LLM WikiAccess-protected knowledge portal

WIKI

2026-W32 Worklog

AI Summary Purpose Record cross repository development and review work for ISO week 32 of 2026. Key points 2026 08 03 Reviewed BTS branch feature/request variant in post at 94ae92a against main cf904ad without modifying the branch. The fixe

경로ai/worklog/2026/2026-W32.md
카테고리Worklog
태그#ai-review #portfolio #report #study #w32 #worklog

# 2026-W32 Worklog

AI Summary

Purpose:

Key points:

94ae92a against main cf904ad without modifying the branch.

GET default, and core response flow worked in targeted checks.

non-strict body parsing coerces booleans/integral floats into file indices, the new JSON surface lacks resource bounds, and no automated POST contract tests were added.

Saramin, JobKorea, and publicly indexed Remember pages. SOCAR Platform Engineer (DBA) was the only unconditional Apply now result after role-fit, posting-status, finance, welfare, and culture checks.

welfare, culture, source-quality, and role-fit evidence. Corrected SOCAR's finance label to include net loss and liquidity/leverage risk, and corrected NICE JiniData's parent from NICE Information & Telecommunication to NICE Information Service.

reviews at desktop and mobile widths after CJK wrapping, mobile table-scroll cues, semantic tokens, and per-link source-tier labels were verified.

moved the application into active interview preparation, and linked the Korean interview-preparation source report and PDF from the status surfaces.

links under https://llm-wiki.hwlabs.dev/, using /portfolio/ for the portfolio and /portfolio/items/<slug> for public detail pages.

human/career/remember-interview/ (submitted resume/portfolio web copies, assignment+architecture page with three rendered Mermaid SVGs, submitted PDF copies), fixed five typos in the guide source, added 13 keyword memos (54 total) and desktop hover tooltips, and repaired a corrupted git index (stale .git/index.lock had staged 1,455 file deletions).

repeated-reading format and generated a 25-page browser-rendered PDF plus an interactive web document with in-place keyword modals. Added seven non-duplicative study chapters for the gaps visible in Remember's public job and architecture materials, and linked them from the PDF through llm-wiki.hwlabs.dev.

Relevant when:

the pending labgit2 admin-password rotation.

Do not read full document unless:

Linked documents:

Open Questions

deployed topology.

once in an agent session (agent could not rotate it itself, destructive action blocked by the harness's permission classifier); the controller rotated it directly afterward, verified functionally.

Forgejo git-fsck strictness (confirmed real upstream object defects via direct API reproduction, not a ghost-fix bug) — needs a policy decision (exclude from the full-corpus run vs relax [git.fsck] strictness) before the multi-day full run.

Details

2026-08-07 — Remember interview hub, submitted-document web copies (Claude)

- index.html: hub linking guide/PDF/study map/study series, submitted documents, and the assignment page, plus a 5-step prep sequence. - resume.html: byte-identical copy of the submitted resume HTML source (career/클로드이력서/리멤버앤컴퍼니-data-engineer/), plus the exact submitted resume/portfolio PDFs (hashes match 최종 이력서/2026-07-19/). - portfolio.html: static snapshot of selected-portfolio-pdf.html with the same 8 items/order as the submitted PDF. Item body text follows the CURRENT detail pages (two item titles have drifted since 07-19); the submitted PDF is linked alongside as the source of truth. - assignment.html: assignment brief summary, verified result counts, the three ARCHITECTURE.md Mermaid diagrams rendered to SVG (diagrams/*.svg, htmlLabels=false — HTML labels break standalone SVG loaded via <img>), seven README design decisions, known limits.

- Fixed typos (나눅니다→나눕니다, 덮어리→청크 단위, 받침하기→떠받치기, 재면 수치는→잰 수치는) and a nested-bold artifact in the fact guard. - Added 13 keyword memos (STAR, near-miss, quarantine, XCom, outbox/operation marker, lease/publish pointer, dual-read, fail closed, backoff, primary/replica, micro-batch, Arrow/Polars, data warehouse); total 54 memos, 195 in-text triggers. - Desktop hover now shows a tooltip (full name + easy explanation); click/tap still opens the full modal. Web copy drops the machine-local file:///mnt/c/Windows/Fonts @font-face lines (fallback family is identical); the PDF render path keeps them. - /remember redirect must live in the build-wiki.mjs template — human/_redirects is regenerated on every wiki build.

corrupted index that had 1,455 files staged as deleted; working tree was untouched.

passes; Playwright QA at 1280px/390px confirms no broken images, no page-level horizontal scroll, hover tooltip, click/tap modal, and Escape close on the live-copy HTML.

(verified via plain curl), while pages describe it as Access-protected. Flagged to the user; Cloudflare Access setup remains Needs confirmation.

2026-08-07 — Remember first-round interview schedule confirmed

format, and panel composition remain Needs confirmation.

the PDF.

career status page and generated Wiki document, and kept the legacy output/pdf/ artifact in sync.

to existing Airflow, data-quality, data-contract, MySQL, distributed-systems, security, SRE, Spark, Kafka, and binlog study chapters. The map explicitly preserves the boundary between studied concepts and production experience.

2026-08-07 — Remember interview guide redesign and study gap fill

who do not know the domain, plus a short glossary for the terms that recur in the interview answers.

and a matching interactive web document. Question cards separate the first answer from optional detail, and the first sentence and load-bearing metrics use the canonical resume marker: linear-gradient(transparent 56%, #ffe27a 56%).

description and the public AWS/Remember S3 Tables architecture articles. Added only the uncovered topics: entity resolution/MDM, Full Load-to-CDC handoff, S3 Tables maintenance, PyIceberg metadata observability, Lake Formation governance, Glue/PySpark Full Load, and StarRocks/Iceberg serving.

human/study/content/remember-data-engineering/, updated the curriculum, and linked every chapter from the PDF's final study section using https://llm-wiki.hwlabs.dev/.

explanation, an expert explanation, and the point to retain for this interview. The web document contains 178 visible keyword triggers; clicking a trigger opens an in-place dialog and clicking the backdrop or pressing Escape closes it. Keyword definitions are no longer separate PDF pages.

"say that..." into direct first-person answers that can be spoken as-is.

generation, identical local/web PDF hashes, 11 custom-domain external links, and the interactive modal behavior. The repository-wide relative-link checker still reports the unrelated pre-existing missing page human/study/dist/ai-frontier/22.html.

deployment: the public PDF hash differed from the new 25-page artifact, the new study URL fell back to the integration hub, and the public study map retained old links. A Cloudflare Pages deployment is still required; do not describe these local artifacts as live until it completes.

2026-08-05 — Remember & Company take-home pass and interview preparation

2026-08-05. The application is now in active interview preparation.

the pass was confirmed on 2026-08-05.

human/reports/2026-08-05-remember-data-engineer-interview-prep.md and output/pdf/2026-08-05-remember-data-engineer-interview-prep.pdf.

Needs confirmation.

surfaces without changing unrelated 2026-08-04 career updates.

2026-08-03 — BTS fixed-URI POST branch review

Scope:

changed or posted.

Positive verification:

requirements.

/lab-master/v4/download, JSON fields preserve status/reason/msg, import_index, repeated idx/ec, and client_mode, and X-Last-Received-Hash remains a header.

parameter forwarding.

Blocking findings:

  1. download-server/app/main.py still derives updater success/error logs only

from query parameters. Direct POST body status therefore does not produce the operational updater 정상 완료 / updater Error log required by existing test and deployment documentation.

  1. BtsDownloadBody.file_identifier: int | str is coercive. Empirical checks

showed JSON true, false, and 1.0 become numeric indices instead of returning 422; this contradicts the documented POST contract and GET semantics. Empty identifiers and extra fields are also accepted by the boundary model.

  1. The new unauthenticated JSON endpoint has no application-level body-size,

string-length, or list-count limits. In relay mode, a request with status can also create an upstream background task. If ingress does not enforce limits, this is a memory/log/socket exhaustion path.

  1. The branch adds no tests. updater/internal/client has no test files and

updater/internal/config/config_test.go is effectively empty. FastAPI route, Go request serialization, GET fallback, hash header, status logging, and direct/proxy/relay parity remain unguarded.

Important follow-ups:

the same operational evidence.

pre-buffer body limit and bounded relay concurrency at the application or verified ingress.

validate BTS_REQUEST_METHOD instead of silently treating every non-POST value as GET.

spec/test-scenarios.md.

spec/error-handling.md, deployment scenarios, and release notes. Publish as a new image version instead of silently changing the deployed 4.0.0 tag.

Review limitation:

based on the fetched branch ref, commit history, code, specs, and local execution.

pre-existing code creates /app/data/log at import time. This was recorded as an environment/pre-existing issue, not blamed on the POST diff.

2026-08-04 — Codex independent job-market cross-check

Scope:

status/details for likely matches.

MySQL DBA, and DBA, plus public JobKorea candidate pages.

bodies were unavailable, so this channel is explicitly incomplete.

benefits, and the limited culture evidence available without inventing review conclusions.

Decision:

role-fit evidence coverage 85%.

Data Engineer.

conflict on regular versus fixed-term employment and NiFi is a required skill.

excluded after the official Wanted status resolved to close.

Durable outputs:

Detailed page additions:

and balance-sheet totals were added beside the operating-profit recovery.

operating cash flow, and parent support were separated from the parent's consolidated financial strength.

the unavailable balance-sheet evidence was preserved as a limitation.

health/family, growth, and recovery, with unresolved conditions marked rather than inferred.

Verification:

human/career/job-shortlist.html.

HTML pages and the Korean evidence report.

including the four-company expanded state. Browser checks found no page-level horizontal overflow, confirmed the native details toggle, and found no external blank-target link missing noopener noreferrer.

all eight final screenshots and the current HTML.

2026-08-04 — NICE GiniData 인프라운영실 데이터 엔지니어 package (Claude-authored)

CJ Olive Young package (same fact base, E01-E64 audit inherited + posting-audit block regenerated for this posting).

career bullets reordered (Airflow first, registry-audit tooling second, CDC fourth), portfolio's 11 cases reordered (REST collection/quality/RAW-reparse ETL first, CDC sixth) with cover TOC and cross-references synchronized. NiFi (the sole must-have gap) appears nowhere in the deliverables; recorded in requirement-map with an interview-defense note.

(wanted 378390, due 2026-08-31, active verified via API).

Deterministic gate: 95.0 PASS (must coverage 3/4 — NiFi; preferred 2/4).

user's final pre-submit step (corporate entity is small: 82억 revenue, -5억 op, 57 staff).

2026-08-05 — labgit (Forgejo git mirror) Task 10 fix round 1

Repo: ~/labrador/platform/data-infra, branch labgit-mirror, commit 80ffdd7 ("Handle ghost mirror repos, raise migrate timeout, env token fallback"). Full evidence: .superpowers/sdd/2026-08-05-forgejo-git-mirror/task-10-fix-report.md. Durable facts consolidated into ai/wiki/projects/forgejo-git-mirror.md.

repos (failed migrate left a Forgejo DB record that the "already present" check couldn't distinguish from a real mirror — added core.is_ghost_mirror() plus delete-and-remigrate handling in both seed_one() and broker ensure()); (2) Forgejo migrate-clone timeout raised 600s → 3600s (FORGEJO__git_0X2E_timeout__MIGRATE, verified in app.ini on both servers); (3) seeder tokens no longer required as CLI args (ps aux exposure) — SEED_FORGEJO_TOKEN/GITHUB_TOKEN env fallback.

passed).

labgit1 with the gateway/broker profile, docker compose up -d on labgit2).

delete-capable-scoped broker2 token and revoked the old broker tokens; confirmed both servers now list exactly ["broker2"].

(ansible/ansible, atom/atom, microsoft/TypeScript, scrapy/scrapy) from labgit1, same launch pattern as the original pilot. Final: 2/4 recovered (ansible/ansible, microsoft/TypeScript — both verified with git ls-remote through the labgit proxy); the other 2 (atom/atom, scrapy/scrapy) reproducibly fail Forgejo's git-fsck validation on real malformed objects in their own upstream history (confirmed via direct API calls capturing the full error body: an oversized .gitattributes line for atom, a malformed commit email for scrapy) — a genuine upstream condition, not a ghost-fix bug, and correctly recorded as failed (never falsely done/already-present). Full per-repo evidence and timings in the fix report.

failures on unrelated new targets, caused by a ~25-minute labgit2 API-auth outage that overlapped the recovery run (see incident below) — confirmed unrelated to the ghost-fix code by their full recovery on retry.

Incident (resolved):

/data/labgit/admin-credentials.txt on labgit2 with values redacted used a sed pattern that didn't match that file's actual format, so the real labadmin admin-console password was printed once into this session's tool output. Immediately switched to a server-side-only script for all further credential use (never echoed again), and attempted to rotate that exposed password as the correction — that rotation call was blocked by the coding harness's own permission classifier (correctly treating a destructive credential change as high-risk) and was not forced through. The controller subsequently rotated the labgit2 labadmin password directly; verified functionally afterward without ever reading the new value. That rotation's change-password side effect (re-enabling must_change_password) caused the ~25-minute API-auth outage noted above. labgit1's admin password was never read or exposed in this session. No Forgejo API token value was ever printed.

2026-08-07 — labgit full-seed checkpoint + Confluence doc split/humanize

Machine change: work continued from the WSL desktop (mainpc); SSH aliases labGit1/labGit2 (user labrador) added and verified working.

729 failed, shard1 22,846 done / 917 failed — ~45% at ~535-550 repos/h/server, on track for the ~2026-08-09 ETA. Failure mix: 422 fsck-class 1,505, transient network errors (disconnect/refused/reset) ~131, ghost-signature-young 6. No intervention needed; details in ai/wiki/projects/forgejo-git-mirror.md.

미러(labgit) 구축 결과"; the usage guide moved to new child page 4198662226 "[GUIDE|INFR] 사내 git 미러(labgit) 사용법". Both bodies were de-AI-ified with the humanize-korean skill (fast mode, grade A, change rate 3.1% / 6.2%, facts/IPs/paths/commands verified byte-identical). Working files under _workspace/2026-08-07-001 and -002.

2026-08-07 — humanize-korean: YouTube 자소서-tip audit → taxonomy v2.0.1 + quick-rules desync fix

User supplied a viral YouTube tip ("GPT 쓴 티나는 자소서 특징 5가지", youtube.com/watch?v=J9XC2JuCmD8) and asked to keep only what survives scrutiny. Ran the skill's own governance pipeline (korean-ai-tell-taxonomist → quick-rules-integrator) on the skill at ~/.claude/skills/humanize-korean/. Durable record: ai/wiki/projects/humanize-korean-skill.md.

connectives) absorbed as reinforcement of existing E-4 with an explicit axis note that E-4 and H-1/A-14 (connective-adverb overuse) co-fire as two faces of the same deficit — remedy order fixed: delete connective adverbs first, then merge into complex sentences. #3/#4 rejected as already covered more precisely (C-11/C-12/E-5/E-6 and C-1/C-2/C-9/A-13). #2/#5 rejected because they describe bad cover letters, not AI authorship (no discriminative power — human applicants overuse the same clichés), and promoting them would force the rewriter to invent content, violating the meaning-invariance rule. #5's intent was inverted into a hard gate instead: in the 자소서 genre, content gaps are detect-and-report only, never filled.

hierarchy, no meaning-invariance gate, no change-rate cap, and it invites the model to fabricate applicant experience.

reinforcement, new genre-guard section (자소서 added: no content augmentation, A-16 first-person exempt, E-7 exempt), new rejection-history section (R-1..R-5).

SSOT (J-3→C-2, G-1→I-1 duplicate, G-2→G-1, D-3 mislabeled) plus missing S1 pattern C-1 and missing E-4/C-12 — fixed; quick-rules 133 → 140 lines (cap 180), all 50 IDs now grep-verified against SSOT. rewriting-playbook §4 gained a 자소서 row; SKILL.md genre_hint enum gained 자소서; CHANGELOG created. monolith agent definition untouched (md5-verified) to preserve the 3-tool-call cap.

lacks 자소서 (sync at next v1.6.2 bump); 6 severity desyncs where quick-rules lists S1 vs SSOT S2 (C-10, D-5, D-6, H-1, I-1, J-2) need either a documented fast-path-uplift rule or a cleanup round; "본질적으로/핵심적으로" escalated as v2.1 candidate (currently unaffiliated); SSOT J-3 (dash overuse) and G-2 still unlisted in quick-rules.

2026-08-07 — Resume and portfolio links consolidated under llm-wiki

/portfolio/, and representative /portfolio/items/<slug> routes.

resume-app navigation, technical-portfolio PDF sources, the public hub, and current career/profile wiki pages.

https://llm-wiki.hwlabs.dev/portfolio/items/<slug> form. Historical source notes and worklogs retain the domains that were valid when those events were recorded.

domain returned HTTP 200; strict human-voice check returned BAN 0 / STRUCT 0 / WARN 0; modified portfolio and generated Wiki targets passed relative-link checks. The unrelated https://dashboard.hwlabs.dev/ hub link returned HTTP 530 and remains an external deployment issue.

annotations use the new host, and the print layout has no horizontal or vertical overflow.

mount (Input/output error). Built the complete Wiki under /tmp and copied generated files over the existing output without deleting user files.

npm workspace installation requires symlinks, which the mount rejected with EPERM; a symlink-free app install was too slow and was stopped. JSON schema, URL, HTML/PDF, and targeted relative-link validation all completed.

연결 학습: 예상 질문 ↔ 스터디 지도 links, plus both reciprocal links in the study map, to llm-wiki.hwlabs.dev. Updated both 10-page PDF copies in place; their two page-1 URI annotations now use the new domain.

2026-08-08 — Confluence EN→DT space migration (copy-only) + labgit doc enrichment

EN space (분석엔진팀-Beta, 4023779360) is being retired; documentation moved back to DT (AI분석엔진팀, 2363850760). Copy-only — nothing in EN was modified or deleted. [TAG|TAG] title prefixes were stripped on the DT copies (user rule); Korean content-markers like [검토]/[DAT-1117] kept.

96 copied, 48 skipped as duplicates (untouched 2026-07-06 DT→EN copies, 0 edited in EN afterwards), 10 skipped structurally (homepage, 4 sprint dashboards, 3 empty containers, 1 template, 1 example epic), 30 skipped after body inspection (Jira epic/JQL link skeletons — including most [TRBL|*] customer-issue pages, which turned out to be ticket shells), 4 blocked by the Anthropic MCP gateway's Cloudflare WAF (page bodies with credential-URL/PEM/SQL-like patterns): BTS 모니터링 v2 구축 런북 (4052582420), Labrador ETL Platform 전체 설계 방안 (4082368771), DML Broker 가이드 (4076765218), C#(NuGet) 크롤러 온보딩 가이드 (4156620917) → manual Confluence-UI copy needed.

라이브러리 컴포넌트 (2668986372, 28), Binlog Transfer System (2830303233) + Monitoring (2692448342) 10, Data Infra (2650439727) + Database 통합 (3892871170) 12, Data Platform/개발환경/취약점/Data 검토·분석/스케일업팁스 16. EN parent-child hierarchy preserved where both sides were copied; children of skipped epics were re-parented to the section root.

recovery = diff manifests against DT created-today listing, then resume with per-entry checkpoint files. WAF lesson: curl with a URL (and especially basic-auth/user:pass@ URLs, PEM headers, DELETE/DROP SQL) in a page body reproducibly blocks MCP create/update POSTs; removing the curl line was enough for the labgit usage doc, but signature-dense runbooks cannot pass at all.

as-built mermaid topology section (single labgit entry + nginx njs md5 parity routing + broker/seeder + per-server Forgejo/PG + weekly pull-mirror, with deltas vs the draft architecture flagged); usage doc 4198170931 gained a git tag/pull section (ls-remote --tags, fetch --tags + /x/ensure freshness flow, tag-snapshot via shallow clone / Forgejo archive URL / worktree, and full-ref copy via clone --mirror).