Skip to content

Latest commit

 

History

History
242 lines (207 loc) · 16.4 KB

13-feature-reference.md

File metadata and controls

242 lines (207 loc) · 16.4 KB

Complete feature reference

This is the compact inventory of current Remote behavior. “Page” means the loaded browser application; “host” means the Remote server; “project” means the durable workspace and its LXD container.

Navigation and project list

Feature How to use it Availability and persistence
Expand or collapse sidebar Use the chevron in the sidebar header Persists in browser local storage
Mobile sidebar Use the menu button; tap outside or close to dismiss Responsive UI
Search projects and chats Type in Search projects and chats; clear with × Filters names and visible groups
Project and chat counts Read the summary below search Live workspace state
Create project Choose the blue +, enter a name Any registered user; project persists
Reorder projects Drag a project group when no search is active Saved project order
Expand project group Use the project chevron Per loaded page
Inspect project Use the gear beside the project name Members and admins
Create project chat Use the + beside a provisioned project Members and admins
Provisioning and errors Read the disabled/spinning new-chat control or project error text Live workspace updates
Chat model and time Read each chat row's model badge and relative time Persisted metadata
Running indicator Spinner beside an active chat Live run state
Unread indicator Green dot beside completed unseen work Persisted lastReadAt
Mark read or unread Hover a chat and use the eye control Persisted
Fork chat Hover a chat and use the fork control New persisted chat
Delete chat Hover a chat and use ×, then confirm Removes chat metadata/history
Loose chats Appear under Unassigned Run on the host and are outside project isolation; avoid
Account settings and sign-out Use the account footer Current user/session

Chat composer

Feature How to use it Important behavior
Provider Choose Codex, MiniMax, Claude, Kimi, or Antigravity MiniMax is project-only; cannot change while streaming
Model Open the provider/model picker and select a discovered model or Auto Stored per chat; choices come from the current host/project CLI catalog
Refresh models Use the refresh action at the bottom of the provider/model picker Force-probes the current scope; use after CLI, configuration, account, entitlement, or terminal-login changes
Thinking Select one of the efforts reported for the current provider/model Hidden when no effort control is advertised; Kimi currently stores but does not forward the selection
Speed Select a service tier reported for the current provider/model Codex tiers and eligible Claude Fast are supported; account/provider may gate them
Mode Choose Default or provider-native Plan Hidden when Plan is unavailable
Approvals Choose when the agent must ask before an action Available for Codex and MiniMax; disabled while a turn is streaming
Sandbox Choose Read only, Workspace write, or Full access Available for Codex and MiniMax; controls filesystem/process isolation for the next turn
Skill picker Open Skill set, search, and select Catalog depends on provider/project
Skill chips Review or remove selected skills Cleared when provider changes
Attach picker Choose + and select one or more files Project chats; resumable uploads
Drag attachment Drag files over the composer Same upload path
Paste image Paste image clipboard data into the composer Same upload path
Attachment chips Monitor or remove before sending Saved file path is added to prompt
Prompt text Type the instruction; Shift+Enter adds a line Draft survives chat switching and same-tab reloads
Send Press Enter or use the arrow One active run per chat
Queue Send while streaming Per-tab sessionStorage; not a server-side job
Remove queued prompt Use × on a queue chip Before it auto-sends
Cancel Use the red square or Escape while running Cancels active provider context

The placeholder mentions @ files and / commands, but the current source has no implemented mention or slash-command picker. Use the attachment control and skill picker instead.

Conversation and messages

Feature Visible behavior
Streaming text Assistant output appears incrementally
Reasoning Consecutive provider reasoning deltas are grouped into collapsed, live-updating blocks that can be expanded at any time
Tool groups Consecutive tools are grouped and expandable
Specialized tools Read, write, edit, search, shell, and questions have tailored cards
Generic tools Unknown tools use a generic renderer
Markdown Headings, lists, links, tables, blockquotes, and code render in messages
Syntax highlighting Code fences receive language-aware highlighting
AskUserQuestion Agent questions become a paged single/multi-select form with Other
Usage Supported providers report accumulated token usage
Working state Header dot, provider label, sidebar spinner, and composer state update
Load older Complete older conversation turns page backward
Jump to latest Appears when reading above the newest output
Reconnect/replay Chat socket resumes from the last event sequence
Automatic title Remote can title a new chat from its early content; it is persisted, and the current UI has no manual rename control
Rewind Deletes the selected point and later events; next run starts fresh
Fork Copies visible history; provider-specific session fork happens on next run
Error block Run and transport failures render in the thread
Scheduled Tasks application Optional project app remembers work and wakes the same chat; its clock action manages tasks

Codex and MiniMax use the shared app-server interaction transport for approval, permission, and user-input requests. The selected approval and sandbox policies are sent when the harness starts or resumes a thread and again for each turn.

Providers and current differences

Capability Claude Codex MiniMax Kimi Antigravity
Sign-in Host authorization URL and pasted code Host device flow Host-managed write-only API-key form Host device flow Run agy in each project Terminal
Model picker Live /model list with attempted version resolution Live paginated app-server list Provider-owned MiniMax-M3 catalog Configured models from the provider catalog Models/variants returned by signed-in agy
Thinking control Forwarded Forwarded Think-Off or Adaptive Displayed/stored per model, not yet forwarded Forwarded as Auto, Low, Medium, or High
Speed/service tier Fast for Auto and Opus Yes No No No
Plan mode Declared native mode Discovered app-server mode Codex-harness native mode Advertised but incompatible with Remote prompt mode in the currently pinned Kimi CLI Discovered native mode
Approval and sandbox controls No Yes Yes No No
Usage telemetry Yes Yes Yes No No
Provider session fork Yes Yes, native app-server fork Yes, native app-server fork No; starts fresh No; starts fresh
Selected skill trigger Slash command Dollar mention Dollar mention Canonical SKILL.md instruction Canonical SKILL.md instruction
Browser MCP Yes Yes Yes No equivalent plumbing No equivalent plumbing
Structured tool stream Yes Yes Yes Yes No; plain streamed text

Antigravity's project-local /root/.gemini/antigravity-cli state survives stop/start and container replacement. After signing in with agy in the project Terminal, choose Refresh models so the picker replaces any signed-out fallback.

MiniMax's /root/.minimax model catalog and sessions also survive replacement, separately from /root/.codex. Its key remains in the project secret store.

Workspace tools

Feature How to use it Limits or lifecycle
Application actions Install a project application to add its chat actions Available only while the application runs in that project
Installable apps Use the Applications page in Remote App web routes check project membership and installation status
Application agent workflows An enabled app can start work in an existing chat, read its results, and offer its own commands to the agent Requires the app's declared capabilities, a running installation, and the owner's current access; uses the chat's existing agent settings
Open Terminal Install the Terminal application, then choose Container terminal Resizable pane; one shell per chat that survives closing the pane, reloads, and dropped connections; ended after 10 minutes with no viewer
Open History Choose History Git repositories only
Open Files Choose Files Lazy workspace tree
Open Browser Choose Open Browser Preview or Agent Browser
Refresh Files Use the drawer refresh control Reloads root
Expand folder Select folder row Lazy-loads children
Open file Select a file or search result Supported media opens in-app; an installed application can open source files, otherwise they download; unsupported media/archives download
Search filenames Type at least two characters 300 results; 200,000 visited-entry cap
Download file Hover a file and choose download Direct file response
Download folder Hover a folder and choose download ZIP up to 1 GiB; two concurrent
Inline media link Open a supported workspace link from chat Images, audio, video, PDF allowlist and full-screen viewer
IDE file link Open a validated workspace path from chat or Files Path-contained redirect with optional exact :line[:column]

The History button is hidden until Remote discovers at least one repository. It checks again after every completed run. The diff view groups files into collapsible cards with line-number gutters, hunk headers, change counts, and new/deleted/binary badges; unparseable patches fall back to raw text.

Scheduled tasks

Install Scheduled Tasks in the project, then ask the agent to remember work for a specific time or repeat it with an explicit timezone. Tasks start active. The application’s clock action lists, pauses, resumes, runs and deletes tasks. Every occurrence wakes a normal turn in the same chat. Stop/start retain tasks; uninstall removes them. See Scheduled tasks.

App preview and inspection

Feature How to use it Boundary
Discover apps Open Browser or choose refresh Non-loopback listeners, ports 1024–65535
Select app Use the process/port picker One preview target at a time
Remember mentioned URL Agent mentions a matching project preview URL UI prefers the recent matching URL
Resize drawer Drag the divider Persists in browser local storage
Reload Use the reload control Reloads selected frame
Open externally Use the external-link control New authenticated tab
Inspect element Toggle crosshair, then click the element Same-origin inspector wrapper
Insert element context Happens after selection Selector, text, HTML, bounds, styles, parents
Preview authentication Sign in to Remote Admin or project member
Preview hostname slug--port.dev.<host> On-demand TLS and known-project check

Agent Browser

Feature How to use it Boundary
Start human browser view Toggle the key control in Browser Starts project Chromium/noVNC as needed
Share login with agent Sign in visually, then select browser skill Claude, Codex, and MiniMax share the same profile/window
Human intervention Type or click in the live pane Same session as agent
Reload view Use reload while ready Reloads noVNC iframe
Close drawer Close Browser Stops only the human view
Stop complete stack Use the square stop control Stops browser core and view; keeps profile
Persist site sessions Reopen the project browser Profile lives in workspace
Automatic reaping Leave view and agent inactive Stops after roughly 20 minutes

There is one fixed 1366×768 browser session per project, unrestricted network egress, and no per-task browser isolation.

Git history

Feature How to use it Behavior
Discover repositories Open History Workspace root through depth 6 with exclusions
Select repository Use repository picker Shows path and dirty/clean state
List commits Select repository UI asks for 100; backend caps at 200
View commit Select a commit Shows subject, author, date, SHA
View diff Select a commit Structured per-file view; truncates at 768 KiB
Refresh Use refresh control Re-discovers repository state
Switch commit Select a commit and choose Switch on a clean tree Detached HEAD checkout
Safety checkpoint backend Checkout API accepts a checkpoint message Backend stages all, commits, then switches; current drawer does not render the required dirty-tree form

Project settings

Tab Features
Info Container/OS/resources/disks/network/mounts/limits/agent versions/auth-bundle state, refresh, network repair
Settings Effective and override limits; start, stop, force restart, admin-only delete
Secrets Add, reveal, hide, edit, and delete environment values, including multiline values
Sharing Add or remove registered project members

Resource defaults are 6 CPUs, 4 GiB memory, and 2,000 processes. Admins alone may change limits or delete a project.

Global settings

Tab Features
Appearance System, Dark, Light
Notifications Per-device Web Push permission, subscription, status, and test notification
Agents Module-driven admin cards: managed sign-in/status/refresh, external instructions, or no-auth status
Users Google OAuth configuration; add/remove users; member/admin roles
Updates Current release, update checks, and administrator-triggered update flow. Open pages reload themselves once the new frontend is live.
Info Host CPU, memory, disks, network, OS/runtime, process, paths, role

Persistence reference

State Persists across page reload Persists across container replacement
Project metadata and membership Yes Yes
Chat metadata and event log Yes Yes
/workspace Yes Yes
Provider homes Yes Yes
Project secrets Yes Yes
Agent Browser profile Yes Yes
Scheduled-task definitions and run state Yes Yes
Antigravity sign-in and conversation state Yes Yes
Container root filesystem additions Yes until replacement No
Active run control and event streaming No; an lxc exec child can remain alive but orphaned after backend restart No reattachment
Active Terminal shell Yes, while the Terminal application keeps running No
Composer draft Yes, in the same browser tab session Not applicable
Prompt queue Yes, in the same browser tab session Not applicable
Active chat and open drawers No Not applicable
Sidebar width/collapse, Browser drawer width, and Terminal pane width Yes, in the same browser Yes, in the same browser
Project-group collapsed state No Recomputed from unread state rather than stored

Features Remote does not currently provide

  • anonymous/public preview sharing;
  • read-only project membership;
  • per-user provider accounts or quotas;
  • enforced approval prompts for agent tools;
  • durable server-side prompt queues;
  • run reattachment after backend restart;
  • the visible Git dirty-tree checkpoint form, despite backend support;
  • project IDE membership enforcement;
  • built-in backup/restore, audit log, metrics endpoint, or high availability;
  • content search in Files;
  • offline access to the live app shell, chats, project data, or agent controls;
  • current application voice dictation;
  • implemented @-mention or slash-command composer menus.
  • a direct “create schedule” form in the current UI; schedule creation starts by asking the agent after installing Scheduled Tasks; its chat-header action manages the definition.

Read Known limitations and the Threat model before using Remote with mutually untrusted users or high-value credentials.