- Added
accept-editsmode to--remote-agentsessions:Shift+Tabnow switches between the default mode andaccept-editsinstead of doing nothing. Sessions still start in the default mode, where the remote agent now asks before each file edit in its workspace, so you may see edit approval prompts you didn't get before; switch toaccept-editsto let edits go ahead without asking. - Added the Orchestrator model to
/modelfor Gemini Enterprise accounts and sign-ins with Application Default Credentials (AGY_ADC_AUTH): it plans with Gemini 4 Carbon (High) and hands the plan's execution to Gemini 4 Hydrogen (High), and it appears when you can select both of those models. - Added a conversation navigator to the Remote Control browser view: a rail with a mark for each of your prompts previews the prompt when you hover over it and jumps to it when you click, and
[and]jump to the previous or next prompt. - Added a
Quotebutton to the Remote Control browser view: select text in an agent response and clickQuoteto add it to your next message as a quote. - Added clickable links to the message box and your sent messages in the Remote Control browser view: a web address you type or paste opens with
Cmd+click(Ctrl+clickon Windows and Linux) or a middle-click, while a plain click still places the cursor. - Added the time left to timers the agent sets, in the Remote Control browser view: each running timer in
Background Tasksshows a countdown such as59m left, and when a timer is all that's running, the status line reads, for example,Waiting for 1h timer · 59m left. - Improved rate-limit handling: when the model API asks the CLI to wait before retrying, the CLI now waits and retries for delays of up to 4 minutes instead of stopping the turn when the delay is longer than 30 seconds, so per-minute rate and capacity limits no longer end your turn, while spend-cap, prepaid-balance, and weekly quota errors now stop the turn right away.
- Improved
/modelwhen your plan offers more than one Orchestrator planner: the separateOrchestrator (Carbon)andOrchestrator (Argon)rows are now a singleOrchestratorrow with aPlannerchoice that you change with←/→orTab, while--modeland saved settings keep using the full names. - Improved switching to a conversation that's already open in long sessions:
/resume <id>back to it and the switch after/forkno longer load the conversation's whole history first, so they no longer pause for a second or more, and/btwand/helpside questions start faster too. - Improved memory use in long sessions: the CLI now keeps only recently used conversation steps in memory and reloads older ones from disk when it needs them, instead of holding every step for as long as the conversation is open, and it no longer loads every earlier subagent's full history at the start of each turn, which could take several GiB in sessions that ran many subagents.
- Improved click-to-expand in the
full screen (altscreen)rendering mode: items you can click, such as(ctrl+o to expand)hints, thoughts, expanded tool groups, and the pinned prompt, now highlight when you hover over them. - Improved the side panel in the Remote Control browser view: files and reviews open in tabs that come back when you reload the page, and a new Overview Panel, opened from the title bar or with
Cmd+Option+O(Ctrl+Alt+Oon Windows and Linux), lists the conversation's artifacts, modified files, subagents, and background tasks. - Changed the terminal sandbox to block sandboxed commands from reading
~/.config(or yourXDG_CONFIG_HOME) and, on Windows,AppData\Roamingby default, since other programs keep settings and tokens there; previously they were readable. The settings of common developer tools such as Git, Go, npm, and Python stay readable,~/.git-credentialsis now blocked, and you can still allow other paths explicitly in your sandbox rules. - Fixed Auto mode asking for approval to read files outside the workspace even when
Non-Workspace Accessis turned on in/config; Auto mode now honors that setting. - Fixed long, high-effort model responses being canceled and restarted from scratch when the model went more than 5 minutes without sending anything while it was still thinking; the CLI now waits up to 10 minutes before it treats a silent response as stalled.
- Fixed turns failing with
model output must contain either output text or tool calls, or the conversation compacting far earlier than needed, after the agent read very large files, such as multi-megabyte PDFs, before the conversation's first compaction. - Fixed a rare race that could leave a new conversation unusable, with every message failing with
trajectory not foundeven after a restart, when the conversation was created while the CLI was listing saved conversations, such as at startup or when an editor connected throughagy acprefreshed its session list. - Fixed
agy acpholding back what the agent did after a turn ended: when a background task or subagent finishes and wakes the agent, its replies, tool calls, and permission requests now reach the editor right away instead of at your next prompt, so the agent no longer waits on a permission request the editor never showed. - Fixed
agy acpleaving a session's subagents and background tasks, such as dev servers, running after the editor canceled a prompt or closed the session, where they could keep working and wake the conversation again; canceling or closing now stops them, and closing does so even when no prompt is running. - Fixed signing in to a plugin right after installing it from
/plugin, where pasting the code from its sign-in page could fail with "The sign-in link changed. Open the new one above, then paste its code." - Fixed plugins running while your organization turns them off with
pluginsDisabledin managed admin settings: plugin MCP servers and other plugin processes no longer start before sign-in finishes, and they now stop as soon as an administrator turns plugins off during a session. - Fixed skills whose
SKILL.mdhas nonamein its frontmatter: they were all namedSKILL, so only the first one loaded and the rest were silently dropped; each is now named after its folder, andagy plugin validatereports that name. - Fixed the agent writing LF line endings when it appends to a file that uses CRLF line endings, as is common on Windows, which left the file with mixed line endings and made later multi-line edits to it fail with
target content not found; appended text now uses the file's own line endings. - Fixed OSC 7501 terminal status showing the session as done while the agent was paused waiting for background subagents or tasks to finish; the terminal's tab or window now shows it as working until that work ends, and daemon tasks such as dev servers don't count.
- Fixed typing in the Remote Control browser view: with Chrome on Android,
Backspaceon the on-screen keyboard could stop deleting characters and later text could land in the middle of a word, and after typing with an input method or a phone keyboard's suggestions,Entercould add a new line instead of sending your message. - Fixed scrolling up through a long agent turn in the Remote Control browser view, where the view could jump back down instead of reaching your prompt.
- Fixed the permission prompt in the Remote Control browser view letting the number or arrow keys select and submit an unavailable "always allow" option when one of your ask rules or a hook requires you to confirm each time; the keyboard now skips those options, as clicking already did.
- Fixed the model selector in the Remote Control browser view switching to
No Model Selectedafter a brief connection or server error; it now keeps your account and model list until the next successful refresh.