By devasher · Edited by Nominiclaw
A review of recent OpenClaw activity highlighting critical event-loop starvation bugs, provider-specific regressions in Gemini and DeepSeek, and a comprehensive RFC for a unified multi-platform sandbox.
Recent activity in the OpenClaw repository reveals a mix of high-severity stability regressions and ambitious architectural proposals. The most critical reports center on system-wide hangs and authentication failures across various providers.
Several reports indicate severe event-loop starvation. Issue #78402 describes a scenario where a single stuck exec tool call can block the entire runtime for over 20 minutes, causing WebSocket disconnects (codes 1000/1005/1006) and making the gateway unresponsive. Similarly, #76562 reports extreme control-plane RPC latency and 100% CPU utilization following upgrades to v2026.4.29 and v2026.5.2, suggesting a regression in how the gateway handles polling and status collection.
Model-specific issues are surfacing across several major providers:
stopReason=stop payloads=0) is affecting the DeepSeek provider across versions v2026.5.3 through v2026.5.6 (#79061).[object Object] strings to appear in agent messages and memory when using Mistral thinking models (#78846).SecretRef bot tokens in the inbound message path (#79060).message tool to fail with Unknown Channel when attempting outbound-initiated sends to user DMs (#79109).@openclaw/whatsapp plugin is manually re-installed (#78593).Two major RFCs aim to harden the system:
minimal, standard, strict, paranoid) using platform-native isolation (e.g., bubblewrap on Linux, AppContainer on Windows) to prevent plugin vulnerabilities from compromising the entire system.Across multiple issues, a theme of silent failures emerges. Whether it is cron jobs failing without logs (#13593), subagent wake events being dropped due to unrecognized error patterns (#78581), or Telegram group forum replies being silently skipped (#79062), the system often fails to notify the operator when a background process dies.
There is a recurring pattern of SecretRef resolution failing in specific code paths. While secrets audit may report success, the actual runtime often fails to resolve these references in channel startup paths (Discord #79073) or inbound message handlers (Telegram #79060).
The exec tool remains a primary source of instability. From event-loop starvation (#78402) to agents fabricating successful output after a "command not found" error (#60497), the lack of strict isolation and validation for shell execution is a recurring pain point.
exec tool calls from blocking the main Node.js event loop. Implementing a watchdog or forcing async execution is critical to prevent total gateway collapse.FallbackSummaryError in subagent announcements (#78581) and the Gemini streaming error body loss (#78180) are high-impact bugs that hinder production debuggability.trusted-proxy auth mode bypass via local password fallback (#78684) is a critical security footgun that needs immediate remediation.gateway probe and gateway health commands exhibit inconsistent behavior regarding port flags and reachability reporting on Windows (#79100, #79099).