human_in_the_loop: duplicate tool-call-name ambiguity guard silently skipped, causing wrong decisions to be applied

[immuhammadfurqan]

Describe the bug

_apply_tool_execution_decisions (in haystack/human_in_the_loop/strategies.py) is supposed to refuse to proceed when it can't unambiguously match a ToolExecutionDecision to a tool call — specifically, when multiple tool calls share the same tool name but have no tool_call_id. There's already a test for this (test_two_teds_same_name_no_ids), and it works only when every tool call/decision in the batch lacks an id.

The guard is:

decision_by_id = {d.tool_call_id: d for d in tool_execution_decisions if d.tool_call_id}
decision_by_name = {d.tool_name: d for d in tool_execution_decisions if d.tool_name}

if not decision_by_id and len(decision_by_name) < len(tool_execution_decisions):
    raise ValueError(...)

The problem: not decision_by_id is checked across the entire batch. If any unrelated tool call elsewhere in the same batch happens to carry an id, decision_by_id becomes non-empty, and the whole guard is skipped — even though the other tool calls (the ones with duplicate names and no ids) are just as ambiguous as before. Once skipped, decision_by_name.get(tc.tool_name) returns whichever decision happens to be last in the dict, and that decision is silently applied to the wrong tool call. This is silent data corruption, not just a missed validation — there's no error, no warning, just a tool getting invoked with someone else's arguments.

Error message

No error is raised — that's the bug. Expected: ValueError: ToolExecutionDecisions are missing tool_call_id fields and there are multiple tool calls with the same name.... Actual: silent success with corrupted arguments.

Expected behavior

The ambiguity check should fire whenever there's a duplicate tool name among the calls/decisions that lack an id, regardless of whether other, unrelated tool calls in the same batch happen to have ids.

To Reproduce

from haystack.dataclasses import ChatMessage, ToolCall
from haystack.human_in_the_loop import ToolExecutionDecision
from haystack.human_in_the_loop.strategies import _apply_tool_execution_decisions

message_with_tool_calls = ChatMessage.from_assistant(
    tool_calls=[
        ToolCall(tool_name="add_database_tool", arguments={"name": "Malte"}),  # no id
        ToolCall(tool_name="add_database_tool", arguments={"name": "Milos"}),  # no id, duplicate name
        ToolCall(tool_name="other_tool", arguments={}, id="xyz"),  # has id, unrelated
    ]
)

rejection_messages, new_tool_call_messages = _apply_tool_execution_decisions(
    tool_call_messages=[message_with_tool_calls],
    tool_execution_decisions=[
        ToolExecutionDecision(tool_name="add_database_tool", execute=True, final_tool_params={"name": "Malte"}),
        ToolExecutionDecision(tool_name="add_database_tool", execute=True, final_tool_params={"name": "Milos"}),
        ToolExecutionDecision(tool_name="other_tool", execute=True, tool_call_id="xyz", final_tool_params={}),
    ],
)

for m in new_tool_call_messages:
    print(m.tool_calls)
# Both `add_database_tool` calls end up with arguments {'name': 'Milos'}.
# No ValueError is raised. The "Malte" call is silently lost/overwritten.

Additional context

This is the same validation already covered by test_two_teds_same_name_no_ids in test/human_in_the_loop/test_strategies.py, but that test only has tool calls without ids at all. The bug only shows up with a mix of id-bearing and id-less calls in the same batch, which isn't currently tested.

I'd like to fix the guard to check ambiguity only among the id-less subset, and add a regression test for the mixed case, then submit a PR.

System:

• OS: Windows
• Haystack version: main