MMihai_LeanZero·Sep 19The goose swarm, self-verifying: small local models shipping real softwareHow we forked Block's goose into a multi-agent swarm running three small local models on LM Studio, with a gate stack that judges work by running it.19 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 19Inside goose-swarm: How We Turned One Local Model Into a Self-Verifying FleetHow goose local edition works: the scheduler, the parallel planner, the CONTRACTS discipline, the model-free judge, and the smoke and AST gates.24 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 19swarm-gym for goose: grading local models by running the code they writeA hands-on swarm-gym tutorial: set up the goose local-edition swarm testbed, run both modes, read every output, and tune the swarm from what you find.17 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 19Making the goose swarm predictable: 602 commits, 100 levers, and three bugs I found writing this602 commits on goose local edition: what changed in the swarm, what the desktop looks like now, and three defects I found in my own honesty checks.23 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 19goose compaction: my 27B re-read the file 95% of the time — quoting it in the summary didn't helpCompaction summarised away the tool output holding the file. Quoting the file into the summary left 18/20 re-reads; the last turns verbatim, 0/20.18 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 18goose swarm: pytest | head -80 exits 0 when nothing ran, and pipefail only trades the liepytest exits 4 for a missing file; piped to head it exits 0 in bash and zsh. Why a goose swarm worker went green, and why pipefail is not the fix.20 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 18goose Local Edition: a benchmark that runs the app your agent builtOur goose fork, its swarm engine, and a scorer that boots the built app and prices what a user would get. On the current board our own fleet is last of 17.25 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 18Building a scorer that can't flatter itselfHow we built an execution-graded benchmark for coding agents, and the six disciplines that caught every number it published wrong the first time.38 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 18OpenAI's Codex spent six days on goose and never completed a single runSix days of OpenAI Codex on goose produced not one completed run. The same model in a different harness worked fine. Why the harness is the product.9 min read·leanzero.hashnode.dev
MMihai_LeanZero·Sep 18goose: the config error we logged 432 timesA duplicate YAML key failed a whole config file. The loader warned 432 times into a log nobody reads, returned Ok, and the run used defaults.8 min read·leanzero.hashnode.dev