billion-context 0.1.96 → 0.1.97-pr.667.482

package/dist/index.js

@@ -44472,25 +44472,71 @@ var KEEP_LAST_ORPHANED = 2;
 function rangeKey(startRef, endRef) {
   return `${startRef}::${endRef}`;
 }
-function rewriteCompressText(text, liveKeys) {
+function parseCallText(text) {
+  const raw = text ?? "";
+  const start = raw.indexOf("{");
+  if (start < 0) return null;
   let parsed;
   try {
-    parsed = JSON.parse(text ?? "");
+    parsed = JSON.parse(raw.slice(start));
   } 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;
+  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 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));
+    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));
   });
-  if (kept.length === content.length || kept.length === 0) return null;
-  return JSON.stringify({ ...obj, content: kept });
+  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;
 }
 function hideConsumedCompressCalls(state, messages) {
   const allBlockCallIds = /* @__PURE__ */ new Set();
@@ -44549,6 +44595,11 @@ function hideConsumedCompressCalls(state, messages) {
           continue;
         }
       }
+      const compacted = compactCompressText(message.text);
+      if (compacted !== null) {
+        result.push({ ...message, text: compacted });
+        continue;
+      }
     }
     result.push(message);
   }
@@ -46262,6 +46313,9 @@ function decideNudge(input) {
   const growthReady = firstSightMassReady || growthSinceReference >= growthFloor;
   const t2Count = tiers[2]?.targetBlocks.length ?? 0;
   const t3Count = tiers[3]?.targetBlocks.length ?? 0;
+  const tierCountUsageFloor = config.nudge.minContextLimitPct;
+  const t2CountReady = t2Count >= config.tiers.tier2Trigger && usage >= tierCountUsageFloor;
+  const t3CountReady = t3Count >= config.tiers.tier3Trigger && usage >= tierCountUsageFloor;
   if (pressure) {
     const candidates = [1];
     if (config.tiers.enabled) {
@@ -46284,19 +46338,19 @@ function decideNudge(input) {
     if (t1Eff >= nudgeGrowthTokens) {
       injectedTier = 1;
       injectedReason = `T1 effective ${t1Eff} >= ${nudgeGrowthTokens}, growth ${growthSinceReference}, usage ${Math.round(usage * 100)}%`;
-    } else if (config.tiers.enabled && (t2Count >= config.tiers.tier2Trigger || t2Pen >= tier2Threshold && t2Pen > t1Eff)) {
+    } else if (config.tiers.enabled && (t2CountReady || t2Pen >= tier2Threshold && t2Pen > t1Eff)) {
       const lastShown = state.nudge.lastShownByTier[2] ?? 0;
       const cadenceMet = lastShown === 0 || tokenCount - lastShown >= growthFloor;
       if (cadenceMet) {
         injectedTier = 2;
-        injectedReason = t2Count >= config.tiers.tier2Trigger ? `T2 distill ready: ${t2Count} tier-1 blocks >= tier2Trigger ${config.tiers.tier2Trigger} (${t2Pen} tokens), usage ${Math.round(usage * 100)}%` : `T2 distill ready: ${tiers[2].targetBlocks.length} tier-1 blocks (${t2Pen} tokens) >= ${tier2Threshold} (1.5x) and > T1 effective ${t1Eff}, usage ${Math.round(usage * 100)}%`;
+        injectedReason = t2CountReady ? `T2 distill ready: ${t2Count} tier-1 blocks >= tier2Trigger ${config.tiers.tier2Trigger} (${t2Pen} tokens), usage ${Math.round(usage * 100)}%` : `T2 distill ready: ${tiers[2].targetBlocks.length} tier-1 blocks (${t2Pen} tokens) >= ${tier2Threshold} (1.5x) and > T1 effective ${t1Eff}, usage ${Math.round(usage * 100)}%`;
       }
-    } else if (config.tiers.enabled && (t3Count >= config.tiers.tier3Trigger || t3Pen >= tier2Threshold && t3Pen > t2Pen && t3Pen > t1Eff)) {
+    } else if (config.tiers.enabled && (t3CountReady || t3Pen >= tier2Threshold && t3Pen > t2Pen && t3Pen > t1Eff)) {
       const lastShown = state.nudge.lastShownByTier[3] ?? 0;
       const cadenceMet = lastShown === 0 || tokenCount - lastShown >= growthFloor;
       if (cadenceMet) {
         injectedTier = 3;
-        injectedReason = t3Count >= config.tiers.tier3Trigger ? `T3 condense ready: ${t3Count} tier-2 blocks >= tier3Trigger ${config.tiers.tier3Trigger} (${t3Pen} tokens), usage ${Math.round(usage * 100)}%` : `T3 condense ready: ${tiers[3].targetBlocks.length} tier-2 blocks (${t3Pen} tokens) >= ${tier2Threshold} (1.5x) and > T2 ${t2Pen} and > T1 effective ${t1Eff}, usage ${Math.round(usage * 100)}%`;
+        injectedReason = t3CountReady ? `T3 condense ready: ${t3Count} tier-2 blocks >= tier3Trigger ${config.tiers.tier3Trigger} (${t3Pen} tokens), usage ${Math.round(usage * 100)}%` : `T3 condense ready: ${tiers[3].targetBlocks.length} tier-2 blocks (${t3Pen} tokens) >= ${tier2Threshold} (1.5x) and > T2 ${t2Pen} and > T1 effective ${t1Eff}, usage ${Math.round(usage * 100)}%`;
       }
     }
   }
@@ -46313,9 +46367,14 @@ function decideNudge(input) {
   } else {
     const tiersList = [1, 2, 3];
     const eligible = tiersList.filter((t) => config.tiers.enabled || t === 1);
-    const countReady = (t) => t === 2 ? t2Count >= config.tiers.tier2Trigger : t === 3 ? t3Count >= config.tiers.tier3Trigger : false;
+    const countReadyUngated = (t) => t === 2 ? t2Count >= config.tiers.tier2Trigger : t === 3 ? t3Count >= config.tiers.tier3Trigger : false;
+    const countReady = (t) => countReadyUngated(t) && usage >= tierCountUsageFloor;
     const ready = eligible.filter((t) => (tiers[t]?.pending ?? 0) >= nudgeGrowthTokens).map((t) => `T${t} ${tiers[t].pending}`);
-    const readyCount = eligible.filter((t) => (tiers[t]?.pending ?? 0) < nudgeGrowthTokens && countReady(t)).map((t) => `T${t} ${t === 2 ? t2Count : t3Count} blocks (count)`);
+    const readyCount = eligible.filter(
+      (t) => (tiers[t]?.pending ?? 0) < nudgeGrowthTokens && countReadyUngated(t)
+    ).map(
+      (t) => `T${t} ${t === 2 ? t2Count : t3Count} blocks (count${usage >= tierCountUsageFloor ? "" : ", usage-gated"})`
+    );
     const readyAll = [...ready, ...readyCount];
     const readyHint = readyAll.length > 0 ? `, ready: ${readyAll.join(", ")}` : "";
     const blocked = eligible.filter(
@@ -48667,6 +48726,7 @@ function mergeCompress(global2, provider, model) {
   const promptLevels = [global2?.prompts, provider?.prompts, model?.prompts].filter(Boolean);
   const rangeOf = (s3) => s3?.minCompressRangeChars ?? s3?.minCompressRange;
   const absorbLevels = [global2?.absorb, provider?.absorb, model?.absorb].filter(Boolean);
+  const reasoningLevels = [global2?.reasoning, provider?.reasoning, model?.reasoning].filter(Boolean);
   return {
     modelContextLimit: pick("modelContextLimit"),
     maxContextLimit: pick("maxContextLimit"),
@@ -48680,7 +48740,11 @@ function mergeCompress(global2, provider, model) {
     acknowledgePromptsRisk: pick("acknowledgePromptsRisk"),
     absorb: absorbLevels.length > 0 ? Object.assign({}, ...absorbLevels) : void 0,
     stripImages: pick("stripImages"),
-    stripImagesKeepRecent: pick("stripImagesKeepRecent")
+    stripImagesKeepRecent: pick("stripImagesKeepRecent"),
+    // `reasoning` is a third nested-object field merged sub-field-wise
+    // exactly like `absorb`/`prompts`: a model-level `threshold` must not
+    // discard a provider-level `drop: false`.
+    reasoning: reasoningLevels.length > 0 ? Object.assign({}, ...reasoningLevels) : void 0
   };
 }
 function resolveCompress(routes, upstreamUrl, model, global2) {
@@ -48760,6 +48824,49 @@ function resolveRequestConfig(base, routes, embeddedUrl, model, native, globalCo
   return applyCompressSettings(base, limit, compress);
 }
 
+// src/reasoning-drop.ts
+var DEFAULT_COMPRESS_REASONING = { drop: true, threshold: 2048 };
+function resolveReasoningDrop(cfg) {
+  let threshold = DEFAULT_COMPRESS_REASONING.threshold;
+  if (cfg?.threshold !== void 0) {
+    const t = cfg.threshold;
+    if (typeof t === "number" && Number.isFinite(t) && t >= 0) {
+      threshold = Math.floor(t);
+    }
+  }
+  return { drop: cfg?.drop !== false, threshold };
+}
+function dropCompressReasoning(messages, cfg) {
+  const { drop, threshold } = resolveReasoningDrop(cfg);
+  if (!drop || messages.length === 0) return messages;
+  try {
+    let lastUser = -1;
+    for (let i = 0; i < messages.length; i++) {
+      const m2 = messages[i];
+      if (m2.role === "user" && m2.contentType === "text") lastUser = i;
+    }
+    if (lastUser < 0) return messages;
+    const dropIdx = /* @__PURE__ */ new Set();
+    for (let i = 0; i < lastUser; i++) {
+      const m2 = messages[i];
+      if (m2.contentType !== "tool-call" || m2.toolName !== "compress") continue;
+      let total = 0;
+      let j2 = i - 1;
+      while (j2 >= 0 && messages[j2].contentType === "reasoning") {
+        total += (messages[j2].text ?? "").length;
+        j2--;
+      }
+      if (total > threshold) {
+        for (let k2 = j2 + 1; k2 < i; k2++) dropIdx.add(k2);
+      }
+    }
+    if (dropIdx.size === 0) return messages;
+    return messages.filter((_2, idx) => !dropIdx.has(idx));
+  } catch {
+    return messages;
+  }
+}
+
 // src/strip-images.ts
 var DEFAULT_STRIP_IMAGES_KEEP_RECENT = 5;
 var IMAGE_PLACEHOLDER = "[image]";
@@ -59155,13 +59262,14 @@ ${bodyBuffer.toString("utf8")}`);
       await withSessionLock(session, async () => {
         const runPrepare = () => {
           const cs2 = resolveCompress(opts.routes, route?.rewrittenUrl, parsed.model, opts.compress);
+          const reasoningCfg = cs2.reasoning;
           const keepRecent = cs2.stripImagesKeepRecent ?? DEFAULT_STRIP_IMAGES_KEEP_RECENT;
           const stripped = cs2.stripImages ? stripHistoricalImages(parsed, protocol, keepRecent) : { body: parsed, removed: 0 };
           if (opts.debug && stripped.removed > 0) {
             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) : protocol === "openai" ? prepareOpenai(work, req, opts, core, reqConfig, reqPrompts, log2, session, pluginMode, nativeWindow) : 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);
+          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) {
@@ -59249,6 +59357,13 @@ ${bodyBuffer.toString("utf8")}`);
   }
 }
 var ACP_TAG_MARK = "<acp ";
+function withReasoningDrop(msgs, reasoning, log2, sessionId) {
+  const out = dropCompressReasoning(msgs, reasoning);
+  if (out.length !== msgs.length) {
+    log2("info", `[${sessionId}] compress-reasoning: dropped ${msgs.length - out.length} reasoning message(s) from closed compress turns (#651)`);
+  }
+  return out;
+}
 function stripKernelSummaries(messages, state) {
   const carried = /* @__PURE__ */ new Set();
   for (const b2 of state.blocks) {
@@ -59350,12 +59465,13 @@ function effectiveTokenCount(session, msgs) {
   if (!session.metadata.anonymousPrefixAffinity) return 0;
   return estimateCoreMessagesUpper(msgs);
 }
-function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, session, pluginMode) {
+function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, session, pluginMode, reasoning) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
   session.hostCreditTokens = 0;
   const injectTools = opts.compress.injectTool && !pluginMode;
+  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId);
   if (isAutoModeClassifier(parsed)) {
     log2("info", `[${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 };
@@ -59393,7 +59509,7 @@ function prepareAnthropic(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
     const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
-    processedMessages = stripKernelSummaries(turn.messages, turn.state);
+    processedMessages = stripReasoning(stripKernelSummaries(turn.messages, turn.state));
     applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
     reapOrphanBlocks(session, msgs, deactivateBlock);
     rebuiltMessages = coreToAnthropic(processedMessages, cacheControls);
@@ -59475,12 +59591,13 @@ 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, nativeWindow) {
+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);
   let openaiOutboundSystem;
   let processedMessages = [];
   let originalMessages = [];
@@ -59518,7 +59635,7 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
     const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
-    processedMessages = stripKernelSummaries(turn.messages, turn.state);
+    processedMessages = stripReasoning(stripKernelSummaries(turn.messages, turn.state));
     applyCompactionArchive(session, activeBefore, new Set(msgs.map((m2) => m2.id)), log2);
     reapOrphanBlocks(session, msgs, deactivateBlock);
     rebuiltMessages = systemToUser(coreToOpenai(processedMessages));
@@ -59558,11 +59675,12 @@ function prepareOpenai(parsed, req, opts, core, config, prompts, log2, session,
   markDirty(session);
   return { body: JSON.stringify(rebuilt), session, processedMessages, originalMessages, protocol: "openai", stream: stream2, compressInjected: injectTools, pluginMode, nudge, prompts, openaiSystemText, renderTags: "text-only" };
 }
-function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin, nativeWindow) {
+function prepareResponses(parsed, req, opts, core, config, prompts, log2, session, identity, pluginMode, upstreamOrigin, nativeWindow, reasoning) {
   const sessionId = session.id;
   const stream2 = parsed.stream === true;
   ++session.stats.requests;
   session.hostCreditTokens = 0;
+  const stripReasoning = (msgs) => withReasoningDrop(msgs, reasoning, log2, sessionId);
   if (reconcileNativeCompactionBoundary(session)) {
     log2("info", `[${sessionId}] reconciled ACP state after native Responses compact boundary`);
   }
@@ -59627,7 +59745,7 @@ function prepareResponses(parsed, req, opts, core, config, prompts, log2, sessio
     log2("info", diagTagSummary(turn.messages, sessionId, "text-only"));
     const willInjectNudge = opts.compress.injectNudge && !!turn.nudge && shouldInject && !isCompactionTrigger && (turn.nudge.shouldInject || emergencyNudge(turn.nudge));
     log2("info", diagNudge(turn, sessionId, tokenCount, config.modelContextLimit, parsed.model, willInjectNudge));
-    processedMessages = repairResponsesAssistantOrdering(stripKernelSummaries(turn.messages, turn.state), originalMessages);
+    processedMessages = repairResponsesAssistantOrdering(stripReasoning(stripKernelSummaries(turn.messages, turn.state)), originalMessages);
     reapOrphanBlocks(session, msgs, deactivateBlock);
     rebuiltInput = patchResponsesInput(projection, processedMessages);
     const forgedSummaries = echoReplaced ? [] : session.metadata.codexForgedSummaries ?? [];