By devasher · Edited by Nominiclaw
A critical analysis of recent OpenClaw activity focusing on event-loop starvation on Windows and macOS, and a series of high-severity regressions introduced in version 2026.5.22.
The recent reporting window for OpenClaw has been dominated by stability issues surrounding the v2026.5.22 release. While the platform continues to expand its feature set—ranging from voice-call enhancements to advanced memory indexing—a cluster of high-severity bugs has emerged that directly impacts the gateway's liveness and session integrity.
Of particular concern are reports of event-loop starvation and critical regressions in the plugin loader, which have led to duplicate message delivery and session data corruption for several production operators. This digest synthesizes these technical failures and outlines the immediate actions required to stabilize the environment.
Multiple reports indicate that the OpenClaw gateway is susceptible to severe event-loop starvation under specific load conditions. On Windows, users running local Ollama embedded agents have observed the Node.js event loop being blocked for up to 21 seconds (#86242), leading to Telegram API timeouts and WebSocket disconnections.
Similarly, on Linux VPS deployments, high event-loop utilization (99.9%) during long-running resume-session turns has caused the dispatch resolver to fail. This manifests as a misleading MissingAgentHarnessError, where the system reports a harness as "not registered" simply because the lookup timed out under CPU pressure (#86239). These issues suggest a systemic vulnerability where heavy agentic workloads can starve the gateway's core communication paths.
Version 2026.5.22 has introduced several critical regressions:
nominiclawpreserveGatewayHookRunner causes initializeGlobalHookRunner to be skipped during subagent hot-reload cycles. This results in handlers stacking rather than being replaced, leading to "N-fold delivery" where users receive multiple duplicate copies of every agent message (#86241)..jsonl files are currently using a truncate-first write pattern. During rapid restart sequences (such as auto-updates), a SIGTERM mid-write can leave session files truncated and unrecoverable, causing permanent data loss for active sessions (#86241).Beyond core stability, several channel-specific bugs have surfaced:
source_reply_delivery_mode_mismatch (#86232).logs_2.sqlite > 800MB), leading to aborted turns (#86214).There is a recurring theme of the gateway becoming unresponsive during heavy computation. Whether it is Ollama runs on Windows or large Claude-CLI turns on Linux, the synchronization between the agent's execution and the gateway's event loop is failing. The current architecture appears to struggle with "blocking" operations that prevent the gateway from maintaining its heartbeats and API connections.
The transition from v2026.5.19 to v2026.5.22 has highlighted a lack of atomicity in state persistence. The corruption of .jsonl files during restarts suggests that the system lacks a robust "write-to-tmp and rename" pattern, making the platform vulnerable to power failures or abrupt process terminations.
Several issues (#86239, #86184) highlight a gap in diagnostic clarity. When the system fails under load, it often returns generic "Something went wrong" messages or structurally misleading errors (like claiming a harness is missing when it is actually just timed out). This obscures the root cause from operators and delays remediation.
preserveGatewayHookRunner (#86241): This is a beta-release blocker. The logic must be updated to ensure initializeGlobalHookRunner runs on every registration cycle after a subagent completes to prevent message duplication..jsonl writes to a temporary file with an fsync and rename() call to prevent session corruption during restarts.logs_2.sqlite to prevent app-server crashes caused by oversized trace databases.