1. What cost the most. Getting the rules block and the context took the most: I read y27-load.md and y29 in full, two long briefs whose Rules and Report sections are identical. Closing out cost the next most. The push was rejected, and then I amended my commit twice to fit a Pushed: line into the report. That was three commits for one line, and the sha in the first try went stale as soon as I amended.

2. What the brief could have handed me.
   • The To build text of #928 (I fetched it with gh and pulled out the section).
   • D6 of ADR-yaml-1.
   • The name of load’s key-comparison helper (key_of in eef3a2e7).
   • Above all, where the lent-parameter inference of ADR-277 D17 lives. I never found it, so the y30 brief only gives a low-confidence rg hint. One line in builder-facts would have done it, something like “lent-parameter inference: file:function”.
3. Checks.
   • Useful: running the #1092 probe on target/fast/nikaia, both the match form (E0507) and the ?? form (prints 3). That confirmed the defect and gave a row that must keep working.
   • Wasted: grepping builder-facts for “lent” (nothing relevant), and the broad grep across tools that listed nine files without narrowing anything.
   • Not run: rows d and e of #1092. I left them to the builder to probe.
4. One change that would save the most. Have the driver write two things into architect-input.txt: whether unpushed.txt is non-empty, which means a push will be rejected, and the path of the current rules block as a file (docs/briefs/yaml/rules.part). The architect could then write Pushed: rejected (unpushed batch) before the one commit and paste the rules with a sed. That saves reading whole briefs and amending.

This site is open source. Improve this page.