OpenClaw v2026.5.28-beta.2 landed on May 29, 2026 at 7:19 AM America/Chicago time, which was May 29, 2026 at 12:19 UTC. I would file this one under: the beta you install when you are tired of half-broken recovery paths and little delivery bugs turning one normal run into ten minutes of cleanup.

The release notes cover a lot of ground, but the practical story is tighter than it first looks. OpenClaw is getting better at surviving restarts, helper failures, timeout aborts, and flaky channel edges without losing the thread or leaving you with ghost state.

What users can now do

Run Codex and subagent-heavy workflows with less fear that one bad hop will poison the whole session. This beta keeps spawned agent working directories separated, preserves shared Codex app-server state when helpers fail, releases session locks on timeout abort, and avoids stale restart continuations. If a restart has ever left you wondering whether the rest of the run is trustworthy, this beta is aimed straight at that problem.

Use chat channels with fewer identity and delivery surprises. Matrix room IDs keep their case, Slack is less likely to lose a delivered final reply during cleanup, iMessage keeps polling after denied reactions, Telegram holds onto polling and prompt config better, and Microsoft Teams gets stricter URL trust checks. None of that is flashy. All of it is the difference between "the agent answered" and "the agent answered in the wrong place, twice, and now I need to explain it."

Push broader provider and media setups without so much manual babysitting. Claude Opus 4.8 support is in, MiniMax streaming music responses are wired up, NVIDIA featured model catalogs got coverage, encrypted PDF extraction is supported, and GitHub Copilot plus a Codex Supervisor plugin path widen the menu for delegated coding workflows. If your setup is not just one model and one chat app, this beta is doing useful plumbing work for you.

Get clearer failures instead of mushy hangs. A lot of CLI, auth, doctor, and startup paths now reject malformed values faster, bound OAuth and local service waits, and give more actionable restart guidance. That may sound like internals, but it matters because a fast failure is cheaper than staring at a terminal wondering whether anything is still happening.

What got safer or less annoying

The recovery fixes are the headline for me. OpenClaw is cleaning up the kinds of bugs that make advanced users lose confidence: event queues waiting on themselves, stale restart state getting reused, missing-thread compaction failures, and helper crashes taking shared runtime state down with them. When those bugs pile up, the product starts to feel haunted. This beta goes after that feeling directly.

The channel cleanup also looks worth your attention if you run OpenClaw in public or team-facing contexts. Outbound hooks now keep canonical session identity better, fallback tool warnings stay mention-inert, Google Chat DM thread behavior stays in bounds, and Discord recovered tool warnings are less noisy. I am glad to see that kind of restraint work getting shipped because user trust gets burned on tiny routing mistakes long before it gets burned on dramatic outages.

There is also a quiet mobile and web refresh here. The iOS Pro app, Gateway chat transport, onboarding, WebChat reconnect behavior, and session picker state all got attention. That is not the main reason to install this beta, but it is a good sign that the day-to-day surfaces are still getting real product work instead of only backend triage.

What I would test right after updating

1. Retry one Codex or subagent workflow that used to go sideways after a restart, helper crash, or timeout. Do not use a toy prompt. Use the path you actually distrusted.
2. Send real messages through your fussiest channel. Slack, Matrix, iMessage, Telegram, Discord, WhatsApp, and Teams are all fair targets here. Look for duplicate prompts, wrong-thread replies, or cleanup that eats a final answer.
3. Open one encrypted PDF and one provider-heavy workflow. This beta touches both document extraction and provider catalogs, so it is worth verifying with a real file and a real model path.
4. If you run mobile or WebChat regularly, bounce the connection on purpose. Reconnect behavior and session preservation are part of the update. I would test that directly rather than assuming the notes tell the whole story.
5. Try one intentionally bad CLI or auth input. The release claims better fail-fast behavior. Make it prove it.

Should you install v2026.5.28-beta.2 now?

Yes, if you lean on advanced workflows or you have been hit by recovery bugs lately. This is a beta, so I would not throw it blindly onto the calmest production setup in the world. But if you rely on Codex, subagents, multiple channels, or fussy provider stacks, this looks like the right kind of beta: less spectacle, more operational sanity.

If your current install is boring and stable, waiting for the next stable tag is still reasonable. If OpenClaw has felt brittle around restarts, delivery, or auth edges, I think this is worth a pass now.

---

Full changelog: View v2026.5.28-beta.2 on GitHub

Generated release page: Read the structured release notes

- Fred