@pixelbyte-software/pixcode 1.54.9 → 1.54.11

package/server/services/provider-models.js

@@ -1,396 +1,396 @@
-/**
- * Provider model catalog — dynamic discovery + persistent cache.
- *
- * Each CLI provider offers a different way to list the models it can run:
- *
- *   - Anthropic: https://api.anthropic.com/v1/models (API key required)
- *   - OpenAI   : https://api.openai.com/v1/models (or a custom baseUrl when
- *                the user points Codex at an OpenAI-compatible proxy)
- *   - Google   : https://generativelanguage.googleapis.com/v1beta/models
- *   - Qwen     : same OpenAI-compat shape when users BYOK through
- *                ModelStudio / ModelScope / OpenRouter etc.
- *
- * The UI still keeps a hardcoded "known good" catalog as a baseline so
- * brand-new installs don't show an empty dropdown; discovery merges the
- * server-reported list on top, dedupes, and persists to
- * `~/.pixcode/provider-models.json` with a 6-hour freshness window.
- */
-import { promises as fs } from 'node:fs';
-import os from 'node:os';
-import path from 'node:path';
-
-import { getProviderCredentials } from './provider-credentials.js';
-
-const CACHE_FILE = path.join(os.homedir(), '.pixcode', 'provider-models.json');
-const CACHE_TTL_MS = 6 * 60 * 60 * 1000; // 6 hours
-
-async function readCache() {
-    try {
-        const raw = await fs.readFile(CACHE_FILE, 'utf8');
-        const parsed = JSON.parse(raw);
-        return parsed && typeof parsed === 'object' ? parsed : {};
-    } catch {
-        return {};
-    }
-}
-
-async function writeCache(next) {
-    await fs.mkdir(path.dirname(CACHE_FILE), { recursive: true });
-    await fs.writeFile(CACHE_FILE, JSON.stringify(next, null, 2), { mode: 0o600 });
-}
-
-/**
- * Cache entry shape:
- *   { models: [{ value, label, source: 'static' | 'api' }],
- *     fetchedAt: '<iso date>',
- *     error?: '...' }
- *
- * `source` tells the UI which entries came from the live API so a
- * refresh can prune stale ones without losing the hand-maintained
- * defaults.
- */
-async function loadCachedEntry(provider) {
-    const cache = await readCache();
-    return cache[provider] || null;
-}
-
-async function saveCacheEntry(provider, entry) {
-    const cache = await readCache();
-    cache[provider] = { ...entry, fetchedAt: new Date().toISOString() };
-    await writeCache(cache);
-}
-
-function normalizeList(list) {
-    if (!Array.isArray(list)) return [];
-    const seen = new Set();
-    const out = [];
-    for (const item of list) {
-        if (!item || typeof item !== 'object') continue;
-        const value = typeof item.value === 'string' ? item.value.trim() : '';
-        if (!value || seen.has(value)) continue;
-        seen.add(value);
-        const label = typeof item.label === 'string' && item.label.trim() ? item.label.trim() : value;
-        const source = item.source === 'api' ? 'api' : 'static';
-        const entry = { value, label, source };
-        if (typeof item.free === 'boolean') entry.free = item.free;
-        out.push(entry);
-    }
-    return out;
-}
-
-function mergeCatalogs(primary, secondary) {
-    const seen = new Map();
-    for (const item of [...primary, ...secondary]) {
-        if (!seen.has(item.value)) seen.set(item.value, item);
-    }
-    return Array.from(seen.values());
-}
-
-function mergeProviderCatalogs(provider, primary, staticCatalog) {
-    const normalizedPrimary = normalizeList(primary);
-
-    // OpenCode Zen free models rotate often. When models.dev succeeds, treat
-    // that live catalog as authoritative; otherwise stale static freebies can
-    // leak back into the UI and fail later with ProviderModelNotFoundError.
-    if (provider === 'opencode') {
-        const liveModels = normalizedPrimary.filter((item) => item.source === 'api');
-        if (liveModels.length > 0) return liveModels;
-    }
-
-    return mergeCatalogs(normalizedPrimary, staticCatalog);
-}
-
-// ---------------- Per-provider live discovery ----------------
-
-async function discoverAnthropic(apiKey, baseUrl) {
-    const endpoint = (baseUrl?.replace(/\/+$/, '') || 'https://api.anthropic.com') + '/v1/models';
-    const response = await fetch(endpoint, {
-        headers: {
-            'x-api-key': apiKey,
-            'anthropic-version': '2023-06-01',
-        },
-    });
-    if (!response.ok) throw new Error(`Anthropic /v1/models returned ${response.status}`);
-    const data = await response.json();
-    const rows = Array.isArray(data?.data) ? data.data : [];
-    return rows
-        .filter((m) => typeof m?.id === 'string')
-        .map((m) => ({
-            value: m.id,
-            label: typeof m.display_name === 'string' && m.display_name.trim() ? m.display_name : m.id,
-            source: 'api',
-        }));
-}
-
-async function discoverOpenAiCompat(apiKey, baseUrl, fallbackBase) {
-    const endpoint = (baseUrl?.replace(/\/+$/, '') || fallbackBase) + '/models';
-    const response = await fetch(endpoint, {
-        headers: { Authorization: `Bearer ${apiKey}` },
-    });
-    if (!response.ok) {
-        // 401 specifically means our key is bad — but for codex/qwen/etc.
-        // users often log in via OAuth (`codex login`, `qwen auth`) which
-        // doesn't expose an OpenAI-compatible API key. Surface a clean
-        // "no live discovery available" rather than a scary 401 trace.
-        if (response.status === 401) {
-            const err = new Error('OpenAI-compatible /v1/models requires an API key. The static catalog is shown instead — that\'s expected when you signed in via OAuth (e.g. `codex login`).');
-            err.code = 'OAUTH_NO_API_KEY';
-            throw err;
-        }
-        throw new Error(`${endpoint} returned ${response.status}`);
-    }
-    const data = await response.json();
-    const rows = Array.isArray(data?.data) ? data.data : [];
-    return rows
-        .filter((m) => typeof m?.id === 'string')
-        .map((m) => ({
-            value: m.id,
-            label: m.id,
-            source: 'api',
-        }));
-}
-
-/**
- * Detect whether the user is authenticated via the provider's OAuth flow
- * (codex login / qwen auth) so we can skip live model discovery silently
- * — those flows don't surface a usable OpenAI-compatible API key, and the
- * SDK calls the upstream APIs through its own internal auth path.
- */
-async function hasProviderOauthAuth(provider) {
-    if (provider === 'codex') {
-        try {
-            await fs.access(path.join(os.homedir(), '.codex', 'auth.json'));
-            return true;
-        } catch { return false; }
-    }
-    if (provider === 'qwen') {
-        try {
-            await fs.access(path.join(os.homedir(), '.qwen', 'oauth_creds.json'));
-            return true;
-        } catch { return false; }
-    }
-    return false;
-}
-
-/**
- * OpenCode is multi-provider — its "model" picker isn't a single API list,
- * it's the union of every provider it can route to (Anthropic, OpenAI,
- * Google, xAI, OpenRouter, OpenCode Zen, Ollama, etc.). The canonical
- * catalog lives at https://models.dev/api.json (no auth, ~1.8 MB JSON, 115
- * providers as of 2026-04). We pull that, filter to providers the user
- * has authenticated with (read `~/.local/share/opencode/auth.json`) plus
- * always include the OpenCode Zen tier (works without explicit auth on
- * the free models), drop deprecated entries, and tag free models.
- */
-async function discoverOpencode() {
-    const url = process.env.OPENCODE_MODELS_URL || 'https://models.dev/api.json';
-    const response = await fetch(url, {
-        // OpenCode itself caches this for hours; we cache for 6h via the
-        // outer wrapper so a single 7s fetch on cold start is acceptable.
-        signal: AbortSignal.timeout(15000),
-    });
-    if (!response.ok) throw new Error(`models.dev/api.json returned ${response.status}`);
-    const data = await response.json();
-    if (!data || typeof data !== 'object') throw new Error('models.dev returned a non-object payload');
-
-    // Read OpenCode's auth.json to know which providers the user can
-    // actually call. Missing file → only show always-free Zen.
-    const authedProviders = new Set(['opencode']);
-    try {
-        const authPath = path.join(os.homedir(), '.local', 'share', 'opencode', 'auth.json');
-        const raw = await fs.readFile(authPath, 'utf8');
-        const auth = JSON.parse(raw);
-        if (auth && typeof auth === 'object') {
-            for (const k of Object.keys(auth)) authedProviders.add(k);
-        }
-    } catch { /* no auth.json → only Zen free models surface */ }
-
-    // Common env-var providers OpenCode picks up automatically. If the user
-    // exported one in their shell, surface those models too even without
-    // auth.json. Mirrors the env list in opencode-auth.provider.ts.
-    const envProviderHints = {
-        anthropic: ['ANTHROPIC_API_KEY'],
-        openai: ['OPENAI_API_KEY'],
-        google: ['GOOGLE_GENERATIVE_AI_API_KEY', 'GEMINI_API_KEY'],
-        'google-vertex': ['GOOGLE_APPLICATION_CREDENTIALS'],
-        xai: ['XAI_API_KEY'],
-        groq: ['GROQ_API_KEY'],
-        cerebras: ['CEREBRAS_API_KEY'],
-        openrouter: ['OPENROUTER_API_KEY'],
-    };
-    for (const [providerId, envVars] of Object.entries(envProviderHints)) {
-        if (envVars.some((v) => process.env[v]?.trim())) authedProviders.add(providerId);
-    }
-
-    const out = [];
-    for (const [providerId, providerCfg] of Object.entries(data)) {
-        if (!authedProviders.has(providerId)) continue;
-        if (!providerCfg || typeof providerCfg !== 'object') continue;
-        const models = providerCfg.models;
-        if (!models || typeof models !== 'object') continue;
-
-        const providerName = typeof providerCfg.name === 'string' && providerCfg.name.trim()
-            ? providerCfg.name
-            : providerId;
-
-        for (const [modelId, modelCfg] of Object.entries(models)) {
-            if (!modelCfg || typeof modelCfg !== 'object') continue;
-            // Skip deprecated entries from the default list — users can
-            // still hand-type them if they really need to.
-            if (modelCfg.status === 'deprecated') continue;
-            const cost = modelCfg.cost && typeof modelCfg.cost === 'object' ? modelCfg.cost : null;
-            const free = !cost || (Number(cost.input) === 0 && Number(cost.output) === 0);
-            const ctx = modelCfg.limit?.context;
-            const ctxLabel = typeof ctx === 'number' && ctx > 0
-                ? ` · ${ctx >= 1_000_000 ? `${(ctx / 1_000_000).toFixed(1)}M` : `${Math.round(ctx / 1000)}K`}`
-                : '';
-            const freeLabel = free ? ' · Free' : '';
-            const modelName = typeof modelCfg.name === 'string' && modelCfg.name.trim()
-                ? modelCfg.name
-                : modelId;
-
-            out.push({
-                value: `${providerId}/${modelId}`,
-                label: `${providerName} · ${modelName}${ctxLabel}${freeLabel}`,
-                source: 'api',
-                free,
-            });
-        }
-    }
-
-    // Sort: free first (handy when the user is unauthed), then by label.
-    out.sort((a, b) => {
-        if (a.free !== b.free) return a.free ? -1 : 1;
-        return a.label.localeCompare(b.label);
-    });
-
-    return out;
-}
-
-async function discoverGoogle(apiKey) {
-    // Google Generative Language API — public models list, API key as query.
-    const endpoint = `https://generativelanguage.googleapis.com/v1beta/models?key=${encodeURIComponent(apiKey)}`;
-    const response = await fetch(endpoint);
-    if (!response.ok) throw new Error(`Google /v1beta/models returned ${response.status}`);
-    const data = await response.json();
-    const rows = Array.isArray(data?.models) ? data.models : [];
-    return rows
-        .filter((m) => typeof m?.name === 'string' && m.name.includes('models/'))
-        .map((m) => {
-            const id = m.name.replace(/^models\//, '');
-            return {
-                value: id,
-                label: typeof m.displayName === 'string' && m.displayName.trim() ? m.displayName : id,
-                source: 'api',
-            };
-        });
-}
-
-/**
- * Returns the merged catalog for a provider.
- *   opts.forceRefresh: ignore cache and hit the upstream API
- *   opts.staticList: hardcoded fallback from shared/modelConstants.js
- */
-export async function getProviderModels(provider, opts = {}) {
-    const { forceRefresh = false, staticList = [] } = opts;
-    const staticCatalog = normalizeList(staticList.map((m) => ({ ...m, source: 'static' })));
-
-    const cached = await loadCachedEntry(provider);
-    const cacheFresh = cached?.fetchedAt
-        ? Date.now() - Date.parse(cached.fetchedAt) < CACHE_TTL_MS
-        : false;
-
-    if (!forceRefresh && cacheFresh && Array.isArray(cached?.models)) {
-        const merged = mergeProviderCatalogs(provider, cached.models, staticCatalog);
-        return {
-            models: merged,
-            fetchedAt: cached.fetchedAt,
-            error: cached.error,
-            fromCache: true,
-        };
-    }
-
-    // OpenCode is the odd one out: its catalog is models.dev, not a per-key
-    // API endpoint. Skip the credential plumbing and dispatch straight.
-    let liveModels = [];
-    let error;
-    if (provider === 'opencode') {
-        try {
-            liveModels = await discoverOpencode();
-        } catch (err) {
-            error = err?.message || String(err);
-        }
-        const merged = mergeProviderCatalogs(provider, liveModels, staticCatalog);
-        const entry = { models: merged, error };
-        await saveCacheEntry(provider, entry).catch(() => { /* non-fatal */ });
-        return { models: merged, fetchedAt: new Date().toISOString(), error, fromCache: false };
-    }
-
-    // Pick up credentials from Pixcode's UI store first, then fall back to
-    // the native env vars so a user who already exported ANTHROPIC_API_KEY
-    // (or authenticated Claude Code via OAuth — the SDK writes the key into
-    // process.env on boot) gets live models without re-entering anything.
-    const creds = await getProviderCredentials(provider);
-    const envKey = {
-        claude: process.env.ANTHROPIC_API_KEY,
-        codex: process.env.OPENAI_API_KEY,
-        qwen: process.env.OPENAI_API_KEY || process.env.QWEN_API_KEY,
-        gemini: process.env.GEMINI_API_KEY,
-    }[provider];
-    const envBase = {
-        claude: process.env.ANTHROPIC_BASE_URL,
-        codex: process.env.OPENAI_BASE_URL,
-        qwen: process.env.OPENAI_BASE_URL,
-        gemini: undefined,
-    }[provider];
-    const apiKey = creds?.apiKey || envKey;
-    const baseUrl = creds?.baseUrl || envBase || undefined;
-
-    if (!apiKey) {
-        // Codex and Qwen support OAuth (`codex login`, `qwen auth`) which
-        // DOESN'T expose a usable API key — the SDK auths against the
-        // upstream API directly. Skip the discovery step silently in that
-        // case; the static catalog is the right answer.
-        const oauthOnly = await hasProviderOauthAuth(provider);
-        if (!oauthOnly) {
-            // Be explicit so the UI can surface a useful hint rather than just
-            // showing the static baseline with no reason given.
-            error = `No ${provider} API key configured. Save one in Settings > Agents > API Key, or sign in via the CLI (e.g. \`codex login\`).`;
-        }
-    } else {
-        try {
-            if (provider === 'claude') {
-                liveModels = await discoverAnthropic(apiKey, baseUrl);
-            } else if (provider === 'codex') {
-                liveModels = await discoverOpenAiCompat(apiKey, baseUrl, 'https://api.openai.com/v1');
-            } else if (provider === 'qwen') {
-                liveModels = await discoverOpenAiCompat(
-                    apiKey,
-                    baseUrl,
-                    'https://dashscope-intl.aliyuncs.com/compatible-mode/v1',
-                );
-            } else if (provider === 'gemini') {
-                liveModels = await discoverGoogle(apiKey);
-            }
-        } catch (err) {
-            // OAuth users get a clean message instead of a raw 401 stack.
-            if (err?.code === 'OAUTH_NO_API_KEY') {
-                error = err.message;
-            } else {
-                error = err?.message || String(err);
-            }
-        }
-    }
-
-    const merged = mergeProviderCatalogs(provider, liveModels, staticCatalog);
-    const entry = { models: merged, error };
-    await saveCacheEntry(provider, entry).catch(() => { /* non-fatal */ });
-    return { models: merged, fetchedAt: new Date().toISOString(), error, fromCache: false };
-}
-
-export async function clearProviderModelCache(provider) {
-    const cache = await readCache();
-    if (provider) delete cache[provider];
-    else Object.keys(cache).forEach((k) => delete cache[k]);
-    await writeCache(cache);
-}
+/**
+ * Provider model catalog — dynamic discovery + persistent cache.
+ *
+ * Each CLI provider offers a different way to list the models it can run:
+ *
+ *   - Anthropic: https://api.anthropic.com/v1/models (API key required)
+ *   - OpenAI   : https://api.openai.com/v1/models (or a custom baseUrl when
+ *                the user points Codex at an OpenAI-compatible proxy)
+ *   - Google   : https://generativelanguage.googleapis.com/v1beta/models
+ *   - Qwen     : same OpenAI-compat shape when users BYOK through
+ *                ModelStudio / ModelScope / OpenRouter etc.
+ *
+ * The UI still keeps a hardcoded "known good" catalog as a baseline so
+ * brand-new installs don't show an empty dropdown; discovery merges the
+ * server-reported list on top, dedupes, and persists to
+ * `~/.pixcode/provider-models.json` with a 6-hour freshness window.
+ */
+import { promises as fs } from 'node:fs';
+import os from 'node:os';
+import path from 'node:path';
+
+import { getProviderCredentials } from './provider-credentials.js';
+
+const CACHE_FILE = path.join(os.homedir(), '.pixcode', 'provider-models.json');
+const CACHE_TTL_MS = 6 * 60 * 60 * 1000; // 6 hours
+
+async function readCache() {
+    try {
+        const raw = await fs.readFile(CACHE_FILE, 'utf8');
+        const parsed = JSON.parse(raw);
+        return parsed && typeof parsed === 'object' ? parsed : {};
+    } catch {
+        return {};
+    }
+}
+
+async function writeCache(next) {
+    await fs.mkdir(path.dirname(CACHE_FILE), { recursive: true });
+    await fs.writeFile(CACHE_FILE, JSON.stringify(next, null, 2), { mode: 0o600 });
+}
+
+/**
+ * Cache entry shape:
+ *   { models: [{ value, label, source: 'static' | 'api' }],
+ *     fetchedAt: '<iso date>',
+ *     error?: '...' }
+ *
+ * `source` tells the UI which entries came from the live API so a
+ * refresh can prune stale ones without losing the hand-maintained
+ * defaults.
+ */
+async function loadCachedEntry(provider) {
+    const cache = await readCache();
+    return cache[provider] || null;
+}
+
+async function saveCacheEntry(provider, entry) {
+    const cache = await readCache();
+    cache[provider] = { ...entry, fetchedAt: new Date().toISOString() };
+    await writeCache(cache);
+}
+
+function normalizeList(list) {
+    if (!Array.isArray(list)) return [];
+    const seen = new Set();
+    const out = [];
+    for (const item of list) {
+        if (!item || typeof item !== 'object') continue;
+        const value = typeof item.value === 'string' ? item.value.trim() : '';
+        if (!value || seen.has(value)) continue;
+        seen.add(value);
+        const label = typeof item.label === 'string' && item.label.trim() ? item.label.trim() : value;
+        const source = item.source === 'api' ? 'api' : 'static';
+        const entry = { value, label, source };
+        if (typeof item.free === 'boolean') entry.free = item.free;
+        out.push(entry);
+    }
+    return out;
+}
+
+function mergeCatalogs(primary, secondary) {
+    const seen = new Map();
+    for (const item of [...primary, ...secondary]) {
+        if (!seen.has(item.value)) seen.set(item.value, item);
+    }
+    return Array.from(seen.values());
+}
+
+function mergeProviderCatalogs(provider, primary, staticCatalog) {
+    const normalizedPrimary = normalizeList(primary);
+
+    // OpenCode Zen free models rotate often. When models.dev succeeds, treat
+    // that live catalog as authoritative; otherwise stale static freebies can
+    // leak back into the UI and fail later with ProviderModelNotFoundError.
+    if (provider === 'opencode') {
+        const liveModels = normalizedPrimary.filter((item) => item.source === 'api');
+        if (liveModels.length > 0) return liveModels;
+    }
+
+    return mergeCatalogs(normalizedPrimary, staticCatalog);
+}
+
+// ---------------- Per-provider live discovery ----------------
+
+async function discoverAnthropic(apiKey, baseUrl) {
+    const endpoint = (baseUrl?.replace(/\/+$/, '') || 'https://api.anthropic.com') + '/v1/models';
+    const response = await fetch(endpoint, {
+        headers: {
+            'x-api-key': apiKey,
+            'anthropic-version': '2023-06-01',
+        },
+    });
+    if (!response.ok) throw new Error(`Anthropic /v1/models returned ${response.status}`);
+    const data = await response.json();
+    const rows = Array.isArray(data?.data) ? data.data : [];
+    return rows
+        .filter((m) => typeof m?.id === 'string')
+        .map((m) => ({
+            value: m.id,
+            label: typeof m.display_name === 'string' && m.display_name.trim() ? m.display_name : m.id,
+            source: 'api',
+        }));
+}
+
+async function discoverOpenAiCompat(apiKey, baseUrl, fallbackBase) {
+    const endpoint = (baseUrl?.replace(/\/+$/, '') || fallbackBase) + '/models';
+    const response = await fetch(endpoint, {
+        headers: { Authorization: `Bearer ${apiKey}` },
+    });
+    if (!response.ok) {
+        // 401 specifically means our key is bad — but for codex/qwen/etc.
+        // users often log in via OAuth (`codex login`, `qwen auth`) which
+        // doesn't expose an OpenAI-compatible API key. Surface a clean
+        // "no live discovery available" rather than a scary 401 trace.
+        if (response.status === 401) {
+            const err = new Error('OpenAI-compatible /v1/models requires an API key. The static catalog is shown instead — that\'s expected when you signed in via OAuth (e.g. `codex login`).');
+            err.code = 'OAUTH_NO_API_KEY';
+            throw err;
+        }
+        throw new Error(`${endpoint} returned ${response.status}`);
+    }
+    const data = await response.json();
+    const rows = Array.isArray(data?.data) ? data.data : [];
+    return rows
+        .filter((m) => typeof m?.id === 'string')
+        .map((m) => ({
+            value: m.id,
+            label: m.id,
+            source: 'api',
+        }));
+}
+
+/**
+ * Detect whether the user is authenticated via the provider's OAuth flow
+ * (codex login / qwen auth) so we can skip live model discovery silently
+ * — those flows don't surface a usable OpenAI-compatible API key, and the
+ * SDK calls the upstream APIs through its own internal auth path.
+ */
+async function hasProviderOauthAuth(provider) {
+    if (provider === 'codex') {
+        try {
+            await fs.access(path.join(os.homedir(), '.codex', 'auth.json'));
+            return true;
+        } catch { return false; }
+    }
+    if (provider === 'qwen') {
+        try {
+            await fs.access(path.join(os.homedir(), '.qwen', 'oauth_creds.json'));
+            return true;
+        } catch { return false; }
+    }
+    return false;
+}
+
+/**
+ * OpenCode is multi-provider — its "model" picker isn't a single API list,
+ * it's the union of every provider it can route to (Anthropic, OpenAI,
+ * Google, xAI, OpenRouter, OpenCode Zen, Ollama, etc.). The canonical
+ * catalog lives at https://models.dev/api.json (no auth, ~1.8 MB JSON, 115
+ * providers as of 2026-04). We pull that, filter to providers the user
+ * has authenticated with (read `~/.local/share/opencode/auth.json`) plus
+ * always include the OpenCode Zen tier (works without explicit auth on
+ * the free models), drop deprecated entries, and tag free models.
+ */
+async function discoverOpencode() {
+    const url = process.env.OPENCODE_MODELS_URL || 'https://models.dev/api.json';
+    const response = await fetch(url, {
+        // OpenCode itself caches this for hours; we cache for 6h via the
+        // outer wrapper so a single 7s fetch on cold start is acceptable.
+        signal: AbortSignal.timeout(15000),
+    });
+    if (!response.ok) throw new Error(`models.dev/api.json returned ${response.status}`);
+    const data = await response.json();
+    if (!data || typeof data !== 'object') throw new Error('models.dev returned a non-object payload');
+
+    // Read OpenCode's auth.json to know which providers the user can
+    // actually call. Missing file → only show always-free Zen.
+    const authedProviders = new Set(['opencode']);
+    try {
+        const authPath = path.join(os.homedir(), '.local', 'share', 'opencode', 'auth.json');
+        const raw = await fs.readFile(authPath, 'utf8');
+        const auth = JSON.parse(raw);
+        if (auth && typeof auth === 'object') {
+            for (const k of Object.keys(auth)) authedProviders.add(k);
+        }
+    } catch { /* no auth.json → only Zen free models surface */ }
+
+    // Common env-var providers OpenCode picks up automatically. If the user
+    // exported one in their shell, surface those models too even without
+    // auth.json. Mirrors the env list in opencode-auth.provider.ts.
+    const envProviderHints = {
+        anthropic: ['ANTHROPIC_API_KEY'],
+        openai: ['OPENAI_API_KEY'],
+        google: ['GOOGLE_GENERATIVE_AI_API_KEY', 'GEMINI_API_KEY'],
+        'google-vertex': ['GOOGLE_APPLICATION_CREDENTIALS'],
+        xai: ['XAI_API_KEY'],
+        groq: ['GROQ_API_KEY'],
+        cerebras: ['CEREBRAS_API_KEY'],
+        openrouter: ['OPENROUTER_API_KEY'],
+    };
+    for (const [providerId, envVars] of Object.entries(envProviderHints)) {
+        if (envVars.some((v) => process.env[v]?.trim())) authedProviders.add(providerId);
+    }
+
+    const out = [];
+    for (const [providerId, providerCfg] of Object.entries(data)) {
+        if (!authedProviders.has(providerId)) continue;
+        if (!providerCfg || typeof providerCfg !== 'object') continue;
+        const models = providerCfg.models;
+        if (!models || typeof models !== 'object') continue;
+
+        const providerName = typeof providerCfg.name === 'string' && providerCfg.name.trim()
+            ? providerCfg.name
+            : providerId;
+
+        for (const [modelId, modelCfg] of Object.entries(models)) {
+            if (!modelCfg || typeof modelCfg !== 'object') continue;
+            // Skip deprecated entries from the default list — users can
+            // still hand-type them if they really need to.
+            if (modelCfg.status === 'deprecated') continue;
+            const cost = modelCfg.cost && typeof modelCfg.cost === 'object' ? modelCfg.cost : null;
+            const free = !cost || (Number(cost.input) === 0 && Number(cost.output) === 0);
+            const ctx = modelCfg.limit?.context;
+            const ctxLabel = typeof ctx === 'number' && ctx > 0
+                ? ` · ${ctx >= 1_000_000 ? `${(ctx / 1_000_000).toFixed(1)}M` : `${Math.round(ctx / 1000)}K`}`
+                : '';
+            const freeLabel = free ? ' · Free' : '';
+            const modelName = typeof modelCfg.name === 'string' && modelCfg.name.trim()
+                ? modelCfg.name
+                : modelId;
+
+            out.push({
+                value: `${providerId}/${modelId}`,
+                label: `${providerName} · ${modelName}${ctxLabel}${freeLabel}`,
+                source: 'api',
+                free,
+            });
+        }
+    }
+
+    // Sort: free first (handy when the user is unauthed), then by label.
+    out.sort((a, b) => {
+        if (a.free !== b.free) return a.free ? -1 : 1;
+        return a.label.localeCompare(b.label);
+    });
+
+    return out;
+}
+
+async function discoverGoogle(apiKey) {
+    // Google Generative Language API — public models list, API key as query.
+    const endpoint = `https://generativelanguage.googleapis.com/v1beta/models?key=${encodeURIComponent(apiKey)}`;
+    const response = await fetch(endpoint);
+    if (!response.ok) throw new Error(`Google /v1beta/models returned ${response.status}`);
+    const data = await response.json();
+    const rows = Array.isArray(data?.models) ? data.models : [];
+    return rows
+        .filter((m) => typeof m?.name === 'string' && m.name.includes('models/'))
+        .map((m) => {
+            const id = m.name.replace(/^models\//, '');
+            return {
+                value: id,
+                label: typeof m.displayName === 'string' && m.displayName.trim() ? m.displayName : id,
+                source: 'api',
+            };
+        });
+}
+
+/**
+ * Returns the merged catalog for a provider.
+ *   opts.forceRefresh: ignore cache and hit the upstream API
+ *   opts.staticList: hardcoded fallback from shared/modelConstants.js
+ */
+export async function getProviderModels(provider, opts = {}) {
+    const { forceRefresh = false, staticList = [] } = opts;
+    const staticCatalog = normalizeList(staticList.map((m) => ({ ...m, source: 'static' })));
+
+    const cached = await loadCachedEntry(provider);
+    const cacheFresh = cached?.fetchedAt
+        ? Date.now() - Date.parse(cached.fetchedAt) < CACHE_TTL_MS
+        : false;
+
+    if (!forceRefresh && cacheFresh && Array.isArray(cached?.models)) {
+        const merged = mergeProviderCatalogs(provider, cached.models, staticCatalog);
+        return {
+            models: merged,
+            fetchedAt: cached.fetchedAt,
+            error: cached.error,
+            fromCache: true,
+        };
+    }
+
+    // OpenCode is the odd one out: its catalog is models.dev, not a per-key
+    // API endpoint. Skip the credential plumbing and dispatch straight.
+    let liveModels = [];
+    let error;
+    if (provider === 'opencode') {
+        try {
+            liveModels = await discoverOpencode();
+        } catch (err) {
+            error = err?.message || String(err);
+        }
+        const merged = mergeProviderCatalogs(provider, liveModels, staticCatalog);
+        const entry = { models: merged, error };
+        await saveCacheEntry(provider, entry).catch(() => { /* non-fatal */ });
+        return { models: merged, fetchedAt: new Date().toISOString(), error, fromCache: false };
+    }
+
+    // Pick up credentials from Pixcode's UI store first, then fall back to
+    // the native env vars so a user who already exported ANTHROPIC_API_KEY
+    // (or authenticated Claude Code via OAuth — the SDK writes the key into
+    // process.env on boot) gets live models without re-entering anything.
+    const creds = await getProviderCredentials(provider);
+    const envKey = {
+        claude: process.env.ANTHROPIC_API_KEY,
+        codex: process.env.OPENAI_API_KEY,
+        qwen: process.env.OPENAI_API_KEY || process.env.QWEN_API_KEY,
+        gemini: process.env.GEMINI_API_KEY,
+    }[provider];
+    const envBase = {
+        claude: process.env.ANTHROPIC_BASE_URL,
+        codex: process.env.OPENAI_BASE_URL,
+        qwen: process.env.OPENAI_BASE_URL,
+        gemini: undefined,
+    }[provider];
+    const apiKey = creds?.apiKey || envKey;
+    const baseUrl = creds?.baseUrl || envBase || undefined;
+
+    if (!apiKey) {
+        // Codex and Qwen support OAuth (`codex login`, `qwen auth`) which
+        // DOESN'T expose a usable API key — the SDK auths against the
+        // upstream API directly. Skip the discovery step silently in that
+        // case; the static catalog is the right answer.
+        const oauthOnly = await hasProviderOauthAuth(provider);
+        if (!oauthOnly) {
+            // Be explicit so the UI can surface a useful hint rather than just
+            // showing the static baseline with no reason given.
+            error = `No ${provider} API key configured. Save one in Settings > Agents > API Key, or sign in via the CLI (e.g. \`codex login\`).`;
+        }
+    } else {
+        try {
+            if (provider === 'claude') {
+                liveModels = await discoverAnthropic(apiKey, baseUrl);
+            } else if (provider === 'codex') {
+                liveModels = await discoverOpenAiCompat(apiKey, baseUrl, 'https://api.openai.com/v1');
+            } else if (provider === 'qwen') {
+                liveModels = await discoverOpenAiCompat(
+                    apiKey,
+                    baseUrl,
+                    'https://dashscope-intl.aliyuncs.com/compatible-mode/v1',
+                );
+            } else if (provider === 'gemini') {
+                liveModels = await discoverGoogle(apiKey);
+            }
+        } catch (err) {
+            // OAuth users get a clean message instead of a raw 401 stack.
+            if (err?.code === 'OAUTH_NO_API_KEY') {
+                error = err.message;
+            } else {
+                error = err?.message || String(err);
+            }
+        }
+    }
+
+    const merged = mergeProviderCatalogs(provider, liveModels, staticCatalog);
+    const entry = { models: merged, error };
+    await saveCacheEntry(provider, entry).catch(() => { /* non-fatal */ });
+    return { models: merged, fetchedAt: new Date().toISOString(), error, fromCache: false };
+}
+
+export async function clearProviderModelCache(provider) {
+    const cache = await readCache();
+    if (provider) delete cache[provider];
+    else Object.keys(cache).forEach((k) => delete cache[k]);
+    await writeCache(cache);
+}