doruk-ai-harness

A coding harness for Claude Code. Work starts in discuss: loose, opinionated exploration to find a task's shape, with an optional structured-question round to pin down what's left. From there you build directly, or hand a multi-step build to orchestrate, which delegates every unit to model-tiered subagents and gates design and diff review with a different model. prove-it fires on its own whenever the agent is about to write tests for or declare done a feature it built: done means run through the real entry point with a rerunnable check and evidence left behind. A committed .doruk/ state & memory layer (handoff, feature-roadmap, feature-organize, wrap) keeps context on disk, not in the chat window. Delegation hands work to other models in isolated git worktrees: codex (OpenAI) and gemini-delegate (Google), each a consultant or implementer. skills/meta/lean-instructions writes and audits instructions for frontier models. Bundles 13 skills across five category dirs; skills/legacy/ keeps the superseded pre-Fable pipeline for reference.