ycode · ysamcode · Jul 28, 2026 · Jul 28, 2026
@@ -29,6 +29,10 @@ OPENAI_API_KEY=
 GEMINI_API_KEY=
 # Optional: override the default model (defaults to claude-sonnet-5).
 ANTHROPIC_MODEL=
+# Optional: seconds one AI chat request may run before the platform kills it.
+# Must match the AI chat route's maxDuration in vercel.json (300 by default —
+# the Vercel hobby-plan maximum). Raise both together on paid Vercel plans.
+AI_CHAT_MAX_DURATION=
 
 # Debug only: set to `true` to skip auth on /ycode/preview routes. Never enable in production.
 DISABLE_PREVIEW_AUTH=
@@ -20,11 +20,13 @@ import type { AgentContentBlock, AgentMessage } from '@/lib/agent/providers/type
 
 export const runtime = 'nodejs';
 export const dynamic = 'force-dynamic';
-// Vercel hard-kills the function at this limit without running catch/finally,
-// so the stream just stops and the turn looks silently truncated. The agent
-// loop stops itself earlier (MAX_RUN_MS in lib/agent/config.ts) to end runs
-// gracefully — keep that budget below this value.
-export const maxDuration = 800;
+// This route's maxDuration is set in vercel.json (300s — the Vercel hobby-plan
+// ceiling), NOT exported here: a route-level export always overrides
+// vercel.json, which would prevent paid deployments (e.g. Ycode Cloud) from
+// raising it in their own vercel.json. Deployments that raise it must also set
+// AI_CHAT_MAX_DURATION to match, so the agent loop (MAX_RUN_MS in
+// lib/agent/config.ts) stops itself before Vercel hard-kills the function
+// without running catch/finally.
 
 /**
  * Lightweight in-process rate limiter with a per-tenant sliding window, bounding

@@ -27,16 +27,27 @@ export const DEFAULT_MAX_TOKENS = 16384;
 export const MAX_TOOL_TURNS = 24;
 
 /**
- * Wall-clock budget for one agent run. The chat route's exported `maxDuration`
- * (app/(builder)/ycode/api/ai/chat/route.ts) hard-kills the function without
- * streaming anything, leaving the turn silently truncated. The agent loop stops
- * starting new turns once this budget is spent, so long runs end gracefully:
- * page/component snapshots and usage are emitted, and the user gets a
- * resumable "ran out of time" error instead of a silent cut. The buffer below
- * `maxDuration` must cover one full provider turn plus its tool calls and the
- * end-of-run snapshot/CSS work.
+ * Wall-clock budget for one agent run. Vercel's `maxDuration` for the chat
+ * route (set in vercel.json, not in the route file) hard-kills the function
+ * without streaming anything, leaving the turn silently truncated. The agent
+ * loop stops starting new turns once this budget is spent, so long runs end
+ * gracefully: page/component snapshots and usage are emitted, and the user
+ * gets a resumable "ran out of time" error instead of a silent cut.
+ *
+ * The default matches vercel.json's 300s (the Vercel hobby-plan ceiling).
+ * Deployments that raise maxDuration in their own vercel.json (e.g. Ycode
+ * Cloud at 800s) must set AI_CHAT_MAX_DURATION (seconds) to the same value.
+ * The 60s buffer below `maxDuration` must cover one full provider turn plus
+ * its tool calls and the end-of-run snapshot/CSS work.
  */
-export const MAX_RUN_MS = 740_000; // maxDuration (800s) minus a 60s buffer
+const DEFAULT_AI_CHAT_MAX_DURATION_SECONDS = 300;
+const RUN_BUFFER_MS = 60_000;
+const configuredMaxDurationSeconds = Number(process.env.AI_CHAT_MAX_DURATION);
+
+export const MAX_RUN_MS =
+  (Number.isFinite(configuredMaxDurationSeconds) && configuredMaxDurationSeconds > 60
+    ? configuredMaxDurationSeconds
+    : DEFAULT_AI_CHAT_MAX_DURATION_SECONDS) * 1000 - RUN_BUFFER_MS;
 
 /**
  * Cross-turn conversation history budget, applied before the agent runs so a long

@@ -9,6 +9,9 @@
     "app/(builder)/ycode/api/**/*.ts": {
       "maxDuration": 60
     },
+    "app/(builder)/ycode/api/ai/chat/route.ts": {
+      "maxDuration": 300
+    },
     "app/(builder)/ycode/mcp/*/route.ts": {
       "maxDuration": 60
     }