Searchable Agent Memory + Smarter Run Evaluation

Planning chat now uses searchable agent-local memory, completed runs can check whether they actually met the goal, and setup friction dropped for model credentials and filesystem browsing.

This release is about better continuity. Agents retain context more usefully, completed runs are judged by outcomes instead of ceremony, and a few rough edges around setup and browsing have been sanded down.

Planning Chat Now Has Searchable Agent Memory

The Memory tab and planning chat now share a more capable agent-local memory layer.

  • Memory search lets you retrieve past context instead of hoping it surfaces on its own
  • Category views make stored context easier to inspect
  • Visible connection status makes the memory system feel less opaque when it is healthy, loading, or unavailable

Completed Runs Can Check Whether The Goal Was Actually Met

Run completion is now more outcome-aware.

  • Success criteria can be evaluated when a run finishes
  • A run can fail because the result missed the goal, not just because a tool crashed
  • Planning becomes more honest about whether meaningful progress actually happened

Small But Important Friction Was Removed

  • OpenRouter now has its own credential flow with clearer account-level status
  • The filesystem view can recover from stale saved paths instead of getting stuck on them
  • Recurring work and site deploys are less likely to wedge into bad states