Agent Changes, Send to many, Remote Redesign

Agent Changes — see which agent is changing which file, while it happens — open the new Agent Changes panel from the right dock or the top bar's Changes menu and the home screen answers three questions: what needs you, where two agents collide, and who is working on what right now.

Every version of every changed file is kept — Recent changes lists the newest version of each file an agent or you touched; open one to read its versions and put an earlier one back, with each version labelled by who wrote it (“You”, “Captured just before an agent wrote”, “Saved outside an agent”). A Map view draws the live picture, each line connecting an agent to a file it is changing, filtered to everyone or just the conflicts.

Pick a safety level per project — Off, Watch (“Show me who's changing what. Never gets in the way”) or Guide (“Also tells agents where others are working so they steer clear”). It is off by default, chosen per project and per working tree from the panel's setup screen or the File Explorer header, and the panel only reads until you pick a level.

No level ever blocks an agent or changes a file — Guide tells a cooperating agent what it is about to write over and lets it decide; the write still lands, and the version it replaced is preserved. Setup says how far it reaches with a “Works for N of M” pill marking which of your agents are told before they write and which are watched after the fact.

Two agents touched the same file? You get the call — conflicts open a review that explains what happened, offers each agent's version, a merged version and an optional AI referee's fix, and states plainly that nothing is deleted whatever you choose. Your calls keeps every decision you made and whether it can still be taken back.

Attention where you are already looking — a count on the top bar's Changes menu, a ring in the File Explorer on files an agent is working in, and a pill over the terminal when the agent in that tab changed files, or wrote over yours.

Storage you control — Settings → AI → Agent Changes shows what the vault keeps and what can be released, with a size budget you set and a Clean up button; open conflicts are never released, and a closed project's record is kept for 30 days before it goes.

Ask an AI to sort it out — hand a recorded conflict to an AI terminal with its context attached, instead of re-explaining it by hand.

Send to many — fan one prompt out from Agent Input — the new Send to many control opens a picker of every AI terminal across every project in the workspace; pick a crew, send once, and each target gets the same prompt after its agent is ready.

Cold terminals start on their own — a tab that was only saved is launched, waited on until its composer is empty, then typed into, and never reported as sent if readiness failed. Your crew is remembered on the toolbar badge, a partial send keeps the draft and offers a retry scoped to the ones that missed so a retry cannot double-send, and focus never gets yanked into another project's tab while the delivery runs.

Phone remote, redesigned around Agents — the phone opens on a kanban of every terminal across every project, grouped Needs you / Working / Idle, each card carrying the agent's mark, the project's avatar and what it is waiting on. Filter chips for status, agent or project carry live counts and compose with each other, plain shells sit in their own bucket, and dismissing a review card moves it to Idle until that terminal runs again.

Four tabs that match how you work — Agents, Projects, Team and Inbox; Browser and Tools left the tab bar and now live inside a project's detail screen, where they belong. Opening a project gives you everything it owns — its terminals, Start a team, git, files, resume, prompt history, browser and tools — without hopping through a global tab that then asks which project.

Start an Agent Team from the phone — write a goal, pick the project, review the plan the desktop authored, drop any members you do not want, then start; nothing spawns until you approve. A Viewer cannot start a team — the composer is inert when the device only has watch access, with the reason spelled out.

Find any terminal without the mouse — ⌘L — open the searchable list from the tab strip's new All N button too; terminals are grouped Running / Done / Idle / Hidden, each row showing its agent, worktree or machine, context use and its last line of output. Search matches agent, model, branch, account, even that last line; move with ↑↓, open with ↵, close with ⌘⌫, and picking a hidden terminal brings it back.

An overview of every terminal — ⌘⇧O — every terminal as a card with its latest output, status, model, worktree, account and context use; filter by state, click to jump, drag to reorder your tabs.

Switch project by number — ⍃1–9 (Alt+1–9 on Windows/Linux) — numbers follow the sidebar exactly as it is shown, groups and filters included. Hold ⌃ to see the numbers, then ↑↓ to browse the whole list, ↵ to open, and ⌃F to jump into the project search for the ones past the ninth.

Phone pairing asks what the device may do — a paired phone starts as a Viewer and the desktop asks “Allow this device to use this desktop?”; nothing beyond watching is granted until you choose a role and confirm, and dismissing the prompt, reloading or missing it all leave the phone read-only.

Four roles, with the powers spelled out — Admin, Operator, Approver and Viewer, each showing a live checklist of what it unlocks (watch terminals and diffs, approve agent requests, type commands and send prompts, edit files and commit) and a plain warning when the role can run commands on your computer. Several phones at once queue up one at a time, and Settings → Remote explains each device in a line.

Session inspector in Analytics — every session in range in one sortable table, filtered by title, project, agent or session id and sorted by start, span, tokens, tools, MCP calls or skills. Open a session for a “Where tokens go” split plus the tools it ran, the MCP servers it called and the skills it fired, with a pill that says how complete the scan was — and pick a second row to compare two sessions side by side.

“AI Diff” is now “Git Diff” — the panel, its settings, the menus and the feature list all use the name people actually say; it sits beside Agent Changes in the top bar's new Changes menu.

Analytics is back to four tabs — Overview, Agents, Sessions and Projects; session inspection moved into Sessions, and orchestration run health now lives in the Orchestration surface instead of a second place.

Analytics reads more of your history — Grok sessions are picked up from their own history file, Grok tool calls count toward MCP usage, Claude slash commands count as skills, and sessions in other formats are counted instead of being reported as skipped.

Oh My Pi, Kiro, Devin and Qoder terminals carry their own badge colour and glyph instead of the generic grey one.

Terminal tabs show the name, not a row of chips — status is drawn on the agent's logo, context use runs as a thin line along the tab's edge, and the model, account and token count move to the active tab's meta line and a hover card. Tabs shrink before the strip scrolls, keeping the start of every name, and ⌘1–9 hints appear only while ⌘ is held.

New shortcuts sit with their category — ⌘L and ⌘⇧O appear next to the other terminal bindings in the shortcut guide and Settings → Shortcuts, instead of landing at the bottom of a long-running profile.

Phone text stays readable in daylight — secondary labels and filled buttons on the remote UI meet contrast rules for body-sized type outdoors.