feat: add /generate endpoint for structured one-shot LLM calls

Made-with: Cursor

src/server.ts

@@ -11,7 +11,7 @@ import { ToolContext } from './tools';
 import { PROTECTED_GITEA_REPOS } from './tools/security';
 import { orchestratorChat, listSessions, clearSession } from './orchestrator';
 import { atlasChat, listAtlasSessions, clearAtlasSession } from './atlas';
-import { LLMMessage } from './llm';
+import { LLMMessage, createLLM } from './llm';
 
 const app = express();
 app.use(cors());
@@ -343,6 +343,27 @@ app.post('/webhook/gitea', (req: Request, res: Response) => {
         });
 });
 
+// ---------------------------------------------------------------------------
+// Generate — thin structured-generation endpoint (no session, no system prompt)
+// Use this for one-shot tasks like architecture recommendations.
+// ---------------------------------------------------------------------------
+
+app.post('/generate', async (req: Request, res: Response) => {
+    const { prompt, model } = req.body as { prompt?: string; model?: string };
+    if (!prompt) { res.status(400).json({ error: '"prompt" is required' }); return; }
+
+    try {
+        const llm = createLLM(model ?? 'A', { temperature: 0.3 });
+        const messages: import('./llm').LLMMessage[] = [
+            { role: 'user', content: prompt }
+        ];
+        const response = await llm.chat(messages, [], 8192);
+        res.json({ reply: response.content ?? '' });
+    } catch (err) {
+        res.status(500).json({ error: err instanceof Error ? err.message : String(err) });
+    }
+});
+
 // ---------------------------------------------------------------------------
 // Error handler
 // ---------------------------------------------------------------------------