Account for Anthropic thinking tokens in usage details by PratikDhanave · Pull Request #554 · microsoft/agent-framework-go

PratikDhanave (PratikDhanave) · 2026-07-19T11:13:25Z

The Anthropic SDK exposes Usage.OutputTokensDetails.ThinkingTokens — "the number of output tokens the model generated as internal reasoning." toUsageDetails mapped input, output, total, and both cache fields but never read it:

details := message.UsageDetails{
    InputTokenCount:       usage.InputTokens,
    OutputTokenCount:      usage.OutputTokens,
    TotalTokenCount:       usage.InputTokens + usage.OutputTokens,
    CachedInputTokenCount: usage.CacheReadInputTokens,
}

For any extended-thinking request (this provider explicitly supports it — it emits ThinkingBlock/ThinkingDelta as TextReasoningContent), those tokens were silently dropped: the framework’s dedicated UsageDetails.ReasoningTokenCount (populated by the OpenAI, Gemini, and Copilot providers, and surfaced as an OTel telemetry attribute) stayed 0. toUsageDetailsDelta additionally rebuilt the Usage from only four scalar fields, omitting OutputTokensDetails, so the streaming path could never recover it.

Fix

toUsageDetails: ReasoningTokenCount: usage.OutputTokensDetails.ThinkingTokens.
toUsageDetailsDelta: carry OutputTokensDetails through when reconstructing the Usage.

Thinking tokens are a subset of output tokens, so TotalTokenCount is unchanged — only the missing breakdown field is added, matching the OpenAI providers.

Test

TestUsageReasoningTokens returns a non-streaming response with output_tokens_details.thinking_tokens: 35 and asserts ReasoningTokenCount == 35. It fails on the old code (0) and passes with the fix.

Copilot

Pull request overview

This PR improves token-usage accounting by ensuring “reasoning/thinking” tokens emitted by providers are surfaced via UsageDetails.ReasoningTokenCount, including in Anthropic streaming deltas, so telemetry/OTel breakdowns are accurate.

Changes:

Anthropic: map Usage.OutputTokensDetails.ThinkingTokens into UsageDetails.ReasoningTokenCount.
Anthropic streaming: preserve OutputTokensDetails when reconstructing anthropic.Usage from MessageDeltaUsage.
Copilot: surface reasoning tokens from assistant.usage events and add a regression test; add an Anthropic non-streaming regression test.

Reviewed changes

Copilot reviewed 4 out of 4 changed files in this pull request and generated 1 comment.

File	Description
provider/copilotprovider/copilot.go	Adds `ReasoningTokenCount` mapping in Copilot usage updates.
provider/copilotprovider/copilot_test.go	Adds coverage asserting Copilot usage events expose reasoning tokens.
provider/anthropicprovider/agent.go	Maps Anthropic thinking tokens into `ReasoningTokenCount` and carries details through streaming delta reconstruction.
provider/anthropicprovider/agent_test.go	Adds a non-streaming Anthropic regression test for thinking tokens.

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

toUsageDetails mapped input, output, total, and cache tokens but dropped Usage.OutputTokensDetails.ThinkingTokens (the tokens the model generated as internal reasoning), so the framework's dedicated ReasoningTokenCount field stayed zero for extended-thinking requests. toUsageDetailsDelta also reconstructed the Usage without OutputTokensDetails, so the streaming path could not recover it either. Map ThinkingTokens to UsageDetails.ReasoningTokenCount and carry OutputTokensDetails through the delta path, matching the OpenAI, Gemini, and Copilot providers. Thinking tokens are a subset of output tokens, so TotalTokenCount is unchanged.

Covers the toUsageDetailsDelta path: thinking tokens arriving on the streamed message_delta usage must surface as ReasoningTokenCount.

# Conflicts: # provider/anthropicprovider/agent_test.go

github-actions · 2026-08-22T04:38:37Z

API Consistency Review — PR #554

Scope: provider/anthropicprovider/agent.go (unexported helpers only), agent_test.go

Public API change: None. UsageDetails.ReasoningTokenCount is an existing exported field; this PR only fixes the mapping inside the unexported toUsageDetails / toUsageDetailsDelta functions.

Cross-repo parity

SDK	`reasoning_token_count` from Anthropic `thinking_tokens`
Go (other providers)	✅ OpenAI, Gemini, and Copilot providers already populate `ReasoningTokenCount`
Go Anthropic (before this PR)	❌ Always `0` — the field was silently dropped
Go Anthropic (after this PR)	✅ Correctly mapped from `OutputTokensDetails.ThinkingTokens`
Python upstream (`_parse_usage_from_anthropic`)	⚠️ Also missing — `thinking_tokens` is not yet mapped in the Python provider; that is a gap in the Python SDK, not a reason to hold back this Go fix
.NET Anthropic	Delegates to `IChatClient` adapter; usage mapping is handled at a different abstraction layer

Conclusion

This is a targeted bug fix that brings the Anthropic provider into alignment with the Go framework's own usage-reporting convention (the same ReasoningTokenCount already surfaced as an OTel attribute from other providers). No new exported API surface is introduced and no behavioral default changes. The Python upstream has a parallel gap, but that does not make the Go fix incorrect.

✅ Parity approved. No cross-repo consistency issues found.

Generated by Go API Consistency Review Agent · sonnet46 · 27.5 AIC · ⌖ 5.81 AIC · ⊞ 6K · ◷

Copilot AI review requested due to automatic review settings July 19, 2026 11:13

PratikDhanave (PratikDhanave) requested a review from a team as a code owner July 19, 2026 11:13

Copilot started reviewing on behalf of PratikDhanave (PratikDhanave) July 19, 2026 11:13 View session

Copilot AI reviewed Jul 19, 2026

View reviewed changes

Comment thread provider/anthropicprovider/agent.go

PratikDhanave (PratikDhanave) force-pushed the fix-anthropic-reasoning-token-count branch 2 times, most recently from fb35d43 to 1753be4 Compare July 23, 2026 15:44