fix: address review feedback for streaming handlers and test robustness

- Fix sync streaming handler: check accumulated_tool_args (from deltas)
  instead of only checking tool_calls from last chunk. This mirrors the
  async streaming handler fix.
- Add isinstance(tool_calls, list) checks in non-streaming handlers for
  type safety (prevents false positives from auto-created MagicMock attrs).
- Update test_handle_streaming_tool_calls_no_available_functions to verify
  tool calls are returned (not discarded) when available_functions is None,
  matching the corrected behavior from issue #4788.
- Emit LLM completion event before returning accumulated tool calls from
  streaming handler.

Co-Authored-By: João <joao@crewai.com>

.github/workflows/nightly.yml

@@ -0,0 +1,127 @@
+name: Nightly Canary Release
+
+on:
+  schedule:
+    - cron: '0 6 * * *' # daily at 6am UTC
+  workflow_dispatch:
+
+jobs:
+  check:
+    name: Check for new commits
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    outputs:
+      has_changes: ${{ steps.check.outputs.has_changes }}
+    steps:
+      - uses: actions/checkout@v4
+        with:
+          fetch-depth: 0
+
+      - name: Check for commits in last 24h
+        id: check
+        run: |
+          RECENT=$(git log --since="24 hours ago" --oneline | head -1)
+          if [ -n "$RECENT" ]; then
+            echo "has_changes=true" >> "$GITHUB_OUTPUT"
+          else
+            echo "has_changes=false" >> "$GITHUB_OUTPUT"
+          fi
+
+  build:
+    name: Build nightly packages
+    needs: check
+    if: needs.check.outputs.has_changes == 'true' || github.event_name == 'workflow_dispatch'
+    runs-on: ubuntu-latest
+    permissions:
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Set up Python
+        uses: actions/setup-python@v5
+        with:
+          python-version: "3.12"
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v4
+
+      - name: Stamp nightly versions
+        run: |
+          DATE=$(date +%Y%m%d)
+          for init_file in \
+            lib/crewai/src/crewai/__init__.py \
+            lib/crewai-tools/src/crewai_tools/__init__.py \
+            lib/crewai-files/src/crewai_files/__init__.py; do
+            CURRENT=$(python -c "
+          import re
+          text = open('$init_file').read()
+          print(re.search(r'__version__\s*=\s*\"(.*?)\"\s*$', text, re.MULTILINE).group(1))
+          ")
+            NIGHTLY="${CURRENT}.dev${DATE}"
+            sed -i "s/__version__ = .*/__version__ = \"${NIGHTLY}\"/" "$init_file"
+            echo "$init_file: $CURRENT -> $NIGHTLY"
+          done
+
+          # Update cross-package dependency pins to nightly versions
+          sed -i "s/\"crewai-tools==[^\"]*\"/\"crewai-tools==${NIGHTLY}\"/" lib/crewai/pyproject.toml
+          sed -i "s/\"crewai==[^\"]*\"/\"crewai==${NIGHTLY}\"/" lib/crewai-tools/pyproject.toml
+          echo "Updated cross-package dependency pins to ${NIGHTLY}"
+
+      - name: Build packages
+        run: |
+          uv build --all-packages
+          rm dist/.gitignore
+
+      - name: Upload artifacts
+        uses: actions/upload-artifact@v4
+        with:
+          name: dist
+          path: dist/
+
+  publish:
+    name: Publish nightly to PyPI
+    needs: build
+    runs-on: ubuntu-latest
+    environment:
+      name: pypi
+      url: https://pypi.org/p/crewai
+    permissions:
+      id-token: write
+      contents: read
+    steps:
+      - uses: actions/checkout@v4
+
+      - name: Install uv
+        uses: astral-sh/setup-uv@v6
+        with:
+          version: "0.8.4"
+          python-version: "3.12"
+          enable-cache: false
+
+      - name: Download artifacts
+        uses: actions/download-artifact@v4
+        with:
+          name: dist
+          path: dist
+
+      - name: Publish to PyPI
+        env:
+          UV_PUBLISH_TOKEN: ${{ secrets.PYPI_API_TOKEN }}
+        run: |
+          failed=0
+          for package in dist/*; do
+            if [[ "$package" == *"crewai_devtools"* ]]; then
+              echo "Skipping private package: $package"
+              continue
+            fi
+            echo "Publishing $package"
+            if ! uv publish "$package"; then
+              echo "Failed to publish $package"
+              failed=1
+            fi
+          done
+          if [ $failed -eq 1 ]; then
+            echo "Some packages failed to publish"
+            exit 1
+          fi

docs/en/changelog.mdx

@@ -4,6 +4,38 @@ description: "Product updates, improvements, and bug fixes for CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="Mar 04, 2026">
+  ## v1.10.1
+
+  [View release on GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## What's Changed
+
+  ### Features
+  - Upgrade Gemini GenAI
+
+  ### Bug Fixes
+  - Adjust executor listener value to avoid recursion
+  - Group parallel function response parts in a single Content object in Gemini
+  - Surface thought output from thinking models in Gemini
+  - Load MCP and platform tools when agent tools are None
+  - Support Jupyter environments with running event loops in A2A
+  - Use anonymous ID for ephemeral traces
+  - Conditionally pass plus header
+  - Skip signal handler registration in non-main threads for telemetry
+  - Inject tool errors as observations and resolve name collisions
+  - Upgrade pypdf from 4.x to 6.7.4 to resolve Dependabot alerts
+  - Resolve critical and high Dependabot security alerts
+
+  ### Documentation
+  - Sync Composio tool documentation across locales
+
+  ## Contributors
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
 <Update label="Feb 27, 2026">
   ## v1.10.1a1
 

docs/ko/changelog.mdx

@@ -4,6 +4,38 @@ description: "CrewAI의 제품 업데이트, 개선 사항 및 버그 수정"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="2026년 3월 4일">
+  ## v1.10.1
+
+  [GitHub 릴리스 보기](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## 변경 사항
+
+  ### 기능
+  - Gemini GenAI 업그레이드
+
+  ### 버그 수정
+  - 재귀를 피하기 위해 실행기 리스너 값을 조정
+  - Gemini에서 병렬 함수 응답 부분을 단일 Content 객체로 그룹화
+  - Gemini에서 사고 모델의 사고 출력을 표시
+  - 에이전트 도구가 None일 때 MCP 및 플랫폼 도구 로드
+  - A2A에서 실행 이벤트 루프가 있는 Jupyter 환경 지원
+  - 일시적인 추적을 위해 익명 ID 사용
+  - 조건부로 플러스 헤더 전달
+  - 원격 측정을 위해 비주 스레드에서 신호 처리기 등록 건너뛰기
+  - 도구 오류를 관찰로 주입하고 이름 충돌 해결
+  - Dependabot 경고를 해결하기 위해 pypdf를 4.x에서 6.7.4로 업그레이드
+  - 심각 및 높은 Dependabot 보안 경고 해결
+
+  ### 문서
+  - Composio 도구 문서를 지역별로 동기화
+
+  ## 기여자
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
 <Update label="2026년 2월 27일">
   ## v1.10.1a1
 

docs/pt-BR/changelog.mdx

@@ -4,6 +4,38 @@ description: "Atualizações de produto, melhorias e correções do CrewAI"
 icon: "clock"
 mode: "wide"
 ---
+<Update label="04 mar 2026">
+  ## v1.10.1
+
+  [Ver release no GitHub](https://github.com/crewAIInc/crewAI/releases/tag/1.10.1)
+
+  ## O que mudou
+
+  ### Recursos
+  - Atualizar Gemini GenAI
+
+  ### Correções de Bugs
+  - Ajustar o valor do listener do executor para evitar recursão
+  - Agrupar partes da resposta da função paralela em um único objeto Content no Gemini
+  - Exibir a saída de pensamento dos modelos de pensamento no Gemini
+  - Carregar ferramentas MCP e da plataforma quando as ferramentas do agente forem None
+  - Suportar ambientes Jupyter com loops de eventos em A2A
+  - Usar ID anônimo para rastreamentos efêmeros
+  - Passar condicionalmente o cabeçalho plus
+  - Ignorar o registro do manipulador de sinal em threads não principais para telemetria
+  - Injetar erros de ferramentas como observações e resolver colisões de nomes
+  - Atualizar pypdf de 4.x para 6.7.4 para resolver alertas do Dependabot
+  - Resolver alertas de segurança críticos e altos do Dependabot
+
+  ### Documentação
+  - Sincronizar a documentação da ferramenta Composio entre locais
+
+  ## Contribuidores
+
+  @giulio-leone, @greysonlalonde, @haxzie, @joaomdmoura, @lorenzejay, @mattatcha, @mplachta, @nicoferdi96
+
+</Update>
+
 <Update label="27 fev 2026">
   ## v1.10.1a1
 

lib/crewai-tools/pyproject.toml

@@ -108,7 +108,7 @@ stagehand = [
     "stagehand>=0.4.1",
 ]
 github = [
-    "gitpython==3.1.38",
+    "gitpython>=3.1.41,<4",
     "PyGithub==1.59.1",
 ]
 rag = [

lib/crewai/src/crewai/agents/agent_builder/base_agent_executor_mixin.py

@@ -30,12 +30,9 @@ class CrewAgentExecutorMixin:
         memory = getattr(self.agent, "memory", None) or (
             getattr(self.crew, "_memory", None) if self.crew else None
         )
-        if memory is None or not self.task or getattr(memory, "_read_only", False):
+        if memory is None or not self.task or memory.read_only:
             return
-        if (
-            f"Action: {sanitize_tool_name('Delegate work to coworker')}"
-            in output.text
-        ):
+        if f"Action: {sanitize_tool_name('Delegate work to coworker')}" in output.text:
             return
         try:
             raw = (
@@ -48,6 +45,4 @@ class CrewAgentExecutorMixin:
             if extracted:
                 memory.remember_many(extracted, agent_role=self.agent.role)
         except Exception as e:
-            self.agent._logger.log(
-                "error", f"Failed to save to memory: {e}"
-            )
+            self.agent._logger.log("error", f"Failed to save to memory: {e}")

lib/crewai/src/crewai/agents/crew_agent_executor.py

@@ -8,6 +8,7 @@ from __future__ import annotations
 
 import asyncio
 from collections.abc import Callable
+import contextvars
 from concurrent.futures import ThreadPoolExecutor, as_completed
 import inspect
 import logging
@@ -755,6 +756,7 @@ class CrewAgentExecutor(CrewAgentExecutorMixin):
                 with ThreadPoolExecutor(max_workers=max_workers) as pool:
                     futures = {
                         pool.submit(
+                            contextvars.copy_context().run,
                             self._execute_single_native_tool_call,
                             call_id=call_id,
                             func_name=func_name,

lib/crewai/src/crewai/experimental/agent_executor.py

@@ -1,6 +1,7 @@
 from __future__ import annotations
 
 import asyncio
+import contextvars
 from collections.abc import Callable, Coroutine
 from concurrent.futures import ThreadPoolExecutor, as_completed
 from datetime import datetime
@@ -728,7 +729,7 @@ class AgentExecutor(Flow[AgentReActState], CrewAgentExecutorMixin):
             max_workers = min(8, len(runnable_tool_calls))
             with ThreadPoolExecutor(max_workers=max_workers) as pool:
                 future_to_idx = {
-                    pool.submit(self._execute_single_native_tool_call, tool_call): idx
+                    pool.submit(contextvars.copy_context().run, self._execute_single_native_tool_call, tool_call): idx
                     for idx, tool_call in enumerate(runnable_tool_calls)
                 }
                 ordered_results: list[dict[str, Any] | None] = [None] * len(

lib/crewai/src/crewai/lite_agent.py

@@ -600,7 +600,7 @@ class LiteAgent(FlowTrackable, BaseModel):
 
     def _save_to_memory(self, output_text: str) -> None:
         """Extract discrete memories from the run and remember each. No-op if _memory is None or read-only."""
-        if self._memory is None or getattr(self._memory, "_read_only", False):
+        if self._memory is None or self._memory.read_only:
             return
         input_str = self._get_last_user_content() or "User request"
         try:

lib/crewai/src/crewai/llm.py

@@ -967,7 +967,43 @@ class LLM(BaseLLM):
                 self._track_token_usage_internal(usage_info)
             self._handle_streaming_callbacks(callbacks, usage_info, last_chunk)
 
-            if not tool_calls or not available_functions:
+            # --- 8) Check accumulated_tool_args from streaming deltas
+            # Streaming responses deliver tool calls via deltas accumulated in
+            # accumulated_tool_args, not via the final chunk's message. When
+            # available_functions is None (native tool handling), we must return
+            # the accumulated tool calls so the caller (e.g., executor) can
+            # handle them. When available_functions is provided, tool execution
+            # already happened during the chunk processing loop via
+            # _handle_streaming_tool_calls.
+            if accumulated_tool_args and not available_functions:
+                tool_calls_list: list[ChatCompletionDeltaToolCall] = [
+                    ChatCompletionDeltaToolCall(
+                        index=idx,
+                        function=Function(
+                            name=tool_arg.function.name,
+                            arguments=tool_arg.function.arguments,
+                        ),
+                    )
+                    for idx, tool_arg in accumulated_tool_args.items()
+                    if tool_arg.function.name
+                ]
+
+                if tool_calls_list:
+                    self._handle_emit_call_events(
+                        response=full_response,
+                        call_type=LLMCallType.LLM_CALL,
+                        from_task=from_task,
+                        from_agent=from_agent,
+                        messages=params["messages"],
+                    )
+                    return tool_calls_list
+
+            # --- 8b) If there are tool calls from last chunk but no available functions,
+            # return the tool calls
+            if tool_calls and not available_functions:
+                return tool_calls
+
+            if not tool_calls and not accumulated_tool_args:
                 if response_model and self.is_litellm:
                     instructor_instance = InternalInstructor(
                         content=full_response,
@@ -994,10 +1030,11 @@ class LLM(BaseLLM):
                 )
                 return full_response
 
-            # --- 9) Handle tool calls if present
-            tool_result = self._handle_tool_call(tool_calls, available_functions)
-            if tool_result is not None:
-                return tool_result
+            # --- 9) Handle tool calls from last chunk if present (execute when available_functions provided)
+            if tool_calls and available_functions:
+                tool_result = self._handle_tool_call(tool_calls, available_functions)
+                if tool_result is not None:
+                    return tool_result
 
             # --- 10) Emit completion event and return response
             self._handle_emit_call_events(
@@ -1234,8 +1271,17 @@ class LLM(BaseLLM):
         # --- 4) Check for tool calls
         tool_calls = getattr(response_message, "tool_calls", [])
 
-        # --- 5) If no tool calls or no available functions, return the text response directly as long as there is a text response
-        if (not tool_calls or not available_functions) and text_response:
+        # --- 5) If there are tool calls but no available functions, return the tool calls
+        # This allows the caller (e.g., executor) to handle tool execution
+        # This must be checked before the text response fallback because some LLMs
+        # (e.g., Anthropic) return both text content and tool calls in the same response.
+        # The isinstance check ensures we have actual tool call data (list), not
+        # auto-generated attributes from mocks or unexpected types.
+        if isinstance(tool_calls, list) and tool_calls and not available_functions:
+            return tool_calls
+
+        # --- 6) If no tool calls or no available functions, return the text response directly as long as there is a text response
+        if not tool_calls and text_response:
             self._handle_emit_call_events(
                 response=text_response,
                 call_type=LLMCallType.LLM_CALL,
@@ -1245,13 +1291,8 @@ class LLM(BaseLLM):
             )
             return text_response
 
-        # --- 6) If there are tool calls but no available functions, return the tool calls
-        # This allows the caller (e.g., executor) to handle tool execution
-        if tool_calls and not available_functions:
-            return tool_calls
-
         # --- 7) Handle tool calls if present (execute when available_functions provided)
-        if tool_calls and available_functions:
+        if isinstance(tool_calls, list) and tool_calls and available_functions:
             tool_result = self._handle_tool_call(
                 tool_calls, available_functions, from_task, from_agent
             )
@@ -1364,7 +1405,16 @@ class LLM(BaseLLM):
 
         tool_calls = getattr(response_message, "tool_calls", [])
 
-        if (not tool_calls or not available_functions) and text_response:
+        # If there are tool calls but no available functions, return the tool calls
+        # This allows the caller (e.g., executor) to handle tool execution
+        # This must be checked before the text response fallback because some LLMs
+        # (e.g., Anthropic) return both text content and tool calls in the same response.
+        # The isinstance check ensures we have actual tool call data (list), not
+        # auto-generated attributes from mocks or unexpected types.
+        if isinstance(tool_calls, list) and tool_calls and not available_functions:
+            return tool_calls
+
+        if not tool_calls and text_response:
             self._handle_emit_call_events(
                 response=text_response,
                 call_type=LLMCallType.LLM_CALL,
@@ -1374,13 +1424,8 @@ class LLM(BaseLLM):
             )
             return text_response
 
-        # If there are tool calls but no available functions, return the tool calls
-        # This allows the caller (e.g., executor) to handle tool execution
-        if tool_calls and not available_functions:
-            return tool_calls
-
         # Handle tool calls if present (execute when available_functions provided)
-        if tool_calls and available_functions:
+        if isinstance(tool_calls, list) and tool_calls and available_functions:
             tool_result = self._handle_tool_call(
                 tool_calls, available_functions, from_task, from_agent
             )
@@ -1513,7 +1558,7 @@ class LLM(BaseLLM):
             if usage_info:
                 self._track_token_usage_internal(usage_info)
 
-            if accumulated_tool_args and available_functions:
+            if accumulated_tool_args:
                 # Convert accumulated tool args to ChatCompletionDeltaToolCall objects
                 tool_calls_list: list[ChatCompletionDeltaToolCall] = [
                     ChatCompletionDeltaToolCall(
@@ -1527,7 +1572,14 @@ class LLM(BaseLLM):
                     if tool_arg.function.name
                 ]
 
-                if tool_calls_list:
+                # If there are tool calls but no available functions, return the tool calls
+                # This allows the caller (e.g., executor) to handle tool execution.
+                # This must be checked before the text response fallback because some LLMs
+                # (e.g., Anthropic) return both text content and tool calls in the same response.
+                if tool_calls_list and not available_functions:
+                    return tool_calls_list
+
+                if tool_calls_list and available_functions:
                     result = self._handle_streaming_tool_calls(
                         tool_calls=tool_calls_list,
                         accumulated_tool_args=accumulated_tool_args,

lib/crewai/src/crewai/memory/memory_scope.py

@@ -3,11 +3,9 @@
 from __future__ import annotations
 
 from datetime import datetime
-from typing import TYPE_CHECKING, Any
+from typing import Any, Literal
 
-
-if TYPE_CHECKING:
-    from crewai.memory.unified_memory import Memory
+from pydantic import BaseModel, ConfigDict, Field, PrivateAttr, model_validator
 
 from crewai.memory.types import (
     _RECALL_OVERSAMPLE_FACTOR,
@@ -15,22 +13,38 @@ from crewai.memory.types import (
     MemoryRecord,
     ScopeInfo,
 )
+from crewai.memory.unified_memory import Memory
 
 
-class MemoryScope:
+class MemoryScope(BaseModel):
     """View of Memory restricted to a root path. All operations are scoped under that path."""
 
-    def __init__(self, memory: Memory, root_path: str) -> None:
-        """Initialize scope.
+    model_config = ConfigDict(arbitrary_types_allowed=True)
 
-        Args:
-            memory: The underlying Memory instance.
-            root_path: Root path for this scope (e.g. /agent/1).
-        """
-        self._memory = memory
-        self._root = root_path.rstrip("/") or ""
-        if self._root and not self._root.startswith("/"):
-            self._root = "/" + self._root
+    root_path: str = Field(default="/")
+
+    _memory: Memory = PrivateAttr()
+    _root: str = PrivateAttr()
+
+    @model_validator(mode="wrap")
+    @classmethod
+    def _accept_memory(cls, data: Any, handler: Any) -> MemoryScope:
+        """Extract memory dependency and normalize root path before validation."""
+        if isinstance(data, MemoryScope):
+            return data
+        memory = data.pop("memory")
+        instance: MemoryScope = handler(data)
+        instance._memory = memory
+        root = instance.root_path.rstrip("/") or ""
+        if root and not root.startswith("/"):
+            root = "/" + root
+        instance._root = root
+        return instance
+
+    @property
+    def read_only(self) -> bool:
+        """Whether the underlying memory is read-only."""
+        return self._memory.read_only
 
     def _scope_path(self, scope: str | None) -> str:
         if not scope or scope == "/":
@@ -52,7 +66,7 @@ class MemoryScope:
         importance: float | None = None,
         source: str | None = None,
         private: bool = False,
-    ) -> MemoryRecord:
+    ) -> MemoryRecord | None:
         """Remember content; scope is relative to this scope's root."""
         path = self._scope_path(scope)
         return self._memory.remember(
@@ -71,7 +85,7 @@ class MemoryScope:
         scope: str | None = None,
         categories: list[str] | None = None,
         limit: int = 10,
-        depth: str = "deep",
+        depth: Literal["shallow", "deep"] = "deep",
         source: str | None = None,
         include_private: bool = False,
     ) -> list[MemoryMatch]:
@@ -138,34 +152,34 @@ class MemoryScope:
         """Return a narrower scope under this scope."""
         child = path.strip("/")
         if not child:
-            return MemoryScope(self._memory, self._root or "/")
+            return MemoryScope(memory=self._memory, root_path=self._root or "/")
         base = self._root.rstrip("/") or ""
         new_root = f"{base}/{child}" if base else f"/{child}"
-        return MemoryScope(self._memory, new_root)
+        return MemoryScope(memory=self._memory, root_path=new_root)
 
 
-class MemorySlice:
+class MemorySlice(BaseModel):
     """View over multiple scopes: recall searches all, remember is a no-op when read_only."""
 
-    def __init__(
-        self,
-        memory: Memory,
-        scopes: list[str],
-        categories: list[str] | None = None,
-        read_only: bool = True,
-    ) -> None:
-        """Initialize slice.
+    model_config = ConfigDict(arbitrary_types_allowed=True)
 
-        Args:
-            memory: The underlying Memory instance.
-            scopes: List of scope paths to include.
-            categories: Optional category filter for recall.
-            read_only: If True, remember() is a silent no-op.
-        """
-        self._memory = memory
-        self._scopes = [s.rstrip("/") or "/" for s in scopes]
-        self._categories = categories
-        self._read_only = read_only
+    scopes: list[str] = Field(default_factory=list)
+    categories: list[str] | None = Field(default=None)
+    read_only: bool = Field(default=True)
+
+    _memory: Memory = PrivateAttr()
+
+    @model_validator(mode="wrap")
+    @classmethod
+    def _accept_memory(cls, data: Any, handler: Any) -> MemorySlice:
+        """Extract memory dependency and normalize scopes before validation."""
+        if isinstance(data, MemorySlice):
+            return data
+        memory = data.pop("memory")
+        data["scopes"] = [s.rstrip("/") or "/" for s in data.get("scopes", [])]
+        instance: MemorySlice = handler(data)
+        instance._memory = memory
+        return instance
 
     def remember(
         self,
@@ -178,7 +192,7 @@ class MemorySlice:
         private: bool = False,
     ) -> MemoryRecord | None:
         """Remember into an explicit scope. No-op when read_only=True."""
-        if self._read_only:
+        if self.read_only:
             return None
         return self._memory.remember(
             content,
@@ -196,14 +210,14 @@ class MemorySlice:
         scope: str | None = None,
         categories: list[str] | None = None,
         limit: int = 10,
-        depth: str = "deep",
+        depth: Literal["shallow", "deep"] = "deep",
         source: str | None = None,
         include_private: bool = False,
     ) -> list[MemoryMatch]:
         """Recall across all slice scopes; results merged and re-ranked."""
-        cats = categories or self._categories
+        cats = categories or self.categories
         all_matches: list[MemoryMatch] = []
-        for sc in self._scopes:
+        for sc in self.scopes:
             matches = self._memory.recall(
                 query,
                 scope=sc,
@@ -231,7 +245,7 @@ class MemorySlice:
     def list_scopes(self, path: str = "/") -> list[str]:
         """List scopes across all slice roots."""
         out: list[str] = []
-        for sc in self._scopes:
+        for sc in self.scopes:
             full = f"{sc.rstrip('/')}{path}" if sc != "/" else path
             out.extend(self._memory.list_scopes(full))
         return sorted(set(out))
@@ -243,15 +257,23 @@ class MemorySlice:
         oldest: datetime | None = None
         newest: datetime | None = None
         children: list[str] = []
-        for sc in self._scopes:
+        for sc in self.scopes:
             full = f"{sc.rstrip('/')}{path}" if sc != "/" else path
             inf = self._memory.info(full)
             total_records += inf.record_count
             all_categories.update(inf.categories)
             if inf.oldest_record:
-                oldest = inf.oldest_record if oldest is None else min(oldest, inf.oldest_record)
+                oldest = (
+                    inf.oldest_record
+                    if oldest is None
+                    else min(oldest, inf.oldest_record)
+                )
             if inf.newest_record:
-                newest = inf.newest_record if newest is None else max(newest, inf.newest_record)
+                newest = (
+                    inf.newest_record
+                    if newest is None
+                    else max(newest, inf.newest_record)
+                )
             children.extend(inf.child_scopes)
         return ScopeInfo(
             path=path,
@@ -265,7 +287,7 @@ class MemorySlice:
     def list_categories(self, path: str | None = None) -> dict[str, int]:
         """Categories and counts across slice scopes."""
         counts: dict[str, int] = {}
-        for sc in self._scopes:
+        for sc in self.scopes:
             full = (f"{sc.rstrip('/')}{path}" if sc != "/" else path) if path else sc
             for k, v in self._memory.list_categories(full).items():
                 counts[k] = counts.get(k, 0) + v

lib/crewai/src/crewai/memory/unified_memory.py

@@ -6,7 +6,9 @@ from concurrent.futures import Future, ThreadPoolExecutor
 from datetime import datetime
 import threading
 import time
-from typing import TYPE_CHECKING, Any, Literal
+from typing import TYPE_CHECKING, Annotated, Any, Literal
+
+from pydantic import BaseModel, ConfigDict, Field, PlainValidator, PrivateAttr
 
 from crewai.events.event_bus import crewai_event_bus
 from crewai.events.types.memory_events import (
@@ -39,13 +41,18 @@ if TYPE_CHECKING:
     )
 
 
+def _passthrough(v: Any) -> Any:
+    """PlainValidator that accepts any value, bypassing strict union discrimination."""
+    return v
+
+
 def _default_embedder() -> OpenAIEmbeddingFunction:
     """Build default OpenAI embedder for memory."""
     spec: OpenAIProviderSpec = {"provider": "openai", "config": {}}
     return build_embedder(spec)
 
 
-class Memory:
+class Memory(BaseModel):
     """Unified memory: standalone, LLM-analyzed, with intelligent recall flow.
 
     Works without agent/crew. Uses LLM to infer scope, categories, importance on save.
@@ -53,116 +60,119 @@ class Memory:
     pluggable storage (LanceDB default).
     """
 
-    def __init__(
-        self,
-        llm: BaseLLM | str = "gpt-4o-mini",
-        storage: StorageBackend | str = "lancedb",
-        embedder: Any = None,
-        # -- Scoring weights --
-        # These three weights control how recall results are ranked.
-        # The composite score is: semantic_weight * similarity + recency_weight * decay + importance_weight * importance.
-        # They should sum to ~1.0 for intuitive scoring.
-        recency_weight: float = 0.3,
-        semantic_weight: float = 0.5,
-        importance_weight: float = 0.2,
-        # How quickly old memories lose relevance. The recency score halves every
-        # N days (exponential decay). Lower = faster forgetting; higher = longer relevance.
-        recency_half_life_days: int = 30,
-        # -- Consolidation --
-        # When remembering new content, if an existing record has similarity >= this
-        # threshold, the LLM is asked to merge/update/delete. Set to 1.0 to disable.
-        consolidation_threshold: float = 0.85,
-        # Max existing records to compare against when checking for consolidation.
-        consolidation_limit: int = 5,
-        # -- Save defaults --
-        # Importance assigned to new memories when no explicit value is given and
-        # the LLM analysis path is skipped (all fields provided by the caller).
-        default_importance: float = 0.5,
-        # -- Recall depth control --
-        # These thresholds govern the RecallFlow router that decides between
-        # returning results immediately ("synthesize") vs. doing an extra
-        # LLM-driven exploration round ("explore_deeper").
-        #   confidence >= confidence_threshold_high  => always synthesize
-        #   confidence <  confidence_threshold_low   => explore deeper (if budget > 0)
-        #   complex query + confidence < complex_query_threshold => explore deeper
-        confidence_threshold_high: float = 0.8,
-        confidence_threshold_low: float = 0.5,
-        complex_query_threshold: float = 0.7,
-        # How many LLM-driven exploration rounds the RecallFlow is allowed to run.
-        # 0 = always shallow (vector search only); higher = more thorough but slower.
-        exploration_budget: int = 1,
-        # Queries shorter than this skip LLM analysis (saving ~1-3s).
-        # Longer queries (full task descriptions) benefit from LLM distillation.
-        query_analysis_threshold: int = 200,
-        # When True, all write operations (remember, remember_many) are silently
-        # skipped. Useful for sharing a read-only view of memory across agents
-        # without any of them persisting new memories.
-        read_only: bool = False,
-    ) -> None:
-        """Initialize Memory.
+    model_config = ConfigDict(arbitrary_types_allowed=True)
 
-        Args:
-            llm: LLM for analysis (model name or BaseLLM instance).
-            storage: Backend: "lancedb" or a StorageBackend instance.
-            embedder: Embedding callable, provider config dict, or None (default OpenAI).
-            recency_weight: Weight for recency in the composite relevance score.
-            semantic_weight: Weight for semantic similarity in the composite relevance score.
-            importance_weight: Weight for importance in the composite relevance score.
-            recency_half_life_days: Recency score halves every N days (exponential decay).
-            consolidation_threshold: Similarity above which consolidation is triggered on save.
-            consolidation_limit: Max existing records to compare during consolidation.
-            default_importance: Default importance when not provided or inferred.
-            confidence_threshold_high: Recall confidence above which results are returned directly.
-            confidence_threshold_low: Recall confidence below which deeper exploration is triggered.
-            complex_query_threshold: For complex queries, explore deeper below this confidence.
-            exploration_budget: Number of LLM-driven exploration rounds during deep recall.
-            query_analysis_threshold: Queries shorter than this skip LLM analysis during deep recall.
-            read_only: If True, remember() and remember_many() are silent no-ops.
-        """
-        self._read_only = read_only
+    llm: Annotated[BaseLLM | str, PlainValidator(_passthrough)] = Field(
+        default="gpt-4o-mini",
+        description="LLM for analysis (model name or BaseLLM instance).",
+    )
+    storage: Annotated[StorageBackend | str, PlainValidator(_passthrough)] = Field(
+        default="lancedb",
+        description="Storage backend instance or path string.",
+    )
+    embedder: Any = Field(
+        default=None,
+        description="Embedding callable, provider config dict, or None for default OpenAI.",
+    )
+    recency_weight: float = Field(
+        default=0.3,
+        description="Weight for recency in the composite relevance score.",
+    )
+    semantic_weight: float = Field(
+        default=0.5,
+        description="Weight for semantic similarity in the composite relevance score.",
+    )
+    importance_weight: float = Field(
+        default=0.2,
+        description="Weight for importance in the composite relevance score.",
+    )
+    recency_half_life_days: int = Field(
+        default=30,
+        description="Recency score halves every N days (exponential decay).",
+    )
+    consolidation_threshold: float = Field(
+        default=0.85,
+        description="Similarity above which consolidation is triggered on save.",
+    )
+    consolidation_limit: int = Field(
+        default=5,
+        description="Max existing records to compare during consolidation.",
+    )
+    default_importance: float = Field(
+        default=0.5,
+        description="Default importance when not provided or inferred.",
+    )
+    confidence_threshold_high: float = Field(
+        default=0.8,
+        description="Recall confidence above which results are returned directly.",
+    )
+    confidence_threshold_low: float = Field(
+        default=0.5,
+        description="Recall confidence below which deeper exploration is triggered.",
+    )
+    complex_query_threshold: float = Field(
+        default=0.7,
+        description="For complex queries, explore deeper below this confidence.",
+    )
+    exploration_budget: int = Field(
+        default=1,
+        description="Number of LLM-driven exploration rounds during deep recall.",
+    )
+    query_analysis_threshold: int = Field(
+        default=200,
+        description="Queries shorter than this skip LLM analysis during deep recall.",
+    )
+    read_only: bool = Field(
+        default=False,
+        description="If True, remember() and remember_many() are silent no-ops.",
+    )
+
+    _config: MemoryConfig = PrivateAttr()
+    _llm_instance: BaseLLM | None = PrivateAttr(default=None)
+    _embedder_instance: Any = PrivateAttr(default=None)
+    _storage: StorageBackend = PrivateAttr()
+    _save_pool: ThreadPoolExecutor = PrivateAttr(
+        default_factory=lambda: ThreadPoolExecutor(
+            max_workers=1, thread_name_prefix="memory-save"
+        )
+    )
+    _pending_saves: list[Future[Any]] = PrivateAttr(default_factory=list)
+    _pending_lock: threading.Lock = PrivateAttr(default_factory=threading.Lock)
+
+    def model_post_init(self, __context: Any) -> None:
+        """Initialize runtime state from field values."""
         self._config = MemoryConfig(
-            recency_weight=recency_weight,
-            semantic_weight=semantic_weight,
-            importance_weight=importance_weight,
-            recency_half_life_days=recency_half_life_days,
-            consolidation_threshold=consolidation_threshold,
-            consolidation_limit=consolidation_limit,
-            default_importance=default_importance,
-            confidence_threshold_high=confidence_threshold_high,
-            confidence_threshold_low=confidence_threshold_low,
-            complex_query_threshold=complex_query_threshold,
-            exploration_budget=exploration_budget,
-            query_analysis_threshold=query_analysis_threshold,
+            recency_weight=self.recency_weight,
+            semantic_weight=self.semantic_weight,
+            importance_weight=self.importance_weight,
+            recency_half_life_days=self.recency_half_life_days,
+            consolidation_threshold=self.consolidation_threshold,
+            consolidation_limit=self.consolidation_limit,
+            default_importance=self.default_importance,
+            confidence_threshold_high=self.confidence_threshold_high,
+            confidence_threshold_low=self.confidence_threshold_low,
+            complex_query_threshold=self.complex_query_threshold,
+            exploration_budget=self.exploration_budget,
+            query_analysis_threshold=self.query_analysis_threshold,
         )
 
-        # Store raw config for lazy initialization. LLM and embedder are only
-        # built on first access so that Memory() never fails at construction
-        # time (e.g. when auto-created by Flow without an API key set).
-        self._llm_config: BaseLLM | str = llm
-        self._llm_instance: BaseLLM | None = None if isinstance(llm, str) else llm
-        self._embedder_config: Any = embedder
-        self._embedder_instance: Any = (
-            embedder
-            if (embedder is not None and not isinstance(embedder, dict))
+        self._llm_instance = None if isinstance(self.llm, str) else self.llm
+        self._embedder_instance = (
+            self.embedder
+            if (self.embedder is not None and not isinstance(self.embedder, dict))
             else None
         )
 
-        if isinstance(storage, str):
+        if isinstance(self.storage, str):
             from crewai.memory.storage.lancedb_storage import LanceDBStorage
 
-            self._storage = LanceDBStorage() if storage == "lancedb" else LanceDBStorage(path=storage)
+            self._storage = (
+                LanceDBStorage()
+                if self.storage == "lancedb"
+                else LanceDBStorage(path=self.storage)
+            )
         else:
-            self._storage = storage
-
-        # Background save queue. max_workers=1 serializes saves to avoid
-        # concurrent storage mutations (two saves finding the same similar
-        # record and both trying to update/delete it). Within each save,
-        # the parallel LLM calls still run on their own thread pool.
-        self._save_pool = ThreadPoolExecutor(
-            max_workers=1, thread_name_prefix="memory-save"
-        )
-        self._pending_saves: list[Future[Any]] = []
-        self._pending_lock = threading.Lock()
+            self._storage = self.storage
 
     _MEMORY_DOCS_URL = "https://docs.crewai.com/concepts/memory"
 
@@ -173,11 +183,7 @@ class Memory:
             from crewai.llm import LLM
 
             try:
-                model_name = (
-                    self._llm_config
-                    if isinstance(self._llm_config, str)
-                    else str(self._llm_config)
-                )
+                model_name = self.llm if isinstance(self.llm, str) else str(self.llm)
                 self._llm_instance = LLM(model=model_name)
             except Exception as e:
                 raise RuntimeError(
@@ -197,8 +203,8 @@ class Memory:
         """Lazy embedder initialization -- only created when first needed."""
         if self._embedder_instance is None:
             try:
-                if isinstance(self._embedder_config, dict):
-                    self._embedder_instance = build_embedder(self._embedder_config)
+                if isinstance(self.embedder, dict):
+                    self._embedder_instance = build_embedder(self.embedder)
                 else:
                     self._embedder_instance = _default_embedder()
             except Exception as e:
@@ -356,7 +362,7 @@ class Memory:
         Raises:
             Exception: On save failure (events emitted).
         """
-        if self._read_only:
+        if self.read_only:
             return None
         _source_type = "unified_memory"
         try:
@@ -444,7 +450,7 @@ class Memory:
         Returns:
             Empty list (records are not available until the background save completes).
         """
-        if not contents or self._read_only:
+        if not contents or self.read_only:
             return []
 
         self._submit_save(

lib/crewai/src/crewai/tools/memory_tools.py

@@ -121,7 +121,7 @@ def create_memory_tools(memory: Any) -> list[BaseTool]:
             description=i18n.tools("recall_memory"),
         ),
     ]
-    if not getattr(memory, "_read_only", False):
+    if not memory.read_only:
         tools.append(
             RememberTool(
                 memory=memory,

lib/crewai/tests/agents/test_lite_agent.py

@@ -1136,7 +1136,7 @@ def test_lite_agent_memory_instance_recall_and_save_called():
         successful_requests=1,
     )
     mock_memory = Mock()
-    mock_memory._read_only = False
+    mock_memory.read_only = False
     mock_memory.recall.return_value = []
     mock_memory.extract_memories.return_value = ["Fact one.", "Fact two."]
 

lib/crewai/tests/memory/test_unified_memory.py

@@ -172,8 +172,8 @@ def test_memory_scope_slice(tmp_path: Path, mock_embedder: MagicMock) -> None:
     sc = mem.scope("/agent/1")
     assert sc._root in ("/agent/1", "/agent/1/")
     sl = mem.slice(["/a", "/b"], read_only=True)
-    assert sl._read_only is True
-    assert "/a" in sl._scopes and "/b" in sl._scopes
+    assert sl.read_only is True
+    assert "/a" in sl.scopes and "/b" in sl.scopes
 
 
 def test_memory_list_scopes_info_tree(tmp_path: Path, mock_embedder: MagicMock) -> None:
@@ -198,7 +198,7 @@ def test_memory_scope_remember_recall(tmp_path: Path, mock_embedder: MagicMock)
     from crewai.memory.memory_scope import MemoryScope
 
     mem = Memory(storage=str(tmp_path / "db5"), llm=MagicMock(), embedder=mock_embedder)
-    scope = MemoryScope(mem, "/crew/1")
+    scope = MemoryScope(memory=mem, root_path="/crew/1")
     scope.remember("Scoped note", scope="/", categories=[], importance=0.5, metadata={})
     results = scope.recall("note", limit=5, depth="shallow")
     assert len(results) >= 1
@@ -213,7 +213,7 @@ def test_memory_slice_recall(tmp_path: Path, mock_embedder: MagicMock) -> None:
 
     mem = Memory(storage=str(tmp_path / "db6"), llm=MagicMock(), embedder=mock_embedder)
     mem.remember("In scope A", scope="/a", categories=[], importance=0.5, metadata={})
-    sl = MemorySlice(mem, ["/a"], read_only=True)
+    sl = MemorySlice(memory=mem, scopes=["/a"], read_only=True)
     matches = sl.recall("scope", limit=5, depth="shallow")
     assert isinstance(matches, list)
 
@@ -223,7 +223,7 @@ def test_memory_slice_remember_is_noop_when_read_only(tmp_path: Path, mock_embed
     from crewai.memory.memory_scope import MemorySlice
 
     mem = Memory(storage=str(tmp_path / "db7"), llm=MagicMock(), embedder=mock_embedder)
-    sl = MemorySlice(mem, ["/a"], read_only=True)
+    sl = MemorySlice(memory=mem, scopes=["/a"], read_only=True)
     result = sl.remember("x", scope="/a")
     assert result is None
     assert mem.list_records() == []
@@ -319,7 +319,7 @@ def test_executor_save_to_memory_calls_extract_then_remember_per_item() -> None:
     from crewai.agents.parser import AgentFinish
 
     mock_memory = MagicMock()
-    mock_memory._read_only = False
+    mock_memory.read_only = False
     mock_memory.extract_memories.return_value = ["Fact A.", "Fact B."]
 
     mock_agent = MagicMock()
@@ -360,7 +360,7 @@ def test_executor_save_to_memory_skips_delegation_output() -> None:
     from crewai.utilities.string_utils import sanitize_tool_name
 
     mock_memory = MagicMock()
-    mock_memory._read_only = False
+    mock_memory.read_only = False
     mock_agent = MagicMock()
     mock_agent.memory = mock_memory
     mock_agent._logger = MagicMock()
@@ -393,7 +393,7 @@ def test_memory_scope_extract_memories_delegates() -> None:
 
     mock_memory = MagicMock()
     mock_memory.extract_memories.return_value = ["Scoped fact."]
-    scope = MemoryScope(mock_memory, "/agent/1")
+    scope = MemoryScope(memory=mock_memory, root_path="/agent/1")
     result = scope.extract_memories("Some content")
     mock_memory.extract_memories.assert_called_once_with("Some content")
     assert result == ["Scoped fact."]
@@ -405,7 +405,7 @@ def test_memory_slice_extract_memories_delegates() -> None:
 
     mock_memory = MagicMock()
     mock_memory.extract_memories.return_value = ["Sliced fact."]
-    sl = MemorySlice(mock_memory, ["/a", "/b"], read_only=True)
+    sl = MemorySlice(memory=mock_memory, scopes=["/a", "/b"], read_only=True)
     result = sl.extract_memories("Some content")
     mock_memory.extract_memories.assert_called_once_with("Some content")
     assert result == ["Sliced fact."]
@@ -670,10 +670,10 @@ def test_agent_kickoff_memory_recall_and_save(tmp_path: Path, mock_embedder: Mag
         verbose=False,
     )
 
-    # Mock recall to verify it's called, but return real results
-    with patch.object(mem, "recall", wraps=mem.recall) as recall_mock, \
-         patch.object(mem, "extract_memories", return_value=["PostgreSQL is used."]) as extract_mock, \
-         patch.object(mem, "remember_many", wraps=mem.remember_many) as remember_many_mock:
+    # Patch on the class to avoid Pydantic BaseModel __delattr__ restriction
+    with patch.object(Memory, "recall", wraps=mem.recall) as recall_mock, \
+         patch.object(Memory, "extract_memories", return_value=["PostgreSQL is used."]) as extract_mock, \
+         patch.object(Memory, "remember_many", wraps=mem.remember_many) as remember_many_mock:
         result = agent.kickoff("What database do we use?")
 
     assert result is not None

lib/crewai/tests/test_crew.py

@@ -36,7 +36,7 @@ from crewai.flow import Flow, start
 from crewai.knowledge.knowledge import Knowledge
 from crewai.knowledge.source.string_knowledge_source import StringKnowledgeSource
 from crewai.llm import LLM
-
+from crewai.memory.unified_memory import Memory
 from crewai.process import Process
 from crewai.project import CrewBase, agent, before_kickoff, crew, task
 from crewai.task import Task
@@ -2618,9 +2618,9 @@ def test_memory_remember_called_after_task():
     )
 
     with patch.object(
-        crew._memory, "extract_memories", wraps=crew._memory.extract_memories
+        Memory, "extract_memories", wraps=crew._memory.extract_memories
     ) as extract_mock, patch.object(
-        crew._memory, "remember", wraps=crew._memory.remember
+        Memory, "remember", wraps=crew._memory.remember
     ) as remember_mock:
         crew.kickoff()
 
@@ -4773,13 +4773,13 @@ def test_memory_remember_receives_task_content():
         # Mock extract_memories to return fake memories and capture the raw input.
         # No wraps= needed -- the test only checks what args it receives, not the output.
         patch.object(
-            crew._memory, "extract_memories", return_value=["Fake memory."]
+            Memory, "extract_memories", return_value=["Fake memory."]
         ) as extract_mock,
         # Mock recall to avoid LLM calls for query analysis (not in cassette).
-        patch.object(crew._memory, "recall", return_value=[]),
+        patch.object(Memory, "recall", return_value=[]),
         # Mock remember_many to prevent the background save from triggering
         # LLM calls (field resolution) that aren't in the cassette.
-        patch.object(crew._memory, "remember_many", return_value=[]),
+        patch.object(Memory, "remember_many", return_value=[]),
     ):
         crew.kickoff()
 

lib/crewai/tests/test_llm.py

@@ -614,6 +614,11 @@ def test_handle_streaming_tool_calls_with_error(get_weather_tool_schema, mock_em
 def test_handle_streaming_tool_calls_no_available_functions(
     get_weather_tool_schema, mock_emit
 ):
+    """When tools are provided but available_functions is not (defaults to None),
+    the streaming handler should return the accumulated tool calls so the caller
+    (e.g., CrewAgentExecutor) can handle them. This is the fix for issue #4788
+    where tool calls were previously discarded and an empty string was returned.
+    """
     llm = LLM(model="openai/gpt-4o", stream=True, is_litellm=True)
     response = llm.call(
         messages=[
@@ -621,7 +626,14 @@ def test_handle_streaming_tool_calls_no_available_functions(
         ],
         tools=[get_weather_tool_schema],
     )
-    assert response == ""
+    # With the fix for #4788, tool calls should be returned as a list
+    # instead of being discarded (previously returned "")
+    assert isinstance(response, list), (
+        f"Expected list of tool calls but got {type(response)}: {response}"
+    )
+    assert len(response) == 1
+    assert response[0].function.name == "get_weather"
+    assert response[0].function.arguments == '{"location":"New York, NY"}'
 
     assert_event_count(
         mock_emit=mock_emit,
@@ -1022,3 +1034,166 @@ async def test_usage_info_streaming_with_acall():
     assert llm._token_usage["total_tokens"] > 0
 
     assert len(result) > 0
+
+
+def test_non_streaming_tool_calls_returned_when_no_available_functions():
+    """Test that tool calls are returned (not text) when available_functions is None.
+
+    This reproduces the bug from issue #4788 where LLMs like Anthropic return both
+    text content AND tool calls in the same response. When available_functions=None
+    (as used by the executor for native tool handling), tool calls should be returned
+    instead of the text content.
+    """
+    from litellm.types.utils import ChatCompletionMessageToolCall, Function
+
+    llm = LLM(model="gpt-4o-mini", is_litellm=True)
+
+    # Mock a response that has BOTH text content AND tool calls
+    mock_tool_call = ChatCompletionMessageToolCall(
+        id="call_123",
+        type="function",
+        function=Function(
+            name="code_search",
+            arguments='{"query": "test query"}',
+        ),
+    )
+    mock_message = MagicMock()
+    mock_message.content = "I will search for the given query."
+    mock_message.tool_calls = [mock_tool_call]
+
+    mock_choice = MagicMock()
+    mock_choice.message = mock_message
+
+    mock_response = MagicMock()
+    mock_response.choices = [mock_choice]
+    mock_response.usage = MagicMock()
+    mock_response.usage.prompt_tokens = 10
+    mock_response.usage.completion_tokens = 5
+    mock_response.usage.total_tokens = 15
+
+    with patch("litellm.completion", return_value=mock_response):
+        # Call WITHOUT available_functions (as the executor does for native tool handling)
+        result = llm.call(
+            messages=[{"role": "user", "content": "Search for something"}],
+            tools=[{"type": "function", "function": {"name": "code_search"}}],
+            available_functions=None,
+        )
+
+    # Result should be the tool calls list, NOT the text response
+    assert isinstance(result, list), (
+        f"Expected list of tool calls but got {type(result)}: {result}"
+    )
+    assert len(result) == 1
+    assert result[0].function.name == "code_search"
+
+
+def test_non_streaming_text_returned_when_no_tool_calls():
+    """Test that text response is still returned when there are no tool calls."""
+    llm = LLM(model="gpt-4o-mini", is_litellm=True)
+
+    mock_message = MagicMock()
+    mock_message.content = "The capital of France is Paris."
+    mock_message.tool_calls = None
+
+    mock_choice = MagicMock()
+    mock_choice.message = mock_message
+
+    mock_response = MagicMock()
+    mock_response.choices = [mock_choice]
+    mock_response.usage = MagicMock()
+    mock_response.usage.prompt_tokens = 10
+    mock_response.usage.completion_tokens = 5
+    mock_response.usage.total_tokens = 15
+
+    with patch("litellm.completion", return_value=mock_response):
+        result = llm.call(
+            messages=[{"role": "user", "content": "What is the capital of France?"}],
+        )
+
+    assert isinstance(result, str)
+    assert result == "The capital of France is Paris."
+
+
+@pytest.mark.asyncio
+async def test_async_non_streaming_tool_calls_returned_when_no_available_functions():
+    """Test async path: tool calls are returned (not text) when available_functions is None.
+
+    Same bug as #4788 but for the async non-streaming handler.
+    """
+    from litellm.types.utils import ChatCompletionMessageToolCall, Function
+
+    llm = LLM(model="gpt-4o-mini", is_litellm=True, stream=False)
+
+    mock_tool_call = ChatCompletionMessageToolCall(
+        id="call_456",
+        type="function",
+        function=Function(
+            name="web_search",
+            arguments='{"query": "test"}',
+        ),
+    )
+    mock_message = MagicMock()
+    mock_message.content = "I will search the web."
+    mock_message.tool_calls = [mock_tool_call]
+
+    mock_choice = MagicMock()
+    mock_choice.message = mock_message
+
+    mock_response = MagicMock()
+    mock_response.choices = [mock_choice]
+    mock_response.usage = MagicMock()
+    mock_response.usage.prompt_tokens = 10
+    mock_response.usage.completion_tokens = 5
+    mock_response.usage.total_tokens = 15
+
+    with patch("litellm.acompletion", return_value=mock_response):
+        result = await llm.acall(
+            messages=[{"role": "user", "content": "Search for something"}],
+            tools=[{"type": "function", "function": {"name": "web_search"}}],
+            available_functions=None,
+        )
+
+    assert isinstance(result, list), (
+        f"Expected list of tool calls but got {type(result)}: {result}"
+    )
+    assert len(result) == 1
+    assert result[0].function.name == "web_search"
+
+
+def test_non_streaming_tool_calls_executed_when_available_functions_provided():
+    """Test that tool calls are still executed when available_functions IS provided.
+
+    This ensures the fix doesn't break the normal tool execution path.
+    """
+    llm = LLM(model="gpt-4o-mini", is_litellm=True)
+
+    mock_tool_call = MagicMock()
+    mock_tool_call.function.name = "get_weather"
+    mock_tool_call.function.arguments = '{"location": "New York"}'
+
+    mock_message = MagicMock()
+    mock_message.content = "I will check the weather."
+    mock_message.tool_calls = [mock_tool_call]
+
+    mock_choice = MagicMock()
+    mock_choice.message = mock_message
+
+    mock_response = MagicMock()
+    mock_response.choices = [mock_choice]
+    mock_response.usage = MagicMock()
+    mock_response.usage.prompt_tokens = 10
+    mock_response.usage.completion_tokens = 5
+    mock_response.usage.total_tokens = 15
+
+    def get_weather(location: str) -> str:
+        return f"Sunny in {location}"
+
+    with patch("litellm.completion", return_value=mock_response):
+        result = llm.call(
+            messages=[{"role": "user", "content": "What's the weather?"}],
+            tools=[{"type": "function", "function": {"name": "get_weather"}}],
+            available_functions={"get_weather": get_weather},
+        )
+
+    # When available_functions is provided, the tool should be executed
+    assert result == "Sunny in New York"

uv.lock

@@ -1426,7 +1426,7 @@ requires-dist = [
     { name = "docker", specifier = "~=7.1.0" },
     { name = "exa-py", marker = "extra == 'exa-py'", specifier = ">=1.8.7" },
     { name = "firecrawl-py", marker = "extra == 'firecrawl-py'", specifier = ">=1.8.0" },
-    { name = "gitpython", marker = "extra == 'github'", specifier = "==3.1.38" },
+    { name = "gitpython", marker = "extra == 'github'", specifier = ">=3.1.41,<4" },
     { name = "hyperbrowser", marker = "extra == 'hyperbrowser'", specifier = ">=0.18.0" },
     { name = "langchain-apify", marker = "extra == 'apify'", specifier = ">=0.1.2,<1.0.0" },
     { name = "linkup-sdk", marker = "extra == 'linkup-sdk'", specifier = ">=0.2.2" },
@@ -2201,14 +2201,14 @@ wheels = [
 
 [[package]]
 name = "gitpython"
-version = "3.1.38"
+version = "3.1.46"
 source = { registry = "https://pypi.org/simple" }
 dependencies = [
     { name = "gitdb" },
 ]
-sdist = { url = "https://files.pythonhosted.org/packages/b3/45/cee7af549b6fa33f04531e402693a772b776cd9f845a2cbeca99cfac3331/GitPython-3.1.38.tar.gz", hash = "sha256:4d683e8957c8998b58ddb937e3e6cd167215a180e1ffd4da769ab81c620a89fe", size = 200632, upload-time = "2023-10-17T06:09:52.235Z" }
+sdist = { url = "https://files.pythonhosted.org/packages/df/b5/59d16470a1f0dfe8c793f9ef56fd3826093fc52b3bd96d6b9d6c26c7e27b/gitpython-3.1.46.tar.gz", hash = "sha256:400124c7d0ef4ea03f7310ac2fbf7151e09ff97f2a3288d64a440c584a29c37f", size = 215371, upload-time = "2026-01-01T15:37:32.073Z" }
 wheels = [
-    { url = "https://files.pythonhosted.org/packages/3c/ae/044453eacd5a526d3f242ccd77e38ee8219c65e0b132562b551bd67c61a4/GitPython-3.1.38-py3-none-any.whl", hash = "sha256:9e98b672ffcb081c2c8d5aa630d4251544fb040fb158863054242f24a2a2ba30", size = 190573, upload-time = "2023-10-17T06:09:50.18Z" },
+    { url = "https://files.pythonhosted.org/packages/6a/09/e21df6aef1e1ffc0c816f0522ddc3f6dcded766c3261813131c78a704470/gitpython-3.1.46-py3-none-any.whl", hash = "sha256:79812ed143d9d25b6d176a10bb511de0f9c67b1fa641d82097b0ab90398a2058", size = 208620, upload-time = "2026-01-01T15:37:30.574Z" },
 ]
 
 [[package]]