Add a global, temporal ai_notes list appended to the classifier prompt
(seeded once from no-PII defaults, documented in README), managed inline
on a new AI tab with a read-only view of the assembled prompt. Every
AI-run upload records the browser-round-tripped suggestion blob + model;
misreads are derived (final field != AI guess) and reviewed one by one
(image + per-field guess-vs-entered + notes-since), attributing which
note fixed each or closing unresolved. Update SPEC (new section 10),
DESIGN item 15, README, and changelog (0.0.3).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Bake the EXIF Orientation rotation into uploaded JPEG pixels (and strip
the tag) so receipts are upright in every consumer, not just EXIF-aware
viewers. Acts only when orientation is known (tag 2..8); images with no
tag, tag 1, non-JPEG, or PDFs pass through byte-for-byte. Wired into the
receipt, attachments, and AI-classify paths. Documented as spec item 13.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Features (see spec.md v2):
- Wire receipt classification into the upload flow; cheap model (Haiku 4.5)
is now the default, shown as a footnote with per-scan cost in cents.
- Skip-AI toggle to enter fields by hand.
- Duplicate-transaction warning: live check on date+amount, gated submit.
- Tally tab: person x year totals with margins and grand total.
- Recent uploads / recent receipts tabs with paging and file serving.
- People reconcile on startup: merge stray partial names (e.g. "Jude" ->
"Jude Tremblay"), reassigning receipts; idempotent seeding.
- scripts/build.sh builds the binary; scripts/run.sh builds and runs with .env.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Go app for capturing and archiving HSA-eligible receipts: OIDC/PKCE auth
against Authelia, SQLite storage with dual-write (filesystem + DB blob),
mobile-first upload, and DB export.
Adds AI receipt classification: a config.json catalog of people and
categories (seeded into the DB on startup), a prompt builder that derives
name-order/initial variants from the data (with same-surname ambiguity
handling), and an Anthropic tool-use client behind POST /classify. Tests
run against a mock endpoint; a live integration test is env-gated to the
cheapest model.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
