Python: Fix AG-UI message handling and MCP tool double-call bug

[moonbox3]

Motivation and Context

Summary

• Fix TextMessageEndEvent not being emitted after tool results (Fixes Python: [AG-UI]: TEXT_MESSAGE_END is not emitted when using the MCP tool #3568)
• Fix tool calls and text content being incorrectly merged in MessagesSnapshotEvent (Fixes Python: [AG-UI]: MESSAGES_SNAPSHOT merges toolCalls and content in single assistant message #3619)
• Fix unhandled JSONDecodeError when parsing malformed tool arguments
• Fix MCP tool double-call bug where second tool call fails with "tool_call_ids did not have response messages" (Fixes Python: [Bug]: Error when calling MCP tool twice via AG-UI with Human-in-the-loop Approvals #3426)
• Fix confirm_changes tool not appearing in MessagesSnapshotEvent (confirmation dialog not rendering)

Description

Issue #3568: TextMessageEndEvent missing after tool results

When a tool-only response was detected, a TextMessageStartEvent was emitted to create message context, but TextMessageEndEvent was not emitted after
tool results. Fixed by emitting the end event in _emit_tool_result().

Issue #3619: MessagesSnapshot merging tool_calls and content

The AG-UI protocol expects tool_calls and content to be in separate messages within MessagesSnapshotEvent. The previous implementation merged them into
a single message. Fixed _build_messages_snapshot() to emit separate messages.

JSONDecodeError crash

Malformed JSON in tool arguments could crash the streaming response. Now we skip the confirmation flow with a warning log instead of crashing.

MCP tool double-call bug

Two root causes:

1. _replace_approval_contents_with_results() placed function_result content in user messages instead of tool messages. OpenAI requires tool results in
   role="tool" messages.
2. _sanitize_tool_history() didn't remove call_ids from pending tracking after seeing their results, causing duplicate synthetic results.

Fixed by:

• Adding _convert_approval_results_to_tool_messages() to extract function_result content from user messages into proper tool messages
• Adding pending_tool_call_ids.discard(call_id) after processing tool results

confirm_changes not in MessagesSnapshotEvent

The confirm_changes tool call events were emitted but not tracked in flow.pending_tool_calls, so the frontend couldn't see them in the snapshot to
render the confirmation dialog. Fixed by tracking confirm_changes in both _emit_approval_request() and the predictive tools path.

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? If yes, add "[BREAKING]" prefix to the title of the PR.

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/ag-ui/agent_framework_ag_ui
_message_adapters.py	453	95	79%	89, 99–100, 109–112, 115–119, 121–126, 129, 138–144, 147, 151–153, 162–164, 184, 190–192, 222, 235–236, 246–247, 284, 287, 289, 292, 295, 311, 328, 350, 381, 386, 397–398, 449, 465–466, 532–535, 537, 543, 551–552, 554, 558–561, 574, 663–666, 668, 733, 768–770, 772–775, 778–779, 781, 787, 790, 792, 795, 797, 803–804, 806
_run.py	472	118	75%	152–159, 302, 321–322, 337–338, 353, 381–383, 408, 411–414, 416–417, 420–426, 429–431, 434, 450–452, 459, 465–467, 471, 476–478, 480–481, 497–501, 512, 525, 527–528, 544, 565–566, 618–620, 632–634, 832, 843–844, 851, 898–900, 917, 923, 931, 933, 969–975, 978–981, 983–992, 995, 1003–1006, 1013, 1016–1017, 1022, 1028–1030, 1034, 1039–1042, 1056–1058
TOTAL	16313	1921	88%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
3943	221 💤	0 ❌	0 🔥	1m 10s ⏱️

[Copilot]

Pull request overview

Fixes several AG-UI streaming and snapshot edge-cases around tool calling, approvals, and MCP integration to align emitted events/messages with protocol and provider constraints.

Changes:

• Emit TEXT_MESSAGE_END when a tool result arrives while a text message context is open.
• Rebuild MESSAGES_SNAPSHOT so tool calls and assistant text appear as separate assistant messages.
• Improve robustness around malformed JSON tool arguments and track confirm_changes in snapshots for UI rendering.

Reviewed changes

Copilot reviewed 5 out of 5 changed files in this pull request and generated 5 comments.

Show a summary per file

File	Description
python/packages/ag-ui/agent_framework_ag_ui/_run.py	Updates streaming event emission, snapshot construction, confirm flow tracking, and adds post-processing for approval/tool history.
python/packages/ag-ui/agent_framework_ag_ui/_message_adapters.py	Adjusts tool-history sanitization to filter confirm_changes and fixes pending tool-call tracking behavior.
python/packages/ag-ui/tests/test_run.py	Adds regression tests for message end balancing and snapshots; introduces a malformed-JSON test case.
python/packages/ag-ui/tests/test_message_hygiene.py	Updates/extends hygiene tests around confirm_changes filtering and approval-result conversion.
python/packages/ag-ui/tests/test_message_adapters.py	Updates expectations/docs for approval-modified arguments behavior (LLM context vs snapshot payload).