Ordewell

Why it's ranked here

The safety claims live in the mechanism, not in prompts. A task counts as done only when its agent reports through a tool bound to that one attempt, so a stale or confused session cannot mark work finished. The planner's shell runs inside an operating-system sandbox on Linux and macOS, underneath a command classifier that already refuses writes. Design choices, including the options rejected, are written up as numbered decision records, and the code is Apache 2.0. For software that points several agents at your repository at once, that is where the engineering should go.

What's good

Plans are data you can change before spending anything on execution: move a task to another agent, swap its model or rewire its dependencies from the command line without asking the planner again. Each code-changing task gets its own worktree, and finished work lands one task at a time with a merge commit per task, so any task's changes can be traced. A conflicting merge is aborted with the work kept rather than failed. An optional supervisor can watch a run and, if you allow it, retry or stop tasks within limits. The planner can run on an agent subscription you already pay for, or on keys from 25 providers.

Tradeoffs

The sandbox has gaps the project documents itself. Windows gets none, and stock Ubuntu 24.04 blocks the unprivileged namespaces the Linux sandbox needs, leaving only the command classifier. On macOS a sandboxed command can still reach existing launchd jobs. The per-task access tokens separate roles but, by the project's own account, do not defend against a hostile process running as the same user. Uncommitted edits to tracked files switch off branch isolation for a run until you stash or opt out. On Windows the terminal interface needs WSL. The author also notes that a job one agent finishes in one context gains nothing from orchestration.

How to use it well

It fits a developer who already uses Claude Code, Codex or OpenCode, ideally more than one, and wants a multi-step change split up, reviewed as a plan, and run in parallel without agents editing the same checkout. Commit or stash before a run so isolation stays on. Use read-only tasks for reviews and audits, and keep the supervisor at Watch until you trust its edits. Install through npm or the editor extension; you need Node.js 20 or newer and git. Skip it for small single-file fixes, where one agent session costs less.

Observed

License

Apache-2.0; the name and logos are excluded and covered by a separate notice

Language

TypeScript, npm workspaces monorepo (core, vscode, cli, web)

Install

npm install -g ordewell; VS Code Marketplace and Open VSX extension that bundles its own core

Interfaces

Terminal UI, VS Code extension, CLI for scripts, local API, sharing one core

Supported coding agents

Claude Code, Codex and OpenCode, mixable within one plan

Requirements

Node.js 20 or newer, git for task isolation, at least one supported coding agent

Platform support

Linux and macOS; on Windows the terminal UI runs under WSL

Planner shell sandbox

bubblewrap on Linux, sandbox-exec on macOS, none on Windows

Agent integration

Loopback-only MCP server with per-attempt task tokens

Design records

Architecture decision records in the repository, including rejected options