Loop guard: also identical call with identical result 3 times in a row; baseline restarts (run base 20500)

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

harness/agents.py

@@ -156,8 +156,11 @@ class LlmAgent:
                     err, text = proxy.call(fn["name"], args)
                     messages.append({"role": "tool", "tool_call_id": c.get("id", ""),
                                      "content": ("ERROR: " if err else "") + text[:12000]})
-                    if self.loop_guard and proxy.same_push_streak >= self.loop_guard:
-                        final = f"Stopped: loop (the same source was pushed {proxy.same_push_streak} times in a row)."
+                    streak = max(proxy.same_push_streak, proxy.same_call_streak)
+                    if self.loop_guard and streak >= self.loop_guard:
+                        final = (f"Stopped: loop (the same source was pushed {proxy.same_push_streak} times in a row)."
+                                 if proxy.same_push_streak >= self.loop_guard else
+                                 f"Stopped: loop (the same call with the same result {proxy.same_call_streak} times in a row).")
                         self.end_reason = "loop"
                         break
                 if self.end_reason == "loop":

harness/proxy.py

@@ -71,6 +71,8 @@ class ToolProxy:
         self.activation_errors = []  # unique error messages of failed writes
         self.same_push_streak = 0  # identical sap_push_source (object + source hash) in a row
         self._last_push = None
+        self.same_call_streak = 0  # identical call (tool, arguments and result) in a row, any tool
+        self._last_call = None
         self.log = open(log_path, "a")
         self.adt = None
         self.fallbacks = []  # writes that needed the ADT activation fallback
@@ -171,6 +173,11 @@ class ToolProxy:
                         for m in activation_messages(text, err):
                             if m not in self.activation_errors and len(self.activation_errors) < 30:
                                 self.activation_errors.append(m)
+        if tool in MODEL_TOOLS and not result[1].startswith("Tool "):
+            ck = (tool, hashlib.md5(json.dumps(args, sort_keys=True).encode()).hexdigest(),
+                  hashlib.md5(result[1].encode()).hexdigest())
+            self.same_call_streak = self.same_call_streak + 1 if ck == self._last_call else 1
+            self._last_call = ck
         entry["is_error"], entry["result"] = result[0], result[1][:20000]
         self.log.write(json.dumps(entry) + "\n")
         self.log.flush()

train/README.md

@@ -57,7 +57,7 @@ G0139, G0151, G0157, G0174, G0167, G0185. Use the same list before and after tra
 | temperature / top_p / top_k / min_p | 0.2 / 0.95 / 20 / 0 |
 | max_tokens per turn | **16384**, sent in each request by `train/baseline.py` (`MAX_TOKENS`); the server limit stays 32768 |
 | Tool-call budget per task | 60 calls, 15 activations (T01 too) |
-| Loop guard | `loop_guard` 3: the run ends when `sap_push_source` pushes the same source (object + md5) 3 times in a row; final report "Stopped: loop ...", `end_reason` "loop". Scores use the final state, so they do not change. Same after training. T01 of the first run (started before the guard) ran without it |
+| Loop guard | `loop_guard` 3: the run ends when `sap_push_source` pushes the same source (object + md5) 3 times in a row, or when any tool is called 3 times in a row with the same arguments and the same result (read loop, e.g. G0105 pulled the same include 18 times); final report "Stopped: loop ...", `end_reason` "loop". Scores use the final state, so they do not change. Same after training. T01 of the first run (started before the guard) ran without it |
 | Per run record | `end_reason` (report, loop, empty_response, tool_budget, time_budget, model_error, max_turns), `activation_failures`, `activation_error_messages` (unique), also in `baseline.json` |
 | Empty turn | retried (2 times), then the run stops ("Stopped: empty model response") |
 | Docker (A4H) | VM memory 36 GB (`MemoryMiB` 36864), container `--memory 32g --memory-swap 32g` (2026-10-04) |

train/baseline_chain.sh

@@ -4,7 +4,7 @@
 cd "$(dirname "$0")/.."
 log() { echo "$(date '+%F %T') $*"; }
 log "start baseline, 11 tasks, thinking off, max_tokens 16384, loop guard 3"
-python3 train/baseline.py --label baseline --run-base 20400 > runs/stage1/baseline.log 2>&1 &
+python3 train/baseline.py --label baseline --run-base 20500 > runs/stage1/baseline.log 2>&1 &
 BP=$!
 log "baseline PID $BP"
 NOTE=0