billion-context 0.1.32 → 0.1.33

package/dist/index.js

@@ -43930,14 +43930,49 @@ function truncateLargeToolOutputs(messages, tokenCount, config, countTokens, opt
   return { messages: updated, truncatedCount, savedTokens };
 }
 var KEEP_LAST_ORPHANED = 0;
+function rangeKey(startRef, endRef) {
+  return `${startRef}::${endRef}`;
+}
+function rewriteCompressText(text, liveKeys) {
+  let parsed;
+  try {
+    parsed = JSON.parse(text ?? "");
+  } catch {
+    return null;
+  }
+  if (!parsed || typeof parsed !== "object") return null;
+  const obj = parsed;
+  const content = obj.content;
+  if (!Array.isArray(content) || content.length === 0) return null;
+  const kept = content.filter((entry) => {
+    if (!entry || typeof entry !== "object") return false;
+    const s3 = typeof entry.startId === "string" ? entry.startId : typeof entry.messageId === "string" ? entry.messageId : "";
+    const e = typeof entry.endId === "string" ? entry.endId : typeof entry.messageId === "string" ? entry.messageId : "";
+    return liveKeys.has(rangeKey(s3, e));
+  });
+  if (kept.length === content.length || kept.length === 0) return null;
+  return JSON.stringify({ ...obj, content: kept });
+}
 function hideConsumedCompressCalls(state, messages) {
-  const activeCallIds = /* @__PURE__ */ new Set();
   const allBlockCallIds = /* @__PURE__ */ new Set();
+  const activeCallIds = /* @__PURE__ */ new Set();
+  const liveRangeKeysByCallId = /* @__PURE__ */ new Map();
+  const legacyLiveByCallId = /* @__PURE__ */ new Set();
   for (const block of state.blocks) {
-    if (block.compressCallId) {
-      allBlockCallIds.add(block.compressCallId);
-      if (block.active) activeCallIds.add(block.compressCallId);
+    if (!block.compressCallId) continue;
+    allBlockCallIds.add(block.compressCallId);
+    if (!block.active) continue;
+    activeCallIds.add(block.compressCallId);
+    if (block.startRef === void 0 || block.endRef === void 0) {
+      legacyLiveByCallId.add(block.compressCallId);
+      continue;
+    }
+    let keys = liveRangeKeysByCallId.get(block.compressCallId);
+    if (!keys) {
+      keys = /* @__PURE__ */ new Set();
+      liveRangeKeysByCallId.set(block.compressCallId, keys);
     }
+    keys.add(rangeKey(block.startRef, block.endRef));
   }
   const lastOrphanedCallIds = [];
   for (let i = messages.length - 1; i >= 0 && lastOrphanedCallIds.length < KEEP_LAST_ORPHANED; i--) {
@@ -43966,6 +44001,16 @@ function hideConsumedCompressCalls(state, messages) {
       hidden++;
       continue;
     }
+    if (message.toolName === "compress" && message.contentType === "tool-call" && message.toolCallId && keepCallIds.has(message.toolCallId)) {
+      const liveKeys = liveRangeKeysByCallId.get(message.toolCallId);
+      if (liveKeys && liveKeys.size > 0 && !legacyLiveByCallId.has(message.toolCallId)) {
+        const rewritten = rewriteCompressText(message.text, liveKeys);
+        if (rewritten !== null) {
+          result.push({ ...message, text: rewritten });
+          continue;
+        }
+      }
+    }
     result.push(message);
   }
   return { messages: result, hidden };
@@ -44199,8 +44244,8 @@ function refNum(ref) {
   const n = parseInt(ref.slice(1), 10);
   return Number.isNaN(n) ? -1 : n;
 }
-function estimateMessageTokens(message) {
-  return Math.ceil((message.text ?? "").length / 4);
+function estimateTextTokens(text) {
+  return Math.ceil(text.length / 4);
 }
 function isToolMessage(message) {
   return message.contentType === "tool-call" || message.contentType === "tool-result";
@@ -44212,7 +44257,7 @@ function isSyntheticOrPruned(message, state) {
   }
   return false;
 }
-function computeProtectedRefs(messages, state, config) {
+function computeProtectedRefs(messages, state, config, countTokens = estimateTextTokens) {
   const preserveN = config.preserveRecentMessages;
   const preserveTokens = config.preserveRecentTokens;
   const result = /* @__PURE__ */ new Set();
@@ -44222,7 +44267,7 @@ function computeProtectedRefs(messages, state, config) {
     if (isNeverPreserveRecent(msg2)) continue;
     const ref = state.messageRefs.byRaw[msg2.id];
     if (!ref || ref === "BLOCKED") continue;
-    visible.push({ ref, tokens: estimateMessageTokens(msg2) });
+    visible.push({ ref, tokens: countTokens(msg2.text ?? "") });
   }
   if (preserveN > 0) {
     for (const m2 of visible.slice(-preserveN)) {
@@ -44247,7 +44292,7 @@ function computeProtectedRefs(messages, state, config) {
   }
   return result;
 }
-function buildCompressibleRanges(messages, state, config, protectedZoneRefs) {
+function buildCompressibleRanges(messages, state, config, protectedZoneRefs, countTokens = estimateTextTokens) {
   const compressibleMsgs = [];
   const protectedMsgs = [];
   const protectedCallIds = collectProtectedToolCallIds(messages, config);
@@ -44260,7 +44305,7 @@ function buildCompressibleRanges(messages, state, config, protectedZoneRefs) {
       protectedMsgs.push({
         ref,
         refNum: rn2,
-        tokens: estimateMessageTokens(msg2),
+        tokens: countTokens(msg2.text ?? ""),
         tools: msg2.toolName ? [msg2.toolName] : []
       });
       continue;
@@ -44271,7 +44316,7 @@ function buildCompressibleRanges(messages, state, config, protectedZoneRefs) {
     compressibleMsgs.push({
       ref,
       refNum: rn2,
-      tokens: estimateMessageTokens(msg2),
+      tokens: countTokens(msg2.text ?? ""),
       isTool: isToolMessage(msg2),
       isUser: msg2.role === "user"
     });
@@ -44358,7 +44403,7 @@ function createCore(ports = {}) {
     let tokensCompressed = 0;
     const errors = [];
     const warnings = [];
-    const protectedMessageIds = input.protectedMessageIds ?? computeProtectedRefs(input.messages, input.state, input.config);
+    const protectedMessageIds = input.protectedMessageIds ?? computeProtectedRefs(input.messages, input.state, input.config, countTokens);
     const preExistingCoverage = collectCoverage(state);
     const rangeIndexSets = [];
     for (const spec of input.ranges) {
@@ -44383,29 +44428,25 @@ function createCore(ports = {}) {
       const bMin = b2.indices.length > 0 ? Math.min(...b2.indices) : Infinity;
       return aMin - bMin;
     });
-    for (let i = 1; i < sortedRanges.length; i++) {
-      const prev = sortedRanges[i - 1];
-      const curr = sortedRanges[i];
-      const prevMax = prev.indices.length > 0 ? Math.max(...prev.indices) : -1;
-      const currMin = curr.indices.length > 0 ? Math.min(...curr.indices) : -1;
-      if (prevMax >= currMin && prevMax >= 0) {
-        return {
-          state: input.state,
-          result: {
-            blocksCreated: 0,
-            tokensCompressed: 0,
-            errors: [
-              `content: range (${prev.spec.startRef}..${prev.spec.endRef}) overlaps (${curr.spec.startRef}..${curr.spec.endRef}). Overlapping ranges cannot be compressed in the same batch.`
-            ],
-            warnings: []
-          }
-        };
+    const skipSpecs = /* @__PURE__ */ new Set();
+    let acceptedMaxIndex = -1;
+    for (const entry of sortedRanges) {
+      const entryMax = entry.indices.length > 0 ? Math.max(...entry.indices) : -1;
+      const entryMin = entry.indices.length > 0 ? Math.min(...entry.indices) : -1;
+      if (entryMin >= 0 && entryMin <= acceptedMaxIndex) {
+        skipSpecs.add(entry.spec);
+        warnings.push(
+          `Skipped range (${entry.spec.startRef}..${entry.spec.endRef}) \u2014 overlaps an earlier range in the batch; the earlier range takes precedence. Keep ranges disjoint.`
+        );
+        continue;
       }
+      if (entryMax > acceptedMaxIndex) acceptedMaxIndex = entryMax;
     }
     if (input.config.compress.minCompressRange > 0 && input.ranges.length > 0) {
       let totalRangeChars = 0;
       let hasBlockBoundaryRange = false;
       for (const spec of input.ranges) {
+        if (skipSpecs.has(spec)) continue;
         let resolved;
         try {
           resolved = resolveBoundaries({
@@ -44441,6 +44482,7 @@ function createCore(ports = {}) {
       }
     }
     for (const spec of input.ranges) {
+      if (skipSpecs.has(spec)) continue;
       try {
         const outcome = applySingleRange({
           spec,
@@ -44581,13 +44623,15 @@ var recommendNode = {
     const protectedRefs = computeProtectedRefs(
       io2.messages,
       io2.state,
-      ctx.config
+      ctx.config,
+      ctx.countTokens
     );
     const contextRanges = buildCompressibleRanges(
       io2.messages,
       io2.state,
       ctx.config,
-      protectedRefs
+      protectedRefs,
+      ctx.countTokens
     );
     const nothingToCompress = contextRanges.compressible.length === 0;
     const recommendation = {
@@ -44765,7 +44809,9 @@ function applySingleRange(input) {
     survivedCount: 0,
     generation: "young",
     active: true,
-    compressCallId: input.spec.compressCallId
+    compressCallId: input.spec.compressCallId,
+    startRef: input.spec.startRef,
+    endRef: input.spec.endRef
   };
   input.state.blocks.push(block);
   for (const consumedId of consumedBlockIds) {
@@ -45256,7 +45302,7 @@ function renderNudgeText(decision) {
         breakdownStr,
         "",
         `[TIER ${decision.tier} ${isT2 ? "DISTILLATION" : "CONDENSATION"} TRIGGER]`,
-        isT2 ? `Your tier-1 compression summaries have accumulated. Distill them into a single denser tier-2 summary. Use block IDs as boundaries.` : `Your tier-2 compression summaries have accumulated. Condense them further into a tier-3 ultra-condensed summary. Use block IDs as boundaries.`,
+        isT2 ? `Your tier-1 compression summaries have accumulated. Distill them into a single denser tier-2 summary. Use block IDs as boundaries (startId and endId as bN). Any raw (uncompressed) messages sitting between the boundary blocks are absorbed into the tier-2 block as well \u2014 apply HOW TO COMPRESS to those raw messages and the TIER 2 distillation rules to the existing summaries, so the whole span is covered and nothing is lost.` : `Your tier-2 compression summaries have accumulated. Condense them further into a tier-3 ultra-condensed summary. Use block IDs as boundaries (startId and endId as bN). Any raw (uncompressed) messages sitting between the boundary blocks are absorbed into the tier-3 block as well \u2014 apply HOW TO COMPRESS to those raw messages and the TIER 3 condensation rules to the existing summaries, so the whole span is covered and nothing is lost.`,
         blockList,
         `Example: compress({ content: [{ startId: "${startId}", endId: "${endId}", summary: "..." }] })`,
         "",
@@ -52018,11 +52064,24 @@ async function handle(req, res, opts, core, config, log2) {
     return;
   }
   let bodyBuffer;
+  let urlPath;
+  let responsesCompact;
+  let route;
+  let upstreamOrigin;
+  let protocol;
   try {
     bodyBuffer = await readBody(req);
-    const decoded = await decodeRequestBody(headerValue(req, "content-encoding"), bodyBuffer, MAX_REQUEST_BYTES);
-    bodyBuffer = decoded.body;
-    if (decoded.decoded) delete req.headers["content-encoding"];
+    const url = req.url ?? "";
+    urlPath = url.split("?", 2)[0];
+    responsesCompact = urlPath.endsWith("/responses/compact");
+    route = resolveUpstream(opts, req.url ?? "", req);
+    upstreamOrigin = route ? route.upstream : opts.upstream;
+    protocol = route?.explicitProtocol ?? (req.method === "POST" && bodyBuffer.length > 0 ? urlPath.endsWith("/chat/completions") ? "openai" : urlPath.endsWith("/v1/messages") || urlPath.endsWith("/messages") ? "anthropic" : urlPath.endsWith("/responses") || responsesCompact ? "responses" : null : null);
+    if (protocol !== null && bodyBuffer.length > 0) {
+      const decoded = await decodeRequestBody(headerValue(req, "content-encoding"), bodyBuffer, MAX_REQUEST_BYTES);
+      bodyBuffer = decoded.body;
+      if (decoded.decoded) delete req.headers["content-encoding"];
+    }
   } catch (err2) {
     if (err2 instanceof BodyTooLargeError) {
       log2("warn", `413: request body exceeds ${err2.limit} bytes`);
@@ -52035,13 +52094,7 @@ async function handle(req, res, opts, core, config, log2) {
     res.end(JSON.stringify({ error: { type: "invalid_request", message: String(err2) } }));
     return;
   }
-  const url = req.url ?? "";
-  const urlPath = url.split("?", 2)[0];
-  const responsesCompact = urlPath.endsWith("/responses/compact");
   const countTokens = isCountTokensRequest(req.method ?? "GET", urlPath, bodyBuffer.length > 0);
-  const route = resolveUpstream(opts, req.url ?? "", req);
-  const upstreamOrigin = route ? route.upstream : opts.upstream;
-  const protocol = route?.explicitProtocol ?? (req.method === "POST" && bodyBuffer.length > 0 ? urlPath.endsWith("/chat/completions") ? "openai" : urlPath.endsWith("/v1/messages") || urlPath.endsWith("/messages") ? "anthropic" : urlPath.endsWith("/responses") || responsesCompact ? "responses" : null : null);
   let parsed = null;
   if (protocol && bodyBuffer.length > 0) {
     try {
@@ -52098,7 +52151,7 @@ async function handle(req, res, opts, core, config, log2) {
   }
   if (!prepared) {
     if (protocol === null && !opts.passthrough) {
-      log2("warn", `unrecognized path ${url} \u2014 not a known protocol (/chat/completions, /v1/messages, /responses, /responses/compact); forwarding unchanged`);
+      log2("warn", `unrecognized path ${req.url ?? ""} \u2014 not a known protocol (/chat/completions, /v1/messages, /responses, /responses/compact); forwarding unchanged`);
     }
     await forward(req, res, opts, bodyBuffer, null, core, reqConfig, log2, route, void 0);
   }