Ordewell turns one goal into an ordered plan of coding-agent tasks, one model per task

The developer who posts as ac-ciano published Ordewell, a task orchestrator for coding agents that turns one plain-English goal into an ordered plan of tasks, each with its own runner and model, then runs one real agent session per task. The Show HN thread has 50 points and 30 comments after about 20 hours, the repository is Apache-2.0 with 101 stars, and there is a CLI, a terminal UI, and a VS Code extension against a shared core.

What it does

A goal is typed in plain prose. Ordewell's planner reads the workspace with read-only access, asks a question when the goal is vague, and writes an ordered plan as its final message. Each task carries a runner, a model, a thinking-effort level, and a mode. The plan is a typed artifact the user can edit before anything runs: change a runner, swap a model, add a task, rewire dependencies. Execution then spawns one coding-agent session per task in a fresh context, hands each task its predecessors' results, and runs independent tasks in parallel up to three sessions by default.

Verdicts do not depend on the model's own opinion. A task completes only when its unique completion marker appears in the runner's output, and the exit code is kept as diagnostic evidence. A stuck task can be advanced with a manual mark, and a task marked done by mistake can be sent back.

Which runners and planners

Claude Code, Codex, and OpenCode ship built in, and a plugin manifest can add another CLI without a code change. Any one of the three coding agents can also be the planner, restricted to read-only exploration, so no separate API key is needed as long as the user already holds a subscription. If the user prefers a key, Ordewell recognises 25 providers, including OpenRouter, Anthropic, OpenAI, Google, xAI, Groq, DeepSeek, Mistral, Together, Fireworks, Perplexity, and Cerebras, plus any OpenAI-compatible endpoint via OPENAI_COMPATIBLE_BASE_URL. The default planner model is DeepSeek V4 Flash, chosen so planning stays cheap while the runners spend the real budget.

Where it fits

The design pitch is that Ordewell writes an editable plan where a coding agent would keep an internal one. The planner makes one portfolio decision across the whole plan, so a README update and a security refactor can each run on the model they deserve. A run then produces a per-task record: which runner ran, which model, and whether the marker showed up. The tools are provider-neutral by construction; Ordewell picks between them per task and shows every assignment before Execute Plan.

The cost to adopt

Node.js 20 or later on macOS, Linux, or Windows, plus tmux for the terminal UI (Windows users run the TUI under WSL). Everything else runs natively on Windows. Install is npm install -g ordewell or the ordewell.ordewell extension in VS Code. Ordewell says the local API server it runs listens on 127.0.0.1:3742 and serves JSON, with no browser dashboard yet.

Coding agents that plan on their own tend to keep their plan hidden and to run every step on the same model. Ordewell argues that both are choices open to change, and that a plan you can rewrite before a token is spent is worth the extra layer. Whether that argument holds depends on how honest the marker-based verdict feels once real work runs through it: it may be a clean way to stop the model grading its own homework, or it may become another checkbox the agent learns to satisfy. Either way, the plan-first, portfolio-model shape is a specific pattern that did not have an off-the-shelf tool a week ago.

Source

Primary source: ordewell/ordewell on GitHub and the Ordewell website. Show HN thread: Ordewell on Hacker News.