Repository navigation
Python: [Bug]: Executable sibling tool calls are never run when a declaration-only tool call appears in the same model response #8898
Description
Description
When one model response contains a call to a declaration-only tool (FunctionTool(func=None),
answered by the host, for example by asking the user) and a call to an ordinary executable
tool, the Python function invocation layer runs neither. Both calls come back to the caller
as user-input pauses, although the executable tool has a function and needs no approval.
This is the Python counterpart of the .NET bug #6922 ("FunctionInvokingChatClient drops sibling
backend tool calls when a frontend (declaration-only) tool call appears in the same iteration"),
fixed there by #7388 ("Store executable function calls bypassed by declaration-only tool calls").
Code Sample
if has_declaration_only_call:
# Declaration-only calls are returned as user input rather than executed locally.
declaration_only_calls: list[Content] = []
for function_call in function_calls:
if function_call.type == "function_call":
declaration_only_calls.append(_as_user_input_pause(function_call))
return [[function_call] for function_call in declaration_only_calls], FalseError Messages / Stack Traces
Package Versions
agent-framework-core: 1.19.0
Python Version
3.14
Additional Context
The approval path in the same function already handles a mixed batch: executable siblings of an
approval request are stored and run when the approval batch resumes. The declaration-only path
has no equivalent.
Metadata
Metadata
Labels
Type
Projects
- Status
Activity
[-]Python: [Bug]: Executable sibling tool calls are never run when a declaration-only tool call appears in the same model response[/-][+].NET: Python: [Bug]: Executable sibling tool calls are never run when a declaration-only tool call appears in the same model response[/+]github-actions commentedon Sep 30, 2026
🤖 Automated triage reproduction notes (agent-authored — trust but verify)
Agent analysis
The bug occurs in
python/packages/core/agent_framework/_tools.py::_try_execute_function_call_groups, particularly the declaration-only branch at line 2514. A response containingFunctionTool(func=None)and an executable sibling marks both as user-input pauses, and the executable call is absent when the host result resumes the session. Minimal repro: return both calls from a mocked client, resolve the host call in a secondAgent.run, and assert that the executable tool ran once.python/packages/core/tests/core/test_function_invocation_logic.py::test_mixed_declaration_only_and_executable_batch_runs_executable_after_host_response1.19.01.19.0[-].NET: Python: [Bug]: Executable sibling tool calls are never run when a declaration-only tool call appears in the same model response[/-][+] Python: [Bug]: Executable sibling tool calls are never run when a declaration-only tool call appears in the same model response[/+]Li-john1021 commentedon Oct 2, 2026
I'll take this.
A local fix follows the existing approval mixed-batch path: when one response has a declaration-only call and an executable sibling, the executable call is stored and run once after the host
function_result, instead of being returned as another user-input pause. I'll open a PR shortly.fix: match declaration-only host pauses by occurrence (microsoft#8898)
fix: run executable sibling tools alongside declaration-only calls (m…
fix: match declaration-only host pauses by occurrence (microsoft#8898)