feat: Agnes上下文512K→1M + Ollama支持配置num_ctx

Agnes AI:
- adapter注释和API文档更新: 上下文512K→1M (官方已升级)

Ollama num_ctx 支持:
- MetonaGenerationParams 新增 contextLength 字段
- OllamaAdapter.toNativeRequest 发送 options.num_ctx
- AgentLoopEngine 从 config 读取并传递到请求参数
- database.service 新增 ollama.numCtx 默认配置(null=由模型决定)
- SettingsModal LLM设置: Ollama 选中时显示上下文长度输入框
- main.ts 启动时从 configService 读取 ollama.numCtx

数据流:
Settings → configService → main.ts → AgentLoopEngine.config
→ MetonaRequest.params → OllamaAdapter → options.num_ctx

apis/agnes-ai-api-docs-20260625.html

@@ -495,7 +495,7 @@
         <p style="margin-top: 12px;">在 Claw-Eval 基准测试中排名 General Leaderboard <strong style="color:var(--accent)">第 9</strong>，Pass³ 分数 <strong style="color:var(--green)">60.9%</strong>。</p>
         <div class="hero-meta">
             <span><span class="dot dot-blue"></span> 模型名称: <strong>agnes-2.0-flash</strong></span>
-            <span><span class="dot dot-green"></span> 上下文: <strong>512K</strong> | 最大输出: <strong>65.5K</strong></span>
+            <span><span class="dot dot-green"></span> 上下文: <strong>1M</strong> | 最大输出: <strong>65.5K</strong></span>
             <span><span class="dot dot-orange"></span> Base URL: <strong>https://apihub.agnes-ai.com/v1</strong></span>
         </div>
     </div>
@@ -1004,7 +1004,7 @@ response = client.chat.completions.<span class="hl-fn">create</span>(
             <table class="params">
                 <thead><tr><th>项目</th><th>数值</th></tr></thead>
                 <tbody>
-                    <tr><td class="param-name">Context</td><td class="param-type">512K</td></tr>
+                    <tr><td class="param-name">Context</td><td class="param-type">1M</td></tr>
                     <tr><td class="param-name">Max Output</td><td class="param-type">65.5K</td></tr>
                 </tbody>
             </table>

electron/harness/adapters/agnes-ai.adapter.ts

@@ -100,7 +100,7 @@ export class AgnesAdapter extends BaseAdapter {
    * Agnes AI 特有参数：
    * - 多模态图片：user 消息的 images[] → OpenAI content 数组 [{type:"text"}, {type:"image_url"}]
    * - chat_template_kwargs: { enable_thinking: true } — 启用思考模式（非 thinking 字段）
-   * - 默认 max_tokens: 65536（512K 上下文）
+   * - 默认 max_tokens: 65536（1M 上下文，65.5K 最大输出）
    */
   private toNativeRequest(request: MetonaRequest, stream: boolean): Record<string, unknown> {
     const messages = buildOpenAICompatibleMessages(request);

electron/harness/adapters/ollama.adapter.ts

@@ -388,6 +388,7 @@ export class OllamaAdapter extends BaseAdapter {
         num_predict: request.params.maxTokens,
         top_p: request.params.topP,
         stop: request.params.stopSequences,
+        num_ctx: request.params.contextLength,
       },
     };
 

electron/harness/agent-loop/engine.ts

@@ -147,6 +147,7 @@ export class AgentLoopEngine extends EventEmitter {
             stream: true,
             thinkingEnabled: true,
             thinkingEffort: 'high',
+            contextLength: this.config.contextLength,
           },
         };
 

electron/harness/agent-loop/types.ts

@@ -63,6 +63,8 @@ export interface AgentLoopConfig {
   contextWindow: number;
   retryCount: number;
   temperature: number;
+  /** Ollama 上下文窗口大小（num_ctx），其他 Provider 忽略 */
+  contextLength?: number;
 }
 
 // ===== Agent Loop 输出 =====

electron/harness/types/metona-request.ts

@@ -53,6 +53,8 @@ export interface MetonaGenerationParams {
   thinkingEnabled?: boolean;
   /** 思考强度（替代 thinkingBudget，各 Provider 映射见 Adapter 规范） */
   thinkingEffort?: 'low' | 'medium' | 'high' | 'max';
+  /** 上下文窗口大小（仅 Ollama 支持 num_ctx） */
+  contextLength?: number;
 }
 
 // ===== 安全约束 =====

electron/main.ts

@@ -147,7 +147,11 @@ async function initialize(): Promise<void> {
   ];
 
   // ===== Agent Loop =====
-  const agentLoop = new AgentLoopEngine({}, adapter, toolRegistry, preToolHooks, postToolHooks);
+  const ollamaNumCtx = configService.get<number>('ollama.numCtx');
+  const agentLoop = new AgentLoopEngine(
+    { contextLength: ollamaNumCtx ?? undefined },
+    adapter, toolRegistry, preToolHooks, postToolHooks,
+  );
   agentLoop.setTools(toolRegistry.listTools());
   agentLoop.setWorkspacePath(workspaceInfo.path);
 

electron/services/database.service.ts

@@ -270,6 +270,9 @@ export class DatabaseService {
       { key: 'logging.auditEnabled', value: 'true', category: 'logging' },
       { key: 'logging.traceEnabled', value: 'true', category: 'logging' },
 
+      // Ollama 配置
+      { key: 'ollama.numCtx', value: 'null', category: 'ollama' },
+
       // Onboarding
       { key: 'onboarding.completed', value: 'false', category: 'general' },
     ];

src/components/settings/SettingsModal.tsx

@@ -101,6 +101,7 @@ function LLMSettings() {
   const [model, setModel] = useConfig('llm.model', '');
   const [apiKey, setApiKey] = useConfig('llm.apiKey', '');
   const [baseURL, setBaseURL] = useConfig('llm.baseURL', '');
+  const [numCtx, setNumCtx] = useConfig('ollama.numCtx', null as number | null);
   const [showKey, setShowKey] = useState(false);
 
   return (
@@ -118,6 +119,20 @@ function LLMSettings() {
       <TextField size="small" label="API Key" type={showKey ? 'text' : 'password'} value={apiKey} onChange={(e) => setApiKey(e.target.value)} placeholder="sk-...（本地模型可留空）"
         slotProps={{ input: { endAdornment: <IconButton size="small" onClick={() => setShowKey(!showKey)}>{showKey ? <EyeOff size={14} /> : <Eye size={14} />}</IconButton> } }}
       />
+      {provider === 'ollama' && (
+        <TextField
+          size="small"
+          label="上下文长度 (num_ctx)"
+          type="number"
+          value={numCtx ?? ''}
+          onChange={(e) => {
+            const v = e.target.value;
+            setNumCtx(v === '' ? null : Number(v));
+          }}
+          placeholder="默认由模型决定（如 2048、4096、128000）"
+          slotProps={{ htmlInput: { min: 512, step: 512 } }}
+        />
+      )}
     </Stack>
   );
 }