I trialled Test and Perf improvers more today in some repos

• https://github.com/dsyme/librapid/ - a fork of a pretty clean repo, though with long build/test times
• https://github.com/fsprojects/FSharp.Control.TaskSeq - a library but with build/test system that needs updating. Not great results

I think "Perf Improver" would benefit from an initial deep SWE research into the repo

• How performance testing is done in the repo?
• How to do micro benchmarks in the repo?
• What are typical workloads for the library/app/product?
• Where are performance bottlenecks?
• Is perf I/O, CPU or Storage bound?
• What do the repo maintainers care about most w.r.t. perf.?
• What are realistic goals for Round 1, 2, 3 of perf improvement?

Doing this agentically and writing a report would probably be good enough.

Once the research is done and written up as some .github/perf-notes.md then subsequent iterative attempts at improvement could read this. I think things would go more smoothly.

Perf is a pretty clear journey for, say, numeric libraries. But much harder for more messy software.