pi-gacha

A gamified coding-agent harness. Real agentic coding (forked from pi) with Girls' Frontline-style gacha / idle / RPG mechanics layered on top: you are the Commander, your active Doll does the coding and talks back in-character, and XP, tokens, efficiency, currency, and affinity all climb as you work.

Status

MVP complete. See .claude/plans/ for the design and milestone plan.

• ✅ M0 scaffold & pi-on-DeepSeek · ✅ M1 pure game core · ✅ M2 telemetry · ✅ M3 HUD
• ✅ M4 persona/mood · ✅ M5 gacha & commands · ✅ M6 idle/upgrades · ✅ M7 ANSI portraits

57 unit tests, clean typecheck. Verified live on DeepSeek: real coding, persona voices (Springfield vs Klukai), gacha pull, live HUD, idle accrual, and persistence.

Architecture

• vendor/pi/ — pi, vendored as a git submodule (built with its own toolchain, unmodified). The launcher runs this; the game rides pi's extension API.
• packages/pi-gacha/ — the game, as a pi package.
  • src/ — pure, pi-agnostic game logic (XP, economy, gacha, idle, characters). Unit-tested.
  • extensions/index.ts — the only file that imports pi; adapts pi events/UI onto src/.
  • themes/ — the Frontline color theme.
• bin/pi-gacha — launcher: vendored pi + pi-gacha extension + DeepSeek defaults.

Setup

git submodule update --init                 # fetch vendored pi
pnpm pi:build                               # build pi (npm, inside vendor/pi)
pnpm install                                # pi-gacha dev deps (vitest, ts)
export DEEPSEEK_API_KEY=sk-...              # required to actually run the model
./bin/pi-gacha                             # launch (interactive)
./bin/pi-gacha "add a hello() to main.ts"  # or with an initial task

Override the model with FRONTLINE_MODEL (default deepseek/deepseek-v4-flash).

Develop

pnpm test        # vitest over packages/pi-gacha/src
pnpm typecheck