The weather widget's icons/layout are hand-drawn PIL primitives -- clean
under quantization but flat, no gradients/shadows. Adds an opt-in
render_style="modern" (current/daily modes only) that instead renders a
Jinja2 template through a persistent headless-Chromium browser
(app/html_render.py), following the approach of Tesserae, an open-source
e-ink dashboard targeting this same panel family.
Key design points:
- The Chromium dependency (Playwright) is lazily imported only when a
weather widget actually uses "modern" style, and the background browser
itself only launches on first use -- every other widget type, and this
one's own classic/hourly/multi_city paths, never pay for it.
- No Frame-level dithering setting needed: html_render dithers its own
rendered widget to exact palette colors (Bayer/ordered, not
Floyd-Steinberg) before compositing, so the shared whole-canvas
Floyd-Steinberg pass sees zero quantization error there and leaves it
untouched -- same trick draw_text/hand-drawn icons already use. Floyd-
Steinberg keeps working unchanged for photos and every other widget.
- A "Load calibrated Spectra 6 preset" button in Advanced configuration
offers a community-measured palette (data ported from
paperlesspaper/epdoptimize, Apache 2.0) as an alternative starting
point to the existing idealized DEFAULT_PALETTE_RGB -- fills the
existing palette table, doesn't save by itself.
Known open risk, not resolved here: a headless Chromium binary is far
larger than the ~100MB single-layer limit that already forced this
project's pip/npm installs into split layers, and (unlike those) is a
single ~180MB file that can't be split across layers by ordinary
Dockerfile restructuring. Flagged prominently in server/Dockerfile and
docs/widgets.md -- treat this render style as experimental/local-only
until that's resolved.
Introduces app/panel_style.py, a shared style module every render
module now draws through instead of independently duplicating margins/
colors/fonts: Inter Bold/Regular (already vendored, previously only
used by widgets/text.py) replace PIL's single-weight bundled default
font everywhere else; a per-widget-kind accent color (calendar=blue,
tasks=green, weather=black header) replaces plain black-on-white chrome
and is centralized in one THEME mapping so a future global theme only
needs to touch panel_style.py; a small per-widget gutter separates
adjacent widgets without touching grid.py's cell math; header bars,
color chips, and the battery icon get rounded corners.
Also drops the MUTED gray text color used throughout calendar_render.py
and weather_render.py -- a non-palette color that has no close match in
the panel's 6-ink palette and dithers into visible speckle once the
composited canvas is quantized. Secondary text now reads through size/
weight alone, always exact black.
widgets/battery.py and manage_overlay.py's previously-duplicated
battery-glyph-drawing code now share one implementation (panel_style.
draw_battery_icon). widgets/_shared.py's placeholder image is fixed to
use exact palette colors and route through image_pipeline.draw_text,
same as everything else -- it was quietly violating both rules already.
image_pipeline.draw_widget_border gains an opt-in radius param (default
0, unused by any call site) for a possible future rounded-border
setting -- doesn't touch the exact-corner-pixel behavior test_widget_
border.py already pins.
Deliberately out of scope: DEFAULT_PALETTE_RGB and the Floyd-Steinberg
quantization pipeline are untouched, per the prior reverted measured-
palette/OKLab attempt (05b417a/dfe9d701).
A new widget type shipping without an entry there fails silently --
no error, no test failure, it just saves/applies with an empty config
forever. Caught for real on the weather widget (37d57a1); adding the
step and a matching test-pattern bullet so the next widget type doesn't
repeat it.
LAYOUT_CONFIG_FIELDS never had a "weather" entry, so saving a layout
captured an empty config for any weather widget -- applying it back
(including via hold-to-cycle) reset mode/provider/city/units/etc to
defaults instead of restoring what was configured.
The combined pip install layer was already over Cloudflare's
single-blob/layer payload-size limit (~113MB unpacked) before any
recent change -- the last two build-and-push CI runs were failing on
it. Isolate the three largest packages (sqlalchemy, pillow, pypdfium2)
into their own layers, same fix already applied to render-service's
npm installs below for the same limit.
DEFAULT_PALETTE_RGB was a guessed approximation of the panel's ink
colors (pure sRGB primaries); swap in epdoptimize's measured spectra6
palette instead, which is far more muted/darker, matching how these
inks actually look.
_quantize now matches against the palette in OKLab space (perceptual
distance) instead of PIL's raw-RGB quantize(), with lightness weighted
down relative to hue/chroma when selecting the nearest color -- this
palette's inks are lit so differently from their sRGB namesakes
(muted dark red, bright yellow) that unweighted distance let lightness
dominate and mismatch hue (pure red nearest "yellow").
Dithering switched from Floyd-Steinberg error diffusion to a Bayer
ordered dither: true error diffusion is an inherently serial per-pixel
loop, and doing that in pure Python for a full 800x480 panel took
~1s, blowing past the render-latency budget the "render widgets
concurrently" fix (previous commit) exists to protect. The ordered
dither finds each pixel's true nearest and second-nearest palette
color and mixes between them (via projection onto that segment, not
distance ratio) using a tiled Bayer threshold -- fully vectorized, no
Python-level pixel loop.
A layout with several network-backed widgets (photos, weather,
calendar) paid their fetch latency serially in one /frame/* request,
which could exceed the firmware's fixed HTTP timeout and show a false
"server failed" status screen even though the server was still
working -- most visibly on the hold-triggered "cycle layouts" action,
which swaps in a whole new, cold-started widget set. Each widget now
renders on its own DB session in a thread pool (a plain Session isn't
thread-safe to share, but the per-frame threading.Lock in
frame_locked/widget_locked already made this kind of concurrency safe
by design -- see app/db.py); regions are still collected in
sort_order so overlapping widgets paint in the same z-order as before.
The admin log viewer only ever showed exceptions from device.py, not
successful requests -- no way to see a request that was slow-but-200,
or a device probing with a stale/wrong token. Adds a middleware that
logs method, path, device id (never the token), status, and wall time
for every /frame/* request.
The root logger previously had no handler at all, so every module's
logger.info() call (user creation, claims, password resets, ...) was
silently dropped, not just unviewable. Adds a RotatingFileHandler
writing into the existing /data volume so log content also survives
container restarts/redeploys, plus /admin/logs (tail + line-count
picker + full-file download) alongside the existing Users & Frames
admin page.
The boxed-card look reads as "still a website" once the app is
running full-screen off the home screen. Scoped to
display-mode: standalone so the regular browser-tab view is
untouched; dialog-internal cards keep their box since they group
subsections of one form rather than acting as page furniture.
Adds a web manifest, hand-drawn cup+frame icons, and a presence-only
service worker (no offline caching) so mobile browsers offer
"Add to Home Screen" for the server UI.
A 1M-ohm divider (way over the ~10k source impedance the ESP32 ADC's
sample-and-hold expects) doesn't always misfire in isolation -- short
bursts of a few consecutive bad readings, and multi-reading drifts,
both slip past the existing step-level MAD outlier rejection since the
steps between two bad readings in the same burst look ordinary. Add a
Hampel-filter smoothing pass (local-neighborhood MAD, same statistical
approach as the existing outlier rejection) ahead of it.
The server now records exactly what was last sent to the device on
every device-facing render (/frame/image, /frame/advance, /frame/back,
and the global hold actions), persisted as Frame.last_displayed_image/
_at and served back via GET /api/frames/{id}/now-displaying. The
header thumbnail is split into that frozen "now displaying" snapshot
and the existing live "up next" re-render, with an arrow between them
-- so editing a layout shows the change immediately on the right while
the left stays exactly what's actually on the panel until the device's
next real wake.
update_available only compared the latest Gitea release against what's
staged, not what the frame is actually running -- so once a release
was staged (manually or via auto-update) but the frame hadn't woken up
and applied it yet, "Check now" reported "Up to date" even though the
device was still on the old version. Report the frame's actual running
version and use it to show a distinct "staged, applies on next wake"
message instead.
Quick reset is now the fast/default action; summoning the management
menu takes a deliberate hold. Factory reset at ~15s is unchanged.
Renamed FRAME_COMBO_SOFT_RESET_HOLD_MS -> FRAME_COMBO_MENU_HOLD_MS to
match its new meaning. Bumps firmware to 1.4.1.
CI builds firmware inside the espressif/idf Docker image, but this
sandbox can't run containers at all -- it strips cap_sys_admin (and
blocks unshare) from the capability set even for root, which container
image-layer extraction and namespace setup both need. Confirmed by
hand: docker.io installs and dockerd starts fine, but even a bare
`docker run hello-world` fails to extract its own layer.
Works around it by installing ESP-IDF natively instead (git clone +
its own install.sh, scoped to just this project's esp32c6 target) --
the same way a developer would set it up on their own machine, needing
nothing this sandbox disallows. Verified end-to-end: both board
variants (devkit, xiao) build clean from a fresh checkout via the
packaged setup.sh/build.sh.
Release build for the per-widget button actions + hold-for-global-
action firmware changes (short/long press detection on next/back,
POST /frame/global-next|back). CI's firmware-build-check.yml already
confirmed both board variants compile clean at this commit.
Next/back button assignment moves from a frame-level "Button
assignments" card into each widget's own gear-icon dialog, prefilled
with a sane default at creation (photos/calendar -> advance/back,
whiteboard/weather -> check_now, others -> none). At most one binding
per (widget, button) now -- cross-widget execution order never
mattered since each widget's action only touches its own state.
New firmware capability: holding NEXT or BACK past a configurable
duration (min 3s, server-side default) triggers a frame-wide action
instead of the per-widget short-press one -- cycling saved layouts,
refreshing all widgets, or freezing/unfreezing every photo widget (see
app/global_actions.py). Firmware next/back checks gain the same
hold-duration polling the combo button already had; the threshold
comes from the previous wake's /frame/config fetch (persisted in NVS),
since this wake's button decision happens before that request.
Not done here: firmware/version.txt is intentionally left unbumped --
this hasn't been built or hardware-tested (no ESP-IDF toolchain in this
environment), so no firmware release build should be triggered yet.
A "Lock this photo" button in the photos widget's dialog toggles
PhotoWidgetConfig.locked, which suppresses both the timer-elapsed
auto-advance and the advance/back button actions until unlocked. The
Layout tab canvas shows a lock badge on any locked photo widget's box.
deb.nodesource.com started intermittently 403ing today on both its
setup_*.x scripts and its GPG key (confirmed directly, not just via CI --
some setup_NN.x paths 403, others 200, no consistent pattern), and the
curl-piped-into-bash install pattern silently swallowed that failure
instead of breaking the build loudly: curl -f exits non-zero on a 403,
but bash then runs on empty stdin and exits 0, so the RUN kept going
into a broken fallback (Debian's own split nodejs package with no
bundled npm) rather than stopping.
This base image now tracks Debian trixie, whose own nodejs package
(20.19.2) is inside jsdom 29's engines range and clears express/
resvg-js's much lower floors -- the version gap that originally required
routing through NodeSource is gone, so this drops that whole external
dependency (and the curl/gnupg install-then-purge dance) rather than
just swapping to a different NodeSource script.
A Widget-level property (border_style/border_thickness/border_color_index),
not a per-type config field, since every widget type can have one -- drawn
once centrally in device.py's _render_widgets before compositing, using
an exact panel palette color so it never dithers. Styles: solid, dashed,
dotted, and a fancy double-line picture-frame-mat look. Configurable from
a shared "Border" card in every widget's gear-icon dialog.
Shows Frame.battery_percent/battery_as_of, already set by every device
wake-on-battery report, plus routers/common.py's existing
battery_estimate_s time-remaining estimate -- nothing new to fetch or
cache. Compact (icon + percent) or detailed (+ estimate, last report
age) display mode. No button actions.
The vendored EC bitmaps looked good but dither into a visible speckle
once quantized to the panel's 6-color palette (their colors are
anti-aliased/arbitrary RGB, essentially never an exact palette match).
Hand-drawn icons filled with the frame's actual ink colors quantize with
zero dithering error to diffuse -- confirmed by running both through the
real quantize pass: the bitmap version speckles, the hand-drawn one is
pixel-identical before and after.
Redrawn to look more like EC's style this time around: pointed
triangular sun rays (the earlier attempt's thin-line rays read as a
crosshair, not a sun) and dendrite snowflakes (tick marks near each tip,
not a bare asterisk), plus the same cloud/raindrop/lightning-bolt shapes
as before. Removed the vendored server/app/weather_icons/ directory
entirely -- no longer used, and removes the icon-image licensing
question along with it.
The hand-drawn glyphs (draw_cloud/draw_sun/draw_raindrop/draw_snowflake/
draw_lightning_bolt) are replaced by 7 vendored bitmaps, one per shared
weather category, sourced from weather.gc.ca's public icon set -- these
are small, flat-shaded images that dither cleanly onto the panel's
6-color palette and read as recognizable weather icons in a way the
hand-drawn attempt (a plain circle-with-ticks "sun") didn't. Used for
every provider's rendering (Open-Meteo, NWS, EC), not just when EC is
selected.
Vendored (not fetched live at render time), matching this project's
existing convention for the Noto Emoji fonts -- server/app/weather_icons/
SOURCE.md documents the source, attribution, and the licensing caveat
(this is a personal, non-commercial project; the icon images' own
copyright terms are less clearly permissive than the weather data's own
End-use Licence, since they're served from the public website rather
than ECCC's data servers).
draw_weather_icon's signature changes from (draw, cx, cy, r, category,
palette_rgb) to (img, cx, cy, r, category): pasting a bitmap needs the
Image object, not just an ImageDraw handle, and palette_rgb is no longer
needed since the shared _quantize step already maps whatever's on the
composited canvas to the frame's actual palette -- no per-icon color
resolution required anymore.
app/weather/ec.py -- api.weather.gc.ca's MSC GeoMet OGC API
(citypageweather-realtime collection), the modern replacement for the
old dd.weatheroffice.gc.ca XML feed (that host no longer resolves).
Unlike Open-Meteo/NWS's simple lat/lon REST, this collection is only
queryable by bounding box, so _nearest_site widens the box
progressively and picks the closest of the ~844 sites by straight-line
distance -- capped at 300km, calibrated against a real bug caught in
development where an unconditional "nearest site, however far" matched
a Miami, FL query to a site in Ontario 1824km away once the box widened
to cover the whole country.
EC's own numeric icon codes get a small confirmed-against-live-data
mapping table plus the same keyword-on-condition-text fallback NWS
already uses for anything unmapped. Daily periods are named ("Today"/
"Tonight"/"Tuesday"/...) rather than dated, so dates are inferred by
walking them in issued order.
Verified end-to-end against the real live API (Toronto, rural
Saskatchewan, a US border city, and a rejected far-away match) and
through the browser (daily mode, composited panel preview). Test
fixtures mirror the actual response shapes captured live. docs/
widgets.md and CLAUDE.md's TODO updated -- EC is no longer a documented
gap.
New widget type with four display modes -- current conditions, an
hourly forecast strip, a multi-day forecast, and several cities' current
day side by side -- backed by a pluggable provider registry (app/weather/,
mirroring the app/widgets/ dispatch pattern): Open-Meteo (worldwide) and
NWS (US-only) both wired up now, Environment Canada documented as the
next one to add given its more involved station/grid-lookup API.
The calendar widget's existing embedded weather strip is untouched and
still Open-Meteo-only; this lifts the same underlying icon-drawing
primitives (now shared via app/weather_render.py, calendar_render.py
still imports draw_weather_row unchanged) into a widget that can be
placed and sized on its own. Icons are redrawn in the panel's actual ink
colors (yellow sun/bolt, blue rain/snow) instead of flat black, and
build_multi_city's icon/font sizing now scales with how many cities need
to fit rather than the box's height alone -- both fixed after catching
them via live browser verification, along with a mode-switch cache-shape
crash and a mobile-width dialog overflow.
New WeatherWidgetConfig table (migration 24), grid footprint, widget
module, common.py fetch/cache helper, router endpoints (location set/
clear, city add/remove, preview), dialog template + JS, and full test
coverage (providers, widget render, HTTP endpoints, migration replay).
docs/widgets.md and CLAUDE.md's TODO updated accordingly.
Two bullets had a stray "a \"coming up this week\" widget" phrase
overwriting their actual sentence ending (the AGPL bullet's "silently
accepting it)" and the mobile-breakpoint bullet's "below it"), likely
from an earlier bad edit. Restored both from git history; the phrase
still exists correctly once, as its own CURRENT TODO bullet.
Standing convention: commit and push once a task is verified working,
without waiting for a separate go-ahead each time -- see the new bullet
under Conventions specific to this repo.
The share QR's URL carried no auth params at all, so it silently fell
back through require_device's legacy-token resolution to whichever
frame happened to still be flagged legacy -- working only by accident
for a single frame, sharing the wrong frame's photos for any other, and
going fully dead once that frame's legacy flag was cleared.
Move the endpoint to manage.py, keyed on the frame's own manage_token
(same pattern /m/<manage_token> already uses) instead of device auth.
Since the server now resolves assets itself instead of trusting a
caller-supplied asset_id, it naturally generalizes to gather every
photo widget's current photo into one Immich share link, not just one
"primary" widget's.
Six more vendored families alongside the existing Noto Sans (Inter,
Source Sans 3, Noto Serif, Crimson Text, Arvo, IBM Plex Mono -- sans/
serif/slab/mono variety), all OFL-licensed with their own per-family
license file in app/fonts/ since each has a different copyright holder.
Static Regular/Bold/Italic/BoldItalic builds only -- variable-font-only
families (Inter and Source Sans's current Google Fonts releases, plus
Playfair Display/Lora/Merriweather) were skipped in favor of static
builds from their own upstream repos, keeping every family's loading
code uniform with what was already there. Considered but deliberately
left out: Georgia -- a proprietary Microsoft core font, not freely
redistributable, unlike everything else vendored here.
Also moves the bold/italic/underline/color toolbar below the
contenteditable box per request, and reorders the dialog's Settings
card to a more natural family-then-size order.
Fixes a latent migration bug this surfaced: migration 20 (static image
widget) used Base.metadata.create_all, which creates every table
declared in Base.metadata that's missing, not just its own new one --
harmless when nothing else pending, but once TextWidgetConfig existed
it would silently pre-create text_widget_configs (in whatever shape
models.py currently declares) before migration 21 got a turn, so
migration 21's own CREATE TABLE (or a later ALTER TABLE adding
font_family) would collide with a table create_all had already leaked
into existence. Both migrations 20 and 21 now use raw, frozen CREATE
TABLE SQL instead, matching migration 17's existing precedent for
exactly this reason.
A new self-contained widget type showing user-authored rich text -- no
live upstream to poll, like the static image widget, just word-wrapped
styled text instead of an uploaded image.
The dialog's contenteditable HTML is never stored or replayed as HTML:
app/text_content.py parses it server-side (on save) into a plain
paragraphs-of-styled-runs structure -- the actual sanitization
boundary, since raw HTML never round-trips back into any browser DOM
(the dialog rebuilds its editor from that same JSON via
createElement/textContent). app/widgets/text.py renders it with a
custom word-wrap/shrink-to-fit layout, using real vendored font weights
(app/fonts/NotoSans-{Regular,Bold,Italic,BoldItalic}.ttf, OFL-licensed
like the emoji fonts already there) rather than every other widget's
single ImageFont.load_default() -- the one widget type where that
distinction matters.
Nested .claude/skills/ dirs (previously under server/) are only
auto-discovered on-demand once a file under that subdirectory is
touched, so /make-widget and /run-server weren't invocable from a
fresh session. Root .claude/skills/ is scanned at session start.
Fixes setup.sh/start-server.sh's relative cd-depth math (was hardcoded
for the old server/.claude/skills/run-server/ depth) to instead
resolve the repo root via git and cd into server/ explicitly, and
updates SKILL.md/driver.py's path references to match the new layout.
The static-image widget shipped without updating docs/widgets.md's
per-type enumeration -- fixed, and added a skill encoding the full
file-by-file checklist a new widget type touches, so future ones (a
text widget is next) don't repeat either gap.
A new widget type showing a single user-uploaded image with no live
upstream to poll -- decoded once at upload time (PDF's first page via
pypdfium2, BSD-3-Clause/Apache-2.0, no copyleft exposure) into plain RGB
PNG bytes, then composed per a crop/stretch/letterbox display mode like
the photos widget.
Adds TaskWidgetConfig.name (migration 19, plain column add) shown at
the top of the widget on the actual panel instead of the hardcoded
"Tasks" header -- e.g. "Chores" or "Mom's list". The only widget type
with its own on-panel title at all, since it's the only one where
"which list is this" isn't already obvious from a calendar/photo/
whiteboard's own content.
Threaded through calendar_render's _draw_tasks/_build_tasks/
render_tasks/render_tasks_preview_png as a `title` param (default
"Tasks", truncated to fit -- a long custom name shouldn't be able to
overflow the widget's box), the config-save endpoint (tasks_name,
truncated server-side to a sane header length rather than rejected),
and a new "Settings" section in the tasks dialog.
Verified live in the browser: the name actually renders at the top of
the real composited panel (not just the dialog preview, which stays
gated on having a configured source), and persists correctly on both
desktop and mobile. Full suite (196 tests) passes.
Tasks widgets could only ever point at one CalDAV task list (a radio-
button picker, owner-only). Now they merge any number of included task
lists across every linked user, same checkbox-inclusion + optional
pinned-color shape a calendar widget already has for its calendars --
FrameTaskList mirrors FrameCalendar exactly, down to the same owner-
adds/anyone-mutes permission split (api_widget_task_list_select/
api_widget_task_list_color). Reused calendar_render._event_colors/
_draw_color_bar as-is for the per-task color bar -- a task dict's
owner_display_name/color_index is exactly that function's single-
source fallback shape.
Also added an opt-in "show tasks completed in the last 24 hours"
toggle (TaskWidgetConfig.show_completed): caldav_client.fetch_tasks
now accepts a completed_since cutoff and returns completed VTODOs
(with their completion time) instead of silently dropping them, and
_draw_tasks gives a completed task a filled checkbox + muted text
instead of the normal empty-box/due-date row.
Migration 18 splits the single-source TaskWidgetConfig columns
(added by 17, splitting tasks out of the calendar widget in the first
place) into frame_task_lists, carrying forward each widget's existing
single source as its first included list -- same shape migration 9
used carrying forward frame_calendars' old single opt-in.
Verified live in the browser (desktop + mobile): the new "Included
task lists" + "Recently completed" dialog sections, the show_completed
toggle actually persisting through a real HTTP round-trip, and no
regression in the calendar widget's own "Included calendars" dialog.
Full suite (192 tests, including new merge_tasks/config_save/migration
coverage) passes.
Task lists used to be a week-view-only sub-feature bolted onto calendar
widgets (CalendarWidgetConfig.tasks_*), so a task list could only exist
tied to a calendar's view and only inside its footprint. Tasks are now
a standalone widget type (TaskWidgetConfig, app/widgets/tasks.py) that
can be placed and sized independently, same as photos/calendar/
whiteboard -- no separate "enabled" flag either, since being on the
grid at all is the on/off switch, matching every other widget type.
Migration 17 creates task_widget_configs, extracts any existing
calendar widget's configured task source into a new sibling tasks
widget (auto-placed in open grid space, source dropped+logged if truly
none is left), then drops calendar_widget_configs' now-dead tasks_*
columns in the same migration -- this project's usual same-migration-
drop convention. Also handles the rarer case of a database jumping
straight from before the widget system existed to after this split in
one boot, via the legacy Frame.calendar_tasks_* columns.
Verified live in the browser at desktop and mobile widths: adding a
Tasks widget, its own dialog (task-list source picker + preview), and
confirming the calendar widget's dialog no longer mentions tasks at
all. Full test suite (180 tests, including new coverage for the widget
render/actions, the migration's data-extraction path, and the
permission-boundary shape for tasks-source) passes.
One request (DELETE /api/frames/{id}/widgets) removes every widget on
the frame in a single locked transaction, cascading their configs and
button-action bindings the same way single-widget delete already does.
Gated behind confirm() like the existing per-widget remove button, and
disabled when there's nothing to clear. Covered by the same owner/
unrelated-user/linked-but-not-controlling permission shape used
elsewhere (test_widget_placement.py).
Also fixes a real, pre-existing mobile bug this surfaced: the Layout
page's .layout grid used a bare `1fr` track on the <860px breakpoint
instead of `minmax(0, 1fr)` like the desktop rule already does, so a
wide enough descendant (previously nothing hit this; the new title-row
button did) would force the whole page into horizontal scroll on phone
widths. Verified before/after with the run-server driver's new
`viewport` command.
Confirmed the theatre-mode preview dialog actually renders correctly
at phone widths (390x844) -- centers properly, backdrop and close
button both fine, nothing overflows. Added a `viewport` command to the
driver (defaults to mobile, since that's the step easy to skip) and a
CLAUDE.md rule to screenshot both breakpoints for future UI changes,
since the 860px sidebar/mobile-bar fork is a real, previously-hit
source of bugs here.
This sandbox ships with no Python/Node/Docker/browser and no sudo, so
the bulk of this is setup.sh: bootstrap Python via uv, then get
Playwright's Chromium (and tmux, also missing) working by extracting
their .deb dependencies non-root instead of apt-get install. driver.py
is a small Playwright REPL standing in for chromium-cli, which isn't
available here either.
Also carves out .claude/skills/ from the blanket .claude/ gitignore --
skills are shared project tooling, not personal/local state.
tfaour
Thomas Faour