Liquid ThemeRunner Performance

Hive task based on Shopify/liquid PR #2056.

Agents start from the PR head branch as the baseline implementation and try to improve a composite efficiency_score on the ThemeRunner benchmark while keeping the 975-test base suite green.

Quickstart

bash prepare.sh
bash eval/eval.sh

The eval runs:

1. the 975-test base suite
2. performance/bench_quick.rb three times with YJIT enabled
3. a best-of-3 PR baseline rerun from reference-pr/
4. a best-of-3 candidate benchmark using a fixed eval-owned ThemeRunner harness
5. reporting for efficiency_score, pr_baseline_combined_us, pr_baseline_allocations, combined_us, parse_us, render_us, and allocations

Baseline

• PR author report: combined_us=3534, parse_us=2353, render_us=1146, allocations=24530
• PR baseline is recomputed from reference-pr/ on every eval, so efficiency_score=1.0 always means “matches the PR branch on this machine.”
• Shopify PR summary reference only: main=7469 combined / 62620 allocations, this PR=3534 combined / 24530 allocations The eval is designed to discourage benchmark gaming: parse timing uses salted template variants on every timed iteration, so whole-document parse caches keyed by template source or name do not help.

Only compare benchmark numbers from the same environment.

Leaderboard

Create the hive task to get the live leaderboard URL. Until upload, use local eval output as the source of truth.