fix: clamp output token count to prevent negative values

[singhvishalkr]

Summary

• Clamp output token count in getUsage to Math.max(0, ...) to prevent negative values when a provider reports reasoningTokens > outputTokens

Fixes #9168

Root cause

packages/opencode/src/session/index.ts:319 subtracts reasoningTokens from outputTokens to derive the 'visible output' count. When the Kilo gateway reports an inconsistent usage breakdown (reasoningTokens > outputTokens), the result goes negative.

Observed with: moonshotai/kimi-k2.5 (thinking variant) via Kilo gateway.

Impact of fix

• Display: no more negative token counts in TUI, VS Code extension, or session exports
• Stats: stats.ts aggregation no longer skewed by negative output values
• Cost: unchanged (Kilo gateway cost is reported separately)

Changes

packages/opencode/src/session/index.ts (line 319):

\\diff

• output: safe(outputTokens - reasoningTokens),

• output: Math.max(0, safe(outputTokens - reasoningTokens)),
  \\

Test plan

• Build passes (pnpm build)
• With a model that has reasoningTokens > outputTokens, output shows 0 instead of negative
• Normal sessions (reasoningTokens < outputTokens) are unaffected

[kilo-code-bot]

Code Review Summary

Status: No Issues Found | Recommendation: Merge

Files Reviewed (1 files)

• packages/opencode/src/session/session.ts

---

_{Reviewed by gpt-5.4-20260305 · 372,255 tokens}

[singhvishalkr]

Rebased onto main at 74df868. The session code moved from packages/opencode/src/session/index.ts (now a barrel re-export) into packages/opencode/src/session/session.ts, so the one-line clamp has been re-applied against the new location at session.ts:321. Diff is unchanged in intent: output: Math.max(0, safe(outputTokens - reasoningTokens)).

[marius-kilocode]

Duplicate of #9178 with the same patch. That one was opened first and references the underlying issue #9168, so closing this in favor of it.