feat(gateway): use gateway service locally if specified (#247)

Author: joshsny

.env.example

@@ -12,4 +12,7 @@ APPLE_CODESIGN_KEYCHAIN_PASSWORD="xxx"
 
 VITE_POSTHOG_API_KEY=xxx
 VITE_POSTHOG_API_HOST=xxx
-VITE_POSTHOG_UI_HOST=xxx
+VITE_POSTHOG_UI_HOST=xxx
+
+# Use new LLM gateway locally (experimental, needs to be started in mprocs)
+LLM_GATEWAY_URL=http://localhost:3308

apps/array/.env.example

@@ -411,7 +411,12 @@ export class SessionManager {
     process.env.POSTHOG_AUTH_HEADER = `Bearer ${credentials.apiKey}`;
     process.env.ANTHROPIC_API_KEY = credentials.apiKey;
     process.env.ANTHROPIC_AUTH_TOKEN = credentials.apiKey;
-    process.env.ANTHROPIC_BASE_URL = `${credentials.apiHost}/api/projects/${credentials.projectId}/llm_gateway`;
+
+    const llmGatewayUrl =
+      process.env.LLM_GATEWAY_URL ||
+      `${credentials.apiHost}/api/projects/${credentials.projectId}/llm_gateway`;
+    process.env.ANTHROPIC_BASE_URL = llmGatewayUrl;
+
     // process.env.ELECTRON_RUN_AS_NODE = "1";
     process.env.CLAUDE_CODE_EXECUTABLE = getClaudeCliPath();
 

apps/array/src/main/services/session-manager.ts

@@ -142,7 +142,9 @@ async function testAgent() {
           POSTHOG_AUTH_HEADER: `Bearer ${process.env.POSTHOG_API_KEY}`,
           ANTHROPIC_API_KEY: process.env.POSTHOG_API_KEY,
           ANTHROPIC_AUTH_TOKEN: process.env.POSTHOG_API_KEY,
-          ANTHROPIC_BASE_URL: `${process.env.POSTHOG_API_URL}/api/projects/${process.env.POSTHOG_PROJECT_ID}/llm_gateway`,
+          ANTHROPIC_BASE_URL:
+            process.env.LLM_GATEWAY_URL ||
+            `${process.env.POSTHOG_API_URL}/api/projects/${process.env.POSTHOG_PROJECT_ID}/llm_gateway`,
         },
       },
     });

packages/agent/example.ts

@@ -673,12 +673,19 @@ export class ClaudeAcpAgent implements Agent {
             throw RequestError.authRequired();
           }
 
-          // For assistant messages, text/thinking are normally streamed via stream_event.
-          // But some gateways (like LiteLLM) don't stream, so we process all content.
+          // When using the gateway service, text/thinking is streamed via stream_event,
+          // so skip them here to avoid duplication.
           const content = message.message.content;
+          const isUsingGatewayService = !!process.env.LLM_GATEWAY_URL;
+          const contentToProcess =
+            isUsingGatewayService && Array.isArray(content)
+              ? content.filter(
+                  (block) => block.type !== "text" && block.type !== "thinking",
+                )
+              : content;
 
           for (const notification of toAcpNotifications(
-            content,
+            contentToProcess as typeof content,
             message.message.role,
             params.sessionId,
             this.toolUseCache,

packages/agent/src/adapters/claude/claude.ts

@@ -134,7 +134,8 @@ export class Agent {
     }
 
     try {
-      const gatewayUrl = this.posthogAPI.getLlmGatewayUrl();
+      const gatewayUrl =
+        process.env.LLM_GATEWAY_URL || this.posthogAPI.getLlmGatewayUrl();
       const apiKey = this.posthogAPI.getApiKey();
       process.env.ANTHROPIC_BASE_URL = gatewayUrl;
       process.env.ANTHROPIC_AUTH_TOKEN = apiKey;