reflectt-node + cloud: describe your team, AI builds it (March 3 sprint)

#ai #agents #opensource #changelog

Big day. The intent pipeline is live, the provisioning sprint closed a bunch of real blockers, and the dashboard works on mobile now.

Here's what shipped.

The headline: describe your team, AI builds it

Until today, setting up a reflectt-cloud team meant picking a template — dev, content, ops, or custom. That was always a placeholder for the thing we actually wanted: you tell us what you need, the AI figures out the rest.

That's now live.

In the wizard (reflectt-cloud#348): Step 3 is now a textarea instead of a template grid. "I run an e-commerce store and need help with support and marketing." That's all you write.

On first boot (reflectt-node#622): When your Fly machine starts, it reads that intent from TEAM_INTENT. It creates a bootstrap agent with the intent embedded in its SOUL.md. The agent's first task: design and provision a team that matches what you described — roles, SOULs, and starter tasks included.

No YAML. No template grid. Just tell it what you need.

Cloud provisioning sprint (8 fixes)

We ran the full provisioning flow end-to-end and hit a wall of infrastructure issues. We fixed them all today:

#353 — Gateway image corrected (ghcr.io/openclaw/openclaw:latest)
#354 — Gateway internal port fixed (3000 → 18789)
#355 — IPv6 allocation failure no longer kills provisioning
#356 — Provisioning errors now surface instead of masking as "pending"
#358 — Gateway IP allocation errors made non-fatal
#359 — Fly secrets format fixed (flat object, not array)
#360 — Gateway secrets moved to machine env vars (Fly's secrets REST API is broken)
#361 — Provision route timeout extended to 300s — gateway machines now actually create

The short version: "pending" used to mean "something went wrong and we're not telling you." Now it means pending. Errors surface. Machines create. Nodes appear in the dashboard.

Also: #349 pre-registers the host so your provisioned node shows up in the cloud dashboard immediately. And #351 ensures gateway machines always get at least a main agent for intent-driven teams.

reflectt-node: what else shipped

Alert-preflight persistent snapshots (#621) — Daily metric snapshots now write to disk instead of staying in memory. Before this, a server restart wiped the observation window. The 7-day canary window now survives restarts.

Mobile 375px fix (#620) — The intensity-control bar overflowed the viewport on small screens. Fixed with flex-wrap. The dashboard works on a phone now.

Insight-task bridge dedup (#619) — User-created tasks were being excluded from the bridge's cluster check, causing duplicate tasks when a reflection clustered to a user-created one. Fixed.