v2.2.0

## What's Changed
## ✨ Added

**QwenPaw Hub**

- Self-host a multi-user QwenPaw Hub with local-process or Docker runtimes, workspace-level access controls, credential management, and reverse-proxy support ([#7112](https://github.com/agentscope-ai/QwenPaw/pull/7112))

**QwenPaw Data**

- Deploy QwenPaw Data as a native PawApp with a durable analysis workspace, packaged runtime, and one-command Docker Compose example ([#6940](https://github.com/agentscope-ai/QwenPaw/pull/6940), [#7190](https://github.com/agentscope-ai/QwenPaw/pull/7190))

**QwenPaw Mail**

- Add an intelligent mail assistant with real-time monitoring, search, thread-aware actions, sender allow/deny rules, and human approval for unknown recipients ([#6800](https://github.com/agentscope-ai/QwenPaw/pull/6800))
- Bundle and launch the Mail MCP with the desktop app ([#7170](https://github.com/agentscope-ai/QwenPaw/pull/7170))
- Keep Mail tool responses readable in Chinese conversations ([#7303](https://github.com/agentscope-ai/QwenPaw/pull/7303))

**Models & Providers**

- Make model setup and selection more consistent with unified provider discovery, capability-aware routing, fallback policies, and thinking controls ([#6302](https://github.com/agentscope-ai/QwenPaw/pull/6302))
- Add Volcengine Agent Plan and MiMo V2.5 providers, with refreshed Volcengine catalogs ([#6515](https://github.com/agentscope-ai/QwenPaw/pull/6515))
- Add prompt-cache hit and write observability to Token Usage ([#7342](https://github.com/agentscope-ai/QwenPaw/pull/7342))

**QwenPaw Creator**

- Creator 1.1.0 expands image and video generation across mainstream providers and improves narrative videos by carrying planned dialogue into generation prompts ([#7167](https://github.com/agentscope-ai/QwenPaw/pull/7167))
- Creator 1.1.1 adds recorded website and desktop operation, Bailian Wan3 video generation, and configurable automated video-quality checks ([#7274](https://github.com/agentscope-ai/QwenPaw/pull/7274))

**Console & Workspaces**

- Organize subagent conversations into groups and surface completion and unread status ([#7035](https://github.com/agentscope-ai/QwenPaw/pull/7035), [#7275](https://github.com/agentscope-ai/QwenPaw/pull/7275))
- Make background work easier to follow with a compact task list and consistent run tracking ([#7083](https://github.com/agentscope-ai/QwenPaw/pull/7083), [#7319](https://github.com/agentscope-ai/QwenPaw/pull/7319))
- Show response artifacts and make generated media directly downloadable ([#7161](https://github.com/agentscope-ai/QwenPaw/pull/7161), [#7036](https://github.com/agentscope-ai/QwenPaw/pull/7036))
- Select system-prompt files from the workspace UI ([#7078](https://github.com/agentscope-ai/QwenPaw/pull/7078))
- Create folders and copy files from the workspace UI ([#7151](https://github.com/agentscope-ai/QwenPaw/pull/7151), [#7314](https://github.com/agentscope-ai/QwenPaw/pull/7314))
- Keep project directories scoped to each session so concurrent work can use different roots safely ([#6976](https://github.com/agentscope-ai/QwenPaw/pull/6976))
- Expand Token Usage with model and tool-call trends, per-Agent breakdowns, and an all-Agent view ([#7207](https://github.com/agentscope-ai/QwenPaw/pull/7207), [#7219](https://github.com/agentscope-ai/QwenPaw/pull/7219))
- Auto-hide the QwenPaw OS dock for maximized windows ([#7233](https://github.com/agentscope-ai/QwenPaw/pull/7233))
- Improve navigation when editor tabs overflow ([#6983](https://github.com/agentscope-ai/QwenPaw/pull/6983))
- Copy assistant responses without including hidden reasoning text ([#7448](https://github.com/agentscope-ai/QwenPaw/pull/7448))

**Skills, Channels & Media**

- Integrate AnySearch as a web SearchProvider and MCP service ([#7081](https://github.com/agentscope-ai/QwenPaw/pull/7081))
- Search Skills and enable or disable them in batches ([#7140](https://github.com/agentscope-ai/QwenPaw/pull/7140), [#6990](https://github.com/agentscope-ai/QwenPaw/pull/6990))
- Keep Skills current with faster file loading, cleaner bindings, and automatic updates for built-ins ([#7097](https://github.com/agentscope-ai/QwenPaw/pull/7097), [#7232](https://github.com/agentscope-ai/QwenPaw/pull/7232))
- Manage apps, plugins, and Skills in one marketplace ([#6880](https://github.com/agentscope-ai/QwenPaw/pull/6880), [#7017](https://github.com/agentscope-ai/QwenPaw/pull/7017), [#7235](https://github.com/agentscope-ai/QwenPaw/pull/7235))
- Open newly installed PawApps without reloading the Console ([#7238](https://github.com/agentscope-ai/QwenPaw/pull/7238))
- Configure plugin channels interactively from the CLI ([#6943](https://github.com/agentscope-ai/QwenPaw/pull/6943))
- Include custom channels in Agent MCP access rules ([#7225](https://github.com/agentscope-ai/QwenPaw/pull/7225))
- Keep Matrix group sessions and memories isolated per sender, with retries for transient login failures ([#7001](https://github.com/agentscope-ai/QwenPaw/pull/7001), [#7109](https://github.com/agentscope-ai/QwenPaw/pull/7109))
- Let DingTalk groups share conversation context when configured ([#7208](https://github.com/agentscope-ai/QwenPaw/pull/7208))
- Configure automatic layout for DingTalk widescreen cards ([#7416](https://github.com/agentscope-ai/QwenPaw/pull/7416))
- Download OneBot media locally before Agent processing ([#6715](https://github.com/agentscope-ai/QwenPaw/pull/6715))
- Let Computer-use observe related windows ([#7037](https://github.com/agentscope-ai/QwenPaw/pull/7037))
- Optionally resize oversized images before sending them to a model ([#7294](https://github.com/agentscope-ai/QwenPaw/pull/7294))
- Propagate Console stop requests to active Agent tools ([#7349](https://github.com/agentscope-ai/QwenPaw/pull/7349))
- Automatically fold completed assistant process messages ([#7374](https://github.com/agentscope-ai/QwenPaw/pull/7374))

## 🔄 Changed

- Start persistent Drivers concurrently for faster workspace startup ([#7174](https://github.com/agentscope-ai/QwenPaw/pull/7174))
- Reduce streaming overhead during long responses ([#7244](https://github.com/agentscope-ai/QwenPaw/pull/7244), [#7239](https://github.com/agentscope-ai/QwenPaw/pull/7239))
- Stabilize reusable prompt-cache prefixes across requests ([#7346](https://github.com/agentscope-ai/QwenPaw/pull/7346))
- Keep long, Markdown-heavy chats responsive ([#7176](https://github.com/agentscope-ai/QwenPaw/pull/7176))
- Use tighter, more consistent spacing for Markdown lists ([#7317](https://github.com/agentscope-ai/QwenPaw/pull/7317))
- Improve the ReMe runtime dashboard and long-term-memory guidance ([#6984](https://github.com/agentscope-ai/QwenPaw/pull/6984), [#6997](https://github.com/agentscope-ai/QwenPaw/pull/6997))
- Stop sending notifications when a scheduled job has not changed ([#7115](https://github.com/agentscope-ai/QwenPaw/pull/7115))
- Remove obsolete Desktop Mode reminders and outdated approval hints ([#7245](https://github.com/agentscope-ai/QwenPaw/pull/7245), [#6981](https://github.com/agentscope-ai/QwenPaw/pull/6981))
- Update AgentScope through 2.0.7.post1 ([#7276](https://github.com/agentscope-ai/QwenPaw/pull/7276), [#7373](https://github.com/agentscope-ai/QwenPaw/pull/7373))
- Refresh the Console chat package used by this release ([#7257](https://github.com/agentscope-ai/QwenPaw/pull/7257), [#7289](https://github.com/agentscope-ai/QwenPaw/pull/7289))
- Clarify the behavior of each approval mode ([#7368](https://github.com/agentscope-ai/QwenPaw/pull/7368))
- Make embedding-model changes explicit and reversible, falling back to BM25 until a scoped reindex succeeds ([#7133](https://github.com/agentscope-ai/QwenPaw/pull/7133))
- Rename Marketplace to Extensions and restore dark-mode styling across the Console ([#7454](https://github.com/agentscope-ai/QwenPaw/pull/7454))

## 🐛 Fixed

**Models, Runtime & Memory**

- Restore complete free-model listings and legacy model names ([#7175](https://github.com/agentscope-ai/QwenPaw/pull/7175))
- Refresh Aliyun and Kimi model catalogs ([#7271](https://github.com/agentscope-ai/QwenPaw/pull/7271), [#7277](https://github.com/agentscope-ai/QwenPaw/pull/7277))
- Separate model output capabilities from request limits so routing and validation use the correct metadata ([#7337](https://github.com/agentscope-ai/QwenPaw/pull/7337))
- Improve compatibility with strict DashScope tool schemas and audio messages ([#7284](https://github.com/agentscope-ai/QwenPaw/pull/7284), [#7024](https://github.com/agentscope-ai/QwenPaw/pull/7024))
- Honor server-provided retry limits during streaming ([#6617](https://github.com/agentscope-ai/QwenPaw/pull/6617))
- Recover from stalled model streams ([#7150](https://github.com/agentscope-ai/QwenPaw/pull/7150))
- Propagate and report model timeouts accurately ([#7268](https://github.com/agentscope-ai/QwenPaw/pull/7268), [#7308](https://github.com/agentscope-ai/QwenPaw/pull/7308))
- Keep Agent behavior compatible with AgentScope 2.0.6 ([#6908](https://github.com/agentscope-ai/QwenPaw/pull/6908))
- Enable the Ollama embedding backend ([#7131](https://github.com/agentscope-ai/QwenPaw/pull/7131))
- Preserve embedding verification while navigating ([#7227](https://github.com/agentscope-ai/QwenPaw/pull/7227))
- Exclude reasoning text from generated chat titles ([#7187](https://github.com/agentscope-ai/QwenPaw/pull/7187))
- Give background tasks a reliable default timeout without overriding explicit values ([#6869](https://github.com/agentscope-ai/QwenPaw/pull/6869))
- Start ReMe memory services before model configuration so local memory features remain available without a configured model ([#7468](https://github.com/agentscope-ai/QwenPaw/pull/7468))
- Include `~`-based Agent workspace paths in model and tool-usage trends ([#7432](https://github.com/agentscope-ai/QwenPaw/pull/7432))

**Console, Files & Media**

- Reopen the sole history session directly from New Chat ([#7104](https://github.com/agentscope-ai/QwenPaw/pull/7104))
- Show the actual completion time for assistant responses ([#6938](https://github.com/agentscope-ai/QwenPaw/pull/6938))
- Render historical data-URL images after reload ([#7069](https://github.com/agentscope-ai/QwenPaw/pull/7069))
- Improve model-selector styling and dark-mode session groups ([#7137](https://github.com/agentscope-ai/QwenPaw/pull/7137), [#7111](https://github.com/agentscope-ai/QwenPaw/pull/7111))
- Hide empty grep badges and unavailable approval controls ([#7165](https://github.com/agentscope-ai/QwenPaw/pull/7165))
- Improve mobile composer controls ([#7334](https://github.com/agentscope-ai/QwenPaw/pull/7334))
- Align cron calendar dates with Sunday-first headers ([#7352](https://github.com/agentscope-ai/QwenPaw/pull/7352))
- Keep embedding verification when switching Agents ([#7353](https://github.com/agentscope-ai/QwenPaw/pull/7353))
- Preserve uploaded and non-ASCII filenames ([#6492](https://github.com/agentscope-ai/QwenPaw/pull/6492), [#7192](https://github.com/agentscope-ai/QwenPaw/pull/7192))
- Preserve Windows CRLF line endings and refresh file previews after changes ([#7091](https://github.com/agentscope-ai/QwenPaw/pull/7091), [#7094](https://github.com/agentscope-ai/QwenPaw/pull/7094))
- Continue remote-file download fallback after a downloader timeout ([#6371](https://github.com/agentscope-ai/QwenPaw/pull/6371))
- Persist remote images before saving tool results ([#7146](https://github.com/agentscope-ai/QwenPaw/pull/7146))
- Deliver tool-result videos through the OpenAI Responses API ([#7061](https://github.com/agentscope-ai/QwenPaw/pull/7061))
- Make inline video limits configurable ([#7071](https://github.com/agentscope-ai/QwenPaw/pull/7071))
- Preserve local file URLs while normalizing media messages ([#7343](https://github.com/agentscope-ai/QwenPaw/pull/7343))
- Save screenshots to the active project directory ([#7439](https://github.com/agentscope-ai/QwenPaw/pull/7439))
- Link Daily Paper to the correct QwenPaw documentation ([#7466](https://github.com/agentscope-ai/QwenPaw/pull/7466))

**Channels**

- Restore Yuanbao message decoding with protobuf 6 and later ([#7421](https://github.com/agentscope-ai/QwenPaw/pull/7421))

**Plugins, Desktop & Operations**

- Restore plugin workspace state before reloads and isolate absolute imports between plugin namespaces ([#6996](https://github.com/agentscope-ai/QwenPaw/pull/6996), [#6688](https://github.com/agentscope-ai/QwenPaw/pull/6688))
- Recover Windows WebView2 after a browser crash ([#7128](https://github.com/agentscope-ai/QwenPaw/pull/7128))
- Avoid false process matches in the Windows installer, including the NSIS caller itself ([#7281](https://github.com/agentscope-ai/QwenPaw/pull/7281), [#7323](https://github.com/agentscope-ai/QwenPaw/pull/7323))
- Handle NSIS uninstall blockers and make application-data cleanup choices explicit ([#7336](https://github.com/agentscope-ai/QwenPaw/pull/7336), [#7354](https://github.com/agentscope-ai/QwenPaw/pull/7354))
- Clean up safely when workspace startup is cancelled ([#7194](https://github.com/agentscope-ai/QwenPaw/pull/7194))
- Protect environment configuration from partial or corrupt writes ([#7135](https://github.com/agentscope-ai/QwenPaw/pull/7135))
- Keep backup jobs running after an SSE connection closes ([#7283](https://github.com/agentscope-ai/QwenPaw/pull/7283))
- Keep Agent configuration persistence consistent across shared filesystems and concurrent writes ([#6767](https://github.com/agentscope-ai/QwenPaw/pull/6767))
- Enforce File Guard paths during active policy evaluation ([#7375](https://github.com/agentscope-ai/QwenPaw/pull/7375))
- Expand sandbox mount paths correctly ([#7116](https://github.com/agentscope-ai/QwenPaw/pull/7116))
- Avoid antivirus interference during sandbox setup ([#6986](https://github.com/agentscope-ai/QwenPaw/pull/6986))
- Isolate QQ conversations and preserve their reply routes ([#7169](https://github.com/agentscope-ai/QwenPaw/pull/7169))
- Report Xiaoyi proactive-send failures reliably and preserve session IDs ([#7147](https://github.com/agentscope-ai/QwenPaw/pull/7147))
- Patch security-sensitive website and Creator dependencies ([#7172](https://github.com/agentscope-ai/QwenPaw/pull/7172))
- Speed up plugin loading by avoiding a full module scan after every plugin ([#7383](https://github.com/agentscope-ai/QwenPaw/pull/7383))
- Restore ReMe and its Auto Fin and Daily Paper plugins in packaged desktop builds ([#7458](https://github.com/agentscope-ai/QwenPaw/pull/7458))
- Prevent desktop packages from mixing different Python runtime sources ([#7372](https://github.com/agentscope-ai/QwenPaw/pull/7372))
- Block shell line-continuation variants from bypassing sensitive-path protections ([#7472](https://github.com/agentscope-ai/QwenPaw/pull/7472))

## New Contributors

- [@cyruszhang](https://github.com/cyruszhang)
- [@Luohh5](https://github.com/Luohh5)
- [@LUOSENGWA](https://github.com/LUOSENGWA)
- [@An-idd](https://github.com/An-idd)
- [@TinyBai](https://github.com/TinyBai)
- [@c020627](https://github.com/c020627)
- [@suantea](https://github.com/suantea)
- [@xiaoka76](https://github.com/xiaoka76)
- [@anysearch-ai](https://github.com/anysearch-ai)
- [@guodaxia103](https://github.com/guodaxia103)
- [@kilowu](https://github.com/kilowu)
- [@kaiwangleo](https://github.com/kaiwangleo)

**Full Changelog**: [v2.1.0...v2.2.0](https://github.com/agentscope-ai/QwenPaw/compare/v2.1.0...v2.2.0)