.NET: Set AgentResponse.CreatedAt from the A2A task status timestamp

[Charles (anneheartrecord)]

What

A2AAgent never sets AgentResponse.CreatedAt / AgentResponseUpdate.CreatedAt, so every response it produces reports a null creation time. CopilotStudioAgent and GitHubCopilotAgent both populate it, so the metadata surface is inconsistent across agent implementations.

This sets CreatedAt from the A2A task status timestamp in the three converters whose input actually carries one.

Why the value is available

AgentTaskStatus in the A2A SDK declares:

/// ISO 8601 datetime string when the status was recorded.
public DateTimeOffset Timestamp { get; set; } = DateTimeOffset.UtcNow;

The surrounding code already reads Status.State from the same object on adjacent lines, so nothing new is fetched or plumbed through — the timestamp is one property away at every site that now uses it.

Scope: three of six converters

A2AAgent has six response converters. Only three can be fixed; the other three are a data limitation rather than an oversight, so I left them alone:

Converter	Input	Set CreatedAt?
ConvertToAgentResponse(AgentTask)	task.Status	✅
ConvertToAgentResponseUpdate(AgentTask)	task.Status	✅
ConvertToAgentResponseUpdate(TaskStatusUpdateEvent)	statusUpdateEvent.Status	✅
ConvertToAgentResponse(Message)	Message	❌ AgentMessage exposes no timestamp
ConvertToAgentResponseUpdate(Message)	Message	❌ same
ConvertToAgentResponseUpdate(TaskArtifactUpdateEvent)	—	❌ type carries only Artifact / Append / LastChunk

Behaviour change worth calling out

Consumers that previously saw CreatedAt == null from an A2A agent will now see a value on the task-backed paths. That is the intent, but it is observable, so flagging it rather than burying it. The message-backed paths still return null, which is accurate — there is no timestamp to report there.

Tests

Two regression tests in A2AAgentTests, covering the non-streaming and streaming task paths. Both use an explicit timestamp distinct from "now", so they cannot pass by accident against AgentTaskStatus.Timestamp's DateTimeOffset.UtcNow default.

Verified red-before-green: with the source change reverted and the tests kept, both fail with Assert.Equal() Failure: Values differ; with the change applied, Microsoft.Agents.AI.A2A.UnitTests is 90/90 on net10.0. dotnet format --verify-no-changes is clean on both changed projects.

Not covered by a test: the TaskStatusUpdateEvent path. Driving a status-update event stream through the existing stub handler was more setup than the one-line change seemed to warrant, so I would rather say so than imply coverage I do not have. Happy to add it if you would prefer the path tested before merge.

One question

#6790 cited A2AAgent and GitHubCopilotAgent as the reference for a complete metadata surface while CopilotStudioAgent was the outlier. After #6791 that has inverted — CopilotStudioAgent now sets all four of CreatedAt / FinishReason / RawRepresentation / AdditionalProperties, and the two references have gaps of their own. This PR closes the A2AAgent one. GitHubCopilotAgent looks like it may be missing FinishReason and AdditionalProperties on its update path — worth a separate pass, or would you rather have the remaining gaps swept in one PR? Happy to do either; I would just rather ask than guess at the scope you want.

[Copilot]

Pull request overview

Populates A2A response timestamps from task status metadata.

Changes:

• Maps task status timestamps onto three task-backed response paths.
• Adds regression coverage for task response conversion.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
A2AAgent.cs	Maps status timestamps to CreatedAt.
A2AAgentTests.cs	Tests non-streaming and fallback-streaming mappings.

---

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

[Luis Rodriguez (luisangelrod)]

Reviewed the full change at 600df790, including all six response converters and the existing streaming status-event test path. The three new assignments are correctly limited to A2A inputs that carry a status timestamp, and leaving the message and artifact paths unset is appropriate because those inputs do not expose one.

I ran the complete Microsoft.Agents.AI.A2A.UnitTests project on net10.0: 146 passed, 0 failed, 0 skipped. I found no additional correctness issue beyond the two existing inline threads.

I agree that the stale AgentTaskStatus/UtcNow comment should be corrected and that the already-existing TaskStatusUpdateEvent test should assert CreatedAt before approval, since that gives direct coverage for the third changed converter with minimal added setup. I will recheck after those are addressed.

Generic disclosure: this review was prepared with AI assistance and verified against the full diff, source, and local test results.

[SergeyMenshykh]

I’m not sure TaskStatus.Timestamp is the right value for AgentResponse.CreatedAt. It records when the current task status was set, not when the response was created.

OpenAI keeps these concepts separate:

• created_at - when the Response object was created.
• completed_at - when generation completed.

An A2A TaskStatus.Timestamp is closer to a status-transition timestamp, or possibly completed_at, than OpenAI’s created_at.

Clients that want the status timestamp can access the original task through RawRepresentation and copy it to CreatedAt themselves. I think we should leave CreatedAt null unless A2A provides an actual response creation time.

[Charles (anneheartrecord)]

That distinction makes sense. Status.Timestamp records when the task entered its current state, so for a completed task it's closer to completed_at than created_at, and AgentResponse has no completed-at slot to put it in. Agreed that leaving CreatedAt null beats teaching clients a wrong meaning, with RawRepresentation as the escape hatch.

One converter I think is different: ConvertToAgentResponseUpdate(TaskStatusUpdateEvent). There the update we emit is the status change itself, so the event's timestamp is when that update came into existence — same shape as #6791 mapping activity.Timestamp for Copilot Studio. I can trim the PR down to just that converter and drop the two task-level mappings, or close it entirely if you'd rather keep CreatedAt untouched across the board. Which do you prefer?

[SergeyMenshykh]

Thanks, that distinction makes sense for a live status event. My concern is that CreatedAt would be patchy across the stream because it would be populated for only one event type.

A typical task stream contains updates in this sequence:

1. Task - current task state; CreatedAt would be unset because the task status timestamp is not the response creation time.

2. TaskStatusUpdateEvent - live status change; CreatedAt would, as you suggested, be populated from the status timestamp.

3. TaskArtifactUpdateEvent - artifact or artifact chunk; CreatedAt would be unset because the event carries no timestamp.

4. TaskStatusUpdateEvent - task completion; CreatedAt would be populated from the status timestamp.

Given that, I think it may be clearer and more consistent to leave CreatedAt unset across these A2A update types rather than have it depend on the event representation.

[Charles (anneheartrecord)]

Consistency across the stream is the stronger argument here — a CreatedAt that only ever shows up on status updates isn't something callers can rely on. Closing this one, and thanks for taking the time to think it through.