@blockrun/franklin 3.8.11 → 3.8.12

package/dist/agent/commands.js

@@ -203,7 +203,7 @@ const DIRECT_COMMANDS = {
                 `  **Analysis:** /security /lint /optimize /todo /deps /clean /migrate /doc\n` +
                 `  **Session:** /plan /ultraplan /execute /compact /retry /sessions /resume /session-search /context /tasks\n` +
                 `  **Power:** /ultrathink [query] /ultraplan /noplan /moa [query] /dump\n` +
-                `  **Info:** /model /wallet /cost /tokens /learnings /brain /mcp /doctor /version /bug /help\n` +
+                `  **Info:** /model /auto /wallet /cost /tokens /learnings /brain /mcp /doctor /version /bug /help\n` +
                 `  **UI:** /clear /exit\n` +
                 (ultrathinkOn ? `\n  Ultrathink: ON\n` : '')
         });
@@ -616,6 +616,17 @@ export async function handleSlashCommand(input, ctx) {
         emitDone(ctx);
         return { handled: true };
     }
+    // /auto — hard-reset to smart routing regardless of current state.
+    // Shortcut for `/model auto`. Fixes the common "I got stuck on a model
+    // and want Franklin to pick again" scenario without typing the long form.
+    if (input === '/auto') {
+        ctx.config.model = 'blockrun/auto';
+        ctx.config.baseModel = 'blockrun/auto';
+        ctx.config.onModelChange?.('blockrun/auto', 'user');
+        ctx.onEvent({ kind: 'text_delta', text: `Model → **Auto** (smart routing re-enabled)\n` });
+        emitDone(ctx);
+        return { handled: true };
+    }
     // /model — show current model or switch with /model <name>
     if (input === '/model' || input.startsWith('/model ')) {
         if (input === '/model') {
@@ -920,7 +931,7 @@ export async function handleSlashCommand(input, ctx) {
         ...Object.keys(DIRECT_COMMANDS),
         ...Object.keys(REWRITE_COMMANDS),
         ...ARG_COMMANDS.map(c => c.prefix.trim()),
-        '/branch', '/resume', '/model', '/wallet', '/cost', '/help', '/clear', '/retry', '/exit', '/session-search', '/ssearch', '/failures',
+        '/branch', '/resume', '/model', '/auto', '/wallet', '/cost', '/help', '/clear', '/retry', '/exit', '/session-search', '/ssearch', '/failures',
     ];
     const cmd = input.split(/\s/)[0];
     const close = allCommands.filter(c => {

package/dist/agent/context.js

@@ -158,9 +158,12 @@ function getToolPatternsSection() {
 - **Understanding code**: Glob for structure → Read key files → Grep for specific symbols/patterns. Don't read every file in a directory.
 - **Making changes**: Read the file → Edit with targeted replacement → verify the edit worked (Read again or run tests). Never Edit without Reading first.
 - **Running commands**: Use Bash for shell operations that have no dedicated tool. Chain commands with && when sequential. Use separate Bash calls when you need to inspect intermediate output.
-- **Research**: WebSearch for discovery → WebFetch for specific URLs from search results. Don't WebFetch URLs you invented.
+- **Financial / market questions**: If the user asks about ANY ticker, price, stock, crypto, FX, commodity, or "should I sell / hold X" — call **TradingMarket** FIRST to get the live quote before answering. Never answer from memory for price / market data. The agent has a wallet and the tool costs \$0 for crypto/FX/commodity and \$0.001 for stocks — use it. Examples that MUST call TradingMarket: "how is CRCL doing", "BTC 现在多少钱", "should I sell AAPL", "EUR-USD rate", "gold price today".
+- **Current events / "what happened to X" / "why did X drop"**: call **ExaAnswer** for a cited synthesized answer. For breadth use **ExaSearch** then **ExaReadUrls** on the best results. Never guess at recent events from training data — the model cutoff is older than the question.
+- **General web research**: WebSearch for discovery → WebFetch for specific URLs from search results. Don't WebFetch URLs you invented.
 - **Complex tasks**: Use Agent to spawn sub-agents for 2+ independent research or implementation tasks. Don't do sequentially what can be done in parallel.
-- **Multiple independent lookups**: Call all tools in a single response. NEVER make sequential calls when parallel calls would work.`;
+- **Multiple independent lookups**: Call all tools in a single response. NEVER make sequential calls when parallel calls would work.
+- **Don't ask the user for what a tool can answer.** If they say "Circle stock" and you don't know the ticker, call ExaAnswer or TradingMarket rather than demanding they supply a ticker symbol.`;
 }
 function getTokenEfficiencySection() {
     return `# Token Efficiency

package/dist/agent/loop.js

@@ -694,6 +694,16 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 routingSavings = routing.savings;
                 lastRoutedModel = routing.model;
                 lastRoutedCategory = routing.signals[0] || '';
+                // Surface the routing decision so users know which concrete model
+                // just got picked. Without this the status bar reads "auto" and
+                // users have no idea what's actually running — or worse, they
+                // believe they're stuck on the last-seen concrete name.
+                if (loopCount === 1) {
+                    onEvent({
+                        kind: 'text_delta',
+                        text: `*Auto → ${routing.model}*\n\n`,
+                    });
+                }
             }
             // Update token estimation model for more accurate byte-per-token ratio
             setEstimationModel(resolvedModel);
@@ -964,7 +974,7 @@ export async function interactiveSession(config, getUserInput, onEvent, onAbortR
                 + (usage.outputTokens / 1_000_000) * OPUS_PRICING.output;
             sessionSavedVsOpus += Math.max(0, opusCost - costEstimate);
             // ── Max-spend guard ──
-            // Session-level cost ceiling. Cron/daily drivers pass this to bound a
+            // Session-level cost ceiling. Batch/scripted callers pass this to bound a
             // single run ("spend at most $0.50 for today's digest"); interactive
             // users can pass it to feel safe walking away. Hits as soon as accumulated
             // cost crosses the cap — the last call that tipped us over still runs,

package/dist/agent/types.d.ts

@@ -158,7 +158,7 @@ export interface AgentConfig {
     /**
      * Hard cap on total USD spend for this session. When accumulated API cost
      * crosses the cap, the loop stops with `reason: 'budget'`. Zero/negative
-     * values disable the cap. Primary use case: cron/daily drivers that must
+     * values disable the cap. Primary use case: batch/scripted callers that must
      * bound a single run to keep autonomous execution inside a known envelope.
      */
     maxSpendUsd?: number;

package/dist/commands/start.d.ts

@@ -1,3 +1,4 @@
+import type { StreamTurnDone } from '../agent/types.js';
 interface StartOptions {
     model?: string;
     debug?: boolean;
@@ -9,8 +10,9 @@ interface StartOptions {
     continue?: boolean;
     /** Hard USD cap on total session spend. Stops the loop when exceeded. */
     maxSpend?: string | number;
-    /** Run a single prompt non-interactively, then exit. For cron/scripted use. */
+    /** Run a single prompt non-interactively, then exit. For batch/scripted use. */
     prompt?: string;
 }
 export declare function startCommand(options: StartOptions): Promise<void>;
+export declare function oneShotExitCodeForTurnReason(reason: StreamTurnDone['reason']): number;
 export {};

package/dist/commands/start.js

@@ -76,7 +76,7 @@ export async function startCommand(options) {
     }
     const workDir = process.cwd();
     // --prompt batch mode: skip all interactive startup UI/side effects so
-    // stdout stays clean for cron/scripts. Keep the capability surface to the
+    // stdout stays clean for scripts and one-shot callers. Keep the capability surface to the
     // built-ins only — no panel, no MCP autoconnect, no wallet/banner chatter.
     if (options.prompt) {
         if (options.resume === true || options.resume === 'picker') {
@@ -282,7 +282,7 @@ export async function startCommand(options) {
         workingDir: workDir,
         // Non-TTY (piped) input = scripted mode → trust all tools automatically.
         // Interactive TTY = default mode (prompts for Bash/Write/Edit).
-        // --prompt is also scripted; the cron driver never sees a TTY.
+        // --prompt is also scripted; batch callers never see a TTY.
         permissionMode: (options.trust || options.prompt || !process.stdin.isTTY) ? 'trust' : 'default',
         debug: options.debug,
         resumeSessionId,
@@ -309,8 +309,12 @@ export async function startCommand(options) {
     }
 }
 // ─── One-shot mode (franklin --prompt "...") ──────────────────────────────
-// Used by cron drivers. Non-interactive, prints text deltas to stdout as
-// they stream, honors --max-spend, exits 0 on completion / 1 on error.
+// Used by batch/scripted callers. Non-interactive, prints text deltas to
+// stdout as they stream, honors --max-spend, exits non-zero for any
+// non-completed terminal state.
+export function oneShotExitCodeForTurnReason(reason) {
+    return reason === 'completed' ? 0 : 1;
+}
 async function runOneShot(agentConfig, prompt) {
     let delivered = false;
     let exitCode = 0;
@@ -325,8 +329,7 @@ async function runOneShot(agentConfig, prompt) {
             process.stdout.write(event.text);
         }
         else if (event.kind === 'turn_done') {
-            if (event.reason === 'error')
-                exitCode = 1;
+            exitCode = oneShotExitCodeForTurnReason(event.reason);
             process.stdout.write('\n');
         }
     });

package/dist/index.js

@@ -44,7 +44,7 @@ program
     .option('-r, --resume [sessionId]', 'Resume a session by ID (or show picker if omitted)')
     .option('-c, --continue', 'Continue the most recent session in this directory')
     .option('--max-spend <usd>', 'Hard USD cap on total session API spend — session stops when exceeded')
-    .option('-p, --prompt <text>', 'Run a single prompt non-interactively (for cron/scripted use)')
+    .option('-p, --prompt <text>', 'Run a single prompt non-interactively (for batch/scripted use)')
     .action((options) => startCommand({ ...options, version }));
 program
     .command('resume [sessionId]')

package/dist/tools/tool-categories.d.ts

@@ -1,15 +1,20 @@
 /**
  * Tool visibility categories.
  *
- * Franklin ships with ~27 capabilities. Exposing all of them to the model on
- * every turn makes the tool inventory large enough that weak models start
- * hallucinating tool names or emitting role-play "[TOOLCALL]" fragments.
- * The fix: keep a minimal always-on core (file ops, shell, ask) and gate the
- * rest behind an `ActivateTool` meta-tool that the agent pulls on demand —
- * the same per-session visibility pattern that OpenBB's MCP server uses.
+ * Franklin ships with ~27 capabilities. Exposing all of them on every turn
+ * makes the tool inventory large enough that weak models start hallucinating
+ * tool names or emitting role-play "[TOOLCALL]" fragments. The compromise:
+ * keep the hero surface always-on (file/shell/search PLUS the trading and
+ * research tools that define Franklin's category), and gate the long tail
+ * (webhook, imagegen, videogen, musicgen, memory, etc.) behind an
+ * `ActivateTool` meta-tool the agent pulls on demand.
  *
- * `CORE_TOOL_NAMES` is the per-session initial active set. Everything else
- * becomes visible only after the agent calls ActivateTool with its name.
+ * History: earlier releases kept only file/shell/search in core, which made
+ * mid-tier models answer stock / market questions from 2022 training data
+ * instead of calling TradingMarket. That's anti-positioning for an agent
+ * whose whole brand is "spends USDC for real market data." Hero tools now
+ * live in the always-on set so the default experience shows the wallet
+ * actually at work.
  */
 export declare const CORE_TOOL_NAMES: ReadonlySet<string>;
 /** True if this tool is always available without activation. */

package/dist/tools/tool-categories.js

@@ -1,15 +1,20 @@
 /**
  * Tool visibility categories.
  *
- * Franklin ships with ~27 capabilities. Exposing all of them to the model on
- * every turn makes the tool inventory large enough that weak models start
- * hallucinating tool names or emitting role-play "[TOOLCALL]" fragments.
- * The fix: keep a minimal always-on core (file ops, shell, ask) and gate the
- * rest behind an `ActivateTool` meta-tool that the agent pulls on demand —
- * the same per-session visibility pattern that OpenBB's MCP server uses.
+ * Franklin ships with ~27 capabilities. Exposing all of them on every turn
+ * makes the tool inventory large enough that weak models start hallucinating
+ * tool names or emitting role-play "[TOOLCALL]" fragments. The compromise:
+ * keep the hero surface always-on (file/shell/search PLUS the trading and
+ * research tools that define Franklin's category), and gate the long tail
+ * (webhook, imagegen, videogen, musicgen, memory, etc.) behind an
+ * `ActivateTool` meta-tool the agent pulls on demand.
  *
- * `CORE_TOOL_NAMES` is the per-session initial active set. Everything else
- * becomes visible only after the agent calls ActivateTool with its name.
+ * History: earlier releases kept only file/shell/search in core, which made
+ * mid-tier models answer stock / market questions from 2022 training data
+ * instead of calling TradingMarket. That's anti-positioning for an agent
+ * whose whole brand is "spends USDC for real market data." Hero tools now
+ * live in the always-on set so the default experience shows the wallet
+ * actually at work.
  */
 export const CORE_TOOL_NAMES = new Set([
     // File operations — nothing else works without these.
@@ -27,8 +32,24 @@ export const CORE_TOOL_NAMES = new Set([
     // so keeping this in the core doesn't leak the full inventory.
     'Task',
     // The meta-tool itself — must always be callable so the agent can
-    // discover and activate the rest.
+    // discover and activate anything not in this core set.
     'ActivateTool',
+    // ── Hero surface: Franklin's reason to exist ────────────────────────
+    // Trading market data — crypto, FX, commodity, stocks (via x402).
+    // "Is NVDA up?" / "Should I sell CRCL?" must never fall back to
+    // training-data guessing.
+    'TradingMarket',
+    'TradingSignal',
+    // Research — synthesized answers with real citations, semantic web
+    // search, and clean URL fetching. Any factual current-events question
+    // ("why did X drop?") should route here rather than the model's prior.
+    'ExaAnswer',
+    'ExaSearch',
+    'ExaReadUrls',
+    // Plain web fetch — specific URL → readable text. Cheap and obvious
+    // enough that every model tends to pick it correctly.
+    'WebFetch',
+    'WebSearch',
 ]);
 /** True if this tool is always available without activation. */
 export function isCoreTool(name) {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.8.11",
+  "version": "3.8.12",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {