Ordewell orchestrates Claude Code, Codex and OpenCode tasks

What happened

Ordewell, an open-source task orchestrator for coding agents, launched as a Show HN post on 15 September 2026. The tool separates planning from execution: a planner reads a repository, turns one goal into an ordered list of tasks, and hands each task to a coding-agent runner such as Claude Code, Codex or OpenCode. The post, by developer ac-ciano, reached 50 points and 30 comments on Hacker News.

Ordewell is a command-line tool, terminal UI and VS Code extension that converts a plain-language goal into a typed, editable plan of coding-agent tasks. Each task carries its own runner, model and thinking effort, and a task passes only when a unique completion marker appears in that runner's own output, not when the model claims it is finished.

What changed

• The planner researches a workspace read-only and asks clarifying questions before it commits to a plan; its final chat message is the plan itself, made of ordered tasks rather than free text.
• Each task holds a runner, a model, a thinking effort and a mode. Changing the runner re-derives the other three from that runner's own catalog, so a combination the runner cannot spawn is never saved.
• A task completes when its unique completion marker shows up in the runner's output. Exit code is kept as diagnostic evidence, but the model is never consulted on its own verdict.
• Enabling more than one runner (Claude Code, Codex, OpenCode, or a plugin such as Aider) lets the planner assign a different one per task; independent tasks run in parallel, three at a time by default and up to five.
• Claude Code, Codex or OpenCode can act as the planner itself, read-only by construction, on the subscription the developer already holds, so no separate API key is required. A key is optional, and Ordewell recognises twenty-five providers by their own *_API_KEY variable, among them OpenRouter, Anthropic, OpenAI, Gemini, xAI, Groq and DeepSeek.
• Optional skill toggles change how the planner behaves: grill-me asks at least three questions before outlining a vague goal, prd drafts a PRD to .scratch/<slug>/PRD.md and waits for sign-off, tdd adds red-green-refactor steps, and review and verify append a final judged pass.
• Install is npm install -g ordewell (or npx ordewell without installing) for the CLI and TUI, and code --install-extension ordewell.ordewell for VS Code. The TUI needs tmux on every platform, WSL on Windows. The project is licensed Apache-2.0.

Why it matters for agent builders

A plan that lives as structured data, not as an agent's internal state, is what lets a developer swap a task's model after the planner runs without losing already-completed work elsewhere in the same plan. That separation of planning and execution also opens up per-task cost control: a security refactor and a README edit no longer have to run on the same model just because one planner picked it for the whole job. Builders who already juggle Claude Code for one class of work and Codex or OpenCode for another get a single place to see every model and runner assignment before any of them spend a token.

What to watch

The Hacker News thread split over whether tools like Ordewell earn their place once frontier coding agents absorb the same planning behaviour natively, a concern several commenters raised directly against the project. A separate thread of comments criticised the author for replying to commenters with AI-generated responses without saying so, after a reader pointed out that the project's docs, code and some author replies were all AI-written; the author confirmed the docs and code but did not address the replies specifically. Whether Ordewell's model-routing and verdict-marker approach holds up against a general-purpose harness update is the open question the thread never settled.

Key takeaways

• Ordewell turns one goal into an editable, typed plan of coding-agent tasks, each with its own runner, model and thinking effort, rather than leaving the plan inside an agent's own context.
• A task is marked done only when its unique completion marker appears in the runner's output; the model itself never decides its own verdict.
• Claude Code, Codex or OpenCode can serve as both the read-only planner and a task runner, so no separate API key is required to start.
• The project also recognises twenty-five model-provider API keys and ships a JSON plugin format for adding a custom coding-agent runner.
• The Show HN launch drew both a substantive debate about meta-frameworks and criticism of undisclosed AI-written replies in the comments.

FAQ

What coding agents does Ordewell support?

Claude Code, Codex and OpenCode ship built in as runners, and the planner can also be one of these three, read-only, on the subscription a developer already holds. Any other CLI agent, such as Aider, is added through a JSON plugin manifest rather than a code change.

Does Ordewell require an API key?

No. The default path uses a coding agent the developer already pays for as both planner and runner, with no key exported anywhere. A key becomes optional if the developer prefers an API-based planner instead; Ordewell recognises twenty-five providers by their own environment variable, including OpenRouter, Anthropic, OpenAI and DeepSeek.

How does Ordewell know a task is actually finished?

Its VerdictEngine looks for a unique completion marker in the runner's output. If the marker is present, the task passes; the exit code is stored as supporting evidence, but the model's own claim of completion is never the deciding signal.

Source: Ordewell on GitHub

Discussion: Hacker News

Developers picking a coding agent to pair with a planner like Ordewell can compare runners in AI Agents Listing's coding category.