Ordewell for VS Code

Plan once, run it across Claude Code, Codex and OpenCode, and review one branch. One goal becomes an editable plan with a runner and model assigned to each AI task. By default, submitted build plans wait for your approval. Run settings let you automate execution and merging. Research tasks and consented script tasks can start during planning.

This README describes 0.11.3. See the published releases for versions available to install and their release notes.

Features

An editable plan. Every agent task carries a runner, model, thinking effort and mode. Change any of them, add or remove tasks, and rewire dependencies without losing completed work.
A planner that talks back. It researches your workspace without modifying it, asks clarifying questions, and then hands you the plan.
Choose models per task. Review and edit each AI task's model and thinking effort before execution. Planning uses the selected planner model too.
Worktree isolation for code tasks. With isolation enabled, code tasks run in separate worktrees and passing changes land on an integration branch for review. Ops and script tasks run at the workspace root. Without a git repository, or with isolation disabled, tasks share your workspace folder.
Skills your team shares. Write a skill once and commit it. The planner attaches it to the tasks it fits, and it shows as a chip on each task card.
More than coding tasks. A script task runs exact commands under your approval and run settings, a check-and-fix loop repeats until its check passes or its rounds run out, and a research task reports back to the planner while you keep talking. A read-only task lands nothing; in a single-repository worktree, it can save a prototype as a patch for a later task.
Limits and context in view. The footers show your Codex, Claude Code or OpenCode Go account's plan limits beside context and cost, and an API planner compacts its own context before it overflows.
Completion is recorded. An AI task's task_complete report with status done passes its completion check. Reports marked failed or blocked, or a runner exit without a report, do not pass. Script tasks use their exit status. Review the diff and test evidence before merging.
No extra API key. Claude Code, Codex or OpenCode can be the planner, using the subscription you already have.

Getting started

Install the extension.
Open the Ordewell panel from the activity bar.
Type a goal. With Claude Code, Codex or OpenCode installed, the first usable one plans for you with no API key. Choose another planner in the planner bar.
Review the plan, then approve it to run.

Autorun and automatic merging are off by default, and the supervisor starts at Watch. In /settings, you can enable autorun, automatic merging or an Act supervisor that can merge. The handoff card lets you review the combined diff, merge it or discard the run.

To plan with an API key instead, set ordewell.openAiApiKey (OpenRouter) or ordewell.apiKey (Gemini), or export OPENROUTER_API_KEY or GEMINI_API_KEY. A key you have set takes priority over the installed agent.

Each runner you want to use (Claude Code, Codex or OpenCode) must be installed and on your PATH. If a runner appears greyed out right after installing it, reload the window so the extension picks up the new PATH.

Skills are SKILL.md folders: type /name in the panel to load one into the conversation, and attach task skills as chips on a task's card. See Writing skills.

The Command Palette offers Ordewell: Rewind Conversation, Fork Conversation and Compact Conversation. Isolation can be configured under ordewell.worktreeIsolation and related settings.

On Windows

The extension and its local daemon run natively on Windows. The ordewell terminal UI is not verified there; run it under WSL. Ordewell finds npm-installed runners (.cmd and .exe shims).

Most of the Windows support is tested on Linux, through the seams that stand in for Windows. On a Windows host, only how npm-installed runners are launched has been measured, and no first run from goal to merge has been watched end to end. Two things are known to be open: whether Codex's read-only sandbox is enforced, and %VAR% expansion on the cmd.exe shim route. The planner's shell has no OS sandbox on Windows and relies on its command rules alone. See Windows support.

What the extension sends

Once a day, the extension asks ordewell.ai whether a newer version is out. The request carries the version you run and a random install ID, kept in ~/.ordewell/version-check.json and shared with the command line. The ID is replaced every 30 days. Nothing else is sent: no code, prompts, paths, file names or account details. Like any server, ordewell.ai sees the IP address the request comes from.

The first time Ordewell starts on a machine, the extension shows a notice saying this and sends nothing. If it cannot write that file, it shows no notice and never sends anything. The Marketplace handles updates, so the extension shows no update line.

Turn it off by turning off VS Code's telemetry (telemetry.telemetryLevel). ORDEWELL_NO_VERSION_CHECK=1 and DO_NOT_TRACK=1 turn it off too. Turned off, nothing is sent and no install ID is written.

Also available

The same project ships a command line and a terminal UI:

npm install -g ordewell

Ordewell