🚀 CodeBuddy Code v2.143.0 Release

📦 Version Information

Component	Version
CodeBuddy Code CLI	v2.143.0
Agent SDK JS	v0.3.250
Agent SDK Python	v0.3.249

✨ New Features

Agent-to-Agent Collaboration (A2A)

• Call other Agents: Agent CLI can now act as a standard A2A (Agent2Agent) client to discover and invoke other authorized Agents. Delegate tasks with @Agent or natural language, invoke multiple Agents in parallel, and see intermediate status updates in real time without repeatedly polling manually.
• Expose services to external callers: A hidden --a2a mode has been added, running as a persistent stdio service using the standard A2A protocol for programmatic callers such as SDK and CloudAgent.
• Send attachments: A2ASendMessage can now attach files by URL and send them to a remote Agent alongside text messages.

Enhanced Background Tasks

• View output in real time: Output from background commands is now streamed to clients in real time. In the Web UI, expand a task in the task panel or open the right-side workspace to view it full-screen in a read-only terminal. Multibyte characters such as Chinese text are reconstructed correctly even when split across chunks.
• Cancel tasks proactively: ACP clients and the Web UI can now cancel running background tasks, including grandchildren spawned with nohup, instead of waiting for them to finish on their own.
• Clearly explain why a task ended: User cancellation, termination due to output limits, session interruption, sandbox blocking, and other outcomes now have distinct messages instead of all appearing as “Ended.”
• PowerShell support: Commands moved to the background on Windows now appear in the client's background task panel with the correct running status.

Streamlined CLI Build for WorkBuddy

Introducing codebuddy-lite-wb: it goes beyond the headless build by removing modules that WorkBuddy scenarios do not use, resulting in faster cold starts and lower resident memory usage. If the build artifact is unavailable, startup falls back automatically without affecting any packaging variant.

Per-Skill Description Limit Overrides

A new CODEBUDDY_SKILL_DESC_MAX_OVERRIDES environment variable allows the description length limit to be raised for specific Skills, making primary Skills with longer descriptions—whose key details might otherwise be lost under the default limit—easier to match correctly.

🔧 Improvements

• A2A protocol alignment and experience: The list of delegatable Agents is rendered in real time from discovery results, and Agent Cards are retrieved by exact name. Canceling a remote task now actually notifies the remote endpoint to stop, preventing it from continuing to consume compute resources.
• Memory optimization: Image-processing and WASM dependencies are now loaded on demand at runtime, further reducing memory use for idle processes and during cold starts.
• Refined Web UI conversation flow: The hierarchy and pacing of reasoning, tool steps, and final answers are now consistent for a cleaner interface. Also fixed the “Thinking” status being attached to the wrong turn.
• Model guidance after task completion: After a background task ends, the model now suggests different next actions based on the completion reason, so canceled tasks are not restarted by mistake.
• Higher Prompt Cache hit rate: Tool definitions are grouped by stability to speed up conversation responses.

🐛 Bug Fixes

• Canceled input no longer persists: Proactively canceled, incomplete conversations no longer pollute the context of the next prompt.
• Custom-model “Extra High” setting is no longer downgraded: Custom models that support the extra-high (xhigh) reasoning effort now retain that setting correctly instead of being downgraded to high.
• Accurate empty-stream error messages: When an upstream model returns an empty response, the message now says “The model service returned no message” instead of incorrectly directing troubleshooting toward the local network.
• Self-recovery from stuck conversations: Fixed an issue where some sessions could spin internally after ending and prevent new messages from being sent.
• Complete Web UI streaming text: Repeated consecutive fragments in model output, such as 1000, are no longer dropped and rendered as 100.
• Narrow terminal rendering: Wrapping very long lines no longer causes cursor misalignment, stale characters, or display corruption.
• macOS input methods: The Pinyin candidate window now follows the input cursor instead of remaining fixed in a corner of the window.
• Complete Enterprise model list: The Web UI model dropdown now matches /model instead of showing only a small set of default models.
• Startup memory reduced by about 35MB: Sandbox-auditing WASM is now loaded on demand, so it no longer remains resident during cold starts in standard scenarios.