@blockrun/franklin 3.9.2 → 3.9.4

package/dist/agent/llm.d.ts

@@ -90,6 +90,7 @@ export declare function modelHasExtendedThinking(model: string): boolean;
  * direct unit testing — the happy path hits it only on stream error.
  */
 export declare function classifyToolCallFailure(toolName: string, rawInput: string, signal: AbortSignal | undefined, model: string): string;
+export declare function isRoleplayedJsonToolCallText(text: string): boolean;
 export declare class ModelClient {
     private apiUrl;
     private chain;

package/dist/agent/llm.js

@@ -170,6 +170,23 @@ export function classifyToolCallFailure(toolName, rawInput, signal, model) {
         `Preview: ${preview}${rawInput.length > 120 ? '…' : ''} — ` +
         `this is usually a model output bug; try \`/model <other>\` or retry.]`;
 }
+export function isRoleplayedJsonToolCallText(text) {
+    const trimmed = text.trim();
+    if (!trimmed.startsWith('{') || !trimmed.endsWith('}'))
+        return false;
+    try {
+        const parsed = JSON.parse(trimmed);
+        return (parsed !== null &&
+            typeof parsed === 'object' &&
+            !Array.isArray(parsed) &&
+            parsed.type === 'function' &&
+            typeof parsed.name === 'string' &&
+            ('parameters' in parsed || 'arguments' in parsed));
+    }
+    catch {
+        return false;
+    }
+}
 function applyAnthropicPromptCaching(payload, request) {
     const out = { ...payload };
     const cacheMarker = { type: 'ephemeral' };
@@ -402,6 +419,7 @@ export class ModelClient {
         let currentToolId = '';
         let currentToolName = '';
         let currentToolInput = '';
+        const textEmission = { mode: 'undecided' };
         // Split inline <think>…</think> emitted by reasoning models (nemotron,
         // deepseek-r1, qwq, etc.) that use the text field instead of the native
         // thinking block. Thinking emitted this way is display-only — we don't
@@ -413,6 +431,24 @@ export class ModelClient {
         // system-prompt guard in loop.ts is responsible for preventing this.
         // Debug-only because the user already sees the literal text in the UI.
         let toolCallRoleplayWarned = false;
+        const appendText = (text) => {
+            if (!text)
+                return;
+            currentText += text;
+            if (textEmission.mode === 'undecided') {
+                const trimmed = currentText.trimStart();
+                if (!trimmed)
+                    return;
+                textEmission.mode = trimmed.startsWith('{') ? 'hold' : 'stream';
+                if (textEmission.mode === 'stream') {
+                    onStreamDelta?.({ type: 'text', text: currentText });
+                }
+                return;
+            }
+            if (textEmission.mode === 'stream') {
+                onStreamDelta?.({ type: 'text', text });
+            }
+        };
         for await (const chunk of this.streamCompletion(request, signal)) {
             switch (chunk.kind) {
                 case 'content_block_start': {
@@ -429,6 +465,7 @@ export class ModelClient {
                     }
                     else if (cblock?.type === 'text') {
                         currentText = '';
+                        textEmission.mode = 'undecided';
                         textStripper = new ThinkTagStripper();
                     }
                     break;
@@ -458,9 +495,7 @@ export class ModelClient {
                         }
                         for (const seg of textStripper.push(raw)) {
                             if (seg.type === 'text') {
-                                currentText += seg.text;
-                                if (seg.text)
-                                    onStreamDelta?.({ type: 'text', text: seg.text });
+                                appendText(seg.text);
                             }
                             else if (seg.text) {
                                 onStreamDelta?.({ type: 'thinking', text: seg.text });
@@ -537,20 +572,30 @@ export class ModelClient {
                         // Flush any partial tag held in the stripper
                         for (const seg of textStripper.flush()) {
                             if (seg.type === 'text') {
-                                currentText += seg.text;
-                                if (seg.text)
-                                    onStreamDelta?.({ type: 'text', text: seg.text });
+                                appendText(seg.text);
                             }
                             else if (seg.text) {
                                 onStreamDelta?.({ type: 'thinking', text: seg.text });
                             }
                         }
                         if (currentText) {
-                            collected.push({
-                                type: 'text',
-                                text: currentText,
-                            });
+                            if (textEmission.mode === 'hold' && isRoleplayedJsonToolCallText(currentText)) {
+                                if (this.debug) {
+                                    console.error(`[franklin] Model ${request.model} emitted a raw JSON function-call object as text. ` +
+                                        'Treating it as non-productive output so recovery can try another model.');
+                                }
+                            }
+                            else {
+                                if (textEmission.mode !== 'stream') {
+                                    onStreamDelta?.({ type: 'text', text: currentText });
+                                }
+                                collected.push({
+                                    type: 'text',
+                                    text: currentText,
+                                });
+                            }
                             currentText = '';
+                            textEmission.mode = 'undecided';
                         }
                     }
                     break;
@@ -588,16 +633,25 @@ export class ModelClient {
         // Flush any remaining text (stream ended without content_block_stop)
         for (const seg of textStripper.flush()) {
             if (seg.type === 'text') {
-                currentText += seg.text;
-                if (seg.text)
-                    onStreamDelta?.({ type: 'text', text: seg.text });
+                appendText(seg.text);
             }
             else if (seg.text) {
                 onStreamDelta?.({ type: 'thinking', text: seg.text });
             }
         }
         if (currentText) {
-            collected.push({ type: 'text', text: currentText });
+            if (textEmission.mode === 'hold' && isRoleplayedJsonToolCallText(currentText)) {
+                if (this.debug) {
+                    console.error(`[franklin] Model ${request.model} emitted a raw JSON function-call object as text. ` +
+                        'Treating it as non-productive output so recovery can try another model.');
+                }
+            }
+            else {
+                if (textEmission.mode !== 'stream') {
+                    onStreamDelta?.({ type: 'text', text: currentText });
+                }
+                collected.push({ type: 'text', text: currentText });
+            }
         }
         return { content: collected, usage, stopReason };
     }

package/dist/agent/loop.js

@@ -822,8 +822,10 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     '\n\n# Available tools\n' +
                     `You have exactly these tools: ${names}.\n` +
                     'Do not invent other tool names. Do not emit literal "[TOOLCALL]", ' +
-                    '"<tool_call>", or similar tokens in your text — call tools via the ' +
-                    'proper API only. If no tool fits, explain plainly in prose.';
+                    '"<tool_call>", raw JSON function-call objects like {"type":"function","name":"Tool","parameters":{}}, ' +
+                    'or similar tokens in your text — call tools via the proper API only. ' +
+                    'If the user asks you to echo a token, marker, or string, echo it as plain text; ' +
+                    'do not call Wallet or any other tool unless the user explicitly asks for that tool-backed information.';
             }
             // Safety net: handled in llm.ts resolveVirtualModel()
             // Sanitize: remove orphaned tool results that could confuse the API

package/dist/agent/tokens.js

@@ -197,6 +197,9 @@ const MODEL_CONTEXT_WINDOWS = {
     'moonshot/kimi-k2.6': 256_000,
     'moonshot/kimi-k2.5': 128_000,
     'minimax/minimax-m2.7': 128_000,
+    // NVIDIA-hosted free tier (2026-04-29 V4 Flash + Omni launch)
+    'nvidia/deepseek-v4-flash': 1_000_000,
+    'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning': 256_000,
 };
 /**
  * Get the context window size for a model, with a conservative default.

package/dist/commands/start.js

@@ -3,7 +3,7 @@ import { getOrCreateWallet, getOrCreateSolanaWallet } from '@blockrun/llm';
 import { loadChain, API_URLS } from '../config.js';
 import { retryFetchBalance } from './balance-retry.js';
 import { flushStats, loadStats } from '../stats/tracker.js';
-import { OPUS_PRICING } from '../pricing.js';
+import { OPUS_PRICING, MODEL_PRICING } from '../pricing.js';
 import { loadConfig } from './config.js';
 import { printBanner } from '../banner.js';
 import { assembleInstructions } from '../agent/context.js';
@@ -126,17 +126,11 @@ export async function startCommand(options) {
         return;
     }
     // Warn when a paid model is active so users know they'll be charged.
-    // Set members = BlockRun gateway's current live free tier (refreshed 2026-04).
-    const FREE_MODELS = new Set([
-        'nvidia/glm-4.7',
-        'nvidia/qwen3-next-80b-a3b-thinking',
-        'nvidia/qwen3-coder-480b',
-        'nvidia/mistral-small-4-119b',
-        'nvidia/llama-4-maverick',
-        'nvidia/deepseek-v3.2',
-        'blockrun/free',
-    ]);
-    if (!FREE_MODELS.has(model)) {
+    // Derive "free" from MODEL_PRICING so adding a new free entry there is enough —
+    // no second hardcoded list to keep in sync.
+    const pricing = MODEL_PRICING[model];
+    const isFree = pricing != null && pricing.input === 0 && pricing.output === 0 && (pricing.perCall ?? 0) === 0;
+    if (!isFree) {
         console.log(chalk.yellow(`  Model: ${model}  (paid — charges from your wallet per call)`));
         console.log(chalk.dim(`  Switch to free with: /model free\n`));
     }

package/dist/pricing.js

@@ -8,7 +8,9 @@ export const MODEL_PRICING = {
     'blockrun/eco': { input: 0.2, output: 1.0 },
     'blockrun/premium': { input: 3.0, output: 15.0 },
     'blockrun/free': { input: 0, output: 0 },
-    // FREE — BlockRun gateway free tier (refreshed 2026-04)
+    // FREE — BlockRun gateway free tier (refreshed 2026-04-29 with V4 Flash + Omni launch)
+    'nvidia/deepseek-v4-flash': { input: 0, output: 0 },
+    'nvidia/nemotron-3-nano-omni-30b-a3b-reasoning': { input: 0, output: 0 },
     'nvidia/glm-4.7': { input: 0, output: 0 },
     'nvidia/qwen3-next-80b-a3b-thinking': { input: 0, output: 0 },
     'nvidia/qwen3-coder-480b': { input: 0, output: 0 },

package/dist/ui/model-picker.js

@@ -118,26 +118,26 @@ export const PICKER_CATEGORIES = [
         ],
     },
     {
+        // Picker trim (v3.9.3): hide superseded / awkward-middle / niche-premium
+        // entries to bring choice paralysis down. Their shortcuts (`opus-4.6`,
+        // `gpt-5.4`, `gpt-5.4-pro`, `grok`, `o1`, `o4`, `nano`) all stay live in
+        // MODEL_SHORTCUTS, so muscle memory keeps working — they just aren't
+        // listed in the visible picker. Same pattern v3.9.0 used to retire dead
+        // free-tier entries and v3.9.2 used to retire Kimi K2.5.
         category: '✨ Premium frontier',
         models: [
             { id: 'anthropic/claude-opus-4.7', shortcut: 'opus', label: 'Claude Opus 4.7', price: '$5/$25', highlight: true },
             { id: 'anthropic/claude-sonnet-4.6', shortcut: 'sonnet', label: 'Claude Sonnet 4.6', price: '$3/$15' },
-            { id: 'anthropic/claude-opus-4.6', shortcut: 'opus-4.6', label: 'Claude Opus 4.6', price: '$5/$25' },
             { id: 'openai/gpt-5.5', shortcut: 'gpt', label: 'GPT-5.5', price: '$5/$30', highlight: true },
-            { id: 'openai/gpt-5.4', shortcut: 'gpt-5.4', label: 'GPT-5.4', price: '$2.5/$15' },
-            { id: 'openai/gpt-5.4-pro', shortcut: 'gpt-5.4-pro', label: 'GPT-5.4 Pro', price: '$30/$180' },
-            { id: 'google/gemini-2.5-pro', shortcut: 'gemini', label: 'Gemini 2.5 Pro', price: '$1.25/$10' },
             { id: 'google/gemini-3.1-pro', shortcut: 'gemini-3', label: 'Gemini 3.1 Pro', price: '$2/$12' },
+            { id: 'google/gemini-2.5-pro', shortcut: 'gemini', label: 'Gemini 2.5 Pro', price: '$1.25/$10' },
             { id: 'xai/grok-4-0709', shortcut: 'grok-4', label: 'Grok 4', price: '$0.2/$1.5' },
-            { id: 'xai/grok-3', shortcut: 'grok', label: 'Grok 3', price: '$3/$15' },
         ],
     },
     {
         category: '🔬 Reasoning',
         models: [
             { id: 'openai/o3', shortcut: 'o3', label: 'O3', price: '$2/$8' },
-            { id: 'openai/o4-mini', shortcut: 'o4', label: 'O4 Mini', price: '$1.1/$4.4' },
-            { id: 'openai/o1', shortcut: 'o1', label: 'O1', price: '$15/$60' },
             { id: 'openai/gpt-5.3-codex', shortcut: 'codex', label: 'GPT-5.3 Codex', price: '$1.75/$14' },
             { id: 'deepseek/deepseek-reasoner', shortcut: 'r1', label: 'DeepSeek R1', price: '$0.28/$0.42' },
             { id: 'xai/grok-4-1-fast-reasoning', shortcut: 'grok-fast', label: 'Grok 4.1 Fast R.', price: '$0.2/$0.5' },
@@ -148,7 +148,6 @@ export const PICKER_CATEGORIES = [
         models: [
             { id: 'anthropic/claude-haiku-4.5-20251001', shortcut: 'haiku', label: 'Claude Haiku 4.5', price: '$1/$5' },
             { id: 'openai/gpt-5-mini', shortcut: 'mini', label: 'GPT-5 Mini', price: '$0.25/$2' },
-            { id: 'openai/gpt-5-nano', shortcut: 'nano', label: 'GPT-5 Nano', price: '$0.05/$0.4' },
             { id: 'google/gemini-2.5-flash', shortcut: 'flash', label: 'Gemini 2.5 Flash', price: '$0.3/$2.5' },
             { id: 'deepseek/deepseek-chat', shortcut: 'deepseek', label: 'DeepSeek V3', price: '$0.28/$0.42' },
             { id: 'moonshot/kimi-k2.6', shortcut: 'kimi', label: 'Kimi K2.6', price: '$0.95/$4' },

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.9.2",
+  "version": "3.9.4",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {