Repo Harness

This tool is an open source command line workflow that turns Claude and Codex coding sessions into a repeatable process kept inside the repository. Instead of relying on chat memory, it stores plans, contracts, handoffs, checks and reviews as files in the repo, so a session can end mid task and the next one resumes at the exact next step with its blockers and changed files. A developer hands the agent an approved plan or sprint, and the human loop becomes review and next. It is built for developers who work with coding agents on real repositories and want structure and auditability. A stable root context of about 12KB plus a CodeGraph index is meant to replace repeated grep and read loops. Eight managed hook routes warn, block, trace and hand off work, with edit gates that hold until a plan is approved and done claim gates that check file backed evidence. Each task produces a single Human Review Card showing the verdict, intended versus actual files, commands that passed, residual risk and rollback. Agents can work in isolated worktrees limited to the contract's allowed paths, and an optional MCP sidecar reads the repo read only. Setup is run with npx, followed by an adopt command that has a dry run preview. The site works with Claude, Codex and ChatGPT Pro, is published in several languages and lists version 0.14.2. No pricing is published, and the only price data on the site lists a cost of zero.