@cjhyy/code-shell-core 0.5.0-rc.2 → 0.6.0-rc.2

package/dist/credentials/use-credential-tool.js

@@ -0,0 +1,205 @@
+/**
+ * UseCredential 工具(凭证模块第二期 §3-§5)。
+ *
+ * 对 AI 暴露**一个统一工具**取用已存凭证(token / link / cookie 同构,描述只说「凭证」)。
+ * - 无 `id` → 返回脱敏清单(权威实时源,兜底动态描述的滞后)。
+ * - 有 `id` → 过 CredentialUseGate(默认问 / 本会话记住 / 全自动)后:
+ *     - token/link → `{ kind: "value", value }`
+ *     - cookie     → 就地写临时 cookies.txt(0600),返回 `{ kind: "cookie", cookiesFile, count }`
+ *
+ * 取值全部 core 直读 CredentialStore(cookie 值第二期已进库),无跨进程。
+ * 集中在 core/src/credentials/ 下,只经 ToolDefinition 注册 + ToolContext.askUser 耦合 core,
+ * 满足设计稿 §1「可整块外移」约束。
+ */
+import { writeFileSync, existsSync, readdirSync, statSync, rmSync } from "node:fs";
+import { join } from "node:path";
+import { tmpdir } from "node:os";
+import { randomUUID } from "node:crypto";
+import { CredentialStore } from "./store.js";
+import { formatNetscapeCookies, parseCookieJar } from "./cookie-jar.js";
+import { credentialUseGate, } from "./use-gate.js";
+import { SettingsManager } from "../settings/manager.js";
+import { logger } from "../logging/logger.js";
+const TOOL_NAME = "UseCredential";
+const COOKIE_FILE_PREFIX = "codeshell-cred-cookie-";
+const COOKIE_FILE_MAX_AGE_MS = 30 * 60 * 1000; // 30min 启动 sweep 上限
+const BASE_DESCRIPTION = "Use a stored credential (token / API key / login cookie) to run a command. " +
+    "Call with NO arguments first to list available credentials (id + label + type); " +
+    "then call again with `id` to fetch one. Token/link credentials return their secret " +
+    "value; cookie credentials are materialized to a temporary Netscape cookies.txt file " +
+    "(use it as `yt-dlp --cookies <cookiesFile>` / `curl -b <cookiesFile>`). " +
+    "Each use is gated by a quick user approval unless auto-approve is on.";
+export const useCredentialToolDef = {
+    name: TOOL_NAME,
+    description: BASE_DESCRIPTION,
+    inputSchema: {
+        type: "object",
+        properties: {
+            id: {
+                type: "string",
+                description: "Credential id to fetch. Omit to list all available credentials.",
+            },
+            purpose: {
+                type: "string",
+                description: "Short reason you need it (shown in the approval prompt).",
+            },
+        },
+    },
+};
+/**
+ * 动态描述:在基础描述末尾附当前可用凭证清单(镜像 generateVideoToolDefFor)。
+ * 空时回退基础描述。AI 一搜出/一看到就知道有哪些可用,减少无谓的无参列举调用。
+ */
+export function useCredentialToolDefFor(cwd) {
+    try {
+        const list = new CredentialStore(cwd).listMasked();
+        if (list.length === 0)
+            return useCredentialToolDef;
+        const names = list.map((c) => `${c.id} (${c.type})`).join(", ");
+        return {
+            ...useCredentialToolDef,
+            description: `${BASE_DESCRIPTION}\nCurrently available: ${names}.`,
+        };
+    }
+    catch {
+        return useCredentialToolDef;
+    }
+}
+/**
+ * 启动期清理上次遗留的临时 cookies.txt(沿用 sweepStaleLeases 思路:扫同目录、按 mtime)。
+ * 不引入 lease 对象/定时器 —— 文件用完靠进程退出 + 这个轻量 sweep。
+ */
+export function sweepStaleCredentialCookies(now = Date.now()) {
+    const dir = tmpdir();
+    try {
+        if (!existsSync(dir))
+            return;
+        for (const f of readdirSync(dir)) {
+            if (!f.startsWith(COOKIE_FILE_PREFIX))
+                continue;
+            const p = join(dir, f);
+            try {
+                if (now - statSync(p).mtimeMs > COOKIE_FILE_MAX_AGE_MS)
+                    rmSync(p, { force: true });
+            }
+            catch {
+                /* skip */
+            }
+        }
+    }
+    catch {
+        /* best-effort */
+    }
+}
+/** 内存会话 allow 集:每个 Engine 一份(从 ctx.sessionId 键)。纯内存,关进程即忘。 */
+const sessionAllowByEngine = new Map();
+function sessionAllowFor(ctx) {
+    // 无 sessionId(headless / 临时 ToolContext / 某些子代理)绝不能共享一个
+    // 全局 "__nosession__" 桶 —— 否则一个上下文里「本会话记住」的批准会被
+    // 任意其它无 sessionId 的上下文复用,造成跨会话凭证串台。这种情况下返回
+    // 一个一次性 Set(不入 map):取用本身正常,但「记住」对它失效,每次重新过门。
+    if (!ctx?.sessionId)
+        return new Set();
+    const key = ctx.sessionId;
+    let set = sessionAllowByEngine.get(key);
+    if (!set) {
+        set = new Set();
+        sessionAllowByEngine.set(key, set);
+    }
+    return set;
+}
+// CredentialStore only distinguishes "full" (user+project) from "project"
+// (project-only). An "isolated" engine is at least as restrictive as project,
+// so it maps to "project" — it must never read the host user's ~/.code-shell.
+function credentialScope(scope) {
+    return scope === "full" || scope === undefined ? "full" : "project";
+}
+function readAutoApprove(cwd, scope) {
+    try {
+        // A project/isolated engine must read autoApprove from its own scope, not
+        // the host user's ~/.code-shell — otherwise a host `credentialUse.autoApprove`
+        // would silently auto-approve credential use inside an isolated engine.
+        const s = new SettingsManager(cwd, scope === "full" ? "full" : "project").get();
+        return s.credentialUse?.autoApprove === true;
+    }
+    catch {
+        return false;
+    }
+}
+export async function useCredentialTool(args, ctx) {
+    const cwd = ctx?.cwd ?? process.cwd();
+    const store = new CredentialStore(cwd);
+    const scope = credentialScope(ctx?.settingsScope);
+    const id = typeof args.id === "string" ? args.id.trim() : "";
+    const purpose = typeof args.purpose === "string" ? args.purpose : undefined;
+    // 无 id → 清单(脱敏,权威实时源)。按 engine scope 过滤:project/isolated
+    // 引擎不得列出宿主 user 层凭证。
+    if (!id) {
+        const credentials = store
+            .listMasked(scope)
+            .map((c) => ({ id: c.id, label: c.label, type: c.type }));
+        return json({ kind: "list", credentials });
+    }
+    const cred = store.resolve(id, scope);
+    if (!cred) {
+        return json({ kind: "error", error: `凭证不存在: "${id}"。调用本工具(无参)可列出可用凭证。` });
+    }
+    // 取值前过门
+    const ask = ctx?.askUser
+        ? (q, opts) => ctx.askUser(q, opts)
+        : undefined;
+    const decision = await credentialUseGate({ id: cred.id, label: cred.label, purpose }, {
+        autoApprove: readAutoApprove(cwd, scope),
+        credentialAutoUse: cred.autoUseByAI === true,
+        sessionAllow: sessionAllowFor(ctx),
+        ask,
+    });
+    if (!decision.allowed) {
+        const msg = decision.reason === "no-ui"
+            ? "无法取用凭证:当前无审批 UI(headless),且未开启 credentialUse.autoApprove。"
+            : `用户拒绝取用凭证「${cred.label}」。可回退 yt-dlp --cookies-from-browser 或提示用户。`;
+        return json({ kind: "error", error: msg });
+    }
+    // token/link → 直接返回值
+    if (cred.type === "token" || cred.type === "link") {
+        if (!cred.secret)
+            return json({ kind: "error", error: `凭证「${cred.label}」没有可用的值。` });
+        return json({ kind: "value", value: cred.secret });
+    }
+    // cookie → 就地写临时 cookies.txt
+    if (cred.type === "cookie") {
+        const jar = parseCookieJar(cred.secret);
+        if (jar.length === 0) {
+            return json({
+                kind: "error",
+                error: `凭证「${cred.label}」的 cookie 为空或已失效,请在凭证页对该账号点「重拓」(重新登录后重新拓取)。`,
+            });
+        }
+        // Unique per write: id + pid alone collides when two concurrent UseCredential
+        // calls (the tool is concurrency-safe) materialize the same credential in the
+        // same process — the second write would clobber the first and a caller could
+        // read another account's cookies. A random component makes each file distinct.
+        // Prefix is unchanged so the 30-min startup sweep still matches & cleans them.
+        const file = join(tmpdir(), `${COOKIE_FILE_PREFIX}${safe(cred.id)}-${process.pid}-${randomUUID()}.txt`);
+        try {
+            writeFileSync(file, formatNetscapeCookies(jar), { mode: 0o600 });
+        }
+        catch (e) {
+            logger.warn(`UseCredential: failed to write cookies.txt: ${String(e)}`);
+            return json({ kind: "error", error: `写临时 cookie 文件失败: ${String(e)}` });
+        }
+        return json({ kind: "cookie", cookiesFile: file, count: jar.length });
+    }
+    return json({ kind: "error", error: `未知凭证类型: ${cred.type}` });
+}
+/** 文件名安全化(凭证 id 用了 `__`,这里只挡路径分隔符/控制字符)。 */
+function safe(s) {
+    return s.replace(/[^a-zA-Z0-9_.-]/g, "_");
+}
+function json(r) {
+    return JSON.stringify(r);
+}
+/** 测试钩子:清空会话 allow 集(避免跨用例污染)。 */
+export function __resetCredentialSessionAllowForTests() {
+    sessionAllowByEngine.clear();
+}

package/dist/credentials/use-gate.d.ts

@@ -0,0 +1,56 @@
+/**
+ * CredentialUseGate — 凭证取用前的轻量三档审批(凭证模块第二期 §5)。
+ *
+ * AI 的 `UseCredential` 工具在把某条凭证的值/cookies.txt 交给命令**之前**过这道门:
+ *   1. **全自动**:settings `credentialUse.autoApprove === true` → 直接放行,不弹。
+ *   2. **本会话记住**:用户上次选了「本会话记住」→ 该凭证 id 进内存 allow 集 → 后续放行。
+ *   3. **默认弹审批**:经 `askUser`(InteractiveApprovalBackend)问一次。
+ *
+ * 关键决策(见设计稿 §5 + memory project_permission_session_cache):
+ * - 本会话记住**按凭证 id 键**,不按工具名 —— 否则一次批准会放行对无关凭证的取用。
+ * - 会话 allow 集是**纯内存**(关 app 即忘),由 Engine 持有并注入,故 gate 本身无模块级单例,
+ *   多 Engine 并发互不干扰、可整块外移。
+ *
+ * 边界:gate 只依赖 `askUser`(既有审批后端)+ 一个内存 Set + settings 读取,不耦合 core 其它部分。
+ */
+/** AI 取用一条凭证的请求(供审批文案 + 记忆键)。 */
+export interface CredentialUseRequest {
+    /** 凭证 id(记忆键)。 */
+    id: string;
+    /** 展示名(审批文案)。 */
+    label: string;
+    /** 可选用途(审批文案)。 */
+    purpose?: string;
+}
+export type CredentialUseDecision = {
+    allowed: true;
+} | {
+    allowed: false;
+    reason: "denied" | "no-ui";
+};
+/** 本会话已批准的凭证 id 集(纯内存,Engine 持有并跨 turn 复用)。 */
+export type SessionCredentialAllow = Set<string>;
+/** 审批问询函数(对接 ToolContext.askUser;返回用户所选 label)。 */
+export type CredentialAskFn = (question: string, opts: {
+    header?: string;
+    options: {
+        label: string;
+        description: string;
+        tone?: "ok" | "danger" | "neutral";
+    }[];
+    optionsOnly: true;
+}) => Promise<string>;
+export interface CredentialUseGateDeps {
+    /** settings `credentialUse.autoApprove`(全局总闸)。 */
+    autoApprove: boolean;
+    /** 该凭证自身的逐条「AI 可自动取用」标志(Credential.autoUseByAI);命中即放行。 */
+    credentialAutoUse?: boolean;
+    /** 本会话 allow 集(内存,可变;gate 命中「本会话记住」时写入)。 */
+    sessionAllow: SessionCredentialAllow;
+    /** 审批问询(undefined → 无 UI,headless 直接拒)。 */
+    ask?: CredentialAskFn;
+}
+/**
+ * 过门并返回是否放行。命中「本会话记住」会写入 `deps.sessionAllow`。
+ */
+export declare function credentialUseGate(req: CredentialUseRequest, deps: CredentialUseGateDeps): Promise<CredentialUseDecision>;

package/dist/credentials/use-gate.js

@@ -0,0 +1,52 @@
+/**
+ * CredentialUseGate — 凭证取用前的轻量三档审批(凭证模块第二期 §5)。
+ *
+ * AI 的 `UseCredential` 工具在把某条凭证的值/cookies.txt 交给命令**之前**过这道门:
+ *   1. **全自动**:settings `credentialUse.autoApprove === true` → 直接放行,不弹。
+ *   2. **本会话记住**:用户上次选了「本会话记住」→ 该凭证 id 进内存 allow 集 → 后续放行。
+ *   3. **默认弹审批**:经 `askUser`(InteractiveApprovalBackend)问一次。
+ *
+ * 关键决策(见设计稿 §5 + memory project_permission_session_cache):
+ * - 本会话记住**按凭证 id 键**,不按工具名 —— 否则一次批准会放行对无关凭证的取用。
+ * - 会话 allow 集是**纯内存**(关 app 即忘),由 Engine 持有并注入,故 gate 本身无模块级单例,
+ *   多 Engine 并发互不干扰、可整块外移。
+ *
+ * 边界:gate 只依赖 `askUser`(既有审批后端)+ 一个内存 Set + settings 读取,不耦合 core 其它部分。
+ */
+const ALLOW_ONCE = "允许本次";
+const ALLOW_SESSION = "本会话都允许";
+const DENY = "拒绝";
+/**
+ * 过门并返回是否放行。命中「本会话记住」会写入 `deps.sessionAllow`。
+ */
+export async function credentialUseGate(req, deps) {
+    // 1. 全自动(全局总闸,或该凭证逐条开了「AI 可自动取用」)
+    if (deps.autoApprove || deps.credentialAutoUse)
+        return { allowed: true };
+    // 2. 本会话记住
+    if (deps.sessionAllow.has(req.id))
+        return { allowed: true };
+    // 3. 默认弹审批
+    if (!deps.ask)
+        return { allowed: false, reason: "no-ui" };
+    const purpose = req.purpose?.trim();
+    const question = purpose
+        ? `AI 想用凭证「${req.label}」(${purpose}),是否允许?`
+        : `AI 想用凭证「${req.label}」,是否允许?`;
+    const choice = await deps.ask(question, {
+        header: "凭证取用",
+        options: [
+            { label: ALLOW_ONCE, description: "仅本次取用该凭证", tone: "ok" },
+            { label: ALLOW_SESSION, description: "本会话内该凭证不再询问(关 app 后失效)", tone: "ok" },
+            { label: DENY, description: "拒绝本次取用", tone: "danger" },
+        ],
+        optionsOnly: true,
+    });
+    if (choice === ALLOW_SESSION) {
+        deps.sessionAllow.add(req.id);
+        return { allowed: true };
+    }
+    if (choice === ALLOW_ONCE)
+        return { allowed: true };
+    return { allowed: false, reason: "denied" };
+}

package/dist/data/model-metadata.d.ts

@@ -0,0 +1,77 @@
+/**
+ * Hand-maintained model metadata for direct-provider models and onboarding
+ * picker tuning, extracted out of onboarding.ts so updating a model's
+ * max-output / context-window or the picker vendor list is a data edit
+ * (model-metadata.json) rather than a code change.
+ *
+ * Scope: only what the build-time OpenRouter snapshot does NOT cover —
+ * direct-provider ids (no "vendor/" prefix) for the metadata tables, plus the
+ * curated OpenRouter vendor ordering for the picker. OpenRouter "vendor/model"
+ * ids get their context/max-output from the snapshot (see data/openrouter-*).
+ *
+ * Loaded with createRequire so the bundled JSON ships in dist (the build copies
+ * src/data/*.json → dist/data/), matching static-catalogs.ts.
+ */
+export interface OpenRouterVendor {
+    prefix: string;
+    take: number;
+}
+/** Per-1M-token pricing (USD). Mirrors cost-tracker's ModelPricing. */
+export interface ModelPricing {
+    input: number;
+    output: number;
+    cacheRead: number;
+    cacheWrite: number;
+}
+/**
+ * Onboarding provider catalog entry. Pure data — the onboarding flow's logic
+ * (key validation, model registration, env detection) lives in onboarding.ts
+ * and consumes this shape. Lives here so the catalog is a data edit.
+ */
+export interface ProviderDef {
+    id: string;
+    name: string;
+    envKey: string;
+    provider: string;
+    baseUrl: string;
+    keyUrl: string;
+    keyPrefix: string;
+    /**
+     * The catalog of model IDs to register for this provider. The first entry
+     * doubles as the zero-config default — auto-activated when the user boots
+     * with just an env key (no onboarding). Users override in the wizard or via
+     * `/model`.
+     */
+    models: string[];
+    /** When true, skip API key prompt (e.g. local providers like Ollama). */
+    noKey?: boolean;
+}
+/** Known max output tokens for direct-provider models (id → tokens). */
+export declare const KNOWN_MAX_OUTPUT: Record<string, number>;
+/** Known context window sizes for direct-provider models (id → tokens). */
+export declare const KNOWN_CONTEXT_WINDOWS: Record<string, number>;
+/**
+ * Curated vendors and how many of their newest models to surface in the
+ * onboarding picker. Order matters — first vendors appear first.
+ */
+export declare const OPENROUTER_VENDORS: readonly OpenRouterVendor[];
+/**
+ * Model-id prefix → env var holding a Vertex region override. First matching
+ * prefix wins (order matters); the resolver reads `process.env[name]` and falls
+ * back to the default region when unset. Consumed by envUtils.getVertexRegionForModel.
+ */
+export declare const VERTEX_REGION_OVERRIDES: ReadonlyArray<readonly [string, string]>;
+/**
+ * Per-model pricing for direct-provider ids (USD per 1M tokens), resolved to
+ * full ModelPricing (input/output from data, cacheRead/cacheWrite derived).
+ * OpenRouter vendor/model ids price from the build-time snapshot instead.
+ */
+export declare const MODEL_PRICING: Record<string, ModelPricing>;
+/** Fallback pricing for models absent from MODEL_PRICING and the snapshot. */
+export declare const DEFAULT_PRICING: ModelPricing;
+/**
+ * Onboarding provider catalog. First entry (openrouter) doubles as the
+ * zero-config default; each entry's models[0] is that provider's default pick.
+ * Consumed by onboarding.ts (key validation / model registration / env detect).
+ */
+export declare const PROVIDERS: readonly ProviderDef[];

package/dist/data/model-metadata.js

@@ -0,0 +1,56 @@
+/**
+ * Hand-maintained model metadata for direct-provider models and onboarding
+ * picker tuning, extracted out of onboarding.ts so updating a model's
+ * max-output / context-window or the picker vendor list is a data edit
+ * (model-metadata.json) rather than a code change.
+ *
+ * Scope: only what the build-time OpenRouter snapshot does NOT cover —
+ * direct-provider ids (no "vendor/" prefix) for the metadata tables, plus the
+ * curated OpenRouter vendor ordering for the picker. OpenRouter "vendor/model"
+ * ids get their context/max-output from the snapshot (see data/openrouter-*).
+ *
+ * Loaded with createRequire so the bundled JSON ships in dist (the build copies
+ * src/data/*.json → dist/data/), matching static-catalogs.ts.
+ */
+import { createRequire } from "node:module";
+const requireJson = createRequire(import.meta.url);
+const METADATA = requireJson("./model-metadata.json");
+/**
+ * Derive a full ModelPricing from an [input, output] pair, applying the
+ * standard cache-price heuristic (read = 10% of input, write = 125% of input).
+ * This is the same derivation the old `pricing()` helper in cost-tracker used;
+ * the previously-explicit Anthropic cacheRead/cacheWrite values equalled this
+ * heuristic exactly, so moving every entry to a pair causes zero value drift.
+ */
+function pricingFromPair([input, output]) {
+    return { input, output, cacheRead: input * 0.1, cacheWrite: input * 1.25 };
+}
+/** Known max output tokens for direct-provider models (id → tokens). */
+export const KNOWN_MAX_OUTPUT = METADATA.knownMaxOutput;
+/** Known context window sizes for direct-provider models (id → tokens). */
+export const KNOWN_CONTEXT_WINDOWS = METADATA.knownContextWindows;
+/**
+ * Curated vendors and how many of their newest models to surface in the
+ * onboarding picker. Order matters — first vendors appear first.
+ */
+export const OPENROUTER_VENDORS = METADATA.openRouterVendors;
+/**
+ * Model-id prefix → env var holding a Vertex region override. First matching
+ * prefix wins (order matters); the resolver reads `process.env[name]` and falls
+ * back to the default region when unset. Consumed by envUtils.getVertexRegionForModel.
+ */
+export const VERTEX_REGION_OVERRIDES = METADATA.vertexRegionOverrides;
+/**
+ * Per-model pricing for direct-provider ids (USD per 1M tokens), resolved to
+ * full ModelPricing (input/output from data, cacheRead/cacheWrite derived).
+ * OpenRouter vendor/model ids price from the build-time snapshot instead.
+ */
+export const MODEL_PRICING = Object.fromEntries(Object.entries(METADATA.pricing).map(([id, pair]) => [id, pricingFromPair(pair)]));
+/** Fallback pricing for models absent from MODEL_PRICING and the snapshot. */
+export const DEFAULT_PRICING = pricingFromPair(METADATA.pricingDefault);
+/**
+ * Onboarding provider catalog. First entry (openrouter) doubles as the
+ * zero-config default; each entry's models[0] is that provider's default pick.
+ * Consumed by onboarding.ts (key validation / model registration / env detect).
+ */
+export const PROVIDERS = METADATA.providers;

package/dist/data/model-metadata.json

@@ -0,0 +1,216 @@
+{
+  "_comment": "Hand-maintained metadata for direct-provider models (ids without a 'vendor/' prefix) and onboarding picker tuning. OpenRouter 'vendor/model' ids get their metadata from the build-time openrouter snapshot instead; these tables only cover what the snapshot doesn't. Edit this file (no code change) when a vendor ships a new direct-provider model or you want to tune the picker. Consumed by data/model-metadata.ts.",
+  "knownMaxOutput": {
+    "anthropic/claude-opus-4.7": 32000,
+    "anthropic/claude-opus-4-7": 32000,
+    "claude-opus-4-7": 32000,
+    "anthropic/claude-sonnet-4.6": 16000,
+    "anthropic/claude-sonnet-4-6": 16000,
+    "claude-sonnet-4-6": 16000,
+    "anthropic/claude-haiku-4.5": 8192,
+    "anthropic/claude-haiku-4-5": 8192,
+    "claude-haiku-4-5": 8192,
+    "openai/gpt-5": 32000,
+    "openai/gpt-5-mini": 32000,
+    "openai/gpt-5-nano": 16000,
+    "gpt-5": 32000,
+    "gpt-5-mini": 32000,
+    "gpt-5-nano": 16000,
+    "openai/gpt-4o": 16384,
+    "gpt-4o": 16384,
+    "openai/o4-mini": 100000,
+    "o4-mini": 100000,
+    "openai/o3": 100000,
+    "o3": 100000,
+    "google/gemini-2.5-pro": 65536,
+    "google/gemini-2.5-flash": 65536,
+    "gemini-2.5-pro": 65536,
+    "gemini-2.5-flash": 65536,
+    "gemini-2.0-flash": 8192,
+    "deepseek/deepseek-v3.2": 8192,
+    "deepseek/deepseek-r1": 8192,
+    "deepseek-v4-flash": 8192,
+    "deepseek-v4-pro": 65536,
+    "deepseek-chat": 8192,
+    "qwen/qwen3-coder": 16384,
+    "meta-llama/llama-4-maverick": 32000
+  },
+  "knownContextWindows": {
+    "deepseek-v4-pro": 1000000,
+    "deepseek-v4-flash": 1000000,
+    "deepseek-chat": 1000000,
+    "deepseek/deepseek-chat": 1000000,
+    "deepseek/deepseek-v4-pro": 1000000,
+    "deepseek/deepseek-v4-flash": 1000000
+  },
+  "openRouterVendors": [
+    { "prefix": "anthropic/", "take": 4 },
+    { "prefix": "openai/", "take": 5 },
+    { "prefix": "google/", "take": 3 },
+    { "prefix": "deepseek/", "take": 3 },
+    { "prefix": "x-ai/", "take": 2 },
+    { "prefix": "qwen/", "take": 2 },
+    { "prefix": "meta-llama/", "take": 2 },
+    { "prefix": "mistralai/", "take": 1 }
+  ],
+  "vertexRegionOverrides": [
+    ["claude-haiku-4-5", "VERTEX_REGION_CLAUDE_HAIKU_4_5"],
+    ["claude-3-5-haiku", "VERTEX_REGION_CLAUDE_3_5_HAIKU"],
+    ["claude-3-5-sonnet", "VERTEX_REGION_CLAUDE_3_5_SONNET"],
+    ["claude-3-7-sonnet", "VERTEX_REGION_CLAUDE_3_7_SONNET"],
+    ["claude-opus-4-1", "VERTEX_REGION_CLAUDE_4_1_OPUS"],
+    ["claude-opus-4", "VERTEX_REGION_CLAUDE_4_0_OPUS"],
+    ["claude-sonnet-4-6", "VERTEX_REGION_CLAUDE_4_6_SONNET"],
+    ["claude-sonnet-4-5", "VERTEX_REGION_CLAUDE_4_5_SONNET"],
+    ["claude-sonnet-4", "VERTEX_REGION_CLAUDE_4_0_SONNET"]
+  ],
+  "_pricingComment": "Per-1M-token USD pricing [input, output] for direct-provider models (no vendor/ snapshot). cacheRead/cacheWrite are derived by the loader: cacheRead = input*0.1, cacheWrite = input*1.25 (matches the original pricing() helper; verified the prior explicit Anthropic cacheRead/cacheWrite equalled this heuristic exactly). 'default' is the fallback for unknown models. OpenRouter vendor/model ids price from the build-time snapshot, not here.",
+  "pricingDefault": [3, 15],
+  "pricing": {
+    "claude-opus-4.6": [15, 75],
+    "claude-opus-4-6": [15, 75],
+    "claude-opus-4.5": [15, 75],
+    "claude-opus-4.1": [15, 75],
+    "claude-opus-4": [15, 75],
+    "claude-sonnet-4.6": [3, 15],
+    "claude-sonnet-4-6": [3, 15],
+    "claude-sonnet-4.5": [3, 15],
+    "claude-sonnet-4": [3, 15],
+    "claude-3.7-sonnet": [3, 15],
+    "claude-3-7-sonnet": [3, 15],
+    "claude-3.5-haiku": [0.8, 4],
+    "claude-haiku-4.5": [1, 5],
+    "claude-haiku-4-5": [1, 5],
+    "gpt-5.4": [5, 20],
+    "gpt-5.4-mini": [1.5, 6],
+    "gpt-5.4-pro": [10, 40],
+    "gpt-5": [5, 20],
+    "gpt-4o": [2.5, 10],
+    "gpt-4o-mini": [0.15, 0.6],
+    "gpt-4.1": [2, 8],
+    "gpt-4.1-mini": [0.4, 1.6],
+    "o4-mini": [1.1, 4.4],
+    "o3": [10, 40],
+    "o3-mini": [1.1, 4.4],
+    "gemini-2.5-pro": [1.25, 10],
+    "gemini-2.5-pro-preview": [1.25, 10],
+    "gemini-2.5-flash": [0.15, 0.6],
+    "gemini-3.1-pro-preview": [2, 10],
+    "gemini-3-flash-preview": [0.15, 0.6],
+    "gemini-2.0-flash": [0.1, 0.4],
+    "deepseek-v4-flash": [0.126, 0.252],
+    "deepseek-v4-pro": [0.5, 2],
+    "deepseek-v3.2": [0.14, 0.28],
+    "deepseek-chat": [0.14, 0.28],
+    "deepseek-r1": [0.55, 2.19],
+    "deepseek-reasoner": [0.55, 2.19],
+    "qwen3-coder": [0.3, 0.6],
+    "qwen3-235b-a22b": [0.3, 1.2],
+    "qwen3-30b-a3b": [0.1, 0.3],
+    "qwen3-max": [0.5, 2],
+    "llama-4-maverick": [0.2, 0.8],
+    "llama-4-scout": [0.15, 0.4],
+    "devstral-medium": [0.5, 1.5],
+    "mistral-large": [2, 6],
+    "devstral-small": [0.1, 0.3]
+  },
+  "_providersComment": "Onboarding provider catalog (consumed by onboarding.ts). First entry (openrouter) doubles as the zero-config default; its model list is ordered so models[0] is the default pick. 'provider' is the adapterKind. Edit this (no code change) to add/retune a provider in the first-run wizard. name 文案随之在此(原 onboarding hardcoded 中文).",
+  "providers": [
+    {
+      "id": "openrouter",
+      "name": "OpenRouter (推荐 — 支持所有模型)",
+      "envKey": "OPENROUTER_API_KEY",
+      "provider": "openai",
+      "baseUrl": "https://openrouter.ai/api/v1",
+      "keyUrl": "https://openrouter.ai/keys",
+      "keyPrefix": "sk-or-",
+      "models": [
+        "anthropic/claude-sonnet-4.6",
+        "anthropic/claude-opus-4.7",
+        "anthropic/claude-haiku-4.5",
+        "openai/gpt-5",
+        "openai/gpt-5-mini",
+        "openai/gpt-4o",
+        "openai/o4-mini",
+        "openai/o3",
+        "google/gemini-2.5-pro",
+        "google/gemini-2.5-flash",
+        "deepseek/deepseek-chat",
+        "deepseek/deepseek-reasoner",
+        "qwen/qwen3-coder",
+        "meta-llama/llama-4-maverick"
+      ]
+    },
+    {
+      "id": "anthropic",
+      "name": "Anthropic (直连 Claude API)",
+      "envKey": "ANTHROPIC_API_KEY",
+      "provider": "anthropic",
+      "baseUrl": "https://api.anthropic.com",
+      "keyUrl": "https://console.anthropic.com/settings/keys",
+      "keyPrefix": "sk-ant-",
+      "models": ["claude-sonnet-4-6", "claude-opus-4-7", "claude-haiku-4-5"]
+    },
+    {
+      "id": "openai",
+      "name": "OpenAI",
+      "envKey": "OPENAI_API_KEY",
+      "provider": "openai",
+      "baseUrl": "https://api.openai.com/v1",
+      "keyUrl": "https://platform.openai.com/api-keys",
+      "keyPrefix": "sk-",
+      "models": ["gpt-5", "gpt-5-mini", "gpt-5-nano", "gpt-4o", "o4-mini", "o3"]
+    },
+    {
+      "id": "deepseek",
+      "name": "DeepSeek (官方直连)",
+      "envKey": "DEEPSEEK_API_KEY",
+      "provider": "openai",
+      "baseUrl": "https://api.deepseek.com/v1",
+      "keyUrl": "https://platform.deepseek.com/api_keys",
+      "keyPrefix": "sk-",
+      "models": ["deepseek-v4-pro", "deepseek-v4-flash", "deepseek-chat"]
+    },
+    {
+      "id": "zai",
+      "name": "Z.AI (GLM 官方)",
+      "envKey": "ZAI_API_KEY",
+      "provider": "openai",
+      "baseUrl": "https://api.z.ai/api/paas/v4",
+      "keyUrl": "https://z.ai/manage-apikey/apikey-list",
+      "keyPrefix": "",
+      "models": ["glm-5.1", "glm-4.6", "glm-4.5-air"]
+    },
+    {
+      "id": "gemini",
+      "name": "Google Gemini (官方直连)",
+      "envKey": "GEMINI_API_KEY",
+      "provider": "openai",
+      "baseUrl": "https://generativelanguage.googleapis.com/v1beta/openai",
+      "keyUrl": "https://aistudio.google.com/apikey",
+      "keyPrefix": "",
+      "models": ["gemini-2.5-pro", "gemini-2.5-flash", "gemini-2.0-flash"]
+    },
+    {
+      "id": "ollama",
+      "name": "Ollama (本地，无需 Key)",
+      "envKey": "",
+      "provider": "openai",
+      "baseUrl": "http://localhost:11434/v1",
+      "keyUrl": "https://ollama.com/library",
+      "keyPrefix": "",
+      "models": ["llama3.1", "qwen2.5-coder", "deepseek-r1", "mistral", "gemma3"],
+      "noKey": true
+    },
+    {
+      "id": "custom",
+      "name": "自定义 (任何 OpenAI 兼容 API)",
+      "envKey": "",
+      "provider": "openai",
+      "baseUrl": "",
+      "keyUrl": "",
+      "keyPrefix": "",
+      "models": []
+    }
+  ]
+}