Everything that landed across the beta cycle, rolled up into the first stable, publicly-installable release since 1.2.4.
🧭 New — guided setup skill. Run openclaw skills install ocuclaw-assist, then ask your OpenClaw agent to install, upgrade or repair OcuClaw — it walks you through each step.
🦞 OcuClaw is now on ClawHub — openclaw plugins install ocuclaw just works.
Grab a coffee, this is a big one. ☕
🧠 New: Neural Modules
The agent can now reach past a plain text reply and shape how the response feels on your glasses.
Neural Emoji Reactor — lets the agent nominate a status-bar emoji for specific spans of its streaming response, swapping it in sync with what you're reading. Reserved for genuinely impactful moments; the usual status-bar behaviour runs the rest of the time. Toggleable in the Display tab.
Neural Pace Modulator — lets the agent dial the reveal speed of specific phrases as it streams: <dwell> to slow a line down so it lands, or <skim> to rush past a recap that doesn't need to breathe. Everything else reveals at your normal reading pace and it combines cleanly with the Emoji Reactor.
Neural Topic Distiller — the agent gives each session a short title from context and quietly renames it as the discussion shifts. Ask it to rename anytime, or edit titles yourself; a manual rename always wins and locks the auto-titler out of that session. It runs as a small background agent after your turn, so it never bloats your prompt or derails a reply.
🛠️ New: Agent Tools
Hardware Info — a new agent tool that surfaces live G2 hardware state, starting with battery percentage, so the agent can factor your glasses' status into its replies.
🎙️ Voice & Speech-to-Text
Cartesia Ink-2 is now a second STT provider — pick your speech-to-text engine in Voice settings. Ink-2 is purpose-built for real-time voice agents, with ~100 ms to a final transcript, built-in turn detection and strong capture of structured data.
Soniox default model → stt-rt-v5 — the retired v4 is auto-migrated. Accuracy improves across 60+ languages, noisy audio, accents, interruptions and overlapping or mixed-language speech.
- Endpoint Sensitivity tuning on glasses and WebUI controls how eagerly the mic decides you've stopped talking.
- Conversation Context primes the recognizer with your agent name and session title.
- The entire voice pipeline is faster and more reliable from capture through stream-start.
- Voice-latched lifecycle keeps the spoken user message visible through capture, commit, thinking and stream start.
- Configurable voice auto-timeout, with a glasses selector and WebUI controls.
- Continue-listening setting for chained Even AI voice turns.
🗂️ Agents & Sessions
- Per-mode and per-session Agent selection — choose which OpenClaw agent runs for each mode and each session.
- WebUI per-session agent switcher — swap to a different specialist mid-work.
- Session multi-select agent filter — filter your session list by one or more agents.
- Faster Sessions tab — history is cached on-device, reconnect syncs by diff and transcript search can auto-run.
- Better mobile session search — the fixed search row uses a native mobile keyboard and keeps focus after the first character.
- Active turn controls — stop, replace or steer an in-flight agent turn.
- New WebUI Sessions browser — OcuClaw and Even AI tabs, Preview, Pin, Rename, Delete, bulk actions, live search and pinned favourites.
👋 First-launch Experience
- A new welcome screen stays on the glasses until you connect, with an animated OCUCLAW wordmark, scanline loader and tagline type-in.
💬 Messaging Screen Redesign
- New text rendering engine — the messaging screen reliably fills more of each page before flowing to the next.
- Faster, more stable, lower-flicker streaming, page-0 transitions, session switches and voice hand-off.
- Voice and typed messages display with parity: the same first-chunk behaviour and reveal pacing.
- Listen-resume holds until you've finished reading the streaming response.
- Unified header and body rendering with independent truth models.
- Assistant typing placeholder before the first response content arrives.
- Manual-send overlay on the glasses after sending from the phone UI, remote control or voice.
🌊 Streaming Modes & Flow
- Continuous mode scrolls the response line-by-line instead of jumping to a fresh page.
- Dynamic flow eases the reveal in and subtly changes pace at paragraph endings.
- Word-atomic reveal prevents mid-word line-wrap flicker.
🪟 Page Indicator & Scroll Bar
- Scroll-bar marks response-start pages with
· and continuation pages with ..
- Fast-scroll v2 ramps page skips as you go; a one-second pause resets the streak.
- The page indicator is pixel-stabilized and right-anchored.
😀 Agent Emoji
- Speaking animation while text streams.
- Idle heartbeats every 30–60 seconds and short mini-sequences every 2–4 minutes.
- Show Agent Emoji master toggle, plus a separate Idle Animations toggle.
🎛️ Agent Modes
- Fast mode toggle — opt supported models into faster output, per session.
- Exec escalation mode toggle — choose whether the agent can request elevated exec access, per session.
🐞 Bug Reports & Diagnostics
- One-tap Send Bug Report — add a note, review what's included, then send the bundle with a copyable reference.
- Works while disconnected — builds a client-only diagnostics bundle with an offline export escape hatch.
- Selectable history window — defaults to the last five minutes, with a Max option.
- Safer contents — secrets are stripped before upload and Review shows exactly what's included.
- Record a window — start and stop a short diagnostic capture from the WebUI with a live timer.
🧭 Glasses Skills Menu
- Redesigned skills browser with a fixed selection box, descriptions, safe truncation and clean swipe or tap navigation.
- Cleaner agent menu labels, with long names truncated safely and the default agent marked directly on glasses.
⚠️ Transient Failure Feedback
Short error labels now appear in the glasses header for about 3.5 seconds, then clear on your next listen, send, chunk or disconnect:
- Assistant run: Out of credits, API key invalid, Rate limited, Provider unavailable, Provider timeout, Request failed, Run failed.
- Voice: No Soniox credits, Soniox key invalid, Soniox rate limited, Soniox unavailable, Soniox key missing, Mic access failed, Voice disconnected.
- Send: Image send failed, Image too large, Connection lost, Send failed.
The relay setup section auto-expands with a typed error for every connection failure — Invalid token, Bad relay address, Wrong server type or Cannot reach relay — and the Connect button changes to Retry.
🔐 Exec Approval
- Approval screens honour server-provided allowed decisions, with improved reliability.
👓 Glasses Display
- Root-menu double-tap opens the SDK exit prompt; the old Exit row becomes a Conversation shortcut.
- Glasses rate-limit indicator — the header shows reset time when limits are hit.
📱 Phone UI, Theme & Settings
- Brand-new phone app theme and layout, with icons across every setting.
- Help button, Help modal and per-setting help across Display, Session, OcuClaw, Even AI and Voice.
- Agent avatar or emoji across the phone UI, cached client-side.
- Session token usage with a compaction button.
- Even AI controls visibly disabled when Even AI isn't enabled — no more dead controls.
- Improved iOS and Android keyboard handling.
- Refreshed Info popup; Report a Bug is easier to find and Discord points at discord.ocuclaw.com.
📊 Provider Usage & Rate Limits
- Adaptive hero card and secondary list in the OpenClaw Status pane.
- Multi-auth provider-pool display suppresses the warning while a fallback is ready or rotating.
- Rate-limit details and reset windows render in your local timezone.
🛡️ Reliability, Performance & Polish
- Removed an O(history) layout cost so long replies stream without build-up stutter.
- Thinking glyph paints about 40% faster on G2; glasses bridge writes are bounded so a stuck call can't hang rendering.
- Page-0 paragraph-break freeze fixed, plus dozens of smaller streaming, voice, session-menu and display fixes.
🔗 Connection & Relay
- Smarter reconnect remembers which addresses have proven good and re-seeds them on cold-start auto-connect.
- Hardened relay worker adds backoff, watchdog, socket liveness, ping reaping, slow-client shedding, queue caps and connect timeouts.
- Plugin and app versions are exchanged in the handshake; the WebUI detects a mismatch and pauses auto-reconnect.
- Reconnect tolerates slower networks and high-latency Tailscale or cellular handshakes.
- Honest “never connected” copy and restored per-error messages in the connection panel.
- The 1008 reconnect storm and several first-connection or iOS first-start lockups are fixed.
📌 Migration note — Connection
The default Tailscale setup now uses dual-port: the relay moves to direct TCP wss://…:8444, while https://…:8443 stays for the Even AI endpoint. Existing users are walked through the swap by the installation support agent.
🚚 Distribution, Install & Onboarding
- OcuClaw is now published on ClawHub — a bare
openclaw plugins install ocuclaw resolves to ClawHub by default; npm remains the beta channel and fallback.
- Minimum host version raised to OpenClaw 2026.6.9.
- Guided setup lane — install
openclaw skills install ocuclaw-assist and let your agent walk through install, upgrade and repair.
- Config schema enforces
relayToken and evenAiToken requirements at config-set time, with friendlier startup errors.
📌 Migration note — Even AI
If your config has evenAiEnabled: true without evenAiToken, updating will fail validation. Set the token, or disable Even AI before updating.
📌 Migration note — Tools profile
If openclaw config get tools shows a profile, grant plugin tools back with openclaw config set tools.alsoAllow '["ocuclaw"]' --strict-json, then restart the gateway.
⚙️ Behind the Scenes
- Website launching soon. The release timeline, setup and support material were on the way.
- Fleet Control Center — a web cockpit for remote development boxes with live glasses streams, input routing, live diagnostics, trace tabs, screenshot galleries and milestone notifications.
- Debug-triage service — uploaded bug reports feed a secured dashboard and agent API for investigating bundles safely.
- remote-debug skill — stand up a public, token-gated instrumented build for an external tester and watch the live boot or trace beacon.
Thanks to everyone who ran the betas and sent reports.
