@exreve/exk 1.0.76 → 1.0.78

package/dist/cli/agentSession.js

@@ -56,7 +56,7 @@ 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-5.1';
+const DEFAULT_AI_MODEL = 'glm-5.2';
 /** TTL cache for ai-config.json reads to avoid hitting disk on every call */
 let _aiConfigCache = null;
 const AI_CONFIG_TTL_MS = 5_000;
@@ -66,12 +66,12 @@ const PROVIDERS = {
     zai: {
         apiKey: process.env.ZHIPU_API_KEY || '',
         baseUrl: process.env.CLI_AI_BASE_URL || 'https://api.z.ai/api/anthropic',
-        models: ['glm-5.1', 'glm-4.7', 'glm-4.5-air'],
+        models: ['glm-5.2'],
     },
     minimax: {
         apiKey: '',
         baseUrl: 'https://api.minimax.io/anthropic',
-        models: ['MiniMax-M2.7', 'MiniMax-M2.7-highspeed'],
+        models: ['MiniMax-M3'],
     },
     openrouter: {
         apiKey: '',
@@ -739,13 +739,13 @@ export class AgentSessionManager {
                         resumeSessionId = session.sdkSessionId;
                     }
                 }
-                // Enable auto-compaction: trigger when context reaches ~150k (of 200k), keep default target
+                // Enable auto-compaction: trigger when context reaches ~800k (of 1M), keep default target
                 if (effectiveSettings) {
                     effectiveSettings.autoCompactEnabled = true;
-                    effectiveSettings.autoCompactWindow = 150000;
+                    effectiveSettings.autoCompactWindow = 800000;
                 }
                 else {
-                    effectiveSettings = { autoCompactEnabled: true, autoCompactWindow: 150000 };
+                    effectiveSettings = { autoCompactEnabled: true, autoCompactWindow: 800000 };
                 }
                 // Build backend config
                 const backendConfig = {

package/dist/cli/benchmark-models-sdk.js

@@ -262,7 +262,7 @@ async function runSdkBenchmark(provider) {
 // ── Main ────────────────────────────────────────────────────────────
 async function main() {
     console.log(`\n${BOLD}╔══════════════════════════════════════════════════════════════════════╗${RESET}`);
-    console.log(`${BOLD}║  SDK Benchmark: MiniMax M2.7-highspeed vs Cerebras zai-glm-4.7     ║${RESET}`);
+    console.log(`${BOLD}║  SDK Benchmark: MiniMax M3 vs Cerebras zai-glm-4.7                 ║${RESET}`);
     console.log(`${BOLD}║  Task: Generate a complete HTML real estate page                    ║${RESET}`);
     console.log(`${BOLD}╚══════════════════════════════════════════════════════════════════════╝${RESET}\n`);
     const config = loadConfig();
@@ -278,8 +278,8 @@ async function main() {
     const providers = [];
     if (minimaxKey) {
         providers.push({
-            name: 'MiniMax M2.7-highspeed',
-            model: 'MiniMax-M2.7-highspeed',
+            name: 'MiniMax M3',
+            model: 'MiniMax-M3',
             apiKey: minimaxKey,
             baseUrl: 'https://api.minimax.io/anthropic',
             needsProxy: false,

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({
@@ -188,6 +216,8 @@ export class ClaudeBackend {
             settingSources: ['project'],
             permissionMode: 'bypassPermissions',
             allowDangerouslySkipPermissions: true,
+            // Enable 1M token context window (GLM-5.2 supports 1M context via z.ai)
+            betas: ['context-1m-2025-08-07'],
             mcpServers: { 'claude-voice-modules': mcpServer },
             ...(pathToClaudeCodeExecutable ? { pathToClaudeCodeExecutable } : {}),
             spawnClaudeCodeProcess: (spawnOptions) => {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@exreve/exk",
-  "version": "1.0.76",
+  "version": "1.0.78",
   "description": "exk - Control Claude CLI with voice and programmable interfaces",
   "type": "module",
   "bin": {