The concepts behind worksmith, one page each. Start with the validation loop: why the harness exists, what 'done' means, and the measurements that bound the claim.
One binary, one bet. The other terminal agents — Codex, Gemini CLI, pi — are thin wrappers around a frontier model that mostly stays on task. Worksmith bets the other way: the harness does the work of keeping a weaker model honest. These pages explain the machinery that bet relies on.
The single most important page is the validation loop.
Everything else hangs off it. In one paragraph: a task carries a check you
named — --until "cargo test" — and the turn is not done when the model says
it is done, but when that check exits 0. When the model spins, the harness
notices and sends it back with the failure output. That is the whole product,
and it is measurable: on a small model (qwen3.5-9b) it was worth +34 points —
52% to 86% — at flat cost per solved task, because all ten of the unguided
failures had outcome done. The model declared itself finished and was wrong.
On a capable 27B the same loop changed nothing — 21/21 either way, for about
18% more tokens — which is why guidance is earned, not assumed, and why the
docs say so out loud.
/spawn into N workers, the supervisor that
is the same nudge/escalate mechanism applied to many, and the worker
validator that closes the same hole in the background..worksmith/config.toml is not
applied until you say it is, and why the decision is keyed by content, so an
edit re-asks.↻ in the footer means, why
max-tokens must cover reasoning and output, and when /fast is the
right call.The reference — CLI flags, config keys, tool descriptions, the event enum — is generated from the binary and the source, not hand-written here. When the two disagree, the generated reference is right.