Hermes (n8n) drops the AGENT_API_KEY on direct REST calls, so anything
n8n needs to run must exist as an MCP tool. This forwards the agent's
JSON-RPC call to POST /api/v1/orders/quote-match/sweep, mirroring the
existing orders.set_quote forwarding pattern.
Updated the tools/list expected-list test to include the new tool so
the catalog stays self-documenting.
Co-Authored-By: Claude <noreply@anthropic.com>
The matcher only runs on orders created from now on. This sweeps the
historical backlog plus any order whose lines changed after creation.
- sweepUnmatchedOrders in order-quote-link-service.ts walks
quote_matching_attempted = false && quote = '' (the codebase's
pattern for an empty relation, verified against orders-service.ts
usage) in batches, projecting only id (the wider order fields are
re-fetched inside suggestQuoteForOrder).
- bucketAgeGaps exposes the histogram that lets FUZZY_WINDOW_DAYS be
re-tuned from real data instead of the 90-day estimate.
- POST /api/v1/orders/quote-match/sweep is the agent-auth surface
with maxDuration=300 for the 500-order worst case.
Tests cover the filter shape, the field projection (load-bearing —
over-projection wastes wire bytes), the no-op on no-match path, the
fault tolerance (one row throwing doesn't kill the batch), and all
six histogram bucket boundaries including the ≤ N vs < N edges.
Co-Authored-By: Claude <noreply@anthropic.com>
The mock for next/server's after() detaches the callback regardless of
what production code does, which left the outer catch's fire-and-forget
fallback untested and the runQuoteMatch inner catch tested for nothing.
- New test 'falls back to an inline match when after() is unavailable'
forces after() to throw and asserts suggestQuoteForOrder was called
anyway, plus that the fallback log fired. Mutation: deleting the
'void runQuoteMatch()' line breaks it.
- 'never fails order creation when the matcher throws' now spies on
console.error and asserts the inner catch's swallow actually logged.
Mutation: deleting the inner console.error breaks it.
- The outer catch now logs [createOrder] after() unavailable... so a
silent switch to fire-and-forget (after upgrade, waitUntil missing on
deploy target) is impossible to miss in production logs.
- Comment names error code E468, Next 16.2.6, and
node_modules/next/dist/server/after/after.js, and corrects the
justification: createOrder is an exported service function reachable
from multiple layers, not 'a smoke script or another background job'.
Verified against the documented baseline — same single pre-existing
failure (splitOrderForContinuation > moves pending items...), no new
failures introduced.
Co-Authored-By: Claude <noreply@anthropic.com>
createOrder previously hard-coded quote_matching_attempted: false and
never scheduled a match, so ~85% of orders that correspond to an
existing quote arrived unmatched and stayed that way until an
operator opened the order and clicked a button. Fire
suggestQuoteForOrder(order.id) via next/server's after() once the
response has been sent, so intake latency is unaffected, and guard
both the matcher call and the after() call itself: after() throws
synchronously outside a request scope (confirmed by reading
node_modules/next/dist/server/after/after.js), so a caller without
one (e.g. a background script invoking createOrder directly) falls
back to firing the match inline instead of crashing order creation.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The previous commit's two "happy path" pairing tests never reset
pairOrderLines' mock call history — vitest.config.ts doesn't set
clearMocks, and the file's global beforeEach only clears the PB
store. A prior passing setOrderQuote test left toHaveBeenCalledWith
already satisfied on entry, so those two tests passed even with the
production call site deleted. Add mockClear() after the dynamic
import in both so they actually exercise the call they claim to.
Also: reworded the idempotent-path comment to note the strict-pending
gate runs first, so pairing only re-runs there while every line is
still pending (not a general refresh mechanism); fixed the
"don't write, don't stamp" line, which the pairing call now
contradicts (order_items.quote_item is a write); and replaced a
comment pointer to a nonexistent plan section with a standalone
statement of the deferred-signal gap.
setOrderQuote is the single choke point for creating orders.quote —
used by the UI, REST API, and Hermes/n8n. Call pairOrderLines(orderId,
quoteId) right after the hard link lands so confident order-line to
quote-line pairs persist automatically instead of needing a separate
manual step.
Pairing also runs on the idempotent no-op path (order already points
at the requested quote) — it's safe to repeat and lets an operator
refresh pairing by re-running the same call after editing order lines.
A pairing failure never fails the link itself (console.error only,
by design) — the hard link is the operator's important intent and
must land even if pairing can't complete.
pairOrderLines' no-overwrite guard only checked whether an order
line's own row already carried a quote_item, not whether the quote
line it was about to claim was already held by a different order
line. A description drift (or a hand-confirmed link) on one line let
a fresh confident match on another line silently double-assign the
same quote_item, so two parts would render the same estimator's note
and attachments — the exact failure this feature exists to prevent,
reached from the other direction. Fixed by seeding a claimed-quote-item
set from the fetched order rows before the write loop, and skipping
any confident pair that targets an already-claimed quote line.
unmatchedOrderItemIds now also excludes any line that already carries
a link, even when this run's pairing pass didn't re-propose it.
Also names the two field projections (PAIR_ORDER_ITEM_FIELDS,
PAIR_QUOTE_ITEM_FIELDS) next to the existing MATCHER_QUOTE_ITEM_FIELDS
constant, and adds a test that asserts on the actual `fields` string
passed to each getFullList call. Previously nothing distinguished
`fields: "id,description"` from `fields: "id,description,quote_item"`
at the order-items query — a one-token trim silently disabled the
no-overwrite guard (every row read as unlinked) with the full suite
staying green, since the existing behavioral tests supply `quote_item`
values directly in their fixtures rather than through the query. Verified
by re-running that exact mutation against the new test: it fails, and
neither pre-existing behavioral test catches it on its own.
Adds pairOrderLines to order-quote-link-service.ts: pairs an order's
lines against a quote's lines via pairQuoteLines, writes quote_item
only for confident (>=0.5 Jaccard) pairs, and never overwrites a
line's existing quote_item. Possible-tier pairs are returned but left
unwritten so an operator has to confirm them before a buyer sees a
note attached on a guess.
Reuses fetchQuoteItemsForQuotesAdmin (already used by the fuzzy pass
in this file) to fetch quote_items rather than reintroducing a raw
QUOTE_ITEMS_COLLECTION query. Writes stay sequential, matching the
existing bulk-update convention in orders-service.ts.
Six review findings on suggestQuoteForOrder, all bite hardest once
order intake and the backlog sweep start calling it:
1. fetchQuoteItemsForQuotesAdmin's `fields` string was a hardcoded
constant owned by the duplicate-quote scan. The matcher now passes
its own `fields` (an added, defaulted parameter) instead of
depending on that string never being trimmed out from under it —
losing `description`/`selling_price` would have collapsed every
confidence score toward zero with no exception and no failing
test.
2. The candidate-quote query in fuzzyPass had no `fields`, so it
pulled full quote records (email_body, ai_insights,
migration_meta) for up to 200 candidates, twice per order when the
window widens, to read five scalars. Added
`fields: "id,quote_number,company,customer,created"`.
3. findExactMatch's bare `catch {}` swallowed everything, not just
the "field doesn't exist on an older PB schema" case it claimed to
guard (all four fields exist per pb_schema-2.json). An expired
admin token or a PB blip during the sweep would silently fall
through to the fuzzy path, and writeSuggestion would then persist
quote_matching_attempted: true / quote_match_method: "no_match" -
permanently hiding an order that had a valid citation from the
sweep's retry filter. Narrowed to rethrow anything that isn't a
ClientResponseError with status 400.
4. Exact-identifier lookups (quote_reference_text, rfq_number are
customer free text) weren't scoped to the order's company, so a
reference-number collision across two customers could return a
confidence-1.0 suggestion pointing at the wrong company's quote.
Added `company = "..."` to the filter when the order has one.
5. SuggestQuoteResult.window used the literal "90d", hardcoding
FUZZY_WINDOW_DAYS. Renamed to "initial" so retuning the constant
can't make the sweep's reporting lie.
6. order-quote-link-service.test.ts's makePb forced every test to
redeclare the full collection map. Gave it a default map merged
with a per-test override and renamed useP to mockAdminPb, since
two more tasks add tests to this file.
Adds tests for #1 (matcher's quote_items fields include description
and selling_price), #2 (candidate query's fields), and #4 (a
same-reference quote belonging to another company is not returned).
fuzzyPass built a single `quote = "x" || quote = "y" || ...` OR-chain
across up to FUZZY_CANDIDATE_CAP (200) candidate quotes to fetch their
items. The predecessor in pb-helpers.ts documents exactly why that
doesn't scale: PB's filter expression has a max length, and a naive
OR-chain 400s past roughly 50 ids. There was no try/catch, so
getFullList's rejection propagated straight out of
suggestQuoteForOrder - the matcher failed outright for precisely the
high-volume companies where auto-matching matters most.
Reuse fetchQuoteItemsForQuotesAdmin (now exported from
quotes-service.ts) instead of a third copy of the same chunking logic.
It already does this for the duplicate-quote scan: try PB's `IN (...)`
syntax first, fall back to 80-id chunked OR-chains on rejection.
Adds a regression test that simulates PB rejecting an oversized filter
across 100 candidate quotes and asserts the fetch recovers via
multiple chunked calls rather than rejecting outright. Confirmed it
fails against the pre-fix code (uncaught rejection propagates out of
suggestQuoteForOrder) before verifying it passes against the fix.
Ports the PocketBase half of the order<->quote matcher off the
browser's PB singleton and onto getAdminPocketBase(), so matching can
run at order intake and from a backlog sweep instead of only on a
button click. Reuses the pure scoring in order-quote-match.ts
untouched (scoreOrderVsQuote, quoteWindowForOrder,
quoteToOrderAgeDays).
suggestQuoteForOrder tries the four exact-identifier fields first
(never date-limited), then falls back to a company/customer/
description/price fuzzy scan windowed to 90 days before the order
date, widening to 365 days only if that comes up empty.
Writes only the soft suggestion (potential_quote,
potential_quote_confidence, quote_matching_attempted,
quote_match_method) - never orders.quote, the hard link. Orders that
already carry a hard link are skipped untouched.
pb-helpers.ts's findMatchingQuotesForOrder is left in place; a later
task migrates its last UI caller and removes it.
Two findings from review of the quote-line-pairing task:
- quoteToOrderAgeDays: Math.round ties toward +Infinity, so a negative
half-day gap rounds to -0. Normalize to a clean 0.
- quoteWindowForOrder / quoteToOrderAgeDays relied on `new Date(...)`
parsing PocketBase's space-separated datetime form
("2026-08-29 10:00:00.000Z") the same as ISO 8601. That happens to
work on V8 but is implementation-defined per spec. Both now go
through a shared parseFlexibleDate helper that normalizes the
separator explicitly.
Adds tests pinning both behaviours; the existing 40 tests stay green.
Adds quoteWindowForOrder (pure) to constrain the fuzzy-match candidate
scan to a date window anchored on the order's own created date, plus
the forward grace period for quotes raised a few days after the PO.
Also adds quoteToOrderAgeDays, used by a later task's age-gap
histogram, with test coverage the plan omitted (normal, inverted, and
unparseable-input cases).
Code review found LINE_PAIR_POSSIBLE was decorative -- pairLinesByDescription
still filtered on a hardcoded 0.3 literal, so editing the constant wouldn't
change behaviour. Also adds a test pinned to an exact 0.5 Jaccard pair to
catch off-by-one mutations on the confident/possible tier boundary.
The original sample scored 0.222 Jaccard — below the 0.3 candidate
floor — so it would have been dropped rather than tiered. Verified
0.333 for the replacement.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
13 TDD tasks across two phases. Phase 1 (backend) ends with orders
matched automatically at intake plus a backlog sweep; Phase 2 surfaces
the paired quote line's note and attachments on both order pages.
Resolves the spec's two open questions: `after` is stable in
next/server on Next 16.2.6, and PocketBase date filters take
`.toISOString()`.
Two deviations from the spec, both noted in the plan: line pairing
hooks into the existing setOrderQuote rather than a new
linkQuoteToOrder, and the per-line context extends the existing
matched-attachments response rather than adding an endpoint.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Design for closing three gaps in the existing order↔quote matcher:
- The matcher only runs when an operator clicks it, so the ~85% of
orders that arrive against a quote stay unlinked. Adds a call from
createOrder plus a catch-up sweep for the backlog.
- order_items.quote_item is only written when an order is created
from a quote, so post-hoc links never pair lines. Adds persistence
for confident pairs.
- findMatchedQuoteItem re-guesses the pairing at render time using
exact/substring matching, so buyer notes and per-line attachments
silently vanish on any wording difference. Replaced with the
tokenised matcher already in the repo.
Fuzzy candidates are windowed to 90 days before the order's own date
(widening to 12 months only when nothing clears the threshold);
exact-identifier lookups stay unwindowed.
No PocketBase schema changes required.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Moving the admin panels below the items table left them at a third of
the row each (~186px of tab strip), which is too narrow for the labels
they carry. "Outbound" plus its count needs ~71px against the ~58px it
was given, so the three Attachments labels overlapped each other; the
same squeeze hit "Email Status" on the Activity card.
Lays the panels out two-up instead: Attachments spans both columns on
its own row (its three tabs, dropzone and file list want the space),
with Linked Orders and Activity sharing the row beneath at half width.
Both strips now have several times the headroom they need, measured at
a 1044px viewport, and mobile is unchanged (single column, no
page-level horizontal scroll).
Also drops the `ml-1.5` from the attachment count spans — TabsTrigger
already applies `gap-1.5` between its children, so the margin was
double-spacing and contributed to the overflow — and adds
`tabular-nums` so the tab widths stay steady as counts tick past 9.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The detail page was a 3-column grid with the Items card in a
lg:col-span-2 and the admin panels beside it. That left the items
table ~430px for a table whose natural width is ~790px, so pricing a
quote meant scrolling the table sideways.
Stacks the page instead: the items table (the actual work surface)
takes the full width, and Attachments / Linked Orders / Activity sit
3-up underneath, where they read fine at a third of the width because
they're reference panels rather than things you act on while pricing.
Two width trims make the table fit outright rather than merely fit
better:
- Description's min-width 360px → 280px. It's a floor, not a target;
the table is table-layout:auto so Description still absorbs any
width the fixed columns leave. 360 was forcing overflow on its own.
- The "Processed" header is whitespace-nowrap, so the word set a
~78px floor on a column whose content is a 2-3 character badge.
Renamed to "Done" (title="Processed") at w-[64px].
Measured at a 1044px viewport: the table went from 110px of horizontal
overflow to 0. Mobile is unaffected — it uses the stacked card list,
and the admin grid collapses to one column.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The table had 9 header cells against 8 body cells: a stray duplicate
"Description" header sat above the one carrying min-w-[360px]. Every
column therefore rendered one position off its label (the "Qty" header
sat over the supplier cell, "Supplier" over cost, and so on), and the
360px reserved for Description landed on Qty instead.
The visible symptom was a long description crushed into ~99px and
wrapped one word per line down a tall thin column, while the Qty cell
sat in 360px of empty space.
Removing the stray header restores the header/cell alignment and gives
Description its intended 360px. Also switches the row action from a
"Update price" text button to an icon-only button with the same title
and an aria-label, which trims the table's natural width from 911px to
789px — the card it lives in is only ~430px wide, so every pixel is
horizontal scrolling the operator has to do.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
When a mine puts a spot out to tender, several engineering companies
each ask us to price the same item list. Before pricing a new one the
operator wants to see what the earlier quote went out at.
The per-line answer already existed: populateItemIntel powers the ⓘ
popover on each row and returns recent_quote (prior quote number,
company, selling price, days ago). This adds no new matching logic —
it runs that same lookup across every line of a quote and rolls the
results up, so the operator gets one panel instead of clicking ⓘ on
each row.
What the roll-up adds over the per-line popover:
- "N of M lines have been quoted before" at a glance
- relatedQuotes ranked by HOW MANY of this quote's lines they also
carried; a prior quote covering 5 of 5 is very likely the same
tender from a different bidder, one covering 1 of 5 is a shared
commodity line
Two honesty details, both driven by what the first live run showed:
- match confidence is surfaced as close/likely/loose. The underlying
matcher accepts anything above 0.5, and a live check returned a
R150 hit against a magnetic base drill — a loose row must not look
as trustworthy as a close one.
- an unpriced prior quote renders "unpriced", never R0.00, which
would read as "we quoted it at zero".
Read-only: passes write:false so viewing the panel doesn't persist
intel onto rows the operator only looked at. On-demand rather than
auto-loading, since it costs one intel computation per line
(~400-650ms each, batched 5 at a time, capped at 40 lines).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Adds findDuplicateQuotes(), which groups quotes representing the same
underlying request. The business case is mining tendering: several
engineering companies each ask us to price the SAME item list because
they're all bidding for one spot on a mine, and the office needs to
find what the earlier quote went out at.
That inverts the usual dedupe rules, so both are deliberate:
- matching is CROSS-COMPANY (two customers sending the same list is
the signal, not a false positive)
- the email-body pass is CROSS-SENDER (bidders forward the same
tender document from different addresses)
Three passes: item_content (normalised line fingerprint, order
independent), email_body, and email_message_id.
Noise filtering is tuned against live data. The n8n intake writes an
"items from email request <name>" placeholder when it can't extract
line items; without excluding those, every unextracted email matches
every other one. A minimum line length drops junk like a bare "quote"
while keeping real one-liners the office genuinely re-quotes.
The completed-quote window measures `created`, not `updated`: a bulk
backfill rewrote `updated` on all 8,964 rows to a single timestamp, so
an `updated >=` cutoff matches the whole collection.
Not wired to the UI. A live run showed the quotes collection has no
trustworthy per-quote date (94% of `created` and 95% of `updated` are
migration stamps; quote_date and email_received_at are ~empty), so the
recency window can't currently bound the scan, and full-history content
matching returns groups of up to 53 for commonly-quoted commodities.
Landing the service on its own so the matching logic is available for
the on-demand "have we quoted this before?" lookup.
Also fixes a latent bug in the test PB mock's filter parser, which
stripped outer parens even when the leading "(" didn't match the
trailing ")", corrupting `(a) || (b)` into `a) || (b`.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
.claude/launch.json lets the Browser pane start the app on :4001 by
name instead of needing a manually-run dev server. The .tmp screenshots
are visual verification captures from mobile dashboard debugging.
The archive's search fetched the entire recency-windowed matching set
(thousands of rows once "Include older" is on) and substring-filtered
it in memory, taking several seconds per keystroke; it now delegates
to the same ranked pushdown search service the main /quotes list uses.
Also fixes pagination being hidden while searching (results past page
1 were unreachable) and the match count showing the current page's
size instead of the true total.
Adds the chevron item-preview toggle to archive rows (extracted into
a shared QuoteItemsExpanded component, backed by a new GET
/api/v1/quotes/[id]/items route), removes a duplicate "Customer or
Company" label in the new-quote form, and drops the stray
warning-colored search box styling that leaked from DataTable onto
every list page that uses it.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- Updated the search input handling to improve responsiveness by integrating a debounced search mechanism, reducing latency from 550 ms to approximately 300 ms.
- Added keyboard shortcut support for focusing the search box, enhancing accessibility and usability for frequent users.
- Improved the search placeholder to reflect all searchable fields, providing clearer guidance to operators.
- Introduced a clear-search button to streamline user interactions, allowing quick resets of the search input.
- Enhanced the display of search results with a live result count badge, providing immediate feedback on search outcomes.
These changes collectively enhance the search experience for operators, making it more intuitive and efficient.
- Added `email_sender`, `email_subject`, and `email_body` fields to the order creation input schema for improved traceability of inbound emails.
- Updated logic to include these fields in the order creation process, allowing operators to track the source of orders without needing follow-up updates.
- Enhanced documentation to reflect the new optional fields and their purpose in the order management workflow.
This update improves the audit trail for orders and enhances user experience by linking orders directly to their source emails.
- Introduced `order_number` field in the order creation input schema, allowing users to provide a custom identifier (e.g., PO/reference).
- Implemented logic to handle empty or whitespace values, defaulting to auto-generated order numbers.
- Added error handling for order_number collisions, returning structured error responses for better client-side handling.
- Updated documentation to reflect the new order_number functionality and its implications.
This enhancement improves order management flexibility and user experience.
Companion to the design spec at docs/superpowers/specs/. Covers all 15
implementation tasks (windowRange helper, aggregations, rep-activity
pagination, route handlers, tests, PWA manifest, mobile layout,
dashboard view with realtime + bottom sheet) plus the final
verification step.
Co-Authored-By: Claude <noreply@anthropic.com>
Fixes react-hooks/set-state-in-effect lint error in the tab-change useEffect.
- Tab buttons now call selectWindow(w) directly instead of relying on a
useEffect that synchronously calls setState. selectWindow flips the
state, closes the sheet, and triggers the refetch from one event handler.
- fetchData now reads activeWindow via a ref synced by a tiny useEffect
(the canonical React pattern — eslint flagged the inline ref-write as
'Cannot update ref during render'). With no useCallback deps, fetchData
identity is stable, so the realtime subscription no longer churns on
every tab switch (a perf concern flagged in Task 15).
- The mount-only initial fetch useEffect is preserved with a justified
eslint-disable on the set-state-in-effect rule (no SSR data pass-through
on this client-only page, so the initial fetch has no user event to
hang off).
Co-Authored-By: Claude <noreply@anthropic.com>
