nolo-cli 0.25.2 → 0.26.1

package/{chunk-UTADKKM7.js → chunk-PTKI7TIN.js}

@@ -3,11 +3,10 @@ import {
   refreshAntigravityOAuthToken,
   refreshOpenAiCodexToken,
   refreshXaiOAuthToken
-} from "./chunk-J3HWN4B5.js";
+} from "./chunk-TWA4L7J7.js";
 import {
-  buildLocalAgentLookupKeys,
-  shouldReadAgentKeyRemotely
-} from "./chunk-OTIS3VY6.js";
+  buildLocalAgentLookupKeys
+} from "./chunk-AZW5AHMW.js";
 import {
   LOCAL_CODEX_AGENT_ID,
   LOCAL_CODEX_AGENT_KEY,
@@ -20,13 +19,13 @@ import {
 import {
   inlineImageUrlsForCustomProvider,
   prepareTools
-} from "./chunk-DZNONG7L.js";
+} from "./chunk-ZFL3LTU7.js";
 import {
   getAntigravityUserAgent,
   isAntigravityOAuthAgent,
   readAntigravityProjectId,
   resolveAntigravityCloudCodeBaseUrl
-} from "./chunk-OMWJXZSZ.js";
+} from "./chunk-XNZJVYFS.js";
 import {
   readXPostFunc,
   readXPostFunctionSchema,
@@ -35,7 +34,7 @@ import {
   rememberMemoryFunctionSchema,
   resolveAgentRequiredPackIds,
   setTodoListFunc
-} from "./chunk-P5AVCQF3.js";
+} from "./chunk-GGB6MPHD.js";
 import {
   createDrainExhaustedResponse,
   isCoreDrainingBody,
@@ -46,15 +45,20 @@ import {
   getDefaultCliLocalRuntimeDb
 } from "./chunk-WJRJDX2W.js";
 import {
+  applyTokenUsageToDayStats,
   prepareTokenUsageData,
-  resolveBuiltinPlatformAgentConfig
-} from "./chunk-THN2YFSR.js";
+  resolveBuiltinPlatformAgentConfig,
+  runKeyed
+} from "./chunk-ADE6577S.js";
 import {
   BUILTIN_TITLE_LLM_CONFIG
 } from "./chunk-XOIF2TNF.js";
 import {
   fetchServerSyncedCredential
-} from "./chunk-QJZJW65A.js";
+} from "./chunk-GD23ZX7Z.js";
+import {
+  require_date_fns
+} from "./chunk-M76ECQRD.js";
 import {
   FORCED_TOOLS,
   addDefaultLightWebToolsForConfiguredAgents,
@@ -76,11 +80,14 @@ import {
   dialogMessageRecordToAgentRuntimeMessage,
   executeLocalToolWithPolicy,
   executeOpenAiCompatibleChatCompletion,
+  isAgentCoolingDown,
   parseSseDataLineJson,
   parseSseDataLineObject,
+  parseSyncServersEnv,
   readOpenAiCompatibleSseCompletion,
+  resolveAgentNextAvailableAt,
   resolveOpenAiCompatibleProviderConfig
-} from "./chunk-5TAM6OX3.js";
+} from "./chunk-SIUEKVK4.js";
 import {
   buildPlatformChatCompletionRequest,
   canUsePlatformChatProvider,
@@ -90,25 +97,25 @@ import {
   readSseDataValues,
   resolvePlatformChatProviderConfig,
   shouldUsePlatformChatProvider
-} from "./chunk-MSBHAQVV.js";
+} from "./chunk-MCXQNEJ3.js";
 import {
   resolveAgentRuntimeConfigFromRecord
 } from "./chunk-OIQPORYI.js";
 import {
   summarizeToolArguments
-} from "./chunk-WV3WZRYZ.js";
+} from "./chunk-P3RZ6DHN.js";
 import {
   convertMessagesToResponsesInput,
   toResponsesTools
 } from "./chunk-D34RI6NI.js";
 import {
   resolvePlatformAuthToken
-} from "./chunk-5UI2CDQG.js";
+} from "./chunk-PXPISYEH.js";
 import {
   buildNoloWorkspaceOpenAiTools,
   parseNoloWorkspaceToolArguments,
   resolveRuntimeToolSurfaceForAgent
-} from "./chunk-UZH44DWK.js";
+} from "./chunk-WCP5FDTF.js";
 import {
   buildDialogFallbackTitleFromUserInput,
   normalizeDialogTitle
@@ -130,7 +137,7 @@ import {
 } from "./chunk-JZPBVBIY.js";
 import {
   isAdaptiveThinkingModelId
-} from "./chunk-J6MEBHAT.js";
+} from "./chunk-LEHTC2FI.js";
 import {
   canonicalizeToolNames
 } from "./chunk-ZZ2MH7LE.js";
@@ -177,6 +184,7 @@ import {
 } from "./chunk-OLZBQWZU.js";
 import {
   createTokenKey,
+  createTokenStatsKey,
   createUserKey,
   dialogMessageRange
 } from "./chunk-NRY4GCKJ.js";
@@ -207,6 +215,9 @@ import {
 import {
   asOptionalTrimmedString
 } from "./chunk-G3SUCT4W.js";
+import {
+  __toESM
+} from "./chunk-4VNS5WPM.js";
 
 // packages/cli/tui/i18n.ts
 var ZH_PATTERNS = /^zh/i;
@@ -862,6 +873,10 @@ Use /switch, /switch list, /switch minimax-m3, or a full agent key.`,
     zh: "nolo {0}\n\u7528 nolo update \u66F4\u65B0\u5F53\u524D\u5B89\u88C5\u3002\n\u5982\u679C\u672C\u5730\u4ED3\u5E93\u8F93\u51FA\u7684\u7248\u672C\u4E0D\u540C\uFF0C\u8BF7\u5148\u53D1\u5E03/\u5B89\u88C5\u6700\u65B0\u7684 npm \u5305\u3002"
   },
   versionUnknown: { en: "unknown version", zh: "\u672A\u77E5\u7248\u672C" },
+  updateAvailable: {
+    en: "New nolo {0} available (you have {1}) \u2014 run /update to upgrade",
+    zh: "\u65B0\u7248\u672C nolo {0} \u53EF\u7528\uFF08\u5F53\u524D {1}\uFF09\u2014 \u8FD0\u884C /update \u5347\u7EA7"
+  },
   // --- Dialog list / timestamps -----------------------------------------
   recentDialogs: { en: "Recent dialogs:", zh: "\u6700\u8FD1\u5BF9\u8BDD\uFF1A" },
   dialogListTip: {
@@ -1576,9 +1591,7 @@ function resolveBuiltinLocalCliAgentConfig(agentRef, userId) {
 }
 async function readAgentFromStore(args) {
   for (const key of buildLocalAgentLookupKeys(args)) {
-    const record = await args.store.read(key, {
-      remote: shouldReadAgentKeyRemotely(key)
-    });
+    const record = await args.store.read(key);
     if (!record || typeof record !== "object") continue;
     return resolveAgentRuntimeConfigFromRecord(key, record);
   }
@@ -2582,8 +2595,8 @@ function createCliControlAgentRunExecutor(deps = {}) {
             runSummaries.push({
               runId: rec.runId,
               status: rec.status,
-              startedAt: readTimestamp(rec.startedAt),
-              finishedAt: readTimestamp(rec.endedAt),
+              startedAt: readTimestamp(rec.startedAt)?.toString(),
+              finishedAt: readTimestamp(rec.endedAt)?.toString(),
               ...rec.agentName ? { agentName: rec.agentName } : {}
             });
           }
@@ -2928,6 +2941,9 @@ import { randomUUID } from "node:crypto";
 
 // packages/agent-runtime/antigravityWireModel.ts
 var WIRE_PROFILES = {
+  "gemini-3.7-flash-low": { maxOutputTokens: 65536 },
+  "gemini-3.7-flash-medium": { maxOutputTokens: 65536 },
+  "gemini-3.7-flash-high": { maxOutputTokens: 65536 },
   "gemini-3.6-flash-low": { maxOutputTokens: 65536 },
   "gemini-3.6-flash-medium": { maxOutputTokens: 65536 },
   "gemini-3.6-flash-high": { maxOutputTokens: 65536 },
@@ -2949,6 +2965,10 @@ function resolveAntigravityWireModel(logicalModelId) {
     const wireModelId = "gemini-3.5-flash-low";
     return { wireModelId, profile: WIRE_PROFILES[wireModelId] };
   }
+  if (lower === "gemini-3.7-flash") {
+    const wireModelId = "gemini-3.7-flash-low";
+    return { wireModelId, profile: WIRE_PROFILES[wireModelId] };
+  }
   if (lower === "gemini-3.6-flash") {
     const wireModelId = "gemini-3.6-flash-low";
     return { wireModelId, profile: WIRE_PROFILES[wireModelId] };
@@ -3547,6 +3567,9 @@ function buildAnthropicMessagesBody(args) {
       input_schema: fn?.parameters && typeof fn.parameters === "object" ? fn.parameters : { type: "object", properties: {} }
     }];
   }) : [];
+  if (tools.length > 0 && !tools.some((tool) => isCacheControl(tool.cache_control))) {
+    tools[tools.length - 1].cache_control = { type: "ephemeral" };
+  }
   const model = stringValue(args.openAiBody.model) ?? args.agentConfig.model ?? "claude-sonnet-5";
   const maxTokensRaw = args.openAiBody.max_completion_tokens ?? args.openAiBody.max_tokens ?? args.agentConfig.max_tokens;
   const maxTokens = typeof maxTokensRaw === "number" && Number.isFinite(maxTokensRaw) ? Math.max(1, Math.floor(maxTokensRaw)) : 8192;
@@ -3654,7 +3677,7 @@ async function fetchAnthropicMessagesCompletion(args) {
   const payload = await response.json().catch(async () => ({
     error: { message: await response.text().catch(() => response.statusText) }
   }));
-  if (!response.ok) return { status: response.status, body: payload };
+  if (!response.ok) return { status: response.status, body: payload, headers: response.headers };
   return { status: 200, body: mapAnthropicMessageToOpenAi(payload) };
 }
 function isAnthropicOAuthAgent(agent) {
@@ -10378,6 +10401,7 @@ function resolveFallbackServers(env) {
   const values = [
     env.NOLO_SERVER_URL,
     env.BASE_URL,
+    ...parseSyncServersEnv(env),
     ...NOLO_CLUSTER_SERVERS
   ].filter((value) => typeof value === "string" && value.trim().length > 0);
   return [...new Set(values.map(normalizeServerOrigin))];
@@ -10563,6 +10587,7 @@ function isRemoteAuthEndpoint(endpoint) {
 }
 
 // packages/cli/client/localRuntimeDialog.ts
+var import_date_fns = __toESM(require_date_fns(), 1);
 async function resolveStore(deps) {
   if (deps.store) return deps.store;
   return createCliHybridRecordStore({
@@ -10662,6 +10687,8 @@ async function writeLocalTokenRecord(args) {
       entry_path: "cli-local"
     });
     const recordKey = createTokenKey.recordForStableCall(args.userId, callId);
+    const existingToken = await args.store.read(recordKey, { remote: false }).catch(() => null);
+    if (existingToken) continue;
     const tokenRecord = {
       id: args.createId(),
       type: "token",
@@ -10669,6 +10696,26 @@ async function writeLocalTokenRecord(args) {
     };
     await args.store.write(recordKey, tokenRecord);
     ops.push({ type: "put", key: recordKey, value: tokenRecord });
+    const dateKey = (0, import_date_fns.format)(timestamp, "yyyy-MM-dd");
+    const statsKey = createTokenStatsKey(args.userId, dateKey);
+    await runKeyed(statsKey, async () => {
+      let existingStats = null;
+      try {
+        existingStats = await args.store.read(statsKey, { remote: false });
+      } catch {
+      }
+      const newStats = applyTokenUsageToDayStats(existingStats, {
+        userId: args.userId,
+        timeKey: dateKey,
+        model: prepared.billedModel,
+        provider: prepared.recordProvider,
+        input_tokens: prepared.usage.input_tokens,
+        output_tokens: prepared.usage.output_tokens,
+        cost: prepared.tokenData.cost
+      });
+      const statsRecord = { ...newStats, id: statsKey, type: "token" };
+      await args.store.write(statsKey, statsRecord);
+    });
   }
   return ops;
 }
@@ -10973,6 +11020,21 @@ function createCliLocalRuntimeAdapter(deps) {
       })
     }),
     resolveProvider: async (agentConfig) => {
+      if (isAgentCoolingDown(agentConfig, now())) {
+        throw new Error(
+          `agent temporarily unavailable until ${new Date(Number(agentConfig.nextAvailableAt)).toISOString()}`
+        );
+      }
+      const recordLocalAvailability = async (status, body) => {
+        const key = typeof agentConfig.key === "string" ? agentConfig.key : "";
+        if (!key || status !== 200 && status !== 429) return;
+        const current = await getOrCreateSharedStore(deps).then((store) => store.read(key, { remote: false })).catch(() => null);
+        if (!current || typeof current !== "object") return;
+        const next = { ...current };
+        if (status === 429) next.nextAvailableAt = resolveAgentNextAvailableAt(body, now());
+        else delete next.nextAvailableAt;
+        await getOrCreateSharedStore(deps).then((store) => store.write(key, next)).catch(() => void 0);
+      };
       if (isCliProviderAgent(agentConfig)) {
         const provider = resolveCliProviderName(agentConfig);
         logLocalRuntimeDiagnostic("provider.selected", {
@@ -11023,7 +11085,20 @@ function createCliLocalRuntimeAdapter(deps) {
       const credentialBroker = createFileCredentialBroker();
       const serverUrl = asOptionalTrimmedString(deps.env.NOLO_SERVER) ?? "https://us.nolo.chat";
       const authToken = asOptionalTrimmedString(deps.env.AUTH_TOKEN);
-      const syncFetcher = authToken ? (ref) => fetchServerSyncedCredential({ currentServer: serverUrl, authToken }, ref) : void 0;
+      const syncFetcher = authToken ? async (ref) => {
+        try {
+          return await fetchServerSyncedCredential(
+            { currentServer: serverUrl, authToken },
+            ref
+          );
+        } catch (error) {
+          logLocalRuntimeDiagnostic("credential.sync.unavailable", {
+            ref,
+            error: toErrorMessage(error)
+          });
+          return null;
+        }
+      } : void 0;
       if (isAntigravityOAuthAgent(agentConfig)) {
         const accessToken = await apiKeyRefResolver("antigravity");
         if (!accessToken) {
@@ -11077,6 +11152,7 @@ function createCliLocalRuntimeAdapter(deps) {
                 loopbackRequest
               })
             });
+            await recordLocalAvailability(result.status, result.body);
             if (result.status < 200 || result.status >= 300) {
               const errMsg = result.body && typeof result.body === "object" && result.body.error && typeof result.body.error.message === "string" ? result.body.error.message : JSON.stringify(result.body);
               throw new Error(
@@ -11145,6 +11221,7 @@ function createCliLocalRuntimeAdapter(deps) {
                 loopbackRequest
               })
             });
+            await recordLocalAvailability(result.status, result.body);
             if (result.status < 200 || result.status >= 300) {
               const errMsg = result.body?.error && typeof result.body.error === "object" && typeof result.body.error.message === "string" ? result.body.error.message : JSON.stringify(result.body);
               throw new Error(
@@ -11215,6 +11292,7 @@ function createCliLocalRuntimeAdapter(deps) {
                 loopbackRequest
               })
             });
+            await recordLocalAvailability(result.status, result.body);
             if (result.status < 200 || result.status >= 300) {
               const errMsg = result.body?.error && typeof result.body.error === "object" && typeof result.body.error.message === "string" ? result.body.error.message : JSON.stringify(result.body);
               throw new Error(

package/{chunk-5UI2CDQG.js → chunk-PXPISYEH.js}

@@ -1,10 +1,10 @@
 import {
   getModelConfig
-} from "./chunk-J6MEBHAT.js";
+} from "./chunk-LEHTC2FI.js";
 import {
   isPlatformHostedClaudeModel,
   isPlatformHostedGrokModel
-} from "./chunk-OHK3HE4G.js";
+} from "./chunk-CZHE6LLV.js";
 import {
   asTrimmedLowercaseString
 } from "./chunk-QSJ3HFEF.js";
@@ -249,7 +249,11 @@ async function buildProviderExecutionPlan(args) {
       }
     }
     if (!apiKey2 && agentConfig.credentialSynced && args.syncFetcher && credentialRef2) {
-      const synced = await args.syncFetcher(credentialRef2);
+      let synced = null;
+      try {
+        synced = await args.syncFetcher(credentialRef2);
+      } catch {
+      }
       const trimmed = asTrimmedString(synced);
       if (trimmed) {
         try {
@@ -316,13 +320,10 @@ async function buildProviderExecutionPlan(args) {
       serverUrl: resolvePlatformServerUrl(env),
       authToken: resolvePlatformAuthToken(env),
       agentKey: agentConfig.key,
-      ...agentConfig.apiSource ? { apiSource: agentConfig.apiSource } : {},
+      ...agentConfig.apiSource ? { apiSource: agentConfig.apiSource } : {}
       // Platform agents use the server-managed provider credential. Never
       // forward a raw credential from a stale/local agent record through the
-      // platform proxy. Custom proxy-backed agents retain their explicit key
-      // for the proxy's configured upstream.
-      ...mode === "custom" && agentConfig.apiKey?.trim() ? { apiKey: agentConfig.apiKey.trim() } : {},
-      ...mode === "custom" && agentConfig.apiKeyHeader?.trim() ? { apiKeyHeader: agentConfig.apiKeyHeader.trim() } : {}
+      // platform proxy; credentials are resolved by the platform server.
     };
   }
   let apiKey = "";
@@ -334,7 +335,11 @@ async function buildProviderExecutionPlan(args) {
     }
   }
   if (!apiKey && agentConfig.credentialSynced && args.syncFetcher && credentialRef) {
-    const synced = await args.syncFetcher(credentialRef);
+    let synced = null;
+    try {
+      synced = await args.syncFetcher(credentialRef);
+    } catch {
+    }
     const trimmed = asTrimmedString(synced);
     if (trimmed) {
       try {

package/{chunk-LWHK7TZR.js → chunk-PY7VGC5M.js}

@@ -1,9 +1,9 @@
 import {
   MODEL_LOOKUP_MAP
-} from "./chunk-J6MEBHAT.js";
+} from "./chunk-LEHTC2FI.js";
 import {
   ALL_MODELS
-} from "./chunk-T3EOQH72.js";
+} from "./chunk-G36SUBHA.js";
 import {
   addContentAction
 } from "./chunk-OXLLFZNF.js";
@@ -463,6 +463,11 @@ var classifyConversationLoad = (msgs) => {
   }
   return "medium";
 };
+var normalizeContextUsageRatio = (value) => {
+  if (typeof value !== "number" || !Number.isFinite(value)) return void 0;
+  const ratio = value > 1 ? value / 100 : value;
+  return ratio >= 0 && ratio <= 1 ? ratio : void 0;
+};
 var emptyPlan = (startIndex) => ({
   shouldCompress: false,
   compressCount: 0,
@@ -478,7 +483,8 @@ function planCompression(input) {
     summary,
     contextWindow,
     force = false,
-    reason
+    reason,
+    realContextUsagePercent
   } = input;
   let startIndex = 0;
   if (summarizedBeforeId) {
@@ -504,8 +510,20 @@ function planCompression(input) {
     summaryTokens: adjustedSummaryTokens,
     recentLoad
   });
+  let shouldTriggerByUsage = false;
+  const usageRatio = normalizeContextUsageRatio(realContextUsagePercent);
+  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;
+  }
   const shouldRunActiveSummary = force && reason === "manual" && !hasOpenEndedToolCall(pendingMsgs[pendingMsgs.length - 1]) && isActiveSummaryWorthDoing(pendingTokens, contextWindow);
-  if (totalUsed < historyBudget && !shouldRunActiveSummary) {
+  if (!shouldTriggerByUsage && !shouldRunActiveSummary) {
     return emptyPlan(startIndex);
   }
   let tokensToKeep = 0;
@@ -516,7 +534,8 @@ function planCompression(input) {
     tokensToKeep += t;
     keepCount++;
   }
-  let compressCount = shouldRunActiveSummary && totalUsed < historyBudget ? Math.max(0, pendingMsgs.length - ACTIVE_SUMMARY_TAIL_KEEP_COUNT) : pendingMsgs.length - keepCount;
+  const triggeredByRealUsage = usageRatio !== void 0 && shouldTriggerByUsage;
+  let compressCount = (shouldRunActiveSummary || triggeredByRealUsage) && totalUsed < historyBudget ? Math.max(0, pendingMsgs.length - ACTIVE_SUMMARY_TAIL_KEEP_COUNT) : pendingMsgs.length - keepCount;
   while (compressCount > 0 && compressCount < pendingMsgs.length && pendingMsgs[compressCount].role === "tool") {
     compressCount--;
   }

package/{chunk-C4JWF4EX.js → chunk-RM464B7I.js}

@@ -1,6 +1,6 @@
 import {
   clearCliLocalRuntimePreparedAgentCache
-} from "./chunk-UTADKKM7.js";
+} from "./chunk-PTKI7TIN.js";
 import {
   getReadableCliDb
 } from "./chunk-WFHXMLHZ.js";
@@ -12,7 +12,7 @@ import {
   resolveAgentRecordFromHybridStore,
   sanitizeAgentRecordForCliOutput,
   writeAgentRecord
-} from "./chunk-6NUG4BJE.js";
+} from "./chunk-J2WT5CEO.js";
 import {
   resolveCliAgentKeyInput
 } from "./chunk-VHBWHNVT.js";

package/{chunk-2KPE5HQM.js → chunk-S7AV3NCK.js}

@@ -19,7 +19,7 @@ import {
   parseAgentRunEvent,
   t,
   toolLabel
-} from "./chunk-UTADKKM7.js";
+} from "./chunk-PTKI7TIN.js";
 import {
   getReadableCliDb
 } from "./chunk-WFHXMLHZ.js";
@@ -27,7 +27,7 @@ import {
   normalizeAgentHandle,
   readDbRecord,
   resolveAgentRecordFromHybridStore
-} from "./chunk-6NUG4BJE.js";
+} from "./chunk-J2WT5CEO.js";
 import {
   NOLO_PROJECT_MANAGER_AGENT_KEY,
   isLocalCliAgentKey
@@ -38,7 +38,7 @@ import {
 } from "./chunk-LTRXGFMS.js";
 import {
   formatAgentListCard
-} from "./chunk-P5AVCQF3.js";
+} from "./chunk-GGB6MPHD.js";
 import {
   isGatewayHttpStatus
 } from "./chunk-RQSOAMPJ.js";
@@ -50,12 +50,12 @@ import {
   chooseMemoryOwners,
   loadMemoryCandidatesFromDb,
   resolveAgentMemoryPolicy
-} from "./chunk-NZXRTSOY.js";
+} from "./chunk-GNGOZMU2.js";
 import {
   API_REPORTED_COST_MULTIPLIER,
   normalizeUsage,
   parseUiAskChoiceContent
-} from "./chunk-THN2YFSR.js";
+} from "./chunk-ADE6577S.js";
 import {
   buildSkillDiscoveryContextBlock
 } from "./chunk-7OBMZCRD.js";
@@ -71,7 +71,7 @@ import {
   LOCAL_TURN_ABORTED_CODE,
   readActionGate,
   readCommandActionGatePayload
-} from "./chunk-WV3WZRYZ.js";
+} from "./chunk-P3RZ6DHN.js";
 import {
   createThinkParserState,
   flushThinkParser,
@@ -79,17 +79,17 @@ import {
 } from "./chunk-N7J63Q6C.js";
 import {
   resolvePlatformAuthToken
-} from "./chunk-5UI2CDQG.js";
+} from "./chunk-PXPISYEH.js";
 import {
   DEFAULT_CONTEXT_WINDOW,
   getModelContextWindow
-} from "./chunk-LWHK7TZR.js";
+} from "./chunk-PY7VGC5M.js";
 import {
   clipCompactText
 } from "./chunk-5MMAK75Q.js";
 import {
   findModelConfig
-} from "./chunk-J6MEBHAT.js";
+} from "./chunk-LEHTC2FI.js";
 import {
   buildMemoryOverlayLayer,
   buildMemoryUseGuidanceLayer
@@ -285,9 +285,20 @@ function formatUsage(usage, dialogId) {
   const parts = [];
   if (typeof dialogId === "string" && dialogId)
     parts.push(`dialog=${dialogId}`);
-  const input = usage?.input_tokens ?? usage?.prompt_tokens ?? 0;
-  const output = usage?.output_tokens ?? usage?.completion_tokens ?? 0;
-  if (input || output) parts.push(`tokens=${input}+${output}`);
+  if (usage && typeof usage === "object") {
+    const normalized = normalizeUsage(usage);
+    const input = Number(normalized.input_tokens || usage?.input_tokens || usage?.prompt_tokens || 0);
+    const output = Number(normalized.output_tokens || usage?.output_tokens || usage?.completion_tokens || 0);
+    const cacheHit = Number(normalized.cache_read_input_tokens || 0);
+    if (input || output) {
+      let tokenStr = `tokens=${input}+${output}`;
+      if (cacheHit > 0 && input > 0) {
+        const hitPercent = (cacheHit / (input + cacheHit) * 100).toFixed(1).replace(/\.0$/, "");
+        tokenStr += ` (cache: ${formatTokenCount(cacheHit)} / ${hitPercent}%)`;
+      }
+      parts.push(tokenStr);
+    }
+  }
   return parts.length ? `  (${parts.join("  ")})` : "";
 }
 
@@ -1405,9 +1416,28 @@ function createRenderAwareStreamWriter(args) {
   let inFence = false;
   let fenceLang = "unknown";
   let lastKind = null;
+  const PARTIAL_FLUSH_THRESHOLD = 48;
   const emitBreathIfNeeded = (nextKind) => {
     if (needsListProseBreath(lastKind, nextKind)) args.write("\n");
   };
+  const flushPartialProse = () => {
+    if (inFence || buffer.length < PARTIAL_FLUSH_THRESHOLD) return;
+    const firstLine = buffer.split("\n", 1)[0] ?? "";
+    if (isTableRow(firstLine) || classifyStreamLine(firstLine) !== "prose") return;
+    const Segmenter = Intl.Segmenter;
+    const graphemes = typeof Segmenter === "function" ? Array.from(new Segmenter(void 0, { granularity: "grapheme" }).segment(firstLine), (part) => part.segment) : Array.from(firstLine);
+    if (graphemes.length < PARTIAL_FLUSH_THRESHOLD) return;
+    let cut = PARTIAL_FLUSH_THRESHOLD;
+    const whitespaceIndex = graphemes.slice(0, cut).findLastIndex((part) => /\s/.test(part));
+    if (whitespaceIndex > 0) cut = whitespaceIndex + 1;
+    const prefix = graphemes.slice(0, cut).join("");
+    const hasUnsafeMarkdownPrefix = /[`*_~\[]/.test(prefix) || /^\s*[#>]/.test(prefix) || /!\[/.test(prefix);
+    if (hasUnsafeMarkdownPrefix) return;
+    emitBreathIfNeeded("prose");
+    emitFormattedAssistantBlock(args.write, prefix, false);
+    lastKind = "prose";
+    buffer = prefix.length === buffer.length ? "" : buffer.slice(prefix.length);
+  };
   const flushCompleteBlocks = () => {
     while (buffer.includes("\n")) {
       const lines = buffer.split("\n");
@@ -1467,6 +1497,7 @@ function createRenderAwareStreamWriter(args) {
       if (!chunk) return;
       buffer += chunk;
       flushCompleteBlocks();
+      flushPartialProse();
     },
     flush() {
       if (!buffer) return;
@@ -3449,7 +3480,7 @@ function findCollapsedPasteSpanAt(buffer, pos, opts = {}) {
 // packages/cli/client/agentRun.ts
 import { ulid } from "ulid";
 async function loadRunLocalAgentTurn() {
-  const { runLocalAgentTurn } = await import("./agentRuntimeLocal-7N2SUWO2.js");
+  const { runLocalAgentTurn } = await import("./agentRuntimeLocal-GVNERRKQ.js");
   return runLocalAgentTurn;
 }
 async function resolveCurrentMachineId(options) {
@@ -4969,7 +5000,9 @@ ${mdContent}`);
       const memoryAgentKey = effectiveAgentKey;
       const memorySpaceId = parsed.spaceId ?? void 0;
       try {
-        if (memoryAuthToken && memoryServerUrl) {
+        const resolvedRuntimeMode = parsed.runtimeMode ?? (env.NOLO_RUNTIME_MODE === "local" || env.NOLO_RUNTIME_MODE === "server" || env.NOLO_RUNTIME_MODE === "auto" ? env.NOLO_RUNTIME_MODE : "auto");
+        const usesExplicitServerRuntime = resolvedRuntimeMode === "server";
+        if (usesExplicitServerRuntime && memoryAuthToken && memoryServerUrl) {
           const memoryResponse = await fetch(`${memoryServerUrl}/api/memory/query`, {
             method: "POST",
             headers: {