fix(oci): also report finish_reason="tool_calls" on streaming terminal chunks

Extends the finish_reason fix to the streaming handlers (per review): the
terminal SSE chunk normalized OCI's "COMPLETE" to "stop" even when tool calls
were emitted. handle_cohere_stream_chunk now reports "tool_calls" when tool
calls are present on the chunk or were emitted earlier (prior_tool_calls_emitted,
since the terminal chunk suppresses already-streamed tool calls), and
handle_generic_stream_chunk reports "tool_calls" when the chunk carries them.

litellm/llms/oci/chat/cohere.py

@@ -375,6 +375,13 @@ def handle_cohere_stream_chunk(
         ]
 
     finish_reason = _normalize_oci_finish_reason(typed_chunk.finishReason)
+    # OpenAI semantics: the terminal chunk must report finish_reason="tool_calls"
+    # when tool calls were emitted — on this chunk or earlier in the stream. OCI
+    # Cohere's terminal finishReason is "COMPLETE" (normalized to "stop"), and the
+    # terminal chunk suppresses already-emitted tool calls, so rely on
+    # prior_tool_calls_emitted as well.
+    if finish_reason is not None and (tool_calls or prior_tool_calls_emitted):
+        finish_reason = "tool_calls"
 
     return ModelResponseStream(
         choices=[

litellm/llms/oci/chat/generic.py

@@ -463,6 +463,9 @@ def handle_generic_stream_chunk(dict_chunk: dict) -> ModelResponseStream:
     finish_reason: Optional[str] = _normalize_oci_finish_reason(
         typed_chunk.finishReason
     )
+    # OpenAI semantics: a chunk carrying tool calls reports finish_reason="tool_calls".
+    if tool_calls:
+        finish_reason = "tool_calls"
 
     return ModelResponseStream(
         choices=[

tests/test_litellm/llms/oci/chat/test_oci_finish_reason.py

@@ -3,6 +3,8 @@ import datetime
 import httpx
 
 from litellm import ModelResponse
+from litellm.llms.oci.chat.cohere import handle_cohere_stream_chunk
+from litellm.llms.oci.chat.generic import handle_generic_stream_chunk
 from litellm.llms.oci.chat.transformation import OCIChatConfig
 
 
@@ -94,3 +96,52 @@ def test_generic_tool_calls_report_finish_reason_tool_calls():
     result = _transform("meta.llama-3.3-70b-instruct", _generic_body_with_tool_calls())
     assert result.choices[0].message.tool_calls is not None
     assert result.choices[0].finish_reason == "tool_calls"
+
+
+# --- streaming handlers (terminal chunk finish_reason) ---
+
+
+def _cohere_terminal_chunk() -> dict:
+    return {
+        "apiFormat": "COHERE",
+        "text": "All done.",
+        "finishReason": "COMPLETE",
+        "chatHistory": [{"role": "CHATBOT", "message": "All done."}],
+        "index": 0,
+    }
+
+
+def test_cohere_stream_terminal_chunk_reports_tool_calls_when_emitted():
+    # On the terminal chunk, OCI Cohere reports finishReason="COMPLETE" and
+    # suppresses already-streamed tool calls; finish_reason must still be
+    # "tool_calls" when tool calls were emitted earlier in the stream.
+    result = handle_cohere_stream_chunk(
+        _cohere_terminal_chunk(), prior_tool_calls_emitted=True
+    )
+    assert result.choices[0].finish_reason == "tool_calls"
+
+
+def test_cohere_stream_terminal_chunk_reports_stop_without_tool_calls():
+    result = handle_cohere_stream_chunk(
+        _cohere_terminal_chunk(), prior_tool_calls_emitted=False
+    )
+    assert result.choices[0].finish_reason == "stop"
+
+
+def test_generic_stream_chunk_with_tool_calls_reports_tool_calls():
+    chunk = {
+        "finishReason": "STOP",
+        "message": {
+            "toolCalls": [
+                {
+                    "id": "call_0",
+                    "type": "FUNCTION",
+                    "name": "get_weather",
+                    "arguments": "{}",
+                }
+            ]
+        },
+        "index": 0,
+    }
+    result = handle_generic_stream_chunk(chunk)
+    assert result.choices[0].finish_reason == "tool_calls"