This `extended-stable` release of OpenClaw, version 2026.9.33, focuses on critical security updates, reliability, performance, and new model support, making it a recommended upgrade for stability. Key highlights include safer September upgrades that preserve state and plugin data, and robust update mechanisms for pnpm, npm, and Windows environments, preventing rollbacks and ensuring service continuity. New model support has been added for Meta Muse Wonder and MiniMax M3.1 Flash Preview, expanding the range of available AI capabilities. Model turns are now more reliable, with agents honoring live thinking levels, local providers receiving response-header deadlines, and improved recovery for ChatGPT WebSockets and Codex replies. Crucially, work now survives Gateway restarts, meaning interrupted automation turns, steered sessions, cloud worker claims, and partial worker replies will resume or persist, significantly improving workflow resilience. Delivery and device stability have also been enhanced, ensuring live sends survive queue recovery and node requests are no longer disrupted by unrelated pairing activity. Numerous fixes address update and Doctor issues, preventing silent failures and ensuring proper cleanup, especially for cron jobs and OAuth profiles. The release also resolves issues with Gateway responsiveness during long streamed replies, Tailscale Serve stability, context engine warnings, and state database persistence. Providers like DeepSeek and Codex have improved recovery mechanisms, and channel tools, ACP agents, and Talk consultations now exhibit greater stability. Overall, this release offers substantial improvements in system stability, update reliability, and model integration, making it a crucial update for all users. It is highly recommended to upgrade to ensure system integrity and leverage the latest model capabilities.