consolidate: HQ TandemCore — Gemma4 reasons, Qwen3.5 implements

hq/tandem_core.py

@@ -0,0 +1,43 @@
+"""
+Agent Q3 [HQ] — TandemCore
+Gemma4-E4B (Reasoner) → Qwen3.5-4B (Coder) chained pipeline.
+Step 1: Reasoner produces a structured plan.
+Step 2: Coder implements the plan.
+"""
+import httpx, os
+
+OLLAMA_URL = os.getenv("OLLAMA_URL", "http://localhost:11434/v1/chat/completions")
+REASONER   = os.getenv("REASONER_MODEL", "gemma4:e4b-instruct-q4_K_M")
+CODER      = os.getenv("CODER_MODEL",    "qwen3.5:4b-instruct-q4_K_M")
+
+REASON_SYSTEM = (
+    "You are a senior architect and researcher. Given the user request, "
+    "produce a concise, structured plan (steps, constraints, edge cases) "
+    "that a coder can implement directly. Be precise. No code yet."
+)
+CODE_SYSTEM = (
+    "You are an expert software engineer. Implement the plan provided exactly. "
+    "Output production-quality code with inline comments. No explanation outside code blocks."
+)
+
+class TandemCore:
+    async def _call(self, model: str, system: str, messages: list) -> str:
+        payload = {
+            "model": model,
+            "messages": [{"role": "system", "content": system}] + messages,
+        }
+        async with httpx.AsyncClient(timeout=120) as client:
+            r = await client.post(OLLAMA_URL, json=payload)
+            r.raise_for_status()
+            return r.json()["choices"][0]["message"]["content"]
+
+    async def run(self, messages: list) -> dict:
+        plan = await self._call(REASONER, REASON_SYSTEM, messages)
+        plan_msg = {"role": "assistant", "content": f"[PLAN]\n{plan}"}
+        impl = await self._call(CODER, CODE_SYSTEM, messages + [plan_msg])
+        return {
+            "tandem": True,
+            "plan":   plan,
+            "implementation": impl,
+            "choices": [{"message": {"role": "assistant", "content": impl}}]
+        }