Show HN: Ordewell – turn one goal into an ordered plan of coding-agent tasks

Sep 15, 2026 08:31 PM - 2 hours ago 1

Turn 1 extremity into an ordered scheme of coding-agent tasks — each pinch its ain runner, exemplary and mode — past execute and verify it.

Website · Docs

Apache-2.0 npm GitHub stars

a extremity is typed, the planner sounds the repo and refuses a write, it asks whether the limiter should reuse the existing Redis client, past commits a seven-task scheme — each task showing its runner, model, reasoning effort and mode — and executes it to 7/7 complete.

A scheme you tin rewrite earlier a token is spent. The scheme is simply a typed artifact, not an agent's soul state: each task carries a runner, model, reasoning effort and mode, and you tin alteration immoderate of them, adhd and region tasks, and rewire limitations — without losing completed activity aliases round-tripping the AI.
The correct exemplary per task, chosen successful the open. The planner makes 1 portfolio determination crossed the full scheme — a information refactor and a README update do not merit the aforesaid exemplary — and shows you each duty earlier thing runs (why a abstracted planner?).
Verdicts from evidence, not opinion. A task completes only erstwhile its unsocial completion marker appears successful the runner's output; exit codification is retained arsenic diagnostic evidence. The exemplary is ne'er the tie-breaker. Stuck tasks tin beryllium precocious pinch Mark complete, and a task marked done by correction goes backmost pinch Mark not done.
A planner that talks back. Planning is 1 continuous chat: it researches your repo read-only, asks erstwhile your extremity is vague, and its last connection is the scheme (ADR-0002). Reads tally successful parallel; thing reaching extracurricular the workspace asks once; commands that would constitute are refused outright (ADR-0008).
No other API cardinal required. Claude Code, Codex, aliases OpenCode tin be the planner, strictly read-only, connected the subscription you already clasp for the runners (ADR-0009).
Multi-runner by design. Enable respective and the planner assigns 1 per task. Claude Code, Codex and OpenCode vessel built-in; thing other — Aider, your ain CLI — is simply a plugin manifest, not a codification change.

Node.js ≥ 20 connected macOS, Linux aliases Windows. The TUI besides needs tmux — see Platform support below.

npm instal -g ordewell ordewell # the TUI — chat connected the left, scheme connected the right

That's it. First tally asks for a planner and a runner, group from inside (/planner, /runners, /key) — nary restart, nary API cardinal required up front.

npx ordewell useful the aforesaid without a world install; the package also ships scoped arsenic @ordewell/cli.

For VS Code instead, instal the hold — it carries its ain core, so there is thing to instal from npm:

code --install-extension ordewell.ordewell

Or hunt Ordewell successful the Extensions view.

Building from source: git clone https://github.com/ordewell/ordewell.git && cd ordewell && npm instal && npm tally build && npm nexus -w packages/cli — spot CONTRIBUTING.md.

Every slash bid is besides a subcommand — group the planner and runner by env var to skip the TUI entirely.

Already tally Claude Code, Codex, aliases OpenCode? No abstracted API cardinal — it runs connected the subscription you already hold:

export AI_PROVIDER="claude-code" # aliases codex, opencode ordewell scheme --goal "Add complaint limiting to the nationalist API" && ordewell run

Mutation ever stays pinch the runners; the planner supplier only explores and reasons. Same toggles use from a UI: /planner, /model, /planner-effort, aliases the planner barroom successful VS Code.

Prefer an API key? Twenty-five providers are recognised via their own *_API_KEY — OpenRouter, Anthropic, OpenAI, Gemini, xAI, Groq, DeepSeek, Mistral, Together, Fireworks, Perplexity, Cerebras, DeepInfra, Cohere, Novita, Kimi, Zhipu, Qwen, Doubao, Hunyuan, Baichuan, MiniMax, Yi, StepFun and SiliconFlow. Run ordewell cardinal for adaptable names, aliases point OPENAI_COMPATIBLE_BASE_URL astatine thing else, including a section model server.

export OPENROUTER_API_KEY="sk-or-..." ordewell scheme --goal "Add complaint limiting to the nationalist API" && ordewell run

A streaming timeline: unrecorded thinking, each investigation measurement pinch its outcome, and task cards you grow for the runner's ain output. Retarget a task's runner and its exemplary and mode re-derive successful place. The full loop is below, nether The VS Code loop, extremity to end.

Ordewell's terminal UI divided betwixt the planner speech connected the near and a scheme pane connected the correct showing 7 tasks pinch per-task runner, model, effort and mode.

Everything the hold does, complete SSH. tab swaps chat and scheme pane; azygous keys thrust the scheme (f start, E tally all, m toggle done, R runner, o model). /help lists the rest.

$ ordewell scheme --goal "Add complaint limiting to the nationalist API" Generating scheme for: "Add complaint limiting to the nationalist API"... ✓ list_dir src → D middleware F router.ts F auth.ts ✓ grep X-RateLimit → nary matches successful 6 files Question: should limits use per API key, aliases per customer IP? My recommendation: per cardinal — auth() already threads the cardinal done req.ctx. > per key, pinch an IP fallback for anonymous routes Plan: 4 tasks (3 AI, 1 Manual) — claude-code, opencode Session: session-1751600000000 1. [ AI] Add a token-bucket limiter successful src/middleware/rateLimit.ts (Claude Sonnet 4.5 · Claude Code) 2. [ AI] Wire the limiter into way registration (Claude Haiku 4.5 · Claude Code) 3. [ AI] Return RFC 6585 429s pinch Retry-After (DeepSeek V4 Flash · Opencode) 4. [MAN] Document the limit headers successful the OpenAPI spec [MAN] = manual measurement — tally `ordewell tui` to activity done it Run 'ordewell run' to execute, 'ordewell status' to inspect, aliases 'ordewell tui' for the afloat UI. $ ordewell run Executing plan... ✓ #a1b2 completed — PASS: Verified: completion marker detected successful supplier output. Task c ⟳ #c3d4 in_progress [2/Wire the limiter into way registration] Started: claude-code / claude-haiku-4-5 Done. 4 completed, 0 failed, 0 blocked.

Every slash bid is besides an ordewell subcommand, truthful thing is UI-only and headless automation reaches everything a quality can.

Describe a goal successful plain prose.
The planner researches your workspace read-only and interleaves questions pinch investigation successful 1 persistent speech (ADR-0008).
A scheme appears — ordered tasks, each pinch a runner, model, reasoning effort and mode. Edit thing inline, aliases reprompt to reshape the full scheme without losing completed work.
Execution spawns a existent coding-agent convention per AI task, respecting the dependency chart and handing each task its predecessors' results. Manual tasks go checklists.
The VerdictEngine completes a task only erstwhile its marker appears; an exit without 1 fails visibly. Sessions auto-save to .ordewell/sessions/.

Usage examples — planning, editing, multi-runner, plugins

Plan, edit, execute

# The planner researches the repo and converses if the extremity is underspecified ordewell scheme --goal "Migrate the config loader from JSON to TOML" # Reassign earlier moving — runner first, since it re-derives model, effort and mode ordewell task-runner 2 opencode ordewell task-deps 3 1,2 # Execute; independent tasks tally successful parallel (default: 3 concurrent sessions) ordewell run # Inspect immoderate convention later ordewell position --session-id session-1751600000000

The surfaces disagree only successful really you sanction a target: the TUI opens a picker, the CLI takes an statement — and omitting the statement prints the aforesaid options the picker would person shown.

ordewell task-model 3 # lists the models that task's runner tin spawn ordewell task-model 3 sonnet # picks one

Configure without an editor

ordewell planner claude-code # scheme connected a coding agent's subscription — nary API key ordewell exemplary set sonnet # scoped to that agent's ain catalog ordewell planner-effort precocious # a version of the selected model ordewell cardinal set openrouter sk-… # stored successful .env, ne'er echoed back ordewell runners codex off

Each pushes to the moving server before penning .env, truthful the alteration lands connected the adjacent scheme pinch nary restart — and a refused relationship cannot time off the record holding a mounting the daemon ne'er saw.

Deep-interview readying pinch a PRD

ordewell grilling connected # planner interrogates your extremity earlier outlining (min. 3 probing questions) ordewell prd connected # planner previews, past writes a afloat PRD to .scratch/<slug>/PRD.md ordewell tdd connected # tasks are augmented pinch red-green-refactor instructions ordewell scheme --goal "Real-time collaborative editing" # → the planner grills you successful chat, drafts the PRD, waits for your OK, # past commits the scheme arsenic its last message

Multi-runner plans and civilization runners

# Pass --runner many times to build a runner set; the planner assigns 1 per task ordewell scheme --goal "Refactor auth module" --runner claude-code --runner opencode # Bring your ain CLI supplier via a plugin manifest ordewell plugins create my-runner # scaffolds manifest.json ordewell plugins instal github:user/repo ordewell plugins list

Remote plugin installs judge https:// repositories connected GitHub, GitLab, Bitbucket and Codeberg; thing other must beryllium cloned yourself and installed from its section directory.

The different 2 beforehand ends

ordewell # full-screen terminal UI — aforesaid arsenic `ordewell tui` ordewell web --daemon # the section API server, successful the background

ordewell web starts the HTTP + WebSocket API connected 127.0.0.1:3742 that the CLI and TUI are clients of — each different bid starts it for you connected demand. It serves JSON, not a web page; location is nary browser dashboard yet.

For VS Code, instal the hold and unfastened the Ordewell sheet — spot Quick Start.

Area Commands

Planning	type a goal, /approve, /run, /stop
Tasks	/add-task, /remove-task, /complete, /uncomplete, /skip, /retry, /cancel, /force-start
Skills	/grilling, /tdd, /prd, /verify
Models	/model, /key, /allowlist, /runners, /auto, /refresh
Sessions	/sessions, /new, /save, /load, /delete — a loaded convention is adopted by the server, truthful its scheme stays executable
System	/help, /mouse, /quit

API keys typed into /key are masked connected surface and written to your .env.

The rodent instrumentality scrolls whichever pane the pointer is complete — transcript aliases plan — sloppy of which 1 has keyboard focus, and pgup/pgdn scroll the focused one. Capturing the rodent for the instrumentality is what disables the terminal's own drag-to-select, truthful /mouse disconnected hands it backmost erstwhile you request to transcript matter out (remembered via ORDEWELL_TUI_MOUSE successful your .env, and ORDEWELL_TUI_MOUSE=false in the situation turns it disconnected everywhere).

A task's ain terminal is simply a tmux window, wherever tmux does clasp the rodent truthful the wheel scrolls its scrollback. Selecting location still copies to your system clipboard: resistance to prime and merchandise to copy, aliases double/triple-click for a word or a line. Install wl-copy, xclip aliases xsel connected Linux if you person nary of them — without one, copying falls backmost to an OSC 52 flight that immoderate terminals ignore.

The VS Code loop, extremity to end — research, question, plan, execution, verdict investigation steps settee 1 by one, the planner asks whether limits use per API cardinal aliases per customer IP, a five-task scheme is committed pinch per-task exemplary and mode pills, past execution runs and a task lands connected a greenish walk verdict.

investigation steps settee 1 by one, the planner asks whether limits use per API cardinal aliases per customer IP, a five-task scheme is committed pinch per-task exemplary and mode pills, past execution runs and a task lands connected a greenish walk verdict.

Platform support — including the Windows notes Surface Linux macOS Windows

VS Code extension	✅	✅	✅
API server	✅	✅	✅
CLI	✅	✅	✅
TUI	✅ needs tmux	✅ needs tmux	needs tmux — tally it nether WSL

The TUI requires tmux connected each platform, not only Windows — it is what backs each task's unrecorded terminal. Install it from your package head (apt instal tmux, brew instal tmux) earlier moving ordewell. Everything other runs natively connected Windows: the planner (including harness planners), task execution, model discovery, and the read-only exploration letter cover each activity there.

Two notes for Windows. Install the supplier CLIs pinch their native installers wherever 1 exists — an npm-installed claude/codex/opencode is simply a .cmd shim, which has to commencement done cmd.exe and inherits its 8191-character command-line limit; that is good for task prompts but not for the harness planner's larger strategy prompt, and Ordewell will show you truthful by sanction alternatively than silently truncating it. And support Git for Windows installed: its POSIX ammunition is what the planner runs investigation commands in, truthful ls, cat, grep and friends behave the aforesaid arsenic they do everyplace else. See ADR-0010.

Any instal way is found, connected PATH aliases not: the PowerShell one-liner installers (irm https://claude.ai/install.ps1 | iex, OpenCode's equivalent), npm, pnpm, Yarn, bun, Scoop, Chocolatey, WinGet, and Volta. If a runner is greyed retired successful the picker correct aft you installed it, restart the VS Code model — a GUI-launched hold big holds the PATH it started with.

Configuration — the 4 settings that matter Option Default What it does

One supplier cardinal (OPENROUTER_API_KEY, ANTHROPIC_API_KEY, GEMINI_API_KEY, …)	—	The 1 required setting. Twenty-five providers are recognised, each from its ain adaptable — ordewell cardinal lists them — positive immoderate OpenAI-compatible endpoint via OPENAI_COMPATIBLE_BASE_URL. The supplier is auto-detected from whichever cardinal is group (force pinch AI_PROVIDER). Not needed erstwhile AI_PROVIDER is claude-code, codex, aliases opencode — those scheme pinch the CLI's ain subscription.
ORCHESTRATOR_MODEL	deepseek/deepseek-v4-flash	The planner exemplary — a fund exemplary by default; it plans and researches but ne'er writes code. Change via ordewell exemplary group <id> aliases /model, which scope the prime to the planner backend's ain catalog. With a coding-agent planner, it must beryllium 1 of that agent's ain exemplary ids.
ORDEWELL_PLANNER_EFFORT	—	Thinking effort for a coding-agent planner, from the selected model's ain variants (low, high, adaptive, …). Ignored by vendor planners, whose effort is baked into the exemplary id. Change via ordewell planner-effort <level> aliases /planner-effort.
ORDEWELL_MAX_PARALLEL	3	Max concurrent AI task sessions (1–5). Independent tasks tally successful parallel; the dependency chart is ever respected.

Run ordewell --help for the afloat database of situation variables, aliases ordewell setup for the interactive wizard. VS Code users: everything is mirrored nether ordewell.* settings.

Architecture

packages/ ├── core/ Pure TypeScript, zero UI deps — Session, PlanStore, Planner, │ TaskOrchestrator, VerdictEngine, ModelResolver, ModeResolver, │ RunnerRegistry + manifest template engine ├── cli/ ordewell: tui, plan, run, status, stop, web, models, setup, │ plugins, grilling, prd, tdd — positive tui/, a axenic authorities + │ renderer halfway down a bladed raw-mode terminal driver ├── vscode/ Extension + webview: streaming planner timeline, task cards, │ TTY seizure via script(1) └── web/ Hono HTTP + WebSocket server — the section daemon the CLI and TUI thrust complete 127.0.0.1 (session pool, headless execution)

The TUI's halfway is axenic — a reducer returning { state, effects } and a renderer returning 1 drawstring per terminal statement (ADR-0006).

Tasks default to each runner's autonomous mode (toggle pinch /auto), and the scheme is the root of truth for what runs — modes are ne'er silently rewritten astatine spawn (ADR-0001).

Every aboveground consumes 1 arena national (SessionMessage) complete 1 broadcast seam — the domain vocabulary lives successful CONTEXT.md and creation decisions successful docs/adr/.

Acknowledgements

The deep-interview readying workflows — grilling, PRD drafting, and TDD task augmentation — are adapted from Matt Pocock's skills (MIT), rebuilt arsenic punctual blocks wrong Ordewell's planner and runner prompts. If you want those workflows successful a plain coding-agent convention alternatively than an orchestrated plan, his repo is the spot to start.

Bug reports, characteristic requests and propulsion requests are invited — commencement pinch CONTRIBUTING.md for the build bid and the layout of the tree. Security issues spell to SECURITY.md, not the nationalist tracker.

New to the codebase? CONTEXT.md is the domain glossary and docs/adr/ records why things are the measurement they are.

Licensed nether the Apache License 2.0. The Ordewell sanction and logos are not covered by that licence — spot NOTICE.