Cas réel : codein 487673 site 579 affichait 38 vendus au lieu de 24,
avec un CA pourtant exact (215,76 €) — mvtart contient des lignes
genremvt=3 portant une quantité mais aucun montant TTC (corrections /
régularisations), que les endpoints officiels de l'API excluent.
Ajout du filtre mntmvtttc IS NOT NULL AND <> 0 dans l'agrégation :
une vente réelle a toujours un montant TTC non nul. Validé sur
PostgreSQL local en reproduisant le cas : sans filtre 38/215,76 €
(= symptôme), avec filtre 24/215,76 € (= réalité). Le CA et les
requêtes Analytics ne changent pas (ces lignes valent 0 €).
La version précédente agrégeait les ventes avec les jointures fournisseur
(LATERAL par ligne de mouvement + fouident) dans la même passe : si
fouident contient des codes dupliqués, chaque mouvement était compté
plusieurs fois → chiffres faux sur les deux magasins, et le LATERAL
s'exécutait par mouvement (très lent).
Restructuration, validée sur 12 000 mouvements réels de l'API chargés
dans un PostgreSQL 16 local :
- CTE ventes : mvtart JOIN articles uniquement (1:1), SUM(-qtemvt) /
SUM(-mntmvtttc) / SUM(margemvt) par (codein, site) — le calcul
canonique de la Grille (pgGetMensuelByFournisseur), signes vérifiés
sur les mouvements réels (ventes négatives, genremvt=3).
- CTE attrs : libellé/fournisseur/nomenclature joints APRÈS agrégation,
dédupliqués par DISTINCT ON — aucune jointure ne peut plus fausser
les sommes (totaux invariants aux doublons artfou1/fouident injectés).
- Même principe pour pgGetCaByFournisseur, pgGetStockNegatif,
pgGetSansVente6Mois : sous-requête artfou1 DISTINCT ON au lieu du
LATERAL par ligne.
- Grille (fallback API mensuel) : Math.abs remplacé par la négation sur
qte_vendue/ca_ht (négatifs côté API) — les retours clients restent
déduits au lieu d'être inversés en positif, même convention que le SQL.
- Hit Parade (client) : ColHeader extrait hors du composant (erreur
'Cannot create components during render'), useMemo déplacés avant
exportToExcel pour préserver la mémoïsation React Compiler, variable
idx inutilisée supprimée. ESLint 0 erreur, build OK.
La requête Hit Parade divergeait du calcul canonique de la Grille :
- SUM(ABS(qtemvt)) additionnait les retours clients comme des ventes au
lieu de les déduire → quantités et CA gonflés sur le magasin ayant des
retours/avoirs sur la période (visible sur 579).
- ABS(SUM(mntmvtttc)) inversait le signe d'un CA net négatif.
- GROUP BY no_id + jointure artfou1 preference=1 non dédupliquée pouvait
produire plusieurs lignes par (codein, site), écrasées par le pivot.
Corrections :
- Hit Parade : ventes nettes SUM(-qtemvt) / SUM(-mntmvtttc), agrégation
par (codein, site), fournisseur préféré via LATERAL ... LIMIT 1,
stock agrégé par codein ; pivot en accumulation (+=) par sécurité.
- Analytics (même classe de bug) : pgGetCaByFournisseur et
pgGetCaByNomenclature passent en CA net ; jointure fournisseur
dédupliquée (LATERAL LIMIT 1) pour éviter le double comptage.
- Gestion de stock : pgGetStockNegatif et pgGetSansVente6Mois dédupliquent
le fournisseur préféré (LATERAL LIMIT 1) pour éviter les lignes en double.
Audit complet du mapping par magasin (292/579) côté client : aucun champ
inversé (hit-parade, analytics, dashboard, grid, heatmap). Build OK.
Large supplier (4816 codes) triggered Qlik Engine "Request aborted"
(code 15): a single SelectValues over all codes forced the hypercube to
calculate too many rows at once. Now select in 1200-code batches and
accumulate, keeping each calc small. 1417 codes worked before, 4816 did
not - chunking scales to any supplier.
Also remove the "Couverture de stock" network column end-to-end (grid,
types, Qlik clients, cache, schema, db-init), shifting the marge measure
index from 6 to 5 and the hypercube width from 7 to 6.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The button was a two-line flex-col (button + date below), making it
taller than the single-line Export button next to it, so its button sat
visibly higher in the items-center row. Lay it out inline instead: the
"MAJ Qlik" date sits to the left (hidden under lg, kept in the tooltip),
the button keeps the 36px btn-action height and now aligns with siblings.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Redesign the grid page shell while preserving the Slate/teal Apple theme:
- header reworked into a branded bar: teal accent marker + uppercase
eyebrow + supplier name as the hero, with "Magasin" and "Références"
metric pills
- loading/error state shown as a teal (or error-red) accent chip
- "Valider" save button restyled with the brand gradient + soft shadow
- grid container radius/shadow aligned to the apple-card system
No change to table cell logic or the network columns.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Stock Rupture % N saturated at 100% for high-presence products (degenerate
at the article x network grain), so remove it end to end: config id,
NetworkMetric, hypercube (width 8->7), cache, schema, enrich and grid
column. db-init drops the column on existing DBs.
Also remove the "Sans vente 6 mois" tab from the grid page: the grid now
renders directly without the tab switcher; delete the NoSalesTab component.
(The separate Stock negatif page keeps its own sans-vente tabs.)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The qData expr path returned empty. Fetch the real master-measure
definition through GetMeasure -> GetProperties (qMeasure.qDef.qDef) and
log rupture, couverture and CA/mag expressions to understand why
Stock Rupture % saturates at 100% for high-presence products.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
3 of 4 new measures now resolve to engine qIds and return varied values
(CA/mag, couverture, marge%). Stock Rupture % N resolves but returns a
constant 1.0 (100%) per article — a measure-semantics issue. Log the
resolved rupture & couverture expressions to understand the definition,
and make the enrich sample pick a non-zero match (the previous sample hit
a no-sales code and looked all-zero).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The new measures returned 0 because their qLibraryId was set to the QRS
object GUID, but the engine references master measures by qInfo.qId
(short codes like JhqJ) — an unknown GUID is silently ignored and yields
null. Resolve CA par Magasin / Couverture / Marge % / Rupture % by title
via the in-page MeasureList (whitespace+case normalized), falling back to
the configured id, and log the resolved ids.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds a sample log of the first extracted rows (all 8 hypercube columns)
and a sample of the network metrics matched from cache in getProductRows,
to pinpoint whether the new measures return null from Qlik, fail to
persist, or fail to render.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Adds CA par Magasin, Couverture stock Qte, Marge % and Stock Rupture %
network measures to the Qlik hypercube extraction (in-page Playwright ws
+ legacy server ws), persists them in qlik_network_metrics, joins them in
getProductRows and renders 4 new grid columns with buy-signal coloring
(low coverage / high margin = green; high rupture = rose).
- qlik-client: config IDs (env-overridable, master GUIDs of app 9872ee6e),
NetworkMetric fields, hypercube width 4->8, page height 2500->1250
- qlik-playwright: same 4 measures in the in-page hypercube + row mapping
- schema + db-init: 4 numeric columns (ALTER ADD COLUMN IF NOT EXISTS for
existing DBs)
- cache + grid types + enrich: plumb the new fields end to end
Note: percent measures stored as raw Qlik ratio; UI renders x100 when |v|<=1.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Restore the _refresh searchParam bump so grid-client refetches
/api/grid/rows with refresh=1 (forceRefresh, bypassing the 10-min
getProductRows cache) and re-renders the network columns from the
store. Also call router.refresh() to update the server-rendered
MAJ Qlik timestamp on the button.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Validated end-to-end. The recent Qlik client authorizes the engine websocket
via ?reloadUri=...&qlik-csrf-token=<token> (not Xrfkey). Capture that token
from the client's own ws, then open ours in-page with it. Use the real engine
master-item qIds (dim Article Code=yesCP, CA N=JhqJ, Quantité N=41516861...,
Magasin Ventes Nb N=yNBLjc) — the QRS object IDs returned error 7001. Select
the supplier's codes on field "Article Code" (the app holds ~1.2M codes).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The recent Qlik client uses qmfe/single-spa (no js/qlik Capability API). Go
back to a raw engine websocket but opened in-page from the authenticated
browser (injected X-Qlik-Session cookie), which the proxy accepts. Use the
session xrfkey and capture ws close code/reason for diagnostics.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Raw and in-page websockets with a self-generated xrfkey are rejected (403).
Switch to the Qlik Capability API (require(['js/qlik']) → openApp → createCube),
which reuses the authenticated page session and manages the engine ws + xrfkey
internally. Auth via injected X-Qlik-Session cookie (Node ticket flow) +
httpCredentials. Logs rich page diagnostics on failure.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
The Qlik proxy refuses a raw server-side websocket (403, even internally), so
extraction now runs through headless Chromium (NTLM via httpCredentials, like
the hermes agent) and opens the Engine websocket in-page — same origin, which
the proxy accepts. Efficient: selects the supplier's article codes on the
"Article Code" field so the hypercube returns only those rows.
- qlik-playwright.ts: browser singleton, in-page hypercube extraction (paginated)
- /api/qlik/sync: use the Playwright extractor
- Dockerfile: install chromium + headless deps, PLAYWRIGHT_CHROMIUM_PATH, copy playwright-core
- deps: playwright-core (lockfiles synced)
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Qlik proxy validates the WebSocket Origin header (whitelist) — a browser
always sends it. Add Origin + User-Agent to the engine ws. Also surface the
403 response (server/www-authenticate/body) for diagnosis.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Add a "Qlik Sense — Données réseau" section in Paramètres (host/user/password
+ Save & Test buttons), stored in data/.db-config.json. qlik-client reads this
file first (env fallback), so no env vars needed for the sync.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Docker build uses `npm ci` which requires package-lock.json in sync with
package.json. Add httpntlm, ws, @types/ws to the npm lockfile (were only in
pnpm-lock.yaml), fixing the failed deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Docker build uses `npm ci` which requires package-lock.json in sync with
package.json. Add httpntlm, ws, @types/ws to the npm lockfile (were only in
pnpm-lock.yaml), fixing the failed deploy.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Le build Docker échouait sur le fetch de la police Inter via next/font/google
(accès réseau/TLS indisponible pendant 'next build').
- layout.tsx : charge Inter via <link> côté navigateur (plus de fetch au build),
avec fallback système ; --font-sans défini dans globals.css (@theme).
- /commandes-auto : export dynamic = 'force-dynamic' (données live DB + API,
jamais prérendues au build).
- listCadences : résilient si la DB est injoignable (retourne []).
Ajoute un onglet "Cadencier / Alertes" dans la page Commandes auto.
Depuis la liste des fournisseurs, on active un fournisseur en alerte de
commande sur un magasin (292/579) avec une fréquence en X semaines,
gérée indépendamment par magasin.
L'échéance est calculée à partir de la dernière réception réelle
(mvtart, genremvt 1/2, fournisseur principal artfou1.preference=1) +
X semaines. Statut: à commander / bientôt / OK, avec temps restant.
- Table commande_cadences (schema + db-init)
- pgGetDerniereReceptionParFournisseur (mvtart par fournisseur/site)
- Server actions CRUD (list/upsert/toggle/remove)
- UI à onglets + cadencier (ajout, édition intervalle, activer/supprimer)
LogiFlow
Claude
R0m1k3
MichaelandClaude Opus 4.8
Michael