kun@MacBook-Pro dev % agent

Kun How should browser feedback reach the agent, and where should the review queue live?

Agent Long-polling and a local file store are the simplest start, with tradeoffs to review.

> enter

Drafting the technical approach...

Creating .lavish/plan.html

$ npx -y lavish-axi .lavish/plan.html

Opened in Lavish Editor.

localhost:4387/session/9a7c2f31

Technical plan

Element-specific feedback without screenshots

Send precise browser annotations from a reviewed HTML artifact back to the coding agent.

Architecture diagram

Initial technical approach

Browser review click element + write prompt

↓

Long-poll API poll waits for feedback

↓

Agent patch apply targeted edits

SDK in iframe selector + snapshot

↓

WebSocket channel push feedback to agent

↓

Agent patch apply targeted edits

Session store

Pick the storage for the review queue

Server restart: poll resumes from persisted state.

Missing session: reopen the same HTML path.

Bad asset path: reject traversal outside the artifact.

Agent

Plan opened. Click any element and tell me what to change.

You

Use WebSocket instead of long-poll

Session store: SQLite

What are the failure modes here?

Working...

Agent

Done. The plan now uses WebSocket, SQLite, and lists failure modes.

use WebSocketchose SQLite

Write a message for the agent...

LavishEditor

For when a rich editor is not rich enough.

/lavish

LavishEditor

For when a rich editor is not rich enough.

/lavish