fix(gemini): surface thought output from thinking models

2026-07-02 21:58:11 +00:00 · 2026-03-02 16:12:54 -05:00
parent b3f8a42321
commit 582799ca05
4 changed files with 82 additions and 15 deletions
--- a/lib/crewai/src/crewai/events/types/llm_events.py
+++ b/lib/crewai/src/crewai/events/types/llm_events.py
@@ -86,3 +86,11 @@ class LLMStreamChunkEvent(LLMEventBase):
    tool_call: ToolCall | None = None
    call_type: LLMCallType | None = None
    response_id: str | None = None
+
+
+class LLMThinkingChunkEvent(LLMEventBase):
+    """Event emitted when a thinking/reasoning chunk is received from a thinking model"""
+
+    type: str = "llm_thinking_chunk"
+    chunk: str
+    response_id: str | None = None
--- a/lib/crewai/src/crewai/llms/base_llm.py
+++ b/lib/crewai/src/crewai/llms/base_llm.py
@@ -26,6 +26,7 @@ from crewai.events.types.llm_events import (
    LLMCallStartedEvent,
    LLMCallType,
    LLMStreamChunkEvent,
+    LLMThinkingChunkEvent,
 )
 from crewai.events.types.tool_usage_events import (
    ToolUsageErrorEvent,
@@ -465,6 +466,35 @@ class BaseLLM(ABC):
            ),
        )

+    def _emit_thinking_chunk_event(
+        self,
+        chunk: str,
+        from_task: Task | None = None,
+        from_agent: Agent | None = None,
+        response_id: str | None = None,
+    ) -> None:
+        """Emit thinking/reasoning chunk event from a thinking model.
+
+        Args:
+            chunk: The thinking text content.
+            from_task: The task that initiated the call.
+            from_agent: The agent that initiated the call.
+            response_id: Unique ID for a particular LLM response.
+        """
+        if not hasattr(crewai_event_bus, "emit"):
+            raise ValueError("crewai_event_bus does not have an emit method") from None
+
+        crewai_event_bus.emit(
+            self,
+            event=LLMThinkingChunkEvent(
+                chunk=chunk,
+                from_task=from_task,
+                from_agent=from_agent,
+                response_id=response_id,
+                call_id=get_current_call_id(),
+            ),
+        )
+
    def _handle_tool_execution(
        self,
        function_name: str,
--- a/lib/crewai/src/crewai/llms/providers/gemini/completion.py
+++ b/lib/crewai/src/crewai/llms/providers/gemini/completion.py
@@ -61,6 +61,7 @@ class GeminiCompletion(BaseLLM):
        interceptor: BaseInterceptor[Any, Any] | None = None,
        use_vertexai: bool | None = None,
        response_format: type[BaseModel] | None = None,
+        thinking_config: types.ThinkingConfig | None = None,
        **kwargs: Any,
    ):
        """Initialize Google Gemini chat completion client.
@@ -93,6 +94,10 @@ class GeminiCompletion(BaseLLM):
                         api_version="v1" is automatically configured.
            response_format: Pydantic model for structured output. Used as default when
                           response_model is not passed to call()/acall() methods.
+            thinking_config: ThinkingConfig for thinking models (gemini-2.5+, gemini-3+).
+                           Controls thought output via include_thoughts, thinking_budget,
+                           and thinking_level. When None, thinking models automatically
+                           get include_thoughts=True so thought content is surfaced.
            **kwargs: Additional parameters
        """
        if interceptor is not None:
@@ -139,6 +144,14 @@ class GeminiCompletion(BaseLLM):
            version_match and float(version_match.group(1)) >= 2.0
        )

+        self.thinking_config = thinking_config
+        if (
+            self.thinking_config is None
+            and version_match
+            and float(version_match.group(1)) >= 2.5
+        ):
+            self.thinking_config = types.ThinkingConfig(include_thoughts=True)
+
    @property
    def stop(self) -> list[str]:
        """Get stop sequences sent to the API."""
@@ -520,6 +533,9 @@ class GeminiCompletion(BaseLLM):
        if self.safety_settings:
            config_params["safety_settings"] = self.safety_settings

+        if self.thinking_config is not None:
+            config_params["thinking_config"] = self.thinking_config
+
        return types.GenerateContentConfig(**config_params)

    def _convert_tools_for_interference(  # type: ignore[override]
@@ -931,15 +947,6 @@ class GeminiCompletion(BaseLLM):
        if chunk.usage_metadata:
            usage_data = self._extract_token_usage(chunk)

-        if chunk.text:
-            full_response += chunk.text
-            self._emit_stream_chunk_event(
-                chunk=chunk.text,
-                from_task=from_task,
-                from_agent=from_agent,
-                response_id=response_id,
-            )
-
        if chunk.candidates:
            candidate = chunk.candidates[0]
            if candidate.content and candidate.content.parts:
@@ -976,6 +983,21 @@ class GeminiCompletion(BaseLLM):
                            call_type=LLMCallType.TOOL_CALL,
                            response_id=response_id,
                        )
+                    elif part.thought and part.text:
+                        self._emit_thinking_chunk_event(
+                            chunk=part.text,
+                            from_task=from_task,
+                            from_agent=from_agent,
+                            response_id=response_id,
+                        )
+                    elif part.text:
+                        full_response += part.text
+                        self._emit_stream_chunk_event(
+                            chunk=part.text,
+                            from_task=from_task,
+                            from_agent=from_agent,
+                            response_id=response_id,
+                        )

        return full_response, function_calls, usage_data

@@ -1329,7 +1351,7 @@ class GeminiCompletion(BaseLLM):
        text_parts = [
            part.text
            for part in candidate.content.parts
-            if hasattr(part, "text") and part.text
+            if part.text and not part.thought
        ]

        return "".join(text_parts)
--- a/uv.lock
+++ b/uv.lock
@@ -1197,7 +1197,7 @@ requires-dist = [
    { name = "crewai-files", marker = "extra == 'file-processing'", editable = "lib/crewai-files" },
    { name = "crewai-tools", marker = "extra == 'tools'", editable = "lib/crewai-tools" },
    { name = "docling", marker = "extra == 'docling'", specifier = "~=2.75.0" },
-    { name = "google-genai", marker = "extra == 'google-genai'", specifier = "~=1.49.0" },
+    { name = "google-genai", marker = "extra == 'google-genai'", specifier = "~=1.65.0" },
    { name = "httpx", specifier = "~=0.28.1" },
    { name = "httpx-auth", marker = "extra == 'a2a'", specifier = "~=0.23.1" },
    { name = "httpx-sse", marker = "extra == 'a2a'", specifier = "~=0.4.0" },
@@ -2249,6 +2249,11 @@ wheels = [
    { url = "https://files.pythonhosted.org/packages/83/1d/d6466de3a5249d35e832a52834115ca9d1d0de6abc22065f049707516d47/google_auth-2.48.0-py3-none-any.whl", hash = "sha256:2e2a537873d449434252a9632c28bfc268b0adb1e53f9fb62afc5333a975903f", size = 236499, upload-time = "2026-01-26T19:22:45.099Z" },
 ]

+[package.optional-dependencies]
+requests = [
+    { name = "requests" },
+]
+
 [[package]]
 name = "google-cloud-vision"
 version = "3.12.1"
@@ -2267,21 +2272,23 @@ wheels = [

 [[package]]
 name = "google-genai"
-version = "1.49.0"
+version = "1.65.0"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
    { name = "anyio" },
-    { name = "google-auth" },
+    { name = "distro" },
+    { name = "google-auth", extra = ["requests"] },
    { name = "httpx" },
    { name = "pydantic" },
    { name = "requests" },
+    { name = "sniffio" },
    { name = "tenacity" },
    { name = "typing-extensions" },
    { name = "websockets" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/82/49/1a724ee3c3748fa50721d53a52d9fee88c67d0c43bb16eb2b10ee89ab239/google_genai-1.49.0.tar.gz", hash = "sha256:35eb16023b72e298571ae30e919c810694f258f2ba68fc77a2185c7c8829ad5a", size = 253493, upload-time = "2025-11-05T22:41:03.278Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/79/f9/cc1191c2540d6a4e24609a586c4ed45d2db57cfef47931c139ee70e5874a/google_genai-1.65.0.tar.gz", hash = "sha256:d470eb600af802d58a79c7f13342d9ea0d05d965007cae8f76c7adff3d7a4750", size = 497206, upload-time = "2026-02-26T00:20:33.824Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/d5/d3/84a152746dc7bdebb8ba0fd7d6157263044acd1d14b2a53e8df4a307b6b7/google_genai-1.49.0-py3-none-any.whl", hash = "sha256:ad49cd5be5b63397069e7aef9a4fe0a84cbdf25fcd93408e795292308db4ef32", size = 256098, upload-time = "2025-11-05T22:41:01.429Z" },
+    { url = "https://files.pythonhosted.org/packages/68/3c/3fea4e7c91357c71782d7dcaad7a2577d636c90317e003386893c25bc62c/google_genai-1.65.0-py3-none-any.whl", hash = "sha256:68c025205856919bc03edb0155c11b4b833810b7ce17ad4b7a9eeba5158f6c44", size = 724429, upload-time = "2026-02-26T00:20:32.186Z" },
 ]

 [[package]]