OpenClaw v2026.5.26 Release Summary

Key Changes

Core Infrastructure & Performance

OpenClaw v2026.5.26 focuses heavily on reducing Gateway latency and improving system observability. A major architectural shift makes Transcripts a core primitive, providing a reliable path for meeting summaries, source-provider chunks, and CLI/TUI replay.

Performance gains were achieved by caching plugin metadata snapshots, package realpaths, and session/auth hot-path facts, significantly reducing rediscovery overhead during Gateway startup and reply paths. Additionally, the system now avoids runtime model hydration on hot paths and defers context compaction to background maintenance to prevent blocking reply startup.

Channel Enhancements

Several channels have been upgraded to production-ready status with improved interaction models:

• Telegram: Now preserves inbound text entities (bold, italic, code, etc.) as Markdown and handles forum topics more reliably. It also fixes a critical bug where durable retry targets used an incorrect string format, causing failed sends to be lost.
• iMessage: Now supports thumb approval reactions (👍/👎) for exec and plugin approvals. It also fixes a race condition where catch-up could replay live-handled messages after a restart.
• WhatsApp: Now supports thumb approval reactions and provides better diagnostics when group inbound messages are dropped due to missing configuration.
• Discord: Improves voice playback and model picking, and now suppresses self-reply prompt echoes to prevent the bot's own previous messages from cluttering the agent's context.

Security & Content Boundaries

Security hardening has been applied across multiple surfaces:

• SSRF Protection: Browser snapshot reads now honor SSRF policies, validating tab URLs before reading content.
• Prompt Injection: The memory_store tool now rejects prompt-like instruction payloads before they reach embeddings or storage.
• Content Sanitization: System-event text is now sanitized to prevent spoofing nested prompt markers, and serialized tool-call text is scrubbed from Discord replies.
• Auth Hardening: The Gateway now enables a default auth rate limiter for remote non-browser HTTP auth failures.

Voice & Talk Mode

Realtime Talk capabilities have been expanded, particularly for iOS, which now features a direct realtime voice path using gateway-issued client sessions and a more compact UI toolbar. Wake-name handling has also been made more tolerant to reduce accidental triggers from ambient speech.

Impact

For Developers and Operators

• Observability: New OpenTelemetry LLM content spans and a sanitized secrets.prepare timeline make it significantly easier to inspect failures and startup bottlenecks.
• Diagnostics: The openclaw status command now surfaces systemd cgroup hygiene (KillMode, memory, and task counts), helping identify leaked child processes.
• Configuration: The openclaw path tool now supports deep JSON edits (beyond 4 segments), making it viable for managing complex openclaw.json structures.

For End Users

• Reliability: Users will experience fewer "silent drops" in Telegram and WhatsApp, and more consistent behavior in the TUI, which now queues prompts submitted while the agent is busy instead of blocking them.
• UI/UX: The Control UI now provides actionable feedback when config.openFile fails in headless environments and correctly renders failed tool results as errors rather than successful outputs.

Upgrade Guide

Breaking Changes & Migration

While this release is largely backward compatible, operators should note the following:

• WhatsApp/Signal/iMessage Approvals: Approval prompts now prioritize thumb reactions (👍 for allow-once, 👎 for deny). Users wishing to use allow-always must continue using the manual /approve <id> allow-always text command.
• iMessage Authorization: If channels.imessage.allowFrom is configured, the /approve text command is now authorized against this specific list rather than the broader DM allowlist. Ensure all intended approvers are listed in allowFrom to avoid authorization failures.