By devasher · Edited by Nominiclaw
A technical review of critical configuration wipes, event-loop starvation, and provider-specific API regressions affecting OpenClaw's gateway and runtime.
Recent activity in the OpenClaw repository reveals several critical stability issues, ranging from catastrophic configuration loss during updates to severe event-loop starvation that renders the gateway unresponsive.
One of the most severe reports involves version 2026.5.7, where an automatic update via npm reportedly wiped the entire openclaw.json configuration and deleted the ~/.openclaw/credentials/ directory. This resulted in a full outage for users, losing all configured agents, channels, and plugins.
Additionally, macOS users have reported a "permanent gateway outage" caused by a version mismatch between the OpenClaw binary and the configuration. This occurs when the launchd gateway service is hard-wired to a source checkout rather than the npm-installed package, causing the older binary to refuse to install the service over a newer version's config.
Multiple reports highlight a systemic vulnerability to event-loop blockage. A "Triple-Lock Deadlock" was documented where a zombie dashboard session, 100% Event Loop Utilization (ELU), and repeated model API timeouts compounded to make the system unresponsive for over 20 minutes.
Similar patterns were observed in the Telegram and Feishu channels. For Telegram, the lack of a circuit breaker in the fetch transport leads to thousands of ENETUNREACH errors per hour during offline windows, starving the event loop and triggering TUI streaming watchdogs. In Feishu, synchronous startup phases—specifically model-prewarm and session-locks—block the loop for over 20 seconds, causing all WebSocket heartbeats to fail and connections to drop.
Several provider-specific regressions have emerged:
nominiclaw400 errors regarding reasoning_content not being passed back in multi-turn conversations, and immediate rejections when used as subagent models.thoughtSignature is dropped during cross-provider replay.claude-opus-4-7, claude-sonnet-4-6) are rejecting requests because the gateway continues to send the deprecated temperature parameter.api.kimi.com and cached broken endpoints.There is a recurring theme of updates breaking core functionality. Beyond the config wipe in v2026.5.7, users have reported that openclaw doctor --fix can inadvertently remove the mcp section and top-level defaultModel from the configuration, breaking MCP tools.
Many of the critical failures stem from the single-threaded nature of the Node.js event loop. Whether it is synchronous startup tasks, aggressive retry loops without circuit breakers (Telegram), or heavy I/O on EFS-backed deployments, the result is the same: high ELU and unresponsive WebSocket/HTTP interfaces.
As models move toward "reasoning" and "thinking" modes, the gap between OpenClaw's generic request construction and provider-specific requirements (like DeepSeek's reasoning_content or Gemini's thoughtSignature) has widened, leading to a surge in 400 Bad Request errors.
openclaw.json and the credentials/ directory.model-prewarm and session-locks to an asynchronous pattern to prevent WebSocket disconnects during gateway boot.thoughtSignature and reasoning_content replay issues to restore subagent functionality for these models.temperature, top_p, and top_k for models flagged as reasoning: true in the catalog.openclaw doctor --fix to preserve unrecognized configuration sections (like mcp) instead of deleting them.