Files
Kokorone/.agents/skills/prisma-cli/references/agent-safety.md
T
Jeffery f44200f543 feat: 完成 B 群組 — 資料層與核心領域模型
- 導入 Prisma 7.9.1 + SQLite(better-sqlite3 driver adapter),prisma.config.ts 管理連線設定
- 完整資料模型:User/Work/Character/CharacterAlias/EpisodicMemory/SemanticMemory/
  ProceduralRule/EmotionState/Relationship/SentimentLedgerEntry
- packages/db:Prisma Client 存取層(ESM,因 Prisma 7 產出的 generated client 僅支援 ESM)
- apps/api 隨之改為 ESM 以相容 packages/db
- packages/shared:新增角色/記憶/情緒/關係共用型別,api 與 web 皆從此匯入
- prisma/seed.ts:建立測試使用者與元氣型測試角色,含完整關係帳本與記憶種子資料
- apps/api:新增 GET /characters
- scripts/smoke/B.mjs:驗證資料表齊全、API 讀出種子角色、關係與記憶可寫入讀出
- 保留 Prisma 官方隨 CLI 附的 agent skill 文件(.agents/skills 等),供後續群組查閱

npm run restart && npm run smoke -- B 皆通過(B-V),A 群組冒煙測試無回歸。
2026-08-13 09:43:15 +08:00

1.1 KiB

AI safety checkpoint for destructive commands

Prisma detects common AI-agent environments and blocks these commands until the user gives explicit consent:

  • prisma migrate reset
  • prisma db push --force-reset
  • prisma db push --accept-data-loss

Required workflow

  1. Inspect the target database/config and explain exactly what can be deleted or reset.
  2. Ask the user for explicit consent immediately before the action.
  3. Run the command only after that consent.

For an agent-run subprocess, Prisma accepts the exact consent text through:

PRISMA_USER_CONSENT_FOR_DANGEROUS_AI_ACTION='<exact user consent message>' prisma migrate reset --force

The value must match the user's message exactly and must not contain added quotes or newlines. Never fabricate consent, reuse an old unrelated approval, or bypass the checkpoint by hiding agent-detection environment variables.

The MCP server has no migrate-reset tool. Use the shell command only after consent.

Reference