fix(zen): stop double-counting reasoning_tokens in oa-compat usage

[tiffanychum]

Issue for this PR

Closes #24268

Re-opens #24367 (auto-closed by needs:compliance for missing template sections); also addresses follow-up review feedback from the reporter.

Type of change

• Bug fix
• New feature
• Refactor / code improvement
• Documentation

What does this PR do?

Stops Zen from double-counting reasoning tokens for OpenAI-compatible providers (Moonshot, Kimi, etc.) so the user-facing usage panel and cost calculation match what the upstream API actually billed.

Per the OpenAI chat-completions usage spec, completion_tokens already includes completion_tokens_details.reasoning_tokens. Zen's oaCompatHelper.normalizeUsage (in packages/console/app/src/routes/zen/util/provider/openai-compatible.ts) was reporting outputTokens = completion_tokens and reasoningTokens = reasoning_tokens without subtracting, so when downstream calculateCost bills outputCost + reasoningCost separately at the same cost.output rate, reasoning was billed twice.

The reporter's evidence:

• API: prompt_tokens=22, completion_tokens=77, completion_tokens_details.reasoning_tokens=78 for a single "Hi"
• Local opencode db: output: 436, reasoning: 790 for a real session (correctly excludes reasoning from output)
• Zen Usage console (broken): output: 1226, reasoning: 790, total: 2016 — i.e. completion_tokens (1226) + reasoning_tokens (790), double-counting
• Cost was being computed off the inflated 2016 total

The fix mirrors openaiHelper.normalizeUsage from the OpenAI Responses helper (openai.ts) and subtracts reasoning_tokens from completion_tokens before returning. It then enforces the invariant outputTokens + reasoningTokens === completion_tokens, which is what the upstream API actually charges against. To make that hold even under the OA-compat provider quirk where reasoning_tokens > completion_tokens (e.g. Moonshot Kimi K2.6 returning reasoning=78, completion=77), the PR clamps reasoningTokens down to completion_tokens. So for the reporter's "Hi" case:

• Before this PR: outputTokens=77, reasoningTokens=78 → bills 155 (double-counts; also exceeds the 77 the upstream API charged)
• Subtracting only: outputTokens=0, reasoningTokens=78 → bills 78 (still 1 unit over what upstream charged)
• This PR: outputTokens=0, reasoningTokens=77 → bills 77 (matches upstream exactly)

Clamping reasoning (not just flooring output at 0) is the refinement raised by the reporter @ceshine in review of the previous attempt at this fix in #24367, and it keeps the invariant the upstream API charges against. Once reasoning is clamped, the Math.max(0, …) on output is no longer needed since reasoning <= completion guarantees output >= 0.

I deliberately did not also touch openaiHelper.normalizeUsage (which already subtracts but does not clamp), since the OpenAI Responses API doesn't exhibit the reasoning > completion quirk and changing it is out of scope.

How did you verify your code works?

Added 4 unit tests in packages/console/app/test/zen-usage.test.ts that lock in the wire-level invariant; all 4 fail against the old code and pass after this fix:

• Reporter's real session: completion=1226, reasoning=790 → outputTokens=436, reasoningTokens=790, sum=1226
• Reporter's "Hi" with inverted ratio: completion=77, reasoning=78 → outputTokens=0, reasoningTokens=77, sum=77 (clamped, matches upstream)
• No reasoning at all: outputTokens=77 left untouched, reasoningTokens undefined
• Parity with openaiHelper.normalizeUsage for the same logical usage

Then ran the full console/app suite and full repo typecheck:

• bun test in packages/console/app/ — 7/7 pass
• bun turbo typecheck from repo root — all 13 tasks successful, no new lint errors on the modified file

Screenshots / recordings

N/A — backend billing-math change, no UI surface.

Checklist

• I have tested my changes locally
• I have not included unrelated changes in this PR

[github-actions]

Thanks for updating your PR! It now meets our contributing guidelines. 👍