kyuno/batchCooking
1
0
Fork 0
Code Issues 5 Pull requests 1 Projects Releases Packages Wiki Activity Actions 3
192 commits 53 branches 4 tags 4.9 MiB
Commit graph

4 commits

Author SHA1 Message Date
Nicolas
fcaecb4e07 feat(admin): tri des corrections + declenchement du gate F1/backfill
Some checks failed
CI / lint (push) Successful in 1m58s
Details
CI / build (push) Successful in 2m35s
Details
CI / e2e (push) Failing after 5m24s
Details
CI / intent-service-test (push) Successful in 22m51s
Details
CI / test (push) Failing after 23m21s
Details 
PR 5 (derniere) du chantier admin. Remplace le duo CLI
list-pending-training-suggestions.ts / retrain-tech-steps.ts par une UI.

API (admin-tech-steps.service.ts, routes /admin/tech-steps/*, requireAdmin) :
- GET /suggestions : TechStepTrainingSuggestion filtrees, groupees par
  technique, enrichies du contexte de la correction source.
- GET /corrections : corrections brutes filtrables, incluant les
  suppressions correctedTechStepId:null invisibles ailleurs.
- PATCH /suggestions/:id : edite synonymes/phrases et/ou status.
- GET /training-data-snippet : bloc training_data.py a coller (lecture
  seule).
- POST /retrain : runTechStepEvalSuite() (gate F1 vs MIN_OVERALL_F1) puis
  si passe backfillTechSteps() + marquage applied/rejected. Verrou memoire
  -> 409 RETRAIN_ALREADY_RUNNING. Gate echoue -> 200 gatePassed:false.
  N'edite pas le .py ni ne redemarre l'intent-service (manuel).

Shared : nouveau ErrorCode RETRAIN_ALREADY_RUNNING (4023, + cle i18n
apps/web), schemas (list*/update*/retrain*/snippet), types
(TrainingSuggestion*/Correction*/RetrainResultView...).

Front : CorrectionsPage (onglets Suggestions / Corrections brutes,
bandeau caveat permanent, cartes editables + Appliquer/Rejeter, panneau
snippet, panneau gate F1). Logique pure corrections.ts. i18n
admin.corrections.*. AdminApiClient : 5 methodes.

Tests : Mocha admin-tech-steps.test.ts (401 partout, groupement+filtre,
PATCH 400/404/ok, corrections incluant removals, snippet, retrain shape +
409 concurrent) ; Cypress corrections.cy.ts (4 verts). Admin-web Cypress
13/13. specs/backend-architecture.md : section tri + retrain.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
 2026-08-28 23:36:31 +02:00
Nicolas
29432e5bf2 feat(admin): monitoring des microservices + heartbeat du worker LLM 
PR 4 du chantier admin. Board de sante temps reel des dependances.

- POST /internal/tech-steps/heartbeat (requireInternalWorker) ->
  recordWorkerHeartbeat : upsert WorkerHeartbeat (cle fixe
  "tech-step-llm-worker"), lastRunAt/lastResult pour un ping "job".
  Schema workerHeartbeatSchema dans packages/shared.
- services/tech-step-llm-worker : api-client.postHeartbeat (best-effort,
  ne throw jamais) appele au boot (index.ts), a chaque tick et apres
  chaque job (scheduler.ts, avec job/ok/counts).
- admin-monitoring.service.ts + GET /admin/monitoring (requireAdmin) :
  sonde active bornee (~2 s) de Postgres (SELECT 1), l'API (uptime/RSS),
  tech-step-intent-service (/health), et le worker via son heartbeat.
  Statut up/degraded/down/unknown ; une sonde down ne casse ni les autres
  ni l'endpoint. Seuils worker : > 8 j degraded, > 21 j down.
- MonitoringView / ServiceHealthView dans packages/shared.
- Front : MonitoringPage (grille de cartes coloree par statut, re-poll
  15 s), logique pure monitoring.ts, i18n admin.monitoring.*,
  AdminApiClient.getMonitoring.
- Tests : Mocha admin-monitoring.test.ts (heartbeat 401/400/upsert
  job+boot ; GET /admin/monitoring 401, board 4 cibles, worker unknown
  sans heartbeat puis up apres) ; Cypress monitoring.cy.ts (2 verts).
  Worker mocha : 6/6 toujours verts.
- specs/backend-architecture.md : section monitoring. .gitignore :
  apps/admin-web/cypress/{screenshots,videos,downloads}.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
 2026-08-28 23:36:31 +02:00
Nicolas
3a416ea955 feat(admin): metriques d'utilisation (derive DB + AnalyticsEvent) 
PR 3 du chantier admin. Tableau de bord metriques : snapshot de compteurs
+ series temporelles journalieres.

Schema (migration admin_metrics) :
- AnalyticsEvent (type String libre, actorType/actorId sans FK, context Json,
  index [type, created_at]) + WorkerHeartbeat (cable en PR 4).
- colonnes createdAt @default(now()) sur UserProfile / Recipe / Planning /
  PlanningItem (lecture admin uniquement ; lignes existantes = timestamp de
  la migration).

Instrumentation (lib/analytics.service.ts, fire-and-forget) :
- analytics.recordEvent(type, {actorId?, context?}) : retourne void, insert
  detache, echec loggue+avale, jamais de latence sur la requete.
- points d'appel : user.signup, recipe.created, recipe.imported,
  planning.item_added, tech_step.correction_submitted, shopping_list.viewed.

API : GET /admin/metrics?days= (7-365, defaut 30, requireAdmin) ->
admin-metrics.service.ts. bucketByDay pur (zero-remplissage, teste sans
base). MetricsView dans packages/shared.

Front : DashboardPage (tuiles KPI + un graphe recharts par serie + listes
recettes-par-source / evenements), logique pure dans dashboard.ts, i18n
admin.dashboard.*. AdminApiClient.getMetrics.

reset-db.ts truncate analytics_events + worker_heartbeats.
Tests : Mocha admin-metrics.test.ts (bucketByDay pur x2 verts ; snapshot,
series zero-remplies, event user.signup fire-and-forget) ; Cypress
dashboard.cy.ts (2 verts). specs/backend-architecture.md : section admin.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
 2026-08-28 23:36:31 +02:00
Nicolas
c7301ea844 feat(admin): scaffold de l'application d'administration apps/admin-web 
Nouvelle app Vite/React independante (workspace apps/*), calquee sur
apps/web : port 5174, sa propre image Docker (nginx statique), son propre
domaine/deploiement.

- AdminApiClient (VITE_ADMIN_API_URL, credentials: include) + ApiError.
- AdminAuthContext / RequireAdmin : restaure la session admin via
  GET /admin/auth/me, garde de route (miroir de AuthContext/RequireAuth).
- LoginPage (/login) : validation cliente via adminLoginSchema partage,
  erreurs traduites via ErrorMessageService.
- AdminLayout : sidebar (Tableau de bord / Monitoring / Corrections) +
  deconnexion, rendu une fois autour du groupe RequireAdmin.
- Pages Dashboard / Monitoring / Corrections en placeholder (remplies aux
  PR 3-5).
- i18n fr (bloc admin.* + sous-ensemble errors.*), tokens _theme.scss
  copies de apps/web (extraction en package partage : suivi separe).
- Dockerfile multi-stage (node build -> nginx:alpine) + nginx.conf (SPA
  fallback). Service admin-web dans docker-compose.yml (port ADMIN_WEB_PORT,
  VITE_ADMIN_API_URL en build arg).
- Cypress : login.feature (KO -> message, OK -> dashboard) + admin-layout
  .cy.ts (redirection /login sans session, nav entre sections, logout).
  Job "Run admin-web E2E tests" ajoute a ci.yml.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
 2026-08-28 23:36:30 +02:00