fix: enable parallel tool calls and prompt LLM to batch them

- Add parallel_tool_calls: true to both agent and setup LLM requests
- Add system prompt instruction to batch multiple tool calls in one
  response (e.g. read 3 files at once instead of one per round-trip)
- Reduces total LLM API calls and the triangular token cost growth

Co-Authored-By: Claude Opus 4.6 (1M context) <noreply@anthropic.com>

Author: hkirat

apps/server/src/services/agent-script.ts

@@ -128,7 +128,7 @@ async function callLLM(msgs) {
         Authorization: \`Bearer \${llmApiKey}\`,
         "X-Title": "Vendi",
       },
-      body: JSON.stringify({ model: llmModel, messages: msgs, tools, max_tokens: 4096 }),
+      body: JSON.stringify({ model: llmModel, messages: msgs, tools, max_tokens: 4096, parallel_tool_calls: true }),
       signal: controller.signal,
     });
     if (!res.ok) throw new Error(\`LLM error \${res.status}: \${await res.text()}\`);

apps/server/src/services/session.service.ts

@@ -144,6 +144,7 @@ function buildSystemPrompt(project: {
     `- Startup commands: ${project.startupCommands.length > 0 ? project.startupCommands.join(" && ") : "check package.json"}`,
     "",
     "RULES:",
+    "- IMPORTANT: Batch multiple tool calls in a single response whenever possible. For example, if you need to read 3 files, call read_file 3 times in one response instead of reading them one at a time. This saves time and cost.",
     "- Only modify files matching these patterns: " +
       (project.allowedFilePatterns.length > 0
         ? project.allowedFilePatterns.join(", ")

apps/server/src/services/setup.service.ts

@@ -103,6 +103,7 @@ OUTPUT FORMAT — when you have everything:
 [/SETUP_COMPLETE]
 
 RULES:
+- IMPORTANT: Batch multiple tool calls in a single response whenever possible. For example, read package.json, .env.example, and docker-compose.yml all in one response instead of one at a time. This saves time and cost.
 - ALWAYS read the codebase FIRST before saying anything to the user
 - Do NOT ask the user about services, ports, startup commands, or file patterns — detect them yourself
 - Do NOT ask about or suggest code changes — this is setup, not development
@@ -160,7 +161,7 @@ async function callLLM(messages: any[]): Promise<any> {
       Authorization: `Bearer ${key}`,
       "X-Title": "Vendi",
     },
-    body: JSON.stringify({ model, messages, tools, max_tokens: 4096 }),
+    body: JSON.stringify({ model, messages, tools, max_tokens: 4096, parallel_tool_calls: true }),
   });
   if (!res.ok) throw new Error(`LLM error ${res.status}: ${await res.text()}`);
   const json = await res.json();