Python: [Bug]: AG-UI starts each tool call twice when an agent calls two declaration-only tools in one response #9638

Observed Behavior

When the model calls two declaration-only (client-side) tools in one response, the AG-UI endpoint starts each tool call twice. The client receives TOOL_CALL_START twice for each call id, and the thread snapshot records each call twice. The next run then fails, because the provider rejects the duplicate call ids.

A single client-side tool call per response is not affected.

Expected Behavior

Each tool call is started once and recorded once in the snapshot, however many declaration-only calls the response contains.

Steps to Reproduce

1. Register two declaration-only tools on an agent served through the AG-UI endpoint.
2. Send a message that makes the model call both tools in one response.
3. Observe two TOOL_CALL_START events per call id, and each call twice in the thread snapshot.
4. Send the tool results back: the provider rejects the run because the call ids repeat.

Minimal Reproduction

from ag_ui.core import ToolCallStartEvent
from agent_framework import Content
from agent_framework_ag_ui._run_common import FlowState, _emit_content

flow = FlowState()
streamed = [
    Content.from_function_call(call_id="call_a", name="get_weather", arguments='{"city": "NYC"}'),
    Content.from_function_call(call_id="call_b", name="get_time", arguments='{"zone": "UTC"}'),
]
# The function loop hands declaration-only calls back as caller-input pauses after they stream.
replayed = [
    Content.from_function_call(call_id="call_a", name="get_weather"),
    Content.from_function_call(call_id="call_b", name="get_time"),
]
for pause in replayed:
    pause.user_input_request = True

events = [e for c in streamed + replayed for e in _emit_content(c, flow)]
print([e.tool_call_id for e in events if isinstance(e, ToolCallStartEvent)])
print([call["id"] for call in flow.pending_tool_calls])

Error Messages and Stack Traces

tool_use ids must be unique

(Anthropic's 400 response on the run after the duplicated snapshot.)

Package Versions

agent-framework-core: 1.20.0, agent-framework-ag-ui: 1.5.0, agent-framework-anthropic: 1.0.0b261002

Python Version

Python 3.13.16

Operating System

Linux

Regression

Unknown

Additional Context

Cause.

1. After the calls stream, the function loop returns each declaration-only call again as a caller-input pause (_as_user_input_pause, python/packages/core/agent_framework/_tools.py:2482, used at :2510 and :2758).
2. The AG-UI stream passes those copies to _emit_content like any other content (python/packages/ag-ui/agent_framework_ag_ui/_agent_run.py:3893), which routes them to _emit_tool_call.
3. _emit_tool_call treats a call as new when its id differs from the last call id (python/packages/ag-ui/agent_framework_ag_ui/_run_common.py:733): if content.name and tool_call_id != flow.tool_call_id:. The stream is A, B, then the replay A, B. Replayed A differs from the last id (B), so A starts again; replayed B then differs from A, so B starts again. Each restart also appends a second entry to flow.pending_tool_calls, which the snapshot is built from.

Proposed fix. Start a call only when its id is not yet registered in this run: if content.name and tool_call_id not in flow.tool_calls_by_id:. A replayed call then takes the existing continuation branch. Every writer of flow.tool_calls_by_id emits TOOL_CALL_START when it registers the id, so a registered id has always been started. I have this change and a regression test ready, and will open a PR once the direction is agreed.

Platform. Observed in the server's Debian 13 (aarch64) container. The cause is platform-independent: the reproduction gives the same output on macOS.

Workaround. We wrap _emit_tool_call so a call id already in flow.tool_calls_by_id is made the current call before the original function runs.

AI assistance. This report was drafted with AI assistance. I verified the reproduction and the analysis.

Acknowledgements

• I searched existing issues and did not find a duplicate.
• I personally verified this behavior and the reproduction details are authentic.
• I will wait for explicit maintainer agreement before starting implementation of a non-trivial change.