nexrall-code 0.5.199 → 0.5.200

package/README.md

@@ -80,7 +80,7 @@ Run `nex --help` for the full list.
 
 ## Issues
 
-Bug reports and feature requests: **[github.com/Maxrall-Inc/nexrall-issue](https://github.com/Maxrall-Inc/nexrall-issue/issues)**.
+Bug reports and feature requests: **[github.com/Maxrall-Inc/nexrall-issues](https://github.com/Maxrall-Inc/nexrall-issues/issues)**.
 
 ## License
 

package/dist/index.js

@@ -10647,6 +10647,13 @@ var require_client = __commonJS({
                     break;
                   }
                   case "context_warning": {
+                    const message = typeof evt.message === "string" ? evt.message : "";
+                    onEvent({
+                      type: "context_warning",
+                      message,
+                      ...typeof evt.reason === "string" ? { reason: evt.reason } : {},
+                      ...typeof evt.trimmed === "number" ? { trimmed: evt.trimmed } : {}
+                    });
                     break;
                   }
                   case "notice": {
@@ -24553,8 +24560,10 @@ var require_compaction = __commonJS({
   "../core/dist/agent/compaction.js"(exports) {
     "use strict";
     Object.defineProperty(exports, "__esModule", { value: true });
-    exports.CompactionUnavailableError = exports.PRUNE_KEEP_RECENT = exports.VERIFY_CMD_RE = exports.WRITE_TOOL_NAMES = exports.MAX_BODY_BYTES = exports.COMPACT_MAX_FAILURES = exports.COMPACT_MIN_RECLAIM_BYTES = exports.COMPACT_KEEP_MIN = exports._lastApiCallEndedAt = exports.PRUNE_MIN_RECLAIM_BYTES = exports.MODEL_CONTEXT_TOKENS = void 0;
+    exports.BYTES_PER_TOKEN_ESTIMATE = exports.CompactionUnavailableError = exports.PRUNE_KEEP_RECENT = exports.VERIFY_CMD_RE = exports.WRITE_TOOL_NAMES = exports.MAX_BODY_BYTES = exports.COMPACT_MAX_FAILURES = exports.COMPACT_MIN_RECLAIM_BYTES = exports.COMPACT_KEEP_MIN = exports._lastApiCallEndedAt = exports.PRUNE_MIN_RECLAIM_BYTES = exports.MODEL_CONTEXT_TOKENS = void 0;
     exports.contextWindowFor = contextWindowFor2;
+    exports.outputReserveFor = outputReserveFor;
+    exports.serverMessageBudgetFor = serverMessageBudgetFor;
     exports.compactionLimits = compactionLimits;
     exports.compactionThresholds = compactionThresholds2;
     exports.pruneReclaimFloor = pruneReclaimFloor;
@@ -24647,17 +24656,34 @@ var require_compaction = __commonJS({
         return fromSettings;
       return void 0;
     }
-    function compactionLimits(contextWindow, settingsRaw) {
+    var SERVER_ESTIMATE_RATIO = 0.875;
+    var SERVER_OVERHEAD_TOKENS = 4e4;
+    var SERVER_BUDGET_SAFETY = 0.9;
+    var MIN_SERVER_AWARE_COMPACT = 15e3;
+    function outputReserveFor(contextWindow, model) {
+      return (0, modelCatalogue_1.liveMaxOutputTokensFor)(model, Math.min(128e3, Math.max(16384, Math.floor(contextWindow * 0.2))));
+    }
+    function serverMessageBudgetFor(contextWindow, model) {
+      const hard = contextWindow - outputReserveFor(contextWindow, model) - SERVER_OVERHEAD_TOKENS - Math.floor(contextWindow * 0.03);
+      return Math.max(0, hard);
+    }
+    function serverAwareCompactLine(contextWindow, model) {
+      const converted = Math.floor(serverMessageBudgetFor(contextWindow, model) * SERVER_ESTIMATE_RATIO * SERVER_BUDGET_SAFETY);
+      return Math.max(MIN_SERVER_AWARE_COMPACT, converted);
+    }
+    function compactionLimits(contextWindow, settingsRaw, model) {
       const cap = compactWindowCap(settingsRaw);
       const eff = cap === void 0 ? contextWindow : Math.min(contextWindow, cap);
       const compactFrac = envFraction("NEXRALL_COMPACT_THRESHOLD", 0);
       const pruneFrac = envFraction("NEXRALL_PRUNE_THRESHOLD", 0);
-      const compact = compactFrac ? eff * compactFrac : Math.max(eff * 0.5, eff - Math.min(COMPACT_OUTPUT_RESERVE, eff * 0.1) - COMPACT_BUFFER_TOKENS);
+      let compact = compactFrac ? eff * compactFrac : Math.max(eff * 0.5, eff - Math.min(COMPACT_OUTPUT_RESERVE, eff * 0.1) - COMPACT_BUFFER_TOKENS);
+      if (model !== void 0)
+        compact = Math.min(compact, serverAwareCompactLine(contextWindow, model));
       const prune = Math.min(pruneFrac ? eff * pruneFrac : eff * DEFAULT_PRUNE_FRACTION, compact * 0.9);
       return { prune: Math.floor(prune), compact: Math.floor(compact) };
     }
-    function compactionThresholds2(contextWindow = 1e6, settingsRaw) {
-      const { prune, compact } = compactionLimits(contextWindow, settingsRaw);
+    function compactionThresholds2(contextWindow = 1e6, settingsRaw, model) {
+      const { prune, compact } = compactionLimits(contextWindow, settingsRaw, model);
       return { prune: prune / contextWindow, compact: compact / contextWindow };
     }
     exports.PRUNE_MIN_RECLAIM_BYTES = 96 * 1024;
@@ -24900,11 +24926,22 @@ ${tail}`;
       }
       return reclaimed;
     }
+    var ORIGINAL_TASK_MARKER = "ORIGINAL TASK (verbatim \u2014 keep working toward this, do not lose sight of it):\n";
+    var COMPACT_PREAMBLE_PREFIX = "[Auto-compacted ";
     function originalTaskText(messages) {
       const first = messages.find((m2) => m2.role === "user");
       if (!first || !Array.isArray(first.content))
         return "";
-      return first.content.filter((b) => b.type === "text" && b.text && !(0, types_1.isRuntimeContextBlock)(b)).map((b) => b.text).join("\n").trim();
+      const text = first.content.filter((b) => b.type === "text" && b.text && !(0, types_1.isRuntimeContextBlock)(b)).map((b) => b.text).join("\n").trim();
+      if (text.startsWith(COMPACT_PREAMBLE_PREFIX)) {
+        const at = text.indexOf(ORIGINAL_TASK_MARKER);
+        if (at >= 0) {
+          const rest = text.slice(at + ORIGINAL_TASK_MARKER.length);
+          const end = rest.search(/\n\n(PROGRESS LEDGER|Summary of the earlier conversation so far:)/);
+          return (end >= 0 ? rest.slice(0, end) : rest).trim();
+        }
+      }
+      return text;
     }
     var CompactionUnavailableError = class extends Error {
       constructor() {
@@ -25003,8 +25040,7 @@ Do NOT call any tools and do NOT continue the task. Reply with text only: a conc
         }
       if (!summary)
         return false;
-      const taskBlock = originalTask ? `ORIGINAL TASK (verbatim \u2014 keep working toward this, do not lose sight of it):
-${originalTask}
+      const taskBlock = originalTask ? `${ORIGINAL_TASK_MARKER}${originalTask}
 
 ` : "";
       const ledgerText = ledger ? ledgerSummary(ledger) : "";
@@ -25053,9 +25089,9 @@ Continue the work from here.`
       } catch {
       }
     }
-    var RESUME_CHARS_PER_TOKEN = 4;
+    exports.BYTES_PER_TOKEN_ESTIMATE = 4;
     function estimateTokensRough3(messages) {
-      return Math.ceil(estimateBodyBytes2(messages) / RESUME_CHARS_PER_TOKEN);
+      return Math.ceil(estimateBodyBytes2(messages) / exports.BYTES_PER_TOKEN_ESTIMATE);
     }
     async function compactMessagesForResume2(messages, opts) {
       if (messages.length <= exports.COMPACT_KEEP_MIN + 2)
@@ -25063,10 +25099,10 @@ Continue the work from here.`
       const settings = (0, rules_1.loadSettings)(opts.workDir);
       if (!resolveAutoCompact(void 0, settings.raw))
         return false;
-      const contextWindow = (0, modelCatalogue_1.liveContextWindowFor)(opts.model, exports.MODEL_CONTEXT_TOKENS[opts.model ?? "claude-sonnet-5-5"] ?? 1e6);
+      const contextWindow = contextWindowFor2(opts.model);
       let bodyBytes = estimateBodyBytes2(messages);
       let tokenGuess = estimateTokensRough3(messages);
-      const limits = compactionLimits(contextWindow, settings.raw);
+      const limits = compactionLimits(contextWindow, settings.raw, opts.model);
       const overPruneThreshold = () => tokenGuess > limits.prune || bodyBytes > exports.MAX_BODY_BYTES;
       const overCompactThreshold = () => tokenGuess > limits.compact || bodyBytes > exports.MAX_BODY_BYTES;
       if (!overPruneThreshold())
@@ -26596,8 +26632,8 @@ var require_loop = __commonJS({
   "../core/dist/agent/loop.js"(exports) {
     "use strict";
     Object.defineProperty(exports, "__esModule", { value: true });
-    exports.stopReasonNotice = exports.executeAgentMemoryWrite = exports.AGENT_MEMORY_TOOL_SCHEMA = exports.AGENT_MEMORY_TOOL = exports._stallLimits = exports.errorRoundSignature = exports._emptyTurnRetry = exports.shouldRetryEmptyTurn = exports.emptyTurnBackoffMs = exports.compactMessagesForResume = exports.estimateTokensRough = exports.makeCachedSummarizer = exports.pruneOldToolResults = exports.ledgerSummary = exports.ledgerRecord = exports.createLedger = exports.transcriptOf = exports.persistRuntimeContext = exports.findSafeCutIndex = exports.VERIFY_CMD_RE = exports.allowsTestOnlyWrite = exports.WRITE_TOOL_NAMES = exports.estimateBodyBytes = exports.pruneReclaimFloor = exports.compactionThresholds = exports.compactionLimits = exports.contextWindowFor = exports.runElicitationResultHooks = exports.runElicitationHooks = exports.hasHookEventListener = exports.setHookEventListener = exports.queueHookDelivery = exports.resetHookOnceState = exports.setHookMcpCaller = exports.setHookStatusListener = exports.setHookWakeListener = exports.formatHookDeliveries = exports.hasPendingHookWake = exports.takeHookWakes = exports.drainHookDeliveries = exports.fireObserverHook = exports.fireManualPreCompactHook = exports.fireManualPostCompactHook = exports.fireNotificationHooks = exports.runPermissionRequestHooks = exports.runLifecycleHooks = exports.classifyStopFailure = exports.parsePromptHookReply = exports.loadHooks = exports.bashNeedsRepoLock = void 0;
-    exports.lastToolResults = exports.summariseSubTaskProgress = exports.capSubTaskText = exports.extractSubTaskText = exports.ToolNotAllowedError = exports.resolveSubtaskTimeoutMs = exports.canSpawnSubAgents = exports.intersectAllowlists = exports.noSpawnReason = exports.resolveMaxSubagentDepth = exports._sessionSubAgentCount = exports.resetSessionSubAgentBudget = exports._resetSubTaskLimiter = exports.createLimiter = exports.resolveMaxSubagentsPerSession = exports.resolveMaxConcurrentSubtasks = exports.lockPathsFor = exports.resolveMaxIterations = void 0;
+    exports.AGENT_MEMORY_TOOL = exports._stallLimits = exports.errorRoundSignature = exports._emptyTurnRetry = exports.shouldRetryEmptyTurn = exports.emptyTurnBackoffMs = exports.BYTES_PER_TOKEN_ESTIMATE = exports.outputReserveFor = exports.serverMessageBudgetFor = exports.compactMessagesForResume = exports.estimateTokensRough = exports.makeCachedSummarizer = exports.pruneOldToolResults = exports.ledgerSummary = exports.ledgerRecord = exports.createLedger = exports.transcriptOf = exports.persistRuntimeContext = exports.findSafeCutIndex = exports.VERIFY_CMD_RE = exports.allowsTestOnlyWrite = exports.WRITE_TOOL_NAMES = exports.estimateBodyBytes = exports.pruneReclaimFloor = exports.compactionThresholds = exports.compactionLimits = exports.contextWindowFor = exports.runElicitationResultHooks = exports.runElicitationHooks = exports.hasHookEventListener = exports.setHookEventListener = exports.queueHookDelivery = exports.resetHookOnceState = exports.setHookMcpCaller = exports.setHookStatusListener = exports.setHookWakeListener = exports.formatHookDeliveries = exports.hasPendingHookWake = exports.takeHookWakes = exports.drainHookDeliveries = exports.fireObserverHook = exports.fireManualPreCompactHook = exports.fireManualPostCompactHook = exports.fireNotificationHooks = exports.runPermissionRequestHooks = exports.runLifecycleHooks = exports.classifyStopFailure = exports.parsePromptHookReply = exports.loadHooks = exports.bashNeedsRepoLock = void 0;
+    exports.lastToolResults = exports.summariseSubTaskProgress = exports.capSubTaskText = exports.extractSubTaskText = exports.ToolNotAllowedError = exports.resolveSubtaskTimeoutMs = exports.canSpawnSubAgents = exports.intersectAllowlists = exports.noSpawnReason = exports.resolveMaxSubagentDepth = exports._sessionSubAgentCount = exports.resetSessionSubAgentBudget = exports._resetSubTaskLimiter = exports.createLimiter = exports.resolveMaxSubagentsPerSession = exports.resolveMaxConcurrentSubtasks = exports.lockPathsFor = exports.resolveMaxIterations = exports.stopReasonNotice = exports.executeAgentMemoryWrite = exports.AGENT_MEMORY_TOOL_SCHEMA = void 0;
     exports.dispatchSubAgent = dispatchSubAgent;
     exports.dispatchBackgroundSubAgent = dispatchBackgroundSubAgent;
     exports.dispatchForkedSubAgent = dispatchForkedSubAgent2;
@@ -26756,6 +26792,15 @@ var require_loop = __commonJS({
     Object.defineProperty(exports, "compactMessagesForResume", { enumerable: true, get: function() {
       return compaction_2.compactMessagesForResume;
     } });
+    Object.defineProperty(exports, "serverMessageBudgetFor", { enumerable: true, get: function() {
+      return compaction_2.serverMessageBudgetFor;
+    } });
+    Object.defineProperty(exports, "outputReserveFor", { enumerable: true, get: function() {
+      return compaction_2.outputReserveFor;
+    } });
+    Object.defineProperty(exports, "BYTES_PER_TOKEN_ESTIMATE", { enumerable: true, get: function() {
+      return compaction_2.BYTES_PER_TOKEN_ESTIMATE;
+    } });
     var iterationPolicy_1 = require_iterationPolicy();
     var iterationPolicy_2 = require_iterationPolicy();
     Object.defineProperty(exports, "emptyTurnBackoffMs", { enumerable: true, get: function() {
@@ -27176,7 +27221,7 @@ ${ctx}
       const autoCompact = (0, compaction_1.resolveAutoCompact)(options.autoCompact, settings.raw);
       const verifyNudgeOn = (0, compaction_1.resolveVerificationNudge)(settings.raw);
       const contextWindow = (0, modelCatalogue_1.liveContextWindowFor)(model, compaction_1.MODEL_CONTEXT_TOKENS[model] ?? 2e5);
-      const compactLimits = (0, compaction_1.compactionLimits)(contextWindow, settings.raw);
+      const compactLimits = (0, compaction_1.compactionLimits)(contextWindow, settings.raw, model);
       const readDedupe = new readDedupe_1.ReadDedupe(options.workDir);
       const drainBackgroundNotifications = () => depth === 0 ? options.backgroundAgents?.takeNotifications() ?? [] : [];
       const collectTaskEventTexts = async () => {
@@ -27213,6 +27258,9 @@ ${extra}` : extra;
         return texts;
       };
       let lastPromptTokens2 = 0;
+      let serverTrimmed = false;
+      let serverTrimNoticeShown = false;
+      let contextWarnNoticeShown = false;
       let compacting = false;
       const hardCap = autoContinue ? Math.max(maxIterations, iterationPolicy_1.MAX_ITERATIONS_CEILING) : maxIterations;
       let stopReason = "unknown";
@@ -27243,8 +27291,9 @@ ${extra}` : extra;
             break;
           }
           let bodyBytes = (0, compaction_1.estimateBodyBytes)(messages);
-          const prunePressure = lastPromptTokens2 > compactLimits.prune;
-          const tokenPressure = lastPromptTokens2 > compactLimits.compact;
+          const localTokens = Math.ceil(bodyBytes / compaction_1.BYTES_PER_TOKEN_ESTIMATE);
+          const prunePressure = lastPromptTokens2 > compactLimits.prune || localTokens > compactLimits.prune;
+          const tokenPressure = lastPromptTokens2 > compactLimits.compact || localTokens > compactLimits.compact || serverTrimmed;
           let bytePressure = bodyBytes > compaction_1.MAX_BODY_BYTES;
           if (autoCompact && !compacting && (prunePressure || bytePressure) && messages.length > compaction_1.PRUNE_KEEP_RECENT + 2) {
             const floor = depth === 0 ? (0, compaction_1.pruneReclaimFloor)(compaction_1._lastApiCallEndedAt.get(options.sessionId || subAgentBudget_1.PROCESS_BUDGET_KEY)) : compaction_1.PRUNE_MIN_RECLAIM_BYTES;
@@ -27275,9 +27324,12 @@ ${extra}` : extra;
               const reclaimed = bytesBefore - bytesAfter;
               if (did) {
                 lastPromptTokens2 = 0;
-                const reason = bytePressure ? `body ~${(bodyBytes / (1024 * 1024)).toFixed(1)}MB` : "context window";
+                serverTrimmed = false;
                 bodyBytes = bytesAfter;
                 bytePressure = bodyBytes > compaction_1.MAX_BODY_BYTES;
+                if (isMainAgent) {
+                  (options.onNotice ?? options.onText)(`\u267B\uFE0F Auto-compacted this chat's earlier history to keep it within the context budget.`);
+                }
               }
               if (unavailable) {
               } else if (did && reclaimed >= compaction_1.COMPACT_MIN_RECLAIM_BYTES) {
@@ -27340,6 +27392,24 @@ ${extra}` : extra;
               case "runtime_context":
                 pendingRuntimeContext = event.text;
                 break;
+              case "context_warning": {
+                if (event.reason === "overflow")
+                  break;
+                if (event.reason === "history-trim") {
+                  serverTrimmed = true;
+                  if (isMainAgent && !serverTrimNoticeShown) {
+                    serverTrimNoticeShown = true;
+                    const what = typeof event.trimmed === "number" && event.trimmed > 0 ? `${event.trimmed} oldest message(s)` : "the oldest turns";
+                    (options.onNotice ?? options.onText)(`\u26A0\uFE0F This chat exceeds what ${model} can carry alongside its output budget \u2014 the server dropped ${what} from what the model saw this turn. ` + (autoCompact ? "Auto-compaction will summarise the history at the next step so it stops recurring." : "Enable autoCompact or run /compact to re-summarise the history."));
+                  }
+                  break;
+                }
+                if (isMainAgent && !contextWarnNoticeShown && event.message) {
+                  contextWarnNoticeShown = true;
+                  (options.onNotice ?? options.onText)(`\u26A0\uFE0F ${event.message}`);
+                }
+                break;
+              }
               case "notice":
                 options.onNotice?.(event.message);
                 break;
@@ -106651,6 +106721,8 @@ ${text}` : "");
   void (0, import_code_core65.getModelCatalogue)().then((cat) => {
     if (!cat || !cat.isFreeTier || !cat.defaultModel)
       return;
+    if (!cat.models.some((m2) => m2.id === cat.defaultModel))
+      return;
     if (modelAlias === cat.defaultModel)
       return;
     const prev = resolveModelLabel(modelAlias);
@@ -106666,6 +106738,8 @@ ${text}` : "");
   void (0, import_code_core65.getModelCatalogue)().then((cat) => {
     if (!cat || cat.isFreeTier || !cat.defaultModel)
       return;
+    if (!cat.models.some((m2) => m2.id === cat.defaultModel))
+      return;
     if (options.modelExplicit !== false)
       return;
     if (agentRunning)
@@ -109841,7 +109915,7 @@ ${text}
           }
           const measure = measureContext(messages, parsed.all ? TOP_ALL : TOP_DEFAULT);
           const window2 = (0, import_code_core65.contextWindowFor)(modelAlias);
-          const thresholds = (0, import_code_core65.compactionThresholds)(window2);
+          const thresholds = (0, import_code_core65.compactionThresholds)(window2, (0, import_code_core67.loadSettings)(workDir).raw, modelAlias);
           for (const line of formatContextPanel({
             measure,
             window: window2,

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "nexrall-code",
-  "version": "0.5.199",
+  "version": "0.5.200",
   "description": "Nexrall Code — AI coding assistant for your terminal (headless agent for scripts, CI and automation)",
   "keywords": [
     "ai",
@@ -13,7 +13,7 @@
   ],
   "homepage": "https://nexrall.com",
   "bugs": {
-    "url": "https://github.com/Maxrall-Inc/nexrall-issue/issues"
+    "url": "https://github.com/Maxrall-Inc/nexrall-issues/issues"
   },
   "license": "MIT",
   "type": "module",