← 提示词库 Meta/muse-agent/skills/replit/SKILL.md 原文 md
🌐 中英双语对照

name: "replit"
description: "Read, create, update, and publish apps through Replit's official MCP server."
icon: "replit"
metadata: { "includeInPrompt": false }

Replit / Replit

Use the installed replit CLI. Start with replit status. If it reports
not_connected, run replit authorize-url and share only the returned
connect_url with the user.

使用已安装的 replit CLI。先运行 replit status。如果报告 not_connected,则运行 replit authorize-url,并且只把返回的 connect_url 分享给用户。

OAuth uses Dynamic Client Registration and PKCE through authd. No shared client
secret enters Muse, and credentials must never be requested in chat.
The first connection requests read-only access. Before a write, or after one
fails for missing access, run replit status --for-command <tool-name>. If it
returns scope_status: not_granted, copy scope_add_url exactly and post it on
its own line as [Additional Replit access](<scope_add_url>) so the client
renders the native access button, then wait for consent to finish before
retrying once. OAuth access does not replace Hatch approval. Never construct a
scope URL or ask for tokens in chat.

OAuth 通过 authd 使用动态客户端注册(Dynamic Client Registration)和 PKCE。任何共享的客户端密钥都不会进入 Muse,且绝不能在对话中索要凭据。首次连接只请求只读权限。在执行写操作之前,或某次写操作因缺少权限而失败之后,运行 replit status --for-command <tool-name>。如果返回 scope_status: not_granted,原样复制 scope_add_url,并将其单独一行以 [Additional Replit access](<scope_add_url>) 的形式发布,以便客户端渲染原生的授权按钮,然后等待授权完成后再重试一次。OAuth 授权不能替代 Hatch 审批。绝不在对话中自行构造权限 URL 或索要令牌。

【评论】此处体现了最小权限设计:默认只读,写操作前逐项申请权限,授权链接交给客户端原生按钮完成,模型不经手令牌。

Run replit list-tools to inspect the live provider catalogue and schemas,
then call an advertised tool with:

运行 replit list-tools 查看实时的提供方目录及各工具的 schema,然后用以下方式调用已公布的工具:

replit call-tool --name <tool> --arguments-json '<json-object>'

list-tools exposes only reviewed Replit tools and includes each tool's
hatch_permission, hatch_action, and hatch_permission_label. Unknown or
new provider tools remain unavailable until reviewed. Read permissions follow
the user's connector settings; creating, editing, and publishing apps use
separate granular approvals. Do not retry a failed or timed-out write
automatically because its side effect may have completed.

list-tools 只暴露经过审核的 Replit 工具,并包含每个工具的 hatch_permission、hatch_action 和 hatch_permission_label。未知或新出现的提供方工具在审核之前保持不可用。读取权限遵循用户的连接器设置;创建、编辑和发布应用则分别使用独立的细粒度审批。不要自动重试失败或超时的写操作,因为其副作用可能已经生效。

【评论】"不要自动重试写操作"是针对非幂等副作用的安全条款——失败可能只是响应丢失,操作实际已经执行,重试会造成重复写入。