@cjhyy/code-shell-core 0.5.0-rc.2 → 0.6.0-rc.1

package/dist/lsp/manager.js

@@ -4,9 +4,11 @@
  */
 import { LSPClient } from "./client.js";
 import { BUILTIN_LSP_SERVERS } from "./servers.js";
-import { execSync } from "node:child_process";
+import { accessSync, constants, statSync } from "node:fs";
+import { delimiter, join } from "node:path";
 import { pathToFileURL } from "node:url";
 import { rootUriToPath } from "./root-path.js";
+import { commandCandidateNames } from "../utils/exec.js";
 export class LSPServerManager {
     servers = new Map();
     rootUri;
@@ -43,7 +45,7 @@ export class LSPServerManager {
         if (!managed)
             return undefined;
         // Check if command is available
-        if (!this.isCommandAvailable(managed.config.command)) {
+        if (!isCommandAvailable(managed.config.command)) {
             managed.state = "error";
             managed.error = `${managed.config.command} not found. Install: ${managed.config.installHint}`;
             return undefined;
@@ -101,17 +103,45 @@ export class LSPServerManager {
             server.client = undefined;
         }
     }
-    isCommandAvailable(command) {
-        try {
-            execSync(`which ${command} 2>/dev/null || where ${command} 2>nul`, {
-                encoding: "utf-8",
-                timeout: 5000,
-            });
-            return true;
+}
+export function isCommandAvailable(command, env = process.env) {
+    const trimmed = command.trim();
+    if (!trimmed)
+        return false;
+    const hasPathSeparator = trimmed.includes("/") || trimmed.includes("\\");
+    // A bare command with whitespace is not executable via spawn(command, args)
+    // anyway. Reject it up front so a configured command can never be interpreted
+    // as shell syntax such as `pylsp; touch /tmp/pwned`.
+    if (!hasPathSeparator && /\s/.test(trimmed))
+        return false;
+    if (hasPathSeparator) {
+        return candidateCommandNames(trimmed, env).some(isExecutableFile);
+    }
+    for (const dir of (env.PATH ?? "").split(delimiter)) {
+        if (!dir)
+            continue;
+        for (const name of candidateCommandNames(trimmed, env)) {
+            if (isExecutableFile(join(dir, name)))
+                return true;
         }
-        catch {
+    }
+    return false;
+}
+function candidateCommandNames(command, env) {
+    return commandCandidateNames(command, env);
+}
+function isExecutableFile(filePath) {
+    try {
+        const st = statSync(filePath);
+        if (!st.isFile())
             return false;
+        if (process.platform !== "win32") {
+            accessSync(filePath, constants.X_OK);
         }
+        return true;
+    }
+    catch {
+        return false;
     }
 }
 // Singleton

package/dist/model-catalog/builtin.d.ts

@@ -0,0 +1,12 @@
+/**
+ * Built-in model catalog (source A) — the official provider templates shipped
+ * with the app. Migrated from the renderer's hardcoded ProviderMeta[] arrays
+ * (ImageGenConnectionsPanel / VideoGenConnectionsPanel) and enriched with
+ * `paramsDoc` + `shape`. See docs/superpowers/specs/2026-06-11-model-catalog-design.md.
+ *
+ * Adding a same-shaped provider = add a CatalogEntry here (or a user entry in
+ * ~/.code-shell/model-catalog.user.json) — no UI / adapter changes, as long as
+ * its adapterKind points at an already-wired adapter.
+ */
+import type { CatalogEntry } from "./types.js";
+export declare const BUILTIN_CATALOG: CatalogEntry[];

package/dist/model-catalog/builtin.js

@@ -0,0 +1,407 @@
+/**
+ * Built-in model catalog (source A) — the official provider templates shipped
+ * with the app. Migrated from the renderer's hardcoded ProviderMeta[] arrays
+ * (ImageGenConnectionsPanel / VideoGenConnectionsPanel) and enriched with
+ * `paramsDoc` + `shape`. See docs/superpowers/specs/2026-06-11-model-catalog-design.md.
+ *
+ * Adding a same-shaped provider = add a CatalogEntry here (or a user entry in
+ * ~/.code-shell/model-catalog.user.json) — no UI / adapter changes, as long as
+ * its adapterKind points at an already-wired adapter.
+ */
+import { paramSpecsFromCapability } from "../llm/capabilities/param-specs.js";
+/**
+ * Build a text model preset, projecting its params from the capability layer
+ * (rules.ts) so reasoning knobs aren't re-hand-written. `params` is omitted
+ * when the model has none, matching the "absent → no knobs" contract.
+ */
+function textPreset(kind, value, label, ctx) {
+    const params = paramSpecsFromCapability(kind, value);
+    return {
+        value,
+        ...(label ? { label } : {}),
+        ...(ctx !== undefined ? { maxContextTokens: ctx } : {}),
+        ...(params.length > 0 ? { params } : {}),
+    };
+}
+/**
+ * OpenRouter normalizes reasoning to a unified effort enum across providers
+ * (xhigh|high|medium|low|minimal|none). We write OpenRouter presets' params
+ * EXPLICITLY (not via paramSpecsFromCapability) because the capability layer's
+ * OpenRouter catch-all would slap reasoning onto *every* model — including ones
+ * that don't support it. A reasoning OpenRouter preset opts in by listing this.
+ */
+const OPENROUTER_REASONING = {
+    name: "reasoning",
+    label: "思考强度",
+    control: "enum",
+    options: ["minimal", "low", "medium", "high", "xhigh"],
+    default: "medium",
+    doc: "Reasoning effort — how hard the model thinks before answering.",
+    wire: { field: "reasoning.effort" },
+};
+/**
+ * Zhipu GLM params (OpenAI-compatible endpoint, but with GLM-specific wire
+ * fields). Promoted from the user catalog after live-verifying glm-5.2 / 5.1 /
+ * 5 / 5-turbo / 4.7 / 4.6 against open.bigmodel.cn (2026-06-23). Written
+ * explicitly because the capability layer has no GLM rules.
+ */
+const ZHIPU_PARAMS = [
+    {
+        name: "reasoning_effort",
+        label: "推理强度",
+        control: "enum",
+        options: ["max", "xhigh", "high", "medium", "low", "minimal", "none"],
+        default: "max",
+        doc: "控制模型推理程度；thinking 开启时生效。GLM-5.2 支持，none/minimal 放弃思考，low/medium 映射为 high，xhigh 映射为 max。",
+        wire: { field: "reasoning_effort" },
+    },
+    {
+        name: "thinking_type",
+        label: "思维链",
+        control: "enum",
+        options: ["enabled", "disabled"],
+        default: "enabled",
+        doc: "控制是否开启思维链；GLM-5.2 开启后为强制思考。",
+        wire: { field: "thinking.type" },
+    },
+    {
+        name: "temperature",
+        label: "Temperature",
+        control: "number",
+        min: 0,
+        max: 1,
+        default: 1,
+        doc: "采样温度，控制输出随机性，范围 0.0 到 1.0。",
+        wire: { field: "temperature" },
+    },
+    {
+        name: "top_p",
+        label: "Top P",
+        control: "number",
+        min: 0.01,
+        max: 1,
+        default: 0.95,
+        doc: "核采样参数，范围 0.01 到 1.0；官方建议不要与 temperature 同时调整。",
+        wire: { field: "top_p" },
+    },
+    {
+        name: "max_tokens",
+        label: "最大输出 Tokens",
+        control: "number",
+        min: 1,
+        max: 131072,
+        default: 4096,
+        doc: "模型输出最大 token 数量，GLM-5.2 最大 131072。",
+        wire: { field: "max_tokens" },
+    },
+];
+/** Zhipu GLM preset — shared ZHIPU_PARAMS, with per-model context window. */
+function glmPreset(value, label, ctx) {
+    return { value, label, maxContextTokens: ctx, params: ZHIPU_PARAMS };
+}
+/** OpenRouter preset with explicit params (no capability-layer projection). */
+function orPreset(value, label, params, ctx) {
+    return {
+        value,
+        label,
+        ...(ctx !== undefined ? { maxContextTokens: ctx } : {}),
+        ...(params && params.length > 0 ? { params } : {}),
+    };
+}
+export const BUILTIN_CATALOG = [
+    // ─── text (LLM) ───
+    {
+        id: "openai",
+        tag: "text",
+        adapterKind: "openai",
+        protocol: "openai-compat",
+        displayName: "OpenAI",
+        description: "OpenAI chat models (GPT-5 系列、GPT-4o 等)。需要 OpenAI key。",
+        defaultBaseUrl: "https://api.openai.com/v1",
+        defaultModel: "gpt-5.5",
+        signupUrl: "https://platform.openai.com/api-keys",
+        needsKey: true,
+        // Verified live (2026-06-23) against this account's key: gpt-5.5 / 5.4 /
+        // 5.4-mini returned text; o4-mini reachable; o3 reachable (slow). New OpenAI
+        // models reject `max_tokens` — must use `max_completion_tokens` (handled by
+        // the client). gpt-5.5-pro is 404 on /chat/completions (not a chat model).
+        modelPresets: [
+            textPreset("openai", "gpt-5.5", "GPT-5.5", 1_050_000),
+            textPreset("openai", "gpt-5.4", "GPT-5.4", 1_050_000),
+            textPreset("openai", "gpt-5.4-mini", "GPT-5.4 Mini", 400_000),
+            textPreset("openai", "gpt-5.4-nano", "GPT-5.4 Nano", 400_000),
+            textPreset("openai", "o4-mini", "o4-mini", 200_000),
+            textPreset("openai", "o3", "o3", 200_000),
+            textPreset("openai", "gpt-4o", "GPT-4o", 128_000),
+            textPreset("openai", "gpt-4o-mini", "GPT-4o mini", 128_000),
+        ],
+    },
+    {
+        id: "anthropic",
+        tag: "text",
+        adapterKind: "anthropic",
+        protocol: "anthropic-style",
+        displayName: "Anthropic",
+        description: "Claude 模型(Opus / Sonnet / Haiku)。需要 Anthropic key。",
+        defaultBaseUrl: "https://api.anthropic.com/v1",
+        defaultModel: "claude-opus-4-8",
+        signupUrl: "https://console.anthropic.com/settings/keys",
+        needsKey: true,
+        modelPresets: [
+            textPreset("anthropic", "claude-opus-4-8", "Claude Opus 4.8", 1_000_000),
+            textPreset("anthropic", "claude-opus-4-7", "Claude Opus 4.7", 1_000_000),
+            textPreset("anthropic", "claude-sonnet-4-6", "Claude Sonnet 4.6", 1_000_000),
+            textPreset("anthropic", "claude-haiku-4-5", "Claude Haiku 4.5", 200_000),
+        ],
+    },
+    {
+        id: "openrouter",
+        tag: "text",
+        adapterKind: "openrouter",
+        protocol: "openai-compat",
+        displayName: "OpenRouter",
+        description: "通过 OpenRouter 路由多家模型;统一 reasoning 配置。需要 OpenRouter key。",
+        defaultBaseUrl: "https://openrouter.ai/api/v1",
+        defaultModel: "~anthropic/claude-opus-latest",
+        signupUrl: "https://openrouter.ai/keys",
+        needsKey: true,
+        // Two kinds of entries per family:
+        //  - "~vendor/model-latest" router aliases — OpenRouter resolves them
+        //    server-side to the newest version, so the slug never goes stale (vs.
+        //    dated slugs like anthropic/claude-opus-4.8-20260528). Verified live
+        //    2026-07-02: ~anthropic/claude-opus-latest→opus-4.8,
+        //    ~anthropic/claude-sonnet-latest→sonnet-5, ~google/gemini-pro-latest→
+        //    gemini-3.1-pro. (~openai/gpt-latest is NOT a real alias — it 400s —
+        //    so it's dropped; pin a concrete openai slug instead.)
+        //  - concrete versioned slugs — so the picker can select a SPECIFIC
+        //    version (e.g. Opus 4.7-fast) instead of only "latest". Every concrete
+        //    slug + context window below verified live 2026-07-02 against this
+        //    account's OpenRouter key.
+        // Explicit params (not capability-projected): reasoning models opt into
+        // OPENROUTER_REASONING; non-reasoning ones (deepseek-chat V3, llama-4)
+        // omit params — no catch-all forcing reasoning onto a model that lacks it.
+        modelPresets: [
+            // Anthropic (Claude) — alias + concrete versions
+            orPreset("~anthropic/claude-opus-latest", "Claude Opus (latest)", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("~anthropic/claude-sonnet-latest", "Claude Sonnet (latest)", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-opus-4.8", "Claude Opus 4.8", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-opus-4.8-fast", "Claude Opus 4.8 (fast)", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-opus-4.7", "Claude Opus 4.7", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-opus-4.7-fast", "Claude Opus 4.7 (fast)", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-sonnet-5", "Claude Sonnet 5", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-sonnet-4.6", "Claude Sonnet 4.6", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("anthropic/claude-fable-5", "Claude Fable 5", [OPENROUTER_REASONING], 1_000_000),
+            // OpenAI (GPT) — concrete only (no working ~latest alias). Led by 5.5 to
+            // match the direct-OpenAI default (state.json model = gpt-5.5).
+            orPreset("openai/gpt-5.5", "GPT-5.5", [OPENROUTER_REASONING], 1_050_000),
+            orPreset("openai/gpt-5.5-pro", "GPT-5.5 Pro", [OPENROUTER_REASONING], 1_050_000),
+            orPreset("openai/gpt-5.4", "GPT-5.4", [OPENROUTER_REASONING], 1_050_000),
+            orPreset("openai/gpt-5.4-mini", "GPT-5.4 Mini", [OPENROUTER_REASONING], 400_000),
+            orPreset("openai/gpt-5.1-codex-max", "GPT-5.1 Codex Max", [OPENROUTER_REASONING], 400_000),
+            // Google (Gemini) — alias + concrete
+            orPreset("~google/gemini-pro-latest", "Gemini Pro (latest)", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("google/gemini-3-flash-preview", "Gemini 3 Flash (preview)", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("google/gemini-2.5-pro", "Gemini 2.5 Pro", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("google/gemini-2.5-flash", "Gemini 2.5 Flash", [OPENROUTER_REASONING], 1_048_576),
+            // xAI (Grok)
+            orPreset("x-ai/grok-4.3", "Grok 4.3", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("x-ai/grok-4.20", "Grok 4.20", [OPENROUTER_REASONING], 2_000_000),
+            // DeepSeek — V4 reasoning + V3 chat (non-reasoning)
+            orPreset("deepseek/deepseek-v4-pro", "DeepSeek V4 Pro", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("deepseek/deepseek-v4-flash", "DeepSeek V4 Flash", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("deepseek/deepseek-chat", "DeepSeek V3 (chat)", undefined, 131_072),
+            // Zhipu (GLM) — glm-5.1 (65k ctx) omitted: below the catalog's ≥100k
+            // context floor; glm-5.2 (1M) and glm-5 (202k) cover the family.
+            orPreset("z-ai/glm-5.2", "GLM 5.2 (Z.ai)", [OPENROUTER_REASONING], 1_048_576),
+            orPreset("z-ai/glm-5", "GLM 5 (Z.ai)", [OPENROUTER_REASONING], 202_752),
+            // Qwen
+            orPreset("qwen/qwen3.7-max", "Qwen3.7 Max", [OPENROUTER_REASONING], 1_000_000),
+            orPreset("qwen/qwen3-coder", "Qwen3 Coder", [OPENROUTER_REASONING], 262_144),
+            // Meta Llama (non-reasoning)
+            orPreset("meta-llama/llama-4-maverick", "Llama 4 Maverick", undefined, 1_048_576),
+            orPreset("meta-llama/llama-4-scout", "Llama 4 Scout", undefined, 327_680),
+        ],
+    },
+    {
+        id: "deepseek",
+        tag: "text",
+        adapterKind: "deepseek",
+        protocol: "openai-compat",
+        displayName: "DeepSeek",
+        description: "DeepSeek 模型。需要 DeepSeek key。",
+        defaultBaseUrl: "https://api.deepseek.com/v1",
+        defaultModel: "deepseek-v4-flash",
+        signupUrl: "https://platform.deepseek.com/api_keys",
+        needsKey: true,
+        modelPresets: [
+            textPreset("deepseek", "deepseek-v4-flash", "DeepSeek V4 Flash", 1_000_000),
+            textPreset("deepseek", "deepseek-v4-pro", "DeepSeek V4 Pro", 1_000_000),
+        ],
+    },
+    {
+        id: "zhipu",
+        tag: "text",
+        adapterKind: "openai",
+        protocol: "openai-compat",
+        displayName: "Zhipu GLM",
+        description: "智谱 GLM 系列(OpenAI 兼容端点)。适合 Coding Agent 与长上下文任务。需要智谱 key。",
+        defaultBaseUrl: "https://open.bigmodel.cn/api/paas/v4",
+        defaultModel: "glm-5.2",
+        signupUrl: "https://open.bigmodel.cn/",
+        needsKey: true,
+        // All slugs verified live (2026-06-23) against open.bigmodel.cn: glm-5.2 /
+        // 5.1 / 5 / 5-turbo / 4.7 / 4.6 each returned a successful chat completion.
+        modelPresets: [
+            glmPreset("glm-5.2", "GLM-5.2", 1_000_000),
+            glmPreset("glm-5.1", "GLM-5.1", 1_000_000),
+            glmPreset("glm-5", "GLM-5", 1_000_000),
+            glmPreset("glm-5-turbo", "GLM-5 Turbo", 1_000_000),
+            glmPreset("glm-4.7", "GLM-4.7", 200_000),
+            glmPreset("glm-4.6", "GLM-4.6", 200_000),
+        ],
+    },
+    {
+        id: "google",
+        tag: "text",
+        adapterKind: "google",
+        protocol: "openai-compat",
+        displayName: "Google Gemini",
+        description: "Gemini 文本模型(OpenAI-compat 端点)。需要 Google key。",
+        defaultBaseUrl: "https://generativelanguage.googleapis.com/v1beta/openai",
+        defaultModel: "gemini-3.5-flash",
+        signupUrl: "https://aistudio.google.com/apikey",
+        needsKey: true,
+        modelPresets: [
+            textPreset("google", "gemini-3.5-flash", "Gemini 3.5 Flash", 1_048_576),
+            textPreset("google", "gemini-3.1-flash-lite", "Gemini 3.1 Flash Lite", 1_048_576),
+            textPreset("google", "gemini-2.5-pro", "Gemini 2.5 Pro", 1_048_576),
+            textPreset("google", "gemini-2.5-flash", "Gemini 2.5 Flash", 1_048_576),
+        ],
+    },
+    {
+        id: "ollama",
+        tag: "text",
+        adapterKind: "ollama",
+        protocol: "openai-compat",
+        displayName: "Ollama",
+        description: "本地 Ollama 模型,无需 key。",
+        defaultBaseUrl: "http://localhost:11434/v1",
+        defaultModel: "llama3.1",
+        needsKey: false,
+        modelPresets: [
+            textPreset("ollama", "llama3.1", "Llama 3.1"),
+            textPreset("ollama", "qwen2.5-coder", "Qwen Coder"),
+            textPreset("ollama", "deepseek-r1", "DeepSeek R1"),
+        ],
+    },
+    {
+        id: "custom",
+        tag: "text",
+        adapterKind: "openai",
+        protocol: "openai-compat",
+        displayName: "Custom (OpenAI-compatible)",
+        description: "任意 OpenAI 兼容端点;baseUrl/model 由连接自带。",
+        defaultBaseUrl: "",
+        needsKey: true,
+        modelPresets: [],
+    },
+    // ─── image ───
+    {
+        id: "openai-images",
+        tag: "image",
+        adapterKind: "openai",
+        shape: "generic-sync",
+        displayName: "OpenAI Images (gpt-image)",
+        description: "OpenAI 图像 API。需要 OpenAI key；baseUrl 默认官方端点。",
+        defaultBaseUrl: "https://api.openai.com/v1",
+        defaultModel: "gpt-image-2",
+        // Live-verified 2026-07-02 against this account's OpenAI key (GET /v1/models):
+        // gpt-image-2 (current), gpt-image-1.5, gpt-image-1, gpt-image-1-mini.
+        modelPresets: [
+            { value: "gpt-image-2", label: "GPT Image 2" },
+            { value: "gpt-image-1.5", label: "GPT Image 1.5" },
+            { value: "gpt-image-1", label: "GPT Image 1" },
+            { value: "gpt-image-1-mini", label: "GPT Image 1 mini" },
+        ],
+        signupUrl: "https://platform.openai.com/api-keys",
+        test: true,
+        paramsDoc: "OpenAI 图像：支持 size (1024x1024 | 1536x1024 | 1024x1536 | auto)、quality (low | medium | high | auto)。" +
+            "支持图生图:传 referenceImages(工作区图片路径,最多 16 张)即按这些图编辑/再创作(走 /images/edits)。",
+    },
+    {
+        id: "google-images",
+        tag: "image",
+        adapterKind: "google",
+        shape: "generic-sync",
+        displayName: "Gemini Images (Nano Banana)",
+        description: "Gemini 图像生成。可直接用你已有的 Google key；OpenAI 兼容 baseUrl（/v1beta/openai）也会被自动规范到原生端点。",
+        defaultBaseUrl: "https://generativelanguage.googleapis.com/v1beta",
+        defaultModel: "gemini-3.1-flash-image",
+        modelPresets: [
+            { value: "gemini-3.1-flash-image", label: "Nano Banana 2 (3.1)" },
+            { value: "gemini-2.5-flash-image", label: "Nano Banana (2.5)" },
+        ],
+        signupUrl: "https://aistudio.google.com/apikey",
+        test: true,
+        paramsDoc: "Gemini 图像 (Nano Banana)：size 会被映射到最接近的支持比例;quality 参数对该后端无效(忽略)。" +
+            "支持图生图/多图融合:传 referenceImages(工作区图片路径)作为参考图一起喂给模型。",
+    },
+    // ─── video ───
+    {
+        id: "fal-video",
+        tag: "video",
+        adapterKind: "fal",
+        shape: "fal-queue",
+        displayName: "fal.ai (Kling / 即梦 Seedance 等)",
+        description: "通过 fal.ai 统一 API 调用 Kling、即梦(Seedance,字节同源)等视频模型。需要 fal key；" +
+            "「默认模型」决定底层模型与文生/图生(传图自动切图生)。即梦 = fal 上的 bytedance/seedance 模型,选它即可。",
+        defaultBaseUrl: "https://queue.fal.run",
+        defaultModel: "fal-ai/kling-video/v3/pro/text-to-video",
+        signupUrl: "https://fal.ai/dashboard/keys",
+        test: false,
+        modelPresets: [
+            { value: "bytedance/seedance-2.0/text-to-video", label: "即梦 Seedance 2.0 · 文生视频" },
+            { value: "bytedance/seedance-2.0/image-to-video", label: "即梦 Seedance 2.0 · 图生视频" },
+            { value: "fal-ai/kling-video/v3/pro/text-to-video", label: "Kling v3 Pro · 文生视频" },
+            { value: "fal-ai/kling-video/v3/pro/image-to-video", label: "Kling v3 Pro · 图生视频" },
+        ],
+        paramsDoc: "fal 视频:用 model 选底层模型(文生 vs 图生)。传 image/images(本地路径自动上传)→ 用图生视频模型;1 张=图生视频,2+ 张=参考生视频(最多 9,prompt 里用 @Image1/@Image2 引用)。" +
+            "续接/参考视频:传 videos(http/https URL,非本地路径;最多 3)续接已有视频,prompt 里用 @Video1/@Video2 引用(如「从 @Video1 结尾继续」);需用 Seedance 模型(走 reference-to-video,Kling 不支持)。异步后台生成。",
+    },
+    // ─── audio (speech-to-text / 语音输入听写) ───
+    {
+        id: "openai-transcribe",
+        tag: "audio",
+        adapterKind: "openai",
+        shape: "generic-sync",
+        displayName: "OpenAI 语音转写 (Whisper / gpt-4o-transcribe)",
+        description: "OpenAI 兼容 /audio/transcriptions,用于语音输入听写。可复用 OpenAI key。",
+        defaultBaseUrl: "https://api.openai.com/v1",
+        defaultModel: "gpt-4o-transcribe",
+        signupUrl: "https://platform.openai.com/api-keys",
+        test: false,
+        modelPresets: [
+            { value: "gpt-4o-transcribe", label: "gpt-4o-transcribe" },
+            { value: "gpt-4o-mini-transcribe", label: "gpt-4o-mini-transcribe" },
+            { value: "whisper-1", label: "Whisper (whisper-1)" },
+        ],
+        paramsDoc: "语音转写:录音(webm/opus)→ 文字,填进输入框。换 baseUrl+model 可指向 Groq/本地 whisper.cpp。",
+    },
+    {
+        id: "groq-transcribe",
+        tag: "audio",
+        adapterKind: "openai",
+        shape: "generic-sync",
+        displayName: "Groq 语音转写 (whisper-large-v3-turbo)",
+        description: "Groq 的 OpenAI 兼容 /audio/transcriptions,极快且便宜。需要 Groq key。",
+        defaultBaseUrl: "https://api.groq.com/openai/v1",
+        defaultModel: "whisper-large-v3-turbo",
+        signupUrl: "https://console.groq.com/keys",
+        test: false,
+        modelPresets: [
+            { value: "whisper-large-v3-turbo", label: "whisper-large-v3-turbo" },
+            { value: "whisper-large-v3", label: "whisper-large-v3" },
+        ],
+        paramsDoc: "语音转写(Groq):同 OpenAI /audio/transcriptions 形状,换 baseUrl+model。",
+    },
+];

package/dist/model-catalog/gen-connections.d.ts

@@ -0,0 +1,20 @@
+/**
+ * genInstancesFromConnections — bridge unified modelConnections (tag=image or
+ * video) + credentials into the GenInstance shape the image/video runtime
+ * resolvers (resolveImageProvider / resolveVideoProvider) consume. Lets
+ * image/video flow through the unified credentials store; key comes from the
+ * referenced credential. Connections whose catalogId doesn't resolve are
+ * skipped. See docs/.../2026-06-15-unified-model-catalog-design.md §6.
+ */
+import { type ModelInstance, type Credential } from "./resolve.js";
+import type { CatalogEntry } from "./index.js";
+/** The shape image/video resolvers expect (mirrors generate-image GenInstance). */
+export interface GenRuntimeInstance {
+    id: string;
+    kind: string;
+    baseUrl: string;
+    apiKey?: string;
+    defaultModel?: string;
+    catalogId: string;
+}
+export declare function genInstancesFromConnections(connections: ModelInstance[], credentials: Credential[], catalog: CatalogEntry[], tag: "image" | "video" | "audio"): GenRuntimeInstance[];

package/dist/model-catalog/gen-connections.js

@@ -0,0 +1,28 @@
+/**
+ * genInstancesFromConnections — bridge unified modelConnections (tag=image or
+ * video) + credentials into the GenInstance shape the image/video runtime
+ * resolvers (resolveImageProvider / resolveVideoProvider) consume. Lets
+ * image/video flow through the unified credentials store; key comes from the
+ * referenced credential. Connections whose catalogId doesn't resolve are
+ * skipped. See docs/.../2026-06-15-unified-model-catalog-design.md §6.
+ */
+import { resolveInstance } from "./resolve.js";
+export function genInstancesFromConnections(connections, credentials, catalog, tag) {
+    const out = [];
+    for (const inst of connections) {
+        if (inst.tag !== tag)
+            continue;
+        const resolved = resolveInstance(inst, credentials, catalog);
+        if (!resolved)
+            continue;
+        out.push({
+            id: inst.id,
+            kind: resolved.adapterKind,
+            baseUrl: resolved.baseUrl,
+            ...(resolved.apiKey !== undefined ? { apiKey: resolved.apiKey } : {}),
+            defaultModel: inst.model || resolved.entry.defaultModel,
+            catalogId: inst.catalogId,
+        });
+    }
+    return out;
+}

package/dist/model-catalog/index.d.ts

@@ -0,0 +1,41 @@
+/**
+ * Model catalog assembly — merge built-in (A) + user (B) templates.
+ *
+ * - A: {@link BUILTIN_CATALOG} (ships with the app).
+ * - B: ~/.code-shell/model-catalog.user.json (user-added templates).
+ * - merge: union by `id`; same id → user B wins (so升级 doesn't clobber the
+ *   user's overrides/additions).
+ *
+ * No remote source this version (design doc §5).
+ */
+import { type CatalogEntry } from "./types.js";
+export type { CatalogEntry } from "./types.js";
+export { BUILTIN_CATALOG } from "./builtin.js";
+export { saveCatalogEntry, deleteUserCatalogEntry } from "./save-entry.js";
+/** Path to the user catalog file (source B). */
+export declare function userCatalogPath(): string;
+/**
+ * Load user-defined catalog entries (source B). Returns [] when the file is
+ * absent or invalid — a bad user file must never break the built-in catalog.
+ */
+export declare function loadUserCatalog(): CatalogEntry[];
+/**
+ * Merged catalog: built-in (A) ∪ user (B), deduped by `id` with user winning.
+ * This is the single source the 连接 page renders from and the tool-description
+ * injector looks up `paramsDoc` in.
+ */
+export declare function getMergedCatalog(): CatalogEntry[];
+/**
+ * Look up a catalog entry by id (e.g. an instance's `catalogId`), falling back
+ * to the first entry whose `adapterKind` matches `kindFallback`. Since an
+ * adapterKind (e.g. "openai") can now back both a text and an image entry,
+ * `tagFallback` disambiguates the fallback to the right group.
+ */
+export declare function findCatalogEntry(catalog: CatalogEntry[], id: string | undefined, kindFallback?: string, tagFallback?: CatalogEntry["tag"]): CatalogEntry | undefined;
+export type CatalogEntryOrigin = "builtin" | "user" | "user-override-of-builtin";
+/**
+ * Per-id provenance of the merged catalog: pure built-in, pure user-added, or a
+ * user override of a built-in. Lets the editor UI show 编辑/删除/重置 correctly
+ * (only an override can be "reset"; only a user-only entry can be truly deleted).
+ */
+export declare function catalogEntryOrigins(): Record<string, CatalogEntryOrigin>;