Blender defaults to the AgX view transform, a photoreal film emulation
that rolls off highlights and desaturates. On the flat-lit item renders
the 3D route produces, it turned gold into beige and washed out every
material -- which is why row C read as unusable in the bake-off.
Measured on an identical coin scene where only view_transform differed:
mean saturation 0.282 (AgX) vs 0.787 (Standard), a 2.8x increase.
This makes row C a fair test, not a good one: the geometry is still
low-poly primitives and ring/bone remain nearly invisible at 32px.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T9rQ12yZ7sJxiPQTCS5dUF
Closes the remaining bake-off cells. The phase's real result was not a route
score but a MISSING GATE.
All twelve SDXL icons measured 100% opaque coverage -- square tiles with no
sprite silhouette -- and passed every gate in the harness. Gate-pass was
measuring pipeline hygiene, not usability. MARGIN_TOO_SMALL inverts the table
from a meaningless 12/12/12 tie to:
diffusion 0/12 (every output full-bleed)
vector 10/12
procedural 11/12
The two non-diffusion failures were true positives in MY shapes, not gate
error: the procedural sword pommel ran off the bottom edge and the vector meat
bone clipped the top. Both fixed; both routes back to 12/12.
Diffusion recovers to 12/12 with flood-fill background removal, but visual
inspection shows the fill eats into the art (bone reduced to a fragment, sword
nearly erased). Reported as inadequate rather than shipped -- proper matting is
needed before route (b) is a usable pipeline.
Blender exits 0 even when its embedded script raises: BLENDER_EEVEE_NEXT does
not exist in 5.2, and the failure produced a clean exit and no file. render()
now asserts the output exists, or the route would report phantom successes.
Model gating verified at execution time, never recalled: SDXL base and
nerijs/pixel-art-xl are ungated, FLUX.1-dev is gated=auto, and
runwayml/stable-diffusion-v1-5 now 307-redirects.
The finishing stage now generalises across three continuous-tone sources
(SVG raster, Blender render, SDXL): each is 0/N raw and N/N finished.
Verified: 114 tests, 100% branch coverage on 13 modules, zero suppressions,
all hooks pass. Note: a local pytest run masked a missing branch that the
pre-commit hook caught -- the hook's run is authoritative.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T9rQ12yZ7sJxiPQTCS5dUF
Closes the two CPU-only cells of the art bake-off and builds the bridge the
diffusion cell will reuse.
Pins the bake-off to the SAME twelve subjects item-icons measured, with its
easy/hard split intact. Without this a route tested on distinctive silhouettes
beats the 6/12 baseline for free and the decision table means nothing. It also
makes the subject-routing rule falsifiable: a generator route is predicted to
lift specifically the six human-prior failures, and if it does not, the rule is
recorded as wrong.
Route (a) procedural: all twelve subjects drawn deterministically from a
7-colour locked palette. Every shape gets a distinguishing feature up front,
because item-icons measured that plain primitives read as placeholders however
correct the geometry is. Machine gate: 12/12 -- which proves "not malformed",
NOT quality. A procedural generator satisfies palette/alpha/canvas constraints
trivially; the human contact sheet is the real verdict.
Route (d) code-as-art: SVG emitted as text and rasterised by rsvg-convert, so
no new Python dependency (cairosvg/svgwrite stay absent).
The structural finding, measured on real rasteriser output rather than a
synthetic stand-in: raw vector renders pass 0/5 -- every one fails
ALPHA_NOT_BINARY, most also TOO_MANY_COLORS. After the finishing stage, 5/5.
Continuous-tone routes are therefore GENERATORS FEEDING A FINISHER, not
pixel-art generators. finish.py binarises alpha before quantising colour,
mirroring the gate order: quantising first blends transparent pixels' arbitrary
RGB into the palette.
Verified: 92 tests, 100% branch coverage across all ten modules, zero
suppressions, all hooks pass. The PLR2004 lint refactor was confirmed
pixel-neutral by byte-comparing all 12 icons before and after (12/12 identical).
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T9rQ12yZ7sJxiPQTCS5dUF
Adds the route-independent spine of the planned art-generation bake-off: a
harness that decides whether generated art is well-formed, so "non-sloppy"
becomes machine-checkable instead of a judgement call.
Findings that shaped the design, all measured rather than assumed:
- Integer-scale invariance separates true pixel art (impurity 0.0000) from
interpolated fakes (1.0000) exactly -- but passes trivially at scale 1, so
it must never be trusted alone on native-resolution output.
- Gate ORDER is load-bearing. A soft-alpha image has no fully-opaque pixels,
so an opaque-only colour count returns 0 and silently passes. ALPHA_NOT_BINARY
now runs first and COLOR_COUNT fails closed on an empty opaque set.
- Counting must be over opaque RGB only: transparent pixels retain arbitrary
RGB and inflated a naive RGBA count, rejecting correct art at 18/16.
- Declaring a gate was not enough to run it. hand_painted.toml shipped
declaring SILHOUETTE_LOW_CONTRAST with no implementation, so it parsed, ran
nothing, and reported a pass. load_spec now fails closed on unimplemented
gates, result-only codes, and gates missing their config.
Thresholds are calibrated against 250 genuine 32x32 CC0 icons, not guessed:
p50=7 p90=13 p99=26 opaque colours; cap 32 rejects 0.0% of real art while raw
generator output measured 34403 colours. An initial cap of 16 wrongly rejected
a real DCSS icon -- re-derived from the corpus rather than nudged.
Verified: 57 tests, 100% branch coverage, zero suppressions, all hooks pass.
58/60 (96.7%) of real CC0 icons pass the full stack; both failures confirmed
true positives.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01T9rQ12yZ7sJxiPQTCS5dUF
The enforcer failed in both directions: nothing blocked at home, and
everything blocked at the office. The comparison was never inverted --
`metres <= threshold` was always right. Three separate causes:
Only ACCESS_COARSE_LOCATION was declared. Android 12+ fuzzes coarse to a
~1-2 km grid against a 150 m fence, so a fix taken at home routinely landed
outside it, read as AWAY, and restored every app the geofence exists to hide.
FINE is now declared and self-granted by the device owner, along with coarse
and background, re-asserted every pass so a reinstall cannot silently drop it.
There were three states but two behaviours. LOCATION_UNKNOWN applies the
identical sweep to AT_HOME -- correct fail-closed design, indistinguishable
from outside. And nothing ever requested a fix: the pass read whatever another
app happened to cache, so at the office with no navigation running the cache
was empty. Passes now acquire a fix inside a location-typed foreground
service, bounded at 20 s, and treat anything older than 30 min as no fix. That
window is deliberately longer than the 15-minute cadence; a shorter one marks
the newest cache stale on every pass of an idle phone and blocks everything.
None of it was diagnosable. `adb logcat -d -s FocusOwner:V` came back empty
while dumpsys showed 82 alarms had fired, and `run-as` is refused on the
release build device owner requires. So the pass now writes a ring-buffered
JSONL record read back in-app: reason, distance vs the fence, fix
age/provider/accuracy, curfew state, and every hidden app with the rule that
hid it. Coordinates are never recorded -- a distance is enough to debug a
geofence, and this is a home address.
Also fixed along the way:
- Hysteresis keyed off `isEnforcing`, which is true under AWAY too because
YouTube stays hidden, so the threshold was permanently radius+hysteresis.
It now keys off whether the previous pass was inside the fence.
- The scheduler's curfew-boundary alarms were documented but never scheduled;
both callers invoked `scheduleNext()` bare, so 05:00 could land 15 min late.
- Installing over the app cancelled its pending alarms, so `adb install -r`
silently stopped enforcement until someone opened the app. MY_PACKAGE_REPLACED
now re-arms it, verified on device without launching the app.
- A wrong-typed log record threw instead of being skipped, and the exception
escaped far enough to leave the page stuck busy -- which disables the
release-device-owner button, the one control that must work when everything
else has gone wrong.
- `lib/enforcement.dart` was a second decision layer with no always-blocked
concept, cited by KDoc as pinned by a `PolicyParityTest` that never existed.
It would have made the new screen report YouTube visible while it is hidden.
Deleted; the UI renders the Kotlin record rather than re-deciding.
- Anchoring home accepted a stale or unknown-precision fix. A wrong home
offsets the fence everywhere and forever, which is what produced a 766 m
reading while standing at home.
Policy: Play Store is allowed by day (not during curfew) so infakt can be
installed and updated; the sweep is default-deny for third-party packages, so
anything installed from Play is hidden on the next at-home pass. Tachiyomi's
per-source extension apks are covered by a new prefix allowlist rather than an
exact list that goes stale whenever an extension is installed.
`export.py` stopped subtracting the allowlist from `blockable_system_packages`.
Sweepable is not hideable: that list decides which system packages are
eligible for a decision, and `isAllowed` is what protects them. Removing an
allowed system package from the sweep freezes it in its last state -- it would
have stranded Play hidden permanently the moment it was allowlisted.
Verified on the Pixel 6a: FINE granted POLICY_FIXED, `applied AT_HOME` with a
0 m distance from a +/-11 m GPS fix, YouTube/Chrome/Firefox absent from
`pm list packages -e`, Play Store, infakt, RunnerUp and all five Tachiyomi
packages present.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HjxL9kdAzZmGqQA5tMNrZ8
The network layer was the wrong place for this. Tested remotely today: opening RethinkDNS and switching off its device blocklists un-blocked YouTube in a few taps, and no device owner API can stop that -- an on-device filter app always owns its own rules.
App hiding is different. 'pm unhide' from adb shell throws a SecurityException and the package stays hidden; only the device owner can reverse it, and reaching that costs a factory reset. So the block moves there, scoped to an explicit allowlist and the home geofence.
Both whitelists are rewritten from the apps that actually exist on this phone. The old ones described the rooted Blackview. Every browser is excluded now, so at home a link does not open at all -- confirmed as intended.
LAUNCHER_PACKAGE was still the Blackview's launcher, which is not installed here. The exporter refused to render the policy twice over it -- precisely what that check is for.
The network layer stays a soft deterrent by choice: pinned and uninstall-blocked, but without lockdown, and its blocklists remain switchable.
Private DNS is implemented but ships disabled. stunnel logged zero connections from the handset across two pin attempts, so the router does not forward 853.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014DqEd1uUhyT52KjaNeLanC
Both were measured, not theorised. `pm uninstall --user 0
com.celzero.bravedns` succeeded while the VPN was pinned and
DISALLOW_CONFIG_VPN was set -- I ran it to check whether the provider
was protected, and it removed the app, took the DNS filter with it, and
left the always-on pin aimed at a package that no longer existed. The
second hole was quieter: without lockdown, stopping the tunnel let
traffic flow normally, so the filter only held while it was running.
So the provider now gets setUninstallBlocked, re-asserted every pass,
and the pin runs with lockdown. Lockdown is a real hazard and is
deliberate: if the VPN app breaks, the device has no connectivity until
ownership is released. That is only survivable because the release path
lives in the enforcer app and needs no network.
releaseDeviceOwner reads the pinned package before unpinning it, then
unblocks its uninstall. Reversed, getAlwaysOnVpnPackage returns null
after the unpin and the provider would stay unremovable forever -- the
same trap the self-uninstall ordering already avoids.
Home coordinates are provisioned. The pass now logs "applied AT_HOME"
rather than LOCATION_UNKNOWN, so enforcement is finally geofenced rather
than permanent-everywhere, driven by a real fix.
`pm uninstall --user 0` on the provider now returns
DELETE_FAILED_OWNER_BLOCKED.
What this does NOT fix: RethinkDNS's blocklists are gone with the
uninstall, since they live in app-private storage. The filter is running
with no rules and youtube.com resolves normally until they are
re-entered. And nothing here stops those rules being switched off from
inside the app -- no device owner API can.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014DqEd1uUhyT52KjaNeLanC
Hiding the YouTube package stops the app. It does nothing about
youtube.com in Firefox, in a webview, or in any client nobody has
enumerated. This is that layer.
setAlwaysOnVpnPackage from the device owner is not the mechanism that
failed on 2026-08-09 -- that was `settings put secure always_on_vpn_app`,
which writes successfully, survives a reboot, and is ignored by
VpnManager. The DPM call takes effect: dumpsys now reports
mAlwaysOnVpnPackage=com.celzero.bravedns.
Lockdown stays off. It drops all traffic whenever the tunnel is down,
and the wipe checklist names it together with the DISALLOW_*
restrictions as the combination most likely to leave an unrecoverable
device. DISALLOW_CONFIG_VPN is deferred for the same reason: locking VPN
configuration before the filter is confirmed working would pin a broken
setup in place.
The exporter emits always_on_vpn_package only when the provider is also
in never_disable_prefixes, and RethinkDNS was added there. Pinning a
package the sweep can hide is the worst case -- the restriction would
point at a hidden app and cost general connectivity, not just the
filter. The pin is re-asserted every pass and, unlike the uninstall
block, does not track the decision: the network filter must not lift
when the geofence says AWAY.
releaseDeviceOwner now unpins the VPN and clears the restriction before
clearing ownership, for the same reason it already unblocked uninstall.
Not yet active, and the commit does not pretend otherwise: the tunnel
needs a one-time in-app VPN consent tap that adb cannot grant, and
RethinkDNS is at first-run state because its wildcard rules live in
app-private storage that the backup never reached. They have to be
re-entered by hand.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014DqEd1uUhyT52KjaNeLanC
Leaving the house was an off switch. The AWAY branch returned early with
packagesToShow = every installed package, so walking 150m from home
restored YouTube -- which is the exact affordance device owner was
provisioned to remove.
always_blocked_packages is exempt from all four branches: AWAY hides
them, the allowlist loop cannot show them, and the workout release skips
them, because otherwise "go for a run" is the off switch instead. Chrome
is included as a second route to the same content; Firefox stays, and
carries the uBlock filters. Play is deliberately left geofenced so apps
can still be installed away from home -- and a hidden package cannot be
reinstalled from Play anyway.
infakt moves from the allowlist to SYSTEM_NEVER_DISABLE. Allowlisting
only held while it stayed in BOTH the day and night lists, and dropping
it from either would silently make it hideable; isAllowed() consults the
never-disable list before the curfew split, so it now holds under every
condition. It is device-paired to a bank over SMS, so losing it strands
the same re-authentication chain a hidden Messages app would.
The exporter now refuses an always-blocked package that is not also
blockable: the runner filters unopted system packages out of
installedPackages, so such an entry would never be hidden and nothing
would say so.
Verified: 128 Python tests at 100% branch coverage, 16 Kotlin tests, the
new ones canary-checked by breaking an assertion and confirming the
suite actually fails. On device, youtube/music/chrome/vending are
hidden=true, infakt hidden=false, uninstall still refused.
The AWAY path itself is unit-tested but NOT verified on hardware: with
no home_location.json every pass decides LOCATION_UNKNOWN, so the branch
is unreachable on this phone until coordinates can be provisioned.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014DqEd1uUhyT52KjaNeLanC
The Device Owner sweep skipped every FLAG_SYSTEM package, which meant it
could not touch the apps it exists to block: YouTube ships at
/product/app/YouTube as a system app, and so does Chrome.
Deleting the filter is not the fix. Measured on the Pixel 6a, that would
expose 320 system packages of which 243 match no allowlist entry and no
never_disable_prefix -- including com.android.cellbroadcastreceiver
(emergency alerts), com.android.credentialmanager and
com.android.devicelockcontroller. So system apps stay default-deny and
blockable_system_packages is the opt-in.
com.android.vending is on the list deliberately: leaving Play reachable
makes every other removal a one-tap undo.
The list lives in the Python exporter rather than config.sh because it is
a safety boundary, and a shell typo should not be able to widen it. The
exporter subtracts anything already allowed, so an asset can never state
that a package is both allowed and blockable. Kotlin reads the field with
a new optionalStringSet: an asset predating the field parses to the empty
set instead of throwing, because a throw makes decide() skip the pass and
silently disable enforcement.
Verified on-device under a profile owner: hid 1 for YouTube, restored 1
when temporarily allowed, and cellbroadcastreceiver, credentialmanager,
devicelockcontroller and settings all left hidden=false.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_014DqEd1uUhyT52KjaNeLanC
Shared across every repo with an Android surface (screen-locker's workout_app,
todo, dufs-cloud/app). Works against a plain release APK on a physical device
over adb — no root, no emulator, no Flutter debug connection, no app changes.
Driving an app by tapping pixel positions read off a screenshot fails in ways
that look like success. All observed in one session (2026-08-10) verifying
workout_app's Firebase restore:
* A tap aimed at a password field landed with the field unfocused. The typed
password went nowhere and NOTHING errored — the next screenshot just showed
an empty box. type_into/type_into_field now re-read the field and raise.
* Opening the soft keyboard moved a field from y=1572 to y=1319, so a
coordinate captured one step earlier was already wrong. Every action
re-resolves the element immediately before acting.
* A widget behind the keyboard reports its LAID-OUT position (1549) not its
real one (1939) — a 390px error that taps a letter key and reports success.
keyboard_is_up() distinguishes "absent" from "behind the keyboard", and tap
closes the keyboard first.
* KEYCODE_BACK is a route pop in Flutter, not a keyboard dismiss: it navigated
out of Settings and discarded typed credentials, twice. dismiss_keyboard
tries ESCAPE then BACK and verifies via dumpsys input_method.
* `input text` inserts at the cursor, so typing over a filled field produced
"canary@example.com321krzychu@gmail.com". Fields are cleared first.
* uiautomator returns ONLY the focused EditText nodes while a Flutter text
field has focus, and stays that way across retries — so a naive retry loop
never converges. Partial trees are retried and the best result kept.
* An EMPTY EditText has no text, content-desc or resource-id, so a
"must be labelled" filter drops the very field you need to fill in. Editable
nodes are kept regardless and addressable by position.
Ambiguity is an error rather than a silent first-match: `tap "Back"` correctly
refuses when "OFFLINE BACKUP" is also on screen.
Its own test suite caught a real bug in _escape: backslash sat inside the
escape loop, so "&" became "\\&" and a password containing shell
metacharacters would have been typed wrong.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01JupKvZCNKRmEeRQY3G1unG
Indexes the arch-wiki-docs dump that is already on disk and exposes it
to the knowledge-rag MCP server, so Arch questions can be asked by
symptom ("why doesn't my systemd user unit see DISPLAY under i3")
rather than by page title, with no network round-trip at query time.
The one Arch-specific MCP server in the wild fetches pages over the
network and bundles pacman operations that duplicate the local yay-mcp.
This does neither: it only searches, against a local corpus.
install.sh installs its own dependencies, wires the MCP config and
registers a pacman hook so the index follows package updates. The
generated index is machine-specific and regenerable, so it is
gitignored rather than committed.
_reindex_entry runs under the knowledge-rag interpreter, where
mcp_server is importable and this repo's packages are not. Resolving
that import with importlib.import_module states the deferral outright
instead of a function-level import statement that reads as an oversight,
and the backend failure path now re-raises as SystemExit, so the
exit-code translation is explicit rather than a swallowed exception.
Both replace lint suppressions; the package carries none.
pylint runs in its own venv, so beautifulsoup4 and markdownify are
declared there too — being in meta/requirements.txt does not make them
resolvable to the hook.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQLwdquDCNmWtvgi3Lk72J
TachiyomiSY 1.13.2 can report a restore as SUCCESS while writing almost
nothing, then sync that near-empty library up and replace the only good
copy. On 2026-08-09 a 14 MB payload holding 2182 manga became a 267 KB
stub, and the real library survived only as stale pages in the SQLite
write-ahead log.
The upstream bug is Kotlin in an APK and cannot be fixed here, so this
targets the fixable half: decode the payload, count what is actually in
it, and refuse to treat a collapse as normal drift. A pure is-it-zero
check would have caught this incident but not the harder case upstream
describes, where a restore keeps some entries and drops the rest, so the
threshold is proportional. Categories get their own rule because all
2182 manga survived the round trip and all 18 categories did not.
Adds the setup script and a runbook covering the WAL-page salvage used
to recover the library on the day.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQLwdquDCNmWtvgi3Lk72J
setApplicationHidden drops a package from the default PackageManager
enumeration even though it stays installed=true, so a sweep querying
with flags 0 could never see anything it had already hidden. That made
packagesToShow permanently empty: measured on device, a pass following
one that hid three apps reported "hid 0, restored 0" and left them
hidden with no way back. Query with MATCH_UNINSTALLED_PACKAGES instead.
Hiding the enforcer itself is the other unrecoverable case, since it
takes the escape hatch and the trigger with it. The exporter now injects
com.kuhy.focus_owner into both the day and night allowlists so the
decision layer cannot propose it, and the runner keeps its point-of-use
skip to cover a policy asset rendered before this fix or hand-edited.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01YQLwdquDCNmWtvgi3Lk72J
The new ~/untools app (guided thinking tools) needs a family launcher icon
rather than the stock Flutter beach one.
The glyph is one node branching into two -- "decompose a problem into parts",
which is the move every one of the app's 25 frameworks makes in some form.
Nodes are hollow rings rather than filled dots so the 72px stroke leaves
visible interior at icon size.
A two-level, seven-node tree was drawn first and rejected on the generator's
own evidence: it reported the glyph overflowing the 560px safe box by
~170x162px, and rendering confirmed the doubled connectors closed up into a
solid mass at launcher size. One clean split survives the mask and still reads
as a hierarchy. Branch spread was then narrowed to clear a residual 44px.
linux=False for this app: its desktop build is the web build in a Chrome
--app window, so there is no GTK app to hand a hicolor PNG to.
Also whitelists com.kuhy.untools in phone_focus_mode/config.sh. That entry is
source-of-truth only -- the current Pixel 6a is not rooted, has no
/data/local/tmp/focus_mode and no running daemon, so nothing was pushed to the
device and no app is being killed there. Kept for the rooted BL9000 phone.
Verified: 44 app_icons tests pass with the package at 100% coverage; the
generator emits 6 assets with no safe-box warning; the rendered icon was
inspected and matches the family; and the resulting APK installed and stayed
alive on 23181JEGR08034.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_019hVZ5HfcrcrVzinmjQuQup
Three rising bars with a plotted point above the tallest: a day-by-day log
that adds up to a trend, which is what the app tracks. Bars are strokes
rather than filled rects so the glyph keeps one stroke weight, and the
168px pitch leaves 96px gaps -- comfortably above MIN_NEGATIVE_SPACE, which
is where these glyphs usually fail.
kuhylog has a Linux desktop target as well as Android, so it is registered
with linux=True and gets the hicolor set.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01HQtqicQHSrvBa4XaNNh9Wy
OctoForge (~/octoforge) is app #10 in the shared launcher-icon family. It was
still shipping the stock Flutter icon, which is not a finished state.
The glyph is a filled silhouette rather than a stroked outline: a stroked
anvil is a long thin horn plus a narrow waist, and at the family's 72px stroke
weight the waist closes up solid at 48dp. An octopus (GitHub trade dress, and
a tangle of thin curves that blurs at launcher size) and a bare hammer (reads
as a generic tool, not a forge) were both rejected.
The horn is a separate wedge projecting off the face, not a rounded corner on
it. Two earlier drafts folded the taper into the slab's left end and both
rendered as a bump, leaving the shape reading as an I-beam. It now tapers to a
40px blunt tip, above MIN_NEGATIVE_SPACE so it survives at 48dp.
Registered with linux=True: unlike the other phone-first apps this one has a
real desktop target with a .desktop entry, so it needs the hicolor set.
Verified: generated for Android and Linux, checked against the family preview
sheet at all three sizes, and confirmed in the device app drawer next to todo.
Committed from a temporary worktree: the working tree held another session's
in-flight kuhylog entry in the same hunk, and git-rules.md forbids stashing
around other agents' uncommitted work. That work is untouched.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01WzEA5srk8VPn6mKKjCaGpb
Two gaps stopped the unrooted path working end to end. The phone had no way to
read the focus policy, and the blocklist was only reachable as a hosts file or
over plain UDP/53 on the LAN -- while the rooted enforcer actively rejected
every DoT connection, so a self-hosted resolver could not have worked even if
one existed.
focus_owner now loads a generated policy.json instead of hardcoding anything,
so config.sh stays the single source of truth across Python, shell and Kotlin.
The committed asset is rendered with --redact-home and a test fails if home
coordinates ever appear in it. An unknown schema_version is rejected outright:
silently accepting a newer schema means enforcing a policy the app has misread,
and a misread allowlist is one that blocks the dialer.
dns_enforcer.sh rejected ALL port 853 traffic and pinned Private DNS off, every
20 seconds. That is right when DoT is a bypass channel and exactly wrong when
DoT is the enforcement channel, which is what an unrooted phone needs -- it has
no hosts file to fall back to. With DNS_TRUSTED_DOT_HOST set it now ACCEPTs the
trusted resolver before the blanket REJECT, and pins Private DNS to that
hostname so the user cannot pick an unfiltered one. Rule order is load-bearing:
iptables takes the first match, so an ACCEPT appended after the REJECT would
never be reached. expected_rule_count() counts the new rules, because a
mismatch there makes chain_intact() rebuild the chain every 20s -- a silent
fork storm that also drops the connection each time. With the variable empty
the behaviour is byte-identical, so the rooted phone is untouched until this is
deployed.
setup_dot_resolver.sh terminates DoT with stunnel rather than Caddy: DoT is a
raw TLS stream and stock caddy:2.8 has no layer4 module, so using Caddy would
have meant an xcaddy rebuild of a container serving five live sites. It reuses
the certificate Caddy already renews, so nothing competes for port 80. It binds
to the WireGuard address, not 0.0.0.0 -- a public DoT endpoint is an open
resolver, and over the existing tunnel it needs no port-forward, no firewall
rule and no rate limiting. Verified: example.com resolves, pornhub.com returns
0.0.0.0, TLS 1.3 against the LE cert, and both the LAN and public addresses
refuse.
The Device Owner checklist is planning only. Checking the device rather than
the config settled the one open risk: Google Wallet is not installed and no NFC
payment default is set, so the unconfirmable tap-to-pay question does not apply
here. The same check found that Aegis, MS Authenticator, Oracle IDM and PKO IKO
are absent -- they are in config.sh's whitelist, which describes the rooted
phone, not this one.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AwepyL5Lxs5QKbKkMpxxxL
phone_focus_mode is built for a rooted phone: the hosts layer is a read-only
bind mount over /system/etc/hosts, chosen so that even root-in-a-terminal
cannot overwrite it. None of that ports to a stock locked device. This splits
out the parts that are device-independent so one policy can drive both.
python_pkg/focus_policy parses config.sh rather than copying it, so the shell
config stays the source of truth and the two cannot drift. It rejects the
shipped REDACTED_* coordinate placeholder instead of defaulting to (0,0),
which would move "home" to the Atlantic and quietly disable the location gate.
blocklist.py converts the hosts file to domains for a DoT resolver or a local
VpnService. Three properties of the real 194k-line file make the obvious parser
wrong, and all three are handled: generate_hosts_file.sh *comments out*
deliberate unblocks rather than deleting them, so stripping '#' re-blocks
facebook/messenger/linkedin/4chan; the StevenBlack header carries loopback and
IPv6 housekeeping that would blackhole localhost; and ~1060 upstream annotation
tokens ('#[adware.zeno]', '"spam') render as address=/"spam/, which makes
dnsmasq refuse to start -- taking DNS down rather than blocking anything.
Verified with dnsmasq --test over all 185681 generated domains.
strip_workout_hosts.py had the same class of bug on the live rooted phone:
exact set membership meant the allowlisted googlevideo.com never matched the
dynamic r*---sn-*.googlevideo.com hosts YouTube actually streams from, so the
workout unblock was only partly working. Now suffix-matched (17 lines stripped,
was 14). The three new tests were checked against the original code in a
throwaway worktree; two of them fail there, so they pin the fix rather than
passing vacuously.
focus_owner is a device-owner app that deliberately does nothing yet. It ships
clearDeviceOwnerApp() implemented and tested first, because dpm
remove-active-admin does not work on a device owner: an app that locks the
device down before proving it can let go leaves a soft-brick recoverable only
through fastboot on a locked bootloader. Built, installed and launched on the
Pixel 6a; dumpsys afterwards confirms Device Owner Type -1 and no user
restrictions, i.e. the phone is unchanged.
Measured while scoping this, and not yet reflected in any enforcement code:
on Android 16 both pm suspend and pm disable-user are cleared by a reboot, so
shell-level app blocking is per-session only. setApplicationHidden is assumed
to persist under Device Owner but is untested.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AwepyL5Lxs5QKbKkMpxxxL
Registers the new Epopeja character-sheet app (~/epopeja_karta, its own
repo at kuhyx/epopeja_karta) in the shared icon family and lets it run
under focus mode.
The quill-nib glyph and the initial APPS entry already landed in c600760
("Move billsplit Flutter app into the monorepo"), which picked them up
from the working tree while they were still in progress. This commit adds
what was still outstanding:
- phone_focus_mode/config.sh: whitelist com.kuhy.epopeja_karta, without
which focus mode kills it within ~1s at home. Not deployed to the device
yet, so config and phone are out of sync until deploy.sh runs.
- apps.py: correct the epopeja_karta comment. It claimed there was no
desktop target; a linux/ target was since added so the app can be driven
on the desktop during development. linux stays False - the app ships no
desktop entry and gets no hicolor set.
- The evidence and contract artifacts the pre-commit hooks require.
Verified: 44 app_icons tests pass, 100% branch coverage on the package,
safe_box_overflow for quill-nib is (0, 0), and the generated mipmaps show
the nib rather than the stock Flutter icon.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
billsplit was a standalone Flutter app in ~/Downloads with one commit and no
git remote — that directory was the only copy of its history.
Three things the move needed beyond copying the tree:
- The root .gitignore silently swallowed 32 of its files (*.png, *.ico,
*.manifest). Ignored files do not error on `git add`, they just vanish, so
this would have committed an app that cannot build. Added a trailing block
of ! overrides plus .binary-allowlist (which did not exist yet) — the
documented in-repo exception for binaries a build genuinely needs. The
../testsAndMisc_binaries/ symlink scheme does not apply: those directories
are gone, and a symlinked PNG breaks a fresh clone.
- billsplit/.github/workflows/ci.yaml was inert once nested; GitHub only reads
workflows at the repository root. Relocated with a working-directory and a
billsplit/** path filter.
- The app's coverage gate moved to python_pkg/billsplit_coverage/ so every
Python file stays under python_pkg/, and gained a test suite at 100% branch
coverage as that location requires. billsplit/ now holds only Dart.
Also replaces the stock Flutter icon with a family icon from the shared
generator. The glyph is a filled silhouette because the family's 72px stroke
cannot carry a fine torn edge, and the split is an evenodd-cut perforation so
the holes stay transparent on the foreground and monochrome layers.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
New package. Walks every thread on 4chan's /wsg/, downloads each video it has
not seen, and reviews them one at a time: mpv embedded in a Tk window via --wid
with Keep/Pass/Undo directly beneath the picture.
Dedupe is structural rather than a convention. Every post carrying a file also
carries md5 -- base64 MD5 of the whole file -- so it is the sqlite primary key.
An already-seen video costs zero bytes to skip and a byte-identical repost in
another thread collapses onto the existing row. Verified against a live file:
the API's Nz9OEKdMuMZEdYE6eLTKmA== matched base64(md5(bytes)) exactly.
Undo is unbounded and survives quitting: the verdict trail lives in the index,
not in memory, so yesterday's decisions can still be taken back.
No downloaded video is ever deleted. A pass moves the file to trash/ and leaves
it there; files.apply_move refuses to overwrite an existing destination, so the
guarantee holds by construction rather than per caller.
Three bugs only running it could find, each caught from evidence rather than
guesswork:
- the video stayed black: a <Configure> handler set mpv's geometry, fighting
the automatic sizing --wid already does
- playback never advanced: refresh() ran every tick and re-sent loadfile ~7x
a second, restarting the video continuously
- it froze on one video around the 26th while filenames kept advancing: mpv
pushes an event stream down the same socket and nothing drained it, so the
buffer filled and mpv stopped reading commands. Isolated to command #35
without a drain thread; all 60 honoured with one.
A five-lens review then found nine more, all reproduced before fixing: a failed
mpv start stranded the non-daemon worker forever; undo could clobber a file in
incoming/; a NUL-bearing md5 poisoned the index permanently; an oversized int
killed the worker with no event; gzip bombs on both hosts were accepted; the
board endpoints were fetched unconditionally forever; and Session.missing could
corrupt the undo trail.
mpv runs as a subprocess over its JSON IPC socket rather than via python-mpv,
which resolves libmpv at import time and would fail on CI while passing here.
Third-party surface is requests alone, already declared.
277 package tests, 100% branch coverage on all 22 modules, no suppressions.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Its desktop app is a Chrome wrapper around the Flutter web build, and it
ships a .desktop entry plus a hicolor icon theme -- so it genuinely has a
Linux target. With linux=False, `generate --linux-out` skipped it silently
and reported success, which meant desktop/icons/ in that repo could not be
regenerated at all.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The new ~/home_inventory Flutter app needs two things from this repo: a
family launcher icon, and a focus-mode whitelist entry so the daemon does
not pm disable-user it ~1s after launch.
The glyph is an isometric carton — a hexagon plus the three edges meeting
at the near-top corner. A front-on rectangle with a lid seam was tried
first and rejected on the contact sheet: a seam plus a divider reads as a
browser window, not a box.
Body is 488x440, not a rounder 500x440, because STROKE_WIDTH straddles the
path and adds 72px to each rendered dimension; 488 is the widest body whose
ink still fits SAFE_BOX. The generator warned about exactly this at 500.
Accent stays the one shared family accent — apps differ by glyph, not
colour.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Two things that did not belong in a monorepo whose actual history is 48MB:
- third_party/mcp: 12GB of 24 third-party clones plus regenerable build output
(node_modules, Rust target/, venvs). Moved out and gitignored -- committing it
would have produced broken gitlinks from 24 nested .git directories. What is
tracked instead is a pinned manifest (upstream URL + exact commit for each) and
a setup.sh that reconstructs the fleet: 20KB in place of 12GB.
- python_pkg/session_autopsy: self-contained, imported nothing else here. Moved
with history via git subtree split.
Verified before removing anything: the SessionEnd hook ingests a real transcript
cleanly from the new location, and all 12 absolute-path references across
~/.claude.json and the three mcp-optional configs were rewritten with every
command path confirmed to resolve.
third_party/agent-skills is unrelated to MCP and stays.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
The solver only ran behind ./run.sh, had one media query in the whole
stylesheet, and kept the inventory in a 46vh scroll box that showed maybe a
third of the 43 dice. Setting it up on the PC did nothing for the phone.
Deployment
dice.kuhy.duckdns.org, behind the existing gitea-caddy edge. DuckDNS already
wildcards *.kuhy.duckdns.org (checked against 1.1.1.1 and 8.8.8.8), so no
dashboard step. deploy_build.sh builds to dist.staging and rsyncs into dist/
with --delete-after --delay-updates; plain --delete removes stale assets
before writing the new ones, which 404s a client mid-load. A post-commit hook
starts a MemorySwapMax=0 systemd user unit, so a rebuild can neither slow a
commit nor thrash zram.
Three traps found by driving the live URL in a browser rather than curling it:
a site address of "127.0.0.1:8089" reads as a HOSTNAME, so Caddy enabled
auto-HTTPS and started an HTTP->HTTPS redirect listener that bound *:80 next
to the edge -- SO_REUSEPORT would then split inbound HTTP and intermittently
break ACME for every domain here. The same address also made a Host matcher,
so the edge's forwarded Host matched nothing and served an empty 200 while the
local health check passed. And watching dist/'s own inode could never detect a
detached bind mount: rsync into an existing directory never changes it. Each
now has a guard that asserts positively rather than failing open.
Dice list
All 43 on screen at every width, no inner scrollbar. Column count and size
tier are chosen together from a container query on the panel -- not the
viewport, which mispredicts by 2x at the point the columns split. Two
invariants govern it: the list fits above the fold, and the name column stays
wide enough to tell Painter's die B, G and R apart. The left column holds the
grid alone, so it is sized to fill the height it owns: one --die-row value
drives the type, buttons and bar strip, giving a 52px row and 44px buttons on
a 1296px-tall screen and 20px rows on a phone.
Three safe-design rules are broken inside .die-list deliberately: sub-16px
body text, square icon buttons, sub-8px intra-cell gaps. 43 items with a name,
a chart and a counter in one viewport is arithmetically incompatible with 16px.
Import / export
A file, the clipboard, or a link carrying the inventory in the URL fragment.
One validator behind all four entry points including localStorage; it keeps
what it can and reports what it dropped, so a die added in a later game patch
costs you that entry, not the inventory. A link never silently overwrites what
is saved here -- it offers Keep or Discard, and only a link that actually
carries dice is applied at all, because one encoding a count of zero parsed
cleanly, blanked the screen, disabled Clear and left Keep writing the emptiness
over the visitor's own loadout.
session_autopsy
parse.py (551) and report.py (504) were over the repo's 500-line limit, so
test_file_length has been failing on main since 5ffd035. Extracted TurnTracker
and the two candidate-cell helpers. Report output is byte-identical on real
data apart from its timestamp, checked against HEAD in a worktree.
Verified: 288 TS tests and 911 Python tests, 100% coverage both, lint and
pre-commit clean, 8 viewports measured on the live HTTPS site, solver confirmed
through the real module worker.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Cache read is charged per API round-trip on the whole conversation so far, so a
long session's cost is driven by TURN COUNT, not message length. One measured
session spent 267M cache-read against 856k output — a ratio of 290:1. The report
priced tokens but never counted the turns spending them.
Across the 263 stored transcripts this now shows 33,508 tool turns at an 8%
batching rate, of which 4,437 (13%) decided nothing: polling a counter, running
a linter, checking whether a push landed. At 606k average context re-read per
turn that is ~2.7B cache-read tokens on turns a script could have taken.
- turns.py classifies a command as poll / lint_test / vcs_check / substantive,
deliberately conservative: state-changing commands (git commit, push) count as
substantive even though they are scriptable, so the reported waste is a floor.
A turn is only mechanical when EVERY command in it is.
- Batching is grouped by requestId, not by transcript line: one API response can
span several lines, and counting per line reports every turn as single-tool
(0% batching where the real rate is 12%).
- records.py gains TurnEfficiency (schema v3, additive — old records read as
zeros) with derived rates written into to_dict so readers need not recompute.
- report.py gains a "Turn efficiency" section pricing each mechanical kind.
- _TurnTracker owns both the totals and the open turn, after pylint flagged
_Accumulator over its attribute budget; _Accumulator.turns is now a property,
so every existing caller is unchanged.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
Per user verdicts: the ⚙ statusline glyph told them nothing (removed on
the ~/.claude side; the plain-words SessionStart line is the channel) and
explain-diff is not wanted at all (skill deleted there, verdict 'dropped').
The count now means 'significant AND not yet handled' — resolving a
candidate visibly shrinks it (7 -> 4), and handled table rows stop
re-suggesting a review.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The ⚙ statusline badge counted all 182 candidates — meaningless. It now
counts candidates ≥100k est weekly tokens (7 today): a number small enough
to mean 'worth your time'. token-savers.zsh gives the compiled scripts
shell entry points (finish-gate / phone-deploy / autopsy / yay-clean) so
the deterministic halves cost zero LLM tokens. Adds the first explanation
doc produced by the newly compiled explain-diff scaffold.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
The transcript analyzer behind /compile-candidate lived only in the working
tree. Untracked meant unlinted — pre-commit only inspects files handed to it —
so nothing had ever checked it, and it had been abandoned mid-refactor with
every parse raising AttributeError: '_Accumulator' object has no attribute
'first_timestamp'.
The accumulator had been moved onto the grouped records dataclasses, but three
increment sites and parse_session still read flat attributes that no longer
existed. parse_session now hands over acc.meta and acc.counts directly; only
subagent_count comes from elsewhere, since no transcript line carries it.
Two modules broke the 500-line limit, so they were split along seams that were
already there. signatures.py takes the text-signature helpers — traces.py was
importing them from parse.py, making parse an accidental hub, and now neither
imports the other. stats.py takes the pure number and time helpers. parse.py
521 -> 458, detectors.py 559 -> 467.
The structural lint findings were refactored, not configured away.
_counter_candidates took eight arguments; the five naming which counter field to
mine and at what thresholds became a frozen _CounterSpec, so both call sites are
one line. _extract_skill carried in_span and a line budget as loop variables
across thirteen branches; a _SpanBuffer owns that state now, leaving the walk at
four. record() took twelve flat kwargs and now takes one override per record
block, with session_meta() so a partial meta override keeps the standard
timestamp — a bare SessionMeta(slug=...) would have nulled started_at and
quietly changed the per-week arithmetic. All 36 call sites rewritten by an AST
pass. tests/factories.py moved into conftest.py, where every other package here
keeps shared builders and where name-tests-test does not object.
965 tests pass at 100% statement and branch coverage, every hook is green, and
no suppression was added anywhere. `candidates` and `traces skill-finish` both
run against the real corpus.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
It walked the working tree and skipped vendored trees by hard-coded directory
name. `.venv` was listed; `.ci-mirror-venv` — created inside the repo by the
pre-push CI mirror — was not, so the walk descended into it and died with a
UnicodeDecodeError on a joblib fixture that is deliberately encoded big5. A
crash, not a report: every real violation was hidden behind it.
Venvs are now found by their pyvenv.cfg marker, so any venv is pruned whatever
it is named, and the walk prunes in place instead of visiting thousands of
third-party files to filter them afterwards. Lines are counted from bytes, so a
line count cannot depend on an encoding. third_party/ joins the vendored skip
set — it holds upstream MCP servers, which this repo's 500-line rule never
governed.
Two regression tests come with it: a venv under an unlisted name holding an
oversized, undecodable module is pruned, and the line counter handles big5.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com>
style.BACKGROUND changing doesn't help if flutter_launcher_icons reads a
stale value from pubspec.yaml's own independent config -- that's exactly
what just happened across all 6 apps during the warm-palette rollout: the
generated PNG assets were correct, but Android's actual home-screen icon
(the adaptive icon: background color + foreground drawable, not the
legacy ic_launcher.png mipmap) stayed on the old color, invisible until a
physical device check caught it.
sync_pubspec_background() now rewrites the line in place before every
run_flutter_launcher_icons() call, so this can't silently drift again.
Verified: 43/43 tests (4 new), 100% line+branch coverage on generate.py.
Charcoal field #1B1D21 -> #211D1B, accent #4C7DF0 -> #B8862E across all 6
apps in the family, matching the unified-design-system palette rework
already applied to diet-guard/dufs-cloud/screen-locker/gatelock. Both are
exact byte-reversed mirrors of the old values, so relative lightness is
unchanged -- only the tint shifts from blue toward red/brown.
Also fixes stale prose in README.md left over from an earlier accent-
unification pass (claimed each app had its own accent; they've shared one
since that pass).
Verified: 39/39 tests, preview contact sheet reviewed, all 6 repos'
assets regenerated (diet-guard's rebuilt APK pixel-verified on device).
New Flutter app (~/habit_stack) gets an on-style launcher icon from the
shared family: two interlocking stadium-ring outlines, sized so the
72px stroke leaves a genuinely hollow interior at launcher size. Accent
matches the app's own ThemeData seedColor (#66BB6A), not already
claimed by a sibling app.
Also whitelists com.kuhy.habit_stack in phone_focus_mode/config.sh (both
the always-on productive-apps list and the night/curfew health-tool
list) and redeploys, since an unwhitelisted com.kuhy.* app is killed
within ~1s and would otherwise look like an install failure.
Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Every Flutter app still shipped the stock Flutter "beach" launcher icon, and
none had an adaptive-icon descriptor at all. Rather than hand-drawing five
one-off icons, define the family once and let each app contribute only an
accent colour and a glyph.
The family: a charcoal #1B1D21 field carrying one glyph in that app's own
theme seed colour, at a single stroke weight. Glyphs are optically centred
automatically - render.centre_offset rasterises a glyph alone, measures its
ink box, and translates (never scales) the fragment, so a row of icons lines
up while every icon keeps the exact family stroke weight.
safe_box_overflow warns when a glyph exceeds the 560px adaptive-icon safe
box; it caught three real overflows that a launcher mask would have clipped.
The registry points dufs_client at ~/dufs-cloud/app: ~/dufs_client was a
stale, remote-less checkout with an identical package name, since deleted.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Moved to https://github.com/kuhyx/build-your-x with full git history, where it
gained a `byox` progress tracker (crdt_sync-backed) and the builds themselves.
- git rm -r python_pkg/byox_ladder/
- CLAUDE.md: add build-your-x to the extracted-repos list
- .pre-commit-config.yaml: drop the now-stale byox_ladder codespell exclude
No remaining package imports byox_ladder.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01Nw9Vw8ufxLjkDox7iXPMH3
Ranks the codecrafters build-your-own-x projects by difficulty and emits two
offline, self-contained HTML pages: a per-guide ladder (~360 tutorials) and a
per-category ladder (30 categories), grouped into Beginner/Intermediate/Advanced/
Expert tiers plus an Unsorted bucket. A guide inherits its category's tier;
language only nudges order within a tier. Includes a Makefile pipeline
(fetch -> parse -> build) and 100%-branch-covered tests. Excludes the generated
guide page from codespell since it inlines upstream titles verbatim.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01QNRTPNYvoBTQoTEWkYtCJD
a12e696 turned CI red two ways, both missed locally because the suite was
run for the brother_printer package alone while CI runs everything in a
random order.
The repo caps Python files at 500 lines and cups_service.py had reached
659. Rather than cut it arbitrarily, extract consumable life tracking -
page counters, replacement baselines, the scale migration, dropped-page
detection - into consumables.py. The seam is real: none of it depends on
CUPS service control, and cups_service needs only estimate_consumable_life
for its fallback query, so the dependency runs one way. The test files are
split along the same seams.
test_main_entry passed or failed depending on execution order: runpy warns
when the target module is already in sys.modules, a sibling test imports
it, and pytest is configured with filterwarnings=error. Drop the module
before running it so the test no longer depends on who ran first.
Structural only; no behaviour change. 692 tests, 100% branch coverage.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RG3eEn8rwZbcXaJrKJxdv8
./run.sh failed outright with ModuleNotFoundError, and once running it
reported a sleeping printer as a critical paper jam. Both are fixed, plus
the defects found while verifying them against the real HL-1110.
Entry point: run.sh ran the script by path, so the absolute
python_pkg.* imports could not resolve. Resolve the repo root and pass
PYTHONPATH as a sudo argument, since sudo's env_reset strips an exported
one. __main__.py had the wrong import prefix and ran main() on import.
Status codes: the table claimed provenance from a "Brother PJL Technical
Reference" that does not appear to exist, and 23 of its 25 entries were
wrong. Rebuilt from the HP PJL Technical Reference Manual, Appendix D,
and verified against six states induced on real hardware (10001 READY,
10003 WARMING UP, 10023 PRINTING, 40000 SLEEP, 40021 TOP COVER OPEN,
41213 NO PAPER) - all six matched the reference, none matched the table.
40000 is SLEEP MODE, which the reference explicitly flags "not an error";
the real jam code 40022 was absent entirely. The inverse also bit: 10006
is TONER LOW, previously shown as a benign "Processing". Codes in no
reference are dropped and now report "Unknown status - check printer
display" rather than a confident wrong answer, and states we infer
ourselves no longer borrow PJL numbers to describe themselves.
Page counts: estimates came from the CUPS page log, which only sees jobs
CUPS logged and counts pages the printer never printed (it logged
total=50 for a job that printed 3). Take them from @PJL INFO PAGECOUNT
instead, migrating stored baselines onto the printer's scale once so the
reported percentages do not jump.
Reading status must not cost anything: a blocking open() hung forever
whenever the printer was unwell - precisely when a status report is
wanted - and the port-status probe stopped CUPS and reset the USB device,
destroying the in-flight job and then misreporting the damage as a
printer error. The fd is now non-blocking with deadline-bounded writes,
and the probe reads the same byte via the free usblp ioctl, skips
entirely while a job prints, and re-attaches the kernel driver.
Dropped pages: the HL-1110 silently discards pages whose 600dpi raster
exceeds its ~1MB memory, stays READY, and CUPS still calls the job
successful - so nothing anywhere reports the failure. Compare CUPS's
claim against the printer's counter between jobs and name the workaround
(lp -o Resolution=300dpi).
402 tests, 100% branch coverage.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01RG3eEn8rwZbcXaJrKJxdv8
test_codebase_fingerprint_stat_oserror_raises passed on 3.14 but failed
on CI's 3.11: there Path.is_file() itself calls stat() and re-raises an
OSError without a recognized errno, so the injected error escaped before
the loop reached the guarded path.stat() call. Force is_file() True for
the target (without calling stat) so the explicit fingerprint-time stat()
deterministically raises and hits the except OSError branch on every
Python version. Verified passing on 3.11.11 and 3.14.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RRCdR4D8SaVrN2xraMQxM
New python_pkg/code_tutor package that turns an unfamiliar codebase into
an interactive lesson: analyse code into items (_analyzer), order them
dependency-first (_plan_builder), pose reimplement challenges
(_challenge + _challenge_support), verify user implementations with tests
and coverage (_verifier), track progress (_progress/_session), and offer
local-LLM hints (_llm), driven by a Typer/Rich CLI (cli).
- meta/pyproject.toml + .pre-commit-config.yaml + meta/requirements.txt:
add anthropic/rich/typer deps, a cli.py per-file-ignore (Typer B008 etc),
and code_tutor-specific ruff/pylint/mypy tuning.
- _challenge.py and its tests are split across _challenge_support.py and
test_challenge_part2/part3.py to stay under the 500-line file cap.
Verification: pytest 631 passed at 100% line+branch coverage; every file
<=500 lines; pre-commit (ruff ALL, mypy strict, pylint 10/10, bandit,
evidence + multi-file contract) green.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_017RRCdR4D8SaVrN2xraMQxM
Full history preserved via git filter-repo; production
wake-alarm.service already cut over to the pip-installed standalone
package and verified clean.
Also fixes morning_routine._orchestrator, which has been silently
broken since screen-locker's extraction on 2026-05-28: it still
referenced python_pkg.screen_locker.screen_lock, causing
ModuleNotFoundError on every morning the workout-lock handoff ran.
Both module references now point at the pip-installed standalone
packages (wake_alarm._alarm, screen_locker.screen_lock); fixing this
required pip-installing screen_locker into system Python for the
first time too (see the separate screen-locker packaging-fix
commits), which it had never been.
Full history preserved via git filter-repo; production
diet-guard-gate.service already cut over to the pip-installed
standalone package and verified clean. No other monorepo file
referenced diet_guard, so removal is a clean delete.
The previous two fix attempts (removing -n 4 xdist parallelism, ruling
out a Python 3.12-vs-3.14 difference) both targeted the wrong cause for
the coverage gap on _alarm_display.py:71. The actual bug:
_restore_display() was never mocked in the _block_extra_devices fixture
shared by test_alarm_part2/3/4.py, so every test exercising
on_close()/_lock.close() ran the REAL _restore_display(), which calls
the REAL shutil.which('xset'). My dev machine has /usr/bin/xset
installed, so it always hit the "found" branch by accident; the CI
runner doesn't, so it always hit the "missing" branch instead, and
there was no dedicated unit test for the "found" path at all.
Fix: mock _restore_display in that fixture (matching its already-mocked
siblings _restore_fans/_restore_alarm_audio/turn_off_plug), and add a
dedicated, hermetic test for the xset-found branch that doesn't depend
on whether the running machine actually has xset on PATH. Side benefit:
the test suite no longer makes a real `xset s on` subprocess call on
the developer's desktop as an untracked side effect of running tests.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7vbgtFfZmfxJtN5DdtJky
test_demo_opens_window and test_bare_gate_due_opens_window never mocked
wait_for_display(), so they silently relied on a real, reachable X
display being present. On a dev machine this is always true, so they
passed; on a true headless CI runner (no X server at all) each test
blocked for the real 60s timeout before failing, explaining the CI
run's 131s duration. Mocking it to True (matching the existing
test_gate_due_but_display_not_ready_defers pattern for the False case)
restores hermeticity with no production code change.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7vbgtFfZmfxJtN5DdtJky
Third and final leg of the diet_guard -> screen-locker -> wake_alarm
gatelock extraction. WakeAlarm now composes gatelock.GateRoot +
LockWindow(mode="soft") instead of driving tk.Tk() directly, and moves
hardware teardown into on_close() so it runs on every exit path
(including SIGTERM), closing the prior gap where killing the process
left fans maxed. _state.py's HMAC import moves to gatelock.log_integrity,
the last call site of python_pkg.shared.log_integrity, which is deleted.
Manually verified on the real display: fullscreen overlay renders,
xdotool-typed input reaches the dismiss-code Entry (mode="soft" keeps
the existing typing-focus behavior), and SIGTERM exits within ~0.4s
while restoring hardware state. Full repo suite: 949 passed, 100%
branch coverage.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01A7vbgtFfZmfxJtN5DdtJky
MealGate now composes gatelock.GateRoot + gatelock.LockWindow instead of
inheriting the deleted _GateWindow/_GateRoot, and its HMAC signing goes
through gatelock.log_integrity. This is the first of three migrations
(diet_guard -> screen-locker -> wake_alarm) extracting the lock-window
mechanics that diet_guard's own _GateWindow proved out into a shared,
reusable package. Window-mechanics tests moved with the code; diet_guard's
suite now only tests its own wiring (LockConfig choice, hook delegation).
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01XCdT46zV8hESDvbgYMGDLt
Split diet_guard/_gatelock.py, wake_alarm/_alarm.py, and the
usage_report.py/_usage_report_parsing.py pair into focused
sub-modules so every Python file is <= 500 lines, satisfying
test_file_length.py. Install python-kasa into .venv (declared in
requirements but missing after the 3.13->3.14 venv upgrade),
fixing 8 failing smart_plug tests and restoring 100% coverage.
Also includes prior in-progress work from the working tree: the
wake_alarm Progress/View/Hardware field-grouping refactor,
brother_printer query module + tests, diet_guard foodbank/state/cli
updates, new shared coerce/logging_setup helpers, morning_routine
orchestrator tweaks, dwm window-manager config, gaming scripts, and
misc maintenance/digital-wellbeing script updates.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
