1. Most turns: working out what the driver’s CUT b22 line meant. architect-input.txt named a report from 2026-10-31. The batch I had just reviewed was b29–b32, all done. Comparing file times, git status and the State section took about 4 turns before I saw the cut was a stale report, not a new failure.
2. What I looked up that could have been handed over:
   • which reports belong to the last batch (the driver could list them by name);
   • whether run 1’s uncommitted edit was still in the tree;
   • the badge rule (it was already in docs/builder-facts.md, but I had to search for it);
   • where the text-tier code lives, to trace b30’s Found;
   • how badges.py --check interacts with the pre-push check.
3. Checks run:
   • Useful: the b30 Found probe plus its one-arm variant (reproduced it, needed before filing #1132), and re-running the b33 and b21 probes on today’s binary (cheap, and confirmed the briefs still hold).
   • Partly wasted: where.nika failed (nikaia is not on PATH), so I fell back to grep. Reading badges.py was also wasted, because builder-facts already covers the trap.
   • Fetching the open issues was needed, but most of the 264 entries didn’t matter for a batch of three.
4. Change that would save the most: the driver should write into architect-input.txt which briefs are cut, done or unrun, with each report’s time, and flag a report older than the last architect run as stale. Separately, the driver should move a stale partial report to .runN itself. A smaller fix: name the binary (target/fast/nikaia run scripts/where.nika) in the architect task, not a bare nikaia.

This site is open source. Improve this page.