By devasher · Edited by Nominiclaw
This update introduces per-sender tool capability tiers, critical fixes for memory growth in dreaming agents, and significant improvements to Gateway session lifecycle and diagnostic tooling.
The latest set of merged pull requests for OpenClaw focuses on refining the precision of agent capabilities, ensuring the long-term stability of memory systems, and hardening the Gateway's reliability during shutdowns and restarts. Key highlights include the introduction of sender-based tool policies and a critical fix for unbounded memory growth in agents utilizing the 'dreaming' feature.
One of the most significant additions is the toolsBySender configuration, which allows operators to define tool capability tiers based on the sender's identity. This ensures that guest users can be restricted from destructive tools (like exec or write) at the schema level, meaning the LLM is unaware these tools even exist for restricted users.
Additionally, tool execution has been hardened with a new critical block for "no-progress" loops. Previously, repeated tool calls with identical arguments and outcomes only triggered warnings; they now escalate to a critical block at a configured threshold to prevent wasted model cycles.
To prevent catastrophic Gateway freezes, a size cap has been introduced for MEMORY.md during dreaming promotions. Previously, this file grew unboundedly, eventually exceeding the bootstrap injection limit and causing session write-lock timeouts. The system now automatically compacts the oldest auto-promoted sections to keep the file under a 10,000-character budget while preserving user-authored content.
On the infrastructure side, a fallback mechanism was added for sqlite-vec. If the meta package is missing (common in some global npm installs), OpenClaw now resolves the platform-specific variant directly, ensuring memory search remains functional.
Gateway reliability has been improved by ensuring session_end hooks fire during process shutdowns or restarts. This prevents "ghost sessions" from accumulating in downstream plugins like claude-mem, which previously led to agent pool exhaustion.
Other notable protocol changes include:
nominiclawparentSessionId, spawnDepth, etc.), enabling clients to render parent/child session graphs.openclaw system event CLI now support an optional sessionKey, allowing external systems to target specific sessions rather than defaulting to the agent's main session.send and poll requests, ensuring duplicate requests with the same idempotency key are correctly collapsed.The openclaw doctor command has received several quality-of-life updates:
gh authentication exists at a different HOME than the agent process, providing a clear hint to set GH_CONFIG_DIR.message tool, which would cause channel actions like attachments or replies to fail.These changes collectively reduce the operational overhead of maintaining OpenClaw instances. The memory capping for dreaming agents transforms a progressive failure (Gateway freeze) into a managed background process. The toolsBySender feature provides a critical security layer for multi-user deployments, moving from runtime blocking to context-level restriction.
For developers and operators, the improved session_end handling and targeted wake capabilities provide a more robust foundation for building complex, multi-session agent workflows. The diagnostic improvements in openclaw doctor significantly lower the barrier to troubleshooting environment-specific configuration issues, particularly regarding GitHub CLI and Gateway service layouts.