Sandbox Expansion and New Integration Releases

Thursday, April 9, 2026 · v1.5.0

This release centers on three themes: a significant expansion of Sandbox execution capabilities, new integration channels with Custom Sub-Agents and WeChat Clawbot, and platform hardening around Castrel Proxy permission controls and automation quotas. A range of long-standing UX issues have also been addressed in this update.

New Features

Custom Sub-Agents: You can now create and manage your own sub-agents under "Contextual Sub-Agents." Once defined, they're instantly accessible via @ mention in any conversation, making AI collaboration more targeted and composable without touching the built-in agents.
WeChat Clawbot Integration: Castrel now connects with WeChat Clawbot. After a quick QR-code pairing, you can interact with your AI agents directly through WeChat, bringing automation and observability workflows into one of the most widely used communication platforms.
Sandbox — Major Capability Upgrade: The Sandbox execution environment has been substantially expanded in this release. Files generated by AI can now be previewed inline (up to 5 MB) or downloaded directly from the conversation (up to 100 MB). A bundled Chromium instance enables real browser rendering, screenshots, and DOM inspection without any additional setup. The built-in Skill Creator lets you initialize, validate, package, and register custom Skills entirely within a conversation. The runtime user and working directory are now standardized, making file I/O and script execution paths consistent and predictable. Finally, when the underlying environment template changes, Sandboxes are automatically rebuilt, and existing files and script artifacts are preserved throughout.
Qingfan Cloud — Process Lookup by Name: A new tool allows AI agents to retrieve workflow details from Qingfan Cloud using a process name, completing the last mile of the automated ticketing pipeline.

Improvements

Automation Quota Management: Automation task limits are now enforced per plan tier (Free: 1 task, Pro: 10 tasks, Enterprise: unlimited). When the limit is reached, users see a clear prompt with an upgrade path rather than a silent failure.
Castrel Proxy — Granular Permission Controls: Proxy nodes now support fine-grained access policies, including command execution allowlists and filesystem permissions (read-only / read-write / full edit). These policies are surfaced directly in the AI context at runtime, so agent behavior always stays within your defined boundaries.
Alert Analysis Reports — Chart Rendering: Charts in troubleshooting and alert analysis reports now render correctly. Previously, only raw tool call data was displayed; diagnostic reports are now fully visual as intended.
Built-in Skills & MCP — Internationalization: Platform-bundled Skills and MCP tools now ship with proper internationalization support. Environments running in non-Chinese locales will automatically initialize with English content.
Usage & Billing — Data Accuracy: Several inconsistencies in usage tracking have been corrected: personal usage no longer shows spurious zeros; per-user and global totals are now reconciled; date-based filtering returns accurate results; and the user filter pill now clears correctly after a global search is confirmed.

Fixes

@ Mention Panel: The Infra category was missing from the mention panel on the chat home page. Additionally, the mention dropdown was being obscured when triggered from within an active conversation thread. Both issues are resolved.
Sidebar — Recent Sessions: Navigating away from and back to the chat page caused the recent sessions list to re-fetch unnecessarily, producing a repeated loading state. The list now caches correctly across navigation.
Connector / Proxy Selection State: Connector and proxy node selections made in the chat input were being reset whenever users navigated to another page. Selections are now persisted in the browser session and restored on return.
Feishu @ Message Display: Fixed a sync issue where @ mention markers from Feishu group chats were displayed incorrectly in Castrel conversations.
Workspace Invite Links: Clicking a workspace invite link was landing users in their previously visited workspace rather than the intended one. Invite links now route correctly to the target workspace.
Shared Links — Guest Access: Unauthenticated users visiting a shared conversation link were being redirected to the login page when interacting with tool outputs. Public share content is now fully accessible without requiring a login.
Visual Consistency: A number of styling inconsistencies across the UI have been corrected, bringing the interface in line with the overall design language.