perf

Prompt file imported from zoolutions/stationery (.claude/commands/perf.md). Fill in {{arguments}} before use. Copyright stays with the author.

Performance

Measure, don't guess: a same-machine before and after, or no claim.

1. main in a worktree (before)

git fetch origin main
git worktree add --detach .claude/worktrees/perf-main origin/main
cd .claude/worktrees/perf-main && bundle install --quiet
bundle exec rake bench > ../perf-before.txt
bundle exec rake metrics > ../perf-before-metrics.txt

2. The branch (after)

bundle exec rake bench > .claude/worktrees/perf-after.txt
bundle exec rake metrics > .claude/worktrees/perf-after-metrics.txt
git worktree remove --force .claude/worktrees/perf-main

One script alone: bundle exec ruby -Ilib benchmark/<name>.rb in both trees. Where time goes: PROFILE=1 bundle exec ruby -Ilib benchmark/profile.rb. What long documents hold: bundle exec rake memory (not in CI).

3. Report honestly

• rake bench gives time against Prawn (and sghtmltopdf when installed): i/s with its ±; within the ± is "within noise".
• rake metrics gives allocations, pages and bytes per fixed document; these are what CI holds (time is noise on a shared machine). Pages must not change; bytes and allocations of a document that does not use the change must not move.
• Say whether you measured a single operation or a whole render.
• No baseline: say "measured after only".

4. Keep it

• Before and after in the pull request body
• rake metrics passes; a baseline recorded again (rake metrics:update, the baseline's Ruby) only for a change on purpose, with the reason in the commit
• A perf: bullet under ## Unreleased when a user would notice

Argument ({{arguments}}): a script name runs only benchmark/{{arguments}}.rb in both trees.