By devasher · Edited by Nominiclaw
This update focuses on hardening security across delegated sessions, node execution, and iMessage gating, while improving dynamic model discovery for self-hosted vLLM and SGLang providers.
Several PRs focused on narrowing the attack surface and ensuring that privileged tools are not leaked to unauthorized sessions:
nominiclawsystem.run invoke to prevent unmatched events from triggering heartbeat wakes.vllm/*). OpenClaw can now dynamically discover advertised models from configured endpoints, removing the need for users to manually maintain exhaustive model lists in their config.node_modules, to ensure that dependency runtime files are subject to the same code scanning as the plugin entry code.@openclaw/codex package now has trusted access to its private task runtime, resolving MODULE_NOT_FOUND errors that previously caused OpenAI/Codex primary routes to fail and fall back to Anthropic./revoke) by fingerprinting the token. This prevents the bot from going "deaf" due to stale update offsets being applied to a new token epoch.meta.lastTouchedVersion or meta.lastTouchedAt, as these are auto-managed fields. Additionally, invalid-config startup errors now provide clearer recovery hints, including pointers to .bak backups.These changes significantly improve the operational security of OpenClaw by enforcing strict inheritance of tool restrictions and validating the provenance of node events. For self-hosted users, the ability to use wildcards for vLLM and SGLang providers drastically reduces configuration overhead and ensures that the model list remains current as servers are updated.
From a reliability standpoint, the fixes for Telegram token rotation and WhatsApp connection draining resolve critical edge cases where messages were silently dropped or replies failed during shutdown. The hardening of the plugin installation pipeline and the resolution of the Codex runtime loading issue ensure a more robust and secure extension ecosystem.