0.4.4 – Docs – StablEdge

StablEdge Desktop · Version 0.4.4 · Covers 0.4.0 – 0.4.3

These notes gather every update since 0.3.8 (public version 0.3.7.2), 0.4.0 – 0.4.3 included, in three parts: new capabilities, improvements, and fixes.

1. New capabilities

Ways of working

Focus mode: Chat has a new Focus option. Choose the display, app window, terminal or log file to watch, write down what you need done and select Start focus: when the screen changes, a command fails or a log line goes wrong, the Agent handles one round and gives you a summary, and stays quiet when nothing needs doing. You can box the area that matters, mask the clock or other moving parts, and switch a display to a plain background for the duration. With Allow execution turned on, the Agent can fix things in the terminal itself, while high-risk commands still need your confirmation every time. Focus results can go to the Lark, DingTalk or Telegram channel you have connected, and from your phone you can check status, end a session, or approve a high-risk action with a one-time code.
Teams (Preview): Define roles in the role library and tick the tools each one may use, put together a team of a Lead and Members, and open an Office for them to work in. The Lead splits tasks and assigns them to Members, who can ask questions, consult each other and work in parallel; you can also @ a Member in the group chat to assign work directly. The task Board has columns for to do, awaiting review and accepted. Each role has its own Memory and private material and can be given a per-round usage budget; three preset teams ship with the app for code, product design and graphic design.
Multiple windows and split panes: Open a new window from the title bar, each with its own project, and split the chat area into panes left and right to push several chats forward at once.
Building tasks inside a chat: Chat has a new Tasks option: describe what you want in plain language and the task is set up in Task Center. The task editor gains a run panel, a Remote delivery node and parameterized test runs.
Island mode: Shrink the main window to a small ball on the desktop: click it to open the chat capsule, double-click to return to the full workspace, and it tucks itself away when you click elsewhere. Click Fix at the head of the capsule to keep watching output. When a reply arrives or something needs your confirmation, a notice slides out beside the ball, and one click opens it. It can be called up from the taskbar, Dock or tray, and stays visible over full-screen apps on Mac.
Launch at login (Windows): Turn on launch at login in Settings; after starting, the app stays in the tray without opening the main window.

Agents and subagents

Working with subagents: Subagents run in the background by default and wake the chat when their result arrives. You can add requirements while one is running, and it can ask the main Agent questions in return; after delivery you can talk to it directly in its tab. Research subagents deliver a Markdown report, a new quick-lookup subagent answers with the conclusion only, and each subagent can use a fast, standard or deep model tier.
Writable subagents edit code in parallel: Each writable subagent works in its own Git worktree, so they never overwrite each other and several can run at once; the main Agent reviews and merges, then writes everything back to your workspace in one step that can be undone as a batch. Projects without Git can use this too. When a project folder holds several repositories, you can point each subagent at one, and several repositories can be changed and written back separately from the same chat.
Agents talking to each other: An Agent can send a message to the Agent in another chat and reply to messages it receives, without that chat being open; an incoming message shows its source chat in a notice of its own.
Agents that open chats: An Agent can open a new ordinary chat and hand it a task; the new chat starts working and can write back when it is done. Every creation asks you to confirm first. The new chat opens a tab in the current pane without taking you away from the chat you are reading, and you can choose its project (or global) and its model.
Automatic failover when the main model fails: When the main model keeps failing, runs out of quota or goes offline, the turn continues on the fallback model chain from Settings, with a notice in the chat, and the next turn goes back to the model you chose. While the network is down there is no retry and no failover; work continues once the connection is back.
Background context cleanup: Context is condensed in the background before it fills up, so the chat no longer stops to compact, and manual compaction runs in the background too.
Cloud text knowledge base: Browse, search, annotate and edit enterprise, department and your own text knowledge bases by folder, and the Agent can search, read, create and change knowledge base documents, showing you a diff to confirm before it writes.
CI/CD: A CI/CD drawer beside the chat shows GitHub Actions and GitLab CI runs, logs and artifacts, and you can start, re-run or cancel them by hand; the Agent can also look at a failure and deal with it.
Git status bar: Click the branch to switch, create or track a remote branch, commit, push, merge and resolve conflicts, and connect a code hosting platform; the commit panel groups files. You can open a branch in a new worktree and work on different branches from several chats at once, and opening a folder no longer creates a repository on its own.
Changed files bar: A summary above the input box lists the files this chat changed and the lines added and removed, and you can keep or undo them one by one; nothing is committed to your repository.
A separate background browser for the Agent: The Agent browses the web in a background browser of its own instead of disturbing the page you are reading; you enter the account and password in a card when a site needs a sign-in, and the password is kept encrypted on your machine only.
Terminal: Commands show their output as it arrives, long commands can wait in the background, the Agent can read terminal logs when you allow it (SSH remote terminal logs included), and you can send selected output to the Agent as a reference.
Creating and reading images: The Agent can generate images (transparent backgrounds included) and keep editing them, and it can read images in the chat and in the workspace.
Enterprise Skills: Ability Center gains an Enterprise section for department-level and enterprise-level Skills, and you can publish Skills of your own.

Chat and input

Drag things into the input box as references: Past chats in the sidebar, chat tabs at the top, and page and file tabs can all be dragged into the input box, and a sidebar chat's context menu offers Quote in new chat.
Keep sending while a reply is being written: Messages queue up and go out one after another when the current turn ends; you can edit them and change their order, and the queue survives a restart.
Chat groups in the sidebar: Chats in a project can be gathered into groups with an icon and a colour, and you can drag to create groups.
Tidying chat tabs: Tabs you are not using fold into a stack beside the search box and open when clicked; running chats and chats with unread replies are never folded away.
Input box improvements: The attachment button adds files or folders; a new chat can pick or create its project beside the input box; drafts are saved automatically; Skill and @ mention tags survive drafts, queue edits and pasting.
Take back a message mid-generation: While a reply is being written, click your own question, the one still without an answer, to withdraw it and cancel the turn; the text goes back into the input box.
Sidebar projects can be dragged into a new order.
Folding parts of a reply: Exploration and tool calls fold away, and tool activity shows a short label.
Markdown: Math is supported; table columns can be widened and tables exported; the Markdown preview highlights annotations, discusses them in threads and sends the selected text to the Agent.
One way to open files and links: Paths, files and URLs in a reply all open, with a choice of built-in preview, the editor, VS Code, Office or the system browser, and the menu shows real app icons. Settings has a new Open with page where you set a default app per file type, or choose another app for one occasion; files on cloud drives open once downloaded.
Clearer model choice: Automatic model selection is gone; Claude models let you choose how long prompts are cached; model slots such as the main model and subagents each have their own reasoning effort.

Memory, rules and settings

Memory and prompts: Settings manages long-term Memory in one place, split into deep Memory that comes along every time and experience Memory that is looked up when needed.
Memory for one chat: Memory can apply to the current chat only and survives compaction; review it in a panel and promote it to project Memory in one click, and each Memory card shows which level it lives at.
Project rules you can edit: AGENTS.md at each level can be read and edited, and when the Agent wants to change a rule an approval card appears.
The Files settings page: Export chats and set the files the Agent may reach.
Secret environment variables: When the Agent needs an API key a card asks you for it; the value is kept in the system keychain only and injected while the command runs, so it never enters the model context.
Data recovery and upgrade backups: Older data folders can be scanned and merged back into the current account, and storage is backed up automatically before an upgrade.
Longer sign-in and several organizations: Choose 7 or 30 days without signing in again, and pick which identity to use when several enterprise accounts are signed in.

Channels and plugins

Channels: DingTalk can send and receive images, files and media; replies stream in Lark; !new, !model and other shortcuts work in chat; files the Agent makes are sent back to the chat automatically; each remote connection can use its own model.
Plugins: A plugin's detail page lists its Skills, workspaces and background programs with their permissions, which you can revoke at any time; This-computer mode can be turned on; the PC control plugin ships with the app; plugins without an interface can be attached to a chat or kept resident.

Interface

A frosted-glass redesign of the chat area; a choice of interface font and size; custom workspace panel layouts; top search that reaches abilities and settings; a new onboarding tour; a sidebar that folds into an icon rail in a narrow window.
Window background and low-graphics mode (Windows): Three levels of window transparency, and a new low-graphics mode that turns off the frosted glass, large shadows and decorative animation for weaker graphics or a stuttering interface.

2. Improvements

Faster and cheaper: Much more context can be reused in long chats and condensed summaries hit the cache too; tool output is leaner, and very long output is saved to disk where it can be searched.
Smoother long chats: Rendering by window and a rewritten streaming display make scrolling and generation smoother, and resource use drops close to zero when the window is idle or unfocused.
Faster lists and file tree: The file tree no longer stalls when you come back to the window, long lists load in batches, and requests no longer block one another.
Cleaner tool cards: Tool cards no longer show the JSON and correction prompts written for the model, and failures are judged consistently.
Incoming messages take less room: Messages from other chats arrive as a single collapsed line that opens to the full text, and expand automatically when you search the chat.
Pasting very long text no longer stalls: Text over ten thousand characters or 300 lines is saved and sent as an attachment.
Context menu: The browser's own Back, Reload and Print menu no longer appears (a stray Reload reloads the whole interface), and the input box offers Cut, Copy, Paste and Select all, plus Copy for selected text.
Fewer tool call mistakes: A wrong argument now comes with an explanation you can act on, and common formatting mistakes are corrected automatically.
The Agent follows project rules more accurately: The right rules load when it works in a subfolder or a remote workspace, and fake system prompts in pages, files or command output are no longer treated as system content.
Steadier under rate limits and very long context: On a rate limit the Agent waits for the time the server suggests and then continues; an error before the first output switches to a non-streaming retry; an over-long context is condensed and retried. Background cleanup keeps the most recent content within the budget actually available.
Lighter on resources: Resizing the window turns the load down, editing a sent message no longer stutters, the storage panel skips node_modules, target and .git when it measures a project folder and pauses while the window is unfocused, and logs are written in the background.
Deleting a project no longer deletes local or remote files; the confirmation card's wait can be set between 60 and 360 seconds; clicking a Windows notification returns you to the chat it came from.

3. Fixes

Fixed tool arguments being lost on delete-style edits, which sent the Agent into a retry loop; fixed long output being rewritten over and over after truncation.
Fixed intermittent crashes from a cross-thread race in the window event loop; node, rg and similar processes no longer linger after the app exits; running browser scripts and reading file icons no longer freeze the interface.
Fixed the input box losing focus after Alt+Tab or after switching back from the taskbar.
Fixed settings such as sidebar groups and the annotation send preference living in memory only and being lost after a restart or sign-out.
Fixed reasoning being swallowed by folding during streaming, and content you expanded by hand folding again when the answer finished.
Fixed answers to a withdrawn question being lost, and right-click Paste behaving differently from Ctrl+V.
Fixed closing from the taskbar or with Alt+F4 quitting outright when the close action was set to minimize to the tray.
Fixed tool cards folding away during streaming and pulling the whole view down; fixed the selection toolbar in the preview being scrolled out of view by the chat area.
Fixed the capsule in Island mode not tucking away on the first click elsewhere after it opened.
Fixed a subagent failing to create a sandbox in a folder that holds several repositories, which reported an oversized snapshot.
When the network is not ready at startup the app now retries the connection instead of reporting that it cannot connect; a removed or deleted repository no longer throws Git status errors on every refresh; fixed a background error after closing the mini diff in a chat.
Fixed subagents stuck in a running state after a crash or restart, and leftover cards when the stop button was used.
Fixed knowledge base read and write problems where the Agent's tool read only some sections but wrote over the whole document, and versions no longer matched after an edit; a document changed by someone else at the same time is now reported clearly.
Fixed editing a message and sending it again wiping the whole chat when the original message could not be found; text put back after a withdrawal now keeps its Skill tags and draft state.
Fixed manual compaction doing nothing; failed compaction, Memory operations and tool calls now show Failed and expand to the reason.
Fixed a large remote GBK file being converted to UTF-8 when written back; fixed content containing NUL characters failing to send; fixed repositories with very long paths failing in subagent sandboxes and in the Agent terminal; ~/ paths in replies now expand to your home folder.
Fixed the input method candidate window drifting after the window moved or changed display; fixed some window actions doing nothing while a browser tab was open; fixed history appearing blank after switching back to a chat tab; fixed the capsule in Island mode tucking away while the Agent was using the computer.
Fixed sidebar group cleanup affecting other projects and chats disappearing from their group after a restart; fixed a branched chat not inheriting the original chat's confirmation level.
Fixed the wrong model name on replies after failing over to a fallback model; fixed a CI/CD authorization card appearing when no code hosting account was connected.
Fixed a bundled plugin not installing again after uninstalling a same-named plugin you had installed yourself; fixed a numbered list following a bulleted list in the knowledge base editor starting from the wrong number; fixed copying selected text failing now and then; the login page background animation no longer reloads over and over and uses up resources; upgrades clear debug logs left by older versions.
Fixed sending and receiving images and files in WeChat; fixed a plugin's Skill name being cut off at a dot.
Fixed the PowerPoint preview canvas shrinking after zooming in; fixed the model-change divider scrolling out of view after sending; a site icon that failed to load is fetched again on a different network.
Fixed connecting failing when the SSH private key path was entered as ~/.ssh/...; display screenshots on Windows are steadier.
Fixed existing chats and Skills disappearing after an upgrade or after switching enterprise; fixed clicking your own message withdrawing a question that had already been answered.
Fixed messages sent in an older chat being swallowed, and an empty image when a pasted image was sent straight away; fixed chats started from a channel erroring out before they reached the model.
A model timeout, a dropped connection or too little quota now says so clearly, and network problems that can recover are retried automatically.
On Windows, processes running from the install folder are released before an install or uninstall, so a locked file no longer breaks an upgrade; fixed the frosted effect sometimes not applying on Windows 11; a failed start now explains why and offers to restart.
Fixed past chats not being found in search, errors when uninstalling a Skill, variable substitution in task nodes, and interface problems with window resizing, drag drop targets, banding in dark-mode dialogs and covered dropdowns.