is_default: false
 })
 
 const fetchProviders = async () => {
@@ -504,7 +584,7 @@ const fetchSettings = async () => {
 const closeModal = () => {
   showModal.value = false
   editing.value = null
-  form.value = { name: '', base_url: '', api_key: '', default_model: '', max_tokens: 8192 }
+  form.value = { name: '', provider_type: 'openai', base_url: '', api_key: '', default_model: '', max_tokens: 8192, is_default: false }
   formError.value = ''
 }
 
@@ -515,10 +595,12 @@ const editProvider = async (p) => {
     if (res.success && res.data) {
       form.value = {
         name: res.data.name,
+        provider_type: res.data.provider_type || 'openai',
         base_url: res.data.base_url,
         api_key: '',  // 不显示原密码
         default_model: res.data.default_model,
-        max_tokens: res.data.max_tokens || 8192
+        max_tokens: res.data.max_tokens || 8192,
+        is_default: res.data.is_default || false
       }
     }
   } catch (e) {
@@ -663,6 +745,19 @@ onMounted(() => {
 .section-title { display: flex; align-items: center; gap: 0.5rem; margin-bottom: 0.75rem; padding: 0.5rem 0; }
 .section-icon { font-size: 1rem; }
 .section-text { font-size: 1rem; font-weight: 700; color: var(--text-primary); }
+.section-count {
+  display: inline-flex;
+  align-items: center;
+  justify-content: center;
+  width: 22px;
+  height: 22px;
+  border-radius: 50%;
+  background: var(--accent-primary);
+  color: white;
+  font-size: 0.7rem;
+  font-weight: 600;
+  margin-left: 0.35rem;
+}
 
 /* 内联输入框 */
 .inline-select, .inline-input { padding: 0.5rem 0.75rem; border: 1px solid var(--border-input); border-radius: 6px; background: var(--bg-input); color: var(--text-primary); font-size: 0.85rem; }
@@ -678,6 +773,12 @@ onMounted(() => {
 
 /* 列宽 */
 .name-col { width: 15%; min-width: 120px; }
+.provider-type {
+  font-size: 0.75rem;
+  color: var(--text-tertiary);
+  margin-top: 0.2rem;
+  font-family: monospace;
+}
 .info-col { width: 60%; min-width: 200px; }
 .switch-col { text-align: center; width: 80px; }
 .ops-col { width: 15%; min-width: 180px; text-align: center; }
@@ -692,6 +793,7 @@ onMounted(() => {
 
 /* 表格底部 */
 .table-footer { text-align: right; padding: 0.75rem 1rem; background: var(--bg-secondary); border-top: 1px solid var(--border-light); }
+.pagination-inline { display: inline-flex; margin-right: 1rem; border-top: none; padding: 0; background: transparent; }
 
 /* Provider 单元格 */
 .provider-name { font-weight: 600; font-size: 0.9rem; color: var(--text-primary); }
@@ -731,6 +833,7 @@ onMounted(() => {
 /* 模态框 */
 .modal-overlay { position: fixed; top: 0; left: 0; right: 0; bottom: 0; background: rgba(0,0,0,0.5); display: flex; align-items: center; justify-content: center; z-index: 1000; }
 .modal { background: var(--bg-primary); border-radius: 16px; padding: 1.25rem; width: 100%; max-width: 400px; }
+.provider-modal { max-width: 480px; }
 .modal h2 { margin: 0 0 1.25rem; color: var(--text-primary); font-size: 1.1rem; }
 
 .result-modal { text-align: center; }
@@ -745,9 +848,37 @@ onMounted(() => {
 .result-modal .result-json { background: var(--bg-code); border: 1px solid var(--border-light); border-radius: 8px; padding: 0.75rem; margin: 0.75rem 0; max-height: 150px; overflow: auto; font-size: 0.75rem; text-align: left; white-space: pre-wrap; word-break: break-all; }
 .modal-actions { display: flex; justify-content: flex-end; gap: 0.75rem; margin-top: 1.25rem; }
 .form-group { margin-bottom: 1rem; }
+.form-group .required { color: var(--danger-color); }
+.form-select {
+  width: 100%;
+  padding: 0.65rem;
+  border: 1px solid var(--border-input);
+  border-radius: 8px;
+  background: var(--bg-input);
+  box-sizing: border-box;
+  color: var(--text-primary);
+  font-size: 0.9rem;
+}
+.checkbox-group { margin-top: 0.5rem; }
+.checkbox-label {
+  display: flex;
+  align-items: center;
+  gap: 0.5rem;
+  cursor: pointer;
+}
+.checkbox-label input[type="checkbox"] {
+  width: auto;
+  cursor: pointer;
+}
 .form-group label { display: block; margin-bottom: 0.5rem; font-weight: 500; color: var(--text-primary); font-size: 0.9rem; }
 .form-group input { width: 100%; padding: 0.65rem; border: 1px solid var(--border-input); border-radius: 8px; background: var(--bg-input); box-sizing: border-box; color: var(--text-primary); font-size: 0.9rem; }
 .form-group .hint { font-size: 0.75rem; color: var(--text-tertiary); margin-top: 4px; display: block; }
+.info-item code {
+  background: var(--bg-code);
+  padding: 0.1rem 0.3rem;
+  border-radius: 3px;
+  font-size: 0.8rem;
+}
 .optional { color: var(--text-tertiary); font-weight: normal; font-size: 0.8rem; }
 .error { color: var(--danger-color); background: var(--danger-bg); padding: 0.75rem; border-radius: 8px; margin-top: 0.75rem; font-size: 0.85rem; }
 
diff --git a/luxx/routes/messages.py b/luxx/routes/messages.py
index 656ce0a..8ecf396 100644
--- a/luxx/routes/messages.py
+++ b/luxx/routes/messages.py
@@ -150,7 +150,8 @@ async def stream_message(
             user_id=current_user.id,
             username=current_user.username,
             workspace=workspace,
-            user_permission_level=current_user.permission_level
+            user_permission_level=current_user.permission_level,
+            skip_user_message=True  # User message already saved in DB above
         ):
             # Chat service returns raw SSE strings (including done event)
             yield sse_str
diff --git a/luxx/routes/tools.py b/luxx/routes/tools.py
index aab7529..26dac0d 100644
--- a/luxx/routes/tools.py
+++ b/luxx/routes/tools.py
@@ -6,13 +6,29 @@ from pydantic import BaseModel
 from luxx.database import get_db
 from luxx.models import User
 from luxx.routes.auth import get_current_user
-from luxx.tools.core import registry
+from luxx.tools.core import registry, CommandPermission
 from luxx.utils.helpers import success_response
 
 
 router = APIRouter(prefix="/tools", tags=["Tools"])
 
 
+def _tool_to_detail(tool) -> Dict[str, Any]:
+    """Convert tool definition to a detailed dict with permission and category info"""
+    # to_openai_format returns {"type": "function", "function": {name, description, parameters}}
+    # Flatten the function properties to top level for easier frontend access
+    base = tool.to_openai_format()
+    func = base.get("function", {})
+    return {
+        "name": func.get("name", tool.name),
+        "description": func.get("description", tool.description),
+        "parameters": func.get("parameters", tool.parameters),
+        "category": tool.category,
+        "required_permission": tool.required_permission.name,
+        "required_permission_level": int(tool.required_permission)
+    }
+
+
 @router.get("/", response_model=dict)
 def list_tools(
     category: Optional[str] = None,
@@ -23,17 +39,17 @@ def list_tools(
     
     if category:
         categorized_tools = [t for t in registry._tools.values() if t.category == category]
-        tools = [t.to_openai_format() for t in categorized_tools]
+        tools = [_tool_to_detail(t) for t in categorized_tools]
     else:
         categorized_tools = list(registry._tools.values())
-        tools = registry.list_all()
+        tools = [_tool_to_detail(t) for t in categorized_tools]
     
     categorized = {}
     for tool in categorized_tools:
         cat = tool.category
         if cat not in categorized:
             categorized[cat] = []
-        categorized[cat].append(tool.to_openai_format())
+        categorized[cat].append(_tool_to_detail(tool))
     
     return success_response(data={
         "tools": tools,
@@ -53,7 +69,7 @@ def get_tool(
     if not tool:
         return {"success": False, "message": "Tool not found", "code": 404}
     
-    return success_response(data=tool.to_openai_format())
+    return success_response(data=_tool_to_detail(tool))
 
 
 @router.post("/{name}/execute", response_model=dict)
diff --git a/luxx/services/agentic_loop.py b/luxx/services/agentic_loop.py
index 79dbaaf..09a3c16 100644
--- a/luxx/services/agentic_loop.py
+++ b/luxx/services/agentic_loop.py
@@ -2,6 +2,7 @@
 
 This module follows the Single Responsibility Principle.
 """
+import json
 import uuid
 import logging
 from typing import List, Dict, AsyncGenerator
@@ -22,6 +23,7 @@ class AgenticLoop:
     
     def __init__(self, tool_executor: ToolExecutor):
         self.tool_executor = tool_executor
+        self._reasoning_content = ""  # DeepSeek thinking mode - persists across iterations
 
     async def execute(
         self,
@@ -36,19 +38,43 @@ class AgenticLoop:
         tool_context: dict = None
     ) -> AsyncGenerator[str, None]:
         total_usage = {"prompt_tokens": 0, "completion_tokens": 0, "total_tokens": 0}
+        
+        # Reset reasoning_content for each execution (DeepSeek thinking mode)
+        self._reasoning_content = ""
 
         for iteration in range(MAX_ITERATIONS):
             # Per-iteration reset, keep previous steps and tool results
             context.reset(full_reset=False)
 
+            # BUG FIX: DeepSeek thinking mode requires reasoning_content to be passed back.
+            # Without this, DeepSeek returns:
+            #   "The `reasoning_content` in the thinking mode must be passed back to the API."
+            # We find the last assistant message and add reasoning_content from the previous
+            # iteration's accumulated thinking content.
+            if iteration > 0 and self._reasoning_content:
+                for i in range(len(messages) - 1, -1, -1):
+                    if messages[i].get("role") == "assistant":
+                        messages[i]["reasoning_content"] = self._reasoning_content
+                        break
+
             async for delta in llm.stream_call(
                 model=model,
                 messages=messages,
-                tools=tools,
+                tools=tools if iteration == 0 else None,  # Only send tools on first iteration
                 temperature=temperature,
                 max_tokens=max_tokens,
                 thinking_enabled=thinking_enabled
             ):
+                # Handle error delta - propagate as SSE error event
+                if delta.has_error():
+                    logger.error(f"[AGENTIC_LOOP] Stream error: {delta.error_msg}")
+                    yield sse_event("error", {"content": delta.error_msg})
+                    return
+                
+                # Accumulate reasoning_content for DeepSeek thinking mode
+                if delta.has_thinking():
+                    self._reasoning_content += delta.thinking
+                
                 events = self._process_delta(delta, context, total_usage)
                 for event in events:
                     yield event
@@ -88,6 +114,19 @@ class AgenticLoop:
                 "total_tokens": delta.usage.get("total_tokens", 0)
             })
 
+        # BUG FIX: Handle DeepSeek reasoning_content (delta.thinking).
+        # DeepSeek sends reasoning_content in separate deltas from normal content.
+        # These must be rendered as thinking steps AND accumulated into
+        # self._reasoning_content for subsequent API calls (line 71-72).
+        # Without this, DeepSeek returns 400:
+        #   "The `reasoning_content` in the thinking mode must be passed back to the API."
+        if delta.thinking:
+            if ctx.current_step_type != StepType.THINKING:
+                ctx.start_step(StepType.THINKING)
+            ctx.full_thinking += delta.thinking
+            events.append(StreamRenderer.render_thinking(ctx))
+            ctx._thinking_offset = len(ctx.full_thinking)
+
         if delta.content:
             result = ctx.process_content(delta.content)
             if result["should_emit"]:
@@ -120,12 +159,30 @@ class AgenticLoop:
         """Execute tools and add results to messages"""
         events = []
 
+        # Check if there's an active streaming step that needs to be finalized
+        # before tool execution (e.g. text step started but not finalized)
+        if ctx.current_step_id is not None:
+            ctx.finalize_step()
+
         for event in StreamRenderer.render_tool_calls(ctx):
             events.append(event)
 
-        tool_results = self.tool_executor.process_tool_calls_parallel(
-            ctx.tool_calls_list, tool_context or {}
-        )
+        # Execute tools and handle any exceptions
+        try:
+            tool_results = self.tool_executor.process_tool_calls_parallel(
+                ctx.tool_calls_list, tool_context or {}
+            )
+        except Exception as e:
+            logger.error(f"[EXECUTE_TOOLS] Tool execution failed: {type(e).__name__}: {e}")
+            # Generate error results for each tool call
+            tool_results = []
+            for tc in ctx.tool_calls_list:
+                tool_results.append({
+                    "tool_call_id": tc.get("id", ""),
+                    "role": "tool",
+                    "name": tc.get("function", {}).get("name", "unknown"),
+                    "content": json.dumps({"success": False, "error": f"Tool execution failed: {type(e).__name__}: {str(e)}"})
+                })
 
         # Build mapping from LLM tool_call_id to step id
         # Use index-based matching as fallback when id is empty
@@ -146,7 +203,7 @@ class AgenticLoop:
                         logger.debug(f"[EXECUTE_TOOLS] Matched by id: tc_id={tc_id} -> step_id={ref_id}")
                         break
             
-            # Fallback: use index-based matching
+            # Fallback: use index-based matching (most reliable for parallel execution)
             if ref_id is None and i < len(tool_call_steps):
                 ref_id = tool_call_steps[i].id
                 logger.debug(f"[EXECUTE_TOOLS] Matched by index: i={i} -> step_id={ref_id}")
@@ -154,16 +211,31 @@ class AgenticLoop:
             # Last resort: generate a step id
             if ref_id is None:
                 ref_id = f"step-{len(ctx.all_steps) - len(tool_results) + i}"
-                logger.debug(f"[EXECUTE_TOOLS] Generated ref_id: {ref_id}")
+                logger.warning(f"[EXECUTE_TOOLS] Could not match tool result, generated ref_id: {ref_id}")
             
             _, event = StreamRenderer.render_tool_result(ctx, tr, ref_id)
             events.append(event)
 
-        messages.append({
+        # When tool_calls exist, content should be null (OpenAI API spec)
+        assistant_msg = {
             "role": "assistant",
-            "content": ctx.full_content or "",
             "tool_calls": ctx.tool_calls_list
-        })
+        }
+        # When tool_calls exist, content must be null (don't include text before tool calls)
+        if not ctx.tool_calls_list and ctx.full_content and not ctx.full_content.isspace():
+            assistant_msg["content"] = ctx.full_content
+        # If tool_calls exist, omit content field entirely (or set to null)
+        
+        # BUG FIX: Include reasoning_content in assistant message for DeepSeek.
+        # DeepSeek's thinking mode REQUIRES reasoning_content to be echoed back
+        # to the API in every subsequent request. Without this, the API returns:
+        #   HTTP 400: "The `reasoning_content` in the thinking mode must be
+        #   passed back to the API."
+        # This field is accumulated from delta.thinking in the stream loop above.
+        if self._reasoning_content:
+            assistant_msg["reasoning_content"] = self._reasoning_content
+        
+        messages.append(assistant_msg)
         messages.extend(ctx.all_tool_results[-len(tool_results):])
 
         return events
diff --git a/luxx/services/chat.py b/luxx/services/chat.py
index 965c151..13192d9 100644
--- a/luxx/services/chat.py
+++ b/luxx/services/chat.py
@@ -187,16 +187,19 @@ class ChatService:
         user_id: int = None,
         username: str = None,
         workspace: str = None,
-        user_permission_level: int = 1
+        user_permission_level: int = 1,
+        skip_user_message: bool = False  # Skip adding user message if already in DB
     ) -> AsyncGenerator[str, None]:
         """Streaming response with Agentic Loop."""
         try:
             # Build initial messages
             messages = self.build_messages(conversation)
-            messages.append({
-                "role": "user",
-                "content": json.dumps({"text": user_message, "attachments": []})
-            })
+            # Only add user message if not already in database
+            if not skip_user_message:
+                messages.append({
+                    "role": "user",
+                    "content": json.dumps({"text": user_message, "attachments": []})
+                })
             
             # Get tools and LLM client via factory
             tools = self._get_tools(enabled_tools)
diff --git a/luxx/services/llm_adapters/openai_adapter.py b/luxx/services/llm_adapters/openai_adapter.py
index fc41c01..36e6927 100644
--- a/luxx/services/llm_adapters/openai_adapter.py
+++ b/luxx/services/llm_adapters/openai_adapter.py
@@ -35,6 +35,15 @@ class OpenAIAdapter(ProviderAdapter):
         if tools:
             body["tools"] = tools
             body["tool_choice"] = "auto"
+        
+        # DeepSeek-specific parameters
+        if "reasoning_effort" in kwargs:
+            body["reasoning_effort"] = kwargs["reasoning_effort"]
+        if "thinking_enabled" in kwargs and kwargs["thinking_enabled"]:
+            body["thinking"] = {"type": "enabled"}
+        if "response_format" in kwargs:
+            body["response_format"] = kwargs["response_format"]
+            
         return body, headers
 
     def reset(self):
@@ -79,9 +88,23 @@ class OpenAIAdapter(ProviderAdapter):
         finish_reason = choice.get("finish_reason")
         content = delta.get("content", "")
         
+        # BUG FIX: Extract DeepSeek reasoning_content (thinking mode).
+        # DeepSeek sends reasoning_content alongside normal content in stream deltas.
+        # This must be yielded as a separate ParsedDelta with only the "thinking" field
+        # set, so the agentic loop can:
+        #   1. Render it as a thinking step in the UI
+        #   2. Accumulate it to echo back in subsequent API calls
+        # Without this, DeepSeek returns 400:
+        #   "The `reasoning_content` in the thinking mode must be passed back to the API."
+        reasoning_content = delta.get("reasoning_content", "")
+        
         # MiniMax may send tool_calls as array in delta
         tool_calls = delta.get("tool_calls", [])
 
+        # Yield thinking (includes DeepSeek reasoning_content)
+        if reasoning_content:
+            yield ParsedDelta(thinking=reasoning_content)
+        
         # Yield content if present
         if content:
             yield ParsedDelta(content=content)
@@ -90,8 +113,10 @@ class OpenAIAdapter(ProviderAdapter):
         for tc in tool_calls:
             yield ParsedDelta(tool_call=tc)
 
-        # Set is_complete for final chunks (DeepSeek may return null, "length", "content_filter")
-        if finish_reason and finish_reason not in (None, ""):
+        # Only set is_complete for actual completion ("stop").
+        # DeepSeek sends "tool_calls" when it wants to call tools (stream continues with tool_call deltas).
+        # "length" / "content_filter" mean max_tokens hit - treat as incomplete, let loop handle it.
+        if finish_reason == "stop":
             yield ParsedDelta(is_complete=True)
 
     def parse_response(self, data: Dict) -> Dict:
diff --git a/luxx/services/llm_client.py b/luxx/services/llm_client.py
index 8c6b5e6..1561843 100644
--- a/luxx/services/llm_client.py
+++ b/luxx/services/llm_client.py
@@ -22,6 +22,7 @@ Usage:
 Extending Providers:
     LLMClient.register_adapter("my_provider", MyAdapter)
 """
+import json
 import logging
 import traceback
 from typing import Dict, List, Any, Optional, AsyncGenerator, Type
@@ -284,6 +285,36 @@ class LLMClient:
             logger.error(f"Sync call error: {e}\n{traceback.format_exc()}")
             raise
     
+    @staticmethod
+    def _parse_error(response: httpx.Response) -> str:
+        """Extract error message from an API error response.
+        
+        Handles various formats: JSON with "error.message", plain text, empty body.
+        This is a static method so it works both with an open stream (aread)
+        and a closed response (text fallback).
+        """
+        error_body = ""
+        try:
+            error_body_bytes = response.read()
+            if error_body_bytes:
+                error_body = error_body_bytes.decode('utf-8', errors='replace')
+        except Exception:
+            try:
+                error_body = response.text
+            except Exception:
+                pass
+        
+        if error_body:
+            try:
+                error_json = json.loads(error_body)
+                detail = error_json.get("error", {}).get("message", "") or str(error_json)
+            except json.JSONDecodeError:
+                detail = error_body
+        else:
+            detail = f"HTTP {response.status_code} (no body)"
+        
+        return detail[:500]
+
     async def stream_call(
         self,
         model: str,
@@ -316,7 +347,6 @@ class LLMClient:
         
         endpoint = self.build_endpoint()
         logger.info(f"Stream call to {endpoint} with model {model}")
-        
         try:
             async with httpx.AsyncClient(timeout=120.0) as client:
                 async with client.stream(
@@ -326,7 +356,16 @@ class LLMClient:
                     json=body
                 ) as response:
                     logger.info(f"Response status: {response.status_code}")
-                    response.raise_for_status()
+                    
+                    # BUG FIX: Read error body BEFORE raising or closing the stream.
+                    # httpx's client.stream() __aexit__ closes the response stream
+                    # when an exception occurs, making e.response.aread() return 0 bytes
+                    # in the except block below. Reading here (stream still open) fixes this.
+                    if response.status_code >= 400:
+                        error_msg = self._parse_error(response)
+                        logger.error(f"HTTP {response.status_code}: {error_msg}")
+                        yield ParsedDelta(error_msg=error_msg)
+                        return
                     
                     async for line in response.aiter_lines():
                         line = line.strip()
@@ -351,23 +390,22 @@ class LLMClient:
                         # Pass clean data to adapter (OpenAIAdapter also handles stripping,
                         # but AnthropicAdapter and others need clean JSON input)
                         async for delta in self.adapter.parse_stream_chunk(event_data):
-                            if delta.content or delta.has_tool_call() or delta.is_complete or delta.usage:
+                            # BUG FIX: Include has_thinking() in filter condition.
+                            # DeepSeek sends reasoning_content as separate deltas with only
+                            # the "thinking" field populated. Without has_thinking() check,
+                            # these deltas were silently dropped, preventing reasoning_content
+                            # accumulation and leading to "must be passed back to the API" error.
+                            if delta.content or delta.has_thinking() or delta.has_tool_call() or delta.is_complete or delta.usage:
                                 yield delta
                                 
         except httpx.HTTPStatusError as e:
-            status_code = e.response.status_code if e.response else "?"
-            error_body = ""
-            if e.response:
-                try:
-                    await e.response.aread()
-                    error_body = e.response.text
-                except Exception:
-                    pass
-            logger.error(f"HTTP error: {status_code} - {error_body}")
-            yield ParsedDelta()
+            # Fallback: httpx.HTTPStatusError with closed stream
+            error_msg = self._parse_error(e.response) if e.response else f"HTTP error"
+            logger.error(f"HTTP error (fallback): {error_msg}")
+            yield ParsedDelta(error_msg=error_msg)
         except Exception as e:
-            logger.error(f"Stream error: {type(e).__name__}: {e}\n{traceback.format_exc()}")
-            yield ParsedDelta()
+            logger.error(f"Stream call failed: {type(e).__name__}: {e}")
+            yield ParsedDelta(error_msg=f"{type(e).__name__}: {str(e)}")
 
 
 # Convenience function
diff --git a/luxx/services/llm_response.py b/luxx/services/llm_response.py
index ae35a73..34b779b 100644
--- a/luxx/services/llm_response.py
+++ b/luxx/services/llm_response.py
@@ -41,11 +41,12 @@ class Step:
 class ParsedDelta:
     """LLM streaming response delta"""
     content: str = ""
-    thinking: str = ""
+    thinking: str = ""  # Includes DeepSeek reasoning_content
     text: str = ""
     tool_call: Optional[Dict] = None
     usage: Dict[str, int] = field(default_factory=dict)
     is_complete: bool = False
+    error_msg: Optional[str] = None
 
     def has_thinking(self) -> bool:
         return bool(self.thinking)
@@ -58,3 +59,6 @@ class ParsedDelta:
 
     def has_content(self) -> bool:
         return bool(self.content) or self.has_thinking() or self.has_text() or self.has_tool_call()
+    
+    def has_error(self) -> bool:
+        return self.error_msg is not None
diff --git a/luxx/services/stream_context.py b/luxx/services/stream_context.py
index 2f35eb2..4cfbb62 100644
--- a/luxx/services/stream_context.py
+++ b/luxx/services/stream_context.py
@@ -91,6 +91,9 @@ class StreamState:
         # Track content offset per step to avoid duplication across steps
         self._text_offset = 0
         self._thinking_offset = 0
+        # Track step start offset for finalize_step to compute complete step content
+        self._step_start_text_offset = 0
+        self._step_start_thinking_offset = 0
     
     def process_content(self, content: str) -> Dict:
         """Process raw content, handling thinking tags."""
@@ -172,8 +175,10 @@ class StreamState:
         # Record content offset so this step only includes content added during it
         if step_type == StepType.TEXT:
             self._text_offset = len(self.full_content)
+            self._step_start_text_offset = len(self.full_content)
         elif step_type == StepType.THINKING:
             self._thinking_offset = len(self.full_thinking)
+            self._step_start_thinking_offset = len(self.full_thinking)
         self.step_index += 1
         return self.current_step_id
     
@@ -190,9 +195,9 @@ class StreamState:
             self.current_step_type = None
             return
         if self.current_step_type == StepType.TEXT:
-            content = self.full_content[self._text_offset:]
+            content = self.full_content[self._step_start_text_offset:]
         elif self.current_step_type == StepType.THINKING:
-            content = self.full_thinking[self._thinking_offset:]
+            content = self.full_thinking[self._step_start_thinking_offset:]
         else:
             content = ""
         step = Step(
@@ -211,15 +216,22 @@ class StreamState:
         """Accumulate tool call delta"""
         idx = tc_delta.get("index", 0)
         if idx >= len(self.tool_calls_list):
+            # BUG FIX: DeepSeek may omit "type" in tool_call deltas.
+            # Without defaulting to "function", the tool call would have
+            # type=None, causing JSON serialization issues or API rejections.
+            tc_type = tc_delta.get("type", "function")
             self.tool_calls_list.append({
                 "id": tc_delta.get("id", ""),
-                "type": "function",
+                "type": tc_type,
                 "function": {"name": "", "arguments": ""}
             })
         else:
             # Update id if provided (LLM may send id in a later delta)
             if tc_delta.get("id"):
                 self.tool_calls_list[idx]["id"] = tc_delta["id"]
+            # Ensure type field is always present
+            if "type" in tc_delta:
+                self.tool_calls_list[idx]["type"] = tc_delta["type"]
         func = tc_delta.get("function", {})
         if func.get("name"):
             self.tool_calls_list[idx]["function"]["name"] += func["name"]
diff --git a/luxx/tools/core.py b/luxx/tools/core.py
index c283446..890cdbf 100644
--- a/luxx/tools/core.py
+++ b/luxx/tools/core.py
@@ -2,7 +2,9 @@
 from dataclasses import dataclass, field
 from typing import Callable, Any, Dict, List, Optional
 from enum import IntEnum
+import logging
 
+logger = logging.getLogger(__name__)
 
 class CommandPermission(IntEnum):
     """Command permission level - higher value means more permissions"""
@@ -98,8 +100,10 @@ class ToolRegistry:
     
     def execute(self, name: str, arguments: dict, context: ToolContext = None) -> Dict[str, Any]:
         """Execute tool with optional context and automatic permission check"""
+ 
         tool = self.get(name)
         if not tool:
+            logger.error(f"[REGISTRY] Tool '{name}' not found")
             return {"success": False, "error": f"Tool '{name}' not found"}
         
         # Automatic permission check (transparent to tool function)
@@ -109,22 +113,27 @@ class ToolRegistry:
             if isinstance(user_level, int):
                 user_level = CommandPermission(user_level)
             if user_level < tool.required_permission:
+                logger.warning(f"[REGISTRY] Permission denied for tool '{name}': requires {tool.required_permission.name}, user has {user_level.name}")
                 return {
                     "success": False, 
                     "error": f"Permission denied: requires {tool.required_permission.name}, you have {user_level.name}"
                 }
         
+        logger.info(f"[REGISTRY] Executing tool '{name}' with args: {arguments}")
         try:
             # Pass context to handler if it accepts it
             if context is not None:
                 result = tool.handler(arguments, context=context)
             else:
                 result = tool.handler(arguments)
+            logger.info(f"[REGISTRY] Tool '{name}' executed successfully")
             if isinstance(result, ToolResult):
                 return result.to_dict()
             return result
         except Exception as e:
-            return {"success": False, "error": str(e)}
+            import traceback
+            logger.error(f"[REGISTRY] Tool '{name}' execution failed: {type(e).__name__}: {str(e)}\n{traceback.format_exc()}")
+            return {"success": False, "error": f"{type(e).__name__}: {str(e)}"}
     
     def clear(self) -> None:
         """Clear all tools"""
diff --git a/luxx/tools/executor.py b/luxx/tools/executor.py
index 0926dfb..920d9e0 100644
--- a/luxx/tools/executor.py
+++ b/luxx/tools/executor.py
@@ -7,11 +7,12 @@ This module follows the Single Responsibility Principle:
 """
 import json
 import time
+import logging
 from typing import List, Dict, Any, Optional
 from threading import Lock
-
 from luxx.tools.core import registry, ToolContext
 
+logger = logging.getLogger(__name__)
 
 class CacheManager:
     """Manages tool result caching"""
@@ -165,6 +166,8 @@ class ToolExecutor:
         not in completion order. This ensures proper matching between tool_call
         and tool_result steps in the frontend.
         """
+
+        
         if len(tool_calls) <= 1:
             return self.process_tool_calls(tool_calls, context)
         
@@ -201,16 +204,27 @@ class ToolExecutor:
                 # Wait for all futures and store results at correct indices
                 for future in futures_with_index:
                     idx, call_id, name, args, cache_key = futures_with_index[future]
-                    result = future.result()
-                    self.cache.set(cache_key, result)
-                    self.history.record(name, args, result)
-                    results[idx] = self._create_tool_result(call_id, name, result)
+                    try:
+                        result = future.result()
+                        self.cache.set(cache_key, result)
+                        self.history.record(name, args, result)
+                        results[idx] = self._create_tool_result(call_id, name, result)
+                    except Exception as e:
+                        logger.error(f"[EXECUTOR] Tool '{name}' execution failed: {type(e).__name__}: {e}")
+                        # Create error result
+                        error_result = {"success": False, "error": f"{type(e).__name__}: {str(e)}"}
+                        self.history.record(name, args, error_result)
+                        results[idx] = self._create_tool_result(call_id, name, error_result)
             
             # Filter out None values (shouldn't happen, but safety check)
             return [r for r in results if r is not None]
                 
         except ImportError:
             return self.process_tool_calls(tool_calls, context)
+        except Exception as e:
+            logger.error(f"[EXECUTOR] Parallel execution failed: {type(e).__name__}: {e}")
+            # Fallback to sequential execution
+            return self.process_tool_calls(tool_calls, context)
     
     def _build_tool_context(self, context: Dict[str, Any]) -> ToolContext:
         """Build ToolContext from context dict"""
@@ -232,11 +246,10 @@ class ToolExecutor:
             return {}
     
     def _create_tool_result(self, call_id: str, name: str, result: Dict) -> Dict[str, Any]:
-        """Create tool result message"""
+        """Create tool result message (OpenAI format: only tool_call_id and content)"""
         return {
             "tool_call_id": call_id,
             "role": "tool",
-            "name": name,
             "content": json.dumps(result, ensure_ascii=False)
         }