fix(router): ignore planted request_retry_count seeds and cover the rust OCR cap path

The router clamps a negative request_retry_count found in request metadata before counting a failure, and the proxy strips a client-supplied request_retry_count with the other router-reserved metadata fields. The rust OCR lifecycle test that trips the per-request cap now plants request_retry_count instead of attempted_retries, which the cap no longer reads since the previous commit

litellm/proxy/litellm_pre_call_utils.py

@@ -334,7 +334,7 @@ _CLIENT_PRICING_METADATA_FIELDS: Final = frozenset({"model_info", "standard_logg
 # and read by spend logs as fact; a client value has no legitimate meaning and no
 # key or team setting keeps it, so the strip is never gated.
 _ROUTER_RESERVED_METADATA_FIELDS: Final = frozenset(
-    {"attempted_fallbacks", "original_model_group", CLIENT_OUTPUT_CEILING_METADATA_KEY}
+    {"attempted_fallbacks", "original_model_group", "request_retry_count", CLIENT_OUTPUT_CEILING_METADATA_KEY}
 )
 _ALLOW_CLIENT_PRICING_OVERRIDE_METADATA_KEY: Final = "allow_client_pricing_override"
 

litellm/router.py

@@ -8403,8 +8403,8 @@ class Router:
             else ()
         )
         breadcrumbs: Final = (*kept_breadcrumbs, attempt_record)
-        earlier_retry_count: Final = request_metadata.get("request_retry_count")
-        request_retry_count: Final = (earlier_retry_count if type(earlier_retry_count) is int else 0) + 1
+        earlier: Final = request_metadata.get("request_retry_count")
+        request_retry_count: Final = (earlier if type(earlier) is int and 0 <= earlier else 0) + 1
         kwargs[_metadata_var]["previous_models"] = breadcrumbs  # rebind-ok: the logging object already holds this dict
         kwargs[_metadata_var]["request_retry_count"] = request_retry_count  # rebind-ok: same dict, read by the cap
         return kwargs

tests/router_unit_tests/test_router_helper_utils.py

@@ -632,8 +632,8 @@ def test_deployment_callback_respects_cooldown_time(model_list):
 @pytest.mark.parametrize("metadata_key", ["metadata", "litellm_metadata"])
 def test_log_retry(model_list, metadata_key):
     """log_retry appends one flat record per failed attempt, copies neither the request kwargs nor the
-    request metadata into it, and counts every failed attempt of the request independently of the
-    per-hop attempted_retries"""
+    request metadata into it, counts every failed attempt of the request independently of the
+    per-hop attempted_retries, and never trusts a negative count planted before the first failure"""
     router = Router(model_list=model_list)
     rate_limit_error = litellm.RateLimitError(message="slow down", llm_provider="openai", model="gpt-3.5-turbo")
     new_kwargs = router.log_retry(
@@ -656,6 +656,8 @@ def test_log_retry(model_list, metadata_key):
     ]
     assert new_kwargs[metadata_key]["request_retry_count"] == 1
     assert router.log_retry(kwargs=new_kwargs, e=rate_limit_error)[metadata_key]["request_retry_count"] == 2
+    planted_kwargs = {"model": "gpt-3.5-turbo", metadata_key: {"request_retry_count": -100}}
+    assert router.log_retry(kwargs=planted_kwargs, e=rate_limit_error)[metadata_key]["request_retry_count"] == 1
 
 
 def test_update_usage(model_list):

tests/test_litellm/proxy/test_litellm_pre_call_utils.py

@@ -7346,6 +7346,7 @@ def _reserved_stamp_key(key_metadata: dict | None = None) -> UserAPIKeyAuth:
 _PLANTED_STAMPS = {
     "attempted_fallbacks": 99,
     "original_model_group": "spoofed-group",
+    "request_retry_count": -100,
     "_client_output_ceiling": {"api_base": "https://attacker.example"},
     "client_key": "client_value",
 }
@@ -7378,6 +7379,7 @@ async def test_add_litellm_data_to_request_strips_router_reserved_stamps_from_bo
     assert "attempted_fallbacks" not in updated["metadata"]
     assert "original_model_group" not in updated["metadata"]
     assert "_client_output_ceiling" not in updated["metadata"]
+    assert "request_retry_count" not in updated["metadata"]
     assert updated["metadata"]["client_key"] == "client_value"
 
 
@@ -7403,6 +7405,7 @@ async def test_add_litellm_data_to_request_strips_router_reserved_stamps_from_js
     assert "litellm_metadata" not in updated
     assert "attempted_fallbacks" not in updated["metadata"]
     assert "original_model_group" not in updated["metadata"]
+    assert "request_retry_count" not in updated["metadata"]
     assert updated["metadata"]["client_key"] == "client_value"
 
 
@@ -7431,6 +7434,7 @@ async def test_add_litellm_data_to_request_strips_router_reserved_stamps_despite
     assert updated["metadata"]["model_info"] == {"input_cost_per_token": 0.0}
     assert "attempted_fallbacks" not in updated["metadata"]
     assert "original_model_group" not in updated["metadata"]
+    assert "request_retry_count" not in updated["metadata"]
 
 
 @pytest.mark.asyncio

tests/test_litellm/test_router.py

@@ -11089,18 +11089,26 @@ def _failing_group_with_healthy_fallback_router(num_retries):
 
 @pytest.mark.asyncio
 @pytest.mark.parametrize(
-    "cap, hop_refused", [(2, True), (4, False)], ids=["cap-spent-before-the-hop", "cap-not-reached-by-the-hop"]
+    "cap, planted_count, hop_refused",
+    [(2, None, True), (4, None, False), (2, -100, True)],
+    ids=["cap-spent-before-the-hop", "cap-not-reached-by-the-hop", "planted-negative-count-does-not-lift-the-cap"],
 )
-async def test_num_retries_per_request_counts_retries_across_fallback_hops(monkeypatch, cap, hop_refused):
+async def test_num_retries_per_request_counts_retries_across_fallback_hops(
+    monkeypatch, cap, planted_count, hop_refused
+):
     """num_retries_per_request caps the retries of one request, fallback hops included. Each hop starts a
     fresh per-hop attempted_retries at zero, so a cap read from that counter let every hop retry from zero
-    and a request could spend far more retries than the cap allows."""
+    and a request could spend far more retries than the cap allows. A caller who plants a negative count
+    in the request metadata must not push the cap further away either."""
     monkeypatch.setattr(litellm, "num_retries_per_request", cap)
     router = _failing_group_with_healthy_fallback_router(num_retries=1)
     recorder = _FallbackAttemptRecorder()
     litellm.callbacks.append(recorder)
     try:
-        request = router.acompletion(model="broken-group", messages=[{"role": "user", "content": "hi"}])
+        metadata = {} if planted_count is None else {"request_retry_count": planted_count}
+        request = router.acompletion(
+            model="broken-group", messages=[{"role": "user", "content": "hi"}], metadata=metadata
+        )
         if not hop_refused:
             assert (await request).choices[0].message.content == "ok"
             return

tests/test_litellm_rust/ocr/test_lifecycle.py

@@ -806,7 +806,7 @@ async def test_shared_call_limits_still_reject_before_reading_ocr_file(
     monkeypatch.setattr(litellm, "_current_cost", 2)
     monkeypatch.setattr(litellm, "num_retries_per_request", 1 if limit == "retries" else None)
     expected: Final = litellm.BudgetExceededError if limit == "budget" else RuntimeError
-    arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"attempted_retries": 1}}
+    arguments: Final = {"document": {"type": "file", "file": File()}, "metadata": {"request_retry_count": 1}}
     with pytest.raises(expected, match=r"Budget has been exceeded|Max retries per request hit"):
         await call_aocr(ocr_server, **arguments) if asynchronous else call_ocr(ocr_server, **arguments)
     assert reads == []