[Phantivia]

中文

删掉统一供应商编辑器已经替代的旧 llm-settings 面板：服务端的声明与专属操作、WebUI 的面板实现与内置面板表、只服务它的文案和测试。@kwk1001 这次按你在 issue 里列的边界做：保留 ProviderSettings、sources()、模块自定义栏目与它们的 invoke 分发、底层的保存、密钥状态、探测与报价能力；报价编辑器连同十一种语言的文案先迁进 features/providers/ 再删旧目录；公共协议的 builtin 字段保留；旧的 #/provider/llm:<kind> 链接照旧跳到供应商页。

两个提交：

• refactor(webui)!：pricing-panel.ts 移到 src/web/client/features/providers/pricing.ts，用到的十九条文案并入供应商页的语言表；删掉 webui-pages/builtins/ 与 builtins.ts。WebUiPageHost 不再接 builtins，面板声明了 builtin 时页签、插槽、端点编辑页三条挂载路径都在那一块显示「内置面板不存在」，不去加载页面扩展。
• refactor(providers/webui)!：llm:<模块> 页不再声明 settings 面板，页面 invoke 只转给模块自己的面板；删掉只为它服务的 ProviderSettings.activate/delete/setSecret 和私有的 assertNewName/directoryExtras/probe，以及七条只有它用的文案。ProviderHub.action 不带草稿时直接用已保存端点的实例（模型列表走 registry.resolve，测试走 ProviderSettings.probeClient），请求和回执形状不变，仍然不要求带草稿；端点没有模型时的报错改成和草稿相同的「请选择模型」。

BREAKING

/api/webui/providers/llm:<模块>/panels/settings/<方法> 删除，方法有 state、create、activate、save、delete、duplicate、setSecret、models、probe、extras。改用 /api/providers：列表 GET /api/providers，详情 GET /api/providers/:name，新建 POST /api/providers（复制是带 copyFrom 的新建），保存 POST /api/providers/:name/save（整条提交，带 expectedRevision，密钥值放 secretValue），设为当前 POST /api/providers/:name/activate，删除 POST /api/providers/:name/delete，测试与模型列表 POST /api/providers/:name/test、/models；模块的推理档、服务档与默认地址在 GET /api/provider-modules。extras（列端点目录里的文件）没有替代。WebUI 不再带 llm-settings 面板，面板声明 builtin: 'llm-settings' 会显示错误卡。不设弃用期。

测试

• 只覆盖旧面板界面的测试删除（llm-settings-open-ui、model-temperature-ui、service-tier-ui 与它们的夹具）。
• 报价编辑器的六条测试改为直接挂编辑器（tests/web/feature-providers-pricing.test.ts）。
• webui-builtins 去掉注册表断言，挂载、插槽、错误隔离与卸载改用页面扩展的面板，并断言声明了 builtin 的面板显示错误卡且不加载扩展。
• 旧数据面测试按能力改挂：密钥状态、模型名与探测提示留在 provider-settings；已保存端点不带草稿的模型列表与探测移到 hub.test.ts；Cormini 与全系统集成测试改走 /api/providers。

验证

• pnpm run typecheck、pnpm typecheck:web、pnpm test（全量）、pnpm build:web 通过。
• scripts/dev-webui.ts 假数据（端口 8871）里打开供应商页：旧链接 #/provider/llm:llamacpp 落到 #/providers；llamacpp 端点的运行时、模型两块模块面板照常挂上；对 local 端点「测试连接」出回执（假 llama-server 不答 /chat/completions，回执是 404 提示）、「获取模型列表」列出 2 个模型；报价三格表填币种和输出单价后完整规则 JSON 同步更新、卡片标草稿，保存后 GET /api/providers/local 读回这份报价；新建 llamacpp 连接 llama-two，取模型列表后选一个保存，列表里显示可连接。
• curl 不带草稿：POST /api/providers/local/models 回 2 个模型，/test 回 404 提示的回执，没选模型的 external 回 400「请选择模型。」；旧的 /api/webui/providers/llm%3Allamacpp/panels/settings/state 回 404；manifest 里 llm:* 页只剩编辑页段落与模块面板。

没验证

• 真实模型服务上的测试连接（dev 假数据里只有本机假 llama-server，它不答 /chat/completions，所以测试回执是 404 的失败提示）。
• 外部客户端是否还在调旧接口：仓内与 extensions/ 下的 provider 扩展都没有调用。

Fixes #365

English

Removes the old llm-settings panel that the unified provider editor replaced: its server-side declaration and operations, the WebUI panel and the built-in panel registry, and the strings and tests that served only it. @kwk1001 this follows the boundary you listed in the issue: ProviderSettings, sources(), module sections with their invoke dispatch, and the underlying save, secret status, probe and pricing capabilities stay; the pricing editor and its strings in all eleven languages move into features/providers/ before the old directory goes; the public builtin protocol field stays; old #/provider/llm:<kind> links still redirect to the providers page.

Two commits:

• refactor(webui)!: pricing-panel.ts moves to src/web/client/features/providers/pricing.ts and its nineteen strings join the providers page tables; webui-pages/builtins/ and builtins.ts are deleted. WebUiPageHost no longer takes builtins; a panel that declares builtin gets an error card on all three mount paths (tab, slot, connection editor) without loading the page's extension.
• refactor(providers/webui)!: llm:<module> pages no longer declare the settings panel and the page invoke forwards only to the module's own panels; ProviderSettings.activate/delete/setSecret, the private assertNewName/directoryExtras/probe and seven strings used only by them are deleted. ProviderHub.action without a draft now uses the saved endpoint's instance directly (registry.resolve for models, ProviderSettings.probeClient for the probe); request and receipt shapes are unchanged and a draft is still optional. An endpoint without a model now reports the same "Model is required." as a draft.

BREAKING

/api/webui/providers/llm:<module>/panels/settings/<method> is removed, covering state, create, activate, save, delete, duplicate, setSecret, models, probe and extras. Use /api/providers: list GET /api/providers, detail GET /api/providers/:name, create POST /api/providers (duplicate is a create with copyFrom), save POST /api/providers/:name/save (whole entry, with expectedRevision, key value in secretValue), activate POST /api/providers/:name/activate, delete POST /api/providers/:name/delete, probe and model list POST /api/providers/:name/test and /models; module reasoning tiers, service tiers and default URL are in GET /api/provider-modules. extras (listing files in the endpoint directory) has no replacement. The WebUI no longer ships the llm-settings panel; a panel declaring builtin: 'llm-settings' shows an error card. No deprecation period.

Verified

• pnpm run typecheck, pnpm typecheck:web, full pnpm test, pnpm build:web.
• With scripts/dev-webui.ts fake data (port 8871): #/provider/llm:llamacpp lands on #/providers; the llamacpp runtime and models module panels mount; on the local endpoint "Test connection" returns a receipt (the fake llama-server does not answer /chat/completions, so it is the 404 hint) and "Fetch models" lists 2 models; filling currency and output rate in the three-rate pricing form updates the full-rules JSON and marks the card as a draft, and after saving GET /api/providers/local returns that pricing; a new llamacpp connection llama-two was created, given a model from the fetched list, saved, and listed as ready.
• curl without a draft: POST /api/providers/local/models returns 2 models, /test returns the 404-hint receipt, external (no model) returns 400 "请选择模型。"; the old /api/webui/providers/llm%3Allamacpp/panels/settings/state returns 404; llm:* pages in the manifest carry only the editor blocks and module panels.

Not verified

• A probe against a real model service (the dev fixture's local fake llama-server does not answer /chat/completions, so the probe receipt is the 404 hint).
• Whether any external client still calls the old routes; no provider extension in this repository or under extensions/ does.

🤖 Generated with Claude Code