@blockrun/franklin 3.27.0 → 3.27.2

package/dist/agent/loop.js

@@ -1348,7 +1348,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 const hasTools = responseParts.some(p => p.type === 'tool_use');
                 const hasThinking = responseParts.some(p => p.type === 'thinking');
                 if (!hasText && !hasTools && !hasThinking) {
-                    const EMPTY_FALLBACK_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/glm-4.7', 'zai/glm-5.1'];
+                    const EMPTY_FALLBACK_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'zai/glm-5.1'];
                     const nextModel = EMPTY_FALLBACK_MODELS.find(m => m !== config.model && !turnFailedModels.has(m));
                     if (nextModel && recoveryAttempts < 2) {
                         recoveryAttempts++;

package/dist/learnings/extractor.js

@@ -11,7 +11,7 @@ import { loadLearnings, mergeLearning, saveLearnings, loadSkills, saveSkill } fr
 const EXTRACTION_MODELS = [
     'nvidia/qwen3-coder-480b', // Agent-tested free model; strong at JSON tasks
     'nvidia/llama-4-maverick', // Agent-tested fallback
-    'nvidia/glm-4.7', // Chat fallback; not default for tool-heavy paths
+    'nvidia/deepseek-v4-flash', // 1M-ctx fallback (was glm-4.7 — NVIDIA NIM hung 2026-06-07)
 ];
 const VALID_CATEGORIES = new Set([
     'language', 'model_preference', 'tool_pattern', 'coding_style',

package/dist/router/index.js

@@ -575,16 +575,19 @@ export function getFallbackChain(tier, profile = 'auto') {
 // models was being handed to qwen3-coder-480b — a coder model trying to
 // do technical analysis. Reported 2026-05-03 with a markets question
 // routed to a coder model on Sonnet failure.
+// 2026-06-07: nvidia/glm-4.7 dropped from every chain — NVIDIA NIM hung, the
+// gateway redirects it to qwen3-coder-480b (already present here), so routing to
+// it just wasted a slot and mislabeled the model. qwen3-coder-480b + llama-4-
+// maverick are both healthy and cover all categories.
 const FREE_MODELS_BY_CATEGORY = {
-    coding: ['nvidia/qwen3-coder-480b', 'nvidia/glm-4.7', 'nvidia/llama-4-maverick'],
-    trading: ['nvidia/glm-4.7', 'nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
-    research: ['nvidia/glm-4.7', 'nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
-    reasoning: ['nvidia/glm-4.7', 'nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick'],
-    chat: ['nvidia/llama-4-maverick', 'nvidia/glm-4.7', 'nvidia/qwen3-coder-480b'],
-    creative: ['nvidia/llama-4-maverick', 'nvidia/glm-4.7', 'nvidia/qwen3-coder-480b'],
+    coding: ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick'],
+    trading: ['nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
+    research: ['nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
+    reasoning: ['nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
+    chat: ['nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
+    creative: ['nvidia/llama-4-maverick', 'nvidia/qwen3-coder-480b'],
 };
 const DEFAULT_FREE_CHAIN = [
-    'nvidia/glm-4.7',
     'nvidia/llama-4-maverick',
     'nvidia/qwen3-coder-480b',
 ];

package/dist/tools/moa.js

@@ -16,7 +16,7 @@ import { ModelClient } from '../agent/llm.js';
 const REFERENCE_MODELS = [
     'nvidia/qwen3-coder-480b', // Free, agent-tested coding
     'nvidia/llama-4-maverick', // Free, agent-tested general chat
-    'nvidia/glm-4.7', // Free chat fallback
+    'nvidia/deepseek-v4-flash', // Free, 1M ctx (was glm-4.7 — NVIDIA NIM hung 2026-06-07)
     'google/gemini-2.5-flash', // Fast, cheap
     'deepseek/deepseek-chat', // Cheap, good reasoning
 ];

package/dist/ui/model-picker.js

@@ -72,10 +72,10 @@ export const MODEL_SHORTCUTS = {
     'deepseek-v4': 'nvidia/deepseek-v4-flash',
     'deepseek-v4-flash': 'nvidia/deepseek-v4-flash',
     dsv4: 'nvidia/deepseek-v4-flash',
-    // V3.2 free fallback for users who specifically want the older Terminus
-    // checkpoint instead of the V4 Flash default.
-    'deepseek-v3.2': 'nvidia/deepseek-v3.2',
-    'deepseek-v3': 'nvidia/deepseek-v3.2',
+    // V3.2 NVIDIA NIM hung (server redirects to V4 Flash) — point the aliases
+    // straight at the live V4 Flash so the picker doesn't hand back a dead id.
+    'deepseek-v3.2': 'nvidia/deepseek-v4-flash',
+    'deepseek-v3': 'nvidia/deepseek-v4-flash',
     // Free (agent-tested BlockRun gateway free tier — refreshed 2026-04)
     free: 'nvidia/qwen3-coder-480b',
     glm4: 'nvidia/qwen3-coder-480b',

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.27.0",
+  "version": "3.27.2",
   "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": {