OpenAI's work on Git for large repositories
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Ted Nyman fixed a Coccinelle static-analysis loop that exhausted CI's six-hour timeout by declaring separate loop indexes, preserving prior branch-deletion behavior while letting the analysis finish normally.
- Taylor Blau rewrote bitmap selection to weight commits by the history they expose rather than by path length, dropping MIDX bitmap write time from 984.73 seconds to 200.06 seconds with the same 578 bitmaps selected.
- Taylor Blau extended Trace2 so parallel checkout now logs per-worker item counts, bytes, failures, elapsed time, and the slowest item, letting users distinguish uneven work distribution from a single slow file.
- Taylor Blau closed a race where bulk content checks and the final filesystem-monitor query saw different file states by holding scanned descriptors and metadata until the final check completes.
- Taylor Blau documented an optional
.cstscache file that records a known-cleangit status --porcelain=v2result, valid only after rechecking the index, HEAD, config, attributes, ignore rules, and filesystem-monitor history.
Why it matters: Large monorepos — the kind run by every major tech firm — get measurable, concrete speedups: the MIDX bitmap write alone drops from ~16 minutes to ~3 minutes, and the new Trace2 checkout counters finally tell operators whether slowness is systemic or one bad file. Caching `git status` results via `.csts` could meaningfully shrink latency for tools that poll repo state constantly.
