fix(streaming): keep the text of a chunk that has an empty reasoning field

CHANGELOG.md

@@ -24,6 +24,7 @@ and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0
 
 ### Bug fixes
 
+* Streaming from `ChatOpenAICompletions()` (and the providers built on it, such as `ChatOpenRouter()` and `ChatGroq()`) no longer drops a chunk's text when the same chunk carries an empty `reasoning` or `reasoning_content` field, as some OpenAI-compatible servers send.
 * `ChatOllama()` can now be created without a `model`, so `ChatOllama().list_models()` works. Sending a message without a model still raises an error that lists the available models. (#166)
 * `ChatGoogle()` now reports `finish_reason == "content_filter"` for Gemini's `IMAGE_SAFETY`, `IMAGE_PROHIBITED_CONTENT`, `IMAGE_RECITATION` and `LANGUAGE` finish reasons, instead of passing the raw Gemini string through, and reports `"tool_use"` (rather than `"success"`) when a turn ends with a tool request, as `ChatOpenAI()` does. (#473)
 * The error raised for a tool request with invalid JSON arguments from `ChatOpenAI()` and `ChatOpenAICompletions()` (and the providers built on them) now shows the offending arguments, instead of the literal text `{args}`.

chatlas/_provider_openai_completions.py

@@ -320,7 +320,7 @@ def stream_content(self, chunk, completion, turns=()) -> list[Content]:
         reasoning = getattr(delta, "reasoning", None) or getattr(
             delta, "reasoning_content", None
         )
-        if reasoning is not None:
+        if reasoning:
             return [ContentThinkingDelta(thinking=reasoning)]
 
         text = delta.content

tests/test_provider_openai_completions.py

@@ -206,6 +206,33 @@ def __init__(self, choices):
     assert result == [ContentText(text="hello")]
 
 
+def test_stream_content_empty_reasoning_keeps_text():
+    """Some backends send `reasoning_content: ""` alongside the text of a chunk."""
+    from openai.types.chat import ChatCompletionChunk
+
+    provider = OpenAICompletionsProvider(model="test")
+
+    def chunk(**delta):
+        return ChatCompletionChunk.model_validate(
+            {
+                "id": "1",
+                "object": "chat.completion.chunk",
+                "created": 0,
+                "model": "test",
+                "choices": [{"index": 0, "delta": delta, "finish_reason": None}],
+            }
+        )
+
+    result = provider.stream_content(chunk(content="Hello", reasoning_content=""), None)
+    assert result == [ContentText(text="Hello")]
+
+    result = provider.stream_content(chunk(content="Hello", reasoning=""), None)
+    assert result == [ContentText(text="Hello")]
+
+    # A chunk with neither reasoning nor text adds nothing.
+    assert provider.stream_content(chunk(reasoning_content=""), None) == []
+
+
 def test_response_as_turn_extracts_reasoning_content():
     from unittest.mock import Mock