feat(router): add context pruner and extend timeout for local models

strix_hub/model_router.py

@@ -176,10 +176,12 @@ def _create_router_handler(router_instance: ModelRouterServer) -> type[BaseHTTPR
                 target_key = cfg["subagent_api_key"]
                 role_label = "Subagent (执行打手)"
 
-            # 2. Rewrite model in payload
+            # 2. Rewrite model in payload and prune bloated historical context
             if isinstance(payload, dict):
                 payload["model"] = target_model
-                forward_body = json.dumps(payload).encode("utf-8")
+                if "messages" in payload and isinstance(payload["messages"], list):
+                    payload["messages"] = _prune_messages_for_local_model(payload["messages"])
+                forward_body = json.dumps(payload, ensure_ascii=False).encode("utf-8")
             else:
                 forward_body = raw_body
 
@@ -208,7 +210,7 @@ def _create_router_handler(router_instance: ModelRouterServer) -> type[BaseHTTPR
 
             try:
                 req = Request(upstream_url, data=forward_body, headers=headers, method="POST")
-                with urlopen(req, timeout=300) as response:
+                with urlopen(req, timeout=600) as response:
                     self.send_response(response.status)
                     for k, v in response.getheaders():
                         if k.lower() not in ["content-length", "transfer-encoding", "content-encoding"]:
@@ -315,3 +317,41 @@ def _parse_qwen_tool_calls(content: str) -> list[dict[str, Any]]:
             pass
 
     return tool_calls
+
+
+def _prune_messages_for_local_model(
+    messages: list[dict[str, Any]], max_tool_output_chars: int = 1500
+) -> list[dict[str, Any]]:
+    """Prune bloated older tool outputs to prevent context explosion on local models."""
+    if not isinstance(messages, list) or len(messages) <= 6:
+        return messages
+
+    pruned: list[dict[str, Any]] = []
+    # Keep the first 2 messages (system prompt + user task description) untouched
+    # Keep the latest 4 messages (recent turns) untouched
+    # For intermediate messages, trim excessively long tool outputs (nmap logs, HTML blobs)
+    cutoff = len(messages) - 4
+    for idx, msg in enumerate(messages):
+        if idx <= 1 or idx >= cutoff:
+            pruned.append(msg)
+            continue
+
+        if not isinstance(msg, dict):
+            pruned.append(msg)
+            continue
+
+        role = msg.get("role")
+        content = msg.get("content")
+
+        if role == "tool" and isinstance(content, str) and len(content) > max_tool_output_chars:
+            trimmed = (
+                content[:max_tool_output_chars]
+                + f"\n\n[... Strix Hub: Output trimmed to {max_tool_output_chars} chars to accelerate local inference ...]"
+            )
+            new_msg = dict(msg)
+            new_msg["content"] = trimmed
+            pruned.append(new_msg)
+        else:
+            pruned.append(msg)
+
+    return pruned