billion-context 0.1.58 → 0.1.59-pr.295.153

package/dist/index.js

@@ -49830,15 +49830,17 @@ function normalizeSseLineEndings(buf) {
 }
 
 // src/loop/tag-echo-filter.ts
-var PAIRED = /\x3cacp\s[^<>]*>([^<>]{0,64})<\/acp>/;
-var LONE = /<\/?acp(?=[\s>])[^<>]*>/;
+var PAIRED = /\x3cacp\s[^<>]{0,256}>([^<>]{0,64})<\/acp>/;
+var LONE_OPEN = /\x3cacp(?:\s[^<>]{0,256})?>/;
+var LONE_CLOSE = /<\/acp(?=[\s>])[^<>]{0,32}>/;
 var PARTIAL_TAIL = /(\x3cacp\s[^<>]*|<\/?a?c?p?)$/;
-var TRUNCATED_TAG = /\x3cacp\s[^<>]*tokens\s*=\s*"[^<>"]*"[^<>]*$/;
+var TRUNC_OPEN = /\x3cacp\s[^<>]*$/;
 var CLOSE_TAG = "</acp";
 var HOLD_LIMIT = 128;
+var TAG_OPEN_CAP = 4096;
 var SWALLOW_CAP = 80;
 function stripAcpTags(text) {
-  return text.replace(new RegExp(PAIRED.source, "g"), "").replace(new RegExp(LONE.source, "g"), "").replace(new RegExp(TRUNCATED_TAG.source), "");
+  return text.replace(new RegExp(PAIRED.source, "g"), "").replace(new RegExp(LONE_OPEN.source, "g"), "").replace(new RegExp(LONE_CLOSE.source, "g"), "").replace(new RegExp(TRUNC_OPEN.source), "");
 }
 function containsRenderTagText(s3) {
   return /\x3c\/?acp(?=[\s>])/.test(s3) || /\\u003c\/?acp(?=[\s>\\])/.test(s3);
@@ -49881,15 +49883,26 @@ function createTagEchoFilter(onDrop) {
         return out;
       }
       const p2 = PAIRED.exec(buf);
-      const l = LONE.exec(buf);
-      let m2;
-      if (p2 && l) m2 = p2.index <= l.index ? p2 : l;
-      else m2 = p2 ?? l;
+      const o = LONE_OPEN.exec(buf);
+      const c = LONE_CLOSE.exec(buf);
+      let m2 = null;
+      for (const cand of [p2, o, c]) {
+        if (cand && (m2 === null || cand.index < m2.index)) m2 = cand;
+      }
       if (!m2) {
         const t = PARTIAL_TAIL.exec(buf);
-        if (t && t[0].length <= HOLD_LIMIT) {
-          held = t[0];
-          out += buf.slice(0, buf.length - t[0].length);
+        if (t) {
+          const definite = /^\x3cacp\s/.test(t[0]);
+          const cap = definite ? TAG_OPEN_CAP : HOLD_LIMIT;
+          if (t[0].length <= cap) {
+            held = t[0];
+            out += buf.slice(0, buf.length - t[0].length);
+          } else if (definite) {
+            drop(t[0]);
+            out += buf.slice(0, buf.length - t[0].length);
+          } else {
+            out += buf;
+          }
         } else {
           out += buf;
         }
@@ -49898,8 +49911,7 @@ function createTagEchoFilter(onDrop) {
       drop(m2[0]);
       out += buf.slice(0, m2.index);
       buf = buf.slice(m2.index + m2[0].length);
-      const wasPaired = m2 === p2;
-      if (!wasPaired && /^\x3cacp\s/.test(m2[0])) {
+      if (m2 === o && /^\x3cacp\s/.test(m2[0])) {
         swallowUntilClose = true;
         swallowed = "";
       }
@@ -49913,14 +49925,19 @@ function createTagEchoFilter(onDrop) {
       return process2(chunk);
     },
     flush() {
-      let rest = swallowed + held;
+      const rest = swallowed + held;
+      const wasSwallowing = swallowUntilClose;
       swallowed = "";
       held = "";
       swallowUntilClose = false;
-      const t = new RegExp(TRUNCATED_TAG.source).exec(rest);
+      if (wasSwallowing) {
+        if (rest.length > 0) drop(rest);
+        return "";
+      }
+      const t = new RegExp(TRUNC_OPEN.source).exec(rest);
       if (t) {
         drop(t[0]);
-        rest = rest.slice(0, t.index);
+        return rest.slice(0, t.index);
       }
       return rest;
     },
@@ -50953,11 +50970,13 @@ function usageTotals(protocol, usage) {
 }
 var CONTEXT_OVERFLOW_PATTERNS = [
   /context_length_exceeded/i,
+  /context_window_exceeded/i,
   /context length exceeded/i,
   /maximum context length/i,
   /max context length/i,
   /maximum context size/i,
   /exceeds the context window/i,
+  /out of room in the model/i,
   /exceeded model token limit/i,
   /prompt is too long/i,
   /prompt_too_long/i,
@@ -55865,7 +55884,16 @@ ${hdrText}
           s3.metadata.learnedContextLimits = learnedMap;
           log2("warn", `[${s3.id}] upstream context overflow \u2014 learned real window ${info.window} for ${reqModel ?? "(unknown model)"} (was ${prev ?? "unset"}); arming emergency shrink`);
         } else {
-          log2("warn", `[${s3.id}] upstream context overflow (window not parseable): ${info.message}`);
+          const payloadEstimate = estimateCoreMessages(prepared.processedMessages);
+          const prev = (reqModel ? learnedMap[reqModel] : void 0) ?? s3.metadata.learnedContextLimit;
+          if (payloadEstimate >= 1e3 && (prev === void 0 || payloadEstimate < prev)) {
+            if (reqModel) learnedMap[reqModel] = payloadEstimate;
+            else s3.metadata.learnedContextLimit = payloadEstimate;
+            s3.metadata.learnedContextLimits = learnedMap;
+            log2("warn", `[${s3.id}] upstream context overflow (window not parseable) \u2014 learned conservative window ${payloadEstimate} for ${reqModel ?? "(unknown model)"} from rejected payload size (was ${prev ?? "unset"}); arming emergency shrink`);
+          } else {
+            log2("warn", `[${s3.id}] upstream context overflow (window not parseable): ${info.message}`);
+          }
         }
         const floor = info.window ?? (reqModel ? learnedMap[reqModel] : void 0) ?? s3.metadata.learnedContextLimit ?? s3.metadata.effectiveContextLimit ?? 0;
         if (floor > 0) s3.stats.lastInputTokens = Math.max(s3.stats.lastInputTokens, floor);