🚀 CodeBuddy Code v2.121.3 Release
📦 Version Information
| Component | Version |
|---|---|
| CodeBuddy Code CLI | v2.121.3 |
| Agent SDK JS | v0.3.216 |
| Agent SDK Python | v0.3.215 |
🔧 Improvements
- stdio initialize Speedup: The SDK
initializehandshake is no longer blocked by slow MCP services when custom system prompts / Agent overrides are provided. Override injection is now performed in the background, and the handshake returns immediately. The first prompt still waits for overrides to take effect before executing. In practice, handshake latency has been reduced from hundreds of milliseconds to millisecond level. - Output Style Lazy Rendering: Built-in output style template content is now rendered on demand. Loading the style list no longer unconditionally renders all templates; rendering and caching occur only when a style is actually applied, speeding up list loading.
🐛 Bug Fixes
- First Prompt Missing Some MCP Tools: Fixed an issue in prewarm mode where the available tools list for the first prompt of a new session occasionally lacked user-level MCP tools that connected late, requiring a second prompt for them to appear. A tool index validation has been added to the runtime path; when late-connecting MCP services become ready, the Agent is rebuilt before execution, with no impact on pre-request performance.
- Sub-Agent Delegation Crash: Fixed a regression where sub-agent delegation failed. The telemetry interceptor crashed due to missing null checks when input was empty. Defensive null checks have been added at both the source and interceptor levels.
- TUI Keyboard Input Crash: Fixed an issue where pasting text or inputting specific escape sequences (such as Ctrl+function key sequences that can be embedded in large pastes) caused key name resolution to read a null value, triggering an uncaught exception that crashed the CLI process. Input boundaries have been uniformly normalized for null values.
- Reverted Deferred Tool Retry Parameter Hint Approach: Reverted the previous approach of injecting active tool parameter structure hints to guide the model in self-correction. This was found to be ineffective for some models after verification. The issue is now resolved via an upstream direct tool connection approach, avoiding the overhead of ineffective hints.