@adhdev/daemon-core 0.9.82-rc.48 → 0.9.82-rc.480

package/src/mesh/mesh-event-forwarding.ts

@@ -0,0 +1,2002 @@
+import type { DaemonComponents } from '../boot/daemon-lifecycle.js';
+import { loadConfig } from '../config/config.js';
+import { getMesh, getMeshByRepo, listMeshes } from '../config/mesh-config.js';
+import { LOG } from '../logging/logger.js';
+import { appendLedgerEntry, buildTaskCompletionEvidence, getSessionRecoveryContext, isIntentionalCleanupStopEntry, readLedgerEntries } from './mesh-ledger.js';
+import type { SessionRecoveryContext } from './mesh-ledger.js';
+import { updateSessionTaskStatus, enqueueTask, updateDirectDispatchStatus, cleanupTerminalDirectDispatches, getActiveDirectDispatches, hasPendingDependents, getQueue } from './mesh-work-queue.js';
+import { markSessionDeliveriesTerminal, updateSessionDeliveryStatus } from './mesh-delivery-policy.js';
+import { MeshRuntimeStore } from './mesh-runtime-store.js';
+import { queuePendingMeshCoordinatorEvent, drainPendingMeshCoordinatorEvents, prunePendingMeshCoordinatorEventsRetention, readV2EnvelopeFromWire, type PendingMeshCoordinatorEvent } from './mesh-events-pending.js';
+import type { ProviderInstance } from '../providers/provider-instance.js';
+import { resolveWorkerDelegateRouting, recordUnroutableDelegateEvent, isUnroutableDelegateRejection } from './mesh-routing.js';
+import { resolveMeshHostStatus } from './mesh-host-ownership.js';
+import { enqueueUnresolvedDelegateForward, peekUnresolvedDelegateForwards, ackUnresolvedDelegateForward } from './mesh-unresolved-forward-outbox.js';
+import { traceMeshEventStage, traceMeshEventDrop } from './mesh-event-trace.js';
+import { getLastDisplayMessage } from '../status/snapshot.js';
+import { resolveDelegatedWorkerAutoApprove } from '../repo-mesh-types.js';
+import { meshNodeIdMatches, daemonIdsEquivalent, expandDaemonIdForms, sessionIdsEquivalent, type MeshNodeIdentified } from '@adhdev/mesh-shared';
+import {
+    findRecentTerminalLedgerEvidence,
+    findTerminalLedgerEvidenceForTask,
+    hasDispatchAfterTerminal,
+    hasUnterminalDirectDispatchLedgerEntry,
+    buildNoProgressCompletionReconciliation,
+} from './mesh-events-stale.js';
+import { endTaskDispatchInFlight } from './mesh-task-inflight.js';
+import {
+    buildMeshSystemMessage,
+    readNonEmptyString,
+    resolveEventSessionId,
+    readRefineJobId,
+    readWorkerResultMetadata,
+    resolveMeshSurfacedSessionPreview,
+    isFalseIdleCompletion,
+    isWeakCompletionEvidence,
+} from './mesh-events-utils.js';
+import { isMeshCoordinatorEvent, shouldForceInjectMeshEvent, EVENT_TO_LEDGER_KIND } from './mesh-event-classify.js';
+import {
+    getMeshWithCache,
+    tryAssignQueueTask,
+    triggerMeshQueue,
+    runIdleMaintenanceThenAssignQueue,
+    maybeAutoFastForwardIdleNode,
+    sessionHasActiveAssignment,
+} from './mesh-queue-assignment.js';
+
+// The set of coordinator-daemon ids this daemon answers to when draining the
+// pending-events queue. Mirrors resolveCoordinatorDaemonIds in mesh-reconcile-loop:
+// a unicast event may be stamped with the status id, the bare machineId, OR the
+// config-form node daemonId (`daemon_<machineId>`) depending on which dispatch path
+// created the worker. We expand to EVERY equivalent form so a `daemon_<machineId>`
+// completion matches a coordinator that knows itself as bare `<machineId>` (the
+// base-node completion-surface bug) and vice versa.
+function resolveCoordinatorDrainDaemonIds(components: DaemonComponents): string[] {
+    const statusInstanceId = readNonEmptyString((components as { statusInstanceId?: string }).statusInstanceId);
+    const machineId = readNonEmptyString(loadConfig().machineId);
+    return expandDaemonIdForms([statusInstanceId, machineId]);
+}
+
+// ---------------------------------------------------------------------------
+// Remote Node Idle Session Tracking
+// ---------------------------------------------------------------------------
+const REMOTE_IDLE_SESSION_TTL_MS = 5 * 60 * 1000; // 5 minutes
+
+// ---------------------------------------------------------------------------
+// Workspace-to-mesh lookup cache
+// ---------------------------------------------------------------------------
+const meshByWorkspaceCache = new Map<string, { mesh: any; cachedAt: number }>();
+const MESH_WORKSPACE_CACHE_TTL_MS = 5_000;
+
+function getCachedMeshByWorkspace(workspace: string): any {
+    const now = Date.now();
+    const cached = meshByWorkspaceCache.get(workspace);
+    if (cached && now - cached.cachedAt < MESH_WORKSPACE_CACHE_TTL_MS) return cached.mesh;
+    const mesh = getMeshByRepo(workspace);
+    meshByWorkspaceCache.set(workspace, { mesh, cachedAt: now });
+    return mesh;
+}
+
+// Deterministic meshId recovery for a forwarded worker event that carries no meshId.
+// An unresolved-mesh worker (forwardUnresolvedDelegateEvent) cannot resolve its own
+// meshId locally, so it pushes the event with nodeId + workspace only and relies on
+// the coordinator — which hosts the mesh — to recover the id. Workspace recovery
+// (getCachedMeshByWorkspace → getMeshByRepo) is the fast path but can miss (a worktree
+// clone whose repoIdentity differs, or a transient cache state), which left the retry
+// permanently rejected with "meshId required". The node-id IS a stable, coordinator-side
+// fact: scan the hosted meshes for the one whose node matches the forwarded nodeId
+// (3-form normalizer). This is timing-independent and never depends on repo lookup.
+function recoverMeshIdByNodeId(nodeId: string): string {
+    if (!nodeId) return '';
+    for (const mesh of listMeshes()) {
+        if (Array.isArray(mesh.nodes) && mesh.nodes.some((n: any) => meshNodeIdMatches(n, nodeId))) {
+            return readNonEmptyString(mesh.id);
+        }
+    }
+    return '';
+}
+
+// MESHID-DROP coordinator-anchor recovery (Fix B): the last-resort meshId recovery for an
+// unresolved-delegate forward whose payload carries the worker's coordinator anchor
+// (meshCoordinatorDaemonId) but no resolvable meshId — neither workspace nor nodeId scan
+// matched (a freshly-cloned worktree node not yet registered under the forwarded id form, or
+// an empty payload nodeId). The receiving daemon IS the coordinator/host, so scope to the
+// meshes IT hosts whose host daemon matches the anchor (daemonIdsEquivalent — never a raw
+// compare, so daemon_mach_/mach_/standalone_ forms all match the same machine). Among those:
+//   • a nodeId → the mesh whose nodes contain it (meshNodeIdMatches 3-form) wins;
+//   • no nodeId → fall back to the anchor's SINGLE hosted mesh only. Ambiguity (the anchor
+//     hosts >1 mesh and no node disambiguates) returns '' rather than guess — a wrong meshId
+//     would inject the completion into an unrelated mesh's ledger, worse than the retry-cap drop.
+export function recoverMeshIdByCoordinatorAndNode(coordinatorDaemonId: string, nodeId: string): string {
+    if (!coordinatorDaemonId) return '';
+    const hosted = listMeshes().filter(mesh => {
+        const host = resolveMeshHostStatus(mesh);
+        return host.role === 'host'
+            && (!host.hostDaemonId || daemonIdsEquivalent(host.hostDaemonId, coordinatorDaemonId));
+    });
+    if (hosted.length === 0) return '';
+    if (nodeId) {
+        const byNode = hosted.find(mesh =>
+            Array.isArray(mesh.nodes) && mesh.nodes.some((n: any) => meshNodeIdMatches(n, nodeId)));
+        if (byNode) return readNonEmptyString(byNode.id);
+        // nodeId present but matched no hosted mesh — do NOT fall through to the single-mesh
+        // guess; the node belongs to a mesh we don't host (or under a different id), and
+        // guessing would misroute. Stay unresolved.
+        return '';
+    }
+    // No nodeId to disambiguate: only safe when the anchor hosts exactly one mesh.
+    return hosted.length === 1 ? readNonEmptyString(hosted[0].id) : '';
+}
+
+// RECONCILE-MESHID-DROP: WORKER-side meshId resolution for an unresolved-delegate
+// forward payload. forwardUnresolvedDelegateEvent omits meshId by design (the worker
+// "can't resolve it") and relies on the COORDINATOR recovering it from workspace/nodeId.
+// That recovery fails when the no_node_binding session's payload has an empty nodeId AND
+// the coordinator's workspace→mesh lookup misses (a worktree clone whose repoIdentity
+// differs / a cache miss) — leaving the reconcile retry rejected with "meshId required"
+// every 4s forever. The worker actually has MORE context than the stripped payload gives
+// the coordinator: it hosts the node as a member and holds the LIVE session, whose
+// settings.meshNodeFor / meshNodeId are authoritative even when they were not stamped
+// onto the original event. Resolve here (worker side) and stamp meshId onto the payload so
+// the coordinator accepts it. Mirrors the receiver's recovery order, then adds the live-
+// session fallback. Returns '' when even the worker cannot resolve it (truly unresolvable —
+// the retry cap then drops it instead of looping). No side effects; safe to call per retry.
+export function resolveForwardEventMeshId(
+    components: DaemonComponents,
+    payload: Record<string, unknown>,
+): string {
+    const direct = readNonEmptyString(payload.meshId);
+    if (direct) return direct;
+    const workspace = readNonEmptyString(payload.workspace);
+    const byWorkspace = workspace ? readNonEmptyString(getCachedMeshByWorkspace(workspace)?.id) : '';
+    if (byWorkspace) return byWorkspace;
+    const byNode = recoverMeshIdByNodeId(readNonEmptyString(payload.nodeId));
+    if (byNode) return byNode;
+    // Live-session fallback: the worker session may carry meshNodeFor / meshNodeId now even
+    // though the original event didn't (a late stamp, or an event that fired before binding).
+    const sessionId = readNonEmptyString(payload.targetSessionId)
+        || readNonEmptyString(payload.sessionId)
+        || readNonEmptyString(payload.instanceId);
+    if (sessionId) {
+        try {
+            const state = components.instanceManager?.getInstance?.(sessionId)?.getState?.();
+            const settings = (state?.settings as Record<string, unknown>) || {};
+            const meshNodeFor = readNonEmptyString(settings.meshNodeFor);
+            if (meshNodeFor) return meshNodeFor;
+            const byStamp = recoverMeshIdByNodeId(readNonEmptyString(settings.meshNodeId));
+            if (byStamp) return byStamp;
+            const sessionWorkspace = readNonEmptyString(state?.workspace);
+            const bySessionWorkspace = sessionWorkspace ? readNonEmptyString(getCachedMeshByWorkspace(sessionWorkspace)?.id) : '';
+            if (bySessionWorkspace) return bySessionWorkspace;
+        } catch { /* best-effort — fall through to unresolved */ }
+    }
+    return '';
+}
+
+
+export function __resetMeshWorkspaceCacheForTests(): void {
+    meshByWorkspaceCache.clear();
+}
+
+// Throttle the pending-event retention DELETE so it runs at most hourly — the
+// remote-idle sweep below fires on every completion/idle transition, but a table
+// scan + DELETE should not. In-process timestamp; a restart re-arms it, which only
+// means one prune shortly after boot (harmless, idempotent).
+let lastPendingEventsPruneAt = 0;
+const PENDING_EVENTS_PRUNE_INTERVAL_MS = 60 * 60 * 1000; // 1 hour
+
+function sweepExpiredRemoteIdleSessions(): void {
+    try {
+        MeshRuntimeStore.getInstance().pruneExpiredRemoteIdleSessions();
+    } catch { /* best-effort */ }
+    // Piggyback the pending-event retention prune on the same periodic sweep, but
+    // hourly — this is the maintenance hook that keeps mesh_pending_events from
+    // accumulating stale drained/orphaned rows without bound.
+    const now = Date.now();
+    if (now - lastPendingEventsPruneAt >= PENDING_EVENTS_PRUNE_INTERVAL_MS) {
+        lastPendingEventsPruneAt = now;
+        prunePendingMeshCoordinatorEventsRetention();
+    }
+}
+
+const INTENTIONAL_CLEANUP_STOP_SUPPRESSION_MS = 30 * 60 * 1000;
+
+function isIntentionalCleanupStopMetadata(event: Record<string, unknown>): boolean {
+    return event.intentional === true
+        || event.intentionalStop === true
+        || event.operatorCleanup === true
+        || event.reason === 'operator_cleanup'
+        || event.stopReason === 'operator_cleanup'
+        || event.cleanupReason === 'operator_cleanup'
+        || event.source === 'mesh_cleanup_sessions'
+        || event.source === 'mesh_remove_node';
+}
+
+function hasRecentIntentionalCleanupStop(meshId: string, sessionId?: string, nodeId?: string): boolean {
+    if (!sessionId && !nodeId) return false;
+    const cutoff = Date.now() - INTENTIONAL_CLEANUP_STOP_SUPPRESSION_MS;
+    const entries = readLedgerEntries(meshId, { tail: 200 });
+    for (let i = entries.length - 1; i >= 0; i--) {
+        const entry = entries[i];
+        const timestamp = new Date(entry.timestamp).getTime();
+        if (!Number.isNaN(timestamp) && timestamp < cutoff) break;
+        if (!isIntentionalCleanupStopEntry(entry)) continue;
+        // Session ids are single-form; sessionIdsEquivalent is the one canonical
+        // exact-match predicate (see its doc for why no form expansion is needed).
+        if (sessionId && sessionIdsEquivalent(entry.sessionId, sessionId)) return true;
+        // Normalized node-id match (P4): the cleanup-stop entry's node id may be stored as
+        // `nodeId` or `node_id` and the `nodeId` arg can be in either form — a raw `===`
+        // would miss a genuine intentional-cleanup entry and fail to suppress the stop event.
+        if (!sessionId && nodeId && meshNodeIdMatches(entry as unknown as MeshNodeIdentified, nodeId)) return true;
+    }
+    return false;
+}
+
+function shouldSuppressIntentionalCleanupStop(args: {
+    event: string;
+    meshId: string;
+    metadataEvent: Record<string, unknown>;
+    sessionId?: string;
+    nodeId?: string;
+}): boolean {
+    if (args.event !== 'agent:stopped' && args.event !== 'monitor:no_progress') return false;
+    if (isIntentionalCleanupStopMetadata(args.metadataEvent)) return true;
+    return hasRecentIntentionalCleanupStop(args.meshId, args.sessionId, args.nodeId);
+}
+
+const RECENT_COMPLETION_FINGERPRINT_TTL_MS = 10 * 60 * 1000;
+
+function hasFingerprintSeen(meshId: string, fingerprint: string): boolean {
+    try {
+        return MeshRuntimeStore.getInstance().hasCompletionFingerprint(meshId, fingerprint);
+    } catch {
+        return false;
+    }
+}
+
+function recordFingerprintSeen(meshId: string, fingerprint: string): void {
+    try {
+        const db = MeshRuntimeStore.getInstance();
+        db.recordCompletionFingerprint(meshId, fingerprint, RECENT_COMPLETION_FINGERPRINT_TTL_MS);
+        db.sweepExpiredFingerprints();
+    } catch { /* best-effort; duplicate events are preferable to a crash */ }
+}
+
+function readEventTimestamp(value: unknown): number | null {
+    if (typeof value === 'number' && Number.isFinite(value)) return value;
+    if (typeof value === 'string' && value.trim()) {
+        const numeric = Number(value);
+        if (Number.isFinite(numeric)) return numeric;
+        const parsed = Date.parse(value);
+        if (Number.isFinite(parsed)) return parsed;
+    }
+    return null;
+}
+
+function buildMeshCompletionFingerprint(args: {
+    meshId: string;
+    event: string;
+    sessionId: string;
+    providerType?: string;
+    providerSessionId?: string;
+    timestamp?: number | null;
+    finalSummary?: string;
+    coordinatorDaemonId?: string;
+}): string {
+    const timestampPart = Number.isFinite(args.timestamp)
+        ? String(args.timestamp)
+        : readNonEmptyString(args.finalSummary).slice(0, 200);
+    return [
+        args.meshId,
+        args.event,
+        args.sessionId,
+        args.providerType || '',
+        args.providerSessionId || '',
+        timestampPart,
+        args.coordinatorDaemonId || '',
+    ].join('::');
+}
+
+function isDuplicateMeshCompletionEvent(args: {
+    meshId: string;
+    event: string;
+    sessionId: string;
+    providerType?: string;
+    providerSessionId?: string;
+    timestamp?: number | null;
+    finalSummary?: string;
+    coordinatorDaemonId?: string;
+    taskId?: string;
+    nodeId?: string;
+}): boolean {
+    const fingerprint = buildMeshCompletionFingerprint(args);
+    if (!fingerprint) return false;
+    if (hasFingerprintSeen(args.meshId, fingerprint)) {
+        // MESH-COMPLEXITY-AUDIT Part 8-2: the dedup DECISION is the fingerprint
+        // match on the line above — that is the no-loss-neutral gate and it is
+        // unchanged. The former mesh_completion_conflicts diagnostic side-record
+        // (which task lost a fingerprint collision) had no production reader and
+        // played no part in the delivery contract, so it was dropped. Suppressing
+        // the duplicate here is unaffected: a genuine second, distinct completion
+        // has a different fingerprint (different taskId/timestamp/summary) and
+        // therefore does not match, so it is NOT suppressed here.
+        return true;
+    }
+    recordFingerprintSeen(args.meshId, fingerprint);
+    return false;
+}
+
+function isDuplicateMeshApprovalEvent(args: {
+    meshId: string;
+    sessionId: string;
+    providerType?: string;
+    timestamp?: number | null;
+    modalMessage?: string;
+    modalButtons?: unknown;
+}): boolean {
+    const modalButtons = Array.isArray(args.modalButtons)
+        ? args.modalButtons.map(button => String(button).trim()).filter(Boolean)
+        : [];
+    const approvalIdentity = Number.isFinite(args.timestamp)
+        ? String(args.timestamp)
+        : JSON.stringify({ message: args.modalMessage || '', buttons: modalButtons });
+    if (!approvalIdentity || approvalIdentity === '{"message":"","buttons":[]}') return false;
+    const fingerprint = [
+        args.meshId,
+        'agent:waiting_approval',
+        args.sessionId,
+        args.providerType || '',
+        approvalIdentity,
+    ].join('::');
+    if (hasFingerprintSeen(args.meshId, fingerprint)) return true;
+    recordFingerprintSeen(args.meshId, fingerprint);
+    return false;
+}
+
+function isDuplicateRefineTerminalEvent(meshId: string, eventName: string, metadataEvent: Record<string, unknown>): boolean {
+    const jobId = readRefineJobId({ metadataEvent });
+    const fingerprint = jobId && new Set(['refine:completed', 'refine:failed']).has(eventName) ? `${meshId}::${eventName}::${jobId}` : '';
+    if (!fingerprint) return false;
+    if (hasFingerprintSeen(meshId, fingerprint)) return true;
+    recordFingerprintSeen(meshId, fingerprint);
+    return false;
+}
+
+// NOTIF-MISS (FIX 2): the set of `source` tags the transcript-reconcile synthesis path stamps
+// onto the terminal ledger entry it writes (mesh-events-stale.reconcileDirectDispatchCompletion
+// FromTranscript). A terminal carrying one of these was NOT produced by an actual provider
+// agent_status_event — it is a coordinator-side reconstruction from the worker's transcript. The
+// authoritative real completion must never be permanently masked by such a synthesized record.
+const RECONCILED_COMPLETION_SOURCES = new Set([
+    'direct_task_transcript_reconciliation',
+    'daemon_reconcile_transcript_completion',
+    'mcp_mesh_status_transcript_reconciliation',
+    'no_progress_reconciliation',
+]);
+
+// True when the recorded terminal ledger entry was SYNTHESIZED by transcript reconciliation
+// rather than emitted by the real provider event. The reconcile path tags its ledger payload
+// with a reconcile `source`; a genuine provider completion carries no such tag (or a normal
+// completionDiagnostic.reason). Both the explicit source AND the diagnostic reason are checked so
+// either carrier identifies the synthesized record.
+function isSynthesizedReconciledTerminal(terminalPayload: Record<string, unknown>): boolean {
+    const source = readNonEmptyString(terminalPayload.source);
+    if (source && RECONCILED_COMPLETION_SOURCES.has(source)) return true;
+    const diagnostic = terminalPayload.completionDiagnostic;
+    if (diagnostic && typeof diagnostic === 'object' && !Array.isArray(diagnostic)) {
+        const reason = readNonEmptyString((diagnostic as Record<string, unknown>).reason);
+        if (reason && RECONCILED_COMPLETION_SOURCES.has(reason)) return true;
+    }
+    return false;
+}
+
+// True when the INCOMING completion event is a REAL provider agent_status_event (an actual
+// agent:generating_completed from the live session) rather than itself a synthesized/reconciled
+// re-injection. We must only let a real event override a synthesized terminal — a second
+// synthesized re-arrival should still dedup normally. A real provider event carries no reconcile
+// `source` tag; a re-injected reconciliation does.
+function isRealProviderCompletionEvent(metadataEvent: Record<string, unknown>): boolean {
+    const source = readNonEmptyString(metadataEvent.source);
+    return !source || !RECONCILED_COMPLETION_SOURCES.has(source);
+}
+
+// The genuine-completion counterpart: a real final summary / worker result is present and
+// the completion is not flagged as a missing-final-assistant false idle. Used to decide
+// whether a new completion may supersede a prior WEAK (false-idle) terminal. NOTE this gates
+// only on isFalseIdleCompletion (the completionDiagnostic subset), NOT the broader
+// isWeakCompletionEvidence — preserving the original semantics where a self-declared
+// weak/insufficient event still counts as genuine if it carries a real final summary.
+function isGenuineCompletionEvidence(metadataEvent: Record<string, unknown>): boolean {
+    if (isFalseIdleCompletion(metadataEvent)) return false;
+    return !!readWorkerResultMetadata(metadataEvent) || !!readNonEmptyString(metadataEvent.finalSummary);
+}
+
+// (FALSEIDLE-BGCHILD-b) A later genuine completion of the SAME task that carries a
+// substantively different — and fuller — final summary than the recorded terminal is the REAL
+// final that an earlier (false-idle) completion pre-empted, not a duplicate. The background-child
+// false idle is the nasty case the plain isWeakCompletionEvidence supersession misses: the
+// early completion's screen parser DID see a prior/intermediate standard assistant, so it is
+// recorded as a STRONG terminal with a non-empty (but truncated) finalSummary. Without this the
+// providerSessionId/finalSummary dedup below swallows the genuine final and the coordinator is
+// stuck with the truncated mid-turn text forever (the one-shot-consumption symptom). Same-task,
+// new event is genuine, prior terminal summary is a strict prefix of (or otherwise shorter than)
+// the new one → treat as the corrected final and let it through. Conservative: requires the new
+// summary to be genuine evidence AND meaningfully longer, so an identical re-arrival or a SHORTER
+// later summary is still deduped.
+function supersedesTruncatedTerminalSummary(args: {
+    terminalPayload: Record<string, unknown>;
+    metadataEvent: Record<string, unknown>;
+    terminalTaskId: string;
+    eventTaskId: string;
+}): boolean {
+    // Only applies when both name the SAME task (a distinct task is handled by distinctTaskCompletion).
+    if (!args.terminalTaskId || !args.eventTaskId || args.terminalTaskId !== args.eventTaskId) return false;
+    if (!isGenuineCompletionEvidence(args.metadataEvent)) return false;
+    const terminalSummary = readNonEmptyString(args.terminalPayload.finalSummary);
+    const eventSummary = readNonEmptyString(args.metadataEvent.finalSummary);
+    if (!eventSummary) return false;
+    // Identical text → genuine duplicate, keep deduping.
+    if (terminalSummary === eventSummary) return false;
+    // The recorded terminal was a known-weak (false-idle) one → already handled by the weak
+    // supersession path; nothing extra to do here.
+    if (isWeakCompletionEvidence(args.terminalPayload)) return false;
+    // No prior summary at all, or the new summary strictly extends / is meaningfully longer than
+    // the recorded one → the recorded terminal was the truncated pre-emption; supersede it.
+    if (!terminalSummary) return true;
+    if (eventSummary.startsWith(terminalSummary)) return true;
+    return eventSummary.length > terminalSummary.length + 32;
+}
+
+// The latest still-active direct-dispatch taskId for a session, resolved BEFORE the
+// completion flips the dispatch row terminal. Direct dispatches (mesh_send_task) have no
+// work-queue row, so this is the only taskId available to attribute the terminal ledger
+// entry (and thus mesh task-stats) to — without it the terminal carries no taskId and the
+// task surfaces as status='unknown' / terminalKind=null in computeMeshTaskStats.
+function resolveActiveDirectDispatchTaskId(meshId: string, sessionId: string): string | undefined {
+    try {
+        // Session ids are single-form; sessionIdsEquivalent is the one canonical
+        // exact-match predicate — no node-id-style form normalization needed.
+        const matches = getActiveDirectDispatches(meshId).filter(d => sessionIdsEquivalent(d.sessionId, sessionId));
+        if (!matches.length) return undefined;
+        // getActiveDirectDispatches returns rows ordered by dispatched_at ASC; the last is
+        // the most recent dispatch (the re-dispatch / nudge whose completion this is).
+        return readNonEmptyString(matches[matches.length - 1].taskId) || undefined;
+    } catch {
+        return undefined;
+    }
+}
+
+// Coordinator-side suppression/reconcile gate for an incoming mesh event. Each clause is a
+// closed dedup/suppression concern that only inspects the event + already-resolved context and
+// either (a) returns a `suppress` result the caller forwards verbatim, (b) returns a `reconcile`
+// signal carrying the rewritten metadataEvent for the caller to re-inject as
+// agent:generating_completed, or (c) returns null to let the event fall through to the
+// terminal/ledger machinery. Extracted verbatim from injectMeshSystemMessage — no behavior
+// change; the only side effects (best-effort remote-idle cleanup, LOG, trace) fire on the same
+// paths as before.
+function evaluateMeshEventSuppression(
+    args: {
+        meshId: string;
+        sourceInstanceId?: string;
+        nodeId?: string;
+        nodeLabel: string;
+        event: string;
+        metadataEvent: Record<string, unknown>;
+    },
+    ctx: {
+        traceCtx: Parameters<typeof traceMeshEventDrop>[1];
+        eventSessionId: string;
+        eventNodeId: string;
+        eventTimestamp: number | null;
+        workerCoordinatorDaemonId: string | undefined;
+    },
+):
+    | { kind: 'suppress'; result: { success: true; forwarded: 0; suppressed: true; [extra: string]: unknown } }
+    | { kind: 'reconcile'; metadataEvent: Record<string, unknown> }
+    | null {
+    const { traceCtx, eventSessionId, eventNodeId, eventTimestamp, workerCoordinatorDaemonId } = ctx;
+
+    const intentionalCleanupStop = shouldSuppressIntentionalCleanupStop({
+        event: args.event,
+        meshId: args.meshId,
+        metadataEvent: args.metadataEvent,
+        sessionId: eventSessionId || undefined,
+        nodeId: eventNodeId || undefined,
+    });
+    if (intentionalCleanupStop) {
+        if (eventSessionId && eventNodeId) {
+            try {
+                MeshRuntimeStore.getInstance().deleteRemoteIdleSession(args.meshId, eventNodeId, eventSessionId);
+            } catch { /* best-effort */ }
+        }
+        LOG.info('MeshEvents', `Suppressed ${args.event} for intentionally cleanup-stopped session ${eventSessionId || '(unknown session)'}`);
+        traceMeshEventDrop('intentional_cleanup_stop', traceCtx);
+        return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, intentionalCleanupStop: true } };
+    }
+
+    if (args.event === 'monitor:no_progress') {
+        const reconciledCompletion = buildNoProgressCompletionReconciliation({
+            meshId: args.meshId,
+            nodeId: args.nodeId,
+            nodeLabel: args.nodeLabel,
+            metadataEvent: args.metadataEvent,
+            sourceInstanceId: args.sourceInstanceId,
+        });
+        if (reconciledCompletion?.source === 'no_progress_reconciliation') {
+            LOG.info('MeshEvents', `Reconciled no-progress monitor to completion for session ${eventSessionId || '(unknown session)'}`);
+            return { kind: 'reconcile', metadataEvent: reconciledCompletion };
+        }
+        if (reconciledCompletion?.source === 'no_progress_terminal_ledger_suppression') {
+            LOG.info('MeshEvents', `Suppressed no-progress monitor because terminal ledger evidence already exists for session ${eventSessionId || '(unknown session)'}`);
+            traceMeshEventDrop('no_progress_terminal_ledger_suppression', traceCtx, `terminalKind=${reconciledCompletion.terminalLedgerKind}`);
+            return {
+                kind: 'suppress',
+                result: {
+                    success: true,
+                    forwarded: 0,
+                    suppressed: true,
+                    terminalLedgerEvidence: true,
+                    terminalLedgerKind: reconciledCompletion.terminalLedgerKind,
+                },
+            };
+        }
+    }
+
+    if (isDuplicateRefineTerminalEvent(args.meshId, args.event, args.metadataEvent)) {
+        LOG.info('MeshEvents', `Suppressed duplicate ${args.event} for refine job ${readRefineJobId({ metadataEvent: args.metadataEvent })}`);
+        traceMeshEventDrop('duplicate_refine_terminal', traceCtx);
+        return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, duplicateRefineTerminalEvent: true } };
+    }
+
+    if (args.event === 'agent:waiting_approval' && eventSessionId) {
+        const duplicateApproval = isDuplicateMeshApprovalEvent({
+            meshId: args.meshId,
+            sessionId: eventSessionId,
+            providerType: readNonEmptyString(args.metadataEvent.providerType) || undefined,
+            timestamp: eventTimestamp,
+            modalMessage: readNonEmptyString(args.metadataEvent.modalMessage) || undefined,
+            modalButtons: args.metadataEvent.modalButtons,
+        });
+        if (duplicateApproval) {
+            LOG.info('MeshEvents', `Suppressed duplicate approval event for mesh ${args.meshId} session ${eventSessionId}`);
+            traceMeshEventDrop('duplicate_approval', traceCtx);
+            return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, duplicateApproval: true } };
+        }
+    }
+    if (args.event === 'agent:generating_completed' && eventSessionId) {
+        const terminal = findRecentTerminalLedgerEvidence({
+            meshId: args.meshId,
+            sessionId: eventSessionId,
+            nodeId: eventNodeId || undefined,
+        });
+        if (terminal?.kind === 'task_completed' && !sessionHasActiveAssignment(args.meshId, eventSessionId)) {
+            const newDispatchAfterTerminal = hasDispatchAfterTerminal(args.meshId, eventSessionId, terminal.id);
+            // Fix B (re-dispatch 2nd-completion routing): a prior terminal recorded from a FALSE
+            // idle (weak evidence / no confirmed final assistant) must NOT permanently suppress a
+            // later GENUINE completion of the same session. providerSessionId is stable across a
+            // session's turns, so the providerSessionId/finalSummary dedup below would otherwise
+            // swallow the real 2nd-turn completion that a coordinator nudge (direct re-dispatch)
+            // drove — exactly the missed-event bug. When the prior terminal was weak and the new
+            // event carries genuine completion evidence, let it through so it is recorded and
+            // re-attributed to the latest task (the normal task_completed path below).
+            const supersedesWeakTerminal = isWeakCompletionEvidence(terminal.payload)
+                && isGenuineCompletionEvidence(args.metadataEvent);
+            // CANON-B (direct-dispatch completion race): a FAST direct dispatch (mesh_send_task)
+            // to an already-idle, previously-used session can have its genuine completion reach
+            // this coordinator handler BEFORE the dispatching side records the new task's dispatch
+            // row / task_dispatched ledger entry — insertDirectDispatch + appendLedgerEntry both run
+            // AFTER the agent_command await resolves, while the worker may already be done. In that
+            // window sessionHasActiveAssignment is false (no active dispatch row, no unterminal
+            // ledger entry yet), so this prior-terminal dedup engages; and because providerSessionId
+            // is STABLE across a reused session's turns, the providerSessionId/finalSummary match
+            // below would suppress the NEW task's completion as a duplicate of the PRIOR task —
+            // silently losing it (the observed intermittent miss; fresh enqueue/autoLaunch is immune
+            // because a fresh session has no prior same-providerSessionId terminal and the queue row
+            // is claimed atomically before dispatch). The echoed taskId is the authoritative
+            // discriminator: when the completion names a DIFFERENT task than the recorded terminal,
+            // it is a genuinely new task's completion, never a duplicate — let it through so it is
+            // attributed to its own taskId. A same-task re-arrival (taskId equal) or a taskId-less
+            // legacy event still falls through to the providerSessionId/finalSummary dedup.
+            const terminalTaskId = readNonEmptyString(terminal.payload.taskId);
+            const eventTaskId = readNonEmptyString(args.metadataEvent.taskId);
+            const distinctTaskCompletion = !!eventTaskId && !!terminalTaskId && eventTaskId !== terminalTaskId;
+            // (FALSEIDLE-BGCHILD-b) Same-task genuine completion carrying a fuller summary than the
+            // recorded (truncated, false-idle-pre-empted) terminal supersedes it — see helper.
+            const supersedesTruncatedTerminal = supersedesTruncatedTerminalSummary({
+                terminalPayload: terminal.payload,
+                metadataEvent: args.metadataEvent,
+                terminalTaskId,
+                eventTaskId,
+            });
+            // NOTIF-MISS (FIX 2): the recorded terminal was SYNTHESIZED by transcript reconciliation
+            // (no real provider event), and THIS incoming event is the authoritative REAL provider
+            // completion. The reconcile path may fire ~1s after a direct dispatch from a reused
+            // session's stale transcript tail, writing a synthesized terminal whose providerSessionId
+            // (stable across a session's turns) and/or finalSummary then match the real completion —
+            // so the providerSessionId/finalSummary dedup below would drop the genuine event and the
+            // coordinator would never learn the task finished. A synthesized record must NEVER
+            // permanently mask the real completion: let the real event through (it is re-recorded by
+            // the normal task_completed path, superseding the synthesized one). A second SYNTHESIZED
+            // re-arrival is NOT a real event and still dedups normally.
+            // RECONCILE-SYNTH-PREEMPTS-COMPLETION: the bar here is deliberately LOWER than the
+            // weak/truncated supersessions above — a synthesized terminal is a coordinator-side
+            // reconstruction, never a real provider event, so ANY genuine real provider completion
+            // for the session must win over it. Requiring the full isGenuineCompletionEvidence
+            // (finalSummary OR workerResult present) dropped the real event whenever the relay did
+            // not re-populate finalSummary at this layer — exactly the observed 71s task whose real
+            // generating_completed hit drop:duplicate_completion_terminal_ledger after a premature
+            // synth. We require only that the incoming event is a REAL provider completion that is
+            // not itself a false idle (missing-final-assistant); a real-but-false-idle event still
+            // does not supersede (it is not trustworthy terminal evidence either).
+            const supersedesSynthesizedTerminal = isSynthesizedReconciledTerminal(terminal.payload)
+                && isRealProviderCompletionEvent(args.metadataEvent)
+                && !isFalseIdleCompletion(args.metadataEvent);
+            if (!newDispatchAfterTerminal && !supersedesWeakTerminal && !distinctTaskCompletion && !supersedesTruncatedTerminal && !supersedesSynthesizedTerminal) {
+                const terminalProviderSessionId = readNonEmptyString(terminal.payload.providerSessionId);
+                const terminalFinalSummary = readNonEmptyString(terminal.payload.finalSummary);
+                const eventProviderSessionId = readNonEmptyString(args.metadataEvent.providerSessionId);
+                const eventFinalSummary = readNonEmptyString(args.metadataEvent.finalSummary);
+                if (
+                    (terminalProviderSessionId && terminalProviderSessionId === eventProviderSessionId)
+                    || (terminalFinalSummary && terminalFinalSummary === eventFinalSummary)
+                    || args.metadataEvent.source === 'no_progress_reconciliation'
+                ) {
+                    LOG.info('MeshEvents', `Suppressed duplicate completion with existing terminal ledger evidence for mesh ${args.meshId} session ${eventSessionId}`);
+                    traceMeshEventDrop('duplicate_completion_terminal_ledger', traceCtx);
+                    return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, duplicateCompletion: true, terminalLedgerEvidence: true } };
+                }
+            }
+        }
+        const duplicateCompletion = isDuplicateMeshCompletionEvent({
+            meshId: args.meshId,
+            event: args.event,
+            sessionId: eventSessionId,
+            providerType: readNonEmptyString(args.metadataEvent.providerType) || undefined,
+            providerSessionId: readNonEmptyString(args.metadataEvent.providerSessionId) || undefined,
+            timestamp: eventTimestamp,
+            finalSummary: readNonEmptyString(args.metadataEvent.finalSummary) || undefined,
+            coordinatorDaemonId: workerCoordinatorDaemonId || undefined,
+            taskId: readNonEmptyString(args.metadataEvent.taskId) || undefined,
+            nodeId: eventNodeId || undefined,
+        });
+        if (duplicateCompletion) {
+            LOG.info('MeshEvents', `Suppressed duplicate completion for mesh ${args.meshId} session ${eventSessionId}`);
+            traceMeshEventDrop('duplicate_completion', traceCtx);
+            return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, duplicateCompletion: true } };
+        }
+    }
+    if (args.event === 'agent:stopped' && eventSessionId) {
+        const duplicateStopped = isDuplicateMeshCompletionEvent({
+            meshId: args.meshId,
+            event: args.event,
+            sessionId: eventSessionId,
+            providerType: readNonEmptyString(args.metadataEvent.providerType) || undefined,
+            providerSessionId: readNonEmptyString(args.metadataEvent.providerSessionId) || undefined,
+            timestamp: eventTimestamp,
+            finalSummary: readNonEmptyString(args.metadataEvent.finalSummary) || undefined,
+            coordinatorDaemonId: workerCoordinatorDaemonId || undefined,
+            taskId: readNonEmptyString(args.metadataEvent.taskId) || undefined,
+            nodeId: eventNodeId || undefined,
+        });
+        if (duplicateStopped) {
+            LOG.info('MeshEvents', `Suppressed duplicate stopped event for mesh ${args.meshId} session ${eventSessionId}`);
+            traceMeshEventDrop('duplicate_stopped', traceCtx);
+            return { kind: 'suppress', result: { success: true, forwarded: 0, suppressed: true, duplicateStopped: true } };
+        }
+    }
+
+    return null;
+}
+
+/**
+ * True when the worker session that emitted this event has auto-approve enabled
+ * (a MAGI/delegated worker is launched with autoApprove:true). Such a worker
+ * resolves its own approval modals locally, so an agent:waiting_approval event
+ * from it is transient noise for the coordinator: forwarding it injects a
+ * "[System] … is waiting for approval, use mesh_approve" turn that the
+ * coordinator cannot usefully act on (the modal is already auto-resolving) and,
+ * mid-MAGI-collect, HIJACKS the coordinator's synthesis turn — the observed
+ * failure where a replica's repeated auto-approvals drowned out the completion
+ * events and the coordinator answered about approvals instead of the RCA. Only
+ * suppress when we can positively confirm the source worker auto-approves; a
+ * worker that genuinely needs a human/coordinator approval (autoApprove off)
+ * still forwards so the coordinator is told.
+ */
+function sourceWorkerAutoApproves(components: DaemonComponents, sessionId: string): boolean {
+    if (!sessionId) return false;
+    try {
+        const state = components.instanceManager?.getInstance?.(sessionId)?.getState?.();
+        const settings = (state?.settings as Record<string, unknown>) || {};
+        return settings.autoApprove === true;
+    } catch {
+        return false;
+    }
+}
+
+function injectMeshSystemMessage(components: DaemonComponents, args: {
+    meshId: string;
+    sourceInstanceId?: string;
+    nodeId?: string;
+    nodeLabel: string;
+    event: string;
+    metadataEvent: Record<string, unknown>;
+    // T4 (B3b): v2 envelope restored from a remote (P2P) relay's flat payload. Only the
+    // relay path (handleMeshForwardEvent) sets it; the in-process forward path leaves it
+    // undefined and the local emit stamp applies as usual. When present with a preserved
+    // eventId, it is spread onto the re-queued pending event so stampPendingEventV2's
+    // already-stamped short-circuit keeps the ORIGINAL eventId (cross-machine idempotency).
+    v2Envelope?: Partial<Pick<PendingMeshCoordinatorEvent,
+        'protocolVersion' | 'eventId' | 'scope' | 'dispatchedBy' | 'intendedFor'>>;
+}) {
+    const eventSessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId);
+    const eventNodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+
+    // EVTTRACE correlation context for this event's coordinator-side lifecycle (queue /
+    // dedup / suppress). Observation only — never read by any decision below.
+    const traceCtx = {
+        taskId: args.metadataEvent.taskId,
+        sessionId: eventSessionId,
+        nodeId: eventNodeId,
+        meshId: args.meshId,
+        event: args.event,
+    };
+
+    const sourceSession = args.sourceInstanceId
+        ? components.instanceManager.getInstance(args.sourceInstanceId)
+        : undefined;
+    const workerCoordinatorDaemonId = readNonEmptyString(
+        (sourceSession?.getState()?.settings as Record<string, unknown>)?.meshCoordinatorDaemonId,
+    );
+    // Session-level routing anchor (multi-coordinator). Prefer the LIVE worker session's
+    // stamp; fall back to a relayed value carried in metadataEvent.meshCoordinatorSessionId
+    // (a remote worker's completion arrives via handleMeshForwardEvent with no local
+    // sourceSession, so the stamp can only ride in the relayed metadata). Empty on legacy /
+    // version-skewed dispatches → the event stays daemon-broadcast (no regression).
+    const workerCoordinatorSessionId = readNonEmptyString(
+        (sourceSession?.getState()?.settings as Record<string, unknown>)?.meshCoordinatorSessionId,
+    ) || readNonEmptyString(args.metadataEvent.meshCoordinatorSessionId);
+
+    // T2: a summary-less completion (and any non-completion status-sync event) carries no
+    // assistant text on the event, so resolveMeshSurfacedSessionPreview had nothing to surface
+    // and the coordinator's inbox mirror stayed stuck on the first dispatched user task. When
+    // THIS daemon hosts the live worker instance (sourceSession present), derive the worker's
+    // latest display message straight from its transcript and attach it to the event as
+    // lastMessagePreview/lastMessageRole/lastMessageAt. resolveMeshSurfacedSessionPreview reads
+    // these as an assistant-only fallback; they also ride the pending-queue + P2P relay
+    // (handleMeshForwardEvent whitelist) so a remote coordinator can surface them. A remote
+    // coordinator has no local instance and keeps relying on the relayed fields — unchanged.
+    const enrichedMetadataEvent = ((): Record<string, unknown> => {
+        const last = sourceSession ? getLastDisplayMessage(sourceSession.getState()) : null;
+        const base = (!last || !last.preview)
+            ? args.metadataEvent
+            : {
+                ...args.metadataEvent,
+                lastMessagePreview: last.preview,
+                lastMessageRole: last.role,
+                ...(last.receivedAt > 0 ? { lastMessageAt: last.receivedAt } : {}),
+            };
+        // MAGI: stamp the queue task's consensusGroupId onto the completion metadata
+        // so the intentional-fan-out dedup exemption (buildPendingEventFingerprint)
+        // can see it. The work queue is owned by THIS host/coordinator daemon — where
+        // both the lookup and the dedup run — so the local lookup covers local and
+        // relayed workers alike. Best-effort: never fail the event path on a miss, and
+        // never clobber a consensusGroupId the worker already relayed.
+        if (readNonEmptyString((base as Record<string, unknown>).consensusGroupId)) return base;
+        const eventTaskId = readNonEmptyString(args.metadataEvent.taskId);
+        if (!eventTaskId) return base;
+        try {
+            const entry = MeshRuntimeStore.getInstance().findQueueEntryById(args.meshId, eventTaskId);
+            const consensusGroupId = readNonEmptyString((entry as { consensusGroupId?: unknown } | null)?.consensusGroupId);
+            if (consensusGroupId) return { ...base, consensusGroupId };
+        } catch { /* queue lookup is best-effort; absence just falls back to the generic fingerprint */ }
+        return base;
+    })();
+
+    // R2: cloud P2P dashboard metadata sync. The cloud daemon used to do this from its own
+    // relay listener; now the single core forwarder invokes the injected hook (no-op on
+    // standalone) so the event path stays single-listener and the local code path is identical
+    // across standalone and cloud.
+    if (components.onMeshCoordinatorEventForwarded) {
+        try {
+            // T: the coordinator surfaces a remote worker's session but holds no local
+            // instance for it, so the status snapshot can't derive a preview and the
+            // mirror would stay stuck on the first dispatched user task. Resolve the
+            // worker's latest assistant reply (carried on the completion event's
+            // finalSummary / workerResult) into a preview the mirror can stamp, so the
+            // mobile inbox reflects the assistant response. Completion events carry assistant
+            // text as finalSummary; a summary-less completion / status sync falls back to the
+            // worker's latest assistant display message (enrichedMetadataEvent.lastMessage*).
+            // For a mid-turn user-only event this is undefined and the prior surfaced preview
+            // is preserved downstream (no clobber).
+            const surfacedPreview = resolveMeshSurfacedSessionPreview(enrichedMetadataEvent);
+            components.onMeshCoordinatorEventForwarded({
+                event: args.event,
+                meshId: args.meshId,
+                nodeId: eventNodeId || undefined,
+                ...enrichedMetadataEvent,
+                // Ensure a `workspace` field reaches updateMeshOwnedSession even when the
+                // worker provider event only carried `workspaceName`. The merge spread of
+                // metadataEvent above wins when it already has a non-empty `workspace`.
+                workspace: readNonEmptyString(args.metadataEvent.workspace)
+                    || readNonEmptyString(args.metadataEvent.workspaceName)
+                    || undefined,
+                ...(surfacedPreview ? {
+                    meshSessionLastMessagePreview: surfacedPreview.preview,
+                    meshSessionLastMessageRole: surfacedPreview.role,
+                    meshSessionLastMessageAt: surfacedPreview.receivedAt || undefined,
+                } : {}),
+            });
+        } catch { /* dashboard metadata sync is best-effort */ }
+    }
+
+    const eventTimestamp = readEventTimestamp(args.metadataEvent.timestamp);
+
+    // Auto-approving worker: never forward its approval prompts to the coordinator.
+    // The daemon resolves the modal locally, so the "[System] … waiting for
+    // approval" injection is pure noise that hijacks the coordinator's turn — the
+    // observed MAGI failure where a replica's repeated auto-approvals flooded the
+    // coordinator and derailed its final synthesis. A worker without auto-approve
+    // (genuinely blocked on a human/coordinator decision) still forwards.
+    if (args.event === 'agent:waiting_approval' && sourceWorkerAutoApproves(components, eventSessionId)) {
+        LOG.info('MeshEvents', `Suppressed agent:waiting_approval for auto-approving worker session ${eventSessionId || '(unknown)'} (mesh ${args.meshId}) — modal is resolved locally, coordinator not notified`);
+        traceMeshEventDrop('waiting_approval_auto_approving_worker', traceCtx);
+        return { success: true, forwarded: 0, suppressed: true, autoApprovingWorkerApproval: true };
+    }
+
+    // Coordinator-side dedup/suppression gate (extracted, behavior-preserving). A non-null
+    // outcome either short-circuits with a forwarded result or signals a no-progress→completion
+    // reconciliation that we re-inject; null lets the event fall through to the ledger machinery.
+    const suppression = evaluateMeshEventSuppression(args, {
+        traceCtx,
+        eventSessionId,
+        eventNodeId,
+        eventTimestamp,
+        workerCoordinatorDaemonId,
+    });
+    if (suppression) {
+        if (suppression.kind === 'reconcile') {
+            return injectMeshSystemMessage(components, {
+                ...args,
+                event: 'agent:generating_completed',
+                metadataEvent: suppression.metadataEvent,
+            });
+        }
+        return suppression.result;
+    }
+
+    function markSessionTerminal(sessionId: string, outcome: 'completed' | 'failed', occurredAtMs?: number | null, opts?: { tentativeIfDirect?: boolean }): { id?: string } | null {
+        // C2: prefer an exact taskId match when the completion event carries one —
+        // it's immune to coordinator↔worker clock skew that can hide the assigned row.
+        const eventTaskId = readNonEmptyString(args.metadataEvent.taskId) || undefined;
+        const task = updateSessionTaskStatus(args.meshId, sessionId, outcome, {
+            occurredAt: occurredAtMs != null ? new Date(occurredAtMs).toISOString() : undefined,
+            taskId: eventTaskId,
+        });
+        // Fix A (early-terminal prevention): a false-idle completion (no confirmed final
+        // assistant) for a DIRECT dispatch — i.e. no work-queue row matched — must not flip the
+        // dispatch row terminal. Leaving it active lets the reconcile loop (PHASE 4) re-read the
+        // transcript and record the genuine completion once the worker truly finishes (commonly
+        // after a coordinator nudge / re-dispatch). A matched queue task, or a completion with
+        // genuine evidence, is marked terminal as before.
+        // WARMUPGAP: a no-taskId completion from a session that holds no active assignment is a
+        // pre-assignment warmup / ghost event (a worker spawns, idles, and emits idle→generating→
+        // completed before any task is dispatched, with meshActiveTaskId unset so the event carries
+        // no taskId). Letting it through would hit the session_id fallback in updateDirectDispatchStatus
+        // and flip a sibling/stale dispatch row this event does not own — the real task later lands on
+        // a corrupted row and never reaches completed. Skip the dispatch update for that case. A
+        // taskId-carrying completion (real task), or any completion whose session currently holds an
+        // active assignment (legacy/relayed worker), still flips as before.
+        const leaveDirectDispatchActive = (!task && opts?.tentativeIfDirect === true)
+            || (!eventTaskId && !sessionHasActiveAssignment(args.meshId, sessionId));
+        if (!leaveDirectDispatchActive) {
+            // CANON-B: flip the exact dispatch row the completion echoed its taskId for; the
+            // session_id fallback (no echoed taskId) still covers legacy/relayed workers.
+            updateDirectDispatchStatus(args.meshId, sessionId, outcome, eventTaskId);
+        }
+        markSessionDeliveriesTerminal(args.meshId, sessionId, outcome);
+        // COMPLETION-PROPAGATION F2 (double safety net): the flip found no matching assigned
+        // row (task === null) but the completion echoed a taskId AND a queue row for that id is
+        // STILL 'assigned'. This is the stranded flip-miss case F1's equivalence match is meant
+        // to reconcile — NOT a direct dispatch (which legitimately has no queue row and is
+        // covered by updateDirectDispatchStatus above). Record a terminal ledger entry keyed by
+        // the echoed taskId and release the single-flight lock, so the reconcile PHASE 2.5
+        // terminal-ledger branch (findTerminalLedgerEvidenceForTask by row.id) has a taskId-based
+        // path to flip the stranded row terminal even if the direct SQL flip could not resolve
+        // it, and a subsequent reclaim/requeue is not blocked by a stale in-flight mark. Gated
+        // to GENUINE terminals (a weak / false-idle completion is left for the transcript
+        // reconcile, matching the leaveDirectDispatchActive philosophy above).
+        const genuineTerminal = outcome === 'failed' || !isWeakCompletionEvidence(args.metadataEvent);
+        if (!task && eventTaskId && genuineTerminal) {
+            try {
+                const strandedRow = MeshRuntimeStore.getInstance().findQueueEntryById(args.meshId, eventTaskId);
+                if (strandedRow && strandedRow.status === 'assigned') {
+                    endTaskDispatchInFlight(args.meshId, eventTaskId);
+                    if (!findTerminalLedgerEvidenceForTask({ meshId: args.meshId, taskId: eventTaskId })) {
+                        appendLedgerEntry(args.meshId, {
+                            kind: outcome === 'completed' ? 'task_completed' : 'task_failed',
+                            sessionId,
+                            nodeId: readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId) || undefined,
+                            providerType: readNonEmptyString(args.metadataEvent.providerType) || undefined,
+                            payload: {
+                                taskId: eventTaskId,
+                                event: args.event,
+                                source: 'flip_miss_safety_net',
+                                finalSummary: readNonEmptyString(args.metadataEvent.finalSummary) || undefined,
+                            },
+                        });
+                    }
+                }
+            } catch { /* best-effort safety net — never fail the completion path */ }
+        }
+        setImmediate(() => cleanupTerminalDirectDispatches());
+        return task ? { id: task.id } : null;
+    }
+
+    let completedTaskForLedger: { id?: string } | null = null;
+    // Fix B: direct-dispatch taskId used to attribute the terminal ledger entry when no
+    // work-queue row matches (resolved BEFORE markSessionTerminal flips the dispatch terminal).
+    let directDispatchTaskIdForLedger: string | undefined;
+    // BOOTSTRAP-MSG: whether a queued task already targets this node, so the
+    // worktree_bootstrap_complete [System] message reflects the auto-claim instead of
+    // advising a manual mesh_launch_session (which would spawn a duplicate session).
+    let worktreeHasQueuedTask = false;
+    if (args.event === 'agent:generating_completed') {
+        const sessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId);
+        const nodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+        const providerType = readNonEmptyString(args.metadataEvent.providerType);
+
+        if (sessionId) {
+            // CANON-B / ARCH-REFACTOR R1: trust the taskId the completion echoed. With R1's
+            // per-turn identity binding the worker stamps the COMPLETING turn's own taskId
+            // (not the racy session scalar), so the echoed id is authoritative and this is
+            // the path that should always be taken for an R1+ worker. The most-recent-by-
+            // session heuristic (resolveActiveDirectDispatchTaskId) is retained ONLY as a
+            // backward-compat fallback for legacy / version-skewed workers that carry no
+            // taskId — it is the very re-derive R1 exists to make unnecessary, and must not
+            // override a present echoed id, hence the `||` short-circuit order.
+            directDispatchTaskIdForLedger = readNonEmptyString(args.metadataEvent.taskId)
+                || resolveActiveDirectDispatchTaskId(args.meshId, sessionId);
+            // A false-idle completion of a direct dispatch is recorded but kept tentative (the
+            // dispatch row stays active for the reconcile fallback); a genuine completion is terminal.
+            const isFalseIdle = isFalseIdleCompletion(args.metadataEvent);
+            completedTaskForLedger = markSessionTerminal(sessionId, 'completed', eventTimestamp, { tentativeIfDirect: isFalseIdle });
+            if (nodeId && providerType) {
+                // OVEREAGER-REMOTE-IDLE (Defect A+B): re-register the now-idle remote session into
+                // the remote-idle store, symmetric with the agent:ready branch. Previously
+                // setRemoteIdleSession ran ONLY on agent:ready, while agent:generating_started
+                // DELETES the entry — so the FIRST turn a remote worker runs permanently evicts it
+                // from the store, and generating_completed never re-added it. A later
+                // mesh_enqueue_task's triggerMeshQueue then read getRemoteIdleSessions() == 0
+                // (remoteIdleSessionsChecked:0) for a session that is genuinely live-idle, needlessly
+                // auto-launching a second worker (Defect A). Worse, the two idle-session sources then
+                // disagreed: the enqueue drain saw 0 (store empty) and auto-launched + left the queue
+                // task pending, while THIS completing session's runIdleMaintenanceThenAssignQueue
+                // claimed the same still-pending row straight from SQL — so the task body injected into
+                // BOTH the reused idle session and the auto-launched one (Defect B). Re-registering here
+                // unifies the source: the next triggerMeshQueue drain sees the live idle session, reuses
+                // it (claimed:true, no auto-launch), and injects exactly once. Skip a false-idle
+                // (mid-turn / no-final-assistant) completion — that session is NOT genuinely idle.
+                if (!isFalseIdle) {
+                    sweepExpiredRemoteIdleSessions();
+                    try {
+                        MeshRuntimeStore.getInstance().setRemoteIdleSession(args.meshId, nodeId, sessionId, providerType, Date.now() + REMOTE_IDLE_SESSION_TTL_MS);
+                    } catch { /* best-effort */ }
+                    setImmediate(() => {
+                        maybeAutoFastForwardIdleNode(components, { meshId: args.meshId, nodeId, sessionId, providerType })
+                            .finally(() => {
+                                try {
+                                    // Claim for THIS session first; on success drop the just-registered
+                                    // idle entry so the enqueue drain doesn't re-pick an already-busy session.
+                                    const assigned = tryAssignQueueTask(components, args.meshId, nodeId, sessionId, providerType);
+                                    if (assigned) MeshRuntimeStore.getInstance().deleteRemoteIdleSession(args.meshId, nodeId, sessionId);
+                                } catch (e: any) {
+                                    LOG.warn('MeshQueue', `Failed to assign idle queue task after completion for ${nodeId}: ${e?.message || e}`);
+                                }
+                            });
+                    });
+                } else {
+                    runIdleMaintenanceThenAssignQueue(components, { meshId: args.meshId, nodeId, sessionId, providerType });
+                }
+            }
+            // M1-3: wake dependents of the completed task. The maintenance path above
+            // only assigns to the completing session; dependents may be claimable by
+            // other idle sessions, so run a full queue trigger when any are waiting.
+            const completedTaskId = completedTaskForLedger?.id;
+            if (completedTaskId && hasPendingDependents(args.meshId, completedTaskId)) {
+                setImmediate(() => {
+                    triggerMeshQueue(components, args.meshId).catch((e: any) => {
+                        LOG.warn('MeshQueue', `Dependent wake after task ${completedTaskId} failed: ${e?.message || e}`);
+                    });
+                });
+            }
+        }
+    } else if (args.event === 'agent:ready') {
+        const sessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId);
+        const nodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+        const providerType = readNonEmptyString(args.metadataEvent.providerType);
+        const providerSessionId = readNonEmptyString(args.metadataEvent.providerSessionId) || undefined;
+        const finalSummary = readNonEmptyString(args.metadataEvent.finalSummary) || undefined;
+        const workerResult = readWorkerResultMetadata(args.metadataEvent);
+        const hasCompletionEvidence = !!finalSummary || !!workerResult;
+        if (sessionId && hasCompletionEvidence) {
+            completedTaskForLedger = markSessionTerminal(sessionId, 'completed');
+            if (completedTaskForLedger) {
+                try {
+                    appendLedgerEntry(args.meshId, {
+                        kind: 'task_completed',
+                        nodeId: nodeId || undefined,
+                        sessionId,
+                        providerType: providerType || undefined,
+                        payload: {
+                            event: args.event,
+                            nodeLabel: args.nodeLabel,
+                            taskId: completedTaskForLedger.id,
+                            completedViaReady: true,
+                            providerSessionId,
+                            finalSummary,
+                            workerResult,
+                            evidence: buildTaskCompletionEvidence({
+                                event: 'agent:ready',
+                                nodeId,
+                                sessionId,
+                                providerType: providerType || undefined,
+                                providerSessionId,
+                                finalSummary,
+                                workerResult,
+                            }),
+                        },
+                    });
+                } catch (e: any) {
+                    LOG.warn('MeshLedger', `Failed to record task_completed from ready: ${e?.message || e}`);
+                }
+            }
+        }
+
+        if (sessionId && nodeId && providerType) {
+            // WORKTREE-BOOTSTRAP-DISPATCH-RACE: a freshly cloned worktree session emits
+            // agent:ready as soon as the CLI process reaches its idle prompt — which happens
+            // BEFORE the worktree bootstrap (npm install + native-addon repair: node-datachannel,
+            // better-sqlite3, ghostty-vt-node) has finished. Claiming a queued task on that early
+            // ready dispatches work into a session whose runtime is half-built: the child daemon
+            // dies loading the absent native addon → the worker session boots empty (no task ever
+            // injected), the queue row reports assigned-but-dead, and the work silently leaks to /
+            // gets re-routed onto the base node. Live evidence (the OPSRULES dispatch): node
+            // node_6df455dd emitted agent:ready at 10:36:28 but worktree_bootstrap_complete only at
+            // 10:36:56 — a 28s window in which a claim produced an empty session.
+            //
+            // Gate the claim on bootstrap state: if this node is a worktree whose bootstrap is still
+            // 'running', register the idle session (so it stays a claim candidate) but DEFER the
+            // claim. The claim re-fires on the next agent:ready / reconcile drain tick once bootstrap
+            // reaches a terminal state. 'failed' is left to flow through unchanged — a failed
+            // bootstrap surfaces its own coordinator event and a dispatch there fails loudly rather
+            // than silently, which is the correct, visible behavior (deferring forever would hide it).
+            let worktreeBootstrapPending = false;
+            try {
+                const mesh = getMeshWithCache(components, args.meshId);
+                const node = mesh?.nodes?.find((n: any) => meshNodeIdMatches(n, nodeId)) as { worktreeBootstrap?: { status?: string } } | undefined;
+                worktreeBootstrapPending = node?.worktreeBootstrap?.status === 'running';
+            } catch { /* best-effort: unknown bootstrap state → do not defer (prior behavior) */ }
+
+            sweepExpiredRemoteIdleSessions();
+            try {
+                MeshRuntimeStore.getInstance().setRemoteIdleSession(args.meshId, nodeId, sessionId, providerType, Date.now() + REMOTE_IDLE_SESSION_TTL_MS);
+            } catch { /* best-effort */ }
+            if (worktreeBootstrapPending) {
+                LOG.info('MeshQueue', `Deferring queue claim for worktree node ${nodeId} (${sessionId}): worktree bootstrap still running — early agent:ready precedes bootstrap completion; the idle session is registered and the claim will re-fire once bootstrap finishes (guards against dispatching into a half-built worktree → empty session / work leaking to the base node)`);
+            } else {
+                setImmediate(() => {
+                    maybeAutoFastForwardIdleNode(components, { meshId: args.meshId, nodeId, sessionId, providerType })
+                        .finally(() => {
+                            try {
+                                const assigned = tryAssignQueueTask(components, args.meshId, nodeId, sessionId, providerType);
+                                if (assigned) MeshRuntimeStore.getInstance().deleteRemoteIdleSession(args.meshId, nodeId, sessionId);
+                            } catch (e: any) {
+                                LOG.warn('MeshQueue', `Failed to assign idle queue task after maintenance for ${nodeId}: ${e?.message || e}`);
+                            }
+                        });
+                });
+            }
+        }
+    } else if (args.event === 'agent:generating_started') {
+        const sessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId);
+        const nodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+        if (sessionId && nodeId) {
+            try {
+                MeshRuntimeStore.getInstance().deleteRemoteIdleSession(args.meshId, nodeId, sessionId);
+            } catch { /* best-effort */ }
+        }
+        if (sessionId) {
+            // CANON-B: a generating_started that echoes its taskId acks exactly the dispatch
+            // and the delivery for THAT task — not every in-flight dispatch/delivery on the
+            // session. A session that already holds a freshly-dispatched (still 'dispatched')
+            // sibling must keep that row 'dispatched' so its own confirm can match it; acking
+            // by session would mark it 'acked' prematurely and hide a genuine non-delivery.
+            const startedTaskId = readNonEmptyString(args.metadataEvent.taskId) || undefined;
+            // WARMUPGAP: only ack a dispatch row when the event names its task, or the session
+            // currently holds an active assignment. A no-taskId generating_started from an
+            // unassigned session is a pre-assignment warmup — the session_id fallback would ack a
+            // sibling/stale dispatch row this event does not own, marking it 'acked' prematurely and
+            // hiding a genuine non-delivery. Skip the dispatch ack for that ghost case (the delivery
+            // acks below are bound to actual deliveries and stay a no-op for a warmup session).
+            //
+            // MESH-DISPATCH-MISROUTE (fix 3, consumer residual): when the event carries no taskId
+            // (a legacy/relayed worker whose producer never stamped meshActiveTaskId) but the
+            // session owns EXACTLY ONE active dispatch, resolve that row's taskId and flip it by PK
+            // instead of the session_id sweep — the sweep flips every non-terminal row for the
+            // session ("may flip a sibling dispatch row"). With ≥2 active rows the owner is
+            // ambiguous, so resolvedAckTaskId stays undefined and we DROP the ack rather than
+            // mis-flip a sibling (the genuine ack arrives once the producer/reconcile names a task).
+            if (startedTaskId) {
+                updateDirectDispatchStatus(args.meshId, sessionId, 'acked', startedTaskId);
+            } else if (sessionHasActiveAssignment(args.meshId, sessionId)) {
+                const soleTaskId = (() => {
+                    try { return MeshRuntimeStore.getInstance().getSoleActiveDirectDispatchTaskId(args.meshId, sessionId); }
+                    catch { return null; }
+                })();
+                if (soleTaskId) {
+                    updateDirectDispatchStatus(args.meshId, sessionId, 'acked', soleTaskId);
+                }
+            }
+            const activeDeliveries = ((): { id: string; taskId: string | null }[] => {
+                try { return MeshRuntimeStore.getInstance().getActiveSessionDeliveries(args.meshId, sessionId); }
+                catch { return []; }
+            })();
+            const deliveriesToAck = startedTaskId
+                ? activeDeliveries.filter(d => d.taskId === startedTaskId)
+                : activeDeliveries;
+            for (const d of deliveriesToAck) {
+                updateSessionDeliveryStatus(d.id, 'acked');
+            }
+        }
+    } else if (args.event === 'agent:stopped') {
+        const sessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId);
+        const nodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+        if (sessionId && nodeId) {
+            try {
+                MeshRuntimeStore.getInstance().deleteRemoteIdleSession(args.meshId, nodeId, sessionId);
+            } catch { /* best-effort */ }
+        }
+        if (sessionId) {
+            // CANON-B: prefer the echoed taskId; session heuristic is the fallback.
+            directDispatchTaskIdForLedger = readNonEmptyString(args.metadataEvent.taskId)
+                || resolveActiveDirectDispatchTaskId(args.meshId, sessionId);
+            completedTaskForLedger = markSessionTerminal(sessionId, 'failed');
+        }
+    } else if (args.event === 'worktree_bootstrap_complete' || args.event === 'worktree_bootstrap_failed') {
+        // WORKTREE-BOOTSTRAP-REFIRE: the agent:ready branch above DEFERS the queue claim while
+        // worktreeBootstrap.status === 'running' — a freshly cloned worktree emits agent:ready at
+        // its idle prompt BEFORE bootstrap finishes, and dispatching then produces a half-built
+        // session. The defer registers the idle session (setRemoteIdleSession) but schedules NO
+        // retry; it relies on "the next agent:ready / reconcile drain tick" to re-fire. But
+        // agent:ready is a one-shot (emitAgentReadyOnce, guarded by agentReadyEmitted) and the
+        // session stays idle→idle after bootstrap, so no new agent:ready edge ever fires → the
+        // deferred claim is stranded → the worker session boots empty (totalMessages=0) and the
+        // coordinator relaunch/stop-loops. Re-fire the queue drain on the terminal bootstrap
+        // transition: the idle session registered at defer time is now claimable (tryAssignQueueTask
+        // has no bootstrap gate of its own). This runs on whichever daemon processes the event —
+        // the local worker (when it co-hosts the coordinator) at emit time, or the remote
+        // coordinator after it pulls the queued event — so the registered local/remote idle session
+        // is drained in every topology. 'failed' is re-fired too so a deferred-then-failed bootstrap
+        // dispatches and fails loudly/visibly rather than stranding silently. Falls through to the
+        // coordinator broadcast below — the bootstrap event is still delivered to the coordinator.
+        // WORKTREE-BOOTSTRAP-COORD-STATE: stamp the terminal bootstrap status onto the
+        // COORDINATOR's mesh view BEFORE re-firing the queue. The clone+bootstrap ran on
+        // the worker daemon (clone_mesh_node forwards to the source node's machine), so
+        // persistWorktreeSetupState only flipped status→'complete' on the worker's mesh
+        // object — the coordinator still holds the 'running' state it stamped from the
+        // forwarded clone reply. Without this, the claim gate (agent:ready defer above +
+        // mesh-queue-assignment) reads getMeshWithCache, sees 'running' forever, and
+        // defers every claim — so this very re-fire loops against a gate that never opens
+        // (claim never lands; idle session re-registered each tick; auto-launch spawns a
+        // fresh session every cycle → runaway worktree-session multiplication). Stamping
+        // the terminal state opens the gate so the deferred claim lands on this re-fire.
+        const bootstrapNodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId);
+        if (bootstrapNodeId) {
+            try {
+                // Fix (3): pass the worktree path so a hydrate-on-miss upsert (when this
+                // coordinator never received the clone reply) can seed an addressable node.
+                const bootstrapWorkspace = readNonEmptyString(args.metadataEvent.worktreePath)
+                    || readNonEmptyString(args.metadataEvent.workspace);
+                (components.router as any)?.markWorktreeBootstrapTerminalState?.(
+                    args.meshId,
+                    bootstrapNodeId,
+                    args.event === 'worktree_bootstrap_failed' ? 'failed' : 'complete',
+                    bootstrapWorkspace ? { workspace: bootstrapWorkspace } : undefined,
+                );
+            } catch (e: any) {
+                LOG.warn('MeshQueue', `Failed to stamp terminal bootstrap state for ${bootstrapNodeId} (mesh ${args.meshId}): ${e?.message || e}`);
+            }
+        }
+        // BOOTSTRAP-MSG: detect whether the queue re-fire below has a task to auto-claim for
+        // this node BEFORE setImmediate(triggerMeshQueue) runs — at this point a deferred task
+        // is still 'pending' (it has not been claimed yet); 'assigned' covers a task already
+        // claimed by an earlier tick. Either case means a session is being/has been spun up by
+        // the queue, so the completion message must NOT advise a manual mesh_launch_session.
+        // Only meaningful for the 'complete' transition (a failed bootstrap claims nothing).
+        if (args.event === 'worktree_bootstrap_complete' && bootstrapNodeId) {
+            try {
+                worktreeHasQueuedTask = getQueue(args.meshId, { status: ['pending', 'assigned'] })
+                    .some((task) => meshNodeIdMatches({ id: task.targetNodeId } as MeshNodeIdentified, bootstrapNodeId));
+            } catch (e: any) {
+                LOG.warn('MeshQueue', `Failed to check queued task for ${bootstrapNodeId} (mesh ${args.meshId}): ${e?.message || e}`);
+            }
+        }
+        setImmediate(() => {
+            triggerMeshQueue(components, args.meshId).catch((e: any) => {
+                LOG.warn('MeshQueue', `Queue re-fire after ${args.event} failed (mesh ${args.meshId}): ${e?.message || e}`);
+            });
+        });
+    }
+
+    const ledgerKind = EVENT_TO_LEDGER_KIND[args.event];
+    if (ledgerKind) {
+        try {
+            const ledgerNodeId = readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId) || undefined;
+            const ledgerSessionId = resolveEventSessionId(args.metadataEvent, args.sourceInstanceId) || undefined;
+            const ledgerProviderType = readNonEmptyString(args.metadataEvent.providerType) || undefined;
+            const providerSessionId = readNonEmptyString(args.metadataEvent.providerSessionId) || undefined;
+            const finalSummary = readNonEmptyString(args.metadataEvent.finalSummary) || undefined;
+            const workerResult = readWorkerResultMetadata(args.metadataEvent);
+            const completionEvidence = ledgerKind === 'task_completed' && ledgerNodeId && ledgerSessionId
+                ? buildTaskCompletionEvidence({
+                    event: 'agent:generating_completed',
+                    nodeId: ledgerNodeId,
+                    sessionId: ledgerSessionId,
+                    providerType: ledgerProviderType,
+                    providerSessionId,
+                    finalSummary,
+                    workerResult,
+                })
+                : undefined;
+            appendLedgerEntry(args.meshId, {
+                kind: ledgerKind,
+                nodeId: ledgerNodeId,
+                sessionId: ledgerSessionId,
+                providerType: ledgerProviderType,
+                payload: {
+                    event: args.event,
+                    nodeLabel: args.nodeLabel,
+                    // Fix B: fall back to the direct-dispatch taskId when no work-queue row
+                    // matched, so the terminal entry is attributable in mesh task-stats
+                    // (otherwise the direct task shows status='unknown' / terminalKind=null).
+                    taskId: completedTaskForLedger?.id || directDispatchTaskIdForLedger || undefined,
+                    providerSessionId,
+                    finalSummary,
+                    workerResult,
+                    completionDiagnostic: args.metadataEvent.completionDiagnostic && typeof args.metadataEvent.completionDiagnostic === 'object'
+                        ? args.metadataEvent.completionDiagnostic
+                        : undefined,
+                    evidence: completionEvidence,
+                    // B2: evidenceLevel lets coordinator know when completion evidence is insufficient.
+                    // NOTIF Defect-2b: ONLY source==='default' (no parseable answer at all) is
+                    // 'insufficient'. 'parseable_answer' (a real JSON answer that just isn't
+                    // worker-result-shaped, e.g. a MAGI envelope) is concrete evidence and must
+                    // resolve to 'sufficient' — resolveWorkerResult now upgrades that case so a
+                    // complete, valid answer is no longer mislabelled insufficient/reviewRecommended.
+                    ...(completionEvidence
+                        ? completionEvidence.workerResult.source === 'default'
+                            ? { evidenceLevel: 'insufficient', reviewRecommended: true }
+                            : { evidenceLevel: 'sufficient' }
+                        : {}),
+                },
+            });
+        } catch (e: any) {
+            LOG.warn('MeshLedger', `Failed to record ${ledgerKind}: ${e?.message || e}`);
+        }
+    }
+
+    let recoveryContext: SessionRecoveryContext | null = null;
+    if (args.event === 'agent:stopped') {
+        try {
+            const mesh = getMesh(args.meshId);
+            const maxRetries = mesh?.policy?.maxTaskRetries ?? 1;
+
+            recoveryContext = getSessionRecoveryContext(args.meshId, {
+                sessionId: resolveEventSessionId(args.metadataEvent, args.sourceInstanceId) || undefined,
+                nodeId: readNonEmptyString(args.nodeId) || readNonEmptyString(args.metadataEvent.meshNodeId) || undefined,
+                maxRetries,
+            });
+            recoveryContext.failedProviderType = readNonEmptyString(args.metadataEvent.providerType) || null;
+
+            if (recoveryContext.retryRecommended && recoveryContext.consecutiveNodeFailures > 0) {
+                appendLedgerEntry(args.meshId, {
+                    kind: 'recovery_attempted',
+                    nodeId: recoveryContext.failedNodeId || undefined,
+                    sessionId: recoveryContext.failedSessionId || undefined,
+                    providerType: recoveryContext.failedProviderType || undefined,
+                    payload: {
+                        consecutiveFailures: recoveryContext.consecutiveNodeFailures,
+                        taskAttemptCount: recoveryContext.taskAttemptCount,
+                        retryRecommended: recoveryContext.retryRecommended,
+                        advice: recoveryContext.advice,
+                    },
+                });
+
+                if (recoveryContext.lastTaskMessage && recoveryContext.failedNodeId && recoveryContext.failedProviderType) {
+                    const autoNodeId = recoveryContext.failedNodeId;
+                    try {
+                        const task = enqueueTask(args.meshId, recoveryContext.lastTaskMessage, {
+                            targetNodeId: autoNodeId
+                        });
+                        LOG.info('MeshRecovery', `Auto-requeued failed task: ${task.id} for node ${autoNodeId}`);
+
+                        const node = mesh?.nodes.find((n: any) => meshNodeIdMatches(n, autoNodeId));
+                        if (node) {
+                            components.cliManager.handleCliCommand('launch_cli', {
+                                cliType: recoveryContext.failedProviderType,
+                                dir: node.workspace,
+                                settings: {
+                                    role: 'worker',
+                                    meshNodeFor: args.meshId,
+                                    meshNodeId: node.id,
+                                    spawnedSessionVisibility: mesh?.policy?.spawnedSessionVisibility || 'hidden',
+                                    // Coordinator-dispatched recovery relaunch: same auto-approve
+                                    // policy as the primary worker launch path.
+                                    autoApprove: resolveDelegatedWorkerAutoApprove(mesh?.policy, node?.policy),
+                                    launchedByCoordinator: true,
+                                }
+                            }).catch((e: any) => LOG.error('MeshRecovery', `Failed to auto-relaunch session for ${node.id}: ${e?.message}`));
+                        }
+                    } catch (e: any) {
+                        LOG.warn('MeshRecovery', `Failed to execute auto-recovery: ${e?.message}`);
+                    }
+                }
+            }
+
+            LOG.info('MeshRecovery', `Recovery context for ${args.nodeLabel}: ${recoveryContext.advice}`);
+        } catch (e: any) {
+            LOG.warn('MeshRecovery', `Failed to build recovery context: ${e?.message || e}`);
+        }
+    }
+
+    const messageText = buildMeshSystemMessage({
+        event: args.event,
+        nodeLabel: args.nodeLabel,
+        metadataEvent: args.metadataEvent,
+        recoveryContext,
+        worktreeHasQueuedTask,
+    });
+    if (!messageText) {
+        // Lifecycle events that carry no coordinator-facing message (agent:ready /
+        // agent:generating_started) still drive the remote-claim state machine: the
+        // coordinator's agent:ready branch above runs setRemoteIdleSession +
+        // tryAssignQueueTask, and agent:generating_started clears the remote-idle entry.
+        // For a LOCAL worker whose coordinator is a REMOTE daemon those side effects ran
+        // on the wrong daemon (this worker's empty queue / store), so the coordinator never
+        // learns the auto-launched session went idle and re-auto-launches it forever
+        // (queue task stuck pending). Queue the silent event so the coordinator pulls it
+        // (PHASE 1 pullRemoteNodeQueues → handleMeshForwardEvent) and re-runs the claim on
+        // the daemon that actually owns the queue. Gate strictly on a present, REMOTE
+        // coordinator daemon id: a co-located worker already ran the claim on the right
+        // daemon, and a coordinator processing a *pulled* event has no sourceSession so
+        // workerCoordinatorDaemonId is empty — neither re-queues, so there is no loop.
+        const isSilentClaimRelevantEvent = args.event === 'agent:ready' || args.event === 'agent:generating_started';
+        const coordinatorIsRemote = !!workerCoordinatorDaemonId
+            && !resolveCoordinatorDrainDaemonIds(components).includes(workerCoordinatorDaemonId);
+        if (!(isSilentClaimRelevantEvent && coordinatorIsRemote)) {
+            return { success: false, error: 'unsupported mesh event' };
+        }
+    }
+
+    // ── Queue-only delivery (single-model: queue + periodic poll) ──────────────
+    // Every mesh coordinator event — terminal or not, local-coordinator or
+    // remote — is persisted to the pending-events queue (SQLite + JSONL) and
+    // NOTHING is pushed here. The old spontaneous-forward paths were removed:
+    //   - F1 remote P2P `mesh_forward_event` dispatch (network/stamp-dependent,
+    //     silently dropped on P2P failure or missing meshCoordinatorDaemonId)
+    //   - F3 live-CLI PTY `send_message` fire-and-forget inject (silently
+    //     dropped when the coordinator was generating)
+    // Delivery to a live CLI coordinator now happens via setupMeshReconcileLoop,
+    // which drains this queue on a fixed interval and injects into the coordinator
+    // only when it is idle. A pure stdio MCP (LLM) coordinator — which has no live
+    // CLI session to inject into — drains the queue itself when it calls a mesh
+    // tool (mesh_status / mesh_read_chat). Either way the queue is the single
+    // source of truth and the only thing this function writes to.
+    //
+    // targetCoordinatorDaemonId scopes the event to a specific coordinator daemon
+    // (unicast) when the worker carries one, so the reconcile loop on the right
+    // daemon drains it and other daemons skip it. Absent → broadcast/backfill.
+    const pendingEvent = {
+        event: args.event,
+        meshId: args.meshId,
+        nodeLabel: args.nodeLabel,
+        nodeId: args.nodeId || undefined,
+        workspace: readNonEmptyString(args.metadataEvent.workspace)
+            || readNonEmptyString(args.metadataEvent.workspaceName),
+        metadataEvent: {
+            ...enrichedMetadataEvent,
+            ...(recoveryContext ? { recoveryContext } : {}),
+            // Stash the coordinator session id INSIDE metadataEvent too, so it survives the
+            // P2P relay serialization (buildForwardPayloadFromPending spreads metadata; the
+            // handleMeshForwardEvent whitelist reads it back) — a top-level field alone would
+            // be dropped when the event crosses a machine boundary.
+            ...(workerCoordinatorSessionId ? { meshCoordinatorSessionId: workerCoordinatorSessionId } : {}),
+        },
+        // Silent lifecycle events (agent:ready / agent:generating_started) carry no
+        // coordinator message; they are queued only so the coordinator re-runs the
+        // remote-claim state machine on pull. injectPendingIntoCoordinator skips
+        // entries without a coordinatorMessage, so a live CLI coordinator is not spammed.
+        ...(messageText ? { coordinatorMessage: messageText } : {}),
+        queuedAt: Date.now(),
+        ...(workerCoordinatorDaemonId ? { targetCoordinatorDaemonId: workerCoordinatorDaemonId } : {}),
+        // Top-level session anchor for the local PHASE 2 strict-match on the coordinator
+        // daemon. Absent → daemon-level broadcast (legacy / single-coordinator path).
+        ...(workerCoordinatorSessionId ? { targetCoordinatorSessionId: workerCoordinatorSessionId } : {}),
+        // T4 (B3b): restore the v2 envelope from the remote relay so the re-queue keeps the
+        // ORIGINAL eventId. Spread LAST so its authoritative eventId/scope/identity win over
+        // any default. queuePendingMeshCoordinatorEvent → stampPendingEventV2 then no-ops
+        // (already-stamped short-circuit) instead of minting a fresh eventId. Empty object
+        // for a v1 relay → unchanged v1 emit-stamp path (version-skew safe).
+        ...(args.v2Envelope ?? {}),
+    };
+    if (queuePendingMeshCoordinatorEvent(pendingEvent)) {
+        LOG.info('MeshEvents', `Queued ${args.event} for coordinator (mesh ${args.meshId}${workerCoordinatorDaemonId ? `, coordinator daemon ${workerCoordinatorDaemonId}` : ''}${workerCoordinatorSessionId ? `, coordinator session ${workerCoordinatorSessionId}` : ''})`);
+        // EVTTRACE: event persisted to the coordinator pending queue (awaiting reconcile drain).
+        traceMeshEventStage('queued', traceCtx, workerCoordinatorDaemonId ? `coordinatorDaemon=${workerCoordinatorDaemonId}` : 'broadcast');
+    } else {
+        // EVTTRACE: queue rejected the event (dedup at queue time / persistence guard).
+        traceMeshEventDrop('queue_dedup', traceCtx);
+    }
+    return { success: true, forwarded: 0 };
+}
+
+// Reconstruct the metadataEvent that injectMeshSystemMessage consumes from a forwarded
+// (cross-machine) mesh event. The remote relay hop arrives as a flat payload, NOT the
+// original provider event object, so this whitelists the fields the coordinator-side
+// pipeline reads and re-projects them. Kept pure + exported so the relay-path field
+// preservation (esp. taskId) is unit-testable without driving injectMeshSystemMessage.
+//
+// IMPORTANT asymmetry: the LOCAL in-process forward path (onMeshCoordinatorEventForwarded)
+// passes the whole event through as metadataEvent, so every field on the event survives
+// there for free. This remote-only path must explicitly mirror each field it needs.
+export function buildRelayMetadataEvent(payload: Record<string, unknown>): Record<string, unknown> {
+    const relayModalMessage = readNonEmptyString(payload.modalMessage);
+    const relayModalButtons = Array.isArray(payload.modalButtons)
+        ? (payload.modalButtons as unknown[]).filter((b): b is string => typeof b === 'string' && b.trim().length > 0)
+        : null;
+    return {
+        // Preserve the dispatch task id across the machine boundary. The `received` trace
+        // stage reads payload.taskId; without mirroring it here the rebuilt metadataEvent
+        // loses it, so injectMeshSystemMessage's traceCtx.taskId and the
+        // updateDirectDispatchStatus(eventTaskId) call go undefined — the EvtTrace
+        // queued/surfaced stages show task=- and the direct-dispatch ledger falls back to a
+        // session_id match (which can flip a sibling row). The local in-process forward path
+        // keeps event.taskId/meshActiveTaskId for free; this mirrors it for the remote relay.
+        // Same taskId/meshActiveTaskId ordering the local unroutable trace uses.
+        taskId: readNonEmptyString(payload.taskId) || readNonEmptyString(payload.meshActiveTaskId),
+        targetSessionId: readNonEmptyString(payload.targetSessionId) || readNonEmptyString(payload.sessionId) || readNonEmptyString(payload.instanceId),
+        providerType: readNonEmptyString(payload.providerType),
+        providerSessionId: readNonEmptyString(payload.providerSessionId),
+        // Preserve the originating coordinator SESSION id across the machine boundary so
+        // the completion routes back to the exact coordinator session (multi-coordinator).
+        // buildForwardPayloadFromPending spreads the worker event's metadata, so the id
+        // arrives as payload.meshCoordinatorSessionId; the top-level targetCoordinatorSessionId
+        // is also accepted as a fallback. injectMeshSystemMessage re-derives the routing
+        // anchors from this. Absent → daemon-level fallback (version-skew safe).
+        meshCoordinatorSessionId: readNonEmptyString(payload.meshCoordinatorSessionId) || readNonEmptyString(payload.targetCoordinatorSessionId),
+        // Carry the session identity fields the worker provider event emits so the
+        // coordinator's mirror (updateMeshOwnedSession) gets a real workspace/title/
+        // settings. Without these the remote-relay hop reconstructs metadataEvent with
+        // an empty workspace, and the dashboard flaps to the generic
+        // "Terminal (Mesh Node)" title (and degrades the provider label) between live
+        // events and the periodic get_status_metadata snapshot. The local in-process
+        // forward path (onMeshCoordinatorEventForwarded) already preserves these; this
+        // mirrors them for the remote-only relay path.
+        workspace: readNonEmptyString(payload.workspace) || readNonEmptyString(payload.workspaceName),
+        workspaceName: readNonEmptyString(payload.workspaceName) || readNonEmptyString(payload.workspace),
+        sessionTitle: readNonEmptyString(payload.sessionTitle),
+        sessionStatus: readNonEmptyString(payload.sessionStatus),
+        sessionChatStatus: readNonEmptyString(payload.sessionChatStatus),
+        providerName: readNonEmptyString(payload.providerName),
+        ...(payload.sessionSettings && typeof payload.sessionSettings === 'object' && !Array.isArray(payload.sessionSettings) ? { sessionSettings: payload.sessionSettings } : {}),
+        finalSummary: readNonEmptyString(payload.finalSummary) || readNonEmptyString(payload.summary),
+        // T2: carry the worker's status-snapshot last-message preview across the machine
+        // boundary so a summary-less completion still surfaces the assistant reply in the
+        // coordinator's inbox mirror. resolveMeshSurfacedSessionPreview reads these
+        // (assistant-role only) when finalSummary is absent.
+        lastMessagePreview: readNonEmptyString(payload.lastMessagePreview),
+        lastMessageRole: readNonEmptyString(payload.lastMessageRole),
+        ...(payload.lastMessageAt !== undefined ? { lastMessageAt: payload.lastMessageAt } : {}),
+        jobId: readNonEmptyString(payload.jobId),
+        interactionId: readNonEmptyString(payload.interactionId),
+        status: readNonEmptyString(payload.status),
+        targetDaemonId: readNonEmptyString(payload.targetDaemonId),
+        startedAt: readNonEmptyString(payload.startedAt),
+        completedAt: readNonEmptyString(payload.completedAt),
+        retryOfJobId: readNonEmptyString(payload.retryOfJobId),
+        ...(relayModalMessage ? { modalMessage: relayModalMessage } : {}),
+        ...(relayModalButtons && relayModalButtons.length > 0 ? { modalButtons: relayModalButtons } : {}),
+        ...(payload.result && typeof payload.result === 'object' && !Array.isArray(payload.result) ? { result: payload.result } : {}),
+        ...(payload.completionDiagnostic && typeof payload.completionDiagnostic === 'object' && !Array.isArray(payload.completionDiagnostic) ? { completionDiagnostic: payload.completionDiagnostic } : {}),
+        ...(payload.workerResult && typeof payload.workerResult === 'object' && !Array.isArray(payload.workerResult) ? { workerResult: payload.workerResult } : {}),
+        ...(payload.meshWorkerResult && typeof payload.meshWorkerResult === 'object' && !Array.isArray(payload.meshWorkerResult) ? { meshWorkerResult: payload.meshWorkerResult } : {}),
+        ...(payload.structuredResult && typeof payload.structuredResult === 'object' && !Array.isArray(payload.structuredResult) ? { structuredResult: payload.structuredResult } : {}),
+        ...(payload.timestamp !== undefined ? { timestamp: payload.timestamp } : {}),
+        intentional: payload.intentional === true,
+        intentionalStop: payload.intentionalStop === true,
+        operatorCleanup: payload.operatorCleanup === true,
+        reason: readNonEmptyString(payload.reason),
+        stopReason: readNonEmptyString(payload.stopReason),
+        cleanupReason: readNonEmptyString(payload.cleanupReason),
+        source: readNonEmptyString(payload.source),
+    };
+}
+
+export function handleMeshForwardEvent(components: DaemonComponents, payload: Record<string, unknown>) {
+    const eventName = readNonEmptyString(payload.event);
+    if (!isMeshCoordinatorEvent(eventName)) {
+        return { success: false, error: 'unsupported mesh event' };
+    }
+    const nodeId = readNonEmptyString(payload.nodeId);
+    const workspace = readNonEmptyString(payload.workspace);
+
+    // The fallback worker-forward path (forwardUnresolvedDelegateEvent) cannot resolve a
+    // mesh id locally on the remote worker, so it forwards the event with nodeId +
+    // workspace only. The coordinator hosting the mesh CAN resolve it. Two recovery
+    // paths, in order:
+    //   1) workspace → mesh (fast path; cached repoIdentity lookup), then
+    //   2) nodeId → mesh (deterministic backstop; scans hosted meshes for the node).
+    // Workspace recovery alone was unreliable — a worktree clone whose repoIdentity
+    // differs, or a transient cache miss, left the reconcile retry permanently rejected
+    // ("meshId required") so the worker's completion never surfaced to the coordinator.
+    // The nodeId is a stable coordinator-side fact and resolves timing-independently.
+    const meshId = readNonEmptyString(payload.meshId)
+        || (workspace ? readNonEmptyString(getCachedMeshByWorkspace(workspace)?.id) : '')
+        || recoverMeshIdByNodeId(nodeId)
+        // Fix B last resort: workspace + nodeId scan both missed, but the worker stamped
+        // its coordinator anchor onto the forward (forwardUnresolvedDelegateEvent). Recover
+        // via the hosted mesh that anchor owns (+ node disambiguation), guarding ambiguity.
+        || recoverMeshIdByCoordinatorAndNode(
+            readNonEmptyString(payload.meshCoordinatorDaemonId) || readNonEmptyString(payload.coordinatorDaemonId),
+            nodeId,
+        );
+    if (!meshId) {
+        // EVTTRACE: forwarded event rejected at receive — no meshId could be resolved
+        // (no payload.meshId, no workspace→mesh, no nodeId→mesh). Observation only.
+        traceMeshEventDrop('meshId_required', {
+            taskId: payload.taskId,
+            sessionId: readNonEmptyString(payload.targetSessionId) || readNonEmptyString(payload.sessionId),
+            nodeId,
+            event: eventName,
+        }, workspace ? `workspace=${workspace} unresolved` : 'no workspace/nodeId');
+        return { success: false, error: 'meshId required' };
+    }
+    // EVTTRACE: forwarded event accepted at receive (meshId resolved).
+    // NOTIF-MISS (FIX 3): mirror buildRelayMetadataEvent's taskId fallback. A worker provider
+    // completion stamps its task id as `meshActiveTaskId` (settings → event), not always the
+    // top-level `taskId`, so reading payload.taskId alone rendered `task=-` at the received stage
+    // (the [stage:queued] task=<id> → [stage:received] task=- loss). Falling back to
+    // meshActiveTaskId keeps the trace task-scoped end-to-end so same-task redelivery is
+    // distinguishable from a different task's real completion.
+    traceMeshEventStage('received', {
+        taskId: readNonEmptyString(payload.taskId) || readNonEmptyString(payload.meshActiveTaskId),
+        sessionId: readNonEmptyString(payload.targetSessionId) || readNonEmptyString(payload.sessionId),
+        nodeId,
+        meshId,
+        event: eventName,
+    });
+    const nodeLabel = nodeId ? `Node '${nodeId}'` : workspace ? `Agent at ${workspace}` : 'Remote agent';
+
+    return injectMeshSystemMessage(components, {
+        meshId,
+        nodeId,
+        nodeLabel,
+        event: eventName,
+        metadataEvent: buildRelayMetadataEvent(payload),
+        // T4 (B3b): restore the v2 envelope carried at the top level of the relayed flat
+        // payload (buildForwardPayloadFromPending → serializeV2EnvelopeToWire) so the
+        // re-queue preserves the original eventId (idempotency) and unicast routing rather
+        // than re-stamping a fresh v1/broadcast event. Empty for a v1 relay (version-skew safe).
+        v2Envelope: readV2EnvelopeFromWire(payload),
+    });
+}
+
+// ---------------------------------------------------------------------------
+// Per-coordinator forward serialization (P2P send-backpressure relief).
+//
+// When several workers finish at once, each completion runs forwardUnresolvedDelegate
+// Event and fires its own `mesh_forward_event` push. Firing the whole burst
+// concurrently dumps it into the single per-peer P2P DataChannel buffer in one tick,
+// which starves the rpc_ack/rpc_res replies the same channel must carry — a
+// coordinator's inbound `git_status` then times out even though the worker's own
+// forward acks return in ~1s. To cap the concurrent burst we serialize the immediate
+// pushes per coordinator: at most one push is in flight to a given coordinator at a
+// time, the rest run in arrival order behind it. A lone event (idle lane) still
+// dispatches immediately — only a genuine burst is paced. Durability is unchanged:
+// every event is already persisted to the outbox before the push runs, so serializing
+// only delays the best-effort fast path; PHASE 0 retry still covers any gap. This pairs
+// with the DataChannel send-buffer gate in daemon-cloud's mesh manager (writeRequest),
+// which is the hard guarantee; this throttle keeps the burst from piling up there.
+interface CoordinatorForwardLane { tail: Promise<unknown>; depth: number; }
+const coordinatorForwardLanes = new Map<string, CoordinatorForwardLane>();
+function enqueueCoordinatorForwardPush(coordinatorDaemonId: string, run: () => Promise<unknown>): void {
+    let lane = coordinatorForwardLanes.get(coordinatorDaemonId);
+    if (!lane) { lane = { tail: Promise.resolve(), depth: 0 }; coordinatorForwardLanes.set(coordinatorDaemonId, lane); }
+    const wasIdle = lane.depth === 0;
+    lane.depth += 1;
+    const dec = (): void => { lane!.depth -= 1; };
+    if (wasIdle) {
+        // Idle lane → dispatch synchronously, so a lone completion (the common case) has
+        // ZERO added latency and the push call happens in-line. Only a genuine burst —
+        // events arriving while a push is still in flight — is paced (else branch).
+        lane.tail = Promise.resolve(run()).catch(() => {}).then(dec, dec);
+    } else {
+        // Burst: queue behind the in-flight push(es) in arrival order so the whole burst
+        // is not dumped into the shared DataChannel buffer at once. The tail is guarded
+        // so one rejecting push never wedges the lane for the next.
+        lane.tail = lane.tail.then(() => run()).catch(() => {}).then(dec, dec);
+    }
+}
+
+// ---------------------------------------------------------------------------
+// Worker-side fallback forward for unresolved-mesh delegates.
+//
+// A REMOTE worker daemon that is being P2P-remote-controlled by a coordinator is
+// NOT a member of the coordinator's mesh — it has no local mesh record. So when its
+// completion event reaches the forwarder, resolveWorkerDelegateRouting() resolves the
+// coordinator anchor (meshCoordinatorDaemonId) from the worker envelope but cannot
+// resolve the mesh id (neither meshNodeFor nor a workspace→mesh lookup yields one) and
+// returns isDelegate=false / mesh_unresolved. Before this fallback the event was dropped
+// (delivery_unroutable) and only recovered later when the coordinator happened to pull
+// the worker's queue — which it can't, because the worker never queued an unroutable
+// event. Live symptom: `WARN [MeshEvents] delivery_unroutable: ... mesh unresolved`.
+//
+// The fix: the routing object still carries coordinatorDaemonId. Forward the raw event
+// straight to that coordinator daemon over P2P (mesh_forward_event). The coordinator
+// hosts the mesh, so it recovers the mesh id by workspace in handleMeshForwardEvent and
+// injects/queues it normally. meshId is intentionally omitted from the payload (the
+// worker has none); workspace is the routing anchor the coordinator resolves from.
+//
+// No loop / no double-delivery:
+//  - This only fires on the WORKER (the coordinator-own session is rejected by the
+//    resolver before reaching here), and the coordinator merely injects — it does not
+//    re-enter this forwarder for the relayed event.
+//  - It fires only when the normal queue path did NOT run (isDelegate=false), so the
+//    event is never both queued locally and forwarded.
+//
+// Returns true when the event was durably accepted for delivery to the coordinator
+// daemon (so the caller skips the delivery_unroutable diagnostic); false when no
+// fallback was possible (no coordinator anchor / no dispatch transport).
+//
+// Durability: the directed push to the coordinator is the ONLY delivery route for an
+// unresolved-mesh worker (it is in no mesh.node the coordinator can pull). So instead
+// of a fire-and-forget push that drops on one transient P2P failure, the event is
+// persisted to the worker-side outbox FIRST and only acked after a successful push.
+// A best-effort immediate push keeps latency low on the happy path; a failed or
+// un-acked push leaves the durable row for setupMeshReconcileLoop's PHASE 0 to retry.
+function forwardUnresolvedDelegateEvent(
+    components: DaemonComponents,
+    routing: ReturnType<typeof resolveWorkerDelegateRouting>,
+    event: Record<string, unknown>,
+): boolean {
+    const coordinatorDaemonId = readNonEmptyString(routing.coordinatorDaemonId);
+    if (!coordinatorDaemonId) return false;
+    if (!components.dispatchMeshCommand) return false;
+
+    const eventName = readNonEmptyString(event.event);
+    if (!eventName) return false;
+
+    // Flat payload mirroring buildForwardPayloadFromPending / what handleMeshForwardEvent
+    // reads. nodeId/workspace come from the worker envelope so the coordinator can name and
+    // locate the node.
+    const payload: Record<string, unknown> = {
+        ...event,
+        event: eventName,
+        nodeId: readNonEmptyString(routing.nodeId) || readNonEmptyString(event.meshNodeId) || undefined,
+        workspace: readNonEmptyString(routing.workspace) || readNonEmptyString(event.workspace) || undefined,
+        // Fix B: carry the resolved coordinator anchor so the coordinator's receive-side
+        // recovery (recoverMeshIdByCoordinatorAndNode) can match this forward to one of the
+        // meshes it hosts when workspace + nodeId both miss. routing.coordinatorDaemonId is
+        // the same anchor this forward is addressed to (coordinatorDaemonId below).
+        meshCoordinatorDaemonId: coordinatorDaemonId,
+    };
+    // RECONCILE-MESHID-DROP: stamp meshId when the WORKER can resolve it (member node /
+    // live-session meshNodeFor). Historically omitted "because the worker can't resolve
+    // it", but for a member-hosted node a no_node_binding session's coordinator-side
+    // recovery (empty payload nodeId + workspace cache miss) fails and the retry is
+    // rejected "meshId required" forever. Resolving here makes the forward self-sufficient;
+    // when unresolvable even here it stays absent and the coordinator's own workspace/nodeId
+    // recovery still runs (unchanged), with the retry cap as the loop backstop.
+    const resolvedMeshId = resolveForwardEventMeshId(components, payload);
+    if (resolvedMeshId) payload.meshId = resolvedMeshId;
+
+    // Self-addressed fallback: the resolved coordinator IS this daemon (a self-
+    // coordinating / single-node mesh, or a delegate whose coordinator anchor resolved
+    // to our own id). A cross-daemon mesh_forward_event to our own id is REFUSED by the
+    // dispatch self-dial guard ("route via the local router instead"), so persisting it
+    // to the outbox would only loop forever in PHASE 0's retry, never acked. Honour the
+    // guard's advice: route the event straight through the local receiver — the exact
+    // path the coordinator runs on receiving a remote push — and skip the outbox entirely.
+    const selfDaemonIds = resolveCoordinatorDrainDaemonIds(components);
+    if (selfDaemonIds.some(self => daemonIdsEquivalent(self, coordinatorDaemonId))) {
+        try {
+            handleMeshForwardEvent(components, payload);
+            LOG.info('MeshEvents', `Self-addressed unresolved-delegate ${eventName} routed via local router (coordinator ${coordinatorDaemonId} is self) — outbox skipped`);
+        } catch (e: any) {
+            LOG.warn('MeshEvents', `Local route of self-addressed unresolved-delegate ${eventName} failed: ${e?.message || e}`);
+        }
+        return true;
+    }
+
+    // 1) Persist durably FIRST. Idempotent on fingerprint, so a re-fired completion
+    //    does not duplicate the outbox row. If persistence fails we still attempt the
+    //    push below (degrades to the old at-most-once behaviour rather than dropping
+    //    the chance entirely).
+    const persisted = enqueueUnresolvedDelegateForward(coordinatorDaemonId, eventName, payload);
+    // EVTTRACE: unresolved-mesh worker persisted its completion to the outbox (no meshId
+    // available locally; coordinator will recover it on receive).
+    const fwdTraceCtx = {
+        taskId: (payload as Record<string, unknown>).taskId,
+        sessionId: readNonEmptyString(payload.targetSessionId) || readNonEmptyString(payload.sessionId),
+        nodeId: readNonEmptyString(routing.nodeId) || readNonEmptyString(event.meshNodeId),
+        event: eventName,
+    };
+    traceMeshEventStage('outbox_enqueue', fwdTraceCtx, `coordinatorDaemon=${coordinatorDaemonId} meshId=absent`);
+
+    // 2) Best-effort immediate push for low latency. On success, ack the outbox row so
+    //    the retry loop won't re-send it. On failure, leave it queued — PHASE 0 retries.
+    traceMeshEventStage('forward_send', fwdTraceCtx, 'immediate push');
+    // Serialize per coordinator so a multi-worker completion burst is paced rather than
+    // dumped concurrently into the shared P2P DataChannel buffer (see coordinator
+    // ForwardLanes). dispatchMeshCommand was null-checked above; capture it for the
+    // deferred closure.
+    const dispatchMeshCommand = components.dispatchMeshCommand;
+    enqueueCoordinatorForwardPush(coordinatorDaemonId, () =>
+        Promise.resolve(dispatchMeshCommand(coordinatorDaemonId, 'mesh_forward_event', payload))
+            .then((result: any) => {
+                if (result && result.success === false) {
+                    LOG.warn('MeshEvents', `Immediate forward of ${eventName} to coordinator ${coordinatorDaemonId} rejected (${readNonEmptyString(result.error) || 'no reason'}) — left queued for retry`);
+                    traceMeshEventDrop('immediate_forward_rejected', fwdTraceCtx, readNonEmptyString(result.error) || 'no reason');
+                    return;
+                }
+                // Acked. Mark the durable copy delivered so the retry loop skips it.
+                if (persisted) ackUnresolvedDelegateForwardByFingerprint(coordinatorDaemonId, eventName, payload);
+            })
+            .catch((e: any) => {
+                // Coordinator momentarily unreachable; the durable row stays queued and the
+                // reconcile loop retries it. Trace so the relay attempt is visible.
+                LOG.warn('MeshEvents', `Immediate forward of ${eventName} to coordinator ${coordinatorDaemonId} failed: ${e?.message || e} — left queued for retry`);
+            }));
+    LOG.info('MeshEvents', `Durably forwarded ${eventName} for unresolved-mesh worker at ${routing.workspace || '(no workspace)'} to coordinator daemon ${coordinatorDaemonId}`);
+    return true;
+}
+
+// Ack a just-pushed outbox entry by re-deriving its row from the same coordinator +
+// event + payload. We don't thread the row id back from enqueue (the immediate push is
+// fire-then-ack), so locate it among the undrained entries by matching coordinator and
+// the flat payload's forward identity. A miss is harmless — the retry loop's own
+// receiver-side dedup suppresses a duplicate delivery.
+function ackUnresolvedDelegateForwardByFingerprint(
+    coordinatorDaemonId: string,
+    eventName: string,
+    payload: Record<string, unknown>,
+): void {
+    const match = peekUnresolvedDelegateForwards().find(entry =>
+        daemonIdsEquivalent(entry.coordinatorDaemonId, coordinatorDaemonId)
+        && readNonEmptyString(entry.payload.event) === eventName
+        && readNonEmptyString(entry.payload.targetSessionId || entry.payload.sessionId || entry.payload.instanceId)
+            === readNonEmptyString(payload.targetSessionId || payload.sessionId || payload.instanceId)
+        && readNonEmptyString(entry.payload.workspace) === readNonEmptyString(payload.workspace),
+    );
+    if (match) ackUnresolvedDelegateForward(match.id);
+}
+
+/**
+ * NOTIF-HELD-DRAIN (Fix 2): event-driven coordinator drain. The reconcile loop delivers a
+ * worker's queued completion to an IDLE local coordinator only on its periodic poll. When a
+ * coordinator is sitting idle awaiting exactly that completion, waiting up to a full poll
+ * interval is the avoidable delivery latency the RCA flags — and combined with the (now-fixed)
+ * modal-park false-positive it stretched into the multi-minute notification stall. So the
+ * MOMENT a worker delegate event is persisted for a mesh, attempt the same idle-coordinator
+ * drain immediately, mirroring the event-driven worker-claim path (agent:ready /
+ * agent:generating_completed → triggerMeshQueue).
+ *
+ * Safety:
+ *  - drainPendingMeshCoordinatorEvents marks rows drained=1 atomically, so this races the
+ *    reconcile poll and the coordinator's own idle auto-flush harmlessly — exactly one consumes
+ *    each row.
+ *  - Only IDLE, non-modal-parked coordinators are delivery targets (never a generating /
+ *    consent-modal PTY).
+ *  - Strict session routing is honoured: an event naming an originating coordinator session is
+ *    delivered only to that live idle session; anything not currently deliverable here
+ *    (wrong/absent session, or a message-less lifecycle event) is RE-QUEUED — never dropped —
+ *    so the reconcile loop's strict hold/expire path remains the single authority for it.
+ */
+export function flushPendingForMeshIdleCoordinators(components: DaemonComponents, meshId: string): void {
+    // O(1) gate: skip the (relatively expensive) per-instance getState scan when the queue is
+    // empty for this mesh.
+    try {
+        const store = MeshRuntimeStore.getInstance();
+        if (store.pendingEventCount(meshId) === 0) return;
+    } catch { /* store unavailable — fall through and let the drain decide */ }
+
+    const idleCoordinators: { instance: ProviderInstance; sessionId: string }[] = [];
+    try {
+        for (const inst of components.instanceManager.getByCategory('cli')) {
+            const state = inst.getState();
+            const settings = state.settings && typeof state.settings === 'object'
+                ? state.settings as Record<string, unknown>
+                : {};
+            if (readNonEmptyString(settings.meshCoordinatorFor) !== meshId) continue;
+            const status = readNonEmptyString(state.status).toLowerCase();
+            const modalParked = typeof (inst as any).isModalParked === 'function'
+                ? (inst as any).isModalParked() === true
+                : (status === 'waiting_choice' || status === 'waiting_approval');
+            // PTY-OVERTRUST-DRAIN (Defect B): decide idle on the RAW adapter turn-state
+            // (getDrainStatus, mask-stripped) to match the reconcile loop — getState().status
+            // overlays the auto-approve hold-idle mask that paints a genuinely-idle coordinator
+            // `generating`, which would make this opportunistic flush skip a real drain target.
+            // Fall back to the masked literal for any instance without getDrainStatus().
+            const drainStatus: string | null = typeof (inst as any).getDrainStatus === 'function'
+                ? (inst as any).getDrainStatus()
+                : null;
+            const idle = drainStatus !== null ? drainStatus === 'idle' : (status === 'idle');
+            if (idle && !modalParked) {
+                idleCoordinators.push({ instance: inst, sessionId: readNonEmptyString(state.instanceId) });
+            }
+        }
+    } catch { return; }
+    if (idleCoordinators.length === 0) return; // no idle target now → leave for the reconcile poll
+
+    const drainDaemonIds = resolveCoordinatorDrainDaemonIds(components);
+    let pendingEvents: PendingMeshCoordinatorEvent[];
+    try {
+        pendingEvents = drainPendingMeshCoordinatorEvents(meshId, drainDaemonIds.length > 0 ? drainDaemonIds : undefined);
+    } catch (e: any) {
+        LOG.warn('MeshEvents', `Event-driven coordinator drain failed for mesh ${meshId}: ${e?.message || e}`);
+        return;
+    }
+    if (pendingEvents.length === 0) return;
+
+    let delivered = 0;
+    for (const pending of pendingEvents) {
+        const wantSession = readNonEmptyString(pending.targetCoordinatorSessionId);
+        const targets = wantSession
+            ? idleCoordinators.filter(c => sessionIdsEquivalent(c.sessionId, wantSession))
+            : idleCoordinators;
+        // Not deliverable into an idle target here (wrong/absent session), or a message-less
+        // lifecycle event (agent:ready / generating_started carry no coordinatorMessage and
+        // must not be injected): re-queue so the reconcile loop owns it (lazy-synth / strict
+        // hold/expire). Re-queue preserves queuedAt so the strict TTL measures true age.
+        if (targets.length === 0 || !pending.coordinatorMessage) {
+            try { queuePendingMeshCoordinatorEvent(pending); } catch { /* best-effort re-queue */ }
+            continue;
+        }
+        const message = pending.coordinatorMessage;
+        const force = shouldForceInjectMeshEvent(pending.event);
+        for (const c of targets) {
+            c.instance.onEvent('send_message', {
+                input: { text: message, textFallback: message },
+                ...(force ? { force: true } : {}),
+            });
+            delivered++;
+        }
+    }
+    if (delivered > 0) {
+        LOG.info('MeshEvents', `Event-driven drain delivered ${delivered} pending event(s) to ${idleCoordinators.length} idle coordinator(s) for mesh ${meshId}`);
+    }
+}
+
+export function setupMeshEventForwarding(components: DaemonComponents) {
+    components.instanceManager.onEvent((event) => {
+        // --- Coordinator idle auto-flush (fast path) ---
+        // When a coordinator session becomes idle, immediately flush any pending
+        // coordinator events that accumulated while it was generating, rather than
+        // waiting up to one reconcile interval for setupMeshReconcileLoop to do it.
+        // Both paths drain the SAME queue via drainPendingMeshCoordinatorEvents,
+        // whose SQLite drained=1 marking is atomic — whichever fires first consumes
+        // the events and the other gets nothing, so there is no double-delivery.
+        // This runs before the delegate routing below so that coordinator-own idle
+        // transitions are handled first.
+        // Exception: a coordinator that is itself a direct-dispatch target still needs
+        // to go through delegate routing so that the dispatching coordinator receives a
+        // pendingCoordinatorEvents entry for the completion.
+        if (event.event === 'agent:ready' || event.event === 'agent:generating_completed') {
+            const flushInstanceId = readNonEmptyString(event.instanceId);
+            if (flushInstanceId) {
+                const flushSource = components.instanceManager.getInstance(flushInstanceId);
+                if (flushSource && flushSource.category === 'cli') {
+                    const flushState = flushSource.getState();
+                    const flushSettings = flushState.settings && typeof flushState.settings === 'object' ? flushState.settings as Record<string, unknown> : {};
+                    const coordinatorMeshId = readNonEmptyString(flushSettings.meshCoordinatorFor);
+                    if (coordinatorMeshId) {
+                        const status = readNonEmptyString(flushState.status).toLowerCase();
+                        if (status === 'idle') {
+                            try {
+                                // Drain with the daemon's full coordinator-id set (status id + machineId).
+                                // The MCP layer stamps the prefixed status id (`standalone_<machineId>` /
+                                // `daemon_<machineId>`) as the worker's meshCoordinatorDaemonId; draining
+                                // with bare machineId alone would miss those unicast events. Mirrors
+                                // resolveCoordinatorDaemonIds in mesh-reconcile-loop.
+                                const drainDaemonIds = resolveCoordinatorDrainDaemonIds(components);
+                                const pendingEvents = drainPendingMeshCoordinatorEvents(coordinatorMeshId, drainDaemonIds.length > 0 ? drainDaemonIds : undefined);
+                                if (pendingEvents.length > 0) {
+                                    LOG.info('MeshEvents', `Auto-flushing ${pendingEvents.length} pending coordinator event(s) for mesh ${coordinatorMeshId} on coordinator idle`);
+                                    for (const pending of pendingEvents) {
+                                        if (!pending.coordinatorMessage) continue;
+                                        const forcePending = shouldForceInjectMeshEvent(pending.event);
+                                        flushSource.onEvent('send_message', {
+                                            input: { text: pending.coordinatorMessage, textFallback: pending.coordinatorMessage },
+                                            ...(forcePending ? { force: true } : {}),
+                                        });
+                                    }
+                                }
+                            } catch (e: any) {
+                                LOG.warn('MeshEvents', `Failed to auto-flush pending coordinator events: ${e?.message || e}`);
+                            }
+                        }
+                        // Skip delegate routing unless this coordinator session is itself
+                        // a direct-dispatch target — in that case fall through so the
+                        // dispatching coordinator gets a pendingCoordinatorEvents entry.
+                        let hasDirectDispatch = false;
+                        try {
+                            hasDirectDispatch =
+                                getActiveDirectDispatches(coordinatorMeshId).some(d => sessionIdsEquivalent(d.sessionId, flushInstanceId))
+                                || hasUnterminalDirectDispatchLedgerEntry(coordinatorMeshId, flushInstanceId);
+                        } catch { /* best-effort */ }
+                        if (!hasDirectDispatch) return;
+                    }
+                }
+            }
+        }
+
+        // --- Delegate event routing ---
+        if (!isMeshCoordinatorEvent(event.event)) return;
+
+        const instanceId = readNonEmptyString(event.instanceId);
+        if (!instanceId) return;
+
+        // R1: all session→node→mesh→coordinator interpretation is folded into the single
+        // resolveWorkerDelegateRouting() resolver. No stamp (meshNodeFor / meshNodeId /
+        // meshCoordinatorDaemonId / meshCoordinatorNodeId / launchedByCoordinator) is read
+        // here to make a routing decision — the resolver is the one authority, and the
+        // forwarder consumes its typed result only.
+        const routing = resolveWorkerDelegateRouting(components, instanceId, {
+            getMeshById: (meshId) => getMeshWithCache(components, meshId),
+            getMeshByWorkspace: (workspace) => getCachedMeshByWorkspace(workspace),
+        });
+        if (!routing.isDelegate) {
+            // Fallback: a REMOTE worker that isn't a member of the coordinator's mesh can't
+            // resolve a mesh id locally (mesh_unresolved), but it still carries the coordinator
+            // daemon anchor. Forward the event straight to that coordinator over P2P instead of
+            // dropping it — the coordinator hosts the mesh and recovers the id by workspace.
+            if (isUnroutableDelegateRejection(routing)
+                && forwardUnresolvedDelegateEvent(components, routing, event)) {
+                return;
+            }
+            // R4: a worker that presented a valid envelope but resolved to no mesh (and could
+            // not be fallback-forwarded — e.g. no coordinator anchor) used to be dropped
+            // silently. Leave a fail-loud diagnostic so the missing completion is traceable.
+            // Benign non-delegate rejections (not_cli / no_workspace / etc.) are no-ops inside
+            // recordUnroutableDelegateEvent.
+            // EVTTRACE: a delegate event that could not be routed AND could not be
+            // fallback-forwarded (no coordinator anchor). Only mesh_unresolved is a real
+            // drop; the benign non-delegate rejections are ordinary non-mesh traffic.
+            if (isUnroutableDelegateRejection(routing)) {
+                traceMeshEventDrop('unroutable', {
+                    taskId: (event as Record<string, unknown>).meshActiveTaskId ?? (event as Record<string, unknown>).taskId,
+                    sessionId: routing.sessionId,
+                    nodeId: routing.nodeId,
+                    event: event.event,
+                }, 'no coordinator anchor / mesh_unresolved');
+            }
+            recordUnroutableDelegateEvent(routing, event.event);
+            return;
+        }
+
+        injectMeshSystemMessage(components, {
+            meshId: routing.meshId,
+            sourceInstanceId: instanceId,
+            nodeId: routing.nodeId,
+            nodeLabel: routing.nodeLabel,
+            event: event.event,
+            metadataEvent: event,
+        });
+
+        // NOTIF-HELD-DRAIN (Fix 2): the worker's event is now persisted in the pending queue.
+        // If a local coordinator for this mesh is sitting idle awaiting it, deliver immediately
+        // instead of waiting up to a full reconcile interval (event-driven, mirrors the
+        // worker-claim path). No-op when no idle coordinator is present (held for reconcile).
+        flushPendingForMeshIdleCoordinators(components, routing.meshId);
+    });
+}