diff --git a/packages/client/README.md b/packages/client/README.md
index 1337c8fe..bd770abf 100644
--- a/packages/client/README.md
+++ b/packages/client/README.md
@@ -628,8 +628,8 @@ HTTP 422 when it will not serve one.
 - **Not the empty case:** an environment with zero skills is served an empty payload that
   commits normally.
 - **Recovery:** once the cause is fixed, call `start()` on the same store. Only `close()` is
-  final: a restart clears `failed`, resets the backoff, and held content stays readable
-  throughout. Restarting the process also works.
+  final: a restart clears `failed`, resets the backoff and `connection_failures`, and held
+  content stays readable throughout. Restarting the process also works.
 
 **Nothing above the store changes.** The accessors, verification, and `write_skills` see raw
 objects through the `SkillStore` interface and cannot tell which store produced them.
diff --git a/packages/client/agents.md b/packages/client/agents.md
index 8e1eb533..a57c3389 100644
--- a/packages/client/agents.md
+++ b/packages/client/agents.md
@@ -264,7 +264,8 @@ payload assigned to it, and chose a non-400 4xx because LD SDKs treat those as t
 `_classify_status` returns a `_FatalTransportError` and the normal give-up path runs:
 `failed` and `last_error` are set, `wait_for_skills` returns `False` at once, and
 `connection_failures` is left untouched (it counts consecutive *recoverable* failures; an
-existing count is kept, not zeroed).
+existing count is kept, not zeroed, until `start()` runs delivery again and `_rearm_waiters`
+resets it).
 
 **The 422 message matches the TypeScript SDK's word for word, and stays short.** It names the
 one cause a customer can fix — a view-scoped SDK key — and refers every other case to
diff --git a/packages/client/src/launchdarkly_ai_server/skills_fdv2.py b/packages/client/src/launchdarkly_ai_server/skills_fdv2.py
index 25f42918..fa19aef6 100644
--- a/packages/client/src/launchdarkly_ai_server/skills_fdv2.py
+++ b/packages/client/src/launchdarkly_ai_server/skills_fdv2.py
@@ -1503,12 +1503,14 @@ def start(self) -> FDv2SkillStore:
     def _rearm_waiters(self) -> None:
         """
         Resets per-run state for a store being started again after it gave up:
-        the ended-delivery flag, ``failed``, and the failure count. A payload
-        already held still answers ``wait_for_skills``. Call with the lock held.
+        the ended-delivery flag, ``failed``, and the failure count, including
+        the one ``diagnostics`` reports. A payload already held still answers
+        ``wait_for_skills``. Call with the lock held.
         """
         self._delivery_ended.clear()
         self._failed_reason = None
         self._failures = 0
+        self._reader.diagnostics.connection_failures = 0
         self._backoff_attempts = 0
         if not self._first_payload.is_set():
             self._released.clear()
@@ -1557,7 +1559,8 @@ def wait_for_skills(self, timeout: float = 10.0) -> bool:
         has skills; see ``diagnostics``.
 
         Returns ``False`` on timeout, or early if delivery ends first (``close``,
-        or a failure that will not be retried).
+        or a failure that will not be retried). A store that gave up waits again
+        once ``start()`` runs delivery again; only ``close`` is final.
         """
         self._released.wait(timeout=timeout)
         return self._first_payload.is_set()
diff --git a/packages/client/tests/test_skills_fdv2.py b/packages/client/tests/test_skills_fdv2.py
index 1864229b..0fbe85b8 100644
--- a/packages/client/tests/test_skills_fdv2.py
+++ b/packages/client/tests/test_skills_fdv2.py
@@ -2168,6 +2168,53 @@ def test_a_restart_resets_the_failure_count(self) -> None:
             # One failure in the new run, not three.
             assert store.diagnostics.connection_failures == 1
 
+    def test_a_restart_reports_no_failures_before_the_new_run_has_any(self) -> None:
+        """The reported count starts over in ``start()``, not at the next failure.
+
+        Read only after the new run fails, a stale count is overwritten and looks
+        reset. Read while the new run's first request is still in flight, it
+        would show the old run's failures on a store whose ``failed`` is None.
+        """
+
+        class _HoldsWhenScriptEnds(_ScriptedRequester):
+            def __init__(self, *outcomes: Any) -> None:
+                super().__init__(*outcomes)
+                self.release = threading.Event()
+
+            def poll(self, basis: str | None, etag: str | None) -> Any:
+                if not self.outcomes:
+                    self.calls.append((basis, etag))
+                    self.release.wait(timeout=10)
+                    raise _RecoverableTransportError("released")
+                return super().poll(basis, etag)
+
+        requester = _HoldsWhenScriptEnds(
+            _RecoverableTransportError("x"),
+            _RecoverableTransportError("x"),
+            _FatalTransportError("401"),
+        )
+        store = FDv2SkillStore(
+            SDK_KEY,
+            mode="poll",
+            initial_backoff=0.001,
+            max_backoff=0.002,
+            _requester=requester,
+        )
+        try:
+            store.start()
+            assert wait_until(lambda: store.failed is not None)
+            assert store.diagnostics.connection_failures == 2
+
+            calls = len(requester.calls)
+            store.start()
+            assert store.diagnostics.connection_failures == 0
+            assert wait_until(lambda: len(requester.calls) > calls)
+            assert store.failed is None
+            assert store.diagnostics.connection_failures == 0
+        finally:
+            requester.release.set()
+            store.close()
+
     def test_a_404_stops_delivery_immediately(self, endpoint: Any) -> None:
         """A 404 means the endpoint does not exist for this credential.