Hallucinating
great code
since 2026.

The product is called gm

You describe something that does not exist. gm invents the whole file set, writes the tests that would catch it lying, runs them, and feeds every failure back into the next attempt until the suite is green. If it cannot get there, it says so.

$ gm "an LRU cache with TTL eviction"

round 1: hallucinating...
round 1: 3 files (package.json, src/lru.ts, test/lru.test.ts)
round 1: verifying with node --test test/lru.test.ts
round 1: failed (exit 1) in 0.4s
round 2: hallucinating...
round 2: verifying with node --test test/lru.test.ts
round 2: passed in 0.3s

PASS  ./lru  (2 rounds, node --test test/lru.test.ts)

The failing round is the point. A model that gets it right first try was answering an easy question; a model that finds its own bug in round two is doing the work. Every round is recorded, so the repair is auditable rather than magic.

How the loop works, flag by flag →

Invent wildly. Verify hard.

1. Hallucinate

   One shot at the entire thing. Every file is complete, with no stubs and no TODOs. Partial output is rejected before it ever reaches your disk.

2. Verify

   The run has to prove itself by executing. The tests run under a hard timeout, with no network and no install step, in a sandbox that cannot escape its own directory.

3. Repair

   A failure is not an apology, it is the next round's input: the exact command, exit code and output go back in, and the model attacks its own mistake.