Do not double-count usage when Response.ToUpdates re-emits per-message UsageContent

[PratikDhanave]

What

(*Response).ToUpdates now filters *message.UsageContent out of each per-message update's Contents, while keeping the single trailing aggregate UsageContent update.

Why

Providers embed usage inside message contents (e.g. anthropicprovider/agent.go puts UsageContent in msg.Contents). ToUpdates previously emitted the full msg.Contents (still holding that UsageContent) and appended a trailing update carrying resp.Usage(), which itself is the sum of per-message usage.

ResponseStream.Collect merges the empty-keyed trailing update into the last message, so Collect(resp.ToUpdates()).Usage() returned sum(per-message usage) + aggregate — double the real total.

This mirrors .NET ToAgentResponseUpdates, which emits usage exactly once. The filtering builds a fresh slice and never mutates msg.Contents, since the Response is caller-owned.

Testing

Added TestResponse_ToUpdates_RoundTripDoesNotDoubleCountUsage: builds a Response with a message containing a TextContent and a UsageContent{TotalTokenCount:100}, then asserts the per-message update carries no UsageContent, the trailing update reports 100, and Collect(resp.ToUpdates()).Usage().TotalTokenCount == 100 (was 200 before the fix). The test fails before the change and passes after.

go build ./..., go vet ./agent/..., and go test ./agent/... all pass.

[Copilot]

Pull request overview

This PR fixes usage double-counting when converting a completed agent.Response into streaming-style ResponseUpdates and then collecting them back into a Response.

Changes:

• Filter *message.UsageContent out of per-message updates produced by (*Response).ToUpdates, while still emitting a single trailing aggregate usage update.
• Add a regression test ensuring Collect(resp.ToUpdates()).Usage() matches the real total (no double-counting).

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
agent/response.go	Filters per-message Contents to omit embedded UsageContent so collection doesn’t double-count usage.
agent/response_test.go	Adds a round-trip test covering the double-counting scenario and asserting correct aggregate usage.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[github-actions]

Parity Review: No Cross-Repo Consistency Issues

This PR fixes a usage double-counting bug in (*Response).ToUpdates() and is in scope for parity review (it changes observable behavior of an exported method).

What changed

ToUpdates() previously included *message.UsageContent items inside per-message update Contents and appended a separate trailing aggregate UsageContent update. This caused Collect(resp.ToUpdates()).Usage() to return per-message usage + aggregate — twice the real total.

The fix strips *message.UsageContent from per-message update Contents, so usage is emitted exactly once via the trailing aggregate update. The unexported helper contentsWithoutUsage never mutates the caller-owned Response.

Upstream parity

The PR description correctly identifies the upstream reference: .NET ToAgentResponseUpdates already emits usage exactly once. The Go fix restores parity with .NET behavior rather than diverging from it.

No equivalent Python surface was found for ToUpdates/ToAgentResponseUpdates, so no Python parity gap applies here.

Conclusion

The exported ToUpdates() signature is unchanged. The behavioral correction aligns Go semantics with the .NET implementation. No cross-repo consistency issues found.

Warning

Firewall blocked 1 domain

The following domain was blocked by the firewall during workflow execution:

• awmgmcpg

To allow these domains, add them to the network.allowed list in your workflow frontmatter:

network:
  allowed:
    - defaults
    - "awmgmcpg"

See Network Configuration for more information.

Generated by Go API Consistency Review Agent · 38.4 AIC · ⌖ 5.54 AIC · ⊞ 5.7K · ◷