Docker defaults to HTTPS for any non-docker.io registry, and
10.0.0.246:3000 (a bare LAN IP:port) isn't serving valid HTTPS -- the
push kept going out over HTTPS and failing. Adds a buildkitd config
telling BuildKit specifically to use HTTP for that host.
This only fixes the actual build+push step (BuildKit). The "Log in"
step runs a plain `docker login` through the classic Docker CLI/daemon
instead, which doesn't read this config -- that one still needs
10.0.0.246:3000 added to "insecure-registries" in the actual Docker
daemon's /etc/docker/daemon.json on whatever host runs the Gitea
Actions runner, followed by a daemon restart. That's runner-host
infrastructure outside this repo.
Cloudflare (fronting git.thumeit.com) started rejecting image pushes
with 413 Payload Too Large once whiteboard mode's Node runtime + native
resvg bindings made the image significantly bigger than before. Switched
the build/push target to a LAN-local registry (10.0.0.246:3000) that has
nothing in front of it to hit that limit, and updated
docker-compose.yml.example to match.
New third mode alongside photos/calendar: fetches a .whiteboard file
over plain WebDAV (Basic auth -- generic, not Nextcloud-specific) and
renders it via a small Node.js sidecar using Excalidraw's own real
export code (@excalidraw/utils + @resvg/resvg-js, no headless browser),
since a .whiteboard file turns out to be Excalidraw scene JSON, not an
image. The sidecar runs as a second process inside this same container
(Dockerfile installs Node, start.sh backgrounds it before exec'ing
uvicorn) rather than a separate docker-compose service -- lightweight,
stateless, reachable only at 127.0.0.1 from the Python process, nothing
worth independently scaling.
The rendered PNG is treated exactly like a photo from there on --
composed/quantized through the existing image_pipeline (letterboxed,
never cropped) rather than a second parallel rendering pipeline.
WebDAV credentials support the common "it's actually the same Nextcloud
account as my CalDAV" case (an explicit opt-in checkbox, not silently
inferred) while still working with any WebDAV server generically.
Frame-level source (URL + owning account) follows the same owner-
controls-their-own-data permission split as calendar sources and the
week view's task list: only the account owner can point a frame at it,
anyone linked can clear it.
Honest limitation: this environment has no Node.js/npm, so
render-service/ is written carefully against each library's documented
API (verified via the npm registry, including transitive dependency
licenses after the CalDAV/AGPL surprise earlier this session) but has
never actually been executed. First real docker build is the first
true test -- see render-service/README.md.
Both are photos-specific settings (rotation order, how a photo's aspect
ratio gets reconciled with the panel), not device-wide configuration --
they belonged on the Photos tab next to the album/queue settings, not
Configuration. No backend change: both already save through the shared
partial-update /config endpoint regardless of which tab's form sends
them.
calendar_week_start's fixed-weekday anchor ("start on the most recent
Monday") stops making sense once the view isn't a literal calendar
week, so a non-7-day week view now starts calendar_week_start_offset
days from today instead (0 = starts today, negative/positive = past/
future) -- calendar_week_start still governs at the default 7 days,
unchanged.
Also hides the Calendar tab's week-only fields (days to show, layout,
start offset) unless View is actually set to Week, and further hides
the new start-offset field specifically when Days to show is 7 (where
it has no effect). "Week starts on" stays visible for Month too, since
it actually still applies there.
- Day count (2-10, was fixed at 7) -- 5 days trims the weekend clutter
without losing the grid format.
- Layout choice: days side by side (original behavior) or stacked
vertically as full agenda-style sections (reuses _draw_agenda_day,
same approach _build_today_tomorrow already used for a fixed 2 days).
- Optional task list (CalDAV VTODO collections only -- a plain ICS
subscription doesn't meaningfully have one) that takes the space of
one day slot instead of adding an extra one. Same owner-controls-
their-own-data permission split as calendar sources: only the
calendar's owner can point a frame's task list at it, but anyone
linked to the frame can clear it.
Browse-offset paging now moves by N days (was hardcoded to weeks),
identical to the old behavior when days=7. Changing the day count
resets the browse offset, same reasoning as changing views already did.
A single noisy ADC/regulator glitch (see firmware/main/battery.c)
survives the existing recharge filter: whichever way it reads, one of
the two steps around it (into a dip, or out of a spike) still looks
like an ordinary drop and got averaged straight into the remaining-
time estimate, letting one bad reading swing it dramatically.
Added a MAD-based modified z-score outlier check on top of the
existing recency-weighted average. Had to special-case the standard
MAD degenerating to exactly 0, which happens whenever more than half
the steps share the same value -- the norm for battery data (most
wakes cost the same small integer percent), and exactly the shape a
single spliced-in glitch among a steady discharge rate has, so the
naive case would have let the outlier this is for sail straight
through. Falls back to mean absolute deviation there instead.
Verified against constructed glitch scenarios in both directions
(spurious dip and spurious spike): estimate now comes out identical
to the same series with the glitch removed entirely.
Switched from the monochrome emoji font to color: NotoColorEmoji's
embedded CBDT bitmap glyphs, rasterized once at their native 109px
size and scaled to the target row height (unlike normal vector text,
color bitmap glyphs aren't stored at arbitrary sizes). Confirmed by
rendering an actual agenda row through the real quantizer that the
dithered-to-6-color result still reads clearly, not just muddy noise.
Falls back to the monochrome font if a deployment's Pillow/FreeType
wasn't built with embedded color bitmap support, so this degrades
instead of crashing or showing nothing.
The earlier fix stripped emoji instead of rendering them, which wasn't
what was asked for. Event titles now draw with two fonts: the usual
default font for text, and a vendored monochrome emoji font (Noto
Emoji, OFL-1.1) for actual emoji runs, so they show up as real glyphs
instead of a tofu box or nothing at all. Monochrome rather than color,
since reliably rendering COLR/CBDT color glyphs depends on how Pillow's
FreeType was built -- not something to depend on across deployments.
Also: events sharing the exact same title and time across different
calendars (e.g. a shared family event synced onto more than one
person's calendar) now collapse into one row instead of showing twice,
with a color bar split between every contributing calendar so it's
still clear whose event it is.
ImageFont.load_default() has no emoji glyphs, and PIL/FreeType don't
skip an unsupported codepoint -- they substitute a visible ".notdef"
tofu box, which read as a rendering glitch rather than "not supported."
Stripped instead (covers the standard emoji Unicode blocks, skin-tone
modifiers, and the ZWJ used to combine them into one glyph), with
surrounding whitespace collapsed.
Each linked person can pin one of the panel's four non-black/white
colors (Yellow/Red/Blue/Green) to their own calendar instead of relying
on calendar_render.py's old auto-cycle-by-owner-name order -- owner-only,
like adding a calendar in the first place. Colors resolve against
whichever palette a frame actually renders with (including a custom
Advanced configuration override), so a pinned "Blue" stays this frame's
actual blue. Event color bars/dots are also bigger and rounded now
across agenda/week/month views, easier to tell apart at a glance.
Weather: multiple cities per frame, geocoded via Open-Meteo (no API
key), shown above the event list on agenda/today & tomorrow/week views
-- never month, no room for it there. Hand-drawn sun/cloud/rain/snow/
thunderstorm icons (no new font/icon asset, same primitives-only
approach the rest of calendar_render.py already uses). City geocoding
handles "City, State" qualifiers Open-Meteo's own search doesn't
(disambiguates same-named cities, e.g. the three "Portland"s).
CalDAV fix: events were never showing despite calendars discovering
fine, because fetch_calendar_events relied on the calendar-query
REPORT's server-side time-range filter, which real servers implement
inconsistently (confirmed against a real server, not just guessed --
reproduced locally with Radicale). Switched to fetching every event
unfiltered and doing all date-window filtering/expansion client-side,
same approach already used for plain ICS feeds.
Calendar rule/grid lines were light gray, which dithers away to
near-invisible on the 6-color e-ink palette -- now black.
CalDAV accounts (Nextcloud, Fastmail, iCloud, ...) can now be linked
alongside the existing single ICS subscription, since one account can
expose several calendars. A frame's Calendar tab now lists calendars
per person rather than one opt-in per person: your own row shows every
calendar you have available with a full add/remove toggle, while other
linked users' rows show only calendars they've included, toggleable
off (mute) but not on -- only a calendar's owner can add it to a
shared frame. FrameCalendar replaces the old single-boolean
UserFrame.calendar_included; existing opt-ins are migrated forward.
The old estimate used a single linear percent/second rate from the
current discharge cycle's battery_history, which resets to empty on
every recharge -- so "not enough data yet" kept showing up despite the
frame having plenty of history overall, and the rate it did compute was
tied to whatever refresh interval produced it (changing the interval
didn't move the estimate until enough new history accumulated under
the new setting).
Now pulls the last 100 rows from the permanent battery_log table
instead, and averages the *per-wake* percent drop (not per-second) --
recharge jumps are skipped rather than counted as negative drain,
flat/zero-drop wakes still count so the rate isn't overstated, and
more recent steps are weighted more heavily. The per-wake rate then
converts to wall-clock time using the frame's current
refresh_interval_s and quiet-hours settings, so halving the refresh
interval roughly halves the estimate immediately, and quiet hours
correctly stretches it out (fewer wakes/day at the same per-wake cost).
Frame name (pencil-icon inline edit) and the Photos/Calendar mode
selector now live in the page header, shared across all four per-frame
pages instead of being buried in the Configuration form -- so renaming
a frame or flipping its mode no longer requires navigating to a
specific tab first.
Calendar settings move out of a conditionally-hidden card on the
Configuration tab into their own dedicated tab (new /frames/{id}/
calendar route), visually greyed out when the frame is in Photos mode
but still fully usable so calendar settings can be configured ahead of
switching modes. Also wires the week-start setting into the UI for the
first time (the column/backend support landed earlier but had no
control anywhere).
Responds to post-launch feedback on calendar mode: configurable
week-start day for week/month views, crisper non-antialiased text
(threshold-masked instead of drawn straight, so Floyd-Steinberg
dithering doesn't speckle glyph edges), a color-coded/proportionally
filled battery icon on the manage overlay, word-wrapped placeholder
text so "Calendar isn't set up yet" no longer clips in portrait, photo
inlay support extended from agenda-only to every view, and a fix so
manage-overlay face labels reposition correctly when a photo inlay is
active (they previously assumed the photo filled the whole canvas).
Also adds a fourth calendar view, "Today & Tomorrow" -- a two-day
agenda that reuses the same per-day row-layout helper the single-day
agenda view already has.
calendar_feed.py/calendar_render.py: fetch/merge per-user ICS feeds,
render agenda/week/month views. manage_overlay.py: composites the
manage-button overlay server-side (QR, battery, location/date,
share-QR, face labels), reused by every render mode. device.py/common.py
wire both together: mode dispatch for /frame/image+advance+back, and
the &manage=1 flag. Plus UI (frame_config.html Calendar card, settings
calendar URL field) and the icalendar/recurring-ical-events deps.
Deletes manage_qr_overlay.c/.h; frame_client.c's manage-button flow is
now one fetch with &manage=1 instead of on-device QR/text generation
plus separate photo-info/face-labels requests.
New columns, all ADD COLUMN with inert defaults -- no existing frame's
behavior changes until mode is explicitly switched to "calendar":
- users.calendar_ics_url: one personal iCal/CalDAV subscription per
user, same shape as the existing per-user immich_url/immich_api_key.
- user_frames.calendar_included: explicit per-(user,frame) opt-in,
default off. Being linked to a frame does not by itself contribute
your calendar to it -- each person's calendar is their own data to
share, not something a frame's controller decides on their behalf.
- frames.calendar_view/calendar_photo_inlay/calendar_browse_offset:
per-frame display settings and NEXT/BACK navigation state.
- frames.calendar_checked_at/calendar_cached_events/calendar_fetch_summary:
the throttled merge-fetch cache, same shape as the existing
firmware_update_checked_at/firmware_gitea_latest_version pattern.
A report was flagged as "the battery got recharged" (resetting
battery_history and stats_recharge_cycles, and re-arming the low-battery
alert) whenever it came in >= RECHARGE_JUMP_PCT above the single
immediately-previous report. That's exactly what a real recharge looks
like, but it's also exactly what a normal reading looks like right
after one noisy low report: e.g. 60, 59, 58, then a stray 53, then back
to a perfectly normal 58 -- 58 >= 53+5 falsely read as a recharge.
Now compared against the max of the last RECHARGE_LOOKBACK (3) reports
instead of just the one before it, so a lone stray reading doesn't get
to set the bar a normal reading then trips. A real recharge still needs
to clear all of them, so genuine recharges are still caught immediately
(verified: 18% -> 90% still triggers, history still resets).
Paired with the firmware-side battery.c change (trimmed-mean ADC
sampling) that reduces how often a stray reading like the 53 above
happens in the first place.
Sometimes a single reading came in noticeably off from the real trend
(a regulator/RF transient during sampling), and the next normal reading
would then look like a big jump relative to that bad one -- server-side,
enough to misfire the recharge-cycle heuristic (see the paired server
commit). Went from 8 raw-averaged samples to 16, sorted, with the 3
extreme samples on each end dropped before averaging the remaining 10 --
a handful of outliers can no longer skew the result the way a plain
average let them.
/api/frames/{id}/firmware/check could silently stage new firmware as a
side effect (the auto-apply path, when firmware_auto_update is on and
a newer release exists) but was gated by require_frame_view instead of
require_frame_control like its sibling firmware routes, and being a
GET, was exempt from the app's CSRF check (which only applies to
non-GET/HEAD/OPTIONS). A linked viewer without control -- or a
cross-site page riding a control-holding victim's session via a plain
GET -- could trigger an unreviewed firmware install. Now POST +
require_frame_control, matching /firmware/apply-latest; the frontend's
two callers (passive poll on page load, "Check now" button) both
already handle a 409 from a non-controller gracefully via the existing
apiError()/control-banner pattern, so this doesn't change UX for a
frame's actual controller.
Separately: Immich has been observed to return a face detection entry
with a null bounding-box field (a still-pending or otherwise
incomplete detection). Both places that do arithmetic on those fields
-- image_pipeline._face_aware_crop_box (crop_faces display mode) and
face_labels.compute_face_labels (manage-menu name labels) -- crashed
with an unhandled TypeError on such an entry, taking down that frame's
whole photo instead of the intended graceful fallback. Both now skip
any face missing a bounding-box field via a shared _has_bounding_box()
check; a face list with zero valid entries already degrades cleanly to
the plain center crop (the existing inf/-inf sentinel math already
handled "no faces" correctly, it just couldn't tell "none passed
Immich" apart from "one broken entry" before).
fetch_face_labels() clamped the server-reported label count against
max_labels by casting the count to int first -- a value >= 2^31 (a
perfectly ordinary decimal in JSON) went negative under that cast, so
the comparison was always false and the clamp never fired. The loop
then ran with the full, unclamped count, writing past the caller's
fixed MANAGE_FACE_LABELS_MAX-element stack array on a crafted
/frame/face-labels response. Reachable by a compromised/malicious
tools server, or a MITM on the default plain-HTTP connection.
Fixed by comparing unsigned instead of casting to int.
Two fixes from a security pass over the server:
- /api/frames/{id}/thumbnail/{asset_id} accepted any asset id and
fetched it via the frame owner's Immich credentials, unscoped to what
that frame actually shows -- a user merely linked to view a frame
could pull thumbnails for any asset in the owner's whole library, not
just the frame's own album. Now scoped to current_asset_id/queue,
matching the check device.frame_share and manage.manage_thumbnail
already both apply.
- firmware_update_repo_url now has to be a plain http(s) URL. Unlike a
one-off manual firmware upload (a deliberate, explicit act -- left
alone), auto-update from a repo is a standing trust relationship: the
frame keeps fetching from it and, with auto-update on, installs
whatever it finds with nobody reviewing it first. Added a plain-
language note next to the checkbox saying exactly that.
Previously the row just disappeared whenever battery_estimate_s
couldn't be computed yet, which looked like the feature was gone.
Now it always shows once there's any battery reading at all, with a
placeholder until enough discharge history accumulates (matches the
"Not enough data yet." wording battery_chart.js already uses for the
same situation on the chart).
"On battery for" was clutter next to the actual number people care
about. Relabeled "Est. remaining" to "Est. battery life left" and
dropped the now-unused on_battery_since field from the /queue response.
battery_estimate_s itself is unchanged -- it still needs 2h of span and
a 2% drop within the current discharge cycle (reset on any 5%+ jump,
i.e. a recharge or reflash) before it'll show anything. A frame that's
been power-cycled/reflashed recently won't have an estimate yet; that's
expected, not a regression.
_migration_1() is Base.metadata.create_all() -- it already builds
today's full schema straight from models.py. Every migration after it
is an incremental ALTER/UPDATE meant to bring an *existing* install
forward from an older version; replaying them against a brand-new
database collided with columns create_all had already added ("duplicate
column name"), crashing on first boot.
Found while testing the device-status-bar change against a scratch DB.
Every real deployment has been migrating forward incrementally since
before this bug existed, so it never showed up in practice -- but any
brand-new install would have hit it. Fresh databases now jump straight
to the latest schema_version after create_all; existing databases keep
applying whichever migrations are still pending, same as before.
Moved out of the Stats tab's side column into a new horizontal bar
shared by every frame page (Photos/Configuration/Stats), sitting
between the page title and the tabs so it's on screen regardless of
which tab is active.
_device_status_bar.html is a new partial included via a device_status
block in app_base.html; device_status_bar.js is the fetch/render/poll
logic extracted from frame_stats.js and adapted to a wrapping row of
label/value pairs instead of a stacked list. frame_stats.html's Device
card and now-single-card .side-col are gone -- Battery history and
Lifetime stats just stack directly.
battery_read_percent() was called once at the very start of boot, before
WiFi even connects, and that value was reused both for the manage-menu
overlay and the server report. Taken right after a reset (e.g. the OTA
reboot that immediately precedes it), the rail may still be settling --
plausible source of noisy jumps in reported battery level.
Now there's a single read, in frame_client_run() right before
report_battery(), after the photo (and manage overlay, if shown) is
already on the panel -- the fetch/display work already done this cycle
is the settle time, no delay to guess. The manage overlay no longer
needs an early local reading at all: it shows the server's last-known
value instead, added to the /frame/photo-info response it already
fetches.
Previously the softAP was torn down (esp_restart) only 1s after sending
the success page, while the page's own redirect timer waited 7s -- so
the AP (and the phone's captive-portal session with it) was gone long
before the redirect could fire. Now the page shows a live 10s countdown
before redirecting, and the device holds the AP up for 11s so the
countdown always completes. Also added a "Redirect now" button for a
phone that's already reconnected to normal WiFi.
Advanced configuration gains three sliders (PIL ImageEnhance factors
for color/contrast, 0-2, 1=unchanged; a 0-1 dithering strength) applied
to every photo this frame renders. Confirmed the parameter conventions
against a similar project (jwchen119/EPF: ImageEnhance.Color/Contrast,
1.0 baseline) before implementing; dithering strength isn't natively
exposed by PIL's quantize(), so it's implemented by blending the source
toward its own flat/undithered quantization before running Floyd-
Steinberg on the blend -- at 0 there's no quantization error left to
diffuse (exactly the flat result), at 1 it's the original unmodified
behavior, with a smooth continuum between rather than dithering being
an on/off toggle.
image_pipeline.py split into composition (_compose), enhancement
(_enhance), quantization (_quantize), and transpose+pack stages so
render_frame (device bytes) and the new render_preview_png (a normal
viewable PNG, upright logical orientation) share the same pipeline
instead of duplicating it. Named-face overlay label math (face_labels.py)
was already routed through the shared _placement_transform, so it
needed no changes for the new params.
Also added the requested before/after comparison: the Configuration
tab's new Preview card shows the current photo's untouched Immich
preview next to that same photo run through the frame's actual saved
rendering pipeline (two new GET endpoints, /preview/original and
/preview/rendered) -- immediate visual feedback for tuning the palette
and these new sliders. "Refresh preview" re-fetches after saving.
Schema migration v6 adds color_boost/contrast_boost/dither_strength,
defaulting to 1.0/1.0/1.0 -- reproduces the exact previous rendering
until a frame's Configuration tab changes one.
Verified against the live-shaped test database: the migration, sliders
persisting and clamping out-of-range input, both preview endpoints
(real JPEG passthrough / real PNG at correct logical size+orientation),
confirmed dither_strength=0 actually changes the rendered bytes vs.
default, and the standing legacy-device curl suite.
Replaces the smart_crop_faces boolean with a 4-way display_mode select
on each frame's Configuration tab (image_pipeline.DISPLAY_MODES):
- Crop to fill / Crop to faces: the previous False/True behavior,
unchanged (center-crop trimming excess, optionally shifted to keep
faces on screen).
- Stretch to fill (new): fills the panel exactly, aspect ratio not
preserved -- a plain resize, no crop.
- Shrink to fit (new): the whole photo visible, letterboxed with white
where it doesn't fill the panel.
Named-face overlay label positioning (face_labels.py, the manage menu's
"who's in this photo") now goes through a shared _placement_transform()
in image_pipeline.py instead of duplicating crop-box math, so label
placement stays correct (and in-bounds) under all four modes, not just
the two crop ones -- letterbox/stretch never crop a face out, so labels
just use straight scale+offset math there.
Schema migration v5 adds display_mode, backfills it from the old
boolean (True/False -> crop_faces/crop_fill), and drops the boolean.
Verified against the live-shaped test database: the migration
(existing frames correctly preserved as crop_faces), the config page's
new 4-option select, actual renders under letterbox (confirmed real
white letterbox padding in the packed panel-code bytes) and
stretch_fill, invalid-input fallback, and the standing legacy-device
curl suite.
A native <input type="color"> swatch can't be typed into precisely --
no way to enter an exact measured value. Replaced with a small table:
a read-only preview swatch, a hex text field, and three 0-255 number
fields (R/G/B) per ink color, kept in sync live in both directions
(editing hex updates R/G/B and the swatch; editing any of R/G/B updates
hex and the swatch). Hex stays the field actually read at save time --
the server-side validation (#rrggbb via hex_to_rgb) is unchanged, this
is a client-side-only swap of the input widget.
Advanced configuration (Configuration tab, collapsed <details> section):
a color picker per ink color (black/white/yellow/red/blue/green),
overriding image_pipeline.DEFAULT_PALETTE_RGB for that frame's actual
panel -- different units can vary enough from the documented
approximations to be worth calibrating once you can compare a rendered
photo against the real hardware. Stored as Frame.palette_rgb (NULL =
default, schema migration v4), threaded through render_frame/
render_placeholder/_quantize_and_pack (which now builds the PIL palette
image per call instead of once at import) so both photos and the
unclaimed/unconfigured placeholder screen respect it. "Reset to
defaults" clears back to NULL. Config-save validates exactly 6 #rrggbb
values, rejecting anything else with a 400.
Also: each frame's sidebar entry now shows its last-reported battery
percent (🔋NN%) next to the name, using the frame_dot's existing
recently-seen indicator conventions -- silent when never reported
(mains-only frames, or before the first report), matching how battery
is hidden everywhere else it's not applicable.
Verified against the same live-shaped database as the SMTP work: the
v3->v4 migration, save/reload/reset round trip through the real HTTP
route, an actual rendered image using a custom palette (confirmed via
its packed panel-code bytes), input validation, and the sidebar badge
against real battery data -- plus the standing legacy-device curl suite.
Two real fixes to app/mail.py, both found by testing against an actual
mail server rather than just a fake stub:
- Replaces the STARTTLS-only smtp_use_tls boolean with a three-way
smtp_encryption ("none"/"starttls"/"ssl"). Implicit TLS (port 465,
what Purelymail and most providers offer alongside 587/STARTTLS) is a
different handshake entirely -- TLS from the first byte, not a
plaintext connection that gets upgraded -- so it needs its own
smtplib.SMTP_SSL code path, not just a skipped starttls() call.
Schema migration v3 adds the column, backfills it from the old
boolean, and drops the boolean (safe on a live, populated DB).
- Outgoing mail was missing Date and Message-ID headers -- email.mime
doesn't set either automatically, and a missing Message-ID in
particular is enough for a strict content filter (confirmed via a
real Postfix+Amavis mail server's logs: SPF/DKIM/DMARC all passed
cleanly, but Amavis quarantined the message as "BAD-HEADER-0" purely
for the missing id) to silently swallow an otherwise-legitimate
email, even though smtplib reports success -- the send genuinely
succeeds to the relay, it just never survives the recipient's own
filtering. Both headers are now set, with the Message-ID's domain
matching the From address.
Verified: SMTP_SSL path against a hand-rolled implicit-TLS fake server
(self-signed cert, client-side verification relaxed only in the test
harness -- production code keeps ssl.create_default_context()'s real
verification), the v2->v3 migration against live data, the full admin
SMTP-save + test-email round trip over HTTP, and the standing legacy-
device curl suite.
Admin-configured SMTP (server/port/username/password/from address/
STARTTLS, a singleton server_settings row set from /admin -- not env
vars, since it's operator infrastructure a household admin sets up
once through the UI) powers two features, both requiring the relevant
user to have an email set in their own Settings:
- "Forgot password?" on /login emails a one-hour single-use reset link
(password_reset_tokens table). The endpoint always returns the same
generic "check your email" response regardless of whether the address
matched an account, so it can't be used to enumerate registered users.
- A frame's Configuration tab can set a battery-alert threshold
(Frame.battery_alert_threshold_pct, -1 = disabled); POST /frame/battery
emails the owner the first time a report drops to or below it, then
stays quiet for the rest of that discharge cycle (battery_alert_sent,
reset alongside battery_history whenever the existing recharge-jump
detection fires) -- not once per wake.
New app/mail.py wraps stdlib smtplib (no new dependency); send_email()
never raises, so a broken mail server can't 500 a battery report or a
password-reset request. Schema migration v2 adds users.email and the
two frame columns via ALTER TABLE (safe against the live, already-
populated database) plus the two new tables via the existing
create_all-based migration runner.
Verified against a real (already-migrated, real user/frame data)
database: the v1->v2 migration, admin SMTP config + test-email button,
full forgot/reset-password roundtrip (including single-use token
invalidation and the no-enumeration response), and the battery alert
firing exactly once per crossing against a hand-rolled fake SMTP
server -- all via curl end-to-end, plus the standing legacy-device
curl suite to confirm the device protocol is untouched.
The web UI grows into the multi-frame world: a left sidebar lists the
user's frames (with an online dot driven by the same overdue math as
the Device panel; collapsible off-canvas with a hamburger on mobile),
and each frame gets three tabs -- Photos (album picker, now displaying,
the drag-to-reorder upcoming grid), Configuration (name/order/
orientation/refresh/quiet hours/timezone/smart crop + the firmware
card), and Stats (device telemetry, lifetime counters, battery chart).
Settings and Admin adopt the same shell. / becomes a routing hub:
first frame, empty-state onboarding page, setup/login, or the
manage-QR redirect.
The JSON API moves to /api/frames/{id}/... behind require_frame_view /
require_frame_control: any linked user (admins see all) can view; 404
for frames outside your view so ids aren't confirmed; mutations 409
with the holder's name unless you hold the soft control lock, and
POST take-control always flips it to you. Config saves are now partial
updates -- each tab posts only its own fields (checkboxes always sent
explicitly), so the split forms can't clobber each other.
All CSS moves to static/theme.css and the old 680-line inline script
block splits into static/*.js -- the Pointer Events drag-drop state
machine and the canvas battery chart ported intact, not rewritten. The
CSRF fetch wrapper now reads a <meta> tag. No build step, still vanilla.
Verified end-to-end: page/static/API suites, control-lock handoff in
both directions, partial-save field preservation, non-admin frame
isolation, and the legacy-device curl suite (still byte-identical
responses for the deployed frame).
The frame-claiming pipeline, end to end. Firmware: every request now
carries ?id=<12-hex STA MAC> via build_url (mirrored in build_ota_url),
and the captive portal's success page became a redirect that hands the
user's browser to <server>/claim?device_id=... after ~7s -- enough time
for the phone to drop the provisioning AP while the device reboots.
The server pushes a per-frame device token through /frame/config during
a one-time handshake; the firmware persists it to NVS (a dedicated
single-key write that deliberately doesn't reset the connected-once
flag or WiFi cache) and prefers it over the provisioned shared token
from the next request on. Config response buffer grows 256->512. Both
board variants compile clean; new firmware also works against an old
server (which ignores ?id=) and old firmware against this server (the
phase A legacy mapping), so either deploy order survives.
Server: /claim lands the captive-portal redirect -- claim-gated signup
(a valid unclaimed/unregistered device id IS the enrollment invitation),
pending claims for the user-beats-the-frame race (auto-attached at
self-registration, 24h expiry), and a waiting page that refreshes until
the frame checks in. Unclaimed/unconfigured frames get a rendered
instruction placeholder with a QR from /frame/image (200, never an
error loop) -- new qrcode dep, placeholder shares the exact
quantize/pack path photos use.
The on-frame manage QR now resolves to a limited no-login page: scans
of / carrying device credentials (new ?id&token or the legacy shared
token) 303 to /m/<manage_token>, which allows exactly view queue,
show-next, advance, back, and scoped thumbnails -- no settings, no
removal, no other frames. Full control means logging in.
One real protocol hole found by simulating full wake cycles: after
self-registration the device could never authenticate again (the wake
cycle fetches the image BEFORE /frame/config delivers its token).
require_device now treats the id itself as the credential until the
first authenticated request flips device_token_ack -- the same trust
level as open registration, closing permanently once the handshake
completes.
Real identity on top of phase A's schema: scrypt-hashed passwords
(stdlib, no new deps -- parameters baked into each stored hash),
server-side sessions (sha256 of the cookie value stored, 30-day rolling
expiry), and per-session CSRF tokens enforced on every mutating
session-authed request -- via X-CSRF-Token for the JSON API (a fetch()
wrapper in base.html injects it, so the existing page scripts didn't
need touching) and a hidden form field for the HTML forms.
/setup runs once while no users exist: creates admin #1, links every
existing frame to them (owner + controller), and inherits the migrated
Immich creds onto their account -- per-user creds are now the primary
source, with env vars still winning as the operator fallback. /login,
/logout, /settings (display name, Immich creds, password change), and
/admin (enroll users, reset passwords, link users to frames, close a
frame's legacy-token window, delete) round out the pages, all in the
existing template/card style.
The legacy shared token stays accepted on browser routes so the
deployed frame's on-panel manage QR keeps working until phase C swaps
it for the limited manage page; token access renders without nav or
CSRF shim and is exempt from CSRF (explicit credential, not an ambient
cookie). Device routes untouched -- the legacy curl suite passes
verbatim.
Identity is provider-pluggable (identity_provider/provider_subject
already modeled) so OIDC can land later without schema surgery.
Replaces the single global config.json (whole-file pydantic model under
one RLock) with SQLite via SQLAlchemy 2.0: users/sessions/frames/links/
pending-claims/battery_log tables (models.py), a per-frame lock registry
(db.frame_locked) succeeding config.locked(), and hand-rolled schema
versioning (migration.py). A pre-database deployment's config.json is
imported verbatim as frame #1 on first boot and left untouched as the
rollback path; the old single firmware.bin slot becomes per-frame
firmware/<id>.bin.
Routes split out of the 900-line main.py into routers/device.py (the
frozen /frame/* protocol) and routers/api.py (web UI, still on the old
single-frame paths for now). Device auth moves to require_device, which
already speaks the full multi-frame protocol: per-frame device tokens
pushed via /frame/config and acknowledged on first use, self-
registration of unknown device ids as unclaimed frames, pending-claim
attachment, and the legacy-token migration window that keeps the
currently-deployed firmware (no id, shared MANAGEMENT_TOKEN) resolving
to frame #1 -- including the one-time binding of its device id when it
first reports one after a future OTA.
Externally identical for existing deployments: same paths, same token
semantics, same response shapes -- verified with a migration fixture,
the legacy-device curl suite, a 20-way concurrent-advance smoke test,
and a mutate-restart-assert persistence check against a fake Immich.
photo_queue.py ports nearly verbatim onto the Frame ORM row (MutableList
JSON columns make its in-place list mutations dirty-track); quiet-hours
math extracted unchanged into quiet_hours.py.
GET /api/firmware/check's 15-minute throttle meant a genuinely new
Gitea release could sit invisible in the UI for up to that long even
though POST /api/firmware/apply-latest (unthrottled) would've picked
it up immediately. New ?force=true bypasses the throttle for an
explicit check; the button wires it up and surfaces errors instead of
failing silently like the passive poll.
Dropped the paragraph explaining the release workflow builds the
binaries -- not something the web UI needs to narrate. The repo URL
field now shows as plain text with an Edit button once a value is
saved, instead of always being an open input.
All firmware-related controls (manual upload, Gitea repo URL,
auto-update checkbox, detected board, Update frame button) now live in
one "Firmware update" card instead of being split across the main
Settings form and a separate card.
The board variant used to pick a Gitea release asset was a dropdown
the user had to set by hand and could get wrong. The device now
reports it itself via a new X-Frame-Board header (CONFIG_FRAME_BOARD_NAME,
"devkit" by default, "xiao" in sdkconfig.xiao) on every /frame/config
poll, stored as device_board_variant -- the server learns it instead.
Update checks/applies are gated on the board being known, since there's
nothing to fetch until a device has checked in at least once.
The build container was pinned to espressif/idf:release-v5.3, but the
firmware is actually developed against v6.0.2 (this machine's
$IDF_PATH). That gap caused real compile failures in CI --
esp_sleep_enable_gpio_wakeup_on_hp_periph_powerdown and
ESP_NETIF_CAPTIVEPORTAL_URI are both newer than v5.3 has, plus a
transitive stdbool.h include that changed between versions -- none of
it a real bug in the source.
The runner's job workspace lives in a named Docker volume, not a real
host path, so a nested docker run -v "\$PWD:/workspace" bind-mounted an
empty directory when talking to the host daemon (which has no notion of
paths inside the calling container) -- build_for_board.sh: No such file
or directory. docker create/cp/start streams the checkout into the
build container through the Docker API instead, which works regardless
of what backs either side's workspace.
espressif/idf:release-v5.3 has no Node.js, and actions/checkout (a Node
action) runs inside whatever container: the job specifies -- it failed
immediately with "node: not found" on the first real run. Checkout now
runs on the plain runner; only the two build steps spin up the ESP-IDF
image themselves via docker run, using the runner's already-mounted
docker socket.
New Gitea Actions workflow builds both board variants and publishes
them as release assets whenever firmware/version.txt is bumped. The
server can now poll that repo's releases (next to the existing manual
upload) and either surface an "Update frame" button or, with
"Automatically apply updates" checked, stage the new build itself --
the frame still only updates on its own next wake either way.
The card being dragged never moved -- it just faded in place while a
static outline highlighted whatever was underneath the finger. Now the
card tracks the pointer 1:1 (translate + a slight scale-up "lift"),
gets a stronger shadow while airborne, and leaves its grid slot looking
like an empty gap until dropped (transform doesn't remove it from
layout flow, so the reserved space stays put -- no reflow needed until
drop). pointer-events:none while dragging so elementFromPoint's
drop-target hit-test sees through to the card underneath instead of
hitting the translated one.
Also: a short vibration tick when the touch hold-to-arm fires and
another on a successful drop (Chrome/Android only, iOS Safari has no
Vibration API -- harmless no-op there), and trimmed the arm delay from
350ms to 250ms now that there's actual feedback confirming the hold
registered.
The right column had grown to 5 stacked cards (Now displaying, Device,
Battery history, Firmware update, Stats) against the left's single
Settings form, running noticeably longer. Moved the collapsed Stats
card into a new left-column stack under Settings -- it's already
de-emphasized by design, so pairing it with the other
occasionally-checked column reads naturally, and it's small enough
(collapsed by default) to roughly balance the two columns' heights
without overcorrecting the other way.
Quiet hours only clamped what refresh_interval_s the *device* is told
to sleep for -- the separate elapsed-time check in
photo_queue.get_current() (used by /frame/image, /frame/photo-info,
/frame/face-labels, and /api/queue) had no quiet-hours awareness at
all, since it runs independent of the device. An open web UI tab
polling /api/queue overnight, or just loading the page during a quiet
window, would silently advance which photo is "current" on raw elapsed
time alone -- nothing reaches the panel until the device wakes after
quiet hours end, but the pointer moving mid-window still isn't what
"don't do anything overnight" implies.
get_current() now takes an in_quiet_hours flag that suppresses only the
elapsed-time trigger; an unset/invalid current photo is still picked
regardless (showing nothing is worse than showing something even at
3am). New _in_quiet_hours() helper in main.py, passed at all four call
sites.
New FrameStats (first_seen, device_wakes, photos_displayed,
photos_removed, battery_reports, recharge_cycles, ota_updates_applied,
config_saves), persisted alongside everything else in config.json.
Incremented at the existing route/photo_queue.py call sites that already
own each event -- no new instrumentation plumbing, no behavior depends
on these, purely informational. GET /api/stats serves them; the web UI
renders them into a native <details> "Stats" card (collapsed by default,
no JS needed for the expand/collapse), fetched once on page load like
the battery-history chart.
Verified: TestClient run through /frame/config (wakes + first_seen +
OTA-applied detection), /frame/battery (reports + recharge detection),
/api/config (saves); direct photo_queue.py unit checks for
advance/back/remove covering the "did the current photo actually
change" distinction (removing a queued-but-not-current photo bumps
photos_removed but not photos_displayed).
A press-and-drag gesture over a card's badge/button text also triggered
the browser's native selection highlight, distracting and occasionally
fighting the pointer-based drag tracking closely enough to break it.
Replaces the ad-hoc inline styling with a shared base template driven by
CSS custom properties (light/dark palettes), a card-based two-column
layout, and a persistent dark-mode toggle. Also moves quiet-hours'
timezone from the container's TZ env var into a proper web UI setting
(zoneinfo-backed), so it no longer needs a docker-compose.yml edit and
restart to change.
After a successful home-WiFi connection, caches BSSID/channel and
IP/netmask/gateway/DNS in NVS. The next wake's first connect attempt
uses the cached BSSID/channel (skips the all-channel scan) and applies
the cached IP directly once the link comes up (skips DHCP) -- a couple
fewer seconds of radio-on time per wake, free every wake since nothing
about the network actually needs renegotiating most of the time.
Falls back to a normal scan+DHCP attempt, and clears the cache, if: the
fast attempt itself fails, or it "succeeds" at the WiFi layer but the
full fetch cycle then fails anyway (a stale cached IP/DNS/gateway that
associates but can't actually reach the server). Also cleared on
(re)provisioning and factory reset, since a new network shouldn't try
to reuse the old one's cache.
The static-IP path needed care to get right without touching untested
territory: esp_netif_set_ip_info() only posts IP_EVENT_STA_GOT_IP (what
the existing connect-wait logic blocks on) once the netif is already
up, which the internal netif-glue's own WIFI_EVENT_STA_CONNECTED
handler guarantees by running first (registered earlier, in
esp_netif_create_default_wifi_sta()) -- confirmed against ESP-IDF's own
static_ip example and esp_netif_handlers.c source rather than assumed.
Falling back after a failed fast attempt also needed an explicit
esp_netif_dhcpc_start() first: esp_netif_dhcpc_stop() leaves the netif's
DHCP status STOPPED rather than resetting to INIT, and left alone the
glue would silently re-post the stale cached IP on the next connect
instead of actually running DHCP (esp_netif_action_connected).
Version bumped to 1.1.0 (real feature, not just a fix); build-verified
clean on both board configs (devkit 8MB, XIAO 4MB), no new warnings.
Purely a server-side decision: GET /frame/config hands back a longer
refresh_interval_s while quiet hours are in effect (exactly the seconds
until they end), and clamps the normal interval so the device's next
wake lands at the boundary instead of wandering into the window, when
outside it but approaching. A device already mid-sleep when quiet hours
begin can still land one wake inside the window -- unavoidable without
touching the firmware, since it has no wall-clock awareness -- but from
that wake on it sleeps straight through to the end.
Window is "HH:MM"-"HH:MM", wrap-past-midnight aware (e.g. 22:00-07:00),
in the server's local timezone -- added tzdata to the Dockerfile since
python:3.12-slim doesn't include it and TZ would otherwise silently
resolve to nothing and fall back to UTC.
Also fixed the "overdue" device-status check to account for quiet hours:
without this it would falsely flag a device sleeping through a long
quiet window as unreachable.
battery_history stays cycle-scoped (reset on recharge, feeds the "on
battery for"/estimate numbers), but nothing kept a permanent record --
added battery_log, appended on every report and never reset, capped at
~2 years of hourly reports as a sanity bound rather than a real limit.
New GET /api/battery-log serves it; the web UI draws it as a plain
canvas line chart (no chart library) under a new "Battery history"
section, loaded once on page load.
Also caught up server/README.md, which never documented the OTA
firmware endpoints or the /api/queue response's current "device" shape
from the earlier status-panel work.
Two independent timers: a 1s tick re-renders "Last seen"/"On battery
for" from the already-fetched device data (so they count up smoothly --
1s ago, 5s ago, 1m ago...) without hitting the server that often, and a
10s poll re-fetches /api/queue to pick up real changes (new photo
displayed, queue edited elsewhere, a battery/firmware report) -- both
reuse the existing render functions, no new endpoints needed. The poll
skips itself while a drag-reorder is in progress so it can't yank the
grid out from under an in-flight drag.
It's a ratio divider -- what matters is the two resistors matching each
other, not hitting 200k exactly. Also note sdkconfig.xiao now enables
FRAME_BATTERY_ADC_GPIO by default, which the doc's "off by default"
line no longer reflected.
FRAME_BATTERY_ADC_GPIO was still at its off-by-default -1 in
sdkconfig.xiao despite the XIAO being the one board this was designed
for -- the status panel showed no battery line because the device never
sent a report at all. Sets it to GPIO0, matching the settled
shared-with-back-button design.
- version.txt + esp_app_desc_t version reporting (X-Frame-Version header);
new ota_update.c checks the server's advertised version against the
running one and streams+applies an update via esp_https_ota, gated by
bootloader rollback (marks the image valid only after a full successful
cycle, so a bad update can't brick a wall-mounted frame).
- Dual-OTA partition tables: partitions.csv (8MB dev board, 2MB slots) and
new partitions_xiao.csv (4MB XIAO, 1.875MB slots -- the dev board's
table doesn't fit the XIAO's flash). New build_for_board.sh gives each
board its own build dir + generated sdkconfig via SDKCONFIG_DEFAULTS
layering, so switching boards never clobbers the other's config.
- fetch_photo_info()/fetch_face_labels() were using the short
reachability-check timeout even though the manage-menu path can be the
first (cold, TLS-handshake-paying) request of a wake cycle -- switched
to the longer fetch timeout to stop spurious ESP_ERR_HTTP_CONNECT
failures.
- XIAO: the RF switch that selects onboard vs. external antenna
(GPIO3/14) isn't initialized by plain ESP-IDF the way Seeed's Arduino
package does it, leaving WiFi unable to reliably reach the antenna at
all -- new board_antenna.c powers the switch and selects the onboard
antenna, gated behind FRAME_XIAO_ANTENNA_INIT (on by default in
sdkconfig.xiao). Also remaps the EPD DC/RST/BUSY pins, since the dev
board's defaults (GPIO9/10/11) aren't physically exposed on the XIAO.
/api/queue now returns a "device" object: last_seen/overdue, running and
available firmware versions, battery percent + on-battery duration +
linear-fit remaining-time estimate (recharge cycles reset the history so
estimates never span a charge). New POST /api/firmware (token-gated
upload, validates the embedded esp_app_desc_t) and GET /frame/firmware
(token-gated download) let a build be pushed to the device without
touching it physically. GET /frame/config now accepts an X-Frame-Version
header and returns the available firmware version, piggybacking the
device's update check on a request it already makes every wake.
Battery (firmware + server, disabled by default): new battery.c reads
a 2x200k voltage divider via ADC oneshot with curve-fitting calibration
(the ESP32-C6's scheme), maps through a piecewise LiPo discharge curve,
and restores the pin to button duty after each read -- the settled
XIAO ESP32-C6 design shares the back button's GPIO0/A0, time-shared per
wake. Skipped entirely when on mains (a 2x100k VBUS divider into a
spare digital pin -- the 5V pin is dead on battery power, so presence =
mains, where the charging voltage would read misleadingly full) or when
the reading is implausible. The manage overlay gains a battery region
(static outline glyph + "NN%", below the manage QR, all menu levels),
and the device POSTs to the new /frame/battery endpoint after a
successful fetch; the server stores percent + as-of timestamp, exposed
via /api/queue and shown in the web UI. FRAME_BATTERY_ADC_GPIO /
FRAME_VBUS_SENSE_GPIO default to -1 (fully inert on the dev board);
compile-verified both disabled and enabled, hardware bring-up deferred
until the ordered XIAO + batteries arrive.
Orientation (server-side only): new config setting + web UI dropdown
(landscape / portrait / landscape_flipped / portrait_flipped). Photos
are composed/cropped at the logical hanging shape (portrait crops at
480x800, so face-aware crops match how the frame actually hangs), then
rotated losslessly into the panel's native 800x480 byte layout after
dithering -- the device never knows. Face-label anchors are transformed
through the same rotation (logical_to_native()) so they stay attached
to faces on rotated frames. Known documented limitation: the on-device
manage overlay still renders in native orientation, so it appears
sideways on a portrait-hung frame (QRs scan at any rotation; text reads
sideways).
Remove from rotation: a new bounded exclude list
(FrameConfig.excluded_asset_ids) that photo_queue._top_up() never
selects from. POST /api/queue/remove scrubs an asset out of
queue/history too so it can't resurface via "Show next" or the back
button, and if it was the current photo, advances away from it
immediately -- without recording it in history, since going back to a
photo you just explicitly removed doesn't make sense. Doesn't touch
Immich or the album itself, just this frame's own selection. Wired into
the web UI as a small "x" button on both the current-photo thumbnail
and every upcoming card.
Mobile scroll fix: touching a card to scroll the page was being
captured as a drag attempt every time (touch-action: none on every
.photo-card, needed for the existing drag-reorder gesture to work at
all), making it too easy to accidentally reorder instead of scroll.
Reworked touch dragging to require a brief hold (350ms, roughly
stationary) before it arms -- touch-action stays "pan-y" (native
scroll allowed) the whole time up to that point, so a normal
touch-and-swipe scrolls the page like anywhere else, and only switches
to "none" once a hold is confirmed as deliberate. Mouse dragging is
unchanged (no hold delay -- no scroll-vs-drag ambiguity with a mouse).
Also made the "Show next" and new remove buttons always visible instead
of hover/focus-revealed, since that was invisible-but-still-tappable on
touch (no hover state) -- a real hazard for a destructive action.
Root-caused the earlier "No matching trusted root certificate found"
failure properly this time by reading ESP-IDF's actual bundle-matching
code (esp_crt_bundle.c): it looks up a trusted root by the ISSUER name
of whatever certificate it can't otherwise validate, not by matching
the presented certificate itself. The live server's chain ends in a
GTS Root R4 certificate cross-signed by the old GlobalSign Root CA R1
(common Cloudflare/Google Trust Services practice, for compatibility
with older/embedded clients) -- and ESP-IDF's current bundle snapshot
has dropped that old GlobalSign root entirely, so the lookup came up
empty. This was a general gap, not something specific to this one
deployment's cert.
Fix: keep the standard public CA bundle (esp_crt_bundle_attach) as the
trust mechanism -- so any normal reverse-proxy cert (Let's Encrypt,
etc.) works out of the box -- and add the one missing root on top via
ESP-IDF's CONFIG_MBEDTLS_CUSTOM_CERTIFICATE_BUNDLE mechanism
(sdkconfig.defaults), which appends a project-supplied cert file to the
bundle at build time. Fetched GlobalSign's official Root CA R1 cert and
cryptographically verified (openssl verify) it actually validates the
live server's certificate before embedding it -- see
firmware/main/certs/additional_root_ca.pem (replaces the old
tools_server_ca.pem, which pinned one exact certificate directly and
would've broken for anyone else's reverse proxy). Confirmed working
against the real deployment on hardware.
Found by a thorough code review:
- server/app/config.py's load()/save() each locked only their own file
I/O, not the full read-modify-write cycle each route does around
them. Since uvicorn dispatches sync routes to a thread pool, two
concurrent requests (e.g. the device's own poll landing alongside a
web UI edit) could each load() the same on-disk state and the
second's save() silently clobber the first's changes. Added
config.locked() (backed by an RLock, since load()/save() also take
the lock internally) and wrapped every mutating route's
load/mutate/save span in it -- kept outside the lock wherever a route
also does slow Immich network I/O, re-loading fresh state right
before the actual mutation instead. Verified with a new concurrency
stress test (many concurrent /api/queue/promote and /api/config
calls) alongside the existing scratch suite.
- firmware/main/root.html's SSID/password/toolsserver/access-token
inputs had no maxlength, so pasting something longer than the
matching NVS buffer (wifi_provisioning.h's FRAME_CFG_*_MAX_LEN) was
silently truncated with no indication why the device later can't
connect or gets 401s.
- frame_client.c's share_url buffer (256 bytes) could be too small in
the worst case -- toolsserver (128) + "/frame/share/" + asset_id (47)
+ "?token=" + access_token (64) can reach ~266 bytes, silently
dropping the token off a request that would then just 401 with no
obvious cause. Widened to 320.
firmware/README.md's HTTP vs HTTPS section still described the
public-CA-bundle trust approach that was tried and abandoned in favor
of pinning one specific certificate -- rewritten to match what's
actually there. docs/architecture.md was missing the back-photo button
entirely (sequence diagram and boot-flow bullets only covered next) and
still said the system talks "over plain HTTP" despite HTTPS support.
docker-compose.yml.example's MANAGEMENT_TOKEN comment understated its
scope (said "the web UI", omitting that every /frame/* endpoint is
gated too).
The native HTML5 Drag-and-Drop API (draggable/dragstart/dragover/drop)
is mouse-only by spec and never fires on phones/tablets, so reordering
was unusable on mobile. Replaced it with the Pointer Events API
(pointerdown/pointermove/pointerup), which unifies mouse, touch, and
pen into one code path, plus touch-action: none on the cards so
touching one to drag it doesn't get hijacked by the browser's default
scroll gesture. Same visual behavior as before (dim the dragged card,
outline the drop target).
Back button (new GPIO0, POST /frame/back): the server now tracks a
bounded history of previously-current photos (photo_queue.py), pushed
to on every advance (auto or forced) and popped by back_forced() --
symmetric with advance, so pressing next afterwards returns to right
where you were. frame_client.c's force_advance bool becomes a 3-way
fetch_action_t (NORMAL/ADVANCE/BACK) threaded through the whole fetch
path.
Also folds the separate reset and manage buttons onto one pin
(combo_button.c, replacing reset_button.c/manage_button.c entirely),
disambiguated by hold duration: quick press shows the management menu
(unchanged), ~3s hold-then-release soft-resets (esp_restart(), config
kept -- new), ~15s hold factory-resets (today's old reset behavior,
extended from 10s for clearer tier separation). Driven by a production
board (Seeed XIAO ESP32-C6) exposing only 3 of the ESP32-C6's 8
deep-sleep-wakeup-capable GPIOs -- next/back keep their own dedicated
pins where instant response matters most, everything else shares the
third pin via timing instead of needing its own. Same three-pin layout
now works on both the dev board and the production board.
Fixed a fast-tap bug in combo_button_check() before shipping: it only
did a live gpio_get_level() read to decide whether the button was
pressed at all, so a press fast enough to already be released by the
time boot reached that check was missed entirely (treated as "never
pressed" rather than "quick press"). Added the same latched
esp_sleep_get_gpio_wakeup_status() check the other buttons already use
for exactly this reason.
The Tools Server hostname turned out to be Cloudflare-proxied, not a
direct connection to nginx -- so the ESP32 (and any browser) sees
Cloudflare's own edge certificate (issued by Google Trust Services),
never the Origin CA cert, which only ever sits on the Cloudflare-to-
origin leg. Confirmed on hardware: ESP_ERR_HTTP_CONNECT.
Tried switching to ESP-IDF's built-in public CA bundle instead
(esp_crt_bundle_attach) as the more general fix, but that also failed
on hardware ("No matching trusted root certificate found") -- the
bundle's copy of the relevant Google root has the same name and public
key as the live one but a different serial/signature (a reissue), and
the bundle does an exact byte-level match, not a semantic one.
Simplest reliable fix: embed the exact certificate the proxy actually
presents (extracted live via openssl s_client, see
firmware/main/certs/tools_server_ca.pem) and trust that directly via
cert_pem, sidestepping bundle-matching semantics entirely. Documented
in firmware/README.md how to re-extract if the proxy's CA ever changes.
The management token only gated / and /api/* -- every device-facing
/frame/* endpoint (including /frame/image, which serves the actual
photo bytes) stayed open regardless. That was fine while the server
was assumed LAN-only, but defeats the point now that HTTPS exists
specifically to let this sit behind a public hostname.
build_url() (frame_client.c) is the one chokepoint all firmware-side
URL construction already went through, so it now appends ?token= to
every request it builds -- device fetches and QR-embedded links alike
-- instead of that being bolted on per-callsite. Server-side, the
former require_management_token dependency (renamed require_access_token)
is applied to /frame/config, /frame/image, /frame/advance,
/frame/photo-info, /frame/face-labels, and /frame/share/{asset_id} too.
/health stays open -- pure liveness, nothing sensitive to protect.
ESP32 side can now reach the tools server over HTTPS: the Tools Server
field accepts an https:// address for a TLS-terminating reverse proxy
in front of the server (which still only ever speaks plain HTTP
itself), trusting Cloudflare's Origin CA root (embedded at build time)
since that's the common way to get a real cert on a private origin.
Every URL the device builds -- image fetch, config check, manage-menu
data, the QR codes' own links -- goes through one build_url() helper
that picks the scheme from what's configured.
Also adds an optional MANAGEMENT_TOKEN (docker-compose.yml) that gates
the web UI (/, /api/*) behind a shared secret -- unset by default, so
existing trusted-LAN deployments are unaffected. The same token is
entered once during the ESP32's captive-portal setup and gets baked
into the manage-menu's QR code (?token=...), so scanning it just works;
visiting the page without a valid token shows a plain entry prompt
instead of the config UI, and a valid query-param hit sets a cookie so
the page's own fetch()/<img> calls stay authorized for the rest of the
visit. Device-facing /frame/* endpoints are unaffected -- a separate,
already-documented trust boundary.
ESP-IDF locks ("holds") every pin armed as a GPIO deep-sleep wakeup
source across the sleep transition, and never releases it automatically
on wake -- confirmed against sleep_modes.c's
esp_sleep_gpio_wakeup_prepare_on_hp_periph_powerdown(), which calls
gpio_hold_en() with no corresponding gpio_hold_dis() anywhere in
ESP-IDF's own wake path. Left held, live gpio_get_level() reads stay
frozen at whatever level the pin had when sleep began (almost always
"not pressed"), which is indistinguishable from a real "not pressed"
reading and silently broke any live poll for a *new* press later in the
same awake session.
This never surfaced before the manage menu's escalation feature, since
every other button check either used the latched wakeup-status register
(unaffected by hold) or only polled once, early in boot, before any
sleep/wake cycle in that session. wait_for_button_press() is the first
code in this project to repeatedly poll a button live *after* having
just woken via that same pin -- exactly the case hold breaks. Fixed by
calling gpio_hold_dis() before gpio_config() in all three buttons' init
functions, not just manage's -- reset and next-photo have the same
latent issue in their own live-read fallback paths, just not yet
exercised the same way.
A plain read-only key (per the original setup instructions) 403s on
POST /api/shared-links -- confirmed against the live instance's
permission enum in /api/spec.json.
Two rounds of follow-up work on the manage-button overlay:
1. Location formatting: US/Canada now show abbreviated state/province
("CA", "ON") instead of the full name, other countries show the full
country name, and each is its own line (was one line, now wraps to
two) so longer international place names have more room without
threatening to overlap the top-right QR box. The bottom-left share QR
also gets a "SCAN TO DOWNLOAD" caption.
2. Escalating menu: pressing the manage button again while its overlay
is already up adds a second level -- each Immich-identified person's
name labeled next to their face in the photo (using Immich's own
face recognition/People data, no detection/recognition added to this
project). A third press exits immediately instead of waiting out the
30s auto-revert timer. No new Immich API needed -- GET /api/faces
already embeds a nullable person.name per face; new
server/app/face_labels.py maps a named face's box into the final
800x480 frame's pixel space (reusing crop-box math extracted from
image_pipeline.py's face-aware cropping). Capped at 4 named faces,
sized to a real firmware RAM budget: each label is its own malloc'd
overlay region on the device, alongside the 4 fixed corner regions
already in use. New GET /frame/face-labels returns a flattened
fixed-slot JSON shape (not a real array) so firmware's existing
flat-scalar parser can read it without needing an actual array
parser. No persistent state needed for the escalation itself -- it's
all local control flow within one continuous awake session
(frame_client.c's run_management_menu()).
Two changes, bundled since they landed in the same session and touch
overlapping files:
1. Fix: "Show next" sent the browser's full queue snapshot to
POST /api/queue/reorder, which hard-rejected if the server's queue
had shifted since the last fetch (e.g. right after a queue-length
trim). New POST /api/queue/promote moves one photo to the front
authoritatively, with no dependency on client staleness. /reorder
itself is now tolerant too -- unrecognized IDs are dropped and
missing ones appended, instead of rejecting the whole request.
2. Feature: the manage button's overlay now also shows the photo's
location (top-left, only if Immich reverse-geocoded it from GPS
EXIF), the date it was taken (bottom-right), and a QR code (bottom-
left) linking to a 30-minute public Immich share link -- created
lazily when someone actually scans it, not when the button's
pressed. New server endpoints GET /frame/photo-info and
GET /frame/share/{asset_id} (scoped to the frame's current/queued
photos, not any arbitrary Immich asset). Firmware-side, the overlay
mechanism generalizes from one spliced region to up to four
(manage_qr_overlay.c), each its own small buffer, still never
holding the full frame in RAM.
Adds "Upcoming photos to show" to the config UI (queue_target_len, 5-50,
default 20, replacing the hardcoded QUEUE_TARGET_LEN constant). Lowering
it trims the queue immediately on next page load rather than waiting for
enough advances to consume the excess naturally; raising it tops back up
the same way, via a new photo_queue.sync_queue_length() called from
GET /api/queue.
Replaces the up/down-button vertical list with a responsive photo grid
(native HTML5 drag-and-drop between cards, reusing the existing
POST /api/queue/reorder endpoint -- no new server route needed). Each
card also gets a "Show next" button that jumps it straight to the front
of the queue.
Also bumps the queue lookahead from 10 to 24 photos (QUEUE_TARGET_LEN in
photo_queue.py) now that the grid has room to show more at once.
Pressing the manage button (GPIO1) overlays a small QR code -- "SCAN TO
MANAGE" -- in the top-right corner of whatever photo is currently on
screen, linking to the server's config page, then reverts to the plain
photo after 30 seconds.
The overlay is spliced into the existing streaming fetch as chunks pass
through (frame_client.c's http_read_fn), rather than buffering the full
192,000-byte frame in RAM: only the small overlay rectangle itself
(~30KB) is ever held in memory, generated via new stride-parameterized
drawing helpers (epd_draw_*_ex in epd_draw.c) that let the existing
QR/text drawing code target an arbitrarily-sized buffer instead of a
full-frame one. epd7in3e.c is untouched -- it has no idea an overlay
exists.
The QR onboarding and "CONNECTING..." status screens write to the panel
through a separate path that never touched the last-displayed-photo CRC
added in the previous commit. That left it stale relative to what's
actually on screen after either one draws -- most visibly after a
factory reset: reprovisioning and reconnecting could fetch a photo whose
CRC happened to match the one from before the reset, skip the refresh,
and leave the QR code frozen on screen indefinitely. Both screens now
invalidate the tracked CRC right after drawing, so the next photo fetch
is always guaranteed to actually refresh.
The panel driver now splits writing a frame into its SPI buffer
(epd_write_frame(), which also computes a CRC32 as it streams) from
actually triggering the physical refresh (epd_turn_on_display()).
frame_client.c compares the new CRC against the last one that was
actually refreshed (persisted in NVS) and skips the refresh entirely
when they match -- e.g. a reboot redisplaying the same photo before the
server's refresh interval elapsed no longer causes a visible flash for
no visual change.
Also reorders the per-wake fetch cycle: the image fetch (15s timeout)
now goes before the config check (3s timeout), instead of after. The
config check's tighter timeout was intermittently tripping on
connection-setup latency that's common on the first request after
waking from a long deep sleep (e.g. stale ARP); putting the more
tolerant request first absorbs that latency, and the config check then
rides the connection it already warmed up.
Factory-reset (GPIO3, hold 10s): clears stored WiFi/server config and
restarts into provisioning -- the deliberate, USB-free replacement for
the earlier reverted RST-based auto-reprovisioning idea.
Next-photo (GPIO2, tap): wakes the device and forces the server to
advance immediately via a new POST /frame/advance, instead of waiting
for the refresh interval. Both buttons arm themselves as deep-sleep GPIO
wakeup sources so a press is noticed promptly even while asleep.
Also makes GET /frame/image side-effect-free: it now only advances once
refresh_interval_s has elapsed since the current photo was set (tracked
server-side), so a device reboot for any reason just redisplays the
current photo instead of silently skipping ahead. The server maintains a
small reorderable upcoming-photos queue, viewable and rearrangeable from
the web UI.
- LICENSE: MIT, with attribution notes for the vendored qrcode/epaper_fonts/
dns_server code and the epd7in3e driver's transcription of Waveshare's
register sequence.
- Top-level README.md: project overview, hardware list, quick-start
pointing at firmware/ and server/, repo layout, license, Claude Code
attribution.
- firmware/README.md: full rewrite (was still the stock ESP-IDF captive
portal example's README) -- build/flash instructions, Kconfig reference
table, first-boot walkthrough, and how to reset to provisioning mode via
NVS erase (the only way in right now; a proper reconfigure trigger is a
future addition).
- docs/hardware.md: wiring table + parts list + strapping-pin/SPI-speed notes.
- docs/architecture.md: sequence diagram and walkthrough of the full
provision -> connect -> fetch -> display -> sleep cycle, plus the
reasoning behind doing image processing server-side and reusing Immich's
face detection instead of bundling a detector.
- server/README.md: fixed stale endpoint docs (missing GET /frame/config,
POST /api/config still describing removed immich_url/api_key fields).
Found on hardware: the Tools Server field was pointed at Immich's own
port instead of the frame server's, so /frame/image was actually hitting
Immich and getting back a small error response (~10KB) instead of a
192,000-byte frame. epd_display_stream() logged a size-mismatch warning
but called epd_turn_on_display() anyway, physically refreshing the panel
with a buffer that was ~95% whatever was left over from before -- visible
as "garbage" on screen, overwriting a previously-good image.
epd_display_stream() now returns ESP_ERR_INVALID_SIZE instead of
refreshing when the stream doesn't supply exactly EPD_FRAME_BYTES. Since
this check happens before epd_turn_on_display() is ever called, the
pixel data that *did* arrive only ever reached the panel's internal RAM
over SPI, not the physically visible display, so aborting here leaves the
screen exactly as it was.
This also means fetch_and_display() failing now always implies the panel
was never touched -- simplified frame_client_run() accordingly (dropped
the now-always-true/false out-param that used to distinguish "failed
before vs. during streaming", and always shows the FAILED status screen
on any fetch/display error, since it's now guaranteed safe to do so).
tfaour