OpenClaw beta.2: Voice Steering, 4,100x Speed, and iMessage Approvals

OpenClaw dropped v2026.5.24-beta.2 late Sunday night, and the changelog is one of the more packed betas in recent memory. Three headline features stand out above the noise — plus a stack of important fixes and a major privacy tightening for sub-agents.

Real-Time Voice Run Steering

The biggest UX win in this release: you can now check your agent's run status, cancel it, steer it, or queue follow-up work — all while a consult is still running — from both the WebUI and Discord voice. PR #84231 adds realtime wake-name gating and raises the profile bootstrap context budget so longer USER.md and SOUL.md files stay fully loaded during voice sessions.

Previously you were stuck waiting for a run to finish before redirecting it. Now you can interrupt mid-flight, which matters when an agent goes off on a long autonomous task and you need to course-correct without waiting it out.

A 4,100× Speedup on /models

This one is almost comically large. PR #84816 pre-warms the provider auth-state map at Gateway startup, so the /models endpoint and every model-listing call short-circuit per-provider plugin and CLI discovery on the hot path. Per-call cost drops from roughly 20 seconds to ~5 milliseconds — a 4,100× improvement.

The warm state resets and re-warms after hot reloads, so it won't serve stale data. If /models has always felt sluggish in your setup, this should fix it for good.

iMessage Tapback Approvals

Continuing the pattern from WhatsApp approval reactions, iMessage users can now 👍 a pending action to approve it (allow-once) or 👎 to deny it. The approver allowlist is read from channels.imessage.allowFrom. The allow-always escalation still requires the manual /approve allow-always text command — a sensible gate for the more powerful option.

Meeting Notes Plugin Ships

The new meeting-notes plugin arrives as an external source-only plugin with auto-start capture, manual transcript imports, and read-only openclaw meeting-notes CLI access. Discord voice is the first live source. Summaries include speaker-labeled transcripts for Discord group voice captures, so you actually know who said what.

Sub-Agent Context Isolation

PR #85283 meaningfully tightens sub-agent privacy. Default bootstrap context for delegated workers is now limited to AGENTS.md and TOOLS.md only — persona files, USER.md, MEMORY.md, heartbeat files, and setup files are kept out of sub-agents by default. This prevents accidental leakage of personal context when spawning workers to handle tasks on your behalf.

Other Notable Additions

• Image quality preference: new agents.defaults.imageQuality setting for token-efficient, balanced, or high-detail image handling in the image tool
• xAI/Grok web search: OAuth auth profiles now carry through to Grok web_search calls (#85182)
• Chat session search: the WebUI session picker gets search and Load More pagination (#85237)
• Classic onboarding: bare openclaw commands before a config exists now trigger the guided setup flow (#72343)
• Embedding providers API: new generic embeddingProviders contract lets plugins register embeddings as a first-class provider surface
• Security fix: agent-to-agent wildcard allowlist regexes replaced with a precompiled linear matcher to eliminate backtracking-prone patterns (#85849)

Under the Hood

The release also ships extensive Windows compatibility fixes across scripts, release checks, and plugin staging — plus a significant Gateway perf pass that lazy-loads startup-idle plugin work and caches immutable plugin metadata snapshots so hot paths avoid repeated filesystem walks.

As a pre-release, beta.2 can be tested now with npm install -g openclaw@beta. The full changelog is on GitHub.