billion-context 0.1.59-pr.308.158 → 0.1.59

package/dist/index.js

@@ -43649,7 +43649,7 @@ function formatRanges(compressible, protectedRanges) {
   if (compressible.length === 0 && protectedRanges.length === 0) {
     return "[No specific ranges detected \u2014 compress any consumed content.]";
   }
-  const refNum4 = (ref) => {
+  const refNum3 = (ref) => {
     const m2 = ref.match(/\d+/);
     return m2 ? parseInt(m2[0], 10) : 0;
   };
@@ -43658,8 +43658,8 @@ function formatRanges(compressible, protectedRanges) {
     entries.push({
       startRef: r.startRef,
       endRef: r.endRef,
-      startNum: refNum4(r.startRef),
-      endNum: refNum4(r.endRef),
+      startNum: refNum3(r.startRef),
+      endNum: refNum3(r.endRef),
       count: r.count,
       tokens: r.tokens,
       toolPct: r.toolPct,
@@ -43676,8 +43676,8 @@ function formatRanges(compressible, protectedRanges) {
     entries.push({
       startRef: r.startRef,
       endRef: r.endRef,
-      startNum: refNum4(r.startRef),
-      endNum: refNum4(r.endRef),
+      startNum: refNum3(r.startRef),
+      endNum: refNum3(r.endRef),
       count: r.count,
       tokens: r.tokens,
       toolPct: 0,
@@ -48016,10 +48016,6 @@ function replayBaseDelayMs() {
   const raw = Number(process.env.BILI_REPLAY_RETRY_BASE_MS);
   return Number.isFinite(raw) && raw >= 0 ? raw : 1500;
 }
-function maxShrinkPerCompress() {
-  const raw = Number(process.env.BILI_MAX_SHRINK_PER_COMPRESS);
-  return Number.isFinite(raw) && raw > 0 && raw <= 1 ? raw : void 0;
-}
 function replayBackoffMs(attempt) {
   return replayBaseDelayMs() * 2 ** (attempt - 1);
 }
@@ -49595,10 +49591,6 @@ function getStore() {
 }
 
 // src/session.ts
-function lastCompressSuffix(info) {
-  if (!info) return "";
-  return ` [after compress: shrink ${Math.round(info.shrinkRatio * 100)}% foldPoint=${info.foldPoint} blocks=${info.blocks} ~${info.tokensCompressed}tok]`;
-}
 var sessions = /* @__PURE__ */ new Map();
 var MAX_SESSIONS = Math.max(1, Number.parseInt(process.env.BILI_MAX_SESSIONS ?? "256", 10) || 256);
 var initialized = false;
@@ -49752,11 +49744,6 @@ function parseCompressInput(input, callId) {
   }
   return ranges;
 }
-var STAGED_COMPRESS_GUIDANCE = "\n\n[Smooth-transition guidance: when you compress, prefer a SMALLER, TAIL-biased range \u2014 compress the most recent large content and keep the stable prefix (the earliest messages) intact. A large single rewrite changes the request shape sharply and can trip provider risk-control; smaller tail-biased folds keep the prefix cache alive and the transition gentle.]";
-function withStagedCompressGuidance(text) {
-  if (maxShrinkPerCompress() === void 0) return text;
-  return text + STAGED_COMPRESS_GUIDANCE;
-}
 
 // src/decompress-shared.ts
 import { mkdirSync as mkdirSync3, unlinkSync, writeFileSync as writeFileSync2 } from "fs";
@@ -49843,17 +49830,15 @@ function normalizeSseLineEndings(buf) {
 }
 
 // src/loop/tag-echo-filter.ts
-var PAIRED = /\x3cacp\s[^<>]{0,256}>([^<>]{0,64})<\/acp>/;
-var LONE_OPEN = /\x3cacp(?:\s[^<>]{0,256})?>/;
-var LONE_CLOSE = /<\/acp(?=[\s>])[^<>]{0,32}>/;
+var PAIRED = /\x3cacp\s[^<>]*>([^<>]{0,64})<\/acp>/;
+var LONE = /<\/?acp(?=[\s>])[^<>]*>/;
 var PARTIAL_TAIL = /(\x3cacp\s[^<>]*|<\/?a?c?p?)$/;
-var TRUNC_OPEN = /\x3cacp\s[^<>]*$/;
+var TRUNCATED_TAG = /\x3cacp\s[^<>]*tokens\s*=\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_OPEN.source, "g"), "").replace(new RegExp(LONE_CLOSE.source, "g"), "").replace(new RegExp(TRUNC_OPEN.source), "");
+  return text.replace(new RegExp(PAIRED.source, "g"), "").replace(new RegExp(LONE.source, "g"), "").replace(new RegExp(TRUNCATED_TAG.source), "");
 }
 function containsRenderTagText(s3) {
   return /\x3c\/?acp(?=[\s>])/.test(s3) || /\\u003c\/?acp(?=[\s>\\])/.test(s3);
@@ -49896,26 +49881,15 @@ function createTagEchoFilter(onDrop) {
         return out;
       }
       const p2 = PAIRED.exec(buf);
-      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;
-      }
+      const l = LONE.exec(buf);
+      let m2;
+      if (p2 && l) m2 = p2.index <= l.index ? p2 : l;
+      else m2 = p2 ?? l;
       if (!m2) {
         const t = PARTIAL_TAIL.exec(buf);
-        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;
-          }
+        if (t && t[0].length <= HOLD_LIMIT) {
+          held = t[0];
+          out += buf.slice(0, buf.length - t[0].length);
         } else {
           out += buf;
         }
@@ -49924,7 +49898,8 @@ function createTagEchoFilter(onDrop) {
       drop(m2[0]);
       out += buf.slice(0, m2.index);
       buf = buf.slice(m2.index + m2[0].length);
-      if (m2 === o && /^\x3cacp\s/.test(m2[0])) {
+      const wasPaired = m2 === p2;
+      if (!wasPaired && /^\x3cacp\s/.test(m2[0])) {
         swallowUntilClose = true;
         swallowed = "";
       }
@@ -49938,19 +49913,14 @@ function createTagEchoFilter(onDrop) {
       return process2(chunk);
     },
     flush() {
-      const rest = swallowed + held;
-      const wasSwallowing = swallowUntilClose;
+      let rest = swallowed + held;
       swallowed = "";
       held = "";
       swallowUntilClose = false;
-      if (wasSwallowing) {
-        if (rest.length > 0) drop(rest);
-        return "";
-      }
-      const t = new RegExp(TRUNC_OPEN.source).exec(rest);
+      const t = new RegExp(TRUNCATED_TAG.source).exec(rest);
       if (t) {
         drop(t[0]);
-        return rest.slice(0, t.index);
+        rest = rest.slice(0, t.index);
       }
       return rest;
     },
@@ -50031,9 +50001,6 @@ ${lines.join("\n\n")}`;
   }
   return `[Unknown proxy tool: ${toolName}]`;
 }
-function refNum2(ref) {
-  return Number(ref.replace(/\D/g, "")) || 0;
-}
 function applyRanges(ranges, ctx) {
   if (ranges.length === 0) {
     ctx.log("[acp-proxy: compress call had no valid ranges; nothing compressed.]");
@@ -50072,17 +50039,8 @@ function applyRanges(ranges, ctx) {
       ctx.log(`[acp-proxy: compress FAILED ${detail} \u2192 0 blocks. ${errs}]`);
       return `[Compression FAILED: ${errs}]`;
     }
-    const preContext = ctx.session.stats.lastInputTokens;
-    const shrinkRatio = preContext > 0 ? r.tokensCompressed / preContext : 0;
-    const foldPoint = [...ranges].sort((a, b2) => refNum2(a.startRef) - refNum2(b2.startRef))[0]?.startRef ?? "unknown";
-    ctx.session.lastCompress = { at: Date.now(), shrinkRatio, foldPoint, blocks: r.blocksCreated, tokensCompressed: r.tokensCompressed };
-    ctx.log(`[acp-compress-obs] shrink ${Math.round(shrinkRatio * 100)}% (~${r.tokensCompressed}/${preContext} tok) foldPoint=${foldPoint} blocks=${r.blocksCreated}`);
     const warn = r.warnings.length > 0 ? ` ${r.warnings.join("; ")}` : "";
-    let msg = `[Compressed ${detail} \u2192 ${r.blocksCreated} block(s), ~${r.tokensCompressed} tokens saved.${warn}]`;
-    const maxShrink = maxShrinkPerCompress();
-    if (maxShrink !== void 0 && shrinkRatio > maxShrink) {
-      msg += ` [Staged-compress: this rewrite shrank context ${Math.round(shrinkRatio * 100)}%, above your ${Math.round(maxShrink * 100)}% per-compress target \u2014 the shape that trips provider risk-control (3007). Next time compress a SMALLER, TAIL-biased range (the most recent large content) and keep the stable prefix intact.]`;
-    }
+    const msg = `[Compressed ${detail} \u2192 ${r.blocksCreated} block(s), ~${r.tokensCompressed} tokens saved.${warn}]`;
     ctx.log(`[acp-proxy: ${msg}]`);
     ctx.session.stats.compressCreditTokens = (ctx.session.stats.compressCreditTokens ?? 0) + r.tokensCompressed;
     ctx.session.stats.lastInputTokens = Math.max(0, ctx.session.stats.lastInputTokens - r.tokensCompressed);
@@ -50140,7 +50098,7 @@ function refMaps(messages, state) {
   });
   return { refToIdx, idxToRef };
 }
-function refNum3(ref) {
+function refNum2(ref) {
   return Number(ref.replace(/\D/g, "")) || 0;
 }
 function estimateCoreMessages(messages) {
@@ -50230,9 +50188,7 @@ TASK: The conversation segment below (messages ${startRef}\u2013${endRef}) must
     },
     void 0,
     deps.signal,
-    (info) => {
-      deps.log("warn", `[preflight] summary attempt ${info.attempt} got HTTP ${info.status}; retrying in ${info.delayMs}ms${lastCompressSuffix(deps.session.lastCompress)}`);
-    }
+    (info) => deps.log("warn", `[preflight] summary attempt ${info.attempt} got HTTP ${info.status}; retrying in ${info.delayMs}ms`)
   );
   try {
     const text = await response.text();
@@ -50277,7 +50233,7 @@ async function preflightCompress(deps, messages) {
     if (currentTokens < limit) break;
     const ranges = viableRanges(turn.nudge?.compressibleRanges ?? []);
     if (ranges.length === 0) break;
-    const range = [...ranges].sort((a, b2) => refNum3(a.startRef) - refNum3(b2.startRef))[0];
+    const range = [...ranges].sort((a, b2) => refNum2(a.startRef) - refNum2(b2.startRef))[0];
     const { refToIdx } = refMaps(messages, deps.session.state);
     const startIdx = refToIdx.get(range.startRef);
     const endIdx = refToIdx.get(range.endRef);
@@ -50907,9 +50863,8 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
         void 0,
         signal,
         (info) => {
-          const lc = lastCompressSuffix(ctx.session.lastCompress);
-          ctx.log(`[acp-proxy: upstream rejected replay (HTTP ${info.status}: ${info.detail.slice(0, 120)}); likely provider risk-control \u2014 retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})${lc}]`);
-          log("warn", `[acp-loop] upstream rejected replay (HTTP ${info.status}); retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})${lc}`);
+          ctx.log(`[acp-proxy: upstream rejected replay (HTTP ${info.status}: ${info.detail.slice(0, 120)}); likely provider risk-control \u2014 retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})]`);
+          log("warn", `[acp-loop] upstream rejected replay (HTTP ${info.status}); retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})`);
         }
       );
       let newBody = adapter.buildRequest(coreMessages, systemPrompt, requestBody);
@@ -52419,9 +52374,8 @@ async function compressLoopResponsesJson(initialResponse, ctx, requestBody, requ
       body: JSON.stringify(requestBody),
       ...ctx.proxyUrl ? { dispatcher: proxyDispatcher(ctx.proxyUrl) } : {}
     }, void 0, void 0, (info) => {
-      const lc = lastCompressSuffix(ctx.session.lastCompress);
-      ctx.log(`[acp-proxy: responses upstream rejected replay (HTTP ${info.status}: ${info.detail.slice(0, 120)}); likely provider risk-control \u2014 retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})${lc}]`);
-      log("warn", `[acp-compress-responses] upstream rejected replay (HTTP ${info.status}); retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})${lc}`);
+      ctx.log(`[acp-proxy: responses upstream rejected replay (HTTP ${info.status}: ${info.detail.slice(0, 120)}); likely provider risk-control \u2014 retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})]`);
+      log("warn", `[acp-compress-responses] upstream rejected replay (HTTP ${info.status}); retrying in ${info.delayMs}ms (attempt ${info.attempt}/${info.maxAttempts})`);
     }).catch((e) => {
       if (e instanceof UpstreamHttpError) {
         const suffix = e.attempts > 1 ? ` after ${e.attempts} attempt(s)` : "";
@@ -55414,7 +55368,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
       try {
         const rendered = renderNudgeText(turn.nudge, prompts);
         if (rendered.text) {
-          rebuiltMessages = [...rebuiltMessages, { role: "user", content: withStagedCompressGuidance(rendered.text) }];
+          rebuiltMessages = [...rebuiltMessages, { role: "user", content: rendered.text }];
         }
       } catch {
       }
@@ -55473,7 +55427,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
       try {
         const rendered = renderNudgeText(turn.nudge, prompts);
         if (rendered.text) {
-          rebuiltMessages = [...rebuiltMessages, { role: "user", content: withStagedCompressGuidance(rendered.text) }];
+          rebuiltMessages = [...rebuiltMessages, { role: "user", content: rendered.text }];
         }
       } catch {
       }
@@ -55556,7 +55510,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
         const rendered = renderNudgeText(turn.nudge, prompts);
         if (rendered.text) {
           const inputItems = typeof rebuiltInput === "string" ? [{ type: "message", role: "user", content: rebuiltInput }] : rebuiltInput;
-          inputItems.push({ type: "message", role: "user", content: withStagedCompressGuidance(rendered.text) });
+          inputItems.push({ type: "message", role: "user", content: rendered.text });
           rebuiltInput = inputItems;
         }
       } catch {