Skip to content
CLI

🚀 CodeBuddy Code v2.134.0 Release

📦 Version Information

ComponentVersion
CodeBuddy Code CLIv2.134.0
Agent SDK JSv0.3.240
Agent SDK Pythonv0.3.239

✨ New Features

Malicious Domain Protection

Added domain hash matching to Sandbox network requests and WebFetch. It reuses the Security Center's network allowlist, automatically blocks access to known malicious domains, and records security audit events across macOS, Windows, and Linux.

Managed Sandbox Security Policy Enforcement

The command sandbox now reads and enforces enterprise-managed security policies at session runtime, ensuring that enterprise file protection and command control rules cannot be bypassed in the local execution path. Program blocklists can also be synchronized from the Security Center to the sandbox, preventing matched programs and their child processes from starting.

Password Authentication Enabled by Default for --serve

codebuddy --serve now enables password authentication by default. At startup, it generates a random password and prints a clickable login link; after following the link, no additional password entry is required for 30 days. Previously, local access was unauthenticated by default, allowing any process on the same machine or malicious web page unrestricted access to APIs for files, terminals, environment variables, and more. This access has now been tightened. For unauthenticated access in CI or scripts, explicitly disable authentication with --auth none.

Upgrade note: Scripted calls to /api/v1/* must now include credentials with -H "Authorization: Bearer $PASSWORD".

Adaptive Retries for Model Rate Limits

Added exponential-backoff retries for model request failures that occur before generation starts (429 / 5xx / 408 / 409), so gateway restarts and temporary overloads no longer interrupt the entire task immediately. Quota-exhaustion errors switch directly to a fallback model to avoid unnecessary waiting. The new CODEBUDDY_MAX_RETRIES environment variable controls the retry count, while CODEBUDDY_RETRY_WATCHDOG=1 enables unlimited retries for unattended scenarios.

Session Rewind

Added the ability to restore workspace files and conversation history to their state before a selected historical message, supporting workflows such as "undo to this message." New command-line options include --rewind-files, --resume-session-at, and --dry-run, along with a runtime control protocol for persistent stream-json mode.

Artifact Tool Upgrade

The ShareLink tool has been renamed to Artifact for publishing local HTML/Markdown files as shareable public links, and ShareLinkUnpublish has been renamed to ArtifactControl. A new existingShareLink parameter supports updating the same link in place; legacy permission rules and feature switches remain compatible.

Initial MCP Connection Status in ACP

ACP clients (IDE / Web UI) and the TUI now display real-time MCP server connection progress (Starting MCP servers (N/M)) during the first prompt of a new session, replacing an uninformative wait.

Monitor Tool Supports WebSocket Event Sources

The Monitor tool can now listen directly to ws:// / wss:// addresses. Every server-pushed message wakes the model as an event, eliminating the need to write a WebSocket client script. Built-in protections block intranet, loopback, and cloud metadata addresses and validate domains against allowlists and blocklists.

Web UI Background Session Management

Added a background sessions page and sidebar with support for dispatching, stopping, restarting, and deleting sessions. Full conversations can be embedded directly in the sidebar, including follow-up messages, cancellation, tool calls, and MCP interfaces. When dispatching, you can select the model, thinking effort, initial permissions, and working directory.

Agent View and Background Task Experience Improvements

  • The Agent View input now supports the full set of slash commands, and ESC consistently returns to the most recently interactive session
  • The Web UI conversation page now includes a background task panel showing runtime duration and completion status, which persist across page refreshes
  • Background session context inheritance now includes the complete history, preventing information loss when handing off long-running tasks

🔧 Improvements

  • Faster startup in persistent stdio mode: Removed several redundant product configuration synchronization paths, reducing time to the first prompt from about 3 seconds to milliseconds
  • Thinking effort fixes: Selecting max thinking effort is no longer incorrectly downgraded to xhigh; custom models automatically fall back to high for unsupported levels
  • Improved /model scenario switching: Model switching for lite and reasoning scenarios has been split into the dedicated /model:lite and /model:reasoning commands, providing direct access through a single-level list
  • Enhanced enterprise endpoint diagnostics: Added logs for resolution results and their sources to help diagnose cases where configuration was written but did not take effect

🐛 Bug Fixes

  • Cold-start stability: Fixed several intermittent crashes (default agent is not registered) caused by Agent template construction not waiting for product configuration synchronization to finish
  • Web UI session switching: Fixed ADMISSION_FAILED caused by duplicate connection requests, stalls during concurrent multi-tab access, and race conditions that could duplicate or lose content when switching sessions
  • Message queue: Fixed issues where "Send now" waited for the current task to finish, queue items spun indefinitely after being clicked, and slash command messages were lost
  • Background task notifications: Fixed completed tasks remaining displayed as "running"
  • Context and compaction: Fixed incorrect auto-compaction threshold calculations after switching models, 400 limit errors in long sessions, and compaction counters leaking across requests and permanently disabling compaction for a session
  • Login experience: Fixed robustness issues in iOA login and an endless loading state during account login when accessing the daemon remotely
  • Terminal display: Fixed terminal character-set corruption that caused garbled TUI rendering, with automatic reset support
  • Cross-platform fixes: Fixed intermittent session deletion failures on Windows, flashing Web UI terminal windows on Windows, Web UI hangs in large Linux repositories caused by exhausted file watchers, and ineffective malicious website blocking on Linux
  • Model call robustness: Fixed repeated idle loops after the model produced malformed tool arguments, truncated sub-agent response recovery being incorrectly reported as successful, and incomplete session context after security review was triggered
  • Plan mode: Fixed missing content when reviewing plans, "Revise plan" incorrectly exiting plan mode, and team members being unable to execute after a collaborative plan was approved

📝 Documentation Updates

  • Added a background sessions section to the HTTP API documentation
  • Added background session usage instructions to the Web UI documentation