Changelog

Insights — per-people view and instant range switching

See consumption by agent and find out which agents people are actually using. Insights gains a By Agent / By People toggle on the Top Spenders table. The people view groups spend by member and distinguishes Unidentified (a real person from an inbound adapter who isn’t linked) from Unattributed (no person — background jobs, cron, SDK calls).

Historical traces from undeployed agents no longer inflate account totals.

Resource sizing and persistent disks for agents

For advanced agents that need control over underlying compute and data resources, you can now size an agent and give it a persistent disk at deploy time from the deploy form. CPU, memory, and storage tier are tunable, with a live cost estimate as you adjust. Data on a mounted volume survives redeploys; turning the volume off releases the disk. The deployment Metrics tab shows pod CPU and memory limits as reference lines so headroom is clear.

Editing only the advanced provisioning fields (CPU, memory, volume mount, storage size) now triggers the Save & Redeploy action like every other redeploy-relevant field.

Network traffic on the Monitor page

See whether your agents are doing what they should and catch unexpected API calls before they become a bill or a security question. The agent Monitor page has a new Network Traffic section alongside Token Usage, Requests & Latency, and Traces. It surfaces inbound, outbound, and database call volume with top peers, response class breakdown, and latency.

Pod workload metrics

Right-size with confidence, see exactly how much CPU and memory each agent is using before you bump its limits. The pod detail panel now has a Metrics tab with live CPU and memory charts across 1h / 6h / 24h / 7d windows.

Agents page — faster loads, clearer state

Open the Agents page and trust what you see: live state, fresh “Update available” pills, no stale red errors. The Agents page now loads in well under a second. The “Update available” pill appears immediately after a new build is published rather than waiting on a refresh. The Launch button stays available as soon as a messaging URL exists, and the Error pill only shows when a deployment is actually in an error state (not when an old error message is still on the row).

Deploy form awareness of agent interfaces

Simpler deploy flow for self-hosted-frontend agents: no adapter picker, no validation step for something they don’t use. The deploy form now reads what an agent declares. Agents that don’t run the messaging sidecar no longer see a Slack/Web adapter picker, and the validation around it is also gone.

Inline secrets survive configure and redeploy

Reconfigure or redeploy without re-typing your API keys. Inline secrets typed directly at deploy time now round-trip through configure and redeploy. Configure shows a masked placeholder for stored values; a redeploy that doesn’t touch the secret carries it forward instead of clearing it.

Public URL readiness

Deployment endpoints now report whether each public URL is actually reachable. The web Launch button reflects this with a clear tooltip while the load balancer finishes provisioning. ast deploy --wait polls until the URL is ready, and ast agent get prints readiness alongside the URL.

Connectors moved to their own settings page

Identity connections are an org-level concern now, not a personal-account afterthought. GitHub and Slack identity links now live at Settings → Connectors. Reauthorize on the GitHub connector works (previously no-op). Because GitHub’s OAuth re-auth doesn’t re-show the org selection screen for already-authorized apps, a “Request access on GitHub” link is now exposed for managing org access directly in GitHub.

CLI

• Every interactive prompt and TUI now accepts esc (and ctrl+c) to cancel cleanly with a uniform footer and a Cancelled. exit.
• ast push and ast secrets recover from mid-command auth expiry with a single refresh + retry — no more token-expired failures during long Docker builds.
• ast push asks about visibility before the build step, so declining doesn’t waste build time.
• ast create no longer falls through to printing an empty Claude prompt when the form is cancelled.
• ast push accepts dev.interfaces.messaging.slack.observe_channel_ids.

Run ast upgrade to pick up 0.13.2.