Repository navigation
Description
Describe the bug
When a confirmation UI returns ConfirmationUIResult(action="modify", new_tool_params={}), BlockingConfirmationStrategy keeps the original tool arguments instead of applying the empty replacement.
For a tool with optional parameters, removing all arguments should let it use its defaults. In the example below, the model requests record_scope(scope="production"), and the UI replaces the arguments with {}. The tool should use its default "dry_run", but it receives "production" instead.
No exception or warning is raised. The original arguments also remain in the conversation history, with no modification feedback.
Expected behavior
An explicit {} should replace the original arguments. The tool should run with its defaults, and the conversation should reflect the change.
To reproduce
This runs offline and only records a string in a local list. No API key or external action is involved. Save as repro_empty_hitl_replacement.py and run from the repository:
hatch --version
HAYSTACK_TELEMETRY_ENABLED=false hatch -e test run python repro_empty_hitl_replacement.pyimport asyncio
from haystack.components.agents import Agent
from haystack.components.generators.chat import MockChatGenerator
from haystack.dataclasses import ChatMessage, ToolCall
from haystack.hooks.human_in_the_loop import (
AlwaysAskPolicy,
BlockingConfirmationStrategy,
ConfirmationHook,
ConfirmationUIResult,
)
from haystack.tools import Tool
class EmptyReplacementUI:
def get_user_confirmation(self, tool_name, tool_description, tool_params):
return ConfirmationUIResult(action="modify", new_tool_params={})
def make_agent():
observed = []
def record_scope(scope: str = "dry_run") -> str:
observed.append(scope)
return scope
tool = Tool(
name="record_scope",
description="Record the scope locally; perform no external action.",
parameters={"type": "object", "properties": {"scope": {"type": "string"}}},
function=record_scope,
)
generator = MockChatGenerator(
responses=[
ChatMessage.from_assistant(
tool_calls=[ToolCall(tool_name="record_scope", arguments={"scope": "production"}, id="call-1")]
),
"done",
]
)
strategy = BlockingConfirmationStrategy(
confirmation_policy=AlwaysAskPolicy(), confirmation_ui=EmptyReplacementUI()
)
agent = Agent(
chat_generator=generator,
tools=[tool],
hooks={"before_tool": [ConfirmationHook(confirmation_strategies={"record_scope": strategy})]},
)
return agent, observed
async def main():
for mode in ("sync", "async"):
agent, observed = make_agent()
messages = [ChatMessage.from_user("Record the scope.")]
result = agent.run(messages=messages) if mode == "sync" else await agent.run_async(messages=messages)
arguments = [tc.arguments for message in result["messages"] for tc in (message.tool_calls or [])]
print(f"{mode}: tool received {observed}; tool-call arguments {arguments}")
asyncio.run(main())Actual output:
sync: tool received ['production']; tool-call arguments [{'scope': 'production'}]
async: tool received ['production']; tool-call arguments [{'scope': 'production'}]
Expected output:
sync: tool received ['dry_run']; tool-call arguments [{}]
async: tool received ['dry_run']; tool-call arguments [{}]
Cause and suggested fix
The check at strategies.py:124 treats {} as if no replacement was supplied:
if confirmation_ui_result.action == "modify" and confirmation_ui_result.new_tool_params:Checking new_tool_params is not None would honor empty replacements while preserving the current handling of None.
I also verified that a custom strategy returning ToolExecutionDecision(..., final_tool_params={}) correctly uses the tool's defaults. The executor already supports empty arguments; the problem is in the built-in strategy.
Related issue
#13046 concerns a custom strategy that omits final_tool_params. Here, the UI explicitly supplies {}, and the built-in strategy ignores it.
FAQ Check
- Have you had a look at our FAQ page?
System
- OS: macOS
- Haystack: latest pull of
main - Reproduced with both
Agent.run()andAgent.run_async()through Hatch
Activity
Aarkin7 commentedon Nov 22, 2026
I'd like to work on this. I'd keep the fix small: check new_tool_params is not None so an empty replacement is honored, add regression tests covering both sync and async execution. Please let me know if this approach sounds good before I open a PR.