@qwen-code /triage

Thanks for the PR!

Template looks good ✓ — headings present, reviewer test plan with a concrete failing-then-passing assertion.

On direction: this is a clean, obvious bug fix. Math.round on each half independently doesn't preserve the invariant that the parts sum to the total — classic rounding error. Token accounting (billing, logging, telemetry) needs that invariant, so this is worth fixing. No direct Claude Code CHANGELOG reference, but the area (usage/token tracking) is clearly core.

On approach: the fix is exactly the minimal change — one line per path (non-streaming + streaming), deriving the completion half from the remainder. No scope creep, no unrelated edits. The test directly asserts the invariant that was broken. Nothing to cut.

Moving on to code review. 🔍

— Qwen Code · qwen3.7-max

Code Review

Independent proposal (written before reading the diff): the fix is obvious — compute one half with Math.round, derive the other from the remainder so the invariant holds. The PR does exactly this, applied consistently to both non-streaming (convertOpenAIResponseToGemini) and streaming (convertOpenAIChunkToGemini) paths. The approach matches perfectly.

No correctness bugs, no security issues, no regressions. Both paths are updated symmetrically. The comment is updated to explain the rationale. The test directly asserts the broken invariant. No AGENTS.md violations — minimal, focused, no abstractions, no scope creep.

Nothing to flag. ✅

Test Results

Before (main branch, without fix)

Running the new test against the original code:

cd packages/core && npx vitest run src/core/openaiContentGenerator/converter.test.ts \
  -t "keeps the estimated prompt/completion split"

 FAIL  src/core/openaiContentGenerator/converter.test.ts > OpenAIContentConverter > keeps the estimated prompt/completion split summing to total tokens

AssertionError: expected 6 to be 5

- Expected  5
+ Received  6

 ❯ src/core/openaiContentGenerator/converter.test.ts:2722:9
     2720|       expect(
     2721|         (usage?.promptTokenCount ?? 0) + (usage?.candidatesTokenCount …
     2722|       ).toBe(5);

 Test Files  1 failed (1)
      Tests  1 failed | 133 skipped (134)

Bug confirmed: Math.round(5 * 0.7) + Math.round(5 * 0.3) = 4 + 2 = 6, not 5.

After (this PR)

cd packages/core && npx vitest run src/core/openaiContentGenerator/converter.test.ts

 ✓ src/core/openaiContentGenerator/converter.test.ts (134 tests) 57ms

 Test Files  1 passed (1)
      Tests  134 passed (134)

All 134 tests pass, including the new invariant check. The fix correctly derives completionTokens = totalTokens - promptTokens so the split always sums to the total.

— Qwen Code · qwen3.7-max

This is a textbook bug fix — clean, minimal, and correct.

The rounding invariant (prompt + completion = total) was broken by independent Math.round calls on each half. The fix derives one half from the remainder, which is the standard technique. Both code paths (streaming and non-streaming) are updated symmetrically. The test directly asserts the invariant and fails convincingly before the fix (expected 6 to be 5) and passes after.

The diff is exactly what this bug needs: 2 production lines changed, 1 test added, nothing else. No abstractions, no drive-by refactors, no scope creep. If I had to maintain this in six months, I'd have nothing to complain about.

Approving. ✅

— Qwen Code · qwen3.7-max

✅ Local verification — recommend merge

Verified this on a clean worktree (real npm ci, full tsc build and esbuild bundle into the shipped dist/cli.js). Exercised the change at three levels: the author's unit test, an A/B on the compiled converter covering both code paths, and a real end-to-end run of the bundled CLI under tmux against a mock OpenAI‑compatible server. The fix is correct, minimal, and complete.

Scope confirmed

True diff vs origin/main: 38 / 4, 2 files — a 4‑line behavior change in convertOpenAIResponseToGemini (non‑streaming) and convertOpenAIChunkToGemini (streaming), plus one unit test. Completion is now derived as the remainder:

finalPromptTokens = Math.round(totalTokens * 0.7);
finalCompletionTokens = totalTokens - finalPromptTokens; // was Math.round(totalTokens * 0.3)

A repo‑wide grep confirms these are the only two sites using the 70/30 estimate, so the fix is complete. The if (totalTokens > 0 && promptTokens === 0 && completionTokens === 0) guard is unchanged, so providers that report a real prompt/completion breakdown are untouched (no regression).

1) Author's Reviewer Test Plan — reproduced exactly

• Revert the 4‑line fix → the new test fails with expected 6 to be 5 ✔ (matches the PR description)
• On the PR branch → full converter.test.ts suite 134/134 passing ✔

2) A/B on the compiled converter — both paths

Imported the built OpenAIContentConverter from dist/ and fed {prompt_tokens:0, completion_tokens:0, total_tokens:N} for N ∈ {1,2,3,5,7,10,15,25,35,100,101} through both functions:

Example N=5: OLD → 4 + 2 = 6 ≠ 5; NEW → 4 + 1 = 5.
⚠️ The PR's added unit test only covers the non‑streaming function. The streaming convertOpenAIChunkToGemini is fixed too and is verified consistent here.

3) Real end-to-end under tmux

Drove the bundled qwen --output-format json (real HTTP, streaming — confirmed stream_options.include_usage in the mock's request log) against a mock server that reports only total_tokens: 5:

The CLI also issues a background managed-auto-memory-extractor call, so the session aggregates two calls. Under the old code both calls overshoot and the inconsistency propagates into the session‑level usageMetadata (promptTokenCount + candidatesTokenCount > totalTokenCount) — exactly the billing/telemetry symptom the PR describes. The fix makes every call and the session reconcile.

Lint / build

eslint clean on both changed files; core tsc build passes.

Optional (non‑blocking)

Consider adding a sibling assertion for the streaming convertOpenAIChunkToGemini path in the unit test, since the fix touches it as well — behavior is already verified correct above.

finalPromptTokens = Math.round(totalTokens * 0.7);
finalCompletionTokens = totalTokens - finalPromptTokens; // 原来是 Math.round(totalTokens * 0.3)