fix(anthropic): preserve thinking token usage

[hafzism]

Description

Fixes #629.

Anthropic returns extended-thinking usage under usage.output_tokens_details.thinking_tokens, but the provider's typed response discarded that nested object during JSON decoding. As a result, thinking-token counts never reached persisted usage metadata.

This change:

• decodes Anthropic's nested thinking-token count;
• preserves it as the existing normalized completion_reasoning_tokens raw-usage field;
• retains thinking-token usage across streaming usage merges and emits standard nested details for both Chat Completions and Responses streams;
• treats the count as a breakdown already included in output tokens, avoiding duplicate charges and false unmapped-token caveats;
• adds table-driven regression coverage for zero and positive counts, streaming propagation, and base/distinct reasoning rates.

Validation:

• go test ./internal/providers/anthropic ./internal/usage -count=1
• go test ./cmd/... ./config/... ./ext/... ./internal/... ./run/... -count=1
• make lint (0 issues)
• make mod-check
• git diff --check

AI Generated (optional)

This PR was developed with Codex assistance. The issue scope, implementation, review follow-ups, tests, and validation results were reviewed before submission.

[coderabbitai]

📝 Walkthrough

Walkthrough

The Anthropic provider parses output_tokens_details.thinking_tokens, stores positive values as completion_reasoning_tokens, exposes them in usage payloads, and includes them in Anthropic cost calculation.

Changes

Anthropic thinking-token accounting

Layer / File(s)	Summary
Capture thinking tokens internal/providers/anthropic/types.go, internal/providers/anthropic/anthropic.go, internal/providers/anthropic/anthropic_test.go	Usage parsing reads thinking_tokens. Raw usage records positive values as completion_reasoning_tokens. Tests validate parsing, omission of zero values, and usage merging.
Emit usage details internal/providers/anthropic/chat_stream.go, internal/providers/anthropic/responses.go	Chat Completions and Responses API payloads use the shared helper for cache and thinking-token details.
Price thinking tokens internal/usage/cost.go, internal/usage/cost_test.go	Anthropic cost mapping recognizes reasoning output tokens. Tests cover base output pricing and a configured reasoning-output rate.

Estimated code review effort: 3 (Moderate) | ~20 minutes

Possibly related PRs

• ENTERPILOT/GoModel#439: Both changes update Anthropic token-cost mapping in internal/usage/cost.go.

Suggested reviewers: santiagodepolonia

Poem

I’m a rabbit; thinking tokens now appear,
Raw usage keeps each count clear.
Payload details follow the trail,
Cost rules map the thought detail.
Hop, hop—the tests confirm!

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Linked Issues check	✅ Passed	The changes satisfy issue #629 by decoding, preserving, and cost-accounting Anthropic thinking tokens without duplicate charges.
Out of Scope Changes check	✅ Passed	All changes support issue #629 and related regression coverage; no unrelated code changes are identified.
Docstring Coverage	✅ Passed	No functions found in the changed files to evaluate docstring coverage. Skipping docstring coverage check.
Title check	✅ Passed	The title clearly and concisely describes preserving Anthropic thinking-token usage, which is the main change.
Description check	✅ Passed	The description explains the problem, implementation, tests, validation, and linked issue in the required sections.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 2

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@internal/providers/anthropic/anthropic_test.go`:
- Around line 2236-2263: The Anthropic thinking-token tests are not table-driven
and do not cover zero-value normalization or pricing variants. In
internal/providers/anthropic/anthropic_test.go:2236-2263, refactor
TestConvertFromAnthropicResponse_WithThinkingTokens into table-driven cases
asserting zero thinking_tokens omits completion_reasoning_tokens and positive
values preserve it. In internal/usage/cost_test.go:96-112, add table-driven
cases for absent reasoning rates and distinct ReasoningOutputPerMtok values,
asserting base-output adjustment prevents duplicate billing.

In `@internal/providers/anthropic/anthropic.go`:
- Around line 352-354: Update mergeAnthropicUsage to copy nonzero
src.OutputTokensDetails.ThinkingTokens into
dst.OutputTokensDetails.ThinkingTokens and set merged = true, ensuring
buildAnthropicRawUsage can emit completion_reasoning_tokens for streaming usage.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: ASSERTIVE

Plan: Pro Plus

Run ID: 668d5aa3-e0eb-4341-b88b-4ea379ba5265

📥 Commits

Reviewing files that changed from the base of the PR and between 9d54827 and 8b5bf86.

📒 Files selected for processing (5)

• internal/providers/anthropic/anthropic.go
• internal/providers/anthropic/anthropic_test.go
• internal/providers/anthropic/types.go
• internal/usage/cost.go
• internal/usage/cost_test.go

[greptile-apps]

Confidence Score: 5/5

The PR is safe to merge; no blocking failure remains.

Focused execution covered normal Anthropic usage conversion, both streaming response formats, and included-token reasoning pricing without identifying a defect.

T-Rex Logs

What T-Rex did

• • Authored and ran a focused Go validation covering normal Anthropic response conversion, simulated Chat Completions streaming, simulated Responses streaming, and granular cost calculation.
• • Compared results to the parent commit 8b5bf86 and found the focused stream assertion failed because final chat usage lacked completion_tokens_details.reasoning_tokens.
• • After updating to PR 8eb0530, all focused assertions passed, including normal conversion retaining completion_reasoning_tokens, both stream formats emitted the expected reasoning-token detail, and the 20,000 thinking tokens cost matched the base and alternative pricing rates.

_{Ran code and verified through T-Rex}

_{Reviews (2): Last reviewed commit: "fix(anthropic): preserve streamed thinki..." | Re-trigger Greptile}

[hafzism]

The streaming P1 reported against 8b5bf86 is addressed in 8eb0530: thinking tokens are retained during usage merges and emitted as completion_tokens_details.reasoning_tokens for Chat Completions and output_tokens_details.reasoning_tokens for Responses. The expanded focused tests, full unit suite, and repository-wide lint pass.

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[SantiagoDePolonia]

LGTM, thank you!

[SantiagoDePolonia]

Merged. Thank you again for your contribution!