DEV Community

Cover image for I got tired of coding agents editing before “done” was defined — so I built a plan-first harness
vikash ruhil
vikash ruhil

Posted on

I got tired of coding agents editing before “done” was defined — so I built a plan-first harness

Chat-coding is fast until it isn’t.

You ask for a feature. The agent jumps into files. You get a pile of edits and a shrug. Two sessions collide on the same paths. Auto-merge is either terrifying or so babysat that you never leave the keyboard.

I wanted the other shape:

messy goal in → reviewed pull request out

You only show up when it actually needs you.

What kept breaking

  1. No shared definition of done — agents optimize for “changed files,” not “accepted outcome.”
  2. Parallel work without isolation — overlapping edits, merge pain.
  3. Untrusted automation — silent merge is worse than no merge.

What Loomwright does instead

Loomwright is a Claude Code plugin (not a Cursor-native plugin). Marketplace id atelier, plugin id loomwright.

  • Launch Pad writes a brief first (feasibility, file impact, subtasks) before Supervisor touches the repo
  • Workers run in git worktrees; overlapping files wait; merges stay ordered
  • Default: park a reviewed PR — it does not silently ship
  • Opt-in --auto-merge only behind a trusted gate

Then in a git repo:

/setup
/autonomous "the next thing your users can feel"
Enter fullscreen mode Exit fullscreen mode

Done when: /setup prints a module-status table (everything optional). /autonomous asks a few discovery questions, saves a brief under .supervisor/jobs/pending/, does the work in worktrees, and ends with an AUTONOMOUS_RUN summary and an open PR — it never merges on its own.

Site: https://vikashruhilgit.github.io/loomwright-site/

GitHub: https://github.com/vikashruhilgit/loomwright

What I’m still unsure about

Curious how others are handling:

  • Should the review gate diff against the brief’s acceptance criteria, or stay presence/test based?
  • Should trust be path-aware (auth/payments/migrations forever manual)?

If you’ve hit the same wall, I’d love the failure mode that burns you most: no plan, merge conflicts, or untrusted auto-merge.
Today that ships as roughly: 14 agent roles, 21 slash commands, 41 skills, 24 quality hooks.

Install (tight path)

claude plugin marketplace add vikashruhilgit/loomwright
claude plugin install loomwright@atelierclaude plugin list   # loomwright@atelier, Status: ✔ enabled
Enter fullscreen mode Exit fullscreen mode

Top comments (0)