@gajae-code/coding-agent 0.13.3 → 0.14.0

package/src/sdk/bus/telegram-daemon-cli.ts

@@ -4,6 +4,7 @@ import { logger, postmortem } from "@gajae-code/utils";
 import { YAML } from "bun";
 import { applyAtomicYamlPatches, setByPath } from "../../config/atomic-yaml-patch";
 import type { Settings } from "../../config/settings";
+import { isProcessIncarnation, processIncarnation } from "../broker/process-incarnation";
 import {
 	getNotificationConfig,
 	isProviderEffectivelyEnabled,
@@ -12,9 +13,11 @@ import {
 	parseNotificationSettingsSnapshot,
 } from "./config";
 import { daemonPaths, HEARTBEAT_TTL_MS } from "./daemon-paths";
+import { type NotificationDebrisSweepReport, sweepNotificationDebris } from "./notification-service";
 import {
 	type DaemonState,
 	FilesystemTopicRegistryCasAuthority,
+	hasSafeDaemonStateShape,
 	loadInstallationHostId,
 	markDaemonOwnerStopped,
 	readDaemonState,
@@ -41,6 +44,7 @@ export interface RunDaemonInternalDeps {
 	DaemonImpl?: TelegramDaemonConstructor;
 	processPid?: number;
 	pidAlive?: (pid: number) => boolean;
+	pidIncarnation?: (pid: number) => string | undefined;
 	/** Clock used by the ownership-progress watchdog; defaults to `Date.now`. */
 	now?: () => number;
 	/** Timer pair backing the ownership-progress watchdog; defaults to globals. */
@@ -50,6 +54,57 @@ export interface RunDaemonInternalDeps {
 	readDaemonState?: (settings: Settings) => Promise<DaemonState | undefined>;
 	/** Loads the verified machine-local identity; injectable so daemon tests do not touch the host. */
 	loadInstallationHostId?: () => Promise<string>;
+	/**
+	 * Startup hygiene sweep over the notifications dir; injectable so tests can
+	 * prove it is fired, is never awaited by startup, and that a rejection is
+	 * logged instead of failing the daemon.
+	 */
+	sweepNotificationDebris?: (input: { dir: string }) => Promise<NotificationDebrisSweepReport>;
+}
+
+type DaemonOwnerTuple = {
+	ownerId: string;
+	acquisitionId: string;
+	pid: number;
+	incarnation: string;
+};
+
+function readyOwnerTuple(state: DaemonState | undefined): DaemonOwnerTuple | undefined {
+	if (
+		!state ||
+		!hasSafeDaemonStateShape(state) ||
+		state.stoppedAt !== undefined ||
+		typeof state.ownerId !== "string" ||
+		state.ownerId.length === 0 ||
+		typeof state.acquisitionId !== "string" ||
+		state.acquisitionId.length === 0 ||
+		!Number.isSafeInteger(state.pid) ||
+		state.pid <= 0 ||
+		!isProcessIncarnation(state.incarnation)
+	)
+		return undefined;
+	return {
+		ownerId: state.ownerId,
+		acquisitionId: state.acquisitionId,
+		pid: state.pid,
+		incarnation: state.incarnation,
+	};
+}
+
+function ownerTuplesEqual(left: DaemonOwnerTuple, right: DaemonOwnerTuple): boolean {
+	return (
+		left.ownerId === right.ownerId &&
+		left.acquisitionId === right.acquisitionId &&
+		left.pid === right.pid &&
+		left.incarnation === right.incarnation
+	);
+}
+
+function daemonOwnerTuple(ownerId: string, deps: RunDaemonInternalDeps): DaemonOwnerTuple | undefined {
+	const pid = deps.processPid ?? process.pid;
+	const incarnation = (deps.pidIncarnation ?? processIncarnation)(pid);
+	if (!Number.isSafeInteger(pid) || pid <= 0 || !isProcessIncarnation(incarnation)) return undefined;
+	return { ownerId, acquisitionId: ownerId, pid, incarnation };
 }
 
 /** Ownership-watchdog cadence while the daemon process is running. */
@@ -236,6 +291,22 @@ export async function runDaemonInternal(argv: string[], deps: RunDaemonInternalD
 	const settings = await resolveDaemonSettings(resolvedAgentDir, deps);
 	const cfg = getNotificationConfig(settings);
 	if (!isProviderEffectivelyEnabled(cfg, "telegram") || !isTelegramComplete(cfg)) return;
+	// Startup hygiene: reclaim inert quarantine/staging debris left by crashed
+	// writers so the notifications dir cannot grow unboundedly and slow every
+	// later endpoint scan. Never awaited by startup; a rejection is logged and
+	// never fails the daemon, because hygiene must not own daemon availability.
+	void (deps.sweepNotificationDebris ?? sweepNotificationDebris)({
+		dir: daemonPaths(resolvedAgentDir).dir,
+	})
+		.then(report => {
+			// A resolved report can still carry operational failures; surfacing only
+			// a rejected promise would hide exactly the evidence the sweep records.
+			if (report.failures > 0 || report.scanFailed === true)
+				logger.warn(
+					`telegram-daemon: startup debris sweep incomplete: removed ${report.removed.length}, failed ${report.failures}${report.scanFailed ? ", scan failed" : ""}`,
+				);
+		})
+		.catch(error => logger.warn(`telegram-daemon: startup debris sweep failed: ${String(error)}`));
 	const installationHostId = await (deps.loadInstallationHostId ?? loadInstallationHostId)();
 	const topicRegistryAuthority = new FilesystemTopicRegistryCasAuthority(
 		path.join(daemonPaths(resolvedAgentDir).dir, "telegram-topics.json"),
@@ -243,6 +314,71 @@ export async function runDaemonInternal(argv: string[], deps: RunDaemonInternalD
 	);
 	const Daemon: TelegramDaemonConstructor = deps.DaemonImpl ?? TelegramNotificationDaemon;
 	const readState = deps.readDaemonState ?? readDaemonState;
+	// The daemon owns periodic orphan-owner reconciliation: it fences stale
+	// pollers before becoming poll-capable and re-runs the reap on a fixed
+	// cadence so late strays converge. The callback reads the current owner
+	// identity from the durable state (not from this process's volatile
+	// ownerId) so the reap targets only superseded owners, never itself.
+	const orphanReap = async (): Promise<void> => {
+		const { reapTelegramDaemonOrphans, writeTelegramOrphanRecoveryReceipt } = await import(
+			"./telegram-daemon-orphan-reap"
+		);
+		const snapshot = await readOwnerFreshnessSnapshot({ settings: settings as Settings });
+		const state = snapshot.state;
+		if (!state || !hasSafeDaemonStateShape(state) || state.ownershipPhase !== "ready") return;
+		const { nativeProcessBindings } = await import("@gajae-code/utils/native-process");
+		const incarnation = state.incarnation;
+		const { receipt } = await reapTelegramDaemonOrphans({
+			agentDir: resolvedAgentDir,
+			currentOwnerId: state.ownerId,
+			currentAcquisitionId: state.acquisitionId ?? state.ownerId,
+			currentPid: state.pid,
+			currentIncarnation: incarnation,
+			fsImpl: fs.promises as unknown as import("./telegram-daemon").TelegramDaemonFs,
+			deps: {
+				pidAlive:
+					deps.pidAlive ??
+					((pid: number) => {
+						try {
+							process.kill(pid, 0);
+							return true;
+						} catch (e) {
+							return (e as NodeJS.ErrnoException).code !== "ESRCH";
+						}
+					}),
+				pidIncarnation: deps.pidIncarnation ?? processIncarnation,
+				processReference: (pid: number) => {
+					try {
+						const r = nativeProcessBindings().Process.fromPid(pid) as {
+							incarnation?: unknown;
+							killTree?: (signal?: number) => number;
+						} | null;
+						if (!r || typeof r.incarnation !== "string" || !isProcessIncarnation(r.incarnation)) return undefined;
+						return {
+							incarnation: r.incarnation,
+							terminateTree: (sig?: NodeJS.Signals) => {
+								const os = require("node:os") as typeof import("node:os");
+								const code = sig === "SIGKILL" ? os.constants.signals.SIGKILL : os.constants.signals.SIGTERM;
+								const count = (r as { killTree: (signal?: number) => number }).killTree(code);
+								return count > 0;
+							},
+						};
+					} catch {
+						return undefined;
+					}
+				},
+				now: deps.now,
+			},
+		});
+		if (receipt.refused > 0) {
+			logger.warn(`notifications: daemon orphan reap had ${receipt.refused} refused candidate(s)`);
+		}
+		await writeTelegramOrphanRecoveryReceipt(
+			fs.promises as unknown as import("./telegram-daemon").TelegramDaemonFs,
+			resolvedAgentDir,
+			receipt,
+		).catch(() => undefined);
+	};
 	const daemon = new Daemon({
 		settings: settings as Settings,
 		ownerId,
@@ -259,6 +395,8 @@ export async function runDaemonInternal(argv: string[], deps: RunDaemonInternalD
 		control: createDaemonControlHooks(settings as Settings),
 		topicRegistryAuthority,
 		installationHostId,
+		requireTelegramTopicEligibility: true,
+		orphanReap,
 	});
 	// Signals are a process concern: install them at the daemon-internal boundary,
 	// not inside the embeddable daemon class. SIGTERM is the reload wakeup path.
@@ -275,16 +413,65 @@ export async function runDaemonInternal(argv: string[], deps: RunDaemonInternalD
 		if (!watchdogActive || watchdogTickInFlight || stopRequested) return;
 		watchdogTickInFlight = true;
 		try {
-			const snapshot = deps.readDaemonState
-				? { state: await readState(settings as Settings), effectiveHeartbeatAt: undefined }
+			const isInjectedState = Boolean(deps.readDaemonState);
+			const snapshot = isInjectedState
+				? {
+						state: await readState(settings as Settings),
+						effectiveHeartbeatAt: undefined,
+						ownerTag: undefined as unknown as {
+							ownerId: string;
+							acquisitionId: string;
+							pid: number;
+							incarnation: string;
+						} | null,
+					}
 				: await readOwnerFreshnessSnapshot({ settings: settings as Settings });
 			const state = snapshot.state;
-			const heartbeatAt = snapshot.effectiveHeartbeatAt ?? state?.heartbeatAt;
+			const heartbeatAt = (snapshot as { effectiveHeartbeatAt?: number }).effectiveHeartbeatAt ?? state?.heartbeatAt;
 			if (!watchdogActive || !state) return;
-			if (state.ownerId !== ownerId) {
-				stopRequested = true;
-				daemon.requestStop("stop");
-				return;
+			// Self-fencing: prove the complete lock/state tuple (ownerId, acquisitionId, pid, incarnation).
+			// Product-owned proof uses the lock/state tuple; a bare ownerId equality alone is not authorization.
+			const publishedOwner = readyOwnerTuple(state);
+			if (isInjectedState) {
+				// Test seam path: deps.readDaemonState injects state without lock. Prove full tuple against daemon's own stable identity.
+				const expectedOwner = daemonOwnerTuple(ownerId, deps);
+				// Without stable self incarnation authority, fail closed (Windows or unavailable probe).
+				if (!publishedOwner || !expectedOwner) return;
+				if (!ownerTuplesEqual(publishedOwner, expectedOwner)) {
+					stopRequested = true;
+					daemon.requestStop("stop");
+					return;
+				}
+			} else {
+				// Production path: prove the published state matches the lock-backed ownerTag; then prove supersession against current tuple.
+				if (!publishedOwner) return;
+				const snapshotTag = (
+					snapshot as {
+						ownerTag?: { ownerId: string; acquisitionId: string; pid: number; incarnation: string } | null;
+					}
+				).ownerTag;
+				if (!snapshotTag) return;
+				// Lock/state mismatch or missing lock is ambiguous — fail closed, do not selfterminate on noise.
+				if (
+					snapshotTag.ownerId !== publishedOwner.ownerId ||
+					snapshotTag.acquisitionId !== publishedOwner.acquisitionId ||
+					snapshotTag.pid !== publishedOwner.pid ||
+					snapshotTag.incarnation !== publishedOwner.incarnation
+				)
+					return;
+				// Positive supersession: published authoritative tuple no longer matches this daemon's acquisition.
+				// We prove via exact tuple comparison; acquisitionId == ownerId for daemon-spawned owners.
+				const selfTuple = daemonOwnerTuple(ownerId, deps);
+				if (selfTuple && !ownerTuplesEqual(publishedOwner, selfTuple)) {
+					stopRequested = true;
+					daemon.requestStop("stop");
+					return;
+				}
+				// No stable self incarnation proof available (e.g. Windows native
+				// unavailable). Fail closed: do NOT self-terminate on ownerId alone.
+				// A bare ownerId equality is not sufficient authority to terminate a
+				// running daemon — a stale or reused PID could match, killing a
+				// foreign process. The stall watchdog below handles non-progress.
 			}
 			if (lastHeartbeatAt === undefined || heartbeatAt !== lastHeartbeatAt) {
 				lastHeartbeatAt = heartbeatAt;

package/src/sdk/bus/telegram-daemon-contract.ts

@@ -6,6 +6,15 @@
 /** Protocol version the daemon advertises in its ClientHello. */
 export const NOTIFICATION_PROTOCOL_VERSION = 3;
 
+/** Telegram transport generation; independent from SDK lifecycle protocol changes. */
+export const TELEGRAM_TRANSPORT_GENERATION = 1;
+
+/** Notification event schema version; payload compatibility is versioned separately. */
+export const NOTIFICATION_EVENT_SCHEMA_VERSION = 1;
+
+/** SDK Router/lifecycle protocol negotiated by core, not by Telegram transport. */
+export const SDK_LIFECYCLE_ROUTER_PROTOCOL_VERSION = 1;
+
 /**
  * Guarded behavior-inventory version for the current daemon build. Bump this
  * on every guarded daemon-behavior change independent of the wire version; it
@@ -93,26 +102,263 @@ export const NOTIFICATION_PROTOCOL_VERSION = 3;
  * later session from creating a topic or replaying notifications. Generation
  * 58 preserves that invariant during durable-fence load promotion and restores
  * the exact grace deadline when a failed archive publication rolls back.
- * Generation 59 publishes the owner's attached OPEN-socket count in the
- * heartbeat sidecar so `gjc notify health` can report a live daemon that is
- * attached to no registered endpoint (#4128). Generation 60 contains transient
- * heartbeat-sidecar publication failures: a Windows EPERM/EACCES/EBUSY on the
- * sidecar rename is retried bounded under the ownership-lock fence and an
- * exhausted publication keeps the daemon alive as `publish_failed` instead of
- * escaping as an uncaught exception, while only a proven state/lock mismatch
- * (`not_owner`) stops the owner (#4200). Generation 61 keeps authenticated
- * same-session transports attached while predecessor archival converges and
- * settles non-forum close responses. Generation 62 rejects replay-gap
- * authority claims that do not match the requested cursor, answer bounds,
- * and retained replay suffix.
- * Generation 63 contains transient steady-heartbeat state and ownership-lock
- * read failures and cleans every staged sidecar before retrying (#4200).
+ * Generation 59 moves lifecycle and attachment authority into SDK core,
+ * removes provider root scanning and direct endpoint credentials, and rebuilds
+ * Telegram presentation bindings from SessionRouter attachments. Generation 60
+ * persists Router publication receipts, awaits inbound attachment dispatch,
+ * restores actor-scoped create throttling, and exact-unlinks ownership locks.
+ * Generation 61 durably claims provider publication identities before dispatch
+ * and retains ambiguous claims across restart to prevent duplicate effects.
+ * Generation 62 confirms receipts only from accepted provider outcomes,
+ * retains queued/fallback claims, and quarantines legacy v1 confirmations.
+ * Generation 63 applies the same deferred confirmation boundary while draining
+ * replay-state and live-held frames during attachment recovery.
+ * Generation 64 additionally defers stale-lease frames, requires authoritative
+ * model message IDs, and retains claims when continuation admission fails.
+ * Generation 65 treats malformed or unknown Bot API responses as ambiguous
+ * rather than accepting them without a provider receipt.
+ * Generation 66 propagates publication identity through selected and ephemeral
+ * direct effects, retries definite pre-send route failures, and aggregates the
+ * worst continuation disposition before confirmation.
+ * Generation 67 quarantines oversized persisted receipt maps and retains claims
+ * for pending-topic frames evicted before provider delivery.
+ * Generation 68 releases claimed BTW publications when shutdown prevents dispatch.
+ * Generation 69 settles selected acknowledgements and queued BTW publications
+ * only from positive Telegram message receipts.
+ * Generation 70 separates queued claims from provider-attempt ambiguity so
+ * pre-send work remains replayable after daemon restart.
+ * Generation 71 preserves queued state through cooldown/shutdown suppression and
+ * never converts replay-suppressed ambiguity into delivered confirmation.
+ * Generation 72 keeps invalid selected publications queued and mirrors attempted
+ * BTW rejection outcomes across duplicate publication identities.
+ * Generation 73 reports queued publication disposition to SessionRouter so
+ * same-process reconnect retains the unadvanced replay cursor.
+ * Generation 74 serializes each Router frame until its durable publication state
+ * becomes attempted or delivered, preserving sequence order without cursor gaps.
+ * Generation 75 routes every direct effect through the suppression-aware adapter
+ * and gives each publication queue item a non-coalescing identity.
+ * Generation 76 adds durable terminal rejection for invalid direct frames and
+ * confirms accepted non-OK BTW status publications before returning.
+ * Generation 77 bounds terminal rejection receipts alongside delivered receipts.
+ * Generation 78 retries stale direct authority before dispatch and terminalizes
+ * intentionally superseded queued live publications.
+ * Generation 79 preserves rejected state through outer/replay confirmation and
+ * settles publication-level continuation and retry barriers.
+ * Generation 80 terminalizes every no-dispatch queue branch and revalidates
+ * terminal publication state after cooldown waits.
+ * Generation 81 terminalizes flat, pending, and stopped-pool no-dispatch work
+ * and removes the raw Router Broker mutation tunnel.
+ * Generation 82 terminalizes expired and out-of-band queue removals, defers BTW
+ * completion across all chunks, and classifies definitive direct rejections.
+ * Generation 83 terminalizes pending frames discarded during topic flush.
+ * Generation 84 closes policy/archive/selected direct waiter cleanup and routes
+ * pre-identity threaded output through flat delivery instead of a blocked queue.
+ * Generation 85 removes the obsolete pending-frame admission path entirely.
+ * Generation 86 terminalizes queued selected acknowledgements during shutdown.
+ * Generation 87 routes every run-loop exit through the publication stop boundary.
+ * Generation 88 rejects all queued publications before Router shutdown, disables
+ * implicit provider retries, and durably joins receipt persistence.
+ * Generation 89 rejects all pool submissions once a stop is requested.
+ * Generation 90 terminalizes replay-queued publications before Router shutdown.
+ * Generation 91 routes selected-ack retries through the stop admission fence.
+ * Generation 92 rejects Router waiters when terminal receipt persistence fails.
+ * Generation 93 aborts provider delivery immediately and resets failed settlement
+ * resolvers before same-daemon replay.
+ * Generation 94 wires the immediate delivery abort into every publication call.
+ * Generation 95 propagates selected-ack receipt persistence failure to Router.
+ * Generation 96 routes rich-draft delivery through the publication abort boundary.
+ * Generation 97 bounds strict tool terminalization with shutdown abort authority.
+ * Generation 98 applies delivery abort authority to every supervised Bot API call.
+ * Generation 99 propagates generic terminal receipt persistence failure to Router.
+ * Generation 100 centralizes terminal receipt failure propagation in state transitions.
+ * Generation 101 isolates batch terminal failures and replays rejected rollback.
+ * Generation 102 isolates expired and superseded pre-batch receipt failures.
+ * Generation 103 fences and joins Router reconciliation during stop.
+ * Generation 104 prevents Router timer installation after concurrent stop.
+ * Generation 105 revokes attachment capabilities immediately when Router stop begins.
+ * Generation 106 isolates every Router client close attempt during shutdown.
+ * Generation 107 joins reconnect replay within Router pending shutdown work.
+ * Generation 108 interrupts replay requests immediately when Router stop begins.
+ * Generation 109 isolates stale attachment client and provider cleanup failures.
+ * Generation 110 installs replay stop handling before starting the request.
+ * Generation 111 isolates replacement client and provider cleanup failures.
+ * Generation 112 preserves provider handoff while isolating replacement close failure.
+ * Generation 113 removes old provider authority when replacement races stop.
+ * Generation 114 removes successor provider authority when handoff races stop.
+ * Generation 115 removes provider authority for every attachment disposed by stop.
+ * Generation 116 creates publication waiters only after claim admission succeeds.
+ * Generation 117 exposes tentative claims to concurrent teardown and aborts revoked admission.
+ * Generation 118 prevents failed concurrent rejection from reviving tentative claims.
+ * Generation 119 bounds Router stop and preserves rejected failed continuations.
+ * Generation 120 awaits provisional rejection durability before replay suppression.
+ * Generation 121 fences bounded-stop continuations from later Router runs.
+ * Generation 122 fences every nested replay suppression on rejection durability.
+ * Generation 123 scopes stale Router errors to their originating run epoch.
+ * Generation 124 detaches restart tails and restores rejected rollback as replayable.
+ * Generation 125 prevents stale run cleanup from deleting restarted attachments.
+ * Generation 126 fences stale callback bookkeeping and distinguishes definitive rejection.
+ * Generation 127 keeps malformed accepted selected sends transport-ambiguous.
+ * Generation 128 bounds Discord provider work before Router revocation.
+ * Generation 129 bounds every Discord REST operation across retries.
+ * Generation 130 invalidates Discord effect lease renewal when shutdown drain expires.
+ * Generation 131 invalidates Discord effect leases at the journal mutation boundary.
+ * Generation 132 invalidates Discord create-intent leases at drain expiry.
+ * Generation 133 serializes Discord creator invalidation with in-flight admission.
+ * Generation 134 durably restores replay after definitive provider rejection.
+ * Generation 135 revokes attachments whose provider publication hook rejects.
+ * Generation 136 keeps provider attachment publication provisional until recovery succeeds.
+ * Generation 137 fences stale reconciliation, frame tails, and identity-less replay rejection.
+ * Generation 138 starts provider handshakes only after Router publication.
+ * Generation 139 privatizes raw SDK clients and removes the Python transport package.
+ * Generation 140 retires failed handshakes and holds provisional frames until publication.
+ * Generation 141 awaits the provider handshake before Router replay to prevent retained-event deadlock.
+ * Generation 142 migrates live adoption reservations and applies create throttling to topic adoption.
+ * Generation 143 preserves topic presentation continuity across exact Router endpoint replacement.
+ * Generation 144 revalidates endpoint authority before exact publication-time requests.
+ * Generation 145 distinguishes transport reconnect from a changed same-generation endpoint successor.
+ * Generation 146 revokes token-zero and callback leases before endpoint replacement while retaining topic presentation.
+ * Generation 147 uses a restart-stable opaque Router endpoint-incarnation identity.
+ * Generation 148 classifies reconnect-time endpoint successors before attachment retirement.
+ * Generation 149 serializes successor attachment behind predecessor provider retirement.
+ * Generation 150 version-fences attaches already in flight when retirement begins.
+ * Generation 151 preserves provider continuity for endpoint-before-index replacement cleanup.
+ * Generation 152 contains a thrown steady heartbeat renewal in the run loop
+ * (a transient EPERM/EACCES/EBUSY on the state or ownership-lock read) the
+ * same way instead of terminating the daemon (#4200).
+ * Generation 153 restricts Telegram forum-topic ownership, replay, routing, callbacks,
+ * and lease renewal to identities with coordinator or lifecycle provenance.
+ * Generation 154 archives private-chat topics through deleteForumTopic, settles
+ * TOPIC_ID_INVALID as definitive, and drains durable archive retries periodically.
+ * Generation 155 unifies the duplicated durable terminal-retention write path
+ * (bus and SDK-only host runtimes) into a single `boundTerminalRetentionState`
+ * helper in `session/terminal-abort.ts` (#4329).
+ * Generation 156 derives Telegram session eligibility from configuration
+ * (Telegram configured and effectively enabled) instead of coordinator or
+ * lifecycle launch provenance, and stops consulting a session's self-declared
+ * eligibility for topic admission: threaded mode always uses threads. Before
+ * this, every ordinary interactive session declared itself ineligible, the
+ * daemon rejected its identity header, no topic was created, and nothing was
+ * delivered while attachments still looked healthy.
+ * Generation 157 adds orphan-topic reconciliation: FORUM_TOPIC_NOT_FOUND is
+ * settled as definitive, and the unsupported-method fallback closes when
+ * deleteForumTopic is unavailable on older Bot API deployments.
+ * Generation 158 updates terminal-abort notification admission and cleanup semantics.
+ * Generation 159 fences the Windows process-incarnation authority: the native
+ * binding fallback no longer spawns powershell.exe when it cannot bind a pid,
+ * so older daemons that may still flash a console window during liveness polling
+ * are fenced off (#4362).
+ * Generation 160 introduces provider-local Telegram subscriptions, detached
+ * cleanup, stable topic bindings, and explicit archive authority. Older daemons
+ * retain attachment-level lifecycle coupling and may not serve this contract.
+ * Generation 161 completes that authority split: archive reasons and cleanup
+ * receipts are durable, topic recovery no longer reconstructs SDK endpoint
+ * authority, and replay generation cannot mutate lifecycle authority.
+ * Generation 162 removes chat-daemon ownership from the session-open critical
+ * path: the host acquires provider ownership only AFTER publishing its core
+ * endpoint, through a single-flight coordinator keyed to the full provider
+ * identity, and never awaits it on a lifecycle path. The daemon also sweeps
+ * inert notification-dir debris at startup. Older daemons assume ownership is
+ * proved before publication and do not perform that sweep. Generation 165
+ * preserves a bounded lean settlement window across autonomous continuations,
+ * so the Telegram terminal receipt cannot lose a prior user-request result.
+ * Generation 166 introduces complete owned process-group cleanup (killTree
+ * instead of root-only signalRoot), identity-fenced orphan-owner reconciliation
+ * in the daemon run loop, pre-poll stale-poller fencing, and fail-closed
+ * watchdog behavior when stable identity authority is unavailable (#4403).
+ * Generation 167 repairs the telegram daemon generation guard bootstrap so the
+ * generator's post-fix manifest check always byte-compares the regenerated disk
+ * manifest against the current tree, and auto-reaps pre-registry legacy stray
+ * Telegram daemons (#4533).
+ * Generation 168 adds per-update inbound acknowledgement authority and monotonic
+ * reaction settlement for Telegram notification delivery (#4528).
+ * Generation 169 delivers every ring-positioned session event live through the
+ * bounded, capability-gated directed subscriber leg used by replay.
  */
-export const DAEMON_GENERATION = 63;
+export const DAEMON_GENERATION = 169;
 
 /**
- * Serving-compatibility boundary for daemon lifecycle requests. Epoch 5
- * requires the complete generation-36 topic authority contract, so older
- * epoch-4 daemons cannot keep serving across an upgrade.
+ * Serving-compatibility boundary for daemon lifecycle requests. Epoch 7
+ * requires durable publication receipts and acknowledged attachment dispatch,
+ * so epoch-6 daemons cannot serve across the delivery-safety cutover. Epoch 8
+ * requires the durable claimed/confirmed publication receipt format. Epoch 9
+ * requires accepted-only confirmation and legacy receipt quarantine. Epoch 10
+ * requires replay admission to retain deferred publication claims. Epoch 11
+ * requires strict direct receipts and fail-closed continuation admission. Epoch 12
+ * requires malformed provider responses to remain unconfirmed claims.
+ * Epoch 13 requires complete direct-effect identity and pre-send retry semantics.
+ * Epoch 14 requires bounded receipt loading and fail-closed pending eviction.
+ * Epoch 15 requires BTW shutdown to preserve Router replay authority.
+ * Epoch 16 requires accepted-only settlement for special provider queues.
+ * Epoch 17 requires durable queued/attempted/confirmed publication states.
+ * Epoch 18 requires suppression-aware attempt transitions and replay confirmation.
+ * Epoch 19 requires direct invalid/rejected outcomes to retain exact receipt state.
+ * Epoch 20 requires Router cursor advancement to follow settled publication state.
+ * Epoch 21 requires ordered Router settlement waiters for queued publications.
+ * Epoch 22 requires unified dispatch guards and non-coalescing publication work.
+ * Epoch 23 requires explicit rejected state and terminal direct settlement.
+ * Epoch 24 requires bounded terminal rejection retention.
+ * Epoch 25 requires stale-dispatch retry and superseded publication settlement.
+ * Epoch 26 requires terminal-state exclusivity and complete publication barriers.
+ * Epoch 27 requires no-dispatch terminalization and post-wait revalidation.
+ * Epoch 28 requires terminal flat/pending admission and list-only Broker routing.
+ * Epoch 29 requires complete removal terminalization and direct rejection state.
+ * Epoch 30 requires terminal settlement for pending-frame flush discards.
+ * Epoch 31 requires complete direct waiter closure and pre-identity delivery.
+ * Epoch 32 removes identity-gated pending publication admission.
+ * Epoch 33 requires selected-ack shutdown removal to settle publication state.
+ * Epoch 34 requires control-file exits to terminalize queued publications first.
+ * Epoch 35 requires complete publication teardown and single-attempt dispatch.
+ * Epoch 36 closes the retry-after-stop queue race.
+ * Epoch 37 closes replay-queue settlement during stop.
+ * Epoch 38 closes selected-ack retry-after-stop settlement.
+ * Epoch 39 propagates terminal receipt persistence failure to Router shutdown.
+ * Epoch 40 closes hanging-delivery shutdown and transient persistence replay.
+ * Epoch 41 closes ordinary publication hangs during shutdown.
+ * Epoch 42 closes selected terminal receipt persistence settlement.
+ * Epoch 43 closes rich-draft transport hangs during shutdown.
+ * Epoch 44 closes terminal tool transport hangs during shutdown.
+ * Epoch 45 closes raw topic and callback transport hangs during shutdown.
+ * Epoch 46 closes timer-finalized publication persistence settlement.
+ * Epoch 47 covers every delivered/rejected persistence rollback path.
+ * Epoch 48 preserves unrelated batch settlement and rejected replayability.
+ * Epoch 49 preserves every removed inventory settlement after individual failure.
+ * Epoch 50 prevents attachment resurrection after provider shutdown.
+ * Epoch 51 fully fences Router startup and reconciliation after stop.
+ * Epoch 52 prevents provider dispatch throughout the Router stop window.
+ * Epoch 53 prevents one transport close timeout from skipping later cleanup.
+ * Epoch 54 prevents Router return before replay callbacks complete.
+ * Epoch 55 closes pending reconnect replay without hanging Router shutdown.
+ * Epoch 56 guarantees provider authority cleanup despite transport close failure.
+ * Epoch 57 prevents orphaned replay request rejection during stop.
+ * Epoch 58 guarantees old authority cleanup before replacement attachment.
+ * Epoch 59 drops old provider authority only when replacement creation fails.
+ * Epoch 60 closes provider cleanup for replacement creation interrupted by stop.
+ * Epoch 61 closes provider cleanup after published replacement handoff.
+ * Epoch 62 covers stop during recovered-frame and replay awaits.
+ * Epoch 63 prevents capacity and claim-persistence waiter leaks.
+ * Epoch 64 closes attachment replacement during claim persistence.
+ * Epoch 65 closes dual persistence failure during tentative claim teardown.
+ * Epoch 66 prevents callback shutdown deadlock and false ambiguous rejection.
+ * Epoch 67 closes concurrent claim/rejection persistence cursor concession.
+ * Epoch 68 prevents pre-stop reconciliation authority after same-instance restart.
+ * Epoch 69 prevents provisional nested replay cursor concession.
+ * Epoch 70 prevents stale run failures from stopping a restarted Router.
+ * Epoch 71 restores bounded Router restart and rejected publication replay.
+ * Epoch 72 identity-fences provider cleanup after bounded restart.
+ * Epoch 73 preserves successor retry state and ambiguous dispatched cleanup.
+ * Epoch 74 prevents replay after accepted responses lacking message identity.
+ * Epoch 75 prevents hung Discord REST work from blocking daemon shutdown.
+ * Epoch 76 cancels hung Discord REST before active-work drain expires.
+ * Epoch 77 fences late Discord effect commits after Router revocation.
+ * Epoch 78 atomically marks timed-out Discord work uncertain before Router revocation.
+ * Epoch 79 releases timed-out Discord creator ownership for successor recovery.
+ * Epoch 80 terminalizes leases admitted after shutdown invalidator snapshot.
+ * Epoch 81 preserves definitive provider rejection across persistence failure and restart.
+ * Epoch 82 fails closed before exposing provider attachments after cleanup recovery failure.
+ * Epoch 83 binds provider commands to the exact opaque attachment identity.
+ * Epoch 84 prohibits provider lifecycle-equivalent controls and detaches replacement work.
+ * Epoch 85 removes direct relay attachment and prohibits lifecycle controls on every adapter.
+ * Epoch 86 removes public discovery/client exports and retires cross-process raw transports.
+ * Epoch 87 preserves accepted Telegram delivery ambiguity and advances poisoned poll cursors.
  */
-export const SERVING_EPOCH = 5;
+export const SERVING_EPOCH = 87;

package/src/sdk/bus/telegram-daemon-control.ts

@@ -27,7 +27,6 @@ import { resolveGjcRuntimeSpawnInfo } from "../../daemon/runtime";
 import { isProcessIncarnation } from "../broker/process-incarnation";
 
 import { getNotificationConfig, isTelegramComplete, tokenFingerprint } from "./config";
-import { exactUnlinkNotificationFile, readNotificationEndpointFile } from "./notification-service";
 import {
 	type AttestedLegacyDaemonOwner,
 	confirmTelegramDaemonSpawn,
@@ -38,7 +37,6 @@ import {
 	isFreshLiveOwner,
 	isSignalableMatchingOwner,
 	readAttestedLegacyDaemonOwner,
-	readDaemonRoots,
 	readDaemonState,
 	readOwnerFreshnessSnapshot,
 	spawnTelegramDaemonOwner,
@@ -49,9 +47,6 @@ import {
 
 const nodeFs: TelegramDaemonFs = {
 	...(fs.promises as unknown as TelegramDaemonFs),
-	readEndpointFile: readNotificationEndpointFile,
-	exactUnlink: async (file, identity) =>
-		exactUnlinkNotificationFile(file, identity, `.gjc-delete-daemon-transition-${crypto.randomUUID()}.json`),
 };
 const DEFAULT_GRACEFUL_TIMEOUT_MS = 8_000;
 const DEFAULT_KILL_TIMEOUT_MS = 3_000;
@@ -227,7 +222,6 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 		}
 		const snapshot = await readOwnerFreshnessSnapshot({ settings: this.settings, fs: this.fsImpl });
 		const state = snapshot.state;
-		const roots = await readDaemonRoots(this.settings, this.fsImpl);
 		const health: DaemonHealth =
 			!state || state.stoppedAt !== undefined || !this.pidAlive(state.pid)
 				? "stopped"
@@ -250,8 +244,6 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 			ownerId: state?.ownerId,
 			startedAt: state?.startedAt,
 			heartbeatAt: snapshot.effectiveHeartbeatAt,
-			roots,
-			rootCount: roots.length,
 			runtime,
 		};
 	}
@@ -270,12 +262,11 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 	}
 
 	private async spawnAndWait(
-		roots: string[],
 		token: string,
 		chatId: string,
 	): Promise<{ spawned: TelegramSpawnOwnerResult; ready: boolean }> {
 		const spawned = await spawnTelegramDaemonOwner(
-			{ settings: this.settings, roots, tokenFingerprint: token, chatId },
+			{ settings: this.settings, tokenFingerprint: token, chatId },
 			this.spawnDeps(),
 		);
 		const ready = await confirmTelegramDaemonSpawn({
@@ -479,7 +470,6 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 		const cfg = getNotificationConfig(this.settings);
 		const fp = tokenFingerprint(cfg.botToken as string);
 		const chatId = cfg.chatId as string;
-		const roots = before.roots ?? (await readDaemonRoots(this.settings, this.fsImpl));
 		const gracefulTimeoutMs = opts.gracefulTimeoutMs ?? DEFAULT_GRACEFUL_TIMEOUT_MS;
 		const killTimeoutMs = opts.killTimeoutMs ?? DEFAULT_KILL_TIMEOUT_MS;
 
@@ -544,7 +534,7 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 			if (!(opts.spawnIfStopped ?? true)) {
 				return this.result(action, true, "no running telegram daemon to reload", before, before, warnings);
 			}
-			const { spawned, ready } = await this.spawnAndWait(roots, fp, chatId);
+			const { spawned, ready } = await this.spawnAndWait(fp, chatId);
 			warnings.push(...spawned.warnings);
 			const after = await this.status();
 			if (spawned.result === "blocked") {
@@ -704,7 +694,7 @@ export class TelegramDaemonController implements BuiltInDaemonController {
 				warnings,
 			);
 		}
-		const { spawned, ready } = await this.spawnAndWait(roots, fp, chatId);
+		const { spawned, ready } = await this.spawnAndWait(fp, chatId);
 		warnings.push(...spawned.warnings);
 		const after = await this.status();
 		if (spawned.result === "attached") {