@blockrun/franklin 3.8.44 → 3.9.1

package/README.md

@@ -105,7 +105,7 @@ Concretely — $1 in USDC gets you roughly:
 - ~13M Gemini Flash tokens
 - ~20 DALL-E 3 images
 - ~40 Exa neural web searches
-- Unlimited NVIDIA GPT-OSS (free tier, no wallet needed)
+- Unlimited agent-tested NVIDIA free tier (Qwen3 Coder + Llama 4 Maverick, no wallet needed)
 
 ---
 
@@ -440,7 +440,7 @@ src/
 Start with **zero dollars**. Franklin defaults to free NVIDIA models that need no wallet funding.
 
 ```bash
-franklin --model nvidia/qwen3-next-80b-a3b-thinking
+franklin --model free
 ```
 
 When you fund the wallet, Franklin gets more purchasing power: Sonnet, Opus, GPT, Gemini, Grok, and paid tools like Exa, DALL-E, and CoinGecko Pro.
@@ -493,6 +493,7 @@ npm install
 npm run build
 npm test              # deterministic local tests — no API calls
 npm run test:e2e      # live e2e tests — free smoke works unfunded; paid tools need network + funded wallet
+npm run test:free-models # live matrix across current free NVIDIA models
 node dist/index.js --help
 ```
 

package/dist/agent/commands.d.ts

@@ -8,6 +8,7 @@
  */
 import type { ModelClient } from './llm.js';
 import type { AgentConfig, Dialogue, StreamEvent } from './types.js';
+import type { Registry } from '../skills/registry.js';
 type EventEmitter = (event: StreamEvent) => void;
 interface CommandContext {
     history: Dialogue[];
@@ -15,6 +16,10 @@ interface CommandContext {
     client: ModelClient;
     sessionId: string;
     onEvent: EventEmitter;
+    /** Skills loaded for this session — see src/skills/. */
+    skillRegistry?: Registry;
+    /** Runtime variables substituted into skill bodies before $ARGUMENTS. */
+    skillVars?: Record<string, string>;
 }
 interface CommandResult {
     handled: boolean;

package/dist/agent/commands.js

@@ -15,6 +15,7 @@ import { forceCompact } from './compact.js';
 import { getStatsSummary } from '../stats/tracker.js';
 import { resolveModel } from '../ui/model-picker.js';
 import { listSessions, loadSessionHistory, } from '../session/storage.js';
+import { matchSkill } from '../skills/invoke.js';
 // ─── Git helpers ──────────────────────────────────────────────────────────
 function gitExec(cmd, cwd, timeout = 5000, maxBuffer) {
     return execSync(cmd, {
@@ -197,6 +198,20 @@ const DIRECT_COMMANDS = {
     },
     '/help': (ctx) => {
         const ultrathinkOn = ctx.config.ultrathink;
+        let skillsBlock = '';
+        if (ctx.skillRegistry) {
+            const visible = ctx.skillRegistry
+                .list()
+                .filter((l) => !l.skill.disableModelInvocation);
+            if (visible.length > 0) {
+                skillsBlock =
+                    `\n  **Skills:**\n` +
+                        visible
+                            .map((l) => `    /${l.skill.name.padEnd(22)} ${l.skill.description}`)
+                            .join('\n') +
+                        `\n`;
+            }
+        }
         ctx.onEvent({ kind: 'text_delta', text: `**RunCode Commands**\n\n` +
                 `  **Coding:** /commit /review /test /fix /debug /explain /search /find /refactor /scaffold\n` +
                 `  **Git:** /push /pr /undo /status /diff /log /branch /stash /unstash\n` +
@@ -205,6 +220,7 @@ const DIRECT_COMMANDS = {
                 `  **Power:** /ultrathink [query] /ultraplan /noplan /moa [query] /dump\n` +
                 `  **Info:** /model /auto /wallet /cost /tokens /learnings /brain /mcp /doctor /version /bug /help\n` +
                 `  **UI:** /clear /exit\n` +
+                skillsBlock +
                 (ultrathinkOn ? `\n  Ultrathink: ON\n` : '')
         });
         emitDone(ctx);
@@ -926,11 +942,23 @@ export async function handleSlashCommand(input, ctx) {
             return { handled: false, rewritten: rewrite(arg) };
         }
     }
+    // File-loaded skills — registered after built-ins so `/security` etc.
+    // are never shadowed by a user-installed skill of the same name.
+    if (ctx.skillRegistry) {
+        const skillResult = matchSkill(input, ctx.skillRegistry, ctx.skillVars ?? {});
+        if (skillResult) {
+            return { handled: false, rewritten: skillResult.rewritten };
+        }
+    }
     // Not a recognized command — suggest closest match
+    const skillNames = ctx.skillRegistry
+        ? ctx.skillRegistry.list().map((s) => `/${s.skill.name}`)
+        : [];
     const allCommands = [
         ...Object.keys(DIRECT_COMMANDS),
         ...Object.keys(REWRITE_COMMANDS),
         ...ARG_COMMANDS.map(c => c.prefix.trim()),
+        ...skillNames,
         '/branch', '/resume', '/model', '/auto', '/wallet', '/cost', '/help', '/clear', '/retry', '/exit', '/session-search', '/ssearch', '/failures',
     ];
     const cmd = input.split(/\s/)[0];

package/dist/agent/compact.js

@@ -427,7 +427,7 @@ function formatCompactSummary(raw) {
 function pickCompactionModel(primaryModel) {
     // Free parent → free compaction (no silent charge)
     if (primaryModel.startsWith('nvidia/') || primaryModel === 'blockrun/free') {
-        return 'nvidia/glm-4.7';
+        return 'nvidia/qwen3-coder-480b';
     }
     // Use cheapest capable model for summarization to save cost
     // Tier down: opus/pro → sonnet, sonnet → haiku, everything else → flash (cheapest capable)

package/dist/agent/context.js

@@ -16,6 +16,7 @@ You are an interactive agent — not a chatbot. Use the tools available to you t
 
 # Franklin has hands
 You run with live tools by default:
+- **Wallet** — read your own chain, address, and USDC balance. Use this for any "what's my balance / how much money / 钱包余额 / wallet status" question instead of running \`franklin balance\` via Bash. Free, one call, never costs USDC.
 - **TradingMarket** — current stock / FX / crypto / commodity prices (BlockRun Gateway / Pyth; wallet pays automatically, $0.001/stock call, free for everything else).
 - **ExaAnswer / ExaSearch / ExaReadUrls** — cited current-events answers, semantic web search, clean URL content.
 - **WebSearch / WebFetch** — live web.

package/dist/agent/llm.js

@@ -256,12 +256,12 @@ export class ModelClient {
         // Static fallback if router is unavailable. Default to FREE model so
         // users aren't silently charged when their intended model can't resolve.
         const FALLBACKS = {
-            'blockrun/auto': 'nvidia/glm-4.7',
-            'blockrun/eco': 'nvidia/glm-4.7',
+            'blockrun/auto': 'nvidia/qwen3-coder-480b',
+            'blockrun/eco': 'nvidia/qwen3-coder-480b',
             'blockrun/premium': 'anthropic/claude-sonnet-4.6',
-            'blockrun/free': 'nvidia/glm-4.7',
+            'blockrun/free': 'nvidia/qwen3-coder-480b',
         };
-        return FALLBACKS[model] || 'nvidia/glm-4.7';
+        return FALLBACKS[model] || 'nvidia/qwen3-coder-480b';
     }
     async *streamCompletion(request, signal) {
         // Resolve virtual models before any API call

package/dist/agent/loop.js

@@ -6,6 +6,7 @@ import { ModelClient } from './llm.js';
 import { autoCompactIfNeeded, forceCompact, microCompact } from './compact.js';
 import { estimateHistoryTokens, updateActualTokens, resetTokenAnchor, getAnchoredTokenCount, getContextWindow, setEstimationModel } from './tokens.js';
 import { handleSlashCommand } from './commands.js';
+import { loadBundledSkills, getSkillVars } from '../skills/bootstrap.js';
 import { reduceTokens } from './reduce.js';
 import { PermissionManager } from './permissions.js';
 import { StreamingExecutor } from './streaming-executor.js';
@@ -225,8 +226,8 @@ function getBackoffDelay(attempt, maxDelayMs = 32_000) {
 export function isWeakModel(model) {
     const m = model.toLowerCase();
     // NVIDIA-hosted open models have been observed confabulating tool calls.
-    // `blockrun/free` and `blockrun/eco` resolve to nvidia/nemotron-ultra in
-    // llm.ts, so catching the `nvidia/` prefix also catches those paths.
+    // `blockrun/free` resolves to an NVIDIA model before the API call, so
+    // catching the `nvidia/` prefix also catches the free-profile path.
     if (m.startsWith('nvidia/'))
         return true;
     if (m.includes('nemotron-ultra'))
@@ -297,6 +298,17 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
     let lastUserInput = ''; // For /retry
     config.baseModel = config.model; // User's intended model — /model command updates this
     let turnFailedModels = new Set(); // Models that failed this turn (cleared each new turn)
+    // ── Skills (file-loaded SKILL.md prompt-rewrite slash commands) ──
+    // Bundled-only in Phase 1 of the skills MVP. User-global and project-local
+    // discovery + the budget-cap-usd / cost-receipt enforcement contract land
+    // in Phase 2 — see docs/plans/2026-04-29-franklin-skills-mvp-design.md.
+    const skillBoot = loadBundledSkills();
+    if (skillBoot.errors.length > 0 && config.debug) {
+        for (const err of skillBoot.errors) {
+            onEvent({ kind: 'text_delta', text: `[skills] ${err.path}: ${err.error}\n` });
+        }
+    }
+    const skillRegistry = skillBoot.registry;
     // Track models that failed with 402 (payment required) across turns.
     // These persist until the session ends — unlike transient errors, payment failures
     // will keep failing until the user adds funds. Map stores failure timestamp for future TTL.
@@ -387,6 +399,19 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
             else {
                 const cmdResult = await handleSlashCommand(input, {
                     history, config, client, sessionId, onEvent,
+                    skillRegistry,
+                    skillVars: getSkillVars({
+                        chain: config.chain,
+                        perTurnCapUsd: (() => {
+                            const raw = loadConfig()['max-turn-spend-usd'];
+                            if (raw == null)
+                                return 1.0;
+                            const n = Number(raw);
+                            if (!Number.isFinite(n))
+                                return 1.0;
+                            return n <= 0 ? Infinity : n;
+                        })(),
+                    }),
                 });
                 if (cmdResult.handled)
                     continue;
@@ -498,15 +523,17 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
         // is treated as a typo and falls back to the safe default rather than
         // silently removing the wallet guard.
         //
-        // Default raised from $0.25 → $1.00 in v3.8.42 — the original ceiling
-        // dated from when Franklin was mostly chat. Real workloads (multi-stage
-        // dashboard scaffolds on sonnet, image-to-image edits, research-heavy
-        // turns) routinely land in the $0.20–$0.80 range on a single legit
-        // prompt. $1.00 is still meaningful as a runaway-protection guardrail
-        // (catches the kind of failure v3.8.41's retry-policy was built for)
-        // but doesn't impose a friction tax on every multi-stage task. Users
-        // who liked the old ceiling can opt back in via the config.
-        const TURN_SPEND_DEFAULT_USD = 1.0;
+        // Default lineage: $0.25 (≤ v3.8.41) → $1.00 (v3.8.42) → $2.00 (v3.9.1).
+        // v3.8.42's $1.00 was tuned for "multi-stage dashboard scaffold" workloads
+        // landing in the $0.20–$0.80 range on a single prompt. Real coding turns
+        // since — full BTC-style dashboards, multi-file refactors that pull in
+        // sonnet/opus on a COMPLEX-tier route — routinely cross $1.00 in their
+        // first planning pass alone, leaving no headroom for the execution call
+        // and tripping the cap mid-task. $2.00 keeps the runaway-protection
+        // promise (catches the buggy-loop drain v3.8.41's retry-policy targets)
+        // while letting a legitimate complex coding task finish in one turn.
+        // Users who liked the old ceiling can pin it via the config.
+        const TURN_SPEND_DEFAULT_USD = 2.0;
         const turnSpendCap = (() => {
             const raw = loadConfig()['max-turn-spend-usd'];
             if (raw == null)
@@ -985,7 +1012,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                     if (lastRoutedCategory) {
                         recordOutcome(lastRoutedCategory, config.model, 'payment');
                     }
-                    const FREE_MODELS = ['nvidia/glm-4.7', 'nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'nvidia/qwen3-next-80b-a3b-thinking'];
+                    const FREE_MODELS = ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick', 'nvidia/glm-4.7'];
                     const nextFree = FREE_MODELS.find(m => !turnFailedModels.has(m));
                     if (nextFree) {
                         const oldModel = config.model;
@@ -1033,7 +1060,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 onEvent({
                     kind: 'text_delta',
                     text: `\n\n⚠️ Turn spend limit reached ($${turnSpend.toFixed(3)} > $${MAX_TURN_SPEND_USD}). Stopping to protect your wallet.\n` +
-                        `Raise the cap with \`franklin config set max-turn-spend-usd 2.0\` (or \`0\` to disable), then \`/retry\`.\n`,
+                        `Raise the cap with \`franklin config set max-turn-spend-usd 4.0\` (or \`0\` to disable), then \`/retry\`.\n`,
                 });
                 onEvent({ kind: 'turn_done', reason: 'budget' });
                 break;

package/dist/agent/verification.js

@@ -112,8 +112,8 @@ export async function runVerification(history, handlers, client, config) {
         },
     ];
     config.onEvent?.({ kind: 'text_delta', text: '\n*Verifying...*\n' });
-    // Use cheap model for verification
-    const verificationModel = 'nvidia/glm-4.7'; // Free model to keep cost zero
+    // Use agent-tested free model for verification.
+    const verificationModel = 'nvidia/qwen3-coder-480b';
     try {
         // Simple single-turn verification call
         const response = await client.complete({

package/dist/commands/balance-retry.d.ts

@@ -0,0 +1,15 @@
+/**
+ * Wallet-balance retry helper.
+ *
+ * Used by the agent UI's startup balance fetch and the post-turn refresh.
+ * Some wallet client paths return `0` transiently — for example, when the
+ * SDK is queried before the chain provider has finished initializing — even
+ * when the on-chain balance is non-zero. A single defensive retry catches
+ * that case without lengthening the path for a genuinely empty wallet:
+ * empty wallets still resolve to `0` in roughly two RPC round-trips.
+ */
+export interface RetryOptions {
+    /** Delay between the first and second attempt, in milliseconds. */
+    delayMs?: number;
+}
+export declare function retryFetchBalance(fetchOnce: () => Promise<number>, opts?: RetryOptions): Promise<number>;

package/dist/commands/balance-retry.js

@@ -0,0 +1,20 @@
+/**
+ * Wallet-balance retry helper.
+ *
+ * Used by the agent UI's startup balance fetch and the post-turn refresh.
+ * Some wallet client paths return `0` transiently — for example, when the
+ * SDK is queried before the chain provider has finished initializing — even
+ * when the on-chain balance is non-zero. A single defensive retry catches
+ * that case without lengthening the path for a genuinely empty wallet:
+ * empty wallets still resolve to `0` in roughly two RPC round-trips.
+ */
+export async function retryFetchBalance(fetchOnce, opts = {}) {
+    const first = await fetchOnce();
+    if (first !== 0)
+        return first;
+    await sleep(opts.delayMs ?? 750);
+    return fetchOnce();
+}
+function sleep(ms) {
+    return new Promise((resolve) => setTimeout(resolve, ms));
+}

package/dist/commands/skills.d.ts

@@ -0,0 +1,8 @@
+/**
+ * `franklin skills` — list and inspect SKILL.md files Franklin would load
+ * during a session. Phase 1 of the skills MVP — bundled only.
+ */
+export interface SkillsCommandOptions {
+    json?: boolean;
+}
+export declare function skillsCommand(action: string | undefined, arg: string | undefined, opts?: SkillsCommandOptions): Promise<void>;

package/dist/commands/skills.js

@@ -0,0 +1,93 @@
+/**
+ * `franklin skills` — list and inspect SKILL.md files Franklin would load
+ * during a session. Phase 1 of the skills MVP — bundled only.
+ */
+import chalk from 'chalk';
+import { loadBundledSkills } from '../skills/bootstrap.js';
+export async function skillsCommand(action, arg, opts = {}) {
+    const sub = action ?? 'list';
+    if (sub === 'list') {
+        runList(opts);
+        return;
+    }
+    if (sub === 'which') {
+        runWhich(arg);
+        return;
+    }
+    console.log(chalk.red(`Unknown skills subcommand: ${sub}`));
+    console.log('Usage: franklin skills [list|which <name>]');
+    process.exit(1);
+}
+function runList(opts) {
+    const { registry, errors } = loadBundledSkills();
+    const skills = registry.list();
+    if (opts.json) {
+        process.stdout.write(JSON.stringify({
+            skills: skills.map((l) => ({
+                name: l.skill.name,
+                description: l.skill.description,
+                source: l.source,
+                path: l.path,
+                warnings: l.warnings,
+                costReceipt: l.skill.costReceipt ?? false,
+                budgetCapUsd: l.skill.budgetCapUsd ?? null,
+                disableModelInvocation: l.skill.disableModelInvocation ?? false,
+            })),
+            errors,
+            shadowed: registry.shadowed().map((s) => ({
+                winner: { name: s.winner.skill.name, source: s.winner.source, path: s.winner.path },
+                loser: { name: s.loser.skill.name, source: s.loser.source, path: s.loser.path },
+            })),
+        }, null, 2) + '\n');
+        return;
+    }
+    if (skills.length === 0) {
+        console.log(chalk.dim('No skills loaded.'));
+    }
+    else {
+        console.log(chalk.bold(`Skills (${skills.length})`));
+        console.log('');
+        const nameWidth = Math.max(...skills.map((l) => l.skill.name.length), 4);
+        for (const l of skills) {
+            const flags = [];
+            if (l.skill.costReceipt)
+                flags.push('receipt');
+            if (typeof l.skill.budgetCapUsd === 'number')
+                flags.push(`cap $${l.skill.budgetCapUsd.toFixed(2)}`);
+            if (l.skill.disableModelInvocation)
+                flags.push('manual-only');
+            const flagStr = flags.length > 0 ? chalk.dim(` [${flags.join(', ')}]`) : '';
+            const sourceTag = chalk.dim(`(${l.source})`);
+            console.log(`  ${chalk.cyan('/' + l.skill.name.padEnd(nameWidth))}  ${l.skill.description}${flagStr} ${sourceTag}`);
+        }
+    }
+    const shadowed = registry.shadowed();
+    if (shadowed.length > 0) {
+        console.log('');
+        console.log(chalk.yellow('Shadowed (lost to a higher-precedence source):'));
+        for (const s of shadowed) {
+            console.log(`  /${s.loser.skill.name} from ${s.loser.source} ` +
+                chalk.dim(`(winner: ${s.winner.source} at ${s.winner.path})`));
+        }
+    }
+    if (errors.length > 0) {
+        console.log('');
+        console.log(chalk.red(`Failed to load (${errors.length}):`));
+        for (const e of errors) {
+            console.log(`  ${e.path}: ${e.error}`);
+        }
+    }
+}
+function runWhich(name) {
+    if (!name) {
+        console.log(chalk.red('Usage: franklin skills which <name>'));
+        process.exit(1);
+    }
+    const { registry } = loadBundledSkills();
+    const skill = registry.lookup(name);
+    if (!skill) {
+        console.log(chalk.red(`Skill not found: ${name}`));
+        process.exit(1);
+    }
+    console.log(skill.path);
+}

package/dist/commands/social.js

@@ -159,7 +159,7 @@ async function runCommand(options) {
     const chain = loadChain();
     const apiUrl = API_URLS[chain];
     const appConfig = loadAppConfig();
-    const model = options.model || appConfig['default-model'] || 'nvidia/glm-4.7';
+    const model = options.model || appConfig['default-model'] || 'nvidia/qwen3-coder-480b';
     console.log(chalk.dim(`  Model: ${model}`));
     console.log('');
     let result;

package/dist/commands/start.js

@@ -1,6 +1,7 @@
 import chalk from 'chalk';
 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 { loadConfig } from './config.js';
@@ -125,7 +126,7 @@ 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 free tier (refreshed 2026-04).
+    // 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',
@@ -133,8 +134,6 @@ export async function startCommand(options) {
         'nvidia/mistral-small-4-119b',
         'nvidia/llama-4-maverick',
         'nvidia/deepseek-v3.2',
-        'nvidia/gpt-oss-120b',
-        'nvidia/gpt-oss-20b',
         'blockrun/free',
     ]);
     if (!FREE_MODELS.has(model)) {
@@ -183,20 +182,25 @@ export async function startCommand(options) {
     console.log(chalk.dim('  Dashboard: ') + (panelUrl ? chalk.cyan(panelUrl) : chalk.cyan('franklin panel') + chalk.dim(' → http://localhost:3100')));
     console.log(chalk.dim('  Help:      ') + chalk.cyan('/help'));
     console.log('');
-    // Balance fetcher — used at startup and after each turn
+    // Balance fetcher — used at startup and after each turn.
+    //
+    // Some wallet client paths return 0 transiently (chain provider not yet
+    // initialized, RPC dust race). Without a defensive retry the UI's status
+    // bar locks at $0.00 USDC for the rest of the session even after the wallet
+    // is provably non-empty. retryFetchBalance does one extra round-trip on a
+    // zero result; genuinely empty wallets still resolve to $0.00 quickly.
     const fetchBalance = async () => {
         try {
-            let bal;
-            if (chain === 'solana') {
-                const { setupAgentSolanaWallet } = await import('@blockrun/llm');
-                const client = await setupAgentSolanaWallet({ silent: true });
-                bal = await client.getBalance();
-            }
-            else {
+            const bal = await retryFetchBalance(async () => {
+                if (chain === 'solana') {
+                    const { setupAgentSolanaWallet } = await import('@blockrun/llm');
+                    const client = await setupAgentSolanaWallet({ silent: true });
+                    return client.getBalance();
+                }
                 const { setupAgentWallet } = await import('@blockrun/llm');
                 const client = setupAgentWallet({ silent: true });
-                bal = await client.getBalance();
-            }
+                return client.getBalance();
+            });
             return `$${bal.toFixed(2)} USDC`;
         }
         catch {

package/dist/commands/telegram.js

@@ -36,7 +36,7 @@ export async function telegramCommand(opts) {
     // Model: --model flag > config default > free default.
     const model = opts.model ||
         config['default-model'] ||
-        'nvidia/glm-4.7';
+        'nvidia/qwen3-coder-480b';
     const workingDir = process.cwd();
     const systemInstructions = assembleInstructions(workingDir, model);
     // Resume the most recent session tagged for THIS owner so a process

package/dist/index.js

@@ -130,6 +130,15 @@ program
     const matches = searchSessions(query, { limit, model: opts.model });
     process.stdout.write(formatSearchResults(matches, query));
 });
+// ─── franklin skills (file-loaded SKILL.md slash commands) ───────────────
+program
+    .command('skills [action] [arg]')
+    .description('Manage Franklin skills — list | which <name>')
+    .option('--json', 'Output the skill list as JSON')
+    .action(async (action, arg, opts) => {
+    const { skillsCommand } = await import('./commands/skills.js');
+    await skillsCommand(action, arg, opts);
+});
 // ─── franklin social (native X bot) ───────────────────────────────────────
 // First-class subcommand. Handles setup / login / run / stats / config
 // subactions. No plugin SDK, no MCP — everything lives in src/social/.

package/dist/learnings/extractor.js

@@ -9,9 +9,9 @@ import { loadLearnings, mergeLearning, saveLearnings, loadSkills, saveSkill } fr
 // Free models for learning extraction — JSON extraction is simple enough.
 // Ordered by reliability: try the best free model first, fall back to others.
 const EXTRACTION_MODELS = [
-    'nvidia/glm-4.7', // Best free model for structured output
-    'nvidia/qwen3-coder-480b', // Strong at JSON tasks
-    'nvidia/llama-4-maverick', // Fallback
+    '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
 ];
 const VALID_CATEGORIES = new Set([
     'language', 'model_preference', 'tool_pattern', 'coding_style',

package/dist/plugin-sdk/workflow.js

@@ -5,7 +5,7 @@
  * Plugins implement Workflow; core orchestrates execution and provides infrastructure.
  */
 export const DEFAULT_MODEL_TIERS = {
-    free: 'nvidia/glm-4.7',
-    cheap: 'nvidia/glm-4.7', // Free by default; opt-in to paid flat-rate via 'zai/glm-5.1'.
+    free: 'nvidia/qwen3-coder-480b',
+    cheap: 'nvidia/qwen3-coder-480b', // Free by default; opt-in to paid flat-rate via 'zai/glm-5.1'.
     premium: 'anthropic/claude-sonnet-4.6',
 };

package/dist/pricing.js

@@ -15,9 +15,9 @@ export const MODEL_PRICING = {
     'nvidia/mistral-small-4-119b': { input: 0, output: 0 },
     'nvidia/llama-4-maverick': { input: 0, output: 0 },
     'nvidia/deepseek-v3.2': { input: 0, output: 0 },
+    // Retired (kept at 0 for legacy session-cost records; gateway no longer serves these).
     'nvidia/gpt-oss-120b': { input: 0, output: 0 },
     'nvidia/gpt-oss-20b': { input: 0, output: 0 },
-    // Retired (kept at 0 for legacy session-cost records; gateway no longer serves these).
     'nvidia/nemotron-ultra-253b': { input: 0, output: 0 },
     'nvidia/devstral-2-123b': { input: 0, output: 0 },
     'nvidia/nemotron-3-super-120b': { input: 0, output: 0 },

package/dist/proxy/fallback.js

@@ -19,7 +19,7 @@ export const DEFAULT_FALLBACK_CONFIG = {
     chain: [
         'deepseek/deepseek-chat', // Direct fallback — cheap & reliable
         'google/gemini-2.5-flash', // Fast & capable
-        'nvidia/glm-4.7', // Free model as ultimate fallback
+        'nvidia/qwen3-coder-480b', // Free model as ultimate fallback
     ],
     retryOn: [429, 500, 502, 503, 504, 529],
     maxRetries: 5,

package/dist/proxy/server.js

@@ -153,18 +153,18 @@ const MODEL_SHORTCUTS = {
     // DeepSeek
     deepseek: 'deepseek/deepseek-chat',
     r1: 'deepseek/deepseek-reasoner',
-    // Free models (gateway free tier — refreshed 2026-04)
-    free: 'nvidia/glm-4.7',
-    glm4: 'nvidia/glm-4.7',
-    'deepseek-free': 'nvidia/deepseek-v3.2',
+    // Free models (agent-tested gateway free tier — refreshed 2026-04)
+    free: 'nvidia/qwen3-coder-480b',
+    glm4: 'nvidia/qwen3-coder-480b',
+    'deepseek-free': 'nvidia/qwen3-coder-480b',
     'qwen-coder': 'nvidia/qwen3-coder-480b',
-    'qwen-think': 'nvidia/qwen3-next-80b-a3b-thinking',
+    'qwen-think': 'nvidia/qwen3-coder-480b',
     maverick: 'nvidia/llama-4-maverick',
-    'gpt-oss': 'nvidia/gpt-oss-120b',
-    'gpt-oss-small': 'nvidia/gpt-oss-20b',
-    'mistral-small': 'nvidia/mistral-small-4-119b',
-    // Retired-gateway-model aliases (map to closest current).
-    nemotron: 'nvidia/glm-4.7',
+    'gpt-oss': 'nvidia/qwen3-coder-480b',
+    'gpt-oss-small': 'nvidia/qwen3-coder-480b',
+    'mistral-small': 'nvidia/llama-4-maverick',
+    // Retired/unreliable gateway-model aliases (map to closest agent-tested current).
+    nemotron: 'nvidia/qwen3-coder-480b',
     devstral: 'nvidia/qwen3-coder-480b',
     // Minimax
     minimax: 'minimax/minimax-m2.7',

package/dist/router/index.js

@@ -66,20 +66,20 @@ const AUTO_TIERS = {
 };
 const ECO_TIERS = {
     SIMPLE: {
-        primary: 'nvidia/glm-4.7',
-        fallback: ['nvidia/gpt-oss-120b', 'nvidia/deepseek-v3.2'],
+        primary: 'nvidia/qwen3-coder-480b',
+        fallback: ['nvidia/llama-4-maverick'],
     },
     MEDIUM: {
         primary: 'google/gemini-2.5-flash-lite',
-        fallback: ['nvidia/glm-4.7', 'nvidia/qwen3-coder-480b'],
+        fallback: ['nvidia/qwen3-coder-480b', 'nvidia/llama-4-maverick'],
     },
     COMPLEX: {
         primary: 'google/gemini-2.5-flash-lite',
-        fallback: ['deepseek/deepseek-chat', 'nvidia/glm-4.7'],
+        fallback: ['deepseek/deepseek-chat', 'nvidia/qwen3-coder-480b'],
     },
     REASONING: {
         primary: 'xai/grok-4-1-fast-reasoning',
-        fallback: ['deepseek/deepseek-reasoner', 'nvidia/qwen3-next-80b-a3b-thinking'],
+        fallback: ['deepseek/deepseek-reasoner', 'nvidia/qwen3-coder-480b'],
     },
 };
 const PREMIUM_TIERS = {
@@ -409,7 +409,7 @@ export function resolveTierToModel(tier, profile = 'auto') {
     // Free profile short-circuits — everything routes to a single free model.
     if (profile === 'free') {
         return {
-            model: 'nvidia/glm-4.7',
+            model: 'nvidia/qwen3-coder-480b',
             tier: 'SIMPLE',
             confidence: 1.0,
             signals: ['free-profile'],
@@ -440,7 +440,7 @@ export function routeRequest(prompt, profile = 'auto') {
     // Free profile — always use free model
     if (profile === 'free') {
         return {
-            model: 'nvidia/glm-4.7',
+            model: 'nvidia/qwen3-coder-480b',
             tier: 'SIMPLE',
             confidence: 1.0,
             signals: ['free-profile'],
@@ -513,7 +513,7 @@ export function getFallbackChain(tier, profile = 'auto') {
             tierConfigs = PREMIUM_TIERS;
             break;
         case 'free':
-            return ['nvidia/glm-4.7'];
+            return ['nvidia/qwen3-coder-480b'];
         default:
             tierConfigs = AUTO_TIERS;
     }