OpenClaw 2026.2.9

A native iOS app for pairing your iPhone as an OpenClaw node. Uses a setup-code onboarding flow to connect your phone to your agent fleet. This opens up camera snaps, location, notifications, and eventually full device control from your agents.

Availability: The iOS app is currently in internal preview, distributed via Apple's TestFlight. It is not publicly available on the App Store yet. The gateway-side pairing infrastructure (the /pair command and device-pair plugin) is fully shipped and ready — once the app goes public, you'll be able to pair immediately.

#11756

New plugin system for pairing devices via Telegram (/pair command) with full iOS and Android node controls. Pairs with the iOS node app to give agents real-world device interaction — take photos, read sensors, push notifications, and more.

#11755

You can now use xAI's Grok as a web_search provider, complete with inline citations. Adds another option alongside Perplexity and other search backends.

#12419

The Gateway now exposes agents.create, agents.update, and agents.delete RPC methods, enabling full agent lifecycle management directly from the Control UI. No more hand-editing config files to add or modify agents.

#11045

Comprehensive BlueBubbles integration and channel cleanup. BlueBubbles is a self-hosted iMessage bridge for non-Mac setups, widening iMessage access for agents running on Linux or Docker hosts.

#11093

New OPENCLAW_HOME env var lets you override the home directory used by internal path resolution. Useful for containerized deployments, multi-user setups, or non-standard directory layouts.

#12091

Chat history now shows a visual divider where compaction occurred, making it clear when context was summarized. Helpful when reviewing long conversations.

#11341

Agent envelopes now include runtime shell information, giving downstream tools and skills better awareness of the execution environment.

#1835

Forwarded commands in exec approval prompts are now rendered in monospace, making it easier and safer to scan what you're approving. Reduces the risk of visually ambiguous or misleading command strings.

#11937

Injected transcript writes after compaction were breaking the session's parentId chain, causing agents to lose memory of everything before compaction. Now fixed — agents remember again after context is compacted.

#12283

Oversized tool results could blow past the context window and crash the session. Now there's pre-emptive capping of tool output size plus a fallback truncation mechanism if something still slips through.

#11579

HTTP 400 errors are now treated as failover-eligible, so automatic model fallback kicks in for a wider range of failures — not just 5xx errors.

#1879

If a conversation mentioned "context overflow" as a topic (e.g., discussing it), the system could falsely trigger overflow handling. Fixed to only detect actual overflow conditions.

#2078

Major overhaul of quote/reply handling: preserves quote context properly, avoids QUOTE_TEXT_INVALID errors, and prevents nested reply quotes from being misclassified.

#12156

Proactive sends that fail due to stale topic thread_ids now automatically retry without the message_thread_id, recovering gracefully instead of silently failing.

#11620

Spoilers now render correctly using HTML tags instead of broken markdown syntax.

#11543

Bot command registration is now truncated to 100 entries to avoid Telegram's BOT_COMMANDS_TOO_MUCH error on startup — relevant for multi-agent setups with many custom commands.

#12356

allowFrom now correctly matches against the sender's user ID (with fallback to chat ID), fixing cases where DM access control wasn't working as expected.

#12779

Routing bindings are now refreshed per-message by reloading config at route resolution time. You can change agent bindings in the config and they'll take effect immediately — no restart needed.

#11372

Hooks have been broken since the 2026.2.2 tsdown migration. Now fixed. If you had custom hooks that stopped working, they should be back.

#9295

Multiple hardening fixes: scheduling/delivery reliability improved, announce flow shared with isolated sessions, and the cron tool now recovers flat parameters when the LLM forgets to wrap them in the job object.

#11641 · #12124

Announce timing for sub-agents is stabilized, and compaction metrics are now preserved across retries instead of being lost.

#11664

sessions.usage discovery for multi-agent setups is fixed — usage stats now correctly reflect all agents.

#11523

When gateway.bind=lan, probe URLs and Control UI links now correctly use a LAN IP instead of 127.0.0.1.

#11448

Provider-specific settings are now included in the web search cache key, preventing cross-provider result contamination. Grok's inlineCitations are passed through correctly.

#12419

Direct Perplexity model IDs are now normalized correctly while OpenRouter model IDs are left unchanged. Fixes routing issues when using Perplexity both directly and via OpenRouter.

#12795

Voyage AI embeddings now set the input_type parameter, improving retrieval accuracy for memory search.

#10818

The default QMD (memory) model cache is now shared across agents instead of re-downloading per agent. Saves disk space and startup time in multi-agent setups.

#12114

Manual refresh now smoothly scrolls to the latest messages and suppresses the new-messages badge flash that occurred during refresh.

QuickStart mode now auto-installs shell completion (previously manual-only). Also strips embedded line breaks from pasted API keys — no more invisible whitespace breaking auth.

Forum and media thread-create starter messages are now supported, with message thread create --message wired up and routing hardened.

#10062

xhigh thinking level is now allowed for github-copilot/gpt-5.2-codex and github-copilot/gpt-5.2.

#11646

.caf audio attachments (Apple's Core Audio Format) are now recognized for transcription. Relevant for iOS voice memos and macOS audio recordings.

#10982

OPENCLAW_STATE_DIR is now honored for default device identity and canvas storage paths.

#4824