@songsid/agend 2.1.3 → 2.1.4-beta.10

package/dist/fleet-manager.js

@@ -4,7 +4,8 @@ import { createServer } from "node:http";
 import { join, dirname, basename } from "node:path";
 import { fileURLToPath } from "node:url";
 import { getAgendHome, ensureWorkspaceGit } from "./paths.js";
-import { clearUpdateMarker, isUpdateInProgress } from "./update-marker.js";
+import { beginUpdateProgress as persistUpdateProgress, clearUpdateMarker, isUpdateInProgress, readUpdateProgress, setUpdateProgressStage, } from "./update-marker.js";
+import { formatUpdateProgress } from "./update-progress.js";
 import { sdNotify, sdNotifyBlocking } from "./sd-notify.js";
 import { readFleetMemory } from "./process-memory.js";
 import { ReplyDeduper } from "./reply-dedup.js";
@@ -29,7 +30,7 @@ import { processAttachments } from "./channel/attachment-handler.js";
 import { routeToolCall } from "./channel/tool-router.js";
 import { Scheduler } from "./scheduler/index.js";
 import { DEFAULT_SCHEDULER_CONFIG } from "./scheduler/index.js";
-import { TopicCommands, saveCommandForBackend, parseSaveFilename, parsePauseWakeCommand, SAVE_FILENAME_RE, SAVE_UNSUPPORTED_MSG, resolveInstanceContext, forgetInstanceContext } from "./topic-commands.js";
+import { TopicCommands, saveCommandForBackend, parseSaveFilename, parsePauseWakeCommand, SAVE_FILENAME_RE, SAVE_UNSUPPORTED_MSG, CLEAR_UNSUPPORTED_MSG, resolveInstanceContext, forgetInstanceContext } from "./topic-commands.js";
 import { DailySummary } from "./daily-summary.js";
 import { WebhookEmitter } from "./webhook-emitter.js";
 import { TmuxControlClient } from "./tmux-control.js";
@@ -134,6 +135,8 @@ const CANCEL_BTN_LEDGER_FILE = "cancel-buttons.json";
  * exactly one progress message per turn, and it is the cancel button itself.
  */
 const PROGRESS_UPDATE_INTERVAL_MS = 60_000;
+/** Floor between tool-progress-driven bubble edits (Telegram flood safety). */
+const TOOL_PROGRESS_EDIT_MIN_MS = 4_000;
 /** Elapsed time is only shown once work has clearly outlasted a quick answer. */
 /**
  * Default delay before the button starts showing elapsed time. Configurable via
@@ -153,6 +156,12 @@ const PROGRESS_ACTIVITY_MAX_CHARS = 48;
  * 🫡 passes too — it reads as a deliberate acknowledgement, not plumbing.
  */
 const DELIVERY_STATUS_EMOJIS = new Set(["👀", "⏳", "✅", "❌"]);
+/**
+ * Reactions that are neither delivery plumbing nor meaningful conversational
+ * feedback. Keep this separate from DELIVERY_STATUS_EMOJIS so adding a UI-only
+ * emoji never changes the documented delivery-state protocol.
+ */
+const IGNORED_REACTION_EMOJIS = new Set(["📷"]);
 /**
  * How long a delivery waits out a disconnected instance IPC before giving up.
  *
@@ -165,6 +174,13 @@ const CLASSIC_BACKEND_SELECTION_TIMEOUT_MS = 60_000;
 const CLASSIC_BACKEND_CALLBACK_PREFIX = "classic-backend:";
 const MODEL_SELECT_CALLBACK_PREFIX = "model-select:";
 const EFFORT_SELECT_CALLBACK_PREFIX = "effort-select:";
+const INTERACTIVE_ASSIST_CALLBACK_PREFIX = "interactive-assist:";
+const EXIT_RESTART_CALLBACK_PREFIX = "exit-restart:";
+const HANG_CALLBACK_PREFIX = "hang:";
+const CLEAR_CONFIRM_CALLBACK_PREFIX = "clear-confirm:";
+const CLEAR_CONFIRM_TIMEOUT_MS = 15_000;
+/** Default lifetime for long-lived nonce prompts (clear overrides this to 15s). */
+const NONCE_BUTTON_TIMEOUT_MS = 15 * 60_000;
 const CLI_ENV_TTL_MS = 24 * 60 * 60 * 1000; // /model reads cached CLI env within 24h
 export class FleetManager {
     dataDir;
@@ -239,6 +255,8 @@ export class FleetManager {
     replyDeduper = new ReplyDeduper();
     /** instanceName → what it is doing right now, when the backend can tell us. */
     instanceActivity = new Map();
+    /** instanceName → this turn's tool list (multi-line), for the bubble. */
+    instanceProgress = new Map();
     /** instanceName → tail of deliveries waiting for its IPC to come back. */
     ipcWaitTails = new Map();
     /** instanceName → restart currently executing; concurrent callers join it. */
@@ -253,6 +271,8 @@ export class FleetManager {
     /** In-flight /effort selections, same coordinator shape as pendingModelSelects. */
     pendingEffortSelects = new Map();
     pendingModelSelects = new Map();
+    /** nonce → pending button prompt (hang restart, interactive assist, clean-exit restart). */
+    pendingNonceButtons = new Map();
     // Model failover state
     failoverActive = new Map(); // instance → current failover model
     // IPC reconnect: tracks instances being intentionally stopped (skip reconnect)
@@ -274,6 +294,9 @@ export class FleetManager {
     healthServer = null;
     healthPortRetried = false;
     updateCheckTimer = null;
+    updateProgressTimer = null;
+    updateProgressEditRunning = false;
+    lastUpdateProgressText = null;
     eventLogPruneTimer = null;
     logRotateTimer = null;
     /** Days of event/activity history to keep. */
@@ -553,6 +576,45 @@ export class FleetManager {
         }
         await data.respond(await this.topicCommands.runPauseWake(target, action));
     }
+    async handleUpdateSlash(data, adapterId) {
+        const allowed = this.fleetConfig?.channel?.access?.allowed_users ?? [];
+        if (allowed.length > 0 && !allowed.some(u => String(u) === String(data.userId))) {
+            await data.respond(t("not_authorized"));
+            return;
+        }
+        const messageId = await data.respond(t("update.progress.preparing", 0));
+        const adapter = this.adapters.get(adapterId) ?? this.adapter;
+        if (messageId && adapter) {
+            const chatId = String(this.getChannelConfig(adapterId)?.group_id ?? data.channelId);
+            this.beginUpdateProgress(adapter, chatId, data.channelId, messageId);
+        }
+        const { spawn } = await import("node:child_process");
+        const currentVersion = JSON.parse(readFileSync(join(dirname(fileURLToPath(import.meta.url)), "..", "package.json"), "utf-8")).version ?? "";
+        const command = currentVersion.includes("beta") ? "agend update --beta" : "agend update";
+        const child = spawn("sh", ["-c", `sleep 2 && ${command}`], { detached: true, stdio: "ignore" });
+        child.once("error", err => this.failUpdateProgress(err.message));
+        child.unref();
+    }
+    /** Admin-only full conversation reset for fleet-topic and Classic instances. */
+    async handleClearSlash(data, adapterId) {
+        if (!this.isModelAdmin(data.userId, data.channelId, adapterId)) {
+            await data.respond(t("permission.denied"));
+            return;
+        }
+        const name = this.resolveSlashTarget(data.channelId, adapterId);
+        if (!name) {
+            await data.respond(t("classic.no_agent"));
+            return;
+        }
+        const adapter = this.adapters.get(adapterId) ?? this.adapter;
+        if (!adapter) {
+            await data.respond(t("clear.prompt_unavailable"));
+            return;
+        }
+        const chatId = String(this.getChannelConfig(adapterId)?.group_id ?? data.channelId);
+        const fallback = await this.promptClearConfirmation(name, data.channelId, adapter, chatId, data.channelId);
+        await data.respond(fallback ?? t("clear.confirm_posted"));
+    }
     /** Get the adapter bound to an instance, falling back to primary adapter */
     getAdapterForInstance(name) {
         const worldId = this.getInstanceAdapterId(name);
@@ -1016,19 +1078,34 @@ export class FleetManager {
             this.logger.info({ name }, "Instance already running, skipping");
             return;
         }
+        const backend = config.backend ?? this.fleetConfig?.defaults?.backend ?? "claude-code";
         if (config.general_topic) {
             // antigravity (agy) does not read MCP instructions — fleet context and
             // routing instructions are not injected, so it cannot act as a dispatcher.
-            const backend = config.backend ?? this.fleetConfig?.defaults?.backend ?? "claude-code";
             if (backend === "antigravity") {
                 this.logger.warn({ name }, "antigravity backend does not support MCP instructions — general dispatcher will not work correctly");
                 this.notifyInstanceTopic(name, "⚠️ antigravity backend is not supported for General instances (no MCP instructions injection). Switch to claude-code or kiro-cli.");
             }
-            this.ensureGeneralInstructions(config.working_directory, config.backend);
+            this.ensureGeneralInstructions(config.working_directory, backend, name);
+        }
+        else if (kind === "fleet-topic") {
+            // Workers get only role-eligible on-demand skills. Classic instances are
+            // deliberately excluded: their workspace and conversation lifecycle are
+            // managed independently from fleet-topic workers.
+            try {
+                const skillsWorkDir = this.resolveKnowledgeWorkDir(config.working_directory, backend, name);
+                this.syncRoleSkills(skillsWorkDir, backend, "worker");
+            }
+            catch (err) {
+                // Skill publishing is additive. A read-only or temporarily unavailable
+                // workspace must not turn an otherwise valid worker startup into a
+                // fleet outage.
+                this.logger.warn({ err, name, backend }, "Failed to sync worker skills — continuing startup");
+            }
         }
         await this.lifecycle.start(name, config, topicMode, {
             kind,
-            backend: config.backend ?? this.fleetConfig?.defaults?.backend ?? "claude-code",
+            backend,
             model: this.resolveInstanceModel(name).display,
         });
         // Only clear a stale process status after a real start succeeded.  Clearing
@@ -1193,7 +1270,18 @@ export class FleetManager {
         this.instanceStateCache.delete(name);
         this.instanceProcessStatus.delete(name);
         this.lastDeliveryAt.delete(name);
-        return this.lifecycle.stop(name);
+        // A pending hang/exit/assist offer refers to the instance being torn down;
+        // left alone it would stay clickable for the rest of its 15 minutes. Clear
+        // again AFTER the stop completes: the teardown itself can emit hang /
+        // interactive-prompt / clean-exit events whose handlers post fresh prompts
+        // while the stop is still awaiting (the TOCTOU sol's review called out).
+        this.clearNoncePromptsForInstance(name);
+        try {
+            return await this.lifecycle.stop(name);
+        }
+        finally {
+            this.clearNoncePromptsForInstance(name);
+        }
     }
     /** Restart a single instance, reloading fleet.yaml first to pick up config changes. */
     async restartSingleInstance(name, opts) {
@@ -1293,6 +1381,65 @@ export class FleetManager {
     getDashboardAccess() {
         return { ready: this.healthServerListening, token: this.webToken };
     }
+    beginUpdateProgress(adapter, chatId, threadId, messageId) {
+        persistUpdateProgress(this.dataDir, {
+            adapterId: adapter.id,
+            chatId,
+            ...(threadId ? { threadId } : {}),
+            messageId,
+        });
+        this.lastUpdateProgressText = null;
+        this.startUpdateProgressMonitor(adapter);
+    }
+    failUpdateProgress(message) {
+        setUpdateProgressStage(this.dataDir, "failed", { error: message });
+    }
+    /** The old fleet edits CLI install stages; the new fleet adopts the same message below. */
+    startUpdateProgressMonitor(initialAdapter) {
+        if (this.updateProgressTimer)
+            clearInterval(this.updateProgressTimer);
+        const tick = async () => {
+            if (this.updateProgressEditRunning)
+                return;
+            const marker = readUpdateProgress(this.dataDir);
+            if (!marker) {
+                if (this.updateProgressTimer)
+                    clearInterval(this.updateProgressTimer);
+                this.updateProgressTimer = null;
+                return;
+            }
+            const target = marker.progress.target;
+            const adapter = this.adapters.get(target.adapterId) ?? (initialAdapter?.id === target.adapterId ? initialAdapter : undefined);
+            if (!adapter)
+                return;
+            const text = formatUpdateProgress(marker);
+            if (text === this.lastUpdateProgressText)
+                return;
+            this.updateProgressEditRunning = true;
+            try {
+                await adapter.editMessage(target.chatId, target.messageId, text, target.threadId);
+                this.lastUpdateProgressText = text;
+                if (marker.progress.stage === "failed" || marker.progress.stage === "complete") {
+                    clearUpdateMarker(this.dataDir);
+                    if (this.updateProgressTimer)
+                        clearInterval(this.updateProgressTimer);
+                    this.updateProgressTimer = null;
+                }
+            }
+            catch (err) {
+                this.logger.warn({ err }, "Failed to edit update progress");
+            }
+            finally {
+                this.updateProgressEditRunning = false;
+            }
+        };
+        void tick();
+        // Poll stages faster than the elapsed-time display. npm verification and
+        // service-file refresh can be short; 200ms prevents those stages from being
+        // skipped while the text cache still limits normal edits to once per second.
+        this.updateProgressTimer = setInterval(() => void tick(), 200);
+        this.updateProgressTimer.unref?.();
+    }
     /** Start all instances from fleet config */
     async startAll(configPath) {
         const startupStartedAt = Date.now();
@@ -1307,6 +1454,15 @@ export class FleetManager {
         rotateLogIfNeeded(join(this.dataDir, "fleet.log"));
         const fleet = this.loadConfig(configPath);
         setLocale(detectLocale(fleet)); // user-facing text language (fleet.yaml defaults.locale / timezone)
+        const savedUpdateProgress = readUpdateProgress(this.dataDir);
+        const pendingUpdateProgress = savedUpdateProgress
+            && savedUpdateProgress.progress.stage !== "failed"
+            && savedUpdateProgress.progress.stage !== "complete"
+            ? savedUpdateProgress
+            : null;
+        if (pendingUpdateProgress) {
+            setUpdateProgressStage(this.dataDir, "starting", { version: pendingUpdateProgress.progress.version });
+        }
         this.initializeWebAuthTokens();
         const topicMode = fleet.channel?.mode === "topic" || !!fleet.channels?.some(ch => ch.mode === "topic");
         // Set tmux socket isolation for custom AGEND_HOME
@@ -1354,9 +1510,8 @@ export class FleetManager {
         // routing does NOT go through the routing engine (single-key, can't hold two
         // bots in one channel) — it resolves per-bot via getInstanceByChannel.
         this.classicChannels = new ClassicChannelManager(this.dataDir, this.logger);
-        const primaryCh = fleet.channels?.[0] ?? fleet.channel;
-        if (primaryCh)
-            this.classicChannels.setPrimaryAdapterId(primaryCh.id ?? primaryCh.type);
+        const classicAdapters = fleet.channels?.length ? fleet.channels : (fleet.channel ? [fleet.channel] : []);
+        this.classicChannels.configureAdapters(classicAdapters);
         // Restore the persisted bot binding so replies/cancel go through the right
         // bot after a restart (before this, inbound would re-bind lazily).
         for (const ch of this.classicChannels.getAll()) {
@@ -1525,7 +1680,7 @@ export class FleetManager {
             const generalDir = join(getAgendHome(), name);
             mkdirSync(generalDir, { recursive: true });
             const backendName = fleet.defaults.backend ?? "claude-code";
-            this.ensureGeneralInstructions(generalDir, backendName);
+            this.ensureGeneralInstructions(generalDir, backendName, name);
             fleet.instances[name] = {
                 ...DEFAULT_INSTANCE_CONFIG,
                 working_directory: generalDir,
@@ -1563,7 +1718,7 @@ export class FleetManager {
         const allEntries = Object.entries(fleet.instances);
         const generals = allEntries.filter(([_, cfg]) => cfg.general_topic);
         const others = allEntries.filter(([_, cfg]) => !cfg.general_topic);
-        const startupProgress = new RestartProgress(this.runnableStartupCount(fleet, topicMode), startupStartedAt, this.logger);
+        const startupProgress = new RestartProgress(this.runnableStartupCount(fleet, topicMode), pendingUpdateProgress?.startedAt ?? startupStartedAt, this.logger, { mode: pendingUpdateProgress ? "update" : "restart" });
         if (generals.length > 0) {
             for (const [name, cfg] of generals) {
                 try {
@@ -1602,7 +1757,19 @@ export class FleetManager {
                     this.logger.error({ err }, "startSharedAdapter failed — fleet continues without some adapters");
                 }
             })();
-            progressStart = adapterStartup.then(() => startupProgress.start(this.restartProgressTarget()));
+            progressStart = adapterStartup.then(() => {
+                if (pendingUpdateProgress) {
+                    const saved = pendingUpdateProgress.progress.target;
+                    const adapter = this.adapters.get(saved.adapterId);
+                    const target = adapter ? {
+                        adapter,
+                        chatId: saved.chatId,
+                        threadId: saved.threadId,
+                    } : null;
+                    return startupProgress.resume(target, saved.messageId);
+                }
+                return startupProgress.start(this.restartProgressTarget());
+            });
         }
         // The systemd watchdog answers exactly one question: is this process still
         // turning its event loop? Pinging from a timer proves that, and after the
@@ -1980,31 +2147,20 @@ export class FleetManager {
             await this.handleInboundReaction(r);
         }, this.logger, "adapter.reaction"));
         this.adapter.on("callback_query", safeHandler(async (data) => {
+            if (await this.handleClearConfirmation(data, adapterId, this.adapter ?? undefined))
+                return;
+            if (await this.handleExitRestartPrompt(data, adapterId, this.adapter ?? undefined))
+                return;
+            if (await this.handleInteractivePromptAssist(data, adapterId, this.adapter ?? undefined))
+                return;
             if (await this.handleClassicBackendSelection(data))
                 return;
             if (await this.handleModelSelection(data))
                 return;
             if (await this.handleEffortSelection(data))
                 return;
-            if (data.callbackData.startsWith("hang:")) {
-                const parts = data.callbackData.split(":");
-                const action = parts[1];
-                const instanceName = parts[2];
-                if (action === "restart") {
-                    await this.stopInstance(instanceName);
-                    const config = this.fleetConfig?.instances[instanceName];
-                    if (config) {
-                        const topicMode = this.fleetConfig?.channel?.mode === "topic";
-                        await this.startInstance(instanceName, config, topicMode);
-                        // startInstance already calls connectIpcToInstance
-                    }
-                    this.adapter?.editMessage(data.chatId, data.messageId, `🔄 ${instanceName} restarted.`, data.threadId).catch(() => { });
-                }
-                else {
-                    this.adapter?.editMessage(data.chatId, data.messageId, `⏳ Continuing to wait for ${instanceName}.`, data.threadId).catch(() => { });
-                }
+            if (await this.handleHangPrompt(data, adapterId, this.adapter ?? undefined))
                 return;
-            }
             if (data.callbackData.startsWith("cancel:")) {
                 this.handleCancelClick(data.callbackData.slice("cancel:".length), this.adapter, data);
                 return;
@@ -2083,6 +2239,29 @@ export class FleetManager {
                 const result = await this.topicCommands.sendCompact(name);
                 await data.respond(result);
             }
+            else if (data.command === "steer") {
+                const name = this.resolveSlashTarget(data.channelId, adapterId);
+                if (!name) {
+                    await data.respond(t("classic.no_agent"));
+                    return;
+                }
+                const steerText = String(data.options?.message ?? "").trim();
+                if (!steerText) {
+                    await data.respond(t("steer.usage"));
+                    return;
+                }
+                // chat_id/message_id stay empty: a slash interaction has no channel
+                // message to react to, and an empty chat_id keeps updateLastChat from
+                // rerouting the instance's replies to the slash context.
+                const result = this.topicCommands.sendSteer(name, steerText, {
+                    chatId: "", messageId: "", username: data.username ?? "user",
+                    userId: data.userId ?? "", threadId: undefined, adapterId, source: "discord",
+                });
+                await data.respond(result);
+            }
+            else if (data.command === "clear") {
+                await this.handleClearSlash(data, adapterId);
+            }
             else if (data.command === "model") {
                 await this.handleModelSlash(data, adapterId);
             }
@@ -2135,17 +2314,7 @@ export class FleetManager {
                     : t("collab.off.classic"));
             }
             else if (data.command === "update") {
-                const allowed = this.fleetConfig?.channel?.access?.allowed_users ?? [];
-                if (allowed.length > 0 && !allowed.some(u => String(u) === String(data.userId))) {
-                    await data.respond(t("not_authorized"));
-                    return;
-                }
-                await data.respond(t("update.running"));
-                const { spawn } = await import("node:child_process");
-                const _cv = JSON.parse(readFileSync(join(dirname(fileURLToPath(import.meta.url)), "..", "package.json"), "utf-8")).version ?? "";
-                const _cmd = _cv.includes("beta") ? "agend update --beta" : "agend update";
-                const child = spawn("sh", ["-c", `sleep 2 && ${_cmd}`], { detached: true, stdio: "ignore" });
-                child.unref();
+                await this.handleUpdateSlash(data, adapterId);
             }
             else if (data.command === "doctor") {
                 const allowed = this.fleetConfig?.channel?.access?.allowed_users ?? [];
@@ -2214,6 +2383,26 @@ export class FleetManager {
                 const result = await this.topicCommands.sendCompact(name);
                 await data.respond(result);
             }
+            else if (data.command === "steer") {
+                const name = this.resolveSlashTarget(data.channelId, adapterId);
+                if (!name) {
+                    await data.respond(t("classic.no_agent"));
+                    return;
+                }
+                const steerText = String(data.options?.message ?? "").trim();
+                if (!steerText) {
+                    await data.respond(t("steer.usage"));
+                    return;
+                }
+                // chat_id/message_id stay empty: a slash interaction has no channel
+                // message to react to, and an empty chat_id keeps updateLastChat from
+                // rerouting the instance's replies to the slash context.
+                const result = this.topicCommands.sendSteer(name, steerText, {
+                    chatId: "", messageId: "", username: data.username ?? "user",
+                    userId: data.userId ?? "", threadId: undefined, adapterId, source: "discord",
+                });
+                await data.respond(result);
+            }
         }, this.logger, "adapter.slash_command"));
         await this.topicCommands.registerBotCommands().catch(e => this.logger.warn({ err: e }, "registerBotCommands failed (non-fatal)"));
         // Background-probe each backend's CLI env (version/models) → cli-env cache.
@@ -2290,30 +2479,20 @@ export class FleetManager {
             await this.handleInboundReaction(r);
         }, this.logger, `adapter[${adapterId}].reaction`));
         adapter.on("callback_query", safeHandler(async (data) => {
+            if (await this.handleClearConfirmation(data, adapterId, adapter))
+                return;
+            if (await this.handleExitRestartPrompt(data, adapterId, adapter))
+                return;
+            if (await this.handleInteractivePromptAssist(data, adapterId, adapter))
+                return;
             if (await this.handleClassicBackendSelection(data))
                 return;
             if (await this.handleModelSelection(data))
                 return;
             if (await this.handleEffortSelection(data))
                 return;
-            if (data.callbackData.startsWith("hang:")) {
-                const parts = data.callbackData.split(":");
-                const action = parts[1];
-                const instanceName = parts[2];
-                if (action === "restart") {
-                    await this.stopInstance(instanceName);
-                    const config = this.fleetConfig?.instances[instanceName];
-                    if (config) {
-                        const topicMode = this.fleetConfig?.channel?.mode === "topic";
-                        await this.startInstance(instanceName, config, topicMode);
-                    }
-                    adapter.editMessage(data.chatId, data.messageId, `🔄 ${instanceName} restarted.`, data.threadId).catch(() => { });
-                }
-                else {
-                    adapter.editMessage(data.chatId, data.messageId, `⏳ Continuing to wait for ${instanceName}.`, data.threadId).catch(() => { });
-                }
+            if (await this.handleHangPrompt(data, adapterId, adapter))
                 return;
-            }
             if (data.callbackData.startsWith("cancel:")) {
                 this.handleCancelClick(data.callbackData.slice("cancel:".length), adapter, data);
                 return;
@@ -2385,6 +2564,9 @@ export class FleetManager {
             else if (data.command === "model") {
                 await this.handleModelSlash(data, adapterId);
             }
+            else if (data.command === "clear") {
+                await this.handleClearSlash(data, adapterId);
+            }
             else if (data.command === "effort") {
                 await this.handleEffortSlash(data, adapterId);
             }
@@ -2434,17 +2616,7 @@ export class FleetManager {
                     : t("collab.off.classic"));
             }
             else if (data.command === "update") {
-                const allowed = this.fleetConfig?.channel?.access?.allowed_users ?? [];
-                if (allowed.length > 0 && !allowed.some(u => String(u) === String(data.userId))) {
-                    await data.respond(t("not_authorized"));
-                    return;
-                }
-                await data.respond(t("update.running"));
-                const { spawn } = await import("node:child_process");
-                const _cv = JSON.parse(readFileSync(join(dirname(fileURLToPath(import.meta.url)), "..", "package.json"), "utf-8")).version ?? "";
-                const _cmd = _cv.includes("beta") ? "agend update --beta" : "agend update";
-                const child = spawn("sh", ["-c", `sleep 2 && ${_cmd}`], { detached: true, stdio: "ignore" });
-                child.unref();
+                await this.handleUpdateSlash(data, adapterId);
             }
             else if (data.command === "doctor") {
                 const allowed = this.fleetConfig?.channel?.access?.allowed_users ?? [];
@@ -2513,6 +2685,26 @@ export class FleetManager {
                 const result = await this.topicCommands.sendCompact(name);
                 await data.respond(result);
             }
+            else if (data.command === "steer") {
+                const name = this.resolveSlashTarget(data.channelId, adapterId);
+                if (!name) {
+                    await data.respond(t("classic.no_agent"));
+                    return;
+                }
+                const steerText = String(data.options?.message ?? "").trim();
+                if (!steerText) {
+                    await data.respond(t("steer.usage"));
+                    return;
+                }
+                // chat_id/message_id stay empty: a slash interaction has no channel
+                // message to react to, and an empty chat_id keeps updateLastChat from
+                // rerouting the instance's replies to the slash context.
+                const result = this.topicCommands.sendSteer(name, steerText, {
+                    chatId: "", messageId: "", username: data.username ?? "user",
+                    userId: data.userId ?? "", threadId: undefined, adapterId, source: "discord",
+                });
+                await data.respond(result);
+            }
         }, this.logger, `adapter[${adapterId}].slash_command`));
         await adapter.start();
         if (channelConfig.group_id) {
@@ -2607,9 +2799,6 @@ export class FleetManager {
                     }
                     await this.handleOutboundFromInstance(name, msg);
                 }
-                else if (msg.type === "fleet_tool_status") {
-                    this.handleToolStatusFromInstance(name, msg);
-                }
                 else if (msg.type === "fleet_schedule_create" || msg.type === "fleet_schedule_list" ||
                     msg.type === "fleet_schedule_update" || msg.type === "fleet_schedule_delete") {
                     this.handleScheduleCrud(name, msg);
@@ -2633,6 +2822,9 @@ export class FleetManager {
                 else if (msg.type === "instance_activity") {
                     this.cacheInstanceActivity(name, msg.activity);
                 }
+                else if (msg.type === "instance_progress") {
+                    this.cacheInstanceProgress(name, msg.progress || null);
+                }
                 else if (msg.type === "instance_state" || msg.type === "instance_state_response") {
                     this.cacheInstanceExecutionState(name, msg);
                     if (msg.type === "instance_state_response") {
@@ -2821,6 +3013,10 @@ export class FleetManager {
             this.logger.debug({ emoji: r.emoji, user: r.username }, "Ignoring delivery-status emoji as a reaction");
             return;
         }
+        if (IGNORED_REACTION_EMOJIS.has(r.emoji)) {
+            this.logger.debug({ emoji: r.emoji, user: r.username }, "Ignoring non-contextual emoji reaction");
+            return;
+        }
         this.eventLog?.logActivity("reaction", r.username, `${r.emoji} ${r.action}`, instanceName);
         if (r.action === "add") {
             this.eventLog?.addReaction(instanceName, r.messageId, r.username, r.emoji);
@@ -3076,6 +3272,43 @@ export class FleetManager {
                     await msgAdapter?.sendText(chatId, result);
                     return;
                 }
+                // /steer — interject into the running turn. Not admin-gated: anyone who
+                // can talk to this agent can send it a message; steer only changes when
+                // it lands, and it keeps the full [user:] formatting (unlike /raw).
+                if (text === "/steer" || text.startsWith("/steer ") || text.startsWith("/steer@")) {
+                    const steerName = this.classicChannels.getInstanceByChannel(chatId, msg.adapterId);
+                    if (!steerName) {
+                        await msgAdapter?.sendText(chatId, t("classic.no_agent_start"));
+                        return;
+                    }
+                    const steerContent = text.replace(/^\/steer(@\S+)?/, "").trim();
+                    if (!steerContent) {
+                        await msgAdapter?.sendText(chatId, t("steer.usage"));
+                        return;
+                    }
+                    const result = this.topicCommands.sendSteer(steerName, steerContent, msg);
+                    await msgAdapter?.sendText(chatId, result);
+                    return;
+                }
+                // Handle /clear command (admin only) — unlike /compact this starts a
+                // fresh conversation and intentionally discards the current history.
+                if (text === "/clear" || text.startsWith("/clear@")) {
+                    if (!this.isModelAdmin(msg.userId, chatId, msg.adapterId)) {
+                        await msgAdapter?.sendText(chatId, t("cmd.admin_required", "/clear"));
+                        return;
+                    }
+                    const clearName = this.classicChannels.getInstanceByChannel(chatId, msg.adapterId);
+                    if (!clearName) {
+                        await msgAdapter?.sendText(chatId, t("classic.no_agent_start"));
+                        return;
+                    }
+                    if (!msgAdapter)
+                        return;
+                    const fallback = await this.promptClearConfirmation(clearName, chatId, msgAdapter, chatId);
+                    if (fallback)
+                        await msgAdapter.sendText(chatId, fallback);
+                    return;
+                }
                 // Handle /cancel command
                 if (text === "/cancel" || text.startsWith("/cancel@")) {
                     const cancelName = this.classicChannels.getInstanceByChannel(chatId, msg.adapterId);
@@ -3466,7 +3699,7 @@ export class FleetManager {
             this.clearCancelButton(instanceName);
         }
         else {
-            void this.sendCancelButton(instanceName).then(() => this.armReplyGrace(instanceName));
+            void this.sendCancelButton(instanceName, undefined, true).then(() => this.armReplyGrace(instanceName));
         }
         this.reactDone(instanceName);
         const replyTo = this.lastInboundUser.get(instanceName) ?? "user";
@@ -3482,34 +3715,6 @@ export class FleetManager {
             ClassicChannelManager.logMessage(instanceName, "bot", args.text ?? "", new Date());
         }
     }
-    /** Handle tool status update from a daemon instance */
-    handleToolStatusFromInstance(instanceName, msg) {
-        const statusAdapter = this.getAdapterForInstance(instanceName) ?? this.adapter;
-        if (!statusAdapter)
-            return;
-        const text = msg.text;
-        const editMessageId = msg.editMessageId;
-        const senderSessionName = msg.senderSessionName;
-        const senderInstanceName = senderSessionName && this.fleetConfig?.instances[senderSessionName]
-            ? senderSessionName
-            : null;
-        const routingConfig = senderInstanceName
-            ? this.fleetConfig?.instances[senderInstanceName]
-            : (senderSessionName ? undefined : this.fleetConfig?.instances[instanceName]);
-        const threadId = routingConfig?.topic_id ? String(routingConfig.topic_id) : undefined;
-        const chatId = statusAdapter.getChatId();
-        if (!chatId)
-            return;
-        if (editMessageId) {
-            statusAdapter.editMessage(chatId, editMessageId, text, threadId).catch(e => this.logger.debug({ err: e }, "Failed to edit tool status message"));
-        }
-        else {
-            statusAdapter.sendText(chatId, text, { threadId }).then((sent) => {
-                const ipc = this.instanceIpcClients.get(instanceName);
-                ipc?.send({ type: "fleet_tool_status_ack", messageId: sent.messageId });
-            }).catch(e => this.logger.warn({ err: e }, "Failed to send tool status message"));
-        }
-    }
     // ===================== Scheduler =====================
     async handleScheduleTrigger(schedule) {
         const { target, reply_chat_id, reply_thread_id, message, label, id, source, silent } = schedule;
@@ -4243,7 +4448,13 @@ export class FleetManager {
         }
         catch (err) {
             const e = err;
-            return stripAnsi(e.stdout ?? e.message ?? "Doctor failed");
+            const output = stripAnsi(e.stdout ?? "").trim();
+            if (output)
+                return output;
+            if (e.code === "ENOENT") {
+                return "Backend doctor unavailable — agend CLI not found in PATH";
+            }
+            return stripAnsi(e.message ?? "").trim() || "Doctor failed";
         }
     }
     /** Drop event/activity rows older than the retention window. Best-effort. */
@@ -4395,6 +4606,371 @@ export class FleetManager {
                 .catch(e => this.logger.warn({ err: e, instanceName }, "Failed to send notification (no topic)"));
         }
     }
+    // ── Nonce-armed button prompts (hang / assist / exit / clear) ──
+    //
+    // One shared lifecycle for every "notification with decision buttons":
+    // post with a 128-bit nonce, arm a bounded expiry, bind the click to the
+    // exact adapter+chat+thread+message that created it, require fleet admin,
+    // consume exactly once. The features differ only in what they post
+    // and what a consumed click does.
+    /**
+     * Post decision buttons whose callback ids are `<prefix><nonce>:<action>`.
+     * The entry is registered before the send and rolled back if the send
+     * fails, so a nonce in the map always refers to a message that exists (or
+     * is about to). Returns the nonce, or null when the alert could not be sent.
+     */
+    async postNonceButtonPrompt(opts) {
+        // 16 bytes = the 128-bit capability the design claims. Telegram's 64-byte
+        // callback_data cap still holds: longest id is "interactive-assist:" (19)
+        // + 32 hex + ":confirm" (8) = 59 bytes.
+        const nonce = randomBytes(16).toString("hex");
+        const entry = {
+            prefix: opts.prefix,
+            instanceName: opts.instanceName,
+            adapterId: opts.adapterId,
+            adapter: opts.adapter,
+            chatId: opts.chatId,
+            threadId: opts.threadId,
+            expiredText: opts.expiredText,
+            ...opts.extra,
+        };
+        entry.timer = setTimeout(() => {
+            const pending = this.pendingNonceButtons.get(nonce);
+            if (pending !== entry)
+                return;
+            this.pendingNonceButtons.delete(nonce);
+            if (entry.messageId && entry.adapter.editMessageRemoveButtons) {
+                entry.adapter.editMessageRemoveButtons(entry.chatId, entry.messageId, entry.expiredText, entry.threadId).catch(err => this.logger.debug({ err, instanceName: entry.instanceName, prefix: entry.prefix }, "Failed to expire button prompt"));
+            }
+        }, opts.timeoutMs ?? NONCE_BUTTON_TIMEOUT_MS);
+        entry.timer.unref?.();
+        this.pendingNonceButtons.set(nonce, entry);
+        try {
+            const sent = await opts.adapter.notifyAlert(opts.chatId, {
+                type: opts.alertType,
+                instanceName: opts.instanceName,
+                message: opts.message,
+                choices: opts.choices.map(c => ({ id: `${opts.prefix}${nonce}:${c.action}`, label: c.label })),
+            }, opts.threadId ? { threadId: opts.threadId } : undefined);
+            entry.messageId = sent.messageId;
+            return nonce;
+        }
+        catch (err) {
+            this.pendingNonceButtons.delete(nonce);
+            if (entry.timer)
+                clearTimeout(entry.timer);
+            this.logger.warn({ err, instanceName: opts.instanceName, prefix: opts.prefix }, "Failed to send button prompt");
+            return null;
+        }
+    }
+    /**
+     * Claim a nonce-armed callback exactly once.
+     *
+     * Returns:
+     *  - null       — the callback is not for this prefix; try the next handler
+     *  - "consumed" — for this prefix but stale/denied/malformed; stop dispatch
+     *  - the entry+action — the click is authorized and now claimed (removed
+     *    from the map before any await, so double clicks cannot act twice)
+     *
+     * A stale click (expired nonce, or a pre-upgrade button whose payload no
+     * longer parses) collapses the clicked message so the dead button stops
+     * inviting clicks — the same courtesy the cancel button extends.
+     */
+    consumeNonceCallback(prefix, actionRe, data, callbackAdapterId, receivingAdapter) {
+        if (!data.callbackData.startsWith(prefix))
+            return null;
+        const match = data.callbackData.match(actionRe);
+        let pending = match ? this.pendingNonceButtons.get(match[1]) : undefined;
+        // The map is shared across prompt kinds. A nonce that resolves to an entry
+        // of a DIFFERENT kind is not a usable capability for this handler — treat
+        // it as stale rather than acting across kinds (fail closed).
+        if (pending && pending.prefix !== prefix)
+            pending = undefined;
+        if (!match || !pending) {
+            const adapter = receivingAdapter ?? this.adapter;
+            adapter?.editMessageRemoveButtons?.(data.chatId, data.messageId, t("buttons.stale"), data.threadId).catch(() => { });
+            return "consumed";
+        }
+        // Bind the capability to the exact message/world that created it. Telegram
+        // keyboards are visible to everyone, so the click also requires fleet admin.
+        const isAuthorized = data.userId
+            ? pending.authChannelId
+                ? this.isModelAdmin(data.userId, pending.authChannelId, callbackAdapterId)
+                : this.isFleetAdmin(data.userId, callbackAdapterId)
+            : false;
+        if (pending.adapterId !== callbackAdapterId
+            || data.chatId !== pending.chatId
+            || (pending.threadId != null && data.threadId !== pending.threadId)
+            || (pending.messageId != null && data.messageId !== pending.messageId)
+            || !isAuthorized) {
+            // Deliberately does NOT consume the nonce: the real admin can still click.
+            this.logger.warn({ instanceName: pending.instanceName, prefix, userId: data.userId }, "Rejected unauthorized or mismatched button callback");
+            return "consumed";
+        }
+        // Claim before any await: double clicks and duplicate callback delivery can
+        // never act twice.
+        this.pendingNonceButtons.delete(match[1]);
+        if (pending.timer)
+            clearTimeout(pending.timer);
+        return { entry: pending, action: match[2] };
+    }
+    /** Collapse a consumed prompt's buttons into a final status line. */
+    async retireNonceButtons(pending, messageId, text) {
+        if (!pending.adapter.editMessageRemoveButtons) {
+            this.logger.warn({ instanceName: pending.instanceName, adapterId: pending.adapterId }, "Adapter cannot remove prompt buttons");
+            return;
+        }
+        try {
+            await pending.adapter.editMessageRemoveButtons(pending.chatId, messageId, text, pending.threadId);
+        }
+        catch (err) {
+            // The action was already atomically consumed. An edit failure must not
+            // undo that or make the button actionable again.
+            this.logger.warn({ err, instanceName: pending.instanceName }, "Failed to retire prompt buttons");
+        }
+    }
+    /** Post the destructive `/clear` confirmation in the invoking channel. */
+    async promptClearConfirmation(instanceName, channelId, adapter, chatId, threadId) {
+        if (!this.topicCommands.supportsClear(instanceName))
+            return CLEAR_UNSUPPORTED_MSG;
+        const adapterId = adapter.id;
+        if (!adapterId)
+            return t("clear.prompt_unavailable");
+        const nonce = await this.postNonceButtonPrompt({
+            prefix: CLEAR_CONFIRM_CALLBACK_PREFIX,
+            alertType: "clear_confirm",
+            instanceName,
+            adapter,
+            adapterId,
+            chatId,
+            threadId,
+            message: t("clear.confirm_message", instanceName),
+            choices: [
+                { action: "confirm", label: t("clear.confirm") },
+                { action: "cancel", label: t("clear.cancel") },
+            ],
+            expiredText: t("clear.expired", instanceName),
+            extra: { authChannelId: channelId },
+            timeoutMs: CLEAR_CONFIRM_TIMEOUT_MS,
+        });
+        return nonce ? null : t("clear.prompt_unavailable");
+    }
+    /** Consume `/clear` Confirm/Cancel exactly once; only Confirm reaches IPC. */
+    async handleClearConfirmation(data, callbackAdapterId, receivingAdapter) {
+        const claimed = this.consumeNonceCallback(CLEAR_CONFIRM_CALLBACK_PREFIX, /^clear-confirm:([0-9a-f]+):(confirm|cancel)$/, data, callbackAdapterId, receivingAdapter);
+        if (claimed === null)
+            return false;
+        if (claimed === "consumed")
+            return true;
+        const { entry: pending, action } = claimed;
+        this.eventLog?.insert(pending.instanceName, "clear_action", { action, userId: data.userId });
+        if (action === "cancel") {
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("clear.cancelled", pending.instanceName));
+            return true;
+        }
+        await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("clear.clearing", pending.instanceName));
+        try {
+            const result = await this.topicCommands.sendClear(pending.instanceName);
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, result, pending.threadId);
+        }
+        catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            this.logger.error({ err, instanceName: pending.instanceName }, "Clear command failed");
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("clear.failed", pending.instanceName, message), pending.threadId).catch(editErr => this.logger.warn({ err: editErr, instanceName: pending.instanceName }, "Failed to show clear command error"));
+        }
+        return true;
+    }
+    /**
+     * Drop every pending prompt that refers to an instance being stopped or
+     * restarted — a restart offer for an instance the operator just restarted
+     * (or an assist for one they stopped) must not stay clickable for the rest
+     * of its 15 minutes.
+     */
+    clearNoncePromptsForInstance(instanceName) {
+        for (const [nonce, entry] of this.pendingNonceButtons) {
+            if (entry.instanceName !== instanceName)
+                continue;
+            this.pendingNonceButtons.delete(nonce);
+            if (entry.timer)
+                clearTimeout(entry.timer);
+            if (entry.messageId && entry.adapter.editMessageRemoveButtons) {
+                entry.adapter.editMessageRemoveButtons(entry.chatId, entry.messageId, entry.expiredText, entry.threadId)
+                    .catch(err => this.logger.debug({ err, instanceName, prefix: entry.prefix }, "Failed to collapse prompt during instance stop"));
+            }
+        }
+    }
+    /** A clean exit is intentional from the CLI's perspective, but often not from
+     * the operator's. Keep the instance notice passive and put the action in the
+     * same-world General topic where an administrator can make the choice. */
+    async notifyNormalExit(instanceName) {
+        this.notifyInstanceTopic(instanceName, t("exit.instance_notice", instanceName));
+        const worldId = this.getInstanceAdapterId(instanceName);
+        const generalName = this.findGeneralInstance(worldId);
+        if (!generalName) {
+            this.logger.warn({ instanceName, worldId }, "Normal CLI exit has no General notification target");
+            return;
+        }
+        const adapterId = this.getInstanceAdapterId(generalName);
+        const adapter = this.getAdapterForInstance(generalName);
+        const chatId = this.getGroupIdForInstance(generalName);
+        const topicId = this.fleetConfig?.instances[generalName]?.topic_id;
+        const threadId = topicId != null ? String(topicId) : undefined;
+        if (!adapter || !adapterId || !chatId) {
+            this.logger.warn({ instanceName, generalName, adapterId, chatId }, "Cannot address normal-exit restart controls");
+            return;
+        }
+        await this.postNonceButtonPrompt({
+            prefix: EXIT_RESTART_CALLBACK_PREFIX,
+            alertType: "exit_restart",
+            instanceName,
+            adapter,
+            adapterId,
+            chatId,
+            threadId,
+            message: t("exit.general_notice", instanceName),
+            choices: [
+                { action: "restart", label: t("exit.restart") },
+                { action: "ignore", label: t("exit.ignore") },
+            ],
+            expiredText: t("exit.expired", instanceName),
+        });
+    }
+    /** Consume a clean-exit Restart/Ignore button exactly once. */
+    async handleExitRestartPrompt(data, callbackAdapterId, receivingAdapter) {
+        const claimed = this.consumeNonceCallback(EXIT_RESTART_CALLBACK_PREFIX, /^exit-restart:([0-9a-f]+):(restart|ignore)$/, data, callbackAdapterId, receivingAdapter);
+        if (claimed === null)
+            return false;
+        if (claimed === "consumed")
+            return true;
+        const { entry: pending, action } = claimed;
+        this.eventLog?.insert(pending.instanceName, "normal_exit_action", { action, userId: data.userId });
+        if (action === "ignore") {
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("exit.ignored", pending.instanceName));
+            return true;
+        }
+        await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("exit.restarting", pending.instanceName));
+        try {
+            await this.restartSingleInstance(pending.instanceName);
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("exit.restarted", pending.instanceName), pending.threadId);
+        }
+        catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            this.logger.error({ err, instanceName: pending.instanceName }, "Normal-exit restart failed");
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("exit.restart_failed", pending.instanceName, message), pending.threadId).catch(editErr => this.logger.warn({ err: editErr, instanceName: pending.instanceName }, "Failed to show normal-exit restart error"));
+        }
+        return true;
+    }
+    interactivePromptLabel(kind) {
+        const key = `interactive.kind.${kind}`;
+        const translated = t(key);
+        return translated === key ? kind.replaceAll("_", " ") : translated;
+    }
+    /**
+     * Surface a stable terminal prompt in both places that matter:
+     * - the blocked instance gets a plain pointer to General;
+     * - General gets one-shot Confirm/Cancel controls on its own bound adapter.
+     *
+     * Only the prompt category crosses channels. The captured terminal line can
+     * contain usernames or secret-adjacent text and stays in daemon.log.
+     */
+    async notifyInteractivePrompt(instanceName, kind) {
+        this.notifyInstanceTopic(instanceName, t("interactive.instance_notice", instanceName));
+        // Multi-world fleets can have one General per bot/platform. Route the assist
+        // request to the General bound to the same adapter as the blocked instance.
+        const generalName = this.findGeneralInstance(this.getInstanceAdapterId(instanceName));
+        if (!generalName) {
+            this.logger.warn({ instanceName }, "Interactive prompt has no General topic notification target");
+            return;
+        }
+        const adapterId = this.getInstanceAdapterId(generalName);
+        const adapter = this.getAdapterForInstance(generalName);
+        const chatId = this.getGroupIdForInstance(generalName);
+        const topicId = this.fleetConfig?.instances[generalName]?.topic_id;
+        const threadId = topicId != null ? String(topicId) : undefined;
+        if (!adapter || !adapterId || !chatId) {
+            this.logger.warn({ instanceName, generalName, adapterId, chatId }, "Cannot address interactive prompt assistance controls");
+            return;
+        }
+        const label = this.interactivePromptLabel(kind);
+        await this.postNonceButtonPrompt({
+            prefix: INTERACTIVE_ASSIST_CALLBACK_PREFIX,
+            alertType: "interactive_prompt",
+            instanceName,
+            adapter,
+            adapterId,
+            chatId,
+            threadId,
+            message: t("interactive.general_notice", instanceName, label),
+            choices: [
+                { action: "confirm", label: t("interactive.confirm") },
+                { action: "cancel", label: t("interactive.cancel") },
+            ],
+            expiredText: t("interactive.expired", instanceName),
+            extra: { generalName, promptKind: kind },
+        });
+    }
+    /** Consume a General assist button exactly once. */
+    async handleInteractivePromptAssist(data, callbackAdapterId, receivingAdapter) {
+        const claimed = this.consumeNonceCallback(INTERACTIVE_ASSIST_CALLBACK_PREFIX, /^interactive-assist:([0-9a-f]+):(confirm|cancel)$/, data, callbackAdapterId, receivingAdapter);
+        if (claimed === null)
+            return false;
+        if (claimed === "consumed")
+            return true;
+        const { entry: pending, action } = claimed;
+        this.eventLog?.insert(pending.instanceName, "interactive_prompt_assist", {
+            action,
+            userId: data.userId,
+            generalName: pending.generalName,
+        });
+        if (action === "cancel") {
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("interactive.ignored", pending.instanceName));
+            return true;
+        }
+        // Fail closed on a malformed entry. generalName is set at posting time for
+        // every interactive-assist entry; its absence means entry confusion, and
+        // falling back to the blocked instance would type the assist text into the
+        // very terminal prompt this feature exists to keep humans in front of.
+        const generalName = pending.generalName;
+        if (!generalName) {
+            this.logger.error({ instanceName: pending.instanceName }, "Interactive-assist entry has no General target — refusing to deliver");
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("interactive.delivery_failed", pending.instanceName));
+            return true;
+        }
+        // Injecting into General while General itself is blocked would type into the
+        // terminal prompt instead of the agent input. Fail safe and require attach.
+        if (pending.instanceName === generalName) {
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("interactive.self_assist", pending.instanceName));
+            return true;
+        }
+        await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("interactive.confirmed", pending.instanceName));
+        try {
+            await this.deliverToInstance(generalName, {
+                type: "fleet_inbound",
+                content: t("interactive.assist_request", pending.instanceName, this.interactivePromptLabel(pending.promptKind ?? "")),
+                targetSession: generalName,
+                meta: {
+                    chat_id: pending.chatId,
+                    message_id: pending.messageId ?? data.messageId,
+                    user: "AgEnD",
+                    user_id: data.userId,
+                    ts: new Date().toISOString(),
+                    thread_id: pending.threadId ?? "",
+                    adapter_id: pending.adapterId,
+                    source: pending.adapter.type,
+                },
+            }, {
+                // This is system-generated work, not a live user message. Queue behind
+                // General's current turn so Kiro/Claude do not lose it while busy.
+                isCrossInstance: true,
+            });
+            this.logger.info({ instanceName: pending.instanceName, generalName: pending.generalName }, "Interactive prompt assistance delivered to General");
+        }
+        catch (err) {
+            this.logger.warn({ err, instanceName: pending.instanceName, generalName: pending.generalName }, "Failed to deliver interactive prompt assistance to General");
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("interactive.delivery_failed", pending.instanceName), pending.threadId);
+        }
+        return true;
+    }
     // ── Cancel button ────────────────────────────────────────────────────
     // Sent after delivering a user message to an instance; clicking it (or
     // /cancel) sends Escape to the instance's pane to interrupt generation.
@@ -4483,7 +5059,7 @@ export class FleetManager {
         }
         return false;
     }
-    async sendCancelButton(instanceName, correlationId) {
+    async sendCancelButton(instanceName, correlationId, preserveProgress = false) {
         // Post first, retire after (see the tail of this method). Retiring up front
         // meant that from the delete until the new message came back — a chat API
         // round trip, and every reply goes through here — the instance had NO live
@@ -4544,6 +5120,10 @@ export class FleetManager {
                 // not re-edit identical text — which put a "(edited)" mark on Discord
                 // with nothing visibly changed.
                 lastProgressText: "👀 處理中…",
+                // A reply can re-post the same in-flight bubble below the reply. Carry
+                // its current list into that replacement; fresh inbound work must start
+                // empty even if the daemon's reset broadcast is still in flight.
+                toolProgress: preserveProgress ? this.instanceProgress.get(instanceName) : undefined,
             };
             this.startProgressTicker(entry);
             // Idle-check backstop: every 5min, if the instance is idle, retire the
@@ -4639,6 +5219,88 @@ export class FleetManager {
         else
             this.instanceActivity.delete(name);
     }
+    /**
+     * Cache the turn's tool-progress list and push it into the instance's live
+     * bubble, coalesced so a burst of tool events cannot flood the Bot API
+     * (Telegram's flood limit is per-chat across ALL forum topics, so edits are
+     * rate-limited per bubble AND ride behind the daemon-side 3s coalescer).
+     */
+    cacheInstanceProgress(name, progress) {
+        if (progress)
+            this.instanceProgress.set(name, progress);
+        else
+            this.instanceProgress.delete(name);
+        for (const entry of this.cancelButtons.values()) {
+            if (entry.instanceName !== name)
+                continue;
+            // Empty means the daemon completed/reset the turn. Clear the live cache
+            // for the next turn, but retain this bubble's last non-empty list so its
+            // retirement can preserve the history instead of deleting the message.
+            if (progress)
+                entry.toolProgress = progress;
+            this.scheduleProgressEdit(entry);
+        }
+    }
+    /** At most one progress-driven edit per bubble per TOOL_PROGRESS_EDIT_MIN_MS. */
+    scheduleProgressEdit(entry) {
+        if (entry.retiring)
+            return;
+        const since = Date.now() - (entry.lastProgressEditAt ?? 0);
+        if (since >= TOOL_PROGRESS_EDIT_MIN_MS) {
+            this.refreshBubble(entry);
+            return;
+        }
+        if (entry.progressEditTimer)
+            return; // trailing edit already scheduled
+        entry.progressEditTimer = setTimeout(() => {
+            entry.progressEditTimer = undefined;
+            this.refreshBubble(entry);
+        }, TOOL_PROGRESS_EDIT_MIN_MS - since);
+        entry.progressEditTimer.unref?.();
+    }
+    /**
+     * The ONE composer for the bubble text. Both writers — the elapsed-time
+     * ticker and the tool-progress push — go through here; two independent
+     * renderers editing the same message is how the progress list used to get
+     * wiped by the next elapsed tick (#528 trap 2).
+     */
+    composeBubbleText(entry) {
+        return FleetManager.bubbleText(Date.now() - (entry.startedAt ?? Date.now()), this.instanceActivity.get(entry.instanceName), this.progressMinElapsedMs(), entry.toolProgress);
+    }
+    /** Pure composition of header + tool list, exposed for tests. */
+    static bubbleText(elapsedMs, activity, minElapsedMs, progress) {
+        const header = FleetManager.progressText(elapsedMs, 
+        // The single-line activity detail is redundant once a tool list exists.
+        progress ? undefined : activity, minElapsedMs);
+        return progress ? `${header}\n${progress}` : header;
+    }
+    /** Recompose and edit the bubble in place; skips when nothing changed. */
+    refreshBubble(entry) {
+        if (entry.retiring || !this.cancelButtons.has(entry.messageId)) {
+            clearInterval(entry.progressTimer);
+            return;
+        }
+        const text = this.composeBubbleText(entry);
+        if (text === entry.lastProgressText)
+            return; // nothing changed — skip the API call
+        const adapter = this.getAdapterForInstance(entry.instanceName) ?? this.adapter;
+        if (!adapter?.editAlert)
+            return;
+        entry.lastProgressText = text;
+        entry.lastProgressEditAt = Date.now();
+        adapter.editAlert(entry.chatId, entry.messageId, {
+            type: "cancel",
+            instanceName: entry.instanceName,
+            message: text,
+            choices: [{ id: `cancel:${entry.instanceName}`, label: t("cancel.button") }],
+        }, entry.threadId ? { threadId: entry.threadId } : undefined)
+            .catch(err => {
+            // A failed progress edit must never escalate: the button still works and
+            // the next tick retries. Common causes are a deleted message or a
+            // rate limit.
+            this.logger.debug({ err, instanceName: entry.instanceName }, "Progress edit failed");
+        });
+    }
     /**
      * Refresh the button's text in place while the instance keeps working.
      *
@@ -4656,31 +5318,7 @@ export class FleetManager {
         return PROGRESS_MIN_ELAPSED_MS;
     }
     startProgressTicker(entry) {
-        const tick = () => {
-            if (!this.cancelButtons.has(entry.messageId)) {
-                clearInterval(entry.progressTimer);
-                return;
-            }
-            const text = FleetManager.progressText(Date.now() - (entry.startedAt ?? Date.now()), this.instanceActivity.get(entry.instanceName), this.progressMinElapsedMs());
-            if (text === entry.lastProgressText)
-                return; // nothing changed — skip the API call
-            const adapter = this.getAdapterForInstance(entry.instanceName) ?? this.adapter;
-            if (!adapter?.editAlert)
-                return;
-            entry.lastProgressText = text;
-            adapter.editAlert(entry.chatId, entry.messageId, {
-                type: "cancel",
-                instanceName: entry.instanceName,
-                message: text,
-                choices: [{ id: `cancel:${entry.instanceName}`, label: t("cancel.button") }],
-            }, entry.threadId ? { threadId: entry.threadId } : undefined)
-                .catch(err => {
-                // A failed progress edit must never escalate: the button still works and
-                // the next tick retries. Common causes are a deleted message or a
-                // rate limit.
-                this.logger.debug({ err, instanceName: entry.instanceName }, "Progress edit failed");
-            });
-        };
+        const tick = () => this.refreshBubble(entry);
         entry.progressTimer = setInterval(tick, PROGRESS_UPDATE_INTERVAL_MS);
         entry.progressTimer.unref?.();
         // One extra tick right when the threshold passes, so a 30s threshold shows
@@ -4736,7 +5374,7 @@ export class FleetManager {
         this.deleteButtonMessage(entry)
             .then(() => {
             this.discardButton(entry);
-            this.logger.info({ instanceName: entry.instanceName, messageId: entry.messageId }, "Cancel button removed");
+            this.logger.info({ instanceName: entry.instanceName, messageId: entry.messageId, historyPreserved: Boolean(entry.toolProgress) }, "Cancel button retired");
         })
             .catch((err) => this.scheduleButtonRetry(entry, err));
     }
@@ -4748,6 +5386,8 @@ export class FleetManager {
             clearInterval(entry.idleCheckTimer);
         if (entry.progressTimer)
             clearInterval(entry.progressTimer);
+        if (entry.progressEditTimer)
+            clearTimeout(entry.progressEditTimer);
         if (entry.replyGraceTimer)
             clearTimeout(entry.replyGraceTimer);
         this.cancelButtons.delete(entry.messageId);
@@ -4825,12 +5465,17 @@ export class FleetManager {
         // Continue the same retire cycle (bypass the retiring-guard in retireButton).
         entry.retryTimer = setTimeout(() => this.attemptButtonDelete(entry), CANCEL_BTN_RETRY_INTERVAL_MS);
     }
-    /** Delete one button's message via its own adapter. Resolves on success,
+    /** Retire one cancel bubble via its own adapter. A bubble that accumulated
+     * tool progress is kept as a read-only history message; a plain bubble keeps
+     * the legacy delete/edit-to-checkmark behavior. Resolves on success and
      * rejects on failure so the caller can retry. */
     deleteButtonMessage(e) {
         const adapter = (e.adapterId ? this.worlds.get(e.adapterId)?.adapter : undefined) ?? this.adapter;
         if (!adapter)
             return Promise.reject(new Error("no adapter for cancel button"));
+        if (e.toolProgress && adapter.editMessageRemoveButtons) {
+            return adapter.editMessageRemoveButtons(e.chatId, e.messageId, this.composeBubbleText(e), e.threadId);
+        }
         if (adapter.deleteMessage)
             return adapter.deleteMessage(e.chatId, e.messageId, e.threadId);
         if (adapter.editMessageRemoveButtons)
@@ -4921,13 +5566,29 @@ export class FleetManager {
     }
     async sendHangNotification(instanceName, unchangedForMs) {
         const adapter = this.getAdapterForInstance(instanceName) ?? this.adapter;
-        if (!adapter)
-            return;
-        const channelCfg = this.getChannelConfig(this.getInstanceAdapterId(instanceName));
-        const groupId = channelCfg?.group_id;
-        if (!groupId)
+        const adapterId = this.getInstanceAdapterId(instanceName);
+        // Same three-way addressing as sendCancelButton: fleet topic → group+thread,
+        // Classic → its own channel (Classic instances are absent from
+        // fleetConfig.instances, so the topic path can never address them), else the
+        // world group flat. getGroupIdForInstance (not getChannelConfig().group_id)
+        // because on channels[]-configured fleets the legacy channel: block is empty.
+        const topicId = this.fleetConfig?.instances[instanceName]?.topic_id;
+        const groupId = this.getGroupIdForInstance(instanceName) || undefined;
+        let chatId;
+        let threadId;
+        if (topicId != null && groupId) {
+            chatId = String(groupId);
+            threadId = String(topicId);
+        }
+        else {
+            chatId = this.classicChannels?.getChannelIdByInstance(instanceName);
+            if (!chatId && groupId)
+                chatId = String(groupId);
+        }
+        if (!adapter || !adapterId || !chatId) {
+            this.logger.warn({ instanceName, adapterId, chatId }, "Cannot address hang notification");
             return;
-        const threadId = this.fleetConfig?.instances[instanceName]?.topic_id;
+        }
         const instanceHangConfig = this.fleetConfig?.instances[instanceName]?.hang_detector;
         const configuredMinutes = instanceHangConfig?.timeout_minutes
             ?? this.fleetConfig?.defaults?.hang_detector?.timeout_minutes
@@ -4936,17 +5597,52 @@ export class FleetManager {
             ? configuredMinutes
             : Math.max(1, Math.floor(unchangedForMs / 60_000));
         this.setTopicIcon(instanceName, "red");
-        await adapter.notifyAlert(String(groupId), {
-            type: "hang",
+        await this.postNonceButtonPrompt({
+            prefix: HANG_CALLBACK_PREFIX,
+            alertType: "hang",
             instanceName,
-            message: `⚠️ ${instanceName} may be stuck — pane unchanged for ${unchangedMinutes}min, ready prompt not recognized`,
+            adapter,
+            adapterId,
+            chatId,
+            threadId,
+            message: t("hang.detected", instanceName, unchangedMinutes),
             choices: [
-                { id: `hang:restart:${instanceName}`, label: "🔄 Force restart" },
-                { id: `hang:wait:${instanceName}`, label: "⏳ Keep waiting" },
+                { action: "restart", label: t("hang.restart") },
+                { action: "wait", label: t("hang.wait") },
             ],
-        }, {
-            threadId: threadId != null ? String(threadId) : undefined,
-        }).catch(e => this.logger.warn({ err: e }, "Failed to send hang notification"));
+            expiredText: t("hang.expired", instanceName),
+        });
+    }
+    /**
+     * Consume a hang Force-restart / Keep-waiting button exactly once. Restart
+     * goes through restartSingleInstance — serialized against concurrent restart
+     * sources, and with a Classic-instance fallback (the previous hand-rolled
+     * stop+start silently left Classic instances stopped while reporting
+     * "restarted").
+     */
+    async handleHangPrompt(data, callbackAdapterId, receivingAdapter) {
+        const claimed = this.consumeNonceCallback(HANG_CALLBACK_PREFIX, /^hang:([0-9a-f]+):(restart|wait)$/, data, callbackAdapterId, receivingAdapter);
+        if (claimed === null)
+            return false;
+        if (claimed === "consumed")
+            return true;
+        const { entry: pending, action } = claimed;
+        this.eventLog?.insert(pending.instanceName, "hang_action", { action, userId: data.userId });
+        if (action === "wait") {
+            await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("hang.waiting", pending.instanceName));
+            return true;
+        }
+        await this.retireNonceButtons(pending, pending.messageId ?? data.messageId, t("hang.restarting", pending.instanceName));
+        try {
+            await this.restartSingleInstance(pending.instanceName);
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("hang.restarted", pending.instanceName), pending.threadId);
+        }
+        catch (err) {
+            const message = err instanceof Error ? err.message : String(err);
+            this.logger.error({ err, instanceName: pending.instanceName }, "Hang force-restart failed");
+            await pending.adapter.editMessage(pending.chatId, pending.messageId ?? data.messageId, t("hang.restart_failed", pending.instanceName, message), pending.threadId).catch(editErr => this.logger.warn({ err: editErr, instanceName: pending.instanceName }, "Failed to show hang restart error"));
+        }
+        return true;
     }
     // ── Topic icon + auto-archive ─────────────────────────────────────────────
     static INSTRUCTIONS_FILENAME = {
@@ -4955,6 +5651,10 @@ export class FleetManager {
         "gemini-cli": "GEMINI.md",
         "opencode": "AGENTS.md",
         "kiro-cli": ".kiro/steering/project.md",
+        // Grok reads AGENTS.md project docs; agy reads .agents/agents.md — the
+        // same files their writeConfig() appends fleet instructions to.
+        "grok": "AGENTS.md",
+        "antigravity": ".agents/agents.md",
         "mock": "CLAUDE.md",
     };
     static GENERAL_INSTRUCTIONS = `# Fleet Coordinator
@@ -4964,146 +5664,33 @@ You route tasks, manage instances, enforce policies, and synthesize results.
 Do NOT modify project files directly — delegate file changes to the project's instance.
 You CAN write code snippets, explain code, and answer technical questions directly.
 
------
+## Task Routing
 
-## Task Classification
+- **Handle directly**: no file/exec access needed, answerable from knowledge, ≤2 reasoning steps (Q&A, translation, status queries, code snippets).
+- **Delegate to 1 instance**: scoped to one project/repo, needs file access or execution.
+- **Coordinate multiple**: spans repos, outputs feed each other, or parallel helps (max 3 per task).
 
-Classify every incoming request before acting.
+Instance discovery order: list_teams() → list_instances() → describe_instance() → create_instance() only if nothing fits. Prefer reuse; never duplicate a running instance.
 
-### Handle Directly (ALL conditions must be true)
+## Reply Contract
 
-- No file system access needed
-- No external execution needed
-- Answerable from static knowledge
-- ≤ 2 reasoning steps
-
-Examples: Q&A, translation, fleet status queries, explaining a concept, writing code snippets.
-
-### Delegate to 1 Instance
-
-- Task scoped to a single project or repo
-- Requires file access, code changes, or execution
-
-### Coordinate Multiple Instances
-
-- Task spans multiple repos or domains
-- Requires outputs from one instance to feed into another
-- Benefits from parallel execution (max 3 instances per task)
-
------
-
-## Instance Discovery (in this order)
-1. list_teams()        → reuse existing teams first
-2. list_instances()    → find by working_directory, description, or tags
-3. describe_instance() → confirm capabilities before delegating
-4. create_instance()   → only if no suitable instance exists
-
-Rules: prefer reuse over creation. Do NOT create duplicates of running instances.
-
------
-
-## Delegation Protocol
-
-Every delegation via send_to_instance() MUST include:
-
-1. Task scope — what exactly to do, bounded clearly
-2. Expected output — what to return and in what form
-3. Policy reminder — "Follow Development Workflow policy" (for code tasks)
-
-### Loop Prevention
-
-- Never re-delegate a task back to the instance that sent it to you
-- If a task has bounced 3 times, stop and solve locally or reduce scope
-
-### Execution Strategy
-
-Parallel — use only when tasks are independent with no shared state
-Sequential — use when one task's output feeds into the next
-
------
-
-## Result Handling
-
-When an instance reports back, classify the outcome:
-
-- Success → Summarize key results for user. Omit internal coordination noise.
-- Partial → State what succeeded, what remains, proposed next steps.
-- Failure → Retry up to 2 times. If still failing: try alternative instance, reduce scope, or return partial result clearly marked.
-- No response → Ping again after reasonable wait. If still silent: report to user with options.
-
-### Output to User
-
-Every final response to the user should contain:
-
-- Result — the actual answer or deliverable
-- Gaps — anything incomplete or unresolved (omit if none)
-
------
-
-## Shared Decisions
-
-Use post_decision() / list_decisions() for any choice that affects more than 1 instance, changes an API contract, introduces a new dependency, or alters deployment process.
-
-When instances disagree, collect both viewpoints, make a decision, and record it via post_decision.
-
------
+Every final response to the user contains: the result (the actual answer or deliverable) and gaps (anything incomplete — omit if none). Summarize instance reports; omit internal coordination noise.
 
 ## After Restart
 
-After a restart, run this sequence BEFORE processing any new messages:
-1. list_instances()   → rebuild fleet awareness
-2. list_teams()       → restore team structure
-3. list_decisions()   → reload policies and conventions
-
-Only then handle incoming requests.
-
------
-
-## Development Workflow Policy
-
-All code changes across the fleet should follow this workflow.
-The coordinator enforces compliance but does not perform these steps directly.
-Remind instances of this policy when delegating code tasks.
-
-### Workflow Stages
-Design Proposed → Design Approved → Implementation → Submit for Review → Under Review → Approved → Merge
-
-### Policy Rules
-
-1. Design before code — developer sends design proposal to reviewer before implementation. Consensus required before proceeding.
-2. Challenger pairing — every code task should have a developer + reviewer. Reviewer actively questions decisions and finds risks.
-3. Verify by execution — backend/CLI changes must be tested by running them. Do not trust documentation alone.
-4. Independent review — every merge requires code review from someone other than the author.
-5. Root cause first — bug fixes require confirmed root cause before proposing a fix.
-6. Merge conditions: tests pass, reviewer approved, branch and worktree cleaned up.
-
-### Specialist Instance Rules
-
-- Execute within defined scope only
-- Return structured output: result, assumptions, uncertainties, verification status
-- Do NOT create new instances without coordinator approval
-
------
-
-## Team Management
-
-- Always check existing teams before creating new ones
-- Default to ephemeral teams (created for a specific task, dissolved after completion)
-- Clean up ephemeral teams and instances after task completion
+BEFORE processing any new messages: 1. list_instances() 2. list_teams() 3. list_decisions(). Only then handle requests.
 
------
+## Playbooks (on-demand skills)
 
-## Instance Configuration Tips
-
-When users create specialized instances, suggest these configurations:
-
-- **Reviewer instances**: Add \`pre_task_command: "/chat load reviewer-base"\` to reset context before each review, preventing influence from previous conversations.
-- **Collab mode**: For multi-bot channels, use \`/collab\` to enable @mention-based triggering.
-- **Cost control**: Set per-instance \`cost_guard\` for expensive backends.
+Detailed procedures live in your skills — consult them when the situation comes up rather than from memory:
+- **delegation-playbook** — delegation protocol, loop prevention, parallel vs sequential, result/failure handling, team management, instance configuration tips.
+- **development-workflow** — the fleet-wide code-change policy you enforce when delegating code tasks.
+Plus the operational skills (fleet-health, instance-lifecycle, scheduling, session-management, …).
 `;
     /** Ensure the general instance has its project instructions file + knowledge */
-    ensureGeneralInstructions(workDir, backendName) {
+    ensureGeneralInstructions(workDir, backendName, instanceName) {
         const backend = backendName ?? "claude-code";
+        workDir = this.resolveKnowledgeWorkDir(workDir, backend, instanceName);
         const filename = FleetManager.INSTRUCTIONS_FILENAME[backend] ?? "CLAUDE.md";
         const filePath = join(workDir, filename);
         mkdirSync(dirname(filePath), { recursive: true });
@@ -5111,60 +5698,238 @@ When users create specialized instances, suggest these configurations:
             writeFileSync(filePath, FleetManager.GENERAL_INSTRUCTIONS, "utf-8");
             this.logger.info({ filePath }, "Created general instance instructions file");
         }
-        // Sync bundled knowledge files to general's steering directory
+        // Sync bundled knowledge files to general's steering and skills directories.
         this.syncGeneralKnowledge(workDir, backend);
     }
-    /** Copy general-knowledge steering + skills to the general instance's workspace */
+    /** Resolve the workspace path a backend actually uses before publishing knowledge. */
+    resolveKnowledgeWorkDir(workDir, backend, instanceName) {
+        // Some backends relocate their real cwd (agy moves hidden paths to
+        // ~/agend-workspaces/<name>). Instructions and skills must land where the
+        // CLI actually runs, not where fleet.yaml points.
+        try {
+            const resolved = createBackend(backend, join(getAgendHome(), "cli-env"))
+                .resolveWorkingDirectory?.(workDir, instanceName);
+            if (resolved)
+                workDir = resolved;
+        }
+        catch { /* unknown backend name — keep the raw path */ }
+        return workDir;
+    }
+    /**
+     * Where each backend natively loads on-demand skills from, relative to the
+     * workspace. Backends without a native skill mechanism (opencode, grok,
+     * antigravity, gemini-cli) are deliberately absent: dropping files a CLI
+     * never reads is clutter, not capability. Unknown directories are ignored
+     * by older CLI versions, so publishing is fail-open across upgrades.
+     */
+    static SKILLS_DIR_SEGMENTS = {
+        "kiro-cli": [".kiro", "skills"],
+        "claude-code": [".claude", "skills"],
+        "codex": [".agents", "skills"],
+        // Live-verified: OpenCode and Antigravity scan .agents/skills; Grok's
+        // vendor-canonical location is .grok/skills.
+        "opencode": [".agents", "skills"],
+        "grok": [".grok", "skills"],
+        "antigravity": [".agents", "skills"],
+    };
+    /** Copy general-knowledge steering + all role-eligible skills to General. */
     syncGeneralKnowledge(workDir, backend) {
         const knowledgeDir = join(dirname(fileURLToPath(import.meta.url)), "general-knowledge");
         if (!existsSync(knowledgeDir))
             return;
-        // Sync steering files → .kiro/steering/ (or workDir root for non-kiro)
-        const steeringDir = backend === "kiro-cli"
-            ? join(workDir, ".kiro", "steering")
-            : workDir;
-        mkdirSync(steeringDir, { recursive: true });
-        const srcSteering = join(knowledgeDir, "steering");
-        if (existsSync(srcSteering)) {
-            for (const file of readdirSync(srcSteering)) {
-                if (!file.endsWith(".md"))
-                    continue;
-                const src = join(srcSteering, file);
+        this.syncGeneralSteering(workDir, backend, join(knowledgeDir, "steering"));
+        this.syncRoleSkills(workDir, backend, "general", knowledgeDir);
+        this.logger.debug({ knowledgeDir, workDir, backend }, "Synced general knowledge files");
+    }
+    /** Publish only the bundled skills eligible for an instance role. */
+    syncRoleSkills(workDir, backend, role, knowledgeDir) {
+        const skillSegments = FleetManager.SKILLS_DIR_SEGMENTS[backend];
+        if (!skillSegments)
+            return;
+        const root = knowledgeDir ?? join(dirname(fileURLToPath(import.meta.url)), "general-knowledge");
+        if (!existsSync(root))
+            return;
+        // Before managed-skill manifests existed, only Kiro General received
+        // bundled skills. Allow that one legacy layout to be adopted so upgrades
+        // can keep those copies current; other backends never had unmanaged
+        // AgEnD-published skills and must retain the normal user-ownership guard.
+        const adoptLegacyUnmanaged = backend === "kiro-cli" && role === "general";
+        this.syncManagedSkills(join(workDir, ...skillSegments), join(root, "skills"), role, adoptLegacyUnmanaged);
+        this.logger.debug({ workDir, backend, role }, "Synced role-based bundled skills");
+    }
+    /**
+     * Steering (always-on rules like core-rules.md). Kiro loads a native
+     * steering directory; every other backend gets the content embedded into
+     * its instructions file (CLAUDE.md / AGENTS.md / …) inside a managed marker
+     * block — the previous behavior dropped bare .md files in the workspace
+     * root, which no CLI ever read. The block is replaced in place on every
+     * sync, so rule updates reach EXISTING workspaces; everything the user
+     * wrote outside the markers is preserved byte-for-byte.
+     */
+    syncGeneralSteering(workDir, backend, srcSteering) {
+        if (!existsSync(srcSteering))
+            return;
+        const files = readdirSync(srcSteering).filter(f => f.endsWith(".md")).sort();
+        if (files.length === 0)
+            return;
+        if (backend === "kiro-cli") {
+            const steeringDir = join(workDir, ".kiro", "steering");
+            mkdirSync(steeringDir, { recursive: true });
+            for (const file of files) {
                 const dest = join(steeringDir, file);
-                const newContent = readFileSync(src, "utf-8");
+                const newContent = readFileSync(join(srcSteering, file), "utf-8");
                 try {
                     if (existsSync(dest) && readFileSync(dest, "utf-8") === newContent)
                         continue;
                 }
-                catch { }
+                catch { /* rewrite */ }
                 writeFileSync(dest, newContent);
             }
+            return;
         }
-        // Sync skills → .kiro/skills/ (kiro-cli only)
-        if (backend === "kiro-cli") {
-            const srcSkills = join(knowledgeDir, "skills");
-            if (existsSync(srcSkills)) {
-                const destSkills = join(workDir, ".kiro", "skills");
-                mkdirSync(destSkills, { recursive: true });
-                for (const skillDir of readdirSync(srcSkills)) {
-                    const skillSrc = join(srcSkills, skillDir);
-                    if (!existsSync(join(skillSrc, "SKILL.md")))
-                        continue;
-                    const skillDest = join(destSkills, skillDir);
-                    mkdirSync(skillDest, { recursive: true });
-                    const src = join(skillSrc, "SKILL.md");
-                    const dest = join(skillDest, "SKILL.md");
-                    const newContent = readFileSync(src, "utf-8");
-                    try {
-                        if (existsSync(dest) && readFileSync(dest, "utf-8") === newContent)
-                            continue;
-                    }
-                    catch { }
-                    writeFileSync(dest, newContent);
-                }
+        const filename = FleetManager.INSTRUCTIONS_FILENAME[backend] ?? "CLAUDE.md";
+        const instructionsPath = join(workDir, filename);
+        const body = files.map(f => readFileSync(join(srcSteering, f), "utf-8").trim()).join("\n\n");
+        const block = `${FleetManager.STEERING_BLOCK_BEGIN}\n${body}\n${FleetManager.STEERING_BLOCK_END}`;
+        let existing = "";
+        try {
+            existing = existsSync(instructionsPath) ? readFileSync(instructionsPath, "utf-8") : "";
+        }
+        catch { /* treat as empty */ }
+        const beginAt = existing.indexOf(FleetManager.STEERING_BLOCK_BEGIN);
+        const endAt = existing.indexOf(FleetManager.STEERING_BLOCK_END);
+        let next;
+        if (beginAt !== -1 && endAt !== -1 && endAt > beginAt) {
+            next = existing.slice(0, beginAt) + block + existing.slice(endAt + FleetManager.STEERING_BLOCK_END.length);
+        }
+        else {
+            next = existing.trimEnd() + (existing.trim() ? "\n\n" : "") + block + "\n";
+        }
+        if (next !== existing) {
+            mkdirSync(dirname(instructionsPath), { recursive: true });
+            writeFileSync(instructionsPath, next);
+        }
+    }
+    static STEERING_BLOCK_BEGIN = "<!-- >>> agend:core-rules — managed by AgEnD; edits inside this block are overwritten -->";
+    static STEERING_BLOCK_END = "<!-- <<< agend:core-rules -->";
+    /**
+     * Publish AgEnD's bundled skills into a CLI's native skills directory,
+     * owning ONLY what we published. A manifest records which skill names AgEnD
+     * wrote; a bundled rename/removal deletes the stale managed copy, while a
+     * skill the user created by hand is never listed and therefore never
+     * touched — even if a future bundle happens to reuse its name (the sync
+     * then skips it and logs, rather than overwrite the user's work). The sole
+     * migration exception is the pre-manifest Kiro General layout explicitly
+     * selected by adoptLegacyUnmanaged.
+     */
+    syncManagedSkills(destSkills, srcSkills, role, adoptLegacyUnmanaged = false) {
+        if (!existsSync(srcSkills))
+            return;
+        mkdirSync(destSkills, { recursive: true });
+        const manifestPath = join(destSkills, ".agend-managed-skills.json");
+        const hadManifest = existsSync(manifestPath);
+        let previouslyManaged = [];
+        try {
+            const parsed = JSON.parse(readFileSync(manifestPath, "utf-8"));
+            if (Array.isArray(parsed))
+                previouslyManaged = parsed.filter(n => typeof n === "string");
+        }
+        catch { /* first sync, or corrupt manifest — treat as owning nothing */ }
+        const bundled = readdirSync(srcSkills)
+            .filter(name => existsSync(join(srcSkills, name, "SKILL.md")))
+            .sort();
+        // Pre-manifest Kiro General copied bundled skills directly into
+        // .kiro/skills. If every existing skill is still a bundled name, this is
+        // the unambiguous legacy layout: adopt it once, update it below, and write
+        // the ownership manifest. Any extra skill name keeps the directory fully
+        // on the user-owned/collision path.
+        if (!hadManifest && adoptLegacyUnmanaged) {
+            const existing = readdirSync(destSkills, { withFileTypes: true })
+                .filter(entry => entry.isDirectory() && existsSync(join(destSkills, entry.name, "SKILL.md")))
+                .map(entry => entry.name)
+                .sort();
+            if (existing.length > 0 && existing.every(name => bundled.includes(name))) {
+                previouslyManaged = existing;
+                this.logger.info({ skills: existing, destSkills }, "Adopted legacy AgEnD skills into managed ownership");
+            }
+        }
+        const eligible = bundled.filter(name => {
+            const roles = this.readManagedSkillRoles(join(srcSkills, name, "SKILL.md"));
+            // General is the coordinator and receives both coordinator and worker
+            // playbooks. Workers receive only skills explicitly marked for workers.
+            return role === "general" || roles.includes("worker");
+        });
+        const managed = [];
+        for (const name of eligible) {
+            const destDir = join(destSkills, name);
+            const dest = join(destDir, "SKILL.md");
+            const isOurs = previouslyManaged.includes(name) || !existsSync(destDir);
+            if (!isOurs) {
+                // Name collision with a user-authored skill: theirs wins, loudly.
+                this.logger.warn({ skill: name, destSkills }, "Skipping bundled skill — a skill of this name exists but was not published by AgEnD");
+                continue;
+            }
+            managed.push(name);
+            const newContent = readFileSync(join(srcSkills, name, "SKILL.md"), "utf-8");
+            try {
+                if (existsSync(dest) && readFileSync(dest, "utf-8") === newContent)
+                    continue;
             }
+            catch { /* rewrite */ }
+            mkdirSync(destDir, { recursive: true });
+            writeFileSync(dest, newContent);
+        }
+        // Remove managed skills that are no longer bundled OR no longer eligible
+        // for this role. This makes a shared → general-only metadata change take
+        // effect on the next worker startup instead of leaving stale capability.
+        for (const stale of previouslyManaged) {
+            if (eligible.includes(stale))
+                continue;
+            try {
+                rmSync(join(destSkills, stale), { recursive: true, force: true });
+                this.logger.info({ skill: stale, destSkills }, "Removed retired AgEnD-managed skill");
+            }
+            catch (err) {
+                this.logger.debug({ err, skill: stale }, "Failed to remove retired managed skill");
+            }
+        }
+        try {
+            writeFileSync(manifestPath, JSON.stringify(managed, null, 2) + "\n");
+        }
+        catch (err) {
+            this.logger.debug({ err, manifestPath }, "Failed to write managed-skills manifest");
+        }
+    }
+    /** Read AgEnD's roles extension from SKILL.md YAML frontmatter. */
+    readManagedSkillRoles(skillPath) {
+        const fallback = ["general"];
+        try {
+            const content = readFileSync(skillPath, "utf-8");
+            const match = content.match(/^---[ \t]*\r?\n([\s\S]*?)\r?\n---(?:\r?\n|$)/);
+            if (!match)
+                return fallback;
+            const document = parseDocument(match[1]);
+            if (document.errors.length > 0)
+                throw document.errors[0];
+            const frontmatter = document.toJS();
+            if (!frontmatter || frontmatter.roles === undefined)
+                return fallback;
+            if (!Array.isArray(frontmatter.roles))
+                throw new Error("roles must be an array");
+            const roles = [...new Set(frontmatter.roles)]
+                .filter((value) => value === "general" || value === "worker");
+            if (roles.length !== frontmatter.roles.length || roles.length === 0) {
+                throw new Error("roles must contain only general and/or worker");
+            }
+            return roles;
+        }
+        catch (err) {
+            // Fail closed for workers: malformed or unknown metadata keeps the
+            // backwards-compatible General-only behavior instead of leaking an
+            // administrative skill into worker workspaces.
+            this.logger.warn({ err, skillPath }, "Invalid bundled skill roles — defaulting to General only");
+            return fallback;
         }
-        this.logger.debug({ knowledgeDir, steeringDir }, "Synced general knowledge files");
     }
     /** Fetch forum topic icon stickers and pick emoji IDs for each state */
     async resolveTopicIcons() {
@@ -6077,13 +6842,11 @@ When users create specialized instances, suggest these configurations:
         }
         catch { /* default restart */ }
         const warn = isModelCompatible(backendName, model) ? "" : `⚠️ "${model}" doesn't match ${backendName}'s usual pattern — passing through anyway.\n`;
-        if (strategy === "runtime") {
-            if (!this.instanceIpcClients.get(instanceName))
-                return `${warn}❌ ${instanceName} is not running.`;
-            this.pasteRawToClassicInstance(instanceName, `/model ${model}`);
-            return `${warn}✅ Switched ${instanceName} to \`${model}\` (runtime).${this.effortSuffix(instanceName)}`;
+        if (strategy === "runtime" && !this.instanceIpcClients.get(instanceName)) {
+            return `${warn}❌ ${instanceName} is not running.`;
         }
-        // restart: persist the model so the respawned CLI launches with it.
+        // Persist either way: a runtime switch must survive the next respawn too,
+        // or the instance silently reverts to the CLI default after a fleet restart.
         let persisted = false;
         if (this.fleetConfig?.instances[instanceName]) {
             this.fleetConfig.instances[instanceName].model = model;
@@ -6095,6 +6858,10 @@ When users create specialized instances, suggest these configurations:
         }
         if (!persisted)
             return `${warn}❌ Could not set model for ${instanceName}.`;
+        if (strategy === "runtime") {
+            this.pasteRawToClassicInstance(instanceName, `/model ${model}`);
+            return `${warn}✅ Switched ${instanceName} to \`${model}\` (runtime).${this.effortSuffix(instanceName)}`;
+        }
         await this.restartSingleInstance(instanceName);
         return `${warn}✅ Set ${instanceName} to \`${model}\` and restarted.${this.effortSuffix(instanceName)}`;
     }
@@ -6432,6 +7199,11 @@ When users create specialized instances, suggest these configurations:
         for (const pending of this.pendingClassicStarts.values())
             clearTimeout(pending.timer);
         this.pendingClassicStarts.clear();
+        for (const pending of this.pendingNonceButtons.values()) {
+            if (pending.timer)
+                clearTimeout(pending.timer);
+        }
+        this.pendingNonceButtons.clear();
         this.topicArchiver.stop();
         this.scheduler?.shutdown();
         // Stop instances in parallel batches to avoid long sequential waits.