feat(serve): Franklin desktop agent server + backend (cloud sync, wallet holdings/swaps, cost-saver)

package.json

@@ -69,6 +69,7 @@
     "@blockrun/llm": "^2.0.0",
     "@colbymchenry/codegraph": "^0.9.7",
     "@modelcontextprotocol/sdk": "^1.29.0",
+    "@slack/bolt": "^4.7.3",
     "@solana/spl-token": "^0.4.14",
     "@solana/web3.js": "^1.98.4",
     "@types/react": "^19.2.14",

src/agent/llm.ts

@@ -520,7 +520,7 @@ export class ModelClient {
    * default model.
    */
   private resolveVirtualModel(model: string): string {
-    if (!model.startsWith('blockrun/')) return model;
+    if (!model || !model.startsWith('blockrun/')) return model;
 
     try {
       const profile = parseRoutingProfile(model);
@@ -563,6 +563,14 @@ export class ModelClient {
     // Reset the per-call charge tracker. signBasePayment / signSolanaPayment
     // will set it when the gateway demands a 402 settlement.
     this.lastPaidUsd = 0;
+    // Guard: a missing/non-string model (e.g. a flaky-gateway fallback that
+    // produced undefined) must not hard-crash with a cryptic
+    // "reading 'startsWith'". Normalize to the routing profile, which resolves
+    // to a concrete model below.
+    if (!request.model || typeof request.model !== 'string') {
+      console.error('[franklin] request.model was missing — defaulting to blockrun/auto');
+      request = { ...request, model: 'blockrun/auto' };
+    }
     // Resolve virtual models before any API call
     const resolvedModel = this.resolveVirtualModel(request.model);
     if (resolvedModel !== request.model) {

src/agent/loop.ts

@@ -4,7 +4,7 @@
  */
 
 import { ModelClient } from './llm.js';
-import { autoCompactIfNeeded, forceCompact, microCompact } from './compact.js';
+import { autoCompactIfNeeded, forceCompact, microCompact, projectCompactionSavings } 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';
@@ -1153,13 +1153,22 @@ export async function interactiveSession(
       // compacting (the compact itself runs on a cheaper model
       // and costs <$0.05).
       const TURN_COST_CAP_FOR_EARLY_COMPACT = 1.00;
+      // ROI gate: forceCompact (used below) has no savings check of its own, so
+      // without this it fires even on a tiny history and reports "saved 1%" —
+      // a wasted summarizer round-trip. Only compact when the projected savings
+      // clear the floor (≥20%), which a small history can never do.
+      // The ROI gate applies ONLY to the call-count trigger: the $1.00 cost cap
+      // is an emergency brake (see the 2026-05-11 note above) and must fire
+      // even when projected savings are low — gating it would reintroduce the
+      // $9.45 runaway it was added to stop.
+      const bloatTriggered =
+        (turnToolCalls > 15 && turnCostUsd > 0.03 && projectCompactionSavings(history).worthIt) ||
+        turnCostUsd > TURN_COST_CAP_FOR_EARLY_COMPACT;
       if (
+        config.costSaver !== false &&
         !bloatCompactedThisTurn &&
         compactFailures < 3 &&
-        (
-          (turnToolCalls > 15 && turnCostUsd > 0.03) ||
-          turnCostUsd > TURN_COST_CAP_FOR_EARLY_COMPACT
-        )
+        bloatTriggered
       ) {
         try {
           const beforeTokens = estimateHistoryTokens(history);

src/agent/streaming-executor.ts

@@ -329,8 +329,33 @@ export class StreamingExecutor {
       case 'WebFetch':
       case 'WebSearch':
         return ((input.url ?? input.query) as string) || undefined;
-      default:
-        return undefined;
+      default: {
+        // Generic fallback so EVERY tool shows what it's doing. For enum/router
+        // tools (e.g. Surf*) the `endpoint` is the real action — show it, paired
+        // with the most relevant param, e.g. "market/etf · BTC". Otherwise pick
+        // the single most meaningful argument.
+        const PARAM_KEYS = [
+          'query', 'q', 'search', 'prompt', 'question', 'text',
+          'symbol', 'pair', 'metric', 'indicator', 'ticker', 'coin', 'asset', 'market',
+          'protocol', 'handle', 'chain', 'address', 'addresses', 'hash', 'conditionId',
+          'url', 'id', 'slug', 'name', 'path', 'pattern', 'to', 'number',
+        ];
+        const firstParam = (): string => {
+          for (const k of PARAM_KEYS) {
+            const v = input[k];
+            if (typeof v === 'string' && v.trim()) return v.trim();
+          }
+          return '';
+        };
+        // The "action" field (endpoint / action) is the real verb — show it even
+        // when there's no param (e.g. PredictionMarket `leaderboard`).
+        const action =
+          (typeof input.endpoint === 'string' && input.endpoint.trim()) ||
+          (typeof input.action === 'string' && input.action.trim()) || '';
+        const combined = [action, firstParam()].filter(Boolean).join(' · ');
+        if (!combined) return undefined;
+        return combined.length > 80 ? combined.slice(0, 80) + '…' : combined;
+      }
     }
   }
 }

src/agent/types.ts

@@ -217,4 +217,8 @@ export interface AgentConfig {
   maxSpendUsd?: number;
   /** Show user-visible harness prefetch status lines (interactive UX only). */
   showPrefetchStatus?: boolean;
+  /** Mid-turn "research-bloat" compaction — summarizes history when a turn
+   *  racks up many tool calls + spend, to cut input-replay cost. Default on;
+   *  set false to disable (the desktop exposes this as a toggle). */
+  costSaver?: boolean;
 }