billion-context 0.1.96-pr.665.479 → 0.1.96

package/dist/index.js

@@ -44472,71 +44472,25 @@ var KEEP_LAST_ORPHANED = 2;
 function rangeKey(startRef, endRef) {
   return `${startRef}::${endRef}`;
 }
-function parseCallText(text) {
-  const raw = text ?? "";
-  const start = raw.indexOf("{");
-  if (start < 0) return null;
+function rewriteCompressText(text, liveKeys) {
   let parsed;
   try {
-    parsed = JSON.parse(raw.slice(start));
+    parsed = JSON.parse(text ?? "");
   } catch {
     return null;
   }
   if (!parsed || typeof parsed !== "object") return null;
   const obj = parsed;
-  let content = null;
-  let contentWasString = false;
-  if (Array.isArray(obj.content)) {
-    content = obj.content;
-  } else if (typeof obj.content === "string") {
-    contentWasString = true;
-    try {
-      const inner = JSON.parse(obj.content);
-      if (Array.isArray(inner)) content = inner;
-    } catch {
-      content = null;
-    }
-  }
-  if (!content || content.length === 0) return null;
-  return { prefix: raw.slice(0, start), obj, content, contentWasString };
-}
-function rewriteCompressText(text, liveKeys) {
-  const parsed = parseCallText(text);
-  if (!parsed) return null;
-  const { prefix, obj, content, contentWasString } = 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 e = entry;
-    const s3 = typeof e.startId === "string" ? e.startId : typeof e.messageId === "string" ? e.messageId : "";
-    const end = typeof e.endId === "string" ? e.endId : typeof e.messageId === "string" ? e.messageId : "";
-    return liveKeys.has(rangeKey(s3, end));
+    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 === 0) return null;
-  return prefix + serializeCompacted(obj, kept, contentWasString).text;
-}
-var SUMMARY_STUB_CHARS = 200;
-function compactEntry(entry) {
-  if (!entry || typeof entry !== "object") return entry;
-  const e = entry;
-  if (typeof e.summary !== "string" || e.summary.length <= SUMMARY_STUB_CHARS) return entry;
-  return { ...e, summary: `${e.summary.slice(0, SUMMARY_STUB_CHARS - 1)}\u2026` };
-}
-function serializeCompacted(obj, content, contentWasString) {
-  let changed = false;
-  const compacted = content.map((entry) => {
-    const out = compactEntry(entry);
-    if (out !== entry) changed = true;
-    return out;
-  });
-  const outContent = contentWasString ? JSON.stringify(compacted) : compacted;
-  return { text: JSON.stringify({ ...obj, content: outContent }), changed };
-}
-function compactCompressText(text) {
-  const parsed = parseCallText(text);
-  if (!parsed) return null;
-  const { prefix, obj, content, contentWasString } = parsed;
-  const { text: out, changed } = serializeCompacted(obj, content, contentWasString);
-  return changed ? prefix + out : null;
+  if (kept.length === content.length || kept.length === 0) return null;
+  return JSON.stringify({ ...obj, content: kept });
 }
 function hideConsumedCompressCalls(state, messages) {
   const allBlockCallIds = /* @__PURE__ */ new Set();
@@ -44595,11 +44549,6 @@ function hideConsumedCompressCalls(state, messages) {
           continue;
         }
       }
-      const compacted = compactCompressText(message.text);
-      if (compacted !== null) {
-        result.push({ ...message, text: compacted });
-        continue;
-      }
     }
     result.push(message);
   }
@@ -51872,32 +51821,16 @@ function splitChunks(messages, startIdx, endIdx, budget, minUnits, countText = d
   }
   return chunks;
 }
-function summaryPayload(protocol, model, system, content, stream2, includeMaxOutputTokens) {
+function summaryPayload(protocol, model, system, content, stream2) {
   if (protocol === "anthropic") {
     return { model, max_tokens: MAX_SUMMARY_OUTPUT_TOKENS, system, messages: [{ role: "user", content }], stream: stream2 };
   }
   if (protocol === "openai") {
     return { model, max_tokens: MAX_SUMMARY_OUTPUT_TOKENS, messages: [{ role: "system", content: system }, { role: "user", content }], stream: stream2 };
   }
-  const payload = { model, instructions: system, input: [{ role: "user", content }], stream: stream2, store: false };
-  if (includeMaxOutputTokens) payload.max_output_tokens = MAX_SUMMARY_OUTPUT_TOKENS;
-  return payload;
+  return { model, max_output_tokens: MAX_SUMMARY_OUTPUT_TOKENS, instructions: system, input: [{ role: "user", content }], stream: stream2, store: false };
 }
 var STREAM_REQUIRED_RE = /\bstream\b[^\n]{0,60}\btrue\b/i;
-var MAX_OUTPUT_TOKENS_REJECTED_RE = /\bmax_output_tokens\b/i;
-function noMaxOutputTokensKey(deps) {
-  return `${deps.url}\0${deps.model}`;
-}
-function hasLearnedNoMaxOutputTokens(deps) {
-  const learned = deps.session.metadata.preflightNoMaxOutputTokens;
-  return typeof learned === "object" && learned !== null && learned[noMaxOutputTokensKey(deps)] === true;
-}
-function rememberNoMaxOutputTokens(deps) {
-  const learned = deps.session.metadata.preflightNoMaxOutputTokens;
-  const map = typeof learned === "object" && learned !== null ? learned : {};
-  map[noMaxOutputTokensKey(deps)] = true;
-  deps.session.metadata.preflightNoMaxOutputTokens = map;
-}
 function extractSummaryFromSse(protocol, text) {
   let out = "";
   for (const line of text.split("\n")) {
@@ -51961,39 +51894,25 @@ async function summarizeRange(deps, content, startRef, endRef) {
   const system = buildCompressSystemPrompt(deps.prompts) + `
 
 TASK: The conversation segment below (messages ${startRef}\u2013${endRef}) must be compressed because the session context exceeds the current model's window. Write a tier-1 compression summary of the segment following every rule above. Output ONLY the summary text \u2014 no preamble, no closing remarks, no tool calls.`;
-  let stream2 = deps.session.metadata.preflightStreamSummary === true;
-  let includeMaxOutputTokens = !(deps.protocol === "responses" && hasLearnedNoMaxOutputTokens(deps));
-  for (; ; ) {
-    try {
-      return await requestSummary(deps, system, content, stream2, includeMaxOutputTokens);
-    } catch (err2) {
-      if (err2 instanceof UpstreamHttpError && err2.status === 400) {
-        let adapted = false;
-        if (!stream2 && STREAM_REQUIRED_RE.test(err2.body)) {
-          deps.session.metadata.preflightStreamSummary = true;
-          stream2 = true;
-          adapted = true;
-          deps.log("info", "[preflight] upstream requires stream for summaries; retrying with SSE (learned for this session)");
-        }
-        if (deps.protocol === "responses" && includeMaxOutputTokens && MAX_OUTPUT_TOKENS_REJECTED_RE.test(err2.body)) {
-          rememberNoMaxOutputTokens(deps);
-          includeMaxOutputTokens = false;
-          adapted = true;
-          deps.log("info", `[preflight] upstream rejects max_output_tokens for summaries (model=${deps.model}); retrying without it (learned for this session+upstream+model)`);
-        }
-        if (adapted) continue;
-      }
-      throw err2;
+  const learned = deps.session.metadata.preflightStreamSummary === true;
+  try {
+    return await requestSummary(deps, system, content, learned);
+  } catch (err2) {
+    if (err2 instanceof UpstreamHttpError && err2.status === 400 && !learned && STREAM_REQUIRED_RE.test(err2.body)) {
+      deps.session.metadata.preflightStreamSummary = true;
+      deps.log("info", "[preflight] upstream requires stream for summaries; retrying with SSE (learned for this session)");
+      return await requestSummary(deps, system, content, true);
     }
+    throw err2;
   }
 }
-async function requestSummary(deps, system, content, stream2, includeMaxOutputTokens) {
+async function requestSummary(deps, system, content, stream2) {
   const { response, clearTimer } = await fetchWithRetry(
     deps.url,
     {
       method: "POST",
       headers: { "content-type": "application/json", ...deps.headers },
-      body: JSON.stringify(summaryPayload(deps.protocol, deps.model, system, content, stream2, includeMaxOutputTokens)),
+      body: JSON.stringify(summaryPayload(deps.protocol, deps.model, system, content, stream2)),
       dispatcher: proxyDispatcher(deps.proxyUrl)
     },
     void 0,