OpenClaw v2026.5.10-beta.1 Release Summary

OpenClaw v2026.5.10-beta.1 brings a significant set of updates focused on expanding the extensibility of the Gateway, refining the reliability of voice and chat integrations, and hardening the security of the CLI and runtime environment. This release is particularly important for operators managing private skills and those utilizing real-time voice capabilities.

Key Changes

Gateway & Skills Extensibility

• Private Skill Archive Uploads: A new admin-gated path now allows trusted Gateway clients to upload private skill archives via a new RPC (skills.upload.begin/chunk/commit). This is disabled by default and must be explicitly enabled via skills.install.allowUploadedArchives in the configuration.
• ACP Runtime Mapping: To improve compatibility with various backend dialects (such as Claude ACP), OpenClaw now maps canonical runtime options (e.g., thinking) to backend-advertised keys (e.g., effort or reasoning_effort) automatically.

Voice & Real-time Capabilities

• Real-time Voice Diagnostics: Added comprehensive diagnostics for Discord voice, including speaker turns, playback resets, barge-in detection, and audio cutoff analysis.
• Custom Talk Instructions: Operators can now append custom real-time voice style instructions via talk.realtime.instructions while preserving built-in agent-consult guidance.
• Discord Voice Optimization: Default installs now use the pure-JS opusscript decoder to avoid slow native addon compiles outside of dedicated performance lanes.

Integration Enhancements

• Telegram:
  • Improved streaming previews in partial mode by coalescing delta-shaped fragments, ensuring the 30-character debounce is respected without blocking previews.
  • Fixed a bug where no-response direct message (DM) turns would synthesize visible "silent-reply" chatter.
  • Added safe handling for managed select callbacks, preserving delimiter-containing values (e.g., env|prod).
• WhatsApp:
  • Fixed an issue where cumulative tool-call preambles were sent as separate, growing messages; only the new suffix is now delivered.
  • Restored echoTranscript functionality for inbound voice notes by passing the full routing context to the transcription preflight.
• Slack: Interactive reply buttons and select menus now reliably wake the resolved session, ensuring conversations continue visibly after a user interaction.
• iMessage: Threaded replies now support attachments (gated by the imsg build capability), allowing images and files to be sent as part of a threaded response.
• Feishu: Fixed a bug where message(action="send") replies in group_topic sessions were posted to the group root instead of inside the topic thread.

Core Agent & CLI Improvements

• Process Persistence: Background exec/process session references are now preserved across embedded compaction and after-turn runtime contexts, preventing subagents from losing track of long-running tasks.
• Execution Transparency: Failed exec/bash tool calls now surface a concise ⚠️ 🛠️ Exec failed warning by default, preventing cases where an assistant claims success despite a tool failure.
• CLI Configuration: Fixed a critical issue where config set and config patch would silently drop values that matched schema defaults. These explicit writes are now persisted.
• Copilot Integration: Fixed an identity mismatch in GitHub Copilot token exchanges by using the vscode-chat integration ID, resolving 400 errors for image-capable models.

Impact

Fixed Issues

• #79284: Subagent compaction no longer orphans running exec processes.
• #78188: Eliminated unwanted "All quiet on my side" messages in Telegram DMs.
• #79856: Resolved silent data loss during configuration updates via CLI.
• #79946: Fixed GitHub Copilot image tool request failures.
• #60497: Prevented agents from fabricating success after a failed exec tool call.
• #74903: Feishu topic-scoped replies now land in the correct thread.
• #79778: Restored WhatsApp voice note transcript echoes.

Security Hardening

• Console Redaction: The Gateway now installs console capture before command parsing to suppress sensitive WhatsApp/libsignal session dumps from appearing in raw terminal logs.
• Test Sandboxing: The audit-exec-surface test suite is now properly sandboxed under a temporary home directory to prevent it from accidentally truncating live exec-approvals.json files on developer machines.
• Lazy Loading: The rich tree-sitter command explainer is now lazy-loaded, ensuring that lightweight calls like exec.approval.list do not unnecessarily load heavy dependencies.

Upgrade Guide

Configuration Changes

If you wish to use the new private skill upload feature, you must add the following to your configuration:

{
  "skills": {
    "install": {
      "allowUploadedArchives": true
    }
  }
}

Dependency Updates

This release updates several critical workspace pins, including:

• OpenAI: 6.37.0
• Google GenAI: 2.0.1
• AWS SDK: 3.1045.0
• Baileys: 7.0.0-rc10
• Kysely: 0.29.0