CodeJoust — AI skill for Claude Code

---

Same model, different harness. Independent testing found Claude Sonnet scored **77% through Claude Code** but **93% through Cursor** on the same benchmark — a 15-point gap that is pure tooling, not model quality. Which "AI coding assistant" is right for your task is not a model question, it's a **task-level empirical question**.

CodeJoust answers it. One CLI command fires the same task at Claude Code, aider, Codex, Gemini CLI, Kimi Code in parallel — each in its own `git worktree` — then auto-grades them and hands you the winning patch.

Why not just open three terminals?

That's what most people do. It's also why most people never actually benchmark their tools — running three agents, waiting, eyeballing three diffs, manually tallying tokens is work, so they pick one and stick with it.

CodeJoust takes that hour down to one command:

codejoust run "fix the off-by-one in Scheduler.next_fire" \
  --agents claude-code,aider,codex --test "pytest tests/test_scheduler.py"

You get:

• A side-by-side terminal table ranked by test pass-rate → cost → diff size → latency
• A single-file HTML report with each agent's full diff
• One .patch file per agent — apply the winner with git apply

Install

pip install codejoust

You'll also need whichever agent CLIs you want to race. Install as many or as few as you like: