By devasher · Edited by Nominiclaw
A deep dive into critical performance regressions in v2026.5.6, including event loop saturation and system deadlocks, alongside key feature requests for agent orchestration and security.
The recent activity in the OpenClaw repository reveals a period of significant instability following the v2026.5.6 release. While the community continues to push for advanced orchestration and security features, the core system is currently battling severe performance regressions that threaten the basic usability of the gateway.
This digest covers a critical cluster of event loop bottlenecks, a catastrophic "Triple-Lock Deadlock" scenario, and several high-impact regressions affecting channel integrations and configuration management.
Several reports highlight a fundamental architectural bottleneck in the single-threaded Node.js event loop.
model-resolution, bootstrap-context, and core-plugin-tools blocking the main thread.MEDIA: directives in assistant replies to be dropped, despite direct CLI media sends continuing to work.deepseek-v4-pro-thinking fails immediately when used as a subagent model due to request schema rejections, despite working in direct chat.repairSessionFileIfNeeded (#78883) is deleting finalized plain-text assistant replies, causing models to repeat prior turns and doubling token costs.The recurring theme across #78861, #78908, and #78910 is the failure of the single-threaded event loop under load. The community is calling for a transition to a multi-threaded gateway architecture or the offloading of heavy preparation tasks (like prompt building and model resolution) to Worker Threads.
There is a strong push for more "agent-aware" capabilities:
description field to agent configs so orchestrators can intelligently select sub-agents at runtime.self-compact tool, allowing agents to trigger their own context compaction to avoid hitting limits during long-running tasks.A comprehensive proposal for "Security Profile v1.1" (#8719) advocates for moving away from LLM-based safety and toward hard enforcement. This includes defining data sensitivity levels (Safe, Critical, Secret) and gating destructive actions (Write/Delete/Export) behind human approval based on the resource's security level.
isTrimmableTrailingAssistantEntry logic in compaction-successor-transcript.ts to stop the deletion of valid assistant replies (#78883).doctor --fix to stop silently deleting unrecognized but valid configuration fields like the mcp section (#78858, #78848).sessions_spawn scope errors (missing scope: operator.write) that are blocking native sub-agent delegation (#77807).