OpenClaw shipped v2026.5.14-beta.2 on May 15th — the second beta in this cycle — and it's a big one. The changelog covers realtime voice calls, a completely revamped WhatsApp status-reaction system, mid-turn steering for active runs, and meaningful transparency improvements for subagent tasks. Here's what landed.
Telnyx Realtime Voice Calls
The headline feature: OpenClaw agents can now participate in realtime media-streaming voice calls via Telnyx. The implementation (#81024, thanks to @dynamite-bud) adds a conversational voice-call mode to the Telnyx channel, distinct from the existing one-off TTS replies. This means your agent can hold a back-and-forth spoken conversation routed through Telnyx infrastructure — a significant step toward phone-native agent workflows.
WhatsApp Gets Expressive Status Reactions
WhatsApp was the last major channel without lifecycle status reactions. That changes in beta.2: the StatusReactionController is now wired into WhatsApp message turns, giving it the same queued → thinking → tool → done/error lifecycle already present on Telegram and Discord (#80612, thanks to @gado-ships-it).
The default emoji set was also replaced across all channels with self-explanatory indicators:
- 🧠 thinking
- 🛠️ tool use
- 💻 coding
- 🌐 web request
- ⏳ stall (soft)
- ⚠️ stall (hard)
- ✅ done
- ❌ error
- 🗜️ compacting
These read as system status rather than emotional commentary — a sensible upgrade if you run a channel-facing agent and want users to understand what's happening without guessing.
Mid-Turn Steering
A long-requested improvement: mid-turn prompts now steer active runs by default via /queue steer (#77023, thanks to @fuller-stack-dev).
Previously if you sent a message while the agent was mid-run, it would queue. Now it steers — meaning the agent can absorb new direction without finishing its current trajectory. The older /queue followup and /queue collect behaviors remain available for users who prefer those semantics. And /steer gracefully falls back to a normal prompt when steering isn't available.
Subagent Tasks Are Now Visible
Subagent delegation got a transparency upgrade. sessions_spawn tasks are now delivered in the child session's first visible message ([Subagent Task]) instead of being hidden in the sub-agent system prompt (fixes #78592, thanks to @bradestes and @stainlu).
The result: delegated work stays auditable without duplicating tokens in the system prompt. You can see what was handed off directly in the conversation thread.
Continue Codex Sessions from OpenClaw Chat
OpenClaw agents can now list and bind existing Codex CLI sessions running on a paired node, letting an OpenClaw conversation continue where a local Codex session left off. Combined with the Codex harness becoming the default for OpenAI model turns (covered in the official blog post), this tightens the Codex/OpenClaw integration considerably.
Canvas Lazy-Loading
Gateway startup time gets a modest improvement: Canvas implementation costs are now deferred until first use, rather than loading at startup (#82001, thanks to @samzong). If you don't use Canvas features, your gateway cold-start skips that overhead entirely.
Notable Fixes
- Structured-output restored:
response_formatfrom OpenAI-compatible/v1/chat/completionsrequests now properly flows through agent stream params (#82004). - Discord streaming: Stale truncated draft previews are now cleaned up when finalization fails (fixes #82035).
- Config numeric keys: Discord guild IDs and other numeric-looking object keys survive config patch merges (#81999).
- Telegram bot init race: Isolated polling no longer retries every update with "Bot not initialized" on startup (#81975).
- macOS screen snapshots: Malformed params are now rejected before capture, and base64 results are bounded against projected frame limits (fixes #68181).
- Cron timeout fix: Isolated agent jobs that have dispatched to a backend are now governed by their configured job timeout instead of the 60s pre-execution watchdog (#81871).
Upgrade
npm update -g openclaw
Or check the full release notes on GitHub for the complete changelog across 80+ contributors.