OpenClaw v2026.7.1 Breakdown: Control UI Overhaul, GPT-5.6 Integration, and Native App Rebuilds

OpenClaw v2026.7.1 is a massive release—3,063 contributions from 532 people across 2,018 public pull requests. Raw numbers don’t tell the whole story. What actually matters is the complete rewrite of the Control UI, the offline capabilities finally reaching the official iOS, Android, and macOS apps, the integration paths for GPT-5.6 and Tencent Hy3, and the fact that a few long-standing stability nightmares finally got fixed.

I’m breaking this down by functional module, focusing on what changed, why it changed, and how it behaves in practice.

图片来源：Unsplash

What exactly changed in the Control UI, and why call it an overhaul?

The entire interaction model shifted from a “single-session chat page” to a “multi-pane browser workspace.”

The core problem with the old Control UI was context fragmentation. If you were juggling multiple tasks, you relied on browser tabs. Sessions, usage stats, cron jobs, and file management lived on separate pages. v2026.7.1 pulls all of this into a unified sidebar. Sessions can sit side-by-side in resizable panes, and the layout survives a page reload.

Multi-pane sessions deserve a closer look. You can drag a session into the current pane to replace it, or drag it to either edge to create a split. A preview shows exactly where it will land. On a high-resolution desktop monitor, this is highly practical—keep a long-running task open on the left, start a new session on the right, no tab-switching required. On narrow phone screens, it degrades to a stacked view, but the active pane remains usable.

Session management is finally complete. The sidebar supports pinning, custom grouping, batch operations (rename, archive, move, delete), shared read/unread status, and transcript forking. You can organize large lists by channel, agent type, or date, and these assignments persist across reloads. Session rows expand to show labels, tuning controls, and checkpoint history without forcing horizontal scrolling.

The new Tasks page lets you inspect active and recent background work. You can open the corresponding session, refresh details, or cancel queued and running tasks if you have authorization. Previously, figuring out what the background was doing required hunting through separate interfaces.

The Usage page is entirely rebuilt. Recent estimated costs, daily averages, and the share broken down by provider, model, agent, or channel are directly comparable without repeated reloads or hover states. It also surfaces provider-reported plans, quotas, reset times, balances, spend, and budgets, covering OpenRouter and Venice. If you configure Anthropic or OpenAI admin credentials, the page pulls in daily spend, token counts, request volumes, top models, and billing categories. The 7-day, 30-day, and 90-day charts now display zero-activity days instead of compressing the timeline to only show active days. I didn’t realize how important this specific detail was until I tried to debug why traffic dropped to zero on a specific date—the blank space on the chart made it immediately obvious.

图片来源：Unsplash

The composer is redesigned for both mobile and desktop. Attachments, provider and model selection, voice input, run state, and message actions stay accessible without crowding the interface. The reasoning slider changed from “Fast to Slow” to a more intuitive “Faster to Smarter.” Quoted replies work via right-click, eliminating the old copy-paste-format routine. Desktop users can choose whether Enter sends or inserts a newline; the preference saves per Gateway.

Tool-heavy conversations now use compact, expandable rows. Inputs, outputs, previews, and errors fit on a single line—click to expand for details. Matching calls and results merge into one row, stripping redundant labels. Image results from tools render inline instead of appearing as text-only mentions. File paths in chats are clickable, opening a syntax-highlighted preview that jumps to the cited line. Hovering over a GitHub issue or PR link previews its status, title, author, and change size without leaving the conversation.

Several long-standing pain points are resolved. Temporary Gateway disconnects no longer kick authenticated pages back to the login screen; the dashboard context persists while it automatically reconnects. The disconnect notification shrank from a full-width banner to a compact “Retry” pill that doesn’t push page content downward. Non-fatal internal tool failures no longer hijack the transcript with a red error banner after the assistant has already delivered a normal final answer.

Summary: The Control UI transitioned from a single-session chat page to a multi-pane browser workspace with live task monitoring, granular usage analytics, and robust session management.

Did the first-time setup experience actually change?

Structurally, yes. The biggest trap for new users was the “configured but can’t send a message” stage. The new onboarding flow eliminates that dead end.

Fresh CLI installs enter a guided setup. Configured installs open directly in the TUI. Reruns preserve existing choices. The flow ends in a usable terminal chat. Mac users can now open the app and reach a working local agent without touching Terminal, Homebrew, administrator privileges, or extra approval prompts for the app’s own node. The onboarding verifies the AI connection before saving credentials. If authentication fails, it offers recovery steps or a manual API key path. It now displays the Gateway’s full text-inference provider catalog instead of limiting options to Anthropic, OpenAI, and Google, and it tests credentials before committing them. It also detects Claude Code, Codex, and Gemini CLI options.

Android new users go through a complete sequence: setup code or QR entry, Gateway pairing, node approval, permissions, and recovery steps. Onboarding only completes when the phone is confirmed ready.

Interrupt recovery is handled properly now. If authentication or channel setup breaks halfway through, previous choices are retained. Users without a working model are guided through credential entry and model selection instead of being dropped into an unusable chat window.

Existing installations receive strong safeguards. Doctor and update commands refuse to replace an unreadable openclaw.json, protecting existing Gateway, agent, channel, and plugin settings. When openclaw doctor --fix merges several retired model names into one current model, it preserves aliases, parameters, streaming choices, and runtime tuning, reporting merge conflicts accurately. Extended-stable updates retain the selected release channel. Container upgrades finish required migrations and plugin repairs before reporting the Gateway as ready. Unsafe runtimes are blocked before they can open SQLite-backed state, preventing data corruption.

Admins and owners can now renew an expired Codex or OpenAI login via private Telegram, Web UI, Discord, or Slack commands—no SSH required. In multi-team scenarios where not everyone has server access but someone needs to fix an expired login, this is genuinely useful.

Summary: Onboarding shifted from “configure and throw into chat” to “verify before unlocking.” Interrupt recovery and existing configuration protections are now in place.

Are the native iOS, Android, and macOS apps worth updating?

Yes. The core reason: the apps are no longer blank slates when you lose connectivity.

All three platforms now display recently cached sessions and transcripts when the Gateway is unreachable. iOS and iPadOS open directly to the last known conversation. Android behaves identically. The Mac app restores recent sessions and keeps them browsable while disconnected.

iOS and Android let you queue text messages while offline. These messages survive app restarts and send in order after reconnection, with clear retry and delete controls. The Mac app queues offline sends tied to the correct Gateway identity.

Preserving scroll position sounds minor. In practice, having a long conversation where new messages and tool activity blast your reading position to the bottom is deeply irritating. iOS and Android chats now lock the reader’s scroll position as new content arrives, providing a clear way to jump back to the newest reply.

Apple Watch users can dictate a message and hear OpenClaw’s final reply, with separate silent-send, cancel, and stop controls. If you need to query something while commuting or your hands are occupied, this works better than expected.

The Mac app saw the largest changes. It now provides a full session sidebar, actual new-session creation, slash-command completion, context and cost visibility, message copying, transcript export, and compact or clear-history actions. This elevates the Mac app from a “terminal wrapper” to a client with its own complete interface.

When the Gateway comes back, all three clients refresh the same conversation. State divergence across endpoints is no longer a thing. Mobile message queues survive app restarts.

Summary: The three native apps gained offline reading and offline queuing. The macOS app now has a full native chat interface, and the Apple Watch gained basic conversational ability.

How do you connect GPT-5.6, Tencent Hy3, and Meta Muse Spark 1.1?

Each model has its own integration path, but they all operate within OpenClaw’s provider routing system, retaining their own authentication, model limits, reasoning controls, tool behavior, image support, and usage reporting.

Organizations with GPT-5.6 preview access can select the Sol, Terra, or Luna variants through supported OpenAI and Codex routes. These come with the expected context metadata, text and image input, reasoning levels, and cache-write accounting. New OpenAI API key and ChatGPT/Codex setups select the intended GPT-5.6 defaults without overwriting an explicit model choice. Supported Ultra selections stay aligned between OpenClaw and native Codex. Codex supervision records native Codex child agents as tasks and returns their results to the parent session.

Tencent Hunyuan Hy3 has a complete setup path via TokenHub or TokenPlan. This includes first-class authentication choices, CLI onboarding, model discovery, and configuration validation. Existing TokenHub configurations retain access to hy3-preview alongside the stable hy3 model.

Meta Model API is configured using MODEL_API_KEY or the onboarding authentication choice. You select meta-model-api/muse-spark-1.1 and get supported streaming, tool calling, image input, and reasoning-effort controls.

Claude Sonnet 5 is selectable through direct Anthropic, Claude CLI, supported Vertex regions, Bedrock inference profiles, and Bedrock Mantle, with its documented context, output, image, thinking, and pricing behavior. Claude Mythos 5 is available through supported Anthropic-family routes with its required context, output, adaptive-thinking, caching, and replay behavior.

On the provider side, Copilot sessions can explicitly use custom OpenAI, Azure OpenAI, Ollama-compatible, or Anthropic providers and restart cleanly when connection or credential settings change. Paired computers can run short tasks on eligible local Ollama models, with a separate toggle to disable node inference. ClawRouter can expose the models granted to a single managed key via one CLAWROUTER_API_KEY and report their usage and budget. You can claim active promotional models using openclaw promos claim <slug> without overwriting existing defaults. LongCat-2.0 installs through the official LongCat plugin. Gemini latest aliases now continue tool-calling conversations across both supported Google transports instead of failing with a missing-thought-signature HTTP 400.

Summary: GPT-5.6 (three variants), Tencent Hy3, and Meta Muse Spark 1.1 each have distinct integration paths. Claude coverage expanded, and Copilot/ClawRouter provider selection is more flexible.

What actually changed with Codex and external coding agent workflows?

The security model shifted. External tools no longer need process-wide credentials or the ability to impersonate another session. They get temporary, revocable, session-scoped access.

openclaw attach launches Claude Code with temporary access to the main or a selected Gateway session. Credentials stay out of command-line arguments, and the grant revokes when the session ends. You can invoke Claude Code for code modifications within an OpenClaw session without scattering Claude Code’s credentials across the entire process environment. I avoided using similar features on shared machines precisely because the credential exposure was too broad.

Codex app-server agents can again list allowed agents, spawn OpenClaw or connected subagents, and yield during delegated work without resumed heartbeats hitting incompatible tool definitions. Native Codex delegation from OpenClaw threads creates native task records again. Unsupported custom Codex app-server versions fail at startup with a clear minimum-version error instead of failing silently—the latter is far more dangerous because you think something is running when nothing is happening.

Copilot sessions can use explicitly selected custom providers and restart cleanly when model, endpoint, credentials, headers, or token limits change.

Session persistence improved. Important sessions can be pinned, renamed, archived, restored, and monitored as ongoing work. An active /goal continues guiding later turns and survives compaction, queues, and interruptions until the goal is paused, completed, blocked, or limited. Previously, running long tasks meant losing the goal after context compaction and having to set it again. That experience was broken.

Summary: openclaw attach narrows Claude Code’s access to the session level. Codex delegation chains regained reliability, and long-running sessions and goals no longer vanish during compaction or interruptions.

What got fixed across Telegram, Slack, Discord, and the rest?

Each channel had different pain points, but the shared direction is: stop losing messages, make progress visible, and recover cleanly.

Telegram

Photo albums now pass every successfully downloaded image to the model in order, instead of sending one photo and a bunch of unusable file references. Failed items are omitted. The live progress view keeps reasoning, commentary, formatting, and tool activity readable and stable while running, then leaves a concise completion summary instead of vanishing. Failed final replies or bot delivery sends during flaky network handshakes are reduced, without increasing duplicate-message risk for failures that may have occurred after a request was sent.

Slack

Replies stay attached to the root thread’s session history instead of accumulating separate, nearly empty child sessions. This problem previously polluted session lists with empty shell threads. When Slack accepts a reply but the confirmation is lost, OpenClaw verifies the existing post and avoids sending a duplicate. Longer conversations avoid unnecessary waits before reaching the agent.

Discord

Completed partial-stream replies arrive as fresh messages that trigger normal unread cues after a user leaves the channel. Replies, cron reports, and other outbound messages are less likely to disappear during a brief Gateway reconnect, without replaying chunks or media Discord already accepted. Bots can recover in place after repeated session-resume rejections, reducing prolonged outages and full Gateway restarts.

Signal

Auto-replies now quote the message that triggered them, making responses easier to follow in groups. Users can assign stable names like signal:me or signal:ops to frequent contacts and groups instead of reusing raw phone numbers, UUIDs, or IDs.

WhatsApp

Logged-out or replaced accounts stay stopped until reconnection instead of entering repeated restart loops that consume resources and slow other accounts. Eligible messages received during a brief reconnect reach OpenClaw for an automatic reply instead of being marked read and silently skipped. Concurrent replies, media, reactions, and queued messages now send one at a time per account, reducing interference and unpredictable delivery order. Completed replies continue delivering when another message arrives mid-delivery, preventing overlapping conversations from going unexpectedly silent.

Apple Messages

Photos bridged from a remote iMessage Mac to a separate Gateway host now arrive in Codex conversations through a readable local path rather than a Mac-only filename. AppleScript-only and SSH iMessage setups deliver replies even when threading is unavailable. Direct iMessage chats using the imsg private API bridge now show typing as soon as a slow turn is accepted, without waiting for a read receipt. Ordinary sentence text no longer gets swallowed just because it happens to end with user:, system:, or assistant:. This bug was absurd—if you wrote something like “please start your reply with assistant:”, the entire text block would vanish.

Summary: The shared improvement across six channels is delivery reliability and progress visibility. The WhatsApp restart loop fix and the Apple Messages role-marker misidentification fix address two very specific, long-standing pain points.

Is the infinite Gateway crash loop finally fixed?

Yes. Previously, a supervised Gateway that repeatedly failed during startup would relaunch indefinitely. The operator got trapped in a restart loop with no room to investigate.

Now, a repeatedly crashing Gateway leaves a stable control process available for inspection and repair. Automatic channel and provider restarts pause until the underlying problem is fixed.

This looks like a small change on paper. If you’ve ever been woken up at 3 AM to find a Gateway in an infinite restart loop, SSH inaccessible, and logs cycling endlessly, you understand how critical this “stop retrying” behavior is. It gives you an operational window to read logs, change configs, or roll back a version, instead of watching it crash and restart in an endless loop.

Summary: Gateway crash behavior shifted from “infinite restart” to “stop and wait for manual intervention.” This is one of the most practical operational improvements in the release.

What’s new with cron jobs, remote browsers, and workspace terminals?

Three separate features, grouped together because they all relate to automated workflows.

Scheduled jobs can now wake only when an external command finishes or a watched condition changes, rather than fully booting the agent on every poll. Reapplying the same declaration updates the intended job in place, preserving identity and history without creating duplicate schedules. Specifically, a cron job can watch a build, deploy, script, or other command and resume the originating workflow with its exit code and recent output when it finishes. Alternatively, a cron job can watch an external condition and run its real payload only when that state changes, avoiding full agent wakeups on unchanged polls. This directly impacts resource consumption and cost—waking the full agent just to poll a state that hasn’t changed is an expensive waste.

For remote browser control, operators can pair the bundled Chrome extension with a remote Gateway, share only selected signed-in tabs, diagnose Windows and WSL2 connection details, and revoke access by removing a tab from the OpenClaw group. Agents can wait for delayed downloads and save the finished file to a guarded location. The “share only selected tabs” design matters—you do not want the agent looking at all your logged-in sessions.

Guarded workspace terminals are now available across web, iOS, and Android, no longer restricted to desktop browsers.

Summary: Cron jobs evolved from “timed full wakeups” to “condition-triggered wakeups.” Remote browser pairing supports selective tab sharing, and terminal capabilities reached mobile devices.

Actionable Checklist

1. After upgrading, check the Control UI sidebar to confirm session grouping, pinning, and multi-pane drag-and-drop work correctly.
2. Configure provider credentials on the Usage page; Anthropic or OpenAI admin credentials unlock granular billing data.
3. Disconnect your network, then open the iOS, Android, or macOS app to verify cached sessions are visible and messages can be queued.
4. Integrate new models: GPT-5.6 via OpenAI/Codex routes, Hy3 via TokenHub/TokenPlan, Muse Spark 1.1 via MODEL_API_KEY.
5. Run openclaw attach in the terminal to let Claude Code access the current session, then verify credentials are revoked on exit.
6. If you have existing cron jobs, reapply their declarations and confirm no duplicate tasks were created.
7. Manually trigger a startup failure condition to verify the Gateway stops in an inspectable state instead of looping.
8. Update the official app on all three platforms, particularly macOS for the new native chat window.

One-Page Reference

FAQ

Will upgrading overwrite my existing openclaw.json?
No. Doctor and update commands refuse to replace an unreadable configuration file and will preserve your existing settings.

Do multi-pane sessions work on mobile?
They work, but degrade to a stacked layout. The currently active pane remains fully usable.

How do I choose between GPT-5.6 Sol, Terra, and Luna?
It depends on your organization’s preview access and use case. You select them directly within OpenClaw through the OpenAI and Codex routing options.

Are offline-queued messages guaranteed to send after reconnecting?
They attempt to send in order. Failed sends expose retry and delete controls, assuming the Gateway eventually becomes reachable.

What are the prerequisites for openclaw attach?
Claude Code must be installed, and the OpenClaw Gateway must be running. Process-wide credentials are not required.

When does the WhatsApp restart loop occur?
In older versions, it triggered when an account was logged out or replaced. v2026.7.1 keeps the account stopped until you manually reconnect.

Does the Usage page need extra configuration for Anthropic/OpenAI billing data?
Yes. You must configure valid admin credentials. Once set, the Usage page pulls daily spend, token counts, and related data automatically.

Do condition-triggered cron jobs conflict with traditional cron?
No. Condition-triggering is an optional declaration format. Traditional time-based cron jobs continue to function normally.