1. Most turns and time: the full tests/compiler run (about 10 minutes), then chasing its 7 failures. Six came from another agent’s rebuild deleting target/fast/nikaia mid-run. When I reran them alone, they failed on both binaries, and that was my own fault: I had named my copies nikaia-old/nikaia-new. The script tests only accept a running binary named exactly nikaia; otherwise they fall back to the relative target/fast/nikaia, which doesn’t exist in a scratch project. That cost three extra runs.

2. Lookups the brief could have handed me:
   • The issue texts: I fetched them with gh api (REST), so the brief should paste the probe and the To build list.
   • Where the Some wrapping is decided: check_function_wrap_tail / check_widened_slot set a wrap of the result, and an unknown type gets no Some. The brief could have named this.
   • Where loops are walked: the loop arms of tier_statement in text_tiers.nika.
   • The words table: how a grammar entry takes it (ADR-388, the grammar_words_required_kept test).
3. Checks run:
   • Probes on the old binary and several shape variants with --tethers. These were the most useful: the d variant showed the real cause was the loop’s back edge, not “the first initializer”.
   • One build plus the fixpoint check, both new tests old vs new, the full tests/compiler, 203 it tests, clippy, fmt, pre-push.
   • Wasted: the first alone rerun of the 7, both binaries, with the wrongly named copies. The full tests/compiler run was arguably more than “the touched tests”, but the tier change affects every loop.
4. One change: in the builder-facts entry Test with your own copy of the binary, add: “keep the copy in its own directory and name it exactly nikaia ($S/new/nikaia); the script tests check the basename and otherwise fall back to target/fast/nikaia.”

This site is open source. Improve this page.