1. Most turns and time. The failure line only named the test, not the offending call. I had to add temporary debug prints and then bisect sysroot_package.nika through about five 55-second release test runs. A debug build also failed on a recursion limit, which cost one detour before I switched to release.

2. What the brief could have handed me.
   • The test’s output holds only a count, so the useful hint was “bisect by truncating the suspect file, because unanswered keeps no names.”
   • The file named in the log was the file the new commit touched, so “check the sysroot_package file first” would have been enough.
   • I had to look up that lower-std must be run after a .nika change, which binary to use, and that nikaia run refuses files inside a project.
   • I also looked up that CI builds release, and that a debug build fails locally.
   • I did not know the shared checkout holds other agents’ uncommitted emit_core_* edits. I had to check they would not collide with my lower-std run.
3. Checks.
   • Useful: the release ceiling test, the 549-test tests/compiler run, the sysroot tests and pre-push.
   • Wasted: the debug-build attempt, and the second compiler test I wrote. It could not compile in Rust (mixed types in an or-pattern), so I deleted it.
   • Weak: the kept test passes without the fix, so only the ceiling test guards the change.
   • Skipped: clippy, because only generated Rust changed.
4. One change. Make the ceiling test print which file and which method call is unanswered when it fails, or add a note to qa.md saying “bisect by truncating the file and rerun with --release.” That would have saved most of the turns.

This site is open source. Improve this page.