OpenClaw v2026.5.16-beta.6 Release Summary

OpenClaw v2026.5.16-beta.6 brings a comprehensive set of updates focusing on user experience, agent reliability, and provider integration. This release is particularly notable for its visual overhaul of the Mac application and deep architectural improvements to how the Codex app-server handles context and subagent communication.

Beyond the UI, this version addresses several high-priority issues regarding session state and message delivery, ensuring that complex agent workflows—especially those involving subagents and multi-channel routing—are more robust and transparent.

Key Changes

Mac App Redesign

The Mac application has received a significant visual and functional update. Settings pages now utilize consistent card layouts with cached navigation to prevent blanking during tab switches. The native sidebar toggle has been moved to the titlebar, and the General and Connection panes have been refined with cleaner status panels and improved spacing.

Codex App-Server & Runtime Enhancements

Several critical updates have been made to the Codex integration to improve stability and context handling:

• Context Budgeting: To prevent slow turns and oversized native context, the system now rotates stale Codex app-server bindings when native rollouts reach 70k tokens. Additionally, dynamic tool results are now capped before entering the native thread.
• Context Preservation: A fix ensures that in long sessions, the newest turns are retained when projected history is clipped, preventing the loss of recent user activity.
• Liveness & Performance: The native hook relay now suppresses subprocesses for post-tool and finalize events when no handlers are registered, significantly reducing CPU load on smaller hosts.
• Routing Fixes: Explicit OpenAI Codex harness selections are now correctly routed through the openai-codex provider, ensuring OAuth profiles are used instead of failing on missing API keys.

Subagent & Session Reliability

Improvements to the subagent lifecycle ensure that work is not lost during transient failures:

• Delivery Persistence: Successful keep-mode subagent completions that fail delivery after exhausting retries are now suspended rather than terminally cleaned up. This preserves the final payload for later recovery.
• Tighter Timing: The completionAnnouncedAt timestamp is now set only after the announcement actually reaches the parent transcript, eliminating a gap where the registry reported a result as "announced" while the parent was still waiting on a debounce window.
• Routing Accuracy: Subagent completion announces are now correctly routed to the originating parent session rather than being hijacked by parallel channel-bound sessions.

Provider & Channel Updates

• Telegram: Added warnings for media groups that silently drop photos due to download failures. Bot reply-chain context is now preserved, allowing the model to understand the history of a conversation when a user replies to a bot's media response.
• WhatsApp: The forceDocument flag is now honored end-to-end, allowing images and videos to be sent as uncompressed documents.
• Google Gemini: Added validation for thought_signature Base64 strings to prevent HTTP 400 errors caused by compaction-truncated signatures.
• Together AI: Video generation now defaults to the v2 video API endpoint.
• Feishu/Lark: Subagent spawning now returns a routable deliveryOrigin, enabling direct delivery of the initial child run.

CLI & Tooling

• Performance: The openclaw models command now reuses a plugin metadata snapshot, reducing wall time by up to 72% in large plugin installations.
• UX Improvements: Added a list alias for openclaw sessions to match other CLI conventions. PowerShell completion onboarding now provides concrete profile paths and runnable reload commands.
• Diagnostics: openclaw plugins doctor now explicitly warns when a configured Codex runtime is missing the required Codex plugin.

Impact

Fixed Issues

• #76552: Resolved high CPU load during Codex runtime tasks by suppressing idle hook relay subprocesses.
• #82911: Fixed a bug where the message tool routed replies to webchat instead of the originating channel due to stale context.
• #82913: Corrected the timing of subagent completion announcements to reflect actual transcript delivery.
• #82841: Improved diagnostics for Gateway protocol mismatches to help identify stale local clients after rollbacks.
• #82780: Fixed a bug where adding a named Telegram account would silently disable the default bot account.
• #55216: Eliminated silent image loss in Telegram media groups.

Security & Stability

• Trusted Proxy Auth: Same-host internal callers in trusted-proxy mode can now use gateway.auth.password as a local direct fallback, restoring documented behavior for internal CLI and Gateway calls.
• Android TLS: The Android app now prompts users to accept new Gateway TLS thumbprints during certificate rotations instead of hard-failing.
• Memory Core: Added a startup catch-up scan for session sources to ensure indexing doesn't fall behind after gateway restarts.

Upgrade Guide

Breaking Changes & Migration

• Trusted Proxy Policy: This release revisits the fail-closed policy for trusted proxies. While password-authenticated fallback is now allowed for clean same-host direct requests, bearer tokens remain rejected. Operators requiring strict proxy-only authentication should avoid configuring gateway.auth.password and ensure direct Gateway access is firewalled.
• Codex OAuth: Legacy Codex OAuth profiles using oauthRef sidecars remain usable. However, refreshing these profiles will now rewrite the credentials inline. Users can run openclaw doctor --fix to explicitly migrate all legacy sidecar credentials to the new inline format.