What Shipped This Week: Chat That Actually Finishes

Three things landed together: chat turns no longer die mid-stream or save as blank bubbles, helper work shows as a compact status line instead of flooding your thread, and Growth/Pro/Max list prices went live at $29/$59/$199 on 1 September.

Chat turns no longer vanish into blank bubbles or false "AI is down" errors

Before this week, three failure modes could make a specialist look broken when it was not. You would send a message, see the thinking indicator, and then get an empty assistant bubble — or a banner claiming the AI provider was down when the real problem was our side losing the draft. Worse: if we shipped a deploy while your turn was still streaming, the connection dropped and the partial answer never saved. You had to re-ask. On a phone between meetings, that is not a minor annoyance — it is the difference between trusting the agent and opening a laptop to check whether the duty actually ran.

We fixed the full chain. The parent draft now persists before helper work starts, pins to your user turn, and we never store an empty assistant bubble. OpenClaw no longer runs a second moderation pass on continuation blobs unless you explicitly opt in — that double-check was silently killing valid replies. When a deploy rolls out, a 30-second drain waits for in-flight chat turns to finish instead of cutting them off; if a turn cannot complete, we write an interrupt flag so you see what happened instead of silence.

Why this matters in the context of how CloudAxis works: specialists run on a persistent cloud computer with scheduled duties and multi-agent handoffs. Chat is how you steer them between runs — approve a report, redirect a scan, ask Cloudia to add a duty. If chat feels unreliable, the whole OS feels unreliable, even when the overnight duty wrote the file correctly to ~/files/. We wrote about agents that keep working while you sleep in our always-on agents guide; this week we made the check-in loop match that promise.

Edge case we closed: custom employees hired with an empty skills override now get search and files tools again, so a misconfigured hire does not look like a dead specialist. Summarizer calls on GPT-5.6 Luna use max_completion_tokens so long helper rounds do not truncate mid-sentence. If the model never starts talking, we silently re-kick the stream instead of leaving you staring at a frozen composer.

Also shipped

Smaller changes that landed in the same window:

• Helper work as a compact line. When Cloudia spawns a helper for web research or browser work, you see a single status chip above the composer — not a second conversation overwriting the ticker. The helper board drops as soon as helpers go idle. Stop stays live (solid red) while helpers run, and your next message sends immediately after Stop instead of hitting a 409.
• List prices live. Growth is $29/mo, Pro $59/mo, Max $199/mo as of 1 September. Founding members keep their current Stripe price through 1 March 2027. Growth checkout trial retired — paid plans bill at checkout. Settings now shows plan upgrade paths on mobile and desktop. Subscriptions fall back to Free when a cycle ends unpaid instead of sticking on a dead paid tier.
• Usage pills show percentage first. Credit and browser-minute counters in the status bar lead with how much of your plan you have used — not raw numbers buried in a tooltip.
• PWA fixes. Desktop install prompt restored for CloudAxis. Double title bar on desktop PWA is gone. The closed More sheet no longer peeks under the bottom nav on phone.
• Duty board cleanup. Duplicate "Scheduled task running" rows no longer stack while a duty is active. Browser sessions opened for scheduled duties close after the run finishes, freeing browser minutes instead of lingering idle.
• Notifications. Mark all read is instant — no waiting on a slow POST before the badge clears.

What's next

We are watching chat completion rates and deploy-interrupt frequency this week. If you still see a blank bubble after a deploy, tell us which specialist and thread — the new interrupt flag should make those reports actionable.

Frequently asked questions

Do I need to refresh after a deploy?

No. In-flight turns now get a 30-second drain window. If your turn was interrupted, you will see an explicit message — not a blank bubble.

Did helper behavior change?

Yes. Helpers still run behind Cloudia, but their progress is a compact chip instead of flooding the thread. Stop works mid-helper and Send is available for your next message immediately after.