Major Order Desk Monitor

The carousel

Every card is built around the planet

The target's biome fills the plate, the planet's identity is set over the artwork rather than in a row above it, and the progress tracks sit beneath. When an order names several planets the card cycles through them. When there is no order, the same card shows the five busiest liberation campaigns in the galaxy instead — so the screen is never idle while the war is on.

Band	What it carries
Header	Objective type, the LIBCON chip, carousel pips, link state
Art	The biome plate, with the planet's name, sector and headline stat set over it; the faction mark on a dark disc; and, on order cards, a countdown plate
Bars	One track for a liberation, two for a defence — each captioned with what it measures, its value and its measured rate
Strip	Diver share, divers present, the players' push, the enemy's regen
Footer	Divers present, the order's reward, last sync

An active order owns the screen

Its targets are the only pages, and the campaigns feed is not even fetched. High Command's orders are the point of the device; rotating away from them to show a planet nobody was told to take buries the one thing that matters. Swiping pages by hand, wrapping at both ends — but the 7 second timer keeps running underneath and any swipe restarts it. Touch is an accelerator, not a replacement.

Event screens

Two things get the whole panel

A new Major Order arriving, and the verdict when one ends. They interrupt the carousel, stay up until they are acknowledged, and are the only screens on the device with no chrome. Each one plays a clip through the board's speaker as it appears.

Full-screen NEW MAJOR ORDER briefing on the device

Full-screen ORDER COMPLETE screen with a green band

The API has no outcome field

An assignment that ends simply stops being listed, so the verdict is inferred from the last state the device saw before the order vanished: gone with every task complete is a success, gone otherwise is a failure. Completion is tested before the deadline, so an order finished with an hour to spare is not called a failure the moment its clock runs out — and the check only runs on a poll that actually succeeded, because a dropped connection must never read as a lost war.

Hardware

One board, no soldering

A Hosyond 4″ ESP32-32E display module — the LCDwiki E32R40T. Panel, touch controller, amplifier and MicroSD slot are all already on it. Power is USB-C.

Part	Detail
Board	Hosyond 4″ ESP32-32E display module (LCDwiki E32R40T)
MCU	ESP32-D0WD-V3, dual core, no PSRAM, 520 KB SRAM, 4 MB flash
Panel	ST7796S, 320×480 native, driven at rotation 1 → 480×320 landscape
Touch	XPT2046 resistive, on the panel's SPI bus (CS 33) — swipe to page, tap to dismiss
Audio	GPIO26 internal DAC → FM8002E amp, GPIO4 enable (active low)
Storage	Onboard MicroSD on VSPI (18/19/23/5). Entirely optional.
Power	USB-C, 5 V. No battery circuitry.

Pin definitions live in platformio.ini under build_flags, not in a TFT_eSPI User_Setup.h — nothing in the library folder needs editing.

Verify the pinout first

Hosyond does not publish its own datasheet, so the mapping is taken from the LCDwiki listing and cross-checked against the ESP32-Cheap-Yellow-Display community pinouts. Blank screen with the backlight on usually means the driver, not the pins; blank with the backlight off means TFT_BL — the single most likely pin to differ. The README walks through all five failure modes.

Under the hood

Decisions worth writing down

Most of what is interesting here is not the feature list — it is what the constraints of a 4 MB part with no PSRAM forced.

The touch panel was never miscalibrated

TFT_eSPI's own touch layer produced a panel where calibration appeared not to hold — taps landed erratically, swipes barely registered. The cause is in validTouch(), which requires two successive samples to agree within 20 ADC counts.

That is a stationarity test, and a moving finger cannot pass it. It is a reasonable filter for buttons and a fatal one for gestures: most of a swipe is rejected outright, and the readings that survive are whichever ones caught the finger momentarily still — which is why the symptom presents as bad calibration rather than as dropped input.

So the firmware keeps TFT_eSPI for bus arbitration and raw conversions, and replaces only the validation layer: a pressure gate first, a median of five instead of a deadband, the burst bracketed by the pressure gate at both ends, and a release that has to stay quiet for 70 ms.

Two app slots, or no OTA

Flash sits at 96.1% — high because the partition table splits 4 MB into two 1.9375 MiB app slots, which is what over-the-air updates need: the running image stays put while an update is written into the other.

Adding the SD/FAT stack pushed the stock min_spiffs.csv to 98.5%, so the project carries its own table that drops a SPIFFS partition it never mounted and gives 64 KB of it to each slot. A unit already in the field needs one USB reflash to actually get that room — an OTA writes the app slot and nothing else.

The previews cannot drift

Every screen on this page was produced by tools/preview.sh, which compiles the firmware's own renderer against a small host shim and rasterises the result to a PNG. Nothing in the shim describes the layout, so there is no second copy to keep in step.

Its predecessor was a hand-maintained Python reimplementation of the same layout. It fell two commits behind and started rendering a HUD the device no longer drew — exactly the thing a preview exists to prevent. Between the renderer harness and check_layout.py, six real layout bugs were caught before any hardware existed.

Direct drawing, not LVGL

A full 480×320 16-bpp framebuffer is 307 KB against 520 KB of SRAM, most already spoken for by the WiFi stack and TLS. The HUD is a fixed, nearly-static arrangement of text and rectangles with no scrolling or animation, so the abstraction buys little.

Instead the renderer redraws only what changed, via a content signature rebuilt each frame; the countdown, WiFi indicator and footer each keep their own. Text boxes are composited in small transient sprites and freed immediately. Result: 48 KB static RAM, leaving roughly 280 KB of heap for TLS and sprites.

Stale data beats no data

The screen is never blanked or replaced by an error message. A failed fetch keeps the last good data up and switches the footer to a dimmed STALE — LAST hh:mm. Repeated failures back off from 15 seconds to a 10-minute ceiling; a dropped WiFi link turns the header indicator red and nudges a reconnect every 15 seconds.

Deploy

USB once, then never again

Everything — platform, board, libraries, display config — is pinned in platformio.ini. The first build downloads the toolchain and takes a few minutes; later builds take seconds.

# requires PlatformIO Core
git clone https://github.com/codydoerfler/hd2-desk-monitor.git
cd hd2-desk-monitor

pio run                 # build
pio run -t upload       # build + flash over USB-C
pio device monitor      # serial log @ 115200 baud

Flash over USB-C

The only time a cable is required. This also writes the partition table, which an OTA cannot.
Join it to WiFi

No credentials are compiled in. On first boot the device raises a captive portal — join HD2-Monitor, password helldive, pick your network and set a UTC offset. Stored in NVS, survives power cycles. If nobody configures it within five minutes it reboots and retries the saved network, so an unattended unit recovers on its own after a router outage.
Set a real contact header

The community API requires every request to identify its caller and returns HTTP 400 without it. HD2_CONTACT_HEADER ships as a placeholder — replace it with your GitHub handle or an email. This is the courtesy that keeps a free community API open.
Leave it alone

Once an hour the device asks GitHub for the newest release and, if the tag is a higher version than the running image, downloads firmware.bin into the inactive slot and reboots into it. A download that fails halfway leaves the running firmware untouched.

Optional — the SD card

Without a card the device behaves exactly as it always has; the only difference is one [sd] no card line in the serial log. With one, the event screens get their own audio from /audio/mo_new.wav, mo_success.wav and mo_failure.wav. A clean checkout synthesises plain terminal tones, so it produces a working card without shipping game audio nobody can redistribute.

Helldivers 2Major Order Desk Monitor

Every card is built around the planet

Two things get the whole panel

New Major Order

Order complete

Order failed

One board, no soldering

Decisions worth writing down

The touch panel was never miscalibrated

Two app slots, or no OTA

The previews cannot drift

Direct drawing, not LVGL

Stale data beats no data

USB once, then never again

Flash over USB-C

Join it to WiFi

Set a real contact header

Leave it alone