omp packages

@micuintus/aid-evolve1.0.0

Ledger-based evolution by Artificial Intelligent Design: code evolution for Pi in which an LLM, not a genetic operator, decides each step.

micuintus Published 7 days ago MIT ~199 downloads/mo

Untested on omp

This is a pi package. No one has reported whether it works on omp yet. If you try it, add your result below.

Did it work for you?

Result

One report per person. A new report replaces your old one.

Reports 0

No reports yet.

Readme

AIDEvolve

Ledger-based evolution by Artificial Intelligent Design, for Pi, sized for one developer on one workstation.

An experiment is one Markdown file, evolve.md: goal, metric, gates, stop conditions, optional models, and a ledger. Each iteration is a fresh subagent that reads the whole ledger, decides whether to refine a prior candidate, combine two, or try something unrelated, builds it on its own git branch, gates it, benchmarks it once, and appends one row. No program database, no islands, no evaluator pool: the ledger and the branches are the whole state, and you can read both.

How it compares to AlphaEvolve, OpenEvolve, ShinkaEvolve, CodeEvolve and Karpathy's autoresearch, with evidence from 325 ledger rows on a production renderer and the measured token cost: technical report.

Install

pi install git:github.com/micuintus/aid-evolve
pi install git:github.com/edxeth/pi-subagents

or pi install npm:@micuintus/aid-evolve. Recommended underneath, so a rate-limited or exhausted model falls through to the next instead of stalling a run: pi install npm:@alexeiled/pi-model-router.

AIDEvolve delegates each iteration to edxeth/pi-subagents through an agent definition without a persona. Create ~/.pi/agent/agents/evolve-iteration.md (the evolve-init skill offers to):

---
name: evolve-iteration
description: Plain agent for exactly one evolve iteration (no persona, auto-exit). Not for other tasks.
mode: interactive
auto-exit: true
---

Try it

sh examples/json-codec/setup.sh ~/json-codec-run
cd ~/json-codec-run && pi

Then type /evolve. The example evolves a deliberately slow JSON encoder against a byte-exact gate; see examples/json-codec.

Your own experiment

  1. Put the code to evolve in target/, as its own git repository.
  2. Ask Pi to set up an evolve experiment. The evolve-init skill interviews you and writes evolve.md next to target/.
  3. /evolve starts the loop, /evolve <hint> biases the first iteration, /evolve stop stops it. The evolve tool does the same for the agent.

Spend more time on the metric than on anything else: a metric with two flanks, frozen ground truth with a holdout set, and a measured noise band (section 5 of the report) is what makes the ledger trustworthy.

Engines

AIDEvolve is built against specific engines and says so. Iterations run through edxeth/pi-subagents (spawn tool subagent, auto-exit agent definitions); the orchestrator checks for a spawn tool first and stops the run instead of stalling if none exists. The engine was chosen for its inspectable child sessions, parent-model inheritance and clean self-termination; the comparison is in the Pi ecosystem subagent survey. The loop driver comes from @micuintus/loop-driver.

History

AIDEvolve was developed from May 2026 inside pi-dacmicu, which this repository was renamed from, and was briefly published as LedgerEvolve (@micuintus/ledgerevolve, now deprecated). The TODO loop moved to EnforcedTodo. The last combined state, including the experimental callback extension, is tagged dacmicu-final.

Development

npm install
npm run check
npm test

Citing

Voigt, M. (2026). AIDEvolve: Ledger-Based Evolution by Artificial Intelligent Design. Technical report. https://github.com/micuintus/aid-evolve/blob/main/paper/aid-evolve.md