@lightcone-ai/daemon 0.23.1 → 0.23.3

package/mcp-servers/_thin-proxy/forward.js

@@ -21,6 +21,11 @@ import { z } from 'zod';
 const SERVER_URL = process.env.SERVER_URL ?? '';
 const MACHINE_API_KEY = process.env.MACHINE_API_KEY ?? '';
 const AGENT_ID = process.env.AGENT_ID ?? '';
+// Daemon sets LIGHTCONE_AGENT_RUNTIME (claude / codex / kimi) when it
+// spawns the MCP server, so the server-side handler can route to a
+// vision/LLM backend matching the calling agent's own LLM stack
+// (avoid e.g. a codex agent silently using a claude vision call).
+const AGENT_RUNTIME = process.env.LIGHTCONE_AGENT_RUNTIME ?? '';
 
 function toTextContent(payload) {
   let text;
@@ -37,12 +42,16 @@ async function forwardToServer(serverId, toolName, args) {
     throw new Error('thin-proxy missing SERVER_URL / MACHINE_API_KEY / AGENT_ID env');
   }
   const url = `${SERVER_URL}/internal/agent/${encodeURIComponent(AGENT_ID)}/mcp/${encodeURIComponent(serverId)}/${encodeURIComponent(toolName)}`;
+  const headers = {
+    'Content-Type': 'application/json',
+    'Authorization': `Bearer ${MACHINE_API_KEY}`,
+  };
+  if (AGENT_RUNTIME) {
+    headers['X-Lightcone-Agent-Runtime'] = AGENT_RUNTIME;
+  }
   const res = await fetch(url, {
     method: 'POST',
-    headers: {
-      'Content-Type': 'application/json',
-      'Authorization': `Bearer ${MACHINE_API_KEY}`,
-    },
+    headers,
     body: JSON.stringify(args ?? {}),
   });
   let body = null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@lightcone-ai/daemon",
-  "version": "0.23.1",
+  "version": "0.23.3",
   "type": "module",
   "main": "src/index.js",
   "bin": {

package/src/agent-manager.js

@@ -707,6 +707,12 @@ export class AgentManager {
 
   _buildCodexMcpArgs(mcpServers) {
     const args = [];
+    // Codex's default tool_timeout_sec for MCP tool calls is short (~120s).
+    // Fine for fast tools, but kills slow vision/analysis tools — notably
+    // `page-understanding/analyze_page` which takes 5-10 min on long
+    // mp.weixin articles. Apply a 900s ceiling to every MCP server here;
+    // fast tools return well within this, slow tools stop getting killed.
+    const TOOL_TIMEOUT_SEC = 900;
     for (const [serverKey, server] of Object.entries(mcpServers)) {
       const normalizedKey = String(serverKey ?? '').trim();
       if (!normalizedKey) continue;
@@ -722,6 +728,7 @@ export class AgentManager {
         '-c', `mcp_servers.${keyExpr}.command=${commandExpr}`,
         '-c', `mcp_servers.${keyExpr}.args=${argsExpr}`,
         '-c', `mcp_servers.${keyExpr}.enabled=true`,
+        '-c', `mcp_servers.${keyExpr}.tool_timeout_sec=${TOOL_TIMEOUT_SEC}`,
       );
       if (server.required === true) {
         args.push('-c', `mcp_servers.${keyExpr}.required=true`);
@@ -788,6 +795,16 @@ export class AgentManager {
       '${WECHAT_MP_PROFILE_DIR}': path.join(profileRoot, `wechat_mp-${userId}`),
     };
     const mcpServers = this._resolveDirectiveMcpServers(directive, baseReplacements);
+    // Inject LIGHTCONE_AGENT_RUNTIME into every MCP server's env so that
+    // server-side handlers (reached via thin-proxy) can route to a backend
+    // matching the calling agent's own LLM stack. Example: analyze_page
+    // uses this to pick codex vision for codex agents and claude vision
+    // for claude agents — avoiding cross-vendor spawn chains.
+    for (const server of Object.values(mcpServers)) {
+      if (server && typeof server === 'object') {
+        server.env = { ...(server.env ?? {}), LIGHTCONE_AGENT_RUNTIME: runtime };
+      }
+    }
 
     if (runtime === 'codex') {
       const mcpKeys = Object.keys(mcpServers);

package/src/drivers/codex.js

@@ -214,13 +214,6 @@ export function buildCodexSpawn({
     authToken: config.authToken || machineApiKey,
   });
 
-  // Codex's default tool_timeout_sec for MCP tool calls is 60-120s. That's
-  // fine for fast tools (chat, db, etc) but breaks slow vision/analysis
-  // tools — notably `page-understanding/analyze_page` which takes 5-10 min
-  // on long mp.weixin articles (4 chunks × 1.6MB PNG → Claude vision OCR).
-  // Apply a generous 900s ceiling to every skill MCP server. Fast tools
-  // return well within this; slow tools no longer get prematurely killed.
-  const SKILL_TOOL_TIMEOUT_SEC = 900;
   for (const [serverKey, mc] of Object.entries(skillMcpServers)) {
     const keyExpr = formatCodexServerKey(serverKey);
     if (!keyExpr) continue;
@@ -229,16 +222,14 @@ export function buildCodexSpawn({
       args.push(
         '-c', `mcp_servers.${keyExpr}.command=${quote('env')}`,
         '-c', `mcp_servers.${keyExpr}.args=${quote([...envPairs, mc.command, ...(mc.args ?? [])])}`,
-        '-c', `mcp_servers.${keyExpr}.enabled=true`,
-        '-c', `mcp_servers.${keyExpr}.tool_timeout_sec=${SKILL_TOOL_TIMEOUT_SEC}`
+        '-c', `mcp_servers.${keyExpr}.enabled=true`
       );
       continue;
     }
     args.push(
       '-c', `mcp_servers.${keyExpr}.command=${quote(mc.command)}`,
       '-c', `mcp_servers.${keyExpr}.args=${quote(mc.args ?? [])}`,
-      '-c', `mcp_servers.${keyExpr}.enabled=true`,
-      '-c', `mcp_servers.${keyExpr}.tool_timeout_sec=${SKILL_TOOL_TIMEOUT_SEC}`
+      '-c', `mcp_servers.${keyExpr}.enabled=true`
     );
   }