rheinard.korf.au

← All projects

Harkness

An AI coding harness that runs offline against a local model, with tmux as its interface.

Harkness is the runtime around a coding model: the loop, the tools, the context and the safety controls. It runs against local models through Ollama or llama-server, and uses tmux for its interface.

Harkness fixing a failing test with qwen3-coder running locally in Ollama, asking before each change. Waits for the model and for approvals are sped up, so the timers on screen run ahead of the video.

Design

  • A headless core. The agent loop emits events and asks for approvals through an interface. The tmux chat and the unattended harkness run are both just clients of it.
  • One event log per session. The transcript, audit trail, stats and undo are all read from the same record.
  • Model profiles. Each model’s quirks, from prompt layers to edit strictness, live in one place.
  • Prompts are files, not code. Everything the model reads is in prompts/, layered over built-in defaults.
  • A sandbox and a policy. Shell commands run in a sandbox, and a policy decides which calls run, which ask first and which are refused.
  • Rules come from failures. A new rule is only added for a failure that was actually observed, and its effect is measured with an eval set against a real model.

It follows the recommendations in Harness Engineering (Barbaste et al., 2026), a source-code study of eleven production coding agents.

Features

Chat with approvals, tmux panes, undo and redo, resuming sessions, switching models, skills, and reading images.

Built with

Bun, TypeScript and tmux.

Status

A rewrite, started in November 2026. Private for now.

← All projects