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

TL;DR Highlight

Turns one goal into an editable plan and runs it across multiple coding agents in parallel

Who Should Read

Developers using Claude Code, Codex or OpenCode who want to split big tasks across agents running in parallel

Core Mechanics

Turns a goal into a plan you can edit
- It explores the repo, asks clarifying questions, then builds tasks with dependencies.
- Before running, you can edit each task's prompt, runner, model, effort and mode.
Each task can use a different agent and model
- Claude Code, Codex and OpenCode can be mixed in one plan.
- A security refactor and a README update get different models, and you see assignments before any tokens are spent.
Branch isolation with ordered merging
- Tasks that change code work on their own branches, and independent ones run in parallel.
- Passing work lands on one integration branch in plan order, and you choose when to merge.
Done means the agent reported it
- A task finishes only when its agent reports done through Ordewell's own tools.
- It is never marked done because a model thinks the work looks finished.
The planner only reads, and a supervisor watches the run

Community Reactions

Wish for a deterministic planner for cheap models
- One commenter wanted a deterministic program that makes confident plans for lower-cost models like DeepSeek.
- They argued the LLM part stays non-deterministic but quality is easier to measure this way.
AGENTS.md alone wasn't enough, per experience
- A commenter said AGENTS.md could be seen as that, but in their experience it doesn't give non-frontier models a high success rate.
Criticism that the project looks entirely AI-written
- One commenter looked for a differentiator but felt everything, including the author's replies, was AI-written.
Suggestion of a similar tool
- Another commenter asked whether the author had tried highflame-ai/codeoid.

How to Apply

Generate just a plan in a small repo
- Run npm install -g ordewell, set AI_PROVIDER=claude-code, and run ordewell plan --goal for one small feature.
- Only check whether the task split and dependencies look sensible before running.
Reassign models per task by cost
- Use ordewell task-model and task-runner to give docs and simple edits a cheaper model and security or design tasks a stronger one.
- Run only after the assignments look right.
Split reviews and audits into read-only tasks
- Add code review or security check tasks to the plan so they run without write access.
- Read the diff with ordewell handoff review, then merge.
Write team rules as a repo skill
- Write your team's test and commit rules as a skill in the repository.
- Check in the plan that the planner attaches it to the right tasks.

Try It

copy & run

# Run in a terminal (needs Node.js 20+, git, Claude Code installed)

npm install -g ordewell

export AI_PROVIDER=claude-code
ordewell plan --goal "Add rate limiting to the public API"

# Edit the plan before running
ordewell task-runner 2 opencode
ordewell task-model 3 sonnet
ordewell task-deps 3 1,2

# Run, then review and merge
ordewell run
ordewell handoff review
ordewell handoff merge

Terminology

OrchestrationLike a site manager, it hands out work to multiple agents and manages order and results.

RunnerThe coding agent that actually performs a task, such as Claude Code, Codex or OpenCode.

Integration branchA staging branch where passing task results pile up in plan order until you merge them into yours.

SupervisorA role that watches a run, reports failures or stalls, and if allowed retries or stops tasks.

WorktreeA git feature that checks out another branch of the same repo into a separate folder at the same time.

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