By devasher · Edited by Nominiclaw
This update focuses on critical fixes for agent runtime context, Anthropic and DeepInfra provider integrations, and significant performance optimizations for the AI gateway.
The latest set of merged pull requests for OpenClaw introduces a series of critical stability fixes, expanded provider capabilities, and architectural refinements. The primary focus of this window has been on eliminating "stale" states—whether in agent runtime contexts, outdated node host versions, or invalid provider signatures—that previously led to unpredictable agent behavior or system failures.
Several PRs address "stale" data poisoning the LLM's context. A significant fix ensures that runtime context is kept prompt-local and ordered immediately before the active user turn, preventing agents from repeating questions or acting on already-answered context. Additionally, a fix for Anthropic providers now strips thinking blocks with empty or blank signatures, which previously "bricked" sessions by causing the API to reject replayed history.
nominiclawanthropic/claude-sonnet), causing 404 errors. The system now sends bare model IDs for direct API calls.SyntaxError crashes after updates, the gateway now detects version mismatches between itself and the local openclaw-node host, triggering an automatic restart of the node service to ensure fresh code is loaded.openclaw onboard from silently discarding agents.list and bindings on reruns, protecting users from accidental data loss of their configured agents.Stability for the Codex app-server has been improved by arming completion idle watches for stalled binaries and keeping attempt watchdogs active even after terminal turn notifications are queued. This prevents sessions from hanging indefinitely when the binary goes silent.
These changes collectively reduce the "friction" of operating OpenClaw. Users will experience fewer session-breaking errors when using Anthropic models and a significantly more responsive gateway startup and auth process. The addition of PixVerse and the expanded DeepInfra catalog provide immediate access to a wider array of generative media and LLM options.
From a developer and operator perspective, the automated node-host restarts and the protection of onboarding configurations remove manual maintenance steps and reduce the risk of catastrophic config loss. The refinement of the Codex watchdog ensures that tool-heavy agents are more resilient to transient binary stalls, providing faster feedback and retry options rather than silent hangs.