This update removed several sources of avoidable friction around starting agents, authenticating coding tools, and interpreting heartbeat health.
Agent Setup Became Simpler
The new-agent flow now asks for fewer confusing choices upfront and uses agent-kind styling more consistently.
The agent prompt system also moved away from older planning hierarchy language, so new agents start with instructions that better match the current workspace model.
Codex And Model Selection Became Clearer
Codex device-auth flows are better protected against duplicate starts and permission failures, and the inner-loop model list now includes newer Codex options.
For users, this means fewer confusing auth attempts and a clearer path to choosing the coding model an agent should use.
Planning And Heartbeats Became More Reliable
Planning chat image attachments were fixed, recurring process retries were tightened, and heartbeat health now treats a running recovery as a recovery instead of immediately escalating it as a failure.
The practical outcome is less false alarm noise and fewer cases where a healthy recovery looks worse than it is.