Personal AI agent project OpenClaw just dropped v2026.5.19, a fast iterative update that focuses on multi-model fault tolerance, secure mobile communication, and better sub-agent lifecycle synchronization.
On the model compatibility front, the system fixes a serious interruption bug with Google Gemini. The new version proactively discards truncated Gemini thought signatures before replaying conversations, preventing malformed Base64 data from killing the assistant's next response.
For Chinese LLM fault tolerance, the system now properly reclassifies HTTP 429 payloads from Moonshot (Kimi) when the account runs out of credits. Instead of being lumped in with generic rate limits, these are now accurately identified as billing blocks, triggering the right prompts and fallback model switching.
On the cross-device communication layer, the Android client introduces smooth TLS certificate rotation. When the gateway TLS fingerprint changes, the system pops up the old and new SHA-256 fingerprints for user verification, replacing the crash-prone public key pinning approach.
Under the hood, sub-agent handoffs get a rigorous race condition fix. The new version requires that a handoff signal in the queue fully lands in the parent context before it's marked as delivered. Combined with a session keepalive mechanism, this eliminates the risk of losing sub-agent results under extreme load.
Additionally, the browser plugin now prioritizes Chrome CDP-managed diagnostics, eliminating false failures from cold-start HTTP probing races. The WhatsApp channel now fully supports sending original images and videos at the document level.
From eliminating cross-generation concurrency losses to precisely identifying vendor error codes, OpenClaw is steadily shoring up the scheduling stability of multi-agent systems under complex networks and long sessions.