lmong11/dsh-game-center--packages-bundle-game-center ↗★ 0

@deepseek-ai/dsh-game-center

Optional DeepSeek Harness bundle layer for the durable game runtime and game-center surfaces 适合想在DSH中启用游戏中心及多种牌类游戏运行时的用户,需叠加在Web应用之后。

패키지
@deepseek-ai/dsh-game-center
호환성
미검증
Harness peer 범위
workspace:^
Cordis peer 범위
workspace:^
버전
0.1.0-rc.5
라이선스
MIT
최근 업데이트
2026. 8. 24.

설치

$npx -p @deepseek-ai/dsh dsh plugin --profile web add github:lmong11/dsh-game-center#e9f0600e785a3dd15e36df58023298012dc228b3&path:packages/bundle/game-center

@deepseek-ai/dsh-game-center

English | 中文

Optional profile overlay for Game Center. Add this bundle after @deepseek-ai/dsh-web-app; its patch inserts the durable ctx.games runtime, cross-game ctx.gameAgentCatalog, isolated ctx.gameAgentPlayers seat controller, replay audit service, persistent bounded local Arena, doudizhu-rules-v1, holdem-rules-v1, guandan-rules-v1, and mahjong-simple-v1 providers and tables, credential-bound Remote, React-independent Client state, sidebar/overlay shell, and keyed fixture renderer without changing the default Web profile. Hold'em contributes the first five original fictional versioned players and the first semantic Agent editor.

From this repository, build once and launch the shipped profile with pnpm dsh --profile game-center. The ordinary pnpm dsh web profile remains unchanged.

Temporary public friend rooms remain a separate, removable layer. Use pnpm dsh --profile friend-room to compose @deepseek-ai/dsh-game-room-bundle after this bundle; ordinary Game Center never opens a Guest listener.

The overlay expects the underlying profile to provide storage-domain. It does not select or own a storage backend.

Model Experience

Bundle layer

What the model sees

Nothing until a match contains an Agent seat. Each such seat receives an independent session whose complete prompt contains only the fixed game-player contract, the game profile, and optional seat strategy. Its tool directory contains only game_act; each turn receives the authorized seat projection and legal actions.

Token effect

Zero for human-only matches. Agent matches add one bounded seat observation and the game_act schema to each model request.

KV Cache effect

Stable within one Agent seat while its game profile, model selection, and sole tool schema remain unchanged. Each observation appends after the reusable prefix.

Known Limitations and Deferred Work

  • The playable providers are fixed to doudizhu-rules-v1, non-cash holdem-rules-v1, one documented guandan-rules-v1, and the explicitly simplified mahjong-simple-v1; other regional Mahjong rules need separate providers.
  • Agent seats accept any provider/model route registered with ctx.llm; unavailable routes fail that model turn and converge through the configured legal fallback.
  • It must be composed over a Web profile that already mounts storage, storage-domain, Typert transport, and the shell slots.
  • Browser identity is local to one Harness Home and is not an Internet account.
  • Public replays are redacted; owner-only referee replays contain full hidden state and must be protected. Catalog bindings expose ids and digests but not private strategy prose. Arena results persist locally as descriptive statistics, not a public benchmark.