billion-context 0.1.101-pr.690.515 → 0.1.101

package/dist/index.js

@@ -43467,7 +43467,7 @@ var require_lib = __commonJS({
   }
 });
 
-// node_modules/acp-kernel/dist/chunk-6TAK7DSI.js
+// node_modules/acp-kernel/dist/chunk-JGNVAFFE.js
 import { createRequire } from "module";
 var require2 = createRequire(import.meta.url);
 function defaultCountTokens(text) {
@@ -43476,12 +43476,6 @@ function defaultCountTokens(text) {
   const cjkCount = cjk?.length ?? 0;
   return cjkCount + Math.ceil((text.length - cjkCount) / 4);
 }
-function thinkingTokenValue(thinking) {
-  return typeof thinking === "number" && Number.isFinite(thinking) && thinking > 0 ? thinking : 0;
-}
-function countMessageTokens(message, countTokens = defaultCountTokens) {
-  return countTokens(message.text ?? "") + thinkingTokenValue(message.thinkingTokens);
-}
 function estimateTokensFast(text) {
   if (!text) return 0;
   return Math.ceil(text.length / 4);
@@ -45289,8 +45283,7 @@ function renderMessage(message, map, countTokens, strategy, snapshot = null) {
     "^" + escapeRegex(TAG_OPEN) + "[^>]*" + GT + escapeRegex(ref) + escapeRegex(TAG_CLOSE) + "\\n?"
   );
   const cleanText = (message.text || "").replace(ownTagRe, "");
-  const textTokens = snapshot ? snapshot[ref] ?? (snapshot[ref] = countTokens(cleanText)) : countTokens(cleanText);
-  const tokens = textTokens + thinkingTokenValue(message.thinkingTokens);
+  const tokens = snapshot ? snapshot[ref] ?? (snapshot[ref] = countTokens(cleanText)) : countTokens(cleanText);
   const type = classifyType(message);
   const prefix = acpTag(ref, tokens, type) + "\n";
   if (!cleanText) return { ...message, text: prefix };
@@ -45369,12 +45362,8 @@ function adjustBoundariesForReasoningPairs(startIndex, endIndex, messages) {
         j2++;
       }
       const companion = messages[j2 + 1];
-      if (companion !== void 0 && companion.role === "assistant" && (companion.contentType === "text" || companion.contentType === "tool-call")) {
-        let e = j2 + 1;
-        while (e + 1 < messages.length && messages[e + 1].role === "assistant" && (messages[e + 1].contentType === "text" || messages[e + 1].contentType === "tool-call")) {
-          e++;
-        }
-        if (e > newEndIndex) newEndIndex = e;
+      if (companion !== void 0 && companion.role === "assistant" && (companion.contentType === "text" || companion.contentType === "tool-call") && j2 + 1 > newEndIndex) {
+        newEndIndex = j2 + 1;
       }
     }
     if (msg.role === "assistant" && (msg.contentType === "text" || msg.contentType === "tool-call")) {
@@ -45390,64 +45379,6 @@ function adjustBoundariesForReasoningPairs(startIndex, endIndex, messages) {
   }
   return { startIndex: newStartIndex, endIndex: newEndIndex };
 }
-function isAssistantAct(msg) {
-  return msg.role === "assistant" && (msg.contentType === "text" || msg.contentType === "tool-call");
-}
-function computeTurnGroups(messages) {
-  const resultIdByCallId = /* @__PURE__ */ new Map();
-  for (const msg of messages) {
-    if (msg.contentType === "tool-result" && typeof msg.toolCallId === "string" && msg.id) {
-      if (!resultIdByCallId.has(msg.toolCallId))
-        resultIdByCallId.set(msg.toolCallId, msg.id);
-    }
-  }
-  const grouped = /* @__PURE__ */ new Set();
-  const groups = [];
-  for (let i = 0; i < messages.length; i++) {
-    const msg = messages[i];
-    if (!msg.id || grouped.has(msg.id)) continue;
-    if (!(msg.contentType === "reasoning" || isAssistantAct(msg))) continue;
-    let reasoningStart = i;
-    if (msg.contentType === "reasoning") {
-      while (reasoningStart > 0 && messages[reasoningStart - 1].contentType === "reasoning") {
-        reasoningStart--;
-      }
-    } else {
-      let s3 = i;
-      while (s3 > 0 && isAssistantAct(messages[s3 - 1])) s3--;
-      reasoningStart = s3;
-      while (reasoningStart > 0 && messages[reasoningStart - 1].contentType === "reasoning") {
-        reasoningStart--;
-      }
-    }
-    const burstStart = (() => {
-      let s3 = reasoningStart;
-      while (s3 < messages.length && messages[s3].contentType === "reasoning")
-        s3++;
-      return s3;
-    })();
-    if (burstStart >= messages.length || !isAssistantAct(messages[burstStart])) {
-      continue;
-    }
-    let burstEnd = burstStart;
-    while (burstEnd + 1 < messages.length && isAssistantAct(messages[burstEnd + 1])) {
-      burstEnd++;
-    }
-    const members = /* @__PURE__ */ new Set();
-    for (let k2 = reasoningStart; k2 <= burstEnd; k2++) {
-      const m2 = messages[k2];
-      if (!m2.id) continue;
-      members.add(m2.id);
-      if (m2.role === "assistant" && m2.contentType === "tool-call" && typeof m2.toolCallId === "string") {
-        const rid = resultIdByCallId.get(m2.toolCallId);
-        if (rid) members.add(rid);
-      }
-    }
-    for (const id of members) grouped.add(id);
-    groups.push([...members]);
-  }
-  return groups;
-}
 function estimateTextTokens(text) {
   return Math.ceil(text.length / 4);
 }
@@ -45471,7 +45402,7 @@ function computeProtectedRefs(messages, state, config, countTokens = estimateTex
     if (isNeverPreserveRecent(msg)) continue;
     const ref = state.messageRefs.byRaw[msg.id];
     if (!ref || ref === "BLOCKED") continue;
-    visible.push({ ref, tokens: countMessageTokens(msg, countTokens) });
+    visible.push({ ref, tokens: countTokens(msg.text ?? "") });
   }
   if (preserveN > 0) {
     for (const m2 of visible.slice(-preserveN)) {
@@ -45514,7 +45445,7 @@ function buildCompressibleRanges(messages, state, config, protectedZoneRefs, cou
       protectedMsgs.push({
         ref,
         gapBefore: skipSinceProtected,
-        tokens: countMessageTokens(msg, countTokens),
+        tokens: countTokens(msg.text ?? ""),
         tools: msg.toolName ? [msg.toolName] : []
       });
       skipSinceProtected = false;
@@ -45529,7 +45460,7 @@ function buildCompressibleRanges(messages, state, config, protectedZoneRefs, cou
     compressibleMsgs.push({
       ref,
       gapBefore: skipSinceCompressible,
-      tokens: countMessageTokens(msg, countTokens),
+      tokens: countTokens(msg.text ?? ""),
       chars: (msg.text ?? "").length,
       isTool: isToolMessage(msg),
       isUser: msg.role === "user"
@@ -46162,42 +46093,6 @@ function applySingleRange(input) {
       )} from compression range (recent/last-user zone).`
     );
   }
-  {
-    const reasoningIds = /* @__PURE__ */ new Set();
-    const callIds = /* @__PURE__ */ new Set();
-    for (const m2 of input.messages) {
-      if (!m2.id) continue;
-      if (m2.contentType === "reasoning") reasoningIds.add(m2.id);
-      if (m2.role === "assistant" && m2.contentType === "tool-call") callIds.add(m2.id);
-    }
-    const withdrawIds = /* @__PURE__ */ new Set();
-    let splitTurnCount = 0;
-    for (const group of computeTurnGroups(input.messages)) {
-      const foldHasReasoning = group.some(
-        (id) => effectiveMessageIds.has(id) && reasoningIds.has(id)
-      );
-      if (!foldHasReasoning) continue;
-      const keptHasCall = group.some(
-        (id) => !effectiveMessageIds.has(id) && callIds.has(id)
-      );
-      if (!keptHasCall) continue;
-      splitTurnCount++;
-      for (const id of group) withdrawIds.add(id);
-    }
-    if (withdrawIds.size > 0) {
-      for (const id of withdrawIds) effectiveMessageIds.delete(id);
-      const beforeWithdraw = filteredIds.length;
-      filteredIds = filteredIds.filter((id) => !withdrawIds.has(id));
-      if (filteredIds.length === 0 && consumedBlockIds.length === 0) {
-        throw new Error(
-          `Range would split ${splitTurnCount} turn(s) at the protected-zone boundary: a visible tool-call must keep its reasoning run (strict-echo providers reject a rebuilt request that lost it). Shrink the range to end before the turn starts, or wait until the whole turn ages out of the protected zone.`
-        );
-      }
-      warnings.push(
-        `Withdrawn ${beforeWithdraw - filteredIds.length} message(s) from compression range to keep ${splitTurnCount} turn(s) intact (visible tool-call would lose its reasoning run).`
-      );
-    }
-  }
   if (!isBlockBoundary && filteredIds.length === 0 && consumedBlockIds.length > 0) {
     const first = consumedBlockIds[0];
     const last = consumedBlockIds[consumedBlockIds.length - 1];
@@ -46211,7 +46106,7 @@ function applySingleRange(input) {
   let compressedTokens = 0;
   for (const id of filteredIds) {
     const message = input.messages.find((entry) => entry.id === id);
-    compressedTokens += message ? countMessageTokens(message, input.countTokens) : 0;
+    compressedTokens += input.countTokens(message?.text ?? "");
   }
   for (const consumedId of consumedBlockIds) {
     const consumed = blockById(input.state, consumedId);
@@ -46541,7 +46436,7 @@ function computeContextBreakdown(messages, total, growth, countTokens) {
   const count = countTokens ?? ((t) => Math.ceil(t.length / 4));
   let system = 0, tool = 0, summaries = 0, code = 0, text = 0;
   for (const msg of messages) {
-    const tokens = countMessageTokens(msg, count);
+    const tokens = count(msg.text ?? "");
     if (msg.text?.startsWith("[Compressed conversation section]")) {
       summaries += tokens;
     } else if (msg.contentType === "tool-call" || msg.contentType === "tool-result") {
@@ -46727,7 +46622,7 @@ function collectVisible(messages, state, countTokens) {
     if (coveredIds.has(message.id)) return;
     const ref = refForRaw(state.messageRefs, message.id);
     if (!ref) return;
-    const tokens = countMessageTokens(message, countTokens);
+    const tokens = countTokens(message.text ?? "");
     const tool = isToolMessage(message) ? message.toolName ?? (message.toolCallId ? toolCallNames.get(message.toolCallId) : void 0) ?? "tool" : "text";
     if (tokens > 0) visible.push({ ref, tokens, tool, index });
   });
@@ -53826,11 +53721,6 @@ async function* runCompressLoop(upstream, ctx, requestBody, requestOptions, adap
         }
       } catch (e) {
         if (!(e instanceof UpstreamHttpError)) throw e;
-        if (e.status === 400 && /reasoning_content/i.test(e.body) && ctx.session.metadata.strictReasoningEcho !== true) {
-          ctx.session.metadata.strictReasoningEcho = true;
-          ctx.log(`[acp-loop] 400 mentions reasoning_content \u2014 learned strict reasoning echo for this session; #651 reasoning-drop disabled (#684)`);
-          log("warn", `[acp-loop] learned strictReasoningEcho (session ${ctx.session.id}); reasoning-drop disabled (#684)`);
-        }
         const suffix = e.attempts > 1 ? ` after ${e.attempts} attempt(s)` : "";
         ctx.log(`[acp-proxy: compress loop upstream error ${e.status}${suffix}: ${e.body.slice(0, 200)}]`);
         log("error", `[acp-loop] upstream error ${e.status}${suffix}: ${e.body.slice(0, 200)}`);
@@ -59646,7 +59536,7 @@ ${bodyBuffer.toString("utf8")}`);
             log2("info", `[debug] strip-images: dropped ${stripped.removed} historical image part(s), kept last ${keepRecent} (session=${session.id})`);
           }
           const work = stripped.body;
-          return countTokens ? prepareCountTokens(work, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(work, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, upstreamOrigin, reasoningCfg) : protocol === "openai" ? prepareOpenai(work, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, upstreamOrigin, nativeWindow, reasoningCfg) : responsesCompact ? prepareResponsesCompact(stripped.removed > 0 ? Buffer.from(JSON.stringify(work)) : bodyBuffer, work, session, req, core, reqConfig, log2) : prepareResponses(work, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin, nativeWindow, reasoningCfg);
+          return countTokens ? prepareCountTokens(work, core, reqConfig, log2, session) : protocol === "anthropic" ? prepareAnthropic(work, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, reasoningCfg) : protocol === "openai" ? prepareOpenai(work, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, nativeWindow, reasoningCfg) : responsesCompact ? prepareResponsesCompact(stripped.removed > 0 ? Buffer.from(JSON.stringify(work)) : bodyBuffer, work, session, req, core, reqConfig, log2) : prepareResponses(work, req, opts, core, reqConfig, reqPrompts, log2, session, responsesIdentity, pluginMode, upstreamOrigin, nativeWindow, reasoningCfg);
         };
         const isCodexCompactTrigger = protocol === "responses" && !responsesCompact && isCodexClient(req.headers) && hasCompactionTrigger(parsed.input);
         if (isCodexCompactTrigger) {
@@ -59734,64 +59624,13 @@ ${bodyBuffer.toString("utf8")}`);
   }
 }
 var ACP_TAG_MARK = "<acp ";
-function withReasoningDrop(msgs, reasoning, log2, sessionId, strictEcho) {
-  if (strictEcho) return msgs;
+function withReasoningDrop(msgs, reasoning, log2, sessionId) {
   const out = dropCompressReasoning(msgs, reasoning);
   if (out.length !== msgs.length) {
     log2("info", `[${sessionId}] compress-reasoning: dropped ${msgs.length - out.length} reasoning message(s) from closed compress turns (#651)`);
   }
   return out;
 }
-function isStrictReasoningEcho(session, upstreamOrigin) {
-  if (session.metadata.strictReasoningEcho === true) return true;
-  return upstreamOrigin !== void 0 && /deepseek/i.test(upstreamOrigin);
-}
-function warnReasoningPairs(wireMessages, log2, sessionId) {
-  let withRc = 0;
-  let split = 0;
-  for (const m2 of wireMessages) {
-    const msg = m2;
-    if (msg?.role !== "assistant") continue;
-    const hasRc = typeof msg.reasoning_content === "string" && msg.reasoning_content.length > 0;
-    if (hasRc) withRc++;
-    else if (Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0) split++;
-  }
-  if (withRc > 0 && split > 0) {
-    log2("warn", `[${sessionId}] reasoning-pair-violated: ${split} assistant tool-call message(s) lack reasoning_content while ${withRc} carry it \u2014 strict-echo upstreams (DeepSeek thinking mode) will reject the request (#684)`);
-  }
-}
-function warnAnthropicThinkingPairs(wireMessages, log2, sessionId) {
-  let withThinking = 0;
-  let split = 0;
-  for (const m2 of wireMessages) {
-    const msg = m2;
-    if (msg?.role !== "assistant" || !Array.isArray(msg.content)) continue;
-    const hasThinking = msg.content.some((b2) => b2?.type === "thinking");
-    if (hasThinking) withThinking++;
-    else if (msg.content.some((b2) => b2?.type === "tool_use")) split++;
-  }
-  if (withThinking > 0 && split > 0) {
-    log2("warn", `[${sessionId}] thinking-pair-violated: ${split} assistant tool_use message(s) lack a thinking block while ${withThinking} carry one \u2014 extended-thinking tool use requires preserved thinking (#684)`);
-  }
-}
-function warnResponsesReasoningPairs(input, log2, sessionId) {
-  let withReasoning = 0;
-  let split = 0;
-  let prevWasReasoning = false;
-  for (const item of input) {
-    const it2 = item;
-    if (it2?.type === "reasoning") {
-      withReasoning++;
-      prevWasReasoning = true;
-      continue;
-    }
-    if (it2?.type === "function_call" && !prevWasReasoning) split++;
-    prevWasReasoning = false;
-  }
-  if (withReasoning > 0 && split > 0) {
-    log2("warn", `[${sessionId}] reasoning-pair-violated: ${split} function_call item(s) lack a preceding reasoning item while ${withReasoning} exist \u2014 strict-echo upstreams will reject the request (#684)`);
-  }
-}
 function stripKernelSummaries(messages, state) {
   const carried = /* @__PURE__ */ new Set();
   for (const b2 of state.blocks) {
@@ -59895,13 +59734,13 @@ function effectiveTokenCount(session, msgs) {
   if (!session.metadata.anonymousPrefixAffinity) return 0;
   return estimateCoreMessagesUpper(msgs);
 }
-function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, session, pluginMode, upstreamOrigin, reasoning) {
+function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, session, pluginMode, reasoning) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
   session.hostCreditTokens = 0;
   const injectTools = opts.compress.injectTool && !pluginMode;
-  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
+  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId);
   if (isAutoModeClassifier(parsed)) {
     log2("info", `[${sessionId}] auto-mode classifier passthrough (skipping compress injection)`);
     return { body: JSON.stringify(parsed), session, processedMessages: [], originalMessages: [], anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: false, pluginMode, nudge: void 0, prompts };
@@ -59964,7 +59803,6 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
   snapshotMessages(session, originalMessages);
   markDirty(session);
   const rebuilt = { ...parsed, messages: rebuiltMessages, system: systemOut, tools: toolsOut };
-  warnAnthropicThinkingPairs(rebuiltMessages, log2, sessionId);
   delete rebuilt.prompt_cache_key;
   armHostUsageCredit(session, originalMessages, processedMessages, req.headers, opts.hostUsageCredit, log2);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, anthropicSystem: parsed.system, protocol: "anthropic", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, renderTags: "text-only" };
@@ -60022,13 +59860,13 @@ function clampOutgoingOutput(rebuilt, field, ctx, sessionId, log2) {
     log2("info", `[${sessionId}] output budget clamped ${raw} -> ${capped} (input~${inputEstimate}, window=${ctx.nativeWindow}); prevents input+output overflow (#453)`);
   }
 }
-function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode, upstreamOrigin, nativeWindow, reasoning) {
+function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session, pluginMode, nativeWindow, reasoning) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
   session.hostCreditTokens = 0;
   let openaiSystemText = "";
-  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
+  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId);
   let openaiOutboundSystem;
   let processedMessages = [];
   let originalMessages = [];
@@ -60093,7 +59931,6 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
     processedMessages = [];
   }
   const rebuilt = { ...parsed, messages: rebuiltMessages, tools: toolsOut };
-  warnReasoningPairs(rebuiltMessages, log2, sessionId);
   clampOutgoingOutput(rebuilt, typeof parsed.max_completion_tokens === "number" ? "max_completion_tokens" : "max_tokens", { systemText: openaiSystemText, tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow, imageTokens: imageTokensInParsedBody("openai", rebuilt) }, sessionId, log2);
   delete rebuilt.prompt_cache_retention;
   if (stream2 && rebuilt.stream_options === void 0) {
@@ -60112,7 +59949,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
   session.hostCreditTokens = 0;
-  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId, isStrictReasoningEcho(session, upstreamOrigin));
+  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId);
   if (reconcileNativeCompactionBoundary(session)) {
     log2("info", `[${sessionId}] reconciled ACP state after native Responses compact boundary`);
   }
@@ -60230,7 +60067,6 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", `[${sessionId}] codex compact intercepted (trigger); forged SSE with ${summaries.length} block summary(s), upstream not contacted`);
   }
   const rebuilt = { ...parsed, input: rebuiltInput, tools: toolsOut };
-  warnResponsesReasoningPairs(Array.isArray(rebuiltInput) ? rebuiltInput : [], log2, sessionId);
   if (!isCompactionTrigger) {
     clampOutgoingOutput(rebuilt, "max_output_tokens", { systemText: (responsesProjection?.systemParts ?? []).join("\n"), tools: toolsOut, processedMessages, lastInputTokens: session.stats.lastInputTokens, nativeWindow, imageTokens: imageTokensInParsedBody("responses", rebuilt) }, sessionId, log2);
   }