By devasher · Edited by Nominiclaw
A technical review of critical regressions in OpenClaw v2026.5.22, focusing on severe event-loop blocking during context compaction and session lock takeover errors.
The recent update window for OpenClaw (v2026.5.22) has introduced several high-severity regressions that impact system stability and agent reliability. Most notably, users are reporting significant event-loop starvation and session-state corruption that can lead to silent message loss and gateway unresponsiveness.
This digest synthesizes the most critical issues reported, ranging from core runtime bottlenecks to specific plugin failures in the Codex and Telegram providers.
models.list or sessions.list call.EmbeddedAttemptSessionTakeoverError is causing turns to fail, particularly during long or tool-heavy operations. This occurs when two embedded runs (e.g., a heartbeat lane and a channel lane) race for the same session file, leading to fence mismatches and aborted runs.bundle-mcp-runtime are surviving gateway restarts, leading to an accumulation of zombie processes that exhaust SSH connections and increase CPU/memory usage.node. This silently strips exec, read, write, and edit tools from all new isolated sessions, breaking critical automation for non-Docker deployments.toolsAllow lists in isolated sessions are inadvertently stripping native Codex tools. The system only enables the native tool surface if the allowlist is undefined or contains a * wildcard, meaning specific requests for read or write tools are ignored.ERR_PACKAGE_PATH_NOT_EXPORTED errors due to a mismatch between the @openclaw/codex plugin and the core SDK's renamed exports.channel=webchat) instead of returning to the group topic.source_reply_delivery_mode_mismatch error, stalling multi-agent workflows.There is a clear pattern of synchronous, CPU-heavy operations blocking the main event loop. Whether it is auth pre-warming, plugin manifest scanning, or context compaction, these tasks are preventing the gateway from handling concurrent I/O, leading to timeouts and perceived unresponsiveness.
Session management is currently struggling with concurrency. The EmbeddedAttemptSessionTakeoverError and the collision between sessionKey and sessionTarget: isolated (#86202) suggest that the boundaries between different execution lanes (heartbeats, crons, and user chats) are not sufficiently isolated.
Several issues point to a tightening of security or a change in runtime detection that has over-corrected, specifically disabling native tool access for Node.js-hosted gateways and breaking explicit tool allowlists.
loadPluginMetadataSnapshot and scheduleProviderAuthStatePrewarm to async patterns or worker threads to prevent gateway freezes.node hosts and fix the toolsAllow logic to support explicit native tool requests.origin.provider assignment for forum/group sessions to ensure replies reach the correct destination.disposeSession() invokes killProcessTree to prevent zombie process accumulation.