The whole pipeline: storage behind one driver interface (local disk and S3),
a portable job queue with a poller, QR and CDC parsing, OCR through the
Anthropic API, dedupe, manual entry, and the bandeja that turns all of it into
one decision per card.
Scanning tries the trustworthy door first: a QR is parsed and prefilled from
its CDC; a photo without one is queued for OCR when a key is configured and
otherwise opens the manual form against the stored file. A QR that will not
parse records an ingest error and falls through rather than losing the photo.
Job claiming is the only dialect divergence, as SPEC allows: FOR UPDATE SKIP
LOCKED on Postgres, a conditional UPDATE against SQLite's single writer. Retry
backoff follows SPEC exactly and a job abandoned by a killed process returns to
the queue once its lock goes stale, which is the phase 3 acceptance case.
OCR uses structured outputs rather than parsing prose, so the model cannot
return anything but the RULES.md schema, and every field is nullable because
unreadable is a real answer.
Web: scan with live QR decoding (BarcodeDetector, ZXing fallback, wasm served
from our own origin), manual entry with the IVA split worked out from the
total, the bandeja with swipe, buttons and keyboard all doing the same thing,
and the documents list and detail with an editable classification.
The seed now carries Maria's 34 purchases and 8 sales and Carlos's 6, all
classified through the real rules, plus the two open ingest errors.
Two defects found and fixed with tests: seeded documents could be dated in the
future, which would corrupt any projection computed from them, and the category
buttons announced their keyboard shortcut as part of their name.
255 vitest tests, 37 Playwright tests, rules coverage still 100%, typecheck and
lint clean.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every algorithm in docs/RULES.md, implemented exactly as written: the RUC
check digit, the calendario perpetuo with weekend and holiday roll forward,
CDC and KUDE QR parsing, keyword classification, computeF120, computeF515 and
the dashboard projections, plus the two form definitions.
Both worked examples reproduce verbatim on the first implementation: F120 at
Gs. 277.273 to pay with the Gs. 350.000 flip case, F515 at Gs. 11.290.000 on a
5,65% effective rate. 162 tests over the package, coverage enforced at 100%
statements, branches, functions and lines; the only exclusions are three
bounded-loop guards marked v8 ignore with a comment saying why.
No tax rule was invented. All six TODO-TAX-VERIFY items from RULES.md have a
test pinning today's behaviour and a row in the DECISIONS.md register, so
verification later is a red/green diff. Where RULES.md was silent the choice is
marked SPEC-GAP in the code and listed too, the notable one being that
taxpayer.hasIrp gates the deduction amount.
No floats anywhere in a money path: money is a branded Pyg of whole guaranies
and percentages go through integer arithmetic rounded half up.
Also: contracts now takes IRP_CATEGORIES from rules rather than declaring the
eight strings twice; classification returns reason codes with catalog strings
in both locales, so the detail sheet localizes; apps/api/.env.example now names
the same web port as apps/web/.env.example, without which a fresh checkout
fails sign in on the origin check.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
