nolo-cli 0.67.0 → 0.68.0-alpha.2

package/{chunk-HQ3ZLBSI.js → chunk-2XENHR3P.js}

@@ -1,12 +1,24 @@
+import {
+  normalizeTimingMs
+} from "./chunk-3E67YXGD.js";
 import {
   MODEL_LOOKUP_MAP
 } from "./chunk-TQC77XVE.js";
 import {
   ALL_MODELS
 } from "./chunk-6IT5ECCC.js";
+import {
+  asOptionalFiniteNumber
+} from "./chunk-WWC7I6TA.js";
+import {
+  asTrimmedNonEmptyStringArray
+} from "./chunk-I42VC5KR.js";
 import {
   BUILTIN_NOLO_AGENT_KEY
 } from "./chunk-4MKPBCTN.js";
+import {
+  asOptionalTrimmedString
+} from "./chunk-G3SUCT4W.js";
 
 // packages/ai/llm/getModelContextWindow.ts
 var DEFAULT_CONTEXT_WINDOW = 256e3;
@@ -181,6 +193,89 @@ var readCacheCreationInputTokens = (usage) => firstPositive(usage, [
   topLevel("prompt_cache_miss_tokens")
 ]);
 
+// 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 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/agent-runtime/emptyAssistantRepair.ts
 var EMPTY_ASSISTANT_REPAIR_PROMPT = "\u8BF7\u7ED9\u51FA\u660E\u786E\u7684\u6587\u5B57\u56DE\u7B54\u6216\u6267\u884C\u4E0B\u4E00\u6B65\uFF1A\u5982\u679C\u4EFB\u52A1\u5DF2\u5B8C\u6210\uFF0C\u8BF7\u76F4\u63A5\u603B\u7ED3\u7ED3\u679C\uFF1B\u5982\u679C\u9700\u8981\u8C03\u7528\u5DE5\u5177\uFF0C\u53EA\u80FD\u8C03\u7528\u672C\u6B21\u8BF7\u6C42\u4E2D\u5DF2\u58F0\u660E\u7684\u5DE5\u5177\u2014\u2014\u4F60\u60F3\u7528\u7684\u5DE5\u5177\u4E0D\u5728\u58F0\u660E\u5217\u8868\u4E2D\u65F6\uFF0C\u4E0D\u8981\u5C1D\u8BD5\u8C03\u7528\u5B83\uFF0C\u76F4\u63A5\u7528\u6587\u5B57\u8BF4\u660E\u8BE5\u80FD\u529B\u5F53\u524D\u4E0D\u53EF\u7528\u3002\u8BF7\u5207\u52FF\u8FD4\u56DE\u7A7A\u5185\u5BB9\u3002";
 var EMPTY_ASSISTANT_FALLBACK_MESSAGE = "\u6A21\u578B\u8FDE\u7EED\u8FD4\u56DE\u7A7A\u6D88\u606F\uFF0C\u5F53\u524D\u4EFB\u52A1\u672A\u5B8C\u6210\u3002\u8BF7\u91CD\u8BD5\u5F53\u524D\u6B65\u9AA4\uFF0C\u6216\u7ED9\u51FA\u66F4\u5177\u4F53\u7684\u4FEE\u6539\u8303\u56F4\u3002";
@@ -294,6 +389,7 @@ export {
   DEFAULT_CONTEXT_WINDOW,
   getModelContextWindow,
   getModelInfo,
+  normalizeUsage,
   EMPTY_ASSISTANT_REPAIR_PROMPT,
   EMPTY_ASSISTANT_FALLBACK_MESSAGE,
   LENGTH_TRUNCATED_FALLBACK_MESSAGE,

package/{chunk-TDTSQVDX.js → chunk-3E67YXGD.js}

@@ -310,11 +310,64 @@ var planContextUsage = (params) => {
   return { historyBudget, rawMessageBudget, minTailTokens };
 };
 
+// packages/ai/context/toolOutputCap.ts
+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 clamp2 = (v, min, max) => Math.min(max, Math.max(min, v));
+var capToolOutputTokens = (contextWindow) => clamp2(
+  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) + clamp2(
+    Math.floor(contextWindow * ROUND_OVERHEAD_WINDOW_RATIO),
+    ROUND_OVERHEAD_MIN_TOKENS,
+    ROUND_OVERHEAD_MAX_TOKENS
+  );
+  return clamp2(
+    1 - headroom / contextWindow,
+    COMPRESSION_TRIGGER_RATIO_MIN,
+    COMPRESSION_TRIGGER_RATIO_MAX
+  );
+};
+var truncateToolOutputForContext = (text, contextWindow) => {
+  if (!text) return text;
+  const capTokens = capToolOutputTokens(contextWindow);
+  if (estimateTokenCount(text) <= capTokens) return text;
+  const build = (keepChars) => {
+    const headChars = Math.floor(keepChars * 0.6);
+    const tailChars = keepChars - headChars;
+    const omittedTokens = estimateTokenCount(
+      text.slice(headChars, text.length - tailChars)
+    );
+    return text.slice(0, headChars) + `
+...[tool output truncated: ~${omittedTokens} tokens omitted]...
+` + text.slice(text.length - tailChars);
+  };
+  let lo = 0;
+  let hi = text.length;
+  while (lo < hi) {
+    const mid = Math.ceil((lo + hi) / 2);
+    if (estimateTokenCount(build(mid)) <= capTokens) lo = mid;
+    else hi = mid - 1;
+  }
+  return build(lo);
+};
+
 // packages/ai/context/planCompression.ts
 var MIN_COMPRESS_COUNT = 5;
 var ACTIVE_SUMMARY_TAIL_KEEP_COUNT = 2;
-var STUB_KEEP_COUNT = 3;
-var TOOL_STUB_TEXT = "[tool output cleared to save context]";
+var MIN_TRIGGER_PENDING_COUNT = MIN_COMPRESS_COUNT + ACTIVE_SUMMARY_TAIL_KEEP_COUNT;
 var getMessageTokenCount = (msg) => {
   const content = serializeMessageContent(msg.content) || "";
   let tokens = estimateTokenCount(content);
@@ -333,13 +386,6 @@ var getMessageTokenCount = (msg) => {
   return tokens;
 };
 var hasOpenEndedToolCall = (msg) => !!msg && Array.isArray(msg.tool_calls) && msg.tool_calls.length > 0;
-var isActiveSummaryWorthDoing = (pendingTokens, contextWindow) => {
-  const minTokens = Math.min(
-    4e4,
-    Math.max(1e4, Math.floor(contextWindow * 0.05))
-  );
-  return pendingTokens >= minTokens;
-};
 var classifyConversationLoad = (msgs) => {
   const N = 20;
   if (!Array.isArray(msgs) || msgs.length === 0) return "light";
@@ -381,38 +427,27 @@ function findPendingMessages(allMsgs, summarizedBeforeId) {
 }
 function shouldTriggerCompaction(args) {
   const {
-    pendingTokens,
-    summaryTokens,
-    contextWindow,
+    pendingMsgCount,
+    totalUsed,
     historyBudget,
+    triggerRatio,
     force,
     reason,
     realContextUsagePercent,
-    lastCompactedTokenCount,
+    coldResume,
     lastMsg
   } = args;
-  const totalUsed = summaryTokens + pendingTokens;
   const usageRatio = normalizeContextUsageRatio(realContextUsagePercent);
-  let shouldTriggerByUsage = false;
-  if (usageRatio !== void 0) {
-    if (usageRatio >= 0.78) {
-      shouldTriggerByUsage = true;
-    } else if (usageRatio >= 0.65 && isActiveSummaryWorthDoing(pendingTokens, contextWindow)) {
-      shouldTriggerByUsage = true;
-    }
-  }
-  if (totalUsed >= historyBudget) {
-    shouldTriggerByUsage = true;
-  }
-  if (typeof lastCompactedTokenCount === "number" && Number.isFinite(lastCompactedTokenCount) && lastCompactedTokenCount >= 0 && !force && !(usageRatio !== void 0 && usageRatio >= 0.78)) {
-    const minNewTokensToCompress = Math.min(4e4, Math.max(5e3, Math.floor(contextWindow * 0.03)));
-    if (totalUsed < lastCompactedTokenCount + minNewTokensToCompress) {
-      return { trigger: false, triggeredByRealUsage: false, shouldRunActiveSummary: false };
-    }
-  }
-  const shouldRunActiveSummary = force && reason === "manual" && !hasOpenEndedToolCall(lastMsg) && isActiveSummaryWorthDoing(pendingTokens, contextWindow);
-  const triggeredByRealUsage = usageRatio !== void 0 && shouldTriggerByUsage;
-  return { trigger: shouldTriggerByUsage || shouldRunActiveSummary, triggeredByRealUsage, shouldRunActiveSummary };
+  const triggeredByRealUsage = usageRatio !== void 0 && usageRatio >= triggerRatio;
+  const triggeredByEstimate = usageRatio === void 0 && totalUsed >= historyBudget;
+  const shouldRunActiveSummary = force && reason === "manual" && !hasOpenEndedToolCall(lastMsg) && pendingMsgCount >= MIN_TRIGGER_PENDING_COUNT;
+  const triggeredByColdResume = coldResume === true && pendingMsgCount >= MIN_TRIGGER_PENDING_COUNT;
+  return {
+    trigger: triggeredByRealUsage || triggeredByEstimate || shouldRunActiveSummary || triggeredByColdResume,
+    triggeredByRealUsage,
+    shouldRunActiveSummary,
+    triggeredByColdResume
+  };
 }
 function calculateCompressCount(pendingMsgs, rawMessageBudget, opts) {
   const { keepTailCount, totalUsed, historyBudget } = opts;
@@ -429,45 +464,6 @@ function calculateCompressCount(pendingMsgs, rawMessageBudget, opts) {
   }
   return pendingMsgs.length - keepCount;
 }
-function tryComputeStubPlan(pendingMsgs, totalUsed, historyBudget, startIndex) {
-  const toolIndices = [];
-  for (let i = 0; i < pendingMsgs.length; i++) {
-    if (pendingMsgs[i].role === "tool") {
-      const content = serializeMessageContent(pendingMsgs[i].content) || "";
-      if (content !== TOOL_STUB_TEXT) toolIndices.push(i);
-    }
-  }
-  if (toolIndices.length <= STUB_KEEP_COUNT) return null;
-  const firstStubbedIdx = toolIndices[toolIndices.length - STUB_KEEP_COUNT - 1];
-  const beforeMsg = pendingMsgs[firstStubbedIdx];
-  if (!beforeMsg) return null;
-  const stubOverhead = estimateTokenCount(TOOL_STUB_TEXT);
-  const stubbedMsgs = pendingMsgs.slice(0, firstStubbedIdx + 1).filter(
-    (m) => m.role === "tool"
-  );
-  const originalTokens = stubbedMsgs.reduce(
-    (s, m) => s + getMessageTokenCount(m),
-    0
-  );
-  const savedTokens = Math.max(0, originalTokens - stubOverhead * stubbedMsgs.length);
-  if (totalUsed - savedTokens > historyBudget) return null;
-  const keptMsgs = pendingMsgs.slice(firstStubbedIdx + 1);
-  return {
-    shouldCompress: true,
-    compressCount: stubbedMsgs.length,
-    msgsToCompress: stubbedMsgs,
-    msgsToKeep: keptMsgs,
-    newSummarizedBeforeId: void 0,
-    startIndex,
-    stub: {
-      // beforeId = 第一条被 stub 的 tool 结果的消息 id；投影端按
-      // 「abs index <= beforeId 的 index 的 tool 消息」判定是否 stub。
-      beforeId: beforeMsg.id,
-      stubbedCount: stubbedMsgs.length,
-      savedTokens
-    }
-  };
-}
 function guardToolChainBoundary(pendingMsgs, compressCount) {
   let count = compressCount;
   while (count > 0 && count < pendingMsgs.length && pendingMsgs[count].role === "tool") {
@@ -489,7 +485,8 @@ function planCompression(input) {
     contextWindow,
     force = false,
     reason,
-    realContextUsagePercent
+    realContextUsagePercent,
+    coldResume
   } = input;
   const { pendingMsgs, startIndex } = findPendingMessages(allMsgs, summarizedBeforeId);
   if (pendingMsgs.length === 0) return emptyPlan(startIndex);
@@ -503,27 +500,26 @@ function planCompression(input) {
     summaryTokens: adjustedSummaryTokens,
     recentLoad
   });
-  const { trigger, triggeredByRealUsage, shouldRunActiveSummary } = shouldTriggerCompaction({
-    pendingTokens,
-    summaryTokens,
-    contextWindow,
+  const triggerRatio = resolveCompressionTriggerRatio(contextWindow);
+  const {
+    trigger,
+    triggeredByRealUsage,
+    shouldRunActiveSummary,
+    triggeredByColdResume
+  } = shouldTriggerCompaction({
+    pendingMsgCount: pendingMsgs.length,
+    totalUsed,
     historyBudget,
+    triggerRatio,
     force,
     reason,
     realContextUsagePercent,
-    lastCompactedTokenCount: input.lastCompactedTokenCount,
+    coldResume,
     lastMsg: pendingMsgs[pendingMsgs.length - 1]
   });
   if (!trigger) return emptyPlan(startIndex);
-  const stubPlan = tryComputeStubPlan(
-    pendingMsgs,
-    totalUsed,
-    historyBudget,
-    startIndex
-  );
-  if (stubPlan) return stubPlan;
   let compressCount = calculateCompressCount(pendingMsgs, rawMessageBudget, {
-    keepTailCount: shouldRunActiveSummary || triggeredByRealUsage,
+    keepTailCount: shouldRunActiveSummary || triggeredByRealUsage || triggeredByColdResume,
     totalUsed,
     historyBudget
   });
@@ -779,7 +775,8 @@ export {
   buildDatabaseFileContentUrl,
   serializeMessageContent,
   normalizeAssistantContentBuffer,
-  TOOL_STUB_TEXT,
+  resolveCompressionTriggerRatio,
+  truncateToolOutputForContext,
   planCompression,
   COMPACTION_SUMMARY_SCHEMA_VERSION,
   COMPACTION_SUMMARY_SYSTEM_PROMPT,