Atmega32u4 Beetle

Beetle — Mini Arduino Leonardo (ATmega32U4)

Board-specific firmware knowledge. The Beetle is an ATmega32U4 with native USB and ten pads, and almost every trap on it comes from one of those two facts: USB means the programming interface is the sketch you are about to replace, and ten pads means most of the Leonardo pin numbering the compiler accepts goes nowhere on this board.

• reference/board-hardware.md — the complete board reference: pad inventory and full pin map, the sixteen I/O lines that are not brought out, power tree, clock and PLL, memory map, fuses and the Caterina bootloader plus a development guide (Part II: §8 toolchain, §9 USB and the two serial ports, §10 peripheral cookbook, §11 gotchas, §12 flashing and recovery, §13 symptom → cause → fix table).
• reference/recipes.md — copy-paste code: platformio.ini, board.h, non-blocking timing, USB-CDC that never blocks, Serial1 on the TX/RX pads, I2C and SPI, the ADC channel map, external and pin-change interrupts, PWM and the timer map, Servo vs tone(), HID keyboard with its safety guards, EEPROM, power-down sleep with USB.
• template/ — a project that builds clean, in three variants, plus a scaffold script. See template/README.md.

Orientation

Rules that prevent the expensive mistakes

Each of these produces a failure that looks like something else.

1. The board is two serial ports, not one. Running the sketch it enumerates as VID:PID 0x2341:0x8036; in the bootloader it enumerates as 0x2341:0x0036 — a different device with a different port name. Uploading means: open the sketch's port at 1200 baud, the chip resets, the sketch's port vanishes, the bootloader's port appears, avrdude speaks avr109 to that one. Consequences: never hard-code upload_port (it breaks the handshake); a serial monitor holding the sketch port open can block the touch; and "the port disappeared mid-upload" is the process working, not failing.
2. The upload route runs through the sketch you are replacing. The reset is performed by the sketch's own USB interrupt handler: on seeing the 1200-baud port close it writes the magic key 0x7777 to RAMEND-1 and calls wdt_enable(WDTO_120MS) on itself (core CDC.cpp). So a sketch that hangs, blocks interrupts or never attached USB cannot be reset by any host — nothing external is listening. The escape is manual and takes one touch, not a double-tap: short ICSP pin 5 (RESET) to pin 6 (GND) on the six dots on the back. Caterina enters the bootloader on any external reset and holds it TIMEOUT_PERIOD = 8000 ms — eight seconds, from its own source. Locate those dots before you need them.
3. while (!Serial); hangs the board forever off a PC. On a 328P that line is a no-op; here Serial is a USB CDC port and the operator is false until a host opens it. On battery, in a prop, in a keyboard — nothing runs. Use if (Serial) { … } around prints, or a timed wait with an exit (template/src/main.cpp). But note the second edge: operator bool() contains a delay(10), so every if (Serial) costs 10 ms whichever way it answers. One test per pass through loop() caps that loop at 100 Hz — cache it in a bool if you check it often.
4. Serial.begin(115200) sets nothing. USB CDC has no baud rate; the number is discarded. Chasing a "baud mismatch" on the USB port is chasing a symptom that cannot exist. The real UART, where baud does matter, is Serial1 on the TX/RX pads.
5. D3 and D11 PWM come out of Timer0 — the millis() timer. This is the opposite of an Uno/Nano, where D3/D11 are Timer2. Touching TCCR0A/ TCCR0B to "change the PWM frequency" on those two pads makes delay(), millis() and every baud-timed bit-bang lie, with no error anywhere. Two of the Beetle's four PWM pads are on that timer. They also run at a different frequency from the other two — ~977 Hz on D3/D11 (Timer0, fast PWM) against ~490 Hz on D9/D10 (Timer1, phase-correct), both at prescaler /64. Do not assume one pitch across the pad set.
6. Servo takes Timer1 — that is D9 and D10, the other two PWM pads. One servo.attach() and analogWrite() on D9/D10 is dead. With a Servo attached the Beetle has exactly two usable PWM pins left, both on the timer you must not reconfigure (rule 5). tone(), by contrast, takes Timer3, whose only output OC3A = D5 is not on a Beetle pad — so tone() costs you no PWM here, unlike on an Uno.
7. analogRead() takes an analog channel index, not a pin number. On this core analogRead(2) samples A2 (PF5), not the SDA pad; the core subtracts 18 only from values ≥ 18 and then runs everything through analog_pin_to_channel_PGM. Always pass the constants: A0, A1, A2, and A9/A10 for the D9/D10 pads. Any other bare integer silently reads the wrong pin, or a channel that reaches no pad at all.
8. Nine of the pin numbers you know from an Uno reach nothing here. D4, D5, D6, D7, D8, D12, D13, A3, A4, A5 all compile, all toggle real port bits, and none of them are on a Beetle pad (D13 drives the onboard LED and stops there). Library examples written for a Leonardo or Pro Micro will happily target them. Check include/board.h before wiring.
9. A HID sketch that types at boot locks you out of your own machine. The Beetle enumerates ~2 s after plug-in and starts typing into whatever window has focus — including the editor you need to fix it in — every time you plug it in. Gate every HID output behind a pin held to GND and a multi-second grace window after reset (template/variants/hid/main.cpp). Recovery once it happens is a RESET short of ICSP pin 5 to pin 6 while the host is being spammed.
10. Keyboard.h / Mouse.h are not in the core PlatformIO installs. Only the low-level HID library is bundled. #include <Keyboard.h> is a compile error until lib_deps = arduino-libraries/Keyboard is added — the Arduino IDE bundles them, which is exactly why the same sketch builds there and not here.
11. The + pad is an unregulated rail and DFRobot say 6 V destroys the board. Their numbers, verbatim: reliable at 4.5–5 V, "may work" at 3–4.5 V with reliability not guaranteed, and "6V will damage the product by overvoltage." A regulated input would absorb that extra volt, so this is as good as a statement that there is no regulator — there is also no VIN and no 3V3 output. Driving + from a bench supply while USB is plugged in ties your supply to the host's VBUS; pick one.
12. A Beetle that reboots when a motor or relay switches is a power problem, not a firmware one. DFRobot flag it explicitly: large loads must take their VCC and GND from the supply directly rather than through the board, with ≥ 10 µF in parallel, "in order to prevent restart or malfunction caused by large load transient". Chasing this as a watchdog bug, a USB bug or a stack overflow is the classic wasted evening — an unregulated 5 V rail feeding both a motor and an MCU browns out the MCU.
13. All I/O are 5 V push-pull, and there is no 3.3 V anywhere. Into a 3.3 V-only sensor that is damage, not a level mismatch. Level-shift or power the peripheral from its own rail with grounds common.
14. wdt_disable() first in setup() (and clear MCUSR). Caterina uses the watchdog itself to jump from bootloader to sketch, so a sketch that leaves the WDT armed with a short timeout reset-loops faster than the 8 s window can be caught — and re-uploading does not fix it, only a caught ICSP reset does.
15. A factory-fresh ATmega32U4 has JTAG enabled, and JTAG is A0–A3. Default JTAGEN is programmed (enabled); the Arduino Leonardo fuse set (HIGH = 0xD8) turns it off. On a re-fused or badly-fused clone, A0/A1/A2 — three of your ten pads — read garbage and cannot be driven. Symptom: exactly PF4–PF7 dead, everything else fine. Fix by re-burning the bootloader, or at runtime by writing JTD in MCUCR twice within four cycles.
16. Budget flash from ~28 KB, and expect ~4 KB gone before you write a line. Caterina keeps 4 KB of the 32 KB, and the core enumerates USB from main() whether or not you mention Serial: the template's bare Blink is 3,956 B here against 924 B for the same sketch on a Nano.

When the task is USB HID

This is what the board is bought for — macropads, foot pedals, cosplay props — and the whole difficulty is that HID and the upload path share one USB device.

• Keyboard.begin() adds a HID interface alongside the CDC one; the board stays a composite device and uploads keep working. It does re-enumerate, so the host may briefly re-assign the port.
• HID carries key positions, not characters. Keyboard.print("y") on a host set to a non-US layout produces whatever key sits at that position. Restrict yourself to letters and digits, or map layouts yourself.
• Keyboard.releaseAll() on every exit path. A key left down by a reset mid-press repeats into the host until it is unplugged.
• The arm-pin + grace-window shape in template/variants/hid/main.cpp is not a nicety; it is the difference between a board you can iterate on and one you recover with tweezers. Keep both guards while developing, even if the shipped firmware drops them.
• Mouse.h is a separate dependency (arduino-libraries/Mouse) from Keyboard.h. Joystick/gamepad needs a third-party descriptor library — none is bundled.

When the task is memory

2,560 B of SRAM, of which the USB stack has already spent ~150 B before setup() runs — the template's blink reports 149 B used with no globals of its own. The arithmetic that matters:

• Every Serial.print("…") literal lands in SRAM unless wrapped in F(). The full template — heartbeat, CDC report, ADC, EEPROM, free-RAM watch — totals 5,214 B flash and 164 B RAM precisely because every literal is F()-wrapped.
• The 832 B of USB DPRAM is separate silicon and does not come out of the 2,560 B. Endpoint buffers are not your SRAM budget; CDC ring buffers are.
• String fragments the heap and there is no MPU: the heap grows up, the stack grows down, and the collision shows up as random corruption minutes or days later, never as an error. Use snprintf() into fixed char[].
• Instrument, don't estimate. The free_ram() probe in template/src/main.cpp costs ~20 B of flash. Below ~250 B free, restructure instead of adding.

Starting a new project

Do not hand-assemble one. template/ builds clean; scaffold from it:

~/.claude/skills/atmega32u4-beetle/template/variants/new-project.sh <target-dir> [--full|--minimal|--hid]
cd <target-dir> && pio run

• --minimal — blink the onboard D13 LED. 3,956 B flash, 149 B RAM. Flash this first on an unfamiliar Beetle: it touches one GPIO and nothing else, so a failure is the toolchain, the port, the bootloader window or a clone with no LED fitted — not your code.
• --full (default) — non-blocking heartbeat, USB-CDC report guarded by if (Serial), free-RAM watch, ADC on A0, EEPROM boot counter. 5,214 B flash, 164 B RAM.
• --hid — USB keyboard with the arm pin and boot grace window from the section above; the script appends lib_deps = arduino-libraries/Keyboard. 6,206 B flash, 231 B RAM.

All three build as-is with platform-atmelavr 5.3.0 (verified — those are the build's own figures). Nothing is generated and no paths are embedded, so copying the tree by hand works identically. template/README.md maps files to subsystems.

When the user already has a project, prefer bringing it in line with the template's platformio.ini and include/board.h over rewriting their code.

Flashing

Over the micro-USB cable, nothing to press while the sketch is healthy:

pio run -t upload -t monitor

PlatformIO opens the sketch's CDC port at 1200 baud (use_1200bps_touch), the chip resets into Caterina, PlatformIO waits for the bootloader's port to appear (wait_for_upload_port) and avrdude flashes it over avr109 at 57600. Leave upload_port unset so that dance can happen (rule 1).

Recovery ladder, in order — stop at the first that works:

1. Retry, watching pio device list before and after. Two different port names appearing is the handshake working.
2. Close the serial monitor and any other program holding the port.
3. Short ICSP pin 5 (RESET) to pin 6 (GND) once, then upload inside eight seconds. No button on this board — the six dots on the back are the reset. One touch is enough; Caterina enters the bootloader on any external reset and waits TIMEOUT_PERIOD = 8000 ms. This is the fix for a crashed sketch, a WDT boot-loop (rule 14) and a HID sketch spamming the host (rule 9). DFRobot's tip for a board whose sketch killed USB: click Upload first, then touch the pads — "you need try several times, because the correct time is not easy to be caught".
4. Pass the bootloader port explicitly for one upload: pio run -t upload --upload-port <the port that appeared>.
5. ISP. The same six dots are a standard ICSP footprint — a second Arduino running 11.ArduinoISP or a USBasp. This bypasses USB and the bootloader entirely, so no firmware permanently bricks this board. Burning the bootloader also restores the fuse set — LOW 0xFF, HIGH 0xD8, EXT 0xCB (Arduino's boards.txt values; some tools read the extended byte back differently because its top four bits are unimplemented) — which is the fix for a clone with JTAG wrongly enabled (rule 15) or a mis-set clock fuse.

Reporting

Three tiers of evidence, in descending order. Pass the distinction on rather than presenting it all in one voice.

1. Read out of the sources the toolchain itself uses — exact. variants/leonardo/pins_arduino.h (pin↔port map, PWM↔timer map, the analog_pin_to_channel_PGM table, the interrupt map), wiring.c (Timer0 fast-PWM and Timer1 phase-correct at /64, hence 977 Hz vs 490 Hz), Tone.cpp (Timer3), ServoTimers.h (Timer1 only), CDC.cpp (the 1200-baud close trigger, magic key, wdt_enable(WDTO_120MS), operator bool and its delay(10), begin() discarding its argument), boards.txt (VID/PIDs, fuses, avr109, 28672/2560), the bundled-library list, and bootloaders/caterina/Caterina.c (TIMEOUT_PERIOD = 8000, bootKey = 0x7777 at 0x0800, the EXTRF/PORF/WDRF reset-cause logic — and the absence of any double-tap detection). Plus the ATmega16U4/32U4 datasheet Atmel-7766J for the memory map, JTAGEN defaulting to programmed (Table 28-4), BOD levels (Table 8-1), the PLL and the package pin numbers (Figure 1-1). All three template variants were built; every flash/RAM figure is that build's own report.

2. From the vendor's documentation — authoritative for the board, but a wiki page, not a schematic. DFRobot's Beetle SKU:DFR0282 documentation supplies the pad↔pin table transcribed in reference/board-hardware.md §2.1, the supply limits (4.5–5 V reliable, 3–4.5 V unguaranteed, 6 V destroys it), the ≥ 10 µF load-transient warning, the six-dot ICSP array on the back and its use as the reset, and the "magic light blue soft BLINK indicator" that establishes the D13 LED exists.

3. Unresolved — say so when it matters.

• Clone variance. DFRobot's board is 20 × 22 × 3.8 mm; the reseller listing for the board this skill was written against says 21 × 28 mm. Same advertised interface counts, so the pin table almost certainly transfers, but the outline is redrawn and the ICSP dots or LED may sit elsewhere. reference/board-hardware.md §2.4 confirms a pad map in ten minutes.
• No schematic exists for any Beetle, DFRobot's included. "No regulator" is inferred from "6 V will damage the product" — sound, but inference.
• TX/RX LEDs. The core drives D30/D17 with known polarity; whether a Beetle fits those diodes is undocumented, and there is little room at 20 × 22 mm. Do not rely on them.

Nothing here was run on a physical Beetle. Everything marked ⚠︎ compile-checked only in reference/recipes.md was compiled against this core with -Wall and no warnings, and not executed. If the board disagrees with this skill, the board is right — that is a gap to report, not a fault to debug around.