[kuse (liwenjie200543)]

Motivation & Context

Fixes #8826. A streamed background response that is still in progress reports itself as complete: the aggregated response loses its continuation_token on the next event that does not repeat it, and consumers that read the aggregate to decide whether to poll or re-POST act on the wrong reading.

Description & Review Guide

• What are the major changes? _process_update — the single aggregation point behind ChatResponse.from_updates / AgentResponse.from_updates — copied update.continuation_token unconditionally, the only field among its siblings without an is not None guard. The copy is now guarded: the aggregate keeps the last token the provider actually sent, and an update that carries finish_reason (a terminal update) withdraws it.
• What is the impact of these changes? A token the provider has not withdrawn survives later streaming events, so the OpenAI Responses client's completion checks (_chat_client.py :964, :4420) no longer see "complete" while the background operation is running, and polling/function-loop continuation stops dropping the token early. A bare guard alone would have left the stale token on the aggregate after response.completed and broken the documented "presence means in progress, None means complete" contract in the other direction — hence the paired withdrawal on finish_reason.
• What do you want reviewers to focus on? Whether finish_reason is the right terminal marker for all providers (the OpenAI Responses parser sets it exactly on response.completed/incomplete/failed; stdio/Linux behaviour of the repro path is unchanged).

Related Issue

Fixes #8826

Contribution Checklist

• The code builds clean without any errors or warnings
• All unit tests pass, and I have added new tests where possible
• The PR follows the Contribution Guidelines
• This PR is linked to an issue and there is no other open PR for this issue (see Related Issue above).
• This is not a breaking change.

[Copilot]

Copilot review overview

🟡 Changes recommended

Some terminal OpenAI and A2A updates lack a finish reason, causing stale continuation tokens to survive completion.

Review effort: Balanced
Findings: 1

Open (1)

• Clear stale continuation tokens on terminal response updates · New

What changed in this PR

Fixes streamed response aggregation so in-progress continuation tokens survive unrelated updates and clear on terminal updates.

Changes:

• Preserves the last non-None continuation token.
• Clears tokens when an update has a finish reason.
• Adds aggregation tests for both paths.

File	Description
python/​packages/​core/​agent_framework/​_types.py	Updates continuation-token aggregation.
python/​packages/​core/​tests/​core/​test_types.py	Tests token preservation and clearing.

---

💡 Add a code-review agent skill for context-aware, tailored reviews. Learn more in the docs.

[Eduard van Valkenburg (eavanvalkenburg)]

As mentioned in the issue, the core team will fix this.