GitHub - ordewell/ordewell: Multi-agent task orchestration for coding agents. Turn one goal into an ordered plan of tasks — each with its own runner, model and mode — then execute and verify the results.

Turn one goal into an ordered plan of coding-agent tasks — each with its own runner, model and mode — then execute and verify it. Website · Docs What this is A plan you can rewrite before a token is spent. The plan is a typed artifact, not an agent's internal state: every task carries a runner, model, thinking effort and mode, and you can change any of them, add and remove tasks, and rewire dependencies — without losing completed work or round-tripping the AI. The right model per task, chosen in the open. The planner makes one portfolio decision across the whole plan — a security refactor and a README update do not deserve the same model — and shows you every assignment before anything runs (why a separate planner?). Verdicts from evidence, not opinion. A task completes only when its unique completion marker appears in the runner's output; exit code is retained as diagnostic evidence. The model is never the tie-breaker. Stuck tasks can be advanced with Mark complete, and a task marked done by mistake goes back with Mark not done. A planner that talks back. Planning is one continuous chat: it researches your repo read-only, asks when your goal is vague, and its final message is the plan (ADR-0002). Reads run in parallel; anything reaching outside the workspace asks once; commands that would write are refused outright (ADR-0008). No extra API key required. Claude Code, Codex, or OpenCode can be the planner, strictly read-only, on the subscription you already hold for the runners (ADR-0009). Multi-runner by design. Enable several and the planner assigns one per task. Claude Code, Codex and OpenCode ship built-in; anything else — Aider, your own CLI — is a plugin manifest, not a code change. Quick Start Node.js ≥ 20 on macOS, Linux or Windows. The TUI also needs tmux — see Platform support below. npm install -g ordewell ordewell # the TUI — chat on the left, plan on the right That's it. First run asks for a planner and a runner, set from inside (/planner, /runners, /key) — no restart, no API key required up front. npx ordewell works the same without a global install; the package also ships scoped as @ordewell/cli. For VS Code instead, install the extension — it carries its own core, so there is nothing to install from npm: code --install-extension ordewell.ordewell Or search Ordewell in the Extensions view. Building from source: git clone https://github.com/ordewell/ordewell.git && cd ordewell && npm install && npm run build && npm link -w packages/cli — see CONTRIBUTING.md. Scriptable / headless Every slash command is also a subcommand — set the planner and runner by env var to skip the TUI entirely. Already run Claude Code, Codex, or OpenCode? No separate API key — it runs on the subscription you already hold: export AI_PROVIDER="claude-code" # or codex, opencode ordewell plan --goal "Add rate limiting to the public API" && ordewell run Mutation always stays with the runners; the planner agent only explores and reasons. Same toggles apply from a UI: /planner, /model, /planner-effort, or the planner bar in 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 key for variable names, or point OPENAI_COMPATIBLE_BASE_URL at anything else, including a local model server. export OPENROUTER_API_KEY="sk-or-..." ordewell plan --goal "Add rate limiting to the public API" && ordewell run Three surfaces, one core VS Code A streaming timeline: live thinking, each research step with its outcome, and task cards you expand for the runner's own output. Retarget a task's runner and its model and mode re-derive in place. The whole loop is below, under The VS Code loop, end to end. Terminal UI Everything the extension does, over SSH. tab swaps chat and plan pane; single keys drive the plan (f start, E run all, m toggle done, R runner, o model). /help lists the rest. CLI $ ordewell plan --goal "Add rate limiting to the public API" Generating plan for: "Add rate limiting to the public API"... ✓ list_dir src → D middleware F router.ts F auth.ts ✓ grep X-RateLimit → no matches in 6 files Question: should limits apply per API key, or per client IP? My recommendation: per key — auth() already threads the key through req.ctx. > per key, with 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 in src/middleware/rateLimit.ts (Claude Sonnet 4.5 · Claude Code) 2. [ AI] Wire the limiter into route registration (Claude Haiku 4.5 · Claude Code) 3. [ AI] Return RFC 6585 429s with Retry-After (DeepSeek V4 Flash · Opencode) 4. [MAN] Document the limit headers in the OpenAPI spec [MAN] = manual step — run `ordewell tui` to work through it Run 'ordewell run' to execute, 'ordewell status' to inspect, or 'ordewell tui' for the full UI. $ ordewell run Executing plan... ✓ #a1b2 completed — PASS: Verified: completion marker detected in agent output. Task c ⟳ #c3d4 in_progress [2/Wire the limiter into route registration] Started: claude-code / claude-haiku-4-5 Done. 4 completed, 0 failed, 0 blocked. Every slash command is also an ordewell subcommand, so nothing is UI-only and headless automation reaches everything a human can. How it works Describe a goal in plain prose. The planner researches your workspace read-only and interleaves questions with research in one persistent conversation (ADR-0008). A plan appears — ordered tasks, each with a runner, model, thinking effort and mode. Edit anything inline, or reprompt to reshape the whole plan without losing completed work. Execution spawns a real coding-agent session per AI task, respecting the dependency graph and handing each task its predecessors' results. Manual tasks become checklists. The VerdictEngine completes a task only once its marker appears; an exit without one 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 goal is underspecified ordewell plan --goal "Migrate the config loader from JSON to TOML" # Reassign before running — runner first, since it re-derives model, effort and mode ordewell task-runner 2 opencode ordewell task-deps 3 1,2 # Execute; independent tasks run in parallel (default: 3 concurrent sessions) ordewell run # Inspect any session later ordewell status --session-id session-1751600000000 The surfaces differ only in how you name a target: the TUI opens a picker, the CLI takes an argument — and omitting the argument prints the same options the picker would have shown. ordewell task-model 3 # lists the models that task's runner can spawn ordewell task-model 3 sonnet # picks one Configure without an editor ordewell planner claude-code # plan on a coding agent's subscription — no API key ordewell model set sonnet # scoped to that agent's own catalog ordewell planner-effort high # a variant of the selected model ordewell key set openrouter sk-… # stored in .env, never echoed back ordewell runners codex off Each pushes to the running server before writing .env, so the change lands on the next plan with no restart — and a refused connection cannot leave the file holding a setting the daemon never saw. Deep-interview planning with a PRD ordewell grilling on # planner interrogates your goal before outlining (min. 3 probing questions) ordewell prd on # planner previews, then writes a full PRD to .scratch/<slug>/PRD.md ordewell tdd on # tasks are augmented with red-green-refactor instructions ordewell plan --goal "Real-time collaborative editing" # → the planner grills you in chat, drafts the PRD, waits for your OK, # then commits the plan as its final message Multi-runner plans and custom runners # Pass --runner repeatedly to build a runner set; the planner assigns one per task ordewell plan --goal "Refactor auth module" --runner claude-code --runner opencode # Bring your own CLI agent via a plugin manifest ordewell plugins create my-runner # scaffolds manifest.json ordewell plugins install github:user/repo ordewell plugins list Remote plugin installs accept https:// repositories on GitHub, GitLab, Bitbucket and Codeberg; anything else must be cloned yourself and installed from its local directory. The other two front ends ordewell # full-screen terminal UI — same as `ordewell tui` ordewell web --daemon # the local API server, in the background ordewell web starts the HTTP + WebSocket API on 127.0.0.1:3742 that the CLI and TUI are clients of — every other command starts it for you on demand. It serves JSON, not a web page; there is no browser dashboard yet. For VS Code, install the extension and open the Ordewell panel — see 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 session is adopted by the server, so its plan stays executable System /help, /mouse, /quit API keys typed into /key are masked on screen and written to your .env. The mouse wheel scrolls whichever pane the pointer is over — transcript or plan — regardless of which one has keyboard focus, and pgup/pgdn scroll the focused one. Capturing the mouse for the wheel is what disables the terminal's own drag-to-select, so /mouse off hands it back when you need to copy text out (remembered via ORDEWELL_TUI_MOUSE in your .env, and ORDEWELL_TUI_MOUSE=false in the environment turns it off everywhere).