This cycle’s biggest change is how CrossMind handles onboarding and background task tracking — it now gives you meaningful context from the first session, not a blank slate. We also shipped a substantial browser reliability fix that eliminates stuck sessions, and cleaned up social authentication across the board.
Revamped Onboarding and Background Task Tracking
Starting a new task or bringing a new agent online now gives you real context immediately. The onboarding flow has been rebuilt to surface relevant background research and task setup in the first session rather than leaving agents to figure out context from scratch on every run.
Background task tracking is also more reliable. Subagent lifecycles now resolve correctly when tasks are running in parallel — previously, background tasks could get stuck in a “running” state even after completion, making it look like work was still in progress when it wasn’t. The subagent session view now maps correctly to the actual execution transcript so you can follow exactly what ran.
The activity feed has also been cleaned up: tool output grouping now stays consistent across thinking blocks, so the session view is easier to follow even on complex multi-step tasks.
Why it matters: When an agent starts blind, the first session is mostly setup. The revamped onboarding front-loads that work so agents can do meaningful work from day one. Reliable task status means you can trust what you see — if something shows as complete, it is.
Self-Healing Local Browser
The local browser engine has been substantially rebuilt to fix a class of stuck-session bugs that had been affecting production.
The root cause: when a browser session failed to release properly, it would silently leave Chrome processes running while marking the session as free. This created “orphan” sessions — Chrome instances that were consuming resources and blocking new sessions with misleading errors (“all slots occupied” when slots weren’t actually in use).
The fix is multi-layered:
- Session release now confirms the browser actually stopped before clearing the database record — no more silent failures that leave orphans behind
- If a session fails to start, the system automatically force-releases any stuck locks and retries before giving up
- A container-level reconciliation pass runs at startup and every minute to find and clean up any orphan sessions that slipped through
- When your dedicated browser profile is temporarily unavailable, agents fall back gracefully to a temporary profile with a clear notification rather than failing silently
Why it matters: Browser automation tasks are among the most complex workloads agents handle. Stuck sessions and misleading errors are particularly disruptive because they’re hard to diagnose from the outside. This fix makes browser-based tasks significantly more reliable and self-recovering.
Also Shipped
- Artifact tool paths visible — Tool cards now show the file path when an agent reads or writes an artifact, making it easier to trace exactly which files an agent is working with during a session
- Social auth reliability — OAuth login flows for social platforms now handle edge cases and retry failures more gracefully, reducing authentication interruptions during outreach tasks
- Session creation refresh loop fixed — A bug where clicking “New Session” in Copilot would trigger a UI refresh loop has been resolved