Activity

1. seonghobae

   self-assigned this

   on Sep 3, 2026
2. seonghobae commented on Sep 3, 2026

   

   seonghobae

   on Sep 3, 2026 – with ChatGPT Codex Connector

   ContributorAuthor

   More actions

   실행 중간 기록 — 지정 장애 run RCA와 진척률

   완료율은 그대로 **4/41 = 9.8%**로 유지합니다. 아래 source fix·RCA·owner 지정은 운영 재검증 전까지 완료로 올리지 않습니다.

   Item 4 — 응답 정합성·구체적 telemetry: 미완료

   • newsdom-api run 33510285421, job 99864028341
     • credential/token/CO sidecar/orchestrator/free preflight 성공
     • 최종 실패: 모델 출력 JSON 문법 오류, line 1 column 2223, 응답 길이 3,713자, SHA-256만 제공
     • schema path, provider/model, structured-output repair 단계, parser context가 최종 오류에 없음
   • html4tree run 33560972491, job 100033086428
     • 초기 실패: JSON 문법 오류, line 1 column 1530, 응답 길이 1,890자
     • 옛 caller가 repair용 두 번째 LLM 요청을 만들고 NoemaRepairDeadlineExceeded 900초에 도달

   보호된 central source의 commit a28fc2f4e185df7847e2f2f5f6ec561d1e84805d는 caller repair·900초 deadline·두 번째 모델 호출을 제거하고 CO에 timeout/repair/failover를 귀속했습니다. 하지만 새 source로 실제 consumer canary가 성공하기 전에는 item 4를 닫지 않습니다. 남은 acceptance는 provider/model/phase/attempt/elapsed/schema violation을 typed receipt로 남기고 malformed raw output을 노출하지 않으면서 운영자가 고칠 수 있는 원인을 제공하는 것입니다.

   Item 39 — 900초 제한: source-fixed, 운영 미검증

   지정 CO run 33580381913, job 100093226133에서:

   1. 첫 모델 verdict가 request_changes인데 adversarial_validation.status=failed가 없어 local semantic contract failure
   2. 옛 caller가 repair request 생성
   3. caller-owned 900초 absolute deadline이 두 번째 요청을 종료

   a28fc2f4...에서 이 경로는 제거됐습니다. 사용자 지시의 “3시간으로 늘리기”보다, 공통 기본 timeout null·CO owner·provider termination이라는 현행 아키텍처가 더 정확합니다. 다만 queued old-source runs가 계속 materialize되므로 운영 canary 전까지 미완료입니다.

   Item 40 — fast-mlsirm Noema failure: 전제 정정, 미완료

   run 33646974279, job 100304078562:

   • step 7 credential selection: 성공
   • step 8 repository-scoped app token: 성공
   • step 12 CO sidecar + orchestrator/free: 성공, 54 routes admitted, 5 ready
   • step 13: 649.5초 후 CO HTTP 500, phase=connecting, served_model=unknown

   따라서 credential failure가 step 13을 유발했다는 전제는 로그와 다릅니다. 실제 문제는 오래 queue된 run이 옛 workflow/sidecar source로 실행된 것, gateway internal 500의 구체적 원인이 unknown으로 남는 것, current source canary 부재입니다.

   Item 41 — startup failure: 미완료

   • Wardnet run 33710719228: CodeQL PR, startup_failure, Jobs API [], job/runner/checkout 없음
   • run head 46ede...와 연결 PR auto-merge 대기 PR의 update-branch 처리 보강 #140 current head 9389a... 불일치
   • .github source-fix run 33761818467: 같은 형태의 job 0개 failure

   #712와 #1150에 startup_failure_before_job_materialization, stale_workflow_source_materialized, runner_id=0, source/test/credential/gateway failure를 분리하도록 증거를 전달했습니다.

   Item 30 — OpenCode startup UX: source-fixed, 운영 미검증

   옛 run 33548447878은 stale event head가 live head와 다르다는 이유로 step 2에서 red 처리했습니다. 보호된 main은 이제 open/ready PR의 moved head를 성공적으로 self-retire하고 fresh dispatch를 기다리도록 변경되어 있습니다. current-source canary가 없는 관계로 완료 체크는 보류합니다.

   다음 owner 작업

   • epic(actions): make Noema, Strix, and OpenCode concurrency exact-PR and current-head safe #1800/fix(noema-review): close item13 stale-head native-cancellation race #1797/fix(noema-review): authorize sibling repository_dispatch cleanup with a target-scoped token #1799: Noema·Strix·OpenCode exact-PR concurrency
   • fix(actions): remove duplicate queue hygiene from the organization sweep #1796: organization sweep의 중복 Actions inventory 제거
   • ops: diagnose and bound organization GitHub Actions queue starvation #712/feat: add read-only Actions queue health evidence #1150: 60-job/startup/stale-workflow-source 분모
   • CO 🛡️ Sentinel: [보안 개선] 커맨드 인젝션 위험 완화를 위해 shell=False 명시 #1007: Models.dev field-level synthesis auto-merge
   • CO fix(redaction): keep Actions group markers from wiping JSON evidence #1031: OpenCode Go shared-key successor, fix(ci): identify trusted uv downloader #1008 유효 delta 승계

   다른 세션은 위 owner 중 파일 충돌이 없는 하나를 맡고 RED→GREEN→exact-head→merge/canary를 이 이슈에 남겨 주십시오.

   기다리는 동안 다른 로컬 세션에게 지속적으로 지시 받아 도우세요. 계속 묻고 계속 피드백하고 계속 지원하세요. 더 주도적이고 더 적극적으로요.
3. seonghobae commented on Sep 3, 2026

   

   seonghobae

   on Sep 3, 2026 – with ChatGPT Codex Connector

   ContributorAuthor

   More actions

   Item 40 fresh causal evidence from ContextualWisdomLab/fast-mlsirm#1536 confirms the Noema credential-lifetime defect remains live and is not a leaf/source failure.

   Exact consumer canary:

   • protected fast base: main@b5a3a0c1057d4b53d7a4bb18e0de69f630c2b45c
   • unchanged PR head: 77bef27cff780b909be484b52be35e97be752780
   • required Noema run/job: 33634978342 / 100263465886
   • trusted central workflow source used by the run: .github@9330d41c92b1e6ab35261f3f5189936ea1ad8bff
   • runner and exact-head validation succeeded; repository-scoped cwl-noema-review installation token was minted at 2026-09-03T00:26:23Z.
   • contextual-orchestrator sidecar came up and orchestrator/free preflight was healthy; Prepare Noema model verdict started at 00:49:57Z.
   • at 01:44:26Z, ~78 minutes after token mint, the prepare phase failed at a GitHub CLI call with gh: Bad credentials (HTTP 401); post-job cleanup independently reported Token expired, skipping token revocation.

   This exactly matches the defect previously RED/implemented in .github#1745 (12b44ee781daf7bcd88954e62c8ecb2d8784d20f), but #1745 is closed, unmerged. Fresh comparison against protected .github/main@269e5bd9e65c38770a827af1291a5657d5cfcd01 is diverged, ahead_by=1, behind_by=75; its unique delta still changes scripts/ci/noema_review_gate.py, two_phase.py, focused tests, doctoring, and CHANGELOG. Default-branch code search finds no NoemaRepairRecheckUnavailableError, so no verified successor carryover is visible.

   A second fast canary (#1690, f811f471..., run 33633945304) shows a related but distinct prepare-path failure: the model contract repair hit the hard 900-second NoemaRepairDeadlineExceeded; cleanup again reported the repository App token expired. Do not collapse that deadline defect into the 401 defect; item 39 owns the 900s policy, item 40 owns credential lifecycle.

   Required owner repair / acceptance for item 40:

   1. Preserve fix(noema): fail closed when repair-retry live-head recheck outlives its token #1745's valid RED and fail-closed tests or explicitly supersede them with byte/contract-equivalent evidence; a closed-unmerged valid delta is not completion.
   2. Do not carry a GitHub App installation token across an unbounded model/provider phase. Any GitHub live-head/review revalidation that can occur after model execution must obtain a fresh repository-scoped credential at that boundary (or equivalently move the GitHub recheck outside the long-lived credential lifetime); the model/CO subprocess must not receive GitHub credentials.
   3. Preserve a separately fresh publication credential and exact live head/base validation before any review is submitted.
   4. Add a realistic >1h token-age regression for both initial prepare and repair-retry paths, plus fail-closed behavior when refresh/revalidation is unavailable; do not broaden permissions, weaken Noema, or impose a shorter model timeout as a workaround.
   5. GREEN is a current protected .github successor plus a fresh unchanged fast-mlsirm canary where the same long-running path survives past the old token lifetime, performs post-model GitHub revalidation with a fresh credential, and publishes only after exact-head/base revalidation.

   No fast-mlsirm source churn or retrigger commit is indicated by this failure.
4. seonghobae commented on Sep 3, 2026

   

   seonghobae

   on Sep 3, 2026 – with ChatGPT Codex Connector

   ContributorAuthor

   More actions

   Item 40 revalidation requested in the program body is now completed against the named fast-mlsirm canary 33646974279 (#1518@b8e72773c34cd2f383bf44f492e52bf61736c680, job 100304078562).

   Result: the designated run does not reproduce the earlier GitHub-App credential-expiry failure. Trusted workflow source was .github@bbe65f08b1ae663c467be343e8fd5a98881eb686. It minted the repository-scoped reviewer token at ~02:32Z, performed live-head validation, provisioned contextual-orchestrator, and later post-job cleanup reported Token revoked at 02:52Z. There was no gh: Bad credentials / expired-token failure. The workflow also contains a separate Refresh repository-scoped Noema GitHub App token for publication step, although that publication refresh was skipped because verdict preparation failed first.

   The run instead failed at a different owner boundary: after healthy orchestrator/free discovery/readiness and a successful chat/completions preflight, the actual Noema request spent 649.5s in phase=connecting and returned gateway HTTP 500 with served_model=unknown. That new causal evidence has been handed to canonical contextual-orchestrator owner issue ContextualWisdomLab/contextual-orchestrator#1016 (comment 5527940637) because the leaf cannot safely identify selected/attempted provider/fallback outcomes from the public error contract.

   Do not mark item 40 fully complete from this canary alone: its credential lifetime was only ~20 minutes, so it does not prove the >1h path that failed on fast-mlsirm#1536 run/job 33634978342/100263465886. The earlier acceptance still stands: prove a long model phase crossing the old installation-token lifetime, obtain/revalidate fresh GitHub authority at the post-model boundary, keep publication on a separately fresh credential, and preserve exact live-head/base validation. But the named 33646974279 recheck itself is now classified: credential failure not reproduced; terminal failure moved to the contextual-orchestrator transport/evidence lane.
5. seonghobae commented on Sep 7, 2026

   

   seonghobae

   on Sep 7, 2026 – with ChatGPT Codex Connector

   ContributorAuthor

   More actions

   2026-09-07 추적 원장 정합성 점검

   현재 본문 완료 체크는 8·9·10·25·31의 5개인데 상단은 **4/41 = 9.8%**입니다. 이를 곧바로 5/41로 올리면 안 됩니다. 본문 근거 중 source/tests·정책 적용과 공통 DoD의 운영 재검증은 범위가 다릅니다. 각 완료 항목의 보호 병합 SHA와 consumer canary/운영 원시 증거를 재확인하기 전까지 이 원장의 전체 완료율은 재검증 필요 상태입니다. 이번 점검으로 새로 완료 인정한 항목은 0개입니다.

   추가 조사 연결: #2000의 실패 후 route 반복 선택 보고는 항목 4·39와 Actions runner 점유에 연결됩니다. 보고된 909회 시도/837회 단일 route/소진 이벤트 0회는 아직 이 작업에서 원본 artifact를 독립 검증하지 않았습니다. CO 담당 작업에 request-ID별 분모, 동일 logical request와 caller 재호출 구분, 기존 owner의 소진 계약 조사를 요청했습니다. 경과 시간만으로 model timeout을 추가하는 근거로 사용하지 않습니다.

   항목 13·16·17의 현재 변경은 #1962 exact f162c6b(Draft, 로컬 전체 2994 passed/1 skipped), #1996 후속 로컬 9f08ed0(282 passed, 미push)입니다. #1996 기존 Strix job101590121689는 API에서 실제 in_progress를 확인해 유지했습니다. 이 수치들은 보호 병합이나 조직 전체 처리량 개선의 증거가 아닙니다.
6. seonghobae

   added

   maintenance

   priority: mediumNormal-priority or P2 workNormal-priority or P2 work

   on Sep 7, 2026