Measured all 54 blocked notes first. Only 27 are agent work:
24 not_duplicate the same Vula invoices uploaded three times in
different groupings. Resolving these would
double-count real deliveries.
22 product_matched genuine decisions -> the agent
9 units_convertible pipes sold by the each, ordered by the metre.
The 2.5 in 'IJ Pipes 450x2.5' is the length, so
1 ea = 2.5 m. A conversion rule, not a judgement.
5 po_resolved misread order numbers -> the agent proposes only
The brief leads with the rule that governs it: a wrong product match is
worse than a block, because it applies the wrong rate to a supplier
payment. It names both real wrong matches from this data, ranks the
evidence that is actually decisive (quantity pairing first, word overlap
never), and routes every action through the API so each decision writes
an alias and compounds.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A tick is fully sequential: one vision call per page, then each note's
checks and promotion in turn. Claiming three batches of a 13-page scan
can approach the route's 300s ceiling, and a tick that times out leaves
its batch to be reclaimed and re-read, paying for those pages twice.
One batch per tick also makes n8n's execution log directly readable —
one run, one document, one outcome. DN_TICK_BATCH_SIZE still overrides
it once the page-count distribution is known.
34 batches at a 1-minute schedule is ~34 minutes rather than ~12.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Constrains the implementer away from the two failure modes that have
already bitten this work: writing to production PocketBase, and guessing
when the plan does not match the code. Also states plainly that a wrong
product match costs money and a block does not, since Task 6 is where
that trade-off lives.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Making mass_unit required breaks typecheck in three test files the task's
file list did not name. Anyone re-running Task 3 in isolation hits the
same failures.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Twelve tasks against the nine defects the first production run exposed.
Task 6 is the core: a three-outcome PO item matcher (matched / ambiguous /
none) with a dimension gate, where the set of integers >= 100 must match
exactly. Two naive relaxations tested during design both produced wrong
matches -- U PIPES-600 resolving to S&S Pipes-1050, Kabelflex to uPVC Pipe
-- and a wrong match applies the wrong rate to money. The task ships a
regression test asserting no known input returns a wrong match, and the
plan tells the implementer not to move the thresholds without extending
that table first.
Task 7 adds the alias table, keyed on material_code so a human decision
carries onto the next purchase order. That is what makes wording like
"Kabelflex Pipe & Coupling" -> "160mm Sleeves Black" tractable at all.
Task 11 deliberately narrows the original page-accounting guarantee: a
stray page becomes a dismissable open question instead of failing the
batch, because batch 5qzcv6tvgb6cqks promoted 3 of 3 notes correctly and
still read as a failure over a tax invoice cover sheet. The guarantee that
every detected note must promote is unchanged.
Task 12 re-runs all 34 batches and audits every ledger link, so a matcher
regression cannot land silently.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Ran the pipeline over 34 real documents: 7 processed, 14 blocked, 13 never
picked up. 13 of the 14 blocks were our own bugs.
Root causes, measured from dn_batches.checks:
- 5 batches are tax invoices or collection notes, which carry no delivery
note number. The pipeline was built assuming every document has one.
- 4 blocked on product_matched; the matcher requires the PO line's first
token to be 3+ chars, so "IJ Pipes 450x2.5 100D" never matches
"IJ PIPES-450 X 2.5(100D)".
- 2 on weighbridge_arithmetic comparing nett in kg against quantity in tons.
- 2 on a single stray page condemning notes that promoted correctly.
Critically, two candidate relaxations of the matcher were tested against the
real descriptions and BOTH produced wrong matches (U PIPES-600 to S&S
Pipes-1050; Kabelflex to uPVC Pipe). A wrong match applies the wrong rate to
money, so the fix is a three-outcome matcher -- matched / ambiguous / none --
with a dimension gate, plus a learned alias table for wording that can never
match textually. The specified algorithm was verified against all 23 failing
descriptions with zero wrong matches.
Also found: dn_batches.created is never populated, so board ordering is
arbitrary; and the tick has never claimed a batch, so the n8n schedule is
not running.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Fixes two bugs in delivery-note processing:
1. 1000x unit conversion error: slips quoted in kg but PO lines priced per ton.
2. Off-by-one that paired delivery lines to the wrong PO items.
All 8 plan tasks complete: units module, dn-item-row builder, route rewire,
dual-unit UI rendering, backfill migration (56/57 rows patched), and operational
cleanup. Final state: PO MEGCPO9165 line bxqbo8rs04pjqwc reconciles to
133.5 ton / R22,494.75 across 4 notes (not 1 515% or R10.9 m).
See docs/superpowers/plans/2026-09-17-delivery-note-unit-correctness.md.
Seventeen tasks covering the batch collections, the pure segmentation and
check modules, page rendering via poppler, per-page extraction, promotion,
the orchestrator, the tick route, the Kanban board and the batch detail view
with note correction.
The design point: pages are read independently and grouped into notes by the
delivery note numbers printed on them, so a note spanning three pages is one
note and a page belonging to no note blocks the batch. segmentPages throws
rather than return a segmentation that loses a page, and runBatch cannot
reach "processed" unless notes_promoted equals notes_detected -- the check
whose absence let two notes vanish from scan_20260915 while the pipeline
reported success at 95% confidence.
Task 16 step 7 is the regression test for exactly that: force a page read to
fail and confirm the batch lands in needs_attention rather than processed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Eight TDD tasks covering the units module, the row builder that removes the
matched-id off-by-one, the PocketBase schema additions, the route rewiring,
dual-unit display, the backfill migration for the 57 existing rows, and the
operational steps to stop n8n minting orphan records and recover the two
lost notes from scan_20260915.
Also corrects the spec to name the migration as Python, matching the
existing scripts/ convention.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Diagnoses why multi-note delivery slips lose data and specifies phases 1-3
of the fix.
Root cause: the n8n workflow's "Match items to PO" Code node runs in
runOnceForAllItems mode and returns a single item, so the workflow makes
exactly two callbacks regardless of how many notes a document holds. A
4-note scan produced 2 promoted notes and 2 empty records in an invisible
status, while reporting auto-verified at 95% confidence.
Also found: delivery notes are weighbridge tickets in kg while PO lines are
priced per ton, and nothing converts. po-fulfilment and delivery-note-value
therefore overstate received quantity 1000x across all 57 existing rows.
Phase 1 fixes the unit model and migrates existing data. Phase 2 moves
ingestion into a staged inbox with a deterministic validation gate, so
nothing reaches delivery_notes unless every page is accounted for. Phase 3
renders that state machine as a Kanban board.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Scans a worker's existing QR badge to clock them in/out, writing
directly into timesheet_entries as a "Direct (QR)" entry — the
timekeeping app foreshadowed in the workers-timesheets-qr spec.
Re-verified every file:line against the current tree. Three were wrong:
the Overview ranges in task 6 (and it listed two sites.items uses where
there are four), the hamburger line in task 5, and a claim that
plant/page.tsx still had perPage: 5000 caps — it does not.
Also records that tasks 1-5 have landed, so the next worker starts at 6,
and keeps the scope-aware Sites sub-label that was implemented in place
of the deletion this plan originally specified.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The migration's docstring pointed at a spec file that was never
written for this feature (the design was worked out via conversation,
not a separate spec doc) — points at the plan instead.
A concurrent project Month/Week feature is landing in the same view
files and has already shifted purchasing-view.tsx by ~10 lines, adding
a periodFilter beside projectFilter. The removal steps now match on
identifiers and snippets, and the plan header says to treat every
file:line as a hint.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
16 TDD tasks: pure scope helpers, the getSiteScope seam, the topbar
switcher, then per-module server-side filtering. Task 15 records why
the chat agent's default-site change is deferred out of this plan.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replaces the four independent per-table project-code dropdowns with a
single global scope carried in the URL as ?site=, resolved server-side,
and applied as a PocketBase site= filter. Defines the getSiteScope()
seam so the separate PocketBase users work plugs in without touching
the 23 Topbar call sites.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Workers becomes a lean roster (name, employee number, position, type,
site, rate) with filters unchanged. Everything currently cluttering the
list row -- total hours, cost, the 8-week sparkline, per-employee entry
history -- moves to a new worker detail page that doesn't exist yet
(today's "detail" is a dialog). Same pattern already used for Suppliers
and Vehicles/Equipment this session: lean list, detail page carries
depth.
Timesheets keeps its flat entries table but gains a genuinely separate
upload-status section -- the timesheets document collection already has
this data, it's just currently squeezed into a per-row badge instead of
being its own thing. The duplicated Import-timesheets button (currently
on both pages) moves here only.
QR badges: encodes employee_no directly, no new PocketBase field --
every worker already has a valid code the moment they exist, with no
generation step. Explicitly a deliberate tradeoff (guessable numbers,
no revocation) the user chose for now, pending seeing how the separate
timekeeping PWA wants identity handled -- moving to a stored opaque
token later is additive, not a rework. Card layout confirmed via the
visual companion (centered, QR-forward). Both single-worker and
batch-sheet print confirmed in scope. Rendering via the qrcode npm
package + the browser's own print dialog (@media print), not a PDF
library -- CSS handles real-world card dimensions more reliably than
fighting a PDF engine's coordinate system.
The timekeeping PWA itself is explicitly out of scope -- a separate
project this spec only needs to stay compatible with.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
13 tasks across 6 phases: additive schema, pure matching/ledger
functions (TDD, tested in isolation), the one irreversible migration
(guarded, dry-run default), OCR wiring, API routes, and UI.
Grounded in the actual codebase rather than the spec's prose: the
existing "unknown supplier" inline-editor on the PO detail page
(SupplierEditor + POST /api/purchase-orders/[id]/supplier) is upgraded
in place rather than duplicated, following the same principle applied
to the sidebar restructure's delivery-note review queue. Every code
sample is written against files actually read during planning, not
guessed -- site-form-dialog.tsx, the /api/sites routes, matchSite,
po-fulfilment.ts, and the live purchase_orders.supplier data.
Self-review caught a real gap before committing: pb-bootstrap-schema.py
and pb-types.ts were named in the file-structure table but no task
actually touched them -- exactly the foot-gun docs/data-model.md
section 7 warns about. Added as Task 1 steps 5-8.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
19 -> ~16 was wrong. Ran the actual normalizeSupplierName() (trim +
whitespace-collapse + uppercase) against the three example pairs cited
as evidence: only the case-variant pair (Builders Warehouse) actually
collapses. The Ikhasi pair keeps a stray space after a hyphen that
differs between variants; the Global Precast pair differs by the
"(Pty) Ltd" suffix, not case or whitespace. Neither collapses under
exact-normalized matching.
This is not a defect -- it's the conservative-matching principle in
section 3 working as designed. The correct expectation against
production is 19 raw strings -> 18 groups, one merge. Both pairs that
don't collapse are exactly the kind of ambiguous case the spec already
says should stay separate for a human to resolve, not the kind that
should be forced together by loosening the normalizer.
Caught while grounding the implementation plan in real data, before
writing a single line of the plan itself.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
New suppliers collection replacing purchase_orders.supplier (free text)
as the authoritative value, kept alongside it as OCR provenance rather
than dropped -- it genuinely is the supplier name, just inconsistently
formatted, unlike project_code which never held a real project code at
all.
Grounded in production: 24 POs, 19 distinct raw supplier strings, at
least 3 confirmed duplicates under a formatting variant. Migration
auto-merges only exact-normalized matches; anything ambiguous (e.g. two
"Vula" entries that may or may not be the same company) stays separate
rather than guessed.
The accounts-payable rollup (ordered/delivered/outstanding per supplier,
in Rand) is computed live from purchase_order_items and
delivery_note_items every read, never stored -- same reasoning as
po-fulfilment.ts: a stored balance has to be kept correct on every PO,
delivery, and correction, and PocketBase has no transactions to
guarantee that.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Reorganizes the flat 7-item sidebar into Workforce/Plant/Financial
groups, each with a landing dashboard and sub-pages split out of
today's combined Employees and Purchasing pages. Delivers the Delivery
Notes page and ingestion-status Uploads view that
2026-09-11-ingestion-restructure-design.md section 8 deliberately left
for later, now that the data model is settled.
Suppliers (new collection, PO relation, free-text-to-record matching
migration) is explicitly out of scope -- split into its own spec per
user request, same rigor as the ingestion restructure.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
.pyc files are generated, not source -- machine/version-specific,
produce spurious diffs, and a routine __pycache__ cleanup can delete
tracked files as a side effect (hit this earlier in the session). Added
__pycache__/ to .gitignore.
Also fixed a boundary error in the plan's checkbox sweep: Task 17's own
6 steps were skipped when checking off Task 18's range. Task 17 has been
done and independently verified since commit 29ee2cc.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
pb-bootstrap-schema.py -- the script a fresh PocketBase instance would
run against -- had fallen ten migrations behind and still described the
pre-restructure world. Rewrote it field-by-field against a live
production dump (17 collections, every field/required-flag/relation/
index verified). Found and fixed two pre-existing gaps unrelated to this
restructure while verifying: fuel_prices and income_claims were entirely
absent from the bootstrap script.
pb-relax-rules.py's collection list updated to match.
No local PocketBase binary is available in this environment, so the
plan's "bootstrap fresh and diff against production" verification step
was replaced with a manual field-by-field comparison against the live
dump -- same goal, different mechanism, documented as a substitution
rather than silently skipped.
Added site scoping to the one agent tool that still lacked it
(list_imports, backed by loadImportBatches) -- every other list_* tool
already had it. Preparation for the future per-user site-allowlist work,
not part of this restructure itself.
Rewrote docs/data-model.md and docs/data-flows.md end to end: the
site-attribution table, the collection reference, the five ingestion
pipelines, and the schema-drift section that was describing the exact
problem this task fixes. Every remaining data_import_batches reference
in both docs is now historical context, not a current description.
This is the last task of the ingestion restructure
(docs/superpowers/plans/2026-09-11-ingestion-restructure.md).
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
timesheet_entries, fuel_logs and delivery_notes had no indexes at all --
fine at today's row counts, not once site count grows toward the 18-20
target. Added 8 indexes across fuel_logs, timesheet_entries,
delivery_notes, purchase_orders and plant_hours_log, unioned into each
collection's existing index array rather than replacing it.
Verified both pre-existing UNIQUE indexes (idx_po_number, load-bearing
for the placeholder-PO-number scheme; idx_plant_hours_log_qr_start,
prevents duplicate weekly imports) are byte-for-byte unchanged.
Idempotent: a second run touches nothing.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Tightened site from optional to required on fuel_logs, timesheet_entries,
delivery_notes and purchase_orders now that every route writes it and
every historical row has been backfilled. plant_hours_log was already
required (matches docs/data-model.md's documented invariant, confirmed
against production). timesheets (the document collection) was already
required from Task 6 -- a no-op here, not a regression.
fuel_sheets.site and plant_hours_sheets.site deliberately stay optional:
those sheets can legitimately span multiple sites.
Verified: creating a fuel_logs row with no site now gets a hard 400
(validation_required) from PocketBase itself.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Task 15 of the ingestion restructure. Edited the live n8n workflow
(005iI4jX5mq9DsTR) directly via MCP and published the new version.
Repointed Watchdog: List Stuck, Watchdog: Reset Stuck Rows, Fetch Pending
Batches, Atomic Claim, Claim Successful?, Download File and Escalate
Failure from data_import_batches/status onto purchase_orders/
processing_status. Verified zero in-flight rows would be orphaned: all
26 old purchase_orders-type batches were already processed or failed,
and all 24 real PO rows had processing_status unset (won't match the new
pending filter).
Two things found on the live workflow that the plan didn't anticipate:
- Atomic Claim had no ?filter=processing_status="pending" guard at all --
an unconditional PATCH, despite the workflow's own name and this doc's
prior description both calling it "atomic". Added while repointing the
same node.
- Build Callback Payload sent the extracted PO number as `po_number`.
process-batch's validator requires the field named exactly
`purchase_order_number` -- the old field name would have failed
validation on every single real upload, a total silent outage. Fixed
before publishing, verified against the route's actual validateExtracted().
Also removed the n8n-side duplicate short-circuit: the old
Build Callback Payload queried purchase_orders and PATCHed the batch row
directly from n8n code to pre-empt process-batch's own dedup logic.
process-batch's merge branch (Task 14) already does this correctly,
server-side, tested. Keeping both meant the same decision made twice in
two places that could disagree. Is Duplicate? node removed; Build
Callback Payload connects straight to Send Callback.
Published as the active version and verified by reading the workflow
back in full -- every repointed node confirmed, connections confirmed,
zero validation warnings.
Rewrote docs/n8n/purchase-order-ingest-implementation.md end to end to
describe the deployed shape rather than the retired one. Flagged (not
fixed, separate decision) a plaintext PB superuser password baked into
the Auth PB Admin node.
Marked Tasks 1-15 complete in the plan's own checkboxes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Three document collections replacing data_import_batches for the
spreadsheet importers. Each owns its source file and carries honest
outcome counters (rows_total/created/failed, row_errors) — the counts the
current importers compute and then discard.
site is required on timesheets (a wage sheet belongs to one site) and
optional on the other two, which legitimately span sites since the fuel
importer resolves a site per row. cascadeDelete is false on all three, so
deleting a site cannot destroy its source spreadsheets.
Widened source_file mimeTypes beyond the plan's three entries to match
the pool being replaced. Browsers label CSVs text/csv, application/csv or
text/plain interchangeably and the importers branch on file extension,
not MIME — so the narrower list would have parsed a CSV successfully then
rejected it on write. The 2026-08-24 migration exists because that
already happened once.
Plan corrected: its TypeScript edit added a second `expand` property to
an interface that already had one, which does not compile.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
147 of 320 fuel_logs had no site relation; their only attribution was the
denormalised project_code text, which Task 18 deletes. Backfilled before
that column goes, resolving each row twice (by project_code and via the
vehicle's own site) and writing only on agreement.
147/147 written, 0 skipped, 0 remaining.
Recorded a caveat the dry run exposed: with one site in production both
resolvers draw from a candidate set of size one, so "both agree" is
guaranteed rather than evidence. The write is still the only possible
correct value, but the guard only bites once more sites exist.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Review found a transient 500 during download landed in the "no stored
file" bucket, so a broken run printed as a clean one and exited 0. Now
counted separately and exits non-zero.
Also: 404 is no longer swallowed on mutations (a PATCH against a deleted
record was counted as written), the target host and mode are printed, the
apply path reports a final written count, and filenames quote with
safe="".
Plan Task 1 Step 4 still carried the superseded copy-from-batch code
block under the rewritten prose, contradicting it. Replaced.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Production disproved the copy-from-batch approach in two ways:
- 10 delivery-note batches carry no file_hash at all (they predate
78fa8da), and all 10 sit inside the 30-day dedup window, so copying
would have left a live gap rather than a historical one
- one PO row was produced from two different files with different
hashes via the dedup-by-PO-number branch, so copying gives
last-write-wins on an arbitrary one
Hashing the stored bytes has neither problem and does not depend on the
collection we are about to delete. 56/56 records covered, none skipped.
Immediately surfaced a real duplicate the old path missed: two delivery
notes numbered IN132202 with identical bytes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
20 tasks across four phases, each phase ending at a shippable state.
Phase 1 additive only, Phase 2 renames and document collections, Phase 3
route cutover, Phase 4 drops. Nothing is deleted before Phase 4, so every
earlier phase is reversible.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Partial deliveries against a PO line are the reason qty_outstanding is
dropped rather than wired: a stored counter must be decremented correctly
on every receipt, OCR re-run and correction, and PocketBase has no real
transactions to guarantee that. The sum over delivery_note_items cannot
drift.
Records the relation chain, the two invariants (unmatched lines count
toward nothing; over-delivery is legal), and the fulfilment test cases.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Replaces the data_import_batches upload pool with five typed document
collections that each own their file, makes upload status reflect what
actually happened, puts a real site on every ingested row, and removes
the project_code duplication.
Audited against production on 2026-09-11:
- 19 plant_hours batches marked 'processed'; plant_hours_log has 1 row
- 147 of 320 fuel_logs have no site relation
- purchase_orders.project_code holds requisition numbers and site names,
not project codes — automatic site matching succeeds ~8% of the time
- PLANT collection and two data_import_batches fields have zero readers
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Updated the operations dashboard to include a rain outlook feature, providing users with a 7-day forecast based on site coordinates. Added latitude and longitude fields to the site form for better weather data integration. Enhanced the fuel log and site management functionalities to utilize dynamic fuel pricing based on historical data. Improved overall data handling and user experience across various dashboard components.
The vehicles PB collection already has a unit_type field (fuel |
plant, default fuel, made optional on 2026-08-28) and the
PATCH /api/vehicles/[id] endpoint accepts it with full test
coverage. Today there is no UI surface for assigning or changing a
vehicle's unit_type on either the Fleet page (/dashboard/fuel) or
the Plant page (/dashboard/plant), so the 53 currently mis-classified
vehicles cannot be triaged from the UI.
This spec adds that surface via the existing inline-edit pattern
already used on both pages for description editing: a Type column
renders a 'Fleet (L)' / 'Plant (h)' badge in display mode and a
Select in edit mode. Saving a flip calls PATCH /api/vehicles/[id];
on success router.refresh() re-pulls the page and the row drops
out of the page's filter, effectively moving it to the other page.
Scope is intentionally tight:
- Server-side filter on each page (Fleet: unit_type=fuel or empty;
Plant: unit_type=plant AND cost_per_hour > 0).
- New Type column on the existing Fleet and Plant tables (no new
pages, no new top-level nav).
- New shared <UnitTypePill> / <UnitTypeSelect> cell components.
- Reuses the generic useInlineEdit<K extends keyof T> hook (no
hook changes needed).
- No schema migration, no API changes, no new tests.
Spec self-review pass: tightened two claims (pb-types.ts UnitType
line 37 not 48; useInlineEdit is already generic so no hook edit).
