#1 axe trop large (0-24) → axe ADAPTATIF calé sur l'amplitude réelle des shifts réguliers de la
semaine (garde n'élargit pas) → barres plus larges, position lisible. Intervalle début–fin REMIS
en texte dans la cellule (☀ 8–16 🛡️) = référence sans survol. Infobulle = capacité offrable
(corrige aussi un bug: shiftH→bookableH).
#2 modèles d'assignation TRIÉS PAR USAGE (les plus utilisés en premier) + infobulle nom.
(Rappel: créer des modèles custom = éditeur « Types de shift ».)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Modèle: champ on_call (Check) sur Shift Template. Un quart garde:
- N'est JAMAIS offert au booking client (techGaps retourne null) — vérifié: tech Jour+Garde
n'offre que la fenêtre Jour, aucun créneau dans la plage de garde.
- Est EXCLU du dénominateur d'occupation (heures offrables), affiché à part.
- Timeline: bande HACHURÉE (vs neutre pour l'offrable) + 🛡️ dans le label + tag (garde) en infobulle.
- Éditeur de modèles: bascule '🛡️ Garde' pour créer/marquer un quart de garde.
hub: fetchTemplates expose on_call; create/update template le gèrent. Champ ajouté à
setup_dispatch_custom_fields.py (persistance). Démo: Garde 18h-minuit marquée on_call.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Avant: 'J8' ne distinguait pas 7-15 de 9-17 → mêmes créneaux apparents, dispo réelle différente.
Maintenant chaque cellule affiche: chip (lettre) + intervalle '7–15', et une mini-timeline sur un
axe de journée (06:00→21:00) où la fenêtre du shift est positionnée (donc 7-15 à gauche, 9-17 à
droite = visuellement distinctes) avec les blocs de jobs pris (couleur selon charge) → les TROUS
restants = créneaux offrables. Infobulle = intervalle + h occupées/h (%).
- hub occupancyByTechDay renvoie {h, blocks:[{s,e}]} (heures de début réelles des jobs).
- ops: cellWindow/axisPos/shiftStyle/blockStyle, rendu .tl/.tl-shift/.tl-blk + tick midi.
- démo 8 juin: modèles Matinal 7-15 + Décalé 9-17, techs alignés (7→13.8, 9→18.6 surbooké).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Chaque cellule tech×jour avec un shift affiche, sous le chip (J8), une mini-barre + % colorés
(vert <70, orange 70-99, rouge >=100 surbooké) + infobulle = intervalle du shift + h occupées/h.
Occupation = Σ duration_h des Dispatch Jobs planifiés assignés ce jour ÷ Σ heures du shift.
- hub: occupancyByTechDay(start,days) + GET /roster/occupancy → map 'TECH|date': heures.
- ops api: getOccupancy ; PlanificationPage: occCells (computed), cellOcc/occColor/cellInterval,
rendu barre + q-tooltip, chargé dans loadStats. Données démo semaine 8 juin (45/85/120%).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Bug: le copilote comprenait l'absence mais n'agissait pas → aucun impact. Causes:
1) prompt « par défaut analyse » → ne déclenchait pas l'action ;
2) marquer une absence n'excluait PAS des créneaux (techGaps ne testait que le statut
global En pause, pas les Tech Availability approuvées par jour) ;
3) loadBookingData calculait unavail mais ne le retournait pas (oubli) → garde inerte.
Fixes:
- roster.js: loadBookingData inclut unavail (buildUnavailability = En pause + absence_from/until
+ Tech Availability approuvées) ; techGaps exclut le tech absent ce jour-là ; export bookingSlots/fetchTemplates.
- roster-assistant.js: nouvel outil gerer_absence = crée l'absence (par jour) + trouve les RDV
impactés + RÉASSIGNE auto à un autre tech libre du même créneau (IROPS), renvoie les
« à reporter ». Nouvel outil ajouter_disponibilite (tech à l'acte ouvre un créneau). Prompt
orienté ACTION (signaler une absence = instruction d'exécution).
- Validé prod (lab): copilote crée l'absence ✓, booking exclut le tech absent (109→104) ✓,
rematch DJ-…→Antoine même créneau ✓ ; données de test nettoyées.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
RendezVousPage:
- Vue segmentée À planifier / À recontacter / Tous.
- Créneaux proposés groupés Semaine → Jour (se situer dans le temps, comme /book).
- Hold à la sélection: bookHold(date,start,10min) → bloque les autres; libéré à la confirmation
ou au changement de job (onBeforeUnmount).
- File À recontacter (jobs À reporter) + actions: Lien client (copie URL self-serve),
Aviser par SMS (notify-reschedule: désassigne + SMS lien /book).
CopilotePage: carte réglages des créneaux offerts (#56) — lead_hours, plage horaire,
horizon, max/jour, hold, jours offerts (chips) → savePolicy({booking}).
api/roster.js: bookHold, bookLink, jobsToReschedule, notifyReschedule.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- getBookingPolicy() (sous-objet 'booking' du fichier policy): lead_hours, day_start/end,
days_offered, horizon_days, max_per_day, hold_minutes. Appliquée dans bookingSlots →
cohérent pour /book + vue agent + fit. ignorePolicy au moment de confirmer (slot encore libre).
- Holds en mémoire (Map TTL): /roster/book/hold {date,start,minutes|release} → fenêtre retirée
des dispos des autres pendant le hold; libérée à la confirmation. Évite le double-booking
pendant qu'un agent/client choisit.
- roster-assistant: DEFAULT_POLICY.booking + booking_fields/weekdays (descripteurs UI), fusion fine.
- Testé: plage 10-14h filtre bien; hold 10:00 dispo 12→11.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Avant: liste plate jj/mm peu lisible. Maintenant: bandeau de semaine ('Semaine du 8 – 14 juin'),
en-têtes de jour complets (lundi 8 juin), sections Matin/Après-midi. Sélection 1-3 préférences inchangée.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Avant: le job gardait scheduled_date → page /book court-circuitait en 'déjà confirmé' sans options.
Maintenant un report = vrai retour au pool → le client choisit un nouveau créneau (vérifié: 50 créneaux proposés, SMS livré).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Avant: ne rechargeait que sur 401/403 stricts → ratait la redirection IdP / page HTML de
login (vrai cas d'expiration) → données vides nécessitant un refresh manuel. Maintenant:
détecte redirection auth.targo.ca + HTML-au-lieu-de-JSON → reload auto (anti-boucle 20s),
+ 1 retry sur coupure réseau transitoire (ex: backend qui redémarre).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Corrige les manques signalés: champ technicien (congés) → autosuggest typeahead; compétences
(demande de shift + éditeur équipe) → chips au lieu de texte libre. Composants réutilisables
pour une UX cohérente partout (et le copilote/réassignation à venir).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Le copilote peut désormais CHANGER les dispos en langage naturel ('marque Simon malade
aujourd'hui' → crée une Tech Availability approuvée) et réassigner un Dispatch Job, sur
instruction explicite, avec confirmation. Testé OK end-to-end. (SMS client + escalade = suite.)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Nouvelle page /copilote : chat texte/voix (Web Speech API fr-CA) vers le copilote Gemini Flash
(impact d'absence + propositions de réassignation), + sélecteur de politique de reprise
(réassign/SMS/escalade) persistée. Route + nav (icône Sparkles ; ajout CalendarRange/CalendarClock
manquantes dans la map d'icônes).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
lib/roster-assistant.js : couche conversationnelle sur le roster (distincte du solveur OR-Tools).
Outils data réels (etat_equipe, jobs_du_technicien) via roster.fetchTechnicians + Dispatch Job.
Ex: 'TECH-4776 malade le 16 juin' → résout le nom, liste les RDV impactés, propose des techs
dispos qualifiés. Routes /roster/assistant + /roster/policy (politique persistée fichier).
Réutilise le moteur geminiChat de lib/agent.js (gemini-2.5-flash). Testé OK avec données réelles.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Affiché barré (gris) si > prix de vente (Item Price), sur produits simples + variantes
(les bundles gardent barré=somme des composants). Catalogue lit le champ; page rend <s>.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Curation = champ Item.show_in_store (case à cocher) ; catégorie=item_group, prix=Item Price, stock=Bin live
- Bundles via Product Bundle (prix vs somme barrée), variantes mono-attribut (delta prix + stock/valeur)
- Fix filtres erp.list (tableaux, pas objets — listUrl teste .length) + child-tables lues via erp.get parent
- Page: fetch /store/catalog au mount, repli démo, bandeau live/démo, stock simple respecté
Prérequis PG corrigé: retrait ORDER BY NULL (MySQLism) dans erpnext utilities/product.py
(validation de variante) — sinon création de variantes impossible sur PostgreSQL.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Presets Standard 360→240 (10$/mois) / Simple 180→120 (5$/mois) / Aucune
- 2 cases éditables Prix original / Prix financé + aperçu live (barré, $/mois, crédité→0$)
- Alimente install_fee/install_regular du Service Contract → page d'acceptation affiche le détail
- Placé sous Code de référence dans la vue résidentielle
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- Prix original (regular_price) disponible sur lignes récurrentes ET ponctuelles (avant: ponctuelles seulement)
- Si Prix original > Prix marketing → barré affiché (sommaire + récap + aperçu live dans l'éditeur)
- Forfait récurrent barré remonte au contrat (monthly_regular) → page d'acceptation affiche <s>99.95</s> 79.95/mois
- Cohérent avec install (install_regular 360→240). Aucun impact CRTC (rabais go-forward, jamais récupéré)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Installation = vraie créance financée (champ install_fee : 240$ std → 10$/mois,
120$ simple → 5$/mois). Résiliation résidentielle ne facture que le solde restant
de l'install (service rendu) ; promotions/carte-cadeau jamais récupérées
(termination_fee_benefits=0). Récap reformulé (financement install, plus de
"promotion non étalée"). Conforme avant l'échéance du 12 juin.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Spec v1 du module de gestion de shifts + approbations dans le dispatch
(décision build-our-own vs Frappe HR déjà actée: PG + multitenant + géofence-job).
Modèle de données, workflows d'approbation (ShiftRequest/Swap/AttendanceRequest),
géofence vs adresse du job dispatch, auto-attendance, auto-roster Timefold,
mapping validé vs Frappe HR, intégrations (Authentik/dispatch/ERPNext API), phasage.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Surfaces ERPNext's Email Queue in Ops (nav « File courriels ») so ops can see
what's queued — important now that mute_emails=1 + scheduler paused mean nothing
flushes — and delete/purge stale entries without the ERPNext desk.
- hub lib/email-queue.js: GET list (by status, recipients read from each row's
full doc since ERPNext ignores fields on child-doctype REST), DELETE :name,
POST /purge {status}. Wired in server.js.
- ops: api/emailQueue.js + EmailQueuePage.vue (status filter, recipients,
reference, error tooltip, per-row delete + « Purger Not Sent »), route + nav.
Verified live: 13 'Not Sent' (old internal test emails, no invoice refs).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
98 business accounts had commercial=0 but a company name (Ferme X Inc.,
Ville de Farnham, Les Jardins Sorel…), leaking into the residential report.
Rule is now: commercial = (commercial flag) OR (company present). Residential
@90$ drops 739→654 (0 company-bearing rows left); commercial 244→276.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
scripts/refresh-legacy-report-tables.sh: dumps only the 4 tables the
overpriced-internet report needs (account/delivery/service/product) from the
live legacy DB with --single-transaction (non-locking), verifies the dump is
complete before importing into the local legacy-db copy. ~20MB, seconds.
Creds read from a prod-only .refresh.env (gitignored; .env.example committed).
Used to refresh the copy from the 2026-05-22 snapshot to current
(service 66741→69233, account 15321→15706).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Adds a sortable 'Ville' column (field city) to the report. Quasar's default
filter scans all columns, so the existing search box now matches city too.
Street address caption drops the now-redundant city (keeps postal code).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The origin story had a duplicated/merged first sentence, an inconsistent
name (Ghislain / Ghislain Guinois / "Ghilas" typo), and a present-tense
verb ("travaille") inside an otherwise past-tense narrative. Rewrote the
three paragraphs into clean, consistent French (passé composé), fixed the
name to "Ghislain Guinois", "18h" → "18 h".
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Tested the hypothesis that a warm Chromium cache (the register GET being
re-served stale) was causing the protected address/search 401. Disabled
the HTTP cache (CDP Network.setCacheDisabled), the on-disk cache
(--disk-cache-size=0) and service workers (serviceWorkers:'block').
Result: identical trace — register=200 (freshly minted, not cached),
autocomplete=200, address/search=401. So cache was NOT the cause; the
401 is a server-side authorization decision on the protected endpoint
(reCAPTCHA Enterprise assertion required). Keeping the cache-disable as
hygiene + to definitively rule it out in future debugging.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Cogeco's address checker is gated by reCAPTCHA Enterprise (risk-score 401
on the protected /boutique/api/address/search call), so per-address
serviceability can't be scraped reliably from a datacenter IP without a
residential proxy. Per product decision, pivot to an assisted spot-check
instead of automated qualification.
- ReportInternetCherPage: add a "Concurrent" column with a one-click
button that copies the full service address and opens Cogeco's
availability page in a new tab (human reads the verdict in ~10s, only
for the leads that matter). fullAddress() builds "addr, city, QC ZIP".
- cogeco-checker: harden the POC anyway — track service-address/search
responses, retry the verdict call on 401 (re-register cadence), and
prioritize the authoritative JSON body in interpret(). Recon confirmed
the wall is reCAPTCHA scoring, not a timing/selector bug.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Playwright/Chromium microservice (mirrors modem-bridge: node:20-slim +
Chromium, token auth, port 3302, serialized + rate-limited) that drives
Cogeco's public address checker to determine if a competitor serves a
given address.
What works (proven on prod):
- Anti-bot bypass: vanilla headless gets 403 on /boutique/api/register
(reCAPTCHA Enterprise blocks datacenter headless). Adding
playwright-extra + stealth flips it to 200 — register + autocomplete
succeed.
- Reaches Cogeco's address system and pulls real autocomplete
suggestions. Confirmed it's Loqate/AddressComplete (id + next:
Retrieve/Find shape).
What's NOT reliable yet (do not use the verdict for decisions):
- The serviceability verdict. The Loqate flow is multi-step
(Find → Retrieve → Cogeco serviceability) and a single option click
doesn't complete it, so the final yes/no API call isn't captured.
- Current interpret() falls back to scanning UI text and produces FALSE
POSITIVES (a rural out-of-Cogeco address returned available=true off
generic marketing copy). Needs the real Retrieve+serviceability
endpoint wired before it can be trusted.
Next: capture the post-selection Retrieve + serviceability call (likely
needs a "continue" step and handling the multi-dwelling "N Addresses"
branch), then parse the real verdict + speeds.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Marc Robidoux flagged as overpriced (129.95$) — he has a loyalty discount
(service id 74448) that should lower it. Investigation: 74448 doesn't
exist in the copy (its max service id is 74393), so the discount was added
after the snapshot. Same freshness issue as Julie Dupuis — not a calc bug.
But this also exposed that the freshness banner was wrong: it read the
newest INVOICE date (Apr 30) while the snapshot actually carries SERVICES
created through May 22 — May's recurring billing run simply hadn't executed
at dump time, so invoices lag services by ~3 weeks. For a report that reads
active services/plans/discounts, the service date is the right freshness
signal.
fetchDataAsOf now returns both {services, invoices}; data_as_of (shown in
the banner) is the service date (May 22), with last_invoice kept for
reference. The copy is ~10 days stale, not ~1 month. Marc's loyalty credit
still won't show until the copy is refreshed (task #38).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
User flagged Claude Bergeron at 99.95$ "including TV". Investigation: the
report already excludes TV (cat 33/34) — his cat-32 Internet subtotal was
94.95. The real issue was the opposite of what it looked like: a -60$
RAB_FTTH_URBA discount on his account lives in cat 26 ("équipement fibre"),
which the report did NOT count. His true net Internet is 44.95$, so he
should drop off the >90$ list entirely (and now does).
Internet equipment categories (26/29 fibre, 7/8 wireless) carry recurring
items that belong on the Internet bill:
- modem/router rentals: FTTH_LOCMOD +10, LOC_TPL +5, LOCRTHG8245 +6.95
- Internet discounts: RAB_FTTH_URBA (hijacked to -60 for Claude)
Added them to CAT_INTERNET_CORE. The existing price_recurr_type=1 filter
still drops one-time install charges (INSTFIBRE -199, etc.) that share
these categories. Verified HVSECTOUR/INSTTELE (odd high-price items in
cat 7/8) have zero active residential services — no aberrations introduced.
Net effect: the report's "net Internet" is now truly net of every
recurring Internet discount, wherever it's categorized. Residential >90$:
554 → 739 (modem rentals legitimately lift borderline bills; deep
equipment-category discounts like Claude's pull others below the line).
TV and téléphonie remain fully excluded.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
User correctly spotted that Julie Dupuis shows 114.95$ but actually pays
69.95$ — investigation revealed the legacy COPY (legacy-db container) is a
one-shot snapshot from 2026-05-05 with data through 2026-04-30 and NO
auto-sync. She renegotiated in May (a -50$ discount on service 50999) which
the copy never received. The report was correct vs the copy, but the copy
is ~1 month stale.
Two changes (data-source strategy still pending operator decision —
prod 10.100.80.100:3306 is reachable for a future live/refresh option):
1. data_as_of — the report now reports MAX(invoice.date_orig) from the
copy and the Ops page shows a banner ("Données legacy au 30 avril —
copie figée, N jours"). Turns orange past 7 days so nobody acts on
stale prices unknowingly.
2. recent_expired_discount column — per-address sum of deactivated credit
lines (status=0, price<0) whose actif_until fell in the last 180 days.
Surfaces clients whose discount just lapsed (Julie's RAB24M -15 + RAB_X
-35 expired 2026-03-01), i.e. the prime retention targets whose bill is
about to jump. Shown in amber with a warning icon + tooltip; included in
the CSV.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
User flagged that several listed accounts are inactive (Or Viande Inc,
Denis Henderson). Root cause: I filtered service.status=1 but NOT the
account, so terminated accounts carrying an orphan active service line
slipped through. The legacy billing job (LEGACY-ACCOUNTING-ANALYSIS.md
§6.1) bills only when BOTH service.status=1 AND account.status=1.
Three account-level filters added:
- account.status = 1 → drops terminated accounts. Or Viande Inc is
status=4, terminated 2014 (terminate_date set), but still had a
service.status=1 row. 8602 accounts are status=4 vs 6537 status=1.
- account.group_id = 5 → "Client" per account_group. Drops 6 Prospect,
7 Fournisseur, 8 Relais (network infra, e.g. Denis Henderson's
REL_CHRY_CHARLES tower account), 10 Équipement motorisé.
- customer_id NOT LIKE 'PROPRIO%' → 59 landowner-hosts-our-gear accounts
that live in group 5 but aren't paying customers (Denis Henderson's
other account PROPRIOH_STCHARLES). A genuine same-name customer
(Robert Henderson, ROBEH...) correctly stays.
Residential >90$/mo: 983 → 554 (was inflated ~44% by dead/non-customer
accounts). Commercial: 255 → 240.
Ops page note updated to state "comptes clients actifs uniquement" and
list what's excluded.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Reviewed against docs/archive/LEGACY-ACCOUNTING-ANALYSIS.md (the migration
audit) which surfaced two things to check in the overpriced-internet report:
1. service.payment_recurrence (0=annual, 2=monthly, 5=semestrial...) —
checked whether per-cycle prices needed /N normalization. They do NOT:
verified a semestrial FTTH1500I carries product.price=109.95, identical
to the monthly one (billed 6×109.95 every 6 months). Per §6.1
"prix = quantité × prix_unitaire", product.price is already the monthly
unit price. The original monthly logic was correct — no division. The
SKU-LIKE-'%ANN' /12 special-case stays (true annual plans where price
IS the yearly amount, e.g. FTTH_ANN @ 480$/yr).
2. Promo credits carry an actif_until end date (§10). A discount line whose
actif_until is past no longer reduces today's bill, so counting it
understates what the client actually pays. Now excluded.
NULL-safety: the exclusion needs an explicit `actif_until IS NOT NULL`
guard — without it, `NOT (price<0 AND actif_until>0 AND actif_until<now)`
evaluates to NULL for permanent credits (actif_until NULL), which SQL
treats as not-true and silently DROPS every permanent credit line. That
briefly inflated the residential count to 3330; with the guard it's a
correct 1000 (vs 983 before — +17 addresses whose only sub-90 reason was
a now-expired credit).
Net effect: the report reflects the *current* real monthly Internet bill.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
New Ops report to surface clients whose net monthly Internet bill
exceeds a threshold — for spotting plans that should be revised.
Hub (lib/legacy-reports.js — new module, read-only MariaDB):
- GET /reports/legacy/overpriced-internet (+ .csv variant)
- Queries the legacy gestionclient DB directly via a small mysql2 pool
(reuses cfg.LEGACY_DB_* — same vars as auth.js sync-legacy; added
LEGACY_DB_PASS to the hub .env which was previously unset).
- Grain = delivery (service address), NOT account: a multi-unit
building (account 13166 has 82 doors / 205 services) would otherwise
show a single bogus $2117 line instead of ~45 per door.
- Net monthly Internet = SUM of effective per-line price across
Internet categories (32 fibre, 4 wireless, 23 camping + optional
add-ons 16/17/21), discounts included (products with price<0 are
recurring credits like RAB24M -15$).
- Effective price = service.hijack ? hijack_price : product.price.
- Only recurring lines (product.price_recurr_type=1) — excludes
one-time equipment/install charges.
- Annual plans (SKU LIKE '%ANN', e.g. FTTH_ANN @ 480$/yr) normalized
/12 so they compare correctly against a monthly threshold (was
falsely showing $480 → now $40, drops below 90$).
- Excludes TV (33,34) and téléphonie (9) entirely.
Validated counts at 90$/mo: 983 residential, 297 commercial addresses.
Ops UI:
- src/pages/ReportInternetCherPage.vue — threshold/segment/add-ons
filters, summary cards (count, total monthly, avg, discounts),
sortable+filterable table (client, address, net, gross, discount,
plan detail with full tooltip, contact), CSV download.
- Card on the Rapports hub + route /rapports/internet-cher.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
The native-block imageBlock factory was emitting img tags wrapped only
by a td with text-align:center. That doesn't actually center the image
because text-align only affects inline content, and the img has
display:block. The result: top header logo and dark-footer logo were
left-aligned despite the textAlign:"center" prop on the block.
Fix: wrap each img in an inner <table align="<textAlign>"> exactly the
way MJML/Litmus/Mailchimp do it. This is the canonical email-client
pattern that works in Outlook 2007-2019 (which ignores margin:0 auto
on inline tables but respects table align attributes).
Also: the AI converter dumped the entire dark footer band into a
SINGLE htmlBlock with malformed table markup (a stray </td> outside
its row). Split into proper image + text native blocks so:
1. The logo inherits the new centered nested-table pattern
2. The URL+copyright text is now individually editable in Unlayer
3. The {{year}} placeholder is in a text block where it belongs
And one AI hallucination correction: the converter assigned
textAlign:"left" to the top header logo (probably because the
surrounding column had align="left" in the MJML output). Original
design intent was centered — fixed in the spec.
Verified live: both logos (140px top, 120px footer) now render with
align="center" on their nested wrapper table.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Until now, every Unlayer-edited template stored as a single giant
"Custom HTML" block (~37 KB). The operator couldn't manipulate the
greeting, the CTA, or the expiry badge independently — they had to
edit raw HTML inside one block.
New scripts/build-native-template.js generates matched .json
(Unlayer design tree) + .html (compiled output) from a JS template
spec under scripts/templates-spec/. Each block becomes a separate
entry in the design tree with its own type:
- 9 text blocks : greeting, urgency, body, expiry, prorata,
Option 2 text, signature, contact, dark footer
- 2 image blocks : header logo, footer logo
- 1 button block : the CTA (🎁 {{amount}})
- 4 html blocks : view-in-browser, Option 1 chip, brand-logo
card, Option 2 chip (kept as raw HTML — too
custom for native equivalents)
gift-email-native-reminder-fr ships as the proof of concept:
- Compiled HTML: 30,867 bytes (vs 39,484 for the MJML-compiled
reminder-fr — saves 22%)
- JSON: 42,274 bytes (essentially same as before, but now broken into
16 individually-editable blocks instead of 1 monster Custom HTML)
What this unlocks in Unlayer:
- Click any text → font / color / size / padding / alignment in the
right panel
- Click the CTA → button-specific controls (corner radius, hover
color, padding)
- Drag-reorder blocks within the email
- Mobile preview reflects each block's responsive defaults
- Save a block to the personal library for reuse in other campaigns
Limitations on the 4 html blocks:
- Chips (Option 1 / Option 2) require raw HTML edit because the
rounded badge styling has no native equivalent
- Brand-logo strip needs precise inline img widths Unlayer can't set
Once the operator validates rendering across Gmail/Outlook/Apple
Mail, we'll port the rest: gift-email-fr/en + the existing reminder
templates can all migrate using the same build script.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
A reminder campaign is a deep-copy of its parent's non-clicked
recipients with NEW gift_tokens. Clicks on the reminder were flagging
the CHILD recipient's gift_link_clicked but the parent campaign's
counters never updated — operators had to check two campaigns to see
the cumulative click rate.
Hub:
- New cascadeClickToParent() helper — when a recipient with
parent_campaign_id is flagged as gift_link_clicked, mirror the flag
+ timestamp onto parent.recipients[parent_row_index] and broadcast
a recipient-update SSE event so the parent's open page refreshes
live. Adds a gift_clicked_via_reminder breadcrumb (the child
campaign id) so the parent UI can show "↩ via la relance XXX".
Idempotent — already-clicked rows are no-op.
- Three cascade call sites: applyWebhookEvent fast path (CustomID),
applyWebhookEvent fallback (msgId scan), handleGiftRedirect wrapper.
- handleGiftRedirect also now sets gift_link_clicked=true on first
successful redirect (Mailjet webhook can lag or drop; the wrapper
redirect is the most reliable click signal we have).
- GET /campaigns/:id now attaches a "reminders" array with summary
counters for every reminder child of the campaign.
Ops UI:
- "Cette campagne est une relance" banner on child detail pages with
a back-link to the parent.
- "N relance(s) envoyée(s)" banner on parent detail pages with
clickable chips showing each child's gift_clicked/total ratio.
- Recipient table: 🔁 icon next to the gift-click indicator when the
click came via reminder, plus a "↩ via la relance XXX" line in the
tooltip so the operator can trace the engagement channel.
One-time backfill applied on prod to mirror clicks that happened
between reminder send and this deploy (1 click cascaded —
cmp-20260522-2d4605 gift_clicked 27 → 28).
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Surfaces the WHY of the gift in the reminder body. The original
campaign was sent ~10 days before the reminder fires — recipients
may have forgotten the loyalty/gratitude context, leading to a
"what is this?" reaction when they see the reminder cold.
Adding two words ("pour te remercier" / "as a thank-you") cheaply
reconnects with the original messaging and reinforces TARGO's
relationship framing.
FR: "La carte-cadeau qu'on t'a envoyée pour te remercier peut
s'utiliser chez des centaines de marques canadiennes..."
EN: "The gift card we sent you as a thank-you can be redeemed at
hundreds of Canadian brands..."
Greeting kept as "Petit rappel pour {{firstname}}," — first-name
personalization beats generic "toi" on engagement metrics, and the
firstname auto-clean covers ~99% of recipients.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The edit-params picker was showing "2026-06-22" for an expiry stored
as 2026-06-22T03:59:59Z because it sliced the UTC string. But that
UTC instant is actually 23:59 EDT on June 21 in Montreal, which is
what the email recipient sees (and what the operator picked).
Fixes both sides of the round-trip:
DISPLAY (UTC → picker)
- Convert stored ISO UTC to YYYY-MM-DD interpreted in America/Montreal
using en-CA locale (which returns ISO-style YYYY-MM-DD).
SAVE (picker → ISO UTC)
- New endOfDayMontreal() helper that probes Montreal's offset for the
target date (noon UTC always lands in morning Montreal, never spans
a day) and anchors at 23:59:59.999 local. Handles EDT/EST swaps
automatically — verified with edge cases 2026-03-08 (post-DST-spring),
2026-06-21 (mid-summer), 2026-11-01 (post-DST-fall), 2026-12-31 (winter).
Previously the save path relied on the BROWSER's local TZ inference
(new Date('YYYY-MM-DDT23:59:59').toISOString()) which is fine for
Quebec operators but quietly wrong for anyone editing from elsewhere.
The bulk email send was already correct because the worker's
toLocaleDateString uses timeZone: 'America/Montreal' (last commit).
This commit only fixes what the OPERATOR sees in the picker.
🤖 Generated with [Claude Code](https://claude.com/claude-code)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
