switchroom 0.15.3 → 0.15.5

package/dist/host-control/main.js

@@ -13825,7 +13825,8 @@ var TelegramChannelSchema = exports_external.object({
   stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI — " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events — stable " + "order, per-tool status emojis, fires only on semantic transitions."),
   stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
   clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message — Reading X, Searching the web for Y, …) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) — no post-then-delete. Per-agent " + "override; cascades defaults → profile → agent (per-key)."),
-  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+  inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes — suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
   orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
   cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
   deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),

package/dist/vault/approvals/kernel-server.js

@@ -11411,7 +11411,8 @@ var init_schema = __esm(() => {
     stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI — " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events — stable " + "order, per-tool status emojis, fires only on semantic transitions."),
     stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
     clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message — Reading X, Searching the web for Y, …) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) — no post-then-delete. Per-agent " + "override; cascades defaults → profile → agent (per-key)."),
-    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes — suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
     orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
     cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
     deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),

package/dist/vault/broker/server.js

@@ -11411,7 +11411,8 @@ var init_schema = __esm(() => {
     stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI — " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events — stable " + "order, per-tool status emojis, fires only on semantic transitions."),
     stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
     clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message — Reading X, Searching the web for Y, …) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) — no post-then-delete. Per-agent " + "override; cascades defaults → profile → agent (per-key)."),
-    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes — suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
     orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
     cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
     deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "switchroom",
-  "version": "0.15.3",
+  "version": "0.15.5",
   "description": "Run Claude Code 24/7 on your Claude Pro/Max subscription over Telegram. Open-source alternative to OpenClaw and NanoClaw — no API keys.",
   "type": "module",
   "bin": {

package/profiles/_base/start.sh.hbs

@@ -620,8 +620,8 @@ APPEND_PROMPT={{#if systemPromptAppendShellQuoted}}{{{systemPromptAppendShellQuo
 # Inject AGENTS.md / SOUL.md / IDENTITY.md / USER.md / TOOLS.md /
 # BOOTSTRAP.md from the agent's workspace/ dir into --append-system-prompt.
 # These files are stable across a session, so baking them into the system
-# prompt is cache-friendly. Dynamic files (MEMORY.md, today's daily,
-# HEARTBEAT.md) are injected via the UserPromptSubmit hook, not here.
+# prompt is cache-friendly. Dynamic files (MEMORY.md, today's daily)
+# are injected via the UserPromptSubmit hook, not here.
 #
 # When channels.telegram.hotReloadStable is true, this injection moves to
 # a UserPromptSubmit hook instead (see workspace-stable-hook.sh).

package/telegram-plugin/dist/gateway/gateway.js

@@ -23936,7 +23936,8 @@ var init_schema = __esm(() => {
     stream_mode: exports_external.enum(["pty", "checklist"]).optional().describe("How live progress is streamed to Telegram during a turn. " + "'pty' (default) surfaces text snapshots of Claude Code's TUI \u2014 " + "compatible but can flicker as Ink re-renders. 'checklist' drives " + "a structured progress card from session-tail events \u2014 stable " + "order, per-tool status emojis, fires only on semantic transitions."),
     stream_throttle_ms: exports_external.number().int().nonnegative().optional().describe("Throttle window in ms between successive stream edits (or " + "sendMessageDraft tics) during a turn. Lower = more responsive " + "stream, higher = fewer API calls. Floored at 250 by draft-stream " + "itself. Default 300 for draft transport (DMs) and 1000 for " + "message transport (groups/forums). Override per-agent if a " + "particular agent needs snappier or quieter streaming."),
     clear_status_on_completion: exports_external.boolean().optional().describe("When true, the live activity/status feed (the in-place 'what it's " + "doing' message \u2014 Reading X, Searching the web for Y, \u2026) is DELETED " + "when the turn's final answer lands, so only the reply remains. " + "Default false: the status message is left in the chat as a record " + "(its last step marked done) \u2014 no post-then-delete. Per-agent " + "override; cascades defaults \u2192 profile \u2192 agent (per-key)."),
-    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md, HEARTBEAT.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    hotReloadStable: exports_external.boolean().optional().describe("If true, the stable workspace prefix (AGENTS.md, SOUL.md, USER.md, " + "IDENTITY.md, TOOLS.md) is re-injected on every turn via " + "the UserPromptSubmit hook instead of baked into --append-system-prompt " + "at session start. Lets workspace edits propagate without a restart. " + "Costs ~5-10% per-turn latency/spend since the stable prefix is no " + "longer prompt-cached."),
+    inject_on_change: exports_external.boolean().optional().describe("Context-efficiency gate for per-turn hook injection (default true). " + "When true (the default), the turn-pacing directive and dynamic " + "workspace content are only re-emitted when their content changes or " + "the session_id changes \u2014 suppressing redundant injection that " + "otherwise triples compaction frequency. Set to false to revert to " + "the legacy always-emit behaviour (every turn injects the full " + "content regardless of whether it changed)."),
     orphan_promotion_ms: exports_external.number().int().nonnegative().optional().describe("How long (ms) a parent turn waits for a sub-agent JSONL watcher " + "to deliver sub_agent_started before the heartbeat promotes the spawn " + "to a synthesised 'running' row. Default 5000. Set to 0 to disable " + "orphan promotion entirely."),
     cold_sub_agent_threshold_ms: exports_external.number().int().nonnegative().optional().describe("JSONL-cold threshold (ms). When a running sub-agent emits no events " + "for this long, the heartbeat synthesises a turn_end for it so the " + "deferred-completion path can proceed. Default 30000. Set to 0 to " + "disable the synthetic close."),
     deferred_completion_timeout_ms: exports_external.number().int().nonnegative().optional().describe("Force-close timeout (ms) for deferred sub-agent completion. After " + "the parent turn_end arrives while sub-agents are still running, the " + "card is force-closed after this many ms even if sub-agents never " + "finish. Watcher-disconnect safety net. Default 180000 (3 min)."),
@@ -42942,7 +42943,8 @@ function formatAuthLine(auth) {
 function formatAgentLine(meta) {
   const m = meta.model && meta.model.length > 0 ? meta.model : "default";
   const topic = meta.topicName ? ` \u00b7 topic: ${escapeHtml6([meta.topicEmoji, meta.topicName].filter(Boolean).join(" "))}` : "";
-  return `<b>${escapeHtml6(meta.agentName)}</b> \u00b7 model: <code>${escapeHtml6(m)}</code>${topic}`;
+  const session = meta.sessionModel && meta.sessionModel.length > 0 ? ` \u00b7 live session: <code>${escapeHtml6(meta.sessionModel)}</code>` : "";
+  return `<b>${escapeHtml6(meta.agentName)}</b> \u00b7 model: <code>${escapeHtml6(m)}</code>${session}${topic}`;
 }
 function startText(agentName3, dmDisabled) {
   if (dmDisabled)
@@ -45016,13 +45018,13 @@ async function buildModelMenu(deps) {
   }
   if (current) {
     const detail = current.detail ? ` \u00b7 ${deps.escapeHtml(current.detail)}` : "";
-    lines.push(`Now: <b>${deps.escapeHtml(current.label)}</b>${detail}`);
+    lines.push(`Default (new sessions): <b>${deps.escapeHtml(current.label)}</b>${detail}`);
   } else {
-    lines.push("Now: <i>unknown (no \u2714 row in picker)</i>");
+    lines.push("Default (new sessions): <i>unknown (no \u2714 row in picker)</i>");
   }
   if (quota)
     lines.push(`Quota: ${deps.escapeHtml(quota)}`);
-  lines.push("", "Tap to switch (applies to the live session):");
+  lines.push("", "Tap a model to switch the <b>live session</b>:");
   lines.push(PERSIST_NOTE);
   return { text: lines.join(`
 `), html: true, keyboard: menuKeyboard(discovered.options) };
@@ -45035,46 +45037,51 @@ async function handleModelMenuCallback(data, deps) {
     return { answer: "Unknown action", reply: await buildModelMenu(deps) };
   }
   if (deps.isBusy()) {
-    return { answer: "Agent is mid-turn \u2014 try again shortly", reply: busyReply(deps) };
+    return {
+      answer: "\u23f3 Agent is mid-turn \u2014 tap again when it\u2019s idle",
+      reply: busyReply(deps),
+      toastOnly: true
+    };
   }
   const tag = data.slice(MODEL_CALLBACK_SELECT.length);
   const discovered = await deps.discover(deps.getAgentName());
   if (!discovered.ok) {
     return {
       answer: "Picker unavailable",
-      reply: {
-        text: `\u274c Could not open the model picker: ${deps.escapeHtml(discovered.reason)}`,
-        html: true
-      }
+      reply: await menuWithBanner(deps, `\u274c Could not open the model picker: ${deps.escapeHtml(discovered.reason)}`)
     };
   }
   const target = discovered.options.find((o) => labelTag(o.label) === tag);
   if (!target) {
-    const fresh2 = await buildModelMenu(deps);
-    return { answer: "Model list changed \u2014 menu refreshed", reply: fresh2 };
-  }
-  if (target.current) {
-    const fresh2 = await buildModelMenu(deps);
-    return { answer: `Already on ${target.label}`, reply: fresh2 };
+    const fresh = await buildModelMenu(deps);
+    return { answer: "Model list changed \u2014 menu refreshed", reply: fresh };
   }
   const result = await deps.select(deps.getAgentName(), target.label);
   if (!result.ok) {
     return {
-      answer: "Switch failed",
-      reply: {
-        text: `\u274c Switch to <b>${deps.escapeHtml(target.label)}</b> failed: ${deps.escapeHtml(result.reason)}`,
-        html: true
-      }
+      answer: "Switch failed \u2014 see the menu",
+      reply: await menuWithBanner(deps, `\u274c Switch to <b>${deps.escapeHtml(target.label)}</b> failed: ${deps.escapeHtml(result.reason)}`)
     };
   }
+  return {
+    answer: deps.escapeHtml(result.confirmation),
+    reply: await menuWithBanner(deps, `\u2705 ${deps.escapeHtml(result.confirmation)}`),
+    selectedModel: sessionModelFromConfirmation(result.confirmation) ?? target.label
+  };
+}
+function sessionModelFromConfirmation(confirmation) {
+  const m = /(?:Set model to|Switched to)\s+(.+?)(?:\s+for (?:this|the) session|\s*\(|\s*$)/i.exec(confirmation.trim());
+  const name = m?.[1]?.trim();
+  return name && name.length > 0 ? name : null;
+}
+async function menuWithBanner(deps, banner) {
   const fresh = await buildModelMenu(deps);
-  const confirmed = {
-    text: [`\u2705 ${deps.escapeHtml(result.confirmation)}`, "", fresh.text].join(`
+  return {
+    text: [banner, "", fresh.text].join(`
 `),
     html: true,
     ...fresh.keyboard ? { keyboard: fresh.keyboard } : {}
   };
-  return { answer: result.confirmation, reply: confirmed };
 }
 
 // ../src/agents/model-picker.ts
@@ -48353,13 +48360,20 @@ function spoolId(msg) {
   const src = msg.meta?.source ?? "-";
   return `s:${msg.chatId}:${src}:${msg.ts}`;
 }
+function escChatKey(msg) {
+  const threadRaw = msg.meta?.threadId;
+  const thread = typeof threadRaw === "string" && threadRaw.length > 0 ? threadRaw : "-";
+  return `${msg.chatId}:${thread}`;
+}
 function createInboundSpool(opts) {
   const { path, fs: fs2 } = opts;
   const now = opts.now ?? Date.now;
   const log = opts.log ?? ((l) => process.stderr.write(l));
   const escalateAfterMs = opts.escalateAfterMs ?? 15 * 60 * 1000;
   const compactAtBytes = opts.compactAtBytes ?? 256 * 1024;
+  const escalateNoticeCooldownMs = opts.escalateNoticeCooldownMs ?? 30 * 60 * 1000;
   const live = new Map;
+  const escAttemptByChat = new Map;
   function parseLine(line) {
     const s = line.trim();
     if (!s)
@@ -48373,8 +48387,15 @@ function createInboundSpool(opts) {
     if (rec == null || typeof rec !== "object")
       return null;
     const r = rec;
-    if (r.t !== "put" && r.t !== "ack")
+    if (r.t !== "put" && r.t !== "ack" && r.t !== "esc")
       return null;
+    if (r.t === "esc") {
+      if (typeof r.chat !== "string" && typeof r.chat !== "number")
+        return null;
+      if (typeof r.at !== "number")
+        return null;
+      return r;
+    }
     if (typeof r.id !== "string" || r.id.length === 0)
       return null;
     if (r.t === "put") {
@@ -48389,6 +48410,7 @@ function createInboundSpool(opts) {
   }
   function hydrate() {
     live.clear();
+    escAttemptByChat.clear();
     if (!fs2.existsSync(path))
       return;
     let raw = "";
@@ -48408,6 +48430,8 @@ function createInboundSpool(opts) {
           msg: rec.msg,
           firstAt: rec.firstAt
         });
+      } else if (rec.t === "esc") {
+        escAttemptByChat.set(`${rec.chat}:${rec.thread ?? "-"}`, rec.at);
       } else {
         live.delete(rec.id);
       }
@@ -48435,6 +48459,17 @@ function createInboundSpool(opts) {
     for (const [id, e] of live) {
       lines.push(JSON.stringify({ t: "put", id, agent: e.agent, msg: e.msg, firstAt: e.firstAt }));
     }
+    for (const [key, at] of escAttemptByChat) {
+      const sep3 = key.lastIndexOf(":");
+      const chat = key.slice(0, sep3);
+      const thread = key.slice(sep3 + 1);
+      lines.push(JSON.stringify({
+        t: "esc",
+        chat,
+        ...thread !== "-" ? { thread } : {},
+        at
+      }));
+    }
     const tmp = path + ".compact.tmp";
     try {
       fs2.writeFileSync(tmp, lines.length ? lines.join(`
@@ -48496,27 +48531,39 @@ function createInboundSpool(opts) {
       return n;
     },
     sweepEscalations(onEscalate) {
-      const cutoff = now() - escalateAfterMs;
-      let n = 0;
+      const tNow = now();
+      const cutoff = tNow - escalateAfterMs;
+      let dropped = 0;
+      let posted = 0;
       for (const [id, e] of [...live.entries()]) {
         if (e.firstAt > cutoff)
           continue;
         live.delete(id);
         appendRecord({ t: "ack", id });
+        const key = escChatKey(e.msg);
+        const lastAttempt = escAttemptByChat.get(key);
+        const postNotice = lastAttempt === undefined || tNow - lastAttempt >= escalateNoticeCooldownMs;
+        escAttemptByChat.set(key, tNow);
+        const threadRaw = e.msg.meta?.threadId;
+        const thread = typeof threadRaw === "string" && threadRaw.length > 0 ? threadRaw : undefined;
+        appendRecord({ t: "esc", chat: e.msg.chatId, thread, at: tNow });
         try {
-          onEscalate({ agent: e.agent, msg: e.msg });
+          onEscalate({ agent: e.agent, msg: e.msg }, { postNotice });
         } catch (err) {
           log(`inbound-spool: onEscalate threw id=${id}: ${err.message}
 `);
         }
-        n++;
+        if (postNotice)
+          posted++;
+        dropped++;
       }
-      if (n > 0) {
-        log(`inbound-spool: escalated+dropped ${n} undelivered entr${n === 1 ? "y" : "ies"} (older than ${escalateAfterMs}ms)
+      if (dropped > 0) {
+        const suppressed = dropped - posted;
+        log(`inbound-spool: escalated+dropped ${dropped} undelivered entr${dropped === 1 ? "y" : "ies"} ` + `(older than ${escalateAfterMs}ms; ${posted} notice${posted === 1 ? "" : "s"} posted` + `${suppressed > 0 ? `, ${suppressed} coalesced` : ""})
 `);
         maybeCompact();
       }
-      return n;
+      return dropped;
     },
     liveCount() {
       return live.size;
@@ -53554,10 +53601,10 @@ function readTurnActiveMarkerAgeMs(stateDir, now) {
 }
 
 // ../src/build-info.ts
-var VERSION = "0.15.3";
-var COMMIT_SHA = "af652154";
-var COMMIT_DATE = "2026-06-10T09:15:23Z";
-var LATEST_PR = 2266;
+var VERSION = "0.15.5";
+var COMMIT_SHA = "fcc9d7b7";
+var COMMIT_DATE = "2026-06-11T21:49:30Z";
+var LATEST_PR = 2282;
 var COMMITS_AHEAD_OF_TAG = 0;
 
 // gateway/boot-version.ts
@@ -57008,11 +57055,13 @@ if (!STATIC) {
       process.stderr.write(`telegram gateway: idle-drain flushed ${r.redelivered}/${r.drained} buffered inbound for ${selfAgent}${r.rebuffered > 0 ? ` (${r.rebuffered} re-buffered)` : ""}
 `);
     }
-    inboundSpool?.sweepEscalations((e) => {
+    inboundSpool?.sweepEscalations((e, { postNotice }) => {
       const chat = e.msg.chatId;
       const escThread = typeof e.msg.meta?.threadId === "string" && e.msg.meta.threadId ? Number(e.msg.meta.threadId) : undefined;
       const threadOpts = escThread != null ? { message_thread_id: escThread } : {};
       reapQueuedStatus(chat, escThread);
+      if (!postNotice)
+        return;
       swallowingApiCall(() => bot.api.sendMessage(chat, "\u26A0\uFE0F I couldn't deliver an earlier message to the agent after repeated retries (it survived restarts but the agent never picked it up). Please resend it.", { ...threadOpts }), { chat_id: chat, verb: "inbound-spool-escalation" });
     });
   }, IDLE_DRAIN_INTERVAL_MS).unref();
@@ -61068,6 +61117,7 @@ function buildAgentAudit(agentName3) {
     return;
   }
 }
+var activeSessionModelOverride = null;
 async function buildAgentMetadata(agentName3) {
   const list2 = switchroomExecJson(["agent", "list"]);
   const brokerState = switchroomExecJson(["auth", "show"]);
@@ -61084,6 +61134,7 @@ async function buildAgentMetadata(agentName3) {
   return {
     agentName: agentName3,
     model: a?.model ?? null,
+    sessionModel: activeSessionModelOverride,
     extendsProfile: a?.extends ?? a?.template ?? null,
     topicName: a?.topic_name ?? null,
     topicEmoji: a?.topic_emoji ?? null,
@@ -63999,9 +64050,19 @@ bot.on("callback_query:data", async (ctx) => {
       }).catch(() => {});
       return;
     }
-    await ctx.answerCallbackQuery({ text: "Working\u2026" }).catch(() => {});
+    const modelDeps = buildModelDeps();
+    if (modelDeps.isBusy()) {
+      await ctx.answerCallbackQuery({ text: "\u23F3 Agent is mid-turn \u2014 tap again when it\u2019s idle", show_alert: false }).catch(() => {});
+      return;
+    }
+    await ctx.answerCallbackQuery({ text: "Switching\u2026" }).catch(() => {});
     try {
-      const outcome = await handleModelMenuCallback(data, buildModelDeps());
+      const outcome = await handleModelMenuCallback(data, modelDeps);
+      if (outcome.selectedModel) {
+        activeSessionModelOverride = outcome.selectedModel;
+      }
+      if (outcome.toastOnly)
+        return;
       await ctx.editMessageText(outcome.reply.text, {
         parse_mode: "HTML",
         reply_markup: modelMenuReplyMarkup(outcome.reply) ?? { inline_keyboard: [] }

package/telegram-plugin/gateway/gateway.ts

@@ -6609,7 +6609,7 @@ if (!STATIC) {
     // promise EXPLICITLY (honest failure) instead of letting it sit
     // forever. This is what makes the guarantee deterministic: every
     // queued message ends either delivered or visibly retracted.
-    inboundSpool?.sweepEscalations((e) => {
+    inboundSpool?.sweepEscalations((e, { postNotice }) => {
       const chat = e.msg.chatId
       const escThread =
         typeof e.msg.meta?.threadId === 'string' && e.msg.meta.threadId
@@ -6620,7 +6620,14 @@ if (!STATIC) {
       // the message is being declared undeliverable, so the queued-status must
       // not dangle beside the "couldn't deliver" notice (idempotent best-effort;
       // a normal turn-start/turn-end reaps far sooner — this is the 15-min edge).
+      // Reaping happens for EVERY dropped entry; only the user-facing notice is
+      // coalesced (postNotice), so a burst of undeliverable inbounds doesn't
+      // leave dangling placeholders even when its notice is suppressed.
       reapQueuedStatus(chat, escThread)
+      // Coalesced per chat by the spool's sliding window — a multi-restart
+      // outage that re-ages a synthetic into the bound every 15 min posts ONE
+      // notice, not one per cycle (the 2026-06-09 marko "please resend" spam).
+      if (!postNotice) return
       void swallowingApiCall(
         () =>
           bot.api.sendMessage(
@@ -13767,6 +13774,14 @@ function buildAgentAudit(agentName: string): AgentAudit | undefined {
 // broker's fleet-wide `ListStateData` payload via
 // `buildAuthSummaryFromBroker`, with billingType pulled from the
 // agent's `.claude.json` (the broker doesn't track plan tier).
+/**
+ * Live session-model override set by the `/model` picker (session-only). Held
+ * in gateway memory so it clears on restart, the same point at which claude's
+ * session reverts to the configured model — keeping `/status` honest without
+ * a persisted store. Null when no session switch is active.
+ */
+let activeSessionModelOverride: string | null = null
+
 async function buildAgentMetadata(agentName: string): Promise<AgentMetadata> {
   type AgentListResp = {
     agents: Array<{
@@ -13795,6 +13810,7 @@ async function buildAgentMetadata(agentName: string): Promise<AgentMetadata> {
   return {
     agentName,
     model: a?.model ?? null,
+    sessionModel: activeSessionModelOverride,
     extendsProfile: (a?.extends ?? a?.template) ?? null,
     topicName: a?.topic_name ?? null,
     topicEmoji: a?.topic_emoji ?? null,
@@ -18409,15 +18425,35 @@ bot.on('callback_query:data', async ctx => {
         .catch(() => {})
       return
     }
-    // Answer the callback IMMEDIATELY — the select path drives the
-    // picker up to three times (multi-second); leaving the tap
-    // spinning invites a double-tap, which queues a second drive
-    // behind the pane lock and confuses the user. The final state is
-    // conveyed by the message edit (a callback can only be answered
-    // once).
-    await ctx.answerCallbackQuery({ text: 'Working…' }).catch(() => {})
+    const modelDeps = buildModelDeps()
+    // Mid-turn refusal is INSTANT (a sync isBusy() check, no picker drive),
+    // so handle it before the "Working…" ack: toast WHY and leave the menu
+    // message untouched (buttons intact) so the operator taps again when
+    // idle. Editing the menu into a button-less "try again" line was the
+    // "nothing happened" report — the menu looked dead.
+    if (modelDeps.isBusy()) {
+      await ctx
+        .answerCallbackQuery({ text: '⏳ Agent is mid-turn — tap again when it’s idle', show_alert: false })
+        .catch(() => {})
+      return
+    }
+    // Ack IMMEDIATELY — the select path drives the picker (multi-second);
+    // leaving the tap spinning invites a double-tap, which queues a second
+    // drive behind the pane lock. A callback can only be answered once, so
+    // the rich result (what was set / why it failed) is conveyed by the
+    // message edit — which now ALWAYS keeps the menu buttons.
+    await ctx.answerCallbackQuery({ text: 'Switching…' }).catch(() => {})
     try {
-      const outcome = await handleModelMenuCallback(data, buildModelDeps())
+      const outcome = await handleModelMenuCallback(data, modelDeps)
+      // Record a successful session switch so /status reflects what's
+      // actually running. In-memory only → clears when the gateway (and thus
+      // claude's session) restarts, exactly matching the session-only scope.
+      if (outcome.selectedModel) {
+        activeSessionModelOverride = outcome.selectedModel
+      }
+      // toastOnly: a no-op outcome that should not disturb the menu (defence
+      // in depth — the isBusy() short-circuit above is the live path).
+      if (outcome.toastOnly) return
       await ctx
         .editMessageText(outcome.reply.text, {
           parse_mode: 'HTML',