fix: record first-output timing for tool-only Responses streams

[xyjk0511]

Summary

• Record first-output timing for nonempty response.function_call_arguments.delta and response.custom_tool_call_input.delta events in native Responses SSE, and nonempty tool_call_delta arguments in the adapter bridge. Tool-only turns otherwise finish with output-token usage but no first-output timestamp.
• Preserve once-only reporting and exclusions for empty deltas, tool scaffolding, lifecycle events, and steering/injection controls. No TPS arithmetic, routing, credentials, or request-body logging changes.
• Add regressions using actual bridge streams and fragmented native SSE, including custom tools, later prose, completion, and invalid/empty input. Update the transport contracts and dashboard documentation to clarify the observation boundary.
• This is a narrow upstream version of a locally verified fix. First-output timing remains a proxy observation; it does not measure the start of hidden reasoning or exact decoding throughput.

Verification

• Before the production fix, the updated tool-only bridge regression and custom-tool regression failed; both native SSE tool-delta regressions also failed at the first-output assertion before later prose.
• bun test tests/adapters/bridge.test.ts tests/server/response-log-inspection.test.ts tests/responses/sse-inspector-bounds.test.ts tests/server/management-api-logs-metrics.test.ts: 183 passed, 0 failed.
• bun run typecheck: passed.
• bun run privacy:scan: passed.
• bun run structure:check: passed.
• cd docs-site && bun install --frozen-lockfile && bun run build: passed; 545 pages built and 74,644 internal links checked.
• git diff --check: passed.
• bun run test:changed: started through the repository wrapper, resolved the dev comparison commit, and launched Bun's import-connected changed test lane. Stopped after several minutes without a returned lane result on this Windows machine; not counted as passing. This broader result and the full repository suite remain outstanding before review readiness.
• The full repository suite has not been run. This PR stays draft while broader validation and repository review-readiness requirements remain outstanding. Focused coverage exercises the changed stream boundaries; it is not a full-suite claim.
• No upstream GUI files changed, so no GUI screenshot is required.

Checklist

• Scope stays focused and avoids unrelated cleanup.
• Docs or release notes were updated when needed.
• Security-sensitive changes were reviewed for secrets, auth, and unsafe defaults. No auth or credential behavior changes; privacy scan passed.

Review readiness checklist

This PR stays in draft until every box below is ticked. Tick all four boxes once the requirements are met:

• Required local validation passed; commands, results, and any full-suite exception are documented.

• I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).

• I resolved all correct Codex and CodeRabbit findings.

• My PR is ready for review.

[coderabbitai]

Important

Draft PR not reviewed

Draft PRs are not automatically reviewed by default.

• Trigger a manual review

To automatically review draft PRs, update your CodeRabbit configuration:

reviews:
  auto_review:
    drafts: true

---

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.}

[github-actions]

✅ Deterministic PR hygiene checks passed.

[github-actions]

⏳ DRAFT

• review readiness checklist open (0/4 boxes ticked).

What to do

• Tick all four boxes in the PR description once you're done (currently 0/4).

Review readiness checklist

• ⬜ Required local validation passed; commands, results, and any full-suite exception are documented.
• ⬜ I pushed my PR to a recent dev commit (at most 10 behind; a maintainer may still ask for the exact tip before merge).
• ⬜ I resolved all correct Codex and CodeRabbit findings.
• ⬜ My PR is ready for review.

0/4 boxes ticked.

This PR stays in draft until every box above is ticked.

[lidge-jun]

리뷰 · 우선순위 62 / 80

이 PR은 Responses 스트림에서 “첫 출력” 시각을 찍을 때, 글자·추론만 보던 구멍을 고칩니다. 도구만 나오는 턴은 예전에 출력 토큰은 있는데 첫 출력 시각이 비었습니다. 네이티브 SSE는 비어 있지 않은 response.function_call_arguments.delta와 response.custom_tool_call_input.delta를, 어댑터 다리는 비어 있지 않은 tool_call_delta 인자를 첫 출력으로 칩니다. 빈 delta, 도구 시작 알림, steer/inject 같은 제어 프레임은 예전처럼 타이머를 안 켭니다. 한 번만 보고하는 규칙도 그대로입니다. 대시보드·transport 문서에 “프록시가 처음 본 출력”이지 숨은 추론 시작이나 정확한 디코딩 속도가 아니라고 적어 두었습니다. base는 dev입니다. 초점은 좁고, TPS·라우팅·자격 증명·요청 본문 로그는 안 건드립니다.

tests/adapters/bridge.test.ts - 빈 인자·공백 " "·"{}"를 넣어, 공백도 length>0이면 첫 출력으로 칩니다. 글자/추론과 같은 규칙이라 일관됩니다. 다만 도구 스트림이 패딩 공백만 먼저 보내면 TTFT가 아주 이르게 보일 수 있습니다.
src/server/responses/native-steering-log.ts - 스티어링 로그 쪽은 이미 *.delta 문자열이면 첫 출력으로 칩니다. 이번 수정 대상은 메인 inspector/firstOutputFromParsed와 다리입니다. 경로는 둘로 나뉘어 있어도, 도구-only 구멍은 이번으로 메워집니다.
작성자 본문 - focused 183테스트·typecheck·privacy·structure·docs 빌드는 통과했다고 적혀 있습니다. test:changed는 중간에 멈췄고, 전체 스위트는 안 돌렸습니다. review readiness 체크리스트는 아직 0/4이고 PR은 draft입니다. tip of dev와는 커밋 1개(무관한 Anthropic threshold #6207)만큼 갈라져 있습니다.

메인테이너의 판단이 필요한 지점

도구 인자의 공백 한 글자도 첫 출력으로 둘지, trim 후에만 칠지. draft·체크리스트·전체 스위트를 머지 전 필수로 둘지, 경계 테스트+호스트 CI만으로 충분한지.

너의 추천

동작 방향은 맞고, 다리·네이티브·문서가 한목소리입니다. 공백 규칙은 지금처럼 length>0으로 두고, tip of dev에 맞춘 뒤 readiness를 채운 다음 머지하세요. 미리보기 배포 이야기는 이 PR과 무관하니 건너뛰세요.

이 댓글은 grok-bot이 작성했습니다