The idea: parallel drafting is why the toolkit arm was 2x faster, so keep it and remove the friction:
- Lock or queue mxbuild across agents in
exec.sh. 2 of 11 failed applies were concurrent-build collisions (rows 16 and fix-39; row 119 was first counted as one but was a glyph error).
- Batch the gate. Run one mxbuild per 3–5 clean scripts, and go back to per-script after a failure. The per-script apply grew from 22–30 s to 55–65 s as the model grew.
- Fix the usage measurement. Read each message's final usage record. Sub-agent output was undercounted in both arms, which briefly led to a false conclusion.
Field evidence: contrib/inbox/2026-09-29-parallel-build-hygiene.md (#165).
Note: exec.sh is a blocking instrument, so the lock has to follow CLAUDE.md rules 6–7.
The idea: parallel drafting is why the toolkit arm was 2x faster, so keep it and remove the friction:
exec.sh. 2 of 11 failed applies were concurrent-build collisions (rows 16 and fix-39; row 119 was first counted as one but was a glyph error).Field evidence:
contrib/inbox/2026-09-29-parallel-build-hygiene.md(#165).Note:
exec.shis a blocking instrument, so the lock has to follow CLAUDE.md rules 6–7.