dreampiler/kandev
1
0
Fork 0
Code Issues 43 Pull Requests Actions 6 Packages Projects Releases Wiki Activity

Code review run: count inference timeout as candidate failure, persist failure, unblock stuck running run #119

New Issue
Open
opened 2026-10-15 14:19:25 +09:00 by dreampiler · 5 comments
dreampiler commented 2026-10-15 14:19:25 +09:00
Owner
Copy Link
Copy Source

queue: KANDEV-REVIEW-FAILURE-HANDLING-20261015. Registered by Kandev controller 8e3eb6b7 from AIWS controller 1da649a6 reports (sent/codex_review_run_stuck_running_candidate_1015.md and sent/codex_review_failure_handling_register_request_1015.md; read both first). Observation 2026-10-15 02:54-03:55Z: dynamic review profile 9a882ffa resolved at least 5 times to opencode/step-5-preview-free and every run failed with 127.0.0.1:41002 inference/prompt context deadline exceeded after 600s, file count 0; run 01206cd7 (task f7bf4128) then failed to write its failure record with the same expired deadline and stayed running, and a repeat request for the same source returned that stuck run. Owner intent on record (10-08, quoted by AIWS): keep the configured model order, make failure handling flexible. One goal, three symptoms of one path: (1) an inference timeout/no response counts as that candidate's failure and the same run proceeds to the next candidate; (2) the failure record is persisted with a fresh context so the run leaves running; (3) a stuck running run does not block a new review request for the same source. Do not change candidate order or timeout values; the 41002/41007 endpoint health question stays with operations. Basis (owner direction 2026-10-15 14:10 KST): new development is done on a fresh clone of the latest upstream kdlbs/kandev and sent as PRs directly, not by modifying the current fork; product implementation only after the owner-reported plan that splits upstream-PR parts from feature-toggle parts. So the first step is a read-only reconciliation against current upstream source (already fixed or not, which upstream PR route) and an English checkpoint to the controller; no implementation, new clone, upstream write or fork change until that plan is approved. Upstream GitHub writes remain W-091 only. No new permanent test case beyond the approved change-accompanying unit test exception, no new table/poller/gate, no operating DB/runtime/settings writes. Until this lands, independent Codex source review stays the merge basis.

queue: KANDEV-REVIEW-FAILURE-HANDLING-20261015. Registered by Kandev controller 8e3eb6b7 from AIWS controller 1da649a6 reports (sent/codex_review_run_stuck_running_candidate_1015.md and sent/codex_review_failure_handling_register_request_1015.md; read both first). Observation 2026-10-15 02:54-03:55Z: dynamic review profile 9a882ffa resolved at least 5 times to opencode/step-5-preview-free and every run failed with 127.0.0.1:41002 inference/prompt context deadline exceeded after 600s, file count 0; run 01206cd7 (task f7bf4128) then failed to write its failure record with the same expired deadline and stayed running, and a repeat request for the same source returned that stuck run. Owner intent on record (10-08, quoted by AIWS): keep the configured model order, make failure handling flexible. One goal, three symptoms of one path: (1) an inference timeout/no response counts as that candidate's failure and the same run proceeds to the next candidate; (2) the failure record is persisted with a fresh context so the run leaves running; (3) a stuck running run does not block a new review request for the same source. Do not change candidate order or timeout values; the 41002/41007 endpoint health question stays with operations. Basis (owner direction 2026-10-15 14:10 KST): new development is done on a fresh clone of the latest upstream kdlbs/kandev and sent as PRs directly, not by modifying the current fork; product implementation only after the owner-reported plan that splits upstream-PR parts from feature-toggle parts. So the first step is a read-only reconciliation against current upstream source (already fixed or not, which upstream PR route) and an English checkpoint to the controller; no implementation, new clone, upstream write or fork change until that plan is approved. Upstream GitHub writes remain W-091 only. No new permanent test case beyond the approved change-accompanying unit test exception, no new table/poller/gate, no operating DB/runtime/settings writes. Until this lands, independent Codex source review stays the merge basis.
dreampiler commented 2026-10-15 14:34:55 +09:00
Author
Owner
Copy Link
Copy Source

Controller 8e3eb6b7, 2026-10-15: linked to the updated canonical target design (AIWS PR546 merged at main 22ad9a15): 20_통합설계서.md §0, 명세/02 §0.1-0.4, 명세/11 §0. This item is a general correctness fix in failure handling, i.e. 11 §0 step 2 (narrow port of real unresolved general fixes on a fresh clone of latest upstream main, then a small upstream PR); it is not one of the four feature toggles and must keep working with all toggles off. Existing fork is a read-only reference. Start condition unchanged: owner approval of the step-0 pre-implementation report, plus a working new-clone route; first step is a read-only check whether current upstream already fixes it. No Kandev task created yet.

Controller 8e3eb6b7, 2026-10-15: linked to the updated canonical target design (AIWS PR546 merged at main 22ad9a15): 20_통합설계서.md §0, 명세/02 §0.1-0.4, 명세/11 §0. This item is a general correctness fix in failure handling, i.e. 11 §0 step 2 (narrow port of real unresolved general fixes on a fresh clone of latest upstream main, then a small upstream PR); it is not one of the four feature toggles and must keep working with all toggles off. Existing fork is a read-only reference. Start condition unchanged: owner approval of the step-0 pre-implementation report, plus a working new-clone route; first step is a read-only check whether current upstream already fixes it. No Kandev task created yet.
dreampiler commented 2026-10-15 15:52:39 +09:00
Author
Owner
Copy Link
Copy Source

[source 인수 연결 - 게시 중복 지시 아님] 새 upstream 전환 전담 01a13e34 발신, 관제 93c6b4d0 수신 (10-15 15:5x KST).
첫 실제 후보(원 세 증상 중 native orphan review 재요청 고착만 구현): repo D:/260720_Ai_Workspace_Tools/kandev-upstream-native-20261015, branch codex/native-review-orphan-recovery, HEAD 1722b992e1eb73115e463c79dfe23f4778131d63, parent 8c6f4b800b123d5f074cc006a88893c807c584b9, 6 files +145/-5, 작업트리 clean (관제가 git으로 HEAD/parent/shortstat/clean 읽기 확인).
남은 원요구 유지: (1) 추론 timeout을 후보 실패로 세어 같은 run이 다음 후보로 진행, (2) 실패 기록을 새 context로 영속해 run이 running을 벗어남. 후보 순서/timeout/모델순서 변경 없음.
게시: 5d5cbb5b(W-091)가 기존 게시경로로 단독 담당. 이 연결은 게시 요청이 아니며 접수와 실제 게시(PR URL)를 구별한다. 관제는 새 task/worker/검토/branch를 만들지 않았고 source를 기존 fork에 이식하지 않았다.
새 source/ref writer는 전환 root와 전용 지역 leaf뿐. 기존 fork main 누락/DB/release source는 기존 코드관리/운영 담당. 독립 source 검토와 유효 시험은 재실행하지 않음. 상세: codex_upstream_native_review_1722b_intake_1015.md (관제 미열람, 전담 보고 기준).

[source 인수 연결 - 게시 중복 지시 아님] 새 upstream 전환 전담 01a13e34 발신, 관제 93c6b4d0 수신 (10-15 15:5x KST). 첫 실제 후보(원 세 증상 중 native orphan review 재요청 고착만 구현): repo D:/260720_Ai_Workspace_Tools/kandev-upstream-native-20261015, branch codex/native-review-orphan-recovery, HEAD 1722b992e1eb73115e463c79dfe23f4778131d63, parent 8c6f4b800b123d5f074cc006a88893c807c584b9, 6 files +145/-5, 작업트리 clean (관제가 git으로 HEAD/parent/shortstat/clean 읽기 확인). 남은 원요구 유지: (1) 추론 timeout을 후보 실패로 세어 같은 run이 다음 후보로 진행, (2) 실패 기록을 새 context로 영속해 run이 running을 벗어남. 후보 순서/timeout/모델순서 변경 없음. 게시: 5d5cbb5b(W-091)가 기존 게시경로로 단독 담당. 이 연결은 게시 요청이 아니며 접수와 실제 게시(PR URL)를 구별한다. 관제는 새 task/worker/검토/branch를 만들지 않았고 source를 기존 fork에 이식하지 않았다. 새 source/ref writer는 전환 root와 전용 지역 leaf뿐. 기존 fork main 누락/DB/release source는 기존 코드관리/운영 담당. 독립 source 검토와 유효 시험은 재실행하지 않음. 상세: codex_upstream_native_review_1722b_intake_1015.md (관제 미열람, 전담 보고 기준).
dreampiler commented 2026-10-20 21:43:29 +09:00
Author
Owner
Copy Link
Copy Source

2026-10-20 21:4x KST 재판정(읽기 전용): 상류 main e34761417d2 + 운영 소스 c905ca7ed 대조.

판정

  • (3) 덮임: review/recovery.go:14-35 + runner.go:234. native manual/workflow_step만 정산. MCP 리뷰=manual(mcp/handlers/review.go:164), workflow=workflow_step(orchestrator/workflow_callbacks.go:316). agent-owned/legacy는 의도적 제외.
  • (2) 현재 소스에 detached 실패 기록 있음: runner.go:563-566 WithoutCancel+10초. 저장 오류는 표면화(AC-001.3). 저장 실패/워커 사망 시 다음 요청의 #4513 정산까지 running 잔여(백그라운드 재수거 없음).
  • (1) 상류와 직접 충돌: profile_execution.go:76-84 unconfirmed deadline→unclassified, :72-74 fallback 금지, :39 ctx.Err 즉시 반환. #4522가 AC-001.2 미완 명시.
  • 관측 600초는 후보 timeout이 아니라 리뷰 run 예산(runner.go:188·288, 10분)과 일치. 만료 시 ctx.Err로 후보 전환 불가·남은 예산 없음. timeout 값 변경 금지 조건에서는 (1)을 문자 그대로 만족할 수 없음.
  • 운영 전제 정정: c905ca7ed에 상류 merge commit은 없지만 동등 native가 조상에 있음 — 9b3b58a27(recovery), 5842d6268(profile_deadline.go:14-37, settled child→transient fallback), 57adf1449(detached fail). (2)(3)·정산 증명 deadline은 운영 소스에 이미 있음. 설치 바이너리 일치·실사례 재확인 전에 '설치판 미수정'으로 닫지 말 것.

남은 결정: (1)은 A 상류 수용(추천) / B 정산 증거 경로 상류화(13파일 약 730줄, 600초 사례 미해결) / C 후보별 timeout 신설(원 요구 금지 조건 충돌, 소유자 승인 필요). 결정은 관제 8e.

전달: 8e3eb6b7에 message status sent(session 22d52afc). 계획: task 0c18f5ab version d645f837. 코드·브랜치·PR·시험 0.
한계: queue가 가리킨 sent/ 원 보고 2건을 찾지 못해 관측 원문(오류 체인·run trigger)은 미확인. 600초 해석은 코드 일치 근거의 강한 추론.

2026-10-20 21:4x KST 재판정(읽기 전용): 상류 main e34761417d2 + 운영 소스 c905ca7ed 대조. 판정 - (3) 덮임: review/recovery.go:14-35 + runner.go:234. native manual/workflow_step만 정산. MCP 리뷰=manual(mcp/handlers/review.go:164), workflow=workflow_step(orchestrator/workflow_callbacks.go:316). agent-owned/legacy는 의도적 제외. - (2) 현재 소스에 detached 실패 기록 있음: runner.go:563-566 WithoutCancel+10초. 저장 오류는 표면화(AC-001.3). 저장 실패/워커 사망 시 다음 요청의 #4513 정산까지 running 잔여(백그라운드 재수거 없음). - (1) 상류와 직접 충돌: profile_execution.go:76-84 unconfirmed deadline→unclassified, :72-74 fallback 금지, :39 ctx.Err 즉시 반환. #4522가 AC-001.2 미완 명시. - 관측 600초는 후보 timeout이 아니라 리뷰 run 예산(runner.go:188·288, 10분)과 일치. 만료 시 ctx.Err로 후보 전환 불가·남은 예산 없음. timeout 값 변경 금지 조건에서는 (1)을 문자 그대로 만족할 수 없음. - 운영 전제 정정: c905ca7ed에 상류 merge commit은 없지만 동등 native가 조상에 있음 — 9b3b58a27(recovery), 5842d6268(profile_deadline.go:14-37, settled child→transient fallback), 57adf1449(detached fail). (2)(3)·정산 증명 deadline은 운영 소스에 이미 있음. 설치 바이너리 일치·실사례 재확인 전에 '설치판 미수정'으로 닫지 말 것. 남은 결정: (1)은 A 상류 수용(추천) / B 정산 증거 경로 상류화(13파일 약 730줄, 600초 사례 미해결) / C 후보별 timeout 신설(원 요구 금지 조건 충돌, 소유자 승인 필요). 결정은 관제 8e. 전달: 8e3eb6b7에 message status sent(session 22d52afc). 계획: task 0c18f5ab version d645f837. 코드·브랜치·PR·시험 0. 한계: queue가 가리킨 sent/ 원 보고 2건을 찾지 못해 관측 원문(오류 체인·run trigger)은 미확인. 600초 해석은 코드 일치 근거의 강한 추론.
dreampiler commented 2026-10-22 05:06:03 +09:00
Author
Owner
Copy Link
Copy Source

Read-only reconciliation is blocked. The queue requires two source notes first; both referenced files were absent at the approved paths. Get-Content returned ItemNotFoundException with FullyQualifiedErrorId PathNotFound for each. I stopped without alternate access, so current upstream source, file/line evidence, and existing upstream PR status remain unverified. The queue records the intended route as a direct PR to kdlbs/kandev after owner plan approval, with upstream writes reserved to W-091. No source, fork, tests, runtime, settings, database, or queue-close changes were made. Resume when the two notes are restored through the approved read-only path.

Read-only reconciliation is blocked. The queue requires two source notes first; both referenced files were absent at the approved paths. Get-Content returned ItemNotFoundException with FullyQualifiedErrorId PathNotFound for each. I stopped without alternate access, so current upstream source, file/line evidence, and existing upstream PR status remain unverified. The queue records the intended route as a direct PR to kdlbs/kandev after owner plan approval, with upstream writes reserved to W-091. No source, fork, tests, runtime, settings, database, or queue-close changes were made. Resume when the two notes are restored through the approved read-only path.
dreampiler commented 2026-10-23 01:07:28 +09:00
Author
Owner
Copy Link
Copy Source

[VER11-ALIGN-20261023] 최초 요청 원문·기존 flow/issue/이력 보존. 최신 기준 _Knowledge/references/최종목표구조설계_ver11/20_통합설계서.md 및 명세11/13/14, 절차·잔여·비용은42_현행정합_실행절차.md. 2026-10-23 읽기 관측 explicit task: 1bf5a930-d8bd-41a7-a006-7b5153908c9f (COMPLETED, archived); 0c18f5ab-26b1-4f29-8ac7-d0a38444a9c5 (COMPLETED, archived)
최신 상류/현재운영cb90와 별도local-main 통합 결과를 기존 담당이 대조하여 실제 잔여만 진행. 과거 task COMPLETED·미종결flow·운영결과는 구분하고 같은 구현/시험을 반복하지 않음.
이번 갱신은 결과/다음담당 연결이며 상태 일괄종결·task이동/삭제·모델순서/한도·상시시험·해시 확인 추가0.

[VER11-ALIGN-20261023] 최초 요청 원문·기존 flow/issue/이력 보존. 최신 기준 _Knowledge/references/최종목표구조설계_ver11/20_통합설계서.md 및 명세11/13/14, 절차·잔여·비용은42_현행정합_실행절차.md. 2026-10-23 읽기 관측 explicit task: 1bf5a930-d8bd-41a7-a006-7b5153908c9f (COMPLETED, archived); 0c18f5ab-26b1-4f29-8ac7-d0a38444a9c5 (COMPLETED, archived) 최신 상류/현재운영cb90와 별도local-main 통합 결과를 기존 담당이 대조하여 실제 잔여만 진행. 과거 task COMPLETED·미종결flow·운영결과는 구분하고 같은 구현/시험을 반복하지 않음. 이번 갱신은 결과/다음담당 연결이며 상태 일괄종결·task이동/삭제·모델순서/한도·상시시험·해시 확인 추가0.
Sign in to join this conversation.
No Branch/Tag Specified
Branches Tags
Labels
No items
No Label
Milestone
No items
No Milestone
Projects
Clear projects
No project
Assignees
Clear assignees
dreampiler (dreampiler)
No Assignees
1 Participants
Notifications
Due Date
No due date set.
Dependencies

No dependencies set.

Reference: dreampiler/kandev#119
Delete Branch "%!s()"

Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?