🚀 CodeBuddy Code v2.127.0 Release

📦 Version Information

Component	Version
CodeBuddy Code CLI	v2.127.0
Agent SDK JS	v0.3.229
Agent SDK Python	v0.3.228

✨ New Features

Background Tasks (stream-json Long-Connection Mode)

In the --input-format stream-json --output-format stream-json long-connection mode, background commands (run_in_background: true) are now supported, allowing tasks to continue running across conversation turns:

• Background task lifecycle is pushed to the output stream as independent system events: task_started on launch, task_progress during execution (with embedded token / tool call count / elapsed time statistics), task_updated on state changes, and task_notification on completion/failure/stop
• Each event carries a unique task_id (to distinguish concurrent tasks) and tool_use_id (to associate with the call that initiated the task). Tasks that complete between turns are proactively pushed back without requiring additional input
• Background Agent subtasks are also connected to this event stream, carrying cumulative usage statistics
• The SDK's query() single-shot usage automatically disables background tasks to prevent tasks from being killed when the subprocess exits early

Subagent Capability Enhancements

• Added the codebuddy agents [--json] command; the list now includes each agent's memory scope and toolset
• frontmatter now supports more configurations: maxTurns / background / initialPrompt / effort / disallowedTools / memory (agent-level persistent memory), with support for hot-reloading subdirectory configurations
• Subagent output anti-injection scanning (proprietary rules); when the turn limit is reached, existing output is returned as a normal result
• Tool whitelist now supports mcp__server / mcp__server__* server-level patterns

Web UI Context Usage Ring and Per-Model Thinking Configuration

• The input box now displays the current session's context usage in real time and supports one-click session compression
• The model dropdown allows per-model thinking level configuration (including thinking toggle and level list); the default level for custom models can be persisted

🔧 Improvements

• Minimal permissions for AI-created agents: When using AI to create agents, a minimal necessary toolset is generated based on the agent's responsibilities (e.g., code review agents are granted only read-only tools), no longer inheriting all tools by default
• Model dropdown visual redesign: More compact narrow cards, right-aligned numeric columns, checkmark for selected state, and edit entry appearing on hover
• TUI horizontal menu: In addition to retaining Tab switching, left/right arrow keys now support cyclic switching between tabs and model configuration scopes, avoiding accidentally entering Team member selection on non-input panels
• Terminal output: Removed redundant diagnostic logs during file encoding detection to avoid interfering with normal terminal output

🐛 Bug Fixes

• Stop task not working: Fixed an issue where, when a sub-agent was batch-generating images (via Workflow background orchestration), clicking "Stop" on the desktop / mini-program client did not terminate the task — it kept running
• Stop/SubagentStop hook continue:false not working: Fixed a regression where continue:false was misinterpreted as "should stop" — the opposite of the documented semantics. It now correctly prevents stopping and lets the Agent continue working
• Remote control enabled by default: When the remote control master switch is not configured, it is no longer misidentified as disabled. The entry is only hidden when the product configuration explicitly turns it off
• SDK auto-compression long-term skip: After tool approval succeeds, the pending approval state is now correctly cleaned up, preventing auto-compression from being long-term skipped and causing context to exceed the limit
• MCP tool concurrency serialized: Fixed an issue where multiple tool calls from the same connector were serialized, so long-running tools no longer slow down other calls from the same connector
• MCP transport type alias compatibility: The CLI now recognizes streamable-http / streamableHttp / streamable_http variants in mcp.json and automatically normalizes them to http, resolving the issue where the same configuration worked in the IDE but the CLI reported no valid transport type
• MCP first-round tools missing: Optimized MCP config scan timing and first-round wait strategy, preventing config changes from being swallowed during plugin refresh and avoiding missing or incomplete tools in the first conversation turn
• Memory file editing: The /memory command now prioritizes the editor configured via $VISUAL or $EDITOR, falling back to the system default application if opening fails