The 2 ms reset of ESPHome's expander support (0.2.81) left the chip silent: "Calibration error" with the PCA9557
answering at 0x18. A component of smart_display's (elecrow_touch_reset.h, setup priority 700: after the I2C bus,
before the GT911 driver) writes the expander as Elecrow's V3.0 ESPHome example does: every pin an output and low,
20 ms, the reset released, 100 ms, the interrupt line back to an input. The driver's reset and interrupt pins are gone.
A unit's touch never answered: "touchscreen is marked FAILED: Calibration error" at setup, the GT911 still in reset
behind the PCA9557 (I2C 0x18) of the V3.0 board. The board file declares the expander as ESPHome's pca9554 (the same
register map) and hands the GT911 driver its reset (IO0) and interrupt (IO1) pins on it, so the driver runs Elecrow's
own sequence: reset low with the interrupt line low, reset high, 56 ms, address 0x5D. A V2.0 without the expander boots
its GT911 by the pull-up as before. The bus is scanned at boot for the log.
The GT911 of the CrowPanel 7.0" sits on GPIO 19 and 20, the ESP32-S3's USB pins. Firmware 0.2.79 freed them once, in a
static constructor; the Wi-Fi PHY library then kept the USB PHY on (CONFIG_ESP_PHY_ENABLE_USB, on by default on an
ESP32-S3) and the touch went quiet once Wi-Fi came up. The board file turns that option off, frees the pins again in
BOOT_TOUCH after Wi-Fi started, and a one-second watchdog frees them whenever anything takes them, with a log line.
A board file for the Elecrow CrowPanel 7.0" (DIS08070H, ESP32-S3-WROOM-1-N4R8): 800x480 through mipi_rgb, GT911 on the
ESP32-S3's USB pins (elecrow_usb_phy.h frees them, behind a flag only this board sets), the Guition's vertical sizes
across a wider page, no rotation. Its 4 MB of flash take the CYD's budget: a partition table of the component's
(smart_display: partitions: 4mb_wide_slots, 1,984 KB slots), the icon fonts at 2 bits a pixel and the TLS client only
bring the firmware to 1,865,440 bytes, 91.8 % of a slot, with camera images kept. docs/ELECROW.md has the numbers and
the trade.
The add-on reads a large board's name from the type sensor's value, the editor offers the board under New screen and
draws the wide preview, tools/check.sh builds it with the budget, render_topbar and capture_ui take the panel's size.
Firmware 0.2.79; the CYD image is unchanged (1,633,760 bytes).
This fork builds from gittea.trippedonmy.name/Xandrossian/homeassistant_espscreen: the packages, the fonts, the
component and the add-on repository point there instead of GitHub.
A single or double-width media player tile takes Display -> Album cover
on a Guition: the cover of what plays in the icon's place, like Home
Assistant's own tile. It rides in the page's strip of 0.2.91: the request
and the answer are the same, a media player is simply among the tiles.
The app fetches a player's picture only when its address changes; a page
with a camera loads at the camera's pace with the cover reused, a page of
media tiles alone loads once, and a new track (the picture's mark in the
state, part of the page's wish) asks again at once. A player without a
picture keeps its icon and is not asked for; the double-width tile keeps
its controls; the tile over the whole page keeps the card's big cover.
Firmware 0.2.78.
Host harness 31 checks (docs/TEST_RESULTS_0292.md). CYD firmware
1,633,760 B, 89.0 %.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
Live picture (issue #4): a camera or image tile on a Guition takes
Display -> Live picture with a pace of 15 or 30 s. The camera tiles of a
page share one download: the screen asks ESP Screens for them together
(esphome.screen_camera with tiles, size and bg) and the app serves one
BMP, a strip of squares with the corners rounded over each tile's own
colour; every tile draws its square out of it with LVGL's image offset
(LV_IMAGE_ALIGN_TOP_LEFT: the default centres a taller source). A third
online_image on the Guition holds the strip; it loads after the alert's
picture, a cover and the camera full screen, never under a card, in
standby or under a finger. The strip is served whole every time: ESPHome's
http_request logs a 304 as a failed request. Firmware 0.2.77.
Alert button: esp_screens_show_alert takes action and data; the app
performs the action once when the button's report (esphome.screen_alert,
ok) comes from any screen. show_alert keeps its seven fields, so every
firmware from 0.2.31 has it.
Media card and full-page media tile: a title or artist line that is too
wide rolls by (LV_LABEL_LONG_SCROLL_CIRCULAR) instead of ending in dots.
Full-page tiles are white or their own pastel like every other tile; the
tint in the state colour of firmware 0.2.62 to 0.2.76 is gone.
Host harness (27 checks) and bench Guition + local add-on on the Yellow:
docs/TEST_RESULTS_0291.md. CYD firmware 1,633,728 B, 89.0 %.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A tile has one line under the name, and LVGL cuts what doesn't fit off the end, which is where that line
often carries what you want to read. Two lines lost it:
- "Yesterday 9:15 PM" (English with a 12-hour clock) read "Yesterday 9:1…" on both boards. The tile now takes
a shorter wording of the same line when the full one doesn't fit, screen.script.yesterday_time_short; every
language whose word is short enough keeps the same text there.
- A robot's battery stood behind Home Assistant's own word for its state, so German, French and Polish showed
"Angedockt / 10…". The battery keeps its room now and the word takes the dots: "Angedo… / 10 %".
`tools/i18n.py check` measures a tile line with a 12-hour clock in its placeholder, so a translation that is
too long for the line fails there.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The climate card's fan and swing rows showed the raw mode names (Auto, Low, Medium, High) in every language.
Home Assistant names the common ones itself (Laag, Mittel, Élevée): ha-words now copies those into screen.ha, and
the card looks a mode up there; an integration's own mode still reads as its name. ha-words also says how many
words it really changed. Found in the second test pass.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The app fetched Home Assistant's state words in the screens' language only, so a screen still on 0.2.75 showed
"Speelt" among its own English words (and would get letters it can't draw). It now keeps them per language, the
screens' and English, and picks the set of the screen it writes for. Found in the second test pass on the boards.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- The CYD's weather columns hold two letters beside the icon: every language has its own two-letter days (the CLDR's
short width, Polish nd pn wt); French mar became nar there. check fails on a longer one.
- The top bar takes the language's abbreviation (sam. 19 sept., Sa., 19. Sept.), English keeps Sa 19 Sep.
- The editor's clock and date follow the screens: 7:05 PM in the top bar, 7:05 on the clock card.
- A long weather word ends in an ellipsis before the days.
- Tile lines that were cut on a screen are shorter (Zum Öffnen, Ouvrir, Otwórz, Sin imagen, ...); check measures
them with the screens' font against the tile width of both boards.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- Numbers in the mockup follow what the app sends the screens (style, grouping from five digits, the space before
%), a state without a unit stays as it is, and built-in cards carry the screens' names (Instellingen, not the
editor's Einstellungen).
- Automatic in Language & region is labelled with what it means now (Home Assistant's language decides).
- A screen's status (Synced, Loading tiles, ...) shows in the editor's language; the screens keep reporting it in
English, which the app reads.
- The doorbell example uses the alert examples of the editor's language; control labels keep their capitals.
- Editor bundle rebuilt with every language's texts.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
From the end-to-end test on the bench boards and an independent review:
- Units spaced as Home Assistant does (54% / 54 % in German and French, 18°, 21.5 °C), in the firmware and the app.
- Grouping from five digits where the language does (1234 but 12.345 in Spanish, Polish, Portuguese): the layout
message carries group_min and percent_space; an unknown number style no longer refuses a layout.
- Automatic follows Home Assistant's own language for the clock and numbers, also without a translation (Swedish:
24 hours, 1 234,5; pt-BR: 1.234,5): regions.json from the CLDR for all 63 languages Home Assistant offers.
- 12-hour clock: the top bar says 3:48 PM, the clock card 3:48 (not 03:48), Last runs follow the clock.
- The climate setpoint font has the decimal comma; the value fonts have ², ³ and µ (m³ was a box).
- Weather days: at most three letters without the dot (mié, sáb, czw); the screen took three bytes.
- Texts fall back per key through the base language (pt-BR -> pt -> en) in the firmware and the app.
- Each screen gets the app's words in the language its firmware speaks; firmware from before the languages gets
English and its own letters, not boxes.
- set_language re-reads what it writes (indentation, comments, quoted keys, CRLF, !include) and every build writes
the language first; the 24-hour pin no longer hits a fresh install and waits for the old clock switches.
- Tests: region and notation, plural parity with the C++ rules, set_language edge cases, the pin, PUT /api/language.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The cached screen list only read the entities named in SCREEN_ENTITY_NAMES, which lacked the new sensor: the
updater waited in vain after a language update, the live page kept offering it, and the nightly round would have
rebuilt the screen every night. Found in the end-to-end test.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Every error, delivery status, update result and label the editor shows is an `addon` text of en.json now, in the language
of the editor's request: its X-ESP-Screens-Language header, or ?language= for the live updates, which an EventSource sends
without headers. A text made in the background and read later (a delivery status, an update result in updates.json, why a
screen can't show an alert) is kept in English with its key (i18n.Text, i18n.english) and written in the editor's language
when it is shown (i18n.shown); updates.json keeps the English message next to the key, for an older app.
The words the app sends to the screens come in the screens' language (Settings -> Language & region): the names of the
built-in tiles, the top bar's words, dates and numbers, the history card's words, the vacuum chips, the forecast's day
names and the Identify card, and so does the notification of a stopped update round. The top bar and the history card
take a binary sensor's words from screen.ha.binary, as the tiles do, which makes test_binary_words pass again; three tests
follow Home Assistant's new words there (Detected, Clear).
What stays English: what Home Assistant gets as data (a tile event's answer, the layout sensors, entity names), the logs
and the firmware jobs' logs, the screens' protocol words, and the Claude skill, which is built from the English tables in
core.py and tile_icons.py; en.json repeats those and tests/test_addon_translations.py keeps both equal.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- The page takes the screens' words its mockup draws from their own texts: screen.ha (Home Assistant's words, filled
from Home Assistant), screen.time, screen.date, screen.number and screen.tile.page, per language next to _meta and
editor. The copies under editor.mockup go; the rest of `screen` and all of `addon` stay out of the build.
- Numbers in the mockup follow the screens' language (screen.number, with its minimum grouping) or the chosen style,
as the firmware writes them.
- @intlify/unplugin-vue-i18n compiles every text at build time, so the page runs vue-i18n's runtime-only build
without the message compiler; every language's _meta comes with the page. Empty texts stay out while building.
- The live updates carry the editor's language (api/events?language=), as an EventSource sends no headers.
- web/ needs Node 22.13 or newer: vue-i18n 11.4, Vite 7 and the unplugin ask for it.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
- vue-i18n 11 with the Composition API: English comes with the page, any other language loads the first time it is
needed from screen_manager/translations, and a text a language lacks shows in English. The build keeps only _meta
and editor of each file (vite.config.ts), never the screens' or the add-on's texts.
- The editor follows the language of the user's Home Assistant profile (<html lang> of Home Assistant's page), else
the browser's; plural forms follow the file's _meta.plural, with the screens' rules.
- Every request tells the add-on the editor's language (X-ESP-Screens-Language), so its own words can follow it.
- The mockup draws the screens' words (Unavailable, Page 2 ›, the date, "5 min ago") in the screens' language, with
their clock and number format.
- Settings → Language & region: the screen language, time format and number format for every screen, with the user's
own profile offered when it differs. The per-screen clock choice goes; the top bar points there. A screen that
still needs the new language says so in its update badge.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
tools/i18n.py gets cldr (day and month names, date orders, numbers and the clock from Node's Intl), ha-words (the
words for states from Home Assistant's own translations, English included) and lint (English left in the firmware's
code). The screens write numbers from Home Assistant in the chosen format, and every text font carries the apostrophe.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The firmware takes its texts from screen_manager/translations/<code>.json at build time (components/smart_display:
one table in main.cpp, keys in screen_text_keys.h), and the fonts carry every letter European languages write with.
The app keeps Settings -> Language & region (language, clock, numbers) for every screen, writes LANGUAGE into the
screens' YAML and sends the clock and the number format in the layout message; the per-screen 24-hour clock goes.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The add-on ships ESPHome 2026.9.0: it builds an ESP32 with ESP-IDF itself
and keeps a compiler cache. firmware.py gives the CLI its own storage in
/data (ESPHOME_DATA_DIR /data/esphome, which the ESPHome Device Builder no
longer wipes on start; ESP-IDF and ccache in /data/idf, at most 1 GB;
compilers limited to the core count as PlatformIO ran them), removes the
unused PlatformIO once before the first ESP-IDF build, and leaves the new
folders out of backups.
Firmware 0.2.75: assertion_level SILENT in every board's esp32 block, the
core logs from INFO, and clock_parts() in runtime_model.h replaces sscanf.
timer_left() caps a running timer at its duration (the screen's clock runs
up to a second behind Home Assistant's). The packages keep min_version
2026.6.2, the top-level online_image and OTA passwords for new screens, so
an add-on or Device Builder that isn't updated yet keeps building every
screen; CI now builds both boards with the add-on's ESPHome and with
min_version. ESP-IDF's error names stay on, and build_flags stay under
platformio_options, because ESPHome 2026.9's LVGL reads only those.
tools/check.sh: 478 Python tests, 19/19 C++, editor; --firmware PASS with
ESPHome 2026.6.2 (CYD 87.7 %), 2026.8.1 (88.3 %) and 2026.9.0 (86.4 %,
was 91.5 %). Both bench screens OTA'd and watched for half an hour;
ESP Screens as a local add-on on the Yellow built both profiles
(CYD 1,037 s first build, 41 s unchanged). Details:
docs/TEST_RESULTS_0289.md.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
packages/core.yaml sets wifi: power_save_mode: none. ESPHome merges it
under the screen's own wifi: block, which keeps the network, the password
and the hotspot, and the screen's own YAML still wins on the mode. The
YAML ESP Screens writes has had the line since app 0.2.24; screens with
an older or hand-written YAML ran ESPHome's ESP32 default light, and
their messages from Home Assistant waited 76-117 ms on the bench.
esphome config on the bench screens' Device Builder YAMLs: NONE for both
without a line of their own, LIGHT when the YAML says light. The bench
Guition, flashed from its Device Builder YAML: ping 15 ms average instead
of 117 ms. tools/check.sh: 474 Python tests (a new one keeps the core's
wifi: block to this one key), 19/19 C++, editor; firmware unchanged in
size (CYD 1,678,816 B, 91.5 %; Guition 2,055,280 B), since the checkout
profiles already had the line.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Starting screen: until the first layout, render() keeps the status text
("Connecting to Home Assistant", "Waiting for ESP Screens") and a spinner
as one block in the middle of the page; the first layout deletes both.
The room label stays empty meanwhile, and ROOM_NAME defaults to "".
Camera: the full screen shows a spinner instead of "Loading image". The
screen asks when the camera opens and loads a link as soon as it comes
(camera_load, never under a finger) instead of waiting for the 250 ms
tick twice, and online_image reads 16 KB per loop instead of 4 KB. On the
bench Guition with the EZVIZ camera: cold 5.8 -> 4.55 s, warm 3.4 -> 2.06 s.
Alert picture: the image takes the frame's radius (18, the alert card's);
LVGL's software renderer clips it row by row, without a layer.
One spinner_create() makes the busy sheet's, the starting screen's and the
camera's spinner; theme role CAMERA_TRACK colours the camera's ring. The
README gets a row of camera pictures; docs/CAMERA.md, RELEASING.md and
TEST_RESULTS_0287.md describe the measurements.
Checked with tools/check.sh (473 Python tests, 19/19 C++, editor) and
--firmware: CYD 1,678,816 B (91.5 %, +1,312 B), Guition 2,055,280 B
(25.3 %). Host renders of both boards in both looks; four OTA builds on
the bench Guition Wallbox, which runs this firmware.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
A tap on the On/off switch of a wide or full-page tile moved the knob to on
and back, and the light stayed off. Since firmware 0.2.59 control_event drew
the new position first (Tile::optimistic writes it into state) and then asked
key_action for the action, which read that new position: an off light got
light.turn_off, an on light light.turn_on. A tap on the tile itself was never
affected.
tile_controls::press_key chooses the action first and then shows the new
position; control_event uses it. Unit tests cover a light, a fan, a sensor
and a cover key.
Checked with tools/check.sh (471 Python tests, 19/19 C++) and compiled for
both boards: CYD 1,677,504 B (91.4 %, +16 B), Guition 2,054,640 B (25.3 %).
On Max's Guition Wallbox (OTA from this branch) every tap on the switch sent
the right action and Home Assistant's logbook followed.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
README_EXTENDED rendered everything from "Direct control" to "Override YAML" as raw text on
GitHub since 0.2.74: the full-page pictures ended without a blank line. It also gains the
media card, Page buttons, the ··· menu's Identify/Copy/Export/Import, Try it, and where
SCREEN_FIRMWARE_VERSION lives now. EASY_SETUP's settings table drops Show clock and gets
Back to page 1; SETTINGS.md points at web/src/store.ts and packages/core.yaml; GUITION.md
had the GT911 mirrored and cache settings that left the profile on 2026-09-12;
TROUBLESHOOTING still named TILE_COUNT and files deleted in 0.2.50; CALIBRATING put
Calibrate touch in ESP Screens instead of Home Assistant. RELEASING gets the compatibility
notes of 0.2.73 to 0.2.85 and the GitHub release step.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The airco tile stayed orange while the airco was off. The firmware kept its own
short list of what turns grey and gave an off climate Home Assistant's orange,
but Home Assistant has no colour for a climate that is off: its cards fall back
to the inactive grey. Other tiles had the same fault: an off fan, a closed
blind, a docked robot, a player in standby, an idle script, a paused timer and
a camera that isn't streaming all kept their colour.
Tile::active() is now a port of the frontend's stateActive() for the domains a
tile can show, and every tile that is not active is grey. An airco that is off
reads Off, with the room's temperature when the device reports one, as Home
Assistant's tile does. tile_controls::accent() follows stateColorCss(): alarm
binary sensors red, a battery green, orange or red by its charge, a running
script or timer and a streaming camera amber, the weather per condition, the
sun amber and indigo, someone in a zone blue. Scenes, selects, numbers and
sensors keep their own colour per kind. A full-page tile lights up only for
tiles that can be off (Tile::lights_up), and a closed blind's slider keeps its
colour. The top bar's alarm classes are Home Assistant's eleven.
Checked with tools/check.sh (471 Python tests, 19/19 C++ with 58 state cases
and every colour rule) and compiled for both boards: CYD 1,677,488 B (91.4 %,
+576 B), Guition 2,054,640 B (25.3 %).
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
The two board profiles (5,000 lines each, 87 % the same text) become packages/core.yaml and one file per
board under packages/boards/: the hardware, a table of 84 sizes and font sizes, 18 hooks with the lines of C++
only that board adds inside a shared lambda, and the board's own sections. packages/cyd.yaml and
packages/guition.yaml keep their names and places (a screen installed from ESP Screens builds as before);
home-like-2432s028.yaml and guition-4848s040.yaml become the same entries for a checkout.
The firmware stays 0.2.70: the C++ ESPHome generates for each board, through the checkout profile and
through the GitHub package route, was compared line by line with the old profiles' (docs/TEST_RESULTS_0284.md).
tools/check_packages.py replaces tools/generate_packages.py; tools/profiles.py gives the tests and tools the
three files as one; docs/PROFILES.md says how the files fit together and how a board is added.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A light that offers effects gets a sparkles key at the top right of its colour
card. It opens a page with a row per thing the lamp offers, all from Home
Assistant: the light's effect and the select entities of its device (a WLED's
colour palette, preset and playlist) with what they are set to, and a slider per
number entity of the device (speed, intensity). Names are Home Assistant's own
English entity translations, icons its frontend's; nothing about a brand is
hardcoded beyond a preferred order of translation keys.
A row opens LVGL's roller with every name Home Assistant lists at that moment,
asked for when the picker opens (esphome.screen_options, answered with
op: options, one page of names per message), the effects alphabetically with
Solid on top. The check key at the top right sends the choice as one action;
the back key sends nothing. A slider sends number.set_value when let go. While
an effect runs, the tile names it instead of the brightness.
Firmware 0.2.70: effects_page.h (built when it opens, gone when it closes),
Extra.effect/option_rows/number_rows, options_request, a hidden roller widget
so ESPHome compiles LVGL's roller, five firmware-only glyphs. App 0.2.83:
light_effects.py, the effect attribute, the device rows in the state message,
card_options_loop. Checked on the host with a virtual finger (35/35 on both
boards) and compiled for both ESP32 boards; CYD 1,675,424 B (91.3 %).
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
GitHub issue #9 asked to drop the Previous/Next bar when Go to page tiles do the navigation and to give the tiles the whole screen. Page buttons is a setting the screen owns (on by default): a row in the Screen group of the settings page, switch.<screen>_page_buttons in Home Assistant, and a row in the editor's Screen settings. Without the bar, and on a screen with one page, place_page moves and sizes the three rows so they reach the bottom margin (Guition 108 -> 122 px, CYD 52 -> 61 px); a card keeps its circle, name and state in its middle and the single analog clock keeps its dial. With page buttons and swiping both off, the editor says where the Go to page tiles lead and which pages that leaves out of reach.
The bar itself: each half of the band under the tiles is one key with an MDI chevron and lights up under a finger, and a dot per page sits between them; the screen's settings page uses the same pager. App 0.2.82, firmware 0.2.69.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
On a CYD the strip at the bottom of a compact card is 8 px high, so a finger landing a little above it pressed the card: the light's card opened, the light toggled, or the page-swipe gesture turned the page without a log line. The strip's click area now covers its card (slider_zone); a press that never moves forwards the card's tap or hold and shows the card pressed; the CYD's page-swipe handler and the Guition's flick handler leave a captured slider alone and log what they do; a slider commit no longer bounces against its own previous send (TouchGuard::accept_within); the CYD's edge band is 20 px because its resistive panel reports up to the bezel. App 0.2.81, firmware 0.2.68.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
A fast swipe off the right side of a Guition ended at 93-98 %: the GT911 reports the finger for the last time 33-48 px inside the edge while a wide tile's slider ends 29 px from it, and LVGL keeps the last point it saw. cyd::edge_snap() takes a release inside the band (SLIDER_EDGE_SNAP_PX: Guition 67, CYD 56) as the end that lies in that band, for the tile sliders and the light card's sliders; the log names each snap. App 0.2.80, firmware 0.2.67.
Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>
On the tile (and the strip of a full-page blind) the small slider filled with the open part, while the blind's card
draws the blind hanging from the top: a closed blind looked open on one and closed on the other. Now both fill with
the closed part and dragging right closes; the text still gives the position (open %). The editor's mockup follows.
Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
Xandrossian