A complete, self-hosting harness of high-quality skills and process for building reliable, long-running AI coding agents that ship production code across many sessions.
Every skill in this repository is written with the same ruthless standards we apply to production Rust: functional purity, stratified design, zero technical debt, and handover-clean artifacts.
The harness exists so that agents can do serious work across days and weeks without losing the thread — stacked PRs, multi-tier verification, self-critique, and HTML as the primary human interface.
Not a collection of prompts. A complete, battle-tested system for long-running agent teams.
11+ production-grade skills following the official agentskills.io spec. Every one reviewed by reviewer + tester + architect agents.
HARNESS-SPEC.md defines the mandatory artifacts, rituals, stacked PR discipline, and multi-tier verification that turns stateless sessions into reliable agents.
Every large effort decomposed into tiny, reviewable PRs (<10 min deep review). Proven on a 16-PR security-critical authz chain in production.
For any human-facing artifact — specs, reviews, plans, dashboards — we generate beautiful self-contained Tailwind HTML. The unreasonable effectiveness of HTML.
Mechanical Generator → Adversary loop using reusable agent definitions. Reviewer, Tester, and Architect personas run on every significant change.
.agents/skills/ following the official agentskills.io specification.
The harness-bootstrap script emits AGENTS.md, features.json, progress.md, justfile, the complete .agents/ structure, and everything else required by the spec.