diff --git a/bench_env/runner/base.py b/bench_env/runner/base.py
index 09fdc8fd..676ee5cd 100644
--- a/bench_env/runner/base.py
+++ b/bench_env/runner/base.py
@@ -161,7 +161,13 @@ def _evaluate_with_vlm(self, task, exec_result, episode) -> JudgeResult:
 
 def _action_fingerprint(action) -> str:
     """Extract action behavioral fingerprint (type + normalized data)."""
-    return f"{action.action_type}|{json.dumps(action.data, sort_keys=True, ensure_ascii=False)}"
+    data = action.data
+    if action.action_type == ActionType.WAIT:
+        # WAIT's "value" is a duration in seconds, not behavior: two waits are the
+        # same repeated action for loop-detection purposes regardless of how long
+        # each one slept, so drop it before hashing.
+        data = {k: v for k, v in data.items() if k != "value"}
+    return f"{action.action_type}|{json.dumps(data, sort_keys=True, ensure_ascii=False)}"
 
 
 def _snapshot_stopwatch(sw) -> tuple[float, dict[str, float], list[dict[str, Any]]]:
diff --git a/bench_env/tests/common/test_loop_detect.py b/bench_env/tests/common/test_loop_detect.py
new file mode 100644
index 00000000..4f4f2d58
--- /dev/null
+++ b/bench_env/tests/common/test_loop_detect.py
@@ -0,0 +1,88 @@
+from __future__ import annotations
+
+import pytest
+
+from bench_env.env.base import Action, ActionType, Observation, StepResult
+from bench_env.env.stopwatch import StopWatch
+from bench_env.runner.base import Controller
+
+
+def _make_obs(step_idx: int = 0) -> Observation:
+    return Observation(
+        screenshot_base64="",
+        route={"app": "demo", "path": "/"},
+        state={"apps": {}, "os": {}},
+        step_idx=step_idx,
+    )
+
+
+class _VaryingWaitAgent:
+    """Emits WAIT actions with a different duration on every call, the way a real
+    agent does when it picks the wait length itself."""
+
+    name = "varying-wait"
+
+    def __init__(self) -> None:
+        self.history: list = []
+        self._n = 0
+
+    def reset(self, task: str) -> None:
+        self.task = task
+
+    def act(self, obs: Observation) -> Action:
+        self._n += 1
+        return Action.wait(seconds=self._n * 0.1)
+
+
+class _WaitTrackingEnv:
+    def __init__(self) -> None:
+        self._agent_answer: str | None = None
+        self._agent_message: str | None = None
+        self.stopwatch = StopWatch()
+
+    async def get_state(self, *, required_apps: list[str] | None = None) -> dict:
+        return {}
+
+    async def step(self, action: Action) -> StepResult:
+        if action.action_type == ActionType.WAIT:
+            return StepResult(observation=_make_obs(1), done=False, info={})
+        raise AssertionError(f"unexpected action: {action.action_type}")
+
+    @property
+    def agent_answer(self) -> str | None:
+        return self._agent_answer
+
+    @property
+    def agent_message(self) -> str | None:
+        return self._agent_message
+
+
+class _TaskForController:
+    id = "demo.WaitTask"
+    description = "等待"
+    suite = "demo"
+    apps: list[str] = []
+
+    def teardown(self, env) -> None:
+        return None
+
+
+@pytest.mark.asyncio
+async def test_controller_run_detects_repetitive_wait_despite_varying_duration() -> None:
+    env = _WaitTrackingEnv()
+    agent = _VaryingWaitAgent()
+    task = _TaskForController()
+
+    exec_result, *_ = await Controller.run(
+        env,
+        agent,
+        task,
+        _make_obs(),
+        max_steps=5,
+        recorder=None,
+        loop_threshold=3,
+    )
+
+    assert exec_result.truncated is True
+    assert exec_result.stop_reason == "REPETITIVE_LOOP"
+    assert exec_result.steps == 3