Python: Fix doubled tool_call arguments in MESSAGES_SNAPSHOT when streaming

[LEDazzio01]

Motivation and Context

Fixes #4194 — MESSAGES_SNAPSHOT events contain doubled function.arguments strings for tool calls when streaming with client-side tools.

When a streaming provider sends a full-arguments replay after incremental deltas complete, _emit_tool_call() unconditionally appends the replay to the accumulated arguments via +=, causing the arguments string to double (e.g., {"todoText":"buy groceries"}{"todoText":"buy groceries"}). This breaks any client or middleware that relies on MESSAGES_SNAPSHOT for state reconstruction, since the doubled string is not valid JSON.

Description

Root cause: In _run_common.py, the _emit_tool_call function tracks accumulated arguments in flow.tool_calls_by_id[tool_call_id]["function"]["arguments"] using +=. Unlike _emit_text() which already has a guard against full-message replays, _emit_tool_call had no equivalent protection.

Fix: Add a duplicate detection guard (mirroring the existing pattern in _emit_text) that checks whether the incoming delta exactly equals the already-accumulated arguments string. When matched, this indicates a full-arguments replay rather than an incremental delta, and the append is skipped:

         if tool_call_id in flow.tool_calls_by_id:
-            flow.tool_calls_by_id[tool_call_id]["function"]["arguments"] += delta
+            accumulated = flow.tool_calls_by_id[tool_call_id]["function"]["arguments"]
+            # Guard against full-argument replay: if the accumulated arguments
+            # already equal the incoming delta, this is a non-delta replay of
+            # the complete arguments string. Skip the append to prevent
+            # doubling in MESSAGES_SNAPSHOT.  (Fixes #4194)
+            if accumulated and delta == accumulated:
+                logger.debug(
+                    "Skipping duplicate full-arguments replay for tool_call_id=%s",
+                    tool_call_id,
+                )
+            else:
+                flow.tool_calls_by_id[tool_call_id]["function"]["arguments"] += delta

Key properties:

• ToolCallArgsEvent deltas are still emitted correctly for real-time streaming — only the internal snapshot accumulator is guarded
• The guard only triggers on exact equality (accumulated == delta), so normal incremental streaming is unaffected
• Follows the same pattern already established in _emit_text() for text content replay detection

Contribution Checklist

• The code builds clean without any errors or warnings
• The PR follows the Contribution Guidelines
• All unit tests pass, and I have added new tests where possible
• Is this a breaking change? No — this is a bug fix that prevents invalid doubled arguments in snapshots.

[LEDazzio01]

I have sole ownership of intellectual property rights to my Submissions and I am not making Submissions in the course of work for my employer.

[Copilot]

Pull request overview

This PR fixes a bug where MESSAGES_SNAPSHOT events contained doubled function.arguments strings when streaming tool calls. The issue occurred when streaming providers send a full-arguments replay after incremental deltas, causing the arguments to be accumulated twice. The fix adds duplicate detection logic (similar to the existing text content handling) to skip re-accumulating arguments when the incoming delta equals the already-accumulated value.

Changes:

• Added duplicate detection guard in _emit_tool_call() to prevent doubling tool call arguments in MESSAGES_SNAPSHOT events
• Follows the existing pattern from _emit_text() for handling provider full-content replays
• Includes detailed inline comments explaining the fix and referencing issue #4194

[LEDazzio01]

@microsoft-github-policy-service agree

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/ag-ui/agent_framework_ag_ui
_run_common.py	199	29	85%	46, 51–52, 54, 61, 65, 69–70, 72, 96, 173, 221–222, 236, 263–265, 290–291, 297–302, 389, 391, 394–395
TOTAL	21565	3093	85%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
4322	246 💤	0 ❌	0 🔥	1m 15s ⏱️

[moonbox3]

LGTM. Thanks for fixing, @LEDazzio01.