openzoo 0.48.33 → 0.48.34

package/lib/proxy.js

@@ -339,6 +339,28 @@ async function spillTranscript(body, log) {
   }
   if (cut <= firstSpillable) return null;         // nothing safely severable
 
+  // TRIM THE TAIL BY BYTES, NOT MESSAGE COUNT.
+  //
+  // MEASURED on the live session: 9 kept turns of Claude Code tool output made
+  // promptTokens swamp the counterfactual and the call scored 1.00x, while the
+  // SAME bound context with a one-line ask scored 8.53x. Nine messages is a
+  // trivial number and an enormous payload — a single Read or grep result is
+  // tens of KB — so counting messages measures the wrong thing entirely.
+  //
+  // Walk backwards from the newest and stop at a byte budget. The newest turns
+  // are the ones the model actually needs verbatim; everything older is already
+  // in the bound corpus and comes back through recall.
+  let tailStart = cut;
+  {
+    const budget = Number(process.env.OPENZOO_TAIL_MAX_CHARS || 24000);
+    let used = 0;
+    for (let i = msgs.length - 1; i >= cut; i--) {
+      used += msgText(msgs[i]).length;
+      if (used > budget && severable(i)) { tailStart = i; break; }
+    }
+  }
+  if (tailStart > cut) cut = tailStart;
+
   const head = msgs.slice(0, firstSpillable);     // system block, always kept
   const corpus = msgs.slice(firstSpillable, cut).map(msgText).filter(Boolean).join('\n\n');
   if (corpus.length <= BIND_MIN_CHARS) return null;

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "openzoo",
-  "version": "0.48.33",
+  "version": "0.48.34",
   "description": "Local x402-paying proxy + MCP server for openzoo.fun — point any OpenAI-compatible harness (Cursor, Claude Code, aider, SDKs) at localhost and it pays per call from a local burner wallet. Solana and Base rails live; Robinhood experimental.",
   "license": "MIT",
   "type": "module",