FULI 0.23.0

更新内容：
解决了一些已知问题。

（比如「换个说法就找不到」这种已知问题。现在你问「上线窗口」，它也认得出那条叫「发布冻结」的知识。）

• 每次提问都先找一找：不用说「还记得吗」，FULI 也会按你这句话去找相关知识；没找到就什么都不加，不多花 token。
• 搜不到时给目录：一次搜索什么都没找到，AI 会拿到这个项目的知识标题，自己挑着再查。
• 按意思找（可选）：给「语义搜索」接一个小向量模型，本机 Ollama 或在线接口都行，换个说法也能找到；不接也照常按字找。
• 老知识一键补检索词：模型页的「补检索词」先免费数一数，再试 20 条，看补之前和补之后各能找回几条、每条花多少 token，由你决定补不补。
• 模型一处管：顶栏设置旁边多了模型入口，接没接模型一眼看得出；模型页每个模型只列一次，写着谁在用、什么模型就够，「查看详情」跳到官网的模型指南。
• Dreaming：给 Astro 接上模型后，它会回看项目最近的会话，提出记忆该怎么改，放进待你确认的队列。
• 接模型更省心：可以真测一次连接；局域网里的 HTTP 地址、需要端口或请求头的网关也能连；除非你自己设，不再有默认上限。
• 桌面客户端：打开客户端就会启动 FULI；FULI 自己退出了会自动拉起来；客户端要重启才能用上更新时会提示你。
• 等你确认数据迁移时，任务照常进行：以前图谱在等确认时每个任务都会出错，现在只读照常用。
• 体检更细：能找出老版本留下的 FULI MCP 进程；客户端改了会话记录格式会提醒。
• 回复里用到的知识算作用过，不额外花模型 token。
• 官网新增「模型指南」（测试数据和 token 花费）、「技术」页（架构、Skills、数据设计、理念）和「快速上手」。
• 小改进：对话框里的下拉框能用了；迁移页把已完成的和可以撤销的分开显示。

升级提醒：从 0.22.1 升级时，图谱有一步迁移（把内置的 Jefa 改叫项目经理、Tonborg 改叫决策员）要你在控制台确认，确认前图谱只读，任务照常进行，也可以撤销。

这一版还包含 0.23 测试版里的全部更新：桌面客户端、自己体检和修复、每个 FLA 的工作台、AI 权限等，详见 0.23.0 更新说明。

What's new:
Fixed some known issues.

(Such as "ask it differently and it's gone". Ask about the "ship window" now and FULI still finds the item called "release freeze".)

• Every prompt looks first: no need to say "remember?"; FULI searches with your own words. When nothing matches it adds nothing, so no tokens are spent.
• An index when a search misses: a search that finds nothing hands your AI the project's knowledge titles to pick from.
• Search by meaning (optional): connect a small embedding model to 语义搜索 / semantic search, local Ollama or an online API, and a differently worded question still finds the item. Without one, word search works as before.
• Retrieval terms for old knowledge: 补检索词 on the model page counts for free, tests 20 items to show how many more are found and what each costs, and you decide whether to fill the rest.
• One place for models: a model button beside Settings shows at a glance whether a model is connected. The model page lists each model once with who uses it and what is enough, and 查看详情 opens the website's model guide.
• Dreaming: with a model for Astro, it reads a project's recent sessions and proposes what memory should change, queued for you to confirm.
• Easier model setup: test a connection for real; reach plain-HTTP endpoints on your network and gateways that need a port or header; no default limit unless you set one.
• Desktop client: opening it starts FULI, a FULI that stopped by itself comes back, and it tells you when a client needs a restart to load an update.
• Work goes on while a migration waits: tasks used to fail while the graph waited for your confirmation; now it stays readable.
• Sharper health check: finds FULI MCP processes left by older versions and notices when a client changes its session record format.
• Knowledge your AI cites in a reply counts as used, at no model cost.
• The website gains a model guide (test results and token costs), a technology page (architecture, Skills, data design, principles) and a quick start.
• Smaller things: dropdowns work inside dialogs; the migration page shows completed migrations apart from those you can undo.

Upgrade note: coming from 0.22.1, one graph migration (renaming built-in Jefa to Project manager and Tonborg to Decision maker) waits for your confirmation in the console. Until then the graph is read-only, tasks keep working, and the step can be undone.

This release also includes everything from the 0.23 betas: the desktop client, self-checks and repairs, a workbench for every FLA, AI permissions and more. See the 0.23.0 release notes.