@adhdev/daemon-core 0.9.82-rc.380 → 0.9.82-rc.382

package/dist/mesh/mesh-event-forwarding.d.ts

@@ -1,4 +1,5 @@
 import type { DaemonComponents } from '../boot/daemon-lifecycle.js';
+export declare function recoverMeshIdByCoordinatorAndNode(coordinatorDaemonId: string, nodeId: string): string;
 export declare function resolveForwardEventMeshId(components: DaemonComponents, payload: Record<string, unknown>): string;
 export declare function __resetMeshWorkspaceCacheForTests(): void;
 export declare function buildRelayMetadataEvent(payload: Record<string, unknown>): Record<string, unknown>;

package/dist/mesh/mesh-runtime-store.d.ts

@@ -128,6 +128,18 @@ export declare class MeshRuntimeStore {
         updatedAt: string;
     }>;
     updateDirectDispatchStatus(meshId: string, sessionId: string, status: 'acked' | 'completed' | 'failed' | 'stale', taskId?: string): void;
+    /**
+     * MESH-DISPATCH-MISROUTE (fix 3, consumer residual): resolve the task_id of the SINGLE
+     * non-terminal direct dispatch a session owns. Returns the task_id only when the session
+     * holds exactly ONE active ('dispatched'/'acked') row — the case where a taskId-less
+     * lifecycle event (a legacy/relayed worker whose producer never stamped meshActiveTaskId)
+     * unambiguously belongs to that one dispatch. With zero rows there is nothing to ack; with
+     * two or more (a re-dispatch/nudge sibling) the firing event's owner is ambiguous, so we
+     * return null and the caller MUST NOT fall back to the session_id sweep that would flip a
+     * sibling row ("may flip a sibling dispatch row"). This narrows the legacy fallback to the
+     * only safe case instead of removing the producer-side TASKIDLESS stamp's safety net.
+     */
+    getSoleActiveDirectDispatchTaskId(meshId: string, sessionId: string): string | null;
     cleanupTerminalDirectDispatches(olderThanMs: number): void;
     deleteDirectDispatches(meshId: string): void;
     /**

package/dist/mesh/mesh-scheduling-runtime.d.ts

@@ -0,0 +1,78 @@
+import type { RepoMeshNodePolicy, RepoMeshSchedulingStrategy } from '../repo-mesh-types.js';
+import type { MeshWorkQueueEntry } from './mesh-work-queue.js';
+/** Per-(node, provider) cap and its current consumption. */
+export interface MeshNodeProviderSchedulingRuntime {
+    providerType: string;
+    /** Declared maxParallel cap for this (node, provider). Omitted when uncapped. */
+    maxParallel?: number;
+    /** Active (status='assigned') tasks on this node claimed by this provider. */
+    activeAssigned: number;
+    /** True when activeAssigned has reached maxParallel (a further claim is refused). */
+    capReached: boolean;
+}
+/** Per-node scheduling-runtime projection. */
+export interface MeshNodeSchedulingRuntime {
+    nodeId: string;
+    /** Active (status='assigned') task count on this node — the least-loaded rank key. */
+    load: number;
+    /** Soft scheduling priority (PRIORITY rank key; higher = preferred). */
+    schedulingPriority: number;
+    /** Per-node concurrent-session cap, when configured. */
+    maxConcurrentSessions?: number;
+    /** Per-(node, provider) caps + consumption, when providerRoles declares any. */
+    providerRoles?: MeshNodeProviderSchedulingRuntime[];
+    /**
+     * True when the node currently cannot claim a NEW write (non-readonly) task —
+     * either the global write cap is exhausted or a node-local gate (active write
+     * assignment / a fully-consumed provider cap / session cap) blocks it.
+     */
+    capReached: boolean;
+    /** Structured reasons backing capReached (empty when the node can take work). */
+    capReasons: string[];
+}
+/** Mesh-level scheduling-runtime projection. */
+export interface MeshSchedulingRuntime {
+    /** Resolved tie-break strategy (defaults to 'first_eligible'). */
+    strategy: RepoMeshSchedulingStrategy;
+    /** Effective global write-task parallel cap (clamped). */
+    maxParallelTasks: number;
+    /** Derived read-only diagnosis cap (max(2, 2× write cap)) — mirrors the claim path. */
+    maxReadonlyParallelTasks: number;
+    /** Current global write (non-readonly) assigned-task load. */
+    activeWriteAssigned: number;
+    /** Current global read-only assigned-task load. */
+    activeReadonlyAssigned: number;
+    /** True when the global write cap is exhausted (no new write task can launch). */
+    globalWriteCapReached: boolean;
+    /** True when the read-only diagnosis cap is exhausted. */
+    globalReadonlyCapReached: boolean;
+    /** Per-node projections, in mesh config order. */
+    nodes: MeshNodeSchedulingRuntime[];
+}
+interface MeshLike {
+    policy?: {
+        schedulingStrategy?: unknown;
+        maxParallelTasks?: unknown;
+    } | null;
+    nodes?: Array<{
+        id?: string;
+        nodeId?: string;
+        node_id?: string;
+        policy?: RepoMeshNodePolicy | null;
+        isLocalWorktree?: boolean;
+    }> | null;
+}
+/**
+ * Build the scheduling-runtime projection for a mesh from its config + a snapshot of
+ * its work queue. Pure and side-effect free — safe to call on any read path.
+ *
+ * The derivation deliberately mirrors maybeAutoLaunchOneQueueSession's gates so the
+ * exposed "capReached/capReasons" match why a real claim would be refused:
+ *   • global write cap   → activeWriteAssigned >= maxParallelTasks
+ *   • node write conflict→ a node already holding an assigned write task (worktree isolation)
+ *   • provider cap       → a (node, provider) at its declared maxParallel
+ * It does NOT re-evaluate eligibility/required-tags (those are task-specific); it only
+ * reports the capacity picture, which is what an operator needs to read load balance.
+ */
+export declare function buildMeshSchedulingRuntime(mesh: MeshLike | null | undefined, queue: MeshWorkQueueEntry[]): MeshSchedulingRuntime;
+export {};

package/dist/providers/cli-provider-instance.d.ts

@@ -150,6 +150,24 @@ export declare class CliProviderInstance implements ProviderInstance {
      * terminal state. Leaving meshNodeFor pinned would route this session's
      * subsequent unrelated turns (e.g. ad-hoc dashboard chats) to the
      * coordinator as if they were task completions.
+     *
+     * MESHID-DROP-ON-DETACH (Fix C): a coordinator-LAUNCHED worker session
+     * (launchedByCoordinator) holds its mesh membership (meshNodeFor / meshNodeId /
+     * meshCoordinatorDaemonId) at the SESSION level — set once at launch
+     * (mesh_launch_session / queue auto-launch), independent of any single task.
+     * The original detach wiped meshNodeFor + meshNodeId together with the
+     * task-level meshActiveTaskId, so the FIRST task completion stripped the
+     * membership and EVERY subsequent completion forwarded with meshId absent —
+     * resolveWorkerDelegateRouting fell to mesh_unresolved and the coordinator
+     * rejected the forward "meshId required". For a launched member we therefore
+     * clear ONLY the task-level marker (meshActiveTaskId) and preserve the
+     * session-level membership so its next task's completion still resolves.
+     * A task-less ad-hoc turn on a preserved-membership session is NOT misrouted:
+     * its completion carries no taskId and the session holds no active assignment,
+     * so the forwarder's WARMUPGAP guard skips the dispatch-row flip (it only
+     * injects a benign task-less notification). A NON-launched session (a plain CLI
+     * session adopted by mesh_send_task --direct, launchedByCoordinator falsy)
+     * keeps the original full clear so an ad-hoc session is never left pinned.
      */
     detachMeshAssignment(): void;
     /**

package/dist/repo-mesh-types.d.ts

@@ -279,6 +279,34 @@ export interface RepoMeshNodePolicy {
     initSubmodulesOnClone?: boolean;
 }
 export declare const DEFAULT_MESH_POLICY: RepoMeshPolicy;
+/** Min/max bounds for the global write-task parallel cap. */
+export declare const MESH_MAX_PARALLEL_TASKS_MIN = 1;
+export declare const MESH_MAX_PARALLEL_TASKS_MAX = 8;
+/**
+ * Resolve the effective global write-task parallel cap from a raw policy value,
+ * clamped to [MESH_MAX_PARALLEL_TASKS_MIN, MESH_MAX_PARALLEL_TASKS_MAX] and
+ * defaulting to DEFAULT_MESH_POLICY.maxParallelTasks for a missing/NaN value.
+ * Both the config write path and the runtime scheduler read the cap through this
+ * helper so they can never disagree on what "max parallel" means.
+ */
+export declare function resolveMaxParallelTasks(value: unknown): number;
+/**
+ * Normalize an autoFastForward sub-policy, filling defaults and dropping an
+ * invalid maxBehind. Mirrors the (previously mesh-config-local) shape so the merge
+ * always emits a fully-populated, valid autoFastForward object.
+ */
+export declare function normalizeAutoFastForwardPolicy(value: unknown): NonNullable<RepoMeshPolicy['autoFastForward']>;
+/**
+ * Canonical merge+normalize for a RepoMeshPolicy. Layers (lowest→highest):
+ * DEFAULT_MESH_POLICY → base (existing persisted policy) → patch (incoming change),
+ * then applies every per-field normalizer so the result is always valid regardless
+ * of what a hand-edited meshes.json or a partial patch contained.
+ *
+ * Persistence economy is preserved: schedulingStrategy is dropped when it
+ * normalizes to the 'first_eligible' default, and autoConvergeCodeChange is dropped
+ * unless explicitly true — so an untouched meshes.json stays byte-for-byte the same.
+ */
+export declare function mergeAndNormalizePolicy(base: RepoMeshPolicy | undefined, patch: Partial<RepoMeshPolicy> | undefined): RepoMeshPolicy;
 /**
  * Resolve whether a delegated worker session launched onto `nodePolicy` (within a mesh
  * governed by `meshPolicy`) should auto-approve. Precedence: node override → mesh policy
@@ -464,6 +492,41 @@ export interface LocalMeshNodeEntry {
     relatedRepos?: RepoMeshRelatedRepo[];
     role?: RepoMeshDaemonRole;
 }
+/**
+ * Per-(node, provider) cap + consumption, as surfaced on a node's scheduling
+ * status. Wire-shape mirror of MeshNodeProviderSchedulingRuntime.
+ */
+export interface RepoMeshNodeProviderSchedulingStatus {
+    providerType: string;
+    maxParallel?: number;
+    activeAssigned: number;
+    capReached: boolean;
+}
+/**
+ * Per-node scheduling runtime exposed on RepoMeshNodeStatus.scheduling. Carried in
+ * full by verbose mesh_status; compact mesh_status sends only {load, capReached}.
+ */
+export interface RepoMeshNodeSchedulingStatus {
+    load: number;
+    schedulingPriority?: number;
+    maxConcurrentSessions?: number;
+    providerRoles?: RepoMeshNodeProviderSchedulingStatus[];
+    capReached: boolean;
+    capReasons?: string[];
+}
+/**
+ * Mesh-level scheduling rollup exposed on RepoMeshStatus.scheduling: which tie-break
+ * strategy is live and how much of the global parallel caps is consumed.
+ */
+export interface RepoMeshSchedulingStatus {
+    strategy: RepoMeshSchedulingStrategy;
+    maxParallelTasks: number;
+    maxReadonlyParallelTasks: number;
+    activeWriteAssigned: number;
+    activeReadonlyAssigned: number;
+    globalWriteCapReached: boolean;
+    globalReadonlyCapReached: boolean;
+}
 export interface RepoMeshStatus {
     meshId: string;
     meshName: string;
@@ -474,6 +537,11 @@ export interface RepoMeshStatus {
     nodes: RepoMeshNodeStatus[];
     queue?: RepoMeshQueueStatus;
     ledger?: RepoMeshLedgerStatus;
+    /**
+     * Mesh-level scheduling rollup (strategy + global cap consumption). Omitted by
+     * daemons predating the scheduling-runtime exposure — treat as optional.
+     */
+    scheduling?: RepoMeshSchedulingStatus;
     /**
      * Mission summaries for the dashboard overview. Active/paused missions plus a
      * capped, newest-first slice of completed/abandoned history. Omitted by older
@@ -545,6 +613,19 @@ export interface RepoMeshNodeStatus {
     lastSeenAt?: string;
     updatedAt?: string;
     connection?: RepoMeshPeerConnectionStatus;
+    /**
+     * Per-node scheduling runtime (load / priority / provider caps / claim-block
+     * reasons). Verbose mesh_status carries the full shape; compact carries only
+     * {load, capReached}. Omitted by daemons predating the exposure.
+     */
+    scheduling?: RepoMeshNodeSchedulingStatus;
+    /**
+     * Stale-daemon-build marker: the live daemon's build commit is a strict ancestor
+     * of this node's workspace HEAD (merged code not yet live). Best-effort, set by
+     * mesh_status when the git probe reports daemonBuildBehind; shape is daemon-defined
+     * (scope/isDaemonAffecting flags). Omitted when the build is current.
+     */
+    staleDaemonBuild?: Record<string, unknown>;
     error?: string;
 }
 export type RepoMeshQueueTaskStatus = 'pending' | 'assigned' | 'completed' | 'failed' | 'cancelled';

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@adhdev/daemon-core",
-  "version": "0.9.82-rc.380",
+  "version": "0.9.82-rc.382",
   "description": "ADHDev daemon core — CDP, IDE detection, providers, command execution",
   "main": "dist/index.js",
   "types": "dist/index.d.ts",
@@ -46,7 +46,7 @@
   "author": "vilmire",
   "license": "AGPL-3.0-or-later",
   "dependencies": {
-    "@adhdev/mesh-shared": "0.9.82-rc.380",
+    "@adhdev/mesh-shared": "0.9.82-rc.382",
     "@adhdev/session-host-core": "*",
     "@agentclientprotocol/sdk": "^0.16.1",
     "ajv": "^8.20.0",

package/src/commands/high-family/mesh-status.ts

@@ -107,6 +107,19 @@ export const meshStatusHandlers: Record<string, HighFamilyHandler> = {
                     const queue = getQueue(meshId);
                     const queueSummary = getMeshQueueStats(meshId);
 
+                    // Scheduling-runtime projection — the load-balancer's live view (tie-break
+                    // strategy, global parallel caps + consumption, per-node load/priority/provider
+                    // caps). Built from the SAME helper + args the MCP `mesh_status` tool uses
+                    // (mesh-tools-status.ts: buildMeshSchedulingRuntime(mesh, getQueue(mesh.id)))
+                    // so the dashboard surface and the coordinator MCP surface render an identical
+                    // runtime. Without this the dashboard Status/Runtime tab showed the
+                    // "not reported by this daemon (older build)" fallback (RepoMeshStatus.scheduling
+                    // was never populated by this daemon-command producer). Computed once so each
+                    // node entry can attach its slice and statusResult can carry the mesh rollup.
+                    const { buildMeshSchedulingRuntime } = await import('../../mesh/mesh-scheduling-runtime.js');
+                    const schedulingRuntime = buildMeshSchedulingRuntime(mesh, queue);
+                    const schedulingByNode = new Map(schedulingRuntime.nodes.map(n => [n.nodeId, n]));
+
                     const { readLedgerEntries, getLedgerSummary } = await import('../../mesh/mesh-ledger.js');
                     const ledgerEntries = readLedgerEntries(meshId, { tail: 20 });
                     const asyncRefineLedgerEntries = readLedgerEntries(meshId, { tail: 100 });
@@ -260,6 +273,16 @@ export const meshStatusHandlers: Record<string, HighFamilyHandler> = {
                             activeSessionDetails: [],
                             launchReady: false,
                         };
+                        // Per-node scheduling slice (load / priority / provider caps / claim-block
+                        // reasons) read by the dashboard's MeshNodeSchedulingBadges. Full shape —
+                        // unlike the MCP compact path this is a dashboard surface, so it keeps the
+                        // whole RepoMeshNodeSchedulingStatus. Redundant nodeId dropped (the entry
+                        // already carries it).
+                        const nodeScheduling = schedulingByNode.get(nodeId);
+                        if (nodeScheduling) {
+                            const { nodeId: _omitNodeId, ...nodeSchedulingRest } = nodeScheduling;
+                            status.scheduling = nodeSchedulingRest;
+                        }
                         if (isSelfNode) {
                             status.connection = {
                                 perspective: 'selected_coordinator',
@@ -511,6 +534,19 @@ export const meshStatusHandlers: Record<string, HighFamilyHandler> = {
                         branchConvergenceSummary: summarizeInlineMeshBranchConvergence(nodeStatuses),
                         ...(previewFreshness ? { previewFreshness, deployFreshness: previewFreshness } : {}),
                         nodes: nodeStatuses,
+                        // Mesh-level scheduling rollup (strategy + global cap consumption). Mirrors
+                        // the MCP `mesh_status` tool's `scheduling` block field-for-field so both
+                        // surfaces read the same runtime; per-node detail lives on each
+                        // nodes[].scheduling above.
+                        scheduling: {
+                            strategy: schedulingRuntime.strategy,
+                            maxParallelTasks: schedulingRuntime.maxParallelTasks,
+                            maxReadonlyParallelTasks: schedulingRuntime.maxReadonlyParallelTasks,
+                            activeWriteAssigned: schedulingRuntime.activeWriteAssigned,
+                            activeReadonlyAssigned: schedulingRuntime.activeReadonlyAssigned,
+                            globalWriteCapReached: schedulingRuntime.globalWriteCapReached,
+                            globalReadonlyCapReached: schedulingRuntime.globalReadonlyCapReached,
+                        },
                         queue: { tasks: queue, summary: queueSummary },
                         ledger: { entries: ledgerEntries, summary: ledgerSummary },
                         ...(missions.length > 0 ? { missions } : {}),

package/src/config/mesh-config.ts

@@ -22,7 +22,7 @@ import type {
     RepoMeshHostMetadata,
     RepoMeshDaemonRole,
 } from '../repo-mesh-types.js';
-import { DEFAULT_MESH_POLICY, normalizeMeshSchedulingStrategy } from '../repo-mesh-types.js';
+import { mergeAndNormalizePolicy } from '../repo-mesh-types.js';
 import { createDefaultMeshHostMetadata } from '../mesh/mesh-host-ownership.js';
 
 // ─── Persistence ────────────────────────────────
@@ -151,63 +151,11 @@ export function normalizeRepoIdentity(remoteUrl: string): string {
 
 // ─── CRUD Operations ────────────────────────────
 
-const SESSION_CLEANUP_MODES = new Set(['preserve', 'stop', 'delete_stopped', 'stop_and_delete']);
-const SPAWNED_SESSION_VISIBILITY_MODES = new Set(['visible', 'hidden']);
-
-function mergeMeshPolicy(base: RepoMeshPolicy | undefined, patch: Partial<RepoMeshPolicy> | undefined): RepoMeshPolicy {
-    const autoFastForward = normalizeAutoFastForwardPolicy({
-        ...DEFAULT_MESH_POLICY.autoFastForward,
-        ...((base?.autoFastForward && typeof base.autoFastForward === 'object') ? base.autoFastForward : {}),
-        ...((patch?.autoFastForward && typeof patch.autoFastForward === 'object') ? patch.autoFastForward : {}),
-    });
-    const policy: RepoMeshPolicy = {
-        ...DEFAULT_MESH_POLICY,
-        ...(base || {}),
-        ...(patch || {}),
-        autoFastForward,
-    };
-    if (!['block', 'warn', 'checkpoint_then_continue'].includes(policy.dirtyWorkspaceBehavior)) {
-        policy.dirtyWorkspaceBehavior = 'warn';
-    }
-    const maxParallelTasks = Number(policy.maxParallelTasks);
-    policy.maxParallelTasks = Number.isFinite(maxParallelTasks) ? Math.max(1, Math.min(8, Math.floor(maxParallelTasks))) : 2;
-    policy.allowAutoPublishSubmoduleMainCommits = policy.allowAutoPublishSubmoduleMainCommits === true;
-    if (!SESSION_CLEANUP_MODES.has(String(policy.sessionCleanupOnNodeRemove))) {
-        policy.sessionCleanupOnNodeRemove = 'preserve';
-    }
-    if (!SPAWNED_SESSION_VISIBILITY_MODES.has(String(policy.spawnedSessionVisibility))) {
-        policy.spawnedSessionVisibility = DEFAULT_MESH_POLICY.spawnedSessionVisibility;
-    }
-    // Load-balancing: normalize the scheduling strategy so an invalid/blank value
-    // falls back to 'first_eligible' (strict no-change). Only persist the field when
-    // it is explicitly a non-default value to keep existing meshes.json untouched.
-    const normalizedStrategy = normalizeMeshSchedulingStrategy(policy.schedulingStrategy);
-    if (normalizedStrategy === 'first_eligible') {
-        delete policy.schedulingStrategy;
-    } else {
-        policy.schedulingStrategy = normalizedStrategy;
-    }
-    // Convergence routing: strict opt-in (default false). Only persist when explicitly
-    // enabled so existing meshes.json stays byte-for-byte untouched.
-    if (policy.autoConvergeCodeChange === true) {
-        policy.autoConvergeCodeChange = true;
-    } else {
-        delete policy.autoConvergeCodeChange;
-    }
-    return policy;
-}
-
-function normalizeAutoFastForwardPolicy(value: unknown): NonNullable<RepoMeshPolicy['autoFastForward']> {
-    const record = value && typeof value === 'object' && !Array.isArray(value)
-        ? value as Record<string, unknown>
-        : {};
-    const maxBehind = Number(record.maxBehind);
-    return {
-        enabled: record.enabled !== false,
-        ...(Number.isFinite(maxBehind) && maxBehind >= 0 ? { maxBehind: Math.floor(maxBehind) } : {}),
-        requireCleanSubmodules: record.requireCleanSubmodules !== false,
-    };
-}
+// Single source of truth for default+merge+per-field normalization is
+// mergeAndNormalizePolicy in repo-mesh-types.ts. This thin alias keeps the local
+// call sites (createMesh/updateMesh) reading naturally while ensuring config
+// writes go through the exact same normalizer the scheduler/display paths use.
+const mergeMeshPolicy = mergeAndNormalizePolicy;
 
 export function listMeshes(): LocalMeshEntry[] {
     return loadMeshConfig().meshes;

package/src/index.ts

@@ -135,6 +135,9 @@ export type {
   RepoMeshLedgerStatus,
   MeshAsyncJobLifecycle,
   RepoMeshSchedulingStrategy,
+  RepoMeshSchedulingStatus,
+  RepoMeshNodeSchedulingStatus,
+  RepoMeshNodeProviderSchedulingStatus,
 } from './repo-mesh-types.js';
 export {
   DEFAULT_MESH_POLICY,
@@ -143,6 +146,12 @@ export {
   DEFAULT_MESH_SCHEDULING_STRATEGY,
   normalizeMeshSchedulingStrategy,
   resolveNodeSchedulingPriority,
+  resolveProviderMaxParallel,
+  mergeAndNormalizePolicy,
+  normalizeAutoFastForwardPolicy,
+  resolveMaxParallelTasks,
+  MESH_MAX_PARALLEL_TASKS_MIN,
+  MESH_MAX_PARALLEL_TASKS_MAX,
   MESH_CONVERGE_REFINE_TAG,
   MESH_CONVERGE_FAST_FORWARD_TAG,
   resolveAutoConvergeCodeChange,
@@ -246,6 +255,10 @@ export type { MeshActiveWorkRecord, MeshActiveWorkStatus, MeshActiveWorkSummary,
 export { buildMeshAsyncRefineJobs, summarizeMeshAsyncRefineJobs, STALE_TERMINAL_REFINE_WINDOW_MS, RECENT_TERMINAL_REFINE_CAP } from './mesh/mesh-refine-status.js';
 export type { MeshAsyncRefineJobStatus, MeshAsyncRefineJobSummary, MeshAsyncRefineJobsSummary } from './mesh/mesh-refine-status.js';
 
+// ── Mesh Scheduling Runtime (observability projection) ──
+export { buildMeshSchedulingRuntime } from './mesh/mesh-scheduling-runtime.js';
+export type { MeshSchedulingRuntime, MeshNodeSchedulingRuntime, MeshNodeProviderSchedulingRuntime } from './mesh/mesh-scheduling-runtime.js';
+
 // ── Mesh Host Ownership ──
 export { buildMeshHostRequiredFailure, createDefaultMeshHostMetadata, isMeshHostOwner, normalizeMeshDaemonRole, requireMeshHostQueueOwner, resolveMeshHostStatus } from './mesh/mesh-host-ownership.js';
 

package/src/mesh/coordinator-prompt.ts

@@ -31,7 +31,7 @@ import type {
     RepoMeshStatus,
     RepoMeshNodeStatus,
 } from '../repo-mesh-types.js';
-import { DEFAULT_MESH_POLICY } from '../repo-mesh-types.js';
+import { mergeAndNormalizePolicy } from '../repo-mesh-types.js';
 
 /**
  * Cheap, locally-derived "what just happened" snapshot for the coordinator
@@ -196,7 +196,7 @@ Repository: \`${mesh.repoIdentity}\`${mesh.defaultBranch ? `\nDefault branch: \`
     if (operatingNotes) sections.push(operatingNotes);
 
     // ── Policy ──
-    sections.push(buildPolicySection({ ...DEFAULT_MESH_POLICY, ...(mesh.policy || {}) }));
+    sections.push(buildPolicySection(mergeAndNormalizePolicy(undefined, mesh.policy)));
 
     // ── Tools ──
     sections.push(TOOLS_SECTION);
@@ -278,7 +278,7 @@ function expandPromptPlaceholders(template: string, ctx: CoordinatorPromptContex
         mission: ctx.missionSection?.trim() || '',
         recentActivity: buildRecentActivitySection(ctx.recentActivity) || '',
         operatingNotes: buildOperatingNotesSection(ctx.operatingNotes) || '',
-        policy: buildPolicySection({ ...DEFAULT_MESH_POLICY, ...(mesh.policy || {}) }),
+        policy: buildPolicySection(mergeAndNormalizePolicy(undefined, mesh.policy)),
         tools: TOOLS_SECTION,
         workflow: WORKFLOW_SECTION,
         rules: buildRulesSection(coordinatorCliType),

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

@@ -9,6 +9,7 @@ import { markSessionDeliveriesTerminal, updateSessionDeliveryStatus, recordCompl
 import { MeshRuntimeStore } from './mesh-runtime-store.js';
 import { queuePendingMeshCoordinatorEvent, drainPendingMeshCoordinatorEvents } from './mesh-events-pending.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';
@@ -92,6 +93,38 @@ function recoverMeshIdByNodeId(nodeId: string): string {
     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.
@@ -886,8 +919,24 @@ function injectMeshSystemMessage(components: DaemonComponents, args: {
             // 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).
-            if (startedTaskId || sessionHasActiveAssignment(args.meshId, sessionId)) {
+            //
+            // 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); }
@@ -1218,7 +1267,14 @@ export function handleMeshForwardEvent(components: DaemonComponents, payload: Re
     // The nodeId is a stable coordinator-side fact and resolves timing-independently.
     const meshId = readNonEmptyString(payload.meshId)
         || (workspace ? readNonEmptyString(getCachedMeshByWorkspace(workspace)?.id) : '')
-        || recoverMeshIdByNodeId(nodeId);
+        || 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.
@@ -1348,6 +1404,11 @@ function forwardUnresolvedDelegateEvent(
         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

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

@@ -903,6 +903,29 @@ export class MeshRuntimeStore {
         `).run({ status, meshId, sessionId, updatedAt: now });
     }
 
+    /**
+     * MESH-DISPATCH-MISROUTE (fix 3, consumer residual): resolve the task_id of the SINGLE
+     * non-terminal direct dispatch a session owns. Returns the task_id only when the session
+     * holds exactly ONE active ('dispatched'/'acked') row — the case where a taskId-less
+     * lifecycle event (a legacy/relayed worker whose producer never stamped meshActiveTaskId)
+     * unambiguously belongs to that one dispatch. With zero rows there is nothing to ack; with
+     * two or more (a re-dispatch/nudge sibling) the firing event's owner is ambiguous, so we
+     * return null and the caller MUST NOT fall back to the session_id sweep that would flip a
+     * sibling row ("may flip a sibling dispatch row"). This narrows the legacy fallback to the
+     * only safe case instead of removing the producer-side TASKIDLESS stamp's safety net.
+     */
+    getSoleActiveDirectDispatchTaskId(meshId: string, sessionId: string): string | null {
+        if (!sessionId) return null;
+        const rows = this.db.prepare(`
+            SELECT task_id FROM mesh_direct_dispatches
+            WHERE mesh_id = ? AND session_id = ?
+              AND status NOT IN ('completed', 'failed', 'stale')
+        `).all(meshId, sessionId) as Array<{ task_id: string }>;
+        if (rows.length !== 1) return null;
+        const taskId = typeof rows[0]?.task_id === 'string' ? rows[0].task_id.trim() : '';
+        return taskId || null;
+    }
+
     cleanupTerminalDirectDispatches(olderThanMs: number): void {
         const cutoff = new Date(Date.now() - olderThanMs).toISOString();
         this.db.prepare(`