Python: handle streamed A2A update events

[sztoplover-bit]

Summary

• preserve A2A task update events when mapping streamed responses
• surface TaskArtifactUpdateEvent chunks as AgentResponseUpdate items so streaming callers receive incremental output
• avoid re-emitting terminal task artifacts after streamed update events have already produced the content

Verification

• ./.venv/bin/pytest python/packages/a2a/tests/test_a2a_agent.py -q
• validated a minimal TaskArtifactUpdateEvent streaming script now yields ['Hello']

Fixes #4901

[markwallace-microsoft]

Python Test Coverage Report •

File	Stmts	Miss	Cover	Missing
packages/a2a/agent_framework_a2a
_agent.py	238	19	92%	352, 357, 359, 501, 513, 521, 537, 558, 579, 599, 613, 627, 639–640, 682–683, 712–714
TOTAL	28040	3415	87%

Python Unit Test Overview

Tests	Skipped	Failures	Errors	Time
5459	20 💤	0 ❌	0 🔥	1m 29s ⏱️

[giles17]

Automated Code Review

Reviewers: 4 | Confidence: 82%

✓ Correctness

This PR correctly fixes the issue where streaming TaskArtifactUpdateEvent events were silently discarded in _map_a2a_stream. The new _updates_from_task_update_event method properly extracts content from both TaskArtifactUpdateEvent (using artifact.parts) and TaskStatusUpdateEvent (using status.message.parts). The deduplication logic in _updates_from_task prevents terminal tasks from re-emitting artifacts that were already streamed incrementally, which is validated by the second test. The code correctly gates the new streaming behavior behind the emit_intermediate flag, preserving backward compatibility for non-streaming calers like poll_task. The implementation is sound and the tests cover both the basic streaming case and the deduplication edge case.

✓ Security Reliability

This PR correctly fixes the core bug where TaskArtifactUpdateEvent and TaskStatusUpdateEvent content was silently dropped during streaming. The new _updates_from_task_update_event method properly extracts incremental content from both event types. However, the deduplication logic that unconditionally suppresses terminal task artifact processing (return []) when emit_intermediate=True introduces a potential silent data-loss path: if an A2A server sends artifacts only on the terminal Task object (without preceding TaskArtifactUpdateEvent chunks), those artifacts are silently discarded because the code bypasses _parse_messages_from_task. The base branch does not have this issue — terminal tasks always fall through to artifact extraction. No security concerns were identified.

✓ Test Coverage

The PR correctly fixes the core bug where TaskArtifactUpdateEvent was silently discarded during streaming. Two new tests cover the primary artifact streaming path and the deduplication scenario (terminal task artifacts not re-emitted after streaming chunks). However, the new _updates_from_task_update_event method has an untested branch for TaskStatusUpdateEvent with message content, and the deduplication guard (if status.state in TERMINAL_TASK_STATES: return []) could suppress terminal task artifacts when no prior events carried content — an edge case with no test coverage.

✓ Design Approach

The PR correctly fixes the core bug: TaskArtifactUpdateEvent chunks were previously discarded entirely (bound to _update_event and ignored). The new _updates_from_task_update_event helper properly maps artifact update events to AgentResponseUpdate objects. The de-duplication logic (early-return [] when the terminal TaskStatusUpdateEvent has no message) is intentional and aligns with A2A spec semantics — the final status event is a completion signal, not a content carrier. One genuine design concern remains: the early-return on TERMINAL_TASK_STATES inside the emit_intermediate guard also short-circuits the background=True terminal-artifact path. If emit_intermediate=True and background=True are both active and the terminal event has no message, the terminal task's parsed artifacts are silently dropped — a behavioral regression vs. the prior code where terminal artifacts were always parsed. Additionally, TaskArtifactUpdateEvent.append (replace vs. append semantics) is not surfaced as a first-class field on AgentResponseUpdate, making it impossible for callers to correctly reconstruct content without reaching into raw_representation.

---

Automated review by giles17's agents

[sztoplover-bit]

@microsoft-github-policy-service agree

[sztoplover-bit]

Addressed the terminal-artifact fallback concern in e1bd156: streaming now tracks artifact IDs per task and only filters terminal artifacts that were already emitted incrementally, so terminal-only artifacts still surface. I also added regression coverage for terminal-only and mixed streamed-plus-terminal artifact cases, and reran python/packages/a2a/tests/test_a2a_agent.py -q locally.

[giles17]

Hey @sztoplover-bit you have a check failing: https://github.com/microsoft/agent-framework/actions/runs/23642363607/job/68909574436?pr=4919