[dialupdisaster]

Description

OpenCode V2 reports tokens.output with reasoning subtracted (Usage.visibleOutputTokens = outputTokens - reasoningTokens), whereas Claude Code reports thinking tokens inside output. Dashboards and cost pipelines built for Claude Code telemetry therefore undercount OpenCode output, and anything summing every token type sees an unfamiliar reasoning type.

This adds an opt-in combineReasoningTokens option / OPENCODE_COMBINE_REASONING_TOKENS env var. When enabled:

• token.usage emits type=output as output + reasoning and no longer emits type=reasoning (so sums never double-count)
• api_request log output_tokens includes reasoning; reasoning_tokens is omitted
• LLM span llm.token_count.completion includes reasoning; llm.token_count.completion_details.reasoning still carries the subset, per OpenInference semantics
• Session totals and llm.token_count.total are unchanged (they already include reasoning once)

Default behaviour is unchanged.

Type of change

• Bug fix (non-breaking change that fixes an issue)
• New feature (non-breaking change that adds functionality)
• Breaking change (fix or feature that would cause existing functionality to not work as expected)
• Documentation update
• Refactoring (no functional changes)
• Chore (dependency updates, etc.)

Checklist

• I have read the CONTRIBUTING.md document
• My code follows the style guidelines of this project
• bun run lint passes with no errors
• bun run check:jsdoc-coverage passes with no errors
• bun run typecheck passes with no errors
• bun test passes with no errors
• I have added tests that prove my fix is effective or that my feature works
• I have updated the documentation accordingly
• My commits follow the Conventional Commits specification

Related issues

None.

Additional context

Found while comparing OpenCode and Claude Code usage of the same Bedrock model in Datadog: OpenCode users had a reasoning token type and lower output counts than Claude Code users.