What's the latest? Tsukuyomi-space Recent Development Record
Hello, everyone. This is Fushi.
Happy Mid-Autumn Festival!
In recent days,tsukuyomi-space Another round of more intensive updates was carried out.
This time, instead of pursuing a single “big function”, the focus was on the actual use experience of the site as a whole: re-engineering part of the page UI, continuing to refine the mobile end-of-the-way, optimising Room’s AI conversation experience, introducing more complete long-term memory mechanisms, and repairing a number of problems that are gradually becoming visible in practical use.
If this round of updates is summarized in several keywords, it would probably be:
It's better-looking, easier, more stable and more like a complete website.
Website UI: Continue to reduce
I've been adjusting the overall visual performance of the website for some time now.
This time, you want to reduce unnecessary information interference and make the design language between different pages more uniform than you want to continue stacking elements on pages.
Room Page Optimization
Room is where this round of UI adjustments is more concentrated.
The move-end re-engineered the layout between Live2D, chat areas and operating areas, allowing the role to remain the visual subject of the page while leaving more space for actual dialogue.
Some panels had previously appeared to be crowded on their phones and had now been reprocessed:
- Chat Area Transparency Effects
- Altitude available for dialogue areas
- Floating window hierarchy
- Live2D ratio of space to chat area
- Move-end input box and control button layout
The floating windows, such as journals, role information, and signers, are also unified into a new semi-transparent glass style that enhances text readability while preserving the original atmosphere of Room.
Sets the details of the information line, the music button, the guide button, etc. on the page.
For example, the music control at the moving end is no longer crowded with the page guide, and the music control is transformed into a stand-alone round button, making the operation logic more intuitive.
Plaza Optimise with article pages
Plaza also repaired some of the long-standing problems with the display.
These include, inter alia:
- Fix problems with long text spills in responsive cards
- Adjust author information to time label position
- Optimizing the internal spacing of rounded tabs
- Only the latest response preview is shown by default
- Reducing the visual burden of a large number of simultaneous responses
The article page fixes the problem of commenting on the lack of regional contrasts in light-colour mode, while at the same time starting to harmonize the semantic colours of the feedback status of success, warning, error, etc.
Markdown has also been further refined.
These changes are not significant on their own, but when combined, the website as a whole appears to be significantly more uniform than before.
Image loading experience optimization
With more and more pictures on the website, the experience of loading the pictures is becoming an issue to be addressed.
This round of updates applied the new photo loading effect to several main pages, including:
- Hub
- Plaza
- Room
- Arena
- Notifications
Pictures now give priority to decoding, re-entering the corresponding display animations and reducing the scintillation and flashes that begin to transition when the picture is not ready.
The goal here is not simply to add a “good-looking animation”, but to make the process from loading to really showing more natural.
Especially when the network is in general or when the picture size is larger, such details become more apparent.
Room: From the chat window to the real AI Companion
Room remains the most recently changed part.
Before that, Room could have completed the basic:
- LLM Dialogue
- Live2D emoticon connection
- TTS voice play
- Session Record
- Role Settings
- Journals and Memory
However, with the increase in actual use, many problems are beginning to emerge.
For example:
- Respond flow output is not stable enough
- Occasionally, the state is not synchronized during the continuous dialogue
- There are differences in the status of memory-keeping conversations
- The order of execution between expression and voice is not natural.
- Declining context coherence after multi-cycle dialogues
The whole chat process was reorganized in this round.
There is now a clearer relationship between fluent responses, dialogue rounds, context synchronization and memory processing, with corresponding reliability tests added.
Dialogue rhythm adjustments
The manner in which role responses are presented has also been adjusted.
Yachiyo's response no longer favours the output of a long whole paragraph at a time, but seeks to maintain a more natural and concise chat.
Persona continues to adjust to role-setting to bring overall performance closer:
"Calling with the character"
Not:
"A Normal Large Model with a Role Indices"
Live2D emoticons synchronized with TTS
The order of implementation of Live2D emoticons and TTS has also been amended.
Previously, in part:
text语音开始播放
↓
过一会儿角色表情才发生变化
Now read:
text收到完整回复状态
↓
切换 Live2D 表情
↓
开始 TTS 播放
This allows for more synchronization between visual feedback and sound.
Long-term memory: Room begins to really remember things
One of the more important features of this round is the reorganization of Room's long-term memory system.
The project is in. Mem0To provide more sustainable room memory.
The memory logic of the past depends more on:
- Local Status
- Project own database
- Session Sync
- Long-term memory maintained manually
Now, the long-term memory is further removed and added:
- Memory Perseverance
- Memory Retrieval
- Context Injecting
- Memory validation
- Local memory fallback
- Long Memory Sync
Simply put, Room is no longer just reading recent rounds of chat.
Some of the information of long-term value can be preserved, retrieved in subsequent dialogues and handed over to the model as appropriate.
For example:
text用户之前提到的偏好
↓
保存为长期记忆
↓
未来再次聊天
↓
检索相关记忆
↓
注入当前对话上下文
↓
模型能够自然地“想起来”
At the same time, a number of memory-related issues have been continuously repaired over the past few days, including:
- Memory editing is unreliable
- Full memory retention anomaly
- Journal Sync Problem
- Old Room Client Memory Switch Compatibility
- Could not close temporary folder: %s
- PUT request to save anomaly
- The session memory is not consistent with the long-term memory state.
The memory system belongs to:
For the first time, it was easy, and real stability was very difficult.
So in this round of development, more energy than just adding a Memoory button is actually spent on how to ensure that it can really be preserved, synchronized and retrieved.
Mobile Room Fix
The moving end has been the most vulnerable place for Room.
Because there's enough space on the end of the desktop, mobile screens need to accommodate both:
- Live2D
- Chat Record
- Input Box
- Control button
- Floating Panel
It is easy to squeeze each other in a layout.
Recent updates have reset the layout of the mobile Companion and chat areas and further validate the real move side.
Fix input box focus leading to page moves
A more particular problem has also been repaired.
The browser may try to scroll the page automatically when the moving-end input box gets the focus.
And there are larger Live2D Canvas in Room, so sometimes:
text点击输入框
↓
浏览器自动滚动
↓
Live2D Canvas 被一起带动
↓
整个 Room 场景位置发生偏移
This action is being addressed to avoid any sudden displacement of the entire Room scene when the message is entered.
Touch input reliability
The tactile simulation also added Canvas access to the status test.
Only when the target Canvas is actually available will the touchup operation be continued, thereby reducing anomalies during the page initialization phase.
Pixel: Re-emphasize the painting itself
In addition to Room, Pixel / Arena has made a more visible adjustment.
The new design directions are:
painting-first
In other words, the canvas itself is more likely to be the subject of the entire workspace than to have a large number of toolbars and functional areas in the visual centre.
The previous logic was closer:
text功能
工具栏
设置
按钮
画布
Now I'd rather be:
text画布
↓
需要时再调用工具
In this direction, the Arena page has been restructured and a large number of styles have been created to allow the tool to revolve around the creation service rather than to allow the user to enter the page and first face a complex set of interfaces.
This is also a clearer direction in the current UI adaptation:
The functionality can be increased, but the default interface should be simpler and simpler.
The mail notification system continues to improve.
The notification system in the back of the website also added a more important capability this time.
A configured mail notification mechanism has been added and redesigned with tsukuyomi-space style the mail template.
The current relevant updates include, inter alia:
- Configure Mail Notifications
- Branded Mail Templates
- Login Location Change Alarm
- Comment Mail Notification
- Reply to Mail Notification
- Mail content preview optimization
- Backstage Notification Settings
Notices relating to article comments have also been improved.
When a new comment or related response to an article appears, an email alert can be sent to the corresponding author, while the preview of the content in the mail is optimized to avoid overloading the message itself with irrelevant information.
Compared to simply “e-mailable”, there is now a gradual consideration that notifications should:
- When?
- To whom?
- Send what?
- Allow users to close
- How to avoid excessive intrusion.
Deployment processes have also begun to become more reliable
There are still a number of changes in this round that ordinary users may not feel at all, but which are important for an operational website.
The project reorganized the deployment process in the production environment and introduced a more secure Release mechanism.
Resource protection
The deployment will now be verified separately:
text前端构建产物
And:
text需要长期保留的静态资源
Avoid unexpected coverage of user resources or available data in the production environment during construction and deployment.
Release and Roll Back
Also joined:
- Release Backup
- Build Results Validation
- Producing Release Verify
- Rollback
- Autoback cleanup
- Docker Smoke Test
- Deployment Safety Test
In the event of an anomaly following the deployment of the new version, the previous version could be returned more safely, instead of being temporarily rescued on the server.
The corresponding sections were also adjusted as follows:
textDocker
GitHub Actions
deploy.sh
safe-release.py
生产环境资源管理
自动化测试
As the project becomes more complex, this part is actually as important as the addition of new functions.
After all, for a really running website:
The “successful release of new versions” does not mean that the deployment system is sufficiently reliable.
More importantly:
"A new version is safe to come back."
Bug, the repairs still make up a large part.
There's a lot of details in the recent committees, Bug.
For example:
- Room settings cannot be correctly sustained
- Diaries Persona interacts with ordinary chats Persona
- Long-term memory editing instability
- The CDN environment cannot keep memory properly
- Moveend Live2D scene offset
- Reply expression after TTS
- Plaza long text spill
- Plaza Time Tab Malformed
- Light color mode comment contrast insufficient
- Music button overlaps with guide button
- Overresponse of regional information
- Insufficient mobile chat space
- Flashing on image loading
- Old client memory switch compatibility anomaly
These issues may not be a new function worthy of separate advocacy.
But they often decide most directly whether a website is:
"Runable"
Or:
"It really works every day."
As a result, the development focus of this recent phase has begun to shift from fast-tracking of new functions to:
Put what already exists really good.
Recent development directions
We'll look back at this latest batch of committees and we'll find out. tsukuyomi-space It is slowly entering a new phase.
I started thinking more:
text网站还能加入什么？
And now more and more the problem becomes:
text这个功能用起来舒服吗？
手机上还能保持同样的体验吗？
刷新之后状态还在吗？
网络不好时会发生什么？
部署失败怎么办？
AI 下一次还记得这次聊天吗？
The number of functions will certainly continue to increase.
But I'd rather just continue to expand the page now. tsukuyomi-space Make a permanent and growing personal space.
Most of the most recent updates are not very visible “large versions”.
They are only repeated:
text调整几个像素
修掉一个 Bug
重新整理一段逻辑
优化一次交互
补一个测试
改善一次部署流程
And let a function that could already work become a little more natural.
But that's what these things are. tsukuyomi-space。
Tsukuyomi Space, still growing.