Release notes
To save you the effort of manually retyping long command strings from scratch, the system has patched the Ctrl+R history search (Ctrl+R history search, a feature that lets you press Ctrl+R to quickly search through commands you typed earlier) tool. Previously, searching for old commands this way could trigger a string-matching bug that crashed the terminal, forcing you to restart the session; now it is resolved so searches run stably. It's like a library catalog search terminal that used to freeze if you typed too fast; now the database search is patched and never locks up.
To prevent the AI from taking risky actions behind your back when you step away from your desk, dialogs now wait for a human by default. Previously, when the AI asked a question via AskUserQuestion ( the dialog prompt where the AI pauses to ask for your feedback before acting), if you left the terminal unattended, it would auto-continue by itself and potentially break things. Now it waits indefinitely until a human responds, unless you manually configure an idle timeout (idle timeout, a timer that lets the AI take action after you've been away for a set time) in /config. It's like a junior builder who used to knock down a wall if you didn't answer his question; now he stands with his hammer waiting for your nod, unless you tell him 'if I'm away for ten minutes, proceed.'
A new configuration allows you to guide the workflowSizeGuideline ( a guideline setting that controls how many sub-helper agents the AI is allowed to summon when breaking down a massive task) in /config. By selecting small, medium, or large sizes, this serves as an advisory guideline for small/medium/large agent counts, not a strict limit or cost cap, but helps guide the AI against spawning dozens of subagents to tackle a complex task, which used to burn through your quota in seconds. It's like hiring a contractor who used to show up with thirty assistants and bill you for all of them; now you can give them an advisory guideline, which is not a strict limit, to keep your budget under control.
Subagents ( smaller assistant AIs spawned by the main agent to work on subtasks) now fail cleanly when rate-limited. Previously, if they hit a rate limit (rate limit, the speed limit rules that prevent calling an API too many times in a minute) they would return empty results instead of an error, causing the main AI to assume the target file or data was empty; now they throw a clear error so the main agent clearly receives the failure. It's like sending an errand boy to get letters, and because the post office was closed, he returns and tells you 'you have no mail'; now he tells you 'the post office was closed so I couldn't get in,' saving you from being misinformed.
System now preserves authentication credentials during daemon restarts. Previously, when tasks ran under background daemons (daemons, background server programs that run tasks quietly in the background), any connection reset caused the restart sequence to lose the secure connection key (socket auth token, the security credential used to communicate securely with the main application), leaving the agent frozen or orphaned in memory; now the token is preserved, and orphans are swept away. It's like an amusement park ride that broke down, and during the queue restart the ride operator took away your fast-pass ticket, leaving you stranded; now the ticket remains valid through restarts so you can finish the ride.