blun-king-cli 9.1.497 → 9.1.499

package/CHANGELOG.md

@@ -1,5 +1,22 @@
 # Changelog
 
+## 9.1.499 - 2026-08-30
+
+- Replaces image, audio, and video payloads older than the newest twelve conversation messages with short model-facing references while preserving the complete original session wire and every recent media attachment.
+- Reduces Fredrik's measured repeated model input by 85,748 characters, approximately 21,437 estimated tokens per subsequent work step, without changing the current image, message text, or persisted history.
+- Runs overdue TodoList maintenance as its own low-effort model step before the ninth work batch and exposes only TodoList during that step, preventing the normal workflow from hitting a visible red policy error first.
+- Keeps fresh user steers ahead of TodoList housekeeping while preserving the hard policy gate as a fallback for an actually invalid update.
+
+## 9.1.498 - 2026-08-30
+
+- Shows prompt-cache hit rate plus cache-read and cache-write tokens per model in `/usage`.
+- Adds `/tokens` as a short alias for the existing detailed context doctor without introducing another model call.
+- Adds `/offload` as an alias for the existing archive-backed context compaction command.
+- Preserves every distinct authorized Telegram bot message instead of dropping same-sender messages that arrive within a fixed 15-second window.
+- Preserves the exact TodoList through full compaction and adds a one-shot recovery reminder to read the archived history path before broad searches when an exact path, command, hash, or requirement is uncertain.
+- Keeps completed TodoList evidence append-only: later uncertainty becomes a separate in-progress step instead of silently reverting or removing completed work.
+- Defers recurring session-loop wakeups while the maintained TodoList still contains active or pending work, coalescing missed intervals until the assignment is genuinely idle without delaying unrelated scheduled reminders.
+
 ## 9.1.497 - 2026-08-30
 
 - Prioritizes Telegram bot messages from shared groups behind direct and explicitly addressed traffic but ahead of ordinary queued context, without forcing a reply.

package/LIESMICH.txt

@@ -9,7 +9,7 @@ Installation
 ------------
 Die geprüfte Version exakt global installieren:
 
-  npm install -g blun-king-cli@9.1.497
+  npm install -g blun-king-cli@9.1.499
 
 Start
 -----
@@ -65,6 +65,36 @@ geschriebenen Entwurf zu löschen. Das gilt auch bei
 Autovervollständigung, Geistervorschlägen, Bash-Eingabe und einer offenen
 Mehrzeileneingabe.
 
+Version 9.1.499 ersetzt Bild-, Audio- und Videodaten ausserhalb der neuesten zwoelf
+Gespraechsnachrichten ausschliesslich in der wiederholten Modellprojektion durch
+kurze Verweise. Aktuelle Medien, Nachrichtentext und vollstaendiger Sitzungsrohverlauf
+bleiben unveraendert. In Fredriks gemessenem Wiederaufnahme-Checkpoint spart der
+Schnitt 85.748 Zeichen beziehungsweise rund 21.437 geschaetzte Eingabetoken je
+weiterem Werkzeugschritt.
+
+Faellige Todo-Pflege laeuft in 9.1.499 vor dem naechsten Sachwerkzeug als eigener
+kleiner Modellschritt. In diesem Schritt bietet die Laufzeit ausschliesslich
+TodoList an; erst nach einer gueltigen Aktualisierung werden Datei-, Such- und
+Shell-Werkzeuge wieder freigegeben. Dadurch entsteht im normalen Ablauf kein
+roter Zwischenfehler. Frische Benutzernachrichten behalten Vorrang.
+
+Version 9.1.498 zeigt in /usage die gemessene Prompt-Cache-Trefferquote sowie
+gelesene und geschriebene Cache-Token je Modell. /tokens ist ein kurzer Alias
+fuer die bereits vorhandene detaillierte Kontextdiagnose; /offload ruft die
+vorhandene archivierende Kontextverdichtung auf. Unterschiedliche autorisierte
+Botnachrichten werden nicht mehr durch ein starres 15-Sekunden-Fenster verworfen.
+Nach einer vollständigen Verdichtung bleibt die exakte Todo-Liste erhalten. Ein
+einmaliger Wiederherstellungshinweis verlangt bei unsicheren Pfaden, Befehlen,
+Hashes oder Anforderungen zuerst das Lesen des archivierten Verlaufs, bevor eine
+breite Dateisuche oder ein Neubeginn zulässig ist.
+Erledigte Todo-Schritte bleiben als Nachweis erhalten. Neue widersprüchliche
+Erkenntnisse werden als eigener laufender Schritt mit Begründung ergänzt, statt
+einen bereits belegten Abschluss still zurückzustufen oder zu entfernen.
+Wiederkehrende Session-Loops warten, solange die gepflegte Todo-Liste noch
+laufende oder ausstehende Arbeit enthält. Verpasste Intervalle werden bis zum
+echten Leerlauf zusammengefasst; andere geplante Erinnerungen bleiben davon
+unberührt.
+
 Version 9.1.497 priorisiert Telegram-Botnachrichten aus gemeinsamen Gruppen
 hinter privaten und ausdrücklich adressierten Nachrichten, aber vor normalem
 Kontext. Die Priorisierung erzwingt keine Antwort. Eine angenommene

package/README.md

@@ -9,7 +9,7 @@ Voraussetzung ist Node.js 24.15 oder neuer. Die geprüfte Version wird exakt
 installiert:
 
 ```powershell
-npm install -g blun-king-cli@9.1.497
+npm install -g blun-king-cli@9.1.499
 ```
 
 ## Reproduzierbares Staging und Packen
@@ -81,6 +81,36 @@ konkurrierenden Start zurückgewiesen und bleiben an der Spitze der
 FIFO-Warteschlange. Für diesen normalen Wartestatus erscheint kein
 `turn.agent_busy`-Fehler.
 
+Version 9.1.499 ersetzt Bild-, Audio- und Videodaten außerhalb der neuesten zwölf
+Gesprächsnachrichten ausschließlich in der wiederholten Modellprojektion durch
+kurze Verweise. Aktuelle Medien, Nachrichtentext und vollständiger Sitzungsrohverlauf
+bleiben unverändert. In Fredriks gemessenem Wiederaufnahme-Checkpoint spart der
+Schnitt 85.748 Zeichen beziehungsweise rund 21.437 geschätzte Eingabetoken je
+weiterem Werkzeugschritt.
+
+Faellige Todo-Pflege laeuft in 9.1.499 vor dem naechsten Sachwerkzeug als eigener
+kleiner Modellschritt. In diesem Schritt bietet die Laufzeit ausschliesslich
+`TodoList` an; erst nach einer gueltigen Aktualisierung werden Datei-, Such- und
+Shell-Werkzeuge wieder freigegeben. Dadurch entsteht im normalen Ablauf kein
+roter Zwischenfehler. Frische Benutzernachrichten behalten Vorrang.
+
+Version 9.1.498 zeigt in `/usage` die gemessene Prompt-Cache-Trefferquote sowie
+gelesene und geschriebene Cache-Token je Modell. `/tokens` ist ein kurzer Alias
+fuer die bereits vorhandene detaillierte Kontextdiagnose; `/offload` ruft die
+vorhandene archivierende Kontextverdichtung auf. Unterschiedliche autorisierte
+Botnachrichten werden nicht mehr durch ein starres 15-Sekunden-Fenster verworfen.
+Nach einer vollständigen Verdichtung bleibt die exakte Todo-Liste erhalten. Ein
+einmaliger Wiederherstellungshinweis verlangt bei unsicheren Pfaden, Befehlen,
+Hashes oder Anforderungen zuerst das Lesen des archivierten Verlaufs, bevor eine
+breite Dateisuche oder ein Neubeginn zulässig ist.
+Erledigte Todo-Schritte bleiben als Nachweis erhalten. Neue widersprüchliche
+Erkenntnisse werden als eigener laufender Schritt mit Begründung ergänzt, statt
+einen bereits belegten Abschluss still zurückzustufen oder zu entfernen.
+Wiederkehrende Session-Loops warten außerdem, solange die gepflegte Todo-Liste
+noch laufende oder ausstehende Arbeit enthält. Verpasste Intervalle werden bis
+zum echten Leerlauf zusammengefasst; andere geplante Erinnerungen bleiben davon
+unberührt.
+
 Version 9.1.497 priorisiert Telegram-Botnachrichten aus gemeinsamen Gruppen
 hinter privaten und ausdrücklich adressierten Nachrichten, aber vor normalem
 Kontext. Die Priorisierung erzwingt keine Antwort. Eine angenommene

package/bin/historical-media-projection-policy.cjs

@@ -0,0 +1,48 @@
+'use strict';
+
+const HISTORICAL_MEDIA_KEEP_RECENT_MESSAGES = 12;
+const HISTORICAL_MEDIA_PLACEHOLDERS = Object.freeze({
+  image_url: '[historical image omitted from active model context; original remains in session wire]',
+  audio_url: '[historical audio omitted from active model context; original remains in session wire]',
+  video_url: '[historical video omitted from active model context; original remains in session wire]',
+});
+
+function projectHistoricalMediaParts(messages, options = {}) {
+  if (!Array.isArray(messages) || messages.length === 0) return messages;
+  const keepRecentMessages = nonNegativeInteger(
+    options.keepRecentMessages,
+    HISTORICAL_MEDIA_KEEP_RECENT_MESSAGES,
+  );
+  const historicalEnd = Math.max(0, messages.length - keepRecentMessages);
+  let changed = false;
+
+  const projected = messages.map((message, messageIndex) => {
+    if (messageIndex >= historicalEnd || !Array.isArray(message?.content)) return message;
+    let nextContent;
+    for (let partIndex = 0; partIndex < message.content.length; partIndex += 1) {
+      const part = message.content[partIndex];
+      const placeholder = HISTORICAL_MEDIA_PLACEHOLDERS[part?.type];
+      if (placeholder === undefined) {
+        nextContent?.push(part);
+        continue;
+      }
+      nextContent ??= message.content.slice(0, partIndex);
+      nextContent.push({ type: 'text', text: placeholder });
+      changed = true;
+    }
+    return nextContent === undefined ? message : { ...message, content: nextContent };
+  });
+
+  return changed ? projected : messages;
+}
+
+function nonNegativeInteger(value, fallback) {
+  const number = Number(value);
+  return Number.isInteger(number) && number >= 0 ? number : fallback;
+}
+
+module.exports = {
+  HISTORICAL_MEDIA_KEEP_RECENT_MESSAGES,
+  HISTORICAL_MEDIA_PLACEHOLDERS,
+  projectHistoricalMediaParts,
+};

package/bin/usage-cache-efficiency-policy.cjs

@@ -0,0 +1,26 @@
+'use strict';
+
+function wholeNonNegative(value) {
+  if (!Number.isFinite(value) || value <= 0) return 0;
+  return Math.floor(value);
+}
+
+function buildCacheEfficiency(usage = {}) {
+  const uncachedTokens = wholeNonNegative(usage.inputOther);
+  const readTokens = wholeNonNegative(usage.inputCacheRead);
+  const writeTokens = wholeNonNegative(usage.inputCacheCreation);
+  const inputTokens = uncachedTokens + readTokens + writeTokens;
+
+  return {
+    hasActivity: readTokens > 0 || writeTokens > 0,
+    hitRate: inputTokens > 0 ? Math.min(1, readTokens / inputTokens) : 0,
+    inputTokens,
+    readTokens,
+    writeTokens,
+  };
+}
+
+module.exports = {
+  buildCacheEfficiency,
+  wholeNonNegative,
+};

package/blun.mjs

@@ -75375,6 +75375,7 @@ var init_full = __esmMin((() => {
 		"permission_mode",
 		"plan_mode",
 		"plugin_session_start",
+		"compaction_recovery",
 		"todo_list_reminder",
 		"tool-awareness"
 	]);
@@ -77053,6 +77054,7 @@ function createCronScheduler(opts) {
 		const now = clocks.wallNow();
 		try {
 			for (const task of tasks) try {
+				if (!isIdle(task)) continue;
 				if (inFlight.has(task.id)) continue;
 				if (task.paused === true) continue;
 				const parsed = getParsed(task.cron);
@@ -77180,6 +77182,10 @@ var init_telemetry_events = __esmMin((() => {
 }));
 //#endregion
 //#region ../../packages/agent-core/src/agent/cron/manager.ts
+function hasUnfinishedTodoWork(agent) {
+	const todos = agent.tools.storeData()["todo"];
+	return Array.isArray(todos) && todos.some((todo) => todo?.status === "in_progress" || todo?.status === "pending");
+}
 var STALE_THRESHOLD_MS, CronManager;
 var init_manager$3 = __esmMin((() => {
 	init_clock();
@@ -77239,7 +77245,7 @@ var init_manager$3 = __esmMin((() => {
 			this.scheduler = createCronScheduler({
 				clocks: this.clocks,
 				source: () => this.store.list(),
-				isIdle: () => !agent.turn.hasActiveTurn,
+				isIdle: (task) => !agent.turn.hasActiveTurn && (task?.owner !== "session-loop" || !hasUnfinishedTodoWork(agent)),
 				isKilled: () => process.env["BLUN_DISABLE_CRON"] === "1",
 				onFire: (task, ctx) => {
 					this.handleFire(task, ctx);
@@ -79043,7 +79049,8 @@ var init_context$2 = __esmMin((() => {
 		project(messages, options) {
 			const anomalies = [];
 			const userOffloaded = this.agent.userMessageOffload.compact(messages);
-			const historicalUserReferencesProjected = compactHistoricalPersistedUserMessageReferences(userOffloaded);
+			const historicalMediaProjected = projectHistoricalMediaParts(userOffloaded);
+			const historicalUserReferencesProjected = compactHistoricalPersistedUserMessageReferences(historicalMediaProjected);
 			const duplicateUserMessagesProjected = dedupeRepeatedUserMessages(historicalUserReferencesProjected);
 			const historicalTelegramProjected = projectHistoricalUnaddressedTelegramMessages(duplicateUserMessagesProjected);
 			const assistantOffloaded = this.agent.assistantMessageOffload.compact(historicalTelegramProjected);
@@ -233121,7 +233128,7 @@ var init_mission_contract_bridge = __esmMin((() => {
 }));
 //#endregion
 //#region ../../packages/agent-core/src/agent/injection/manager.ts
-var ACTIVE_BACKGROUND_TASK_GUIDANCE, InjectionManager;
+var ACTIVE_BACKGROUND_TASK_GUIDANCE, POST_COMPACTION_RECOVERY_GUIDANCE, InjectionManager;
 var init_manager$2 = __esmMin((() => {
 	init_task_list();
 	init_action_style();
@@ -233136,6 +233143,7 @@ var init_manager$2 = __esmMin((() => {
 	init_validated_learning_signal();
 	init_tool_awareness();
 	ACTIVE_BACKGROUND_TASK_GUIDANCE = "The conversation was compacted, so the earlier messages that started these background tasks are gone — but the tasks are still running from before. Do not start duplicates. Use TaskOutput to fetch a task’s result, TaskList to list them, TaskUpdate to send additional instructions to a running agent task, and TaskStop to cancel one.";
+	POST_COMPACTION_RECOVERY_GUIDANCE = "The conversation was compacted. Continue from the persisted TodoList and summary. If an exact file path, command, hash, requirement, or other working detail is missing or uncertain, use Read on the summary's history_path before broad filesystem searches, guessing, or starting work over.";
 	InjectionManager = class {
 		agent;
 		injectors;
@@ -233193,9 +233201,16 @@ var init_manager$2 = __esmMin((() => {
 		}
 		async injectAfterCompaction() {
 			await this.injectGoal();
+			this.injectCompactionRecoveryGuidance();
 			this.injectActiveBackgroundTasks();
 			await this.inject();
 		}
+		injectCompactionRecoveryGuidance() {
+			this.agent.context.appendSystemReminder(POST_COMPACTION_RECOVERY_GUIDANCE, {
+				kind: "injection",
+				variant: "compaction_recovery"
+			});
+		}
 		/**
 		* Post-compaction only: re-surface still-running background tasks. Folding the
 		* live context to [recent user prompts, summary] drops the messages that
@@ -261265,7 +261280,7 @@ function renderPersistedUserMessage(text, outputPath) {
 		createUserMessagePreview(text)
 	].join("\n");
 }
-var USER_MESSAGE_MAX_CHARS, USER_MESSAGE_OFFLOAD_MARKER, shouldOffloadUserMessage, shouldOffloadHistoricalUserMessage, createUserMessagePreview, compactHistoricalPersistedUserMessageReferences;
+var USER_MESSAGE_MAX_CHARS, USER_MESSAGE_OFFLOAD_MARKER, shouldOffloadUserMessage, shouldOffloadHistoricalUserMessage, createUserMessagePreview, compactHistoricalPersistedUserMessageReferences, projectHistoricalMediaParts;
 var init_user_message_offload = __esmMin((() => {
 	const policy = createRequire(import.meta.url)("./bin/user-message-offload-policy.cjs");
 	USER_MESSAGE_MAX_CHARS = policy.USER_MESSAGE_MAX_CHARS;
@@ -261273,8 +261288,9 @@ var init_user_message_offload = __esmMin((() => {
 	shouldOffloadUserMessage = policy.shouldOffloadUserMessage;
 	shouldOffloadHistoricalUserMessage = policy.shouldOffloadHistoricalUserMessage;
 	createUserMessagePreview = policy.createUserMessagePreview;
-	compactHistoricalPersistedUserMessageReferences = policy.compactHistoricalPersistedUserMessageReferences;
-}));
+		compactHistoricalPersistedUserMessageReferences = policy.compactHistoricalPersistedUserMessageReferences;
+		({ projectHistoricalMediaParts } = createRequire(import.meta.url)("./bin/historical-media-projection-policy.cjs"));
+	}));
 var UserMessageOffload = class {
 	agent;
 	replacements = /* @__PURE__ */ new Map();
@@ -262770,6 +262786,24 @@ var init_turn = __esmMin((() => {
 								currentStepHadFailure = false;
 								const pendingSteers = this.matchingSteers(turnId);
 								if (pendingSteers.length === 0) {
+									const todoMaintenanceMode = goalTodoMaintenanceMode(this.agent);
+									if (todoMaintenanceMode !== null) {
+										const todoTool = eligibleTools.find((tool) => tool.name === "TodoList");
+										if (todoTool === void 0) return {
+											block: true,
+											reason: "TodoList maintenance is due, but the TodoList tool is unavailable."
+										};
+										this.agent.log.info("proactive TodoList maintenance step", {
+											stepNumber,
+											mode: todoMaintenanceMode,
+											workCallsSinceRefresh: this.agent.goalTodoPolicyState?.workCallsSinceRefresh ?? 0
+										});
+										const todoSystemPrompt = buildGoalTodoMaintenanceSystemPrompt(this.agent, turnSystemPrompt ?? this.agent.effectiveSystemPrompt, todoMaintenanceMode);
+										return {
+											llm: this.agent.llmForTurn("low", todoSystemPrompt),
+											tools: [todoTool]
+										};
+									}
 									if (workStepThinkingEffort === void 0) return;
 									this.agent.log.info("work step thinking effort", {
 										stepNumber,
@@ -403833,7 +403867,7 @@ const BUILTIN_SLASH_COMMAND_DEFINITIONS = [
 	},
 	{
 		name: "compact",
-		aliases: [],
+		aliases: ["offload"],
 		descriptionKey: "command.compact.description",
 		priority: 80,
 		argumentHint: "<instruction>"
@@ -403976,7 +404010,7 @@ const BUILTIN_SLASH_COMMAND_DEFINITIONS = [
 	},
 	{
 		name: "context-doctor",
-		aliases: [],
+		aliases: ["tokens"],
 		descriptionKey: "contextDoctor.description",
 		priority: 66,
 		availability: "always"
@@ -417895,9 +417929,11 @@ function buildSessionUsageSection(usage, error, value, muted, errorStyle) {
 	for (const [model, row] of entries) {
 		const input = usageInputTotal$2(row);
 		const output = usageNumber(row.output);
+		const cache = buildCacheEfficiency(row);
 		totalInput += input;
 		totalOutput += output;
 		lines.push(`  ${muted(sanitizeTerminalText(model))}  ${uiText("usage.input")} ${value(formatExactTokenCount(input, locale))}  ${uiText("usage.output")} ${value(formatExactTokenCount(output, locale))}  ${uiText("usage.total")} ${value(formatExactTokenCount(input + output, locale))}`);
+		if (cache.hasActivity) lines.push(`    ${muted("CACHE")} ${value(`${Math.round(cache.hitRate * 100)}%`)} | R ${value(formatExactTokenCount(cache.readTokens, locale))} | W ${value(formatExactTokenCount(cache.writeTokens, locale))}`);
 	}
 	if (entries.length > 1) lines.push(`  ${muted(uiText("usage.total"))}  ${uiText("usage.input")} ${value(formatExactTokenCount(totalInput, locale))}  ${uiText("usage.output")} ${value(formatExactTokenCount(totalOutput, locale))}  ${uiText("usage.total")} ${value(formatExactTokenCount(totalInput + totalOutput, locale))}`);
 	return lines;
@@ -417964,6 +418000,7 @@ var { buildApprovalRejectionStop } = createRequire(import.meta.url)("./bin/appro
 var { reconcileContextBudget } = createRequire(import.meta.url)("./bin/context-budget-ledger.cjs");
 var { buildContextInsight } = createRequire(import.meta.url)("./bin/context-insight-policy.cjs");
 var { buildContextDoctor } = createRequire(import.meta.url)("./bin/context-doctor-policy.cjs");
+var { buildCacheEfficiency } = createRequire(import.meta.url)("./bin/usage-cache-efficiency-policy.cjs");
 var { buildValidatedLearningInsight } = createRequire(import.meta.url)("./bin/validated-learning-insight-policy.cjs");
 var { readValidatedLearningCandidates } = createRequire(import.meta.url)("./bin/validated-learning-signal.cjs");
 function buildUsageReportLines(options) {
@@ -423177,6 +423214,7 @@ async function handleGoalCommand(host, args) {
 }
 const IDEA_CONTRACT_MARKER = "Work as a self-directing employee:";
 const IDEA_ALLOWED_CHANNELS = Object.freeze([]);
+const GOAL_TODO_REFRESH_WORK_CALL_LIMIT = 8;
 const IDEA_TERMINAL_TODO_STATUSES = Object.freeze([
 	"done",
 	"blocked",
@@ -423204,6 +423242,30 @@ function validInitialIdeaPlan(value) {
 	}
 	return active === 1;
 }
+function goalTodoMaintenanceMode(agent) {
+	if (isIdeaGoal(agent)) return null;
+	const todos = ideaTodos(agent);
+	if (todos.length === 0) return agent.goal.getActiveGoal() === null ? null : "initial";
+	const progress = agent.goalTodoPolicyState ??= {
+		workCallsSinceRefresh: 0,
+		refreshRequired: false
+	};
+	return progress.refreshRequired || progress.workCallsSinceRefresh >= GOAL_TODO_REFRESH_WORK_CALL_LIMIT ? "refresh" : null;
+}
+function buildGoalTodoMaintenanceSystemPrompt(agent, basePrompt, mode) {
+	const todos = ideaTodos(agent);
+	const visibleTodoList = todos.length === 0 ? "The visible TodoList is empty." : ["Current visible TodoList:", ...todos.map((todo) => `- [${String(todo?.status ?? "pending")}] ${String(todo?.title ?? "").trim()}`)].join("\n");
+	const action = mode === "initial" ? "Create the visible task-specific TodoList now. Keep exactly one item in_progress and every later item pending." : "Update the visible TodoList now, before any more task work. Mark only evidenced steps done, keep the actual current step in_progress, and leave later steps pending. If the current title is stale, refine it to the concrete verified substep.";
+	return `<todo-maintenance>
+TODO MAINTENANCE IS THE ONLY ACTION FOR THIS STEP.
+${action}
+Call TodoList as the first and only tool. Do not call a work, search, file, shell, web, messaging, or goal tool in this step. Do not merely describe the update.
+
+${visibleTodoList}
+</todo-maintenance>
+
+${basePrompt}`;
+}
 function enforceGoalTodoPolicy(agent, context) {
 	if (isIdeaGoal(agent)) return;
 	if (TELEGRAM_DELIVERY_TOOL_RE.test(String(context.toolCall.name ?? ""))) return;
@@ -423231,6 +423293,12 @@ function enforceGoalTodoPolicy(agent, context) {
 	if (context.toolCall.name === "TodoList") {
 		const nextTodos = context.args?.todos;
 		if (!Array.isArray(nextTodos)) return;
+		const normalizeTitle = (value) => String(value ?? "").trim().toLowerCase().replace(/\s+/gu, " ");
+		const missingCompleted = todos.filter((todo, index) => todo?.status === "done" && !nextTodos.some((nextTodo) => nextTodo?.status === "done" && normalizeTitle(nextTodo.title) === normalizeTitle(todo.title)) && nextTodos[index]?.status !== "done");
+		if (missingCompleted.length > 0) return {
+			block: true,
+			reason: `Completed TodoList evidence is append-only. Keep each completed item present and done. If later evidence reopens work, retain the completed item and add a separate in_progress item with the reason. Missing completed item: ${String(missingCompleted[0]?.title ?? "").trim()}`
+		};
 		if (nextTodos.length === 0) {
 			if (activeGoal !== null) return {
 				block: true,
@@ -423272,7 +423340,7 @@ function enforceGoalTodoPolicy(agent, context) {
 	}
 	if (context.toolCall.id === context.toolCalls[0]?.id) {
 		progress.workCallsSinceRefresh += 1;
-		if (progress.workCallsSinceRefresh > 8) progress.refreshRequired = true;
+		if (progress.workCallsSinceRefresh > GOAL_TODO_REFRESH_WORK_CALL_LIMIT) progress.refreshRequired = true;
 	}
 	if (!progress.refreshRequired) return;
 	return {

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "blun-king-cli",
-  "version": "9.1.497",
+  "version": "9.1.499",
   "description": "BLUN CLI - your own AI agent with a Telegram channel. Get it done. With BLUN.",
   "license": "MIT",
   "bin": {

package/telegram-plugin/dist/bridge.mjs

@@ -130,25 +130,6 @@ function matchesAny(text, patterns) {
 	} catch {}
 	return false;
 }
-/**
-* Bot-to-bot loop guard (Telegram's own safety requirement for Bot API 10.0
-* bot-to-bot mode): once King can receive other bots' messages, an unthrottled
-* exchange could ping-pong forever. We throttle deliveries FROM a given bot in
-* a given group to at most one per BOT_COOLDOWN_MS, so two bots can never fast-
-* loop. In-memory per bridge process — resets on restart, which is fine.
-*/
-const BOT_COOLDOWN_MS = 15e3;
-const lastBotDelivery = /* @__PURE__ */ new Map();
-/** Exposed for tests; drops stale entries so the map can't grow unbounded. */
-function botLoopGuardAllows(senderId, now = Date.now(), cooldownMs = BOT_COOLDOWN_MS) {
-	const last = lastBotDelivery.get(senderId);
-	if (last !== void 0 && now - last < cooldownMs) return false;
-	lastBotDelivery.set(senderId, now);
-	if (lastBotDelivery.size > 200) {
-		for (const [key, ts] of lastBotDelivery) if (now - ts > cooldownMs) lastBotDelivery.delete(key);
-	}
-	return true;
-}
 function gate(ctx, botUsername) {
 	const access = readAccess();
 	if (pruneExpired(access)) saveAccess(access);
@@ -227,7 +208,6 @@ function gate(ctx, botUsername) {
 		if (groupAllowFrom.length > 0 && !groupAllowFrom.includes(senderId)) return { action: "drop" };
 		const mentionBypass = from.isBot !== true && (policy.alwaysAllowFrom ?? []).includes(senderId);
 		const addressed = !matchesAny(ctx.text, access.ignorePatterns) && (mentionBypass || isMentioned(ctx, botUsername, access.mentionPatterns));
-		if (!addressed && from.isBot === true && !botLoopGuardAllows(senderId)) return { action: "drop" };
 		return {
 			action: "deliver",
 			access,