Adds create_giftbit_campaign.js — Node CLI that POSTs to the Giftbit
API (testbed or production), creates a campaign with
delivery_type=SHORTLINK so Giftbit does NOT send their own English
template emails, polls /gifts?campaign_uuid=... until the redemption
shortlinks are generated, then writes a gifts CSV ready to feed into
send_gift_campaign.js.
Two non-obvious things learned while wiring it up:
1. The right endpoint to get the shortlinks is /gifts (not /links).
/links/{uuid} returned 0 rows on our sandbox account; /gifts has
a `shortlink` field on each gift once delivery_status transitions
from QUEUED → LINKCREATED. Polled with 2s interval, up to 20 tries.
2. delivery_type=SHORTLINK is mandatory. Default is GIFTBIT_EMAIL,
which fires their English template immediately — defeating the
whole point of bridging through our French Mailjet template.
Confirmed in the campaign GET response that delivery_type echoes
back correctly when we send "SHORTLINK".
Validated end-to-end (entirely synthetic data — Alice/Bob/Charlie at
@example.com, no real customer info in the sandbox):
✓ Auth probe via /ping returns 200
✓ POST /campaign returns campaign UUID
✓ After ~12s, /gifts returns 3 gifts each with a working shortlink
✓ send_gift_campaign.js consumes the gifts CSV + the contacts CSV
✓ FR template renders: "Bonjour Alice", http://gtbt.co/7TKGFDBNVZq
embedded in the CTA button href, address in the footer line
The --sandbox flag does double duty: routes the API to
api-testbed.giftbit.com AND replaces every recipient email with
louis@targo.ca so we can't accidentally hit real customer inboxes
with the non-redeemable test gifts.
README updated with the two-stage pipeline (create → send), explicit
warnings about the customer-matching gap (only 25% of source rows
resolve via legacy_delivery_id — the rest use a different ID space
from the source Map tool), and the sandbox-quirk where Giftbit
collapses recipient_name when emails are duplicated.
Token NOT committed — pulled from GIFTBIT_TOKEN env var per the
script's contract. In production we'll store it in the hub's
.env alongside SMTP_USER / SMTP_PASS.
User context: needs to send Giftbit gift cards to 203 customers with a
branded French email instead of Giftbit's English-only default delivery.
Giftbit's own UI/API can issue the gifts but its email is English; this
MVP bridges the gap by taking the gift URLs back from Giftbit, pairing
them with our contact CSV, and sending personalized FR emails through
the Mailjet SMTP that's already wired up for ERPNext invoice mail.
Three files in scripts/campaigns/:
1. send_gift_campaign.js — Node CLI. Two CSV inputs (gifts + contacts),
matches by row order (default) or email key, renders the HTML
template with mustache-style {{firstname}} / {{gift_url}} / etc.,
sends via nodemailer with configurable SMTP + throttle.
--dry-run writes per-recipient previews to disk for visual review
before flipping to live mode. Results CSV with per-row status
(sent / failed / dry-run) + error message + timestamp is written
next to the script for follow-up on failures.
2. templates/gift-email-fr.html — branded French email. Table-based
layout (the only thing that renders consistently in Gmail / Outlook /
iOS Mail / Apple Mail / Bell Sympatico). Indigo gradient header,
centered CTA button, contextual {{description}} line citing the
service address, support contact in the footer, no inline images
(defers to text + colour blocks to dodge image-blocking).
3. contacts_from_legacy.py — replaces the ad-hoc /tmp Python I ran
earlier with a proper repo'd version. Same multi-email handling
options (first / split / skip) as I offered the user; defaults to
"first" = 1 gift per household, which is what they chose. Title-
cases the address with French article rules (de / du / la / aux
stay lowercase, 1re / 2e ordinals stay lowercase too).
4. README.md — end-to-end usage with the actual SMTP env vars from
/opt/targo-hub/.env and the matching strategy decision matrix.
Validated end-to-end with a 5-row dry run: matching works, accents
preserved (Amélie, Geneviève, Marc-André), {{firstname}} interpolates,
gift URLs land in the rendered button href, address shows in the
contextual footer line. Previews written to disk for visual QA.
NOT in this MVP (out of scope, can come next if we end up running
gift campaigns regularly):
- No persistence to ERPNext doctype (no Gift Campaign / Recipient
records — pure CLI, results CSV is the audit trail)
- No click-tracking redirect (the gift_url goes verbatim to the
recipient; Giftbit's own API/dashboard reports redemption status,
which is the more relevant signal than "clicked the link")
- No ops UI page (CLI is fine for one-shot; if this becomes regular
we wrap it in services/targo-hub/lib/gift-campaign.js + a Vue page)
Done what the docs suggested in c31a9e0 — actually installed
excel-azmin/frappe_pg on prod erp.gigafibre.ca (1.0.0, master pinned
at commit a237f5995b). Hit one compat bug along the way and fixed it.
The bug
frappe_pg monkey-patches PostgresDatabase.commit() and .rollback()
with wrappers that have the OLD `(self)` signature. Frappe 16.12+
now calls `db.rollback(chain=True)` from app.py:sync_database(),
which makes every HTTP request crash with:
TypeError: patched_rollback() got an unexpected keyword argument 'chain'
Symptom: HTTP 500 on /api/method/ping, Sales Invoice list, etc.
Customer Server Scripts that don't return through sync_database
(like our customer_balance) still worked, which is why the bug
only surfaced after the post-install restart.
The fix
Two-part: signatures take `*args, **kwargs`, and the wrapped call
forwards them to the original. Idempotent sed.
-def patched_rollback(self):
+def patched_rollback(self, *args, **kwargs):
- return _original_rollback(self)
+ return _original_rollback(self, *args, **kwargs)
Both files in frappe_pg need it: postgres/database_patches.py and
patches/postgres_fix.py. Same fix for patched_commit.
Saved as patches/fix_frappe_pg_signatures.sh so we can re-apply after
every `bench update` or fresh install. The comment block in the
script documents why it exists and links the upstream issue (TODO:
file PR at excel-azmin/frappe_pg). docs/SETUP.md §7 was updated to
mention the post-install patch step, the nginx-IP-cache gotcha that
will produce a confusing 502 if you only restart the backend, and
the correct repo (excel-azmin, not the-commit-company that I had
hallucinated in the previous commit).
Smoke test results post-install:
ping, Customer list, Sales Invoice list, Service Location list,
customer_balance Server Script, ops SPA, hub /health — all 200.
ERPNext was built for MariaDB; we run it on PostgreSQL because that's
what fit the legacy migration. Frappe's SQL generator is loose on
MariaDB (missing GROUP BY columns OK, double-quoted strings OK,
HAVING without GROUP BY OK) but strict on Postgres, so we end up
hand-patching files in `patches/fix_pg_groupby.py` after every
ERPNext upgrade. The community has packaged a comprehensive fix as
a Frappe app — `frappe_pg` — that covers the same bugs in one
place. The cleaner path long-term is to install that app instead
of growing our own patch set.
Two doc updates:
- docs/architecture/overview.md §6 item 8 — full background:
the 3 SQL patterns that break (GROUP BY, HAVING, double-quoted
string literals), the 12 hotspots we've already patched, the
4 known remaining (bank_clearance, bank_reconciliation_tool,
accounts/utils L1660, gross_profit), and the install
recommendation with trade-offs (pin a commit, validate on
staging, keep our patches as backup for 4-6 weeks).
- docs/SETUP.md §7 — quick-start install commands for whoever
decides to flip the switch, plus the warning about pinning
rather than tracking main. Also notes that custom Server
Scripts with raw SQL (like `customer_balance`) need the same
single-quote vs double-quote vigilance even after installing
frappe_pg, and the export-fixtures hint to version-control
them.
Three legacy data-quality issues that were leaking into ERPNext on every
import run. Caught while auditing C-LPB4's mis-pinned dispatch job.
1. **Postal code embedded in address_line.** Legacy `gestionclient` had
rows like `2200-3 chemin de la riviere de la guerre J0S1B0` with
the postal code concatenated at the end (and the same code repeated
in the dedicated zip column). Caused 48-char address_line on what
should have been a 39-char address. Now stripped at import: a regex
matches `\\s+<FSA><LDU>\\s*$` (with or without space) and removes
it; the dedicated postal_code field carries the canonical form.
2. **Abbreviations + Cobol-style capitalization.** Legacy stored
`2066 Ch De La 1Re-Concession` instead of the canonical
`2066 Chemin de la 1re-Concession`. ABBREV_MAP expands `Ch` →
`Chemin`, `Av` → `Avenue`, `Bd`/`Boul` → `Boulevard`, `Rte` →
`Route`, `St-` → `Saint-`, `Ste-` → `Sainte-`, `Mtl` → `Montréal`.
Title-casing rule preserves French articles lowercase (`de`, `du`,
`des`, `la`, `le`, `les`, `au`, `aux`, `à`, `et`, `sur`, `en`)
and ordinal markers (`1re`, `2e`, `3e`). 96 SLs in production had
the `1Re-Concession` style; they'll be re-normalized on next
migration run.
3. **`connection_type` left empty even when ONT/CPE devices existed.**
Pre-loads device→delivery mapping at import start; if the legacy
delivery has any device whose category/name/model contains "ont",
"onu", "cpe", "fibre", "gpon", or "ftth", we set
connection_type='Fibre FTTH'. Without devices on file, the field
stays empty (rep fills it later) — we don't guess.
4. **`postal_code` normalized too** — `j0s1b0` → `J0S 1B0` (uppercase
+ canonical space). Was being inserted in lowercase no-space form.
Self-tested on 8 representative cases including the actual broken
records found in production (LOC-15903, LOC-6227, LOC-4 / C-LPB4).
These changes affect only re-imports of locations. Existing data
needs a separate backfill script — a follow-up will cover that
either as a one-shot migration or by running the existing
`reimport_subscriptions.py` after this script.
Three connected dispatcher-facing issues from C-LPB4 audit:
1. **Monthly total was wrong on customer cards.** Section subtotal and
`locSubsMonthlyTotal` summed `actual_price` for ALL subscriptions
regardless of status, so cancelled rows (rendered with strikethrough)
still pumped up the displayed billing figure. C-LPB4 showed
"Total mensuel: 86,10$" computed as `196.05 - 109.95 = 86.10`,
where 196.05 included 3 cancelled internet plans (Megafibre 80,
TEST-E2E-FTTH, FTTH100 — all struck through in the UI). Real
active monthly is 5.00$ (109.95 active + 5 frais réseau − 109.95
loyalty rebate). Fixed both `sectionTotal` and `locSubsMonthlyTotal`
/`locSubsAnnualTotal` to filter on `status === 'Active'`.
2. **"Lieu" link from a dispatch task pointed to ERPNext desk** which
shows a raw doctype form (no abonnements, no totals, no contacts —
just the bare fields). Now points in-app to
`#/clients/<customer>?location=<SL>`. ClientDetailPage reads the
query string on mount and:
• scrolls the matching `loc-card` into view
• pulses an indigo halo around it for ~2s so the rep finds it
immediately even when the customer has many service locations.
3. **The shipping/billing distinction was invisible** on the customer
page. Added an "Adresses de livraison" badge next to the "Lieux de
service" section title — clarifies that this section IS the
shipping address, distinct from the (future) billing address that
will live on the Customer record. Cosmetic for now; the data
migration to formalize that distinction is the next step.
These three round out the C-LPB4 audit triggered by the wrong
mapbox-pin location: now the customer card on the dispatcher's
screen shows correct totals, the dispatch link drops them right at
the spot they're trying to reach, and the role of each address-bearing
record is named explicitly.
Two things ride together because the user noticed the URL bug while
testing the work-in-progress address validation:
1. **Broken Frappe URL pattern.** Three places in the dispatch UI
were generating `/desk/Service Location/<id>` and
`/desk/Dispatch Technician/<id>` links — both return "Page not
found" on Frappe v14+ (= our v16) because the modern desk URL
format is `/app/<slug>/<id>` where slug is lowercase + hyphens.
Fixed in:
• RightPanel.vue (Lieu link in the job details panel)
• DispatchPage.vue (Lieu in the job ctx menu)
• DispatchPage.vue (Ouvrir dans ERPNext in the tech ctx menu)
2. **`POST /address/validate` endpoint** on the hub. Wraps the
existing RQA Supabase search (`address-search.js`) with a
confidence-scored output:
• exact_match (boolean) — score >= 0.7
• best (the top RQA candidate with aq_address_id, lat, lng)
• candidates[] (top 5 ranked)
• confidence (0..1)
• recommendation: validated | review | unmatched
Score combines civic-number exact match, road-name fuzzy overlap,
FSA+full postal-code bonuses, and city-name bonus. The endpoint
is called from ops UI when adding/editing a Service Location to
auto-populate aq_address_id + canonical lat/lng instead of
trusting human typing or Mapbox geocode.
(Custom Fields aq_address_id, address_validation_status,
address_validated_at, linked_address have been added on Service
Location via the Frappe REST API in a separate operation — not in
this commit since they're DB-only.)
Two related issues, one PR:
1. **Bad coords** on customer C-LPB4's "Wifi buggy" job (DJ-MNP8WIKT).
Address on file is `691 rue des Hirondelles, Saint-Michel J0L2J0`,
but the saved lat/lng (-73.677086, 45.159206) reverse-geocodes to
`2336 rue René-Vinet, Sainte-Clotilde J0L1W0` — ~9 km away. The
delta matches the Gigafibre HQ default fallback (-73.6756, 45.1599)
pretty closely, suggesting the geocoder either failed silently at
Service Location creation time or got pinned to the HQ centroid.
Fixed live in DB (UPDATE on tabService Location LOC-0000000004 +
tabDispatch Job DJ-MNP8WIKT to lng=-73.5792377, lat=45.2408452,
verified via Nominatim against the typed address). The job pin
should now show on the correct house.
2. **No way to jump from a job to the client** — the dispatcher had
to memorize/type the customer ID. Now both the RightPanel and the
job context-menu surface clickable shortcuts:
• Client → `#/clients/<id>` (opens ClientDetailPage in-app)
• Lieu → `/desk/Service Location/<id>` (opens ERPNext in a new
tab; the ops SPA doesn't have a dedicated SL detail page)
Required wiring `customer` + `serviceLocation` into the job map in
`stores/dispatch.js` — the API (`fetchJobsFast` uses `["*"]`) was
already returning the fields, the store just wasn't surfacing them.
Note on the deeper bug: the SL lat/lng is the source of truth and the
job currently *copies* it at creation time (rather than reading from
the SL link dynamically). If a Service Location's coords are corrected
after a job exists, the job retains stale coords. A follow-up could
either (a) re-read on render, or (b) trigger a backfill when SL coords
change. Out of scope for this fix — for now, the dispatcher who fixes
an SL must also update any open jobs at that location.
Three places in last week's docs refresh got the DocuSeal URL wrong —
must have been a copy-paste glitch since other parts of the same docs
(roadmap.md, module-interactions.md, archive snapshots) had it right.
Verified against the deployed Traefik labels:
traefik.http.routers.docuseal.rule = Host(`sign.gigafibre.ca`)
`docs.gigafibre.ca` doesn't even resolve in DNS. Fixed in:
• README.md (services table)
• docs/README.md (services tree)
• docs/architecture/overview.md (infra ASCII diagram)
Mass refresh — the docs were last touched 2026-04-22, two weeks behind
shipped reality. This commit updates 9 files to reflect current truth.
WHAT CHANGED IN THE PRODUCT (since 22 Apr) THAT THE DOCS NOW REFLECT:
• Oktopus CE / TR-369 stack decommissioned (containers + volumes +
images all removed; broker had filled /dev/sdb with 75 GB of debug
logs and took ERPNext down for 4 days). Hub gates the integration
behind OKTOPUS_DISABLED=1 — modules retained, no-op'd at runtime.
• dispatch.gigafibre.ca (legacy PHP SPA) replaced by an nginx 301
redirect to /ops/#/dispatch.
• Top toolbar of the dispatch module: collapsed to single-color
Lucide icons + ⋯ overflow menu + "Vue principale ▾" + "[👥 N ▾]"
resource type chip (defaults to techs, materials in the dropdown
only when relevant).
• Tech home base / departure point: editable per-tech via 📍 button,
address geocode (Nominatim) or click-on-map picker, right-click
on tech pin opens the same actions. Map defaults centered on
Gigafibre HQ (1867 chemin de la Rivière, Sainte-Clotilde) instead
of downtown Montreal.
• POST /auth/users invite flow on the hub: creates the Authentik
user, sets a temp password, mails it via Mailjet (Authentik's
own recovery flow isn't configured), creates the matching ERPNext
System User. Surfaced in ops Settings → Utilisateurs → Inviter.
• Two Authentik instances clarified as parallel-and-permanent (not
a migration): auth.targo.ca for staff, id.gigafibre.ca for clients.
FILES TOUCHED:
README.md — service table refreshed, arch diagram redrawn (no
Oktopus row), auth section explains the invite flow + two
parallel instances.
docs/architecture/overview.md — new "Decommissioned" section,
correct retirement status for dispatch-app + apps/field, two
Authentik instances explicitly distinguished, dev-gotchas list
rewritten (drops MongoDB AVX, adds log-rotation hard-learned
lesson, adds note about Authentik recovery flow).
docs/architecture/data-model.md — Step 5 hardware provisioning
now describes the GenieACS path (TR-069 Inform → preset push)
instead of the dead TR-369 path.
docs/architecture/module-interactions.md — oktopus.js and
oktopus-mqtt.js entries marked as gated, provision.js note
updated, GenieACS row in external-integrations updated, MQTT
row removed from real-time channels, interaction matrix loses
the Oktopus column and gains an Authentik admin REST cell.
docs/features/dispatch.md — Top bar section completely rewritten
to match the current chrome (left/center/right regions,
single-color Lucide, dropdowns); new Tech home base section
documenting the 📍 + map-pick + right-click flows; retirement
note now reads as a status, not a plan.
docs/features/cpe-management.md — full rewrite. Oktopus migration
plan replaced by a "decommissioned" note + the existing GenieACS
+ modem-bridge architecture as the steady state. TP-Link XX230v
deep-dive sections preserved (still accurate).
docs/README.md, docs/features/README.md, docs/roadmap.md —
intent-table descriptions and live-URLs table corrected.
The docs/archive/ snapshots (2026-04-18, 2026-04-19) are untouched —
they're historical and should remain that way.
A few connected fixes around the invite UI shipped in 81d61aa:
1. **Bug in 81d61aa**: `auth.js` referenced `erpFetch` without importing
it, so every invite returned `erpnext.ok=false` with the silent
"erpFetch is not defined" error in the catch. Imported it from
./helpers alongside the other helpers we already used.
2. **Authentik recovery flow not configured** (caught while smoke-testing):
the brand `auth.targo.ca` has `flow_recovery=None` and no SMTP, so
`POST /core/users/{pk}/recovery_email/` returned 400 "No recovery
flow set." Rather than build out a full Authentik recovery flow
via API (multiple stages, brand patch, SMTP env var changes), the
hub now generates a strong-but-readable temp password
(`X7K2-9NQB-4GHM-3RTW` style — no look-alike chars), POSTs it via
`/core/users/{pk}/set_password/`, and emails it via the existing
Mailjet SMTP (already wired into lib/email.js for invoice sends).
Returns `{temp_password, password_set, email_sent}` so the admin
has a fallback if Mailjet drops the message.
3. **Settings dialog** now shows a credentials panel after submit:
• Green banner "✓ Courriel envoyé" when email_sent=true
• Yellow "⚠ transmettez manuellement" when email_sent=false
• The temp password as a copyable field either way
• ERPNext User creation status
4. **Dev onboarding**: added `apps/ops/.env.example`,
`services/targo-hub/.env.example`, and a top-level `docs/SETUP.md`
that explains the local-dev flow (clone → cp .env.example .env →
npm install → npx quasar dev). The example envs are commented
per-section so a new dev knows which keys correspond to which
external integration. None of the real secrets are checked in —
the .gitignore already covers .env files.
Surfaces a "Inviter" button in Settings → Utilisateurs that, in one
round-trip:
1. Creates the Authentik user (random password, requested OPS_GROUPS,
auto username from local-part of email with collision suffix).
2. Triggers Authentik's recovery email so the user picks their own
password on first login. If the Email stage isn't configured,
falls back to /core/users/{pk}/recovery/ which returns a one-time
URL the admin can copy + send via SMS or Slack.
3. Creates the matching ERPNext System User with the requested
roles (default: Employee) and `social_logins=[{provider:authentik,
userid:email}]` so OAuth2 finds them on first SSO login.
send_welcome_email=1 also fires Frappe's invite mail.
Idempotent on both sides: if the Authentik user already exists, we
PATCH the requested groups; if the ERPNext User exists, we skip the
POST and return existing=true. Lets the admin re-invite somebody
after a botched first try without breaking anything.
UI:
• "Inviter" button next to the user search bar, gated by the
`manage_users` capability (existing pattern).
• q-dialog with full_name + email + chip-pickable Authentik groups
(admin/sysadmin/tech/support/comptabilite/facturation/dev) + a
comma-separated ERPNext roles input (defaults to Employee).
• Optimistic insert into the visible list on success; the next
search reconciles.
Two issues conflated in the same PR because they touch the same pixels:
1. **Resource filter no longer treats techs and materials as equals.**
Was a 3-button inline toggle [Tous][👤 45][🔧 6] with all three
visually similar — and the wrench glyph clashed with the wrench
used for the filter-settings button. Now:
• Default = 'human' (techs only). Materials are secondary
resources; they don't deserve front-of-bar real estate.
• Single chip [👥 45 ▾] in the toolbar. Click → dropdown:
· Techs (45) ← active by default
· Matériel (6) (only shown if materialCount > 0)
· Tous (51) (only shown if materialCount > 0)
• Defaults to localStorage 'sbv2-filterResType' if previously
persisted, otherwise 'human' instead of '' (was '').
2. **Mixed-style icons (emoji + Lucide SVG) replaced with consistent
single-color Lucide-style strokes.**
Each is a stroke-only inline <svg> with stroke="currentColor", so
they inherit the surrounding text color (no green/red/yellow
tinting). Added to the existing ICON set in useHelpers.js:
user, users, package, sliders, chevDown, map, clipboard,
sparkles, signal, rotateCw, alertTri, moreH, pause, play,
externalLink, target, calendar
Replaced in the dispatch top toolbar:
⚠️ → ICON.alertTri (overload alert)
📋 → ICON.clipboard (unscheduled jobs)
🗺 → ICON.map (map toggle)
🗓 → ICON.calendar (planning toggle)
👥 → ICON.users (team-jobs button + Ressources menu)
🔧 → ICON.sliders (filter-settings — was wrench, which
collided with the materials filter)
👤/🔧 → ICON.users / .package (resource type dropdown)
↻ → ICON.rotateCw (refresh in ⋯ menu)
✨ → ICON.sparkles (AI in ⋯ menu)
📡 → ICON.signal (offers in ⋯ menu)
↗ → ICON.externalLink (ERPNext link in ⋯ menu)
⋯ → ICON.moreH (the ⋯ button itself)
.sb-icon-svg gives them consistent sizing (14px in buttons, 15px
in dropdown items, 16px in the ⋯ trigger) so they're crisp at all
the spots they appear.
Emojis still in place elsewhere (job-tile chips, status badges, etc.)
will be migrated incrementally — out of scope for this pass which
only targeted the user's visible header.
Four fixes around the dispatch header following dispatcher feedback:
1. **⋯ overflow menu was invisible**: .sb-header had `overflow:hidden`,
which clipped the absolutely-positioned dropdown right at the
header's bottom edge. Switched the header to `overflow:visible`
(children all have flex-shrink:0 + a flex:1 center, so the layout
doesn't actually overflow horizontally). Bumped z-index to 5000
for safety on top of map/calendar layers.
2. **NLP/Assistant IA bar hidden by default**: was eagerly rendering
on every page load, with the long French placeholder polluting
the header below the toolbar. The user just wanted the icon. Now
`nlpVisible` defaults to false, persisted in localStorage so power
users who flip it on keep it open across sessions. Toggle still
lives in the ⋯ menu.
3. **Click a tech in the resource list now flies the map to them**:
selectTechOnBoard previously only opened the map panel. Now it
also `map.flyTo({ center })` using `tech.gpsCoords ?? tech.coords`
— live Traccar position wins when the tech is online; falls back
to the saved home base. Animated, deferred a tick so map.resize()
happens first, otherwise flyTo can land on garbage coords during
the panel's open transition.
4. **Board view tabs collapsed into a "Vue principale ▾" dropdown**:
was [Vue principale][Par région][+] inline. Now a single button
showing the active view; click reveals the others + the future
"+ Nouvelle vue" entry. Same dropdown component as the ⋯ menu
(shared CSS, click-outside + ESC close).
The header right-side was getting noisy — 8 buttons + 2 indicators
all competing for screen width, with two visually-similar 📡 icons
(offer pool vs GPS settings) that confused dispatchers. On narrower
laptops the bar wrapped or icons overflowed.
New layout:
[⚠ overload] [📋 unassigned + count] [🗺 Carte] [Publier + count] [+ WO] [⋯]
Everything else dropped into the ⋯ dropdown:
• ↻ Actualiser
• ✨ Assistant IA
• 📡 Offres aux techs (with green count badge)
• 👥 Ressources & GPS ← was 📡 in the bar; this is also where
the tech-management UI (rename, deactivate,
home location, Traccar device link) lives
• ↗ Ouvrir ERPNext (with the inline status dot)
The ⋯ menu closes on Escape, on click outside, and after picking an
item. Same close-handler chain that already serves the job/tech
context menus.
The kept-up-front buttons all have either a status badge (counts,
overload alert) or are the primary CTAs (Publier, + WO) — so the
dispatcher's eye stays on workflow signal, not on chrome.
Three connected UX changes:
1. **Map centered on Gigafibre HQ on first load** —
Sainte-Clotilde (lng=-73.6756, lat=45.1599), zoom 10 — covers the
service area (Sainte-Clotilde + Châteauguay + Napierville +
Hemmingford). Was downtown Montréal.
2. **Right-click on a tech pin** opens the existing techCtx menu
(already used from the calendar via @ctx-tech). New entries:
• 📍 Adresse de départ… → openTechHomeDialog
• 🎯 Choisir sur la carte → startTechGeoFix (mirrors the existing
geoFixJob flow used for jobs)
3. **The 📍 button in the GPS sidebar** now offers a 2-option chooser
first: "Saisir une adresse" or "Cliquer sur la carte". Picking the
map option drops the user into geoFixTech mode.
Implementation:
• useMap.js: new geoFixTech ref + startTechGeoFix/cancelTechGeoFix
+ a contextmenu listener on each tech outer wrapper that calls
openTechCtx(e, tech). The map's main click handler now branches:
if geoFixTech is set, persist the lng/lat via saveTechHome (passed
in via deps as a forward-bound arrow because saveTechHome is
destructured below the useMap call in DispatchPage).
• DispatchPage.vue: new banner shown while in pick mode (animated
indigo bar at top, "Cliquez sur la carte pour {tech}", with a
cancel button); ESC also cancels.
• dispatch-styles.scss: .sb-geofix-banner styles + reusing the
existing pulse keyframe.
Two changes around tech "departure point" coords (used for route
optimization when the tech has no live GPS yet):
1. New default fallback = 1867 chemin de la Rivière, Sainte-Clotilde
(Gigafibre HQ, lng=-73.6756, lat=45.1599). Was downtown Montréal,
which never made sense — every tech started the day with a 70 km
imaginary commute.
2. Per-tech editable home base via a 📍 button on each row of the
tech sidebar. Clicking it opens a dialog that accepts either:
• a free-text address — geocoded via OpenStreetMap Nominatim
(browser-side, sane User-Agent, no hub proxy needed)
• or a literal "lat, lng" pair pasted directly
On confirm: PUT to ERPNext (Dispatch Technician.latitude /
.longitude), patch the local store row, and trigger a route
recompute since the start point changed.
The geocode hits Nominatim public — fine for a low-volume
internal tool. If we ever exceed their fair-use limits, swap to
the existing /address-search hub route which already has the
AQ + RQA pipeline.
The map marker container was being created with an inline
`position:relative`, which overrode Mapbox GL's `.mapboxgl-marker`
class (which applies `position:absolute`). Mapbox writes
`transform: translate(<x>, <y>)` to that exact element on every
zoom/pan frame to project lat/lng → screen coordinates. With the
element kept in the document flow (relative), the transform is
interpreted against the document origin instead of the map pane,
so the pin visually drifts as the user zooms in on a tech.
Removing the inline `position:relative` lets the Mapbox class win.
The SVG ring and the avatar div are children with `position:absolute`
inside outer; absolute children only need a positioned ancestor to
form a containing block — `position:absolute` (Mapbox's value)
qualifies just as well as relative, so the avatar stays centered.
The Oktopus TR-069 stack is being decommissioned (broker + ACS + Mongo
+ NATS + adapters). Its MQTT broker was running with debug logging and
spammed 75 GB of "failed publishing packet" lines into a single Docker
log over 13 days — that's what just took ERPNext down for 4 days when
/dev/sdb hit 100 %.
Surface here: hub no longer pulls in the oktopus / oktopus-mqtt modules
when OKTOPUS_DISABLED is set (default = disabled). Keeps the modules
in the tree so we can re-enable later by flipping the env var to 0,
but stops them attempting reconnects to a stack that no longer exists.
• server.js: late-load oktopus + oktopus-mqtt only when enabled.
Routes /oktopus/* now return 410 Gone with a clear message.
• provision.js: same gate. The on-scan handler already had a soft
`if (oktopus && ...)` guard so it naturally no-ops when the
module isn't loaded — no logic change needed there.
Server-side env (set in /opt/targo-hub/.env on prod):
OKTOPUS_DISABLED=1
Contract termination is a fee-bearing, auditable workflow — it belongs
on the contract, not buried in a sub's delete dialog. Standard SaaS /
telecom practice: subs are an immutable event stream, contracts
orchestrate their lifecycle.
ServiceContractDetail.vue (new)
• Status banner: contract type, dates, status — "Résilier" button
when actionable, termination invoice link when already résilié.
• Term progress bar: months_elapsed / duration_months with color
ramp (primary → amber near end → positive when done).
• Financial summary grid: mensualité, abonnement (clickable), devis,
lieu, total avantages, résiduel, signature method & date.
• Benefits detail table: per-row description, regular_price vs
granted_price, économie, reconnu à date, et "À rembourser"
(valeur résiduelle) — this is what the rep needs to see before
deciding to break a contract.
• Termination recap (only when status=Résilié): date, raison,
penalty breakdown, link to the termination invoice.
• "Résilier" action runs a 2-step dialog: first calls
/contract/calculate-termination for the preview, then prompts for
a reason (textarea, min 3 chars) before firing /contract/terminate.
On success: cascade-cancels the linked sub (status=Annulé +
end_date + cancellation_date — no hard delete), mutates the
local doc so the modal refreshes in place, and emits
contract-terminated so the parent page updates its sub + contract
rows + drops an audit comment on the customer.
DetailModal
• SECTION_MAP now routes Service Contract → ServiceContractDetail.
Also added 'Service Subscription' → SubscriptionDetail (same
template fits; was falling through to the generic grid).
• Re-emits contract-terminated so the parent can listen.
ClientDetailPage
• confirmDeleteSub: when a live contract references the sub, the
dialog now simply redirects the rep to the contract modal
("Voir le contrat") instead of trying to do termination from
the sub row. Terminal-state contracts (Résilié/Complété/Expiré)
still get the inline link-scrub path so stale refs don't block
a legit delete.
• onContractTerminated: reflects the cascade locally — contract
row → Résilié, sub row → Cancelled + end_date, audit Comment
posted to the customer's notes feed.
The raw DELETE on Service Subscription was blowing up with
LinkExistsError because Service Contract.service_subscription still
referenced the sub. Worse: silently unlinking a live contract would
cost the business the break fee (résidentiel = avantages résiduels,
commercial = mensualités restantes).
Now when the user clicks 🗑 on a sub:
1. loadServiceContracts pulls `service_subscription` so the client
can spot the link without a round-trip.
2. If a non-terminal contract is linked, the dialog upgrades to:
• header: Contract name + type
• term bar: start → end, months elapsed / months remaining
(pulled live from /contract/calculate-termination)
• penalty breakdown box: total fee, split into benefits to
refund + remaining months, plus a warning that a termination
invoice will be created
• radio: "Désactiver seulement (conserver le contrat)" vs
"Résilier + facturer X$ + supprimer"
Suspend-only route goes through toggleSubStatus (no fee).
Terminate route hits /contract/terminate (status→Résilié +
invoice), then unlinks + deletes the sub, and drops an audit
line referencing the generated invoice.
3. If the linked contract is already Résilié/Complété we just scrub
the stale link inline in the plain confirm path so the
dispatcher isn't forced into the termination UI.
- InlineField on monthly row price (dblclick) + annual row monthly base
price. Saves via Service Subscription.monthly_price → mirrored back
into the UI row's actual_price; drops an audit line on the customer
timeline.
- Delete button (confirm dialog, v-if=can('delete_records')) on both
monthly + annual rows. Uses deleteDoc + local splice + invalidates
location + section caches.
- display_order custom Int field on Service Subscription, persisted in
10-step increments on drag reorder (so manual inserts have room to
squeeze between without a full re-number pass). loadSubscriptions
sorts by display_order first so the dispatcher-controlled order
survives a page reload and can drive invoice print ordering later.
- Rebate rows nested visually: 32px indent + arrow glyph + lighter
red background + smaller type + inherited red color on the inline
price input. Matches the invoice PDF grouping dispatchers expect.
Adding a forfait from the client detail dialog failed with `Update
failed: 417` because the code path manipulated ERPNext's stock
Subscription doctype — a parent/child (Subscription Plan rows) model
with tight validation ("Subscription End Date is mandatory to follow
calendar months"), and whose `plan` field expects an `SP-<hash>` doc
name rather than a free-form string.
Meanwhile all new subscription work — contract signing, chain
activation, prorated invoicing — already writes to our flat custom
`Service Subscription` doctype. The two systems were not talking to
each other: the Service Subscription created for CTR-00008 was
invisible in the client UI (which only read stock Subscription), and
the stock Subscription created by "Ajouter un service" was invisible
to the contract/chain system.
This commit makes Service Subscription the canonical doctype for
everything the ops UI does:
- useClientData.loadSubscriptions: read Service Subscription directly
(flat doc → UI row) instead of reading stock Subscription + joining
its Subscription Plan child rows to Items. Legacy stock Subscription
rows (~39k from the 2026-03-29 migration) stay as audit records
but are no longer surfaced.
- ClientDetailPage.createService: POST a Service Subscription doc
(category inferred from item_group). No parent/child logic, no
calendar-month coupling, no SP-<hash> plan reference. Manual
description + price entry now works without a catalog pick.
- useSubscriptionActions.updateSub: drop the bogus `ASUB-*` name-based
doctype detection (ASUB is not a real prefix — both stock and
Service subs are named SUB-<hex|digits>) and always target Service
Subscription. Also surface ERPNext's exception one-liner instead of
raw HTML when an update fails.
- searchPlans: empty/short query now returns top-50 of the Subscription
Plan catalog so dispatchers can browse instead of being forced to
guess a name prefix.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The hub responds with `Access-Control-Allow-Origin: *`, and the CORS
spec forbids the wildcard + credentials combination. Firefox rejects
the preflight before any response reaches JS, surfacing as
"NetworkError when attempting to fetch resource" when a user clicks
"Supprimer cette tâche" on an already-completed step (or any step).
Every other HUB_URL call in the ops SPA already omits credentials —
aligning TaskNode with the rest of the codebase is the simplest fix.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- contracts.js: _inferPlanName now reads the Quotation's first positive-rate
item ("Internet Megafibre 80 Mbps") instead of generic fallback.
- contracts.js: subPayload writes service_contract back-ref so an active/
pending sub blocks its parent contract's deletion (LinkExistsError).
- contracts.js: GET /contract/audit-orphans[?fix=1] scans for orphaned subs
(dangling contract link or no link at all) and contracts without a sub;
filters out 2026-03-29 legacy-migration batch via LEGACY_CUTOFF.
- dispatch.js: deleteJobSafely() rewires children's depends_on to the
victim's parent, re-parents descendants if victim was chain root, then
deletes. POST /dispatch/job-delete exposes it. Fixes LinkExistsError
when users delete a middle step in the UI.
- TaskNode.vue: confirmDelete calls /dispatch/job-delete and surfaces a
warning when dependents will be rewired.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Root cause of CTR-00008: _createBuiltInInstallChain only created Issue +
Dispatch Jobs. It never created a pending Service Subscription, so when
the chain's terminal job Completed, activateSubscriptionForJob found
nothing matching customer+service_location+status='En attente' to flip.
Result: 4/4 tasks done, no sub activation, no prorated invoice.
Changes:
- contracts.js: after chain creation, create Service Subscription with
status='En attente' (plan_name + service_category inferred from the
contract). Back-link it on Service Contract.service_subscription (a
new custom field — the stock 'subscription' field on Service Contract
points at the built-in ERPNext Subscription doctype, not ours).
- project-templates.js: add test_single (1-step) and test_parallel
(diamond: step0 → step1 ∥ step2) for faster lifecycle testing.
Extract chooseTemplate(contract) with precedence:
contract.install_template → contract_type mapping → fiber_install.
- contracts.js: chain builder now uses chooseTemplate instead of
hardcoded fiber_install, logs the chosen template per contract.
- _inferServiceCategory/_inferPlanName helpers map contract metadata
into the Service Subscription's required fields.
Companion changes on ERPNext (custom fields, no code):
Service Contract.service_subscription Link → Service Subscription
Service Contract.install_template Select (fiber_install,
phone_service, move_service, repair_service, test_single,
test_parallel)
Retroactive repair for CTR-00008 applied directly on prod:
→ SUB-0000100003 (Actif), SINV-2026-700014 (Draft, $9.32 prorata).
Smoke test of test_single path on prod (CTR-00010 synthetic, cleaned up):
template=test_single ✓ sub created ✓ activated on completion ✓
prorated invoice emitted ✓
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- lib/types.js: single source of truth for Dispatch Job status + priority enums.
Eliminates hard-coded 'In Progress'/'in_progress'/'Completed'/'done' checks
scattered across tech-mobile, acceptance, dispatch. Includes CLIENT_TYPES_JS
snippet for embedding in SSR <script> blocks (no require() needed).
- lib/tech-mobile.js: applies types.js predicates (isInProgress, isTerminal,
isDone, isUrgent) both server-side and client-side via ${CLIENT_TYPES_JS}
template injection. Single aliasing point for future status renames.
- lib/acceptance.js: migrated 7 erpFetch + 2 erpRequest sites to erp.js wrapper.
Removed duplicate "Lien expiré" HTML (now ui.pageExpired()). Dispatch Job
creation uses types.JOB_STATUS + types.JOB_PRIORITY.
- lib/payments.js: migrated 15 erpFetch + 9 erpRequest sites to erp.js wrapper.
Live Stripe flows preserved exactly — frappe.client.submit calls kept as
erp.raw passthroughs (fetch-full-doc-then-submit pattern intact). Includes
refund → Return PE → Credit Note lifecycle, PPA cron, idempotency guard.
- apps/field/ deleted: transitional Quasar PWA fully retired in favor of
SSR tech-mobile at /t/{jwt}. Saves 14k lines of JS, PWA icons, and
infra config. Docs already marked it "retiring".
Smoke-tested on prod:
/payments/balance/:customer (200, proper shape)
/payments/methods/:customer (200, Stripe cards live-fetched)
/dispatch/calendar/:tech.ics (200, VCALENDAR)
/t/{jwt} (55KB render, no errors)
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Replaces hand-rolled `erpFetch` + `encodeURIComponent(JSON.stringify(...))`
URL building with a structured wrapper: erp.get/list/listRaw/create/update/
remove/getMany/hydrateLabels/raw.
Key wins:
- erp.list auto-retries up to 5 times when v16 rejects a fetched/linked
field with "Field not permitted in query" — the field is dropped and the
call continues, so callers don't have to know which fields v16 allows.
- erp.hydrateLabels batches link-label resolution (customer_name,
service_location_name, …) in one query per link field — no N+1, no
v16 breakage.
- Consistent {ok, error, status} shape for mutations.
Migrated call sites:
- otp.js: Customer email lookup + verifyOTP customer detail fetch +
Contact email fallback + Service Location listing
- referral.js: Referral Credit fetch / update / generate
- tech-absence-sms.js: lookupTechByPhone, set/clear absence
- conversation.js: Issue archive create
- magic-link.js: Tech lookup for /refresh
- ical.js: Tech lookup + jobs listing for iCal feed
- tech-mobile.js: 13 erpFetch sites → erp wrapper
Remaining erpFetch callers (dispatch.js, acceptance.js, payments.js,
contracts.js, checkout.js, …) deliberately left untouched this pass —
they each have 10+ sites and need individual smoke-tests.
Live-tested against production ERPNext: tech-mobile page renders 54K
bytes, no runtime errors in targo-hub logs post-restart.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Introduces services/targo-hub/lib/ui/ as the shared kit for every magic-link
page served from the hub (tech mobile, acceptance, payments):
design.css tokens (--brand, --success, etc) + reset + all primitives
components.js server-side HTML builders (badge/section/card/panel/statRow/
tabBar) + shared date helpers (fmtTime/dateLabelFr/montrealDate)
+ canonical STATUS_META
client.js client-side api wrapper ($, toast, api.get/post, router.on/go)
baked into every page — no more hand-rolled fetch+hashchange
scanner.js Gemini field-scan overlay (window.scanner.open(field,label,cb,ctx))
shell.js ui.page({title, body, bootVars, cfg, script, includeScanner})
inlines everything into one self-contained HTML doc
index.js barrel
Migrates tech-mobile.js to the kit:
- drops inline esc/toast/fmtTime/dlbl/STATUS_META/badge helpers
- api.post('/status', {...}) instead of fetch(H+'/t/'+T+'/status', {...})
- router.on('#job/:name', handler) instead of hand-rolled route()
- scanner.open(field, label, cb, ctx) instead of ~60 lines of field-scan logic
Behavior preserved — rendered HTML keeps tabs, detail view, notes editor,
photo upload, per-field Gemini scans, Montreal-TZ date labels, v16 link-label
resolution. Verified live at msg.gigafibre.ca with a real TECH-4 token.
Sets up acceptance.js and payments.js to drop from ~700 → ~300 lines each
in the next commits by consuming the same primitives.
Rewrote msg.gigafibre.ca (tech magic-link page) from a today-only flat list
into a proper 4-tab SPA:
- Aujourd'hui: In Progress / En retard / Aujourd'hui / Sans date / À venir
- Calendrier: placeholder (phase 4)
- Historique: searchable + filter chips (Tous/Terminés/Manqués/Annulés)
- Profil: tech info, support line, refresh
Job detail view (hash-routed, #job/DJ-xxx):
- Customer + tap-to-call/navigate block
- Editable notes (textarea → PUT /api/resource/Dispatch Job)
- Photo upload (base64 → File doctype, is_private, proxied back via /photo-serve)
- Equipment section (inherited from overlay)
- Sticky action bar (Démarrer / Terminer)
Equipment overlay extended with per-field Gemini Vision scanners. Each
input (SN, MAC, GPON SN, Wi-Fi SSID, Wi-Fi PWD, model) has a 📷 that opens
a capture modal; Gemini is prompted to find THAT field specifically and
returns value+confidence. Tech confirms or retries before the value fills in.
Root cause of the "tech can't see his job" bug: page filtered
scheduled_date=today, so jobs on any other day were invisible even though
the token was tech-scoped. Now fetches a ±60d window and groups client-side.
vision.js: new extractField(base64, field, ctx) helper + handleFieldScan
route (used by new /t/:token/field-scan endpoint).
Also fixes discovered along the way:
- Frappe v16 blocks fetched/linked fields (customer_name, service_location_name)
and phantom fields (scheduled_time — real one is start_time). Query now
uses only own fields; names resolved in two batch follow-up queries.
- "Today" is Montreal-local, not UTC. Prevents evening jobs being mislabeled
as "hier" when UTC has already rolled to the next day.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Three bugs combined to make CTR-00008 (and likely others) land silently:
1. Fallback was count-based, not outcome-based.
_fireFlowTrigger returned >0 when a broken Flow Template (FT-00005)
"matched" on_contract_signed but did nothing. We took that as success
and skipped the built-in install chain. Now we ALWAYS run the built-in
chain; the idempotency check inside (look up existing Issue linked to
contract) lets a healthy Flow Template short-circuit us naturally.
2. scheduled_date was null on all chained jobs.
createDeferredJobs passed '' when no step.scheduled_date was set, and
the fiber_install template doesn't set one. Jobs with null dates are
filtered out of most dispatch board views, giving the user-visible
"Aucune job disponible pour dispatch" symptom even after the chain was
built. Default to today (via ctx.scheduled_date) so jobs appear on the
board; dispatcher reschedules per capacity.
3. No post-sign acknowledgment to the customer.
Previously the Flow Template was expected to send the confirmation SMS;
since the template was broken, the customer got nothing after signing.
Add _sendPostSignAcknowledgment that sends a "Bon de commande reçu"
SMS with contract ref + service details + next steps. Fires only when
the chain is actually created (not on idempotent skip) so we never
double-notify.
Also:
- Resolve phone/email from cell_phone + email_billing (legacy-migrated
Customer records use those fields, not Frappe defaults mobile_no /
email_id) — otherwise we'd keep skipping SMS with "no phone on file".
- _createBuiltInInstallChain now returns { created, issue, jobs,
scheduled_date, reason } so callers can branch on outcome.
- Export sendPostSignAcknowledgment so one-shot backfill scripts can
re-notify customers whose contracts were signed during the broken
window.
- Set order_source='Contract' (existing Select options patched separately
to include 'Contract' alongside Manual/Online/Quotation).
Backfilled CTR-00008: ISS-0000250003 + 4 chained Dispatch Jobs all with
scheduled_date=2026-04-23, ack SMS delivered to Louis-Paul's cell.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- contracts.js: built-in install chain fallback when no Flow Template matches
on_contract_signed — every accepted contract now creates a master Issue +
chained Dispatch Jobs (fiber_install template) so we never lose a signed
contract to a missing flow config.
- acceptance.js: export createDeferredJobs + propagate assigned_group into
Dispatch Job payload (was only in notes, not queryable).
- dispatch.js: chain-walk helpers (unblockDependents, _isChainTerminal,
setJobStatusWithChain) + terminal-node detection that activates pending
Service Subscriptions (En attente → Actif, start_date=tomorrow) and emits
a prorated Sales Invoice covering tomorrow → EOM. Courtesy-day billing
convention: activation day is free, first period starts next day.
- dispatch.js: fix Sales Invoice 417 by resolving company default income
account (Ventes - T) and passing company + income_account on each item.
- dispatch.js: GET /dispatch/group-jobs + POST /dispatch/claim-job for tech
self-assignment from the group queue; enriches with customer_name /
service_location via per-job fetches since those fetch_from fields aren't
queryable in list API.
- TechTasksPage.vue: redesigned mobile-first UI with progress arc, status
chips, and new "Tâches du groupe" section showing claimable unassigned
jobs with a "Prendre" CTA. Live updates via SSE job-claimed / job-unblocked.
- NetworkPage.vue + poller-control.js: poller toggle semantics flipped —
green when enabled, red/gray when paused; explicit status chips for clarity.
E2E verified end-to-end: CTR-00007 → 4 chained jobs → claim → In Progress →
Completed walks chain → SUB-0000100002 activated (start=2026-04-24) →
SINV-2026-700012 prorata $9.32 (= 39.95 × 7/30).
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Techs reported cloudflare.com showing 300+ms on the diagnostic page
while OS-level ICMP ping returned 5ms. The gap is entirely protocol
overhead:
- fetch() ≠ ICMP. Every call pays DNS + TCP + TLS + HTTP on top of
the real RTT, which is easily 150–300ms cold on mobile LTE when the
radio has to wake the RRC connection.
- Bare cloudflare.com redirects 301 → www.cloudflare.com, forcing a
second DNS + TCP + TLS handshake for every "ping" and doubling
the measured latency.
- TechDiagnosticPage.vue was also labeling the full 10MB download
time as "Latence", so the number on the speed-test card was never
a latency measurement at all.
Fixes, applied to both surfaces (Ops /j/diagnostic + Field /diagnostic):
- Swap cloudflare.com → 1.1.1.1/cdn-cgi/trace. 88-byte response, no
redirect, no keepalive games — canonical "internet is up" endpoint.
- Warm-up fetch before every measurement. First call absorbs DNS +
TCP + TLS + LTE wake; second call reports steady-state RTT. This
applies to checkHosts() (ops) and resolveHost() (field composable).
- Split runSpeed() into separate ping + throughput measurements. Ping
hits speed.cloudflare.com/cdn-cgi/trace (88 bytes on a warm
connection); throughput hits /__down on the same origin so the TLS
session is reused.
Deployed to production; smoke-verified:
- ops bundle TechDiagnosticPage.b925e02c.js contains
'1.1.1.1/cdn-cgi/trace'
- field bundle DiagnosticPage.38a45f65.js contains the same
- zero bare 'cloudflare.com' hostname in either hosts array
Files:
- apps/ops/src/modules/tech/pages/TechDiagnosticPage.vue
- apps/field/src/composables/useSpeedTest.js
- apps/field/src/pages/DiagnosticPage.vue
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The ops tech module at /ops/#/j/* had drifted from the field app in two ways:
1. Scanner — a prior "restoration" re-added html5-qrcode, but the
design has always been native <input capture="environment"> → Gemini
2.5 Flash via targo-hub /vision/barcodes (up to 3 codes) and
/vision/equipment (structured labels, up to 5). Revert useScanner.js
+ ScanPage.vue + TechScanPage.vue to commit e50ea88 and drop
html5-qrcode from both package.json + lockfiles. No JS barcode
library, no camera stream, no polyfills.
2. Equipment UX — TechJobDetailPage.vue was a 186-line stub missing the
Ajouter bottom-sheet (Scanner / Rechercher / Créer), the debounced
SN-then-MAC search, the 5-field create dialog, Type + Priority
selects on the info card, and the location-detail contact expansion.
Port the full UX from apps/field/src/pages/JobDetailPage.vue (526
lines) into the ops module (458 lines after consolidation).
Rebuilt and deployed both apps. Remote smoke test confirms 0 bundles
reference html5-qrcode and the new TechJobDetailPage.1075b3b8.js chunk
(16.7 KB vs ~5 KB stub) ships the equipment bottom-sheet strings.
Docs:
- docs/features/tech-mobile.md — new. Documents all three delivery
surfaces (legacy SSR /t/{jwt}, transitional apps/field/, unified
/ops/#/j/*), Gemini-native scanner pipeline, equipment UX, magic-link
JWT, cutover plan. Replaces an earlier stub that incorrectly
referenced html5-qrcode.
- docs/features/dispatch.md — new. Dispatch board, scheduling, tags,
travel-time optimization, magic-link SMS, SSE updates.
- docs/features/customer-portal.md — new. Plan A passwordless magic-link
at portal.gigafibre.ca, Stripe self-service, file inventory.
- docs/architecture/module-interactions.md — new. One-page call graph
with sequence diagrams for the hot paths.
- docs/README.md — expanded module index (§2) now lists every deployed
surface with URL + primary doc + primary code locations (was missing
dispatch, tickets, équipe, rapports, telephony, network, agent-flows,
OCR, every customer-portal page). New cross-module edge map in §4.
- docs/features/README.md + docs/architecture/README.md — cross-link
all new docs.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Topology clarification:
- portal.gigafibre.ca = standalone nginx container serving /opt/client-app/
(the actual Vue SPA). This is the real customer portal.
- client.gigafibre.ca = ERPNext frontend (exposes Frappe's password login
form — dead-end UX, legacy MD5 attack surface).
Changes:
- apps/client/deploy.sh: target /opt/client-app/ directly with DEPLOY_BASE=/
(was uploading into ERPNext's /assets/client-app/, which nothing serves).
Atomic stage-and-swap + docker restart so the nginx bind-mount picks up
the new inode.
- apps/portal/traefik-client-portal.yml: replace per-path /login and /desk
blocks on client.gigafibre.ca with a catch-all 307 to portal.gigafibre.ca.
Old bookmarks, old invoice links, and in-flight SMS all end up on the
Vue SPA instead of Frappe's password page.
- apps/ops/package-lock.json: sync to include html5-qrcode transitive deps
so `npm ci` in deploy.sh works from a clean checkout.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Customers no longer authenticate with passwords. A POST to the hub's
/portal/request-link mints a 24h customer-scoped JWT and sends it via
email + SMS; the /#/login Vue page sits on top of this and a navigation
guard hydrates the Pinia store from the token on arrival.
Why now: legacy customer passwords are unsalted MD5 from the old PHP
system. Migrating hashes to PBKDF2 would still require a forced reset
for every customer, so it's simpler to drop passwords entirely. The
earlier Authentik forwardAuth attempt was already disabled on
client.gigafibre.ca; this removes the last vestige of ERPNext's
password form from the customer-facing path.
Hub changes:
- services/targo-hub/lib/portal-auth.js (new) — POST /portal/request-link
• 3-requests / 15-min per identifier rate limit (in-memory Map + timer)
• Lookup by email (email_id + email_billing), customer id (legacy +
direct name), or phone (cell + tel_home)
• Anti-enumeration: always 200 OK with redacted contact hint
• Email template with CTA button + raw URL fallback; SMS short form
- services/targo-hub/server.js — mount the new /portal/* router
Client changes:
- apps/client/src/pages/LoginPage.vue (new) — standalone full-page,
single identifier input, success chips, rate-limit banner
- apps/client/src/api/auth-portal.js (new) — thin fetch wrapper
- apps/client/src/stores/customer.js — hydrateFromToken() sync decoder,
stripTokenFromUrl (history.replaceState), init() silent Authentik
fallback preserved for staff impersonation
- apps/client/src/router/index.js — PUBLIC_ROUTES allowlist + guard
that hydrates from URL token before redirecting
- apps/client/src/api/auth.js — logout() clears store + bounces to
/#/login (no more Authentik redirect); 401 in authFetch is warn-only
- apps/client/src/composables/useMagicToken.js — thin read-through to
the store (no more independent decoding)
- PaymentSuccess/Cancel/CardAdded pages — goToLogin() uses router,
not window.location to id.gigafibre.ca
Infra:
- apps/portal/traefik-client-portal.yml — block /login and
/update-password on client.gigafibre.ca, redirect to /#/login.
Any stale bookmark or external link lands on the Vue page, not
ERPNext's password form.
Docs:
- docs/roadmap.md — Phase 4 checkbox flipped; MD5 migration item retired
- docs/features/billing-payments.md — replace MD5 reset note with
magic-link explainer
Online appointment booking (Plan B from the same discussion) is queued
for a follow-up session; this commit is Plan A only.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
The Apr 22 refactor (41d9b5f) collapsed the tech scanner to Gemini-only
photo capture, dropping the live camera viewport and client-side multi-
barcode detection. Techs lost the fast point-and-scan flow that handles
90% of routine installs.
Restored as a hybrid: html5-qrcode as the primary path (instant, offline,
standard QR/barcode), Gemini kept as a second-chance fallback for hard
labels (damaged stickers, text-only serials, unusual symbologies). Offline
queue + scanEquipmentLabel() preserved unchanged.
Three tabs, defaulting to live camera:
- Caméra — continuous html5-qrcode stream, detection auto-beeps
- Photo — native camera; full-image + 3-strip local scan, Gemini fallback
- Manuel — plain text input
Both apps/field and apps/ops updated in lockstep so nothing drifts while
apps/field is being folded into apps/ops/j.
Run `npm install` in apps/ops/ to pull in html5-qrcode before the next build.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
All docs moved with git mv so --follow preserves history. Flattens the
single-folder layout into goal-oriented folders and adds a README.md index
at every level.
- docs/README.md — new landing page with "I want to…" intent table
- docs/architecture/ — overview, data-model, app-design
- docs/features/ — billing-payments, cpe-management, vision-ocr, flow-editor
- docs/reference/ — erpnext-item-diff, legacy-wizard/
- docs/archive/ — HANDOFF-2026-04-18, MIGRATION, status-snapshots/
- docs/assets/ — pptx sources, build scripts (fixed hardcoded path)
- roadmap.md gains a "Modules in production" section with clickable
URLs for every ops/tech/portal route and admin surface
- Phase 4 (Customer Portal) flipped to "Largely Shipped" based on
audit of services/targo-hub/lib/payments.js (16 endpoints, webhook,
PPA cron, Klarna BNPL all live)
- Archive files get an "ARCHIVED" banner so stale links inside them
don't mislead readers
Code comments + nginx configs rewritten to use new doc paths. Root
README.md documentation table replaced with intent-oriented index.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Records what shipped in e50ea88 (scan + device pages, offline store,
Gemini vision pipeline) and lays out the remaining phases: PWA
hardening, auth unification, magic-link tech access, flow-runtime
integration, and final apps/field removal.
Fixes stale `/t/{token}` route reference in Phase 2 → `/j/`.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
- Invoice OCR migrated from Ollama (GPU-bound, local) to Gemini 2.5
Flash via new targo-hub /vision/invoice endpoint with responseSchema
enforcement. Ops VM no longer needs a GPU.
- Ops /j/* now has full camera scanner (TechScanPage) ported from
apps/field with 8s timeout + offline queue + auto-link to Dispatch
Job context on serial/barcode/MAC 3-tier lookup.
- New TechDevicePage reached via /j/device/:serial showing every
ERPNext entity related to a scanned device: Service Equipment,
Customer, Service Location, active Subscription, open Issues,
upcoming Dispatch Jobs, OLT info.
- New docs/VISION_AND_OCR.md (full pipeline + §10 relationship graph
+ §8.1 secrets/rotation policy). Cross-linked from ARCHITECTURE,
ROADMAP, HANDOFF, README.
- Nginx /ollama/ proxy blocks removed from both ops + field.
Co-Authored-By: Claude Opus 4.7 <noreply@anthropic.com>
Tech mobile view (erp.gigafibre.ca/ops/#/j):
- TechLayout with bottom nav tabs (tasks, scanner, diagnostic, more)
- TechTasksPage: rich header with tech name/stats, job cards with
priority dots, time, location, duration badges, bottom sheet detail
with En route/Terminer buttons + scanner/detail access
- TechJobDetailPage: editable fields, equipment list, GPS navigation
- TechScanPage: device lookup by SN/MAC, create/link to job
- TechDiagnosticPage: speed test + host reachability checks
- Route /j replaces legacy dispatch-app tech view
Dispatch unassign confirmation:
- Dialog appears when unassigning published or in-progress jobs
- Warns that tech has already received the task
- Cancel/Confirm flow prevents accidental removal
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
- New JobDetailPage: full-screen job view with editable fields (subject,
type, priority, time, duration, description), status transitions
(en route / terminer / rouvrir), GPS navigation to service location
- Equipment section: list equipment at location, add via scanner/search/create
- TasksPage: jobs now navigate to detail page instead of inline expand,
quick status buttons remain accessible from the list
- Offline support: all edits queued when offline, cached job data
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Before: techDayJobsWithTravel(tech), periodLoadH(tech), techPeriodCapacityH(tech)
were called as functions in the template v-for — recalculated on EVERY render
for every tech (10 techs × 3 functions = 30 expensive recomputations per render).
After: Pre-computed as Vue computed Maps (segmentsMap, loadMap, capMap) that
only recompute when their reactive dependencies actually change. Template
reads from map[tech.id] — instant O(1) lookup, no recalculation.
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
Dispatch performance:
- Replace sequential batch fetches (batches of 15, one after another)
with full parallel Promise.all — all doc fetches fire simultaneously
- With 20 jobs: was ~3 sequential round-trips, now ~2 (1 list + 1 parallel)
Order traceability:
- Add sales_order (Link) and order_source (Select) fields to Dispatch Job
- checkout.js sets order_source='Online' + sales_order link on job creation
- acceptance.js sets order_source='Quotation' on quotation-sourced jobs
- Store maps new fields: salesOrder, orderSource
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>
