[sjrl]

Related Issues

• Follow up to fix: preserve multiple Responses API reasoning items #13366 Found while testing the linked PR. Realized that the default wording of the template led to the model still re-calling the original tool call or issuing another unnecessary tool call with updated parameters.

Proposed Changes:

• Reword the default modify_template of BlockingConfirmationStrategy to tell the LLM the user wants the modified result used and not to call the tool again with the original arguments.
• Update the matching tests, docs default and add a release note.

How did you test it?

Unit tests, plus the live script below on main vs this branch with gpt-5-mini: 2/5 runs repeated the tool call on main, 0/5 on this branch.

Notes for the reviewer

With the old wording the model often treated the modified result as a tool failure ("I tried to fetch Berlin, but the weather tool keeps returning Munich") and called the tool again.

Reproduction script (needs OPENAI_API_KEY)

from typing import Annotated

from haystack.components.agents import Agent
from haystack.components.generators.chat import OpenAIResponsesChatGenerator
from haystack.dataclasses import ChatMessage
from haystack.hooks.human_in_the_loop import (
    AlwaysAskPolicy,
    BlockingConfirmationStrategy,
    ConfirmationHook,
    ConfirmationUIResult,
)
from haystack.tools import tool

calls = []


@tool
def get_weather(city: Annotated[str, "The city to get the weather for"]) -> str:
    """Get the current weather for one city."""
    calls.append(city)
    return f"Sunny, 21C in {city}"


class ChangeToMunichUI:
    """Simulates a user who changes every Berlin lookup to Munich."""

    def get_user_confirmation(self, tool_name, tool_description, tool_params):
        if tool_params.get("city") == "Berlin":
            return ConfirmationUIResult(action="modify", new_tool_params={"city": "Munich"})
        return ConfirmationUIResult(action="confirm")


agent = Agent(
    chat_generator=OpenAIResponsesChatGenerator(model="gpt-5-mini"),
    tools=[get_weather],
    hooks={
        "before_tool": [
            ConfirmationHook(
                confirmation_strategies={
                    "get_weather": BlockingConfirmationStrategy(
                        confirmation_policy=AlwaysAskPolicy(), confirmation_ui=ChangeToMunichUI()
                    )
                }
            )
        ]
    },
    max_agent_steps=5,
)

runs = 5
repeated = 0
for _ in range(runs):
    calls.clear()
    result = agent.run(messages=[ChatMessage.from_user("What's the weather in Berlin right now?")])
    # Every tool call reaches the UI with the arguments the model chose, so more than one call means the model
    # asked for the weather again after the user's modification.
    repeated += len(calls) > 1
    print(f"tool calls: {calls} | {result['last_message'].text[:80]!r}")

print(f"\n{repeated}/{runs} runs called the tool again after the modification")

Checklist

• I have read the contributors guidelines and the code of conduct.
• I have updated the related issue with new insights and changes.
• I have added unit tests and updated the docstrings.
• I've used one of the conventional commit types for my PR title.
• I have documented my code.
• I have added a release note file, following the contributors guidelines.
• I have run pre-commit hooks and fixed any issue.

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

Project	Deployment	Actions	Updated
haystack-docs	Ready	Preview	Oct 22, 2026 6:25am UTC

[github-actions]

Coverage report

Click to see where and how coverage changed

File	Statements	Missing	Coverage	Coverage (new stmts)	Lines missing
haystack/hooks/human_in_the_loop
strategies.py
Project Total

This report was generated by python-coverage-comment-action