A near-miss used to be priced as if it were the item the customer
asked for. Now only an exact match writes a price; a closest match
writes none and records the nearest item with what differs, so the
rep can see it and decide.
The "AI: no match" badge reads "Closest: {name} ({differs})" when
there is a closest item, and the popover shows its price and the
differences.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The AI could only say "this is the item" or "no idea" — it had no way
to say "this is the nearest one, but it is 5mm too short". So a
near-miss either got priced as if it were right, or was thrown away
with nothing shown.
The submit_line_matches tool now takes a match of "exact", "closest"
or "none" plus the list of what differs. The user text shows the
line's parsed size and each candidate's parsed size with the
differences we computed, and the prompt spells out when to use each
kind.
A response with no match field still reads as "exact" when it named a
candidate — the old tool had no match, so those responses keep
working.
pickPosItems() in pos-match.ts runs the same shortlist and the same
tool for any list of lines, one AI call per 15 lines.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The POS shortlist ranked by shared words only, so "BOLT HEX M16X60
GR8.8 ZP" put the M22x65, the M27x170 and the M20x160 all on 0.75 —
they share "BOLT 8.8 ZP" — and the M16 the customer asked for never
reached the AI.
pos-match.ts is now the shared POS matcher: the word pull plus a
size-targeted pull (name ~ "M16" || description ~ "M16"), ranked by
the size comparison and then by word score. A line with no size
information keeps today's word-only behaviour.
- shortlistStockItems delegates to it, so the quote assistant gets
the better list for free.
- the intel popover's fuzzy stock section uses the same ranking, so
the right size wins there too.
- tokenizeIntel / informativeTokens / combinedScore moved to
item-match-text.ts so the matcher and the intel service can share
one implementation without an import cycle.
PocketBase has no IN operator: every filter here is an OR-chain of
~ and = clauses, escaped with escapeForPbString, capped at 300 rows.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The POS shortlist ranked by shared words only, so "BOLT HEX M16X60
GR8.8 ZP" put the M22x65 above the M16x60 — every bolt shares
"BOLT 8.8 ZP" and scores 0.75. The right size never reached the AI
and the price check showed the wrong item.
parseItemSize pulls the type, thread, length, grade, finish and other
dimensions out of free text; compareSizes says whether a candidate is
the same size and, when it is not, in plain words what differs.
Pure and client-safe: no PocketBase, no AI.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Seen on the TEST order: an M16x60 line with no POS item got an M22x65
bolt as its POS price, and "ordered before" listed the order itself.
- the badge summary starts with "Similar items:" when the intel is a
fuzzy match, so it no longer reads like this item's own price
- quoted/ordered before skip the line's own quote / order
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
buildPriceCheck(intel, line) turns the Item Intel payload into the two
things the blue price-check badge needs:
- a one-line summary, e.g. "POS R 0,45 · 120 in stock · last bought
R 0,42 Fowkes (12 days ago) · 2 other suppliers"
- the flags worth a look: cost_above_last_paid, cheaper_supplier,
in_stock, price_list_cheaper — each with a plain-English sentence
Pure and client-safe: no PocketBase, no server imports. Money is
formatted the way the rest of the app formats it (R 0,45).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
An order line the AI filled from its quote now gets its Item Intel
immediately, so the blue price-check marker on the order Process page
has something to show.
The anchor is the SKU the line ended up with. The quote assistant
deliberately keeps the POS SKU off the line's own `sku` column — it
lives on the quote line's `ai_fill` — so that is the fallback when the
order line has no SKU of its own. `quote_items.ai_fill` is only
projected once the schema shows the column exists, because PocketBase
rejects a projection naming a column the collection does not have.
The step is additive and never throws: a failure is logged and the line
still counts as filled.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A line the AI filled now gets its Item Intel immediately, so the blue
price-check marker on the Process page has something to show instead of
"No intel available" until someone opens the ⓘ popover.
The intel is anchored to the POS item the AI actually picked
(ai_fill.stock_item_id / ai_fill.stock_item_sku), so it points at that
exact catalogue row rather than a description-fuzzy guess.
The step is strictly additive: it runs inside the same audit actor as
the fill, never throws (a failure is logged and the fill still counts
as filled), and it is skipped once the run's TICK_TIME_BUDGET_MS is
spent — the remaining lines keep their fills and get their intel on the
next ⓘ / Refresh.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
populateItemIntel now accepts an optional anchor ({ stockItemId?, sku? }).
When it resolves, the anchored stock_items row is forced to the top of
stock_list and its purchase history always feeds supplier_history, so a
line filled by the AI assistant points at the catalogue row it actually
came from instead of whatever the description-fuzzy ranking happened to
surface.
match_kind records how it resolved: "stock_item" (found by PB id),
"sku" (found by SKU — exact, then canonical form via the SKU's
alphabetic stem) or "fuzzy" when nothing anchored or the anchor no
longer resolves. The field is optional so blobs written before this
change stay readable.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Purple when the order is linked to the quote the AI chose and every
line was paired "sure"; yellow when the AI matched a quote but not
every line for sure, or it is only a suggestion. New tab only.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The rule-based matcher no longer writes or shows suggestions: on
2026-10-24 it had left 2,888 on unlinked orders, one quote suggested
to 309 of them.
- createOrder runs the AI order assistant (matchOrder) straight after
intake when it is switched on, instead of the rule-based suggestion
- order page: Re-match / Try again and the rule-based suggestion card
are gone; Match with AI is the action, and "Match by reference" saves
the reference then runs the AI (the run route takes { reference })
- the unauthenticated /orders/[id]/quote-match route is removed
- the sweep route and MCP orders.quote_match_sweep run the AI batch
(matchUnlinkedOrders, shared with the tick)
- scripts/2026-10-24-clear-rule-quote-suggestions.ts clears the old
suggestions (dry run unless --apply); replaces the rematch script,
which would have written rule-based ones again
- also fixes the setState-in-effect lint error on the order page
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A line that sells below the quote's cost produced a negative markup.
Every other write to an order line's markup stores
Math.max(0, computedMarkup), so the fill now does the same. The margin
chip is unaffected — it works from cost and selling, not from this
field, so a loss-making line still shows its real margin.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A line the AI filled now shows up purple until a person checks it, so
nothing the assistant priced is ever taken on trust.
- order-assistant-review.ts: the four client helpers — a line is
pending only while it is filled, unreviewed and still priced from
the quote; a human save keeps "ai_quote" / "quote" / "catalogue" and
turns anything else into "manual"; the review patch records who
checked the line and whether the prices were accepted unchanged.
- order-assistant-save.ts: the save-time merge, the order-side twin of
quote-assistant-save.ts. The value the operator committed wins over
the one the page was loaded with, so the one-click "use this cost"
copy really does end the AI's provenance.
- The Process page applies the purple row class, shows the bot marker
with the quote it was filled from, the reason and the doubts, and
routes every line save through the merge.
- The Orders board shows an "AI x/y" badge: how many lines the AI
filled, and how many still need a review. One query per page, and
never fatal — the board has to render whether or not the assistant's
columns exist yet.
- AiFillMarker grows an "order" variant instead of assuming a POS item.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The assistant now prices the order's lines from the quote they are
paired with, as purple drafts a person then reviews.
- order-assistant-core.ts: isOrderLineWritable + buildOrderLineFillPatch.
Only a pending line with no cost of its own, no person-set pricing
source and no reviewed fill may be written. Supplier / SKU /
matched_price_list_id only when the line has none; selling_price only
when the line has none; markup + markup_percent written together.
pricing_source becomes "ai_quote". A customer price that differs
from the quoted price by more than 5% becomes a doubt.
- order-assistant-lines.ts (new): the shared, cycle-free helpers —
idFilter, aiQuotePairs, writeAiPairs, clearAiPairs and
fillOrderFromQuote. It lives in its own module so orders-service.ts
can use it without an import cycle, since order-assistant-service.ts
already imports setOrderQuote.
- The fill re-reads every line right before writing it, so a cost a
person typed in the meantime wins. It stays quiet until the migration
has added order_items.ai_fill — PB rejects a projection that names a
column the collection does not have.
- setOrderQuote writes the AI's line pairs before pairOrderLines runs,
so the AI's pairing wins over word matching, then fills the lines.
A failure there is logged and never fails the link.
- The tick gains a second pass that fills the lines of orders already
linked to a quote. Only orders the fill actually wrote to count
towards the budget, so a skipped order cannot block the run.
- The run route fills after the match and reports the counts.
PocketBase filters have no IN operator, so every id list is an
OR-chain. fetchQuoteItemsForQuotesAdmin no longer tries an IN filter
first — that was a guaranteed 400 on every call.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- id lists are OR-chains: PB filters have no IN operator, so the
shortlist and supplier lookups would 400 on every real run
- AI line pairs are written before the auto-link (and undone if it is
refused), so setOrderQuote's word pairing can't replace them
- the tick looks past orders it skips; otherwise the same oldest
already-handled orders filled every run and newer ones were never reached
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Task 8 of the AI order assistant plan.
- new AiQuoteMatchPanel: "Linked by AI" / "AI suggests Q…", the
reason, the doubts and the line-pair table (order line -> quote line
with same_item)
- the quote match card gains a "Match with AI" button in every state,
posts to /api/v1/orders/{id}/assistant/run with the PB token, toasts
the 409 message verbatim, then reloads the order
- new card state: the AI suggested a quote -> the AI panel + Accept
(links the AI's quote) / Match with AI / Dismiss
- the AI suggestion marks the order as quote_matching_attempted so the
deterministic sweep cannot overwrite it
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Task 6 of the AI order assistant plan.
- matchOrder(orderId): load order + lines, isOrderEligible, shortlist
from computeQuoteSuggestion (best + alternatives, first 5), empty
shortlist writes ai_no_match with no AI call, otherwise ask the AI in
batches of 30 lines against the same shortlist
- auto_link only through setOrderQuote inside
runWithAuditActor({ name: "AI order assistant" }); a 409 falls back to
a suggestion, never an error
- suggest writes potential_quote + potential_quote_confidence
(0.9 / 0.7 / 0.5) + quote_match_method "ai_suggest"
- always writes orders.ai_quote_match (lines_hash, model, run_at); an
AI failure is recorded with status "error" and never thrown out
- runOrderAssistantTick: same due/lease/stamp logic as the quote
assistant, oldest eligible order first, max 10 per tick
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Task 5 of the AI order assistant plan.
- DEFAULT_ORDER_ASSISTANT_PROMPT: a quote match needs matching items,
not just the same customer; quantities may differ; prices may differ
slightly; a line may have no quote line; answer only with given ids
- SUBMIT_ORDER_QUOTE_MATCH_TOOL (forced tool) + MatchInput
- buildMatchUserText — ids and line facts only, never the email body
- parseOrderQuoteMatch — unknown ids dropped/nulled, a quote line
claimed twice gives the second order line null, skipped order lines
come back unpaired, bad confidence -> low
- matchOrderWithAi — throws on a cut-off response or a missing tool call
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Task 4 of the AI order assistant plan.
Pure core, no I/O:
- PLACEHOLDER_LINE_PATTERNS, isRealOrderLine
- orderLinesHash (sha1 over sorted id|description|quantity)
- isOrderEligible (no quote, >=1 real pending line, not already matched
on the same lines unless forced, not in the AI-error cooldown)
- decideOutcome — the auto-link rule: high confidence, every real line
paired "sure" on the chosen quote, same company, auto-link on
- confidenceToScore (0.9 / 0.7 / 0.5)
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
suggestQuoteForOrder is split into computeQuoteSuggestion (all the
logic, no writes) and suggestQuoteForOrder (compute + the existing
writeSuggestion). Behaviour is otherwise unchanged: the same order
ends up with the same suggestion, written the same way.
computeQuoteSuggestion also drops candidate quotes that another order
already hard-links. A quote belongs to one order, and setOrderQuote
refuses a second link with a 409, so offering one invites a dead end.
The check is one query on `orders` (fields id,quote), chunked at 50
ids per OR-chain to stay under PB's filter-length limit, and it
applies to the exact-match citation too — an exact hit linked
elsewhere falls through to the fuzzy pass instead of being suggested.
scripts/2026-10-24-rematch-quote-suggestions.ts re-runs the matcher
over every order with a soft suggestion and no hard link, which is
where the stale 16x/11x repeats live. Dry run unless --apply; the dry
run only reads.
The script has NOT been run — the owner runs the dry run against live
PocketBase.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
scoreOrderVsQuote awarded up to 0.5 confidence from the header alone
(same company +3, same customer +2), which clears the 0.4 threshold
with zero paired lines. One quote was suggested to 16 unrelated
orders that way.
A candidate now needs at least one paired order/quote line; with no
pairs the confidence is 0. The header points and raw score stay in
the breakdown so the UI can still show why a quote looked close.
No existing test relied on the header-only confidence, so no fixture
needed changing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Collection Confirm cuts a short line to what was collected (that part
invoices and closes) and creates the rest as a new pending line on the
same order, with Back order (same PO / straight to pool) on the process
page. Invoicing subtracts qty_invoiced so a part-invoiced line can't bill
twice. Supplier-side lines close at confirm when their order is billed,
and stragglers get a close button on Invoicing.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
POST /api/v1/driver/invoice-match returns exact matches (same product,
invoice qty = ordered qty) and lines to review. Read-only for the lines.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Confirm stock now sends the collected lines to stock booking and moves
lines not yet collected to a follow-on order (<PO>-2, -3, ...). The
invoice matcher follows a closed PO to its open follow-on.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
POST /api/v1/driver/collection-check runs the collection AI check for a PO
and returns a per-line verdict (right item, right quantity) for drivers.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Stock rows on Collection Confirm were built without cost_price, so the
Cost column was blank (card total R0) and the AI price tick red ("no
cost") on every stock line, even when the line had a cost (e.g. all ten
lines of PO26100702STK after the B&R invoice was checked).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The first iPhone driver photos (Thabo, 2026-10-20, PO202610003TEL and
PO26100702STK) arrived as HEIC. The AI only takes JPEG/PNG/WebP/GIF, so
the reads were saved as "skipped — Not an image (.heic)" and Check with
AI kept answering 0 files read, 0 lines checked; skipped reads were never
retried.
- heic-to-jpeg.ts: HEIC/HEIF → JPEG with heic-convert (pure JS/WASM;
sharp's prebuilt binaries can't decode HEVC HEIC), downscaled to 1600px
like the PB thumb. defaultFetchFile uses it (PB has no HEIC thumbs).
- HEIC/HEIF count as readable images; a read an older version "skipped"
for a now-readable type is re-read, and the tick treats it as unread.
- next.config: heic-convert / heic-decode / libheif-js are server
externals (libheif's WASM loader uses a dynamic require).
Verified on the live photo: 1 MB HEIC → 315 KB JPEG in ~1 s, readable.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
The driver app records the full quantity as received when it collects a
stock pickup, and Collection Confirm only listed stock lines that were
not fully received — so collected stock orders skipped it and sat at
"sent", outside the accountant's booking queue, until someone clicked
Mark received (29 stuck on 2026-10-19).
- STOCK_ON_PAGE_FILTER / stockLineOnPage: a collected stock line stays on
Collection Confirm until its stock order is received; cancelled orders
never show. Shared by the list and the AI check (so those lines get
photos read and ticks too).
- Confirm stock: markStockOrdersReceived flips each stock order to
"received" once every line is collected/booked (→ booking queue);
orders with lines still out stay open, reported as stillOpen. The
route returns { updatedIds, received, stillOpen }; the toast says so.
- Stock photos: `npm run pb:migrate:stock-photos` makes
stock_purchases.additional_attachments multi-file;
addStockPurchaseAttachment, Move photo and Delete photo add/remove with
`+`/`-` instead of replacing the single slot.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PO numbers were "read the month's highest, add one", so POs for several
suppliers on one order created together got the same number — 34 numbers
by 2026-10-19 (e.g. PO202610008LIF for BarloWorld CAT and John Ratcliffe,
created 71 ms apart).
- createPurchaseOrderWithNumber (po-number.ts): numbers and creates POs
one at a time in the server, retries (up to 5) when the unique index
rejects a clash; ignores any po_number passed in.
- buyer-tasks / MCP generate_po use it.
- The Process page's Send PO and in-stock PO buttons now call
POST /api/v1/orders/[id]/purchase-orders (browser session) instead of
numbering in the browser; the client generatePONumber is removed from
pb-helpers.
- npm run pb:migrate:po-number-unique adds a partial unique index on
purchase_orders.po_number for POs created from then on; the 34 existing
duplicates are left as they are.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A PocketBase file PATCH under the bare field name replaces the whole
list. On 2026-10-19 a buyer's "Add supplier invoice" on PO202610008LIF
deleted the driver's delivery-note photo (uploaded 14:38, read by the AI
14:40, replaced 14:41). Generating or regenerating a delivery-note PDF
did the same to every line on the DN.
- New src/lib/server/order-item-files.ts: addAttachmentForm() adds with
`order_item_attachements+` and removes (`-`) only older stored copies
of the same document; pickStoredCopy() finds a file by its uploaded
name (PB stores "DN26101913.pdf" as "dn26101913_<10>.pdf").
- Used by the Collection Confirm invoice upload, DN create, DN regenerate
(replaces only its own previous PDF) and /api/invoicing/delivery-note.
- /api/invoicing/delivery-note-pdf finds the DN PDF by name instead of
taking the line's first file (falls back to the old behaviour for
legacy names).
- AGENTS.md: always add, never replace.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Delete photo next to "Add supplier invoice" removes the photo open in
the previewer (wrong upload, or one of several for the same PO) after an
inline "Delete …? Yes, delete" confirm; the confirm resets when you
switch photos. Not offered for the PO PDF.
- POST /api/v1/collection-confirm/delete-photo (browser session) →
deleteCollectionPhoto: takes the file off every order line carrying it
(or the stock purchase's photo slot) and deletes the AI's reads of it
plus the checks built on them (shared forgetReads with Move photo).
- Removed the header's blue "Run agent scan" button and its results panel
(not in use); the /api/v1/agent/scan route stays for the agent.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Drivers photograph invoices per item and can't see which order a photo
lands on. B&R invoice 00282979 (stock order PO26100702STK) ended up on
the thinners line of PO202610017BLY; the AI read "PO26100702STK" off it
but the card only said "not found".
- The card list now carries each read's po_refs and source file, and sets
ai_doc.other_po when the document names a PO that exists and isn't the
row's PO (none named, or this card's PO among them: no hint).
- PO card: "Photo may be for <PO>" in the header; when open, "Invoice <no>
is for <PO>, not this PO" with a Move photo button. Nothing moves on its
own.
- POST /api/v1/collection-confirm/move-photo (browser session) →
movePhotoToPo: copies the photo to the target (a collected line of a
customer PO, or the stock purchase's single photo slot — refused when
taken), removes it from the source, deletes the AI's read of it and the
checks built on it so it is re-read where it now belongs.
- normalizePoReference moves to the client-safe src/lib/po-reference.ts
(re-exported from the matching service).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Plumblink hands out delivery notes with prices on them. The Price tick
was always grey on a delivery note, so a card could go purple without
its prices ever being checked, and the AI's reading notes ("delivery
note, not a tax invoice", "year not printed") were never shown.
- Price tick: compare whenever the document line has a price, invoice
or delivery note; grey only when no price is printed. The AI still
only corrects a cost from a tax invoice.
- ai_doc now carries the read's notes; cardNotes() unions them per card
and the AI findings box lists them under "Reading notes" (the box now
shows when there are notes but no findings).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The collapsed PO card showed white instead of purple: it always carried
`bg-card`, which beat `bg-ai/10` in the CSS (but lost to the green).
Now the card gets exactly one background.
- Purple now means "the AI checked every line and no tick is red or
amber" (grey "couldn't check" ticks are fine). A card the AI checked
with a line off, or with unchecked lines, goes amber and says
"AI checked N of M lines · K need a look". Green still means "photo,
not checked yet".
- cardAiState() in collection-ai-ticks.ts works this out from the same
computeLineTicks the table uses.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
HYCO (I009753) prints full prices on the lines and "Discount 25.00%" in
the totals; Swagelok (622831) has a "Disc %" column on each line. The AI
compared our cost (already the discounted price) with the printed price,
so the Price tick went red on both.
- The read now captures a line's discount_percent and a whole-document
doc_discount_percent (prompt rule 5b) and stores net_unit_price on each
line (both discounts stacked).
- The Price tick compares the cost with the price after discount; the
tooltip and the document-lines panel spell it out
("R15.10 less 25% invoice discount = R11.33").
- Owner's call: when a "sure" tax-invoice line has a discount and our
cost isn't the discounted price, the AI sets the cost to it (customer
lines through updateOrderItemCost, which adds a cost note; stock lines
through updateStockPurchaseItem) and records why in ai_check.cost_changed,
shown in purple under the line. No discount, a delivery note or a
"likely" match: the cost is never touched.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- The AI and POS columns sat at the far right of the items table, hidden
behind the document preview in the split view. Ticks (now labelled SKU /
Qty / Desc / Price) and the POS status render under each description,
and the table is back to its original six columns and width.
- Owner's call: fill an empty SKU on a sure pairing even if the code may
be misread; the rep edits it and a near-miss still helps.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Dates: the model also returns the date as printed; code parses SA
day-first formats (16-10-26 -> 2026-10-16) and drops dates more than a
year old or a month ahead (the Bestobell read gave 2016-10-26).
- SKU: a new code_certain flag per line; the AI only fills an empty SKU
when it read every character clearly (it filled 0515315 for 051315).
- Supplier: names sharing a distinctive word (Bestobell, Belting) no
longer raise a 'supplier differs' finding.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
A sideways Bestobell invoice was read as doc 10045064 / a bank-details
line; turned 270 deg the same photo reads correctly (10245084,
PO202610010MRE, rubber insertion @ R365). When no printed PO number
resolves, retry the photo at 270 and 90 deg and keep the first read whose
PO resolves (within the tick's time budget).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
M2: the confirm dialog pre-fill uses the card's newest document (as the
chip) and only when it is a tax invoice.
M3: the AI ticks column (and its min-width) shows only when a row has
ai_check; the POS column stays whenever in_pos is known.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
F1: outstanding = round(ordered - collected, 3) > 0 and the order-line
mismatch compares round(order_qty, 3) with round(quantity, 3), so float
noise never produces "0 are still outstanding" or a false mismatch.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
I7: a hash whose read failed this run is recorded as a failed copy with no
second AI call.
M4: the card's own PO lines are always offered for pairing alongside the
POs the document names; the fallback finding still applies when none resolve.
M5: a failed findings write no longer leaves the lines unchecked.
M6: the schema guard requires ai_check on order_items and stock_purchase_items.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
I5: a customer row without a SKU returns needs_sku and creates nothing; the
Add to POS button is disabled for it ("Set a SKU first").
I6: a createStockItem failure is caught per row (status failed + error), the
route still 200s, and the page toasts "N added, M failed, K need a SKU".
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
I4: ConfirmRow.line_sku carries stock_purchase_items.sku; the SKU tick,
the AI robot and (for stock rows) the SKU column use line_sku before the
catalogue sku via effectiveSku. The SKU pencil keeps line_sku in step.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
I3: ai_check.pair_failed marks lines whose covering document couldn't be
paired; computeLineTicks returns all grey for them. The card findings box
now also lists each row's ai_check.findings after the document findings.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
I1: a 'different'/unrated answer never claims a PO line in parsePairs, is
stored as doc_line null (verdict + reason kept), loses to a real pairing,
is not a pairing for the findings, and gives all-red ticks.
I2: document lines are always numbered 1..n.
M1: the SKU tick goes green only through the normalised code match.
Also adds effectiveSku (line_sku before sku) for I4.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
PocketBase drops unknown fields on update, so before the migration runs
the AI's prices would land with no purple marker and no notes. The
manual button / MCP now return 409 'not set up yet' instead.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- M1: fillQuote and the tick refuse templates ("quote is not open
(template)"); both quote pages gate the fill button with
canAiFillQuote (new/follow_up/claimed, not a template).
- M4: run summary keeps <= 20 errors of <= 300 chars (boundRunErrors).
- M5: forced runs offer never-tried lines before tried ones.
- M6: counts route 400s on ids that aren't PB record ids.
- M7: a fill never replaces a supplier already on the line.
- M8: warn dot on the AI marker when a pending fill has doubts.
- M9: MCP hint states the real eligibility rule.
- M10: detail-page price dialog re-reads the line before recording an
AI review.
- M11: exported QUOTE_NOT_OPEN_ERROR_PREFIX / AI_DISABLED_ERROR used by
the run route.
- M14: prompt textarea maxLength 50000.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- With a SKU on the line, shortlistStockItems runs a second narrow query
(sku = raw || sku ~ normalised) and merges by id, so the capped broad
pull can't crowd out the exact SKU row. Quotes and backslashes are
stripped from the raw SKU before any filter is built.
- fillQuote catches a shortlist failure per line and marks only that line
error instead of abandoning the quote.
- Supplier names load once per fillQuote and are passed to each
shortlist call (new optional parameter; default unchanged).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
After "Let AI fill this quote" the page re-pulls the quote's lines and
merges them with mergeItemsAfterAiRun: rows unchanged since the last load
take the server copy (so fills show), rows with unsaved changes stay as
typed, with a toast saying how many were kept. Previously onDone ran
fetchData, which replaced every row.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- decideWithAi throws on stop_reason max_tokens instead of returning a
partial answer; fillQuote marks a requested line the AI did not answer
as error (retried after 1 h), never a permanent no_match.
- QUOTE_MAX_LINES 30 -> 15; the tick also caps lines per quote so the
60-line budget reaches other quotes.
- runAssistantTick stops starting quotes after a 90 s time budget and
returns reason running while a newer start than the last summary is
under 10 minutes old (lease).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
AuditActor gains ignoreRequestUser: buildAuditHeaders then omits the
forwarded PB token (X-Audit-Auth) and request-derived email hint, so the
PB audit hook falls back to X-Audit-Actor-Name. fillQuote sets it for
all its writes.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
- Orders now match on customer invoice numbers and delivery-note
numbers carried on their lines (tagged "invoice" / "delivery note").
- New Delivery Notes source: DN number, customer PO, customer and
company names; preview pulls lines and invoices for visible hits only.
- Quick-overview dialog moved to search-preview.tsx and redesigned:
type-tinted header, status pills, stat tiles, invoice chips with
QuickBooks links, line list, icon timeline with dates and gaps.
Fixes the dialog being pinned at 24rem by the primitive's sm:max-w-sm.
- Result list: per-type colour (icon tile, accent border, badge, row
hover), status pills per row, and a jump bar; mobile-tightened.
- Stock-order preview now fetches ordered_value/created (Value row and
age chip were always empty); supplier names keep their original case.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Third and last place that read `quotes.company` directly. The list page
showed "Mre" next to Pierre Visser, who works for Harmony — the same
drift already reported on the detail page.
This one was a deliberate perf shortcut rather than an oversight.
`QUOTE_LIST_EXPAND` dropped `customer.company` to "save one IN query per
page load" and the page read the quote's own `company` relation
instead. That relation is a denormalised copy that PocketBase never
re-cascades, so the shortcut read a field nothing keeps correct — and
the column it fed is what a rep reads to decide whether to approve, so
a stale company there is worse than a slightly larger query.
The 3-level expand is back, on both the list and the search path (the
search results feed the same columns, so they needed it too). Search
now also RANKS on the customer's company rather than the stale copy —
otherwise a query for "Mre" would surface quotes whose results then
render "Harmony", contradicting themselves on screen. It falls back to
the quote's own copy for the ~4% of customers with no company relation,
where the two are equal by construction.
Applies to both the desktop table column and the mobile card.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Aligns the tool policy, the system prompt, the MCP route and the skill
doc so they describe the same boundary, with tests pinning the policy
and prompt text.
The docs and the executable policy had drifted apart — the skill doc
advertised capabilities the policy does not grant, which is worse than
not documenting them: an agent reading the doc will plan around a tool
it isn't allowed to call.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Gives the operator two new controls on a collection-pool stop, and fixes
the audit feed they write into.
Back-to-pending (`items/[id]/back-to-pending`)
A stop has to come back when the driver can't take it — no-show,
wrong date, supplier pushed the delivery — or when the supplier
confirmation was premature. A row can be assigned to a driver and/or
marked ready at the same time, so the route clears both in one write
rather than making the operator unwind them in the wrong order.
Collection date (`items/[id]/collection-date`)
`order_items.collection_date` was a PB date field that existed but
nothing read or wrote it. On a PENDING stop it is a forecast — the
supplier said goods land on the 20th, so the stop stays
un-assignable because promising a driver a pickup for goods that
don't exist yet would be a lie. On a READY stop it is the confirmed
collection day.
Notes feed (`src/lib/note-feed.ts`)
`notes` on order_items / stock_purchase_items is NOT free text — it
is a JSON array of NoteEntry objects, and every consumer parses it as
one. The assign route's "mark ready" path wrote a plain string into
it, which silently destroys the entire history: the next parse falls
back to an empty feed. All writers now go through `appendNote` with
the row's current value. Lives in `src/lib` rather than a page view
model because both the client pages and the server routes need it,
and an API route must never import from `app/(app)`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Two quote-flow defects found while auditing Q202610315.
1. Delivery-note quotes were being dumped into the rep approval queue.
A DN quote is never sent on its own — it goes out with the note, via
the note's "Send the delivery note and quote to the customer" button.
But `generateQuotePdf` auto-flipped every non-terminal quote to
`approval_pending`, which overwrote the DN status before the
skip-approval branch in `sendQuoteEmail` could ever run. Ten DN
quotes were sitting in Pending Approval as a result.
- `createQuoteFromDeliveryNote` now creates the quote `approved`
rather than `delivery_note`.
- The PDF render no longer flips DN quotes at all, as a backstop.
- The quote page gets a non-sending action row pointing back at the
note. No "send" button — a second way to email a customer the same
invoice is a liability, not a feature.
Also fixes a latent duplicate-invoice bug: the duplicate guard
pinned `status = "delivery_note"`, so once the quote had been sent
it fell through and a second quote could be raised for the same
note. Verified by test — the old guard silently minted a duplicate.
2. `quotes.company` never followed a customer change.
`quotes.company` is a denormalised copy of `customer.company` and
PB does not re-cascade relations, so swapping the customer left the
old company behind forever (Q202610315 read "Mre" while its
customer sat at Harmony — two contradictory companies on one row).
The customer swap ran as a raw client-side PB PATCH that wrote only
`customer`, which also bypassed the server's placeholder guard. It
now goes through `PATCH /api/v1/quotes/[id]/pdf`, which cascades
both relations atomically.
The customer-facing PDF already read `customer.expand.company`,
which is why this never reached a customer — but every consumer of
the quote's own `company` got the wrong answer. A read-only audit
(`scripts/audit-quote-company-drift.mjs`) puts it at 2,591 of 9,950
quotes. That data is NOT backfilled here.
Also drops the duplicate standalone "COMPANY" column on the quote
detail page, and the stale-company fallback in approvals, so this
class of bug can't hide in the UI again.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Operator request. Bigger effect than expected.
THREE constants, not one
"90" appeared in three places, and two of them are load-bearing:
universal-search-service.ts UNIVERSAL_SEARCH_WINDOW_DAYS orders scope
quotes-search-service.ts SEARCH_RECENT_WINDOW_DAYS quotes pushdown
quotes-service.ts SEARCH_RECENT_WINDOW_DAYS exported; shared
with the /quotes page + archive
The last two must stay equal — the /quotes page renders "Show recent
(last N days)" from the exported constant while the server filters on
the private one, so moving them apart desyncs the label from the
behaviour with no error anywhere. All three moved together. Universal
search calls searchQuotes without `includeOlder`, so it inherits the
quotes window; changing only the universal constant would have left
global search reading 60 days of orders and 90 days of quotes behind a
single "last 60 days" label.
The UI label reads windowDays straight off the service, so it follows.
MEASURED
Order candidate set collapses — the window was carrying far more than
the linear slice it looks like:
90d -> 5,394 orders / 16,640 items / 6,696ms scope scan
60d -> 972 orders / 3,307 items / 1,606ms scope scan
Cold search "tiaan": 17-22s -> 4.6s measured. That is most of the
mitigation the TTL bump was papering over, so the long-deferred pushdown
rewrite is now worth less than it was this morning.
Quotes genuinely drop, and it is real rather than a regression: Tiaan's
own quotes go 22 -> 6 between a 90- and 60-day window, because that
band is where most of his quotes sit. Verified against PB directly.
ALSO
Pinned the clock in the age-chip test. It asserted "4d" against a
fixture dated four days before the wall clock, so it expired the day
the date rolled past — it failed on this run for exactly that reason.
A delivery note's cost price is what the business paid. Editing it
while setting a margin rewrites that silently, so a stray keystroke
changes every downstream number on the note without ever looking like
a pricing decision.
The rule lives in quote-pricing.ts as a `costLocked` flag captured at
load time, and repriceLine DROPS a cost patch on a locked line. That
is the guarantee; the read-only cell in the UI is only the
convenience, so no caller can quietly rewrite the cost.
- "Edit note" and "Create quote" grids: cost is read-only with a lock
icon once the line arrived costed; a blank cost stays fillable and
still seeds the 25% default markup on first entry.
- readOnly, not disabled, so the number can still be selected and
checked against a supplier invoice.
- Locked at load, not per keystroke: a cost typed in this session
stays correctable until save (an operator choice), and locks on the
next open because the row then carries the cost.
- The DN detail page's edit mode is untouched — it stays the
deliberate "that cost is simply wrong" correction path.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
ORDERING — recognition first, then recency
Not recency-first, deliberately. Sorting purely by date buries an exact
order number under every newer order sharing a weaker match, which is
the opposite of what you typed the number for. `score` already encodes
the tier (identifier 100 > name 60 > body 20, with a <=25 exact/prefix
bonus), so comparing it pins precise matches above prose matches for
free; the date then breaks what's left.
That "what's left" is the case that matters day to day: searching a
company name matches every order they have, all with an identical score,
and newest-first is the only useful order for those. Verified live on
"teleg" — 28 orders, identifier-prefix matches first, everything else
descending by date; and "121026-1 ENG" still returns exactly one hit.
Quotes already sort newest-first inside quotes-search-service, so they
inherit this without change.
AGE CHIP
Every row now carries a compact age chip — today / 3d / 2w / 1mo / 2y —
in three colour bands: emerald ≤7d, sky ≤30d, slate beyond. The numeric
label is always rendered, so colour is a scanning aid and never the only
carrier of meaning; the exact date is the chip's tooltip.
No created date from a source means the chip is omitted rather than
rendered blank, matching how the timeline drops undated steps.
ageLabel/ageBand/ageClass moved to ./age because Next.js App Router
pages may only export a fixed set of names — an extra named export from
page.tsx fails the generated route-type check.
That split also surfaced a formatting bug worth recording: the original
bands capped weeks at 60 days, so 30–59 days rendered as "4w"–"8w" and
60+ as "2mo" — "1mo" was unreachable. Month band now starts at 30 days.
Adds `createdAt` to all eight hit sources and `created` to the field
lists that lacked it.
Two reports, one root cause each.
1. "teleg" didn't find order 121026-1 ENG
Same defect class as the supplier gap in efb5a53. `partyNames` was
resolved and used to BUILD the preview, but never added to the MATCH
set — so an order was findable by its number and displayed with its
company name, yet not findable BY that name. `orders.customer` and
`orders.company` are PB relation ids; without the resolved names in
the match set, partial company searches returned nothing.
Now matched on both, with the matched surface reported accurately
(customer / company / contact / rep / order) rather than a blanket
"order". Verified live: "teleg" → ENG found, via=company.
2. Preview had no lifecycle trail
Built from timestamps already on the order and its items, so it costs
no extra round-trip — an audit-log query per hit would undo the
point of the preview being free. Verified live on a progressed order:
2026-08-04 Order raised
2026-08-12 Driver assigned
2026-08-12 Collected
2026-10-14 Invoiced
Steps with no timestamp are OMITTED, not rendered blank — a rail with
holes reads as broken data rather than unfinished work. That is why
121026-1 ENG shows a single step: it is only at po_generated, so
nothing after "Order raised" has happened yet.
Column population on live PB, for anyone extending this:
assigned_at 1371 · drv_collected_at 1358 · invoiced_date 1626
delivered_at 2 · drv_assigned_at 0 · collected_date 0
`delivered_at` being effectively empty is the same defect fixed in
918f73e; here it means the Delivered step will rarely appear.
HITS_PER_GROUP 25 → 50
Searching a company name returns every order that company has, and they
all score identically (same matched surface), so the ENG order fell off
a 25-row list that was one row short of showing it. 50 covers a single
customer's order volume. Truncation is still surfaced as "showing N of
TOTAL", so this stays a UX choice rather than a hidden guarantee.
Dialog restyled: type as an eyebrow, a "matched: <surface>" pill so the
reason a hit is present is stated plainly, a bordered label/value table
instead of a loose grid, and the timeline as a vertical rail.
Clicking a result now opens a summary dialog instead of navigating away,
for all eight result types.
Deliberately NOT an iframe of the detail page and NOT eight bespoke
preview components. Both were on the table:
- Iframe: cheap, but `src/app/(app)/layout.tsx:37` renders the
<Sidebar /> for every page in the group, so the modal would show a
sidebar inside a sidebar.
- Per-type components: 8 fetch + 8 field maps, drifting out of date
as the pages they mirror change.
Instead the SERVICE returns a flat `preview: [{label, value}]` list on
each hit, built from data it already holds in the snapshot and reference
caches. One generic renderer then serves every type, opening a preview
costs no extra round-trip, and a ninth source needs no new UI.
Verified live against PB — e.g. searching "pro-tech" now returns, per
hit: order number, customer, company, rep, status mix, item count,
created date, suppliers, first line descriptions.
Two gaps found while verifying the payload rather than assuming it was
populated:
- `item_status` was not in the item fetch, so the order's Status row
was silently absent — the single most useful field in the preview.
Added.
- `customers.company` is a relation id; the preview would have printed
a raw 15-char PB id. Now resolved to the company name.
Rows became <button>, not <a>/next/link: there is no navigation to
perform, and a link would imply a middle-click affordance that doesn't
apply. "Open full page" in the dialog is the escape hatch — a preview
must never be the only way to reach a record.
Tests: the two new-tab tests are replaced by three dialog tests (opens
instead of navigating, escape hatch present with correct attrs,
dismissing leaves query and results untouched).
Asked what else search excludes, and measured it rather than reading the
code. The terminal-state window filtered on:
(item_status = "delivered" && delivered_at >= "<90d cutoff>")
Live PB, 2026-10-14: 12,957 delivered items, and NOT ONE carries a
`delivered_at` value. That clause matches exactly zero rows. Every
delivered order was invisible to search — 4,360 of them. `cancelled`
was unaffected because it filters on `updated`, which is always
populated, which is the entire reason the two states behaved
differently.
Fix: each terminal state gets two alternatives — the semantically
correct timestamp when it exists, and `updated` as the fallback. `updated`
is always populated and is >= the delivery moment for a delivered row,
so it can only ever be slightly INCLUSIVE, which is the safe direction
for a "must come up" guarantee. `removed` (struck-from-order lines) now
joins the terminal group too; an order whose only line is `removed` is
in effect a cancelled order, which is how /orders already groups it.
Coverage gain, measured live:
"tiaan" orders 96 -> 107
"450NB" orders 1 -> 2
COST, and it is real
Widening the window made the candidate set 5x bigger — 1,075 -> 5,361
orders, 3,5xx -> 16,5xx line items — and the snapshot rebuild is O(scope),
so cold went from ~6s to a measured 17-22s. At the old 60s TTL that was
17-22s of PB work every minute per instance, which is not shippable.
Mitigated for now by raising both the scope-id and snapshot TTLs to 5
minutes (5x less exposure) and relying on the page's mount pre-warm, now
that the pre-warm is actually authenticated and stops being a no-op.
Warm searches remain 0ms.
Stated plainly: this is a mitigation, not a fix. The architecture is
still O(scope) when it should be O(matches) — the quotes search already
does it correctly by pushing `~` filters down to PB. That rewrite is the
proper follow-up and should happen before the data grows further.
Accepted trade: a status flip or new item can take up to 5 minutes to
appear in search. Fine for "find the thing I half remember"; not fine
if this ever backs anything needing immediate consistency.
STOCK ORDERS
The Stock Orders source matched only `purchase_number` and `status`,
which made stock POs findable only by our own internal sequence number.
Live probe on 2026-10-14: searching "saw chain", "protech" or any line
description returned ZERO stock-order hits, while the literal
"PO26101404STK" matched. An operator hunting a stock order thinks in
product or supplier, not in our numbering, so the source now also
matches the resolved supplier name and every line's description, sku,
supplier_sku and product_code.
Verified live:
"pro-tech" -> 13 stock orders (supplier name)
"clamp" -> 1 (line description)
"chelsea boot" -> 4 (line description)
"M36 NUT" -> 2 (line description)
Also fixed the matched-surface label: a header match was always tagged
"PO number", so a supplier-name hit claimed to have matched the PO
number. A "why did this match" tag that's wrong teaches the operator the
tags are noise and stops them reading any of them.
Both new collections join the cache warm so the first search doesn't pay
for them.
RESULTS OPEN IN A NEW TAB
Switched from next/link to a plain `<a target="_blank" rel="noopener
noreferrer">`.
next/link does a client-side route change, which UNMOUNTS the search
page — so the query, the expanded result set and the scroll position are
all gone when you come back. A real anchor in a new tab leaves this tab
mounted and untouched, which is the behaviour that was asked for:
click a result, look at it, click back to the search tab, and you are
exactly where you were.
Verified by test: dropping `target` fails the new-tab assertion while
the rest of the suite stays green.
One box at /search (sidebar, under Dashboard) covering orders, quotes,
both purchase-order flavours, suppliers, companies, customers and stock
items. Replaces guessing which screen held the answer.
Window: 90 days (UNIVERSAL_SEARCH_WINDOW_DAYS) on orders and quotes —
the two that grow without bound. Suppliers, companies, customers and
stock items are searched in full because they don't expire and an
operator searching for one expects to find it at any age. The UI states
the window so a miss reads as "aged out", not "broken".
THE INVARIANT — a cap must never look like emptiness
Every group reports `total` (uncapped) separately from `hits` (capped),
plus `truncated`. The UI renders "showing 25 of 1,240 — refine". This is
the property the 500-cap bug in orders-service violated: the newest 110
orders silently vanished and it read as a missing search field. The type
makes `total` structurally independent of `hits.length`.
Symmetrically, a source that fails or times out renders as an explicit
"couldn't search quotes" row. "We couldn't search X" and "no X matched"
are different statements and must never look alike. All nine sources
fan out in parallel and settle independently.
PERFORMANCE
Profiling the first build showed 7.8s per search — ~3.5s of it pure
data-fetching (1,070 orders / 4,288 line items re-read on every
keystroke) against ~10ms of actual matching. Added:
- order snapshot (headers + items + supplier names), 60s TTL
- reference-collection cache, 5min TTL
- per-query result cache, 60s TTL
Cold ~4.5s, warm 250ms-1.6s, repeat query 0ms. The page pre-warms on
mount via an empty-query request, so the ~4.5s cold build happens while
the operator is still reading the box rather than in front of them.
Orders match supplier NAMES, not relation ids — verified live that
"truco" returns the Truco order, the 6 matching quotes AND the supplier.
NO Reps group, deliberately
A rep is a filter, not a page: this app has no user detail view and no
page accepts ?search=. A Reps group would be rows whose every link
404s. `orders.rep` is a matched field on the Orders source, so
searching "tiaan" still returns exactly his 96 orders — which is what
was actually asked for.
PO hits link to their parent order, since there is no standalone PO
detail page; orphaned POs fall back to the PO PDF.
Tests: 15 cases over the pure layer — ranking tiers, best-of-not-sum
scoring, filter escaping against injection, the window maths, group
ordering, and the cap/failure invariants. One of them caught a real
defect: scoreField lowercased the haystack but not the query, so an
uppercase query scored 0 against every row.
Searching for a supplier returned nothing even though the order existed
with that supplier set and "TRUCO" in the line description.
Root cause is not search coverage — it's a stale safety cap. The search
path fetched the active candidate set with a fixed `perPage: 500`, on the
comment's reasoning: "91 active orders out of 4,782, so the cap is
unreachable in practice." Measured on live PB 2026-10-14 the active set
is 610 orders. Because /api/v1/orders/list sorts `created` ASCENDING
(oldest first, so the buyer works the queue top-down), the slice kept
the 500 oldest and silently discarded the 110 newest.
Verified against live PB before the fix:
truco -> 0 matches
skirting -> 0 matches
tiaan -> 57 matches
After:
truco -> 1 match (ORD-26-10-13-SCALPERS, collection_pool)
skirting -> 1 match
tiaan -> 76 matches (19 were being dropped too)
It presented as "some searches work and some don't": only matches on
OLD orders survived, so it looked like a missing supplier field rather
than truncation, and the supplier lookup added in efb5a53 inherited the
bug rather than causing it.
Fix: `perPage: activeIds.length` so the candidate set is never sliced.
The set is already bounded by COMPLETED_WINDOW_DAYS plus the non-terminal
statuses, so it grows with daily throughput, not total history.
Regression test seeds 520 active orders with the needle on the newest,
so an ascending sort puts it past the old cap. Verified it FAILS against
the old code and passes against the new.
The /orders search box only matched order number, company, customer and
line description, and only did price matching when the whole query parsed
as a clean number. Typing "afrisam" returned nothing because
order_items.supplier stores a PB record id and wasn't even fetched.
Scope is unchanged and still correct: the candidate set is the ACTIVE
order set (getActiveOrderIds) — items not completed plus anything
completed in the last 30 days. Supplier search inherits that window for
free, so "what did Afrisam supply for us?" now returns the live work plus
the last 30 days, and nothing older (that stays on /orders/archive).
Now searchable: supplier name, SKU, contact name, rep, RFQ number, order
description/notes, company address + vendor, customer email/buyer email/
phone, item notes, customer invoice number, quantity, cost_price.
Prices are matched as a substring of the plain digit string, so "12.2"
finds 12.22 and "R 1 250" finds 1250.00. Two guards keep the widened set
usable: a 1-2 char numeric query no longer probes prices ("50" is a
substring of 1250, 350 and 50, which made every price a hit), and
zero-priced placeholder lines never match a price probe. Quantity is an
exact compare for the same reason.
Supplier ids resolve to names via a targeted, chunked lookup (PB 400s on
an inline id filter past ~106 ids) rather than an expand, which would add
an IN query to every page load and realtime refetch. The supplier fetch
and the widened item field list (sku/supplier/notes) are both gated
behind an active search, so plain page loads keep the lean payload.
Tests: 9 new cases covering supplier name/partial/case, the 30-day
window, SKU, price shapes, the short-numeric and zero-price guards,
metadata fields, quantity exactness, and the no-search/no-query guard.
Full suite 3902 passing.
Two reported symptoms that turned out to be one bug, plus the dark
mode the user described as "too black".
The `hidden` attribute does nothing on the discount controls
`hidden` is implemented by the UA stylesheet as `[hidden]{display:none}`.
Any author rule that sets `display` outranks it — and `.totals` and
`.discount-body` both set `display:flex`. So `el.hidden = true` was a
no-op on both:
- the discount tier panel could not be collapsed, making "Add
discount" look like a dead control;
- the Discount totals row stayed on screen after the discount was
cleared, still showing the last figure (R 98.37) under a total
that no longer contained it. The JS only ever set `hidden` and
never cleared the text, so the operator had no way to tell stale
text from a live one.
`.cat-menu` and `.cat-backdrop` already carried per-element
`[hidden]{display:none}` overrides for exactly this reason; the two
discount elements were missed. Replaced with one global rule so the
whole class is covered rather than relying on each element to
remember. The clear path now also blanks the amount.
Verified statically rather than in a browser: jsdom does not model
the UA-vs-author cascade and reported the broken and fixed cases
identically, so the cascade was checked by resolving which rules
apply. Regression test fails without the fix.
Dark palette
The POS canvas was #0f0f0f (L* 4.3 — effectively black) with
surfaces only 3-4 L* above it, mixing a warm #1c1416 against a
neutral #181818 in the same UI. That near-black base plus two
competing surface tones is what read as muddy. Replaced with a
single cool-neutral ramp with four real steps, and menus/modals now
sit on a `--raised` layer instead of reusing the surface colour.
`#717171` was the secondary text colour in 25 places and measured
3.93:1 on the old canvas — below WCAG AA. It now measures 6.49:1.
Both the text greys and the surfaces were hardcoded per component
across seven POS pages; they are now tokens, with the dark block
rebinding them.
The shared `.dark` block in themes.css (drives all 24 themes) sat at
oklch(0.15) ≈ L* 3.1. Lifted +0.04 across the ramp, preserving its
shape and separation; muted-foreground goes 6.32:1 → 5.95:1, still
above AA. Cursor/Supabase, which override it darker, were at L* 1.3
and lifted too.
Service worker cache bumped to pos-v36 — pos.html and pos-app.js are
served cache-first, so without the bump the tills keep serving the
broken files.
3863 tests pass (3 new); build clean.
Three changes to the POS sale pipeline, all in one commit because
pos-service.ts / types.ts / pos-app.js are each touched by all three and
splitting them would only produce intermediate broken states.
VAT is now charged on top, not backed out
Catalogue prices are ex-VAT: `selling_price` is derived as
`cost_price x markup` from the supplier's ex-VAT cost, with no VAT
uplift anywhere (verified — median selling/cost across the catalogue is
exactly 1.8). The till nevertheless treated them as VAT-inclusive and
recovered the net with `total / 1.15`, so every sale was recorded with
an understated net and an understated VAT line.
An R100 ex-VAT line now rings up as R115. The rate comes from the till's
letterhead, read server-side rather than accepted off the wire, and is
stored per sale so a later letterhead rate change cannot rewrite what
was charged. All 3,623 catalogue prices are untouched.
Ad-hoc "custom" lines
A cashier can now ring up something that was never a catalogue item —
"bolt cutting" at R20 agreed at the counter — via a new
"+ Add a custom item" composer. The line prints on the slip and lands in
pos_sale_items, but carries no stock_item relation and takes no stock
movement: no ledger row, no current_qty decrement, no catalogue entry.
pos_sale_items.stock_item is no longer required at the schema level and
an explicit `is_custom` flag marks the line, rather than inferring it
from a blank relation — so "how much did we make on bolt cutting this
month" is a plain query, and a future stock-less line type cannot be
confused with a corrupt one.
Custom lines are refused on account sales: those raise a delivery note
the customer signs for and feed their statement, so an ad-hoc charge
must not slip onto one. Enforced server-side, not just in the UI.
Discounts are now 20% or 40%, nothing else
The R-off / %-off pills and the free-typed number box are gone. The
server treats DISCOUNT_PERCENT_OPTIONS as an allowlist and refuses
anything outside it rather than coercing it, so a stale till cannot
invent a discount. The discount lands on the ex-VAT subtotal, making
"20% off" a true 20% off both the net and the VAT on it.
Two defects found while doing this
pos_sales was missing six fields that checkoutSale has been writing all
along (gross_subtotal, gross_vat, gross_total, discount_amount,
discount_kind, discount_input). PocketBase accepts unknown fields
without error and silently drops them, so all 192 historical sales lack
them and a discounted sale reprinted as though no discount had been
applied. Added, plus vat_percent.
History also passed gross_total (VAT-inclusive) where the receipt
derives the gross VAT from an ex-VAT base — a reprint double-counted
VAT. Corrected.
Verification: 3860 tests pass (14 new); typecheck, lint and production
build clean; live PocketBase probe confirms a pos_sale_items row with
stock_item="" and is_custom=true round-trips, with the probe removed
afterwards.
The DN->quote batch copied cost and selling but never wrote markup, so
the approval screen showed 0% beside differing cost/selling. Write it at
creation, and derive it on the approval page for older lines.
Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Operator report: "If I double click on it and I put in 20 there, goes
to 18... if I take everything out, I start typing, automatically change
it to 18. I need to be able to add any value there and when I click
save and the value is less than 18, it should go to 18."
The floor (`2387a71`) was correct in principle and wrong in WHERE it
sat. It was applied inside `repriceLine`, which runs on every
keystroke in both pricing grids — so the moment a sub-floor digit was
typed the state was raised to 18 and the box snapped.
Typing "20" is the clearest demonstration: the "2" is sub-floor, so it
became 18 mid-keystroke; the next character then APPENDED to "18" and
produced 180. What the operator saw as "goes to 18" was the first
keystroke, and "it jumps around" was the second.
She asked for the floor on SAVE, not while typing. So:
- `repriceLine` no longer floors. Markup is also rounded to 2dp on
the way in, so the box holds `80.06` rather than the float noise
a back-solved markup produces (`80.05999755859375`).
- The floor applies on BLUR and on SAVE, in both the delivery-note
editor and the quote review modal, via `applyMarkupFloor` — the
same function, still the only thing between a typed value and the
database, so the -100-row hole stays closed.
- `step="0.5"` replaced with `step="any"`. A back-solved 80.06 is
not a multiple of 0.5, which left those boxes `:invalid` — and a
number input in an invalid state fights the caret. This is very
likely a second contributor to "I can't do anything there".
- The bulk "set all markups" field gets the same treatment.
Verified with real keystrokes in the browser, not synthetic events
(which had masked this — they set a value in one shot, so the
mid-keystroke floor never fired):
type "20" -> stays 20, selling R605,40 (504,50 x 1,20)
type "15" -> stays 15 while typing
blur -> 18, selling R595,31 (504,50 x 1,18)
Three existing tests asserted the per-keystroke floor. They are
repointed, not deleted: the commit path is now pinned directly, and
`applyMarkupFloor` has its own tests so the database protection is
still asserted.
No data changed — the editor work above was dialog-local state only.
Full suite 3833 passing, tsc clean.
A counter note is raised when the customer takes goods off the shelf.
"They said make it five more" is edited on the NOTE. Previously that
edit changed the note and left the quote showing the old number, so
the same physical goods carried two different quantities across two
screens and the invoice was whichever screen the operator opened last.
`updateDeliveryNoteItemPricing` now pushes the note's quantity onto
the quote lines that came from it. A sync failure is logged, not
fatal — the edit already succeeded and the quote can be re-synced by
re-saving.
## Matching without a back-link
`quote_items` has no field pointing at its source `order_item`, so
the operator's call was to make the note authoritative and work around
it. Lines are paired by `sku + description`, and when a key repeats,
by ORDER OF OCCURRENCE.
That is sound because both sides derive from the same ordered list
of note lines: `createQuoteFromDeliveryNote` iterates the note's
items in query order and emits one quote line each, so the Nth of a
repeated key on the note is the Nth on the quote. On the live data
across 500 note lines there are ZERO groups sharing
(note, sku, description) — the key is unique in practice. The
occurrence counter exists so that if two identical lines are ever
added, the second still lands on the right row instead of both
writing to the first.
That is the failure the tests exist for: matching on the key alone
silently writes the wrong number onto a customer invoice. Verified by
removing the occurrence counter — exactly that test fails.
## Also carried through
- The note's COLLECTED qty is the billing qty, not the ordered qty, so
a short pickup (10 ordered, 8 collected) puts 8 on the quote.
- A cancelled, rejected or declined quote keeps its own numbers.
Rewriting a quote the customer has already seen is worse than a
stale number.
- A line that cannot be matched is left alone, never guessed at.
- A note line with neither a collected nor an ordered qty is skipped,
so this can never reintroduce the zero-quantity invoice line.
8 new tests. Full suite 3832 passing, tsc clean, no new lint errors.
`qty_collected` is read by the invoice path to decide how many units
to bill. A counter delivery note has no parent order: the customer is
standing at the counter taking goods off the shelf, so the collection
has ALREADY happened — there is no driver pickup to record, and it can
even have happened before the system existed. Those notes were being
created with `qty_collected` left at 0, so every one of them invoiced
as a zero-quantity line, which QuickBooks rejects (n8n 400 -> the
opaque 502 the operator saw).
Two writes, both scoped to parentless notes:
1. `createAdHocDeliveryNote` stamps `qty_collected = quantity` on each
new item. The goods have been handed over; that IS the collection.
2. Editing a counter note's quantity keeps them in step, so when the
customer says "actually make it five more" the DN editor, the quote
and the invoice all agree. This is the loop the operator described:
edit -> quote -> invoice, with the quantity carried through.
Order-linked notes are deliberately left alone. There
`qty_collected` means "the driver has this many in the van", and
overwriting it with the order qty would bill for goods that have
not arrived. `orders/process` already stamps collected=quantity for
the in-stock flip; this is the same rule applied to the counter
path that had been missed.
Four tests, including that an order-linked note's collected qty is
NOT touched. Verified they bite by removing both guards: 2 fail.
Not included: 315 existing parentless items still have
qty_collected <= 0. A backfill is a production data change, so it
waits for the operator's go-ahead.
Full suite 3824 passing, tsc clean.
Quoteers reported markup changing to a random number after saving or
clicking away, and lines carrying markup = -100 that nobody had
edited. 51 quote_items rows were affected (40 of them exactly -100),
all with a real cost and selling_price = 0. All 51 reproduced exactly
from ((selling - cost) / cost * 100) — none were typed by a human.
Two independent defects, both price-rounding:
1. The approvals screen priced with Math.round(cost * (1 + m/100)) and
no precision argument — round to WHOLE RAND. A 30% markup on a
R0.07 washer became R0.00, and a selling price of 0 re-derives as
exactly -100. Rounding is now a 2dp ceiling: R0.07 @ 30% -> R0.10.
A ceiling can never collapse a positive price to zero, so that
failure mode is structurally unreachable rather than merely patched.
2. Round-to-nearest also undercut cheap lines generally: 30% on R0.07
rounded to R0.09, which re-derives as 28.6% — the "random number".
Rounding up means the price can only drift in the customer's favour.
Markup now floors at MIN_MARKUP_PERCENT (18) once a line has a cost,
applied in one place so every screen inherits it. Scoped to priced
lines: an unpriced draft (no cost) is undecided, not 0% margin, and
stays writable — 1,181 live rows depend on that. Applied at derivation
(not just on save) so -100 can no longer be computed, plus
server-side in review-approve as the authoritative gate.
order_items already carried markup min=0 plus a Math.max(0, …) clamp,
which is why no order row was ever affected. quote_items had neither.
Verified: 3,817 tests pass, typecheck clean, lint delta zero, and all
16 of the real broken rows re-priced through the new logic resolve to a
valid markup and a non-zero selling price (was 16/16 broken).
Existing negative rows are left untouched — re-pricing live quotes is a
separate, deliberate decision.
A customer who brings goods back needs the sale reversed: the items go
back into stock and the money comes back off the till. There was no way
to mark a POS sale as refunded, so nothing stopped the same sale being
refunded twice — which credits the customer twice and puts the same
stock back on the shelf twice.
Refunds are WHOLE-SALE only by operator decision; no partial refunds.
That is what lets `pos_sales.refunded` be a boolean instead of a
per-line refund table. Three columns: `refunded`, `refunded_at`,
`refund_amount`, stamped from the sale's post-discount total, which is
what the customer actually paid.
DOUBLE-REFUND GUARD, twice over, because this is the one failure that
costs real money:
- `refunded` is read before anything else, and a flagged sale returns
409 with no stock writes at all
- the flag is stamped BEFORE any stock write, so a second concurrent
tap sees it and bails. Writing stock first and flagging last would
let both taps through.
A stock write that fails part-way does NOT roll the flag back. The
customer has their money either way, so re-running would double-credit
them; the failed lines come back in `stockErrors` for a manual count
correction, which is strictly better than a second refund.
Stock needs no schema: `return` was already in STOCK_REASON_VALUES,
just never used. Rows are written with POSITIVE change_qty and
reference_type "pos_sale_refund", leaving the original
customer_fulfillment rows untouched so the sale's own audit trail still
shows what was originally rung up.
cashup.html had expected drawer = sum of cash sales, so a cash refund
would have shown as a false variance on every cash-up after it. It now
subtracts cash refunds, and refunded sales contribute nothing to
takings for any tender.
The tender is read off the sale, never off the request — an earlier
version accepted it in the body, which would have let a caller assert
that a cash sale was really an EFT and talk the drawer out of balance.
At the till: tap still reprints, hold a sale row for 600ms to refund —
moving money sits behind a gesture you cannot do by accident. Refunded
sales stay in the list, dimmed and badged, because deleting them would
destroy the audit trail. The sheet names the items and says plainly what
happens next ("Hand back R10.57 in cash. The drawer count for this day
will drop by that amount.").
The credit slip reuses the sale renderer and the whole printer path with
`refund: true`, so there is no second printer to configure — but note
there were THREE render paths to fix (on-screen, Bluetooth text, ePOS
XML), and missing the third would have printed "TOTAL" on a credit note.
Migration applied to live PB and verified idempotent; field shapes
round-tripped on a real sale and restored. 10 new tests, 148 POS tests
pass, tsc clean, build passes. sw.js v34 -> v35 (history, cashup,
receipt and epos-print are all cache-first in the shell).
The account-sale picker filtered on three customer columns — name,
number, email. `company` on a customer row is a RELATION id, not a
name, so it could never be text-matched; the company name was only ever
resolved afterwards by `expand` for display. Typing "Sibanye" returned
nothing even though 52 people work there.
- the filter now traverses the relation: `company.company ~ q`.
Verified live: "Sibanye - Rustenburg" 0 → 52, "Adal" 0 → 2,
"Alano" 0 → 1, while "Botha" and "Smit" are unchanged at 6 and 4.
- each row carries `viaCompany`, and name matches sort above
company-only matches. PB's own sort is by name across the whole
result set, which would interleave the two bands.
- /customers returns { items, total, shown } and the till renders a
count header: "52 matches", or "Showing 100 of 122 — keep typing to
narrow". The old `slice(0, 8)` showed 8 of 52 people with nothing
to suggest 44 were missing; 15 companies have more staff than that.
The picker click handler bound to every `div` in the results box, so
adding a count header would have made it selectable and picked a
customer with an empty id. Now scoped to `.cust-row`.
Two bugs the tests caught rather than the code:
- the display cap was first set to 50 while its own comment claimed
it sat above the largest company — which has 52 staff. That
silently re-created the truncation this change removes. Now 100,
with a test asserting a 52-person company comes back whole.
- PocketBase has no escape for a literal backslash: `\'` is
accepted, but one, two or three backslashes all fail to parse and
500 the request, so typing `\` in the search box errored. No
escaping survives, so backslashes are stripped before the filter
(`M20\Nut` searches `M20Nut`) and a backslash-only query returns
nothing rather than matching every customer.
13 new tests cover the company clause, the 52-staff case, the cap,
ranking order, backslash stripping and the empty-after-strip case.
Verified live against PocketBase across 10 queries — the ranking
guarantee held on every one.
sw.js v33 -> v34: pos-app.js and pos.html are cache-first in the shell.
Ringing up 800 of something means typing 800, and the only route to the
quantity field was a 600ms long-press with no affordance anywhere on
screen. The quantity itself — the thing a cashier reaches for — was a
plain <div class="v">, so it rendered inert.
- the quantity is now a real button that opens the Edit line sheet,
with an aria-label and a 44px touch target (it was 30px, too narrow
to even show "800")
- holding +/- no longer pops the sheet open mid-tap: the pointerdown
handler bails for .qty/.rm before starting the timer
- the sheet is quantity-only. The Unit price field is gone on request;
a mis-keyed price is expensive and easy to miss. The per-line
override it fed is removed with it — pos-service.ts only ever read
unit_price, the "price edited" label and the sale record's `edited`
flag were unreachable, and nothing else could set an override,
including a cart reload (persistCart saves only {id, qty, p})
The discount trigger moved out of its own 54px footer row to ride
inside the Subtotal row, so it costs no vertical height at all while
staying visible — a discount control that appears from nowhere is one
nobody finds under pressure. The applied-discount red totals row
remains the always-on indicator.
Fixes a pre-existing bug found while testing: long-pressing a cart row
showed a grey screen with no dialog. #editSheet is a SIBLING of
#editBackdrop, but the CSS revealed it with
`.edit-backdrop.show .edit-sheet` — a descendant selector that can
never match a sibling — so the sheet sat at opacity:0 while the
backdrop went opaque. Nothing threw and the build was green, which is
why it survived; it is also why manual quantity entry looked missing.
The sheet now has its own z-index above the backdrop and its own .show
rule. They stay siblings on purpose: nesting the sheet inside the
backdrop would make every tap on Save bubble to the backdrop's
onclick and close the dialog.
Verified in jsdom against the real markup: old CSS reproduces the grey
screen (backdrop opacity 1, sheet opacity 0, not clickable), the fix
clears it (sheet opacity 1, z-index 71 > 70, position fixed). 125 POS
tests pass, including 23 new ones asserting the tap target exists, the
long-press guard fires before the timer, the sheet stacks above its
backdrop, and the popup stays quantity-only.
sw.js v30 -> v33: pos-app.js and pos.html are both served cache-first out
of the shell, so without the bump tills keep the old ones.
