🚀 CodeBuddy Code v2.164.0 Release

📦 Version Information

Component	Version
CodeBuddy Code CLI	v2.164.0
Agent SDK JS	v0.3.276
Agent SDK Python	v0.3.275

✨ New Features

Web UI Unified on the AI Colleague Directory

The Web UI no longer offers switching between solo and team modes and now uses the AI colleague directory throughout. Letting the Agent create colleagues, use channels, and wake up members on its own is now controlled by separate toggles on the Settings page for centralized management.

When the same session switches back and forth between the terminal and the embedded Web UI, the tools for Agent self-organized teams only appear in turns sent from the Web UI. Continuing the conversation back in the terminal no longer inherits these tools, and the same applies after closing the browser. Capability boundaries for the TUI, WorkBuddy, IDE, and unattended entry points remain unchanged.

Mods Enhancements

• New Svg and Client elements: Svg is displayed as an isolated image (interactive mode uses a script-free sandboxed frame). The surface module of Client runs in a sandbox and supports local state, a frame clock, and pointer and keyboard input, with interaction results passed back to the plugin.
• New ui.resolve hook: Plugins can rewrite or remove elements for other plugins at load time, making it easy to compose and override existing interfaces.
• classic.* hook results take effect: classic.PreToolUse can rewrite tool parameters before execution, or deny the call or require confirmation. Additional context and tool output rewrites from hooks such as classic.PostToolUse are passed back to the host. Mods still cannot allow calls the host has already denied, nor skip permission confirmation.
• Run Claude Code sample mods directly: Use --plugin-dir to run sample mods such as Blast Radius, Replay Theater, and Token Weather. Mods can draw strips above the prompt, open panels via buttons, read actual context usage, and hold back high-risk shell commands.

Web UI Font Size Levels

The "Appearance" section of the Settings page adds UI font size and conversation font size levels (Small / Standard / Large / Extra Large). The UI font size adjusts the overall text size, while the conversation font size further adjusts the conversation stream on top of it. Settings are remembered and synced to other open pages.

🔧 Improvements

• Reduced session resume traffic: During stream-json history replay, the length of tool call arguments, tool results, and reasoning text is limited, reducing transfer volume when resuming cloud sessions while keeping live output and model context complete.
• Enhanced Agent SDK regression coverage: Added end-to-end regression tests in JavaScript and Python for the Agent SDK, covering authentication, cancellation, session lifecycle, tool hooks and permissions, MCP integration, subagents, and long-running task tools, boosting release confidence.

🐛 Bug Fixes

• Slimmer session file checkpoints: New files are recorded only once when written, and rewriting already-tracked files no longer appends duplicate records. Session checkpoint data for batch file writes is reduced by about half (a session creating 300 new files drops from about 7.3MB to about 3.6MB). The record format is unchanged, so rewind, file change display, and opening sessions in older versions are unaffected.
• More thorough ESC cancellation: Fixed occasional cases in the interactive TUI where pressing ESC failed to fully abort and left the session state stuck. After cancellation, runtime resources and session state now converge correctly.
• Plugin hooks not running on the first turn: Fixed the UserPromptSubmit hook of installed plugins possibly being skipped on the first conversation turn. Plugins newly installed within the current process also trigger correctly on subsequent turns.
• Subagents losing task content: Fixed default-type subagents losing task content and having requests rejected in the Web interface, IDE integration, and -p non-interactive mode.
• Subagent instructions and return values in Code Mode: Subagents no longer receive descriptions of tools they cannot use. The return type hint for subagent Agent() is more accurate, preventing task handles from being treated as strings and dispatched repeatedly.
• Retrievable truncated results in Code Mode: When a script's return value exceeds the limit, the full content is kept in the sandbox, so the model can filter and summarize it in the next call and return a conclusion without rerunning the entire logic.
• Complete Web UI history replay: Fixed the initial user question disappearing after opening, refreshing, or switching to a long session. Earlier history is now preserved, and only the overlapping latest tail is updated.
• Web UI subagent history restoration: Fixed subagent details not being recognized or expandable after refreshing or reopening a session.
• Web UI agent configuration isolation: The permission mode, model, and engine selected when creating a new agent no longer write back to the host user's defaults, and agents that inherit defaults keep following the global settings. Also fixed the first message not being echoed after creating an agent. In addition, "Full Access" is now restricted to the local Web UI only; calls via remote, Tunnel, LAN, or proxy forwarding are rejected.
• Web UI new agent permission inheritance: Fixed new agents not inheriting the user-level default permission mode.
• Closable Web UI image preview: The enlarged image preview adds a close button in the top-right corner and supports closing with the Esc key or by clicking the overlay, with focus returning to its original position after closing.
• Brand images under CNB subpaths: Fixed brand images on the welcome and login pages failing to load when accessed via a CNB instance subpath.
• Errors reading images with injection protection: With "Tool content injection protection" enabled, the request following an image read no longer fails, and image results are still wrapped as untrusted content before being sent to the model.
• Backups for document and Excel subagents: Document script edits and Excel subagent autosaves are now integrated with session file backups, preventing the original file content needed for rewind from being missed on save.
• WeChat remote control reconnection: After reconnection triggered by waking from sleep or network recovery, multiple concurrent message polling loops no longer linger, preventing logs from multiplying during network outages and messages from being fetched repeatedly after reconnecting.
• Pinyin candidate window in Windows terminals: In the VSCode terminal on Windows, the candidate window is no longer pinned to the bottom-left corner and now follows the input cursor position.