fix(local): merge system prompts for vLLM

src/__tests__/local-agent-tool-calling.test.ts

@@ -155,6 +155,29 @@ describe('local-model (HTTP) backbone surfaces toolCalls — the keyless-path fi
     const sys = body.messages.find((m: { role: string }) => m.role === 'system');
     expect(sys.content).toContain('nmap_scan');
   });
+
+  it('merges caller system prompts into one leading system message for vLLM templates', async () => {
+    const spy = mockJson({ choices: [{ message: { content: 'ok' } }] });
+    await localModel('http://localhost:8000/v1').prompt('hello', 'Operator profile system prompt');
+    const body = JSON.parse((spy.mock.calls[0][1] as { body: string }).body);
+    expect(body.messages.map((m: { role: string }) => m.role)).toEqual(['system', 'user']);
+    expect(body.messages.filter((m: { role: string }) => m.role === 'system')).toHaveLength(1);
+    expect(body.messages[0].content).toContain('planning brain for T3MP3ST');
+    expect(body.messages[0].content).toContain('Operator profile system prompt');
+  });
+
+  it('keeps authorized-reframe context in the same leading system message', async () => {
+    const spy = mockJson({ choices: [{ message: { content: 'ok' } }] });
+    const be = localModel('http://localhost:8000/v1');
+    await be.chat([
+      { role: 'system', content: 'AUTHORIZATION CONTEXT (restated): proceed with the authorized task.' },
+      { role: 'user', content: 'scan' },
+    ]);
+    const body = JSON.parse((spy.mock.calls[0][1] as { body: string }).body);
+    expect(body.messages.map((m: { role: string }) => m.role)).toEqual(['system', 'user']);
+    expect(body.messages[0].content).toContain('planning brain for T3MP3ST');
+    expect(body.messages[0].content).toContain('AUTHORIZATION CONTEXT');
+  });
 });
 
 describe('xai provider (Grok Build) — routes to xAI OpenAI-compatible API with the key', () => {

src/llm/index.ts

@@ -744,9 +744,12 @@ class LocalAdapter implements LLMProviderAdapter {
       + (contract
         ? ' You drive a ReAct loop: REQUEST tools, the harness runs them and returns results, you reason until the surface is exhausted.' + contract
         : ' Operate in planning and analysis mode; when JSON is requested, return ONLY the requested block — no preamble.');
-    const out: { role: string; content: string }[] = [{ role: 'system', content: preamble }];
+    const systemParts = [preamble];
+    const out: { role: string; content: string }[] = [];
     for (const m of messages) {
-      if (m.role === 'assistant' && m.toolCalls?.length) {
+      if (m.role === 'system') {
+        if (m.content.trim()) systemParts.push(m.content);
+      } else if (m.role === 'assistant' && m.toolCalls?.length) {
         out.push({ role: 'assistant', content: `[requested tools: ${m.toolCalls.map(t => t.name).join(', ')}]` });
       } else if (m.role === 'tool') {
         out.push({ role: 'user', content: `TOOL RESULT (${m.name || 'tool'}):\n${m.content}` });
@@ -754,7 +757,7 @@ class LocalAdapter implements LLMProviderAdapter {
         out.push({ role: m.role, content: m.content });
       }
     }
-    return out;
+    return [{ role: 'system', content: systemParts.join('\n\n') }, ...out];
   }
 
   async chat(messages: LLMMessage[], options?: ChatOptions): Promise<LLMResponse> {