switchroom 0.15.45 → 0.16.4

package/telegram-plugin/gateway/gateway.ts

@@ -75,7 +75,8 @@ import {
 } from './permission-timeout.js'
 import { pickRecoveredPermissionOrigin } from './permission-card-origin.js'
 import { isTelegramReplyTool, isTelegramSurfaceTool } from '../tool-names.js'
-import { appendActivityLabel, renderActivityFeedWithNested } from '../tool-activity-summary.js'
+import { appendActivityLabel, clipNarrative, renderActivityFeedWithNested, type SessionActivityHeader } from '../tool-activity-summary.js'
+import { REPLY_TOOLS, isDraftOfReply } from '../narrative-dedup.js'
 import { toolLabel } from '../tool-labels.js'
 import { createTypingWrapper } from '../typing-wrap.js'
 import { type DraftStreamHandle } from '../draft-stream.js'
@@ -97,15 +98,16 @@ import {
   shutdownAnalytics,
 } from '../analytics-posthog.js'
 import { emitRuntimeMetric } from '../runtime-metrics.js'
-import { decideOverPing } from '../over-ping-safety-net.js'
+import { decideOverPing, type OverPingDecision } from '../over-ping-safety-net.js'
 import { decideSilentReplyAnchor } from '../silent-reply-anchor.js'
 import { classifyInbound } from '../inbound-classifier.js'
 import * as silencePoke from '../silence-poke.js'
 import * as pendingProgress from '../pending-work-progress.js'
 import { writeSilentEndState, clearSilentEndState, recordUndeliveredTurnEnd } from '../silent-end.js'
-import { isFinalAnswerReply, isSubstantiveFinalReply } from '../final-answer-detect.js'
+import { isFinalAnswerReply, isSubstantiveFinalReply, FINAL_ANSWER_MIN_CHARS } from '../final-answer-detect.js'
+import { deriveTurnRole, decideTerminalReason, parsePostAnswerLivenessMs, evaluatePostAnswerLiveness, type LoopRole } from '../turn-liveness-floor.js'
 import { createAnswerStream, type AnswerStreamHandle } from '../answer-stream.js'
-import { parseVisibleAnswerStreamEnabled, parseDraftLaneRetiredEnabled, resolveAnswerLaneConfig } from '../answer-stream-flag.js'
+import { parseVisibleAnswerStreamEnabled, resolveAnswerLaneConfig } from '../answer-stream-flag.js'
 import { type SessionEvent } from '../session-tail.js'
 import {
   shouldSuppressToolActivity,
@@ -132,6 +134,7 @@ import {
 } from './microsoft-connect-flow.js'
 import { resolveAuthBrokerSocketPath } from '../../src/auth/broker/client.js'
 import { createFleetFallbackGate } from '../fleet-fallback-gate.js'
+import { createFleetFallbackResumeGate } from '../fleet-fallback-resume.js'
 import { resolveExhaustUntil } from './exhaust-until.js'
 import {
   pendingAuthAddFlows,
@@ -165,7 +168,7 @@ import {
   formatModelUnavailableCard,
   resolveModelUnavailableFromOperatorEvent,
 } from '../model-unavailable.js'
-import { runFleetAutoFallback, renderFallbackFailureNotice, evaluateFallbackFailureNotice, type FallbackFailureNoticeState } from '../auto-fallback-fleet.js'
+import { runFleetAutoFallback, renderFallbackFailureNotice, evaluateFallbackFailureNotice, evaluateAllBlockedNotice, type FallbackFailureNoticeState, type FallbackAllBlockedNoticeState } from '../auto-fallback-fleet.js'
 import { startRestartWatchdog } from './restart-watchdog.js'
 import { validateStringArray } from './access-validator.js'
 
@@ -221,6 +224,7 @@ import {
   isContextExhaustionText,
   shouldArmOrphanedReplyTimeout,
   ORPHANED_REPLY_TIMEOUT_MS,
+  ORPHANED_REPLY_MAX_REARMS,
 } from '../context-exhaustion.js'
 import {
   decideTurnFlush,
@@ -326,11 +330,24 @@ import {
 } from './obligation-ledger.js'
 import { loadObligations, persistObligations } from './obligation-store.js'
 import { driveEscalation } from './escalation-drive.js'
+import { shouldSuppressRepresent } from './represent-guard.js'
 import { createInboundSpool } from './inbound-spool.js'
 import { purgeStaleTurnsForChat } from './turn-state-purge.js'
 import { decideInboundDelivery } from './inbound-delivery-gate.js'
 import { mayDrainBufferedInbound, shouldArmNoReplyDrain } from './serialize-drain-gate.js'
 import { decideFeedReopen } from './feed-reopen-gate.js'
+import {
+  mayOpenActivityCard,
+  computeCrossTurnAnswerDelivered,
+  type FeedOpenProducer,
+  type FeedOpenGateDeps,
+} from './feed-open-gate.js'
+import {
+  EmissionAuthority,
+  EMISSION_AUTHORITY_ENABLED,
+  type CardDrainGateCtx,
+} from './emission-authority.js'
+import { CurrentTurnMap } from './current-turn-map.js'
 import { resolveAnswerThreadId } from './answer-thread-resolve.js'
 import {
   createDeliveryQueue,
@@ -411,6 +428,7 @@ import {
   // preceding shutdown only" semantics.
   clearCleanShutdownMarker,
   shouldSuppressRecoveryBanner,
+  shouldSuppressBootResume,
   resolveShutdownMarker,
   DEFAULT_MAX_AGE_MS as CLEAN_SHUTDOWN_MAX_AGE_MS,
 } from './clean-shutdown-marker.js'
@@ -468,8 +486,10 @@ import {
   resolveQuotaWatchTuning,
   buildQuotaClaimKey,
   QUOTA_WATCH_CLAIM_WINDOW_MS,
+  isLiveCorroboration,
 } from '../quota-watch.js'
 import { buildSnapshotsFromState, buildSnapshotsFromCachedState } from '../auth-snapshot-format.js'
+import { maskUsername, maskVaultKey } from '../demo-mask.js'
 import {
   writeTurnActiveMarker,
   touchTurnActiveMarker,
@@ -739,20 +759,6 @@ const AGENT_ADMIN = process.env.SWITCHROOM_AGENT_ADMIN === 'true'
 const bot = new Bot(TOKEN)
 installTgPostLogger(bot)
 
-// Draft-answer-lane retirement (2026-06-05): default RETIRED so the live answer
-// lane uses a real, mtcute-observable message instead of the invisible
-// compose-box draft. Declared HERE (above the boot-probe block) because
-// `sendMessageDraftFn` below reads it — keep it above its first use to avoid a
-// temporal-dead-zone ReferenceError at boot. Kill switch
-// SWITCHROOM_DRAFT_ANSWER_LANE=0 restores the legacy draft.
-const DRAFT_ANSWER_LANE_RETIRED = parseDraftLaneRetiredEnabled(process.env.SWITCHROOM_DRAFT_ANSWER_LANE)
-
-// ─── sendMessageDraft boot probe ──────────────────────────────────────────
-// grammY 1.x exposes all Telegram Bot API methods through bot.api.raw.
-// bot.api.sendMessageDraft (the typed wrapper) takes chat_id as number, but
-// answer-stream passes chatId as string, so we bridge through raw with an
-// explicit Number() cast and positional → object param translation.
-const _rawSendMessageDraft = (bot.api.raw as unknown as Record<string, unknown>).sendMessageDraft
 const GRAMMY_VERSION: string = (() => {
   try {
     const raw = readFileSync(new URL('../../node_modules/grammy/package.json', import.meta.url), 'utf8')
@@ -761,22 +767,6 @@ const GRAMMY_VERSION: string = (() => {
     return 'unknown'
   }
 })()
-const sendMessageDraftFn: (
-  (chatId: string, draftId: number, text: string, params?: { message_thread_id?: number; parse_mode?: 'HTML' }) => Promise<unknown>
-) | undefined =
-  // When the draft lane is retired (default), force this undefined so BOTH
-  // consumers (the answer-stream config + the stream_reply handler) drop the
-  // draft transport and fall back to visible message transport — the single
-  // chokepoint for the retirement.
-  !DRAFT_ANSWER_LANE_RETIRED && typeof _rawSendMessageDraft === 'function'
-    ? (chatId, draftId, text, params) =>
-        (_rawSendMessageDraft as (args: Record<string, unknown>) => Promise<unknown>)({
-          chat_id: Number(chatId),
-          draft_id: draftId,
-          text,
-          ...(params ?? {}),
-        })
-    : undefined
 
 // ─── sendChecklist / editMessageChecklist boot probes ─────────────────────
 // grammY 1.x exposes new Telegram Bot API methods via bot.api.raw before the
@@ -1157,43 +1147,70 @@ try {
   const pending = findLatestTurnIfInterrupted(turnsDb)
   const selfAgent = process.env.SWITCHROOM_AGENT_NAME ?? ''
   if (pending != null && selfAgent) {
-    // 3h staleness failsafe (operator spec, 2026-06-03): never AUTO-resume
-    // interrupted work older than RESUME_MAX_AGE_MS — selectResumeBuilder
-    // downgrades a stale 'resume' to the passive 'report' so the user is told
-    // ("I was working on X ~Nh ago") but nothing replays unprompted. Env
-    // override SWITCHROOM_RESUME_MAX_AGE_MS (ms); set very high to disable.
-    const RESUME_MAX_AGE_MS = (() => {
-      const v = Number(process.env.SWITCHROOM_RESUME_MAX_AGE_MS)
-      return Number.isFinite(v) && v > 0 ? v : 10_800_000 // 3h
-    })()
-    const kind = selectResumeBuilder(pending.ended_via, {
-      ageMs: Math.max(0, Date.now() - pending.started_at),
-      maxAgeMs: RESUME_MAX_AGE_MS,
+    // Clean-shutdown gate: suppress auto-resume when the prior shutdown was
+    // operator/roll/CLI-initiated (clean). A clean-shutdown marker present and
+    // fresh means the agent was asked to stop; the "interrupted" turn was
+    // abandoned by that decision. Replaying it on every planned restart wastes
+    // subscription quota for no user benefit. Only unclean exits (crash/OOM/
+    // unexpected kill) should auto-resume.
+    //
+    // NOTE: GATEWAY_CLEAN_SHUTDOWN_MARKER_PATH is defined lower in this file
+    // (module-init order); we compute the path inline here using the same
+    // formula so we can read it at boot-resume time.
+    // SWITCHROOM_BOOT_RESUME_ALWAYS=1 is an escape hatch that restores
+    // unconditional resume if needed.
+    const bootResumeMarkerPath =
+      process.env.SWITCHROOM_GATEWAY_CLEAN_SHUTDOWN_MARKER ?? join(STATE_DIR, 'clean-shutdown.json')
+    const bootResumeCleanMarker = readCleanShutdownMarker(bootResumeMarkerPath)
+    const bootResumeForceAlways = process.env.SWITCHROOM_BOOT_RESUME_ALWAYS === '1'
+    const bootResumeSuppressed = shouldSuppressBootResume(bootResumeCleanMarker, Date.now(), {
+      forceAlways: bootResumeForceAlways,
     })
-    if (kind === 'resume') {
-      bootResumeInbound = { agent: selfAgent, msg: buildResumeInterruptedInbound({ turn: pending }) }
-    } else if (kind === 'report') {
-      // idleMs: this boot's measured marker age if it just classified this
-      // turn; otherwise recover it from the persisted interrupt_reason (a
-      // later boot, marker already swept); else fall back to total runtime.
-      let idleMs = pending.turn_key === timeoutTurnKey && markerAgeMs != null ? markerAgeMs : null
-      if (idleMs == null && pending.interrupt_reason) {
-        try {
-          const parsed = JSON.parse(pending.interrupt_reason) as { idleMs?: unknown }
-          if (typeof parsed.idleMs === 'number' && Number.isFinite(parsed.idleMs)) idleMs = parsed.idleMs
-        } catch { /* malformed snapshot — fall through */ }
-      }
-      if (idleMs == null) idleMs = Math.max(0, Date.now() - pending.started_at)
-      bootResumeInbound = {
-        agent: selfAgent,
-        msg: buildResumeWatchdogReportInbound({ turn: pending, idleMs }),
-      }
-    }
-    if (bootResumeInbound != null) {
+    if (bootResumeSuppressed) {
       process.stderr.write(
-        `telegram gateway: boot-resume queued kind=${kind} turnKey=${pending.turn_key} ` +
-        `endedVia=${pending.ended_via ?? 'open'} chat=${pending.chat_id}\n`,
+        `telegram gateway: boot-resume suppressed (clean shutdown` +
+        `${bootResumeCleanMarker?.reason ? ` reason=${JSON.stringify(bootResumeCleanMarker.reason)}` : ''}` +
+        `) — unclean exits still resume turnKey=${pending.turn_key}\n`,
       )
+    } else {
+      // 3h staleness failsafe (operator spec, 2026-06-03): never AUTO-resume
+      // interrupted work older than RESUME_MAX_AGE_MS — selectResumeBuilder
+      // downgrades a stale 'resume' to the passive 'report' so the user is told
+      // ("I was working on X ~Nh ago") but nothing replays unprompted. Env
+      // override SWITCHROOM_RESUME_MAX_AGE_MS (ms); set very high to disable.
+      const RESUME_MAX_AGE_MS = (() => {
+        const v = Number(process.env.SWITCHROOM_RESUME_MAX_AGE_MS)
+        return Number.isFinite(v) && v > 0 ? v : 10_800_000 // 3h
+      })()
+      const kind = selectResumeBuilder(pending.ended_via, {
+        ageMs: Math.max(0, Date.now() - pending.started_at),
+        maxAgeMs: RESUME_MAX_AGE_MS,
+      })
+      if (kind === 'resume') {
+        bootResumeInbound = { agent: selfAgent, msg: buildResumeInterruptedInbound({ turn: pending }) }
+      } else if (kind === 'report') {
+        // idleMs: this boot's measured marker age if it just classified this
+        // turn; otherwise recover it from the persisted interrupt_reason (a
+        // later boot, marker already swept); else fall back to total runtime.
+        let idleMs = pending.turn_key === timeoutTurnKey && markerAgeMs != null ? markerAgeMs : null
+        if (idleMs == null && pending.interrupt_reason) {
+          try {
+            const parsed = JSON.parse(pending.interrupt_reason) as { idleMs?: unknown }
+            if (typeof parsed.idleMs === 'number' && Number.isFinite(parsed.idleMs)) idleMs = parsed.idleMs
+          } catch { /* malformed snapshot — fall through */ }
+        }
+        if (idleMs == null) idleMs = Math.max(0, Date.now() - pending.started_at)
+        bootResumeInbound = {
+          agent: selfAgent,
+          msg: buildResumeWatchdogReportInbound({ turn: pending, idleMs }),
+        }
+      }
+      if (bootResumeInbound != null) {
+        process.stderr.write(
+          `telegram gateway: boot-resume queued kind=${kind} turnKey=${pending.turn_key} ` +
+          `endedVia=${pending.ended_via ?? 'open'} chat=${pending.chat_id}\n`,
+        )
+      }
     }
   }
 
@@ -1425,6 +1442,21 @@ const activeTurnStartedAt = new Map<string, number>()
 // reading activeTurnStartedAt because they want the receipt timestamp.
 const claudeBusyKeys = new Set<string>()
 
+/**
+ * #2527 observability: count emoji transitions per status-reaction controller
+ * so `turn_no_reply_warn` can report how many reaction changes happened while
+ * producing zero text. Keyed by statusKey(chatId, threadId); cleared in
+ * purgeReactionTracking alongside the controller itself.
+ */
+const reactionTransitionCounts = new Map<string, number>()
+
+/**
+ * #2527 observability: tracks which (chatId:threadId) keys have already emitted
+ * a `turn_reply_timing` event this turn so we only fire it on the FIRST text
+ * reply. Cleared in purgeReactionTracking at turn-end alongside the controller.
+ */
+const firstTextReplyLogged = new Set<string>()
+
 /**
  * Helper: stamp a claudeBusyKeys entry for an inbound about to be
  * handed to claude. Pulls the thread id from the top-level field if
@@ -1487,6 +1519,19 @@ const deliveryQueue = createDeliveryQueue<InboundMessage>()
 // SWITCHROOM_OBLIGATION_LEDGER=0 → every hook below is a no-op → zero change.
 const OBLIGATION_LEDGER_ENABLED = process.env.SWITCHROOM_OBLIGATION_LEDGER !== '0'
 const OBLIGATION_REPRESENT_MAX = 2
+// Minimum reply length (chars) the duplicate-represent guard (#2472/#2474) treats
+// as "the user was answered". DECOUPLED from the escalate branch's 200-char proxy:
+// for the represent guard ANY genuine assistant reply — even a terse "Yes — done."
+// — satisfies the obligation, so suppressing the duplicate re-ask must not require
+// 200 chars. Default 1 (any non-empty real reply; empty/whitespace rows are
+// clamped out inside hasOutboundDeliveredSince). Override via env for tuning. Safe
+// because only recordOutbound writes role='assistant' rows — progress-card edits
+// and typing indicators are never counted.
+const OBLIGATION_REPRESENT_GUARD_MIN_REPLY_CHARS = (() => {
+  const raw = process.env.SWITCHROOM_OBLIGATION_REPRESENT_GUARD_MIN_REPLY_CHARS
+  const n = raw != null ? Number.parseInt(raw, 10) : NaN
+  return Number.isFinite(n) && n >= 1 ? n : 1
+})()
 const OBLIGATION_SWEEP_MS = 5_000
 // Bound on escalation SEND attempts. The escalation now closes only AFTER a
 // successful send (a transient failure stays OPEN and retries next sweep), so a
@@ -1710,6 +1755,38 @@ const FEED_REOPEN_AFTER_ACK_ENABLED =
 const FEED_HEARTBEAT_ENABLED = process.env.SWITCHROOM_FEED_HEARTBEAT !== '0'
 const FEED_HEARTBEAT_TICK_MS = 6_000
 const FEED_HEARTBEAT_MIN_STALE_MS = 6_000
+// Liveness-driven feed open. The activity feed is otherwise TOOL-driven — it
+// opens only when a tool emits a non-null label. A turn dominated by thinking
+// or by suppressed-by-design tools (typing / memory recall / reply) emits no
+// label, so the feed never opens and a long turn reads as pure silence until
+// the 300s silence-poke (the #680 dark-turn). When a turn has been alive >=
+// FEED_LIVENESS_OPEN_MS with no feed yet, open a minimal "Working…" feed so the
+// user always has a live indicator; the first real tool label edits it with
+// real content. Runs on the heartbeat interval, so the effective open lands in
+// [threshold, threshold + FEED_HEARTBEAT_TICK_MS). Kill switch:
+// SWITCHROOM_FEED_LIVENESS_OPEN=0. Default on.
+const FEED_LIVENESS_OPEN_ENABLED = process.env.SWITCHROOM_FEED_LIVENESS_OPEN !== '0'
+const FEED_LIVENESS_OPEN_MS = (() => {
+  const raw = process.env.SWITCHROOM_FEED_LIVENESS_OPEN_MS
+  const n = raw ? Number(raw) : NaN
+  return Number.isFinite(n) && n > 0 ? n : 12_000
+})()
+
+// Post-answer background-agent liveness STALENESS CAP (Fix 2 / #2587 supersede,
+// concern 3). The `feedHeartbeatTick` post-answer branch re-renders a "background
+// agent still working" card every FEED_HEARTBEAT_TICK_MS while the sub-agent
+// watcher keeps advancing `turn.subagentActivityAt`. Without a cap that card kept
+// emitting `state:'running'` with an ever-climbing `elapsed` FOREVER — even after
+// the worker's `onFinish` froze the timestamp — because (unlike the pre-answer
+// path's `FEED_LIVENESS_OPEN_MS` recency cap) the post-answer branch had no
+// staleness bound. This cap mirrors that pre-answer pattern: once the worker's
+// last advance is older than the cap, the heartbeat stops re-rendering and the
+// card freezes at its last state. Parsed via the same pure `parsePostAnswerLivenessMs`
+// helper (positive int or 0); `|| 30_000` supplies a default-ON 30s cap, so an
+// unset env keeps the cap active. Override with SWITCHROOM_POST_ANSWER_LIVENESS_STALE_MS.
+const POST_ANSWER_LIVENESS_STALE_MS = parsePostAnswerLivenessMs(
+  process.env.SWITCHROOM_POST_ANSWER_LIVENESS_STALE_MS,
+) || 30_000
 
 /** Compact mm/ss-ish elapsed for the live feed suffix: "18s", "1m05s". */
 function formatFeedElapsed(ms: number): string {
@@ -1884,6 +1961,27 @@ type CurrentTurn = {
   sourceMessageId: number | null
   startedAt: number
   gatewayReceiveAt: number
+  // #2527 — the single turn-provenance discriminator, stamped once at
+  // enqueue from the channel envelope `source`. `user` (a human is waiting:
+  // never-silent guarantee + mid-turn floor), `system` (cron/scheduled:
+  // silence is legitimate). The gateway never builds a turn atom for a
+  // sub-agent, so `sub-agent` never appears here. Read by the mid-turn floor
+  // eligibility and the role-aware terminal reaction gate. Replaces the
+  // scattered `chatType`/`chatId==null`/`source==='cron'` predicates.
+  role: LoopRole
+  // PR1 (cross-turn stale-card guard, design `docs/message-emission-determinism.md`
+  // §9 lever 4 / race C/D). Present ONLY when this turn is a cross-turn SYNTHETIC
+  // surface whose card OPEN must be gated against an answer already delivered in
+  // an EARLIER turn — i.e. an `obligation_represent` re-delivery (and the
+  // liveness/heartbeat timer firing on it). `sinceMs` is the obligation's
+  // `openedAt` — the moment the obligation was RAISED — so the card-OPEN gate
+  // asks `hasOutboundDeliveredSince(chat, openedAt)`: did a substantive answer
+  // already land since then? Stamped at the turn ctor from a pending marker that
+  // `obligationSweep` writes when it pushes the represent inbound (see
+  // `pendingCrossTurnGate`). `undefined` for a normal foreground turn → the
+  // cross-turn lever-4 gate is inert there (the foreground turn's own card is
+  // governed only by the per-turn `finalAnswerEverDelivered` latch, lever 1).
+  crossTurnGate?: { sinceMs: number }
   replyCalled: boolean
   // #1664 — whether the model has delivered its *final answer* this turn
   // (as opposed to only an interim ack). `replyCalled` flips on the first
@@ -1912,6 +2010,18 @@ type CurrentTurn = {
   // Reset to false on every fresh-turn enqueue alongside
   // `finalAnswerDelivered`.
   finalAnswerSubstantive: boolean
+  // Sticky "a substantive final answer has been delivered this turn" latch
+  // (design `docs/message-emission-determinism.md` §9 preamble / R0). Distinct
+  // from the MUTABLE `finalAnswerDelivered`, which the ack-reopen path clears
+  // mid-turn (`feed-reopen-gate.ts:157`) so an "On it…" ack keeps a live feed
+  // (#2141). Ordering gates (the no-OPEN-after-final card gate, lever 1) MUST
+  // key on this sticky latch, not the mutable flag — keying on the mutable flag
+  // is a no-op on exactly the ack-first turn where the reorder originates. Set
+  // true ONLY at the points that set `finalAnswerDelivered = true` AND only when
+  // the reply is `isSubstantiveFinalReply`; NEVER cleared by reopen. Reset to
+  // false ONLY at turn start, mirroring `activityEverOpened`'s sticky-true
+  // contract.
+  finalAnswerEverDelivered: boolean
   // #1675 (over-ping safety net): wall-clock ms of the first reply
   // this turn that landed with `disable_notification: false` (a real
   // device ping). The conversational-pacing contract
@@ -1923,6 +2033,18 @@ type CurrentTurn = {
   // the framework. Null until the first ping lands. Reset on every
   // fresh-turn enqueue.
   firstPingAt: number | null
+  // Notification ownership (R8 / PR-2 — design `docs/message-emission-
+  // determinism.md` §over-ping). Whether the send that CLAIMED this turn's
+  // ping slot (`firstPingAt`) was itself a *substantive* final answer
+  // (`isSubstantiveFinalReply`) as opposed to a short interim ACK. The
+  // over-ping safety net keys on this so a substantive answer pinging AFTER
+  // an ack already pinged is UPGRADED (let through, owns the slot) rather
+  // than silenced — otherwise "the reply is last but the phone never buzzed
+  // for the answer." Set ATOMICALLY with `firstPingAt` (same synchronous
+  // block, no await between) on a claim/upgrade so a racing second reply
+  // reads a consistent pair. Init false; reset to false on every fresh-turn
+  // enqueue alongside `firstPingAt`.
+  firstPingWasSubstantive: boolean
   // #1677 silent-reply auto-edit. The first silent reply of a turn
   // captures `silentAnchorMessageId` + `silentAnchorText`; subsequent
   // silent replies in the SAME turn editMessageText that anchor
@@ -1935,6 +2057,13 @@ type CurrentTurn = {
   silentAnchorText: string
   capturedText: string[]
   orphanedReplyTimeoutId: ReturnType<typeof setTimeout> | null
+  // How many times the orphaned-reply backstop timer has been re-armed
+  // mid-tool-call instead of firing a synthetic turn_end. Bounded so a
+  // genuinely wedged single long-running tool still surfaces: the cap is
+  // ORPHANED_REPLY_MAX_REARMS (20 × 30 s = 10 min of genuine tool activity).
+  // Reset to 0 on a fresh enqueue; NOT reset on text/tool_label re-arms —
+  // only a new turn resets the budget.
+  orphanedReplyRearmCount: number
   // Component 3 (turn-origin reply routing). A stable per-turn identity,
   // `${registryKey-or-chatKey}#${startedAt}`, assigned when the turn
   // starts and stamped into the inbound meta (`origin_turn_id`) so a reply
@@ -1962,6 +2091,14 @@ type CurrentTurn = {
   // Phase 1 of #332: count of tool_use events in the current turn, for
   // the tool_call_count column in the turns registry.
   toolCallCount: number
+  // Count of tool_label events that passed the isTelegramSurfaceTool guard
+  // this turn — the deterministic, surface-tool-excluded step count used by
+  // the `✓ N steps` activity-feed total and the `tools=` lifecycle log.
+  // Incremented in `case 'tool_label':` AFTER the surface-tool guard so
+  // reply/stream_reply/edit_message/react are never counted. send_typing and
+  // sync_retain are suppressed at the hook (computeLabel returns null) and
+  // never arrive as tool_label events — excluded automatically.
+  labeledToolCount: number
   // Tool-activity summary — mirrors Claude Code's native chat-UI
   // rendering ("Ran 5 commands, read a file"). Counters are
   // incremented in `case 'tool_use'`; `activityMessageId` holds the
@@ -2002,11 +2139,45 @@ type CurrentTurn = {
   // step that emits no new label doesn't read as frozen (the feed is otherwise
   // pull-only). undefined until the first label of the turn renders.
   lastToolLabelAt?: number
+  // Fix 2 (post-answer background-agent liveness): wall-clock timestamp last
+  // updated by the sub-agent/workflow watcher's onProgress callback whenever
+  // it surfaces a NEW sub-agent step AFTER this turn's substantive answer was
+  // delivered. Written INDEPENDENTLY of the tool_label path so the drop-guard
+  // (`shouldReopenFeedAfterAck` / `finalAnswerSubstantive`) cannot gate it.
+  // `feedHeartbeatTick` reads THIS (not `lastToolLabelAt`, which is frozen by
+  // the drop-guard) to drive the post-answer liveness card — the core fix for
+  // #2587's inert state. undefined until the first post-answer watcher advance.
+  subagentActivityAt?: number
+  // Sticky wall-clock timestamp when finalAnswerEverDelivered first latched
+  // true this turn. Allows the heartbeat to distinguish "tool label arrived
+  // before the answer" (lastToolLabelAt ≤ finalAnswerDeliveredAt, inert) from
+  // "sub-agent active after the answer" (subagentActivityAt >
+  // finalAnswerDeliveredAt, liveness card warranted). undefined until the
+  // first substantive final answer of the turn.
+  finalAnswerDeliveredAt?: number
   // Accumulating friendly-action feed for this turn. Each non-surface
   // tool_label appends a line via `appendActivityLabel`; the feed renders
   // (via `renderActivityFeed`) as a capped chronological list into the
   // in-place edited activity message and clears on reply. Reset per turn.
   mirrorLines: string[]
+  // Narrative-dedup gate state (JSONL-text-narrative primitive). A `text`
+  // block is held here for ONE lookahead step so the next event (a tool_use
+  // or turn_end) can decide draft-then-send (SUPPRESS, it duplicates the
+  // reply) vs working-narration (SHOW it as a transient mirrorLines step).
+  // Null when nothing is pending. The pure decision lives in
+  // narrative-dedup.ts; this slot is the per-turn cursor. Reset per turn.
+  // Invariant `chat-is-the-single-source-of-truth`: a SHOWN narrative is
+  // rendered through the SAME appendActivityLabel→renderStepFeed path as a
+  // tool step — a transient, clipped, rolling-window line replaced by the
+  // next event, never a persisted parallel mirror.
+  pendingNarrative: { text: string } | null
+  // Most-recently-seen reply/stream_reply `input.text` for this turn — the
+  // ACTUAL delivered answer surface. Set wherever a REPLY_TOOL tool_use is
+  // handled in the reducer. `flushPendingNarrativeAtTurnEnd` compares a
+  // trailing narrative block against THIS (not capturedText.join(''), which
+  // can mis-suppress when the model emits the same short string twice in a
+  // turn). Empty string until the turn delivers a reply. Reset per turn.
+  lastReplyText: string
   // Model A — foreground sub-agent nesting. A foreground sub-agent (Task/Agent
   // with no run_in_background) runs INSIDE this turn while the parent blocks at
   // the Task tool, so its live steps nest under the parent's activity feed
@@ -2019,9 +2190,169 @@ type CurrentTurn = {
   // gates on minInitialChars). Materialized and cleared at turn_end.
   answerStream: AnswerStreamHandle | null
   isDm: boolean
+  // PR-4a (message-emission-determinism, `emission-authority.ts`). The
+  // per-foreground-turn emission-authority façade the foreground-lane card/ping
+  // emission call sites route through. Constructed ONCE per turn in the ctor
+  // with the chat/thread key passed in explicitly (the PR-4e seam). Per-turn
+  // only — a fresh `CurrentTurn` literal gets a fresh façade, so it never
+  // persists across turns. Optional in the type so the bounded recently-ended
+  // registry's older entries (and any hand-built test turn) tolerate its
+  // absence; `emissionAuthorityFor` lazily backfills one when missing.
+  emissionAuthority?: EmissionAuthority
+}
+
+// PR-4e — the singleton `currentTurn` is RETAINED as (a) the flag-OFF store and
+// (b) the flag-ON "most-recent-set" MIRROR. Every GLOBAL-liveness read in this
+// file (`isBusy`, the `if (currentTurn != null) return` poke guards, the
+// orphaned-reply guard, the synchronous-to-live-turn `const turn = currentTurn`
+// captures) keeps reading this variable, so under the sequential-CLI invariant
+// (the most-recently-set turn IS the live turn) those reads stay byte-identical.
+// The per-topic isolation lives in `currentTurnMap.byKey`: a LATE async event
+// captured for topic A resolves A's authority by ITS OWN key even after the live
+// turn flipped to topic B (see current-turn-map.ts). Under the flag OFF the map
+// is never written and this is exactly the old singleton.
+let currentTurn: CurrentTurn | null = null
+const currentTurnMap = new CurrentTurnMap<CurrentTurn>()
+
+/**
+ * Set the live turn for `key`. Flag-branches in ONE place (inside the map):
+ * flag-OFF assigns the singleton only; flag-ON sets the per-topic entry AND
+ * updates the most-recent mirror. We keep the module-scope `currentTurn`
+ * variable in lock-step with the map's mirror so the 140 unchanged global reads
+ * see the same value.
+ */
+function setCurrentTurn(turn: CurrentTurn, key: string): void {
+  currentTurnMap.set(turn, key)
+  currentTurn = currentTurnMap.get() // mirror most-recent-set (== `turn`)
 }
 
-let currentTurn: CurrentTurn | null = null
+/**
+ * End (delete) the live turn for `key`, iff `key` still maps to `turn`. Routes
+ * the clear through the keyed accessor (leak-close-at-origin) and re-syncs the
+ * module-scope mirror to the map's mirror.
+ */
+function endCurrentTurnForKey(turn: CurrentTurn, key: string): boolean {
+  const ended = currentTurnMap.endTurnForKey(turn, key)
+  currentTurn = currentTurnMap.get() // re-sync mirror (null iff it pointed at turn)
+  return ended
+}
+
+/**
+ * Clear the ENTIRE per-topic store + mirror (disconnect-flush / bridge-died:
+ * every entry is a ghost).
+ */
+function clearAllCurrentTurns(): void {
+  currentTurnMap.clearAll()
+  currentTurn = null
+}
+
+/**
+ * Is `turn` still the live turn FOR ITS OWN topic? Flag-OFF: `currentTurn ===
+ * turn` (the ambient check, verbatim). Flag-ON: `byKey.get(turn'sKey) === turn`,
+ * so a B-flip never falsifies A's own liveness. The callsites keep the literal
+ * `currentTurn === turn` in source (the silence-liveness-wiring oracle) by
+ * inlining the flag-OFF branch and delegating the flag-ON branch here.
+ */
+function turnLiveForItsTopic(turn: CurrentTurn): boolean {
+  return currentTurnMap.isLiveForKey(
+    turn,
+    statusKey(turn.sessionChatId, turn.sessionThreadId),
+  )
+}
+
+/**
+ * Accessor for a turn's per-foreground-turn emission-authority façade (PR-4a).
+ * Returns the façade constructed at the turn ctor; lazily backfills one (keyed
+ * on the turn's chat/thread) for any turn that predates the field or was built
+ * outside the ctor. Per-turn: the memoized instance lives on the turn object,
+ * so it is discarded with the turn and never persists across turns.
+ */
+function emissionAuthorityFor(turn: CurrentTurn): EmissionAuthority {
+  if (turn.emissionAuthority == null) {
+    turn.emissionAuthority = new EmissionAuthority(
+      statusKey(turn.sessionChatId, turn.sessionThreadId),
+    )
+  }
+  // PR-4b — CENTRALIZE the OPEN-gate wiring here, the single accessor every
+  // routed call site already funnels through, so all 6 `openOrEditCard(...)`
+  // sites stay byte-identical `(producer, apply)`. The façade reads the LIVE
+  // turn view (a thunk — the card id / latch / tool-count mutate during the
+  // turn) + the injected history deps from this one place, not per-call.
+  // Idempotent; harmless under the flag OFF (the disabled branch never reads
+  // it). The turn IS a structural `FeedOpenGateView`.
+  turn.emissionAuthority.wireFeedOpenGate(() => turn, feedOpenGateDeps())
+  return turn.emissionAuthority
+}
+
+/**
+ * The injected history dependencies the PR-4b OPEN gate needs (the real
+ * `hasOutboundDeliveredSince` predicate + `HISTORY_ENABLED` + the substantive
+ * `FINAL_ANSWER_MIN_CHARS` floor). Centralized so both the façade's enabled
+ * branch AND the drain's own (now-redundant) inline gate consume the SAME deps
+ * via the SAME pure helpers — flag-ON and flag-OFF cannot diverge. Keeps
+ * `feed-open-gate.ts` sqlite-free (it never imports `history.js`).
+ */
+function feedOpenGateDeps(): FeedOpenGateDeps {
+  return {
+    hasOutboundDeliveredSince,
+    historyEnabled: HISTORY_ENABLED,
+    finalAnswerMinChars: FINAL_ANSWER_MIN_CHARS,
+  }
+}
+
+/**
+ * The deliver-before-drain inputs the PR-4d card-drain gate threads into the
+ * façade's pure `mayDrainCardNow`. Centralized so the card-drain helper sources
+ * the SAME `turnInFlightForGate()` + kill-switch the buffer-drain gate uses.
+ *
+ * `endingTurnFinalAnswerDelivered` is FIXED to `null` for the card path (§5
+ * modeling decision): the live foreground card single-flight is governed by
+ * `turn.activityInFlight` (via `mayDrain`), NOT by an ending turn's delivery
+ * state — so `mayDrainBufferedInbound` degenerates to `!turnInFlight` and a
+ * synthetic represent turn (finalAnswerDelivered=false) can never wedge the card.
+ */
+function cardDrainGateCtx(): CardDrainGateCtx {
+  return {
+    turnInFlight: turnInFlightForGate(),
+    endingTurnFinalAnswerDelivered: null,
+    enabled: SERIALIZE_UNTIL_REPLIED_ENABLED,
+  }
+}
+
+/**
+ * PR-4d centralized card-drain gate (Option A). The 6 foreground card-drain
+ * sites pass their EXISTING single-flight-guarded block (the `mayDrain` guard +
+ * the `openOrEditCard(producer, …)` thunk that assigns `turn.activityInFlight =
+ * drainActivitySummary(…)`) in VERBATIM as `run`, so those load-bearing literals
+ * stay byte-identical (the wiring oracles still see them).
+ *
+ *  - **Flag OFF (default):** runs the guarded block directly — NO `chatLock`
+ *    wrapper, byte-equivalent to base.
+ *  - **Flag ON:** acquires `chatLock` AROUND the deliver-before-drain decision
+ *    (`mayDrainCardNow`) + the synchronous block, unifying the card path with
+ *    the #2137 serialization gate. The lock spans ONLY the gate decision + the
+ *    synchronous `openOrEditCard` kick-off inside `run` (which only ASSIGNS
+ *    `turn.activityInFlight = drainActivitySummary(...)`; the async send is NOT
+ *    awaited inside the lock). The lock is released before any drain
+ *    `await sendMessage` suspends, so a card OPEN never holds `chatLock` across
+ *    the gate's release — a synthetic represent turn can never wedge the gate,
+ *    and `mayDrain` stays callable lock-free.
+ *
+ * LOCK ORDERING (no-deadlock invariant): `chatLock` is acquired EXCLUSIVELY
+ * here, around the gate; never the reverse. `mayDrainCardNow` is a pure read.
+ */
+function cardDrainGate(turn: CurrentTurn, ea: EmissionAuthority, run: () => void): void {
+  if (EMISSION_AUTHORITY_ENABLED) {
+    void chatLock.run(
+      statusKey(turn.sessionChatId, turn.sessionThreadId),
+      async () => {
+        if (ea.mayDrainCardNow(turn, cardDrainGateCtx())) run()
+      },
+    )
+    return
+  }
+  run()
+}
 
 // Component 3 (turn-origin reply routing). Recently-ended turns retained
 // by `turnId` so a LATE reply (the Brevo answer landing ~42s after
@@ -2115,7 +2446,22 @@ function deriveTurnId(
  */
 function findTurnByOriginId(originTurnId: string | null | undefined): CurrentTurn | null {
   if (originTurnId == null || originTurnId === '') return null
-  if (currentTurn != null && currentTurn.turnId === originTurnId) return currentTurn
+  // PR-4e — resolve the LIVE turn by ITS OWN topic key under the flag. The
+  // turnId encodes the key: `deriveTurnId` builds `${chatKey}#${messageId}`, so
+  // the substring before `#` IS the statusKey. Flag-ON does an O(1)
+  // `byKey.get(key)` and matches on turnId — so a reply whose origin turn is
+  // STILL live for topic A resolves A even after the singleton mirror flipped to
+  // B. Flag-OFF keeps the singleton check, verbatim. The recentTurnsById
+  // registry fallback is UNCHANGED in both branches.
+  if (EMISSION_AUTHORITY_ENABLED) {
+    const hashIdx = originTurnId.indexOf('#')
+    if (hashIdx > 0) {
+      const live = currentTurnMap.get(originTurnId.slice(0, hashIdx))
+      if (live != null && live.turnId === originTurnId) return live
+    }
+  } else if (currentTurn != null && currentTurn.turnId === originTurnId) {
+    return currentTurn
+  }
   return recentTurnsById.get(originTurnId) ?? null
 }
 
@@ -2368,7 +2714,10 @@ function postQueuedStatus(chatId: string, bufferedThread: number, inFlightThread
   void (async () => {
     const sent = await swallowingApiCall(
       () =>
-        bot.api.sendMessage(chatId, text, { message_thread_id: bufferedThread }),
+        // Queued-placeholder status, not the user's answer — silence the
+        // open ping (BORDERLINE: it's a "your message is queued" notice;
+        // see PR description).
+        bot.api.sendMessage(chatId, text, { message_thread_id: bufferedThread, disable_notification: true }),
       { chat_id: chatId, verb: 'queued-status.post', threadId: bufferedThread },
     )
     const messageId = (sent as { message_id?: number } | undefined)?.message_id
@@ -2542,6 +2891,16 @@ const preambleSuppressor = new PreambleSuppressor({
     // long-lived and flushes can occur outside any session-event
     // handler's scope. If the turn has been cleared, the update is
     // dropped (no chat to send to, no stream to mutate).
+    //
+    // PR-4e — the module-scope suppressor carries NO per-topic key (it is a
+    // single global instance reset/flushed per turn), so there is no key to
+    // scope by here: the correct resolution under BOTH flag states is the
+    // most-recent-set live turn — exactly the singleton mirror `currentTurn`.
+    // The per-turn `reset()` / `flushNow()` / `dropNow()` lifecycle (driven from
+    // the live turn's own handlers) keeps the suppressor aligned with whichever
+    // topic is currently live, so the mirror read is byte-identical to base and
+    // cannot leak A's answer text into B (a flush for A runs while A is still the
+    // most-recent turn; once B flips, A's stream is already force-superseded).
     const stream = currentTurn?.answerStream ?? null
     if (stream != null) stream.update(cumulative)
   },
@@ -2590,6 +2949,28 @@ function streamKey(chatId: string, threadId?: number | null): string {
   return chatKey(chatId, threadId)
 }
 
+// PR1 (cross-turn stale-card guard, design §9 lever 4 / race C/D).
+// `obligationSweep` writes one entry here, keyed on the obligation's
+// `originTurnId`, the instant it pushes an `obligation_represent` inbound —
+// carrying the obligation's `openedAt` (when the obligation was RAISED). The
+// represent inbound reuses the obligation's chat/thread/messageId, so the
+// `enqueue` that spawns the synthetic represent turn reconstructs the SAME
+// `deriveTurnId` value as the key. That represent turn — and ONLY that turn —
+// consumes and clears it, stamping `turn.crossTurnGate = { sinceMs: openedAt }`.
+// That turn's first card-OPEN then consults `hasOutboundDeliveredSince(chat,
+// openedAt)` and is suppressed iff a SUBSTANTIVE answer already landed since the
+// obligation was raised — so a "thinking…" card never narrates beneath an answer
+// the user already received in the original turn. Keying on `originTurnId` (not
+// chat/thread) means an unrelated later foreground turn on the same chat/thread
+// derives a different turn id, finds no entry, and is never mis-gated — closing
+// the residual cross-contamination window where a represent that was armed but
+// never enqueued (degenerate bridge-death) left a stale chat/thread-keyed gate
+// that the next foreground turn could wrongly consume. A normal foreground turn
+// never has an entry here, so the gate stays scoped to the synthetic surface.
+// The map holds at most one entry per obligation; re-arming the same obligation
+// overwrites its own entry with the latest openedAt.
+const pendingCrossTurnGate = new Map<string, { sinceMs: number }>()
+
 /**
  * Component 1 — deliver-before-drain. The single chokepoint that both
  * turn-end drain sites (`purgeReactionTracking`, `releaseTurnBufferGate`)
@@ -2729,6 +3110,10 @@ function purgeReactionTracking(key: string, endingTurn?: CurrentTurn): void {
   // the markClaudeBusyForInbound on the delivery path. Safe no-op
   // when the key was never marked (synthetic purge from a sweep).
   claudeBusyKeys.delete(key)
+  // #2527: clear the per-key reaction-transition counter and first-reply
+  // sentinel alongside the controller so we don't leak state across turns.
+  reactionTransitionCounts.delete(key)
+  firstTextReplyLogged.delete(key)
   // Human-feel UX: stop the turn-long `typing…` indicator started in
   // the turn-start block. `purgeReactionTracking` is the canonical
   // turn-end, so this is the single owner of the stop. (If an abnormal
@@ -2933,8 +3318,16 @@ function releaseTurnBufferGate(key: string, endingTurn?: CurrentTurn): void {
  * gone — handlers that already purge elsewhere are unharmed.
  */
 function endCurrentTurnAtomic(turn: CurrentTurn): void {
-  if (currentTurn !== turn) return
-  currentTurn = null
+  // PR-4e — keyed liveness + keyed clear (leak-close-at-origin). Flag-OFF: the
+  // guard is `currentTurn === turn` and the clear nulls the singleton, verbatim.
+  // Flag-ON: the guard becomes `byKey.get(turn'sKey) === turn` (so a flip to
+  // another topic doesn't spuriously short-circuit THIS topic's teardown) and
+  // the clear does `byKey.delete(key)` + nulls the mirror iff it still points at
+  // `turn`. `endCurrentTurnForKey` returns false (no delete) when the entry no
+  // longer matches — the same early-return semantics as the old `!== turn` guard.
+  const key = statusKey(turn.sessionChatId, turn.sessionThreadId)
+  if (!turnLiveForItsTopic(turn)) return
+  endCurrentTurnForKey(turn, key) // currentTurnByKey.delete(key) + mirror clear
   // Status-surface observability: one line at every turn CLEAR (with how far
   // the turn got), plus a DEGRADED warning when the turn did tool work but the
   // live feed never opened because its sends failed (the resume-400 signature).
@@ -3199,6 +3592,7 @@ async function postIdleClearNotice(idleClearMs: number): Promise<void> {
       () =>
         bot.api.sendMessage(chatId, text, {
           parse_mode: 'HTML',
+          disable_notification: true,
           ...(threadId != null ? { message_thread_id: threadId } : {}),
         }),
       { chat_id: chatId, verb: 'idleAutoClear.notice' },
@@ -3340,7 +3734,7 @@ async function resolveCompactCard(
 function finalizeStatusReaction(
   chatId: string,
   threadId: number | undefined,
-  reason: 'done' | 'error' = 'done',
+  reason: 'done' | 'undelivered' | 'error' = 'done',
 ): void {
   const key = statusKey(chatId, threadId)
   const ctrl = activeStatusReactions.get(key)
@@ -3354,6 +3748,17 @@ function finalizeStatusReaction(
   if (reason === 'done' && deferredDoneReactions.tryDefer(key, ctrl)) return
   deferredDoneReactions.drop(key)
   ctrl.finalize(reason)
+  // #2527: log controller dispose so the lifecycle end is observable. Use
+  // activeReactionMsgIds to reconstruct the turnId token before purge clears it.
+  const msgInfo = activeReactionMsgIds.get(key)
+  if (msgInfo != null) {
+    logStreamingEvent({
+      kind: 'status_reaction_dispose',
+      chatId,
+      turnId: `${chatId}:${msgInfo.messageId}`,
+      reason,
+    })
+  }
   purgeReactionTracking(key)
 }
 
@@ -4960,10 +5365,10 @@ function postLegacyBanner(
 // short-circuit to no-ops at runtime. `progressDriver` is typed `any`
 // so TS doesn't resolve `progressDriver?.X` to `never`.
 const streamMode = process.env.SWITCHROOM_TG_STREAM_MODE ?? 'checklist'
-// PR B: per-agent stream throttle override via channels.telegram.stream_throttle_ms.
-// When unset, draft-stream.ts applies transport-aware defaults (300 ms draft,
-// 1000 ms message). Parsed once at boot; sub-zero / NaN values fall back to
-// undefined so the per-transport default wins. See `src/agents/scaffold.ts`
+// Per-agent stream throttle override via channels.telegram.stream_throttle_ms.
+// When unset, draft-stream.ts applies DM/group defaults (400 ms DMs, 1000 ms
+// groups). Parsed once at boot; sub-zero / NaN values fall back to undefined
+// so the per-chat-type default wins. See `src/agents/scaffold.ts`
 // `channelsToEnv()` for the yaml → env wiring.
 const STREAM_THROTTLE_MS_OVERRIDE: number | undefined = (() => {
   const raw = process.env.SWITCHROOM_TG_STREAM_THROTTLE_MS
@@ -4973,74 +5378,30 @@ const STREAM_THROTTLE_MS_OVERRIDE: number | undefined = (() => {
 })()
 const TURN_FLUSH_SAFETY_ENABLED = isTurnFlushSafetyEnabled()
 
-// #869-Phase1 / openclaw-pattern. When SET, the answer-lane stream
-// (telegram-plugin/answer-stream.ts) renders the model's transcript
-// text as a USER-VISIBLE edit-in-place message instead of writing to
-// Telegram's invisible compose-box draft (which is the default and
-// supports the #1664 "retract + re-prompt" contract). With this flag
-// on:
-//   1. createAnswerStream is instantiated without `sendMessageDraft`,
-//      so it falls back to `sendMessage` + `editMessageText` for a
-//      real chat-timeline message (`answer-stream.ts:212-214`).
-//   2. minInitialChars is set to 1 — the first text chunk pushes a
-//      visible message immediately (TTFO under 5s for short turns).
-//   3. At turn_end, if the model never called reply / stream_reply
-//      AND the streamed message has substantive captured text, the
-//      gateway DOES NOT retract (which would delete a user-visible
-//      message the user has been reading live); it calls
-//      `stream.stop()` to freeze the current text as the final
-//      answer, records the message in dedup + history, and marks
-//      `turn.finalAnswerDelivered = true` so the #1664 silent-end
-//      re-prompt does not fire. Turn-flush is suppressed for this
-//      branch — its job (deliver captured text) is structurally
-//      already done by the visible stream.
-//   4. The reply-tool / stream_reply path is unchanged — when the
-//      model uses an explicit reply tool the prior streamed message
-//      is retracted (delete) and the reply takes over as before.
-// Trade-off: a stream-as-final-answer turn does NOT push a device
-// notification (Telegram does not notify on edits, and we choose
-// not to send a duplicate fresh message for the ping). For short
-// turns where the user is actively watching, this is the right
-// shape — they see the answer materialise live. For longer waits,
-// the cross-turn pending-progress system (#1445/#1669) is the
-// canonical surface and DOES ping at the appropriate boundaries.
-//
-// 2026-05-25: default flipped ON after a fleet-log audit showed a ~19%
-// framework-fallback ("still working…") rate — the visible stream gave an
-// immediate in-timeline signal that suppressed the silence-poke.
-//
-// 2026-06-03: default flipped back OFF (operator request). In practice the
-// visible stream delivered ~none of its intended benefit while imposing a
-// jarring cost:
-//   - Telegram rate-limits editMessageText to roughly once/second, so real
-//     "watch it type" streaming is impossible; and the model emits almost no
-//     interstitial assistant.text (it thinks → tool → reply), so the
-//     preliminary was a near-empty bubble (observed: 5–13 byte edits).
-//   - On every turn where the model calls the reply tool (≈always), the reply
-//     posts a SEPARATE canonical message and the stream RETRACTS (deletes) its
-//     preliminary — the user sees a raw bubble appear then vanish, replaced by
-//     the formatted reply. In supergroup topics it also mis-routed (preliminary
-//     → General, reply → topic). Net: an unformatted flash + a delete, no
-//     streaming value.
-// The anti-silence role the visible stream once filled is now covered by the
-// live ACTIVITY FEED (tool-use streaming, below), the "…typing" chat-action
-// loop, and `thinking_effort: low` (fast tool-less turns) — so off-by-default
-// no longer regresses the framework-fallback rate. With the flag off the lane
-// uses the invisible compose-box draft (the original default, #1664-compatible)
-// and the reply tool is the single canonical, formatted message.
+// When SET, the answer-lane stream (telegram-plugin/answer-stream.ts) renders
+// the model's transcript text as a USER-VISIBLE edit-in-place message. Default
+// OFF: the lane stays dormant and the reply tool is the single canonical
+// formatted message — no unformatted preliminary that flashes and gets deleted.
+// With this flag on, minInitialChars is set to 1 and the first text chunk opens
+// a visible preview immediately. At turn_end, if the model never called reply /
+// stream_reply AND the streamed message has substantive captured text, the
+// gateway materializes it as a pinged final answer (materialize()) and deletes
+// the silent preview. When the model uses an explicit reply tool the prior
+// streamed message is retracted instead.
+// The draft transport (sendMessageDraft) is permanently retired — both ON and
+// OFF use sendMessage + editMessageText; the difference is whether a visible
+// preview is opened at all.
 // Opt back IN per agent with SWITCHROOM_VISIBLE_ANSWER_STREAM=1.
 const ANSWER_STREAM_VISIBLE_ENABLED = parseVisibleAnswerStreamEnabled(
   process.env.SWITCHROOM_VISIBLE_ANSWER_STREAM,
 )
-// Single source of truth for the answer-lane behaviour (flash-decouple,
-// 2026-06-05). The visible preview gates on the visible flag ALONE; the draft
-// flag controls only the transport. Resolved here once and consulted at the
-// createAnswerStream config, the materialize-as-answer guard, and the boot log,
-// so all three can never drift back into the `visible || retired` conflation
-// that re-opened the flash. Total-enumerated in answer-stream-flag.test.ts.
+// Single source of truth for the answer-lane behaviour. The draft transport
+// (sendMessageDraft) is permanently retired — the lane is either VISIBLE
+// (opt-in) or DORMANT (the unconditional default: reply tool is the only
+// message). Resolved here once and consulted at the createAnswerStream config,
+// the materialize-as-answer guard, and the boot log.
 const ANSWER_LANE = resolveAnswerLaneConfig({
   visibleEnabled: ANSWER_STREAM_VISIBLE_ENABLED,
-  draftFnAvailable: sendMessageDraftFn != null,
 })
 
 // Whether to DELETE the activity/status feed when the final answer lands.
@@ -5076,6 +5437,11 @@ const completeProgressCardTurn:
 // #1122 PR3: flushProgressCardsForShutdown deleted with the card. No
 // replacement needed — there are no pinned progress messages to flush.
 let subagentWatcher: SubagentWatcherHandle | null = null
+// Background-worker activity feed manager. Module-scoped so shutdown can stop()
+// its internal heartbeat interval (mirrors subagentWatcher). Recreated per
+// bridge connect; the stale handle's interval is unref'd, so a missed stop()
+// can't keep the process alive, but we stop() on shutdown for cleanliness.
+let workerActivityFeed: ReturnType<typeof createWorkerActivityFeed> | null = null
 
 // ─── IPC server ───────────────────────────────────────────────────────────
 const SOCKET_PATH = process.env.SWITCHROOM_GATEWAY_SOCKET ?? join(STATE_DIR, 'gateway.sock')
@@ -5237,8 +5603,12 @@ let inFlightUpdate: { requestId: string; startedAt: number } | null = null
 //   SWITCHROOM_SILENCE_FALLBACK_MS         — base threshold (default 300000)
 //   SWITCHROOM_SILENCE_FALLBACK_HARD_MS    — hard ceiling for the in-flight-tool
 //                                            defer (default 900000 = 15min)
-//   SWITCHROOM_SILENCE_DEFER_INFLIGHT_TOOLS=1 — enable the defer (default OFF;
-//                                            canary on marko against #2162 telemetry)
+//   SWITCHROOM_SILENCE_DEFER_INFLIGHT_TOOLS=1 — enable the defer. NOTE: this is
+//                                            now set fleet-wide in defaults.env
+//                                            (was a marko canary against #2162;
+//                                            promoted to the fleet default). The
+//                                            code default below is still OFF, so
+//                                            the live behaviour comes from config.
 function parsePositiveMsEnv(name: string, fallbackMs: number): number {
   const raw = process.env[name]
   if (raw == null || raw === '') return fallbackMs
@@ -5247,6 +5617,20 @@ function parsePositiveMsEnv(name: string, fallbackMs: number): number {
 }
 const SILENCE_FALLBACK_MS = parsePositiveMsEnv('SWITCHROOM_SILENCE_FALLBACK_MS', 300_000)
 const SILENCE_FALLBACK_HARD_MS = parsePositiveMsEnv('SWITCHROOM_SILENCE_FALLBACK_HARD_MS', 900_000)
+// #2527 — mid-turn liveness floor threshold (default 45s). The early, quiet
+// beat: a `user` turn working silently this long without a substantive answer
+// gets ONE honest "still on it" interim, so the ambient 👀 never masquerades
+// as "done". Strictly below SILENCE_FALLBACK_MS (the loud 300s unwedge).
+// Whole floor is kill-switchable via SWITCHROOM_TG_LIVENESS_FLOOR=0.
+const SILENCE_FLOOR_MS = parsePositiveMsEnv('SWITCHROOM_SILENCE_FLOOR_MS', 45_000)
+// #2527 — role-aware terminal reaction honesty (the "thumbs-up false done"
+// fix). Default ON; SWITCHROOM_TG_TERMINAL_HONESTY=0 reverts to always-👍.
+const LIVENESS_TERMINAL_HONESTY = process.env.SWITCHROOM_TG_TERMINAL_HONESTY !== '0'
+// SILENCE_DEFER_INFLIGHT_TOOLS: previously an opt-in (=1). The new
+// isLegitimatelyWorking callback supersedes this — defer is now the DEFAULT
+// when the callback is wired. The legacy flag is kept so `=0` still lets
+// operators force-disable the defer (handled inside silence-poke.ts tick()).
+// The old `=1` path is kept for back-compat but is now redundant.
 const SILENCE_DEFER_INFLIGHT_TOOLS = process.env.SWITCHROOM_SILENCE_DEFER_INFLIGHT_TOOLS === '1'
 // Production-liveness (2026-06-05 UAT finding). Count an activity-feed render or
 // an answer-stream draft update as liveness for the silence clock, so a long
@@ -5255,13 +5639,109 @@ const SILENCE_DEFER_INFLIGHT_TOOLS = process.env.SWITCHROOM_SILENCE_DEFER_INFLIG
 // restores the legacy "only a real reply resets the clock" behaviour.
 const SILENCE_LIVENESS_PRODUCTION = process.env.SWITCHROOM_SILENCE_LIVENESS_PRODUCTION !== '0'
 
+/**
+ * Feed-survival predicate — the single source of truth for "is this turn
+ * legitimately working?" used by BOTH teardown timers (orphaned-reply fuse
+ * and silence-poke framework fallback).
+ *
+ * Returns true if ANY of the following hold for the given chat key:
+ *
+ *   (a) A foreground tool call is in flight in the current turn
+ *       (`toolFlightTracker.isMidToolCall()`). This covers most tools
+ *       including ask_user while it blocks awaiting a tap.
+ *
+ *   (b) Detached background work was dispatched in the current turn and has
+ *       not yet resolved — `pendingProgress.hasPendingAsyncDispatch(key)`.
+ *       Covers `Bash run_in_background:true` (which returns a near-instant
+ *       handle, emptying inFlight, while the background process keeps
+ *       running) and `Agent` / `Task` dispatches.
+ *
+ *   (c) A human-wait tool (`ask_user`) is open for this chat. A pending
+ *       ask_user IS already captured by (a) while the tool_use is in flight,
+ *       but we include the explicit pendingAskUser check for defence-in-depth
+ *       (e.g. after an unlikely inFlight clear without a tool_result).
+ *
+ * The key is `statusKey(chatId, threadId)` — the same key used by
+ * silencePoke / pendingProgress.
+ */
+function isLegitimatelyWorking(key: string): boolean {
+  // (a) foreground in-flight tool.
+  // NOTE: toolFlightTracker is GLOBAL, not per-key. In a hypothetical
+  // multi-chat agent a tool in flight for chat A would make this return
+  // true for chat B's key. Accepted: the gateway runs one Claude session
+  // (one turn in flight at a time); true multi-chat concurrency is not
+  // currently supported. (b) and (c) below are correctly per-key.
+  if (toolFlightTracker.isMidToolCall()) return true
+  // (b) detached background work dispatched this turn
+  if (pendingProgress.hasPendingAsyncDispatch(key)) return true
+  // (c) ask_user open for this chat (defence-in-depth)
+  const { chatId: keyChatId } = parseKeyForSurvival(key)
+  for (const entry of pendingAskUser.values()) {
+    if (entry.chatId === keyChatId) return true
+  }
+  return false
+}
+
+/** Parse `<chatId>:<threadIdOrEmpty>` — mirrors silence-poke's parseKey.
+ *  Local copy so we don't need to re-export from silence-poke. */
+function parseKeyForSurvival(key: string): { chatId: string } {
+  const idx = key.indexOf(':')
+  return { chatId: idx < 0 ? key : key.slice(0, idx) }
+}
+
 silencePoke.startTimer({
-  thresholdsMs: { fallback: SILENCE_FALLBACK_MS, fallbackHardCeiling: SILENCE_FALLBACK_HARD_MS },
+  thresholdsMs: { fallback: SILENCE_FALLBACK_MS, fallbackHardCeiling: SILENCE_FALLBACK_HARD_MS, floor: SILENCE_FLOOR_MS },
   deferFallbackWhileToolInFlight: SILENCE_DEFER_INFLIGHT_TOOLS,
+  isLegitimatelyWorking: (key) => isLegitimatelyWorking(key),
   emitMetric: (event) => {
     // Re-emit through the unified runtime-metrics fan-out (PostHog + JSONL).
     emitRuntimeMetric(event)
   },
+  // #2527 — the gateway-owned half of the mid-turn-floor decision: only the
+  // live turn knows its loop role + whether a substantive answer has landed.
+  // Keyed on statusKey so a DM (threadId null) and a forum topic are identical.
+  floorState: (key) => {
+    const turn = currentTurn
+    if (turn == null) return null
+    if (statusKey(turn.sessionChatId, turn.sessionThreadId) !== key) return null
+    return { role: turn.role, finalAnswerDelivered: turn.finalAnswerDelivered }
+  },
+  // #2527 — the early, quiet liveness beat. Honest text from the longest
+  // in-flight tool (model-free, claude-native), routed through the SAME send
+  // path as the 300s fallback; pings OFF (this is the gentle beat, not the
+  // loud unwedge) and the turn is NOT torn down — it keeps working.
+  onMidTurnFloor: async (ctx) => {
+    // Late-fire guard, mirroring the fallback: a clean turn-end can race the
+    // tick. If the turn is gone, stay silent.
+    if (activeTurnStartedAt.get(ctx.key) == null && currentTurn == null) return
+    const blockedOnApproval = activeStatusReactions
+      .get(statusKey(ctx.chatId, ctx.threadId))
+      ?.isAwaiting() ?? false
+    const text = silencePoke.formatFrameworkFallbackText(
+      'working',
+      ctx.silenceMs,
+      ctx.inFlightTools,
+      blockedOnApproval,
+    )
+    try {
+      await robustApiCall(
+        () => bot.api.sendMessage(ctx.chatId, text, {
+          ...(ctx.threadId != null ? { message_thread_id: ctx.threadId } : {}),
+          // The quiet beat: visible in-thread, no device buzz. (The 300s
+          // fallback pings; the floor must not train the user to mute.)
+          disable_notification: true,
+        }),
+        { chat_id: ctx.chatId, ...(ctx.threadId != null ? { threadId: ctx.threadId } : {}) },
+      )
+      // Count it as production so the silence clock resets — the user just
+      // saw a real message, so the 300s loud fallback is measured from here.
+      silencePoke.noteProduction(ctx.key, Date.now())
+    } catch (err) {
+      process.stderr.write(
+        `silence-poke mid-turn floor sendMessage failed chat=${ctx.chatId} thread=${ctx.threadId}: ${err}\n`,
+      )
+    }
+  },
   onFrameworkFallback: async (ctx) => {
     // Late-fire short-circuit (2026-05-23 audit finding). The fallback
     // can race a clean turn-end: the model's actual reply lands inside
@@ -5281,6 +5761,14 @@ silencePoke.startTimer({
         `turn ended cleanly during silence window ` +
         `chat=${ctx.chatId} thread=${ctx.threadId ?? '-'} silence_ms=${ctx.silenceMs}\n`,
       )
+      // #2527: structured skip event so the late-fire race is machine-readable.
+      logStreamingEvent({
+        kind: 'silence_poke_skip',
+        chatId: ctx.chatId,
+        threadId: ctx.threadId ?? undefined,
+        silenceMs: ctx.silenceMs,
+        skipReason: 'turn_ended_cleanly_during_window',
+      })
       // Tell silence-poke this chat-thread is finished so the next
       // arming doesn't carry stale state.
       silencePoke.endTurn(ctx.key)
@@ -5294,6 +5782,15 @@ silencePoke.startTimer({
     // get_status snapshot → pure formatter. Any hostd unavailability
     // degrades silently to the existing generic text (zero regression).
     let text: string | null = null
+    // Hoisted out of the generic-fallback branch below because the send site
+    // gates `disable_notification` on it: when the turn is parked on an
+    // approval card, the fallback TEXT is a user-gating re-ping ("waiting for
+    // your approval — tap Approve or Deny …"), and that must stay LOUD so the
+    // user knows the ball is in their court. The reaction controller tracks the
+    // park via setAwaiting on the permission-request.
+    const blockedOnApproval = activeStatusReactions
+      .get(statusKey(ctx.chatId, ctx.threadId))
+      ?.isAwaiting() ?? false
     const upd = inFlightUpdate
     if (upd != null) {
       try {
@@ -5315,9 +5812,6 @@ silencePoke.startTimer({
       // benign "wedge" class — claude is alive, waiting on the operator's
       // tap), say so instead of "still working…". The reaction controller
       // already tracks this (setAwaiting on the permission-request park).
-      const blockedOnApproval = activeStatusReactions
-        .get(statusKey(ctx.chatId, ctx.threadId))
-        ?.isAwaiting() ?? false
       text = silencePoke.formatFrameworkFallbackText(
         ctx.fallbackKind,
         ctx.silenceMs,
@@ -5325,12 +5819,26 @@ silencePoke.startTimer({
         blockedOnApproval,
       )
     }
+    // #2527: log the actual poke fire with structured data before sending,
+    // so the event is visible even if the send fails.
+    logStreamingEvent({
+      kind: 'silence_poke_fire',
+      chatId: ctx.chatId,
+      threadId: ctx.threadId ?? undefined,
+      silenceMs: ctx.silenceMs,
+      fallbackKind: ctx.fallbackKind,
+    })
     try {
       await robustApiCall(
         () => bot.api.sendMessage(ctx.chatId, text, {
           ...(ctx.threadId != null ? { message_thread_id: ctx.threadId } : {}),
-          // Framework fallback pings — user genuinely needs to know.
-          disable_notification: false,
+          // Conditional: the pure-liveness "still working…" notice is a status
+          // surface and stays SILENT. But when the turn is parked on an
+          // approval card, this same fallback carries a user-gating re-ping
+          // ("waiting for your approval — tap Approve or Deny …") — that must
+          // PING, because the user is the one being waited on. Gate on the same
+          // `blockedOnApproval` signal that selects the re-ping text above.
+          disable_notification: blockedOnApproval ? false : true,
         }),
         { chat_id: ctx.chatId, ...(ctx.threadId != null ? { threadId: ctx.threadId } : {}) },
       )
@@ -5464,13 +5972,31 @@ silencePoke.startTimer({
         return sib == null || sib >= silencePoke.DEFAULT_THRESHOLDS.fallback
       },
     )
+    // PR-4e self-heal backstop: drop any per-topic `currentTurnByKey` entries
+    // for fbChatId whose turn is stale-by-the-same-silence-gate the sibling
+    // sweep above used — the same precedent as `purgeStaleTurnsForChat`'s
+    // activeTurnStartedAt sweep, so a leaked map entry (a turn whose keyed
+    // delete somehow never ran) can never outlive its chat. Gated identically:
+    // the firing key is always stale; a sibling is stale iff it's silent ≥ the
+    // fallback threshold (or has no silence state). No-op under the flag OFF
+    // (the map is empty).
+    currentTurnMap.purgeChatStale(fbChatId, (siblingKey) => {
+      if (siblingKey === fbKey) return true
+      const sib = silencePoke.silenceMsForKey(siblingKey, fbNow)
+      return sib == null || sib >= silencePoke.DEFAULT_THRESHOLDS.fallback
+    })
     // Null `currentTurn` if it's still pointing at the wedged turn —
     // when claude eventually fires a late `turn_end` for this session
     // (or never does), the handler's `const turn = currentTurn` snapshot
     // returns null and the regular teardown short-circuits. Without
     // this, the late event would re-emit `turn_ended` AND clobber
     // whatever fresh turn the next inbound started.
-    if (turnMatchesFallback && currentTurn === wedgedTurn && wedgedTurn != null) {
+    // PR-4e — keyed liveness for the guard. Flag-OFF: `turnLiveForItsTopic`
+    // reduces to `currentTurn === wedgedTurn` (singleton mirror), verbatim.
+    // Flag-ON: `byKey.get(fbKey) === wedgedTurn`, so the keyed delete still
+    // fires when the LIVE mirror has already flipped to another topic B (a bare
+    // `currentTurn === wedgedTurn` would falsely skip and leak A's byKey entry).
+    if (turnMatchesFallback && wedgedTurn != null && turnLiveForItsTopic(wedgedTurn)) {
       // Status-surface observability: emit the lifecycle CLEAR for the
       // silence-poke teardown so a fallback-nulled turn has a turn-lifecycle
       // line like every other clear path (the framework-fallback line below is
@@ -5478,7 +6004,12 @@ silencePoke.startTimer({
       process.stderr.write(
         `telegram gateway: ${formatTurnLifecycle('clear', 'silence_fallback', wedgedTurn, Date.now())}\n`,
       )
-      currentTurn = null
+      // PR-4e — keyed delete for the wedged turn's OWN key (fbKey == the
+      // statusKey this fallback fired for, == the wedgedTurn's key since
+      // turnMatchesFallback gated chat+thread equality). Flag-OFF nulls the
+      // singleton, verbatim; flag-ON deletes only this topic's entry and clears
+      // the mirror iff it still points here — a live sibling topic is untouched.
+      endCurrentTurnForKey(wedgedTurn, fbKey)
     }
     // Best-effort: clear any pending silent-end marker so the Stop hook
     // doesn't double-block when claude eventually exits the wedged turn.
@@ -5789,6 +6320,43 @@ function obligationSweep(): void {
     return
   }
   if (decision.action === 'represent') {
+    // Fix #2472 — duplicate-represent guard. Before re-presenting AGAIN, check
+    // whether the agent has ALREADY delivered a substantive outbound reply to
+    // this chat SINCE the obligation was most recently re-presented. If so the
+    // obligation is satisfied-but-misdetected (the reply landed but its routing
+    // didn't resolve back to this origin, so the normal close path missed it) —
+    // close silently and do NOT re-fire, which is what produced the near-identical
+    // duplicate in #2472 (reply 10608 answered represent_count=1, yet
+    // represent_count=2 fired anyway → duplicate 10609).
+    //
+    // The cutoff is `lastRepresentedAt` (the time of the PREVIOUS represent), NOT
+    // `openedAt`. This is load-bearing: the genuine "agent wrote a plain-text
+    // answer and never called reply" case must still represent ONCE. On the first
+    // represent `lastRepresentedAt` is undefined, so this guard is a no-op and the
+    // single represent fires as before. Only the SECOND-and-later represent is
+    // gated — exactly where a reply that landed between fires must suppress the
+    // re-ask. Falls back to false (never suppresses) if history is unavailable.
+    if (
+      shouldSuppressRepresent(o, {
+        historyEnabled: HISTORY_ENABLED,
+        // Pass the represent-guard's OWN low threshold — a terse-but-real reply
+        // must suppress the duplicate (#2472/#2474), unlike the escalate branch
+        // below which keeps the 200-char default.
+        hasOutboundDeliveredSince: (chatId, sinceMs, threadId) =>
+          hasOutboundDeliveredSince(
+            chatId,
+            sinceMs,
+            threadId,
+            OBLIGATION_REPRESENT_GUARD_MIN_REPLY_CHARS,
+          ),
+      })
+    ) {
+      process.stderr.write(
+        `telegram gateway: obligation closed silently — reply delivered since last represent (no re-fire) origin=${o.originTurnId}\n`,
+      )
+      obligationLedger.close(o.originTurnId)
+      return
+    }
     // Re-present goes through the bridge → buffer. Only the represent path is
     // gated on an empty buffer (let the existing drain run first, avoid
     // double-presenting). Escalation below is NOT gated on the buffer — it is a
@@ -5796,6 +6364,25 @@ function obligationSweep(): void {
     // behind a dead bridge can never block the operator nudge.
     if (pendingInboundBuffer.depth(agent) > 0) return
     pendingInboundBuffer.push(agent, buildObligationRepresentInbound(o, Date.now()))
+    // PR1 (cross-turn stale-card guard, §9 lever 4 / race C/D). Arm the
+    // card-OPEN gate for the synthetic turn this represent inbound will spawn:
+    // carry the obligation's `openedAt` so that turn's first card-OPEN can ask
+    // "was a substantive answer already delivered since the obligation was
+    // raised?" and, if so, suppress the card (it would otherwise narrate beneath
+    // the answer the user already received). Keyed on the obligation's
+    // `originTurnId` — the SAME id the represent inbound carries
+    // (`buildObligationRepresentInbound` reuses `o.messageId`/`o.chatId`/
+    // `o.threadId`, so the enqueue-time `deriveTurnId` reconstructs exactly
+    // `o.originTurnId`). Keying on the turn id (not chat/thread) means ONLY the
+    // exact represent turn this gate was armed for can consume it; an unrelated
+    // later foreground turn on the same chat/thread has a different originTurnId
+    // → finds no entry → its card opens normally. This closes the residual
+    // cross-contamination window where a never-enqueued represent's stale gate
+    // could suppress an unrelated turn's card (the represent/duplicate-reply
+    // family). This does NOT gate the represent SEND — the represent guard above
+    // already owns suppressing an already-satisfied represent; this only governs
+    // the decorative card.
+    pendingCrossTurnGate.set(o.originTurnId, { sinceMs: o.openedAt })
     const attempt = obligationLedger.markRepresented(o.originTurnId)
     process.stderr.write(
       `telegram gateway: obligation re-presented origin=${o.originTurnId} attempt=${attempt}/${OBLIGATION_REPRESENT_MAX}\n`,
@@ -6196,7 +6783,10 @@ const ipcServer: IpcServer = createIpcServer({
           process.stderr.write(
             `telegram gateway: disconnect-flush nulled currentTurn (bridge died with turn in flight)\n`,
           )
-          currentTurn = null
+          // PR-4e — the bridge DIED with a turn in flight: EVERY per-topic entry
+          // is a ghost, not just the mirror's. Clear the whole map + mirror.
+          // Flag-OFF: this nulls the singleton only (the map is empty), verbatim.
+          clearAllCurrentTurns()
         }
       },
       log: (msg) => process.stderr.write(`${msg}\n`),
@@ -6722,10 +7312,10 @@ const ipcServer: IpcServer = createIpcServer({
           ...(cfgTopic != null ? { threadId: cfgTopic } : {}),
         }
       },
-      buildKeyboard: (requestId) =>
+      buildKeyboard: (requestId, epoch) =>
         new InlineKeyboard()
-          .text('✅ Approve', `cfg:${requestId}:approve`)
-          .text('🚫 Deny', `cfg:${requestId}:deny`),
+          .text('✅ Approve', `cfg:${requestId}:${epoch}:approve`)
+          .text('🚫 Deny', `cfg:${requestId}:${epoch}:deny`),
       postCard: async (args) => {
         try {
           const sent = await robustApiCall(
@@ -6757,6 +7347,9 @@ const ipcServer: IpcServer = createIpcServer({
             () =>
               bot.api.editMessageText(args.chatId, args.messageId, args.text, {
                 parse_mode: 'HTML',
+                // Strip the inline keyboard on a terminal/interim edit so the
+                // [Approve]/[Deny] buttons stop being tappable on a resolved card.
+                ...(args.stripKeyboard ? { reply_markup: { inline_keyboard: [] } } : {}),
               }),
             { chat_id: String(args.chatId), verb: 'config-approval-edit' },
           )
@@ -6826,6 +7419,8 @@ const ipcServer: IpcServer = createIpcServer({
             () =>
               bot.api.editMessageText(args.chatId, args.messageId, args.text, {
                 parse_mode: 'HTML',
+                // Finalize is terminal — drop the keyboard so buttons are gone.
+                ...(args.stripKeyboard ? { reply_markup: { inline_keyboard: [] } } : {}),
               }),
             { chat_id: String(args.chatId), verb: 'config-approval-finalize' },
           )
@@ -7377,6 +7972,23 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
     }
   }
   process.stderr.write(`telegram channel: reply: invoked chatId=${chat_id} charCount=${text.length} preview=${JSON.stringify(text.slice(0, 80))}\n`)
+  // #2527: emit time_to_first_text_reply_ms on the FIRST text reply of each
+  // turn so operators can see how long users waited for any visible output.
+  // Only fires once per turn (firstTextReplyLogged guards the repeat).
+  if (turn != null) {
+    const threadId = args.message_thread_id != null ? Number(args.message_thread_id) : undefined
+    const replyKey = statusKey(chat_id, threadId)
+    if (!firstTextReplyLogged.has(replyKey)) {
+      firstTextReplyLogged.add(replyKey)
+      logStreamingEvent({
+        kind: 'turn_reply_timing',
+        chatId: chat_id,
+        threadId,
+        turnId: turn.turnId,
+        timeToFirstTextReplyMs: Date.now() - turn.gatewayReceiveAt,
+      })
+    }
+  }
 
   // #546 dedup check: was this content just sent via turn-flush or
   // a sibling reply path? Skip the actual send and return a
@@ -7411,6 +8023,15 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
   // existing call-sites and the typical "final answer" reply keep their
   // current behaviour without an explicit flag.
   let disableNotification = args.disable_notification === true
+  // #2527/#1664 — the over-ping safety net below may downgrade
+  // `disableNotification` ping→silent for ANTI-SPAM (one ping per turn). That
+  // delivery-channel decision must NOT pollute final-answer CLASSIFICATION: a
+  // final answer the model intended to ping is STILL the final answer even when
+  // the framework silences the actual ping. Classify on the model's original
+  // intent (what executeStreamReply already does), so an over-ping-silenced
+  // final answer sets finalAnswerDelivered=true — fixing both a spurious
+  // silent-end re-prompt and a false 'undelivered' (😐) terminal reaction.
+  const modelDisableNotification = args.disable_notification === true
 
   // #1675 over-ping safety net. The conversational-pacing contract
   // (`reference/rfcs/conversational-pacing.md` beat 5) says EXACTLY ONE
@@ -7441,32 +8062,92 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
     const turn = currentTurn
     if (turn != null) {
       const now = Date.now()
-      const decision = decideOverPing({
-        modelRequestedPing: !disableNotification,
-        firstPingAt: turn.firstPingAt,
-        nowMs: now,
+      // Notification ownership (R8 / PR-2): on the `reply` path,
+      // substantiveness is purely the ≥200-char (or `done`) backstop —
+      // `isSubstantiveFinalReply` is `done === true || text.length >= 200`
+      // and ignores the notification flag entirely. `reply` carries no
+      // `done`, so it reduces to the ≥200-char length test. We still pass
+      // `modelDisableNotification` (the MODEL's original intent, not the
+      // possibly-downgraded `disableNotification`) to mirror the #2533
+      // final-answer decoupling call shape, but that arg does NOT
+      // participate in classification here — it is inert on this path.
+      const replySubstantive = isSubstantiveFinalReply({
+        text: rawText,
+        disableNotification: modelDisableNotification,
       })
-      if (decision.suppress) {
-        process.stderr.write(
-          `telegram gateway: reply over-ping safety net — ` +
-          `downgrading disable_notification:false → true ` +
-          `(chat=${chat_id} thread=${args.message_thread_id ?? '-'} ` +
-          `firstPingAt=${turn.firstPingAt} sinceFirstPing_ms=${decision.sinceFirstPingMs})\n`,
-        )
-        // Observability: surface to the unified runtime-metrics
-        // fan-out so the cadence dashboard can track fleet-wide
-        // over-ping rate (leading indicator of model pacing drift).
-        emitRuntimeMetric({
-          kind: 'over_ping_suppressed',
-          key: statusKey(chat_id, args.message_thread_id != null
-            ? Number(args.message_thread_id) : undefined),
-          sinceFirstPingMs: decision.sinceFirstPingMs ?? 0,
-        })
-        disableNotification = true
-        wasOverPingSuppressed = true
-      } else if (decision.claimSlot) {
-        turn.firstPingAt = now
+      // PR-4c: the over-ping DECISION relocates into the emission-authority
+      // façade, behind the kill-switch (default OFF), the same structural way
+      // PR-4b moved the OPEN gate. `decideOverPing` is already pure, so PR-4c
+      // extracts NOTHING new — it relocates the *call* into the façade's enabled
+      // branch and keeps the *effects* (stderr, metric, the atomic
+      // `firstPingAt`/`firstPingWasSubstantive` pair-set, the
+      // `disableNotification`/`wasOverPingSuppressed` outer-scope writes) HERE,
+      // parameterized by the decision the façade hands back via `applyDecision`.
+      //
+      //  - Disabled branch runs `disabledOverPing()` — its own LITERAL
+      //    `decideOverPing(...)` call + the full effects block, VERBATIM from
+      //    PR-4b-base (the disabled-path-is-byte-identical proof).
+      //  - Enabled branch: the façade computes the decision and hands it to
+      //    `applyOverPingDecision(decision)`, which performs the IDENTICAL
+      //    effects. Same pure inputs ⇒ same decision ⇒ flag-ON ≡ flag-OFF ≡ base.
+      //
+      // The effects block is shared between both thunks by closing over `decision`
+      // — but the disabled thunk computes it via its OWN literal `decideOverPing(`
+      // first, so the disabled path never depends on the façade for the decision.
+      const applyOverPingDecision = (decision: OverPingDecision): void => {
+        if (decision.suppress) {
+          process.stderr.write(
+            `telegram gateway: reply over-ping safety net — ` +
+            `downgrading disable_notification:false → true ` +
+            `(chat=${chat_id} thread=${args.message_thread_id ?? '-'} ` +
+            `firstPingAt=${turn.firstPingAt} sinceFirstPing_ms=${decision.sinceFirstPingMs})\n`,
+          )
+          // Observability: surface to the unified runtime-metrics
+          // fan-out so the cadence dashboard can track fleet-wide
+          // over-ping rate (leading indicator of model pacing drift).
+          emitRuntimeMetric({
+            kind: 'over_ping_suppressed',
+            key: statusKey(chat_id, args.message_thread_id != null
+              ? Number(args.message_thread_id) : undefined),
+            sinceFirstPingMs: decision.sinceFirstPingMs ?? 0,
+          })
+          disableNotification = true
+          wasOverPingSuppressed = true
+        } else if (decision.claimSlot) {
+          // Claim (first ping) OR upgrade (substantive answer pinging over an
+          // ack's slot). Set firstPingAt AND firstPingWasSubstantive ATOMICALLY
+          // (no await between) so a racing second reply reads a consistent pair.
+          turn.firstPingAt = now
+          turn.firstPingWasSubstantive = replySubstantive
+          if (decision.upgrade) {
+            process.stderr.write(
+              `telegram gateway: reply over-ping safety net — ` +
+              `UPGRADE: substantive answer pings over an ack's slot ` +
+              `(chat=${chat_id} thread=${args.message_thread_id ?? '-'})\n`,
+            )
+          }
+        }
       }
+      emissionAuthorityFor(turn).claimOrDowngradePing(
+        { modelRequestedPing: !disableNotification, substantive: replySubstantive },
+        {
+          firstPingAt: turn.firstPingAt,
+          firstPingWasSubstantive: turn.firstPingWasSubstantive,
+          nowMs: now,
+        },
+        applyOverPingDecision,
+        () => {
+          // Disabled-path: literal `decideOverPing(` + effects, VERBATIM base.
+          const decision = decideOverPing({
+            modelRequestedPing: !disableNotification,
+            firstPingAt: turn.firstPingAt,
+            substantive: replySubstantive,
+            firstPingWasSubstantive: turn.firstPingWasSubstantive,
+            nowMs: now,
+          })
+          applyOverPingDecision(decision)
+        },
+      )
     }
   }
 
@@ -7653,10 +8334,38 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
   // clear; the main turn-end path also re-writes the state when
   // finalAnswerDelivered=false, so this is a belt-and-braces gate
   // for the turn_end-missing case (#1741).
-  if (isFinalAnswerReply({ text: rawText, disableNotification })) {
+  if (isFinalAnswerReply({ text: rawText, disableNotification: modelDisableNotification })) {
     clearSilentEndState(statusKey(chat_id, threadId))
   }
 
+  // Lever 2 (design §9 lever 2): finalize the activity card BEFORE the reply
+  // chunks send, so the card keeps its (lower) message_id and the reply is
+  // structurally last on screen. ONLY for a *substantive* final — for an ack
+  // (non-substantive) do NOTHING: finalizing an ack early would
+  // close → reopen → emit MORE messages (the #2141 ack-then-work feed, R3).
+  // `clearActivitySummary` edits the existing card in place (no new send) and
+  // nulls `activityMessageId`; combined with the sticky latch set here it
+  // prevents any post-reply re-OPEN below the answer. Idempotent with the
+  // tool_use-event clear at the first-reply handoff (the existing backstop).
+  {
+    const finalizeTurn = currentTurn
+    if (
+      finalizeTurn != null
+      && isSubstantiveFinalReply({ text: rawText, disableNotification: modelDisableNotification })
+    ) {
+      // PR-4a: routed through the emission-authority façade (no-op delegates —
+      // the latch-set and the finalize run exactly as before).
+      const ea = emissionAuthorityFor(finalizeTurn)
+      ea.markSubstantiveFinalDelivered(() => {
+        finalizeTurn.finalAnswerEverDelivered = true
+        finalizeTurn.finalAnswerDeliveredAt = Date.now()
+      })
+      ea.finalizeCard(() => {
+        clearActivitySummary(finalizeTurn)
+      })
+    }
+  }
+
   if (previewMessageId != null && reply_to != null && replyMode !== 'off') {
     await deleteStalePreview(previewMessageId)
     previewMessageId = null
@@ -7764,7 +8473,7 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
             turn != null
             && isFinalAnswerReply({
               text: decision.mergedText,
-              disableNotification,
+              disableNotification: modelDisableNotification,
             })
           ) {
             turn.finalAnswerDelivered = true
@@ -7772,8 +8481,12 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
             // answer must NOT re-open the feed on post-answer housekeeping.
             turn.finalAnswerSubstantive = isSubstantiveFinalReply({
               text: decision.mergedText,
-              disableNotification,
+              disableNotification: modelDisableNotification,
             })
+            // Sticky ordering latch (lever 1): a substantive final closes the
+            // card OPEN gate for the rest of the turn. NEVER cleared by reopen.
+            if (turn.finalAnswerSubstantive) turn.finalAnswerEverDelivered = true
+            if (turn.finalAnswerSubstantive && turn.finalAnswerDeliveredAt == null) turn.finalAnswerDeliveredAt = Date.now()
             if (turn.finalAnswerSubstantive) closeObligationOnSubstantiveReply(args, turn, replyRoutedOriginTurn)
           }
           outboundDedup.record(
@@ -8112,12 +8825,17 @@ async function executeReply(args: Record<string, unknown>): Promise<{ content: A
     //
     // #1664 — `turn.finalAnswerDelivered = true` keeps the silent-
     // end re-prompt from spuriously firing on a delivered final.
-    if (turn != null && isFinalAnswerReply({ text: rawText, disableNotification })) {
+    if (turn != null && isFinalAnswerReply({ text: rawText, disableNotification: modelDisableNotification })) {
       turn.finalAnswerDelivered = true
       // Feed-reopen refinement: track whether this final was substantive
       // (≥200 chars or stream-done — not a short pinging ack) so post-answer
       // housekeeping tool work does NOT re-open the feed / trip silent-end.
-      turn.finalAnswerSubstantive = isSubstantiveFinalReply({ text: rawText, disableNotification })
+      turn.finalAnswerSubstantive = isSubstantiveFinalReply({ text: rawText, disableNotification: modelDisableNotification })
+      // Sticky ordering latch (lever 1): set once a SUBSTANTIVE final lands;
+      // never cleared by reopen. The card OPEN gate keys on this, not the
+      // mutable finalAnswerDelivered above (which reopen toggles).
+      if (turn.finalAnswerSubstantive) turn.finalAnswerEverDelivered = true
+      if (turn.finalAnswerSubstantive && turn.finalAnswerDeliveredAt == null) turn.finalAnswerDeliveredAt = Date.now()
       // #1728: release the buffer gate + emit terminal 👍. Mid-turn
       // acks bypass this branch and remain non-events for the
       // reaction (preserves #1713). The full turn-state teardown
@@ -8278,9 +8996,8 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
   }
 
   const access = loadAccess()
-  // Detect chat type for draft-transport selection.
+  // Detect chat type for throttle-default selection.
   // Private (DM) chats have positive numeric IDs; groups/channels are negative.
-  // Forum topics have a message_thread_id set — sendMessageDraft is unsupported there.
   const streamChatId = args.chat_id as string
   const streamIsPrivate = isDmChatId(streamChatId)
   const streamIsForumTopic = args.message_thread_id != null && args.message_thread_id !== ''
@@ -8322,6 +9039,19 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
       // PR3b-cutover: feed lastOutboundAt to the delivery machine (see
       // executeReply) so its TTL tick suppresses an active-turn fallback.
       shadowEmit({ kind: 'modelOutbound', key: sKey as _ChatKey, at: Date.now() })
+      // #2527: emit turn_reply_timing on the first stream_reply of the turn,
+      // mirroring the same gate in executeReply. Guards with firstTextReplyLogged
+      // so a turn that calls reply first and stream_reply second doesn't double-emit.
+      if (turn != null && !firstTextReplyLogged.has(sKey)) {
+        firstTextReplyLogged.add(sKey)
+        logStreamingEvent({
+          kind: 'turn_reply_timing',
+          chatId: streamChatId,
+          threadId: streamThreadId,
+          turnId: turn.turnId,
+          timeToFirstTextReplyMs: Date.now() - turn.gatewayReceiveAt,
+        })
+      }
       // #1741 — see executeReply for the rationale: only a plausibly-
       // final stream_reply clears the silent-end state. An interim
       // ack via stream_reply must NOT clear; the Stop hook needs
@@ -8338,6 +9068,33 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
     }
   }
 
+  // Lever 2 (design §9 lever 2): finalize the activity card BEFORE the stream
+  // send so the card keeps its lower message_id and the reply is structurally
+  // last. ONLY for a *substantive* final (a stream_reply done=true or ≥200
+  // chars) — for a short pinging interim chunk do NOTHING (finalizing an ack
+  // early would close → reopen → emit more, the #2141 ack-then-work feed, R3).
+  // `clearActivitySummary` edits in place + nulls activityMessageId; the sticky
+  // latch set here blocks any post-reply re-OPEN below the answer.
+  if (
+    turn != null
+    && isSubstantiveFinalReply({
+      text: (args.text as string | undefined) ?? '',
+      disableNotification: args.disable_notification === true,
+      done: args.done === true,
+    })
+  ) {
+    // PR-4a: routed through the emission-authority façade (no-op delegates —
+    // the latch-set and the finalize run exactly as before).
+    const ea = emissionAuthorityFor(turn)
+    ea.markSubstantiveFinalDelivered(() => {
+      turn.finalAnswerEverDelivered = true
+      if (turn.finalAnswerDeliveredAt == null) turn.finalAnswerDeliveredAt = Date.now()
+    })
+    ea.finalizeCard(() => {
+      clearActivitySummary(turn)
+    })
+  }
+
   const result = await handleStreamReply(
     {
       chat_id: streamChatId,
@@ -8370,7 +9127,6 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
       logStreamingEvent,
       isPrivateChat: streamIsPrivate,
       isForumTopic: streamIsForumTopic,
-      ...(sendMessageDraftFn != null ? { sendMessageDraft: sendMessageDraftFn } : {}),
       // Issue #310: deliver the outbound count bump BEFORE forceCompleteTurn
       // so the terminal render sees outboundDeliveredCount > 0. The handler
       // calls this dep in that order internally.
@@ -8390,12 +9146,10 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
       recordOutbound,
       ...(HISTORY_ENABLED ? { getLatestInboundMessageId } : {}),
       writeError: (line) => process.stderr.write(line),
-      // PR B: drop the legacy 600 ms compromise. When the operator sets
-      // `channels.telegram.stream_throttle_ms` in yaml, the env override
-      // wins; otherwise draft-stream's transport-aware default fires
-      // (300 ms draft / 1000 ms message). `throttleMs: undefined` is a
-      // signal — handlers downgrade to `?? undefined`, which then
-      // passes through to draft-stream where the default applies.
+      // When the operator sets `channels.telegram.stream_throttle_ms` in yaml,
+      // the env override wins; otherwise draft-stream's DM/group defaults apply
+      // (400 ms for DMs, 1000 ms for groups). `throttleMs: undefined` passes
+      // through to draft-stream where the per-chat-type default applies.
       ...(STREAM_THROTTLE_MS_OVERRIDE != null ? { throttleMs: STREAM_THROTTLE_MS_OVERRIDE } : {}),
       progressCardActive: streamMode === 'checklist',
     },
@@ -8495,6 +9249,10 @@ async function executeStreamReply(args: Record<string, unknown>): Promise<unknow
       disableNotification: args.disable_notification === true,
       done: args.done === true,
     })
+    // Sticky ordering latch (lever 1): set once a SUBSTANTIVE final lands;
+    // never cleared by reopen. The card OPEN gate keys on this sticky latch.
+    if (turn.finalAnswerSubstantive) turn.finalAnswerEverDelivered = true
+    if (turn.finalAnswerSubstantive && turn.finalAnswerDeliveredAt == null) turn.finalAnswerDeliveredAt = Date.now()
     if (turn.finalAnswerSubstantive) closeObligationOnSubstantiveReply(args, turn, streamRoutedOriginTurn)
     // #1744 follow-up — stream_reply edge case. The first-emit gate at
     // L5178 only clears silent-end state on the FIRST emit of a stream.
@@ -9832,6 +10590,46 @@ function resetOrphanedReplyTimeout(): void {
         replyCalled: t.replyCalled,
         progressCardActive: progressDriver != null,
       })) {
+        // Feed-survival guard: re-arm the fuse while the turn is
+        // legitimately working — an in-flight tool, a detached background
+        // process (Bash run_in_background), or a human-wait tool (ask_user).
+        // This extends the original "isMidToolCall" guard to cover the
+        // detached-work cases that empty inFlight prematurely.
+        //
+        // Cap logic:
+        //  • Foreground tools / detached background work: bound by
+        //    ORPHANED_REPLY_MAX_REARMS (20 × 30 s = 10 min). A genuinely
+        //    hung tool still surfaces after the cap.
+        //  • Human-wait tools (ask_user): NEVER forcibly backstop while
+        //    ask_user is open for this chat — the human simply hasn't
+        //    tapped yet. We keep re-arming unconditionally until the prompt
+        //    resolves (TTL or tap) and inFlight empties.
+        const turnKey = statusKey(t.sessionChatId, t.sessionThreadId)
+        const working = isLegitimatelyWorking(turnKey)
+        const humanWaiting = (() => {
+          for (const entry of pendingAskUser.values()) {
+            if (entry.chatId === t.sessionChatId) return true
+          }
+          return false
+        })()
+        if (working || humanWaiting) {
+          const underCap = t.orphanedReplyRearmCount < ORPHANED_REPLY_MAX_REARMS
+          if (humanWaiting || underCap) {
+            t.orphanedReplyRearmCount++
+            process.stderr.write(
+              `telegram gateway: orphaned-reply fuse expired — re-arming` +
+              ` (rearm ${t.orphanedReplyRearmCount}/${ORPHANED_REPLY_MAX_REARMS},` +
+              ` in_flight=${toolFlightTracker.inFlightCount()},` +
+              ` human_wait=${humanWaiting},` +
+              ` bg_work=${pendingProgress.hasPendingAsyncDispatch(turnKey)})\n`,
+            )
+            resetOrphanedReplyTimeout()
+            return
+          }
+          process.stderr.write(
+            `telegram gateway: orphaned-reply rearm cap reached (${ORPHANED_REPLY_MAX_REARMS}) — forcing backstop despite working state\n`,
+          )
+        }
         process.stderr.write(
           `telegram gateway: orphaned-reply timeout (${ORPHANED_REPLY_TIMEOUT_MS}ms) — forcing backstop\n`,
         )
@@ -9876,13 +10674,112 @@ const FOREGROUND_SUBAGENT_ACCUM_MAX = 12
  * foreground sub-agents (rare — parallel Task dispatch) flatten in insertion
  * order; the single-sub-agent common case nests precisely under its
  * Delegating line.
+ *
+ * The header (elapsed + tool count) is now threaded into the render so the
+ * main-session card matches the worker card's two-line header style. This
+ * fixes the missing header regression where the worker card showed elapsed/
+ * tool-count metadata but the main-session card rendered step-lines only.
  */
 function composeTurnActivity(turn: CurrentTurn, final = false, liveSuffix = ''): string | null {
   const childLines: string[] = []
   for (const narrative of turn.foregroundSubAgents.values()) {
     childLines.push(...narrative)
   }
-  return renderActivityFeedWithNested(turn.mirrorLines, childLines, final, liveSuffix)
+  // Pass labeledToolCount as stepCount only on the terminal (final) render so
+  // the persisted feed record shows a `✓ N steps` total. The live in-progress
+  // feed omits it (stepCount undefined) to stay clean and minimal.
+  const stepCount = final ? turn.labeledToolCount : undefined
+  // Build the session header so the main-session card renders the same two-line
+  // elapsed/tool-count header as the worker card.
+  const header: SessionActivityHeader = {
+    label: 'Agent',
+    elapsedMs: turn.startedAt > 0 ? Date.now() - turn.startedAt : 0,
+    toolCount: turn.labeledToolCount,
+    state: final ? 'done' : 'running',
+  }
+  return renderActivityFeedWithNested(turn.mirrorLines, childLines, final, liveSuffix, stepCount, header)
+}
+
+/**
+ * Render a SHOWN narrative text block as a transient liveness step — the
+ * same path a tool label takes (appendActivityLabel → renderStepFeed), so
+ * the narrative line is rolling-window-clipped and replaced by the next
+ * event exactly like a tool step. NOT a new message, NOT persisted as a
+ * parallel mirror (invariant `chat-is-the-single-source-of-truth`,
+ * reference/invariants.md). Clipped to a single 120-char line via
+ * clipNarrative so it reads as a step, not a paragraph.
+ */
+function showNarrativeStep(turn: CurrentTurn, text: string): void {
+  const rendered = appendActivityLabel(turn.mirrorLines, clipNarrative(text))
+  if (rendered == null) return
+  turn.activityPendingRender = composeTurnActivity(turn) ?? rendered
+  const ea = emissionAuthorityFor(turn)
+  // PR-4d: route the deliver-before-drain decision through the centralized
+  // card-drain gate (chatLock-serialized under the flag; verbatim block OFF).
+  cardDrainGate(turn, ea, () => {
+  if (ea.mayDrain(turn)) {
+    // Producer A (narrative SHOW): may only EDIT an already-open card, never
+    // OPEN one on a 0-tool turn (design §9 lever 5 base case — the
+    // triplication). The OPEN gate in the drain enforces this; accumulation
+    // into mirrorLines still happens so the narration renders once a tool
+    // label or liveness opens the card.
+    // PR-4a: routed through the emission-authority façade (no-op delegate).
+    ea.openOrEditCard('narrative', () => {
+      turn.activityInFlight = drainActivitySummary(turn, 'narrative')
+    })
+  }
+  })
+}
+
+/**
+ * Narrative-dedup gate, step 2 (reducer-side): a tool_use just arrived while
+ * a narrative block was pending. Decide SHOW vs SUPPRESS and clear the
+ * pending slot. SUPPRESS only when the tool is reply/stream_reply AND the
+ * pending text is a draft-then-send of that reply's `input.text`. Everything
+ * else (a working tool, or a reply whose text differs — post-action
+ * narration) is SHOWN. See narrative-dedup.ts §2b.
+ */
+function resolvePendingNarrativeOnTool(
+  turn: CurrentTurn,
+  toolName: string,
+  input: Record<string, unknown> | undefined,
+): void {
+  const pending = turn.pendingNarrative
+  if (pending == null) return
+  turn.pendingNarrative = null
+  if (REPLY_TOOLS.has(toolName)) {
+    const replyText = typeof input?.text === 'string' ? (input.text as string) : ''
+    if (isDraftOfReply(pending.text, replyText)) return // draft of the answer → SUPPRESS
+  }
+  showNarrativeStep(turn, pending.text) // working preamble / post-action narration → SHOW
+}
+
+/**
+ * Narrative-dedup gate, step 1 (reducer-side): a new narrative block
+ * arrived. A previously-pending block had nothing reply-shaped immediately
+ * after it (pure narration) → flush it as SHOWN, then stage the new one for
+ * one lookahead step. See narrative-dedup.ts §2b.
+ */
+function stagePendingNarrative(turn: CurrentTurn, text: string): void {
+  if (turn.pendingNarrative != null) {
+    showNarrativeStep(turn, turn.pendingNarrative.text)
+  }
+  turn.pendingNarrative = { text }
+}
+
+/**
+ * Narrative-dedup gate, step 3 (reducer-side): the turn is ending with a
+ * trailing narrative block and nothing after it. SUPPRESS only when the turn
+ * already delivered its answer via reply/stream_reply and the trailing text
+ * is a draft of that answer; otherwise SHOW (genuine trailing narration like
+ * "Done — all green."). See narrative-dedup.ts §2b.
+ */
+function flushPendingNarrativeAtTurnEnd(turn: CurrentTurn, lastReplyText: string): void {
+  const pending = turn.pendingNarrative
+  if (pending == null) return
+  turn.pendingNarrative = null
+  if (lastReplyText.length > 0 && isDraftOfReply(pending.text, lastReplyText)) return // trailing duplicate of the answer
+  showNarrativeStep(turn, pending.text)
 }
 
 /**
@@ -9902,11 +10799,69 @@ function composeTurnActivity(turn: CurrentTurn, final = false, liveSuffix = ''):
  * doesn't corrupt the next turn's atom — late writes land on the
  * captured `turn` (already-completed turn, harmless).
  */
-async function drainActivitySummary(turn: CurrentTurn): Promise<void> {
+async function drainActivitySummary(
+  turn: CurrentTurn,
+  // Which producer triggered this drain (design §9 levers 1 + 5). Gates the
+  // OPEN (first sendMessage) branch via `mayOpenActivityCard`; EDITs of an
+  // already-open card are never gated. Defaults to 'tool' — the historically
+  // unconditional OPEN behaviour — so any caller that does not opt into the
+  // gate is unaffected. Narrative-SHOW and liveness callers pass their producer
+  // explicitly.
+  producer: FeedOpenProducer = 'tool',
+  // Optional flags forwarded to `mayOpenActivityCard`.
+  openFlags?: { postAnswerSubagentActivity?: boolean },
+): Promise<void> {
   try {
     while (turn.activityPendingRender !== turn.activityLastSentRender) {
       const target = turn.activityPendingRender
       if (target == null) break
+      // OPEN gate (design §9 levers 1 + 5): when this drain would OPEN a fresh
+      // card (activityMessageId == null), consult the pure gate. Refusing an
+      // OPEN must NOT advance activityLastSentRender — the accumulated render
+      // stays pending so a later OPEN-eligible producer (a tool label, or
+      // liveness) renders it. An EDIT (activityMessageId != null) is never
+      // gated. Enforced HERE so it covers BOTH the inline producers AND the
+      // detached heartbeat setInterval drain (R7/concurrency). The gate guards
+      // gate EVALUATION, not an in-flight send: it is not a hard mutex — a send
+      // already PAST this check and suspended at its `await robustApiCall(
+      // sendMessage)` when a substantive final lands still completes and opens a
+      // card; that residual is reconciled by lever-2's `clearActivitySummary`
+      // chaining its finalize onto `turn.activityInFlight` (the suspended drain)
+      // and editing the card in place, not by this gate blocking it.
+      // Lever 4 (cross-turn / race C/D): a synthetic represent/owed-reply turn
+      // (and the liveness/heartbeat timer firing on it) starts with a CLEARED
+      // per-turn `finalAnswerEverDelivered` latch even when a substantive answer
+      // already reached the user in an EARLIER turn — so without this its first
+      // drain opens a card BELOW that prior reply. Only such a turn carries
+      // `crossTurnGate`; reuse the represent guard's delivered-since check
+      // (`hasOutboundDeliveredSince`) with the obligation's `openedAt` cutoff and
+      // the SUBSTANTIVE 200-char threshold (so an ack never trips it → #2141
+      // stays green). Computed ONLY when about to OPEN (activityMessageId ==
+      // null) AND only for a turn with a cross-turn gate — no history query on
+      // the common foreground path. Scoped to the synthetic surface by the
+      // presence of `crossTurnGate`, so it can never fire on a foreground turn.
+      // PR-4b: the cross-turn predicate is now the PURE, shared helper extracted
+      // into feed-open-gate.ts (body lifted verbatim) — the SAME function the
+      // emission-authority façade calls in its enabled branch, so flag-ON and
+      // flag-OFF compute an identical verdict. History deps injected (the module
+      // stays sqlite-free). The pure-gate consult + the `break` below stay
+      // LITERALLY in the drain (disabled-path byte-identity).
+      const crossTurnAnswerDelivered = computeCrossTurnAnswerDelivered(
+        turn,
+        feedOpenGateDeps(),
+      )
+      if (
+        turn.activityMessageId == null
+        && !mayOpenActivityCard({
+          producer,
+          finalAnswerEverDelivered: turn.finalAnswerEverDelivered,
+          labeledToolCount: turn.labeledToolCount,
+          crossTurnAnswerDelivered,
+          postAnswerSubagentActivity: openFlags?.postAnswerSubagentActivity,
+        })
+      ) {
+        break
+      }
       // `renderActivityFeed` already emitted ready Telegram HTML with per-line
       // markup (<b>→ current</b> / <i>✓ done</i>) and escaped each label's
       // <,>,& itself (#1942 class) — send verbatim, do NOT re-escape or
@@ -9980,17 +10935,125 @@ async function drainActivitySummary(turn: CurrentTurn): Promise<void> {
 function feedHeartbeatTick(): void {
   const turn = currentTurn
   if (turn == null) return
+  if (turn.finalAnswerDelivered) {
+    // Fix 2: post-answer background-agent liveness. When the sub-agent/workflow
+    // watcher has surfaced a new step AFTER the substantive final answer, drive
+    // a liveness card so the operator can see "background agent still working".
+    //
+    // Gate: `turn.subagentActivityAt` must be set (watcher fired) AND it must
+    // exceed `turn.finalAnswerDeliveredAt` (the watcher advanced AFTER the answer
+    // was delivered — not just any pre-answer label). This is the key fix:
+    // #2587 read `lastToolLabelAt`, which is frozen by the drop-guard after a
+    // substantive answer and therefore never crosses the threshold. `subagentActivityAt`
+    // is written by the watcher's onProgress callback INDEPENDENTLY of the
+    // tool_label / drop-guard path, so it correctly advances post-answer.
+    //
+    // Idle-gap suppression + staleness cap (concern 3) — the single pure decision
+    // `evaluatePostAnswerLiveness`:
+    //   - 'idle'  → no watcher activity after the answer (`subagentActivityAt`
+    //               undefined or ≤ finalAnswerDeliveredAt). Stay silent; the
+    //               reply-is-last invariant is fully preserved for idle turns.
+    //   - 'stale' → the worker's last advance is older than POST_ANSWER_LIVENESS_STALE_MS
+    //               (its `onFinish` froze `subagentActivityAt` and no new step has
+    //               arrived). STOP re-rendering so the card doesn't climb `running`
+    //               forever — mirrors the pre-answer FEED_LIVENESS_OPEN_MS cap. The
+    //               worker's own terminal card (workerActivityFeed.finish) is the
+    //               durable record once it completes.
+    //   - 'emit'  → genuine in-flight post-answer activity; render the card below.
+    const subagentAt = turn.subagentActivityAt
+    const livenessVerdict = evaluatePostAnswerLiveness({
+      subagentActivityAt: subagentAt,
+      finalAnswerDeliveredAt: turn.finalAnswerDeliveredAt,
+      now: Date.now(),
+      staleCapMs: POST_ANSWER_LIVENESS_STALE_MS,
+    })
+    if (livenessVerdict !== 'emit' || subagentAt == null) return // idle gap or stale worker → stay silent (the `== null` also narrows subagentAt for the elapsed below)
+    // A background worker is genuinely active after the answer. Open or maintain
+    // a liveness card below the reply. Route through `mayOpenActivityCard` with
+    // `postAnswerSubagentActivity:true` so Lever 1 is lifted for 'tool' producer
+    // (Fix 2's Lever 1 exception in feed-open-gate.ts). The card renders the
+    // turn's accumulated mirrorLines (which may be empty — in that case the drain
+    // opens a "Working…" placeholder matching the pre-answer liveness path).
+    if (turn.sessionChatId == null) return
+    const age = Date.now() - turn.startedAt
+    const livenessHeader: SessionActivityHeader = {
+      label: 'Agent', elapsedMs: age, toolCount: turn.labeledToolCount, state: 'running',
+    }
+    const lines = turn.mirrorLines.length > 0 ? turn.mirrorLines : ['Working in background…']
+    const elapsed = Date.now() - subagentAt
+    const rendered = renderActivityFeedWithNested(lines, [], false, ` · ${formatFeedElapsed(elapsed)}`, undefined, livenessHeader)
+    if (rendered == null) return
+    turn.activityPendingRender = rendered
+    const ea = emissionAuthorityFor(turn)
+    cardDrainGate(turn, ea, () => {
+    if (ea.mayDrain(turn)) {
+      // Producer 'tool' with postAnswerSubagentActivity=true: the Lever 1
+      // exception allows this OPEN. Lever 4 (cross-turn) and idle-liveness
+      // blocks are still respected by the drain. The card surfaces BELOW the
+      // reply showing the background agent's live activity.
+      ea.openOrEditCard('tool', () => {
+        turn.activityInFlight = drainActivitySummary(turn, 'tool', { postAnswerSubagentActivity: true })
+      })
+    }
+    })
+    return
+  }
+
+  // Liveness feed (open + maintain). `mirrorLines.length === 0` means no tool
+  // has ever produced a label this turn — pure thinking, or only suppressed
+  // tools. Open a minimal "Working…" feed once the turn passes the threshold,
+  // and keep its elapsed climbing until a real label arrives. The first label
+  // makes mirrorLines non-empty, so the labelled-feed heartbeat below takes
+  // over and its edit cleanly replaces the placeholder. drainActivitySummary
+  // sends (opens) when activityMessageId is null and edits (maintains) once set
+  // — so this one branch handles both the open and the climb.
+  if (turn.mirrorLines.length === 0) {
+    if (!FEED_LIVENESS_OPEN_ENABLED || turn.sessionChatId == null) return
+    const age = Date.now() - turn.startedAt
+    if (age < FEED_LIVENESS_OPEN_MS) return
+    const livenessHeader: SessionActivityHeader = {
+      label: 'Agent', elapsedMs: age, toolCount: 0, state: 'running',
+    }
+    const rendered = renderActivityFeedWithNested(['Working…'], [], false, ` · ${formatFeedElapsed(age)}`, undefined, livenessHeader)
+    if (rendered == null) return
+    turn.activityPendingRender = rendered
+    const ea = emissionAuthorityFor(turn)
+    // PR-4d: route through the centralized chatLock-serialized card-drain gate.
+    cardDrainGate(turn, ea, () => {
+    if (ea.mayDrain(turn)) {
+      // Producer C (liveness timer): the genuine ≥12s thinking-gap open. Now
+      // that Lever 5 is inert (narrative may open pre-answer — #2588), liveness
+      // remains the natural open for 0-tool pre-answer turns that are silent.
+      // The sticky-latch (lever 1) still gates it in the drain.
+      // PR-4a: routed through the emission-authority façade (no-op delegate).
+      ea.openOrEditCard('liveness', () => {
+        turn.activityInFlight = drainActivitySummary(turn, 'liveness')
+      })
+    }
+    })
+    return
+  }
+
+  // Labelled-feed heartbeat: keep a stale in-progress step visibly advancing.
   if (turn.activityMessageId == null) return // no live feed yet / already cleared
-  if (turn.finalAnswerDelivered) return // feed handed off to the answer
   if (turn.lastToolLabelAt == null) return // feed not driven by a labelled step
   const elapsed = Date.now() - turn.lastToolLabelAt
   if (elapsed < FEED_HEARTBEAT_MIN_STALE_MS) return // step is fresh; feed advancing normally
   const rendered = composeTurnActivity(turn, false, ` · ${formatFeedElapsed(elapsed)}`)
   if (rendered == null) return
   turn.activityPendingRender = rendered
-  if (turn.activityInFlight == null) {
-    turn.activityInFlight = drainActivitySummary(turn)
+  const ea = emissionAuthorityFor(turn)
+  // PR-4d: route through the centralized chatLock-serialized card-drain gate.
+  cardDrainGate(turn, ea, () => {
+  if (ea.mayDrain(turn)) {
+    // Maintains an already-open card (guarded above on activityMessageId !=
+    // null) → only ever EDITs. 'liveness' is correct either way.
+    // PR-4a: routed through the emission-authority façade (no-op delegate).
+    ea.openOrEditCard('liveness', () => {
+      turn.activityInFlight = drainActivitySummary(turn, 'liveness')
+    })
   }
+  })
 }
 if (!STATIC && FEED_HEARTBEAT_ENABLED) {
   setInterval(feedHeartbeatTick, FEED_HEARTBEAT_TICK_MS).unref()
@@ -10043,8 +11106,19 @@ function clearActivitySummary(turn: CurrentTurn, finalHtmlOverride?: string | nu
     }
     // Default: leave the status message as a record, edited to a terminal
     // all-done state so it doesn't freeze on a misleading "→ in-progress" line.
-    const finalHtml =
+    let finalHtml =
       finalHtmlOverride !== undefined ? finalHtmlOverride : composeTurnActivity(turn, true)
+    // Liveness-only feed: opened on the timer for a turn that never labelled a
+    // tool (pure thinking / suppressed tools), so mirrorLines is empty and the
+    // terminal render is null. Finalize to a done "✓ Working…" record instead
+    // of leaving the message frozen on the live "→ Working…" line.
+    if (finalHtml == null && turn.mirrorLines.length === 0 && turn.activityEverOpened) {
+      const livenessElapsed = turn.startedAt > 0 ? Date.now() - turn.startedAt : 0
+      const livenessHeader: SessionActivityHeader = {
+        label: 'Agent', elapsedMs: livenessElapsed, toolCount: turn.labeledToolCount, state: 'done',
+      }
+      finalHtml = renderActivityFeedWithNested(['Working…'], [], true, '', undefined, livenessHeader)
+    }
     if (finalHtml == null) return
     try {
       await robustApiCall(
@@ -10111,6 +11185,19 @@ function handleSessionEvent(ev: SessionEvent): void {
         const turnId =
           deriveTurnId(ev.chatId, enqThreadIdNum ?? null, ev.messageId)
           ?? `${chatKey(ev.chatId, enqThreadIdNum ?? null)}#synthetic-${startedAt}`
+        // PR1 (cross-turn stale-card guard, §9 lever 4 / race C/D). Consume any
+        // pending cross-turn gate `obligationSweep` armed for THIS exact turn
+        // when it pushed an `obligation_represent` inbound. The gate is keyed on
+        // the obligation's `originTurnId`, and the represent inbound reuses the
+        // original chat/thread/messageId, so this turn's `turnId` (derived just
+        // above) equals that key iff this turn IS the represent surface armed for.
+        // An unrelated foreground turn on the same chat/thread derives a
+        // different `turnId` → finds no entry → no gate → its card opens normally
+        // (correct). Consume-once: delete on read so the matched gate can't leak
+        // forward, and a never-matched stale gate can never suppress another turn.
+        const xTurnGateKey = turnId
+        const consumedCrossTurnGate = pendingCrossTurnGate.get(xTurnGateKey)
+        if (consumedCrossTurnGate != null) pendingCrossTurnGate.delete(xTurnGateKey)
         const next: CurrentTurn = {
           sessionChatId: ev.chatId,
           sessionThreadId: enqThreadIdNum,
@@ -10123,20 +11210,34 @@ function handleSessionEvent(ev: SessionEvent): void {
           sourceMessageId: parseSourceMessageId(ev.messageId),
           startedAt,
           gatewayReceiveAt: startedAt,
+          // #2527 — stamp the loop role once, from the enqueue envelope.
+          role: deriveTurnRole(ev.rawContent),
+          // PR1 (cross-turn stale-card guard, §9 lever 4 / race C/D). Only a
+          // synthetic represent/owed-reply turn carries this; a foreground turn
+          // leaves it undefined and the cross-turn card-OPEN gate is inert.
+          ...(consumedCrossTurnGate != null ? { crossTurnGate: consumedCrossTurnGate } : {}),
           replyCalled: false,
           finalAnswerDelivered: false,
           finalAnswerSubstantive: false,
+          // Sticky latch — reset ONLY here (turn start), never by reopen.
+          finalAnswerEverDelivered: false,
           firstPingAt: null,
+          // Notification ownership (R8 / PR-2): no slot claimed yet, so the
+          // "claimer was substantive" flag starts false. Set atomically with
+          // firstPingAt at the over-ping decision site.
+          firstPingWasSubstantive: false,
           silentAnchorMessageId: null,
           silentAnchorText: '',
           capturedText: [],
           orphanedReplyTimeoutId: null,
+          orphanedReplyRearmCount: 0,
           turnId,
           registryKey: null,
           noReplyDrainTimer: null,
           lastAssistantMsgId: null,
           lastAssistantDone: false,
           toolCallCount: 0,
+          labeledToolCount: 0,
           activityMessageId: null,
           activityInFlight: null,
           activityPendingRender: null,
@@ -10144,11 +11245,24 @@ function handleSessionEvent(ev: SessionEvent): void {
           activityEverOpened: false,
           activityDrainFailures: 0,
           mirrorLines: [],
+          pendingNarrative: null,
+          lastReplyText: '',
           foregroundSubAgents: new Map(),
           answerStream: null,
           isDm: isDmChatId(ev.chatId),
+          // PR-4a — construct ONE emission-authority façade per turn, passing
+          // the chat/thread key in EXPLICITLY (the PR-4e seam; today equal to
+          // the singleton-sourced key). Per-turn: born with this turn literal,
+          // discarded with it — never persists across turns.
+          emissionAuthority: new EmissionAuthority(
+            statusKey(ev.chatId, enqThreadIdNum),
+          ),
         }
-        currentTurn = next
+        // PR-4e — route the turn-SET through the keyed accessor: flag-OFF assigns
+        // the singleton (byte-identical to `currentTurn = next`); flag-ON sets the
+        // per-topic `byKey[statusKey]` entry AND the most-recent mirror. The key is
+        // the SAME statusKey the ctor's façade was constructed with just above.
+        setCurrentTurn(next, statusKey(ev.chatId, enqThreadIdNum))
         markIdleActivity() // any turn start (main session) is activity — re-arm idle clear
         // Status-surface observability: one line at every turn SET so a later
         // dark card is traceable to which turn/topic key it belonged to.
@@ -10278,6 +11392,14 @@ function handleSessionEvent(ev: SessionEvent): void {
     case 'tool_use': {
       const turn = currentTurn
       if (turn == null) return
+      // Narrative-dedup gate step 2 (JSONL-text-narrative primitive): a
+      // narrative block was pending; this tool_use is the lookahead event
+      // that decides it. reply/stream_reply with near-identical text ⇒
+      // draft-then-send ⇒ SUPPRESS (the reply prints the canonical answer);
+      // anything else ⇒ SHOW as a transient liveness step. Runs BEFORE the
+      // normal tool handling so a working preamble surfaces just ahead of
+      // its tool step.
+      resolvePendingNarrativeOnTool(turn, ev.toolName, ev.input)
       // Phase 1 of #332: count every tool_use in the current turn.
       turn.toolCallCount++
       // #412: bump turn-active marker mtime so the watchdog sees this
@@ -10300,6 +11422,15 @@ function handleSessionEvent(ev: SessionEvent): void {
       // placeholder-heartbeat label, which has been retired.
       if (isTelegramReplyTool(name)) {
         turn.replyCalled = true
+        // NIT 2 (reply-proxy precision): capture the ACTUAL delivered reply
+        // text so flushPendingNarrativeAtTurnEnd compares a trailing
+        // narrative block against the real answer surface, not
+        // capturedText.join('') (which mis-suppresses when the model emits
+        // the same short string twice in a turn). REPLY_TOOLS ('reply',
+        // 'stream_reply') carry the answer in input.text; only those count.
+        if (REPLY_TOOLS.has(name) && typeof ev.input?.text === 'string') {
+          turn.lastReplyText = ev.input.text as string
+        }
         if (turn.orphanedReplyTimeoutId != null) {
           clearTimeout(turn.orphanedReplyTimeoutId)
           turn.orphanedReplyTimeoutId = null
@@ -10339,6 +11470,12 @@ function handleSessionEvent(ev: SessionEvent): void {
       // where the JSONL tool_use rows arrive too late.
       const turn = currentTurn
       if (turn == null) return
+      // SECONDARY FIX: an active tool_label means the model is producing work
+      // right now — re-arm the orphaned-reply fuse so a multi-phase tool turn
+      // (write → compile → test → fix) that regularly emits labels doesn't let
+      // the 30 s timer run down between labels. Mirrors how `case 'text':` calls
+      // resetOrphanedReplyTimeout() at ~line 10786.
+      resetOrphanedReplyTimeout()
       // Surface tools (reply/stream_reply/react) are the conversation, not
       // activity — the hook labels them ("Replying"), so filter by name.
       if (isTelegramSurfaceTool(ev.toolName)) return
@@ -10400,6 +11537,14 @@ function handleSessionEvent(ev: SessionEvent): void {
       }
       const rendered = appendActivityLabel(turn.mirrorLines, ev.label)
       if (rendered != null) {
+        // Count surfaced tool steps — the single source of truth for the `tools=`
+        // lifecycle field and the `✓ N steps` total. Incremented HERE (not at the
+        // top of the case) so the count stays consistent with what the feed
+        // actually surfaces: an empty label (appendActivityLabel → null) or a
+        // label dropped by the post-final-answer reopen guard never inflates it.
+        // Surface tools (reply/react) returned earlier; send_typing/sync_retain
+        // are suppressed at the hook (computeLabel → null) so they never arrive.
+        turn.labeledToolCount++
         // A new tool label = a new live step → re-anchor the heartbeat clock so
         // the " · Ns" elapsed restarts from this step (and the feed itself just
         // advanced, so it isn't stale).
@@ -10411,16 +11556,34 @@ function handleSessionEvent(ev: SessionEvent): void {
         // and would falsely reset the clock forever on a hung-mid-tool turn,
         // reintroducing the #1556 dangling-turn wedge. Only the model emitting a
         // fresh label reaches here.
-        if (SILENCE_LIVENESS_PRODUCTION && currentTurn === turn) {
+        // PR-4e — keyed liveness under the flag. Flag-OFF keeps the literal
+        // `currentTurn === turn` (a late tool-label for topic A must reset A's
+        // silence clock, not topic B's); flag-ON resolves A by ITS OWN key so a
+        // flip to B doesn't falsify A's liveness here.
+        if (
+          SILENCE_LIVENESS_PRODUCTION &&
+          (EMISSION_AUTHORITY_ENABLED ? turnLiveForItsTopic(turn) : currentTurn === turn)
+        ) {
           silencePoke.noteProduction(statusKey(turn.sessionChatId, turn.sessionThreadId), Date.now())
         }
         // Recompose so any active foreground sub-agent's nested block (Model A)
         // is preserved when the parent appends its own step. composeTurnActivity
         // == the flat render when no foreground sub-agent is active.
         turn.activityPendingRender = composeTurnActivity(turn) ?? rendered
-        if (turn.activityInFlight == null) {
-          turn.activityInFlight = drainActivitySummary(turn)
+        const ea = emissionAuthorityFor(turn)
+        // PR-4d: route through the centralized chatLock-serialized card-drain gate.
+        cardDrainGate(turn, ea, () => {
+        if (ea.mayDrain(turn)) {
+          // Producer B (tool label): always OPEN-eligible (labeledToolCount was
+          // incremented just above). A turn that started conversational and now
+          // dispatches a tool opens here, rendering any narration accumulated
+          // by the suppressed narrative-SHOW drains (design §9 lever 5 / R4).
+          // PR-4a: routed through the emission-authority façade (no-op delegate).
+          ea.openOrEditCard('tool', () => {
+            turn.activityInFlight = drainActivitySummary(turn, 'tool')
+          })
         }
+        })
       }
       return
     }
@@ -10447,58 +11610,33 @@ function handleSessionEvent(ev: SessionEvent): void {
       const turn = currentTurn
       if (turn != null) {
         turn.capturedText.push(ev.text)
+        // Narrative-dedup gate step 1 (JSONL-text-narrative primitive):
+        // stage this text block for one lookahead step. If a previous block
+        // was pending with nothing reply-shaped after it, it flushes here as
+        // a SHOWN transient liveness step. The eventual SHOW/SUPPRESS of THIS
+        // block is decided by the next tool_use / turn_end. Invariant
+        // `chat-is-the-single-source-of-truth` (reference/invariants.md): a
+        // SHOWN line rides the same renderStepFeed path as a tool step —
+        // transient + clipped, never a persisted parallel mirror. This is a
+        // separate lane from the answer-stream wiring below (which owns the
+        // canonical reply), so the two never fight over the same text.
+        stagePendingNarrative(turn, ev.text)
         // Issue #195: feed the answer-lane stream. The stream itself
         // gates on minInitialChars and throttles edits — short replies
         // stay below the threshold and never spawn a message.
         if (turn.answerStream == null) {
           turn.answerStream = createAnswerStream({
             chatId: turn.sessionChatId,
-            isPrivateChat: turn.isDm,
             threadId: turn.sessionThreadId,
-            // Transport selection:
-            // #869-Phase1 visible-answer-stream: omit the draft API so
-            // the lane edits a user-visible chat-timeline message
-            // (minInitialChars:1 opens it on the first chunk). The
-            // draft-mirror does NOT touch this lane — the canary proved
-            // the model emits almost no interstitial assistant.text
-            // (it thinks→tool→reply), so routing it to the draft just
-            // emptied the preview. The draft-mirror instead renders the
-            // tool_use stream (case 'tool_use' above) where the real
-            // signal lives. assistant.text keeps its visible-message
-            // home; the reply tool stays the canonical answer.
-            // Flag OFF (default): use the compose-box draft for DMs, and set
-            // minInitialChars effectively-infinite so the lane NEVER opens a
-            // visible chat message. This matters in supergroup TOPICS, where
-            // draft transport is unsupported (gateway.ts:6422) so the lane
-            // would otherwise fall to message transport and post a visible
-            // preview once interstitial text passed the default 50-char gate
-            // — which retract() then deletes (the unformatted flash, marko
-            // General). With the gate unreachable the only posted message is
-            // the canonical reply. (The gate is bypassed for DM draft
-            // transport, so DM draft streaming is unaffected.)
-            // VISIBLE preview gating decoupled from the draft-transport flag
-            // (2026-06-05 flash regression fix). The visible flag ALONE decides
-            // whether a user-visible preview opens; DRAFT_ANSWER_LANE_RETIRED
-            // controls only the TRANSPORT (whether sendMessageDraftFn exists).
-            // The earlier `|| DRAFT_ANSWER_LANE_RETIRED` here meant retiring the
-            // draft (the default since v0.14.68) silently forced minInitialChars:1
-            // → a visible preliminary opened on every streaming turn and was then
-            // retracted (deleted) when the reply tool fired — the exact "raw bubble
-            // appears, formatted reply lands, raw bubble vanishes" flash that
-            // turning the visible stream OFF (v0.14.52) was meant to remove. So
-            // v0.14.68 silently undid v0.14.52 fleet-wide. Now:
-            //   - VISIBLE on (opt-in) → minInitialChars:1, a real edit-in-place
-            //     preview (observable by UAT, silence-liveness reset on its sends).
-            //   - VISIBLE off (default) → minInitialChars:MAX so NO visible preview
-            //     ever opens; the reply tool is the single canonical formatted
-            //     message (no flash). With the draft retired (default) there is no
-            //     transport either, so the lane stays dormant; with the kill switch
-            //     DRAFT_ANSWER_LANE=0 the legacy compose-box draft transport is
-            //     restored (sendMessageDraftFn defined above, gate bypassed for DM
-            //     draft so #1664 DM draft streaming is unaffected).
-            ...(ANSWER_LANE.usesDraftTransport
-              ? { sendMessageDraft: sendMessageDraftFn, minInitialChars: ANSWER_LANE.minInitialChars }
-              : { minInitialChars: ANSWER_LANE.minInitialChars }),
+            // VISIBLE on (opt-in, SWITCHROOM_VISIBLE_ANSWER_STREAM=1) →
+            //   minInitialChars:1 opens a user-visible edit-in-place preview on the
+            //   first text chunk. At turn_end the preview is materialized as a pinged
+            //   final answer (materialize()) when the model never called reply.
+            // VISIBLE off (default) → minInitialChars:MAX so NO visible preview ever
+            //   opens; the reply tool is the single canonical formatted message
+            //   (no flash). The draft transport is permanently retired — both modes
+            //   use sendMessage + editMessageText for any message that does open.
+            minInitialChars: ANSWER_LANE.minInitialChars,
             // #1075: route through robustApiCall so flood-wait,
             // benign-400, and THREAD_NOT_FOUND are handled uniformly
             // instead of crashing the answer-stream loop on a deleted
@@ -10589,7 +11727,11 @@ function handleSessionEvent(ev: SessionEvent): void {
             // skip the tick (the new turn has its own answer stream).
             onMetric: (metricEv) => {
               logStreamingEvent(metricEv)
-              if (currentTurn === turn) {
+              // PR-4e — keyed liveness under the flag. Flag-OFF keeps the literal
+              // `currentTurn === turn` (a draft-update metric for topic A's stream
+              // must tick A's signal/silence clock); flag-ON resolves A by its own
+              // key so a flip to B doesn't skip A's tick.
+              if (EMISSION_AUTHORITY_ENABLED ? turnLiveForItsTopic(turn) : currentTurn === turn) {
                 signalTracker.noteSignal(
                   statusKey(turn.sessionChatId, turn.sessionThreadId),
                   Date.now(),
@@ -10717,6 +11859,28 @@ function handleSessionEvent(ev: SessionEvent): void {
       return
     }
     case 'turn_end': {
+      // DEFENSIVE FIX: belt-and-braces guard against the synthetic backstop
+      // (`durationMs: -1`) racing live work. durationMs >= 0 is the
+      // authoritative signal from system/turn_duration; -1 is ONLY ever set
+      // by the orphaned-reply backstop. Reject the synthetic event here so that
+      // even if the PRIMARY fix's re-arm logic is bypassed (e.g. a very fast
+      // fire before isLegitimatelyWorking() is sampled) we still don't tear
+      // down a live feed mid-work. Extended from the original isMidToolCall()
+      // check to the full isLegitimatelyWorking predicate so detached background
+      // work and human-wait tools (ask_user) are also protected.
+      // INVARIANT: a REAL turn_end (durationMs >= 0) is NEVER suppressed.
+      if (ev.durationMs === -1) {
+        const turn = currentTurn
+        const key = turn != null ? statusKey(turn.sessionChatId, turn.sessionThreadId) : ''
+        if (isLegitimatelyWorking(key)) {
+          process.stderr.write(
+            `telegram gateway: synthetic turn_end suppressed — legitimately working` +
+            ` (in_flight=${toolFlightTracker.inFlightCount()},` +
+            ` bg_work=${turn != null ? pendingProgress.hasPendingAsyncDispatch(key) : false})\n`,
+          )
+          return
+        }
+      }
       // Drain any still-pending tool dispatch typing entries — covers
       // transcript truncation or a Claude Code crash mid-tool.
       typingWrapper.drainAll()
@@ -10733,6 +11897,33 @@ function handleSessionEvent(ev: SessionEvent): void {
         clearTimeout(turn.orphanedReplyTimeoutId)
         turn.orphanedReplyTimeoutId = null
       }
+      // Narrative-dedup gate step 3 (JSONL-text-narrative primitive): a
+      // trailing narrative block with nothing after it. When the turn
+      // delivered its answer via reply (replyCalled) the trailing text is
+      // almost always a draft of that answer — compare against the ACTUAL
+      // delivered reply text and SUPPRESS the duplicate; otherwise SHOW
+      // genuine trailing narration ("Done — all green."). Must run BEFORE
+      // clearActivitySummary so a SHOWN line lands in the feed's final
+      // render. Always clears turn.pendingNarrative so it can't leak across
+      // turns.
+      //
+      // NIT 2 (reply-proxy precision): use `turn.lastReplyText` (the
+      // most-recent reply/stream_reply input.text) rather than
+      // `capturedText.join('')`. The old proxy concatenated every captured
+      // text block, so a turn that emitted the same short string twice
+      // (e.g. "Done." as working narration, then "Done." as the reply) would
+      // compare the trailing narration against a doubled "DoneDone" — still
+      // a high-prefix match — and wrongly suppress genuine trailing
+      // narration. Comparing against the actual reply text is exact. When
+      // the turn delivered WITHOUT a reply tool (turn-flush emits
+      // capturedText as the answer), fall back to capturedText.join('') so
+      // that path's trailing-draft suppression is preserved.
+      if (turn != null) {
+        const deliveredText = turn.lastReplyText.length > 0
+          ? turn.lastReplyText
+          : (turn.replyCalled ? turn.capturedText.join('') : '')
+        flushPendingNarrativeAtTurnEnd(turn, deliveredText)
+      }
       // Clear the activity feed at the real end of the turn. This is the
       // no-reply safety net — a turn that ends without ever calling reply
       // (the answer is delivered by turn-flush / silent-end) still has its
@@ -10934,6 +12125,17 @@ function handleSessionEvent(ev: SessionEvent): void {
             ` chat=${chatId} turnStartedAt=${turn.startedAt} replyCalled=false capturedText=empty` +
             ` — the progress card steps were the only thing the user saw (#45)\n`,
           )
+          // #2527: emit structured WARN so the reaction-only failure mode is
+          // machine-readable in the streaming-metrics channel.
+          const tKey = statusKey(chatId, threadId)
+          logStreamingEvent({
+            kind: 'turn_no_reply_warn',
+            chatId,
+            threadId,
+            turnId: turn.turnId,
+            turnDurationMs: turn.startedAt > 0 ? Date.now() - turn.startedAt : 0,
+            reactionCount: reactionTransitionCounts.get(tKey) ?? 0,
+          })
         }
       }
 
@@ -11291,10 +12493,39 @@ function handleSessionEvent(ev: SessionEvent): void {
       }
 
       // #1713: turn_end is THE terminal trigger. Finalize via the
-      // single terminal path (👍). Any prior intermediate states
-      // pending in the debounce window are flushed by `finalize()`
-      // before the terminal emoji emits.
-      finalizeStatusReaction(chatId, threadId, 'done')
+      // single terminal path. Any prior intermediate states pending in
+      // the debounce window are flushed by `finalize()` before the
+      // terminal emoji emits.
+      //
+      // #2527 — role-aware terminal honesty: a USER turn that ends without
+      // a delivered answer must NOT paint 👍 (the operator's "thumbs up so
+      // it feels like you're done" report). It finalizes to the gentle
+      // 'undelivered' terminal (😐) instead; the silent-end fallback below
+      // carries the apology text. system/cron turns and NO_REPLY/HEARTBEAT_OK
+      // turns (which return earlier) keep 👍 — their silence is legitimate.
+      let terminalReason = decideTerminalReason({
+        enabled: LIVENESS_TERMINAL_HONESTY,
+        role: turn.role,
+        finalAnswerDelivered: turn.finalAnswerDelivered,
+      })
+      // #2527 review note 1 — worker-hold carve-out: if the turn is STILL
+      // legitimately working at turn_end (a background sub-agent the parent
+      // dispatched is running on), don't prematurely paint 😐. Fall back to
+      // 'done' so the existing deferred-done path holds ✍️ until the worker
+      // completes (then 👍) — the worker-activity feed carries the progress.
+      // Only a turn that genuinely ended undelivered AND is not still working
+      // gets the honest 😐.
+      if (terminalReason === 'undelivered' && isLegitimatelyWorking(statusKey(chatId, threadId))) {
+        terminalReason = 'done'
+      }
+      if (terminalReason === 'undelivered') {
+        process.stderr.write(
+          `telegram gateway: WARN turn_no_reply — user turn ended with an ` +
+          `ambient ack but no delivered answer; painting 😐 not 👍 ` +
+          `chat=${chatId} thread=${threadId ?? '-'} turnId=${turn.turnId} (#2527)\n`,
+        )
+      }
+      finalizeStatusReaction(chatId, threadId, terminalReason)
       {
         const sKey = streamKey(chatId, threadId)
         const turnDurationMs = turn.startedAt > 0 ? Date.now() - turn.startedAt : 0
@@ -11773,6 +13004,9 @@ async function handleInboundCoalesced(
   //   - msgId present (always true for `bot.on('message:*')` paths but
   //     defensive against future routers that might call this without one).
   maybeEarlyAckReaction(ctx, from)
+  // #2527 — if this lands mid-turn, the user is asking "what's happening?";
+  // fire the liveness floor immediately (DM + supergroup alike).
+  maybePokeFloorForMidTurnInbound(ctx, from)
 
   const key = inboundCoalesceKey(
     String(ctx.chat!.id),
@@ -11801,6 +13035,14 @@ function maybeEarlyAckReaction(ctx: Context, from: NonNullable<Context['from']>)
   const msgId = ctx.message?.message_id
   if (msgId == null) return
   const chatType = ctx.chat?.type
+  // Intentionally DM-only (#2527 surface-parity note): pre-acking a GROUP
+  // message risks reacting to one the full gate (requireMention / topic
+  // scoping) would later DROP. The SUBSTANTIVE liveness parity a supergroup
+  // needs — the mid-turn floor and the role-aware terminal reaction — is
+  // surface-agnostic (keyed on statusKey + loop role, no chat-type branch),
+  // so a forum topic gets identical never-silent guarantees without this
+  // sub-second 👀 optimisation. See `maybePokeFloorForMidTurnInbound` for
+  // the surface-agnostic "Status?" short-circuit.
   if (chatType !== 'private') return
   const chatId = String(ctx.chat!.id)
   const threadId = ctx.message?.is_topic_message ? ctx.message.message_thread_id : undefined
@@ -11810,6 +13052,9 @@ function maybeEarlyAckReaction(ctx: Context, from: NonNullable<Context['from']>)
   void bot.api.setMessageReaction(chatId, msgId, [
     { type: 'emoji', emoji: '👀' as ReactionTypeEmoji['emoji'] },
   ]).catch(() => {})
+  // #2527: log the early-ack fire so operators can see how often the
+  // fast pre-coalesce DM path triggers vs. the controller path.
+  logStreamingEvent({ kind: 'early_ack_reaction', chatId, messageId: msgId, emoji: '👀' })
   // #553 PR 3: also fire the native "typing…" indicator. Bridges the
   // visual gap between the early-ack 👀 reaction and the first real
   // model text. No fake content — Telegram clients render this natively
@@ -11818,6 +13063,26 @@ function maybeEarlyAckReaction(ctx: Context, from: NonNullable<Context['from']>)
   void bot.api.sendChatAction(chatId, 'typing').catch(() => {})
 }
 
+/**
+ * #2527 — "Status?" short-circuit. A message arriving DURING an active turn
+ * (the user explicitly asking what's happening) fires the mid-turn liveness
+ * floor immediately, bypassing the timer/working gates. Surface-agnostic:
+ * works identically in a DM and a forum-supergroup topic (keyed on statusKey).
+ * Idempotent per turn (the floor's fire-once latch) and kill-switch-gated.
+ */
+function maybePokeFloorForMidTurnInbound(ctx: Context, from: NonNullable<Context['from']>): void {
+  const rawChatId = ctx.chat?.id
+  if (rawChatId == null) return
+  const chatId = String(rawChatId)
+  const threadId = ctx.message?.is_topic_message ? ctx.message.message_thread_id : undefined
+  const key = statusKey(chatId, threadId)
+  // Only mid-turn: a turn must already be in flight for this (chat, thread).
+  if (!activeTurnStartedAt.has(key)) return
+  const access = loadAccess()
+  if (!access.allowFrom.includes(String(from.id))) return
+  silencePoke.pokeFloorNow(key, Date.now())
+}
+
 async function handleInbound(
   ctx: Context,
   text: string,
@@ -12723,17 +13988,42 @@ async function handleInbound(
         if (!chatAvailableReactions.has(chat_id)) {
           probeAvailableReactions(chat_id)
         }
+        // #2527: use inbound msgId as a stable per-turn reaction identifier.
+        // The controller is created before currentTurn.turnId is assigned
+        // (that happens in handleSessionEvent's enqueue branch), so we capture
+        // msgId here and use it as the reaction-session token in log events.
+        const ctrlTurnToken = `${chat_id}:${msgId}`
         const ctrl = new StatusReactionController(async (emoji) => {
           await bot.api.setMessageReaction(chat_id, msgId, [
             { type: 'emoji', emoji: emoji as ReactionTypeEmoji['emoji'] },
           ])
           // #203: every status-reaction transition is a user-visible signal.
           signalTracker.noteSignal(key, Date.now())
-        }, allowedReactions)
+        }, allowedReactions, {
+          // #2527: emit a structured transition event on each emoji change so
+          // the reaction lifecycle is visible in streaming-metrics logs. Also
+          // increment the per-key counter for the turn_no_reply_warn metric.
+          onTransition: (emoji) => {
+            reactionTransitionCounts.set(key, (reactionTransitionCounts.get(key) ?? 0) + 1)
+            logStreamingEvent({
+              kind: 'status_reaction_transition',
+              chatId: chat_id,
+              turnId: ctrlTurnToken,
+              emoji,
+            })
+          },
+        })
         activeStatusReactions.set(key, ctrl)
         activeReactionMsgIds.set(key, { chatId: chat_id, messageId: msgId })
         activeTurnStartedAt.set(key, Date.now())
         progressUpdateTurnCount.set(key, 0)  // Reset turn counter
+        // #2527: log controller install so the lifecycle start is observable.
+        logStreamingEvent({
+          kind: 'status_reaction_install',
+          chatId: chat_id,
+          turnId: ctrlTurnToken,
+          messageId: msgId,
+        })
         ctrl.setQueued()
         // #203: time-to-ack metric — setQueued() triggers the initial 👀 reaction
         // asynchronously through the controller chain.
@@ -13300,6 +14590,18 @@ function getCommandArgs(ctx: Context): string {
   return m ? m[1].trim() : ''
 }
 
+/**
+ * True when a slash command's argument string carries a trailing `demo`
+ * token — the per-command PII-mask modifier for screen recordings
+ * (`/usage demo`, `/auth demo`, `/status demo`, `/whoami demo`). Matches
+ * `demo` as the last whitespace-delimited token, case-insensitively, so
+ * `/auth show demo` and `/usage demo` both flip the flag while a label
+ * literally named `demo-foo` does not.
+ */
+function hasDemoFlag(args: string): boolean {
+  return /(?:^|\s)demo$/i.test(args.trim())
+}
+
 /** Validate that a string looks like a safe agent/resource name.
  *  Agent names should be alphanumeric with hyphens/underscores only.
  *  This prevents shell metacharacter injection even though both exec
@@ -13790,6 +15092,9 @@ function notifyDetachedFailure(
         lockedBot.api.sendMessage(chatId, text, {
           parse_mode: 'HTML',
           link_preview_options: { is_disabled: true },
+          // Detached restart/update child-failure notice — status, not
+          // the user's answer. Silence the open ping.
+          disable_notification: true,
           ...(threadId != null ? { message_thread_id: threadId } : {}),
         }),
       {
@@ -14577,9 +15882,10 @@ bot.command('status', async ctx => {
   const { access, senderId } = gated
   const from = ctx.from!
   if (access.allowFrom.includes(senderId)) {
+    const demo = hasDemoFlag(getCommandArgs(ctx))
     const userTag = from.username ? `@${from.username}` : senderId
     const meta = await buildAgentMetadata(getMyAgentName())
-    await ctx.reply(buildStatusPairedText({ user: userTag, meta }), { parse_mode: 'HTML' })
+    await ctx.reply(buildStatusPairedText({ user: userTag, meta, demo }), { parse_mode: 'HTML' })
     return
   }
   for (const [code, p] of Object.entries(access.pending)) {
@@ -14712,10 +16018,12 @@ bot.command('model', async ctx => {
 // `/effort` — show or switch the reasoning effort for the live session.
 // The effort sibling of `/model`: bare form renders a five-button menu
 // (low/medium/high/xhigh/max, the live level ✅), a typed form
-// `/effort <level>` sets it directly. Both ride the allowlisted inject
-// primitive (claude's own `/effort` REPL command), session-scoped — boot
-// re-pins the configured default via start.sh's `--effort`. Implementation
-// in effort-command.ts so it's unit-testable without booting the bot.
+// `/effort <level>` sets it directly. Both ride the dedicated `applyEffort`
+// driver (claude's own `/effort` REPL command, with the confirmation modal
+// answered so the pane never wedges — NOT the bare inject primitive, which
+// is blocklisted for `/effort` since #2471), session-scoped — boot re-pins
+// the configured default via start.sh's `--effort`. Implementation in
+// effort-command.ts so it's unit-testable without booting the bot.
 function buildEffortDeps(): EffortCommandDeps {
   return {
     applyEffort: (agent, level) => applyEffort(agent, level),
@@ -14799,6 +16107,9 @@ bot.command('restart', async ctx => {
         (tid) =>
           lockedBot.api.sendMessage(chatId, ackText, {
             parse_mode: 'HTML', link_preview_options: { is_disabled: true },
+            // Restart acknowledgement is a status notice — silence the
+            // open ping (the "restarted — ready" follow-up is what matters).
+            disable_notification: true,
             ...(tid != null ? { message_thread_id: tid } : {}),
           }),
         { threadId, chat_id: chatId, verb: 'restart.ack' },
@@ -14940,6 +16251,9 @@ async function handleNewCommand(ctx: Context): Promise<void> {
       (tid) =>
         lockedBot.api.sendMessage(chatId, ackText, {
           parse_mode: 'HTML', link_preview_options: { is_disabled: true },
+          // /new /reset acknowledgement is a status notice — silence the
+          // open ping (the post-restart greeting card is what matters).
+          disable_notification: true,
           ...(tid != null ? { message_thread_id: tid } : {}),
         }),
       { threadId, chat_id: chatId, verb: 'new-or-reset.ack' },
@@ -15142,6 +16456,9 @@ bot.command('update', async ctx => {
         lockedBot.api.sendMessage(chatId, ackText, {
           parse_mode: 'HTML',
           link_preview_options: { is_disabled: true },
+          // "update started" acknowledgement is a status notice — silence
+          // the open ping (the post-restart greeting card is what matters).
+          disable_notification: true,
           ...(tid != null ? { message_thread_id: tid } : {}),
         }),
       { threadId, chat_id: chatId, verb: 'update.ack' },
@@ -15603,6 +16920,36 @@ const fleetFallbackGate = createFleetFallbackGate({
   brokerReachable: isAuthBrokerSocketReachable,
 })
 
+/**
+ * Resume-after-swap gate (auth-failover-stall fix). Owns the single-flight +
+ * staleness decision for re-running the turn a mid-turn 429 killed. See
+ * fleet-fallback-resume.ts. Wired into doFireFleetAutoFallback below: on a
+ * 'switched' outcome we restart so the boot-resume path replays the dead turn
+ * on the freshly-active account. 3h staleness mirrors the boot-resume
+ * RESUME_MAX_AGE_MS failsafe (gateway boot path); single-flight stops a 429
+ * storm from loop-restarting the agent.
+ */
+const fleetFallbackResumeGate = createFleetFallbackResumeGate({
+  maxAgeMs: (() => {
+    const v = Number(process.env.SWITCHROOM_RESUME_MAX_AGE_MS)
+    return Number.isFinite(v) && v > 0 ? v : 10_800_000 // 3h, matches boot-resume
+  })(),
+})
+
+/**
+ * The start time (epoch-ms) of the most-recently-started active turn — the
+ * staleness signal for the resume gate. `activeTurnStartedAt` is stamped on
+ * inbound receipt (see its declaration), so the newest entry is the turn the
+ * 429 just killed. Returns null when no turn is tracked (then the resume gate
+ * defers staleness to the boot-resume 3h failsafe). */
+function newestActiveTurnStartedAtMs(): number | null {
+  let newest: number | null = null
+  for (const ms of activeTurnStartedAt.values()) {
+    if (newest == null || ms > newest) newest = ms
+  }
+  return newest
+}
+
 function wouldFireFleetAutoFallback(): boolean {
   return fleetFallbackGate.wouldFire()
 }
@@ -15658,6 +17005,17 @@ async function fireFleetAutoFallback(triggerAgent: string, untilMs?: number): Pr
  */
 let fallbackFailureNoticeState: FallbackFailureNoticeState = { lastSentAtMs: 0 }
 
+/**
+ * Bug 2 — per-gateway cooldown for the "All accounts blocked" card. The
+ * all-blocked outcome is a no-op swap (doFireFleetAutoFallback returns false),
+ * so the fleetFallbackGate dedup window never arms for it, and the ~60s
+ * quota_wall_detected re-trigger would otherwise re-broadcast the identical card
+ * every minute for the life of the wall. This bounds it to one card per window.
+ * Reset on a successful swap so a fresh all-blocked after a recovery (a real new
+ * transition) is not stale-suppressed.
+ */
+let fallbackAllBlockedNoticeState: FallbackAllBlockedNoticeState = { lastSentAtMs: 0 }
+
 function broadcastFleetFallbackFailure(triggerAgent: string, reason: string): void {
   if (process.env.SWITCHROOM_FLEET_FALLBACK_FAILURE_NOTICE === '0') return
   // Notice-level cooldown (30 min, per gateway). The fleetFallbackGate's
@@ -15743,19 +17101,66 @@ async function doFireFleetAutoFallback(triggerAgent: string, untilMs?: number):
         (outcome.kind === 'switched' ? ` old=${outcome.oldLabel} new=${outcome.newLabel}` : '') +
         '\n',
     )
+    // Bug 2 — the all-blocked card is a no-op outcome, so the gate's dedup
+    // window never arms for it and the ~60s quota_wall_detected re-trigger would
+    // re-broadcast the identical card every minute. Gate it behind a per-gateway
+    // cooldown; a successful swap resets the window so a later (genuinely new)
+    // all-blocked still emits promptly.
+    if (outcome.kind === 'switched') {
+      fallbackAllBlockedNoticeState = { lastSentAtMs: 0 }
+    } else if (outcome.kind === 'all-blocked') {
+      const verdict = evaluateAllBlockedNotice(fallbackAllBlockedNoticeState, Date.now())
+      if (!verdict.send) {
+        process.stderr.write(
+          `telegram gateway: [fleet-fallback] all-blocked card suppressed (cooldown) agent=${triggerAgent}\n`,
+        )
+        return false
+      }
+      fallbackAllBlockedNoticeState = verdict.next
+    }
     // Post the announcement to every authorized chat. Mirrors the
     // operator-event broadcast pattern (line ~2290) — DM-only opts
     // (no message_thread_id) so THREAD_NOT_FOUND can't fire here;
     // wrap in swallowingApiCall anyway per the codebase rule.
     const access = loadAccess()
     if (access.allowFrom.length === 0) return outcome.kind === 'switched'
-    const opts = { parse_mode: 'HTML' as const }
+    // Account-switch / all-blocked announcement is a system status notice,
+    // not the user's answer — silence the open ping.
+    const opts = { parse_mode: 'HTML' as const, disable_notification: true }
     for (const chat_id of access.allowFrom) {
       void swallowingApiCall(
         () => bot.api.sendMessage(chat_id, outcome.announcement, opts),
         { chat_id, verb: 'fleet-fallback:notify' },
       )
     }
+    // ── Resume the dead turn (auth-failover-stall fix) ──────────────────────
+    // A mid-turn 429 killed a turn; the swap above moved the fleet to a healthy
+    // account, but that only takes effect on the NEXT claude invocation. Re-run
+    // the dead turn via triggerSelfRestart: the boot-resume path (gateway boot,
+    // findLatestTurnIfInterrupted → buildResumeInterruptedInbound) replays the
+    // LATEST interrupted turn on the freshly-active account. We restart rather
+    // than redeliver because the failed inbound was already DELIVERED (the turn
+    // started, then the model 429'd) so it is NOT in pendingInboundBuffer —
+    // redeliverBufferedInbound would find nothing. Guards live in
+    // fleetFallbackResumeGate: single-flight (a 429 storm cannot loop-restart)
+    // + 3h staleness (an ancient interrupted turn is not resurrected). Only
+    // reached on 'switched'; all-blocked / no-op outcomes never get here, so the
+    // all-blocked cooldown path above is preserved.
+    if (outcome.kind === 'switched') {
+      const verdict = fleetFallbackResumeGate.decide(newestActiveTurnStartedAtMs())
+      if (verdict === 'resume') {
+        const selfAgent = process.env.SWITCHROOM_AGENT_NAME ?? triggerAgent
+        process.stderr.write(
+          `telegram gateway: [fleet-fallback] resuming dead turn via self-restart ` +
+          `agent=${selfAgent} (swap ${outcome.oldLabel}→${outcome.newLabel})\n`,
+        )
+        triggerSelfRestart(selfAgent, 'fleet-fallback-resume')
+      } else {
+        process.stderr.write(
+          `telegram gateway: [fleet-fallback] resume suppressed (${verdict}) agent=${triggerAgent}\n`,
+        )
+      }
+    }
     return outcome.kind === 'switched'
   } catch (err) {
     process.stderr.write(
@@ -15815,6 +17220,9 @@ async function runCreditWatch(): Promise<void> {
         bot.api.sendMessage(chat_id, decision.message, {
           parse_mode: 'HTML',
           link_preview_options: { is_disabled: true },
+          // Credit/quota warning is a system status notice — silence the
+          // open ping (the user isn't waiting to tap anything).
+          disable_notification: true,
         }),
       { chat_id, verb: 'credit-watch.notify' },
     )
@@ -15928,6 +17336,10 @@ async function runQuotaWatch(opts: { bootTick?: boolean } = {}): Promise<void> {
       accounts: listStateData.accounts,
       prev: fleetPrev,
       now,
+      // #2478: the same staleness ceiling the per-account loop uses. Gates the
+      // `entered` alert behind live corroboration so a probe blackout's stale
+      // marks can't false-fire 🔴 All accounts exhausted.
+      tuning,
     })
     if (fleetDecision.kind === 'notify') {
       for (const chat_id of access.allowFrom) {
@@ -15950,6 +17362,8 @@ async function runQuotaWatch(opts: { bootTick?: boolean } = {}): Promise<void> {
             bot.api.sendMessage(chat_id, fleetDecision.message, {
               parse_mode: 'HTML',
               link_preview_options: { is_disabled: true },
+              // Quota status notice — silence the open ping.
+              disable_notification: true,
             }),
           { chat_id, verb: 'quota-watch.fleet-all-exhausted' },
         )
@@ -16022,11 +17436,21 @@ async function runQuotaWatch(opts: { bootTick?: boolean } = {}): Promise<void> {
   // numbers for the notification message bodies. One batched RPC for all
   // crossing accounts (typically 1, rarely 2+).
   const crossingLabels = pendingTransitions.map(t => t.accountLabel)
-  let freshProbeMap = new Map<string, Awaited<ReturnType<typeof brokerClient.probeQuota>>['results'][number]['result']>()
+  // #2495 BLOCKER fix — store the FULL entry (result + `served` tag), not just
+  // `entry.result`. The corroboration gate below needs `served` to tell a true
+  // live probe apart from a failed-probe cache fallback (which is `ok:true`
+  // but `served:"cache"` — vacuous corroboration).
+  let freshProbeMap = new Map<string, Awaited<ReturnType<typeof brokerClient.probeQuota>>['results'][number]>()
   try {
-    const probeData = await brokerClient.probeQuota(crossingLabels, 8000)
+    // #2495 Change 3 — forceLive bypasses the broker's probe-on-open TTL so the
+    // DECISION to alarm is corroborated by a TRUE live probe, never a cache hit.
+    // Only the transition-to-alarm pays for this; steady-state polls stay on the
+    // cheap cached listState read (no probe). Honors the existing fleet/consumer
+    // probe knobs upstream — this re-evaluation never fires without a detected
+    // transition.
+    const probeData = await brokerClient.probeQuota(crossingLabels, 8000, true)
     for (const entry of probeData.results) {
-      freshProbeMap.set(entry.label, entry.result)
+      freshProbeMap.set(entry.label, entry)
     }
   } catch (err) {
     process.stderr.write(`telegram gateway: quota-watch: probe for crossing accounts failed: ${err}\n`)
@@ -16058,17 +17482,25 @@ async function runQuotaWatch(opts: { bootTick?: boolean } = {}): Promise<void> {
   for (const { accountLabel, snapIndex, decision } of pendingTransitions) {
     // Re-evaluate with fresh probe data to get an accurate message body.
     // If the fresh probe succeeded, replace the snap's quota with live data.
-    const freshResult = freshProbeMap.get(accountLabel)
+    const freshEntry = freshProbeMap.get(accountLabel)
     let enrichedDecision = decision
     // pendingTransitions only ever holds notify decisions (pushed under
     // `decision.kind !== 'skip'` / `!== 'reconcile'`). Narrow explicitly so
     // `decision.transition` type-checks below; this continue never fires
     // at runtime.
     if (decision.kind !== 'notify') continue
-    if (freshResult && freshResult.ok && snapIndex >= 0) {
+    // #2495 BLOCKER fix — only a GENUINE live probe corroborates the alarm. A
+    // forceLive entry that is `ok:true` but `served:"cache"` (the broker's
+    // failed-probe cache fallback) is NOT corroboration: the upstream probe
+    // failed, so we have no live confirmation that the throttling crossing is
+    // real right now. Treat it exactly like a probe failure → fall through to
+    // the defer branch below (state untouched, re-evaluated next tick). This
+    // also guarantees the "Live-probe corroborated (#2495)" footnote is only
+    // ever stamped on a real live probe.
+    if (isLiveCorroboration(freshEntry) && freshEntry!.result.ok && snapIndex >= 0) {
       // Live numbers replace the cache — and capturedAtMs is cleared so the
       // staleness gate never misfires on data we JUST probed.
-      const enrichedSnap = { ...snapshots[snapIndex]!, quota: freshResult.data, capturedAtMs: undefined }
+      const enrichedSnap = { ...snapshots[snapIndex]!, quota: freshEntry!.result.data, capturedAtMs: undefined }
       const prev = watchState[accountLabel] ?? emptyAccountState()
       const re = evaluateQuotaWatchAccount({ agentName, snap: enrichedSnap, prev, now, bootTick, tuning })
       // If the fresh probe still shows the same transition, use the
@@ -16153,6 +17585,8 @@ async function runQuotaWatch(opts: { bootTick?: boolean } = {}): Promise<void> {
           bot.api.sendMessage(chat_id, message, {
             parse_mode: 'HTML',
             link_preview_options: { is_disabled: true },
+            // Quota throttling status notice — silence the open ping.
+            disable_notification: true,
           }),
         { chat_id, verb: 'quota-watch.notify' },
       )
@@ -16424,7 +17858,13 @@ bot.command("auth", async ctx => {
     }
     return
   }
-  const text = ctx.message?.text ?? ""
+  const rawText = ctx.message?.text ?? ""
+  // `/auth demo` (and `/auth show demo` / `/auth list demo`) — the trailing
+  // `demo` token masks account-email labels on the default dashboard view for
+  // screen recordings. Strip it before parsing so `demo` isn't mistaken for a
+  // verb/agent argument; it's honored only on the show/list path downstream.
+  const authDemo = hasDemoFlag(getCommandArgs(ctx))
+  const text = authDemo ? rawText.replace(/(\s+)demo\s*$/i, "") : rawText
   const parsed = parseAuthCommand(text)
   if (!parsed) return
   const currentAgent = getMyAgentName()
@@ -16486,11 +17926,12 @@ bot.command("auth", async ctx => {
       return
     }
     try {
-      const { loginUrl, scratchDir, child } = await startAccountAuthSession(parsed.label)
+      const { loginUrl, scratchDir, tmuxSocket, tmuxSession } = await startAccountAuthSession(parsed.label)
       pendingAuthAddFlows.set(authAddKey, {
         label: parsed.label,
         scratchDir,
-        child,
+        tmuxSocket,
+        tmuxSession,
         startedAt: Date.now(),
       })
       await switchroomReply(
@@ -16522,6 +17963,7 @@ bot.command("auth", async ctx => {
     isAdmin,
     client,
     chatId,
+    demo: authDemo,
     // Format 2 enricher — live quota probe via the broker (#1336).
     // Pre-broker this read `~/.switchroom/accounts/<label>/credentials.json`
     // off the agent's HOME, which post-RFC-H is never populated (broker
@@ -16533,18 +17975,27 @@ bot.command("auth", async ctx => {
     liveQuotas: async (accounts) => {
       try {
         const { results } = await client.probeQuota(accounts.map((a) => a.label))
+        // #2495 Change 2 — the broker tags each result `served:"live"|"cache"`
+        // (TTL hit or failed-probe fallback). When ANY account was served from
+        // cache, surface the OLDEST snapshot's capturedAt so the card stamps
+        // "⚠ cached Nm ago" instead of a false live stamp.
+        let staleCachedAtMs: number | undefined
         // Preserve input order (broker also preserves it, but be defensive).
-        return accounts.map((a) => {
+        const quotas = accounts.map((a) => {
           const hit = results.find((r) => r.label === a.label)
           if (!hit) return { ok: false as const, reason: "broker returned no result for account" }
+          if (hit.served === 'cache' && hit.capturedAt != null) {
+            staleCachedAtMs = staleCachedAtMs == null ? hit.capturedAt : Math.min(staleCachedAtMs, hit.capturedAt)
+          }
           return hit.result
         })
+        return { quotas, staleCachedAtMs }
       } catch (err) {
         // Surface a uniform per-account failure so the dashboard renders
         // gracefully (label badge stays UNKNOWN) instead of falling back
         // to the legacy table.
         const reason = `broker probe-quota failed: ${(err as Error)?.message ?? String(err)}`
-        return accounts.map(() => ({ ok: false as const, reason }))
+        return { quotas: accounts.map(() => ({ ok: false as const, reason })) }
       }
     },
     tz: process.env.SWITCHROOM_TIMEZONE ?? process.env.TZ,
@@ -18899,6 +20350,7 @@ bot.command('issues', async ctx => {
 
 bot.command('usage', async ctx => {
   if (!isAuthorizedSender(ctx)) return
+  const demo = hasDemoFlag(getCommandArgs(ctx))
   // Format 2 path: enumerate every account in the broker's known set,
   // probe live quota in parallel, render the health-grouped snapshot.
   // Falls back to the legacy single-agent shape when the broker is
@@ -18911,9 +20363,17 @@ bot.command('usage', async ctx => {
       const state = await client.listState()
       if (state.accounts.length > 0) {
         // Broker-routed probe (#1336) — see gateway.ts:8910 for diagnosis.
+        // #2495 Change 2 — the broker applies a probe-on-open TTL + single-
+        // flight; a TTL-hit or failed-probe fallback is tagged served:"cache",
+        // which we surface as a "⚠ cached Nm ago" footer instead of a false
+        // live stamp.
         const probeResp = await client.probeQuota(state.accounts.map((a) => a.label)).catch(() => ({ results: [] }))
+        let staleCachedAtMs: number | undefined
         const quotas = state.accounts.map((a) => {
           const hit = probeResp.results.find((r) => r.label === a.label)
+          if (hit?.served === 'cache' && hit.capturedAt != null) {
+            staleCachedAtMs = staleCachedAtMs == null ? hit.capturedAt : Math.min(staleCachedAtMs, hit.capturedAt)
+          }
           return hit?.result ?? { ok: false as const, reason: 'broker returned no result for account' }
         })
         const { renderAuthSnapshotFormat2, buildSnapshotsFromState } = await import(
@@ -18924,7 +20384,8 @@ bot.command('usage', async ctx => {
         const text = renderAuthSnapshotFormat2(snapshots, {
           tz,
           now: new Date(),
-          liveProbedAtMs: Date.now(),
+          demo,
+          ...(staleCachedAtMs != null ? { staleCachedAtMs } : { liveProbedAtMs: Date.now() }),
         })
         await switchroomReply(ctx, text, { html: true })
         return
@@ -19091,13 +20552,14 @@ bot.command('version', async ctx => {
 // see at a glance what this agent is authorized for.
 bot.command('whoami', async ctx => {
   if (!isAuthorizedSender(ctx)) return
+  const demo = hasDemoFlag(getCommandArgs(ctx))
   try {
     let raw: string
     try { raw = switchroomExecCombined(['config', 'whoami'], 10000) }
     catch (err: unknown) { raw = (err as any).stdout ?? (err as any).message ?? 'whoami failed' }
     const trimmed = stripAnsi(raw).trim()
     let card: string
-    try { card = formatWhoamiCard(JSON.parse(trimmed.split('\n').pop() ?? trimmed)) }
+    try { card = formatWhoamiCard(JSON.parse(trimmed.split('\n').pop() ?? trimmed), demo) }
     catch { card = preBlock(formatSwitchroomOutput(trimmed || 'whoami: no output')) }
     await switchroomReply(ctx, card, { html: true })
   } catch (err: unknown) {
@@ -19105,14 +20567,17 @@ bot.command('whoami', async ctx => {
   }
 })
 
-/** Compact HTML card from the `config whoami` JSON view. Names/booleans only. */
+/** Compact HTML card from the `config whoami` JSON view. Names/booleans only.
+ *  `demo` (the `/whoami demo` suffix) masks the vault key NAMES via maskVaultKey
+ *  for screen recordings — agent/MCP/model/skills topology is left untouched
+ *  (out of scope). Off by default. */
 function formatWhoamiCard(v: {
   name?: string; persona?: string | null; model?: string | null; tier?: string;
   tools?: { allow?: string[]; deny?: string[] }; mcpServers?: string[]; skills?: string[];
   vault?: { key: string; readable: boolean }[];
   powers?: { admin?: boolean; root?: boolean; configEdit?: boolean; crossAgentHostVerbs?: boolean };
   scheduleCount?: number; memoryBackend?: string | null;
-}): string {
+}, demo = false): string {
   const esc = escapeHtmlForTg
   const yn = (b?: boolean) => (b ? '✓' : '✗')
   const lines: string[] = []
@@ -19125,7 +20590,7 @@ function formatWhoamiCard(v: {
   if ((v.mcpServers ?? []).length) lines.push(`MCP: ${esc(v.mcpServers!.join(', '))}`)
   if ((v.skills ?? []).length) lines.push(`Skills: ${esc(v.skills!.join(', '))}`)
   if ((v.vault ?? []).length) {
-    lines.push(`Vault keys (names only): ${v.vault!.map(k => `${esc(k.key)} ${yn(k.readable)}`).join(', ')}`)
+    lines.push(`Vault keys (names only): ${v.vault!.map(k => `${esc(demo ? maskVaultKey(k.key) : k.key)} ${yn(k.readable)}`).join(', ')}`)
   }
   const p = v.powers ?? {}
   lines.push(`Powers: admin ${yn(p.admin)} · root ${yn(p.root)} · config-edit ${yn(p.configEdit)} · cross-agent verbs ${yn(p.crossAgentHostVerbs)}`)
@@ -19350,6 +20815,8 @@ bot.on('callback_query:data', async ctx => {
             await robustApiCall(() =>
               bot.api.editMessageText(args.chatId, args.messageId, args.text, {
                 parse_mode: 'HTML',
+                // Resolved on tap — strip the keyboard so it can't be re-tapped.
+                ...(args.stripKeyboard ? { reply_markup: { inline_keyboard: [] } } : {}),
               }),
             )
           } catch {
@@ -19359,6 +20826,9 @@ bot.on('callback_query:data', async ctx => {
         log: (m) =>
           process.stderr.write(`telegram gateway: config-approval cb — ${m}\n`),
       },
+      // Verify the per-card epoch from the callback_data against the live
+      // pending entry — a stale tap (mismatched epoch) is rejected.
+      parsed.epoch,
     )
     await ctx.answerCallbackQuery({
       text: resolved
@@ -21136,6 +22606,11 @@ async function shutdown(signal: string): Promise<void> {
   subagentWatcher?.stop()
   subagentWatcher = null
 
+  // Worker-activity feed runs an internal heartbeat interval; stop it so no
+  // re-render fires during drain (mirrors subagentWatcher above).
+  workerActivityFeed?.stop()
+  workerActivityFeed = null
+
   // Issues watcher polls issues.jsonl on a setInterval (default 2s) and
   // edits the issues card on every tick. Without an explicit stop() the
   // poll keeps firing for the lifetime of the process and accumulates
@@ -21869,7 +23344,8 @@ void (async () => {
             // or the turn ended while it kept running — extended autonomous
             // work) is surfaced via the worker feed instead of vanishing.
             const orphanStatusEnabled = isOrphanSubagentStatusEnabled(process.env.SWITCHROOM_ORPHAN_SUBAGENT_STATUS)
-            const workerActivityFeed = createWorkerActivityFeed({
+            workerActivityFeed?.stop()
+            workerActivityFeed = createWorkerActivityFeed({
               bot: {
                 sendMessage: async (cid, text, sendOpts) => {
                   const sent = await robustApiCall(
@@ -22063,9 +23539,18 @@ void (async () => {
                       const rendered = composeTurnActivity(turn)
                       if (rendered != null) {
                         turn.activityPendingRender = rendered
-                        if (turn.activityInFlight == null) {
-                          turn.activityInFlight = drainActivitySummary(turn)
+                        // PR-4a: routed through the emission-authority façade
+                        // (no-op delegate). Producer made explicit ('tool' — the
+                        // drain default this foreground sub-agent render used).
+                        const ea = emissionAuthorityFor(turn)
+                        // PR-4d: route through the centralized card-drain gate.
+                        cardDrainGate(turn, ea, () => {
+                        if (ea.mayDrain(turn)) {
+                          ea.openOrEditCard('tool', () => {
+                            turn.activityInFlight = drainActivitySummary(turn, 'tool')
+                          })
                         }
+                        })
                       }
                     }
                     return
@@ -22083,7 +23568,7 @@ void (async () => {
                       orphanStatusEnabled,
                     }) === 'worker-feed'
                   ) {
-                    void workerActivityFeed.finish(agentId, {
+                    void workerActivityFeed?.finish(agentId, {
                       description: dispatch.feedDescription,
                       lastTool: null,
                       toolCount,
@@ -22100,7 +23585,7 @@ void (async () => {
                 // 'orphan' is a stale boot row, not a fresh completion — map
                 // it to 'done' so an already-posted message still finalizes.
                 if (workerFeedEnabled) {
-                  void workerActivityFeed.finish(agentId, {
+                  void workerActivityFeed?.finish(agentId, {
                     description: dispatch.feedDescription,
                     lastTool: null,
                     toolCount,
@@ -22235,7 +23720,7 @@ void (async () => {
                   })
                   if (surface === 'worker-feed') {
                     const origin = resolveSubagentOriginChat(agentId)
-                    void workerActivityFeed.update(
+                    void workerActivityFeed?.update(
                       agentId,
                       origin?.chatId || fleetChatId || (loadAccess().allowFrom[0] ?? ''),
                       {
@@ -22270,7 +23755,13 @@ void (async () => {
                   // feed (the foreground blindspot) — mirroring the
                   // main-turn activity feed, which surfaces both tool labels
                   // and prose.
-                  const child = (progressLine ?? latestSummary).trim().slice(0, 120)
+                  // Route through the SHARED clipNarrative so multi-line
+                  // narration first-line-collapses identically to the main
+                  // tier (the main path at showNarrativeStep already does
+                  // this). Previously this inlined `.trim().slice(0, 120)`
+                  // omitted the first-line collapse, so a multi-line
+                  // narrative rendered DIFFERENTLY here than on the main feed.
+                  const child = clipNarrative(progressLine ?? latestSummary)
                   if (child.length === 0) return
                   let narrative = turn.foregroundSubAgents.get(agentId)
                   if (narrative == null) {
@@ -22288,13 +23779,70 @@ void (async () => {
                   const rendered = composeTurnActivity(turn)
                   if (rendered != null) {
                     turn.activityPendingRender = rendered
-                    if (turn.activityInFlight == null) {
-                      turn.activityInFlight = drainActivitySummary(turn)
+                    // PR-4a: routed through the emission-authority façade (no-op
+                    // delegate). Producer made explicit ('tool' — the drain
+                    // default this foreground sub-agent render used).
+                    const ea = emissionAuthorityFor(turn)
+                    // PR-4d: route through the centralized card-drain gate.
+                    cardDrainGate(turn, ea, () => {
+                    if (ea.mayDrain(turn)) {
+                      ea.openOrEditCard('tool', () => {
+                        turn.activityInFlight = drainActivitySummary(turn, 'tool')
+                      })
+                    }
+                    })
+                    // A foreground sub-agent's nested activity IS user-visible
+                    // production — count it so the silence-poke clock resets,
+                    // exactly like the parent activity-render path (10665). Without
+                    // this, a long tools-only foreground sub-agent (no prose) lets
+                    // the 300s framework fallback (and the #2527 mid-turn floor)
+                    // measure silence against a turn that is visibly working,
+                    // risking a premature tear-down / unwanted liveness beat.
+                    // PR-4e — keyed liveness under the flag (a foreground
+                    // sub-agent's nested render for topic A is A's production).
+                    // Flag-OFF keeps the literal `currentTurn === turn`; flag-ON
+                    // resolves A by its own key.
+                    if (
+                      SILENCE_LIVENESS_PRODUCTION &&
+                      (EMISSION_AUTHORITY_ENABLED ? turnLiveForItsTopic(turn) : currentTurn === turn)
+                    ) {
+                      silencePoke.noteProduction(statusKey(turn.sessionChatId, turn.sessionThreadId), Date.now())
                     }
                   }
                   return
                 }
 
+                // Fix 2 (post-answer background-agent liveness): when the
+                // watcher surfaces a new step for a background worker, update
+                // the current turn's `subagentActivityAt` timestamp IF the turn
+                // has already delivered its substantive answer. This signal is
+                // written HERE — NOT in the tool_label path — so the drop-guard
+                // (`shouldReopenFeedAfterAck` / finalAnswerSubstantive) cannot
+                // gate it. `feedHeartbeatTick`'s post-answer branch reads
+                // `subagentActivityAt` (not `lastToolLabelAt`, which is frozen
+                // after the answer) to decide whether to open a liveness card.
+                // Only stamp when the turn is alive AND post-answer: pre-answer
+                // activity is already surfaced by the normal tool-label feed.
+                //
+                // SCOPE — this is the IN-TURN-WINDOW surface only. The
+                // `feedHeartbeatTick` post-answer card is driven off `currentTurn`,
+                // which `endCurrentTurnAtomic` nulls at `turn_end`. A genuinely
+                // DECOUPLED background worker keeps running PAST the parent
+                // turn's teardown, so `currentTurn` is null when its later
+                // onProgress ticks arrive → this stamp is inert and the
+                // heartbeat is silent for that worker. That is BY DESIGN, not a
+                // gap: a decoupled worker's ongoing activity is surfaced by the
+                // dedicated, currentTurn-independent `workerActivityFeed` (the
+                // edit-in-place worker message, driven below at `workerFeedEnabled`
+                // and bounded by its own non-running/`finish` teardown). So the
+                // currentTurn card covers the brief post-answer/pre-teardown
+                // window; the worker feed covers everything after teardown. Both
+                // are proven in telegram-activity-visibility-integration.test.ts.
+                const stampTurn = currentTurn
+                if (stampTurn != null && stampTurn.finalAnswerEverDelivered) {
+                  stampTurn.subagentActivityAt = Date.now()
+                }
+
                 // #PR2 live worker-feed: when ON, the worker's live chat
                 // message owns the progress beat. Push a running cue and
                 // return BEFORE the legacy bucket relay so the same activity
@@ -22306,7 +23854,7 @@ void (async () => {
                 // is gone — see resolveSubagentOriginChat).
                 if (workerFeedEnabled) {
                   const origin = resolveSubagentOriginChat(agentId)
-                  void workerActivityFeed.update(
+                  void workerActivityFeed?.update(
                     agentId,
                     origin?.chatId || fleetChatId || (loadAccess().allowFrom[0] ?? ''),
                     {
@@ -22370,11 +23918,9 @@ void (async () => {
 
       // Lane state (post flash-decouple): VISIBLE only when the visible flag is
       // Lane state from the single-source-of-truth resolver: 'visible' (preview
-      // on), 'draft' (compose-box transport), or 'dormant' (the default: no
-      // preview, no draft — reply tool is the only message). The old label
-      // wrongly reported 'visible(draft-retired)' for the dormant default, which
-      // masked the flash regression.
-      process.stderr.write(`telegram gateway: answer-stream lane=${ANSWER_LANE.state} draftFn=${sendMessageDraftFn != null ? 'available' : 'off'} visible=${ANSWER_STREAM_VISIBLE_ENABLED} draftRetired=${DRAFT_ANSWER_LANE_RETIRED} grammy=${GRAMMY_VERSION}\n`)
+      // Lane state: 'visible' (opt-in preview) or 'dormant' (default: reply
+      // tool is the only message). The draft transport is permanently retired.
+      process.stderr.write(`telegram gateway: answer-stream lane=${ANSWER_LANE.state} visible=${ANSWER_STREAM_VISIBLE_ENABLED} grammy=${GRAMMY_VERSION}\n`)
       process.stderr.write(`telegram gateway: starting bot polling pid=${process.pid} agent=${process.env.SWITCHROOM_AGENT_NAME ?? '-'} stateDir=${STATE_DIR} historyEnabled=${HISTORY_ENABLED} streamMode=${process.env.SWITCHROOM_TG_STREAM_MODE ?? 'checklist'}\n`)
       runnerHandle = run(bot, {
         runner: {