diff --git a/python/freetoken/engine/cache_budget.py b/python/freetoken/engine/cache_budget.py
index ab7c0a9f1..8f85cf44d 100644
--- a/python/freetoken/engine/cache_budget.py
+++ b/python/freetoken/engine/cache_budget.py
@@ -118,7 +118,9 @@ def resolve_moe_cache_auto(
     """
     budget_bytes = net_cache_budget_bytes(memory_ratio, baseline_free, weights_bytes, fixed_cache_size)
     max_slots = 992 if quant_format == "nvfp4_marlin" else total_experts
-    kv_reserve_pages = div_ceil(kv_reserve_tokens, page_size)
+    # Every pool keeps page 0 as an unreachable dummy/sentinel. The CLI floor is expressed in
+    # usable tokens, so reserve that internal page in addition to the user-visible capacity.
+    kv_reserve_pages = div_ceil(kv_reserve_tokens, page_size) + 1
     return plan_cache_budget(
         budget_bytes=budget_bytes,
         per_expert_bytes=per_expert_bytes,
diff --git a/python/freetoken/server/args.py b/python/freetoken/server/args.py
index 5b4db587d..ab0e0cd63 100644
--- a/python/freetoken/server/args.py
+++ b/python/freetoken/server/args.py
@@ -537,7 +537,10 @@ def _infer_reasoning_parser(model_path: str) -> str | None:
         "--kv-reserve-tokens",
         type=int,
         default=ServerArgs.kv_reserve_tokens,
-        help="KV-cache token floor reserved before --moe-cache-auto fills experts.",
+        help=(
+            "Usable KV-cache token floor reserved before --moe-cache-auto fills experts "
+            "(the internal dummy page is additional)."
+        ),
     )
 
     parser.add_argument(
diff --git a/tests/engine/test_cache_budget.py b/tests/engine/test_cache_budget.py
index 9ac2a4f4c..40e1055b2 100644
--- a/tests/engine/test_cache_budget.py
+++ b/tests/engine/test_cache_budget.py
@@ -112,6 +112,26 @@ def test_resolve_auto_applies_ratio_once_and_marlin_cap():
     assert size == 8 and pages == 40 and overlap is True
 
 
+def test_resolve_auto_reserves_usable_tokens_beyond_the_dummy_page():
+    size, pages, overlap = resolve_moe_cache_auto(
+        baseline_free=940,
+        weights_bytes=0,
+        memory_ratio=1.0,
+        cache_per_page=10,
+        fixed_cache_size=0,
+        per_expert_bytes=100,
+        num_experts=2,
+        total_experts=50,
+        prefill_overlap=False,
+        kv_reserve_tokens=256,
+        page_size=64,
+        quant_format="bf16",
+    )
+    assert overlap is False
+    assert (pages - 1) * 64 >= 256
+    assert size == 8 and pages == 14
+
+
 def test_resolve_auto_marlin_caps_slots():
     size, _, _ = resolve_moe_cache_auto(
         baseline_free=10_000_000, weights_bytes=0, memory_ratio=1.0,