How it's built
A solo project, built with Claude Code as a team of assistants, and measured by tests and bots instead of faith.
1,271 tests
The whole rules engine is covered by 1,271 automated tests that run in about 40 seconds. Items never touch combat code: they only subscribe to an event bus, so a new object is a new file and a new test. The randomness of every room is seeded, so any run can be replayed exactly.
Working with Claude Code
mainonly holds what is validated. Everything new lives in a branch, with green tests, and is merged after my explicit approval.- Agents write code, tests and documentation. They cannot approve art: pixel art goes in two phases (static master, then one animation at a time) and I sign off each step.
- Cheap models do the mechanical searching and routine edits; the expensive one does architecture and hard debugging.
The bot simulator
A headless simulator plays full runs with the same logic as the game (about 0.02 s per run). It does not say whether the game is fun, but it is very good at finding what is broken. What it found in its first pass (1,500 runs per policy):
| Finding | Result |
|---|---|
| Maps with no way to the boss without a key | 4.4 to 5.8% of maps; fixed, covered by tests on every floor and 250 seeds |
| Skill gap, greedy vs random | 7 times more rooms won |
| Attrition | 5.7 HP lost per won fight; bots reach the first boss with 9 to 13 of 33 HP |
| Fights that never end | 84 of 1,500 runs hit the 80-turn cap |
| Same seed, different run | 9 modules used global randomness; fixed, guarded by a test |
Balance decisions
The Garden Gnome boss won 0 of 381 fights against the simple bots, so it looked overtuned. A rollout planner (it tries moves ahead before choosing) won 48% of them, against 91% for the Cursed Rosebush and 92% for the Iron Toad. Verdict: the Garden Gnome is beatable with good positioning, and the old number was mostly the bot's limit. The same planner won 9.3 rooms per run on average against 3.35 for the greedy bot, and it exposed a real game crash that now has a test.
Source: the project's own simulation report, 12 November 2026. The simulator measures regressions and outliers; absolute win rates still need human players.