From 07de96ca98994289d3ee717aeb368425b66fe142 Mon Sep 17 00:00:00 2001
From: mateo-berri <277851410+mateo-berri@users.noreply.github.com>
Date: Sun, 6 Sep 2026 02:45:56 -0700
Subject: [PATCH] fix(oci): keep an aborted tool turn on stop instead of
 tool_calls

The finish-reason upgrade fired on every reason that normalized to stop,
so a Cohere turn OCI ended with ERROR, ERROR_TOXIC, ERROR_LIMIT, USER_CANCEL
or CANCELLED after streaming a tool delta came back as tool_calls and invited
the client to run a call the provider had just blocked. Gate the upgrade on
the reasons that mean the turn ran to completion, which is the only case it
was ever needed for: GENERIC already sends tool_calls itself.
---
 litellm/llms/oci/chat/generic.py              | 16 ++++++--------
 .../oci/chat/test_oci_cohere_tool_calls.py    | 21 +++++++++++++++++++
 .../llms/oci/chat/test_oci_generic_chat.py    | 10 +++------
 3 files changed, 30 insertions(+), 17 deletions(-)

diff --git a/litellm/llms/oci/chat/generic.py b/litellm/llms/oci/chat/generic.py
index 38b6684374c..85cd063cf68 100644
--- a/litellm/llms/oci/chat/generic.py
+++ b/litellm/llms/oci/chat/generic.py
@@ -252,22 +252,18 @@ _OCI_FINISH_REASONS: Final = MappingProxyType(
 
 _OPENAI_FINISH_REASONS: Final = frozenset({"stop", "length", "tool_calls", "content_filter", "function_call"})
 
+_OCI_RAN_TO_COMPLETION_REASONS: Final = frozenset({"COMPLETE", "stop"})
+
 
 def _normalize_oci_finish_reason(raw: str | None, *, has_tool_calls: bool = False) -> str | None:
-    """Map an OCI finish reason to its OpenAI-standard equivalent.
-
-    OCI uses two vocabularies. Cohere sends ``COMPLETE`` / ``MAX_TOKENS`` /
-    ``TOOL_CALL(S)`` plus a long tail of error and cancel reasons, while GENERIC
-    already sends the OpenAI names and has to be passed through rather than
-    collapsed. Anything outside both sets becomes ``"stop"``, ``None`` stays
-    ``None`` so intermediate stream chunks stay open, and ``has_tool_calls``
-    upgrades ``stop`` to ``tool_calls`` for a Cohere tool turn, which OCI still
-    labels ``COMPLETE``.
+    """Cohere speaks ``COMPLETE`` / ``MAX_TOKENS`` / ``TOOL_CALL(S)`` plus a tail of error and
+    cancel reasons while GENERIC already speaks OpenAI, and only a turn that ran to completion
+    may be upgraded to ``tool_calls``, so an aborted one never invites a client to run the call.
     """
     if raw is None:
         return None
     normalized: Final = _OCI_FINISH_REASONS.get(raw, raw if raw in _OPENAI_FINISH_REASONS else "stop")
-    return "tool_calls" if has_tool_calls and normalized == "stop" else normalized
+    return "tool_calls" if has_tool_calls and raw in _OCI_RAN_TO_COMPLETION_REASONS else normalized
 
 
 def _synthesize_oci_tool_call_id(position: int, name: str, arguments: str) -> str:
diff --git a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
index 0b6b375cbf1..995cc1ce562 100644
--- a/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
+++ b/tests/test_litellm/llms/oci/chat/test_oci_cohere_tool_calls.py
@@ -502,6 +502,27 @@ class TestOCICohereToolCalls:
 
         assert chunks[-1].choices[0].finish_reason == "tool_calls"
 
+    @pytest.mark.parametrize(
+        "aborted_reason", ["ERROR", "ERROR_TOXIC", "ERROR_LIMIT", "USER_CANCEL", "CANCELLED"]
+    )
+    def test_cohere_aborted_tool_turn_still_reports_stop(self, aborted_reason: str):
+        """A turn OCI cut short is not a tool turn, however many tool deltas already went out."""
+        wrapper = OCIStreamWrapper(
+            completion_stream=MagicMock(), model="cohere.command-a-03-2025", logging_obj=MagicMock()
+        )
+        tool_calls_event = {
+            "apiFormat": "COHERE",
+            "text": "I will use the get_weather tool.",
+            "toolCalls": [{"name": "get_weather", "parameters": {"city": "Paris"}}],
+        }
+
+        chunks = [
+            wrapper.chunk_creator(f"data: {json.dumps(event)}")
+            for event in (tool_calls_event, {"apiFormat": "COHERE", "finishReason": aborted_reason})
+        ]
+
+        assert chunks[-1].choices[0].finish_reason == "stop"
+
     def test_cohere_non_streaming_tool_call_reports_tool_calls(self):
         """Same COMPLETE-on-a-tool-turn story on the non-streaming Cohere response."""
         body = {
diff --git a/tests/test_litellm/llms/oci/chat/test_oci_generic_chat.py b/tests/test_litellm/llms/oci/chat/test_oci_generic_chat.py
index d81ec0c6812..0f43a54b18e 100644
--- a/tests/test_litellm/llms/oci/chat/test_oci_generic_chat.py
+++ b/tests/test_litellm/llms/oci/chat/test_oci_generic_chat.py
@@ -472,13 +472,9 @@ class TestGpt5MaxCompletionTokens:
 
 
 class TestGenericFinishReasonFidelity:
-    """OCI GENERIC sends the OpenAI finish-reason names verbatim.
-
-    Captured against ``oci/meta.llama-3.3-70b-instruct`` in ``us-chicago-1``: a
-    tool-calling turn ends with ``"finishReason": "tool_calls"`` and a
-    max-token cut-off with ``"finishReason": "length"``, both lowercase.
-    Collapsing either to ``stop`` leaves an agent loop with no signal to run
-    the tool and hides truncation from the caller.
+    """``_TOOL_CALL_WIRE`` was captured against ``oci/meta.llama-3.3-70b-instruct`` in
+    ``us-chicago-1``, where a GENERIC tool-calling turn ends with ``"finishReason": "tool_calls"``
+    and a max-token cut-off with ``"finishReason": "length"``, both already lowercase.
     """
 
     _TOOL_CALL_WIRE = [