@blockrun/franklin 3.8.12 → 3.8.13

package/dist/agent/context.js

@@ -158,12 +158,16 @@ 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.
-- **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.
+- **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.
-- **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.`;
+
+# Grounding Before Answering
+Your training data is frozen in the past. Live-world questions MUST be answered from tool results, not memory.
+- Any question about a current price, quote, market state, or "should I buy/sell/hold X" → use **TradingMarket** (crypto/FX/commodity are free; stocks cost \$0.001 via the wallet).
+- Any "what happened / why did it change / latest news on X" → use **ExaAnswer** for a cited synthesized answer, or **ExaSearch** + **ExaReadUrls** when you need more depth.
+- If the user names a thing you don't recognize (a company, ticker, project), don't demand clarification — call the research tools and figure it out. You have a wallet to spend on exactly this.
+- If a tool returns an error (rate-limit, 404, insufficient funds), say so plainly and suggest the next action. Don't silently fall back to memory.`;
 }
 function getTokenEfficiencySection() {
     return `# Token Efficiency

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.8.12",
+  "version": "3.8.13",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {