agency-orchestrator 0.6.4 → 0.6.6

package/README.en.md

@@ -211,7 +211,24 @@ analyze ──→ tech_review  ──→ summary
 | Claude API | `provider: "claude"` | `ANTHROPIC_API_KEY` |
 | OpenAI | `provider: "openai"` | `OPENAI_API_KEY` |
 
-All API providers support custom `base_url` and `api_key`, compatible with any OpenAI-compatible API.
+**Custom API (any OpenAI-compatible endpoint):**
+
+```bash
+ao init --provider openai --model model-name \
+  --base-url https://your-api-endpoint/v1 \
+  --api-key your-key
+```
+
+Or edit `.env` manually:
+
+```env
+AO_PROVIDER=openai
+AO_MODEL=model-name
+OPENAI_BASE_URL=https://your-api-endpoint/v1
+OPENAI_API_KEY=your-key
+```
+
+> ⚠️ Use `provider: "openai"` for third-party APIs, not `provider: "ollama"`. Ollama is for local models only and does not send API keys.
 
 ## CLI Reference
 

package/README.md

@@ -210,7 +210,33 @@ analyze ──→ tech_review  ──→ summary
 | Claude API | `provider: "claude"` | `ANTHROPIC_API_KEY` |
 | OpenAI | `provider: "openai"` | `OPENAI_API_KEY` |
 
-所有 API 提供商支持自定义 `base_url` 和 `api_key`，兼容智谱、月之暗面等 OpenAI 兼容 API。
+**自定义 API（火山引擎、智谱、月之暗面、硅基流动等 OpenAI 兼容 API）：**
+
+```bash
+ao init --provider openai --model 模型名 \
+  --base-url https://你的API地址/v1 \
+  --api-key 你的key
+```
+
+或手动编辑 `.env`：
+
+```env
+AO_PROVIDER=openai
+AO_MODEL=模型名
+OPENAI_BASE_URL=https://你的API地址/v1
+OPENAI_API_KEY=你的key
+```
+
+常见示例：
+
+| 平台 | base_url | model |
+|------|----------|-------|
+| 火山引擎 | `https://ark.cn-beijing.volces.com/api/coding/v3` | `ark-code-latest` |
+| 智谱 AI | `https://open.bigmodel.cn/api/paas/v4` | `glm-4` |
+| 硅基流动 | `https://api.siliconflow.cn/v1` | `deepseek-ai/DeepSeek-V3` |
+| 月之暗面 | `https://api.moonshot.cn/v1` | `moonshot-v1-8k` |
+
+> ⚠️ 注意：这些平台请使用 `provider: "openai"`，不要用 `provider: "ollama"`。Ollama 仅用于本地模型，不发送 API Key。
 
 ## CLI 命令
 

package/dist/cli/compose.js

@@ -66,6 +66,9 @@ function buildComposeSystemPromptEn(catalog, options) {
     const autoRun = options?.autoRun ?? false;
     const provider = options?.provider || 'deepseek';
     const model = options?.model;
+    const isLocal = provider === 'ollama';
+    const maxTokens = isLocal ? 8192 : 4096;
+    const timeoutMs = isLocal ? 600000 : 120000;
     const inputsSection = autoRun
         ? `
 ## Important: Direct Run Mode
@@ -109,8 +112,8 @@ agents_dir: "agency-agents"
 llm:
   provider: ${provider}
   ${model ? `model: ${model}` : ''}
-  max_tokens: 4096
-  timeout: 120000
+  max_tokens: ${maxTokens}
+  timeout: ${timeoutMs}
   retry: 2
 
 concurrency: 2
@@ -155,6 +158,9 @@ function buildComposeSystemPromptZh(catalog, options) {
     const autoRun = options?.autoRun ?? false;
     const provider = options?.provider || 'deepseek';
     const model = options?.model;
+    const isLocal = provider === 'ollama';
+    const maxTokens = isLocal ? 8192 : 4096;
+    const timeoutMs = isLocal ? 600000 : 120000;
     const inputsSection = autoRun
         ? `
 ## 重要：直接运行模式
@@ -198,8 +204,8 @@ agents_dir: "agency-agents-zh"
 llm:
   provider: ${provider}
   ${model ? `model: ${model}` : ''}
-  max_tokens: 4096
-  timeout: 120000
+  max_tokens: ${maxTokens}
+  timeout: ${timeoutMs}
   retry: 2
 
 concurrency: 2

package/dist/connectors/ollama.js

@@ -5,6 +5,8 @@ export class OllamaConnector {
         this.baseUrl = raw.replace(/\/+$/, '');
     }
     async chat(systemPrompt, userMessage, config) {
+        const numPredict = config.max_tokens || 8192;
+        const numCtx = estimateNumCtx(systemPrompt, userMessage, numPredict);
         let response;
         try {
             response = await fetch(`${this.baseUrl}/api/chat`, {
@@ -18,7 +20,8 @@ export class OllamaConnector {
                     ],
                     stream: false,
                     options: {
-                        num_predict: config.max_tokens || 2048,
+                        num_predict: numPredict,
+                        num_ctx: numCtx,
                     },
                 }),
             });
@@ -47,3 +50,20 @@ export class OllamaConnector {
         };
     }
 }
+/**
+ * 根据实际输入长度自适应计算 num_ctx，避免 Ollama 默认 2048 截断，
+ * 同时小输入不分配多余显存，提升推理速度。
+ */
+function estimateNumCtx(systemPrompt, userMessage, numPredict) {
+    const text = systemPrompt + userMessage;
+    // CJK 字符约 1.5-2 token/字，ASCII 约 0.25 token/字，按实际比例加权
+    let tokens = 0;
+    for (let i = 0; i < text.length; i++) {
+        tokens += text.charCodeAt(i) > 0x7F ? 1.5 : 0.3;
+    }
+    const estimatedInputTokens = Math.ceil(tokens);
+    const buffer = 512;
+    const computed = estimatedInputTokens + numPredict + buffer;
+    // 下限 4096（部分模型低于此值行为异常），上限 131072（主流模型极限）
+    return Math.min(Math.max(computed, 4096), 131072);
+}

package/dist/core/executor.js

@@ -10,7 +10,8 @@ export async function executeDAG(dag, options) {
     const startTime = Date.now();
     const stepResults = [];
     const isCLI = llmConfig.provider.endsWith('-cli') || llmConfig.provider === 'claude-code';
-    const timeout = llmConfig.timeout || (isCLI ? 600_000 : 120_000); // CLI 10分钟（gateway/MiniMax 等可能单步 5+ 分钟），API 2分钟
+    const isLocal = llmConfig.provider === 'ollama';
+    const timeout = llmConfig.timeout || (isCLI ? 600_000 : isLocal ? 600_000 : 120_000);
     const maxRetry = llmConfig.retry ?? 5;
     // CLI provider 强制串行：共享同一账户额度，并发会触发限速反而更慢
     const effectiveConcurrency = isCLI ? 1 : concurrency;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "agency-orchestrator",
-  "version": "0.6.4",
+  "version": "0.6.6",
   "description": "Multi-agent YAML workflow engine — 211 AI roles, auto DAG parallelism, zero code. One sentence → multiple AI roles collaborate → complete plan in minutes. 10 LLM providers, 7 need no API key.",
   "keywords": [
     "multi-agent",