fix: don't retire a completed batch from cost recovery while output_file_id is still lagging

litellm/proxy/openai_files_endpoints/common_utils.py

@@ -1288,6 +1288,25 @@ def batch_cost_poller_is_active() -> bool:
         return False
 
 
+def _completed_batch_safe_to_retire(response) -> bool:
+    """Whether a "completed" batch may be retired from cost recovery.
+
+    ``batch_processed=True`` is the sole re-pickup gate for CheckBatchCost's
+    cost-recovery poller, so setting it retires the batch permanently. A batch can
+    reach ``status="completed"`` while ``output_file_id`` is still ``None`` (the
+    provider response briefly lags before the output id populates). Retiring in that
+    window loses the spend record forever. Retire only once we can prove there is
+    nothing left to recover: the output file has actually arrived, or the provider
+    reports no successful request lines. When counts are unknown, stay eligible so
+    the next poller pass revisits it. (#37713)
+    """
+    if getattr(response, "output_file_id", None) is not None:
+        return True
+    request_counts = getattr(response, "request_counts", None)
+    completed = getattr(request_counts, "completed", None)
+    return completed == 0
+
+
 async def update_batch_in_database(
     batch_id: str,
     unified_batch_id: str | Literal[False],
@@ -1369,7 +1388,7 @@ async def update_batch_in_database(
         }
 
         poller_owns: Final = batch_cost_poller_is_active() if poller_owns_accounting is None else poller_owns_accounting
-        if db_status == "complete" and not poller_owns:
+        if db_status == "complete" and not poller_owns and _completed_batch_safe_to_retire(response):
             update_data["batch_processed"] = True
 
         try:

tests/test_litellm/proxy/openai_files_endpoint/test_files_common_utils.py

@@ -431,3 +431,45 @@ def test_add_internal_model_credentials_survives_a_failing_deployment_lookup():
     add_internal_model_credentials(data=data, llm_router=router, model_id="deployment-gone")
 
     assert data == {"batch_id": "unified-batch-id"}
+
+
+from litellm.proxy.openai_files_endpoints.common_utils import (
+    _completed_batch_safe_to_retire,
+)
+
+
+def _completed_batch(output_file_id, completed=None) -> LiteLLMBatch:
+    kwargs = dict(
+        id="batch-1",
+        completion_window="24h",
+        created_at=1234567890,
+        endpoint="/v1/chat/completions",
+        input_file_id="file-in",
+        object="batch",
+        status="completed",
+        output_file_id=output_file_id,
+        error_file_id=None,
+    )
+    if completed is not None:
+        kwargs["request_counts"] = {"total": completed, "completed": completed, "failed": 0}
+    return LiteLLMBatch(**kwargs)
+
+
+class TestCompletedBatchSafeToRetire:
+    """A completed batch is only safe to retire from cost recovery once its output
+    file has arrived or the provider proves no successful lines (#37713)."""
+
+    def test_output_file_present_is_safe(self):
+        assert _completed_batch_safe_to_retire(_completed_batch("file-out")) is True
+
+    def test_no_output_and_no_successful_lines_is_safe(self):
+        # Every request line errored -> nothing left to recover.
+        assert _completed_batch_safe_to_retire(_completed_batch(None, completed=0)) is True
+
+    def test_no_output_but_successful_lines_is_not_safe(self):
+        # The bug: output_file_id is lagging; retiring here loses the spend record.
+        assert _completed_batch_safe_to_retire(_completed_batch(None, completed=5)) is False
+
+    def test_no_output_and_unknown_counts_is_not_safe(self):
+        # Counts unknown -> stay eligible so the next poller pass revisits it.
+        assert _completed_batch_safe_to_retire(_completed_batch(None)) is False