A star button beside the log button sends the typed text to
estimate_meal and logs the result like a typed meal: the locked slot,
source "claude", food bank, unlock. Icon states only (idle star, spinner,
warning); no visible text. It is composed (ClaudeFlow), not a new MRO
class, since the gate chain is at pylint's max-parents cap.
Idle uses U+2726, not the sparkles emoji: U+2728 is only in FreeSerif and
Noto Color Emoji here, and a colour-emoji fallback can crash Tk 8.6.
Failure paths (all fail-closed: never unlock, block or freeze):
- estimator error, no login, bad JSON, any exception: warning icon,
nothing logged, manual log path untouched.
- no answer in 45 s: the poll gives up (warning icon); a late answer is
discarded.
- slot already logged by hand meanwhile: the result is dropped.
- empty description: error status, no call.
- seam: the estimator name is patched to a raising stub in conftest; a
mistyped target errors every test at setup (verified once).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
kuhy's 2026-11-17 override of "delivered is not eaten": before the lock
opens, the PC logs the due slots' dishes by id uuid5(kuchnia📅key),
shared with the phone (fixture-pinned).
Failure paths:
- fetch error, no credentials, > 5 s: fail-closed, nothing is logged,
the lock opens as before, reason emitted and logged.
- log write error (OSError/ValueError): fail-closed, same.
- gate --check: never logs (a decision, not a write).
- an undone (tombstoned) entry counts as logged: fail-closed against
resurrecting a meal kuhy removed.
- publish outage after a clearing auto-log: fail-open by design, the
local entry stays and the reason is emitted.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
The narrow pull, the log write and the publish each read, merge and write
the food log. Run at once (a publish in flight while a second request
logs) they lose the append: the phone is told "logged" and the only copy
is gone. The new race test lost an entry before this change (1 of 2) and
keeps both after it; the shared temp-file name of the log writer also
collided between threads.
One asyncio lock now covers the pull, "resolve slot + log" and the
publish. A request arriving during a publish waits for it (up to one full
tick); the model call stays outside the lock.
Also: a test that the relay unlocks the header diet_guard's devices store
wrote (same key, a blob sealed by one opens on the other), the skill
returns grams 0 for nothing to estimate (parse fails closed, nothing is
logged), and the doc's session note is corrected: the pull and publish
refresh the gate's session file as the CLI's publish does.
Failure paths: a lost append is now impossible in-process (fail closed by
waiting); the pull failure, locked key, estimator and cleanup paths are
unchanged.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
python -m diet_guard claude-worker serves the phone's meal requests from
RTDB (rtdb_relay RequestServer, one SSE stream, reconnects with backoff):
validate, estimate with Claude, log through the signed log with
source=claude under the oldest due slot (normal mode: nearest slot of the
phone's hint), write the sealed result, delete only the request payload,
then publish in the background. diet-guard-claude.service (Restart=on-failure,
WantedBy=default.target) is installed by install.sh and started once the
worker's own Firebase session exists. Nothing is deployed by this commit.
Failure paths:
- narrow peer pull fails (sync locked/unreachable): fail closed, no claim, no
answer, no model call; the request stays pending, retried every 30 s
- e2e key locked, no passphrase, no session: fail closed, nothing read or
written; the worker waits and retries every 60 s
- bad body, request older than 24 h (or no created): fail closed, an error
result, nothing logged
- Claude failure/timeout/invalid JSON: fail closed, an error result with
retry, nothing logged
- crash between log and result, or a duplicate event: idempotent, the entry
id is uuid5 of the rid and is looked up (tombstones included) before the
model call
- request payload delete, old-answer cleanup and the post-answer publish
failing: fail open on purpose (logged only); the answer and the local
entry are already durable
- due mode with nothing due: falls back to slot_for_log, not an error
New log helpers log_meal_as / find_entry; log_meal's body is split into
_entry_fields / _store_entry with the same output. The gate and CLI import
none of the worker's dependencies at load time.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
estimate_meal(text, photo) -> Estimate runs Claude through
claude_agent_sdk in an isolated CLAUDE_CONFIG_DIR: a copy of the access
token (no refresh token, no MCP tokens) and the diet-food skill, never
kuhy's ~/.claude. The skill text is the system prompt; the closest curated
and derived food-bank rows ride in the user message; the photo is an image
block; the answer is constrained by a JSON schema and validated again.
Failure paths, all fail closed (an EstimateError code, nothing logged):
- no host login, CLI/API error, bad transport: estimate_failed
- more than 45 s: estimate_timeout
- no object, missing/mistyped field, NaN/inf/negative/absurd number,
grams 0, blank or long desc: estimate_invalid
- neither text nor photo: ValueError before any model call
- failed token refresh is logged, the call then fails as estimate_failed
Live smoke (scratch HOME, nothing logged): "2 jajka sadzone i kromka
chleba" -> 315 kcal in 4.3 s.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
Dependencies (exact pins): claude-agent-sdk 0.2.183, rtdb-relay 0.1.1,
httpx 0.28.1, hypothesis 6.170.15 (dev). The pylint hook env gets the
same pins.
Entry points, each reachable only through one module attribute:
_claude_connect.connect_sealed (RTDB + e2e unlock), _claude_sdk.run_model
(the model call), _claude_sync.pull_peers / publish (narrow pull, full
tick). SDK and rtdb_relay are imported only by these modules, never by the
gate or CLI at load time.
Test isolation, all fail closed: connect_sealed and run_model are replaced
by a function that raises; pull/publish return None; the sandbox dir and
kuhy's real Claude login path point into tmp_path; _test_guard also
protects ~/.config/diet_guard_claude and ~/.cache/diet_guard. Each of the
six redirects was removed once and its canary went red (fail closed).
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWZHWP4KZ6x5TcXDBFC
wait_for_turn sleeps its 1.5 s startup grace for real, so every test
that builds a gate paid it. With the grace zeroed in the hermetic
gatelock fixture the 1427 tests run in 11 s instead of overrunning
the 60 s action cap, so the ci-mirror-pytest gate step can finish.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWWfb63jncj9Lca2jeX
pre-commit --all-files failed the file-length gate on AGENTS.md (262),
test_sync_e2e.py (309) and test_sync_refresh.py (254), which would turn
CI red. AGENTS.md prose is rewrapped (same words), test_sync_e2e.py
moves its last three classes to test_sync_e2e_copy.py unchanged, and
two docstrings in test_sync_refresh.py are tightened.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWWfb63jncj9Lca2jeX
A meal logged on the phone while the gate window was up changed nothing
until Fetch was pressed: the new credit watch only fires on a local
write to the food log, and the PC never pulled on its own.
While a non-demo window is open it now pulls peer logs every 30 s, with
the Fetch button's own narrow pull on a worker thread. One pull is in
flight at a time (shared with the button), failures are logged and
retried, and the window stays up. The pull's write trips the watch,
which closes through the usual reconcile path. gatelock moves to 0.14.0
for watch_credit.
refresh_peer_logs also read the local log before the peer downloads and
wrote it after, so a meal logged during the pull was overwritten. It now
merges the peers first and reads the local log just before the write.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWWfb63jncj9Lca2jeX
Every sync payload (food log, banks, budget, catering credential,
Body, export keyring) now lives on diet-guard-sync/devices-e2e, sealed
with the Argon2id key from the shared crdt-sync passphrase; record
ids, tombstones and clocks stay clear. Old plaintext nodes are copied
once and never written or deleted. Locked = no pull, no push.
- PC: _sync_e2e.open_e2e over crdt-sync 0.11.0, key and copy memo in
~/.local/share/diet_guard/e2e; undecryptable peers are skipped.
- Phone/desktop: sync_e2e.dart, crdt_sync_flutter 0.8.1, lock icon
in Settings; pullExportKeys goes through the same open.
- wake-alarm weigh-ins read from wake-alarm-sync/devices-e2e (own
header), on both sides.
Todo: fda71ed1
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
No lock and no owed slot while holds.covers("diet") is true; slots
inside a hold are excused; an open lock lifts within 30 s of a hold
starting. Fail closed (normal rules) if holds is missing or raises.
Pin holds-v0.1.0; bump coverage 7.17.1.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzxdWVSoczW2NHQd887FUm
due_slots asks sickdays.stands_down(LOCKER_DIET, day) instead of
is_sick_day, so a sick day that keeps diet is a normal day for the gate.
Bump sickdays pins to v0.2.0.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01PAhbNszs5JKdtnD4uBi3xp
due_slots() returns nothing on a sick day, like a free day: no lock, no debt;
logging still works. Unreadable pool reads as not sick (fail closed). Pins
sickdays-v0.1.0 and freedays-v0.4.0 (pyproject, requirements, pylint hook env).
Phone gap documented in docs/DOCS-sick-days.md.
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Xfj1sxKH9VBPqpcAni8Jry
An inert dropdown arrow looked broken. ttk.Spinbox posts nothing; every
spin path (arrow press, Up/Down, wheel) generates <<Increment>>/
<<Decrement>>, bound here to step_time from the typed value (07:23 -> up
07:30, down 07:15) with "break" so ttk's values/wrap logic never runs.
First clamps to 00:00-23:45, last to 00:15-23:45. Popdown-blocking code
removed; FakeSpinbox replaces FakeCombobox.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015npsYVB2PdtWkG7MPAi4tn
- The popdown maps beneath the gatelock surface yet holds ttk's grab
(verified on Xvfb: invisible, swallows the next click, steals keys), so
arrow clicks are intercepted as Down already was. Typing and Up/Down stay.
- Catering dedup compares the entry's recorded minute (parity with the
app); occupancy (fill_plan) still snaps.
- Test: a corrupt peer slot (slot 24, slot_min 5000) re-signs and reads
without raising.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015npsYVB2PdtWkG7MPAi4tn
Every Python consumer now speaks slot minutes:
- satisfaction is nearest-slot snapping (_slot_wire.satisfied_by_entries)
at every "is this slot logged?" site: gate, CLI status, MCP, catering
occupancy and dedup. Nothing on disk is rewritten; 07:00/10:00 entries
keep covering 07:15/10:15 after a schedule move.
- log_meal writes slot_fields(minute) before signing (slot + slot_min).
- schedule store and sync encode via schedule_to_wire/from_wire; a
whole-hour schedule stays byte-identical (f/l/n only).
- CLI ate --hour takes HH or HH:MM; MCP log_meal takes "HH:MM" or a
legacy int hour, validated against today's slots; get_slots returns
{minute, label}.
- gate History tab: first/last are HH:MM ttk.Comboboxes (quarter-hour
list, typed times, Up/Down step without posting the popdown), locked
as disabled; validation mirrors normalized().
- gate timer fires every 15 min; docs rewritten for minutes.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015npsYVB2PdtWkG7MPAi4tn
Meal-slot checkpoints move from whole hours to minute-of-day ints
(0..1439) in the pure model and slot arithmetic of both languages:
MealSchedule(first_minute, last_minute, count) / (firstMinute:,
lastMinute:, count:), interior slots snapped to the 15-minute grid,
endpoints exact, integer-only. New: nearest_slot, satisfied_slots,
slot_fields / entry_slot_minute (additive "slot_min" wire field) and
schedule_to_wire / schedule_from_wire ("fm"/"lm" only when off the hour,
so whole-hour data stays byte-identical), parse_hhmm.
tests/fixtures/meal_schedule_vectors.json (scripts/
build_meal_schedule_fixture.py) is read by both suites; kuchnia_day.json
slot values regenerated as minutes.
Consumers (CLI, gate, storage, sync, app screens, kuchnia, notifications)
are intentionally broken until phase 2. patch-targets skipped: it fails
only because diet_guard._gatelock no longer imports.
Co-Authored-By: Claude Opus 5.5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_015npsYVB2PdtWkG7MPAi4tn
LockConfig's colours moved into a nested LockPalette record; every
_COLORS.<colour> read follows. Nothing about the gate's behaviour changed.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HHCbkJWXHjtTUaqvnHRAqA
Third app onto freedays (~/utils/freedays), the shared pool of
pre-declared days when every gate app stands down. due_slots() returns
nothing on a free day, so no slot is ever due, the lock never opens and
nothing is said about the missing meals.
Read from one local file, never the network, so this cannot hang the
systemd-fired gate -- and an unreadable pool reads as "not free", which
means the normal rules apply rather than the gate quietly switching
itself off. Tested explicitly: fail closed, not open.
The conftest redirect is load-bearing. The pool lives under
~/.local/share, outside every diet_guard path the suite already isolates,
so without it a real free day on this machine would fail every "the gate
is due" test here for a reason that looks nothing like the cause.
gatelock moves v0.7.2 -> v0.7.3, and the pylint hook's own pins (which
had drifted further, to gatelock v0.7.1 and crdt-sync v0.6.0) come with
it. Not cosmetic: pip refuses two different direct URLs for one
distribution, so any stale pin alongside freedays -- which requires
v0.7.3 -- is an unresolvable environment rather than a warning.
requirements.txt also picks up pylint 4.0.8, pytest-randomly 5.0.0 and
ruff 0.16.6, which the freshness gate flagged once this file was touched.
1043 tests pass at 100% coverage across the pytest-randomly major bump.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
`requires-python` and the CI pin already said 3.14, but ruff, mypy and pylint
were still grading against 3.10 -- three tools judging an interpreter nothing
runs any more, and ruff withholding the pyupgrade fixes that only become safe
at the newer target. `pin_ci_python.py` rewrites exactly these three knobs.
The fixes ruff then had to make are what the stale target was hiding: 18 files
on `datetime.timezone.utc` instead of `datetime.UTC` (UP017), plus ruff's
stable formatter applying PEP 758 -- `except (A, B):` becomes `except A, B:`.
That second one makes the source 3.14-only at the syntax level, which is
exactly what `requires-python = ">=3.14"` already promised; kuhy chose it
deliberately over holding ruff a release back (2026-08-29).
ruff, ruff-format, mypy --strict and pylint all clean at the new target.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01SfnJmxxsZ534AKk16rC2Gr
The Dart half of the credential sync: a `KuchniaCredentialService` that owns
the value locally, a `sync_merge_kuchnia.dart` adapter mirroring the Python
one, and a `_syncKuchniaCredential` tick alongside the banks and budget. The
phone can now receive the panel password, which is the prerequisite for it
fetching anything itself.
One source of truth, per the answered question: the value lives in the synced
document store and nowhere else. Deliberately *not* mirrored into
`flutter_secure_storage` -- two homes for one value is how a stale password
survives a change made on the other device, and the credential has to be
readable synchronously at merge time anyway.
Gated by a second shared fixture, `tests/fixtures/kuchnia_credential.json`,
asserted by both suites. Two cases carry the weight:
- an **unparsable** edit time must fall back to the same clock on both sides
(Python catches `ValueError` from `fromisoformat`, Dart's `tryParse` returns
null). Verified they agree on `wall_time_ms == 0`; mutating that expectation
fails both suites.
- a **blank half** contributes nothing, so a half-filled settings form cannot
push an empty-string password that wins LWW against a real one.
`KuchniaCredentialService.init()` is called in the WorkManager isolate as well
as `main()`. A fresh isolate has its own static state and `_syncKuchniaCredential`
no-ops when the singleton is missing, so without it a background tick would
neither publish this device's credential nor receive a peer's -- and on the
phone the background tick is most of the syncing there is. Same class of trap
as the `initSyncDeviceId()` one already documented there.
`.toDouble()` came off the bank record: the fields are already `double`, so it
was a no-op the analyzer flags. The field *types* are what keep the encoding
byte-identical to Python's, and `kuchnia_parity_test.dart` asserts that.
654 Dart tests, 1040 Python tests, 100% Python branch coverage.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QQpCiPaSbyRpwsLkH3dgsX
`_local_credential_log` fell back to the hand-written `kuchnia_credentials`
and used its **mtime** as the edit clock. That clock is not trustworthy:
`git checkout`, a backup restore, and re-running the `install -m 600` setup
line all bump it forward without the credential changing.
Reproduced before fixing: a PC whose file had merely been touched, syncing
against a phone that had set a new password an hour earlier, resolved to the
PC's stale password. The phone user would have seen their edit silently revert.
The bootstrap is still needed -- someone has to publish first, and the PC is
where the password is typed -- so it moves to `_bootstrap_credential_log` and
runs only when the merge came back completely empty. It can no longer outrank
a real edit from a peer; it can only fill a vacuum. Once anything has been
published, the synced cache (which carries a real edit time) takes over
permanently.
`test_a_touched_handwritten_file_cannot_clobber_a_peer` pins it, and was
checked against the pre-fix code: reintroducing the fallback fails it.
Also rewrites the two remaining places that still claimed a credential "must
never leave the machine it was entered on" -- `_constants.py`'s comment and
the doc's Credentials section, which now has a table of which of the three
files syncs and which does not. Rewritten rather than given an exception
clause.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QQpCiPaSbyRpwsLkH3dgsX
The phone is getting its own Kuchnia importer, and it cannot fetch anything
until the panel password reaches it. So the credential syncs, and a phone that
has been wiped and reinstalled gets it back without the user digging out the
original.
**It travels in plaintext.** Nothing here encrypts it and the rest of the
synced state is not encrypted either, so the docs no longer say a credential
"must never leave the machine it was entered on" -- that paragraph is rewritten
rather than given an exception clause. What still never leaves is the session
cookie: it is regenerable from the password, so syncing it would widen exposure
and buy nothing.
Its own document, not a field on `budget`. `budget.json` is written back at
default permissions, so a password riding inside it would be readable by
anything that reads the budget. Two independently-clocked fields, so a device
that only ever set the username cannot blank a peer's password.
The hand-written `kuchnia_credentials` keeps its contract -- written by the
user, never by this package -- and now acts as a local override that wins over
the synced copy, plus the bootstrap for the very first push. Without that
fallback no device would ever publish, since the synced cache starts empty on
every device.
Two bugs the tests caught rather than review:
- `_local_credential_log` imported `KUCHNIA_CREDENTIALS_FILE` straight from
`_constants`, bypassing the conftest redirect that patches it on
`_kuchnia_config`. The suite was reading the real ~/.config credentials. It
now goes through the module attribute, which is what the redirect replaces.
This showed up only as a coverage hole -- the branch was unreachable because
the real file existed.
- `log_to_credential` guarded `password_hlc is None`, which cannot happen: a
field's value and clock arrive as one tuple. Dead code, removed rather than
covered by a contrived test.
The relay canary is a real test now, not a scratch script: a field from a
future release survives merge and re-push untouched, so this is shippable
without a coordinated release. `PASSWORD_FIELD_NAME` is spelled as a join
because ruff's S105 reads the literal as a hardcoded secret and the repo bans
`noqa` outright; the wire name is still "password".
100% branch coverage, 1036 tests.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QQpCiPaSbyRpwsLkH3dgsX
The phone is getting its own Kuchnia Wikinga importer so the feature works
with the PC switched off. That turns a risk `_kuchnia_spread.py` called
"theoretical" into a live one: two implementations must now agree exactly on
which dishes are dropped, which slot each lands on, and the JSON a banked
record encodes to. Disagreement means each device re-adds what the other
dropped, and since `add_manual_entry` restamps `t` unconditionally, the whole
curated bank republishes to every peer on every refresh. A slot mismatch is
worse -- a checkpoint one device offers and the other does not can never be
satisfied.
So the port lands with its gate rather than after it. One shared fixture,
`tests/fixtures/kuchnia_day.json`, is asserted by both `pytest` and
`flutter test`: same input, same expected dishes, drops, slots, bank keys and
records. Two suites written independently from the same prose is not a gate.
The fixture is five real dishes (a scrubbed 2026-08-22 capture) plus synthetic
meals that each pin one drop or fallback path -- a per-100 g mix-up, an absurd
portion, a stringly-typed number, a non-dict entry, a missing `nutrition`, a
missing `mealPriority`, and a twin-dish pair. Without those, "the same meals
are dropped" is a vacuous assertion against an all-happy-path capture. Both
halves were mutation-checked: flipping an expected slot, drop count or record
value fails each suite, and flipping `435.0` to `435` fails the byte-identity
test while the value comparison still passes, which is why both exist.
Three traps the port had to survive, each caught before it shipped:
- `_ENERGY_TOLERANCE` is **0.35**. The prose said "~1%", which describes how
closely the captured day agreed, not the threshold; porting 0.01 would drop
dishes the PC keeps. That paragraph now says so.
- `jsonEncode(435)` emits `435` where Python emits `435.0`, so every Dart
macro is forced to `double` and a canonical-encoding test pins it.
- Dart's `List.sort` is unstable where Python's `sorted` is stable, so the
Dart comparator carries the payload index as a final tiebreak.
The real dishes live in their own file rather than as a literal in the
generator: transcribing them by hand corrupted three of the five, and the
error was invisible until the generated fixture was diffed against the
capture.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QQpCiPaSbyRpwsLkH3dgsX
Walking a real 5-dish delivery through the gate by hand showed the
"(N more to go)" the lock promises was a dead letter: `_prefill_next_dish`
had exactly one caller (the initial poll), so logging the first dish cleared
the form and stranded the rest. Reaching them meant clicking "Today's
delivery" again -- and the button's fetch is unguarded, so that was a fresh
login-plus-three-request walk per dish.
`_finish_slot` now prefills the next queued dish, carrying the
"Logged HH:00 ..." confirmation into the same status line so advancing the
queue does not swallow it.
The caterer plans five meals against four default slots, so a dish is still
queued when the last slot unlocks. It is already banked, so `_unlock` names
it rather than dropping it silently.
Also pins the phone conclusion that was previously reasoning only: a verbatim
capture of a real imported bank now round-trips through `FoodBankRecord` in
`app/test/kuchnia_bank_interop_test.dart`, covering the diacritics and the
`t` stamp, and the casefold/toLowerCase agreement the two devices depend on.
Verified against the live panel with state redirected to a scratch HOME
(real ~/.local/share/diet_guard byte-identical before and after): the
autoload prefills silently on window open, a failed autoload stays silent
while the button reports, the after-log warm-up runs off-thread (0.62s vs
0.28s guarded) and nothing is logged until "Log & Continue".
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QQpCiPaSbyRpwsLkH3dgsX
Pulls the day's delivered dishes and their real per-portion macros out of
panel.kuchniavikinga.pl, so a catering meal is one keypress instead of five
retyped numbers.
A delivered meal is not an eaten meal, so nothing is ever logged unattended:
dishes land in the *curated* bank, and entries are written only after an
explicit confirmation. Auto-logging would let the gate satisfy its own
checkpoint from a delivery note, and would record what the courier dropped off
rather than what was eaten.
Three surfaces:
- `diet-guard kuchnia [--log] [--yes]`
- the lock screen's "Today's delivery" button, which prefills the meal form
dish by dish (the user still clicks Log & Continue)
- a warm-up after each logged meal, on the existing background thread
All event-driven; no new systemd timer. The two automatic paths go through
`refresh_delivery_once`, guarded to one fetch per day.
The panel is an undocumented white-labelled Dietly API; the endpoint map was
recovered from its JS bundles and confirmed against live responses. Notes,
including the traps, are in docs/kuchnia-wikinga.md.
Load-bearing details:
* Macros are per portion and `weight` is that portion in grams -- proven, not
assumed: 4P+4C+9F reproduces the stated calories within ~1% on every meal of
the captured day. `_energy_is_consistent` re-runs that check per dish and
drops any that fails; a per-100g mix-up is invisible in a payload and would
silently skew every logged meal by a factor of several.
* `bank_dishes` compares nutritional fields before calling `add_manual_entry`,
which restamps `t` unconditionally. Without the compare, every refresh
republishes the whole curated bank to every peer. Tests assert on call count.
* Duplicate log suppression reads today's synced log, not a local marker:
`log_meal` mints a fresh uuid, so a re-import after marker loss would merge
as duplicates on every peer.
* The gate hook is in `on_focus_ready`, not `_cli_gate._should_lock`:
`gate_is_due()` never reads the food bank, so a refresh there could not
change the lock decision while adding a third-party round trip to the ~105ms
`gate --check` fast path.
* `requests` is lazy-imported so the gate's not-due tick does not pay for it.
* Credentials and the cached session cookie live in ~/.config/diet_guard at
mode 600, outside the synced tree. The credentials file is written by hand,
never by the package -- the same contract as sync_token.
`_GateDashboard` is merged into `_GateMealFlow`, and both "pull it in from
elsewhere" flows share one `_PullFlows` mixin: pylint's max-parents counts
every MRO entry, so the gate's controller chain had no room for a new class.
Also adds a side-by-side row for the two pull buttons -- stacking them cost
42px and pushed the tab past the viewport at 1366x768.
989 tests, 100% branch coverage, pre-commit clean.
Pruning 20 dead devices made the narrow pull *slower* -- ~90ms to ~15s --
which is the opposite of the point, and I caught it only because I measured
after running rather than assuming the win.
`_device_paths` listed the four data files a device publishes but not its
entry under `revs/`. `_sync_refresh._candidate_peers` deliberately enumerates
peers from the revision map instead of calling `list_directory` (that call is
~445ms mirrored), so every orphaned marker resurrected a deleted device as a
candidate and bought a round trip to discover it was gone. 19 orphans, 19
wasted probes, every tick.
The marker is now part of the device's path set, so it is backed up and
deleted with everything else, and a regression test pins it by name rather
than by count. The 19 already-orphaned markers on the remote were cleaned up
separately.
Verified after the cleanup: narrow pull ~84ms, full tick 8.9s (was ~27s),
gate --check ~105ms, and the log is untouched at 441 entries / 117 foods with
peer_revs correctly down to 3.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
Every install mints a uuid and pushes under it, and nothing ever cleans those
directories up. This remote holds 25: two live devices and 23 that stopped
writing. Each dead one still costs a round trip in the log pull and in each of
the three bank syncs -- roughly 400ms per peer per tick -- so they dominate
what is left of the background tick.
Diagnosis first, since the obvious theory was wrong. These are not the residue
of phone reinstalls: seventeen of them are frozen at the *same* instant
(2026-08-13T16:45:04) holding near-identical ~400-record copies of one log,
which is the signature of repeated fresh-install runs, not organic use. That
also means `adb install -r` preserving SharedPreferences was never the issue,
so no Android backup change is warranted on this evidence.
Dry-run by default, and the safety rails are the point:
* `plan_prune` only reads; it is what the default run prints.
* A device is prunable only when every record it holds is already held by a
device being kept. Against the live remote that is 20 devices and **zero**
records lost -- verified before writing the command, not after.
* `pc`/`phone`/`desktop` are never proposed: they carry pre-migration history
and `SYNC_LEGACY_DEVICE_ID` still names one as this device's own former path.
* `--apply` refuses without `--backup-dir`, and every peer's raw bytes are
written to disk before a single delete.
Nothing has been deleted; this commit only adds the tool.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
`diet-guard log ...` blocked for ~15.5s after every meal. The publish itself
has to happen -- until it lands the phone still believes the slot is unlogged
and nags for it -- but only ~2.5s of that is the 293KB push the phone needs;
the rest is pulling peers' banks, which publishing does not require.
The local write is already durable before the publish starts, so the worst
case is a late publish, which is the trade `_sync_events` already documents
for the event-driven model. Moving it to a worker thread lets the caller
print and return at once. Measured: 15.5s -> 2ms before the prompt comes back.
The thread is deliberately non-daemon: the CLI's output is already flushed,
so the process lingers only as long as the push, and the publish is never
killed halfway. A failure still surfaces, via a callback the caller supplies,
just after its own output rather than in the middle of it.
Deliberately NOT applied to the gate. `diet-guard-gate.service` is
`Type=oneshot`, so systemd reaps the unit the moment the main thread exits
and a background publish would be killed mid-push. The gate keeps calling
`publish_after_log` and waiting, which is correct there -- its publish
already runs after the window is gone, where nobody is waiting on it.
`conftest._isolate_state` follows the rename to
`publish_after_log_detached`; without that every meal-logging test would
reach the network through a newly unpatched seam. Verified with
`scripts/check_patch_targets.py` and by timing the real command against a
temporary data dir, never the live log.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
Two independent costs in the background tick, both pure waste.
`list_directory` was called four times per tick -- once by the log pull and
once by each of the three bank syncs. It is the single most expensive call
in a tick at ~445ms mirrored (the GitHub half alone is ~360ms), so that was
~1.3s spent asking the same question four times. It is now listed once in
`run_sync` and passed down.
The curated-bank loop was worse: 17s and 62 requests to fetch *zero* useful
bytes. Firebase returns None for every peer's `food_bank_manual.json`, so
`MirrorSyncClient.get_file_text` treated each miss as "this device has not
migrated yet" and fell through to GitHub, which answers `{}` after ~320ms.
`_primary_only_text` skips that per-peer fallthrough for this one artifact.
Checked before narrowing rather than assumed, because CLAUDE.md requires
both halves of the food bank to sync: reading every peer's curated bank
straight off the mirror found 23 absent, 2 empty, and zero non-empty. No
curated food exists on the mirror alone. The probe is written into the
docstring so it can be re-run before anyone widens this again.
Measured end to end: manual-bank loop 17.0s -> 3.4s, full tick ~27s -> 15.2s,
log/bank/peer state unchanged (441 entries, 117 foods, 22 peers).
`_pull_remote_logs` now returns the revisions it saw instead of mutating a
caller-supplied dict, which keeps it under the argument limit and makes the
data flow visible at the call site.
Deliberately NOT done: revision-skipping the derived food bank. `run_sync`
rebuilds that bank from the merged log immediately before syncing it, so
unlike the log the peer records are not durably held locally -- skipping a
peer would drop its counts from the pushed bank and oscillate under
max-count-wins. Verified empirically: the on-disk bank and a pure log-derived
rebuild are identical (117 foods, 0 differing counts), and the 2 foods peers
know that the local bank lacks are discarded by the existing `if name in
local` filter anyway. The loop is off every interactive path now, so the
remaining cost is background-only.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
The button ran a full sync tick inline on the Tk thread, so the fullscreen
lock was dead for its whole duration -- ~18-27s measured against the live
remote, and unbounded in the worst case.
Two changes, and the second is what makes the first cheap:
* the pull moves to a daemon worker thread, with the result marshalled back
through a queue polled on the Tk thread. The worker touches no widget, by
construction: Tcl is not thread-safe and that includes `after`. Polling
also makes the completion path testable with no thread at all.
* the button now runs the *narrow* pull. Its question is identical to the
gate's -- "did a peer log this slot?" -- and `_reconcile_after_fetch`
reads only `due_slots()`, so the budget and food banks it used to fetch
were never consulted here. The full tick still runs after the window
closes. Measured 84ms against the live remote.
Narrowing also shrinks the write race from tens of seconds to imperceptible:
the pull persists its merge, so a meal logged mid-fetch could be overwritten
by a snapshot taken before it. A second click while a fetch is in flight is
now ignored, which is the reentrancy guard and the race fix in one.
That guard started out as a disabled button, which is how it surfaced a
latent bug worth recording: the fetch button widget is never stored on
`GateWidgets`, so `self._widgets.fetch_button` did not exist and the disable
was a silent no-op -- hidden because the suppression tuple included
`AttributeError`. Coverage would have reported 100% on a guard that did
nothing. Guarding on `_fetch_result` needs no widget at all.
Also adds `INTERACTIVE_TIMEOUT_SECONDS = 2.0` for the paths a user waits on.
`SYNC_TIMEOUT_SECONDS` only ever reached GitHub: crdt_sync v0.6.0's
`firebase_client_for` takes no timeout, so Firebase used its 15s default and
one hung session refresh stalled the lock window -- the journal has five such
refresh failures in five days. Verified to fail closed to the local decision.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
The gate's pre-lock refresh ran a *full* sync tick -- peer logs, budget,
both food banks, a rebuild and a push -- in front of the fullscreen lock.
Measured against the live remote: ~18-27s, of which the curated-bank loop
alone was 17s spent fetching nothing (absent on Firebase for every peer, so
each of ~21 peers fell through to a GitHub round trip returning `{}`).
It was also failing: five `sync unavailable (network error trying to refresh
the session)` entries in the last five days, one of them a meal that logged
locally but never published.
The gate has one question -- "has a peer already logged this slot?" -- and
only the peer logs can answer it. `_sync_refresh` does that and nothing
else: one bulk revision-map read, a fetch of only the peers whose revision
moved, merge, re-sign, write. No budget, no banks, no push. The full tick
still runs, after the window closes, so nothing is permanently skipped --
the exhaustive pass just moves off the path the user waits on.
Three state invariants have their own regression tests, because each is
silent when broken: `pushed_rev` survives verbatim (this pass never pushes,
and overwriting it would make the device stop publishing forever),
`peer_revs` is merged rather than replaced (only changed peers are visited),
and the log is written before the state.
Separately, `gate --check` was importing an HTTP stack and a GUI it never
uses. `requests` (~78ms) arrived via `_cli_sync` -> `crdt_sync` and via
`_resolve` -> `_estimator_off`; both are now resolved through module-level
`__getattr__` (PEP 562) and `import_module`, which keeps every existing
`patch.object` target and needs no lint suppression.
Measured: `gate --check` 216ms -> ~110ms; narrow pull 92ms against the live
remote. Accepted trade, documented in `_sync_refresh`: a budget edited on
the phone is not visible in the lock dashboard until the post-window tick.
It cannot affect the lock decision, which is why it is safe to defer.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01LJ5YnA3CERitQ57HjMu2Qz
The second edit surface, so the schedule is editable from either device the
way the budget already is. A row on the History tab beside the budget row,
with the same Edit/Save toggle: first hour, last hour, meals per day, and the
derived checkpoint times underneath so the effect of a change is visible
before it is saved.
Validated rather than clamped. MealSchedule.normalized() would happily turn
"20 to 8" into a two-meal day ending at 09:00, which is a sensible defence
against a peer's corrupt sync data and a terrible response to a typo at the
keyboard. Each rejection says what is wrong and leaves the row in edit mode so
the value can be corrected instead of silently becoming a different schedule.
The narrow-window case has its own message, because that one is genuinely
non-obvious: six meals between 08:00 and 11:00 would round two checkpoints
onto the same hour, and a duplicate slot silently drops one.
Three structural notes:
- The row's widgets are their own module. _gatelock_calendar_widgets.py was at
216 of the 250-line cap and is now 233; adding the row inline would have
breached it.
- The five new StringVars are grouped in a ScheduleVars dataclass rather than
flattened onto CalendarVars, which pylint flagged at 11 attributes. Grouping
two unrelated concerns separately is the better shape anyway.
- _gate_fixtures.py's _GATE_TK_MODULES is a hand-listed patch target, so both
new modules are registered there. Left out, they build real tk widgets under
a fake parent -- which is how it failed the first time, loudly, rather than
silently opening a window.
785 tests, 100% branch coverage.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159b4E5hh512jYUeaJVNxfM
Without this the schedule would be device-local, which is the failure this
whole feature has to avoid: the phone sets five meals, the PC still derives
four, and the PC's gate nags for a 12:00 checkpoint the phone never offered
and can never satisfy. So sync lands before either edit surface, not after.
The history rides the existing `budget` CRDT record as one
`sched:<YYYY-MM-DD>` field per entry, exactly like the budget's own `hist:`
fields. merge_record is per-field LWW over the *union* of field names and both
devices push the merged record, so a device that predates the feature relays
the fields untouched -- no coordinated release. There is a test for precisely
that: merging our push with a budget-only peer keeps both schedule entries.
Verified the two languages agree on the wire, not just in their own suites --
Python and Dart both emit
{"sched:1970-01-01":{"f":8,"l":20,"n":4},"sched:2026-08-16":{"f":8,"l":20,"n":5}}
The Hlc is derived from the *parsed* timestamp on both sides, which matters
because the epoch fallback formats differently in each language
(1970-01-01T00:00:00+00:00 vs ...000Z) and a string comparison would diverge.
Two guards worth naming:
- A device with no history contributes no `sched:` fields at all, so an unset
default can never outrank a peer's real schedule; and an empty merge result
is never written back, so a pre-feature peer cannot erase a real history.
- Dart's `_syncBudget` synthesises a local record when the schedule has been
edited but the budget never has -- `budgetToLog(null, ...)` returns an empty
Log and would otherwise drop the schedule fields too. Its budget value
carries an epoch stamp, so it loses every LWW race against a real edit.
sync_merge/_schedule.py and sync_merge_schedule.dart are separate modules
rather than additions to the budget adapters, which were at 186 and 170 lines
against the 250 cap. Each declares its own record-id constant so the Dart
pair does not import each other.
772 Python tests at 100% branch coverage, 619 Dart.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159b4E5hh512jYUeaJVNxfM
Wires the Python side to the configurable schedule. No edit surface yet, so
nothing writes a schedule and current_schedule() returns the default -- the
derived hours are still 08/12/16/20 and the cutoff is still 22:00. Verified by
running it: before any edit the slots are unchanged, after recording 08-20 x5
they become 08:00, 11:00, 14:00, 17:00, 20:00, and a 2020 day still resolves
to the four-meal schedule.
_slots.py now takes the schedule as a *required* argument rather than reading
module constants. Required, not defaulted, on purpose: a default would let a
call site missed in this refactor keep deriving the old fixed hours on one
device, and a slot one device offers while the other does not is a checkpoint
that can never be satisfied. Making it required let mypy enumerate the nine
call sites instead of trusting a grep.
The history is forward-only and seeds the default at the epoch *before*
recording today's value, so switching to five meals leaves past days judged
against the four-meal schedule they actually ran under. Same shape and same
reasoning as _budget_history.py; it will ride the same CRDT record.
Two details worth naming:
- MEAL_SCHEDULE_FILE lands in DATA_DIR, so conftest's _isolate_state redirect
is patched in this same commit rather than a later one. Confirmed with a
canary that the redirect is live and that ~/.local/share/diet_guard/ has no
.meal_schedule after a full run.
- _mcp_read imports _budget.current_schedule already, so the meal-schedule
lookup is aliased rather than shadowing it.
_gatelock._pending_slots is spelled `is None` rather than `or` now. The two are
in fact equivalent -- if 0 is a slot it is necessarily the first slot, checked
exhaustively -- but the intent should not rest on that coincidence.
761 tests, 100% branch coverage.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_0159b4E5hh512jYUeaJVNxfM