STAGS

cherry-studio v2.2.3

patch PublishedNov 19, 2026 SourceCherryHQ/cherry-studio Chinese translation

cherry-studio v2.2.3

Downloads (v2.2.3)

Platform Architecture Download
Windows x64 Installer · Portable
Windows arm64 Installer · Portable
macOS Apple silicon (arm64) DMG · ZIP
macOS Intel (x64) DMG · ZIP
Linux x64 AppImage · DEB · RPM
Linux arm64 AppImage · DEB · RPM
Release Notes

Cherry Studio 2.2.3 - Conversation Sorting, New Models, and Agent Reliability

✨ New Features

  • [Conversations] Grouping and sorting are now independent. Sort by last updated, creation date, name (with natural number order), or manual order, and existing custom orders are preserved.
  • [Agent] Import DeepSeek v3 exports into an automatically created DeepSeek Agent, keeping history, branches, reasoning, search records, and available attachments.
  • [Agent] A short next prompt is suggested after replies: press Tab to accept without sending, or Escape to dismiss. Suggestions can be turned off in input settings.
  • [Agent] You can upload an image avatar while creating an Agent.
  • [Agent] Automatic command approval now parses Bash and PowerShell with shared risk rules and asks before each risky command. Approval cards show the requesting task, working directory, and reason. (action required: review permission modes if you rely on the existing prompts; unmatched commands run automatically, while mandatory denials still apply in Full Access.)
  • [Browser] Conversation sidebars support multiple retained tabs with unified styling and context menus, use the configured homepage, and close with their last page.
  • [File Preview] PPTD slide decks now preview in the app with page navigation, zoom, thumbnails, refresh, and a source view.
  • [Files] File paths and Markdown links with line numbers jump to and highlight the referenced line in the preview.
  • [Files] The file selection dialog opens in the folder of the last file you selected during the current session.
  • [Chat] Added an Appearance → Messages preference, “Scroll to the latest message on send” (on by default).
  • [Chat] Long deliverable file lists show the first three files and let you expand or collapse the rest.
  • [Chat] Ctrl + mouse wheel zooms the interface, shows the zoom level while scrolling, and Ctrl+0 resets it.
  • [Knowledge] The URL source accepts several links at once (separated by newlines, commas, or spaces), adds each valid URL separately, and reports invalid entries individually.
  • [Paintings] Generated images show the model and returned size, and are saved in their original format instead of always being converted to PNG.
  • [Computer Use] A new Linux-only “Global input” switch (off by default) lets agents control the shared mouse pointer and system keyboard.
  • [Models] Added GPT-6.2 Astra and Sol, Gemini 4 Carbon and Hydrogen with reasoning controls, Tencent Hunyuan 4 Flash, and Mistral Large 4 reasoning controls.
  • [Models] DeepSeek first-party cost estimates now use weekday peak and off-peak rates while keeping custom prices. (action required: review local cost estimates if you relied on peak-only estimates.)
  • [Models] Enabled providers now appear first in the model service list, and primary model pickers can open in a remembered dialog mode.
  • [Marketplace] Curated skills can be filtered by category, with search and horizontal category navigation.
  • [Composer] Pasting copied folders into Agent conversations adds folder references instead of failing file attachments.
  • [Export] Obsidian exports label each assistant message with the model that produced it.
  • [Shortcuts] Number shortcuts (1–9) have an explicit Off preset in Settings → Shortcuts.
  • [Voice] The recognition model now has an Off option that disables local dictation without removing the saved language or hotkeys.
  • [Code CLI] Qwen Code 1.0 can be installed from its GitHub release.
  • [App] Added Czech (cs-CZ) language support.

🐛 Bug Fixes

  • [Chat] Nested scrollable content such as code previews, editors, and embedded artifacts scrolls independently while the message list scrolls.
  • [Chat] Conversations restored from Settings → Archive open with their messages instead of an empty conversation.
  • [Chat] Interrupted streamed responses are reported as connection errors with partial output kept for retry, instead of an invalid AI response.
  • [Chat] Multi-model fold tabs show every model name again, with a toggle back to compact tabs.
  • [Chat] Chat requests that exceed the model context window recover more reliably.
  • [Notes] LaTeX written with [ … ] and ( … ) renders again in the Notes editor.
  • [Chat] The browser side-panel resize bar is draggable along its full height.
  • [Agent] Assistant text emitted before or between tool calls is visible again in Agent sessions.
  • [Agent] Pi Durable: Stop keeps the session open and the reply paused; reopening after quitting continues an interrupted reply or shows its real result; messages sent while the agent works are saved and answered in order; a withdrawn message is marked “Stopped before sending” with Resend and Edit.
  • [Agent] Messages sent while a turn is running are acknowledged and acted on instead of being silently ignored.
  • [Agent] A pending approval is cleared when its background task ends, so the reply no longer stays on Pending.
  • [Agent] Model selection is scoped to each session, so changing one session no longer changes sibling sessions.
  • [Agent] Claude Code conversations can be forked or edited after tool retries or compaction when the SDK can preserve their context.
  • [Agent] In Accept Edits mode, Claude Code asks before running an escalated Bash command instead of auto-running it by its first word.
  • [Agent] Agent session history repairs message trees during upgrade and prevents edits to in-flight deliveries.
  • [Agent] An Agent turn recovers when a directory occupies its durable store name instead of failing with a network error.
  • [Agent] Reasoning-effort, service-tier, and Fast controls stay available when editing and resending a historical message.
  • [Agent] Historical messages follow the Agent’s current avatar, and an avatar chosen late during creation is applied.
  • [Agent] Selection assistant: reused windows no longer open at full conversation size, a closed window no longer leaves a stray capsule, and the panel opens above the active app.
  • [Agent] Agent requests delivered through HTTP or in-process gateways keep protocol metadata, Gemini configuration, generated files, and continuation state.
  • [Agent] Unreadable stored automation tool results are reported to the model as a note instead of breaking the conversation.
  • [Agent] NO_PROXY no longer includes bracketed IPv6 or IPv6 CIDR entries that broke Python HTTP clients in Agent sessions.
  • [Models] Model connection checks now use streaming for streaming-capable models and detect incomplete or failed responses.
  • [Models] Unsupported sampling parameters and penalties are no longer sent to providers, and declared sampling ranges apply after request overrides.
  • [Models] Gemini tool calls no longer fail when a schema contains propertyNames.
  • [Models] A single Cherry Cloud quota refusal no longer locks a model while plan quota remains; the server decides refusals and shows its suggested wait.
  • [Models] CherryIN saved model endpoints refresh automatically from the server while keeping local values when discovery fails.
  • [Models] CherryAI appears first with a paid subscription and last without one.
  • [Models] Provider request rejections without an HTTP status are reported as request errors instead of network interruptions.
  • [Models] Usage stats recover provider cache buckets, and locally computed costs are labeled as estimates.
  • [Models] Agent sessions on OpenAI-compatible, vLLM, or llama.cpp models report how far a request exceeds the context window.
  • [Models] Aborted provider turns now fail instead of reporting a successful stop.
  • [Files] File type icons render offline and stay consistent across file trees, previews, and attachments.
  • [Files] File trees and previews stay current, show save progress and failures, and support scoped refresh without losing local drafts.
  • [Knowledge] The Add URL dialog rejects non-http(s) addresses immediately.
  • [Knowledge] Self-hosted MinerU 4.1 Tasks API parsing is supported while older protocols still work.
  • [Feedback] Problem reports again include logs, activity records, and chat history for the last 24 hours; individual attachment switches were removed.
  • [Scheduled Tasks] Run history recorded before 2.2.0, or whose conversation was deleted, can be opened and shows its stored result.
  • [Skills] Cherry-only built-in skills are no longer exported to the shared ~/.agents/skills directory.
  • [Feishu] Initial channel connections retry with backoff instead of staying stopped after a failed handshake.
  • [UI] Empty assistant, agent, and notes sidebars explain that they are empty instead of showing a blank panel.
  • [Agent] Fixed agent sessions with background subagents slowing down a little more every round.
  • [Agents] Preserve subagent rate-limit error details and show consistent retry status for Pi and Pi Durable tasks.
  • [Agents] Reduce out-of-memory crashes when subagents process many images in parallel.
  • [Cherry AI] Load the model catalog from the server and show sign-in, subscription or AMD authorization requirements; detect delisted models. Built-in Qwen now requires Cherry AI sign-in (action required for users who previously used it signed out).

💄 Improvements

  • [Conversations] Grouping choices are clearer, and list controls are separated from history and management navigation in conversation and Agent session menus.
  • [Agent] The service tier picker explains each tier’s trade-offs and shows when a saved tier is unavailable, with provider-specific names in a collapsible details view.
  • [Agent workspace] The folder “Open with” menu now lists every application registered on the system for opening folders, with system icons, instead of only the built-in editors and terminal.
  • [UI] Brand icons, mini-app labels, settings layouts, help, skill controls, and marketplace visuals are more consistent.
  • [UI] Launchpad and Mini Apps icons use consistent sizing and corners with refreshed dark-mode artwork, and inset scrollbars in file trees and lists align with pane edges.
  • [Diagnostics] Native crash reports upload automatically to Sentry when data collection is enabled.
  • [Data upgrade] v1 users must migrate through Cherry Studio 2.2.2; eligible installations offer one-click installation of the migration release, while others keep direct platform-matched downloads. (action required)

⚡ Performance

  • [Agent] Agent sessions with many images or subagents use much less main-process memory and CPU; finished subagents are released and Durable history is read incrementally.
  • [Agent] Restoring large Claude Code sessions and loading their history uses less main-process memory.
  • [Chat] Streaming responses and conversation history loading use fewer copies of conversation snapshots.

What’s Changed

New Contributors

Full Changelog: https://github.com/CherryHQ/cherry-studio/compare/v2.2.2…v2.2.3


Source CherryHQ/cherry-studio · v2.2.3 · 2026-11-19 · upstream release

Mirrored verbatim for indexing. All rights belong to the upstream project.

Release notes are published by the upstream project and mirrored here for indexing. All rights belong to their authors.