@blockrun/franklin 3.7.5 → 3.7.6

package/dist/agent/planner.js

@@ -17,25 +17,29 @@ const MULTI_STEP_PATTERN = /first.*then|step\s+\d|\d+\.\s|and\s+then|after\s+tha
  * the overhead of an extra planning call.
  */
 export function shouldPlan(tier, profile, userText, ultrathink, planDisabled) {
-    // Gate 1: only COMPLEX or REASONING tiers benefit from planning
-    if (tier !== 'COMPLEX' && tier !== 'REASONING')
-        return false;
-    // Gate 2: only auto or premium profiles (eco/free already cost-optimized)
-    if (profile !== 'auto' && profile !== 'premium')
-        return false;
-    // Gate 3: skip short queries — planning overhead not worth it
-    if (userText.length < 80)
+    // User disabled planning for this session
+    if (planDisabled)
         return false;
-    // Gate 4: ultrathink already provides deep reasoning
+    // Ultrathink already provides deep reasoning
     if (ultrathink)
         return false;
-    // Gate 5: user disabled planning for this session
-    if (planDisabled)
+    // Only auto or premium profiles (eco/free are cost-constrained)
+    if (profile !== 'auto' && profile !== 'premium')
         return false;
-    // Gate 6: must have agentic or multi-step signals
-    const hasAgenticKeyword = AGENTIC_KEYWORDS.test(userText);
-    const hasMultiStep = MULTI_STEP_PATTERN.test(userText);
-    return hasAgenticKeyword || hasMultiStep;
+    // Explicit multi-step language always plans, regardless of tier / length
+    // ("first ... then ...", "step 1 ... step 2 ...", numbered lists, etc.)
+    if (MULTI_STEP_PATTERN.test(userText))
+        return true;
+    // Planning is high-ROI on COMPLEX / REASONING tiers for agentic verbs,
+    // even when the prompt is short ("refactor the wallet module", "migrate to TS")
+    if (tier === 'COMPLEX' || tier === 'REASONING') {
+        return AGENTIC_KEYWORDS.test(userText) || userText.length >= 60;
+    }
+    // On MEDIUM tier: plan only if long AND agentic
+    if (tier === 'MEDIUM' && userText.length >= 120 && AGENTIC_KEYWORDS.test(userText)) {
+        return true;
+    }
+    return false;
 }
 // ─── Planning Prompt ─────────────────────────────────────────────────────
 /**

package/dist/agent/tool-guard.d.ts

@@ -16,11 +16,14 @@ export declare class SessionToolGuard {
     private toolErrorCounts;
     private recentGreps;
     private recentGlobs;
+    private recentBash;
     startTurn(): void;
     beforeExecute(invocation: CapabilityInvocation, scope: ExecutionScope): Promise<CapabilityResult | null>;
+    private beforeBash;
     private beforeGrep;
     private beforeGlob;
     afterExecute(invocation: CapabilityInvocation, result: CapabilityResult): void;
+    private afterBash;
     private afterGrep;
     private afterGlob;
     cancelInvocation(invocationId: string): void;

package/dist/agent/tool-guard.js

@@ -94,6 +94,7 @@ export class SessionToolGuard {
     // five times in a row when they're confused. Tell them once that it already failed.
     recentGreps = new Map();
     recentGlobs = new Map();
+    recentBash = new Map();
     startTurn() {
         this.turn++;
         this.webSearchesThisTurn = 0;
@@ -123,10 +124,33 @@ export class SessionToolGuard {
                 return this.beforeGrep(invocation);
             case 'Glob':
                 return this.beforeGlob(invocation);
+            case 'Bash':
+                return this.beforeBash(invocation);
             default:
                 return null;
         }
     }
+    beforeBash(invocation) {
+        const cmd = String(invocation.input.command ?? '').trim();
+        if (!cmd)
+            return null;
+        // Only dedup deterministic read-only commands. Skip anything writing/network/long-running.
+        const writeKeywords = /\b(rm|mv|cp|mkdir|touch|chmod|chown|write|install|build|publish|push|pull|curl|wget|fetch|npm|pnpm|yarn|pip|cargo|go\s+(build|run|test)|docker|kubectl|tar|zip|unzip|tee|>\s|>>\s)\b/;
+        if (writeKeywords.test(cmd))
+            return null;
+        const key = cmd;
+        const cached = this.recentBash.get(key);
+        if (cached) {
+            const lead = cached.isError
+                ? 'That exact Bash command was already run this session and FAILED:'
+                : 'That exact Bash command was already run this session and returned:';
+            return {
+                output: `${lead}\n${cached.preview}\n\n` +
+                    'Do not re-run the same command. If the output was insufficient, run a different command or use a dedicated tool (Read for files, Grep/Glob for searching).',
+            };
+        }
+        return null;
+    }
     beforeGrep(invocation) {
         const pattern = String(invocation.input.pattern ?? '').trim();
         const path = String(invocation.input.path ?? '').trim();
@@ -181,10 +205,26 @@ export class SessionToolGuard {
             case 'Glob':
                 this.afterGlob(invocation, result);
                 break;
+            case 'Bash':
+                this.afterBash(invocation, result);
+                break;
             default:
                 break;
         }
     }
+    afterBash(invocation, result) {
+        const cmd = String(invocation.input.command ?? '').trim();
+        if (!cmd)
+            return;
+        const writeKeywords = /\b(rm|mv|cp|mkdir|touch|chmod|chown|write|install|build|publish|push|pull|curl|wget|fetch|npm|pnpm|yarn|pip|cargo|go\s+(build|run|test)|docker|kubectl|tar|zip|unzip|tee|>\s|>>\s)\b/;
+        if (writeKeywords.test(cmd))
+            return;
+        const output = String(result.output ?? '');
+        const preview = output.length > MAX_PREVIEW_CHARS
+            ? output.slice(0, MAX_PREVIEW_CHARS) + '…'
+            : output;
+        this.recentBash.set(cmd, { preview, turn: this.turn, isError: !!result.isError });
+    }
     afterGrep(invocation, result) {
         const pattern = String(invocation.input.pattern ?? '').trim();
         const path = String(invocation.input.path ?? '').trim();

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@blockrun/franklin",
-  "version": "3.7.5",
+  "version": "3.7.6",
   "description": "Franklin — The AI agent with a wallet. Spends USDC autonomously to get real work done. Pay per action, no subscriptions.",
   "type": "module",
   "exports": {