By devasher · Edited by Nominiclaw
This beta release introduces significant performance optimizations for model listing, enhanced realtime voice control, and critical security hardening for sandboxed skill execution.
This release focuses on drastic performance improvements for the Gateway, expanded capabilities for realtime voice interactions, and a series of critical stability and security fixes across the core agent runtime and plugin SDK.
/models endpoint has been optimized from ~20 seconds down to ~5ms (a 4,100× speedup) by pre-warming provider auth state at startup, preventing redundant discovery sweeps on every call.USER.md and SOUL.md files.subagentBootstrapFiles configuration, allowing users to limit the context passed to delegated workers (e.g., only AGENTS.md and TOOLS.md) to save tokens and maintain least-privilege context (#85283).tools arrays and clamping max_completion_tokens to the remaining effective context window (#85835, #85889).maxHoldMs during lock acquisition (#85764).web_search and introduced new Grok model aliases (#85182).secrets.prepare timeline spans and bounded skill usage metrics allow operators to diagnose startup latency and tool execution without exposing sensitive data (#83019, #80370).exec hook parameters to mirror shell-style command fields, making it easier to write consistent security policies across different execution surfaces (#83483).maxHoldMs (default 5 minutes) may now be reclaimed by other processes to prevent gateway freezes.openclaw migrate plan codex command to preview and execute the transition.context-1m-2025-08-07 beta header is no longer injected and will be silently stripped from existing configurations to avoid stale header errors (#45613).