Agent: max_tokens 32k for cloud models and retry on empty turns (runaway reasoning); proxy: sap_activate fallback for PROG/FUNC

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014aUaQeLnwbb1zTpN7kHeat

harness/agents.py

@@ -79,10 +79,16 @@ class LlmAgent:
         self.temperature = temperature
         self.max_seconds = max_seconds
         self.request_timeout = 3600  # local models are slow; a hung request still ends
+        # Runaway reasoning: 35 of 1965 DeepSeek turns produced 393k output tokens and no content (46 % of the
+        # run cost, 2026-10-03). Normal turns: p95 17.5k, max 82k. A cut turn is retried (see run).
+        self.max_tokens = 32000 if ":cloud" in (model or "") else None
+        self.empty_retries = 2
 
     def _chat(self, messages, tools):
         body = {"model": self.model, "messages": messages, "tools": tools,
                 "temperature": self.temperature, "parallel_tool_calls": False}
+        if self.max_tokens:
+            body["max_tokens"] = self.max_tokens
         req = urllib.request.Request(f"{self.base_url}/chat/completions", json.dumps(body).encode(),
                                      {"Content-Type": "application/json",
                                       "Authorization": f"Bearer {self.api_key}"})
@@ -110,6 +116,7 @@ class LlmAgent:
         if ":cloud" in self.model:
             check_budget()
         final = ""
+        empty = 0
         start = time.time()
         for _ in range(self.max_turns):
             if self.max_seconds and time.time() - start > self.max_seconds:
@@ -125,8 +132,12 @@ class LlmAgent:
                                      "tool_calls": msg.get("tool_calls"), "usage": usage})
             messages.append({k: v for k, v in msg.items() if k in ("role", "content", "tool_calls")})
             calls = msg.get("tool_calls") or []
+            if not calls and not (msg.get("content") or "").strip() and empty < self.empty_retries:
+                empty += 1  # empty turn (output limit reached in reasoning): ask the same question again
+                messages.pop()
+                continue
             if not calls:
-                final = msg.get("content") or ""
+                final = msg.get("content") or ("Stopped: empty model response." if empty else "")
                 break
             try:
                 for c in calls:

harness/proxy.py

@@ -52,6 +52,7 @@ class ToolProxy:
         self.log = open(log_path, "a")
         self.adt = None
         self.fallbacks = []  # writes that needed the ADT activation fallback
+        self.last_source = {}  # last main source pushed per PROG/FUNC, for the sap_activate fallback
 
     def schemas(self):
         tools = [t for t in self.mcp.list_tools() if t["name"] in MODEL_TOOLS]
@@ -126,7 +127,13 @@ class ToolProxy:
                     self.activations += 1
                 err, text = self.mcp.call(tool, args)
                 if tool == "sap_push_source":
+                    if str(args.get("objectType", "")).upper() in ("PROG", "FUNC") and not args.get("includeType"):
+                        self.last_source[str(args.get("objectName", "")).upper()] = args.get("source", "")
                     err, text = self._activation_fallback(args, err, text)
+                elif tool == "sap_activate":  # a write with activate=false, then sap_activate (G0014, G0124)
+                    src = self.last_source.get(str(args.get("objectName", "")).upper())
+                    if src is not None:
+                        err, text = self._activation_fallback(dict(args, source=src), err, text)
                 result = (err, self._filter(tool, text))
                 if tool in WRITE_TOOLS:
                     failed = err or '"success":false' in text.replace(" ", "")