Break coding goals into verified multi-agent task plans

Ordewell turns one goal into an ordered plan of coding-agent tasks, each with its own runner, model and mode, then executes and verifies it. You give a goal, and the plan is a file you can rewrite before anything runs. Every task carries its runner, model, thinking effort and mode. You can change any of them, add and remove tasks and rewire dependencies, completed work survives, and nothing runs until you say go. It is Apache-2.0 licensed, with the name and logos excluded from the license.

Its core design choices:

• One runner and model per task, all visible at once. A security refactor and a README update do not deserve the same model, so the planner assigns across the whole plan and shows every assignment first.
• A task is done only when its completion marker appears in the runner's output. Exit code is kept as diagnostic evidence, no model grades its own work, and a task that ends without its marker fails visibly.
• The planner is read-only and conversational. It researches the repo, asks when the goal is vague, and its final message is the plan. Reads run in parallel, anything reaching outside the workspace asks once, and commands that would write are refused outright.
• No extra API key is required. Claude Code, Codex or OpenCode can act as the planner, strictly read-only, on the subscription you already hold.
• Multi-runner by design. Claude Code, Codex and OpenCode ship built in, and any other CLI such as Aider is added through a plugin manifest and not a code change.

Use it for goals that span several steps and would benefit from different models and modes, where you want to review and edit the plan before spending anything, and where you want verified completion. Manual steps become checklists inside the plan.

It has requirements to note. It needs Node.js 20 or newer, and the terminal UI needs tmux on every platform because tmux backs each task's live terminal. On Windows the TUI runs under WSL, while the CLI, API server and VS Code extension work natively, and Git for Windows should stay installed because the planner runs research commands in its POSIX shell. There is no browser dashboard yet: the web command serves a local JSON and WebSocket API on port 3742 that the CLI and TUI use as clients.

The flow has five steps. Describe a goal in plain prose. The planner researches your workspace read-only and interleaves questions with research in one persistent conversation. A plan appears, which you can edit inline or reprompt to reshape without losing completed work. Execution spawns a real coding-agent session per AI task, respecting the dependency graph and passing each task its predecessors' results. Finally, the verdict engine completes a task only once its marker appears, and sessions auto-save to a local folder.

The CLI mirrors every slash command. Plan and run are one line: plan with a goal, then run. Related commands reassign a task's runner, model and dependencies, show status for a session, and toggle skills: grilling makes the planner interrogate your goal with at least three probing questions, prd writes a full PRD, and tdd adds red-green-refactor instructions to tasks. Independent tasks run in parallel, with a default of three concurrent sessions, adjustable from 1 to 5.

For planning without a coding-agent subscription, twenty-five providers are recognised through their own API key variables, including OpenRouter, Anthropic, OpenAI, Gemini, xAI, Groq, DeepSeek and Mistral, and any OpenAI-compatible endpoint can be set through a base URL variable. The default planner model is a budget model, since the planner never writes code.

Three surfaces share one core: a VS Code extension with a streaming timeline and task cards, a terminal UI with a chat pane and plan pane, and the CLI. Plugins install from https repositories on GitHub, GitLab, Bitbucket and Codeberg, or from a local directory. The architecture splits into a pure TypeScript core, the CLI, the VS Code extension and a Hono-based local server, with a pure-reducer TUI design documented in the project's decision records.

Configuration centers on four settings. One provider key is the only required setting, and the provider is auto-detected from whichever key is set, or forced with AI_PROVIDER. The planner model defaults to a budget model, deepseek/deepseek-v4-flash, set through ORCHESTRATOR_MODEL or the model command, and with a coding-agent planner it must be one of that agent's own model IDs. ORDEWELL_PLANNER_EFFORT sets thinking effort for a coding-agent planner and is ignored by vendor planners. ORDEWELL_MAX_PARALLEL defaults to 3. The TUI is driven by single keys, such as f to start, E to run all, m to toggle done, R for runner and o for model, and the mouse wheel scrolls whichever pane the pointer is over, with a mouse-off command to restore drag-to-select.

Install globally with npm and start the TUI, where first run asks for a planner and runner from inside, with no restart or up-front key needed. The package also runs through npx and ships scoped as a CLI package, and the VS Code extension carries its own core.

npm install -g ordewell

Developers who use several coding agents and want a reviewable, editable, verified plan instead of one long unsupervised session. The planning skills are adapted from Matt Pocock's MIT-licensed skills.