Daily updates, newest first. Choose a date to explore what changed.
Versions released this day
macOS v0.5.210 · Windows v0.5.210
macOS v0.5.209 · Windows v0.5.209
New
Create sessions directly in a project, pin or rename projects, and reveal their folders.
Improved
Open the app in your last workspace and choose another from the account menu.
Smaller project actions appear when you hover over a project.
Add members by phone while keeping the current workspace name.
Billing follows the same responsive content width as other settings pages.
The sidebar distinguishes selected, running, unread, waiting, and failed sessions while preserving scroll position.
A more aligned model menu, with edited messages kept visible while rerunning.
Fixed
Locally processed images can be delivered as final results, keeping internal candidates and dependency caches out of deliverables.
Default and session models restore from their own records after account changes or sign-in.
Clearer knowledge content, recently used spaces, localized dates, and team member navigation.
Windows improves dependency installation under Chinese paths and handles read-only bundled files during uninstall.
Versions released this day
macOS v0.5.208 · Windows v0.5.208
New
Add DingTalk and WeCom office skill suites from the market, with pagination for installed skill groups.
Ask Torra to create, edit, and run workflows, then continue from saved progress.
Improved
Clearer sign-in actions for office services and a more compact layout for channels and plugins.
Per-turn costs are hidden in chat and remain available in billing.
Fixed
Imported skills retain headings and content without accumulating escape characters on repeated saves.
The knowledge library recognizes bundled components, and internal communication skills show clearer descriptions.
After moving a session to another project, its file list and previews keep using the session's original working directory.
Windows uninstallation continues past a locked file while preserving user files.
When parallel tasks finish, their follow-up task continues without a false decision prompt.
Versions released this day
macOS v0.5.207 · Windows v0.5.207
macOS v0.5.206 · Windows v0.5.206
macOS v0.5.205 · Windows v0.5.205
macOS v0.5.204 · Windows v0.5.204
New
Manage WeChat, Feishu, WeCom, and DingTalk under Plugins → Message Channels, with cards that open the connection flow directly.
Improved
Image revisions stay internal. At the per-request iteration limit, Torra delivers the best available result.
Automatic model routes adapt to rejections seen on this device and retry later.
Automatic model routing remembers temporarily unavailable services and retries them later, reducing repeated failures on the same route.
Image revisions stay internal instead of appearing as final results. When the iteration limit is reached, Torra delivers the best available candidate.
Set a default assistant for each project. Project conversations use the current project's choice and update when you switch projects.
Conversations show confirmed model charges by turn. If a charge is interrupted, Torra checks the billing record to avoid duplicate charges.
Image revisions keep the original for comparison so the result can be checked against its appearance.
Fixed
Fixed scheduled-task failure and delivery records, and clarified wallet balance errors in media tasks.
Fixed the default agent appearing in the new-session composer.
Fixed scheduled tasks using the current account's authorization and corrected outdated navigation guidance.
Fixed scheduled tasks that could fail to use the current account's model authorization.
Scheduled runs now clearly show when work or delivery did not complete.
Shift selection works again across chat replies, making it easier to select several passages.
Fixed an issue that prevented some workbench websites from building again after source changes.
Idle browser connections no longer stall opening or saving workbench websites.
Fixed a Windows installer error that could appear when the installer window opened slowly.
Versions released this day
macOS v0.5.201 · Windows v0.5.201
New
Workspace managers can add members by their registered phone number, without a separate invitation acceptance step. Unregistered numbers show an account-not-found message.
Improved
When an image task has no deliverable result, Torra attempts recovery and checks again within the existing permissions and task limits. If it remains incomplete, Torra says so.
Image results and the file library show the images selected for final delivery, rather than intermediate candidates. Existing files are preserved.
Created groups appear in the session sidebar. Choose a project at the top of a conversation, or use Move Project in the context menu separately from grouping.
View the built-in browser’s page activity in your conversation. The Chrome connection guide shows extension setup steps and the actual connection status.
Consecutive internal operations appear in expandable activity groups. Failures and requests for confirmation remain visible separately.
The knowledge library has a clearer list and reading layout. In narrow windows, read the content first and then return to the list.
The chat plus menu now includes an entry for Feishu and messaging channels. Duplicate download buttons in PowerPoint previews have been consolidated.
This Mac release requires macOS 13 or later, with separate installers for Apple Silicon and Intel Macs.
Fixed
File-library listings and previews follow the current workspace when you switch, preventing files from another workspace from appearing.
Scheduled tasks retain a failure status in the conversation and run history when execution or delivery fails, instead of leaving blank feedback or reporting success.
Fixed an issue that prevented some local web builds from running again after their source errors were corrected.
Versions released this day
macOS v0.5.200 · Windows v0.5.200 / 0.5.199
macOS v0.5.198 · Windows v0.5.198
Improved
0.5.200 uses Torra’s existing gold for input focus and list selection, with a subtle selected background.
0.5.200 for Apple Silicon Macs and Windows groups the file library, knowledge library and workbenches under Space, with navigation on the left. More moves to the end of the main navigation.
0.5.200 removes the duplicate knowledge-library entry from Settings. Existing files and features remain available through Space.
From Windows 0.5.199 and Mac 0.5.200, file links are easier to distinguish from body text, long deliverable lists can be expanded or collapsed, and common dialogs use a soft gradient background.
Mac 0.5.200 for Apple Silicon updates the built-in updater to support larger future updates. Older releases continue to use the in-app update entry.
On Apple Silicon Macs and Windows, recharge entries open the billing page in Settings. The wider dialog also scrolls on smaller windows.
The image-model selector is removed from the chat composer. Model selection remains available in Settings.
Fixed
Available first in Windows 0.5.199 and included in Mac 0.5.200: newly generated deliverable lists exclude internal cache files.
Mac 0.5.200 fixes missing or blank Chinese text when rendering some Office files as PDFs or previews.
Versions released this day
macOS v0.5.187 · Windows v0.5.187
macOS v0.5.186 · Windows v0.5.186
New
Discover MCP Services in Skills and Connectors lets you add Microsoft Learn and DeepWiki, then use them in conversations to look up public technical documentation and open-source repositories.
Local personal workbenches let you search, add, edit, and delete records. The app and business data stay on this computer.
Add your role, field of study, and commonly used tools to the current account for more relevant skill suggestions. You can edit, clear, or disable these preferences in Settings.
Improved
The limit for a single chat attachment is now 50 MB. Files over the limit are clearly identified and are not submitted.
The composer + menu now groups files, agents, skills, and connectors. The selected agent appears in the composer and can be removed directly.
Project controls on the home screen and in the conversation header are clearer, including the current project when choosing a project or moving a conversation.
Type size, weight, and spacing are refined across conversation lists, question cards, approval cards, tool activity, and file lists for clearer hierarchy.
Windows installation and in-app update windows now share a consistent visual hierarchy, with progress, details, and actions kept together.
Versions released this day
macOS v0.5.183 · Windows v0.5.183
New
Describe a scheduled task in plain language, then confirm its timing and action before creating it. Manual entry remains available.
Skills include version history, with controls to accept or reject proposed changes and restore an earlier version.
Improved
The image and video model lists include the latest flagship models from each provider.
Settings, prompts and task creation dialogs use consistent buttons and input controls.
Fixed
Windows installation no longer opens the English verification window first. Confirming a nonempty folder allows installation to continue.
Versions released this day
macOS v0.5.182 · Windows v0.5.182
macOS v0.5.181 · Windows v0.5.181
New
Files in the file panel now have type-specific icons and colors; images, videos and audio preview right on the card, and long names keep their ending with the full name on hover.
While an image is generating you now see the model, which image it is, elapsed time, and how long that model usually takes.
Mac: scheduled tasks can run in the background without opening a Torra window and quit when done (a sleeping Mac still can't run on time).
The trend chart on the workspace home now shows daily token usage for the last 30 days.
Name a model and it is used: say "generate with Grok" and Grok is used directly. Models that aren't available are reported with the list of ones that are.
After a scheduled task catches up, its run record says how many minutes late it ran.
Meeting-minutes transcripts can now be imported straight from Word and PDF files.
Improved
Group headings everywhere (model picker, command palette, sidebar, settings, skills, file panel…) now use one clear style: no all-caps and no stretched letter spacing.
Downloads are smoother: packing finishes quickly, the progress notice is dismissed before the save dialog appears and no longer spins after cancelling, buttons are disabled while downloading, and the file panel defaults to a list.
The nightly skill check is now off by default; you can turn it on per account in Settings with a nightly cap, and it only checks your own skills.
Each bill line now says where it came from (your message, scheduled task, background follow-up, channel auto-reply, memory upkeep…); the old "won't change balance" text now explains each type; the top-up dialog says which account the money goes to.
Every automatic path that can call a model without you typing (background follow-ups, old scheduled tasks, channel messages, startup memory recovery…) now has a retry cap, and background helper calls wait at most 60 seconds.
The completion check can now see the finished image, so it no longer asks for rewrites just because it can't see the picture.
Typography was revisited across the app: font, size, weight, text contrast, letter spacing and line height are more consistent and easier to read.
Downloads are simpler: clicking Download opens the system save dialog; cancelling leaves nothing behind, and the zip progress notice shows once per click and disappears on success.
The file panel now defaults to a list instead of cramming everything to the left.
Chats that finish in the background are marked unread, the list is ordered by last activity, and relative times keep updating.
When you ask for Chinese text on an image, the result is checked character by character, and mismatches are reported honestly instead of claiming success.
Image generation is steadier and faster: results are picked up sooner after the provider finishes, even on a poor network route.
Skills that run scripts (such as Office documents) no longer fail on the first run because they couldn't read the skill library.
Fixed
After a successful image, the same sentence was repeated three times followed by a false "not complete" notice.
Two images from one request showed only one, and the card's model label showed your default instead of the model actually used.
The notice after an engine restart showed both languages side by side with a raw timestamp.
Explanatory "nothing here yet" texts on empty pages were removed, keeping only the needed action buttons.
Clicking Library in the sidebar no longer opens the page in your browser.
Fixed duplicated replies, repeated runs, and documents from other chats being treated as this chat's material.
After answering a question card, switching away and back no longer asks again.
MCP servers added in Settings no longer fail on first use until a restart.
Local dictation now writes Chinese in Simplified characters.
Billing now updates right after a charge, chat tabs no longer leak between accounts, and nickname and avatar stay in sync.
Background follow-ups that keep failing could be retried and charged repeatedly; they now stop after 3 failures.
Fixed slow spreadsheet previews, preview arrows covering content, unclickable file links in chat, and unclear state after adding a skill.
A brief freeze when starting a new chat.
Versions released this day
macOS v0.5.180 · Windows v0.5.180
macOS v0.5.179 · Windows v0.5.179
macOS v0.5.178 · Windows v0.5.178
New
Image, video and audio cards whose download did not finish now have a Download again button, free of charge. You only need to generate again if the original file is no longer available. Progress is shown while it downloads.
New accounts get a set of official featured skills installed on first sign-in. Skills you remove are not added back.
Improved
All chats now stays at the bottom of the sidebar's chat section, just above your account, and is always visible.
The image model switcher is back under image cards. When one turn produces several images, it is shown once rather than on every image.
Video generation prices have been recalibrated against the provider's official price list, and Grok video is now priced per second.
The billing view grouped by task now covers the last 180 days instead of 30.
Knowledge base: when a file in your folder changes, its previous content stays searchable until the new version has been processed. If processing fails, the previous content is kept and Torra retries later. The same applies to new versions of uploaded files.
Business graph: customers, suppliers and contacts mentioned in a file you explicitly ask Torra to read are added to your business graph. Files you did not ask it to read are left out.
The model picker no longer shows image / text-only and price-tier badges.
When a scheduled task's timing is written incorrectly, the message now says what is wrong in your interface language instead of showing raw program output.
If the dictation model download stalls, the message now says it is most likely a network problem: check your connection and click again, and the download resumes where it stopped.
The workspace page has been redesigned: its stat cards match the Usage page, and a new chart shows chat activity over the last 30 days.
Removed unneeded explanatory text from the sign-in page, the empty chat list, the cloud sign-in expired notice and the grouped transactions view.
Fixed
If you switch to another chat or the project home page while an image is being generated, the chat no longer goes blank when you come back, and the image still arrives.
When a new message interrupts a turn, images, videos and audio that were already generated and paid for in that turn are posted back to the chat instead of being lost.
Image and video downloads that are slow or break off now resume from where they stopped instead of giving up partway.
When Grok generates two images at once, both are delivered.
Grok image generation no longer fails at ratios such as 3:4 and 4:3. Torra switches to the closest supported ratio instead.
A question card that is waiting for your answer no longer asks again after you switch away and come back.
After switching the interface language you can switch back without restarting.
If you change the model while a question card is waiting and then send a message, the new model picks it up instead of hanging.
When you ask whether the current model can see images, it now answers truthfully.
When image generation succeeds, replies no longer end with a mistaken ⚠️ correction note.
Excel previews in chats now show formatting such as header shading and auto-filters.
Importing skills from GitHub now works with skills.sh links, github.com addresses without https, and links that point directly to a SKILL.md.
Skills that were in the wrong group in the skill market are back in the right one.
After you change your nickname or avatar, the sidebar, top bar and workspace all show the update. If you set a nickname, it is shown.
Times such as “updated X hours ago” are no longer off by eight hours.
Clicking Download in the image viewer now opens a save dialog so you can choose where to save, instead of appearing to do nothing or saving several copies.
Closing the Download as zip notice now cancels it instead of letting packaging continue in the background.
When a slow network made a new chat look like it had failed even though it was created, Torra no longer shows a failure or sends you back to the home page.
After generating an image, the next message reuses the previous turn's cache instead of starting from scratch each time.
On Mac 0.5.177, every message failed because the chat could not be found. This is fixed.
An occasional failed background refresh no longer closes and reopens the chat you are working in.
The setup steps after your first sign-in load and save again.
Business graph: cleaning up wrong links such as every image being linked to Torra now removes only those wrong links. A customer or supplier in your business that happens to have the same name is left untouched.
Versions released this day
macOS v0.5.135 · Windows v0.5.135
macOS v0.5.134 · Windows v0.5.134
macOS v0.5.133 · Windows v0.5.133 / 0.5.132
macOS v0.5.131 · Windows v0.5.131
New
Business graph: before Torra starts a Feishu approval for an item your business graph already tracks, it checks the current approver, the item's status and its supporting materials. If any of them falls short (for example, no approver is recorded, the item isn't in a ready-to-submit state, or materials are missing), the approval isn't started and Torra tells you what's still needed. If the records disagree on the approver, just say in the conversation who it should be: Torra records it, checks again and carries on. It only accepts a person you named in your latest message and never picks one for you.
Business graph: rules saved by earlier versions that may not be your company's own (for example, the original wording was a one-time request or the other party's requirement) are marked Needs review in the list. Open one and choose Yes, it's ours or No, stop applying it, and Torra follows your answer when deciding whether to apply it in new chats.
WeChat: images and files you send to WeChat ClawBot are now read by Torra once you confirm the task, so you can ask it to sort out a PDF you just sent. Unsupported content such as video and attachments that fail to download are called out, and if you sent only attachments and none could be read, the task is not run and the reply tells you how to resend.
The file panel now previews PowerPoint (.pptx) decks slide by slide without downloading them first; very long decks show only their first slides, and you can download the file to see the rest.
Below Recently finished, the home page now shows Runs next with the scheduled tasks due soonest; click one to open Schedules.
If you haven't signed in to your Torra cloud account, a corner notice asks you to sign in once when you open the app: Sign in opens Settings › Manage › Account, Later dismisses it, and you can keep working locally either way. It stops appearing once you have signed in.
A redesigned home page: the input box now matches the width of a conversation, with For You and Worth a try side by side below it. Further down, In progress and Recently finished list running tasks and recently ended conversations one click away, and sections with nothing to show stay hidden. Shortcuts sit in tidy columns.
Conversations started by a scheduled task or webhook now have an “Edit this scheduled task” (or “Edit this webhook task”) button in the header that opens the task directly.
Improved
When the agent uses the managed browser, it is now explicitly told to open each new page in the current tab (a new tab only when two pages must stay open together) and to read pages from their structure, taking a screenshot only when it needs to check how a page looks, to cut down on repeated new tabs and unneeded screenshots.
When the Live meeting assistant can't reach its speech service, the notice now says plainly that the service is outside mainland China and needs a VPN, and that the rest of Torra works without one.
Business graph: Graph view now fetches the relationships for the whole graph in one request instead of entity by entity. When there are too many to show them all, a “Showing some of the relationships” note appears in the top-left corner, and relationships that have ended are drawn more faintly.
Hosted models now connect straight from your computer to the model provider: following your Route setting, the China-first route prefers a domestic provider and the International-first route an international one, and your conversations no longer pass through Torra's relay servers. If one provider fails another is tried, and every call is billed as before.
In Feishu groups, only messages that @mention the bot become tasks; other group messages no longer get a reply.
WeChat replies are clearer: when an image fails to send, the reply says which step went wrong and points you to Torra on your computer; generated images given as links are sent as images, other links to files on your computer become a note to view them there; and a message sent while the previous one is still being handled is acknowledged as queued behind it.
Channel settings: WeChat's Last disconnected reason is now a plain explanation (such as a QR code that expired several times or a scan timeout) instead of an English engine message, and the WeChat card now matches the note at the top: messages to WeChat ClawBot run only after you confirm them in Torra. The Feishu setup guide drops internal testing wording, adds the permission needed to read images and files, and notes that only @mentions of the bot are handled in groups.
Business graph: conversations with a single exchange, and the final exchange of a conversation, are now learned once the conversation goes idle, and quantities, deadlines and phrases like “this batch” are no longer treated as entities.
The model list is tidier: with a single group there is no group header or provider suffix, and Auto fits on one line that names the model it prefers.
Before reporting a task as done, Torra checks the files it produced and goes back to finish any that are empty or hold only placeholder content. A search that found nothing along the way no longer tacks a false correction onto a correct answer, and a file with the same name elsewhere in the workspace no longer triggers a false “the file was not changed” warning and an extra round.
Companies, terms and similar entities in uploaded files can now be learned into the business graph, labeled “Seen in a document”, and requirements such as “must” or “must not” written by the other party in an attachment are no longer treated as your own company's rules.
Skills › Installed is now a categorized list with a description for each skill; the old folder tree moved under “View source files”.
Reference costs in My usage, Gateway and team usage now explain that they are estimated from each model's public price, for reference only and not what your wallet is charged, with a link to your billing. The requests column is now called Model calls, and the message count is now Chat messages.
Attachments over the upload limit are handled better: PNG, JPG and WebP images are compressed automatically first, and if a file is still too large the message names it, nothing looks sent, and your text and attachments go back into the input box.
Account pages are clearer: Settings › Manage › Account shows only your cloud account's name and phone number, with the billing tier as Standard, Top-up or Enterprise instead of a raw code; the local account settings page no longer shows an unusable balance and credit ledger and points you to Settings › Manage › Account instead; and the Members page hides email invitations and other options the local edition does not support.
When a task arrives from a channel such as Feishu while you are in a conversation, Review jumps straight to that conversation and a notice you close with × stays closed; new messages in a conversation you already reviewed now show a notice with Open session instead of just a small green dot in the sidebar.
Pending channel messages show what the sender actually wrote instead of a message ID, failed replies explain why in plain words (with the platform error code for Feishu), and channel notices and system notifications for finished, failed or permission-needed tasks follow the interface language.
macOS: a downloaded update is now prepared for installation in the background right away instead of only when you quit (this takes effect from the update after 0.5.133).
Pages in the Customize panel — Agents, Skills, Memory, Business graph, Connectors, Files, Orchestration, Meetings, Members, Account and more — share a new layout: in-page tabs live in the page header and lists sit beside their details. The Files page drops its redundant location bar, Connectors is split into browser, MCP and related capabilities, and Business graph switches between a list and a graph.
The source label in meeting details now follows the interface language instead of always showing “audio”.
Fixed
Pages in the managed browser panel are now rendered at the panel's size and your screen's pixel density, so they're no longer blurry on high-resolution displays and no longer have black bars above and below.
In Runs next on the home page, each task now shows when it will run, such as “in 5m” or “tomorrow 08:00”, instead of always showing “now”.
macOS: when you quit Torra, install an update, or Torra crashes or is force-quit, the background engine now exits within seconds instead of staying behind, using memory and still running scheduled tasks on time (which could make the same scheduled run happen twice). This applies to quits, updates and crashes after this version is installed.
macOS: after an update is installed, Torra now deletes the full copy of the installed version left in its update cache, freeing about 800 MB of disk space; the files the next update needs are kept.
When a hosted model fails before the request is sent (you're not signed in to your Torra cloud account, connection credentials can't be fetched right now, some usage hasn't been reported to Torra cloud yet, or the model has no available route), the conversation now says which it is, in your interface language, instead of a generic “Model access is not authorized” or “Service is temporarily unavailable”.
When a model request fails before producing anything and is retried automatically, the retry no longer counts against the steps a task can take, so a network hiccup no longer leaves a task with fewer steps; if the steps run out with no result, the final message gives the real cause instead of “The model returned an empty response.”
When the agent searches files with an invalid pattern or a path that doesn't exist, it now gets the actual error instead of “No matches”, so it no longer draws the wrong conclusion.
When a read-only lookup fails in a conversation (for example, viewing a skill that doesn't exist) or a call has invalid arguments, it's no longer treated as an action with an unknown outcome that blocks later actions in that conversation, such as sending a message or starting an approval; those now go ahead as usual.
When one conversation starts a new turn and has to look up the model's context length, other conversations no longer freeze for a few seconds.
Scheduled tasks now do the work and deliver the result when they run, instead of replying with a plan such as “then set it to run every day…”, and they no longer create or change scheduled tasks mid-run.
Updates: a version that is downloaded and waiting for a restart no longer blocks a newer release; Torra switches to the newest one so you don't have to update one version at a time, and a stalled download is dropped and the latest version checked for and downloaded again (this applies to updates after 0.5.134).
Permissions: Allow always now remembers the complete command, so the later part of a long command can no longer slip past confirmation.
The pre-completion check no longer treats a short but correct file as empty, reworking it and then saying completion is not confirmed, and an older file with the same name elsewhere in the workspace no longer causes a false failure for the file just written.
If it's unclear whether an external action such as sending a message went through (for example, the network dropped mid-send), Torra won't run it again on its own until you explicitly ask it to retry, so nothing is sent twice.
Download model under Dictation now runs in the background with a progress bar, so a slow connection no longer ends in a timeout and repeated clicks don't start several downloads; error messages no longer show an internal request path.
The Live meeting assistant no longer tells everyone “Cannot reach the meeting service”: when your network can reach the speech service, Start works.
When you ask Torra in a conversation to open a panel such as Memory, it now gives the real location, for example Settings (bottom left) › Build › Memory, instead of pointing at a sidebar item that doesn't exist; setup instructions for Feishu and other channels use the same paths.
Without an enterprise license, the model list now shows Enter enterprise license at the bottom, so you can always find where to enter your license.
When a Feishu, WeCom or DingTalk task ends before it starts (for example, the confirmation timed out or an attachment couldn't be read), the conversation now keeps the sender's message and what happened, and on errors points you to pending messages under Settings › Connect › Channels to see why and retry, instead of opening a blank page. New conversations from these channels are titled with the start of the sender's message instead of an internal ID.
Images sent from Feishu, WeCom and DingTalk are now read as images and documents by their content, instead of images coming out as garbled text, and replies sent back to channels no longer include file paths from your computer or stray tool parameters.
After sending a first message with several images from the home page, a slow network, an error or a window refresh no longer leaves the conversation blank: sent messages are kept, and anything that did not arrive puts your text and finished attachments back into the input box (without resending), with a note above it listing the files that still need to be added again. Images put back into the input box, including those carried over when you fork a conversation, show proper thumbnails.
Sending a message when your wallet balance is too low is no longer misreported as a model provider quota problem that recharging cannot fix: a notice card appears in the conversation, and Top up opens Settings › Manage › Account.
Re-running in place with the pencil (Edit prompt → Save & re-run) no longer leaves the old reply under the new one, even after a refresh, and you can re-run without changing the prompt.
The Changes tab of a conversation now lists files the AI created, modified or deleted with Python, the terminal or other tools, instead of always showing “No changes yet” (changes are recorded from this version on; very large workspaces are skipped).
Several Settings pages work again: any account can switch Dictation on or off, and Local knowledge assets and Let AI edit my own files under Data & storage no longer fail with HTTP 404 — you can list uploaded files and their processing status, and mount, unmount and undo changes to folders.
After adding, removing or updating agents and skills from the market, the input box's + menu and agent list update right away without a restart, and agent details no longer expose internal IDs.
Channel tasks: Open task in Pending channel messages now opens the conversation inside the app instead of a 404 page in your browser, and Feishu, WeCom and DingTalk tasks that failed before they actually started (for example because an image or file could not be read) can be run again with Retry unfinished delivery; tasks that had already started are not run twice.
In the sidebar filter, the Scheduled count now matches its list, and conversations started by scheduled tasks no longer show up under My Chats.
File paths in replies that include an extra account folder now open, preview and download from the file panel instead of showing “File not found”; files belonging to other accounts still cannot be opened.
A message resent because of a network hiccup is no longer run twice.
macOS 26: the Torra icon in the Dock no longer sits on an extra grey background.
Fixed a bug where, once a task laid out its checklist, that round and every round after it failed, so long tasks can now continue through their checklist.
Uploading a larger folder while starting a conversation from the home page now shows a single folder card with its directory structure intact, just like inside a conversation, instead of dozens of thumbnails and a flattened hierarchy.
“Fork to new session” now keeps the model, route and retrieval settings you chose in the original conversation instead of resetting them to defaults.
When the app has just started or the machine is slow, the first message sent from the home page no longer shows an early “Send confirmation was not received” warning while it is in fact being processed.
Versions released this day
macOS v0.5.130 / 0.5.129 / 0.5.128
macOS v0.5.127
New
New conversations now pick a model automatically based on the task at hand, while any model you chose yourself is preserved instead of quietly reverting to a default.
Improved
New versions no longer interrupt you with a modal: the entry point sits next to your avatar in the bottom left, a finished download is only a quiet hint, and you decide when to restart.
The usage page now distinguishes historical totals from details recorded after an upgrade instead of presenting misleading empty charts.
The project home now presents work entry points and first-use guidance more clearly, including in narrower windows.
The account menu now brings plan, usage, settings, help, and update actions together for quicker access.
Fixed
Mainland users now default to a domestic route, and machines upgraded from older versions switch over automatically instead of failing with an upstream error. Changing the route in Settings takes effect in the conversation you already have open.
When a managed model hits a provider quota or outage, Torra falls back to one that works instead of retrying the same model until it fails.
For questions that need a live lookup, such as today's weather, the answer that was actually found is no longer wiped out by repeated self-checking.
Images, video, and audio produced by Python or other tools now preview and play inline instead of showing only a grey file path.
The app icon is back to the official brand mark (the T with a gold vertical stroke on its right), matching the website and social profiles — 0.5.129 briefly shipped a different one.
Interrupted account requests restart after reconnection, so account and subscription details no longer remain stuck loading.
Video tasks use sensible first-run settings and return the finished result to the active conversation when background rendering completes.
Versions released this day
macOS v0.5.124 · Windows v0.5.126
Improved
When a trusted capability is missing, Torra repairs it inside the private local runtime and retries the original task.
Installed skills and agent lists now use a clearer neutral selection style without distracting accent lines.
The Installed and Market switcher on the skills page is more compact and makes the current section easier to identify.
Fixed
Live questions such as weather now proceed to an online lookup and return a result instead of repeatedly stopping at tool discovery.
Stop, follow-up send, and recording actions now match the current task state without showing conflicting controls.
macOS and Windows packages now include the correct architecture-specific runtime dependencies, reducing startup and page-loading failures after installation.
Versions released this day
macOS v0.5.123 · Windows v0.5.125
New
Documents, images, audio, and video can be used as personal knowledge, with assets, indexes, and revocation state isolated by account.
Improved
macOS and Windows packages now include the required local runtime dependencies for more reliable offline startup and cross-architecture installation.
Fixed
The first image uploaded in a new conversation now selects a vision-capable model immediately without requiring a retry.
Large images are prepared within per-image and per-turn limits before sending, reducing oversized-request failures and timeouts.
Personalization settings now include editable instructions and local memory, while settings, project tabs, and local data remain isolated by account.
Built-in offline product help explains account-scoped setup and acceptance steps for Feishu, WeCom, and other channels.
Document tasks can deliver editable results with revision history, conflict handling, PPTX export, and charts selected and validated for their data purpose.
The desktop app now includes the video production runtime so tasks can deliver video results directly.
Improved
New conversations select the highest available model by default, with ordering and defaults managed from the active server catalog.
Start conversations from a unified composer with model selection, attachments, drag-and-drop, and nearby shortcuts.
Conversation tabs use cleaner selection styling and improved spacing between titles and close buttons.
Fixed
Conversation starts show an immediate offline state and recover after reconnection, while previews wait for the current account runtime to become ready.
Feishu reauthorization, group mentions, document links, and connection state are more reliable, and expired WeChat sessions provide clear recovery steps.
Scheduled tasks retain local execution evidence, catch up missed one-time runs, and expose their actual results for review.
Delivery files are checked individually, and long comma-containing paths, failure states, and post-stop recovery are no longer mistaken for success.
Windows can resume an interrupted installation in the original location and reveal the current account workspace from the app.
Drafts and attachments remain available after a failed first send, and attachments stay with their intended conversation.
Account requests report timeouts and invalid responses, allow retries, and refresh after connectivity returns.
Expired or already answered questions show a clear response instead of leaving duplicate submissions waiting.
Tasks that fail their final checks remain marked incomplete, and results for multiple files are checked individually.
Failed or incomplete scheduled tasks retain their actual outcome instead of being reported as successful by a later response.
Feishu group messages preserve link labels and addresses, and empty replies are not recorded as delivered.
Invalid skill parameters return actionable feedback, and reading a skill includes its file location.
After an unsuccessful update installation, a later download does not immediately quit the app; users retain control over when to restart.
Versions released this day
macOS v0.5.117 · Windows v0.5.121
macOS v0.5.116 · Windows v0.5.120
macOS v0.5.115 · Windows v0.5.119
macOS v0.5.114 · Windows v0.5.118
macOS v0.5.113 · Windows v0.5.117
macOS v0.5.112 · Windows v0.5.116
macOS v0.5.111 · Windows v0.5.115
New
Feishu and WeCom now support a sender allowlist (off by default): when turned on, only people on the list can ask Torra to run tasks. You can copy a sender's ID with one click from the approval card or User Settings → Permissions to fill in the list.
A new Permissions page in User Settings: senders you set to Always allow can be revoked one at a time, and you will be asked again the next time they send something. You can also open it by searching "Permissions" in the command palette.
Improved
Interrupted long tasks retain completed actions and pending steps, so continuing a task does not describe finished work as not yet started.
Improved backups and undo checks for workspace edits, with execution restrictions also checked for background commands.
Channel tasks waiting for your approval in the same conversation now share one notice — "N tasks are waiting for you" — instead of stacking up one per task. It closes once they are approved, declined or time out, and "Review" takes you straight to that conversation.
Business graph and memory: one-off requests such as "only look at high-speed rail this time" no longer turn into standing rules for every later conversation; instructions aimed at the AI inside emails or web pages are no longer recorded as if you said them; different people or companies are no longer merged into one; and when a contact changes, who it used to be is kept.
Each rule in the business graph panel now says whether it is "Our rule · applied automatically" or "Their request · not applied automatically", with a one-line reason.
When generating a series of illustrations, later images use the first one as a reference to keep characters and style consistent, and requests that put the count last ("illustrations, 5 of them") are now understood as five images.
Fixed
When a check after editing a file fails or returns no clear result, the agent attempts another check or repair and reports when verification remains incomplete.
An unavailable completion review is no longer treated as approval. One additional review is attempted; if it remains inconclusive, results are preserved and marked unverified without replaying completed actions.
Text and tool activity stay in the order they occurred, including after reopening a conversation or stopping a response.
Text between two uses of the same tool stays visible instead of being folded into one group.
Updates that fail because storage is full now ask you to free space instead of reporting a network problem.
New conversations keep drafts and attachments after a failed send, and confirming text with an input method no longer sends it accidentally.
Scheduled tasks keep a valid interval when switching units, and Excel dates display consistently across time zones.
Follow-up messages can reference images in the conversation, and download links correctly recognize filenames containing commas.
Channel task approval cards: after a reconnect, cards that have already expired close by themselves instead of blocking new ones, and clicking an expired card tells you it was already handled or timed out. Very long message text on a card is now truncated.
The sender shown on channel approval cards is now their actual platform ID instead of an internal prefix.
DingTalk: after Torra restarts, memories from DingTalk conversations that had not been processed yet are picked up again instead of being skipped.
WeCom: a message that is too large, or arrives when the queue is full, now only affects that message (the sender gets a short notice) instead of dropping and reconnecting the whole connection.
Scheduled tasks described in Chinese ("every day at 9…", "every Monday…") are now set up properly; scheduled runs keep their email and web search tools, and document and WeChat tools are there whenever a document is open or WeChat is connected.
Your own skills are now used when you ask in Chinese too (previously they were left out for Chinese requests).
Asking about images you already have ("what are these 5 images about") is no longer taken as a request for 5 more, and "one image for each of these 3 scenes" is now recognized as an image request.
When Torra points you to something it cannot do from the chat (such as opening the Memory panel), it now names the real menu in your interface language.
With the English interface, image, voice and video tool messages no longer mix in Chinese.
Feishu, WeCom and DingTalk: records of handled messages are now cleaned up instead of piling up, and a DingTalk message that can never be processed is no longer redelivered over and over.
Built-in browser: a browser process left over from an unclean exit is now cleaned up when you switch accounts or when startup fails, so it no longer keeps the browser from starting.
Windows: if repairing a shortcut fails during an update (for example when the desktop is synced by OneDrive), the update no longer fails as a whole — Torra finishes installing and opens as usual, and the Start menu and desktop shortcuts are repaired independently.
Relative times such as "3 days ago" now follow your interface language, and statuses, plans and budget periods in the admin area are localized.
Much more leftover untranslated text has been cleaned up (notifications, tickets, assignees, statistics pages, task progress, member counts and more), and a few English sentences with words run together are fixed.
Within a conversation, abilities that worked a moment ago — memory, asking you a question, writing documents, generating images — no longer vanish for a turn. After generating an image, asking for a change no longer gets "no image tool available".
Windows: if the Start menu entry for Torra went missing, updating now puts it back so Windows Search finds Torra again; a desktop shortcut you deleted yourself stays deleted.
Always-allow for a sender now works on Feishu, WeCom and DingTalk too (it previously only took effect on WeChat), and remembered senders can be revoked one at a time instead of all at once.
Feishu sign-in: if you have not finished authorizing in Feishu yet, Torra now tells you to finish there and try again, instead of a red error showing a command and device code.
Windows: Chinese folder names no longer turn into garbled characters in Python output.
Windows: improved the built-in browser failing to start after switching accounts — leftover browser processes from the previous account are now cleaned up first.
"Show in folder" in the file preview now opens the folder instead of downloading; large-file and timeout messages are localized; the account Members page no longer shows "Not Found" and roles are localized; new scheduled tasks show localized time zones and units.
Account page: phone numbers are checked before sending a code; errors and bill types and statuses are localized; "Sign in" responds when you are signed out.
Torra no longer claims to have opened a panel it cannot open; it tells you where to find it instead.
Scheduled tasks no longer claim there is no tool to write files when their prompt asks for one.
Error messages from Torra follow your interface language.
Mentioning a time such as "every day at 9:00" or asking to reword some text is no longer mistaken for an image request, and the reply no longer repeats its confirmation several times.
After you press Stop, the reply no longer dumps its remaining text all at once.
WeChat: messages Torra has already handled are no longer run again after Torra restarts.
When an attachment cannot be read, Torra now says so and asks you to re-attach it, instead of answering as if your own question were the file's contents.
If image generation keeps failing without a reason, Torra stops after the third attempt and gives you a task number to send us, instead of retrying for minutes.
Uploading a scanned, image-only PDF no longer breaks the chat that follows.
Images generated through skills now show up directly in the chat.
Torra no longer promises a wait time it cannot keep.
Versions released this day
macOS v0.5.110 · Windows v0.5.115
macOS v0.5.109 · Windows v0.5.114
macOS v0.5.108 · Windows v0.5.113
macOS v0.5.107 · Windows v0.5.112
macOS v0.5.106 · Windows v0.5.111
macOS v0.5.105 · Windows v0.5.110
macOS v0.5.104 · Windows v0.5.109
macOS v0.5.103 · Windows v0.5.108
macOS v0.5.102 · Windows v0.5.107
macOS v0.5.101 · Windows v0.5.106
macOS v0.5.100 · Windows v0.5.105
macOS v0.5.99 · Windows v0.5.104
macOS v0.5.98 · Windows v0.5.103
New
Right-click a conversation to add “Send feedback”: it writes a report into your project in one click.
When the same customer or product shows up as several entries, Torra now merges them into one. A new “recent automatic tidy-ups” block appears in the graph panel, each with one-click undo — it only shows once something has actually been merged.
Improved
On long jobs like working through a document, the notice now says how many steps are done instead of only how long you have waited — a ticking clock alone made it look stuck.
The English-label cleanup is finished: accounts, members, groups, projects, marketplace, API keys, deployments, tunnels, instances and the ops console, including hover tooltips, placeholders and count suffixes earlier passes could not see.
The workspace home page is reorganized: section headers sit above their content, the two kinds of suggestions look different, and a brand accent was added.
The English-label cleanup now covers accounts, members, groups, projects, marketplace, API keys and the ops console, including hover tooltips, placeholders and count suffixes that earlier passes could not see.
The update dialog now says the right thing for each state — checking, downloading, ready and up to date no longer borrow each other's wording. When you are already on the latest version it only offers Close, instead of asking you to check again.
Around 180 more English labels across projects, members, admin and the file browser are localized, including many hover tooltips.
The new-session page now shows the same “For You” starters and suggestions that used to appear only on the workspace home.
Some models were spending about 1,470 extra tokens per turn; no longer.
A Windows update used to take over twenty minutes; it now finishes in under five by reusing files that did not change.
The Windows installer now runs in a real window with true progress and clear success/failure pages.
Connected WeChat / Feishu / WeCom chats no longer need a click on the computer for every single message — you can always-allow a specific sender.
The update prompt is now a small sidebar icon with details on hover; the top banner is gone — it was adding a stray scrollbar.
Fresh installs now open maximized (if you resize it yourself, that is remembered). Note: existing installs do not get this in 0.5.99 — fixed in the next version.
Replies no longer slow down as the graph grows.
The workspace home is rearranged: the top bar shows just the one thing to do next, followed by a single row each for “For You” and “You might find useful”, so the whole page fits on one screen.
Fixed
When answering from a document, Torra no longer presents its own guesses as if they came from the file. If the file is unclear it quotes the actual wording, says what is missing, and suggests who to ask — instead of filling the gap with a plausible-sounding reading.
Task lists in replies used to render as checkboxes you could not tick. They are now plain done/not-done markers, and when Torra actually needs you to choose it offers clickable options.
Top tab bar: two sessions with the same title now get numbered, the active tab stands out, and its close button is always visible instead of appearing only on hover.
More Chinese across the interface: the two pages a shared link opens, the editor and canvas, gateway budgets, the admin area, connect-link pages and sandbox templates.
Asking how to connect your Feishu account no longer shows a red error on every turn. Not being set up yet is a state, not a failure: Torra now points you to Customize → Channels → Feishu to fill in the App ID and App Secret — no terminal needed.
The instance picker now reads in Chinese: custom instance, edit instance, no matches, copy ssh, add & connect, generate/regenerate.
The subscription end date used to print in US format regardless of language; it now follows the interface language.
Instance settings now reads in Chinese throughout: no instance selected, untitled instance, version checking/unavailable, auto-update state and channel, backup and restore.
Project apps: add app, save changes, the three deploy states, copy URL, the edit/logs titles, “no commits”, captured and summarized.
The file panel's context menu, zip-download state, clipboard hints and change types now read in Chinese.
Instance actions now confirm in Chinese (“restart initiated” and friends), along with health summaries and the invite dialog.
The first-run provider screen — sign in with your subscription, auto-create a key, and the four validation messages for a custom provider — now reads in Chinese.
The ticket board, triggers tab, the sidebar's conversion progress and deletion countdown, and Settings → API keys.
The step-by-step record of what Torra did showed our internal tool identifiers (Get Workspace, Ls). Those now read in plain language; third-party tools you connect still show their own names.
The Scheduled tasks screen was English throughout — when it runs, when it runs next, its status and its buttons. All of it now reads in Chinese.
Relative times (just now, 3m ago, 2d ago) were still English in the session list, file history, checkpoints, change requests and the audit log. They now read in Chinese everywhere.
Expiry labels under accounts and permissions, the “inherited via group” note on members, and several buttons and notices on Settings → API keys are now in Chinese.
Total spend showed $0.0000 when this machine had no usage ledger, which read as “I spent nothing”. It now shows a dash and says why — unknown is unknown, not zero.
In the Chinese interface, the status Torra shows while working stayed in English every single turn. It is now Chinese throughout, from “searching the web” to “generating an image”.
A few leftover English labels on the Usage and Budget pages — the usage window, Set budget — are now in Chinese.
On a shared computer, a new account could open a website and find the previous person still signed in — email, merchant back-office, anything. Each account now gets its own browser.
Signing out did not actually invalidate the session ticket. It does now.
The writing style Torra uses for your email is learned from your own sent mail; it used to be shared across accounts on one computer. It is now per account.
For everyday asks like “make me a spreadsheet”, Torra could end up with no usable tool and simply say it could not. It now looks for one itself, and says what is missing when it truly cannot.
Saying “don't generate an image” could still trigger image generation.
The update dialog could show “downloading” and “download complete” one after the other in the same session.
After an update downloads, choosing Later again spells out that closing the window is not quitting — without it, some people closed the window and the update never installed.
Closing the update dialog opened from Settings dropped you back to the chat page instead of the settings page you came from.
Deleting the session you were in could leave the whole page unclickable.
The context ring could read “100% used” when it was nowhere near full. When the model's context limit is unknown it now shows no percentage instead of guessing one.
Sending an image or file turned the session title into raw `<file path=…>` markup. It now uses the filename or what you actually typed.
Requests like “make me a deck” or “put this into a spreadsheet” were not recognized as asking for a file, so you could get a claim of success with nothing produced. They are recognized now.
Typing a Latin word right against Chinese stopped Torra from picking the right tool; it now recognizes them without a space.
The brand accent turned grey in dark mode; it now keeps its color in both themes.
When you choose Later after an update downloads, the dialog again spells out that closing the window is not quitting — without that line some people closed the window and the update never installed.
Typing a Latin word right against Chinese, as in “打开chrome看一下”, stopped Torra from picking the right tool. It now recognizes them without a space.
The file browser showed an internal identifier in the type column for system folders; it now reads properly.
Member roles no longer render as raw English enum values.
Questions like “ten common support scenarios” were misread as a request for ten images, leaving the assistant unable to answer in words. Prompts unrelated to images are no longer treated as image requests.
Asking about “the 8th image” is no longer read as a request for eight images.
When image generation fails, the turn now ends instead of retrying silently for minutes.
Around 140 English labels across the session view, sidebar, files, Markdown toolbar and scheduled tasks are now localized, including many hover tooltips.
Clicking “New session” while already on an empty one no longer creates a duplicate.
The “/” and “@” pop-up lists in the composer now close when you click elsewhere, instead of only on Esc.
The “New scheduled task” dialog is now localized: frequency, time, weekdays and the summary line.
Another 18 English labels in the session view are localized (side-panel tabs, changes, diagnostics, tool cards).
WeChat and WeCom reconnect themselves after an update; previously messages sent from your phone got no response until you reconnected by hand.
Sending just a link in Feishu or WeCom no longer flips the reply to English.
A slow turn now says what it is waiting on instead of a generic “long task” message.
If an image tool fails to start, the turn now ends with a reason instead of hanging forever.
A failed image send now carries an error code and retries once in another format.
A turn recovered after an unexpected exit no longer loses its token usage.
The update prompt no longer contradicts itself, and is now a small round icon in the sidebar.
The “open maximized” behaviour announced in 0.5.99 did not reach existing installs; fixed.
Buttons such as “Restart”, “Save” and “Create” showed in English on a Chinese UI; 20 of them are fixed in this batch.
Updates no longer delete files you put in the install folder; anything you edited is kept as a backup copy.
After an update, the shortcut opens the version you just installed.
After scanning the WeChat code the screen now says it is waiting for your confirmation, and reconnecting no longer creates a duplicate conversation.
A message sent from WeChat shows up on the computer immediately, instead of only after the whole task finishes.
The closing line after generating an image is no longer repeated three times.
When an image cannot reach WeChat, the message now says why.
Words like “4K” or “cinematic” in a prompt no longer silently upgrade video or image resolution to a pricier tier.
Conversations started from WeChat / Feishu / WeCom now show token usage too.
The occupation prompt no longer overlaps the title.
Once your own company rules are in the business graph, quotes and similar work automatically carry things like “over this amount needs that person's approval”. The other side's contract terms are never mistaken for your rules.
Ask “which other files mention this order number” and you now get a cross-file answer with sources.
New “system picks”: skills and agents recommended for the occupation you set.
Search the web from chat: ask it to look something up and the answer comes with source cards that say whether each page was read in full or only summarised. They survive refresh and restart.
Search works out of the box with no key to configure. If you already set up your own search provider, your setting is kept.
New “Log feedback” in Settings sends your local logs and description to us in one click. You can also export a session problem report as Markdown and send it over.
One account per person: your phone number and account now share one wallet, you can bind a phone number from the account page, and signing in no longer splits you into multiple accounts.
Improved
When you send a link, Torra now fetches the page itself before answering, instead of sometimes answering as if it had read a page it never opened.
Photo edits are routed better: swapping a product, changing the background, restyling or a virtual try-on now go through real generation instead of a touch-up tool that only does four fixed actions.
Paste a link and it no longer answers with a menu of “what would you like me to do with this?” options.
The update entry in the sidebar is now a labelled button — “Update now” / “Downloading N%” — instead of a 20-pixel icon.
In-app updates on mac now use the same China-based source as Windows, instead of a slow route.
Video now defaults to 720p with 1080p available, settable as your default under Settings → Default models. If you state an aspect ratio, it stops asking again.
The workspace home is rearranged: the input box leads, six large cards collapse into a single row of entries, the occupation prompt moves to the top, and the session context dialog is rebuilt.
Fixed
In WeCom groups, @-mentioning a colleague used to delete their name, and mentioning two people removed the person while keeping the bot. The message now reaches Torra exactly as you typed it.
Answers sent to WeCom, Feishu and WeChat used to swallow double underscores in code (`__init__` became `init`), and filenames containing them too. Fixed.
On the same path, `2 * 3 * 4` came out as `2 3 4`.
On Apple Silicon Macs a component needed for video links was missing from the package (Intel builds always had it). Both architectures now match.
When you spelled out your requirements across several sentences, everything after the first one was silently dropped and only sentence one reached the generator. All of it is carried through now.
The aspect ratio you ask for now actually applies. Previously it only made it into the description text while the canvas stayed at the default.
For video, ratios like 4:5, 2:3 and 3:2 were silently replaced with portrait 9:16. They now map to the closest supported ratio instead of quietly becoming portrait.
Asking for a 10-second video is no longer ignored — it always fell back to the 5-second default.
The “For You” starter suggestions and the to-do block on the workspace home now actually render — in earlier builds both were simply empty.
The occupation prompt at the top no longer wraps, the send button uses the brand colour, and the Chinese label on the collapsed web-search card is corrected.
Windows: the previous build could download an update but not install it (the installer quit after a few seconds and the version never changed). Fixed and re-verified on a real machine: a silent update from the old build took 245 seconds, landed on the right version, and left your own files in the install folder untouched.
Windows: dragging the window on a 125% or 150% display-scaling machine made it slowly grow.
Cleared roughly 250 leftover English strings in the Chinese UI — tool card titles, audit log, sign-in errors, startup stages, checkout and more.
The MCP tools panel went from “clicking does nothing” to actually working — add and edit are dialogs now, on both entry points.
The usage ring next to the input box showed “no usage yet” during long conversations.
Windows: an update interrupted by shutdown now resumes on next boot, and the installer no longer pops up while shutting down.
Failure messages when creating a chat are properly translated instead of relaying raw server text; the Windows edit menu follows your UI language; clicking away or pressing Esc clears a text selection in chat.
Ledger-only rows in billing are no longer drawn as incoming money.
Versions released this day
macOS v0.5.89 · Windows v0.5.93
Fixed
Long chats could cross wires or forget what was said earlier — four root causes found and fixed.
Three points where a long answer could stall, drop the connection, or falsely report completion are fixed.
Image and video parameters you specify now actually reach the generator instead of being dropped on the way.
Versions released this day
macOS v0.5.88 · Windows v0.5.92
macOS v0.5.87 · Windows v0.5.91
macOS v0.5.86
New
When an update is ready, a persistent bar appears at the bottom — no more one-shot dialog you can lose.
Torra now checks for updates every 2 hours and whenever you switch back to its window (previously only at launch and every 6 hours).
Occupation settings now cover 30 roles, 9 student majors, a new “stage” field (high school through PhD and continuing education), and 24 everyday tools, grouped by category.
If you haven't picked an occupation yet, the workspace now nudges you to — recommendations depend on it.
Improved
The scheduled-tasks panel now shows last result, failure reason and run count, with a red badge on failures.
Denser occupation panel, dictation split into its own section, and a redemption-code entry.
Updates found in the background only show in the bar — no dialog interrupts your work. Manual “Check for updates” still shows one.
Cleared a batch of leftover English strings in the Chinese UI; “agent” is now consistently translated.
Fixed
Root cause of “the engine looks frozen” on Windows: every small chunk of streamed text triggered a disk write that held up the whole loop. A side request went from 16 seconds to 0.066, and writes per turn from 616 to 5.
Word documents work for the first time: a .docx no longer lands as a .docx.txt.
The wrap-up message after generating an image was repeated three times.
Pressing stop left text still streaming in.
On Windows paths, deliverable filenames could not be parsed, so “said it made the file but it isn't there” went unreported.
English conversations drifted into Chinese partway through.
Weak passwords like `123456` could be used to register.
Files offered in chat could open to “not found in the workspace”.
When Windows stalled twice in a row, diagnostics only captured the first occurrence.
Five scheduled-task problems: permanently silent after an error, missed slots not caught up, invalid schedules not validated, and runs stuck showing as running after being killed.
The first click on “Check for updates” failed; update sources are now picked by measured throughput.
Chinese leftovers in the English UI, an update dialog that ignored the system language, and a half-sentence English banner during silent updates on Windows.
Large images no longer fail to preview in the Files panel with “File is too large to preview”.
Fixed a restart path that could spawn two engines and then take the whole app down.
Removed the “image key” prompt — image generation is built in. The project-secrets page now only shows for licensed accounts.
Versions released this day
macOS v0.5.85
macOS v0.5.84
Improved
Synchronous network calls at the start of each turn moved off the event loop — network hiccups no longer freeze the engine.
Added a loop watchdog that records what held the loop. A ~1s watchdog note at startup is expected and harmless.
Fixed
A corrupted tool index now rebuilds itself once instead of retrying every 30 seconds forever.
Interrupted turns now end cleanly instead of hanging without a result.
Answers over ~60s were killed as “stuck”. Liveness is now judged by whether the process is alive, so long answers finish.
MCP servers (including keys and OAuth tokens) were shared between admin accounts on one machine; they are now per-account.