diet-guard/docs/DOCS-kuchnia-wikinga-phone.md
Krzysztof kuhy Rudnicki 26d6a68ac0 Due-mode log screen, Claude button, photos
Action row 1A, due mode, catering auto-log rows with undo, relay interface, photo input.

Co-Authored-By: Claude Sonnet 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
2026-11-17 21:26:13 +01:00

4.6 KiB

Kuchnia Wikinga on the phone

The Dart half of the catering import. Read kuchnia-wikinga.md first — the API map, the traps and the delivered-is-not-eaten rule all live there and are not repeated here.

app/ carries a Dart mirror of the whole walk, so the feature works with the PC switched off. The files map one-to-one:

Python Dart
_kuchnia_client.py services/kuchnia_client.dart
_kuchnia_orders.py services/kuchnia_orders.dart
_kuchnia_parse.py services/kuchnia_parse.dart
_kuchnia_spread.py services/kuchnia_spread.dart
_kuchnia_import.py services/kuchnia_import.dart
_gatelock_kuchnia.py services/kuchnia_queue.dart (fetched-today marker; the form prefill is gone)
_kuchnia_log.py (fill_plan, log_dishes) services/kuchnia_fill.dart
_gatelock_fillall.py services/kuchnia_fill_flow.dart (plan, no UI) + services/kuchnia_autolog.dart

Auto-log (replaces "Fill all" on the phone)

kuhy overrode "delivered is not eaten" on 2026-11-17 ("LOG EVEN WITHOUT TAP"). When the log screen opens or resumes with a slot due, KuchniaAutoLog.logDue fetches today's delivery and logs the dishes assigned to the due slots, no tap. The Fill all button, its confirm tap and the delivery prefill are gone from the phone; FillAllFlow stays as the tested plan logic.

  • Plan = Fill all's plan minus the tap: assignSlots by the caterer's mealPriority against the user's schedule, fillPlan over occupied slots, then only the due slots. Written by logDishes, source kuchnia wikinga.
  • Deterministic ids. UUID v5, namespace 6ba7b811-9dad-11d1-80b4-00c04fd430c8 (RFC URL), over kuchnia:<delivery date YYYY-MM-DD>:<dishKey>; dishKey is the panel's deliveryMealId as a string, else the dish's bank key. A repeat open and every other device derive the same id, so the merge keeps one entry. tests/fixtures/kuchnia_autolog_ids.json (made by Python's uuid.uuid5, scripts/build_kuchnia_autolog_fixture.py) pins it for both languages.
  • Undo is a tombstone and it sticks. logMeal(id:) and logDishes(idFor:) skip an id that exists in the full log, deleted entries included; otherwise undo plus the next resume would log the dish again.
  • The fetch is cached. KuchniaQueueService.refreshOnce returns nothing after the day's first clean fetch, so a slot falling due at 12:00 would never see its dish. KuchniaAutoLog keeps the day's dishes in memory and retries a failed or empty fetch at most every 10 minutes (a login per try).
  • Each logged dish shows as a row: restaurant glyph, name, kcal and macro dots, a 48 dp undo. One publish per batch (publishAfterLog).
  • Unverified: whether deliveryMealId is stable across fetches on the live panel (the fixtures carry scrubbed synthetic ids).

Android only. The panel sends no CORS headers, so a browser cannot call it; refreshDelivery checks kIsWeb first and returns a reason. The settings fields still show on web — the desktop app is where a password is most likely to be typed, and the credential syncs from there to the phone.

package:http has no cookie jar, so the SESSION cookie is read off the login response and replayed by hand. Response.headers folds duplicate set-cookie values into one comma-joined string, so the parser splits on both , and ;.

What the parity gate covers, and what it does not

tests/fixtures/kuchnia_day.json is one payload with one expected result, read by both diet_guard/tests/test_kuchnia_parity.py and app/test/kuchnia_parity_test.dart. It pins the dish list, the dropped dishes, the slot assignment, the bank keys and the encoded records. A second fixture, kuchnia_credential.json, does the same for the credential adapter.

Three traps it exists for, each of which passed review once:

  • _ENERGY_TOLERANCE is 0.35, not the "~1%" quoted above for the captured day. Porting 0.01 would drop dishes the PC keeps, and then each device re-adds what the other dropped.
  • jsonEncode(435) emits 435 where Python emits 435.0. Every Dart macro is declared double, and the fixture compares canonicalised encodings, not just values — 435 == 435.0 is true, so a value-only test misses this.
  • Dart's List.sort is unstable; Python's sorted is stable. The Dart comparator carries the payload index as a final tiebreak, and the fixture contains two dishes sharing both a priority and a name.

The gate proves the two implementations agree on a fixed payload. It does not prove they agree against the live panel on a given day; that is a separate, on-device observation.