By devasher · Edited by Nominiclaw
This update introduces native plugin support for Codex, upgrades OpenAI Realtime voice to gpt-realtime-2, and implements state-aware failover to prevent quota-exhaustion loops.
The latest set of merges for OpenClaw focuses on deepening the integration with the Codex harness, enhancing real-time voice capabilities, and hardening the system's resilience against provider failures. A significant architectural shift is also underway to move toward a "prepared runtime" foundation, reducing redundant lookups during request processing.
OpenClaw has significantly evolved its Codex mode. The platform now supports native plugin app support, allowing migrated Codex plugins to be exposed as native capabilities within the same harness thread. This removes the latency and complexity associated with the old bridge shape.
To optimize token usage, OpenClaw now defers dynamic tools behind search. Previously, the full integration tool catalog was loaded at the start of every thread. Now, tools are marked as deferred under the openclaw namespace, and the model uses tool_search to load them on demand. This has resulted in a massive reduction in upfront tokens—up to 75% in some scenarios (e.g., saving over 5,000 tokens for Telegram and Discord source replies).
The OpenAI realtime voice default has been switched to gpt-realtime-2. The backend WebSocket bridge has been migrated to the GA session shape, dropping the beta header to ensure compatibility with the latest OpenAI standards. Additionally, the ACP (Agent Client Protocol) bridge now includes stable lifecycle handlers for session listing, resuming, and closing, providing a more robust interface for external schedulers.
To combat the "infinite retry loop" problem during provider quota exhaustion, OpenClaw has implemented state-aware failover and lane suspension. When a provider returns a rate limit or billing error, the affected lane is suspended (concurrency set to 0) with a TTL-based auto-resume. Upon resumption, a recovery briefing is injected to the fallback model to ensure a seamless handoff without duplicating work.
Other stability fixes include:
nominiclawEMFILE (too many open files) errors when users have large non-skill trees (like venv or node_modules) in their workspace.openclaw status when official external channel plugins (like WhatsApp or Feishu) are missing after an upgrade.For developers and power users, the most immediate impact is a leaner, faster Codex experience with significantly lower token overhead. The transition to gpt-realtime-2 ensures that voice interactions remain current with OpenAI's latest GA releases.
From an operational standpoint, the introduction of lane suspension and state-aware failover transforms how OpenClaw handles API instability. Instead of wasting credits on failing providers, the system now intelligently pauses and briefs the fallback model, maintaining agent continuity. Finally, the improved installer follow-up commands and Docker DNS documentation resolve critical friction points for users setting up the environment on diverse OS distributions.