nolo-cli 0.213.0 → 0.222.0-alpha.2

package/{chunk-7BT37FK3.js → chunk-RDH5CQ4T.js}

@@ -1,16 +1,32 @@
+import {
+  normalizeTimingMs
+} from "./chunk-H57T64RA.js";
+import {
+  isAbortError
+} from "./chunk-7DPNBQSV.js";
 import {
   ALL_MODELS
-} from "./chunk-T43B7LB2.js";
+} from "./chunk-WKGQSYTB.js";
+import {
+  asTrimmedNonEmptyStringArray
+} from "./chunk-I42VC5KR.js";
 import {
   estimateTokenCount
 } from "./chunk-PSCNXBYA.js";
 import {
   MODEL_LOOKUP_MAP
-} from "./chunk-SM3NSZJV.js";
+} from "./chunk-2UCBGYO4.js";
 import {
+  PLATFORM_HOSTED_DEEPSEEK_FLASH_MODEL,
   isPlatformManagedProvider,
   resolvePlatformEffectiveModel
-} from "./chunk-GOO6G4PY.js";
+} from "./chunk-CKG3PRZ7.js";
+import {
+  asOptionalFiniteNumber
+} from "./chunk-WWC7I6TA.js";
+import {
+  asOptionalTrimmedString
+} from "./chunk-G3SUCT4W.js";
 
 // packages/core/chat/messageContentSerialize.ts
 var serializeMessageContent = (content, imagePlaceholder = "[\u56FE\u7247]") => {
@@ -34,6 +50,277 @@ var serializeMessageContent = (content, imagePlaceholder = "[\u56FE\u7247]") =>
   return null;
 };
 
+// packages/ai/token/cacheTokenFields.ts
+var finiteTokenCount = (value) => {
+  if (typeof value !== "number" || !Number.isFinite(value)) return void 0;
+  return Math.max(0, Math.floor(value));
+};
+var readNested = (value, path) => {
+  let cursor = value;
+  for (const key of path) {
+    if (!cursor || typeof cursor !== "object") return void 0;
+    cursor = cursor[key];
+  }
+  return finiteTokenCount(cursor);
+};
+var firstPositive = (usage, readers) => {
+  for (const read of readers) {
+    const value = read(usage);
+    if (value !== void 0 && value !== 0) return value;
+  }
+  return 0;
+};
+var topLevel = (field) => (usage) => {
+  if (!usage || typeof usage !== "object") return void 0;
+  return finiteTokenCount(usage[field]);
+};
+var nested = (path) => (usage) => readNested(usage, path);
+var firstPresentAsNumber = (usage, readers) => {
+  let sawField = false;
+  for (const read of readers) {
+    const value = read(usage);
+    if (value === void 0) continue;
+    sawField = true;
+    if (value !== 0) return value;
+  }
+  return sawField ? 0 : null;
+};
+var CACHE_READ_READERS = [
+  topLevel("cache_read_input_tokens"),
+  topLevel("prompt_cache_hit_tokens"),
+  nested(["input_tokens_details", "cached_tokens"]),
+  nested(["prompt_tokens_details", "cached_tokens"])
+];
+var CACHE_CREATION_READERS = [
+  topLevel("cache_creation_input_tokens"),
+  topLevel("prompt_cache_miss_tokens"),
+  nested(["input_tokens_details", "cache_write_tokens"])
+];
+var readCacheReadInputTokens = (usage) => firstPositive(usage, CACHE_READ_READERS);
+var readCacheReadInputTokensOrNull = (usage) => firstPresentAsNumber(usage, CACHE_READ_READERS);
+var readCacheCreationInputTokens = (usage) => firstPositive(usage, CACHE_CREATION_READERS);
+
+// packages/ai/token/normalizeUsage.ts
+var normalizeStringArray = (value) => {
+  if (!Array.isArray(value)) return void 0;
+  const normalized = [...new Set(asTrimmedNonEmptyStringArray(value))];
+  return normalized.length > 0 ? normalized : void 0;
+};
+var finiteTokenCount2 = (value) => {
+  const finite = asOptionalFiniteNumber(value);
+  if (finite === void 0) return void 0;
+  return Math.max(0, Math.floor(finite));
+};
+var readFiniteNumberField = (usage, field) => {
+  if (!usage || typeof usage !== "object") return void 0;
+  if (!(field in usage)) return void 0;
+  const candidate = usage[field];
+  return asOptionalFiniteNumber(candidate);
+};
+var readOptionalOutputTokens = (usage) => {
+  if (!usage || typeof usage !== "object") return void 0;
+  const field = "output_tokens" in usage && usage.output_tokens != null ? "output_tokens" : "completion_tokens";
+  return readFiniteNumberField(usage, field);
+};
+var readCostInUsdTicks = (usage) => readFiniteNumberField(usage, "cost_in_usd_ticks");
+var normalizeUsage = (usage) => {
+  if (!usage || typeof usage !== "object") {
+    return {
+      input_tokens: 0,
+      output_tokens: 0,
+      total_tokens: 0,
+      cache_creation_input_tokens: 0,
+      cache_read_input_tokens: 0,
+      cost: 0
+    };
+  }
+  const inputTokens = "input_tokens" in usage ? usage.input_tokens ?? 0 : "prompt_tokens" in usage ? usage.prompt_tokens : 0;
+  const outputTokens = "output_tokens" in usage ? usage.output_tokens ?? 0 : "completion_tokens" in usage ? usage.completion_tokens : 0;
+  const cacheCreationInputTokens = readCacheCreationInputTokens(usage);
+  const cacheReadInputTokens = readCacheReadInputTokens(usage);
+  let cost = 0;
+  const providedCost = readFiniteNumberField(usage, "cost");
+  if (providedCost !== void 0) {
+    cost = providedCost;
+  }
+  const xaiTicks = readCostInUsdTicks(usage);
+  if (xaiTicks !== void 0 && providedCost === void 0) {
+    cost = xaiTicks / 1e10;
+  }
+  const billingProvider = asOptionalTrimmedString(usage.billing_provider);
+  const billingModel = asOptionalTrimmedString(usage.billing_model);
+  const billingServiceTier = asOptionalTrimmedString(usage.billing_service_tier);
+  const billingEstimated = "billing_estimated" in usage && usage.billing_estimated === true;
+  const serverBilled = "server_billed" in usage && usage.server_billed === true;
+  const providerCallId = asOptionalTrimmedString(usage.provider_call_id);
+  const imageGenerationCount = "image_generation_count" in usage ? asOptionalFiniteNumber(usage.image_generation_count) : void 0;
+  const providerResponseIds = normalizeStringArray(
+    usage.provider_response_ids
+  );
+  const providerRequestIds = normalizeStringArray(
+    usage.provider_request_ids
+  );
+  const firstOutputMs = normalizeTimingMs(
+    readFiniteNumberField(usage, "firstOutputMs") ?? readFiniteNumberField(usage, "firstTokenMs")
+  );
+  const callDurationMs = normalizeTimingMs(
+    readFiniteNumberField(usage, "callDurationMs") ?? readFiniteNumberField(usage, "durationMs")
+  );
+  return {
+    input_tokens: inputTokens,
+    output_tokens: outputTokens,
+    total_tokens: finiteTokenCount2(usage.total_tokens) ?? inputTokens + outputTokens,
+    cache_creation_input_tokens: cacheCreationInputTokens,
+    cache_read_input_tokens: cacheReadInputTokens,
+    cost,
+    ...typeof imageGenerationCount === "number" ? { image_generation_count: imageGenerationCount } : {},
+    ...providerResponseIds ? { provider_response_ids: providerResponseIds } : {},
+    ...providerRequestIds ? { provider_request_ids: providerRequestIds } : {},
+    ...billingProvider ? { billing_provider: billingProvider } : {},
+    ...billingModel ? { billing_model: billingModel } : {},
+    ...billingServiceTier ? { billing_service_tier: billingServiceTier } : {},
+    ...billingEstimated ? { billing_estimated: true } : {},
+    ...serverBilled ? { server_billed: true } : {},
+    ...providerCallId ? { provider_call_id: providerCallId } : {},
+    ...xaiTicks !== void 0 ? { cost_in_usd_ticks: xaiTicks } : {},
+    ...firstOutputMs !== void 0 ? { firstOutputMs } : {},
+    ...callDurationMs !== void 0 ? { callDurationMs } : {}
+  };
+};
+
+// packages/ai/context/summaryContract.ts
+var SUMMARY_TOTAL_BUDGET_MS = { automatic: 6e4, manual: 12e4 };
+var realClock = {
+  now: Date.now,
+  setTimeout: (callback, ms) => setTimeout(callback, ms),
+  clearTimeout: (handle) => clearTimeout(handle)
+};
+function createSummaryBudget(options) {
+  const clock = options.clock ?? realClock;
+  if (options.deadlineMs !== void 0 && !Number.isFinite(options.deadlineMs)) {
+    throw new RangeError("Summary deadline must be finite");
+  }
+  const deadlineMs = Math.min(clock.now() + SUMMARY_TOTAL_BUDGET_MS[options.mode], options.deadlineMs ?? Infinity);
+  const controller = new AbortController();
+  const cancel = () => controller.abort(new DOMException("Summary cancelled", "AbortError"));
+  const expire = () => controller.abort(new DOMException("Summary deadline exceeded", "TimeoutError"));
+  const remainingMs = () => Math.max(0, deadlineMs - clock.now());
+  const check = () => {
+    if (!controller.signal.aborted && remainingMs() === 0) expire();
+    if (controller.signal.aborted) throw controller.signal.reason;
+  };
+  let timer;
+  const cleanup = () => {
+    if (timer !== void 0) clock.clearTimeout(timer);
+    options.signal?.removeEventListener("abort", cancel);
+  };
+  controller.signal.addEventListener("abort", cleanup, { once: true });
+  options.signal?.addEventListener("abort", cancel, { once: true });
+  if (options.signal?.aborted) cancel();
+  else if (remainingMs() === 0) expire();
+  else timer = clock.setTimeout(expire, remainingMs());
+  const budget = {
+    deadlineMs,
+    signal: controller.signal,
+    remainingMs,
+    async run(operation) {
+      check();
+      let onAbort;
+      const stopped = new Promise((_, reject) => {
+        onAbort = () => reject(controller.signal.reason);
+        controller.signal.addEventListener("abort", onAbort, { once: true });
+      });
+      try {
+        const value = await Promise.race([Promise.resolve().then(() => {
+          check();
+          return operation({ deadlineMs, signal: controller.signal });
+        }), stopped]);
+        check();
+        return value;
+      } finally {
+        if (onAbort) controller.signal.removeEventListener("abort", onAbort);
+      }
+    },
+    dispose() {
+      cancel();
+      cleanup();
+    }
+  };
+  return budget;
+}
+function record(value) {
+  return value !== null && typeof value === "object" ? value : {};
+}
+function identifier(value) {
+  return typeof value === "string" && value.length <= 128 && /^[a-zA-Z0-9][a-zA-Z0-9._:/-]*$/.test(value) && !/https?:|^bearer|^sk-|^vault-cred|^nolo_cred/i.test(value) ? value : void 0;
+}
+function classifySummaryError(error, context = {}) {
+  const top = record(error);
+  const nested2 = record(top.error);
+  const rawStatus = top.status ?? top.statusCode ?? nested2.status;
+  const status = typeof rawStatus === "number" && Number.isInteger(rawStatus) && rawStatus >= 100 && rawStatus <= 599 ? rawStatus : void 0;
+  const code = identifier(top.code ?? nested2.code);
+  const errorType = identifier(top.type ?? nested2.type);
+  const balanceSource = errorType === "insufficient_balance_error" || top.balanceSource === "upstream" ? "upstream" : code === "INSUFFICIENT_BALANCE" ? "platform-user" : void 0;
+  const category = top.name === "TimeoutError" || code === "ETIMEDOUT" || code === "DEADLINE_EXCEEDED" ? "timeout" : isAbortError(error) ? "cancelled" : status !== void 0 ? "http" : code ? "provider" : "unknown";
+  return {
+    category,
+    message: `Summary ${category} failure`,
+    ...status === void 0 ? {} : { status },
+    ...code === void 0 ? {} : { code },
+    ...errorType === void 0 ? {} : { errorType },
+    ...balanceSource === void 0 ? {} : { balanceSource },
+    model: identifier(top.model) ?? identifier(context.model),
+    provider: identifier(top.provider) ?? identifier(context.provider),
+    attemptId: identifier(top.attemptId) ?? identifier(context.attemptId),
+    stage: identifier(top.stage) ?? identifier(context.stage) ?? "attempt",
+    ...Object.fromEntries(["finishReason", "responseStatus", "incompleteReason"].flatMap((key) => {
+      const value = identifier(top[key]);
+      return value === void 0 ? [] : [[key, value]];
+    })),
+    ...Object.fromEntries(["maxOutputTokens", "completionTokens", "outputChars"].flatMap((key) => typeof top[key] === "number" && Number.isSafeInteger(top[key]) && top[key] >= 0 ? [[key, top[key]]] : []))
+  };
+}
+function isTruncatedCompletion(completion) {
+  return completion.truncated === true || completion.finishReason === "length" || completion.finishReason === "max_tokens" || completion.responseStatus === "incomplete" || completion.incompleteReason !== void 0;
+}
+function summaryCompletionTokens(usage) {
+  const tokens = readOptionalOutputTokens(usage);
+  return typeof tokens === "number" && Number.isSafeInteger(tokens) && tokens >= 0 ? tokens : void 0;
+}
+
+// packages/ai/context/summaryRouting.ts
+function selectSummaryRoute(input) {
+  return input.authenticated && input.platformAvailable && !!input.userId && input.userId !== "local" && input.userId !== "guest" && typeof input.credits === "number" && Number.isFinite(input.credits) && input.credits > 0 ? { kind: "platform", model: PLATFORM_HOSTED_DEEPSEEK_FLASH_MODEL } : { kind: "current" };
+}
+async function runSummaryWithContract(args) {
+  const attempts = [];
+  const execute = async (route) => {
+    const attempt = args[route];
+    try {
+      const completion = await args.budget.run((context) => attempt.run(context));
+      if (typeof completion?.text !== "string") {
+        throw { code: "INVALID_SUMMARY_COMPLETION", stage: "decode" };
+      }
+      const text = completion.text.trim();
+      const kind = isTruncatedCompletion(completion) ? "truncated" : text ? "success" : "empty";
+      attempts.push({ route, outcome: kind });
+      if (kind === "empty") return { kind, route, attempts, completion };
+      return { kind, text, route, attempts, completion };
+    } catch (cause) {
+      const error = classifySummaryError(args.budget.signal.aborted ? args.budget.signal.reason : cause, attempt);
+      const kind = error.category === "cancelled" ? "cancelled" : "error";
+      attempts.push({ route, outcome: kind, error });
+      return { kind, error, route, attempts };
+    }
+  };
+  if (selectSummaryRoute(args.eligibility).kind === "platform") {
+    const result = await execute("platform");
+    if (result.kind !== "error" || result.error.code !== "INSUFFICIENT_BALANCE" || result.error.balanceSource !== "platform-user" || result.error.category === "timeout" || args.budget.signal.aborted) return result;
+  }
+  return execute("current");
+}
+
 // packages/core/sha256.ts
 import SHA256 from "crypto-js/sha256.js";
 function sha256Hex(text) {
@@ -44,32 +331,14 @@ function sha256Hex(text) {
 var TOOL_OUTPUT_CAP_WINDOW_RATIO = 0.12;
 var TOOL_OUTPUT_CAP_MIN_TOKENS = 4e3;
 var TOOL_OUTPUT_CAP_MAX_TOKENS = 5e4;
-var COMPRESSION_TRIGGER_RATIO_MAX = 0.78;
-var COMPRESSION_TRIGGER_RATIO_MIN = 0.5;
-var ROUND_OVERHEAD_WINDOW_RATIO = 0.08;
-var ROUND_OVERHEAD_MIN_TOKENS = 4e3;
-var ROUND_OVERHEAD_MAX_TOKENS = 24e3;
+var COMPRESSION_TRIGGER_RATIO_MAX = 0.9;
 var clamp = (v, min, max) => Math.min(max, Math.max(min, v));
 var capToolOutputTokens = (contextWindow) => clamp(
   Math.floor(contextWindow * TOOL_OUTPUT_CAP_WINDOW_RATIO),
   TOOL_OUTPUT_CAP_MIN_TOKENS,
   TOOL_OUTPUT_CAP_MAX_TOKENS
 );
-var resolveCompressionTriggerRatio = (contextWindow) => {
-  if (typeof contextWindow !== "number" || !Number.isFinite(contextWindow) || contextWindow <= 0) {
-    return COMPRESSION_TRIGGER_RATIO_MAX;
-  }
-  const headroom = capToolOutputTokens(contextWindow) + clamp(
-    Math.floor(contextWindow * ROUND_OVERHEAD_WINDOW_RATIO),
-    ROUND_OVERHEAD_MIN_TOKENS,
-    ROUND_OVERHEAD_MAX_TOKENS
-  );
-  return clamp(
-    1 - headroom / contextWindow,
-    COMPRESSION_TRIGGER_RATIO_MIN,
-    COMPRESSION_TRIGGER_RATIO_MAX
-  );
-};
+var resolveCompressionTriggerRatio = (_contextWindow) => COMPRESSION_TRIGGER_RATIO_MAX;
 var truncateToolOutputForContext = (text, contextWindow) => {
   if (!text) return text;
   const capTokens = capToolOutputTokens(contextWindow);
@@ -96,8 +365,10 @@ var truncateToolOutputForContext = (text, contextWindow) => {
 
 // packages/ai/context/planCompression.ts
 var MIN_COMPRESS_COUNT = 5;
+var INPUT_ITEM_TRIGGER_RATIO = 0.9;
+var INPUT_ITEM_RETENTION_RATIO = 0.6;
 var ACTIVE_SUMMARY_TAIL_KEEP_COUNT = 2;
-var COMPACTION_RETENTION_CUSHION_RATIO = 0.18;
+var RECENT_ORIGINAL_TOKEN_BUDGET = 2e4;
 var MIN_TRIGGER_PENDING_COUNT = MIN_COMPRESS_COUNT + ACTIVE_SUMMARY_TAIL_KEEP_COUNT;
 var getMessageTokenCount = (msg) => {
   const content = serializeMessageContent(msg.content) || "";
@@ -151,7 +422,7 @@ function shouldTriggerCompaction(args) {
   const usageRatio = normalizeContextUsageRatio(realContextUsagePercent);
   const triggeredByRealUsage = usageRatio !== void 0 && usageRatio >= triggerRatio;
   const triggeredByEstimate = usageRatio === void 0 && totalUsed >= historyBudget;
-  const shouldRunActiveSummary = force && reason === "manual";
+  const shouldRunActiveSummary = force && (reason === "manual" || reason === "budget_rescue");
   const triggeredByColdResume = false;
   return {
     trigger: triggeredByRealUsage || triggeredByEstimate || shouldRunActiveSummary || triggeredByColdResume,
@@ -172,28 +443,17 @@ function calculateCompressCount(pendingMsgs, rawMessageBudget) {
   return pendingMsgs.length - keepCount;
 }
 function resolveEffectiveCompactionBudgets(args) {
-  const {
-    contextWindow,
-    triggerRatio,
-    historyBudget,
-    rawMessageBudget,
-    summaryTokens,
-    cushionRatio = COMPACTION_RETENTION_CUSHION_RATIO
-  } = args;
-  const maxRetainedRatio = Math.max(
-    0.2,
-    Number((triggerRatio - cushionRatio).toFixed(4))
-  );
-  const maxRetainedHistoryBudget = Math.floor(contextWindow * maxRetainedRatio);
-  const effectiveHistoryBudget = Math.min(historyBudget, maxRetainedHistoryBudget);
-  const effectiveRawMessageBudget = Math.min(
-    rawMessageBudget,
-    Math.max(0, effectiveHistoryBudget - summaryTokens)
-  );
+  const effectiveHistoryBudget = Math.max(0, Math.min(
+    Math.floor(args.contextWindow),
+    Math.floor(args.historyBudget)
+  ));
+  const newSummaryTokenReserve = args.newSummaryTokenReserve ?? Math.min(8192, Math.max(1, Math.floor(args.contextWindow / 4)));
   return {
     effectiveHistoryBudget,
-    effectiveRawMessageBudget,
-    maxRetainedRatio
+    effectiveRawMessageBudget: Math.max(0, Math.min(
+      RECENT_ORIGINAL_TOKEN_BUDGET,
+      effectiveHistoryBudget - Math.max(0, newSummaryTokenReserve)
+    ))
   };
 }
 function guardToolChainBoundary(pendingMsgs, compressCount, preserveByProjection) {
@@ -212,6 +472,10 @@ function guardToolChainBoundary(pendingMsgs, compressCount, preserveByProjection
   }
   return count;
 }
+function calculateItemCompressCount(pendingMsgs, maxInputItems, preserveByProjection) {
+  const keepCount = Math.max(0, Math.floor(maxInputItems * INPUT_ITEM_RETENTION_RATIO) - 2 - (preserveByProjection ? 1 : 0));
+  return Math.max(0, pendingMsgs.length - keepCount);
+}
 function planCompression(input) {
   const {
     allMsgs,
@@ -249,6 +513,9 @@ function planCompression(input) {
       }
     }
   }
+  const maxInputItems = typeof input.maxInputItems === "number" && Number.isFinite(input.maxInputItems) && input.maxInputItems > 0 ? input.maxInputItems : void 0;
+  const itemCount = pendingMsgs.length + (summary ? 1 : 0) + 1 + (coveredUserTokens > 0 ? 1 : 0);
+  const triggeredByItemLimit = maxInputItems !== void 0 && itemCount >= maxInputItems * INPUT_ITEM_TRIGGER_RATIO;
   const totalUsed = summaryTokens + coveredUserTokens + pendingTokens;
   const triggerRatio = resolveCompressionTriggerRatio(contextWindow);
   const historyBudget = Math.floor(contextWindow * triggerRatio);
@@ -268,6 +535,7 @@ function planCompression(input) {
   });
   const diagnostics = {
     pendingMsgCount: pendingMsgs.length,
+    ...maxInputItems !== void 0 ? { itemCount, maxInputItems, triggeredByItemLimit } : {},
     totalUsed,
     historyBudget,
     triggerRatio,
@@ -277,7 +545,7 @@ function planCompression(input) {
     triggeredByColdResume,
     shouldRunActiveSummary
   };
-  if (!trigger) {
+  if (!trigger && !triggeredByItemLimit) {
     return emptyPlan(startIndex, {
       ...diagnostics,
       emptyReason: "not-triggered"
@@ -286,17 +554,26 @@ function planCompression(input) {
   const { effectiveRawMessageBudget } = resolveEffectiveCompactionBudgets({
     contextWindow,
     triggerRatio,
-    historyBudget,
+    historyBudget: Math.min(historyBudget, input.availableHistoryTokens ?? historyBudget),
     rawMessageBudget,
-    summaryTokens: Math.max(summaryTokens, Math.min(1e3, Math.floor(contextWindow * 0.1)))
+    summaryTokens,
+    newSummaryTokenReserve: input.newSummaryTokenReserve
   });
   const minimumPrefixTokens = Math.min(1e3, Math.floor(contextWindow * 0.1));
-  const tailBudget = Math.min(effectiveRawMessageBudget, Math.max(0, pendingTokens - minimumPrefixTokens));
-  let compressCount = guardToolChainBoundary(
-    pendingMsgs,
-    calculateCompressCount(pendingMsgs, tailBudget),
-    preserveLatestUserByProjection
-  );
+  const tailBudget = triggeredByItemLimit ? effectiveRawMessageBudget : Math.min(effectiveRawMessageBudget, Math.max(0, pendingTokens - minimumPrefixTokens));
+  const itemCompressCount = maxInputItems !== void 0 ? calculateItemCompressCount(pendingMsgs, maxInputItems, preserveLatestUserByProjection) : 0;
+  const guardedCut = (tokenCompressCount) => {
+    let candidate = Math.max(tokenCompressCount, itemCompressCount);
+    let guarded = guardToolChainBoundary(pendingMsgs, candidate, preserveLatestUserByProjection);
+    while (guarded < itemCompressCount && candidate < pendingMsgs.length - 1) {
+      candidate++;
+      const next = guardToolChainBoundary(pendingMsgs, candidate, preserveLatestUserByProjection);
+      if (next === guarded && pendingMsgs[next]?.role === "user" && !preserveLatestUserByProjection) break;
+      guarded = next;
+    }
+    return guarded;
+  };
+  let compressCount = guardedCut(calculateCompressCount(pendingMsgs, tailBudget));
   if (preserveLatestUserByProjection) {
     let latestUser = -1;
     for (let i = pendingMsgs.length - 1; i >= 0; i--) {
@@ -307,10 +584,8 @@ function planCompression(input) {
     }
     const preservedTokens = latestUser >= 0 && latestUser < compressCount ? getMessageTokenCount(pendingMsgs[latestUser]) : latestUser < 0 ? coveredUserTokens : 0;
     if (preservedTokens > 0) {
-      compressCount = guardToolChainBoundary(
-        pendingMsgs,
-        calculateCompressCount(pendingMsgs, Math.max(0, tailBudget - preservedTokens)),
-        true
+      compressCount = guardedCut(
+        calculateCompressCount(pendingMsgs, Math.max(0, tailBudget - preservedTokens))
       );
     }
   }
@@ -336,7 +611,12 @@ function planCompression(input) {
     msgsToKeep,
     newSummarizedBeforeId,
     startIndex,
-    diagnostics: { ...diagnostics, compressCount }
+    diagnostics: {
+      ...diagnostics,
+      compressCount,
+      recentOriginalTokenBudget: tailBudget,
+      retainedTailOverBudget: msgsToKeep.reduce((sum, msg) => sum + getMessageTokenCount(msg), 0) > tailBudget
+    }
   };
 }
 
@@ -368,7 +648,7 @@ var COMPACTION_SUMMARY_SYSTEM_PROMPT = `\u4F60\u662F\u5BF9\u8BDD\u4E0A\u4E0B\u65
    - \u300C\u5361\u70B9\u4E0E\u9700\u7528\u6237\u51B3\u5B9A\u300D\uFF1A\u963B\u585E\u9879\u5361\u5728\u54EA\u3001\u9700\u8981\u7528\u6237\u63D0\u4F9B\u4EC0\u4E48\uFF08\u6388\u6743 / \u51ED\u636E / \u51B3\u5B9A\uFF09\uFF0C\u5199\u6E05\u695A\uFF0C\u4E0D\u5F97\u542B\u7CCA\u3002
    - \u300C\u8F93\u5165\u672A\u8986\u76D6\u300D\uFF1A\u538B\u7F29\u8303\u56F4\u5185\u67E5\u4E0D\u5230\u7684\u5173\u952E\u4FE1\u606F\uFF08\u5982\u521D\u59CB\u539F\u8BDD\u3001\u6307\u5B9A id\uFF09\uFF0C\u5FC5\u987B\u5199"\u8F93\u5165\u4E2D\u672A\u89C1"\uFF0C\u7981\u6B62\u8111\u8865\u3002
 3) \u5BF9\u8BDD\u8FDB\u5C55\u4E0E\u5F85\u529E\u4EE5\u7B2C\u4E00\u4EBA\u79F0\u7EED\u4F5C\u7B14\u8BB0\u7684\u53E3\u543B\u5199\uFF08"\u6211\u505A\u4E86\u2026""\u7528\u6237\u8981\u6C42\u2026""\u4E0B\u4E00\u6B65\u6211\u9700\u8981\u2026"\uFF09\uFF0C\u4E0D\u8981\u5199\u6210\u7B2C\u4E09\u65B9\u89C2\u5BDF\u62A5\u544A\u3002\u5148\u6781\u7B80\u6982\u62EC\u65E7\u4E0A\u4E0B\u6587\uFF0C\u518D\u66F4\u8BE6\u7EC6\u8BB0\u5F55\u6700\u8FD1\u8FDB\u5C55\u3001\u7ED3\u8BBA\u3001\u5206\u6B67\u4E0E\u4E0B\u4E00\u6B65\u3002\u5BF9\u4E8E\u58F0\u79F0\u5DF2\u5B8C\u6210\u4F46\u5C1A\u672A\u9A8C\u8BC1\u7684\u5DE5\u4F5C\uFF08\u5982\u5DE5\u5177\u8C03\u7528\u8FD4\u56DE\u6210\u529F\u4F46\u7ED3\u679C\u672A\u4EBA\u5DE5\u786E\u8BA4\uFF09\uFF0C\u660E\u786E\u6807\u6CE8"\uFF08\u5F85\u9A8C\u8BC1\uFF09"\u2014\u2014\u4E0D\u8981\u9ED8\u8BA4\u4FE1\u4EFB\u5DE5\u5177\u8FD4\u56DE\u503C\u3002
-4) \u6587\u4EF6\u64CD\u4F5C\u6E05\u5355\uFF1A\u5217\u51FA\u672C\u6B21\u538B\u7F29\u8303\u56F4\u5185\u8BFB\u53D6\u3001\u5199\u5165\u3001\u7F16\u8F91\u8FC7\u7684\u6587\u4EF6\u8DEF\u5F84\u53CA\u64CD\u4F5C\u7C7B\u578B\uFF08\u5982: - \u8BFB\u53D6: path/to/file\uFF1B\u82E5\u672A\u6D89\u53CA\u6587\u4EF6\u64CD\u4F5C\u5199"\u65E0"\uFF09\u3002
+4) \u6587\u4EF6\u64CD\u4F5C\u6E05\u5355\uFF1A\u6309\u8BFB\u53D6\u3001\u5199\u5165\u3001\u7F16\u8F91\u5217\u51FA\u8DEF\u5F84\uFF08\u5982: - \u8BFB\u53D6: path/to/file\uFF1B\u672A\u6D89\u53CA\u5219\u5199"\u65E0"\uFF09\u3002\u6709\u957F\u5EA6\u7EA6\u675F\u65F6\u5408\u5E76\u65E7\u64CD\u4F5C\uFF0C\u7701\u7565\u65E0\u5173\u8BFB\u53D6\u3002
 5) \u5FFD\u7565\u5BD2\u6684\u3001\u91CD\u590D\u5C1D\u8BD5\u548C\u65E0\u4EF7\u503C\u5E9F\u8BDD\uFF1B**\u5DF2\u653E\u5F03\u7684\u65B9\u6848/\u5C1D\u8BD5**\u53EF\u4EE5\u7701\u7565\uFF0C\u4F46\u7528\u6237\u5F53\u521D\u7684\u610F\u56FE\u4E0E\u7406\u7531\u4E0D\u53EF\u7701\u7565\uFF08\u89C1\u7B2C 2 \u6761\uFF09\u3002
 6) \u4E0D\u8981\u7F16\u9020\u672A\u51FA\u73B0\u7684\u4FE1\u606F\uFF1B\u4E0D\u8981\u5F00\u573A\u767D\u3001\u7ED3\u675F\u8BED\u3001markdown \u4EE3\u7801\u5757\u6216\u989D\u5916\u7AE0\u8282\uFF1B\u4E0D\u8981\u8C03\u7528\u4EFB\u4F55\u5DE5\u5177\u3002`;
 function formatMessageContentFallback(msg) {
@@ -380,8 +660,16 @@ function formatMessageContentFallback(msg) {
   return "[\u975E\u6587\u672C\u5185\u5BB9]";
 }
 function buildCompactionUserContent(args) {
-  const parts = [`\u3010\u73B0\u6709\u8BB0\u5FC6\u3011\uFF1A
-${args.previousSummary || "(\u65E0)"}`];
+  const parts = [];
+  if (args.outputTokenBudget !== void 0) {
+    const budget = args.outputTokenBudget;
+    if (!Number.isFinite(budget) || budget < 1) throw new RangeError("Invalid summary output budget");
+    const limit = Math.floor(budget);
+    const target = Math.max(1, Math.min(2e3, Math.floor(limit / 2)));
+    parts.push(`Summary target: ${target} tokens; hard output limit: ${limit} tokens. Complete 3 sections; merge old progress/files.`);
+  }
+  parts.push(`\u3010\u73B0\u6709\u8BB0\u5FC6\u3011\uFF1A
+${args.previousSummary || "(\u65E0)"}`);
   if (args.fileOpsText !== void 0) {
     parts.push(`\u3010\u6587\u4EF6\u64CD\u4F5C\u6E05\u5355\u3011\uFF1A
 ${args.fileOpsText || "\u65E0"}`);
@@ -449,17 +737,46 @@ function makeBoundedExcerpt(content) {
     return content.slice(0, headLength) + marker + (selected.length ? selected.join("\n") + "\n" : "") + (tailLength ? content.slice(-tailLength) : "");
   };
 }
+var OMITTED_ROLE_ORDER = ["user", "assistant", "tool", "system"];
+var MIN_INLINE_CONTENT_TOKENS = 12;
+function formatOmittedRun(start, end, counts) {
+  const ordered = [
+    ...OMITTED_ROLE_ORDER,
+    ...[...counts.keys()].filter((role) => !OMITTED_ROLE_ORDER.includes(role))
+  ];
+  const parts = ordered.filter((role) => (counts.get(role) ?? 0) > 0).map((role) => `${role} ${counts.get(role)}`);
+  const range = start === end ? `\u6D88\u606F ${start + 1}` : `\u6D88\u606F ${start + 1}\u2013${end + 1}`;
+  return `[... ${range} \u5DF2\u7701\u7565\uFF1A${parts.join("\u3001")} ...]`;
+}
+var SUMMARY_HEAD_SHARE = 0.4;
+var DECLARATION_RESERVE = 40;
+var USER_MESSAGE_MAX_DEMAND = 600;
+var SUMMARY_USER_SHARE = 0.2;
+var DECLARATION_SPLIT_COST = 20;
 function buildBoundedCompactionRequest(args) {
   const { msgsToCompress: msgs, previousSummary, contextWindow } = args;
-  if (!Number.isFinite(contextWindow) || contextWindow <= 0) return { status: "blocked", reason: "invalid-window" };
-  if (!msgs.length) return { status: "blocked", reason: "empty-input" };
-  const window = Math.floor(contextWindow);
-  const outputTokenBudget = Math.min(8192, Math.max(1, Math.floor(window / 4)));
-  const cap = Math.min(computeSummaryInputTokenCap(contextWindow), window - outputTokenBudget);
+  const windowValid = Number.isFinite(contextWindow) && contextWindow > 0;
+  const window = windowValid ? Math.floor(contextWindow) : 0;
+  const outputTokenBudget = windowValid ? Math.min(8192, Math.max(1, Math.floor(window / 4))) : 0;
+  const cap = windowValid ? Math.min(computeSummaryInputTokenCap(contextWindow), window - outputTokenBudget) : 0;
+  const blocked = (reason, extra = {}) => ({
+    status: "blocked",
+    reason,
+    diagnostics: {
+      contextWindow: window,
+      cap,
+      outputTokenBudget,
+      previousSummaryTokens: estimateTokenCount(previousSummary ?? ""),
+      messageCount: msgs.length,
+      ...extra
+    }
+  });
+  if (!windowValid) return blocked("invalid-window");
+  if (!msgs.length) return blocked("empty-input");
   const systemTokens = estimateTokenCount(COMPACTION_SUMMARY_SYSTEM_PROMPT);
   for (let i = 0; i < msgs.length; ) {
     const msg = msgs[i];
-    if (msg.role === "tool") return { status: "blocked", reason: "no-safe-group" };
+    if (msg.role === "tool") return blocked("no-safe-group");
     const calls = msg.role === "assistant" && Array.isArray(msg.tool_calls) ? msg.tool_calls : [];
     if (!calls.length) {
       i++;
@@ -468,20 +785,20 @@ function buildBoundedCompactionRequest(args) {
     const ids = calls.map((call) => call.id);
     const hasIds = ids.every((id) => typeof id === "string" && id.length > 0);
     if (!hasIds && ids.some((id) => id !== void 0) || hasIds && new Set(ids).size !== ids.length) {
-      return { status: "blocked", reason: "no-safe-group" };
+      return blocked("no-safe-group");
     }
-    const remaining2 = new Set(ids);
+    const remaining = new Set(ids);
     let count = 0;
     i++;
     while (i < msgs.length && msgs[i].role === "tool") {
       const id = msgs[i].tool_call_id;
-      if (hasIds ? !id || !remaining2.delete(id) : id !== void 0) {
-        return { status: "blocked", reason: "no-safe-group" };
+      if (hasIds ? !id || !remaining.delete(id) : id !== void 0) {
+        return blocked("no-safe-group");
       }
       count++;
       i++;
     }
-    if (count !== calls.length || hasIds && remaining2.size) return { status: "blocked", reason: "no-safe-group" };
+    if (count !== calls.length || hasIds && remaining.size) return blocked("no-safe-group");
   }
   const canonicalize = args.canonicalizeToolName ?? ((name) => name);
   const operations = /* @__PURE__ */ new Map();
@@ -493,44 +810,45 @@ function buildBoundedCompactionRequest(args) {
   }
   const labelsByOperation = { read: "\u8BFB\u53D6", write: "\u5199\u5165", edit: "\u7F16\u8F91" };
   const fileOpsText = [...operations.values()].map((op) => `- ${labelsByOperation[op.type]}: ${op.path}`).join("\n") || "\u65E0";
-  const note = "\u4EE5\u4E0B\u662F\u6709\u635F\u6295\u5F71\uFF0C\u90E8\u5206\u6D88\u606F\u53EF\u80FD\u88AB\u622A\u53D6\uFF1B\u539F\u6587\u4ECD\u5B58\u5386\u53F2\u3002";
+  const note = "\u4EE5\u4E0B\u662F\u6709\u635F\u6295\u5F71\uFF1A\u90E8\u5206\u6D88\u606F\u88AB\u622A\u53D6\uFF0C\u8FDE\u7EED\u7701\u7565\u7684\u533A\u95F4\u5408\u5E76\u4E3A\u4E00\u884C [...] \u58F0\u660E\uFF1B\u539F\u6587\u4ECD\u5B58\u5386\u53F2\u3002";
   const omitted = "[... \u5185\u5BB9\u5DF2\u622A\u53D6 ...]";
   const contents = msgs.map((msg) => serializeMessageContent(msg.content) || formatMessageContentFallback(msg));
-  const excerpts = contents.map((content, i) => msgs[i].role === "user" ? void 0 : makeBoundedExcerpt(content));
+  const excerpts = contents.map((content) => makeBoundedExcerpt(content));
   const raw = contents.map((content, i) => msgs[i].role === "tool" ? excerpts[i](TOOL_RESULT_TRUNCATE_CHARS) : content);
-  const labels = msgs.map((msg) => `${msg.role} [id=${JSON.stringify(msg.id)}]: `);
-  const projected = msgs.map((msg, i) => msg.role === "user" ? raw[i] : omitted);
-  const makePrompt = () => buildCompactionUserContent({
+  const promptFor = (messagesText) => buildCompactionUserContent({
     previousSummary,
     fileOpsText,
-    messagesText: `${note}
-${labels.map((label, i) => label + projected[i]).join("\n")}`
+    outputTokenBudget,
+    messagesText
   });
-  const base = makePrompt();
-  const baseTokens = systemTokens + estimateTokenCount(base);
-  if (baseTokens > cap) return { status: "blocked", reason: "no-budget" };
-  const remaining = cap - baseTokens;
-  const demands = raw.map((text, i) => msgs[i].role === "user" ? 0 : estimateTokenCount(text));
-  const weights = msgs.map((msg) => msg.role === "assistant" ? 2 : 1);
-  let low = 0;
-  let high = remaining + 1;
-  while (low + 1 < high) {
-    const mid = Math.floor((low + high) / 2);
-    const used = demands.reduce((sum, demand, i) => sum + Math.min(demand, mid * weights[i]), 0);
-    if (used <= remaining) low = mid;
-    else high = mid;
-  }
-  let extra = remaining - demands.reduce((sum, demand, i) => sum + Math.min(demand, low * weights[i]), 0);
-  for (let i = 0; i < msgs.length; i++) {
-    if (msgs[i].role === "user") continue;
-    let allowance = Math.min(demands[i], low * weights[i]);
-    const bonus = Math.min(extra, demands[i] - allowance);
-    allowance += bonus;
-    extra -= bonus;
-    if (demands[i] <= allowance) {
-      projected[i] = raw[i];
-      continue;
-    }
+  const demands = raw.map((text) => estimateTokenCount(text));
+  const lineOverheads = msgs.map((msg) => estimateTokenCount(`${msg.role}: `) + 1);
+  const fixedTokens = systemTokens + estimateTokenCount(promptFor(`${note}
+`));
+  const availableTokens = cap - fixedTokens;
+  const diagnosticsOf = (expandedMessages, omittedMessages) => {
+    const previousSummaryTokens = estimateTokenCount(previousSummary ?? "");
+    const baseTokens = systemTokens + estimateTokenCount(promptFor(""));
+    return {
+      contextWindow: window,
+      cap,
+      outputTokenBudget,
+      previousSummaryTokens,
+      messageCount: msgs.length,
+      systemTokens,
+      templateTokens: fixedTokens - systemTokens - previousSummaryTokens,
+      availableTokens,
+      expandedMessages,
+      omittedMessages,
+      baseTokens,
+      fileOpsCount: operations.size
+    };
+  };
+  if (availableTokens <= 0) {
+    return blocked("no-budget", diagnosticsOf(0, msgs.length));
+  }
+  const projectContent = (i, allowance) => {
+    if (demands[i] <= allowance) return raw[i];
     let lo = 0;
     let hi = Math.min(contents[i].length, msgs[i].role === "tool" ? TOOL_RESULT_TRUNCATE_CHARS : contents[i].length);
     let best = omitted;
@@ -544,19 +862,175 @@ ${labels.map((label, i) => label + projected[i]).join("\n")}`
         hi = mid - 1;
       }
     }
-    projected[i] = best;
+    return best;
+  };
+  const selectRegions = (budget2) => {
+    const expanded = msgs.map(() => false);
+    const allowances = msgs.map(() => 0);
+    let used = 0;
+    const declarationReserve = Math.min(DECLARATION_RESERVE, Math.max(4, Math.floor(budget2 / 4)));
+    const room = () => budget2 - used - declarationReserve;
+    const takeFull = (i) => {
+      expanded[i] = true;
+      allowances[i] = demands[i];
+      used += lineOverheads[i] + demands[i];
+    };
+    const drop = (i) => {
+      if (!expanded[i]) return;
+      expanded[i] = false;
+      used -= lineOverheads[i] + allowances[i];
+      allowances[i] = 0;
+    };
+    const headBudget = Math.floor(budget2 * SUMMARY_HEAD_SHARE);
+    let headEnd = -1;
+    let headUsed = 0;
+    for (let i = 0; i < msgs.length; i++) {
+      const cost = lineOverheads[i] + demands[i];
+      const limit = i === 0 ? Math.max(headBudget, room()) : headBudget - headUsed;
+      if (cost > limit) break;
+      takeFull(i);
+      headUsed += cost;
+      headEnd = i;
+    }
+    const userBudget = Math.floor(budget2 * SUMMARY_USER_SHARE);
+    let userSpend = 0;
+    for (let i = msgs.length - 1; i > headEnd; i--) {
+      if (expanded[i] || msgs[i].role !== "user") continue;
+      if (demands[i] > USER_MESSAGE_MAX_DEMAND) continue;
+      const cost = lineOverheads[i] + demands[i] + DECLARATION_SPLIT_COST;
+      if (userSpend + cost > userBudget || cost > room()) continue;
+      takeFull(i);
+      userSpend += cost;
+    }
+    let tailStart = msgs.length;
+    for (let i = msgs.length - 1; i > headEnd; i--) {
+      if (expanded[i]) {
+        tailStart = i;
+        continue;
+      }
+      if (lineOverheads[i] + demands[i] > room()) break;
+      takeFull(i);
+      tailStart = i;
+    }
+    const boundaryCap = Math.floor(room() * 0.5);
+    const takeBoundary = (i) => {
+      const inline = Math.min(room(), boundaryCap) - lineOverheads[i];
+      if (inline < MIN_INLINE_CONTENT_TOKENS) return false;
+      expanded[i] = true;
+      allowances[i] = inline;
+      used += lineOverheads[i] + inline;
+      return true;
+    };
+    if (tailStart - 1 > headEnd && !expanded[tailStart - 1] && takeBoundary(tailStart - 1)) tailStart--;
+    if (headEnd + 1 < tailStart && !expanded[headEnd + 1] && takeBoundary(headEnd + 1)) headEnd++;
+    let grew = true;
+    while (grew) {
+      grew = false;
+      for (let i = tailStart - 1; i > headEnd; i--) {
+        if (expanded[i]) {
+          tailStart = i;
+          continue;
+        }
+        if (lineOverheads[i] + demands[i] > room()) break;
+        takeFull(i);
+        tailStart = i;
+        grew = true;
+        break;
+      }
+      for (let i = headEnd + 1; i < tailStart; i++) {
+        if (expanded[i]) {
+          headEnd = i;
+          continue;
+        }
+        if (lineOverheads[i] + demands[i] > room()) break;
+        takeFull(i);
+        headEnd = i;
+        grew = true;
+        break;
+      }
+    }
+    const callOfResult = /* @__PURE__ */ new Map();
+    const resultsOfCall = /* @__PURE__ */ new Map();
+    for (let i = 0; i < msgs.length; i++) {
+      const msg = msgs[i];
+      const calls = msg.role === "assistant" && Array.isArray(msg.tool_calls) ? msg.tool_calls : [];
+      if (!calls.length) continue;
+      const ids = calls.map((call) => call.id);
+      const byId = ids.every((id) => typeof id === "string" && id.length > 0);
+      const results = [];
+      for (let j = i + 1; j < msgs.length && msgs[j].role === "tool"; j++) {
+        const id = msgs[j].tool_call_id;
+        if (!(byId ? typeof id === "string" && ids.includes(id) : id === void 0)) break;
+        results.push(j);
+        callOfResult.set(j, i);
+      }
+      resultsOfCall.set(i, results);
+    }
+    for (const [callIndex, results] of resultsOfCall) {
+      if (!expanded[callIndex] || !results.length) continue;
+      if (results.every((j) => expanded[j])) continue;
+      drop(callIndex);
+    }
+    for (let i = 0; i < msgs.length; i++) {
+      if (!expanded[i] || msgs[i].role !== "tool") continue;
+      const callIndex = callOfResult.get(i);
+      if (callIndex === void 0 || !expanded[callIndex]) drop(i);
+    }
+    return { expanded, allowances };
+  };
+  const render = (budget2) => {
+    const { expanded, allowances } = selectRegions(budget2);
+    const pieces = [];
+    let runStart = -1;
+    let runCounts = /* @__PURE__ */ new Map();
+    const flushRun = (end) => {
+      if (runStart < 0) return;
+      pieces.push(formatOmittedRun(runStart, end, runCounts));
+      runStart = -1;
+      runCounts = /* @__PURE__ */ new Map();
+    };
+    let expandedCount = 0;
+    for (let i = 0; i < msgs.length; i++) {
+      if (!expanded[i]) {
+        if (runStart < 0) runStart = i;
+        runCounts.set(msgs[i].role, (runCounts.get(msgs[i].role) ?? 0) + 1);
+        continue;
+      }
+      flushRun(i - 1);
+      expandedCount++;
+      pieces.push(`${msgs[i].role}: ${projectContent(i, allowances[i])}`);
+    }
+    flushRun(msgs.length - 1);
+    const promptContent = promptFor(`${note}
+${pieces.join("\n")}`);
+    return {
+      promptContent,
+      inputTokens: systemTokens + estimateTokenCount(promptContent),
+      expandedCount,
+      omittedCount: msgs.length - expandedCount
+    };
+  };
+  let budget = availableTokens;
+  let last = render(budget);
+  for (let attempt = 0; attempt < 12 && last.inputTokens > cap; attempt++) {
+    const overflow = last.inputTokens - cap;
+    budget = Math.max(0, Math.floor(budget - Math.max(overflow, Math.ceil(budget * 0.15))));
+    last = render(budget);
+  }
+  if (last.inputTokens > cap || last.expandedCount === 0) {
+    return blocked("no-budget", diagnosticsOf(last.expandedCount, last.omittedCount));
   }
-  const promptContent = makePrompt();
-  const inputTokens = systemTokens + estimateTokenCount(promptContent);
-  if (inputTokens > cap) return { status: "blocked", reason: "no-budget" };
   return {
     status: "ready",
-    promptContent,
+    promptContent: last.promptContent,
     coveredMessages: msgs.slice(),
     summarizedBeforeId: msgs[msgs.length - 1].id,
-    inputTokens,
+    inputTokens: last.inputTokens,
     outputTokenBudget,
-    partial: false
+    partial: false,
+    capTokens: cap,
+    expandedMessages: last.expandedCount,
+    omittedMessages: last.omittedCount
   };
 }
 var TOOL_RESULT_TRUNCATE_CHARS = 4e3;
@@ -602,16 +1076,29 @@ function buildCompactionMetrics(args) {
     estimateTokens,
     estimateMessageTokens
   } = args;
+  const previousSummaryTokens = estimateTokens(previousSummary || "");
+  const compressedTokens = msgsToCompress.reduce((s, m) => s + estimateMessageTokens(m), 0);
+  const newSummaryTokens = estimateTokens(newSummary || "");
+  const summaryInput = args.summaryInput;
   return {
     reason,
-    previousSummaryTokens: estimateTokens(previousSummary || ""),
-    compressedTokens: msgsToCompress.reduce((s, m) => s + estimateMessageTokens(m), 0),
+    previousSummaryTokens,
+    compressedTokens,
     retainedTokens: msgsToKeep.reduce((s, m) => s + estimateMessageTokens(m), 0),
-    newSummaryTokens: estimateTokens(newSummary || ""),
+    newSummaryTokens,
     compressedCount: msgsToCompress.length,
     retainedCount: msgsToKeep.length,
     summaryUsage,
-    hadPreviousSummary: previousSummary.trim().length > 0
+    hadPreviousSummary: previousSummary.trim().length > 0,
+    residentReductionTokens: previousSummaryTokens + compressedTokens - newSummaryTokens,
+    ...summaryInput ? {
+      summaryInputTokens: summaryInput.inputTokens,
+      ...summaryInput.capTokens !== void 0 ? { summaryInputCapTokens: summaryInput.capTokens } : {},
+      ...summaryInput.expandedMessages !== void 0 ? { summaryExpandedMessages: summaryInput.expandedMessages } : {},
+      ...summaryInput.omittedMessages !== void 0 ? { summaryOmittedMessages: summaryInput.omittedMessages } : {}
+    } : {},
+    ...args.planCounts?.pendingMessageCount !== void 0 ? { pendingMessageCount: args.planCounts.pendingMessageCount } : {},
+    ...args.planCounts?.plannedCompressCount !== void 0 ? { plannedCompressCount: args.planCounts.plannedCompressCount } : {}
   };
 }
 function buildCompactionMetricsFromPlan(args) {
@@ -622,6 +1109,8 @@ function buildCompactionMetricsFromPlan(args) {
     msgsToKeep: args.plan.msgsToKeep,
     newSummary: args.newSummary,
     summaryUsage: args.summaryUsage,
+    ...args.summaryInput ? { summaryInput: args.summaryInput } : {},
+    ...args.planCounts ? { planCounts: args.planCounts } : {},
     estimateTokens: estimateTokenCount,
     estimateMessageTokens: getMessageTokenCount
   });
@@ -631,7 +1120,13 @@ function formatCompactionMetricsLog(metrics) {
   const totalAfter = metrics.newSummaryTokens + metrics.retainedTokens;
   const ratio = totalBefore > 0 ? (totalAfter / totalBefore).toFixed(2) : "N/A";
   const llmInfo = metrics.summaryUsage ? `, summary_llm_in=${metrics.summaryUsage.input_tokens ?? "?"} out=${metrics.summaryUsage.output_tokens ?? "?"}` : "";
-  return `[Compaction] reason=${metrics.reason} compressed=${metrics.compressedCount}->${metrics.retainedCount} msgs, tokens=${totalBefore}->${totalAfter} (ratio=${ratio})${llmInfo}`;
+  const netInfo = `, net=${metrics.residentReductionTokens}`;
+  const omittedTotal = (metrics.summaryExpandedMessages ?? 0) + (metrics.summaryOmittedMessages ?? 0);
+  const capSuffix = metrics.summaryInputCapTokens !== void 0 ? `/${metrics.summaryInputCapTokens}` : "";
+  const omittedSuffix = metrics.summaryOmittedMessages !== void 0 ? ` omitted=${metrics.summaryOmittedMessages}/${omittedTotal}` : "";
+  const projectionInfo = metrics.summaryInputTokens !== void 0 ? `, summary_in=${metrics.summaryInputTokens}${capSuffix}${omittedSuffix}` : "";
+  const planInfo = metrics.pendingMessageCount !== void 0 ? `, pending=${metrics.pendingMessageCount}${metrics.plannedCompressCount !== void 0 ? ` fold=${metrics.plannedCompressCount}/${metrics.pendingMessageCount}` : ""}` : "";
+  return `[Compaction] reason=${metrics.reason} compressed=${metrics.compressedCount}->${metrics.retainedCount} msgs, tokens=${totalBefore}->${totalAfter} (ratio=${ratio})${llmInfo}${netInfo}${projectionInfo}${planInfo}`;
 }
 function computeSummarySourceMetadata(messagesWithIds, summarizedBeforeId) {
   if (!summarizedBeforeId) return void 0;
@@ -743,11 +1238,11 @@ function fuzzyContextWindow(normalizedName) {
   if (normalizedName.includes("gemini")) return 1048576;
   if (normalizedName.includes("qwen3.6") || normalizedName.includes("qwen3p6")) return QWEN_3_6_CONTEXT_WINDOW;
   if (normalizedName.includes("qwen3.7") || normalizedName.includes("qwen3p7")) return 262144;
-  if (normalizedName.includes("minimax-m2") || normalizedName.includes("minimax_m2")) return 262144;
-  if (normalizedName.includes("mistral-large-4")) {
-    const rest = normalizedName.slice(
-      normalizedName.indexOf("mistral-large-4") + "mistral-large-4".length
-    );
+  if (/minimax[-_ ]m2(?![\d.])/.test(normalizedName)) return 262144;
+  const MISTRAL_LARGE_4 = /mistral[-_ ]large[-_ ]4/;
+  const large4 = MISTRAL_LARGE_4.exec(normalizedName);
+  if (large4) {
+    const rest = normalizedName.slice(large4.index + large4[0].length);
     if (rest === "" || /^[-:_@/\s]/.test(rest)) return 524288;
   }
   if (normalizedName.includes("mistral-vibe-cli") || normalizedName.includes("mistral-medium-3.5") || normalizedName.includes("mistral-medium-3-5") || normalizedName.includes("mistral-medium-latest") || normalizedName.includes("mistral-medium-2604") || normalizedName.includes("mistral-large-latest") || normalizedName.includes("mistral-large-2512") || normalizedName.includes("mistral-small-latest") || normalizedName.includes("mistral-small-2603") || normalizedName.includes("magistral-medium-latest") || normalizedName.includes("magistral-small-latest") || // 展示名（TUI 传 displayName 的路径；Mistral 不在 MODEL_LOOKUP_MAP，无法走精确表）：
@@ -772,7 +1267,7 @@ function fuzzyContextWindow(normalizedName) {
   if (normalizedName.includes("llama3") || normalizedName.includes("llama-3") || normalizedName.includes("gemma2") || normalizedName.includes("gemma-2") || normalizedName.includes("gemma")) {
     return 8192;
   }
-  if (normalizedName.includes("llama2") || normalizedName.includes("llama-2") || normalizedName.includes("llama")) {
+  if (/(^|[^a-z])llama/.test(normalizedName)) {
     return 4096;
   }
   return void 0;
@@ -782,7 +1277,8 @@ var findCodexOAuthPinnedContextWindow = (modelName, provider) => {
   if (isPlatformManagedProvider(provider, modelName)) return void 0;
   return channelPinnedContextWindow(modelName.toLowerCase());
 };
-var findModelContextWindow = (modelName, provider) => {
+var findModelContextWindow = (modelName, provider) => findModelContextWindowImpl(modelName, provider, false);
+var findModelContextWindowImpl = (modelName, provider, allowHyphenRetry) => {
   if (!modelName) return void 0;
   const onPlatformChannel = isPlatformManagedProvider(provider, modelName);
   const normalizedName = modelName.toLowerCase();
@@ -834,8 +1330,19 @@ var findModelContextWindow = (modelName, provider) => {
       if (pinned !== void 0) return pinned;
     }
   }
-  return fuzzyContextWindow(normalizedName);
+  const fuzzy = fuzzyContextWindow(normalizedName);
+  if (fuzzy !== void 0) return fuzzy;
+  if (!allowHyphenRetry) {
+    const canonical = normalizeSeparators(modelName);
+    if (canonical.toLowerCase() !== normalizedName) {
+      return findModelContextWindowImpl(canonical, provider, true);
+    }
+  }
+  return void 0;
 };
+function normalizeSeparators(raw) {
+  return raw.replace(/\s+/g, "-").replace(/(?<=[a-z0-9])_(?=[a-z0-9])/g, "-");
+}
 var getModelContextWindow = (modelName, provider) => findModelContextWindow(modelName, provider) ?? DEFAULT_CONTEXT_WINDOW;
 var getModelInfo = (modelName) => {
   if (!modelName) return null;
@@ -844,18 +1351,28 @@ var getModelInfo = (modelName) => {
 };
 
 export {
-  DEFAULT_CONTEXT_WINDOW,
-  findCodexOAuthPinnedContextWindow,
-  findModelContextWindow,
-  getModelContextWindow,
-  getModelInfo,
-  sha256Hex,
+  readCacheReadInputTokens,
+  readCacheReadInputTokensOrNull,
+  readCacheCreationInputTokens,
+  normalizeUsage,
   serializeMessageContent,
   resolveCompressionTriggerRatio,
   truncateToolOutputForContext,
   getMessageTokenCount,
   resolveEffectiveCompactionBudgets,
   planCompression,
+  DEFAULT_CONTEXT_WINDOW,
+  findCodexOAuthPinnedContextWindow,
+  findModelContextWindow,
+  getModelContextWindow,
+  getModelInfo,
+  createSummaryBudget,
+  classifySummaryError,
+  isTruncatedCompletion,
+  summaryCompletionTokens,
+  selectSummaryRoute,
+  runSummaryWithContract,
+  sha256Hex,
   COMPACTION_SUMMARY_SCHEMA_VERSION,
   COMPACTION_SUMMARY_SYSTEM_PROMPT,
   buildBoundedCompactionRequest,