@blockrun/franklin 3.24.0 → 3.24.1

package/dist/agent/llm.js

@@ -79,6 +79,11 @@ function getModelRequestTimeoutMs() {
         180_000);
 }
 function getModelStreamIdleTimeoutMs() {
+    // Inter-chunk idle budget: the max gap allowed *between* SSE chunks once the
+    // stream is flowing. It does NOT cover time-to-first-token — that first read
+    // uses the larger request budget (see getModelRequestTimeoutMs + the
+    // firstRead branch in parseSSEStream). Conflating the two regressed #74:
+    // reasoning models taking 60–120s to first token aborted at this 90s wall.
     return (parseTimeoutEnv('FRANKLIN_MODEL_STREAM_IDLE_TIMEOUT_MS') ??
         parseTimeoutEnv('FRANKLIN_MODEL_IDLE_TIMEOUT_MS') ??
         90_000);
@@ -597,8 +602,11 @@ export class ModelClient {
                 yield* this.parseNonStreamingMessage(response, request.model);
                 return;
             }
-            // Parse SSE stream
-            yield* this.parseSSEStream(response, requestController, streamTimeoutMs, request.model);
+            // Parse SSE stream. The first read waits for time-to-first-token (which
+            // the gateway does *not* cover with the request timeout — it flushes SSE
+            // headers before the first content chunk), so it gets the larger request
+            // budget; subsequent reads use the tighter stream-idle budget.
+            yield* this.parseSSEStream(response, requestController, streamTimeoutMs, request.model, requestTimeoutMs);
         }
         finally {
             unlinkAbort();
@@ -1087,7 +1095,7 @@ export class ModelClient {
         return header;
     }
     // ─── SSE Parsing ───────────────────────────────────────────────────────
-    async *parseSSEStream(response, controller, timeoutMs, model) {
+    async *parseSSEStream(response, controller, timeoutMs, model, firstReadTimeoutMs = timeoutMs) {
         const reader = response.body?.getReader();
         if (!reader) {
             yield { kind: 'error', payload: { message: 'No response body' } };
@@ -1097,12 +1105,17 @@ export class ModelClient {
         let buffer = '';
         // Persist across read() calls — event: and data: may arrive in separate chunks
         let currentEvent = '';
+        // The first read waits for time-to-first-token (60–120s for reasoning
+        // models on cache-cold prompts); only later reads measure inter-chunk idle.
+        let firstRead = true;
         const MAX_BUFFER = 1_000_000; // 1MB buffer cap
         try {
             while (true) {
                 if (controller.signal.aborted)
                     break;
-                const { done, value } = await withAbortableTimeout(() => reader.read(), controller, createModelTimeoutError('stream', model, timeoutMs), timeoutMs);
+                const budgetMs = firstRead ? firstReadTimeoutMs : timeoutMs;
+                firstRead = false;
+                const { done, value } = await withAbortableTimeout(() => reader.read(), controller, createModelTimeoutError('stream', model, budgetMs), budgetMs);
                 if (done)
                     break;
                 buffer += decoder.decode(value, { stream: true });

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.24.0",
+  "version": "3.24.1",
   "description": "Franklin Agent — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {