How it was built.

A personal need, a weekend spike, then a six-day measured campaign — a loop of overseers and a fleet of experimenters, with a real compiler as the referee.

3.7 B
tokens through the worker fleet, at a 97.6% cache-hit rate
943
mandates dispatched — all but 11 authored machine-to-machine
281
worker lanes across 220 isolated worktrees

The timeline.

  1. Aug 22

    A weekend project that refused to stay one

    FoxLift started as a personal need: compiled Visual FoxPro 9 binaries that had outlived their source, and existing tools that did not fit a modern CLI and CI workflow. The first weekend spike cracked the containers — .fxp, .app, .exe structure — and lifted the first bytecode statements back to source.

    Rule zero was set the same day: no finding survives on model consensus. Every hypothesis is settled the same way — compile an authored probe on a real VFP9 and compare the bytecode. That compiler rig has refereed every claim since.

  2. Aug 22–28

    The fleet and the overseers

    Fleets of GLM-5.3-flash workers ran the experiments: each got one narrow task — one bytecode family, one parser region, one test file — in an isolated worktree, and stopped after one verified commit. Claude Code and Codex ran the oversight loop: plan a wave, review every lane, integrate serially, re-run the gates, re-rank what blocks next, dispatch the next wave.

    Measured across the campaign: more than 55,000 tool calls, 168 hours of model time inside 92 wall-clock hours, and a watchdog shepherding crashed lanes back to work.

  3. Aug 26

    Coverage closes

    Every non-empty method section of the development corpus lifts: 10,514 of 10,514, none blocked, none silently dropped. The benchmark corpus is entirely public VFP code on GitHub — development on luhongbin/YiFeiERP, one vendor's real ERP estate of forms and class libraries.

  4. Aug 27

    Correctness, then v0.1.0 ships

    A round dedicated to honesty: recompile every emission on the real compiler, fix what disagrees, and prove what can never come back — comments, DECLARE parameter names, compiler-canonicalised clause orders. FoxLift documents those classes and never invents them. The same day, the repo went public, v0.1.0 landed on PyPI, and this site went live.

  5. Aug 28

    Compiled programs — v0.2.0

    Standalone compiled programs joined the slice: DEFINE CLASS scaffolds in .fxp, compiled menus (.mpx), SQL decode, and non-Latin code pages. Recompiling the decompiled output on real VFP9 reproduces the original bytecode frames for 95% of the measured population.

  6. Now

    Artifact reconstruction

    The campaign continues: writing binaries back as buildable projects, so a recovered application rebuilds in VFP9. Current status, with numbers, lives on the front page.

22–28 Aug

Thanks, OpenRouter and GLM-5.3-flash.

Most of the experimental legwork ran on z.ai’s GLM-5.3-flash — available free through OpenRouter as the stealth model “Ox Alpha” from August 22 to 28. That made many long reverse-engineering loops possible.

FoxLift is created and maintained by timo-bejan.