1. What cost the most time: the test runs. The it filters took about 4 minutes, tests/compiler --both-settings about 10, and the single it re-run another 3. Following up the one it failure took three turns: my first run used the wrong path (tests/build_time instead of tests/build-time), then I ran the package on my binary copy, then the test alone. The cause was another agent rebuilding the binary mid-run, not my change.
2. What I looked up that the brief could have given me:
   • nikaia test has no name filter. Step 4’s “by name” can’t be done, so the brief could just say “run the whole project”.
   • The package directory is tests/build-time, with a hyphen.
   • The brief did not say other agents might change .nika files during the run. Their lowered files then show up in git status and look like the drift that step 2 says should stop the brief.
3. Checks run:
   • Run and useful: the build before applying the patch, lower-std twice plus an md5 comparison to show the lowered files were stable, cargo check nikaia-std --features compiler, fmt, clippy on 1.97 (all targets) and 1.99.1 (-p nikaia only), the it filters, the whole tests/compiler, and pre-push-check.
   • Partly wasted: the single it re-run (3 minutes). Running tests/build-time on my binary copy had already shown the failure was not mine.
   • One correction: I wrote Gate: green in the report, but I never ran the gate. That field only reflects the cheap checks, so it should be read as “not run”.
4. One change that would save the most: builder.md should say to snapshot the lowered files’ md5 right after the fixpoint, and to compare later drift against that snapshot. Drift that appears after it, beside .nika files another agent has modified, is theirs: leave it unstaged and do not stop. The brief’s step 4 should name the real commands (the whole tests/compiler, tests/build-time) instead of name filters that don’t exist.

This site is open source. Improve this page.