diff --git a/src/agentos/provider/failures.py b/src/agentos/provider/failures.py
index 0d447c288..4c050182a 100644
--- a/src/agentos/provider/failures.py
+++ b/src/agentos/provider/failures.py
@@ -87,10 +87,15 @@ def _is_context_overflow(text: str) -> bool:
             "input exceeds",
             "provider_request_budget_exhausted",
             "too many tokens",
+            # Gemini's canonical input-token-limit error, e.g. "The input token
+            # count (5911388) exceeds the maximum number of tokens allowed
+            # (1048576)." — does not contain "input exceeds" or "maximum
+            # context" since the token counts sit between the fixed phrases.
+            "input token count",
+            "exceeds the maximum number of tokens allowed",
         )
     )
 
-
 def _is_policy_refusal(text: str) -> bool:
     return any(
         marker in text
diff --git a/tests/test_provider_failures.py b/tests/test_provider_failures.py
index ca4cbf856..c15fb7a2e 100644
--- a/tests/test_provider_failures.py
+++ b/tests/test_provider_failures.py
@@ -13,3 +13,43 @@ def test_provider_request_budget_exhausted_is_context_overflow() -> None:
         )
         is ProviderFailureKind.CONTEXT_OVERFLOW
     )
+
+def test_gemini_input_token_count_message_is_context_overflow() -> None:
+    # Gemini's real, canonical context-overflow error. Numbers vary per
+    # request but the surrounding phrasing is fixed. Before the fix this
+    # matched none of the context-overflow markers and fell through to
+    # BAD_REQUEST, so the runtime never triggered COMPACT_AND_RETRY.
+    message = (
+        "The input token count (5911388) exceeds the maximum number of "
+        "tokens allowed (1048576)."
+    )
+
+    assert (
+        classify_provider_error(
+            provider_name="gemini",
+            status_code=400,
+            raw_code="400",
+            message=message,
+        )
+        is ProviderFailureKind.CONTEXT_OVERFLOW
+    )
+
+
+def test_gemini_input_token_count_message_is_context_overflow_regardless_of_token_counts() -> (
+    None
+):
+    # Same shape, different digit counts — guards against a marker that
+    # accidentally depends on a specific number of digits.
+    message = (
+        "The input token count (132478) exceeds the maximum number of "
+        "tokens allowed (131072)."
+    )
+
+    assert (
+        classify_provider_error(
+            provider_name="gemini",
+            status_code=400,
+            message=message,
+        )
+        is ProviderFailureKind.CONTEXT_OVERFLOW
+    )