@exreve/exk 1.0.77 → 1.0.79

package/dist/cli/agentSession.js

@@ -56,7 +56,27 @@ const symlinkAsync = promisify(fsSymlink);
 // AI config - loaded from server after registration, stored in ~/.talk-to-code/ai-config.json
 // (Do not read ANTHROPIC_* / CLAUDE_MODEL from the host environment — only this file + code default model.)
 const AI_CONFIG_PATH = path.join(os.homedir(), '.talk-to-code', 'ai-config.json');
-const DEFAULT_AI_MODEL = 'glm-5.2';
+const DEFAULT_AI_MODEL = 'GLM-5.2';
+/**
+ * Normalize a model ID before sending it to the upstream API.
+ *
+ * z.ai's API only accepts UPPERCASE model IDs (GLM-5.2, GLM-5.1, etc.). The
+ * lowercase form 'glm-5.2' triggers a 529 overloaded_error from z.ai, even
+ * though it's a real model on their side. The session schema in MongoDB and
+ * existing ai-config.json files may still contain the legacy lowercase form,
+ * so we normalize defensively at every read point.
+ */
+function normalizeModelId(model) {
+    if (!model)
+        return DEFAULT_AI_MODEL;
+    const m = String(model).trim();
+    if (!m)
+        return DEFAULT_AI_MODEL;
+    // z.ai GLM models must be uppercase
+    if (/^glm[-.]/i.test(m))
+        return m.toUpperCase();
+    return m;
+}
 /** TTL cache for ai-config.json reads to avoid hitting disk on every call */
 let _aiConfigCache = null;
 const AI_CONFIG_TTL_MS = 5_000;
@@ -66,7 +86,7 @@ const PROVIDERS = {
     zai: {
         apiKey: process.env.ZHIPU_API_KEY || '',
         baseUrl: process.env.CLI_AI_BASE_URL || 'https://api.z.ai/api/anthropic',
-        models: ['glm-5.2'],
+        models: ['GLM-5.2'],
     },
     minimax: {
         apiKey: '',
@@ -97,10 +117,11 @@ function rebuildProvidersFromConfig(models) {
     const providerModels = {};
     const providerUrls = {};
     for (const m of models) {
+        const id = normalizeModelId(m.id);
         if (!providerModels[m.provider])
             providerModels[m.provider] = [];
-        if (!providerModels[m.provider].includes(m.id))
-            providerModels[m.provider].push(m.id);
+        if (!providerModels[m.provider].includes(id))
+            providerModels[m.provider].push(id);
         if (m.providerBaseUrl)
             providerUrls[m.provider] = m.providerBaseUrl;
     }
@@ -129,15 +150,17 @@ function resolveProvider(model, providerId) {
     PROVIDERS.kimi.apiKey = aiConfig.kimiApiKey || process.env.KIMI_API_KEY || '';
     if (!PROVIDERS.zai.apiKey)
         PROVIDERS.zai.apiKey = aiConfig.apiKey || '';
+    // Normalize model name for z.ai case-sensitivity
+    const normalizedModel = normalizeModelId(model);
     // 1. Explicit provider selection
     if (providerId && PROVIDERS[providerId]?.apiKey) {
         const provider = PROVIDERS[providerId];
-        return { provider: providerId, apiKey: provider.apiKey, baseUrl: provider.baseUrl, model };
+        return { provider: providerId, apiKey: provider.apiKey, baseUrl: provider.baseUrl, model: normalizedModel };
     }
     // 2. Match model name to a provider
     for (const [id, config] of Object.entries(PROVIDERS)) {
-        if (config.models.includes(model) && config.apiKey) {
-            return { provider: id, apiKey: config.apiKey, baseUrl: config.baseUrl, model };
+        if (config.models.includes(normalizedModel) && config.apiKey) {
+            return { provider: id, apiKey: config.apiKey, baseUrl: config.baseUrl, model: normalizedModel };
         }
     }
     // 3. Fallback: use ai-config.json credentials (z.ai default)
@@ -145,7 +168,7 @@ function resolveProvider(model, providerId) {
         provider: 'zai',
         apiKey: aiConfig.apiKey,
         baseUrl: aiConfig.baseUrl || PROVIDERS.zai.baseUrl,
-        model,
+        model: normalizedModel,
     };
 }
 function loadAiConfig() {
@@ -158,7 +181,7 @@ function loadAiConfig() {
         const config = JSON.parse(data);
         const apiKey = typeof config.authToken === 'string' ? config.authToken.trim() : '';
         const baseUrl = typeof config.baseUrl === 'string' ? config.baseUrl.trim() : '';
-        const model = typeof config.model === 'string' && config.model.trim() ? config.model.trim() : DEFAULT_AI_MODEL;
+        const model = normalizeModelId(typeof config.model === 'string' && config.model.trim() ? config.model.trim() : DEFAULT_AI_MODEL);
         const proxy = typeof config.proxy === 'string' ? config.proxy.trim() : '';
         const minimaxApiKey = typeof config.minimaxApiKey === 'string' ? config.minimaxApiKey.trim() : '';
         const openrouterApiKey = typeof config.openrouterApiKey === 'string' ? config.openrouterApiKey.trim() : '';
@@ -239,6 +262,17 @@ function envForClaudeCodeChild(_localModel, resolvedProvider) {
         env.ANTHROPIC_API_KEY = effectiveApiKey;
     if (effectiveBaseUrl)
         env.ANTHROPIC_BASE_URL = effectiveBaseUrl;
+    // For z.ai specifically: override ALL model aliases so the SDK
+    // sends the correct (UPPERCASE) model ID. z.ai's API is case-sensitive
+    // and rejects lowercase 'glm-5.2' with 529 overloaded_error.
+    if (resolvedProvider?.provider === 'zai') {
+        const m = normalizeModelId(resolvedProvider.model);
+        env.ANTHROPIC_MODEL = m;
+        env.ANTHROPIC_DEFAULT_SONNET_MODEL = m;
+        env.ANTHROPIC_DEFAULT_OPUS_MODEL = m;
+        env.ANTHROPIC_DEFAULT_HAIKU_MODEL = m;
+        env.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC = '1';
+    }
     // For MiniMax specifically: override ALL model aliases so the SDK
     // sends the correct model ID to the Anthropic-compatible endpoint
     if (resolvedProvider?.provider === 'minimax') {
@@ -686,6 +720,15 @@ export class AgentSessionManager {
                         ANTHROPIC_API_KEY: resolved.apiKey,
                         ANTHROPIC_BASE_URL: resolved.baseUrl,
                     };
+                    // For z.ai: also override all model aliases in settings (case-sensitive)
+                    if (resolved.provider === 'zai') {
+                        const m = normalizeModelId(resolved.model);
+                        settingsEnv.ANTHROPIC_MODEL = m;
+                        settingsEnv.ANTHROPIC_DEFAULT_SONNET_MODEL = m;
+                        settingsEnv.ANTHROPIC_DEFAULT_OPUS_MODEL = m;
+                        settingsEnv.ANTHROPIC_DEFAULT_HAIKU_MODEL = m;
+                        settingsEnv.CLAUDE_CODE_DISABLE_NONESSENTIAL_TRAFFIC = '1';
+                    }
                     // For MiniMax: also override all model aliases in settings
                     if (resolved.provider === 'minimax') {
                         settingsEnv.ANTHROPIC_MODEL = resolved.model;

package/dist/cli/claudeBackend.js

@@ -169,6 +169,34 @@ export class ClaudeBackend {
         return CACHED_CLAUDE_PATH;
     }
     async *executePrompt(prompt, config) {
+        const { cwd, apiKey, model, env, settings, signal, attachmentDir, routingSessionId, routingPromptId, resumeSessionId } = config;
+        // Retry config for 529 overloaded errors
+        const MAX_RETRIES = 2;
+        const RETRY_DELAY_MS = 3000;
+        for (let attempt = 0; attempt <= MAX_RETRIES; attempt++) {
+            if (attempt > 0) {
+                console.log(`[ClaudeBackend] Retry ${attempt}/${MAX_RETRIES} after 529 delay...`);
+                await new Promise(r => setTimeout(r, RETRY_DELAY_MS * attempt));
+            }
+            try {
+                const stream = this._executePromptOnce(prompt, {
+                    cwd, apiKey, model, env, settings, signal, attachmentDir, routingSessionId, routingPromptId, resumeSessionId,
+                });
+                for await (const event of stream) {
+                    yield event;
+                }
+                return; // Success — exit retry loop
+            }
+            catch (err) {
+                const is529 = err?.message?.includes('529') || err?.message?.includes('overloaded_error');
+                if (!is529 || attempt === MAX_RETRIES) {
+                    throw err; // Not retryable or out of retries
+                }
+                console.log(`[ClaudeBackend] 529 overloaded detected, will retry: ${err.message}`);
+            }
+        }
+    }
+    async *_executePromptOnce(prompt, config) {
         const { cwd, apiKey, model, env, settings, signal, attachmentDir, routingSessionId, routingPromptId, resumeSessionId } = config;
         // Build MCP server for this query
         const mcpServer = createModuleMcpServer({

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@exreve/exk",
-  "version": "1.0.77",
+  "version": "1.0.79",
   "description": "exk - Control Claude CLI with voice and programmable interfaces",
   "type": "module",
   "bin": {