@blockrun/franklin 3.15.11 → 3.15.12

package/dist/agent/loop.js

@@ -24,7 +24,7 @@ import { logger, setDebugMode } from '../logger.js';
 import { estimateCost, OPUS_PRICING } from '../pricing.js';
 import { maybeMidSessionExtract } from '../learnings/extractor.js';
 import { extractMentions, buildEntityContext, loadEntities } from '../brain/store.js';
-import { routeRequestAsync, resolveTierToModel, parseRoutingProfile, getFallbackChain } from '../router/index.js';
+import { routeRequestAsync, resolveTierToModel, parseRoutingProfile, getFallbackChain, pickFreeFallback } from '../router/index.js';
 import { recordOutcome } from '../router/local-elo.js';
 import { shouldPlan, getPlanningPrompt, getExecutorModel, isExecutorStuck, toolCallSignature } from './planner.js';
 import { shouldVerify, runVerification } from './verification.js';
@@ -1098,8 +1098,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     if (lastRoutedCategory) {
                         recordOutcome(lastRoutedCategory, config.model, 'payment');
                     }
-                    const FREE_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'nvidia/glm-4.7'];
-                    const nextFree = FREE_MODELS.find(m => !turnFailedModels.has(m));
+                    const nextFree = pickFreeFallback(lastRoutedCategory, turnFailedModels);
                     if (nextFree) {
                         const oldModel = config.model;
                         config.model = nextFree;
@@ -1120,8 +1119,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     if (lastRoutedCategory) {
                         recordOutcome(lastRoutedCategory, config.model, 'rate_limit');
                     }
-                    const FREE_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'nvidia/glm-4.7'];
-                    const nextFree = FREE_MODELS.find(m => !turnFailedModels.has(m));
+                    const nextFree = pickFreeFallback(lastRoutedCategory, turnFailedModels);
                     if (nextFree) {
                         const oldModel = config.model;
                         config.model = nextFree;

package/dist/router/index.d.ts

@@ -49,6 +49,12 @@ export declare function routeRequest(prompt: string, profile?: RoutingProfile):
  * Get fallback models for a tier
  */
 export declare function getFallbackChain(tier: Tier, profile?: RoutingProfile): string[];
+/**
+ * Pick the next free model to try given the question category and which
+ * free models have already failed this turn. Returns undefined when every
+ * candidate has been exhausted (caller should surface an error to user).
+ */
+export declare function pickFreeFallback(category: string, alreadyFailed: Set<string>): string | undefined;
 /**
  * Parse routing profile from model string
  */

package/dist/router/index.js

@@ -483,10 +483,45 @@ function computeSavings(model) {
  */
 export function getFallbackChain(tier, profile = 'auto') {
     if (profile === 'free')
-        return ['nvidia/qwen3-coder-480b'];
+        return FREE_MODELS_BY_CATEGORY.chat;
     const config = AUTO_TIERS[tier];
     return [config.primary, ...config.fallback];
 }
+// ─── Free-tier fallback (used when paid models 402 / rate-limit) ───
+// Free fallback chains by question category. Used when a paid model fails
+// mid-turn (402 payment, rate-limit) and we need a zero-cost replacement
+// to keep the user moving without waiting for funding.
+//
+// The lists are ordered: best-fit free model first, then degraded fallbacks.
+// Coding goes to qwen3-coder; everything else (chat / trading / research /
+// reasoning / creative) prefers general-purpose free models that aren't
+// coder-tuned. Without this split, a BTC question that exhausted paid
+// 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.
+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'],
+};
+const DEFAULT_FREE_CHAIN = [
+    'nvidia/glm-4.7',
+    'nvidia/llama-4-maverick',
+    'nvidia/qwen3-coder-480b',
+];
+/**
+ * Pick the next free model to try given the question category and which
+ * free models have already failed this turn. Returns undefined when every
+ * candidate has been exhausted (caller should surface an error to user).
+ */
+export function pickFreeFallback(category, alreadyFailed) {
+    const chain = FREE_MODELS_BY_CATEGORY[category]
+        ?? DEFAULT_FREE_CHAIN;
+    return chain.find(m => !alreadyFailed.has(m));
+}
 /**
  * Parse routing profile from model string
  */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.15.11",
+  "version": "3.15.12",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {