News

What's new in Ordewell

Ordewell turns one goal into a plan of coding-agent tasks, runs them with Claude Code, Codex and six more agents, and lets you review the work before it reaches your branch. This is where each release gets its short story.

Latest Version 0.13.0

Install without Node.js

On Linux and macOS the terminal UI and the CLI now install as one standalone binary, with no Node.js on the machine. The installer puts it at ~/.local/bin/ordewell, and running it again upgrades. Homebrew users can run brew install ordewell/tap/ordewell.

ordewell update, and /update in the terminal UI, updates Ordewell the way it was installed. See updating.

npm install -g ordewell and npx ordewell still work, and are the way to install on Windows for now. They need Node.js 20.3 or newer.

git is now optional. Without git, or outside a git repository, tasks run directly in the folder. git 2.38 or newer is still recommended: it gives each task its own worktree, delegated tasks and Merge all. A coding agent, installed and signed in, is still required.

0.13.1 makes the installer name any other copy of ordewell on your PATH, highlighted, with the step to switch. A loop's review now reports what it finds as a failed check, so the fix runs.

bash
curl -fsSL https://ordewell.ai/install.sh | sh

Version 0.12

More agents, one plan

0.12.6 adds GitHub Copilot CLI, Qwen Code, Cline, Gemini CLI and Cursor CLI. Each can plan, run tasks and take on read-only work alongside Claude Code, Codex and OpenCode. See runners for setup and version requirements.

One list of what waits on you

The inbox gathers everything that needs you into one list, built from the plan, the run and the approvals pending: each failed task, checkpoint and question, each merge conflict and approval, and each task or run held on you, with what you can do about it. /inbox in the terminal UI lists it, numbered, and answers in place or opens the task. In VS Code it is an N waiting control on the usage line that opens its list over the conversation, and /inbox in the chat turns it off and on. To be told when something new waits, set ORDEWELL_NOTIFY_COMMAND or VS Code's ordewell.notifyCommand to a command and it runs once for each new item.

Revert takes one task's landed work off the branch. /revert <task>, the Revert button in VS Code and ordewell revert <task> show what would go, the task and every landed task that built on it, ask, and then revert those merges together, in every repository or none. A task whose work cannot be taken off cleanly is refused, and the refusal says why. It needs git 2.38 or newer. An activity line above the composer says what the runner or planner is doing and for how long, and Claude Code tasks can load only what they need with taskRunnerConfig set to minimal.

Some behaviour changed. A failed task stays failed after a reload, so a restart keeps what you came back for in the inbox. The supervisor wakes in batches, so a failure can reach it later than before, though the inbox and the notify command still show it at once. Merge all on a run with no base branch always makes a merge commit. Each sandbox profile has its own scratch directory, and a runner stops within seconds when its daemon is killed. The changelog lists the rest.

bash
npm install -g ordewell@latest

Earlier releases

0.11

A plan that checks its own work

A task can be the check of a check-and-fix loop: Ordewell runs it, and while it fails with rounds left it runs the tasks that fix what it found, then the check again. Research tasks are read-only tasks whose report wakes the planner at every supervision level. The footers draw account limits from the accounts you already pay with, and a planner whose window cannot fit a call pauses with the turn held instead of failing it. Script tasks, richer handoffs and Guarded as the default came with it.

Release notes

0.10

The planner stays with your run

Until now the planner wrote the plan and left. In 0.10 it stays in the conversation while the run goes on, as a supervisor. At Watch, the default, it wakes on a failure, a task waiting for you, a landing conflict or a usage-limit pause, reads what it needs and writes you a note. At Act, which you choose, it can also run, stop and retry tasks, edit the plan and merge. It decides what to do next, never whether a task passed, and it can narrow a runner's permissions but never widen them.

The planner can run the plan too, not only write it. Turn on autorun in /settings and each plan it submits starts as if you had pressed Execute; turn on merge automatically and it merges at each gate and at the end of a run. A failed task blocks only its dependents, skip lands nothing, usage is recorded per attempt, and a task chat speaks the runner's own commands. Breaking: the ordewell.autonomousMode setting is gone from VS Code, and @ordewell/core exports less; the changelog lists them.

Release notes

0.9

Planners that can look things up, inside rules you set

Every planner can run read-only lookups like gh issue list or kubectl get pods and use your MCP tools. Ordewell runs each command itself against an allowlist you extend, asks on Guarded and refuses on Full, and the planner still never changes a file. 0.9 also supports OpenCode 2.

Release notes

0.8

Skills: teach your agents how your team works

A skill is a short set of written instructions, like how your team writes tests. Write it once, keep it in your repository, and the planner attaches it to the tasks that need it, whichever agent runs them. Every agent also runs through its own native protocol now, so tmux is no longer needed. 0.8.1 runs five tasks at once by default.

Release notes

0.7

Tasks talk to Ordewell directly

Ordewell now ships a small tool server that every agent connects to. Tasks use it to say when they're done and to ask you questions, and a message you send reaches a task while it's still working. If it can't wait, force send interrupts the current step. 0.7.2 followed with security hardening and a Stop that ends everything a task started.

Release notes

0.6

Every agent, driven directly

Claude Code, Codex and OpenCode are each driven through their own protocol, so Ordewell knows whether a task is working, waiting for approval or waiting for you, instead of guessing. Plans can also include ops work now, like a redeploy or a release tag, which waits until the code it depends on is merged.

Release notes

0.5

Every task on its own branch

Each task works in its own git worktree, so agents running side by side never step on each other's files. Finished work is merged in plan order, a conflict is handed back to the task to repair twice at most by default, and only a repair git confirms counts, so nothing reaches your branch until you've read the diff and said merge.

Release notes

Something broken, or an idea for the next one? Open an issue on GitHub.