@youweichen/pi-web-ui 0.45.0

package/dist/server/model-admin.js

@@ -0,0 +1,821 @@
+/**
+ * model-admin — 模型/服务商配置管理，从 agent-service.ts 抽出。
+ *
+ * 职责：auth.json 的 provider api-key 存取（set/clear）、models.json 读写
+ * （listModelsConfig/saveModelConfig/deleteModelConfig）、自定义服务商「自动获取
+ * 模型列表」（fetch_models：服务端探测 OpenAI 兼容 /models 端点，绕开 CORS；
+ * anthropic/google 鉴权头各不同；裸 /models 404 回退 /v1/models）与已保存供应商
+ * 的一键刷新（refresh_provider_models，凭据不出浏览器）。改动后热更新 runtime
+ * （refresh/setRuntimeApiKey）并推 models/models_config。
+ *
+ * 经 ModelAdminHost 与 ClientSession 解耦（同 settings/goal/slash 服务模式）。
+ * UI 文案直接中文（服务端 notice 约定）。apiKey/headers 绝不下发浏览器。
+ */
+import { existsSync, mkdirSync, readFileSync, writeFileSync } from "node:fs";
+import { dirname, join } from "node:path";
+/** Strip // and /* *\/ comments without touching string literals (URLs contain //). */
+function stripJsonComments(src) {
+    let out = "";
+    let inString = false;
+    let i = 0;
+    while (i < src.length) {
+        const c = src[i];
+        const next = src[i + 1];
+        if (inString) {
+            out += c;
+            if (c === "\\") {
+                out += next ?? "";
+                i += 2;
+                continue;
+            }
+            if (c === '"')
+                inString = false;
+            i++;
+            continue;
+        }
+        if (c === '"') {
+            inString = true;
+            out += c;
+            i++;
+            continue;
+        }
+        if (c === "/" && next === "/") {
+            while (i < src.length && src[i] !== "\n")
+                i++;
+            continue;
+        }
+        if (c === "/" && next === "*") {
+            i += 2;
+            while (i < src.length && !(src[i] === "*" && src[i + 1] === "/"))
+                i++;
+            i += 2;
+            continue;
+        }
+        out += c;
+        i++;
+    }
+    return out;
+}
+/** Numeric metadata value (NaN/string "unknown" → undefined). */
+function numMeta(v) {
+    return typeof v === "number" && Number.isFinite(v) ? v : undefined;
+}
+function boolMeta(v) {
+    return typeof v === "boolean" ? v : undefined;
+}
+function strArrMeta(v) {
+    return Array.isArray(v)
+        ? v.filter((x) => typeof x === "string")
+        : undefined;
+}
+/** Best-effort extraction of model metadata from an OpenAI-compatible
+ *  /models `data[]` item. Most endpoints only return `{ id }` — the extra
+ *  fields (context_window / max_model_len / modalities / supports_vision /
+ *  reasoning / display_name) come from vLLM and other extended
+ *  implementations, and are filled into the form when present. */
+function parseOpenAiModel(m) {
+    const r = (m ?? {});
+    const id = typeof r.id === "string" ? r.id : "";
+    const name = (typeof r.name === "string" && r.name.trim() ? r.name : undefined) ??
+        (typeof r.display_name === "string" && r.display_name.trim()
+            ? r.display_name
+            : undefined);
+    const modalities = strArrMeta(r.modalities) ??
+        strArrMeta(r.input_modalities);
+    const vision = modalities?.includes("image") === true ||
+        boolMeta(r.supports_vision) === true ||
+        boolMeta(r.vision) === true ||
+        strArrMeta(r.input)?.includes("image") === true;
+    const reasoning = boolMeta(r.reasoning) === true ||
+        boolMeta(r.supports_reasoning) === true ||
+        modalities?.includes("reasoning") === true;
+    const contextWindow = numMeta(r.context_window) ??
+        numMeta(r.context_length) ??
+        numMeta(r.max_model_len) ??
+        numMeta(r.max_context_length);
+    const maxTokens = numMeta(r.max_tokens) ??
+        numMeta(r.max_output_tokens) ??
+        numMeta(r.max_completion_tokens);
+    return {
+        id,
+        ...(name ? { name } : {}),
+        ...(reasoning ? { reasoning: true } : {}),
+        ...(vision ? { input: ["text", "image"] } : {}),
+        ...(contextWindow ? { contextWindow } : {}),
+        ...(maxTokens ? { maxTokens } : {}),
+    };
+}
+/** google-generative-ai /models shape:
+ *  { models: [{ name: "models/gemini-flash", displayName, inputTokenLimit,
+ *               outputTokenLimit, supportedGenerationMethods }] } */
+function parseGoogleModel(m) {
+    const r = (m ?? {});
+    const rawName = typeof r.name === "string" ? r.name : "";
+    const id = rawName.replace(/^models\//, "");
+    const displayName = typeof r.displayName === "string" ? r.displayName : undefined;
+    return {
+        id,
+        ...(displayName && displayName !== id ? { name: displayName } : {}),
+        ...(numMeta(r.inputTokenLimit)
+            ? { contextWindow: numMeta(r.inputTokenLimit) }
+            : {}),
+        ...(numMeta(r.outputTokenLimit)
+            ? { maxTokens: numMeta(r.outputTokenLimit) }
+            : {}),
+    };
+}
+export class ModelAdminService {
+    host;
+    constructor(host) {
+        this.host = host;
+    }
+    /** Persist an api-key credential for a provider (auth.json) and apply it now. */
+    async setProviderApiKey(provider, apiKey) {
+        const key = apiKey.trim();
+        if (!provider.trim()) {
+            this.host.emit({ type: "notice", level: "error", text: "请填写服务商 ID" });
+            return;
+        }
+        if (!key) {
+            this.host.emit({ type: "notice", level: "error", text: "请填写 API 密钥" });
+            return;
+        }
+        try {
+            // Persist to auth.json (auth.json shape: { <provider>: { type: "api_key", key } }).
+            const authPath = join(this.host.agentDir, "auth.json");
+            mkdirSync(this.host.agentDir, { recursive: true });
+            let data = {};
+            try {
+                data = JSON.parse(readFileSync(authPath, "utf8"));
+            }
+            catch {
+                // no file yet / unparsable — start fresh
+            }
+            data[provider.trim()] = { type: "api_key", key };
+            writeFileSync(authPath, JSON.stringify(data, null, 2) + "\n");
+            // Apply immediately for this session (runtime credentials are cached), then
+            // refresh models. allowNetwork downloads the provider's official model
+            // catalog (openai/anthropic/… are dynamic providers with no built-in list).
+            const mr = this.host.modelRuntime();
+            await mr.setRuntimeApiKey(provider.trim(), key);
+            await mr.refresh({ allowNetwork: true });
+            this.host.invalidatePiConfig();
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: `✅ 已保存 ${provider.trim()} 的 API 密钥并刷新模型列表`,
+            });
+            await this.host.pushModels();
+            await this.listProviders();
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `保存 API 密钥失败：${err.message}`,
+            });
+        }
+        this.host.flushSnapshot();
+    }
+    /**
+     * Clear a built-in provider's stored API key (auth.json entry + runtime
+     * override) so it returns to the unconfigured state — its models disappear
+     * from the picker until a key is set again. Only meaningful for keys that
+     * were stored via set_provider_api_key (source "stored"); env-var sourced
+     * credentials can't be cleared from here.
+     */
+    async clearProviderApiKey(provider) {
+        const pid = provider.trim();
+        if (!pid) {
+            this.host.emit({ type: "notice", level: "error", text: "请填写服务商 ID" });
+            return;
+        }
+        try {
+            // Remove from auth.json ({ <provider>: { type: "api_key", key } }).
+            const authPath = join(this.host.agentDir, "auth.json");
+            let data = {};
+            try {
+                data = JSON.parse(readFileSync(authPath, "utf8"));
+            }
+            catch {
+                // no file yet / unparsable — nothing stored to clear
+            }
+            if (!(pid in data)) {
+                this.host.emit({
+                    type: "notice",
+                    level: "info",
+                    text: `${pid} 没有已保存的密钥`,
+                });
+                return;
+            }
+            delete data[pid];
+            writeFileSync(authPath, JSON.stringify(data, null, 2) + "\n");
+            // Drop the runtime override too, then re-read credentials so the
+            // provider goes back to unconfigured and its models leave the list.
+            const mr = this.host.modelRuntime();
+            await mr.removeRuntimeApiKey(pid);
+            await mr.refresh();
+            this.host.invalidatePiConfig();
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: `🗑  已清除 ${pid} 的密钥，该服务商回到未配置状态`,
+            });
+            await this.host.pushModels();
+            await this.listProviders();
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `清除密钥失败：${err.message}`,
+            });
+        }
+        this.host.flushSnapshot();
+    }
+    /**
+     * Copy a BUILT-IN provider (baseUrl + current model catalog) into an
+     * editable custom-provider draft and return it via clone_provider_result.
+     * Nothing is persisted — the user renames the draft, pastes a DIFFERENT
+     * API key in the form, then saves via save_model_config. Credentials are
+     * never copied: the whole point is running a second key alongside the
+     * built-in one without touching it.
+     */
+    async cloneProvider(providerId, reqId) {
+        const pid = providerId.trim();
+        const fail = (error) => this.host.emit({ type: "clone_provider_result", reqId, ok: false, error });
+        try {
+            if (!pid) {
+                fail("请填写服务商 ID");
+                return;
+            }
+            const mr = this.host.modelRuntime();
+            const p = mr.getProvider(pid);
+            if (!p) {
+                fail(`供应商 ${pid} 不存在`);
+                return;
+            }
+            if (!p.baseUrl) {
+                fail(`${pid} 没有 baseUrl（OAuth/环境变量型供应商），无法复制为自定义服务商`);
+                return;
+            }
+            // Map runtime models → models.json rows; dynamic providers ship an
+            // empty catalog until refreshed over the network.
+            const readModels = () => {
+                try {
+                    return mr.getModels(pid).map((m) => ({
+                        api: m.api,
+                        entry: {
+                            id: m.id,
+                            ...(m.name && m.name !== m.id ? { name: m.name } : {}),
+                            ...(m.reasoning ? { reasoning: true } : {}),
+                            ...(m.input?.includes("image")
+                                ? { input: ["text", "image"] }
+                                : {}),
+                            ...(m.contextWindow ? { contextWindow: m.contextWindow } : {}),
+                            ...(m.maxTokens ? { maxTokens: m.maxTokens } : {}),
+                        },
+                    }));
+                }
+                catch {
+                    return [];
+                }
+            };
+            let models = readModels();
+            if (models.length === 0) {
+                await mr.refresh({ allowNetwork: true });
+                models = readModels();
+            }
+            if (models.length === 0) {
+                fail(`${pid} 的模型列表为空，无法复制（请稍后重试）`);
+                return;
+            }
+            // models.json 的 api 是 provider 级：取占比最高的 api，只复制该 api 的模型。
+            const counts = new Map();
+            for (const m of models)
+                counts.set(m.api, (counts.get(m.api) ?? 0) + 1);
+            let api = models[0].api;
+            for (const [k, v] of counts)
+                if (v > (counts.get(api) ?? 0))
+                    api = k;
+            const kept = models.filter((m) => m.api === api).map((m) => m.entry);
+            // Suggest a free id (<pid>-2, -3, …) — save_model_config would silently
+            // overwrite an existing custom entry with the same id.
+            const taken = new Set([
+                ...Object.keys(this.readModelsConfig().providers),
+                ...mr.getRegisteredProviderIds(),
+            ]);
+            let newId = `${pid}-2`;
+            for (let n = 2; taken.has(newId); n++)
+                newId = `${pid}-${n}`;
+            const config = {
+                providerId: newId,
+                name: p.name,
+                api,
+                baseUrl: p.baseUrl,
+                models: kept,
+            };
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: `📋 已复制 ${pid} → ${newId}（${kept.length} 个模型），请填入新的 API 密钥后保存`,
+            });
+            this.host.emit({ type: "clone_provider_result", reqId, ok: true, config });
+        }
+        catch (err) {
+            fail(`复制服务商失败：${err.message}`);
+        }
+        this.host.flushSnapshot();
+    }
+    /** Enumerate pi's built-in providers with auth status (key-only config). */
+    async listProviders() {
+        const mr = this.host.modelRuntime();
+        let providers;
+        try {
+            providers = mr.getProviders().map((p) => {
+                try {
+                    const st = mr.getProviderAuthStatus(p.id);
+                    return {
+                        id: p.id,
+                        name: p.name,
+                        configured: st?.configured ?? false,
+                        source: st?.source,
+                    };
+                }
+                catch {
+                    // One odd provider must not blank the whole list.
+                    return { id: p.id, name: p.name, configured: false };
+                }
+            });
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `获取服务商列表失败：${err.message}`,
+            });
+            return;
+        }
+        if (providers.length === 0) {
+            this.host.emit({
+                type: "notice",
+                level: "warning",
+                text: "服务商列表为空——pi 运行时未注册任何提供商",
+            });
+        }
+        this.host.emit({ type: "providers_status", providers });
+    }
+    // ---------------------------------------------------------------------------
+    // Custom model config (agentDir/models.json)
+    // ---------------------------------------------------------------------------
+    modelsConfigPath() {
+        return join(this.host.agentDir, "models.json");
+    }
+    /** Strip // and /* *\/ comments without touching string literals (URLs contain //). */
+    static stripJsonComments(src) {
+        let out = "";
+        let inString = false;
+        let i = 0;
+        while (i < src.length) {
+            const c = src[i];
+            const next = src[i + 1];
+            if (inString) {
+                out += c;
+                if (c === "\\") {
+                    out += next ?? "";
+                    i += 2;
+                    continue;
+                }
+                if (c === '"')
+                    inString = false;
+                i++;
+                continue;
+            }
+            if (c === '"') {
+                inString = true;
+                out += c;
+                i++;
+                continue;
+            }
+            if (c === "/" && next === "/") {
+                while (i < src.length && src[i] !== "\n")
+                    i++;
+                continue;
+            }
+            if (c === "/" && next === "*") {
+                i += 2;
+                while (i < src.length && !(src[i] === "*" && src[i + 1] === "/"))
+                    i++;
+                i += 2;
+                continue;
+            }
+            out += c;
+            i++;
+        }
+        return out;
+    }
+    /** Read + parse models.json (tolerating // and /* *\/ comments like the SDK). */
+    readModelsConfig() {
+        const path = this.modelsConfigPath();
+        try {
+            const raw = readFileSync(path, "utf8");
+            const parsed = JSON.parse(stripJsonComments(raw));
+            return { providers: parsed?.providers ?? {} };
+        }
+        catch {
+            return { providers: {} };
+        }
+    }
+    /** Send the current models.json custom providers to the client. */
+    async listModelsConfig() {
+        const { providers } = this.readModelsConfig();
+        const list = Object.entries(providers).map(([providerId, p]) => {
+            const models = Array.isArray(p.models)
+                ? p.models.map((m) => ({
+                    id: String(m.id ?? ""),
+                    name: m.name,
+                    reasoning: m.reasoning,
+                    input: Array.isArray(m.input) ? m.input : undefined,
+                    contextWindow: m.contextWindow,
+                    maxTokens: m.maxTokens,
+                }))
+                : [];
+            return {
+                providerId,
+                name: p.name,
+                api: p.api,
+                baseUrl: p.baseUrl,
+                apiKey: p.apiKey,
+                authHeader: p.authHeader,
+                // headers are intentionally NOT sent to the browser — they may
+                // contain Authorization / API-key values; kept server-side only.
+                models,
+            };
+        });
+        this.host.emit({ type: "models_config", providers: list });
+    }
+    /** Numeric metadata value (NaN/string "unknown" → undefined). */
+    static numMeta(v) {
+        return typeof v === "number" && Number.isFinite(v) ? v : undefined;
+    }
+    static boolMeta(v) {
+        return typeof v === "boolean" ? v : undefined;
+    }
+    static strArrMeta(v) {
+        return Array.isArray(v)
+            ? v.filter((x) => typeof x === "string")
+            : undefined;
+    }
+    /** Best-effort extraction of model metadata from an OpenAI-compatible
+     *  /models `data[]` item. Most endpoints only return `{ id }` — the extra
+     *  fields (context_window / max_model_len / modalities / supports_vision /
+     *  reasoning / display_name) come from vLLM and other extended
+     *  implementations, and are filled into the form when present. */
+    static parseOpenAiModel(m) {
+        const r = (m ?? {});
+        const id = typeof r.id === "string" ? r.id : "";
+        const name = (typeof r.name === "string" && r.name.trim() ? r.name : undefined) ??
+            (typeof r.display_name === "string" && r.display_name.trim()
+                ? r.display_name
+                : undefined);
+        const modalities = strArrMeta(r.modalities) ??
+            strArrMeta(r.input_modalities);
+        const vision = modalities?.includes("image") === true ||
+            boolMeta(r.supports_vision) === true ||
+            boolMeta(r.vision) === true ||
+            strArrMeta(r.input)?.includes("image") === true;
+        const reasoning = boolMeta(r.reasoning) === true ||
+            boolMeta(r.supports_reasoning) === true ||
+            modalities?.includes("reasoning") === true;
+        const contextWindow = numMeta(r.context_window) ??
+            numMeta(r.context_length) ??
+            numMeta(r.max_model_len) ??
+            numMeta(r.max_context_length);
+        const maxTokens = numMeta(r.max_tokens) ??
+            numMeta(r.max_output_tokens) ??
+            numMeta(r.max_completion_tokens);
+        return {
+            id,
+            ...(name ? { name } : {}),
+            ...(reasoning ? { reasoning: true } : {}),
+            ...(vision ? { input: ["text", "image"] } : {}),
+            ...(contextWindow ? { contextWindow } : {}),
+            ...(maxTokens ? { maxTokens } : {}),
+        };
+    }
+    /** google-generative-ai /models shape:
+     *  { models: [{ name: "models/gemini-flash", displayName, inputTokenLimit,
+     *               outputTokenLimit, supportedGenerationMethods }] } */
+    static parseGoogleModel(m) {
+        const r = (m ?? {});
+        const rawName = typeof r.name === "string" ? r.name : "";
+        const id = rawName.replace(/^models\//, "");
+        const displayName = typeof r.displayName === "string" ? r.displayName : undefined;
+        return {
+            id,
+            ...(displayName && displayName !== id ? { name: displayName } : {}),
+            ...(numMeta(r.inputTokenLimit)
+                ? { contextWindow: numMeta(r.inputTokenLimit) }
+                : {}),
+            ...(numMeta(r.outputTokenLimit)
+                ? { maxTokens: numMeta(r.outputTokenLimit) }
+                : {}),
+        };
+    }
+    /** Probe a custom provider's OpenAI-compatible /models endpoint (server-side
+     *  because the baseUrl is often a LAN/loopback host the browser can't reach
+     *  cross-origin) and return the advertised models. reqId is echoed back
+     *  in fetch_models_result so the UI can match concurrent requests. */
+    async fetchModelsList(reqId, baseUrl, apiKey, authHeader, api) {
+        const emitError = (error) => this.host.emit({ type: "fetch_models_result", reqId, ok: false, error });
+        try {
+            const models = await ModelAdminService.probeModelsEndpoint(baseUrl, apiKey, authHeader, api);
+            this.host.emit({ type: "fetch_models_result", reqId, ok: true, models });
+        }
+        catch (err) {
+            emitError(err.message);
+        }
+    }
+    /**
+     * Probe a custom provider's model-list endpoint (OpenAI-compatible /models
+     * with a /v1 retry; Google {models:[…]} shape supported). Throws Error with
+     * a user-facing message on any failure; returns deduped+sorted entries.
+     * Shared by the edit-form "auto fetch" and the saved-provider refresh.
+     */
+    static async probeModelsEndpoint(baseUrl, apiKey, authHeader, api, extraHeaders) {
+        const base = (baseUrl ?? "").trim().replace(/\/+$/, "");
+        if (!base)
+            throw new Error("请先填写 baseUrl");
+        let url;
+        try {
+            url = new URL(base);
+        }
+        catch {
+            throw new Error(`baseUrl 无效：${base}`);
+        }
+        if (url.protocol !== "http:" && url.protocol !== "https:") {
+            throw new Error("baseUrl 仅支持 http/https");
+        }
+        const headers = {
+            ...(extraHeaders ?? {}),
+        };
+        // Per-api auth conventions (mirror pi's built-in provider configs):
+        //   openai-*:      Authorization: Bearer <key>
+        //   anthropic:     x-api-key + anthropic-version
+        //   google:        x-goog-api-key
+        // authHeader=false → no auth header at all (custom gateways).
+        if (apiKey?.trim() && authHeader !== false) {
+            const key = apiKey.trim();
+            if (api === "anthropic-messages") {
+                headers["x-api-key"] = key;
+                headers["anthropic-version"] = "2023-06-01";
+            }
+            else if (api === "google-generative-ai") {
+                headers["x-goog-api-key"] = key;
+            }
+            else {
+                headers["Authorization"] = `Bearer ${key}`;
+            }
+        }
+        const tryFetch = async (u) => {
+            const ac = new AbortController();
+            const timer = setTimeout(() => ac.abort(), 15000);
+            try {
+                return await fetch(u, { headers, signal: ac.signal });
+            }
+            catch (err) {
+                if (err.name === "AbortError") {
+                    throw new Error("请求超时（15 秒）");
+                }
+                throw new Error(`请求失败：${err.message}`);
+            }
+            finally {
+                clearTimeout(timer);
+            }
+        };
+        let res = await tryFetch(`${base}/models`);
+        // BaseUrls that omit the /v1 prefix (e.g. https://api.openai.com) 404 on
+        // the bare path — retry under /v1.
+        if (res && res.status === 404 && !/\/v\d+[a-z-]*$/.test(base)) {
+            res = await tryFetch(`${base}/v1/models`);
+        }
+        if (!res)
+            throw new Error("请求失败");
+        if (!res.ok) {
+            let detail = "";
+            try {
+                detail = (await res.text()).slice(0, 200);
+            }
+            catch {
+                // response body already consumed / not text — ignore
+            }
+            throw new Error(`接口返回 HTTP ${res.status}${detail ? `：${detail}` : ""}`);
+        }
+        let models = [];
+        try {
+            const json = (await res.json());
+            const data = Array.isArray(json.data) ? json.data : null;
+            if (data) {
+                // OpenAI-compatible: { data: [{ id, context_window, modalities, … }] }
+                models = data
+                    .map((m) => parseOpenAiModel(m))
+                    .filter((m) => m.id);
+            }
+            else if (Array.isArray(json.models)) {
+                // Google: { models: [{ name: "models/…", displayName, … }] }
+                models = json.models
+                    .map((m) => parseGoogleModel(m))
+                    .filter((m) => m.id);
+            }
+        }
+        catch {
+            throw new Error("响应不是有效的 JSON");
+        }
+        // Dedupe by id (keep the first, most complete entry) and sort by id.
+        const seen = new Set();
+        models = models
+            .filter((m) => (seen.has(m.id) ? false : (seen.add(m.id), true)))
+            .sort((a, b) => a.id.localeCompare(b.id));
+        if (models.length === 0)
+            throw new Error("接口未返回任何模型");
+        return models;
+    }
+    /**
+     * Re-probe a SAVED custom provider's model list and merge it into its
+     * models.json entry — credentials never leave the server (unlike the
+     * edit-form fetch, which sends whatever the browser typed). Merge rules:
+     * existing ids keep all manually-entered fields and only gain metadata
+     * they were missing; brand-new ids are appended. Hot-reloads the runtime.
+     */
+    async refreshProviderModels(providerId, reqId) {
+        const done = (ok, extra = {}) => this.host.emit({ type: "refresh_provider_result", reqId, ok, ...extra });
+        try {
+            const pid = providerId.trim();
+            const { providers } = this.readModelsConfig();
+            // models.json 原始形状是 Record<string, unknown>——按已保存条目的结构断言
+            const saved = providers[pid];
+            if (!saved?.baseUrl?.trim()) {
+                this.host.emit({
+                    type: "notice",
+                    level: "warning",
+                    text: `服务商 ${pid} 不存在或未配置 baseUrl，无法刷新`,
+                });
+                return done(false, { error: "provider missing or no baseUrl" });
+            }
+            const fetched = await ModelAdminService.probeModelsEndpoint(saved.baseUrl, saved.apiKey, saved.authHeader === true ? true : undefined, saved.api, saved.headers);
+            // Merge: manual values win; fetched fills blanks and appends new ids.
+            const prev = new Map((saved.models ?? []).map((m) => [m.id, m]));
+            let added = 0;
+            for (const f of fetched) {
+                const cur = prev.get(f.id);
+                if (!cur) {
+                    prev.set(f.id, f);
+                    added += 1;
+                    continue;
+                }
+                prev.set(f.id, {
+                    ...f,
+                    ...cur, // 手填字段优先：cur 覆盖 f 的同名字段
+                });
+            }
+            const merged = [...prev.values()].sort((a, b) => a.id.localeCompare(b.id));
+            await this.saveModelConfig(pid, {
+                providerId: pid,
+                name: saved.name,
+                api: saved.api,
+                baseUrl: saved.baseUrl,
+                // apiKey/headers 不回传浏览器——saveModelConfig 会保留旧值
+                authHeader: saved.authHeader === true ? true : undefined,
+                models: merged,
+            });
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: added > 0
+                    ? `🔄 已刷新 ${pid}：新增 ${added} 个模型，共 ${merged.length} 个`
+                    : `🔄 已刷新 ${pid}：无新增模型（共 ${merged.length} 个）`,
+            });
+            return done(true, { added, total: merged.length });
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `刷新模型列表失败：${err.message}`,
+            });
+            return done(false, { error: err.message });
+        }
+    }
+    /** Upsert one provider into models.json and hot-reload the model runtime. */
+    async saveModelConfig(providerId, config) {
+        const pid = providerId.trim();
+        if (!pid || !/^[\w.-]+$/.test(pid)) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: "服务商 ID 无效（仅字母/数字/._-）",
+            });
+            return;
+        }
+        const models = (config.models ?? [])
+            .filter((m) => m.id && m.id.trim())
+            .map((m) => ({
+            id: m.id.trim(),
+            ...(m.name?.trim() ? { name: m.name.trim() } : {}),
+            ...(m.reasoning ? { reasoning: true } : {}),
+            ...(m.input?.length ? { input: m.input } : {}),
+            ...(m.contextWindow ? { contextWindow: Number(m.contextWindow) } : {}),
+            ...(m.maxTokens ? { maxTokens: Number(m.maxTokens) } : {}),
+        }));
+        if (models.length === 0) {
+            this.host.emit({ type: "notice", level: "error", text: "至少需要一个模型" });
+            return;
+        }
+        try {
+            const { providers } = this.readModelsConfig();
+            // headers never reach the browser, so the incoming config can't carry
+            // them — preserve the previously stored values when they are absent.
+            const prevHeaders = providers[pid]?.headers;
+            providers[pid] = {
+                ...(config.name?.trim() ? { name: config.name.trim() } : {}),
+                ...(config.api?.trim() ? { api: config.api.trim() } : {}),
+                ...(config.baseUrl?.trim() ? { baseUrl: config.baseUrl.trim() } : {}),
+                ...(config.apiKey?.trim() ? { apiKey: config.apiKey.trim() } : {}),
+                ...(config.authHeader ? { authHeader: true } : {}),
+                ...(prevHeaders && Object.keys(prevHeaders).length > 0
+                    ? { headers: prevHeaders }
+                    : {}),
+                models,
+            };
+            mkdirSync(this.host.agentDir, { recursive: true });
+            writeFileSync(this.modelsConfigPath(), JSON.stringify({ providers }, null, 2) + "\n");
+            // Allow a custom models.json entry to reuse the provider credential
+            // already stored in auth.json.  Seed the shared runtime too, because
+            // older pi-ai versions did not always fall back to stored credentials
+            // for a newly-created custom provider.  Never copy the secret into
+            // models.json.
+            try {
+                const auth = JSON.parse(readFileSync(join(this.host.agentDir, "auth.json"), "utf8"));
+                const credential = auth[pid];
+                if (credential &&
+                    typeof credential === "object" &&
+                    "key" in credential &&
+                    typeof credential.key === "string" &&
+                    credential.key.trim()) {
+                    await this.host.modelRuntime().setRuntimeApiKey(pid, credential.key);
+                }
+            }
+            catch {
+                // auth.json is optional; models.json can still use its own apiKey.
+            }
+            await this.host.modelRuntime().refresh();
+            await this.listModelsConfig();
+            await this.host.pushModels();
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: `✅ 已保存服务商 ${pid}（${models.length} 个模型）并刷新模型列表`,
+            });
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `保存模型配置失败：${err.message}`,
+            });
+        }
+        this.host.flushSnapshot();
+    }
+    /** Remove a provider from models.json and hot-reload. */
+    async deleteModelConfig(providerId) {
+        try {
+            const { providers } = this.readModelsConfig();
+            if (!(providerId in providers)) {
+                this.host.emit({
+                    type: "notice",
+                    level: "info",
+                    text: `服务商 ${providerId} 不存在`,
+                });
+                return;
+            }
+            delete providers[providerId];
+            writeFileSync(this.modelsConfigPath(), JSON.stringify({ providers }, null, 2) + "\n");
+            await this.host.modelRuntime().refresh();
+            await this.listModelsConfig();
+            await this.host.pushModels();
+            this.host.emit({
+                type: "notice",
+                level: "info",
+                text: `🗑  已删除服务商 ${providerId}`,
+            });
+        }
+        catch (err) {
+            this.host.emit({
+                type: "notice",
+                level: "error",
+                text: `删除模型配置失败：${err.message}`,
+            });
+        }
+        this.host.flushSnapshot();
+    }
+}