🚀 CodeBuddy Code v2.179.0 Release

📦 Version Info

Component	Version
CodeBuddy Code CLI	v2.179.0
Agent SDK JS	v0.3.296
Agent SDK Python	v0.3.295

✨ New Features

View and manage MCP servers directly in plugin management

Under the MCP filter in plugin management, servers configured in mcp.json are now listed with their connection status, transport, and source, and can be enabled/disabled, reconnected, deleted, or authenticated. Servers bundled with a plugin are read-only here.

Plugin kinds are now determined by the capabilities a plugin declares: plugins that really provide MCP servers are found even if their description never says "mcp", and plugins that offer both skills and MCP servers appear under both the Skills and MCP filters.

Live inner tool calls while the REPL runs

Long tasks no longer collapse into a single running line: in the collapsed view, inner calls are revealed one by one along the call tree, and the call in progress shows a spinner plus live elapsed time.

Goal supports attachments and longer conditions

• Files, images, resource references, and long-text context can be sent along with a goal; attachments survive pause and resume, and sensitive content still goes through the existing review.
• The goal condition limit rises from 4000 to 8000 characters.
• The Web UI goal editor supports uploading, pasting, and dragging attachments, plus preview and removal; a new draft is preserved while waiting to send, and switching sessions cancels the old request.

Switch tool modes while a session is idle

The terminal and the CLI's built-in Web UI can switch between standard, PTC, minimal, and creative modes within the same idle session without clearing history. The next turn uses the new calling guidance and tool declarations, while the original role, model, and permissions stay unchanged; switching is still blocked while running or waiting for approval.

🔧 Improvements

• Smoother channel message streams: Agent Home channels now render virtually, mounting only the message rows inside the viewport and keeping the latest 1000 messages in memory per channel, with older history loaded automatically when you scroll to the top. Scrolling and re-rendering a long channel no longer slow down as messages pile up, and merging bubbles, jumping to quotes, and answering cards keep working as before.
• Longer channel messages: the body limit rises from 2000 to 20000 characters, with a clear error instead of silent truncation; trimmed room history now reports how many characters were omitted.
• Automatic fallback for auxiliary models: when a request to your chosen auxiliary task model fails, it retries once on the original model, then switches to the default model for one more attempt, and later requests in the same session use the default model directly; failures no longer hang for a long time.
• No more repeated file reads when sending a message: automatic snapshots skip files already confirmed unchanged, so messages send noticeably faster after tracking large files.
• Mods API alignment: session, tool, command, config, model, and MCP APIs are available in headless runs (SDK / server mode), Mods can register tools, commands, and sub-agents within a session, and settings can be read per source with notifications supported.

🐛 Bug Fixes

• Session history is no longer deleted by mistake: interactive entry points no longer write the cleanup switch into the environment, so nested codebuddy processes started by tools, hooks, MCP, or plugins do not inherit it and delete sessions under the built-in 30-day rule; when the user's retention period cannot be read, the cleanup is skipped entirely instead of falling back to 30 days.
• Web UI status no longer jumps back to idle: while a live turn is still running, the channel, sidebar, and top status bar keep showing the live state instead of being overwritten by a stale snapshot.
• Web UI live messages stay connected: a 409 on the notification channel now releases the old connection and rebuilds it instead of retrying indefinitely; pages that only carry Bearer credentials can also establish the bus connection.
• Remote host preparation failures show what to fix: local asset problems, network problems, and authentication problems are reported separately; the prepare API returns 409 for deterministic failures instead of being retried as if the gateway were flaky.
• Sandboxed commands are no longer executed twice: commands that already succeeded do not trigger approval or run again; in automatic mode a blocked command keeps its first result and suggests explicitly requesting sandbox-free execution; clicking Stop after an approval no longer freezes the session.
• PTC questions no longer time out: a question left unanswered for ten minutes stays answerable when you return to the terminal, and is withdrawn correctly on cancel or stop; inner tool calls in PTC mode now also honor Hook allow/deny and argument rewrites.
• More accurate tool-mode context accounting: tool usage is estimated from actual model declarations and MCP tools are no longer double-counted; PTC and minimal modes get their dedicated REPL calling guidance back.
• No repeated image path injection in the second turn: image-bearing sessions no longer re-inject local image paths and break the model prefix cache.
• WeChat Pay card displays again: after the host takes over WeChat Pay, the payment card renders and query results flow back, and failures keep the error reason and distinguish merchant order from payment status to avoid duplicate orders.
• Agent Home channel collaboration fixes: system inbound messages are treated as read-only receipts and replaying them is explicitly rejected; after a cross-process dispatch the task board refreshes and shows the new task; sub-agent cards no longer all fall back to general-purpose.
• Agent Home colleagues use default permissions: creation, wake-up, and scheduled tasks now use default permissions and inherit the creator's currently effective model.
• Sub-agent tool allowlist: a sub-agent that explicitly declares tools: [] now runs with zero tools instead of inheriting every tool.
• Native ACP deferred MCP calls no longer require approval again: already authorized calls keep their original authorization and execution path.
• MCP log spam removed: when a tool schema cannot be converted losslessly it still runs as non-strict, but the log is now debug level.
• Cloud sandboxes no longer wait on the old config: the legacy Galileo configuration is no longer consumed, so the first request no longer waits up to 10 seconds for it.
• Scheduled task return-to-chat navigation: returning from an execution record to the chat now jumps to the triggering message.
• Terminal backspace works again: holding or rapidly pressing backspace deletes characters instead of inserting invisible control characters.