Repository navigation
Releases
0.18 — 네이티브 질문 대기와 터미널 검증
0.18은 네이티브 질문 대기 계약과 Codex·Claude Code 터미널 검증을 묶은 릴리즈입니다.
인터뷰에서 질문을 보낸 뒤 답변을 받기 전에 다음 결정으로 넘어가던 문제를 다룬다. Claude Code는 AskUserQuestion, Codex는 사용 가능한 모드의 request_user_input으로 질문하고 실제 응답을 기다린다. 접수 성공이나 미리 선택된 추천 항목은 답변으로 취급하지 않는다.
변경
- 네이티브 답변 대기를 기본으로 하고, 비동기 질문을 보낸 뒤 반복 대기하는 방식은 기본 경로에서 제외했다.
- 호스트가 질문 도구를 Plan 모드로 제한하면 그 조건을 알린다. 스킬이 모드를 임의로 바꾸거나 도구 제한을 우회하지 않는다.
- 비동기·평문 질문은 사용자가 그 방식을 명시적으로 받아들인 경우에만 쓴다. 기존 답변과 질문 ID를 보존한다.
검증과 한계
Codex CLI 0.153.4의 Plan 모드와 Claude Code 2.1.263 터미널에서 네이티브 선택지·직접 입력·응답 반환을 실행 검증했다. Claude의 첫 직접 입력은 값이 전달되지 않아 재질문했고, 재입력 후 정상 반환을 확인했다. 에이전트가 조작한 합성 시험이며, 스킬 자동 발견이나 실제 업무 전체 여정을 증명하지 않는다. 자세한 절차는 터미널 검증 기록에 있다.
데스크톱 사이드바의 ‘답변 요청’ 라벨은 실화면 검증을 완료하지 않았다. 질문 표시·도구 접수·응답 수신·사이드바 상태를 서로 다른 증거로 다룬다. 질문 대기 계약과 검증 범위를 참고한다.
Codex 상류 구현에는 일반 모드 질문을 허용하는 실험 설정도 있지만, Plan 요청만 is_blocking: true로 처리한다. 따라서 설정만 켜면 무기한 답변 대기가 보장된다고 안내하지 않는다. 이 릴리즈는 사용자 설정을 변경하지 않는다. 근거: 공식 요청 처리 코드.
0.17 — 두더띵 단일 진입점과 Claude Code 실검증
0.17 — 두더띵 단일 진입점과 실제 Claude Code 검증
v0.16 이후 main에만 있던 O’Reilly 맥락 학습, Ponytail 판단 기준, 도메인 모델링·테스트 증명·명세/규칙 분리 리뷰를 새 릴리스로 묶었습니다. 두더띵 하나를 호출하면 필요한 내부 지침을 선택합니다. 별도 API나 백그라운드 실행기는 아닙니다.
실제 Claude Code 2.1.263 / claude-sonnet-5 터미널에서 용어 충돌 질문과 승인된 정산 CLI 구현을 시험했습니다. 초기 시험에서 리뷰 지침 읽기 누락, low effort 재시험에서 원본 변이·손상 기록 이동을 발견해 실행 단계 지침을 보강했습니다. 실패 사례도 평가 문서에 포함했습니다.
최종 기본 설정 시험: 내부 테스트/리뷰 계약 읽기, 테스트 7개 통과, 별도 사본 변이 4개 검출, 독립 CLI 사례 5개 통과, 명세·손상 기록 원본 보존. 저장소 테스트 40개 및 변이 23개도 통과했습니다.
범위 제한: 합성 로컬 fixture이며 실업무 전체의 안정성, low effort 안정성, 모든 계정, Linear·O’Reilly 실연동이나 배포를 검증한 것은 아닙니다.
상세: docs/claude-code-evaluation-0.17.md
0.16 — 그릴·스펙 인수와 안전한 재개
기존 그릴·스펙을 인수하면 합의·ID·판·실행 권한을 유지하고 첫 미완료 단계부터 진행하도록 보강했습니다. 손상 체크포인트는 보존하며, 상태 복구와 완료 검증을 구분합니다.
변이 검사는 임시 사본에서 실행하고 테스트 0건 및 실행 건수 불일치를 실패로 처리합니다. 기존 브랜딩 변경은 이번 릴리스에 포함하지 않았습니다.
검증: 단위 테스트 40개 통과, 변이 23개 모두 검출, 예제 validator 통과 및 fictional 경고 확인. 별도 CLI 작업의 구현·테스트 2개와 원본 기록 보존도 확인했습니다. 실제 Linear 장애와 프로세스 강제 종료는 미검증입니다.
0.15 — one question per round, and Spec Kit when it is there
0.15 — one question per round, and Spec Kit when it is there
The interview contract now asks one frontier question per round and ends the turn on it. A live run on 2026-09-11 batched six questions with recommended answers into one message and then proceeded as if they were decided; the user never answered one. The contract already said to wait, but a batched list with recommendations reads as a proposal. No answer now pauses the interview — no Big, plan, Minis, or execution on an assumed answer — and a self-supplied answer is recorded as open. "Proceed without answering" is the only way a recommendation becomes a decision, with that instruction as its evidence. This is the third deliberate departure from grilling, recorded in docs/method.md. (#14)
When a repository already carries GitHub's Spec Kit — .specify/ plus installed speckit.* commands or speckit-* skills — the skill calls those commands wherever a step produces the artifact they write: constitution as a source, clarify into the frontier one question at a time, specify/plan/tasks for the artifacts, implement scoped to one Mini, analyze after checks, converge as one input to whole-journey verification. specify init is never run inside the workflow. Spec Kit's [X] marks are not proof; the Mini's own check is. Mapping is in references/speckit.md. (#13)
Not yet exercised: a live run under the one-question rule, and any run in a repository with Spec Kit on. Both are the next trials.
0.14 — the description parses again
0.13 moved the Korean triggers to the front of the description, which made the value begin with a double quote. To YAML that is a quoted scalar, and the English text after its closing quote is a parse error in any strict loader; the lenient loaders in use did not object, and the regex-based frontmatter test did not look. A one-word label now precedes the first quote so the value is a plain scalar, and the test asserts the three shapes a plain scalar cannot take: opening quote, colon-space, space-hash. The triggers still sit inside the first 60 characters.