CLAUDE.md was 1101 lines and loaded into every session in this repo. Three
passes, in order: automate what a gate can check, rewrite what is left as
instructions rather than rationale, extract what still does not fit.
Automated. Five prohibitions were enforced by nothing but a paragraph asking a
future reader not to do something; they are now repo_invariants_test.dart:
- no app/linux/ (a GTK embedder target manages ~20fps at 4K)
- dart:io only in *_io.dart seams and lib/desktop/ -- this one is why the
test exists at all, since dart:io does not fail a web *compile*. It throws
at runtime, so a bad import is a blank white window that the widget tests
happily pass through.
- no meal builder / repeat-last-meal / reward prompt
- no image_picker, no CAMERA permission
- wrapper port 8732 and the diet-guard-desktop Chrome profile, both pinned
because IndexedDB is keyed by origin and moving either hides the whole
local food log
Each guard was canary-tested by breaking the invariant and confirming the
suite goes red -- a guard that cannot fail is decoration.
Rules whose behaviour was already covered (budget epoch seeding, sync merge,
food bank) shrink to a pointer at the test that owns them.
Where a rule's entire value was its rationale ("the removal is the feature"),
the why moved into the name and reason string of the gate that now enforces
it, so it lives somewhere executable instead of somewhere quotable.
Extracted. The vendored flutter/flutter rules were 789 of the 1101 lines --
72% of the file, none of it about this repo. Split into five docs/ topic files
under the cap and linked. Verified content-complete: 664 non-blank lines in,
664 out, zero missing.
Python 692 tests / 100% coverage unchanged; Flutter 548 -> 555 (+7 new).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01MK3HCZrM75sRHL1gL8usve
Meals logged on the phone stopped reaching the PC, so the gate kept locking
for slots that were already logged and the only way through was to spam
"nothing". The PC's fetch was never broken: it was faithfully pulling a
Firebase that did not contain the phone's meals.
The phone had no working publish backstop on either transport:
- `checkAndNotify` gated its sync on `due.isNotEmpty`. Logging every meal
promptly means nothing is ever due, so the ~15-min periodic task never
published. It was a notification job that happened to sync, not a sync
timer. It now syncs on every tick it is asked to, then computes due slots
from the merged result, so a meal pulled from another device suppresses a
nag instead of triggering one.
- `backgroundSyncPush` never called `initSyncDeviceId()`. A fresh isolate has
its own static state, so `currentSyncDeviceId` fell back to the compile-time
'phone' constant and published to `devices/phone/` rather than the persisted
uuid -- which then makes the next foreground push see `unchanged` and skip
writing the uuid directory entirely.
Separately, on the "Fetch from sync" path itself: `FirebaseSyncError` and
`FirebaseAuthError` are siblings of `GitHubSyncError` under `RemoteSyncError`,
not subclasses, so every per-request failure of the *primary* backend escaped
`pull_shared_log`'s catch tuple and raised a traceback out of a helper
documented to fail closed with a reason string. `ConfigError` needed the same
treatment by a different route: it subclasses `Exception` directly, so it is
outside `RemoteSyncError` too, and is now translated at `_client_for_run`.
The PAT is no longer mandatory either. Firebase is primary, but `_read_token`
ran before the client was ever built, so a Firebase-only machine could not
sync at all -- the commits that fixed this touched only the Dart side.
A stalled sync was invisible: the config gates returned silently and
`openFirebase` swallowed every error. `dart:developer`'s `log()` does not
reach logcat on a release build (verified on device), so logging alone would
not have surfaced it. `SyncHealth` records each tick's outcome and the log
screen shows a banner when publishing has actually stopped.
Also make the real-Tk probes bypass the window manager. A tiling WM retiles a
managed toplevel instead of honouring the requested geometry, so `winfo_height`
reported 0/1px and the layout and focus-ring assertions failed on noise rather
than on any regression -- 8 failures on the dev machine that only passed under
Xvfb. `overrideredirect` makes the measurement deterministic with or without a
WM; the probes still detect a genuine overflow at 1366x300.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RT8X32S9yDLW1Xf1PDq8z7
diet_guard pushed as "pc" (Python), "phone" (Android) and "desktop" (web).
A role constant collides the moment a second machine takes the same role,
and after a reinstall the new install inherits the previous one's CRDT
identity -- the same reasoning home_inventory already records for choosing a
uuid. Both halves now mint a uuid once and persist it: the PC in
~/.local/share/diet_guard/.device_id, the app under SharedPreferences
'crdt.nodeId' (the key todo and home_inventory already use).
The old constant survives as the legacy id and is passed to sync_log /
syncLog, so devices/<legacy>/ is skipped as this device's own. Without that
every tick would re-download and re-merge this device's own pre-migration
log as though a peer had written it -- idempotent under LWW, but hundreds of
KB, 96 times a day. It drops to None once that path is reclaimed; a test
covers both states, including a negative control proving the skip is the
legacy id's doing.
All four documents move together (food_log, food_bank, food_bank_manual,
budget), and the hist:<date> fields riding on the budget record are
untouched -- merge_record unions field names, so they relay through
unchanged.
The resolver lives in its own module (_device.py / sync_device_id.dart)
because _sync_merge needs it for HLC stamping too and importing _sync there
would be a cycle. Dart resolves it asynchronously at startup, before
LogStorageService.init(); currentSyncDeviceId falls back to the compile-time
role constant until then, so a too-early Hlc stamp gets a valid id rather
than crashing.
conftest redirects the new id file in the same change that introduces it --
without that a test mints a uuid into the real state dir and this machine's
live sync identity is decided by whichever test ran first. Verified by
checking ~/.local/share/diet_guard/ before and after: untouched.
Verified live, not just against mocks: the sync timer ran mid-change and
migrated this machine to f8f46da6. All 358 food_log records and all 100
food_bank records are present under the new path (0 missing vs devices/pc/),
and the new log's HLCs carry only its own id -- confirming it did not
re-merge the old path as a peer.
Python: 687 tests, 100% branch coverage. App: 518 tests, analyzer clean.
Pins bumped to crdt-sync-v0.6.0 / crdt_sync_dart-v0.7.0 (pyproject,
requirements.txt, .pre-commit-config.yaml, pubspec), and the new library is
installed into /usr/bin/python's user site-packages so the production timer
does not ImportError on its next tick.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GZH89WEqAjuJT8WZdqc7AP
The calendar already said which individual days went over budget, but not
whether the week as a whole did. A run of small overages reads as four
harmless yellow squares; averaged, it is the thing that actually moves the
needle.
Three rules are load-bearing, and are documented in CLAUDE.md because getting
any of them wrong silently changes what "under budget" means:
- The denominator is logged days, not elapsed days. An unlogged day is not a
zero-kcal day, and averaging gaps in as zeros would make a badly-tracked
week look like the best week of the year.
- The yardstick is the mean of the per-day budgets over those same days,
resolved through BudgetSchedule -- never today's budget, which would
retroactively reclassify past weeks on every budget edit.
- Today is excluded. A half-logged today drags the mean far enough to flip
the band, so every period ends at yesterday.
Band boundaries reuse the calendar's own OVER_BUDGET_YELLOW_CEILING by
import rather than by copy, so a period and its days can never disagree
about what "over" means; that constant loses its underscore to say so.
The MCP get_averages tool returns the average and the band but drops the
average budget, keeping the module's existing invariant that the budget
number never leaves this machine. The CLI, the gate and the phone all show
both, since it is hidden from the network, not from the user.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01DjJqpcEsaV5WRhPXyxCThd
Four asks, plus the fixes two review passes turned up.
Removals (all unused):
- "Repeat last meal" button, and the reward/temptation-bundling prompt it
was the sole trigger for. Replaced with a post-log progress card showing
kcal against budget, macros, and the adherence streak.
- The multi-item meal builder, on both the gate and the phone. The stored
`components` field stays: nothing writes it now, but historical composite
entries and the food bank still read it and it is part of the wire format.
- Meal photos entirely: 13 files, `imagePath`, the photo filter, the
wrapper's /blobs/ route, image_picker, and the CAMERA permission.
Reminder false alarms had three real causes, all fixed:
- Logging never reconciled notifications, so a reminder sat on screen up to
15 minutes after the meal was logged -- on the phone itself.
- The background check never pulled, so a PC-logged meal nagged forever
until the app was opened. Now mirrors _cli_gate._should_lock: cheap local
check, network only when a slot looks due, then re-check.
- Meals before 08:00 / after 22:00 were attributed to no slot at all. New
slot_for_log/slotForLog clamps them to the first/last slot, byte-identical
across Python and Dart. Deliberately separate from elapsed_slots, which
would otherwise make every slot fall due at 23:00.
Budget history, so lowering the budget stops rewriting the past:
- Forward-only (effective_from, kcal) entries; each day is judged against
the budget in force then. Verified live: setting 2000 left a 36-day
adherence streak and the whole calendar pixel-identical.
- Rides as hist:<date> fields on the existing budget record. crdt_sync
merges per-field over the union of field names and both devices push the
merged log, so a device that predates the feature relays them untouched --
no coordinated rollout, no second document.
- The seed guard tests emptiness, not file presence: _sync can write an
empty history after merging with a pre-feature peer, and a presence check
would then never seed again, silently reclassifying every past day.
One shared source of truth for every piece of data:
- Body weight now travels as its own `weight` CRDT field. Inside `value` it
was collateral damage of whole-map LWW -- any phone budget edit deleted it,
and the protein target with it.
- Both halves of the food bank sync. The curated half (foods added without
eating them) was the last device-local state. The derived half merges with
count as the clock, so LWW means max-count-wins; the log stays
authoritative for membership, because a CRDT union never shrinks and a
peer's stale copy would otherwise resurrect a deleted food permanently.
- Bank adapters honour tombstones, and Dart parses bank records defensively:
a cast failure there is an Error, not an Exception, so one malformed
remote record would have broken phone sync permanently.
_test_guard.py blocks any pathlib write to real user state while pytest is
in the process, whatever conftest is loaded. A throwaway test run outside
the package conftest destroyed a real GitHub token during this work; this
makes that class of accident impossible rather than merely discouraged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
main.dart now builds MaterialApp's theme from buildAppTheme() instead of
ColorScheme.fromSeed(Colors.teal) -- theme.dart existed but was never
actually used until this commit. photo_viewer_screen.dart,
settings_screen.dart, and day_status_calendar.dart's raw Colors.*
literals (Colors.black, Colors.red, Colors.white38, Colors.grey.shade800,
...) now read from Theme.of(context) instead.
Also rewrites CLAUDE.md's Visual Design & Theming section, which
previously contradicted the unified-design-system skill (noise texture,
multi-layered shadows, ColorScheme.fromSeed -- all things that skill
explicitly forbids).
This was the other prerequisite the on-fill/warm-palette fixes actually
needed: theme.dart existing but unwired meant the app was still running
Flutter's default seeded theme.
Completes the move: `linux/` is gone, replaced by the web build served by the
wrapper in a Chrome `--app` window. The icons and the .desktop installer move
out of `linux/` first -- they are not Flutter scaffolding and are still
needed; the entry now points at the packaged launcher and matches the
`--class=diet_guard_app` the wrapper passes, so the taskbar shows the app's
icon rather than a browser's.
install_arch.sh carries over the three traps each of which produced an
install that looked fine until launched: `dart build cli` rather than
`dart compile exe` (which refuses any package whose dependency graph has
native build hooks), `options=('!strip')` (makepkg would otherwise discard
the AOT snapshot, leaving a binary that runs as the bare Dart VM printing a
usage message), and no browser in `depends` (this machine's policy uninstalls
`chromium` seconds later, failing dependency resolution).
Also adds the frame-cadence instrument the performance claim is measured
with, armed by `--dart-define=DIET_GUARD_FRAME_STATS=1` and read with
`String.fromEnvironment` -- `bool.fromEnvironment` only accepts the literal
`true`, so `=1` would evaluate to false and a whole run would silently
produce no data. It reports frame-to-frame cadence rather than build/raster
durations, since work that fits the budget says nothing about the frame rate
actually achieved.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Both platforms gain a calendar tab/screen (green/yellow/red/not-logged
day coloring, logging/adherence streaks, year-to-date tally), and
tapping a day drills into that day's logged entries. This requires the
daily budget to become a normal, freely-editable value synced between
devices like the food log — replaces the previous chattr +i seal
(deliberately, per user decision) with plain last-edit-wins sync via a
new budget.json alongside food_log.json. The budget field defaults to
2200 kcal, starts read-only, and unlocks via an Edit/Save toggle on
both the PC gate's History tab and the phone app's Calendar screen.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_016dHac1iYQKTMZEtL7xDJvA
gatelock and diet-guard-sync were folded into shared monorepos:
gatelock -> github.com/kuhyx/utils (subdirectory), diet-guard-sync ->
github.com/kuhyx/syncs (subdirectory). Update the gatelock pip URL,
SYNC_REPO_NAME, the devices-dir path prefix (PC and phone), and add a
persisted-value migration so an already-configured phone picks up
'syncs' on next load.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Installs the official flutter/skills and dart-lang/skills packs into
app/.agents/skills/ and appends Flutter's AI rules.md to CLAUDE.md, so
Claude Code has task-specific playbooks alongside the dart MCP server
(registered separately at user scope).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_013NqCvrbFnoNNmqCwZKntBK
Pulls every other device's pushed log from GitHub-backed dumb storage,
merges it with the local log, and pushes this device's own merged copy
back -- the PC half of the diet-guard-app sync plan.
- _sync_merge.py: pure union-by-id merge, tombstone always wins, legacy
(time, desc) dedup for pre-id entries. Commutative and idempotent.
- _sync_github.py: minimal GitHub Contents API client (list/get/put),
distinguishing a 404 on an unused path from the repo itself being
unreachable.
- _sync.py: orchestration -- pull, merge, re-sign every persisted entry
regardless of origin, write, rebuild the food bank, push. Re-signing
unconditionally is load-bearing: an unsigned phone-origin entry would
otherwise be silently dropped on the very next read once a machine
holds the shared HMAC key.
- _foodbank.rebuild_food_bank(): the "replay a full log into a fresh
bank" entrypoint the Python side was missing (the Dart port already
had its equivalent). Backs sync's bank-rebuild step.
- New diet-guard-sync.service/.timer (15-minute cadence, headless, a
separate unit from the gate so a held lock can't stall sync) and a
new install.sh step to install them.
- Created the private kuhyx/diet-guard-sync GitHub repo for storage.
Incidental to this feature: adding the `sync` subcommand pushed _cli.py
past the repo's 500-line cap, so `gate`'s CLI glue moved out alongside
sync's into _cli_gate.py/_cli_sync.py -- same split pattern already used
for the gate window logic itself, not a sync-specific design choice.
338 tests, 100% branch coverage. Verified importing and running cleanly
under /usr/bin/python (the production interpreter), not just the dev
venv -- the gap that caused the earlier 3-day outage.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FU3f5KQ1GHXsbbSecfVEyF
A one-off `pip install "diet_guard @ git+..."` from a scratch
directory was used during the original cutover; that's a non-editable
snapshot that a later git push silently never reaches. Document the
correct deployment convention (install.sh from a durable ~/diet-guard
clone) so this doesn't drift again.
Rewrites python_pkg.diet_guard imports to diet_guard, vendors the
shared as_float coercion helper, drops the monorepo PYTHONPATH from
install.sh and the systemd unit (package is now pip-installed), and
scaffolds standalone lint/test config matching testsAndMisc's real
enforced bar (pylint --fail-under=10 with tests excluded and the
use-implicit-booleaness/consider-using-with disables, mypy's actual
disabled-error-code set, ruff ALL, bandit, 100% branch coverage).
