diff --git a/sentinel/enrichment/circuit_breaker.py b/sentinel/enrichment/circuit_breaker.py
index 0e6f628..9fae868 100644
--- a/sentinel/enrichment/circuit_breaker.py
+++ b/sentinel/enrichment/circuit_breaker.py
@@ -47,6 +47,9 @@ def __init__(
         self._state: State = "closed"
         self._opened_at: float | None = None
         self._lock = asyncio.Lock()
+        # True while a single half-open trial is in flight; gates concurrent
+        # callers so only one probes the unhealthy dependency (see `call`).
+        self._half_open_in_flight = False
 
     @property
     def name(self) -> str:
@@ -75,6 +78,7 @@ def _prune(self, now: float) -> None:
             self._failures.popleft()
 
     async def call(self, fn: Callable[[], Awaitable[T]]) -> T:
+        probing = False
         async with self._lock:
             now = self._time()
             if self._state == "open":
@@ -85,12 +89,29 @@ async def call(self, fn: Callable[[], Awaitable[T]]) -> T:
                     self._transition("half_open")
                 else:
                     raise CircuitOpenError(f"breaker {self._name} open")
+            if self._state == "half_open":
+                # Single-flight the trial: only the first post-cooldown caller
+                # probes the unhealthy dependency. Concurrent callers short-circuit
+                # until that probe closes or re-opens the breaker, instead of all
+                # hammering the dependency at once.
+                if self._half_open_in_flight:
+                    raise CircuitOpenError(f"breaker {self._name} half-open probe in flight")
+                self._half_open_in_flight = True
+                probing = True
         try:
             result = await fn()
         except asyncio.CancelledError:
+            # Release the probe slot without counting a failure. A bare assignment
+            # has no await point, so it is race-free even off-lock — important here
+            # because acquiring the lock during cancellation could itself be
+            # cancelled and leak the slot, wedging the breaker half-open forever.
+            if probing:
+                self._half_open_in_flight = False
             raise
         except BaseException:
             async with self._lock:
+                if probing:
+                    self._half_open_in_flight = False
                 now2 = self._time()
                 if self._state == "half_open":
                     self._failures.append(now2)
@@ -103,6 +124,8 @@ async def call(self, fn: Callable[[], Awaitable[T]]) -> T:
             raise
         else:
             async with self._lock:
+                if probing:
+                    self._half_open_in_flight = False
                 if self._state == "half_open":
                     self._transition("closed")
             return result
diff --git a/tests/unit/enrichment/test_circuit_breaker.py b/tests/unit/enrichment/test_circuit_breaker.py
index e0e6ac8..7e1f278 100644
--- a/tests/unit/enrichment/test_circuit_breaker.py
+++ b/tests/unit/enrichment/test_circuit_breaker.py
@@ -130,6 +130,117 @@ async def cancelled() -> None:
     assert cb.state == "closed"
 
 
+@pytest.mark.asyncio
+async def test_half_open_single_flights_concurrent_probes() -> None:
+    """In half_open, exactly ONE caller probes the unhealthy dependency; callers
+    arriving while the probe is in flight short-circuit with CircuitOpenError
+    instead of all hammering the dependency (issue #65)."""
+    clock = _Clock()
+    cb = CircuitBreaker("x", threshold=1, window_s=60.0, cooldown_s=30.0, time_fn=clock)
+
+    async def boom() -> None:
+        raise RuntimeError("boom")
+
+    with pytest.raises(RuntimeError):
+        await cb.call(boom)  # threshold=1 → opens immediately
+    assert cb.state == "open"
+    clock.advance(31.0)  # cooldown elapsed → next caller may probe
+
+    probe_in_flight = asyncio.Event()
+    release = asyncio.Event()
+    probe_calls = 0
+
+    async def slow_probe() -> str:
+        nonlocal probe_calls
+        probe_calls += 1
+        probe_in_flight.set()
+        await release.wait()
+        return "ok"
+
+    async def must_not_run() -> str:  # pragma: no cover - asserts it never runs
+        raise AssertionError("concurrent caller should have been short-circuited")
+
+    probe_task = asyncio.create_task(cb.call(slow_probe))
+    await asyncio.wait_for(probe_in_flight.wait(), timeout=1.0)
+
+    # While the single probe is in flight, concurrent callers must be rejected.
+    results = await asyncio.gather(
+        cb.call(must_not_run), cb.call(must_not_run), return_exceptions=True
+    )
+    assert all(isinstance(r, CircuitOpenError) for r in results), results
+
+    release.set()
+    assert await probe_task == "ok"
+    assert probe_calls == 1
+    assert cb.state == "closed"  # type: ignore[comparison-overlap]
+
+
+@pytest.mark.asyncio
+async def test_half_open_probe_reopens_then_allows_next_probe_after_cooldown() -> None:
+    """Guards the *failure-path* flag release: a failed half-open probe re-opens
+    the breaker AND clears the in-flight flag, so a fresh probe is admitted after
+    the next cooldown. (Drop the flag-clear from the failure branch and the final
+    call short-circuits with CircuitOpenError instead of probing.)"""
+    clock = _Clock()
+    cb = CircuitBreaker("x", threshold=1, window_s=60.0, cooldown_s=30.0, time_fn=clock)
+
+    async def boom() -> None:
+        raise RuntimeError("boom")
+
+    async def ok() -> int:
+        return 1
+
+    with pytest.raises(RuntimeError):
+        await cb.call(boom)
+    clock.advance(31.0)
+    # Single-flight probe fails → reopen (and in-flight flag must be released).
+    with pytest.raises(RuntimeError):
+        await cb.call(boom)
+    assert cb.state == "open"
+    # Next cooldown → a fresh probe is allowed (proves the flag was cleared).
+    clock.advance(31.0)
+    assert await cb.call(ok) == 1
+    assert cb.state == "closed"  # type: ignore[comparison-overlap]
+
+
+@pytest.mark.asyncio
+async def test_half_open_cancelled_probe_releases_slot() -> None:
+    """Guards the cancellation-path flag release: if the single half-open probe
+    is cancelled (e.g. the caller's task is torn down), the in-flight slot must be
+    freed so a later caller can probe — never wedged half-open forever (issue #65).
+    """
+    clock = _Clock()
+    cb = CircuitBreaker("x", threshold=1, window_s=60.0, cooldown_s=30.0, time_fn=clock)
+
+    async def boom() -> None:
+        raise RuntimeError("boom")
+
+    async def ok() -> int:
+        return 1
+
+    with pytest.raises(RuntimeError):
+        await cb.call(boom)
+    clock.advance(31.0)
+
+    probe_in_flight = asyncio.Event()
+
+    async def blocking_probe() -> None:
+        probe_in_flight.set()
+        await asyncio.Event().wait()  # block until cancelled
+
+    probe_task = asyncio.create_task(cb.call(blocking_probe))
+    await asyncio.wait_for(probe_in_flight.wait(), timeout=1.0)
+    probe_task.cancel()
+    with pytest.raises(asyncio.CancelledError):
+        await probe_task
+
+    # State is still half_open (cancellation isn't a failure); only the flag gates
+    # the next caller. If it leaked True this would raise CircuitOpenError.
+    assert cb.state == "half_open"
+    assert await cb.call(ok) == 1
+    assert cb.state == "closed"  # type: ignore[comparison-overlap]
+
+
 @pytest.mark.asyncio
 async def test_state_change_callback_fires_on_transitions() -> None:
     clock = _Clock()