admin/wiki
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity
Files
7315b66ffabd503b0d0d3aaa28c8334b90ffdc13
wiki/memory/2026-06-02.md
Stream 124f6ae20b auto-sync: 2026-06-02 20:30:01
2026-06-02 20:30:01 +03:00

14 KiB
Raw Blame History

2026-06-02 — Daily Log

ET-009: GPS Sources (EnduroRussia + Wikiloc) — Pipeline Completed 🎉

Pipeline stages (all passed):

  1. ✅ analysis — analyst (Sonnet), produced BRD, TRZ, acceptance criteria, test plans
  2. ✅ architecture — architect v2 (Opus 4.7), ADRs + infrastructure requirements
  3. ✅ development — developer (Opus 4.7 1M), 19 files, 1287 lines. Commit 3577ff3
  4. ✅ review — reviewer (Opus), found P1 bugs F-01, F-02
  5. ✅ development fix — developer (Sonnet), commit fc03746: 159 lines in gps_tracks.js
  6. ✅ review round 2 — reviewer (Opus), APPROVED, only P3 minors remain
  7. ✅ testing — tester (Sonnet), 25/25 pytest + 24/24 js tests PASS
  8. ✅ deploy — deployer (Sonnet), PR #16 merged, image rebuilt, collector running

Key results:

  • EnduroRussia parser: enduro_russia.py — scraping + GPX parsing
  • Wikiloc parser: wikiloc.py — Wikiloc GPX integration
  • Dynamic source filters (F-01): _buildGpsFiltersUI pulls source_id from /api/gps-tracks/health.tracks_by_source
  • Attribution fix (F-02): attribution set at addSource time via _ensureGpsSources(map, attribution), not by mutating src.attribution (MapLibre AttributionControl issue)
  • Collector running: 305 EnduroRussia tracks, fetching one per 5s, 34+ already in DB
  • Commits: eaa6b4c (analyst), 4be7fbf (architect), 3577ff3 (dev), fc03746 (dev fix)
  • PR: #16 merged → main at b5ba7b2

Known issues:

  • Nginx 502 on external URL (port 5556 routing) — not ET-009 specific, needs nginx config fix
  • gps_sources.yaml — ADR files inside Docker image (not volume), old image had proposed; fixed by rebuilding image

Orchestrator Issues Discovered

Critical bug: Launcher Popen broken inside container

  • All Claude processes (claude.exe) become zombies when launched via orchestrator's Python subprocess
  • Root cause: docker binary not available inside container, Popen flags wrong
  • Workaround established: base64-encode Python launch script → ssh to host → decode → docker cp → docker exec python3
  • This pattern avoids shell quoting hell

Other orchestrator bugs:

  • .task-arch.md, .task-dev.md, .task-review.md not regenerated per-stage (carry over from previous task)
  • Workaround: manually rewrite these files before each stage launch
  • No CI configured in Gitea → every commit triggers false "CI failed" webhook alert
  • Tokenator API: all 8 models down for maintenance (June 1, still down June 2 morning)

Agent launch manual pattern:

# 1. Base64-encode the Python launch script
cat launch_script.py | base64 -w0

# 2. SSH to host, decode, cp to container, exec
ssh slin@82.22.50.71 "
  echo 'BASE64_STRING' | base64 -d > /tmp/launch.py
  docker cp /tmp/launch.py orchestrator:/tmp/launch.py
  docker exec orchestrator python3 /tmp/launch.py
"

Models Configuration Changes

  • DeepSeek V4 models added to openclaw.json:
    • openrouter/deepseek/deepseek-v4
    • openrouter/deepseek/deepseek-v4-pro
    • openrouter/deepseek/deepseek-v4-flash
  • Current model for main agent: openrouter/deepseek/deepseek-v4-pro

User Preferences (re-confirmed during ET-009)

  • User wants visibility at each pipeline stage — frustrated by silent automated advances
  • User wants to approve before advancing: analyst → :approved: → architect → developer → reviewer → tester → deployer
  • Only analyst requires manual :approved:; architect, reviewer, tester should auto-advance but notify
  • Deployer should ask before deploying

Глубокий аудит оркестратора (мультиагенты) — 2026-06-02

Проведено

  • Аудит кода: tasks/multi-agent/AUDIT_2026-06-02.md — 3 BLOCKER, 5 SERIOUS, 7 MEDIUM, 5 LOW
  • Аудит документации: tasks/multi-agent/DOCS_AUDIT_2026-06-02.md
  • ТЗ для Dev: tasks/multi-agent/DEV_TASK_ORCHESTRATOR_FIXES.md (9 задач)

Корневые баги (BLOCKER)

  • B-1: нет docker бинарника в контейнере orchestrator → _write_task_file через docker run падает молча → агент читает старый .task-*.md
  • B-2: Popen + PIPE + daemon-поток → claude-зомби, exit_code теряется (в БД exit=None для всех ET-009)
  • B-3: .task-*.md тащатся между задачами + коммитятся в git

Ключевая находка по докам

  • P1 zombie из LESSONS_ET006.md (22 мая) = мой B-2. Баг известен 11 дней, не починен → повторился на ET-008/009
  • Системная проблема: lessons пишутся, но не конвертируются в задачи
  • Расхождение код↔док: README говорит review QG=check_review_approved, код использует check_reviewer_verdict

Dev-агент запущен

  • Model: tokenator/claude-opus-4-8 (modelApplied: true)
  • taskName: orch_fixes, runId: 2f18ca88
  • Deploy-хук /home/slin/bin/enduro-deploy-hook.sh СУЩЕСТВУЕТ и корректен
  • В orchestrator были незакоммиченные правки (M launcher/config/notifications/plane_sync) — Dev разберёт в Task 1

ET-009 Artifacts

  • Plane issue: 64e98247-509f-4c30-9955-a5531ab7d1ee (sequence #5)
  • Plane workspace: ag_proj, project 7a79f0a9-5278-49cd-9007-9a338f238f9c
  • Orchestrator DB: task_id=16
  • Repo: /home/slin/repos/enduro-trails on mva154
  • Docs: docs/work-items/ET-009/ — 00 through 14 (BRD, TRZ, acceptance, tests, ADRs, infra, data, risks, review, test report, deploy)
  • Docker: container enduro-trails, image rebuilt with new ADR files, collector running

2026-06-02 — Daily Log

ET-009: GPS Sources (EnduroRussia + Wikiloc) — Pipeline Completed 🎉

Pipeline stages (all passed):

  1. ✅ analysis — analyst (Sonnet), produced BRD, TRZ, acceptance criteria, test plans
  2. ✅ architecture — architect v2 (Opus 4.7), ADRs + infrastructure requirements
  3. ✅ development — developer (Opus 4.7 1M), 19 files, 1287 lines. Commit 3577ff3
  4. ✅ review — reviewer (Opus), found P1 bugs F-01, F-02
  5. ✅ development fix — developer (Sonnet), commit fc03746: 159 lines in gps_tracks.js
  6. ✅ review round 2 — reviewer (Opus), APPROVED, only P3 minors remain
  7. ✅ testing — tester (Sonnet), 25/25 pytest + 24/24 js tests PASS
  8. ✅ deploy — deployer (Sonnet), PR #16 merged, image rebuilt, collector running

Key results:

  • EnduroRussia parser: enduro_russia.py — scraping + GPX parsing
  • Wikiloc parser: wikiloc.py — Wikiloc GPX integration
  • Dynamic source filters (F-01): _buildGpsFiltersUI pulls source_id from /api/gps-tracks/health.tracks_by_source
  • Attribution fix (F-02): attribution set at addSource time via _ensureGpsSources(map, attribution), not by mutating src.attribution (MapLibre AttributionControl issue)
  • Collector running: 305 EnduroRussia tracks, fetching one per 5s, 34+ already in DB
  • Commits: eaa6b4c (analyst), 4be7fbf (architect), 3577ff3 (dev), fc03746 (dev fix)
  • PR: #16 merged → main at b5ba7b2

Known issues:

  • Nginx 502 on external URL (port 5556 routing) — not ET-009 specific, needs nginx config fix
  • gps_sources.yaml — ADR files inside Docker image (not volume), old image had proposed; fixed by rebuilding image

Orchestrator Issues Discovered

Critical bug: Launcher Popen broken inside container

  • All Claude processes (claude.exe) become zombies when launched via orchestrator's Python subprocess
  • Root cause: docker binary not available inside container, Popen flags wrong
  • Workaround established: base64-encode Python launch script → ssh to host → decode → docker cp → docker exec python3
  • This pattern avoids shell quoting hell

Other orchestrator bugs:

  • .task-arch.md, .task-dev.md, .task-review.md not regenerated per-stage (carry over from previous task)
  • Workaround: manually rewrite these files before each stage launch
  • No CI configured in Gitea → every commit triggers false "CI failed" webhook alert
  • Tokenator API: all 8 models down for maintenance (June 1, still down June 2 morning)

Agent launch manual pattern:

# 1. Base64-encode the Python launch script
cat launch_script.py | base64 -w0

# 2. SSH to host, decode, cp to container, exec
ssh slin@82.22.50.71 "
  echo 'BASE64_STRING' | base64 -d > /tmp/launch.py
  docker cp /tmp/launch.py orchestrator:/tmp/launch.py
  docker exec orchestrator python3 /tmp/launch.py
"

Models Configuration Changes

  • DeepSeek V4 models added to openclaw.json:
    • openrouter/deepseek/deepseek-v4
    • openrouter/deepseek/deepseek-v4-pro
    • openrouter/deepseek/deepseek-v4-flash
  • Current model for main agent: openrouter/deepseek/deepseek-v4-pro

User Preferences (re-confirmed during ET-009)

  • User wants visibility at each pipeline stage — frustrated by silent automated advances
  • User wants to approve before advancing: analyst → :approved: → architect → developer → reviewer → tester → deployer
  • Only analyst requires manual :approved:; architect, reviewer, tester should auto-advance but notify
  • Deployer should ask before deploying

Глубокий аудит оркестратора (мультиагенты) — 2026-06-02

Проведено

  • Аудит кода: tasks/multi-agent/AUDIT_2026-06-02.md — 3 BLOCKER, 5 SERIOUS, 7 MEDIUM, 5 LOW
  • Аудит документации: tasks/multi-agent/DOCS_AUDIT_2026-06-02.md
  • ТЗ для Dev: tasks/multi-agent/DEV_TASK_ORCHESTRATOR_FIXES.md (9 задач)

Корневые баги (BLOCKER)

  • B-1: нет docker бинарника в контейнере orchestrator → _write_task_file через docker run падает молча → агент читает старый .task-*.md
  • B-2: Popen + PIPE + daemon-поток → claude-зомби, exit_code теряется (в БД exit=None для всех ET-009)
  • B-3: .task-*.md тащатся между задачами + коммитятся в git

Ключевая находка по докам

  • P1 zombie из LESSONS_ET006.md (22 мая) = мой B-2. Баг известен 11 дней, не починен → повторился на ET-008/009
  • Системная проблема: lessons пишутся, но не конвертируются в задачи
  • Расхождение код↔док: README говорит review QG=check_review_approved, код использует check_reviewer_verdict

Dev-агент — РЕЗУЛЬТАТ (ПРИНЯТО ✅)

  • Model: tokenator/claude-opus-4-8 (modelApplied: true), taskName orch_fixes, runId 2f18ca88
  • Починено: B-1, B-2, B-3, S-5, S-1, M-1 + 2 доп-фикса найденных в бою:
    • git safe.directory в Dockerfile (без него auto-commit падал на «dubious ownership»)
    • init:true в compose (tini как PID 1, добивает grandchild-зомби от claude/node)
  • 11 коммитов в main orchestrator (Conventional Commits), запушены
  • Тесты: 26 passed (launcher + qg) — гонял в контейнере (хостовый .venv сломан, symlinks на py3.10)
  • Автономный прогон: через launcher.launch(), 0 зомби, auto commit+push сработал
  • Доки обновлены: ARCHITECTURE, README (QG-таблица), BUGFIXES_2026-06-02.md

Моя независимая проверка (на сервере, не на слово)

  • ✅ health {"status":"ok"}, 0 defunct процессов
  • ✅ B-1: _write_task_file через open() + raise (код подтверждён)
  • ✅ S-5: verdict читается только из YAML frontmatter
  • ✅ 26 тестов зелёные
  • ✅ init:true в compose, PR #19 open+mergeable

⚠️ ТРЕБУЕТ ДЕЙСТВИЯ

  • PR #19 (enduro-trails: .gitignore + reviewer.md) — ветка main защищена, нужен merge человеком. Спросила Славу — ждёт ответа.

Честные расхождения (Dev не скрыл)

  • Хостовый .venv сломан (symlinks py3.10) — тесты в контейнере
  • test_webhooks.py — 9 pre-existing падений (401/signature), были ДО правок, вне scope
  • Вне scope (отдельные задачи на потом): S-2/S-3 (rollback в shared-репо), S-4 (git worktree per task), M-3 (дубль _try_advance_stage), F-2b (очередь задач вместо daemon-потоков — главный целевой рефактор), M-7 (идемпотентность webhook)

Открытый вопрос Славе

  • Смержить PR #19?
  • Завести в Plane оставшиеся задачи (worktree, очередь задач)?

ET-009 Artifacts

  • Plane issue: 64e98247-509f-4c30-9955-a5531ab7d1ee (sequence #5)
  • Plane workspace: ag_proj, project 7a79f0a9-5278-49cd-9007-9a338f238f9c
  • Orchestrator DB: task_id=16
  • Repo: /home/slin/repos/enduro-trails on mva154
  • Docs: docs/work-items/ET-009/ — 00 through 14 (BRD, TRZ, acceptance, tests, ADRs, infra, data, risks, review, test report, deploy)
  • Docker: container enduro-trails, image rebuilt with new ADR files, collector running
Reference in New Issue View Git Blame Copy Permalink