You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
Scheduled triggers that run on a calendar recurrence, with time zones.
Conversation continuity: a new arrival in the same issue, pull request, or thread continues the agent already working there instead of starting another, with explicit keys and an opt-out.
A new arrival for a running agent steers it, including agents that run with scoped execution credentials.
A Home page for new organizations with a setup checklist, a notice when provider events arrive with no trigger listening, and a seven-day overview.
Saving a trigger validates its agent against the target daemon, so a model, mode, or provider the daemon cannot run is refused at save instead of failing the first run.
A startup_timeout setting for machines that need longer than the default two minutes to start an agent.
Changed
Upgrade Paseo daemons before upgrading Hub: conversation continuity needs the current daemon.
One consistent interface across sign-in, organization administration, projects, connections, triggers, and instance settings.
The trigger editor walks four numbered steps: what arrives, who may send it, where it lands, and what runs there.
Trigger forms list only the events Hub handles, and label events require the label they match.
Alerts say what happened, and empty pages no longer draw empty tables and panels.
Paseo shows Hub executions under their trigger's name instead of an execution id.
Fixed
Active agents keep their execution credentials across Hub restarts and reconnect to the same daemon agent.
Long rendered prompts reach the agent in full.
New GitHub agents receive their reply tool, and preserved legacy workflows dispatch again.
The sidebar and breadcrumb stay on the page that is visible while the next one loads.
Unsupported methods on the execution MCP route return 405 instead of an HTML page or a 500, and a daemon sending an invalid WebSocket close code no longer crashes Hub.