Only write non-zero AdditionalCounts keys in OpenAI chat addUsage

[PratikDhanave (PratikDhanave)]

What

addUsage in provider/openaiprovider/chat.go unconditionally allocated AdditionalCounts and wrote the four audio/prediction token details regardless of value. Every text-only response therefore carried four zero-valued entries (PromptTokensDetails.AudioTokens, CompletionTokensDetails.AudioTokens, .AcceptedPredictionTokens, .RejectedPredictionTokens).

This change guards each key with a small helper so it is added only when non-zero, leaving AdditionalCounts nil in the common text-only case.

Why

• Parity with the sibling Anthropic provider, which already guards its extra count (if usage.CacheCreationInputTokens != 0 { ... }) before allocating AdditionalCounts.
• Parity with the Python SDK, which writes each usage detail only when the underlying token value is truthy.
• Downstream impact: otelprovider's hasUsage returns true whenever len(AdditionalCounts) > 0, so all-zero OpenAI usage previously reported hasUsage == true and emitted four zero-valued additional-count span attributes for plain text responses. With the guard, text-only responses no longer inflate the usage map or the span.

How it is tested

Existing black-box tests in chat_test.go are updated to reflect the corrected behavior: text-only responses now expect a nil AdditionalCounts, and the audio-token case asserts only the two non-zero keys are present (the two zero prediction-token keys are dropped). These assertions fail before the fix and pass after. go build ./..., go vet ./provider/openaiprovider/..., and go test ./provider/openaiprovider/... all pass.

[Copilot]

Pull request overview

This pull request refines OpenAI chat usage reporting so UsageDetails.AdditionalCounts is only allocated and populated when OpenAI provides non-zero “detail” token counts, avoiding noisy zero-valued attributes for text-only responses and improving downstream OpenTelemetry span emission behavior.

Changes:

• Updated addUsage to lazily allocate AdditionalCounts and only write non-zero usage-detail keys.
• Updated OpenAI provider black-box tests to expect AdditionalCounts == nil for text-only responses and to assert only non-zero additional-count keys are present.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated no comments.

File	Description
provider/openaiprovider/chat.go	Lazily populates UsageDetails.AdditionalCounts with non-zero OpenAI usage-detail counts only.
provider/openaiprovider/chat_test.go	Adjusts expectations to match the new “only non-zero additional counts” behavior.

---

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

[github-actions]

Parity Review — parity-approved ✅

This PR makes an internal behavioral fix to the unexported addUsage helper in provider/openaiprovider/chat.go. No exported function or method signatures change.

Cross-repo consistency assessment:

• Go Anthropic provider: Already uses the same guard pattern (if usage.CacheCreationInputTokens != 0 { ... }) before allocating AdditionalCounts. This PR brings the OpenAI provider into alignment with that existing pattern.
• Python SDK: The PR description cites that Python writes each usage detail only when the underlying token value is truthy — the Go fix mirrors that behavior.
• Behavioral effect: AdditionalCounts will now be nil (not an empty non-nil map) for text-only responses, which correctly prevents otelprovider's hasUsage from returning true on zero-valued entries. This is the expected semantics.

API surface: No exported symbols were added, removed, or changed — UsageContent, AdditionalCounts, and all public provider types remain identical. The public-api-change label is not warranted.

This change preserves and improves cross-SDK semantic parity. No 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 · 38.3 AIC · ⌖ 5.58 AIC · ⊞ 5.9K · ◷