Elution — Timer-Bound AI Research Loop for Code Optimization

Elution is a skills-native, timer-bound AI research loop with BM25 memory for benchmark-driven code optimization. Type /elution 120 in OpenCode, Claude Code, or Pi: subagents research → propose → review → implement → measure until the deadline.

Born for the Yukon competition, extensible to any repo with a measure script.

Yukon result

Record-holder run on matrices.fast (sparse elimination ordering): 0.86136× cheapest ordering, rank 01/11.

Install

pip install git+https://github.com/gdonninelli/elution.git
# or: uv tool install git+https://github.com/gdonninelli/elution.git
# dev: uv sync --group dev

Update

pip install -U git+https://github.com/gdonninelli/elution.git
# dev: git pull && uv sync --group dev

Uninstall

pip uninstall elution
rm -rf .elution  # generated run state only

Usage

elution init --minutes 120 --baseline 1.0 --direction minimize
elution search "sparse carry path" --top 8
elution add --kind knowledge --text "..." --source "https://..."
elution deadline   # running | expired — timer is the only stop rule
elution best --show

In your agent harness:

/elution 120

How it works: researchers gather facts → lead proposes ONE change (after search --top 8) → reviewer approves/refines/rejects → coder makes the smallest diff → you measure (yukon run or your benchmark) → save to memory, repeat. Submit only on >0.10% gain over best-so-far, else once at deadline. Details: skills/elution/SKILL.md, docs/features.md, examples/yukon-generic.md.

For AI agents: read AGENTS.md (repo guide) and AGENT.md (quick pointer), then skills/elution/SKILL.md.

License

MIT. See LICENSE.