volcengine · yaozheng-fang · Jul 22, 2026 · Jul 22, 2026 · Jul 22, 2026 · Jul 22, 2026
diff --git a/README.md b/README.md
@@ -104,6 +104,14 @@ between this live metadata and control-plane information without exposing prompt
 or credentials. The same metadata advertises mounted smart-search sources, so
 Studio can disable unavailable sources up front and query the Agent's web-search
 tool, KnowledgeBase, or long-term memory without exposing component credentials.
+Studio also provides an isolated Insight Sandbox for temporary Codex
+conversations. It reuses a dedicated AgentKit CodeEnv tool, creates a fresh
+user-owned Sandbox session, and deletes that session on exit without adding the
+conversation to normal Studio history. Reloading may create another temporary
+session; AgentKit reclaims abandoned sessions automatically when their TTL ends.
+When configuring skills, Studio can also browse account-scoped AgentKit Skill
+Spaces and their paginated skill lists by region and project. These requests are
+signed on the server, so browser clients never receive Volcengine credentials.
 
 The Studio deployment flow lists Feishu, knowledge-base, short-/long-term
 memory, and observability settings in their feature sections. Values entered
@@ -198,13 +206,17 @@ VeADK provides several useful command line tools for faster deployment and optim
   `--region`, automatically locate the Identity user pool across Beijing and
   Shanghai, and select the VeFaaS project with `--project` (default `default`);
   custom local or remote logo images are bundled into the deployment; the
-  deployed client skips the second OAuth consent confirmation after login
+  deployed client skips the second OAuth consent confirmation after login;
+  two dedicated AgentKit CodeEnv Tools are created automatically for temporary
+  chats and Skill creation unless their IDs are supplied with
+  `--sandbox-chat-codex-tool-id` and `--sandbox-skill-creator-tool-id`
 - `veadk studio update --vefaas-app-name <app-name>`: build the frontend from a
   local VeADK source checkout and release it through the existing VeFaaS
   Application and Function. Omit `--region` and `--project` to search Beijing,
   Shanghai, and all visible projects. Existing URL, SSO, IAM, gateway,
   environment variables, title, and logo are preserved; pass `--site-title` or
-  `--site-logo` only when those branding values should be replaced
+  `--site-logo` only when those branding values should be replaced. Sandbox Tool
+  IDs are also preserved unless the corresponding deploy option is supplied
 
 Studio can assign comma-separated local usernames or OAuth emails to the
 `admin` and `developer` roles:

diff --git a/docs/content/docs/framework/frontend.en.mdx b/docs/content/docs/framework/frontend.en.mdx
@@ -19,14 +19,28 @@ VeADK ships a React frontend that renders [A2UI](/en/docs/framework/a2ui) (agent
   and replay both user and model media from history. Chat images use compact
   thumbnails and open in a zoomable full-screen viewer.
 - **Skill and sub-agent invocation**: type `/` to select a mounted skill or `@` to route the turn to an eligible sub-agent.
+- **New-chat modes**: retain the Agent conversation path, start a temporary
+  Codex Agent conversation in an AgentKit Sandbox, or generate two real Skill
+  candidates in isolated AgentKit CodeEnv sessions for comparison, ZIP
+  download, and AgentKit publishing.
 - **Agent picker**: switch agents from the top-left; long lists scroll within
   the viewport, and hovering an agent shows its model and mounted tools.
 - **Session history**: auto-saved, time-sorted, reopen or delete.
+- **Temporary session**: start an isolated temporary AgentKit Sandbox from the
+  new-chat composer. Studio reuses its dedicated CodeEnv tool,
+  creates a fresh user-owned Sandbox session, and runs multi-turn Codex chat
+  with live public reasoning summaries, commands, file changes, MCP calls, and
+  final replies.
+  Exiting deletes the cloud session without writing it to normal chat history;
+  after a reload, users can start another session and AgentKit reclaims the
+  abandoned session automatically when its TTL ends.
 - **Smart search**: the *Session* source full-text-searches the current agent's history; the *Web* source calls the agent's mounted web-search tool live (using credentials from the server's environment variables).
 - **Add an AgentKit agent**: paste a URL + API key to connect a remote agent over the ADK protocol; it then appears in the picker.
 - **Custom-agent workbench**: author system prompts in a rich Markdown editor
   with heading and list shortcuts. Local skills accept a dropped folder or ZIP
-  and detect the format automatically. Then debug the agent beside its configuration.
+  and detect the format automatically. You can also browse account-scoped
+  AgentKit Skill Spaces and their skills by region and project. Then debug the
+  agent beside its configuration.
   Initialization, session, and conversation failures show credential-redacted
   error details and Runner logs that can be expanded and copied. The deployment
   page shows the Agent topology on the left, with hover details, configuration
@@ -45,6 +59,14 @@ credentials to local debug runs and deployed AgentKit runtimes, allowing VeADK
 to resolve Ark, embedding, media, speech, VeSearch, and APMPlus keys without
 asking for them again.
 
+Insight Sandbox requires server-side `VOLCENGINE_ACCESS_KEY`,
+`VOLCENGINE_SECRET_KEY`, `MODEL_AGENT_API_KEY`, and `MODEL_AGENT_NAME` values.
+Volcengine credentials, model credentials, and the authorized Sandbox endpoint
+remain on the Studio server and are never returned to the browser.
+Temporary Sandbox state is process-local. Run Studio with one server process,
+or configure session affinity so create, message, and exit requests from one
+browser reach the same instance.
+
 When the Feishu channel is enabled or disabled on the deployment page, Studio
 regenerates the project so `app.py`, the `extensions` dependency, and runtime
 environment variables stay aligned before deployment.
@@ -146,6 +168,70 @@ After selecting a sub-agent, the `/` menu shows skills mounted on that target. C
 
 The frontend sends structured `veadkInvocation` metadata instead of inferring invocation intent from the message string. The backend plugin directs ADK to call the skill tool or invoke `transfer_to_agent` one tree edge at a time until it reaches the target. The same metadata is written to the first Google GenAI `Part`, so reopening session history restores the `/skill` and `@agent` chips.
 
+### Skill creation mode
+
+The new-chat **Skill creation** mode runs
+`doubao-seed-2-0-pro-260215` and `deepseek-v4-flash-260425` in parallel through
+Ark Responses. Each candidate gets an independent AgentKit CodeEnv session.
+The server validates the generated directory name, `SKILL.md` frontmatter,
+file count, size, and safe paths before packaging a ZIP. AgentKit credential
+hosting keeps the real Ark API key; Studio and Sandbox receive only its
+revocable gateway ticket. Per-job Session variables cannot override that
+hosted ticket, and Studio accepts only Volcengine HTTPS credential relay URLs.
+Model credentials are never returned to the browser, and the Skill creation
+APIs require the Studio developer or admin role.
+
+After submission, both candidate conversations appear immediately and
+independently render public reasoning summaries, tool calls, and assistant
+messages returned by the generator. Private chain-of-thought and credentials
+never enter the page. Completed candidates still support preview, ZIP download,
+and AgentKit publishing. Once the real job ID is available, the frontend resumes
+polling directly from Sandbox state. An interrupted creation stream or a brief
+AgentKit control-plane delay no longer turns an intermediate state into an
+expired job.
+
+Before starting a local server, configure separate `Ready` CodeEnv Tools for
+temporary chats and Skill creation:
+
+```bash
+export SANDBOX_CHAT_CODEX=<chat-code-env-tool-id>
+export SANDBOX_SKILL_CREATOR=<skill-code-env-tool-id>
+veadk frontend --agents-dir examples
+```
+
+For local Studio, run the AgentKit `credential-hosting` command and choose to
+write its result into both dedicated CodeEnv Tools. Adding a
+candidate to AgentKit uploads its ZIP to TOS, then creates or updates the Skill
+through the Skills API. It is published to a SkillSpace only when the user
+provides SkillSpace IDs.
+
+| Environment variable | Default | Description |
+| :-- | :-- | :-- |
+| `SANDBOX_CHAT_CODEX` | — | AgentKit CodeEnv Tool ID dedicated to temporary chats. |
+| `SANDBOX_SKILL_CREATOR` | — | AgentKit CodeEnv Tool ID dedicated to Skill creation. |
+| `VEADK_SKILL_CREATOR_TOS_BUCKET` | AgentKit account default | TOS bucket used for published artifacts. |
+| `VEADK_SKILL_CREATOR_TOS_PREFIX` | `agentkit/skills` | TOS object-key prefix. |
+| `VEADK_SKILL_CREATOR_PROJECT_NAME` | — | Project name used when creating a Skill. |
+
+Candidate sessions have a 30-minute TTL and are deleted immediately when the
+user starts over or leaves the job. Job state is stored in Sandbox rather than
+frontend-process memory, so polling and downloads continue to work when FaaS
+requests reach different instances.
+
+On the first cloud deployment, the command creates two independent CodeEnv
+Tools with the deployer's AK/SK when Tool IDs are omitted. Existing Tools can be
+selected with `--sandbox-chat-codex-tool-id` and
+`--sandbox-skill-creator-tool-id`. AgentKit credential hosting stores the Ark
+credential in KMS and binds only the relay URL and revocable ticket to both
+Tools. The VeFaaS Function receives only Tool IDs, never model credentials:
+
+```bash
+veadk studio deploy \
+  --user-pool-id <pool-id> \
+  --allowed-client-id <client-id> \
+  --vefaas-app-name <app-name>
+```
+
 ### Custom Studio branding
 
 Use `--site-title` for a system name of up to six characters and `--site-logo`
@@ -199,6 +285,9 @@ and `--site-logo` replace cloud branding only when explicitly provided;
 otherwise the current values are retained. Use `--path` to select a source
 checkout; it defaults to the current directory. Node.js and npm must be
 installed locally.
+`--sandbox-chat-codex-tool-id` and `--sandbox-skill-creator-tool-id` likewise
+replace only the explicitly supplied environment value. Omitting either option
+keeps that Tool ID from the deployed Function instead of clearing it.
 
 When deployment registers the callback URL, it keeps the VeIdentity login page
 enabled and enables skip-consent for that client. This avoids showing a second

diff --git a/docs/content/docs/framework/frontend.mdx b/docs/content/docs/framework/frontend.mdx
@@ -18,12 +18,19 @@ VeADK 自带一个 React 前端，用于渲染智能体经由 Google ADK API Ser
 - **多模态消息**：上传图片、TXT/Markdown、PDF 与视频；用户附件和模型返回的媒体均支持预览与历史回放。
   对话中的图片默认以紧凑尺寸展示，点击后可全屏缩放预览。
 - **技能与子 Agent 调用**：在输入框输入 `/` 选择挂载技能，输入 `@` 将本轮交给可选的子 Agent。
+- **新会话模式**：保留 Agent 对话模式；临时会话会在 AgentKit Sandbox 中启动 Codex Agent；Skill 创建模式会在两个独立的 AgentKit CodeEnv Session 中使用固定模型做真实 A/B 生成，完成后可对比、下载 ZIP 或添加到 AgentKit。
 - **Agent 选择器**：左上角切换 agent；长列表在视口内独立滚动，悬停可查看该 agent 的模型与挂载的工具。
 - **历史会话**：自动保存、按时间排序，可重新打开或删除。
+- **临时会话**：从新会话输入框启动独立的 AgentKit 沙箱临时会话。
+  Studio 会复用专属 CodeEnv Tool、为当前用户创建全新的 Sandbox Session，并通过
+  Codex 进行多轮对话，实时展示公开思考摘要、命令、文件修改、MCP 调用和最终回复。
+  退出后云端 Session 会被删除，内容不会写入普通历史会话；
+  页面刷新后可直接新建，遗留 Session 由 AgentKit 在 TTL 到期后自动回收。
 - **智能搜索**：「会话」源在当前 agent 的历史消息中做全文检索；「网页」源调用该 agent 挂载的联网搜索工具实时检索（使用服务端环境变量里的凭据）。
 - **添加 AgentKit 智能体**：填入访问地址 + API Key，按 ADK 协议接入远程 agent，接入后出现在选择器中。
 - **自定义 Agent 工作台**：系统提示词支持所见即所得的 Markdown
-  编辑与标题、列表等快捷输入；本地技能可直接拖入文件夹或 ZIP，格式会自动识别。
+  编辑与标题、列表等快捷输入；本地技能可直接拖入文件夹或 ZIP，格式会自动识别，
+  也可按地域和项目浏览账号下的 AgentKit Skill Space 及其中的技能。
   完成配置后可在同页调试。调试初始化、
   会话或对话失败时会展示经过凭据脱敏的具体错误和 Runner 日志，
   并支持展开与复制。部署页左侧展示 Agent 拓扑，悬停可核对节点信息与配置，
@@ -39,6 +46,13 @@ VeADK 自带一个 React 前端，用于渲染智能体经由 Google ADK API Ser
 VeADK 会据此获取 Ark、Embedding、图像、视频、语音、VeSearch 和 APMPlus 的
 API Key，挂载组件时无需重复填写。
 
+“灵光一现”需要 Studio 服务端配置 `VOLCENGINE_ACCESS_KEY`、
+`VOLCENGINE_SECRET_KEY`、`MODEL_AGENT_API_KEY` 和 `MODEL_AGENT_NAME`。
+火山引擎凭证、模型凭证以及带授权信息的 Sandbox Endpoint 始终保留在服务端，
+不会返回浏览器。
+临时沙箱状态保存在 Studio 进程内；部署时应使用单个服务进程，或配置会话亲和，
+确保同一浏览器的创建、对话和退出请求到达同一个实例。
+
 在部署页打开或关闭飞书渠道时，Studio 会重新生成项目，确保 `app.py`、
 `extensions` 依赖和运行时环境变量在部署前保持一致。
 
@@ -134,6 +148,59 @@ veadk frontend --agents-dir examples
 
 前端不会从消息字符串中猜测调用意图，而是发送结构化的 `veadkInvocation` metadata。后端插件据此要求 ADK 调用技能工具，或沿 Agent 树逐级调用 `transfer_to_agent`，直到到达目标节点。同一份 metadata 也会写入首个 Google GenAI `Part`，因此重新加载历史会话后仍能恢复 `/skill` 与 `@agent` chip。
 
+### Skill 创建模式
+
+新会话输入框中的「Skill 创建」会固定使用
+`doubao-seed-2-0-pro-260215` 和 `deepseek-v4-flash-260425`，通过 Ark
+Responses API 并行生成两个候选方案。每个候选使用独立的 AgentKit CodeEnv
+Session，生成结果会校验目录名、`SKILL.md` frontmatter、文件数量、大小和路径安全，
+再打包为 ZIP。真实 Ark API Key 由 AgentKit 凭据托管保存，Studio 和 Sandbox
+只使用可撤销的网关票据；每个任务的 Session 环境变量不会覆盖托管票据，且服务端
+只接受 Volcengine HTTPS 凭据中继地址。浏览器不会收到模型凭据。Skill 创建接口仅允许
+Studio developer 和 admin 使用。
+
+提交需求后会立即显示两个候选对话流，各自持续渲染生成器返回的公开思考摘要、工具
+调用和助手回复。私有思维链和凭据不会进入页面。候选完成后仍可预览、下载 ZIP 或
+添加到 AgentKit。前端拿到真实任务 ID 后会直接从 Sandbox 状态恢复轮询；创建流中断
+或 AgentKit 控制面短暂未发现新 Session 时，不会把中间状态误判为任务过期。
+
+本地启动前需要分别配置临时会话和 Skill 创建使用的两个 `Ready` CodeEnv Tool：
+
+```bash
+export SANDBOX_CHAT_CODEX=<chat-code-env-tool-id>
+export SANDBOX_SKILL_CREATOR=<skill-code-env-tool-id>
+veadk frontend --agents-dir examples
+```
+
+本地运行前使用 AgentKit 的 `credential-hosting` 命令，并将结果分别写入上述 CodeEnv
+Tool。添加到 AgentKit 时，服务端会先将
+ZIP 上传到 TOS，再调用 Skills API 创建或更新 Skill；只有用户填写 SkillSpace ID
+时才发布到对应空间。以下变量均为可选覆盖项：
+
+| 环境变量 | 默认值 | 说明 |
+| :-- | :-- | :-- |
+| `SANDBOX_CHAT_CODEX` | — | 临时会话专用 AgentKit CodeEnv Tool ID。 |
+| `SANDBOX_SKILL_CREATOR` | — | Skill 创建专用 AgentKit CodeEnv Tool ID。 |
+| `VEADK_SKILL_CREATOR_TOS_BUCKET` | AgentKit 账户默认 Bucket | 发布产物使用的 TOS Bucket。 |
+| `VEADK_SKILL_CREATOR_TOS_PREFIX` | `agentkit/skills` | TOS 对象 Key 前缀。 |
+| `VEADK_SKILL_CREATOR_PROJECT_NAME` | — | 创建 Skill 时使用的项目名称。 |
+
+候选 Session 的 TTL 为 30 分钟，用户重新创建或离开任务时会立即清理。任务状态保存
+在 Sandbox 而非 Frontend 进程内存，因此 FaaS 请求命中不同实例时仍可继续轮询和下载。
+
+云上首次部署时，若没有指定 Tool ID，部署命令会使用部署 AK/SK 自动创建两个独立
+CodeEnv Tool。也可以分别使用 `--sandbox-chat-codex-tool-id` 和
+`--sandbox-skill-creator-tool-id` 指定已有 Tool。部署命令会通过 AgentKit 凭据托管
+将 Ark 凭据保存到 KMS，并把代理地址和可撤销票据绑定到两个 Tool；VeFaaS Function
+只接收 Tool ID，不接收模型凭据：
+
+```bash
+veadk studio deploy \
+  --user-pool-id <pool-id> \
+  --allowed-client-id <client-id> \
+  --vefaas-app-name <app-name>
+```
+
 ### 自定义 Studio 品牌
 
 使用 `--site-title` 设置不超过 6 个字符的系统名称，使用
@@ -179,6 +246,8 @@ Application。未指定 `--region` 和 `--project` 时，会在北京、上海
 更新不会改变 Application/Function ID、访问 URL、SSO、IAM、网关或已有环境
 变量。`--site-title` 和 `--site-logo` 仅在显式传入时覆盖，省略时保留云上品牌
 设置。可用 `--path` 指定源码目录，默认为当前目录；本地需安装 Node.js 与 npm。
+`--sandbox-chat-codex-tool-id` 和 `--sandbox-skill-creator-tool-id` 也只在显式传入时
+更新对应环境变量；未传的 Tool ID 保留云上已有值，不会被清空。
 
 部署注册回调地址时会保留 VeIdentity 登录页，并为该客户端开启跳过授权确认，
 避免用户登录后再次确认授权。