OpenClaw v2026.5.22 Release Summary

Key Changes

OpenClaw v2026.5.22 introduces a series of deep technical optimizations and stability improvements aimed at reducing latency, improving memory management, and hardening the system across different operating systems.

Gateway Performance & Latency

One of the most significant updates is the massive reduction in /models listing latency. By pre-warming the provider auth-state map at startup, the per-call cost for model listing has dropped from approximately 20 seconds to 5 milliseconds—a 4,100× speedup. This ensures that model pickers in Discord and Telegram remain responsive and no longer trigger interaction timeouts.

Further performance gains include:

• Plugin Registry Reuse: The Gateway now reuses compatible startup plugin registries during dispatch, avoiding redundant loads and reducing first-dispatch latency.
• Lazy Loading: Startup-idle plugin work, core gateway method handlers, and the embedded ACPX runtime are now lazy-loaded, allowing Gateway health and ready signals to fire without waiting for unused handler trees.
• Resource Caching: The system now caches plugin SDK public-surface alias maps and reuses immutable plugin metadata snapshots to avoid repeated filesystem stats on hot paths.

Agent & Sub-agent Coordination

Coordination between parent agents and delegated workers has been refined for better token efficiency and reliability:

• Bootstrap Optimization: Users can now configure a specific list of bootstrap files for sub-agents (e.g., using SUBAGENTS.md instead of the full SOUL.md and USER.md stack), reducing token overhead for delegated tasks.
• Completion Delivery: Codex-native spawn_agent completions are now proactively delivered back to the parent session, ensuring that background worker results are visible in Discord without requiring a manual user poke.
• Sub-agent Handoff: The native sub-agent completion handoff has been simplified; children now report their latest visible assistant result directly to the requester.

Memory & Session Management

To support high-load deployments, the session store has been re-engineered to reduce memory retention:

• Immutable Read APIs: New snapshot and entry-level read APIs replace mutable full-store clones, preventing the duplication of large skillsSnapshot.prompt strings that previously led to OOM errors in high-session environments.
• Locking Hardening: Session write locks now independently enforce the maxHoldMs policy during acquisition, preventing the Gateway from freezing when large session files (>300KB) cause slow writes.
• Context Budgeting: The /status command now exposes estimated context budget snapshots when fresh provider usage is unavailable, providing better visibility into prompt pressure.

OS-Specific Hardening

• macOS (launchd): The reload handoff script now polls for unload completion to prevent Input/output error races during bootstrap. Additionally, concurrent bootstrap races are now handled gracefully by probing the service state before reporting failure.
• Windows: The Codex app-server now validates command overrides to prevent malformed strings (e.g., combining an executable and arguments) from being treated as a single module path. The Windows installer has also been hardened to handle Git checkout failures and bootstrap portable Node.js on raw hosts.

Impact

Fixed Issues

• Stability: Fixed a critical crash in bootstrap context building caused by undefined filenames in workspace bootstrap entries (#85523).
• Reliability: Resolved a bug where openclaw agent ignored SIGTERM under cron, leading to hung process chains and RAM exhaustion (#71710).
• Connectivity: Fixed a loopback CDP issue where the managed proxy blocked the browser's WebSocket handshake, breaking openclaw browser start for users behind proxies (#83255).
• Accuracy: Fixed a precision loss bug in the Anthropic transport where large integer IDs (e.g., Discord snowflakes) were rounded during JSON parsing (#47229).
• User Experience: Fixed a bug where /verbose on was ignored in normal group sessions, hiding tool and progress summaries (#85488).

Integration Improvements

• xAI/Grok: Added support for xAI OAuth profiles for Grok web_search and introduced new Grok model aliases.
• Telegram: Improved media delivery by allowing local attachment paths to be sent as actual media rather than text.
• Discord: Added a configurable agentComponents.ttlMs for component registry lifetimes, supporting longer-running asynchronous workflows.

Upgrade Guide

Breaking Changes & Migration

• Workspace Provider Plugins: For security hardening, workspace-origin provider plugins are no longer auto-discovered during setup mode unless they are explicitly trusted. Users must add these plugin IDs to the config.plugins.allow list to maintain existing behavior.
• Codex App-Server (Windows): Manual appServer.command overrides that combine an executable and arguments in a single string are now rejected. Users should remove these overrides and rely on the managed startup path or provide a valid executable path.