潛 IDE · Vibe Coding (applied layer)

· Context · Files文件

▾ src/

index.ts

routes/approvals.ts

db/schema.ts

rbac.ts

▾ web/

App.tsx

▾ migrations/

001_init.sql

package.json

② Artifact · Code代码 routes/approvals.ts

import { Hono } from "hono"
import { db } from "../db"
import { requireRole } from "../rbac"

// approval state machine: draft→submitted→approved/rejected→paid
const app = new Hono()

app.post("/requests", requireRole("employee"), async (c) => {
  const body = await c.req.json()
  return c.json(await db.requests.create(body))
})

app.post("/requests/:id/approve", requireRole("manager"), /* … */)

· Activity · terminal / tool calls (peripheral · auditable)终端 / 工具调用（外围 · 可审计）

scaffold · generated 7 files + migration + seed生成 7 文件 + 迁移 + 种子

run_migration · created requests/users/approvals tablesrequests/users/approvals 建表

smoke_test · /health 200 · boot 0.3s启动 0.3s

bun run dev · watching…

In most AI IDEs, this plane is a stream of tool calls (“installing dependencies… editing App.tsx…”). 潛 IDE replaces it with the agent's understanding of what to build — requirements, architecture decisions, plan, and risks. The code sits in the middle and the terminal below, both receding to the periphery.

在多数 AI IDE，这块是工具调用流（“正在装依赖…正在编辑 App.tsx…”）。潛 IDE 把它换成 agent 对“要建什么”的理解——需求、架构决策、计划与风险。代码在中间、终端在下方，都退为外围。

① the agent's understandingUnderstand the ask

Your question

Wanted a small expense-approval tool and gave a one-line requirement. Get a working v0 and know what to decide next

My read

still forming — don't trust me yet…

Why I think so

Requirement as understood: three roles (employee/manager/finance) + attachments + state transitions + auth fairly sure

Because roles and flow unpacked from the one-line requirement.

how I confirmed this ▾

I considered but ruled out 1 idea ▾