fix(router): clear basedpyright operator error and the import-cycle alert

- narrow the accumulated stream delta content to str before appending, so the
  shared helper no longer trips reportOperatorIssue (str += object)
- keep the continuation marker in the pre-call-check module and lazy-import it
  at the two router call sites instead of routing it through constants, so the
  router has no module-level dependency that closes an import cycle

litellm/constants.py

@@ -2039,7 +2039,3 @@ BATCH_ENQUEUED_TOKEN_LIMIT_METADATA_KEY: Final = "batch_enqueued_token_limit"
 # Shared read-only empty mapping, for defaulting optional Mapping parameters without
 # constructing a fresh mutable dict at each call site.
 EMPTY_MAPPING: Final = MappingProxyType({})
-
-# Marks a fallback re-entry as a mid-stream continuation, read by the deployment
-# pre-call filter. Lives here so router and the filter share it without an import cycle.
-MID_STREAM_CONTINUATION_KWARG: Final = "_mid_stream_continuation"

litellm/litellm_core_utils/streaming_handler.py

@@ -2414,8 +2414,8 @@ class CustomStreamWrapper:
         # Shared by the sync, async, and non-aiohttp iteration sites so answer
         # text and the disqualifying-content latch stay in step across all three.
         get: Final = getattr(delta, "get", None)
-        content: Final = get("content", "") if callable(get) else ""
-        self.response_uptil_now += content or ""
+        content: Final = get("content") if callable(get) else None
+        self.response_uptil_now += content if isinstance(content, str) else ""
         if not self._emitted_disqualifying_content and self._delta_disqualifies_continuation(delta):
             self._emitted_disqualifying_content = True
 

litellm/router.py

@@ -62,7 +62,6 @@ from litellm.constants import (
     DEFAULT_HEALTH_CHECK_STALENESS_MULTIPLIER,
     DEFAULT_MAX_LRU_CACHE_SIZE,
     INTERNAL_CALL_ORIGIN_METADATA_KEY,
-    MID_STREAM_CONTINUATION_KWARG,
     OUTPUT_TOKEN_CEILING_PARAMS,
     ROUTING_REQUEST_TAGS_METADATA_KEY,
     RUNTIME_UPDATABLE_ROUTER_SETTINGS,
@@ -2853,6 +2852,10 @@ class Router:
                     )
                     initial_kwargs["original_function"] = self._acompletion
                     if continue_after_content:
+                        from litellm.router_utils.pre_call_checks.continuation_prefill_check import (
+                            MID_STREAM_CONTINUATION_KWARG,
+                        )
+
                         initial_kwargs["messages"] = self._build_completion_continuation_input(
                             messages, e.generated_content
                         )
@@ -3634,6 +3637,10 @@ class Router:
                 "client": model_client,
                 **kwargs,
             }
+            from litellm.router_utils.pre_call_checks.continuation_prefill_check import (
+                MID_STREAM_CONTINUATION_KWARG,
+            )
+
             input_kwargs.pop("silent_model", None)
             input_kwargs.pop("include_fallback_errors", None)
             input_kwargs.pop(MID_STREAM_CONTINUATION_KWARG, None)

litellm/router_utils/pre_call_checks/continuation_prefill_check.py

@@ -11,11 +11,14 @@ from typing import Final
 
 from pydantic import TypeAdapter, ValidationError
 
-from litellm.constants import MID_STREAM_CONTINUATION_KWARG
 from litellm.integrations.custom_logger import CustomLogger, Span
 from litellm.types.llms.openai import AllMessageValues
 from litellm.utils import supports_assistant_prefill
 
+# Marks a fallback re-entry as a mid-stream continuation. Router sets it (via a
+# lazy import) and this filter reads it; kept here to avoid a module-level cycle.
+MID_STREAM_CONTINUATION_KWARG: Final = "_mid_stream_continuation"
+
 _STR_KEYED_DICT_ADAPTER: Final = TypeAdapter(dict[str, object])