🚀 CodeBuddy Code v2.159.0 Release

📦 Version Information

Component	Version
CodeBuddy Code CLI	v2.159.0
Agent SDK JS	v0.3.267
Agent SDK Python	v0.3.266

✨ New Features

Grouped Skill List Display

The skill list is now grouped by the directory where each skill resides, and entries are simplified to "name + description" without exposing internal file paths. Individual descriptions can show more complete content. When the overall list exceeds the budget, descriptions are compressed first, and if necessary, entries are omitted from the end with a clear note of how many were omitted.

🔧 Improvements

• More observable ACP auto-retry: When an ACP connection automatically retries due to network issues, integrators can identify the retry reason directly from the session state. Prompt content can also be marked as "model-only", so it is no longer displayed or replayed as a user message.
• Livelier agent avatars: Avatars keep their original vivid colors, show gentle motion when idle, and use faster motion rhythms and status dots to distinguish working and needs-attention states, without overlaid glows or reduced opacity. The system's "Reduce motion" setting continues to be respected.
• Easier-to-distinguish avatar selection: Regular avatars have been expanded to 15 styles. New agents prefer styles that are currently less used and remember the selection. When duplicate avatars appear in rooms and channels, they are automatically reassigned so members can be told apart at a glance.
• Improved associated model routing: Improved lightweight / reasoning tier mapping for GPT and Gemini, and optimized DeepSeek V4 series routing in product configurations that offer new-generation lightweight models. Models lacking a mapping get a safe fallback, and routing for other models is unchanged.
• Fixed skill list budget: The overall character budget for the skill list is fixed at 8000 and is no longer affected by host environment variables, so the list length is consistent across environments.

🐛 Bug Fixes

• Web UI agent settings no longer revert: Fixed agent model and permission mode changes not persisting; opening the model dropdown or refreshing the page no longer reverts to user defaults. New agents inherit the currently effective configuration, older agents automatically get their configuration filled in, and embedded agent directories show the real working directory.
• Concurrent edits to the same file no longer hang: Fixed an issue where concurrently editing the same file in the same conversation turn left the session stuck at "Modifying" for a long time, maxing out a single CPU core and making it impossible to stop.
• More reliable file rollback: When a backup cannot be created, the modification is blocked immediately with an explanation and the file is left unchanged, avoiding missing pre-modification content during rollback. When a path points to a directory or special file, an accurate path error is returned, and empty files are no longer left in the backup directory.
• /rewind can roll back files in PTC / Code Mode: File edits in these modes now create history snapshots. The /rewind panel lists changed files with +N/-M line changes, and you can restore them via "Restore code".
• Auto mode permission decisions: Fixed forced-reasoning models being unable to complete permission decisions in Auto mode, and safe operations being wrongly denied in non-interactive Auto mode.
• More stable Web UI streaming rendering: The message stream no longer jumps up when tool execution ends, and the thinking process stays expanded after completion. Body text containing math formulas is no longer misidentified as a table and broken apart. Long answers are rendered incrementally by completed paragraphs, with body text flowing out at a steady pace with a cursor, instead of jumping out in chunks.
• ESC cancels on the first press: Fixed needing to press ESC twice to stop the current turn.
• Background service registration: Avoids writing invalid entries such as test runners into system services, which caused repeated startup failures, and limits the retry frequency when macOS background services fail to start, so error logs are no longer filled rapidly.
• Web UI panels and file links: Fixed two right-side panels possibly appearing at once when switching back to an agent conversation, and clicking local file links causing a black screen. file:// Markdown links now consistently open in the right-side workspace, and SVGs use the same preview mode as images.
• Default model selection: When no model is explicitly selected, the default no longer lands on a preview model. Instead, the product-marked default is preferred; if none exists, the first available model in the current environment's order is used. Behavior is unchanged when an agent declares a specific model itself.
• First configuration fetch failing after startup: Fixed the first remote configuration fetch being rejected by the server because the client identifier lacked the product brand name, so the available model list no longer degrades as a result.
• Conversation submission failing with multiple screenshots: Fixed conversations with multiple screenshots (about 10 MB or more) failing immediately after submission with Request body larger than maxBodyLength limit. When content truly exceeds the server limit, oversized images are automatically removed or earlier conversation history is truncated before answering, and the AI explains what was omitted.
• Long session stability: Limited the memory usage of trace data in long sessions, reducing the risk of the service process exiting abnormally due to continuously growing trace data.
• In-process Agent Harness: Fixed Bash file access being wrongly denied when using sandbox-cli.