fix: bug fixes from PR review

- streaming_iterator: don't set sent_content_block_finish during compaction
  block lifecycle; that flag tracks the regular text/tool_use/thinking block
  state machine, conflating the two leaks bad state to introspection paths.
- compact._call_summary_model: send propagated proxy auth/spend-attribution
  fields as 'litellm_metadata' instead of 'metadata' so the router's post-call
  hooks attribute summary tokens to the caller's key/team budget.

Co-authored-by: Yassin Kortam <yassin@berri.ai>

litellm/llms/anthropic/experimental_pass_through/adapters/streaming_iterator.py

@@ -245,12 +245,14 @@ class AnthropicStreamWrapper(AdapterCompletionStreamWrapper):
             "type": "content_block_stop",
             "index": compaction_index,
         }
-        # Advance state atomically with returning the terminal event so
-        # outside observers never see ``sent_content_block_finish=True``
-        # before the client has received ``content_block_stop``.
+        # Don't touch ``sent_content_block_finish`` here: that flag is the
+        # state machine for the regular text/tool_use/thinking block and is
+        # independent of the synthetic compaction block lifecycle. Conflating
+        # them would let outside observers (subclass overrides, introspection
+        # hooks, exception paths) see ``sent_content_block_finish=True``
+        # without any regular content block ever having started.
         self._increment_content_block_index()
         self.sent_compaction_block = True
-        self.sent_content_block_finish = True
         return stop_event
 
     def _create_initial_usage_delta(self) -> UsageDelta:

litellm/llms/anthropic/experimental_pass_through/context_management/editors/compact.py

@@ -536,11 +536,16 @@ async def _call_summary_model(
     # accepted by providers that don't strictly require it (OpenAI etc.).
     # Setting a sensible default here means the feature works regardless of
     # which model an admin configures as ``context_management_summary_model``.
+    # The propagated proxy auth/spend-attribution fields (``user_api_key`` etc.)
+    # must travel as ``litellm_metadata`` — that is the parameter the proxy's
+    # post-call spend hooks read for budget attribution. The provider-level
+    # ``metadata`` kwarg corresponds to the upstream API request body and would
+    # not flow into spend tracking.
     call_kwargs: Dict[str, Any] = {
         "model": summary_model,
         "messages": summary_messages,
         "max_tokens": COMPACT_SUMMARY_MAX_TOKENS,
-        "metadata": metadata,
+        "litellm_metadata": metadata,
     }
     if llm_router is not None and hasattr(llm_router, "acompletion"):
         return await llm_router.acompletion(**call_kwargs)

tests/test_litellm/llms/anthropic/experimental_pass_through/context_management/test_compact.py

@@ -1295,7 +1295,7 @@ async def test_prepare_context_managed_request_forwards_proxy_litellm_metadata()
 
     class _RouterStub:
         async def acompletion(self, **kwargs):
-            captured_summary_metadata.update(kwargs.get("metadata", {}))
+            captured_summary_metadata.update(kwargs.get("litellm_metadata", {}))
             return _make_mock_response("<summary>s</summary>")
 
     with (