🚀 CodeBuddy Code v2.163.0 Release

📦 Version Information

Component	Version
CodeBuddy Code CLI	v2.163.0
Agent SDK JS	v0.3.275
Agent SDK Python	v0.3.274

✨ New Features

Refreshed Web UI Management Pages

Pages such as Tasks, Plugins, Remote Control, Settings, Shortcuts, Docs, Processes, Stats, Traces, and Logs now share a unified style for page headers, cards, buttons, tab switching, and toggles. Skeleton screens are shown while loading, and empty or error states offer next steps such as "Retry", "Browse Marketplace", and "Clear Filters". Subtle entrance and feedback animations have also been added (automatically disabled when the system's "Reduce motion" setting is on).

• Easier-to-use Shortcuts page: Shortcuts are displayed in groups, with key combinations shown as individual keycaps. When recording a new shortcut, keys are echoed in real time and conflicts are flagged; "Reset All" requires a second confirmation.
• More reliable QR code binding: QR codes for WeChat and WeCom binding are now generated locally instead of relying on an external service. If fetching fails, retry and manual entry options are provided, along with a clear list of the commands and configuration file locations for manual setup.
• Unified conversation font size: Body text, headings, code, tables, tool names, and process information in conversations are all unified at 13px for a more compact and consistent reading experience.

Web UI Generative UI

• Mods generative UI: Assistant messages can now be rendered by a mod. A mod can inject custom output rules into the model and then render tags in the reply as interactive components such as cards, buttons, and banners. Button clicks are routed back to the mod for execution, and already-displayed cards refresh to the latest state on re-render. A sample "order food → place order → cancel order" mod is included.
• General-purpose generative UI sample mod: The model outputs genui JSON code blocks in its reply, and the Web UI renders them as cards, tables, progress indicators, forms, and buttons. Clicking a button sends the corresponding content as the next message; each card can only be answered once, and the selected result remains visible after a page refresh.
• Editable inputs and dropdowns: Input boxes and dropdowns rendered by mods can now be filled in and selected directly, with values passed back to the mod.

Automatic Agent Naming in the Web UI

When a new agent is created without being renamed, it first shows "New Agent" as a placeholder name. Once a title is generated for its first conversation, that title is automatically used as the agent's name, and this happens only once. Manually assigned names are never overwritten, and it does not matter whether the page is open when the title is generated.

Persistent Privacy Marking for WeChat / WeCom Chat Records

Once the chat record sharing marker is matched, subsequent model requests from the current session and its derived Agents remain marked as private data, and this state persists after the standalone session is resumed.

New Toggle for Sensitive Data Protection

A new environment variable CODEBUDDY_DISABLE_SENSITIVE_PROTECTION=1 disables sensitive data protection for the current process. It has the same effect as the toggle in the settings file, but does not modify the shared settings file; values delivered by enterprise managed policies still take precedence.

🔧 Improvements

• Mods (function hooks) enabled by default: Mods specified via --plugin-dir and mods in installed plugins now load without setting an environment variable. To disable them, set CBC_ENABLE_FUNCTION_HOOKS=0.
• Network requests and process execution for Mods: Mods can now issue http/https requests through the host and run local commands by argv (not through a shell, with a default 30-second timeout and a maximum of 10 minutes). Administrators can disable both capabilities.
• Lower context usage: The Bash and PowerShell tool descriptions have been trimmed to remove duplicated content and rarely used details, reducing the resident per-turn context by about 51% and 33% respectively. It is also now made explicit that once a command is rejected by the user, permission rules, or security policies, the model should not retry with a different command or tool.
• More accurate PowerShell syntax hints: On Windows, the CLI detects whether the current shell is Windows PowerShell 5.1 or PowerShell 7+ and only suggests syntax available in that version. PowerShell 7 users are no longer told they cannot use &&.
• Smoother content search: Unbalanced brackets in regular expressions, such as foo( or struct X {, no longer cause an immediate failure. Instead, the search is retried once with the brackets treated as literal text and the model is informed, saving an extra round trip.
• Responses API support for custom models: Models in models.json can opt into the OpenAI Responses API, with encrypted reasoning content replayed as-is in multi-turn conversations. Requests within the same task can be pinned to the same upstream account to improve cache hits. Rate limits trigger automatic backoff and retries, and when quota is exhausted, requests are handed over to the fallback model.
• Faster historical image processing: Image preprocessing reuses validation results for identical inputs, so historical images are no longer rescanned on every request (a 3 MB PNG goes from about 32ms to 3ms).

🐛 Bug Fixes

• Subagent dispatch in Code Mode: Subagents now run according to their own preset role and task boundaries, instead of starting with the main session's identity and drifting off task until timing out. Results returned by synchronously dispatched subagents can be parsed directly as data, and clear correction hints are given when background task queries are misused.
• /clear preserves session settings: /clear only clears the conversation and execution state, preserving model selection, thinking and context window settings, main Agent mode, workspace and tool configuration, and security review settings.
• More accurate retry messages: When a response times out and automatic retries are exhausted, "Retrying automatically" is no longer shown; instead, users are prompted to check the network, retry, or switch models. Empty responses now show a clear "The model did not return a valid response" message.
• Correct session rewind selection: Fixed rewind targeting the first prompt after asynchronous updates. Only user prompts are used as rewind targets, and the selection is preserved after reloading or sending an edited prompt.
• Images in messages inserted during task execution: Messages inserted during execution keep their attached images, so the model can see the image content directly. Inserted messages containing only images are also delivered to the current turn.
• More reliable checkpoint rewind: Fixed first-turn rewind previews failing to locate their target, available alternative versions not actually being restored, and the session still being truncated after file restoration failed.
• Backups for image and video outputs: With file backup enabled, the original file is backed up before an existing output is overwritten. Also fixed outputs from cloud image and video services failing to save.
• Historical images displayed after resuming a session: Fixed images in historical messages not displaying after a session is resumed. Original images and file names are preserved, and a text hint is shown when an image is unavailable.
• Automatic loading of the mobile file explorer: On mobile, entering the file explorer automatically loads and watches the current workspace directory, without needing to tap "Open Folder" manually.
• AI code adoption rate statistics: Fixed code written by the CLI not being counted toward the adoption rate for SaaS enterprise accounts with adoption rate statistics enabled. The behavior is now consistent with the IDE.
• Complete AgentLens trace reporting: Fixed telemetry being re-initialized mid-conversation under self-hosted reporting, which caused entire categories of Agent and tool call spans to be dropped.
• Sandbox tracing lifecycle: Sessions with sandbox tracing enabled now sync the collection toggle and save directory at the start of each turn, and flush trace records after the turn ends and backup processing completes. Collection failures do not affect the conversation or file backups.
• Web UI page details: Fixed issues including each log line being split into multiple columns with text squeezed onto multiple lines, the view being pulled back to the bottom while browsing historical logs, labels and bars in the trace timeline becoming misaligned when the window width changed, "Retry" not working when a single marketplace was selected in the plugin marketplace, and extra gray borders on some buttons in the dark theme.