feat: context-aware routing + compaction signals

- Added GPU_CONTEXT map (MoE 131K, VLM 131K, Dense 65K)
- Heavy tier now prefers MoE/VLM (131K) over Dense (65K) for large requests
- Response headers: X-Context-Remaining, X-Context-Model
- Routing data includes context_remaining field
- Agents can use this to trigger compaction when nearing limits

router/router.py

@@ -24,6 +24,13 @@ GPU_MAX_CONCURRENT = {
     "qwen3.5-9b-vlm": 2,       # 2 slots (12GB VRAM, 4GB headroom)
 }
 
+# Context window sizes (tokens) — used for compaction signals
+GPU_CONTEXT = {
+    "qwen3.6-35B-A3B": 131072,
+    "qwen3.6-27B-code": 65536,
+    "qwen3.5-9b-vlm": 131072,
+}
+
 TIER_MODELS = {
     "starter": ["qwen3.5-9b-vlm"],
     "professional": ["qwen3.6-35B-A3B", "qwen3.6-27B-code", "qwen3.5-9b-vlm"],
@@ -183,7 +190,8 @@ def route(rd, tier):
     
     # TIER 3: Heavy reasoning — extremely large context or very long conversations
     if t > 50000 or turns > 25:
-        candidates = [m for m in ["qwen3.6-27B-code","qwen3.6-35B-A3B","qwen3.5-9b-vlm"] if m in avail]
+        # Prefer models with larger context windows (MoE/VLM at 131K, Dense at 65K)
+        candidates = [m for m in ["qwen3.6-35B-A3B","qwen3.5-9b-vlm","qwen3.6-27B-code"] if m in avail]
         result = select_best_gpu(candidates, "heavy_reasoning")
         if result: return result
     
@@ -299,15 +307,23 @@ def chat():
                 for raw in resp.iter_content(chunk_size=None, decode_unicode=True):
                     if raw: yield clean_unicode(raw)
             bcast()
-            return Response(stream_with_context(gen()), mimetype="text/event-stream")
+            ctx_remaining = GPU_CONTEXT.get(model, 65536) - estimate_tokens(rd.get("messages",[]))
+            r = Response(stream_with_context(gen()), mimetype="text/event-stream")
+            r.headers["X-Context-Remaining"] = str(max(0, ctx_remaining))
+            r.headers["X-Context-Model"] = model
+            return r
         data = clean_response(resp.json())
         for c in data.get("choices",[]):
             msg = c.get("message",{})
             if not msg.get("content") and msg.get("reasoning_content"):
                 msg["content"] = msg["reasoning_content"]
-        data["routing"] = {"model":model,"reason":reason,"gpu":url,"tier":tier,"agent":agent,"latency_ms":lat,"active_gpu":gpu_active_count(model)}
+        ctx_remaining = GPU_CONTEXT.get(model, 65536) - estimate_tokens(rd.get("messages",[]))
+        data["routing"] = {"model":model,"reason":reason,"gpu":url,"tier":tier,"agent":agent,"latency_ms":lat,"active_gpu":gpu_active_count(model),"context_remaining": max(0, ctx_remaining)}
+        resp = jsonify(data)
+        resp.headers["X-Context-Remaining"] = str(max(0, ctx_remaining))
+        resp.headers["X-Context-Model"] = model
         bcast()
-        return jsonify(data)
+        return resp
     except requests.Timeout:
         gpu_decr(model)
         log.error("TIMEOUT: %s -> %s", agent, model)