A browser roguelite auto-battler about an AI agent stuck in an engineering loop — built almost entirely by AI agents working in a self-maintaining harness.
You play an agent. Pick a harness (model, tools, skills), choose a system prompt, then fight your way through a map of tasks, standups and package registries to ship the release before someone presses Ctrl+C. Fights resolve automatically; your decisions happen between them. The core tension is the context window: every tool fills it, enemies inject noise, and an overflowing context forces a costly compaction.
Status: work in progress, milestone M1 (vertical slice: phase 1 "Implement"). A full run from title to boss is playable. See forge/BOARD.md for live progress and docs/game/milestones.md for the roadmap (M1 slice → M2 full content → M3 polish).
Requires Node 24 (TypeScript runs natively, no build step for tools).
npm install
npm run dev # play in the browser (http://localhost:5173)?sandbox on the dev server opens a combat sandbox to replay single fights.
| Command | What it does |
|---|---|
npm run check |
The single quality gate: typecheck, Biome, Vitest + coverage, build, Playwright e2e, harness lint |
npm test |
Unit and property tests (Vitest, fast-check) |
npm run e2e |
Browser tests against the production build (Playwright) |
npm run balance |
Bots play thousands of seeded runs and report win rates |
npm run golden:update |
Regenerate the golden combat logs after an intended sim change |
- Deterministic simulation (
src/sim): integer-only maths, a seeded forkable RNG, no clock or DOM. A fight resolves instantly into an event log; the same seed always produces a byte-identical log. - Run reducer (
src/run): the whole run is a pure reducer over serializable actions, so a save is just seed + setup + action log and can be replayed exactly. - Content as typed data (
src/content): tools, enemies, events and lessons are TypeScript data with a small trigger/effect DSL; tooltips are generated from the data. - UI (
src/ui): Preact + signals. The combat screen replays the event log at 1x/2x/4x with seeking. The visual style is a terminal/IDE look in "Crimson" red. - Saves (
src/save): versioned schema, checksum, migrations, export as a string.
Stack: Vite, TypeScript (strict), Preact, Biome, Vitest, fast-check, Playwright. Architecture decisions are recorded as ADRs in docs/architecture/adr.
This repository is developed by an orchestrator agent (Claude) that plans, delegates and verifies; implementer and reviewer subagents do the work. The repo is the single source of truth — every plan, decision and review lives in Markdown files here.
- Knowledge base: every doc has YAML frontmatter (title, summary, keywords, …) and every
folder has a generated
INDEX.md, so agents navigate by summaries first. - Forge (forge/): an in-repo kanban — epics, tasks with acceptance criteria, reviews and retrospectives — with a generated BOARD.md.
- Workflow: one task = one subagent run = one independent review = one commit
(
T###: title). Parallel tasks run in git worktrees. - Budgets: file sizes, folder sizes, diff sizes, WIP limits, test coverage and more are limits in harness.config.json, enforced by a linter and hooks.
- Self-improvement: regular retrospectives (forge/retros) change the harness itself — skills, rules and budgets — based on evidence from reviews.
Start with CLAUDE.md and docs/harness/overview.md.
| Path | Contents |
|---|---|
src/ |
The game: sim, run, content, ui, save |
tests/ |
Architecture rules and Playwright e2e specs |
tools/ |
Harness tooling, check runner, golden logs, balance bots |
docs/game/ |
Game design document (systems, content, UX) |
docs/architecture/ |
Architecture, testing strategy, ADRs |
docs/harness/ |
How the agent harness works |
docs/research/ |
Research notes behind the design and the harness |
forge/ |
Board, epics, tasks, reviews, retros |
.claude/ |
Agent roles, skills and hooks |
CONCEPT.md |
The original one-page vision (German) |
