1. Time and turns. The run was short, about 12 tool calls. The most expensive step was the first full read of describe.nika and project_members.nika. One Read returned a 53 KB dump I had to re-read in pieces. The build then failed on #1144, which the brief said b36 would already have fixed.

2. What I looked up that the brief could have given me.
   • The type of ledger.sources and Sources.files. They turned out to be plain BTreeMap<String, String>, but I only checked this late.
   • Where the probe files were. They sit in another session’s scratchpad (b7e2d91b…), and I found them with find /.
   • Whether the b36 fix was in my binary. The brief said it would be, but nothing told me how to check, and I only learned it wasn’t from the failed build.
   • Two forms the brief didn’t cover: comparing words.next() against a String (NK1102), and assigning an owned string to a let mut that started as a literal (NK1105). Each cost one lower-std cycle.
3. Checks.
   • lower-std with the old binary: it caught the NK1102, NK1105 and NK2702 refusals in seconds. That was useful.
   • cargo build: it found the #1144 failure. Useful, but it came after I had written the Rust side too.
   • I never reached the tests or clippy.
   • Wasted: the Rust forwards and the self_hosting.py row, which I wrote before knowing the build would fail. I reverted them.
   • Wasted: the large describe.nika output.
4. One change. Give a brief that depends on an unlanded fix a pre-flight line, like the “match by subject” rule: git log --grep=1144 plus a 3-line probe against the current target/fast/nikaia. If the probe fails, the brief should say stop before any edits. That would have shown the block in 2 turns instead of 12, and left no revert.

This site is open source. Improve this page.