feat: 0.6.5 — ollama timeout 600s & raise default max_tokens

- ollama timeout: 120s → 600s (local models are slow)
- ollama default num_predict: 2048 → 8192
- compose generated YAML: ollama gets max_tokens 8192, timeout 600s
- cloud API providers keep existing 4096/120s defaults

Author: jnMetaCode

package-lock.json

@@ -1,6 +1,6 @@
 {
   "name": "agency-orchestrator",
-  "version": "0.6.4",
+  "version": "0.6.5",
   "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",

package.json

@@ -82,6 +82,9 @@ function buildComposeSystemPromptEn(catalog: string, options?: { autoRun?: boole
   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
     ? `
@@ -129,8 +132,8 @@ agents_dir: "agency-agents"
 llm:
   provider: ${provider}
   ${model ? `model: ${model}` : ''}
-  max_tokens: 4096
-  timeout: 120000
+  max_tokens: ${maxTokens}
+  timeout: ${timeoutMs}
   retry: 2
 
 concurrency: 2
@@ -176,6 +179,9 @@ function buildComposeSystemPromptZh(catalog: string, options?: { autoRun?: boole
   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
     ? `
@@ -223,8 +229,8 @@ 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

src/cli/compose.ts

@@ -25,7 +25,7 @@ export class OllamaConnector implements LLMConnector {
           ],
           stream: false,
           options: {
-            num_predict: config.max_tokens || 2048,
+            num_predict: config.max_tokens || 8192,
           },
         }),
       });

src/connectors/ollama.ts

@@ -50,7 +50,8 @@ export async function executeDAG(dag: DAG, options: ExecutorOptions): Promise<Wo
   const stepResults: StepResult[] = [];
 
   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