@adhdev/daemon-core 0.9.82-rc.292 → 0.9.82-rc.294

package/src/mesh/mesh-events-coordinator.ts

@@ -13,6 +13,7 @@ import { MeshRuntimeStore } from './mesh-runtime-store.js';
 import { queuePendingMeshCoordinatorEvent, drainPendingMeshCoordinatorEvents } from './mesh-events-pending.js';
 import type { PendingMeshCoordinatorEvent } from './mesh-events-pending.js';
 import { resolveWorkerDelegateRouting, recordUnroutableDelegateEvent, isUnroutableDelegateRejection } from './mesh-routing.js';
+import { enqueueUnresolvedDelegateForward, peekUnresolvedDelegateForwards, ackUnresolvedDelegateForward } from './mesh-unresolved-forward-outbox.js';
 import { resolveDelegatedWorkerAutoApprove } from '../repo-mesh-types.js';
 import { normalizeMeshNodeId, meshNodeIdMatches } from '@adhdev/mesh-shared';
 import {
@@ -1539,8 +1540,16 @@ export function handleMeshForwardEvent(components: DaemonComponents, payload: Re
 //  - 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 handed off to the coordinator daemon (so the caller
-// skips the delivery_unroutable diagnostic); false when no fallback was possible.
+// 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>,
@@ -1564,16 +1573,52 @@ function forwardUnresolvedDelegateEvent(
         workspace: readNonEmptyString(routing.workspace) || readNonEmptyString(event.workspace) || undefined,
     };
 
+    // 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);
+
+    // 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.
     Promise.resolve(components.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`);
+                return;
+            }
+            // Acked. Mark the durable copy delivered so the retry loop skips it.
+            if (persisted) ackUnresolvedDelegateForwardByFingerprint(coordinatorDaemonId, eventName, payload);
+        })
         .catch((e: any) => {
-            // The coordinator may be momentarily unreachable; the diagnostic was already
-            // skipped, so leave a trace here so an operator can see the relay attempt failed.
-            LOG.warn('MeshEvents', `Fallback forward of ${eventName} to coordinator ${coordinatorDaemonId} failed: ${e?.message || e}`);
+            // 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', `Fallback-forwarded ${eventName} for unresolved-mesh worker at ${routing.workspace || '(no workspace)'} to coordinator daemon ${coordinatorDaemonId}`);
+    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 =>
+        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);
+}
+
 export function setupMeshEventForwarding(components: DaemonComponents) {
     components.instanceManager.onEvent((event) => {
         // --- Coordinator idle auto-flush (fast path) ---

package/src/mesh/mesh-missions.ts

@@ -47,6 +47,25 @@ export interface MeshMissionSummary extends MeshMissionRecord {
     tasks: MeshMissionTaskAggregate;
 }
 
+/**
+ * Slim mission summary for the mesh_status compact (default) surface. Drops the
+ * full `goal` text — which can be hundreds of chars per mission and is repeated
+ * for every mission on every status call — keeping only a short preview plus a
+ * `goalTruncated` flag when the original was longer. The stored goal is never
+ * mutated; this is an output-only projection. Coordinators that need the full
+ * goal call mesh_status with verbose=true, or read the mission directly via
+ * mesh_mission_upsert / getMeshMission.
+ */
+export interface MeshMissionSlimSummary extends Omit<MeshMissionSummary, 'goal'> {
+    /** Short preview of the goal (≤ GOAL_PREVIEW_MAX chars), '' when goal empty. */
+    goalPreview: string;
+    /** True when the stored goal was longer than the preview (full text elided). */
+    goalTruncated: boolean;
+}
+
+/** Max chars of goal text retained in the slim (compact) mission summary. */
+export const GOAL_PREVIEW_MAX = 120;
+
 function normalizeMissionStatus(value: unknown): MeshMissionStatus {
     return MESH_MISSION_STATUSES.includes(value as MeshMissionStatus)
         ? value as MeshMissionStatus
@@ -125,17 +144,35 @@ export function getActiveMeshMissionSummaries(meshId: string): MeshMissionSummar
     return getMeshMissions(meshId, ['active']).map(mission => summarizeMeshMission(meshId, mission));
 }
 
+/** Project a full mission summary down to the slim (goal-elided) shape. */
+function slimMissionSummary(summary: MeshMissionSummary): MeshMissionSlimSummary {
+    const goal = typeof summary.goal === 'string' ? summary.goal : '';
+    const goalTruncated = goal.length > GOAL_PREVIEW_MAX;
+    const { goal: _omitGoal, ...rest } = summary;
+    return {
+        ...rest,
+        goalPreview: goalTruncated ? goal.slice(0, GOAL_PREVIEW_MAX) : goal,
+        goalTruncated,
+    };
+}
+
 /**
  * Mission summaries for the mesh_status dashboard surface: every active/paused
  * mission plus a capped, newest-first slice of completed/abandoned history so
  * the dashboard can render a collapsible "history" section without unbounded
  * payload growth. Returned newest-first within each group (active/paused first,
  * then history), so the frontend can split on `status` directly.
+ *
+ * Compact mode (the default) elides each mission's full `goal` text — which is
+ * repeated verbatim on every status poll and dominates the payload when a mesh
+ * has many missions — returning only a short `goalPreview` + `goalTruncated`
+ * flag. Pass `verbose: true` to get the full `goal` text per mission. The stored
+ * goal is untouched in both modes; this is an output-only projection.
  */
 export function getMeshStatusMissionSummaries(
     meshId: string,
-    options?: { historyLimit?: number },
-): MeshMissionSummary[] {
+    options?: { historyLimit?: number; verbose?: boolean },
+): MeshMissionSummary[] | MeshMissionSlimSummary[] {
     const historyLimit = Math.max(0, options?.historyLimit ?? 10);
     const all = getMeshMissions(meshId);
     const live = all.filter(m => m.status === 'active' || m.status === 'paused');
@@ -143,7 +180,8 @@ export function getMeshStatusMissionSummaries(
         .filter(m => m.status === 'completed' || m.status === 'abandoned')
         .sort((a, b) => (b.updatedAt || '').localeCompare(a.updatedAt || ''))
         .slice(0, historyLimit);
-    return [...live, ...history].map(mission => summarizeMeshMission(meshId, mission));
+    const full = [...live, ...history].map(mission => summarizeMeshMission(meshId, mission));
+    return options?.verbose ? full : full.map(slimMissionSummary);
 }
 
 /**

package/src/mesh/mesh-reconcile-loop.ts

@@ -49,6 +49,11 @@ import { drainPendingMeshCoordinatorEvents } from './mesh-events-pending.js';
 import type { PendingMeshCoordinatorEvent } from './mesh-events-pending.js';
 import { MeshRuntimeStore } from './mesh-runtime-store.js';
 import { handleMeshForwardEvent, shouldForceInjectMeshEvent, MESH_FORCE_INJECT_EVENTS } from './mesh-events-coordinator.js';
+import {
+    peekUnresolvedDelegateForwards,
+    ackUnresolvedDelegateForward,
+    expireStaleUnresolvedDelegateForwards,
+} from './mesh-unresolved-forward-outbox.js';
 import { readNonEmptyString } from './mesh-events-utils.js';
 
 // Default reconcile cadence. approval/completion notifications to a live CLI
@@ -213,6 +218,20 @@ export async function runMeshReconcileTick(components: DaemonComponents): Promis
         try { return MeshRuntimeStore.getInstance(); } catch { return undefined; }
     })();
 
+    // ── PHASE 0: retry the worker-side unresolved-delegate forward outbox ──────
+    // Cloud-only (needs dispatchMeshCommand). A worker that is NOT a member of the
+    // coordinator's mesh cannot be reached by the coordinator's PHASE 1 pull (it is
+    // in no mesh.node), so its completion must be PUSHED to the coordinator. This
+    // drains the durable outbox enqueued by forwardUnresolvedDelegateEvent and retries
+    // any push that has not yet been acked. See mesh-unresolved-forward-outbox.ts.
+    if (dispatchMeshCommand) {
+        try {
+            await retryUnresolvedDelegateForwards(components);
+        } catch (e: any) {
+            LOG.warn('MeshReconcile', `Unresolved-delegate forward retry failed: ${e?.message || e}`);
+        }
+    }
+
     // ── PHASE 1: pull remote node queues for every mesh this daemon hosts ──────
     // Cloud-only (dispatchMeshCommand present). Runs whether or not a live CLI
     // coordinator exists — this is what lets an MCP/LLM coordinator ever see a
@@ -295,6 +314,42 @@ export async function runMeshReconcileTick(components: DaemonComponents): Promis
     }
 }
 
+// Cloud-only: retry the worker-side unresolved-delegate forward outbox. For each
+// durably-queued entry, push it to its coordinator daemon over P2P (mesh_forward_event)
+// and ack (mark drained) ONLY on a successful, non-rejected response. A failed or
+// rejected push leaves the entry queued for the next tick — at-least-once delivery.
+// Stale entries (coordinator unreachable past the max age) are expired first so the
+// outbox can't grow without bound. The coordinator dedups duplicate deliveries on its
+// own fingerprint, so a retry that races the original immediate push is harmless.
+async function retryUnresolvedDelegateForwards(components: DaemonComponents): Promise<void> {
+    const dispatchMeshCommand = components.dispatchMeshCommand;
+    if (!dispatchMeshCommand) return;
+
+    // Drop entries that have exhausted their retry budget (fail-loud inside).
+    expireStaleUnresolvedDelegateForwards();
+
+    const entries = peekUnresolvedDelegateForwards();
+    if (entries.length === 0) return;
+
+    for (const entry of entries) {
+        let result: any;
+        try {
+            result = await dispatchMeshCommand(entry.coordinatorDaemonId, 'mesh_forward_event', entry.payload);
+        } catch (e: any) {
+            // Coordinator unreachable — keep the entry queued and try again next tick.
+            LOG.warn('MeshReconcile', `Retry forward to coordinator ${entry.coordinatorDaemonId} failed: ${e?.message || e} — left queued`);
+            continue;
+        }
+        if (result && result.success === false) {
+            LOG.warn('MeshReconcile', `Retry forward to coordinator ${entry.coordinatorDaemonId} rejected (${readNonEmptyString(result.error) || 'no reason'}) — left queued`);
+            continue;
+        }
+        // Acked — mark the durable copy delivered.
+        ackUnresolvedDelegateForward(entry.id);
+        LOG.info('MeshReconcile', `Retried+delivered unresolved-delegate ${readNonEmptyString(entry.payload.event)} to coordinator ${entry.coordinatorDaemonId}`);
+    }
+}
+
 // Cloud-only: poll each remote worker node daemon for pending coordinator events
 // and re-inject them locally via handleMeshForwardEvent (which re-queues +
 // surfaces to the live coordinator on the next tick / immediately if idle).

package/src/mesh/mesh-runtime-store.ts

@@ -1396,4 +1396,34 @@ export class MeshRuntimeStore {
         ).get(meshId) as { cnt: number } | undefined;
         return row?.cnt ?? 0;
     }
+
+    /**
+     * Mark specific pending-event rows drained by id (ack). Used by the
+     * unresolved-delegate durable-forward outbox: an event is peeked (not drained)
+     * while its push to the coordinator is unconfirmed, then marked drained ONLY
+     * after the push is acked. A failed push leaves the row undrained so the next
+     * reconcile tick retries it. Returns the number of rows newly marked drained.
+     */
+    markPendingEventsDrainedById(ids: ReadonlyArray<string>): number {
+        const idList = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+        if (idList.length === 0) return 0;
+        const now = Date.now();
+        return this.db.prepare(
+            `UPDATE mesh_pending_events SET drained = 1, drained_at = ? WHERE drained = 0 AND id IN (${idList.map(() => '?').join(',')})`
+        ).run(now, ...idList).changes;
+    }
+
+    /**
+     * Hard-delete pending-event rows by id (including the dedup fingerprint history).
+     * Used to expire an unresolved-delegate outbox entry that has exhausted its retry
+     * budget — fully removing it frees the fingerprint so a genuinely new completion
+     * for the same task could be re-queued later. Returns the number of rows deleted.
+     */
+    deletePendingEventsById(ids: ReadonlyArray<string>): number {
+        const idList = ids.filter((id): id is string => typeof id === 'string' && id.length > 0);
+        if (idList.length === 0) return 0;
+        return this.db.prepare(
+            `DELETE FROM mesh_pending_events WHERE id IN (${idList.map(() => '?').join(',')})`
+        ).run(...idList).changes;
+    }
 }

package/src/mesh/mesh-unresolved-forward-outbox.ts

@@ -0,0 +1,185 @@
+// ---------------------------------------------------------------------------
+// mesh-unresolved-forward-outbox — durable retry for unresolved-delegate forwards
+// ---------------------------------------------------------------------------
+// A REMOTE worker daemon that is 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 setupMeshEventForwarding, resolveWorkerDelegateRouting
+// returns mesh_unresolved: the worker carries the coordinator daemon anchor but no
+// resolvable mesh id, so the normal queue-then-coordinator-pull path can't be used
+// (the coordinator's reconcile PHASE 1 only pulls mesh.nodes, and this worker is in
+// none of them — see docs/refactoring/2026-06-16-mesh-completion-polling-single-model.md).
+//
+// The ONLY delivery route for this case is a directed push to the coordinator
+// daemon (mesh_forward_event over P2P). Previously that push was fire-and-forget:
+// a single dispatchMeshCommand whose rejection was only logged, so one transient
+// P2P failure dropped the completion forever (delivery_unroutable).
+//
+// This outbox makes that push DURABLE without inventing a new persistence layer:
+// it reuses the existing mesh_pending_events SQLite table (the same store every
+// other mesh coordinator event is persisted to), under a synthetic, reserved
+// "mesh id" namespace so it never collides with a real mesh's queue. The worker's
+// reconcile tick (PHASE 0) peeks the outbox, pushes each entry to its coordinator,
+// and marks it drained (acked) ONLY on a successful push — a failed push leaves the
+// row undrained for the next tick. Entries that exceed a max age are expired so the
+// outbox can't grow unbounded when a coordinator stays permanently unreachable.
+//
+// Idempotency:
+//   - Enqueue is idempotent on the event fingerprint (mesh_pending_events has a
+//     UNIQUE (mesh_id, fingerprint) index + INSERT OR IGNORE), so the same
+//     completion fired twice on the worker queues once.
+//   - The receiver dedups independently: handleMeshForwardEvent → injectMeshSystemMessage
+//     → queuePendingMeshCoordinatorEvent recomputes the fingerprint and suppresses a
+//     duplicate, so an at-least-once retry that double-delivers is harmless.
+// ---------------------------------------------------------------------------
+
+import { randomUUID } from 'crypto';
+import { LOG } from '../logging/logger.js';
+import { MeshRuntimeStore } from './mesh-runtime-store.js';
+import { buildPendingEventFingerprint, type PendingMeshCoordinatorEvent } from './mesh-events-pending.js';
+import { readNonEmptyString } from './mesh-events-utils.js';
+
+// Reserved synthetic mesh id for the worker-side unresolved-forward outbox. The `::`
+// and leading `__` make it impossible to collide with a real mesh id (which are
+// `mesh_*` / `mreg_*` slugs). Scoping rows by coordinator_daemon_id inside this one
+// namespace lets a single worker forward to multiple coordinators.
+export const UNRESOLVED_FORWARD_OUTBOX_MESH_ID = '__unresolved_forward_outbox__';
+
+// Entries older than this are expired (dropped) rather than retried forever — a
+// coordinator that has been unreachable this long is treated as gone. Generous
+// enough to ride out a long coordinator outage / restart, bounded enough that a
+// permanently-dead coordinator can't accumulate outbox rows indefinitely.
+const UNRESOLVED_FORWARD_MAX_AGE_MS = 30 * 60 * 1000; // 30 minutes
+
+export interface UnresolvedForwardEntry {
+    /** Row id in mesh_pending_events; pass back to ack/expire after a push attempt. */
+    id: string;
+    /** The coordinator daemon to push this event to (mesh_forward_event target). */
+    coordinatorDaemonId: string;
+    /** The flat payload to forward (already shaped for handleMeshForwardEvent). */
+    payload: Record<string, unknown>;
+    /** When the entry was first enqueued (epoch ms) — used for age-based expiry. */
+    queuedAt: number;
+}
+
+function getStore(): MeshRuntimeStore | undefined {
+    try { return MeshRuntimeStore.getInstance(); } catch { return undefined; }
+}
+
+/**
+ * Durably enqueue an unresolved-delegate forward for a coordinator daemon. The
+ * `forwardPayload` is the flat shape handleMeshForwardEvent reads on the coordinator.
+ * Idempotent on the event fingerprint. Returns true when a new row was written (or a
+ * duplicate was harmlessly ignored), false on a hard persistence failure.
+ */
+export function enqueueUnresolvedDelegateForward(
+    coordinatorDaemonId: string,
+    eventName: string,
+    forwardPayload: Record<string, unknown>,
+): boolean {
+    const target = readNonEmptyString(coordinatorDaemonId);
+    const event = readNonEmptyString(eventName);
+    if (!target || !event) return false;
+    const store = getStore();
+    if (!store) return false;
+
+    const queuedAt = Date.now();
+    // Reuse the standard pending-event fingerprint so enqueue is idempotent AND the
+    // receiver's own dedup keys on a comparable identity. We synthesise the minimal
+    // PendingMeshCoordinatorEvent shape buildPendingEventFingerprint needs.
+    const fingerprintSource: PendingMeshCoordinatorEvent = {
+        event,
+        meshId: UNRESOLVED_FORWARD_OUTBOX_MESH_ID,
+        nodeLabel: readNonEmptyString(forwardPayload.nodeId) || readNonEmptyString(forwardPayload.workspace) || 'unresolved-delegate',
+        nodeId: readNonEmptyString(forwardPayload.nodeId) || undefined,
+        workspace: readNonEmptyString(forwardPayload.workspace) || undefined,
+        metadataEvent: forwardPayload,
+        queuedAt,
+        targetCoordinatorDaemonId: target,
+    };
+    // Scope the fingerprint to the coordinator so two coordinators awaiting the same
+    // worker session don't collapse into one outbox row.
+    const fingerprint = `${target}::${buildPendingEventFingerprint(fingerprintSource)}`;
+
+    try {
+        const inserted = store.insertPendingEvent({
+            id: randomUUID(),
+            meshId: UNRESOLVED_FORWARD_OUTBOX_MESH_ID,
+            coordinatorDaemonId: target,
+            event,
+            // Store the flat forward payload + the queue timestamp so the retry tick can
+            // rebuild the push args and apply age-based expiry without a schema change.
+            payload: { forwardPayload, coordinatorDaemonId: target, queuedAt },
+            fingerprint,
+            queuedAt,
+        });
+        if (inserted) {
+            LOG.info('MeshEvents', `Durably queued unresolved-delegate ${event} for coordinator ${target} (outbox)`);
+        }
+        return true;
+    } catch (e: any) {
+        LOG.warn('MeshEvents', `Failed to persist unresolved-delegate forward to outbox: ${e?.message || e}`);
+        return false;
+    }
+}
+
+/** Peek (non-destructive) every undrained outbox entry across all coordinators. */
+export function peekUnresolvedDelegateForwards(): UnresolvedForwardEntry[] {
+    const store = getStore();
+    if (!store) return [];
+    let rows: Array<{ id: string; event: string; payload: unknown }>;
+    try {
+        rows = store.peekPendingEvents(UNRESOLVED_FORWARD_OUTBOX_MESH_ID);
+    } catch {
+        return [];
+    }
+    const out: UnresolvedForwardEntry[] = [];
+    for (const row of rows) {
+        const stored = row.payload && typeof row.payload === 'object' ? row.payload as Record<string, unknown> : {};
+        const coordinatorDaemonId = readNonEmptyString(stored.coordinatorDaemonId);
+        const forwardPayload = stored.forwardPayload && typeof stored.forwardPayload === 'object'
+            ? stored.forwardPayload as Record<string, unknown>
+            : undefined;
+        if (!coordinatorDaemonId || !forwardPayload) continue;
+        const queuedAt = typeof stored.queuedAt === 'number' ? stored.queuedAt : 0;
+        out.push({ id: row.id, coordinatorDaemonId, payload: forwardPayload, queuedAt });
+    }
+    return out;
+}
+
+/** Mark an outbox entry delivered (acked) after a successful push. */
+export function ackUnresolvedDelegateForward(id: string): void {
+    const store = getStore();
+    if (!store) return;
+    try { store.markPendingEventsDrainedById([id]); } catch { /* best-effort */ }
+}
+
+/**
+ * Drop outbox entries that have exceeded the max retry age. Returns the count
+ * expired so the caller can log a fail-loud trace (a dropped completion is a real
+ * loss; it must be visible, not silent).
+ */
+export function expireStaleUnresolvedDelegateForwards(nowMs: number = Date.now()): number {
+    const entries = peekUnresolvedDelegateForwards();
+    const staleIds = entries
+        .filter(e => e.queuedAt > 0 && nowMs - e.queuedAt >= UNRESOLVED_FORWARD_MAX_AGE_MS)
+        .map(e => e.id);
+    if (staleIds.length === 0) return 0;
+    const store = getStore();
+    if (!store) return 0;
+    try {
+        const removed = store.deletePendingEventsById(staleIds);
+        if (removed > 0) {
+            LOG.warn('MeshEvents', `Expired ${removed} unresolved-delegate forward(s) after ${Math.round(UNRESOLVED_FORWARD_MAX_AGE_MS / 60000)}m of failed retries — coordinator unreachable, completion dropped`);
+        }
+        return removed;
+    } catch {
+        return 0;
+    }
+}
+
+/** Test helper: purge the entire outbox. */
+export function __clearUnresolvedDelegateForwardOutboxForTests(): void {
+    const store = getStore();
+    if (!store) return;
+    try { store.clearPendingEventsForMesh(UNRESOLVED_FORWARD_OUTBOX_MESH_ID); } catch { /* nothing to clear */ }
+}

package/src/providers/cli-provider-instance.ts

@@ -37,6 +37,18 @@ type PersistableCliHistoryMessage = {
     receivedAt?: number;
 };
 
+// Status snapshots only ever surface the newest messages: the cloud 'live'
+// profile drops chat messages entirely (loaded lazily via read_chat on
+// subscribe) and the 'full' profile caps activeChat.messages to the last 60
+// (see status/normalize.ts). Unread/completion markers walk only the tail.
+// So getState()'s saved-history hydration — which runs once per resume/manual
+// CLI session on every status report — must read only a bounded tail, not the
+// entire transcript. A full MAX_SAFE_INTEGER read here makes the initial
+// status report O(transcript) × N(sessions), which is the real cold first-
+// connection bottleneck on chat-heavy machines. The window comfortably exceeds
+// the 60-message snapshot cap so dedup/collapse at the boundary stays stable.
+const STATUS_HYDRATION_TAIL_LIMIT = 200;
+
 type CompletedDebouncePending = {
     chatTitle: string;
     duration: number;
@@ -829,22 +841,27 @@ export class CliProviderInstance implements ProviderInstance {
     }
 
     updateSettings(newSettings: Record<string, any>): void {
-        const runtimeMeshSettings: Record<string, any> = {};
-        for (const key of [
-            'meshNodeFor',
-            'meshNodeId',
-            'meshActiveTaskId',
-            'meshCoordinatorFor',
-            'meshCoordinatorDaemonId',
-            'meshCoordinatorNodeId',
-            'spawnedSessionVisibility',
-            'launchedByCoordinator',
-        ]) {
-            if (this.settings[key] !== undefined && newSettings[key] === undefined) {
-                runtimeMeshSettings[key] = this.settings[key];
-            }
-        }
-        this.settings = { ...newSettings, ...runtimeMeshSettings };
+        // Merge semantics: a key omitted from newSettings preserves its existing
+        // value, a key present in newSettings (even as false) overrides it.
+        //
+        // This is required because updateSettings has two callers with opposite
+        // intent:
+        //   1. Full re-injection — the dashboard toggle path (handleSetProviderSetting
+        //      → getSettings → updateInstanceSettings) sends the COMPLETE settings
+        //      object, so an explicit autoApprove:false must win.
+        //   2. Partial stamp — the mesh relay-safety stamp (router.ts agent_command,
+        //      buildMeshWorkerRelayStamp) sends ONLY {meshNodeFor, meshNodeId,
+        //      meshCoordinatorDaemonId, launchedByCoordinator} on every coordinator
+        //      re-dispatch. It carries no autoApprove, so a full replacement would
+        //      wipe the launch-time autoApprove:true the worker was started with,
+        //      silently dropping every later approval to a manual gate until the
+        //      machine-page toggle re-injected the full settings.
+        //
+        // A plain merge satisfies both: undefined keys fall through to the existing
+        // value (preserving launch-stamp settings like autoApprove + the mesh routing
+        // keys), explicit keys override. This subsumes the previous mesh-key preserve
+        // list, which only protected the routing keys and not autoApprove.
+        this.settings = { ...this.settings, ...newSettings };
         this.adapter.updateRuntimeSettings?.(this.settings);
         this.monitor.updateConfig({
             approvalAlert: this.settings.approvalAlert !== false,
@@ -2185,13 +2202,21 @@ export class CliProviderInstance implements ProviderInstance {
         return newestMessageAt === 0;
     }
 
-    private syncCanonicalSavedHistoryIfNeeded(): boolean {
+    private syncCanonicalSavedHistoryIfNeeded(options: { full?: boolean } = {}): boolean {
         if (!this.providerSessionId) return false;
         const canonicalHistory = this.provider.nativeHistory;
         if (!canonicalHistory) return false;
 
+        // Per-status-report hydration reads only a bounded tail (snapshot needs at
+        // most the newest 60). The once-per-resume restore path passes full:true
+        // because seedSessionHistory needs the COMPLETE transcript to seed dedup
+        // state. The read-cache key encodes the window so the bounded and full
+        // reads don't share/clobber each other's 2s cache entry.
+        const limit = options.full ? Number.MAX_SAFE_INTEGER : STATUS_HYDRATION_TAIL_LIMIT;
+        const windowTag = options.full ? 'full' : `tail:${STATUS_HYDRATION_TAIL_LIMIT}`;
+
         if (isNativeSourceCanonicalHistory(canonicalHistory)) {
-            const cacheKey = [this.type, this.providerSessionId, this.workingDir].join('\0');
+            const cacheKey = [this.type, this.providerSessionId, this.workingDir, windowTag].join('\0');
             const now = Date.now();
             if (cacheKey === this.lastNativeSourceCanonicalCacheKey && now - this.lastNativeSourceCanonicalCheckAt < 2_000) {
                 return true;
@@ -2204,7 +2229,7 @@ export class CliProviderInstance implements ProviderInstance {
                 historySessionId: this.providerSessionId,
                 workspace: this.workingDir,
                 offset: 0,
-                limit: Number.MAX_SAFE_INTEGER,
+                limit,
                 historyBehavior: this.provider.historyBehavior,
                 scripts: this.provider.scripts as any,
             });
@@ -2221,7 +2246,7 @@ export class CliProviderInstance implements ProviderInstance {
         }
 
         try {
-            const cacheKey = [this.type, this.providerSessionId, this.workingDir, canonicalHistory.mode || 'materialized-mirror'].join('\0');
+            const cacheKey = [this.type, this.providerSessionId, this.workingDir, canonicalHistory.mode || 'materialized-mirror', windowTag].join('\0');
             const now = Date.now();
             if (cacheKey === this.lastNativeSourceCanonicalCacheKey && now - this.lastNativeSourceCanonicalCheckAt < 2_000) {
                 return true;
@@ -2232,15 +2257,14 @@ export class CliProviderInstance implements ProviderInstance {
             if (!materializeProviderNativeHistory(this.type, canonicalHistory, this.providerSessionId, this.workingDir, this.provider.scripts as any)) {
                 return false;
             }
-            // Full read is intentional: lastPersistedHistoryMessages is the COMPLETE
-            // session transcript — emitted as statusMessages and used as the
-            // prefix-comparison base for incremental appends — so a bounded tail
-            // would both truncate output and break prefix dedup. This is gated to
-            // once-per-2s (cache key above) for resume/manual launches only, so it
-            // does not run on the per-subscribe/per-poll dashboard tail path (that
-            // path goes through handleReadChat → readChatHistory with a bounded
-            // tailLimit, which is now O(tail)).
-            const restoredHistory = readChatHistory(this.type, 0, Number.MAX_SAFE_INTEGER, this.providerSessionId, 0, this.provider.historyBehavior);
+            // Bounded by default: the per-status-report path only needs the newest
+            // STATUS_HYDRATION_TAIL_LIMIT messages because the snapshot caps
+            // activeChat.messages to the last 60 (status/normalize.ts) and loads
+            // the rest lazily via read_chat on subscribe. The once-per-resume
+            // restore path passes full:true so seedSessionHistory still sees the
+            // COMPLETE transcript for prefix-dedup seeding. readChatHistory serves
+            // a bounded limit as an O(tail) read.
+            const restoredHistory = readChatHistory(this.type, 0, limit, this.providerSessionId, 0, this.provider.historyBehavior);
             this.lastPersistedHistoryMessages = restoredHistory.messages.map((message) => ({
                 role: message.role,
                 content: message.content,
@@ -2256,7 +2280,10 @@ export class CliProviderInstance implements ProviderInstance {
 
     private restorePersistedHistoryFromCurrentSession(): void {
         if (!this.providerSessionId) return;
-        this.syncCanonicalSavedHistoryIfNeeded();
+        // Restore is the once-per-resume seeding path: it needs the COMPLETE
+        // transcript so seedSessionHistory can prime dedup state. Pass full so the
+        // hydration read is unbounded here (and only here).
+        this.syncCanonicalSavedHistoryIfNeeded({ full: true });
         const restoredHistory = isNativeSourceCanonicalHistory(this.provider.nativeHistory)
             ? readProviderChatHistory(this.type, {
                 canonicalHistory: this.provider.nativeHistory,

package/src/providers/extension-provider-instance.ts

@@ -171,7 +171,11 @@ export class ExtensionProviderInstance implements ProviderInstance {
     }
 
     updateSettings(newSettings: Record<string, any>): void {
-        this.settings = { ...newSettings };
+        // Merge (not replace): a key omitted from newSettings preserves its existing
+        // value, a key present (even false) overrides. Mirrors cli/ide-provider-instance
+        // so partial updates do not wipe launch-time settings (e.g. autoApprove) while a
+        // full settings object still applies explicit values.
+        this.settings = { ...this.settings, ...newSettings };
         this.monitor.updateConfig({
             approvalAlert: this.settings.approvalAlert !== false,
             longGeneratingAlert: this.settings.longGeneratingAlert !== false,

package/src/providers/ide-provider-instance.ts

@@ -259,7 +259,12 @@ export class IdeProviderInstance implements ProviderInstance {
     }
 
     updateSettings(newSettings: Record<string, any>): void {
-        this.settings = { ...newSettings };
+        // Merge (not replace): a key omitted from newSettings preserves its existing
+        // value, a key present (even false) overrides. Mirrors cli-provider-instance —
+        // a partial mesh relay-stamp must not wipe launch-time settings such as
+        // autoApprove, while the dashboard toggle's full settings object still applies
+        // an explicit autoApprove:false. See cli-provider-instance.updateSettings.
+        this.settings = { ...this.settings, ...newSettings };
         this.monitor.updateConfig({
             approvalAlert: this.settings.approvalAlert !== false,
             longGeneratingAlert: this.settings.longGeneratingAlert !== false,