1. What cost the most. The shared checkout. Other sessions rebuilt or deleted target/fast/nikaia while my tests ran, so I made binary copies. One copy picked up another agent’s half-finished lowered files and gave 37 false failures. After that I re-ran the touched tests, ran the gate again, and sorted each failure into vanished binary, flaky NK2302 or real. That took more time than the move itself. The step-5 lowering defect came second: probing it down to 12 lines took several rounds.
2. What the brief could have handed me.
   • That deleting a_wall leaves check/mod.rs with an unused import, a file outside the brief’s list.
   • That the badge moves (82.4 → 82.5) and badges.py must run before pre-push.
   • Which rs2nika.names [types] rows the new types need.
   • That grammar is a reserved Nikaia word, which is why the field is outer. I worked that out rather than reading it.
   • Whether the ?? chain in step 5 lowers cleanly. One check before handing out the brief would have caught the defect.
3. Checks.
   • Useful: the it filters (comptime, grammar_at_build_time, dsl_grammar, build_time), tests/build-time on its own, clippy and the guards inside the gate, and the probe for step 5.
   • Wasted: the first tests/compiler run on the 11:05 binary copy (stale lowered files from another agent), and the gate’s test step, which died when the binary vanished. A --keep-going gate also repeats work I had already checked by hand.
4. The change that would save most. Let each builder build its own binary. The gate and the tests should run on that binary (or a pinned copy), not the shared target/fast/nikaia. gate.nika --binary PATH would do it. If that is not possible, builder.md should say to copy the binary right after the build and point every test run at the copy. Close behind: a brief should only include a one-line expression whose lowering someone has already checked. Otherwise it should say “may hit a lowering defect; stop at once”.

This site is open source. Improve this page.