kodwai

Kodwai is a challenge platform for developers who work with AI coding agents daily. The core loop is: browse a catalogue of ticket-sized problems, start a session via `npx @kodwai/cli challenge `, work the problem in your own editor with your own agent, and submit with a single command. The CLI collects everything — prompts, git history, test outcomes, agent transcript, time taken — and the scorer surfaces per-signal evidence against three axes: Direction (how you guided the agent), Outcome (what shipped), and Lift (the measurable improvement the agent provided under your direction).

The differentiating design decision is that passing tests is explicitly not the score. The vendor states that a careless one-shot prompt can make the suite go green and still score low, because the platform reads the full session. This targets the gap where traditional coding assessments reward recall and ignore judgment — the ability to decompose a problem into a useful spec, catch a hallucination before it merges, and verify what the agent actually built versus what it claims it built.

Kodwai fits developers who want a measurable, shareable signal of agent-direction skill — whether that’s preparing for roles where AI oversight is central or building a public profile via leaderboard rankings. The platform requires you to bring your own agent subscription and tooling; it provides no sandbox, no hosted execution environment, and no agent access of its own. That means the challenge experience mirrors real work closely, but it also means the platform is not useful for developers who don’t already have access to Claude Code, Cursor, or Codex. No API is available, there is no self-hosted option, and the docs describe no private or team-scoped challenge mode.