Ordewell Multi‑Agent Task Orchestration Tool – Overview and Key Features

What Ordewell Does — Turn One Goal into an Editable, Multi‑Agent Plan

Ordewell converts a plain‑language objective into a typed plan where every task specifies a runner, model, thinking effort, and mode. The plan can be edited before any token is spent, allowing developers to choose the right model per task, rewire dependencies, and add or remove steps without losing completed work.

• The planner operates in a read‑only exploration envelope: it can query the repository, ask clarification questions, and then emits the final plan as its last message.
• Execution spawns a separate coding‑agent session for each AI task, respecting the dependency graph and verifying completion via a unique marker in the runner’s output.
• Manual tasks appear as checklist items that the user can mark complete.
• No extra API key is required for the planner; it can reuse the subscription you already hold for Claude Code, Codex, or OpenCode.
• Multiple runners are supported out of the box, and additional runners can be added via a plugin manifest.

---

Quick Start – Install and Run the Terminal UI

npm install -g ordewell
ordewell               # launches the full‑screen TUI

• The first launch prompts you to select a planner and one or more runners (e.g., Claude Code, Codex, OpenCode). No API key is needed up front.
• The same workflow works with npx ordewell (no global install) or the VS Code extension (code --install-extension ordewell.ordewell).
• For headless automation, every slash command is also a subcommand; set AI_PROVIDER or the appropriate *_API_KEY environment variable to run ordewell plan … && ordewell run without the UI.

---

Core Concepts and Workflow

1. Goal Definition – Provide a natural‑language description, e.g., "Add rate limiting to the public API".
2. Planner Research – The planner reads the repo read‑only, asks clarification questions, and gathers evidence.
3. Plan Generation – An ordered list of tasks is produced, each annotated with runner, model, effort, and mode. The plan is a typed artifact that can be edited.
4. Execution – Independent tasks run in parallel (default max 3 concurrent sessions). Each AI task runs in its own session; manual tasks become checklists.
5. Verdict Engine – A task is marked PASS only when its unique completion marker appears in the runner’s output. Exit codes are retained as diagnostic evidence; tasks can be manually marked complete or not done.

---

Three User Interfaces, One Core Engine

Interface	Primary Use Case
VS Code Extension	Integrated streaming timeline, live research steps, and expandable task cards. Allows in‑place retargeting of runners and models.
Terminal UI (TUI)	Full‑screen, keyboard‑driven workflow over SSH; tab swaps chat and plan panes, single‑key shortcuts drive the plan (e.g., f start, E run all).
CLI	Scriptable, headless automation; every UI action is also a subcommand (ordewell plan, ordewell run, ordewell task-runner).

All interfaces share the same core packages (core/, cli/, vscode/, web/) and communicate via a local daemon (ordewell web --daemon) that exposes a JSON‑over‑WebSocket API.

---

Configuration Essentials

Setting	Default	Effect
Provider key (OPENROUTER_API_KEY, ANTHROPIC_API_KEY, …)	—	Required for the chosen AI provider; Ordewell auto‑detects the provider from the set key.
ORCHESTRATOR_MODEL	deepseek/deepseek-v4-flash	Planner model used for research and plan generation.
ORDEWELL_PLANNER_EFFORT	—	Effort level for coding‑agent planners (e.g., high, low).
ORDEWELL_MAX_PARALLEL	3	Maximum concurrent AI task sessions (1–5).

Environment variables can be set via ordewell key set …, ordewell model set …, or the VS Code settings panel. Changes take effect immediately without restarting the daemon.

---

Extensibility – Custom Runners and Plugins

• Built‑in runners include Claude Code, Codex, and OpenCode.
• Additional runners are added by creating a plugin manifest (ordewell plugins create my-runner) and installing it from a Git URL (ordewell plugins install github:user/repo).
• Plugins can be hosted on GitHub, GitLab, Bitbucket, or Codeberg; other sources require a local clone.

---

Community Feedback from Hacker News

@ramon156: “I would love a deterministic program that can confidently make plans for lower‑cost models like DeepSeek. The LLM part won’t be deterministic, but it’s easier to measure quality this way.” – Highlights demand for reproducible planning with budget models.

@Lalabadie: “I opened the page looking for a differentiator, but everything about the project is AI‑written (including author replies).” – Points out perception that the project’s documentation is heavily AI‑generated, which may affect trust.

@sharathr: “Have you tried https://github.com/highflame-ai/codeoid” – Suggests an alternative code‑generation tool for comparison.

@ac‑ciano: “[dead]” – Indicates the original author’s comment was removed, offering no additional insight.

---

Platform Support and Windows Notes

Surface	Linux	macOS	Windows
VS Code extension	✅	✅	✅
API server	✅	✅	✅
CLI	✅	✅	✅
TUI (requires tmux)	✅	✅	Needs WSL

• The TUI relies on tmux for per‑task terminals; install it via your package manager (apt install tmux, brew install tmux).
• On Windows, use native installers for agent CLIs to avoid .cmd shim limitations on large system prompts.
• Ensure Git for Windows is installed so the planner’s read‑only shell commands (ls, grep, etc.) behave consistently.

---

Architecture Snapshot

packages/
├─ core/      # Pure TypeScript: Session, PlanStore, Planner, Orchestrator, VerdictEngine
├─ cli/       # TUI, plan/run/status commands, thin terminal driver
├─ vscode/    # Extension UI, streaming planner timeline, task cards
└─ web/       # Hono HTTP + WebSocket daemon (127.0.0.1:3742)

The TUI core is a pure reducer that returns {state, effects} and a renderer that emits one string per terminal row, ensuring UI logic is testable and UI‑agnostic.

---

License

Ordewell is released under the Apache License 2.0. The name and logos are excluded from the license per the NOTICE file.