1. What cost the most: pinning down the two defects. The NK2605-after-close() probe took about 8 calls, because the cleanup_points it test looked as if it covered the pattern. In fact it lowers without the checker. The &'r str lowering probe took about 3 more. Next came the shared checkout: another agent renamed an untracked test and rebuilt target/fast/nikaia while my run was going, so tests/compiler had to run three times.
2. What I looked up that the brief could have handed me:
   • The calling shapes: how comptime constants lower (pub(crate), so Rust can’t see them), how String? and fn(..) parameters lower (&Option<String>, &impl Fn), and that Manifest.build is private.
   • That std::fs::open exists with append.
   • The refuse! versus anyhow! error mapping.
   • That a text slice needs .bytes[..].text(), which throws.
3. Checks run:
   • Useful: lower-std with the old binary (twice, caught refusals in seconds), one cargo check, one build cycle, and one gate run (lib tests, the it filters plus the guard set, clippy, tests/compiler).
   • Wasted: two tests/compiler runs killed by the other agent’s changes, and an early nikaia check on a single tools file, which can’t resolve the tools-wide names.
4. The one change that would save the most:
   • Flag drafts that come out as methods. The draft was method-shaped (impl Settings) although the brief said “free functions over words”. At 6 % kept, it was close to useless.
   • Either convert with a free-function profile, or have the brief say “the draft’s signatures are wrong; keep only the bodies”.
   • Also, builder.md should tell testers to copy target/fast/nikaia to their scratch directory before running long tests, because other agents rebuild it in the shared checkout.

This site is open source. Improve this page.