Builder · Public · built in 2026 · I run it on tunyt daily
melon
Most coding agents tell you they're done. melon shows you: the checks that passed, a second model's review of the real diff, and every file reference in the answer checked against your tree.
- Offline tests
- 449passing in about 27 seconds
- Cached input in a real run
- 85%one 21-call session on tunyt's web app
- System prompt
- ~0.7k tokensdown from ~3k
How it works
- Your checks run on the changed files only. A failure goes back to the model as a repair request.
- A separate reviewer on a different model reads the real diff, and a verifier tries to disprove each finding. A confirmed HIGH is run in a throwaway copy of the change.
- Every path and line number in the answer is looked up on disk before you see it.
Why it's cheap
An agent run re-sends the whole conversation on every call, so melon keeps a map of which lines of each file the model has seen, returns only what's new, and finds code with ranked search and outlines before it reads anything. It runs Claude, GPT and Codex, DeepSeek and anything behind an OpenAI-style API, and routes each request to a cheap or strong model by difficulty.
Safety
The rules live in code. File tools are confined to your folder, shell commands are scanned before they run, and on macOS an OS sandbox sits under both. Before each request it records the working tree so /undo works.
What it can't do yet
No image input, no JSON output for CI, and the OS sandbox is macOS only. No head-to-head benchmark against other harnesses has been run, so I don't claim it beats them. I use it every day on tunyt, and that's my evidence.