v1.8.0-beta.3

What's Changed

New Features

• Split Code Executor into Local and E2B Cloud tools — The single "Code Executor" tool has been separated into two independent tools. The local tool shows CPU/memory/network config; the E2B Cloud tool only requires an API key. E2B errors are now surfaced explicitly instead of silently falling back to local execution.
• MCP Server credential management — New "Edit Server" UI and PUT /tools/mcp-server API endpoint for bulk-updating MCP server URLs and API keys across all tools sharing the same server.
• Feishu Wiki document creation — feishu_doc_create now supports creating documents directly inside Wiki knowledge bases, with automatic detection of Wiki node tokens.
• Feishu permission JSON UI redesign — Two-tier segmented control (Basic / Full) with i18n support for Feishu app permission configuration.
• Live Preview auto-sizing — AgentBay Live Preview panel now auto-sizes to 50% of the chat container width.

Bug Fixes

• Plaintext SMTP relay support — STARTTLS is now auto-negotiated based on server ESMTP capabilities instead of being forced on port 25/587. AUTH is skipped for unauthenticated IP-whitelisted internal relays. Password is no longer a required field in email configuration.
• Unified context window size — Introduced DEFAULT_CONTEXT_WINDOW_SIZE = 100 constant and unified all 9 communication channels (WebSocket, Feishu, Discord, WeCom, DingTalk, Teams, Slack) to use consistent fallback values.
• LLM stream retry — Added httpx.RemoteProtocolError to the stream retry logic to handle upstream connection resets.
• Tool config double-encryption — Fixed a bug where already-encrypted sensitive config fields were encrypted again on save.
• Loguru format collision — Replaced logger.error(..., exc_info=True) with logger.exception(...) across all channel handlers to prevent crashes when error messages contain special characters.
• WeCom message handler — Fixed NameError (agent vs agent_obj) and migrated user creation to channel_user_service to avoid AssociationProxy errors.
• Duplicate tool definition — Removed send_channel_message from _ALWAYS_INCLUDE_CORE to prevent "Tool names must be unique" LLM errors.
• AgentBay connection test — Fixed test image name (linux_latest) and api_key lookup in global tool config fallback.
• FastAPI route ordering — Reordered /tools/mcp-server/bulk before /tools/{tool_id} to prevent 422 validation errors on older FastAPI versions.
• Other fixes — LLM model temperature persistence, org_admin access to GitHub/ClawHub tokens, MCP tool import tenant scoping.

UI / i18n

• Context Window Size terminology — Corrected misleading "Max Rounds" / "Context Rounds" labels to industry-standard "Context Window Size" with accurate descriptions.
• MCP Server group header — Displays hostname instead of full URL for cleaner display.

Upgrade Notes

This is a drop-in upgrade from v1.8.0-beta.2. No breaking changes.

• No database migrations required
• No new dependencies
• No environment variable changes
• The new execute_code_e2b tool will be automatically created by the tool seeder on startup. It is not a default tool — agents will not have it unless explicitly added.
• The existing execute_code tool's config schema will be auto-synced (the sandbox type dropdown is removed since it's now always "subprocess").

Docker Deployment

:::bash
git pull origin main
docker compose down && docker compose up -d --build

Source Deployment

:::bash
git pull origin main
# Backend
pip install -r backend/requirements.txt  # no changes expected, but safe to run
# Frontend (pre-built dist.zip is included)
cd frontend && unzip -o dist.zip -d dist/
# Restart services