Merge branch 'main' into fix/mcp-allowed-tools-empty-list-handling

python/packages/core/AGENTS.md

@@ -56,7 +56,7 @@ agent_framework/
 - **`AgentMiddleware`** - Intercepts agent `run()` calls
 - **`ChatMiddleware`** - Intercepts chat client `get_response()` calls
 - **`FunctionMiddleware`** - Intercepts function/tool invocations
-- **`AgentContext`** / **`ChatContext`** / **`FunctionInvocationContext`** - Context objects passed through middleware
+- **`AgentContext`** / **`ChatContext`** / **`FunctionInvocationContext`** - Context objects passed through middleware. A tool can declare a `FunctionInvocationContext` parameter to receive it; `context.tools` is the live, mutable tools list for the run, and `context.add_tools(...)` / `context.remove_tools(...)` enable progressive tool exposure (changes apply on the next function-calling iteration).
 
 ### Sessions (`_sessions.py`)
 

python/packages/core/agent_framework/_feature_stage.py

@@ -58,6 +58,7 @@ class ExperimentalFeature(str, Enum):
     FOUNDRY_PREVIEW_TOOLS = "FOUNDRY_PREVIEW_TOOLS"
     FUNCTIONAL_WORKFLOWS = "FUNCTIONAL_WORKFLOWS"
     HARNESS = "HARNESS"
+    PROGRESSIVE_TOOLS = "PROGRESSIVE_TOOLS"
     SKILLS = "SKILLS"
     TO_PROMPT_AGENT = "TO_PROMPT_AGENT"
 

python/packages/core/agent_framework/_middleware.py

@@ -11,6 +11,7 @@ from enum import Enum
 from typing import TYPE_CHECKING, Any, Generic, Literal, TypeAlias, cast, overload
 
 from ._clients import SupportsChatGetResponse
+from ._feature_stage import ExperimentalFeature, experimental
 from ._types import (
     AgentResponse,
     AgentResponseUpdate,
@@ -214,6 +215,12 @@ class FunctionInvocationContext:
         result: Function execution result. Can be observed after calling ``call_next()``
                 to see the actual execution result or can be set to override the execution result.
         kwargs: Additional runtime keyword arguments forwarded to the function invocation.
+        tools: The live, mutable list of tools available to the model for the current
+                agent run, or ``None`` when the function is invoked outside of a
+                function-calling loop (for example via ``FunctionTool.invoke`` directly).
+                Tools can add or remove tools during execution using :meth:`add_tools`
+                and :meth:`remove_tools` (progressive tool exposure). Mutations take
+                effect on the **next** model iteration, not the in-flight batch.
 
     Examples:
         .. code-block:: python
@@ -232,6 +239,18 @@ class FunctionInvocationContext:
 
                     # Continue execution
                     await call_next()
+
+        Progressive tool exposure from inside a tool:
+
+        .. code-block:: python
+
+            from agent_framework import FunctionInvocationContext, tool
+
+
+            @tool(approval_mode="never_require")
+            def load_math_tools(ctx: FunctionInvocationContext) -> str:
+                ctx.add_tools([factorial, fibonacci])
+                return "Math tools are now available."
     """
 
     def __init__(
@@ -242,6 +261,7 @@ class FunctionInvocationContext:
         metadata: Mapping[str, Any] | None = None,
         result: Any = None,
         kwargs: Mapping[str, Any] | None = None,
+        tools: list[ToolTypes] | None = None,
     ) -> None:
         """Initialize the FunctionInvocationContext.
 
@@ -252,6 +272,9 @@ class FunctionInvocationContext:
             metadata: Metadata dictionary for sharing data between function middleware.
             result: Function execution result.
             kwargs: Additional runtime keyword arguments forwarded to the function invocation.
+            tools: The live, mutable list of tools for the current agent run. When provided,
+                this is the same list object the model sees on the next iteration, so
+                appending or removing tools changes the model's available tools.
         """
         self.function = function
         self.arguments = arguments
@@ -259,6 +282,96 @@ class FunctionInvocationContext:
         self.metadata: dict[str, Any] = dict(metadata) if metadata is not None else {}
         self.result = result
         self.kwargs: dict[str, Any] = dict(kwargs) if kwargs is not None else {}
+        self.tools = tools
+
+    @experimental(feature_id=ExperimentalFeature.PROGRESSIVE_TOOLS)
+    def add_tools(
+        self,
+        tools: ToolTypes | Callable[..., Any] | Sequence[ToolTypes | Callable[..., Any]],
+    ) -> None:
+        """Add one or more tools to the current agent run (progressive tool exposure).
+
+        Callable inputs are converted to :class:`FunctionTool`, and tool collections are
+        flattened, using the same normalization as the rest of the framework. Added tools
+        become available to the model on the **next** iteration of the function-calling
+        loop; they do not affect tool calls already requested in the in-flight batch.
+
+        Adding a tool whose name already exists is a no-op when it is the same object, and
+        raises ``ValueError`` when it is a different object with a duplicate name.
+
+        Args:
+            tools: A single tool/callable or a sequence of tools/callables to add.
+
+        Raises:
+            RuntimeError: If the context has no live tools list (for example when the
+                function is invoked outside of a function-calling loop).
+            ValueError: If a different tool with a duplicate name is added.
+        """
+        from ._tools import _append_unique_tools, normalize_tools  # type: ignore[reportPrivateUsage]
+
+        if self.tools is None:
+            raise RuntimeError(
+                "Cannot add tools: this FunctionInvocationContext is not bound to a live "
+                "agent run. add_tools is only available for functions invoked within an "
+                "agent's function-calling loop."
+            )
+        # Validate the whole batch against a throwaway copy first, so a duplicate-name
+        # clash partway through the batch raises before the live tool list is mutated
+        # (all-or-nothing semantics).
+        merged = _append_unique_tools(list(self.tools), normalize_tools(tools))
+        self.tools[:] = merged
+
+    @experimental(feature_id=ExperimentalFeature.PROGRESSIVE_TOOLS)
+    def remove_tools(
+        self,
+        tools: ToolTypes | Callable[..., Any] | Sequence[ToolTypes | Callable[..., Any]] | str | Sequence[str],
+    ) -> None:
+        """Remove one or more tools from the current agent run (progressive tool exposure).
+
+        Tools may be specified by name, by tool object, or by the original callable. Names
+        that are not currently present are ignored. Removals take effect on the **next**
+        iteration of the function-calling loop; tool calls already requested in the
+        in-flight batch still execute.
+
+        Args:
+            tools: A tool name, tool/callable, or a sequence of any of these to remove.
+
+        Raises:
+            RuntimeError: If the context has no live tools list (for example when the
+                function is invoked outside of a function-calling loop).
+        """
+        from ._tools import _get_tool_name, normalize_tools  # type: ignore[reportPrivateUsage]
+
+        if self.tools is None:
+            raise RuntimeError(
+                "Cannot remove tools: this FunctionInvocationContext is not bound to a live "
+                "agent run. remove_tools is only available for functions invoked within an "
+                "agent's function-calling loop."
+            )
+
+        names_to_remove: set[str] = set()
+        raw_items: list[Any]
+        if isinstance(tools, str):
+            raw_items = [tools]
+        elif isinstance(tools, Sequence) and not isinstance(tools, (bytes, bytearray)):
+            raw_items = list(cast("Sequence[Any]", tools))
+        else:
+            raw_items = [tools]
+        for item in raw_items:
+            if isinstance(item, str):
+                names_to_remove.add(item)
+                continue
+            for normalized in normalize_tools(item):
+                if name := _get_tool_name(normalized):  # type: ignore[reportPrivateUsage]
+                    names_to_remove.add(name)
+
+        if not names_to_remove:
+            return
+        self.tools[:] = [
+            tool
+            for tool in self.tools
+            if _get_tool_name(tool) not in names_to_remove  # type: ignore[reportPrivateUsage]
+        ]
 
 
 class ChatContext:

python/packages/core/agent_framework/_tools.py

@@ -1418,6 +1418,7 @@ async def _auto_invoke_function(
     sequence_index: int | None = None,
     request_index: int | None = None,
     middleware_pipeline: FunctionMiddlewarePipeline | None = None,
+    live_tools: list[ToolTypes] | None = None,
 ) -> Content:
     """Invoke a function call requested by the agent, applying middleware that is defined.
 
@@ -1432,6 +1433,8 @@ async def _auto_invoke_function(
         sequence_index: The index of the function call in the sequence.
         request_index: The index of the request iteration.
         middleware_pipeline: Optional middleware pipeline to apply during execution.
+        live_tools: The live, mutable tools list for the current agent run, exposed on
+            the FunctionInvocationContext so tools can add/remove tools at runtime.
 
     Returns:
         The function result content.
@@ -1523,6 +1526,7 @@ async def _auto_invoke_function(
                     arguments=args,
                     session=invocation_session,
                     kwargs=runtime_kwargs.copy(),
+                    tools=live_tools,
                 )
             function_result = await tool.invoke(
                 arguments=args,
@@ -1537,6 +1541,10 @@ async def _auto_invoke_function(
         except UserInputRequiredException:
             raise
         except Exception as exc:
+            logger.warning(
+                f"Function '{tool.name}' raised an exception; returning an error result to the "
+                f"model. Set include_detailed_errors=True for the full detail. Exception: {exc!r}"
+            )
             message = "Error: Function failed."
             if config.get("include_detailed_errors", False):
                 message = f"{message} Exception: {exc}"
@@ -1552,6 +1560,7 @@ async def _auto_invoke_function(
         arguments=args,
         session=invocation_session,
         kwargs=runtime_kwargs.copy(),
+        tools=live_tools,
     )
 
     call_id = function_call_content.call_id
@@ -1608,6 +1617,10 @@ async def _auto_invoke_function(
     except UserInputRequiredException:
         raise
     except Exception as exc:
+        logger.warning(
+            f"Function '{tool.name}' raised an exception; returning an error result to the "
+            f"model. Set include_detailed_errors=True for the full detail. Exception: {exc!r}"
+        )
         message = "Error: Function failed."
         if config.get("include_detailed_errors", False):
             message = f"{message} Exception: {exc}"
@@ -1659,6 +1672,9 @@ async def _try_execute_function_calls(
     from ._types import Content
 
     tool_map = _get_tool_map(tools)
+    # The live tools list (when tools is the run-local list) is exposed on the
+    # FunctionInvocationContext so tools can add/remove tools during the run.
+    live_tools: list[ToolTypes] | None = cast("list[ToolTypes]", tools) if isinstance(tools, list) else None
     approval_tools = [tool_name for tool_name, tool in tool_map.items() if tool.approval_mode == "always_require"]
     logger.debug(
         "_try_execute_function_calls: tool_map keys=%s, approval_tools=%s",
@@ -1733,6 +1749,7 @@ async def _try_execute_function_calls(
                 request_index=attempt_idx,
                 middleware_pipeline=middleware_pipeline,
                 config=config,
+                live_tools=live_tools,
             )
             return (result, False)
         except MiddlewareTermination as exc:
@@ -2371,6 +2388,13 @@ class FunctionInvocationLayer(Generic[OptionsCoT]):
                 function_invocation_kwargs=function_invocation_kwargs,
                 client_kwargs=filtered_kwargs,
             )
+        # Establish a single, run-local mutable tools list so that tools can add or remove
+        # tools during the run (progressive tool exposure). A fresh list is created via
+        # normalize_tools so the caller's original tools container is never mutated, while
+        # the same list object is shared with the model (options["tools"]) and the tool map
+        # rebuilt on every loop iteration.
+        if mutable_options.get("tools"):
+            mutable_options["tools"] = normalize_tools(mutable_options["tools"])
         if not stream:
 
             async def _get_response() -> ChatResponse[Any]:

python/packages/core/tests/core/test_function_invocation_logic.py

@@ -3975,3 +3975,425 @@ async def test_user_input_request_empty_contents_returns_fallback(chat_client_ba
     ]
     assert len(function_results) >= 1
     assert any("user input" in (fr.result or "").lower() for fr in function_results)
+
+
+# region Progressive tool exposure (FunctionInvocationContext.add_tools / remove_tools)
+
+
+def _pte_function_call_response(call_id: str, name: str, arguments: str = "{}") -> ChatResponse:
+    return ChatResponse(
+        messages=Message(
+            role="assistant",
+            contents=[Content.from_function_call(call_id=call_id, name=name, arguments=arguments)],
+        )
+    )
+
+
+def _pte_text_response(text: str = "done") -> ChatResponse:
+    return ChatResponse(messages=Message(role="assistant", contents=[text]))
+
+
+@tool(name="factorial", approval_mode="never_require")
+def _pte_factorial(n: int) -> int:
+    """Compute the factorial of n."""
+    result = 1
+    for value in range(2, n + 1):
+        result *= value
+    return result
+
+
+async def test_context_exposes_live_tools(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    seen_names: list[str] = []
+
+    @tool(name="inspect_tools", approval_mode="never_require")
+    def inspect_tools(ctx: FunctionInvocationContext) -> str:
+        assert ctx.tools is not None
+        seen_names.extend(t.name for t in ctx.tools if isinstance(t, FunctionTool))
+        return "inspected"
+
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "inspect_tools"),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [inspect_tools]},
+    )
+    assert "inspect_tools" in seen_names
+
+
+async def test_add_tools_available_next_iteration(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "math tools loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    response = await chat_client_base.get_response(
+        [Message(role="user", contents=["compute 5!"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+    )
+    assert exec_counter == 1
+    assert response.messages[-1].text == "done"
+
+
+async def test_add_tools_model_sees_added_tools_in_options(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["compute 5!"])],
+            options={"tool_choice": "auto", "tools": [load_math]},
+        )
+    finally:
+        monkey.undo()
+
+    assert recorded[0] == ["load_math"]
+    assert "factorial" in recorded[1]
+
+
+async def test_remove_tools_next_iteration(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="get_weather", approval_mode="never_require")
+    def get_weather(location: str) -> str:
+        return "sunny"
+
+    @tool(name="drop_weather", approval_mode="never_require")
+    def drop_weather(ctx: FunctionInvocationContext) -> str:
+        ctx.remove_tools("get_weather")
+        return "removed"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "drop_weather"),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["hi"])],
+            options={"tool_choice": "auto", "tools": [get_weather, drop_weather]},
+        )
+    finally:
+        monkey.undo()
+
+    assert set(recorded[0]) == {"get_weather", "drop_weather"}
+    assert "get_weather" not in recorded[1]
+
+
+async def test_add_tools_does_not_mutate_caller_tools_list(chat_client_base: SupportsChatGetResponse):
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    original_tools: list[Any] = [load_math]
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": original_tools},
+    )
+    assert original_tools == [load_math]
+
+
+async def test_add_tools_persists_across_iterations(chat_client_base: SupportsChatGetResponse):
+    from agent_framework import FunctionTool
+
+    recorded: list[list[str]] = []
+    client_cls = type(chat_client_base)
+    original = client_cls._get_non_streaming_response
+
+    async def recording(self: Any, *, messages: Any, options: dict[str, Any], **kwargs: Any) -> ChatResponse:
+        tools = options.get("tools") or []
+        recorded.append([t.name for t in tools if isinstance(t, FunctionTool)])
+        return await original(self, messages=messages, options=options, **kwargs)
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(_pte_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 4  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_function_call_response("3", "factorial", '{"n": 3}'),
+        _pte_text_response(),
+    ]
+    monkey = pytest.MonkeyPatch()
+    monkey.setattr(client_cls, "_get_non_streaming_response", recording)
+    try:
+        await chat_client_base.get_response(
+            [Message(role="user", contents=["hi"])],
+            options={"tool_choice": "auto", "tools": [load_math]},
+        )
+    finally:
+        monkey.undo()
+
+    assert "factorial" in recorded[1]
+    assert "factorial" in recorded[2]
+
+
+async def test_add_tools_through_function_middleware(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    class PassthroughMiddleware(FunctionMiddleware):
+        async def process(self, context: FunctionInvocationContext, call_next: Any) -> None:
+            await call_next()
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+        middleware=[PassthroughMiddleware()],
+    )
+    assert exec_counter == 1
+
+
+async def test_add_tools_with_approval_required_tool(chat_client_base: SupportsChatGetResponse):
+    @tool(name="secure_tool", approval_mode="always_require")
+    def secure_tool(value: str) -> str:
+        return f"secure: {value}"
+
+    @tool(name="load_secure", approval_mode="never_require")
+    def load_secure(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(secure_tool)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_secure"),
+        _pte_function_call_response("2", "secure_tool", '{"value": "x"}'),
+        _pte_text_response(),
+    ]
+    response = await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_secure]},
+    )
+    assert any(item.type == "function_approval_request" for msg in response.messages for item in msg.contents)
+
+
+async def test_add_tools_accepts_plain_callable(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    def plain_factorial(n: int) -> int:
+        """Compute factorial."""
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(plain_factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.run_responses = [
+        _pte_function_call_response("1", "load_math"),
+        _pte_function_call_response("2", "plain_factorial", '{"n": 5}'),
+        _pte_text_response(),
+    ]
+    await chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        options={"tool_choice": "auto", "tools": [load_math]},
+    )
+    assert exec_counter == 1
+
+
+async def test_add_tools_streaming(chat_client_base: SupportsChatGetResponse):
+    exec_counter = 0
+
+    @tool(name="factorial", approval_mode="never_require")
+    def factorial(n: int) -> int:
+        nonlocal exec_counter
+        exec_counter += 1
+        return 120
+
+    @tool(name="load_math", approval_mode="never_require")
+    def load_math(ctx: FunctionInvocationContext) -> str:
+        ctx.add_tools(factorial)
+        return "loaded"
+
+    chat_client_base.function_invocation_configuration["max_iterations"] = 3  # type: ignore[attr-defined]
+    chat_client_base.streaming_responses = [
+        [
+            ChatResponseUpdate(
+                contents=[Content.from_function_call(call_id="1", name="load_math", arguments="{}")],
+                role="assistant",
+            )
+        ],
+        [
+            ChatResponseUpdate(
+                contents=[Content.from_function_call(call_id="2", name="factorial", arguments='{"n": 5}')],
+                role="assistant",
+            )
+        ],
+        [ChatResponseUpdate(contents=[Content.from_text("done")], role="assistant", finish_reason="stop")],
+    ]
+    async for _ in chat_client_base.get_response(
+        [Message(role="user", contents=["hi"])],
+        stream=True,
+        options={"tool_choice": "auto", "tools": [load_math]},
+    ):
+        pass
+    assert exec_counter == 1
+
+
+def test_add_tools_duplicate_same_object_is_noop():
+    @tool(name="dup", approval_mode="never_require")
+    def dup(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=dup, arguments={}, tools=[dup])
+    ctx.add_tools(dup)
+    assert ctx.tools is not None
+    assert len(ctx.tools) == 1
+
+
+def test_add_tools_duplicate_name_different_object_raises():
+    @tool(name="dup", approval_mode="never_require")
+    def dup_a(x: int) -> int:
+        return x
+
+    @tool(name="dup", approval_mode="never_require")
+    def dup_b(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=dup_a, arguments={}, tools=[dup_a])
+    with pytest.raises(ValueError):
+        ctx.add_tools(dup_b)
+
+
+def test_add_tools_batch_with_duplicate_is_atomic():
+    """A duplicate-name clash partway through a batch must leave the live list unchanged."""
+
+    @tool(name="existing", approval_mode="never_require")
+    def existing(x: int) -> int:
+        return x
+
+    @tool(name="fresh", approval_mode="never_require")
+    def fresh(x: int) -> int:
+        return x
+
+    @tool(name="existing", approval_mode="never_require")
+    def clashing(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=existing, arguments={}, tools=[existing])
+    with pytest.raises(ValueError):
+        ctx.add_tools([fresh, clashing])
+    assert ctx.tools is not None
+    # The valid "fresh" tool must not have been committed before the clash raised.
+    assert ctx.tools == [existing]
+
+
+def test_remove_tools_by_name_and_object():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    @tool(name="b", approval_mode="never_require")
+    def b(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={}, tools=[a, b])
+    ctx.remove_tools("a")
+    assert ctx.tools is not None
+    assert [t.name for t in ctx.tools] == ["b"]
+    ctx.remove_tools(b)
+    assert ctx.tools == []
+
+
+def test_remove_tools_unknown_name_is_noop():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={}, tools=[a])
+    ctx.remove_tools("nonexistent")
+    assert ctx.tools is not None
+    assert [t.name for t in ctx.tools] == ["a"]
+
+
+def test_progressive_tools_helpers_raise_without_live_tools():
+    @tool(name="a", approval_mode="never_require")
+    def a(x: int) -> int:
+        return x
+
+    ctx = FunctionInvocationContext(function=a, arguments={})
+    assert ctx.tools is None
+    with pytest.raises(RuntimeError):
+        ctx.add_tools(a)
+    with pytest.raises(RuntimeError):
+        ctx.remove_tools("a")
+
+
+# endregion