@adhdev/daemon-core 0.9.82-rc.32 → 0.9.82-rc.320

package/src/mesh/mesh-work-queue.ts

@@ -1,28 +1,160 @@
-import { existsSync, writeFileSync, readFileSync, openSync, closeSync, unlinkSync } from 'fs';
-import { join } from 'path';
 import { randomUUID } from 'crypto';
-import { getLedgerDir } from './mesh-ledger.js';
+import { requireMeshHostQueueOwner } from './mesh-host-ownership.js';
+import type { RepoMeshDaemonRole, LocalMeshNodeEntry } from '../repo-mesh-types.js';
+import { MESH_CONVERGE_REFINE_TAG, resolveAutoConvergeCodeChange, resolveTaskAffinityRole } from '../repo-mesh-types.js';
+import { MeshRuntimeStore } from './mesh-runtime-store.js';
+import { getMesh } from '../config/mesh-config.js';
 
 export type MeshTaskStatus = 'pending' | 'assigned' | 'completed' | 'failed' | 'cancelled';
 export type MeshActiveTaskStatus = Extract<MeshTaskStatus, 'pending' | 'assigned'>;
 export type MeshHistoricalTaskStatus = Extract<MeshTaskStatus, 'completed' | 'failed' | 'cancelled'>;
+export type MeshTaskMode = 'code_change' | 'validation' | 'live_debug_readonly' | 'launch_app' | 'convergence';
 
 export const ACTIVE_MESH_QUEUE_STATUSES: MeshActiveTaskStatus[] = ['pending', 'assigned'];
 export const HISTORICAL_MESH_QUEUE_STATUSES: MeshHistoricalTaskStatus[] = ['completed', 'failed', 'cancelled'];
+export const MESH_TASK_MODES: MeshTaskMode[] = ['code_change', 'validation', 'live_debug_readonly', 'launch_app', 'convergence'];
+
+export interface MeshTaskModeValidationResult {
+    valid: boolean;
+    taskMode?: MeshTaskMode;
+    violations: string[];
+    allowedOperations?: string[];
+}
+
+const LIVE_DEBUG_READONLY_FORBIDDEN: Array<{ label: string; pattern: RegExp }> = [
+    { label: 'source_edit', pattern: /\b(edit|modify|patch|apply\s+patch|write\s+(?:to\s+)?(?:file|source)|overwrite|delete\s+file|remove\s+file|create\s+file|touch\s+file)\b/i },
+    { label: 'checkpoint', pattern: /\b(checkpoint|mesh_checkpoint)\b/i },
+    { label: 'deploy_or_version_bump', pattern: /\b(deploy|wrangler\s+deploy|version[-\s]?bump|npm\s+version|release|npm\s+publish|yarn\s+publish|pnpm\s+publish)\b/i },
+    { label: 'destructive_shell', pattern: /\b(rm\s+-rf|mv\s+\S+\s+\S+|truncate\s|tee\s+\S+|sed\s+-i|shred\b)\b/i },
+    { label: 'package_install', pattern: /\b(npm\s+(?:install|i|add|link|uninstall|remove)|yarn\s+(?:add|remove|link)|pnpm\s+(?:add|remove|link)|pip\s+install|brew\s+install|apt\s+install|cargo\s+install)\b/i },
+    { label: 'container_mutation', pattern: /\b(docker\s+(?:build|run|exec|push|tag|rmi|rm|create|start|stop|kill)|kubectl\s+(?:apply|delete|patch|replace|create|scale))\b/i },
+];
+
+/**
+ * Git subcommands that mutate the working tree, index, refs, or remote.
+ * `stash` and `checkout` are intentionally absent here: they have read-only
+ * variants (`git stash list`/`show`, `git checkout-index`) and are classified
+ * token-by-token in {@link detectGitMutation} rather than by bare keyword.
+ */
+const GIT_MUTATION_SUBCOMMANDS = new Set([
+    'add', 'commit', 'push', 'reset', 'rebase', 'clean', 'switch', 'merge',
+    'tag', 'restore', 'rm', 'mv', 'cherry-pick', 'revert', 'pull', 'fetch',
+    'am', 'apply', 'gc', 'prune',
+]);
+
+/**
+ * Read-only `git stash` variants. Any other `git stash <x>` (pop/apply/drop/
+ * push/save/clear, or bare `git stash` which defaults to push) is a mutation.
+ */
+const GIT_STASH_READONLY_SUBCOMMANDS = new Set(['list', 'show']);
+
+/**
+ * Detects a true git mutation in free-text task message, token-aware so that
+ * read-only diagnostics (`git stash list`, `git stash show --stat`,
+ * `git checkout-index`, `git status`, `git diff`, `git log`, ...) are allowed.
+ * Returns true only when a genuine mutating git invocation is present.
+ */
+function detectGitMutation(message: string): boolean {
+    const re = /\bgit\s+([a-z][a-z0-9-]*)/gi;
+    let match: RegExpExecArray | null;
+    while ((match = re.exec(message)) !== null) {
+        const sub = match[1].toLowerCase();
+        if (GIT_MUTATION_SUBCOMMANDS.has(sub)) return true;
+        if (sub === 'stash') {
+            // Token following `git stash`; read-only only for list/show.
+            const after = message.slice(re.lastIndex).match(/^\s+([a-z][a-z0-9-]*)/i);
+            const next = after ? after[1].toLowerCase() : '';
+            if (!GIT_STASH_READONLY_SUBCOMMANDS.has(next)) return true; // bare stash = push, or pop/apply/drop/...
+        } else if (sub === 'checkout') {
+            // `git checkout <ref/path>` mutates; `git checkout-index` is matched
+            // as its own token by the regex (sub === 'checkout-index') and is read-only.
+            return true;
+        } else if (sub === 'submodule') {
+            // `git submodule update` mutates; `git submodule status` is read-only.
+            const after = message.slice(re.lastIndex).match(/^\s+([a-z][a-z0-9-]*)/i);
+            const next = after ? after[1].toLowerCase() : '';
+            if (next === 'update' || next === 'add' || next === 'sync' || next === 'deinit') return true;
+        } else if (sub === 'worktree') {
+            const after = message.slice(re.lastIndex).match(/^\s+([a-z][a-z0-9-]*)/i);
+            const next = after ? after[1].toLowerCase() : '';
+            if (next === 'add' || next === 'remove' || next === 'move' || next === 'prune') return true;
+        }
+        // checkout-index, stash-with-no-next-already-handled, status/diff/log/show/
+        // rev-parse/branch/submodule status fall through as read-only.
+    }
+    return false;
+}
+
+export function normalizeMeshTaskMode(value: unknown): MeshTaskMode | undefined {
+    if (typeof value !== 'string') return undefined;
+    const normalized = value.trim() as MeshTaskMode;
+    return (MESH_TASK_MODES as string[]).includes(normalized) ? normalized : undefined;
+}
+
+export function validateMeshTaskModeRequest(mode: unknown, message: string): MeshTaskModeValidationResult {
+    const taskMode = normalizeMeshTaskMode(mode);
+    if (!taskMode) {
+        return { valid: true, violations: [] };
+    }
+    if (taskMode !== 'live_debug_readonly') {
+        return { valid: true, taskMode, violations: [] };
+    }
+    const text = message || '';
+    const violations = LIVE_DEBUG_READONLY_FORBIDDEN
+        .filter(rule => rule.pattern.test(text))
+        .map(rule => rule.label);
+    if (detectGitMutation(text)) {
+        violations.push('git_mutation');
+    }
+    return {
+        valid: violations.length === 0,
+        taskMode,
+        violations,
+        allowedOperations: [
+            'process/log/window/port/session inspection',
+            'read-only filesystem listing/reading',
+            'status probes and keep-running handle reporting',
+            'diagnostic summaries without source edits, commits, checkpoints, pushes, deploys, resets, rebases, or destructive cleanups',
+        ],
+    };
+}
 
 export interface MeshWorkQueueEntry {
     id: string;
     meshId: string;
     message: string;
     status: MeshTaskStatus;
+    taskMode?: MeshTaskMode;
     /** If specified, only this node can claim the task (used by legacy mesh_send_task) */
     targetNodeId?: string;
     /** If specified, only this runtime session can claim the task */
     targetSessionId?: string;
+    /** If specified, a node must expose all tags before it can claim the task. */
+    requiredTags?: string[];
+    /**
+     * M1: ids of tasks that must reach 'completed' before this task is claimable.
+     * Forward references (ids not yet enqueued) are allowed for batch flows and
+     * simply keep the task waiting until the referenced task exists and completes.
+     */
+    dependsOn?: string[];
+    /** M1/M3: mission this task belongs to (joins mesh_missions). */
+    missionId?: string;
+    /**
+     * M1: why this task is held back (e.g. "dependency_failed:<taskId>").
+     * Only set by the system on dependency failure under the 'block' policy;
+     * waiting-on-dependency state is computed at view time, not stored.
+     */
+    blockedReason?: string;
     /** The node that actually claimed and is executing the task */
     assignedNodeId?: string;
     /** The session currently executing the task */
     assignedSessionId?: string;
+    /**
+     * Provider type of the session that claimed the task. Recorded so the queue
+     * can enforce per-(node, provider) maxParallel caps (RepoMeshNodePolicy
+     * providerRoles) by counting active assignments grouped by node + provider.
+     */
+    assignedProviderType?: string;
     /** Human/operator reason for terminal cancellation. */
     cancelReason?: string;
     cancelledAt?: string;
@@ -30,6 +162,8 @@ export interface MeshWorkQueueEntry {
     requeueReason?: string;
     requeuedAt?: string;
     requeueCount?: number;
+    /** Max automatic requeue attempts. When requeueCount reaches this, task is auto-failed. */
+    maxRetries?: number;
     /** Last automatic queue session spin-up attempt, for mesh_view_queue/debug visibility. */
     autoLaunch?: {
         status: 'skipped' | 'started' | 'failed' | 'completed';
@@ -45,50 +179,268 @@ export interface MeshWorkQueueEntry {
     updatedAt: string;
 }
 
-function getQueuePath(meshId: string): string {
-    const safe = meshId.replace(/[^a-zA-Z0-9_-]/g, '_');
-    return join(getLedgerDir(), `${safe}.queue.json`);
+export interface MeshQueueMutationOptions {
+    ownerRole?: RepoMeshDaemonRole;
 }
 
-function getLockPath(meshId: string): string {
-    const safe = meshId.replace(/[^a-zA-Z0-9_-]/g, '_');
-    return join(getLedgerDir(), `${safe}.queue.lock`);
+export function normalizeMeshCapabilityTags(value: unknown): string[] {
+    if (!Array.isArray(value)) return [];
+    const seen = new Set<string>();
+    return value
+        .map(tag => typeof tag === 'string' ? tag.trim() : '')
+        .filter(Boolean)
+        .filter(tag => {
+            if (seen.has(tag)) return false;
+            seen.add(tag);
+            return true;
+        });
+}
+
+function firstProviderPriority(policy: unknown): string | undefined {
+    const raw = policy && typeof policy === 'object' && !Array.isArray(policy)
+        ? (policy as Record<string, unknown>).providerPriority
+        : undefined;
+    if (!Array.isArray(raw)) return undefined;
+    return raw.find(type => typeof type === 'string' && type.trim())?.trim();
 }
 
 /**
- * Simple advisory file lock using O_EXCL (atomic create) for queue mutations.
- * Retries up to 10 times at 30 ms intervals; proceeds without lock on timeout
- * to prevent deadlock (best-effort — far better than no locking at all).
+ * Synthetic `role=<x>` capability tags advertised by a node's policy.providerRoles.
+ *
+ * When a specific `providerType` is being evaluated, only that provider's declared
+ * role is emitted — so role-based routing (requiredTags: ["role=validation"]) gates
+ * the *selected* provider through the ordinary capability-tag filter. When no
+ * provider is selected (node-level eligibility scan), every declared role is emitted
+ * so the node passes the filter if ANY of its providers could satisfy the role; the
+ * per-provider tag set then narrows it during provider selection.
+ *
+ * Roles are lowercased and deduped. Missing/empty providerRoles emits nothing, so a
+ * node that never declares roles advertises no `role=` tags and is therefore only
+ * matched by role-unconstrained tasks (full backward compatibility).
  */
-function withQueueLock<T>(meshId: string, fn: () => T): T {
-    const lockPath = getLockPath(meshId);
-    let fd = -1;
-    for (let i = 0; i < 10; i++) {
-        try { fd = openSync(lockPath, 'wx'); break; } catch {
-            const deadline = Date.now() + 30;
-            while (Date.now() < deadline) { /* spin */ }
-        }
+function roleCapabilityTags(policy: unknown, providerType: string | undefined): string[] {
+    const roles = policy && typeof policy === 'object' && !Array.isArray(policy)
+        ? (policy as Record<string, unknown>).providerRoles
+        : undefined;
+    if (!Array.isArray(roles)) return [];
+    const wantedProvider = typeof providerType === 'string' && providerType.trim()
+        ? providerType.trim().toLowerCase()
+        : '';
+    const out: string[] = [];
+    for (const entry of roles) {
+        if (!entry || typeof entry !== 'object') continue;
+        const type = typeof (entry as any).providerType === 'string'
+            ? (entry as any).providerType.trim().toLowerCase()
+            : '';
+        const role = typeof (entry as any).role === 'string'
+            ? (entry as any).role.trim().toLowerCase()
+            : '';
+        if (!role) continue;
+        // When narrowing to a selected provider, only emit that provider's role.
+        if (wantedProvider && type && type !== wantedProvider) continue;
+        out.push(`role=${role}`);
     }
-    try { return fn(); } finally {
-        if (fd !== -1) try { closeSync(fd); } catch { /* noop */ }
-        try { unlinkSync(lockPath); } catch { /* already removed */ }
+    return out;
+}
+
+export function buildMeshNodeCapabilityTags(
+    node: { capabilities?: unknown; policy?: unknown; isLocalWorktree?: unknown; worktreeBranch?: unknown } | undefined,
+    providerType?: string,
+): string[] {
+    const provider = typeof providerType === 'string' && providerType.trim()
+        ? providerType.trim()
+        : firstProviderPriority(node?.policy);
+    const worktreeBranch = typeof node?.worktreeBranch === 'string' && node.worktreeBranch.trim()
+        ? node.worktreeBranch.trim()
+        : null;
+    return normalizeMeshCapabilityTags([
+        ...(Array.isArray(node?.capabilities) ? node.capabilities : []),
+        `os=${process.platform}`,
+        `arch=${process.arch}`,
+        ...(provider ? [`provider=${provider}`] : []),
+        // Worktree nodes automatically expose a "worktree=<branch>" tag so that
+        // mesh_enqueue_task with required_tags: ["worktree=<branch>"] routes
+        // only to the matching worktree node.
+        ...(node?.isLocalWorktree === true && worktreeBranch ? [`worktree=${worktreeBranch}`] : []),
+        // Convergence routing: advertise how this node can land its work onto base.
+        //   - converge=refine: local worktree nodes (on ANY machine — refine_mesh_node
+        //     now forwards to the owning daemon) can run the Refinery merge → push →
+        //     cleanup against their own checkout, so they accept code_change tasks.
+        //   - converge=fast_forward: non-worktree nodes (the machine itself) can only
+        //     ff/push an already-converged branch; they are NOT a destination for
+        //     code_change work (a worktree is created first, and that worktree node
+        //     receives the task instead). Reuses the ordinary required-tags filter —
+        //     the load-balancing scheduler auto-injects converge=refine for code_change
+        //     so such work is hard-filtered onto refine-capable nodes.
+        ...(node?.isLocalWorktree === true ? ['converge=refine'] : ['converge=fast_forward']),
+        // Role-based routing: advertise role=<x> for each (node, provider) role
+        // declared in policy.providerRoles. Narrowed to the selected provider when
+        // one is given so the chosen provider must match a task's required role;
+        // when no provider is selected, all declared roles are advertised for the
+        // node-level eligibility scan. Reuses the ordinary required-tags filter —
+        // no separate role field/gate.
+        ...roleCapabilityTags(node?.policy, providerType),
+    ]);
+}
+
+export function nodeSatisfiesRequiredTags(requiredTags: unknown, capabilityTags: unknown): boolean {
+    const required = normalizeMeshCapabilityTags(requiredTags);
+    if (required.length === 0) return true;
+    const available = new Set(normalizeMeshCapabilityTags(capabilityTags));
+    return required.every(tag => available.has(tag));
+}
+
+/**
+ * Convergence-aware required-tags resolution (load-balancing scheduler, opt-in).
+ *
+ * When the mesh enables policy.autoConvergeCodeChange, a `converge=refine` required
+ * tag is merged into a code_change task's required tags at enqueue time, so the
+ * scheduler hard-filters the task onto refine-capable worktree nodes only (on any
+ * machine — refine_mesh_node forwards to the owning daemon). Because the tag is
+ * persisted on the queue entry, BOTH the eligibility scan (maybeAutoLaunchOneQueueSession)
+ * and the claim transaction (claimNextQueueTask → nodeSatisfiesRequiredTags) enforce
+ * it consistently.
+ *
+ * Strict backward compatibility — the injection is skipped (returns the explicit tags
+ * unchanged) when ANY of:
+ *   - the mesh does not opt in (autoConvergeCodeChange !== true), or
+ *   - the task is not code_change (validation / live_debug_readonly / launch_app /
+ *     convergence carry no merge cost and may run anywhere), or
+ *   - the task is explicitly targeted (targetNodeId): the operator chose the node, so
+ *     we do not second-guess it by filtering on convergence capability.
+ * Idempotent: normalizeMeshCapabilityTags dedupes, so re-injection is a no-op.
+ */
+export function resolveConvergeRequiredTags(
+    meshId: string,
+    taskMode: MeshTaskMode | undefined,
+    explicitRequiredTags: string[],
+    opts?: { targetNodeId?: string },
+): string[] {
+    if (taskMode !== 'code_change') return explicitRequiredTags;
+    if (typeof opts?.targetNodeId === 'string' && opts.targetNodeId.trim()) return explicitRequiredTags;
+    let optedIn = false;
+    try {
+        optedIn = resolveAutoConvergeCodeChange(getMesh(meshId)?.policy as any);
+    } catch {
+        optedIn = false;
     }
+    if (!optedIn) return explicitRequiredTags;
+    return normalizeMeshCapabilityTags([...explicitRequiredTags, MESH_CONVERGE_REFINE_TAG]);
 }
 
-function readQueue(meshId: string): MeshWorkQueueEntry[] {
-    const path = getQueuePath(meshId);
-    if (!existsSync(path)) return [];
+/**
+ * Whether any node in the mesh advertises the given `role=<role>` capability tag.
+ * Reuses buildMeshNodeCapabilityTags (node-level scan: every declared providerRole is
+ * advertised) so the membership test matches the exact tag the claim/eligibility filter
+ * would see. Used to keep task-affinity SOFT — if no node can satisfy the role, the
+ * injection is skipped so the task falls back to ordinary least_loaded eligibility
+ * rather than being blocked.
+ */
+function meshHasNodeAdvertisingRole(nodes: LocalMeshNodeEntry[] | undefined, role: string): boolean {
+    if (!Array.isArray(nodes) || nodes.length === 0) return false;
+    const wanted = `role=${role}`;
+    return nodes.some(node => buildMeshNodeCapabilityTags(node as any).includes(wanted));
+}
+
+/**
+ * Declarative task_mode → role affinity resolution (precedent: resolveConvergeRequiredTags).
+ *
+ * At enqueue time, auto-inject a `role=<role>` required tag derived from the task's
+ * taskMode via the mesh's policy.taskAffinity (falling back to DEFAULT_TASKMODE_ROLE_MAP),
+ * so the task hard-filters onto nodes advertising that role through the ordinary
+ * required-tags path (nodeSatisfiesRequiredTags) — no claim/scheduler change needed.
+ *
+ * Strict precedence / backward compatibility — the injection is skipped (returns the
+ * explicit tags unchanged) when ANY of:
+ *   - the caller pinned an explicit targetNodeId (operator chose the node), or
+ *   - the caller supplied their own non-empty required_tags (caller's routing wins), or
+ *   - affinity resolves to no role (policy disabled, blank override, or unknown mode), or
+ *   - SOFT fallback: no node in the mesh advertises the resolved role=<role> — injecting
+ *     would block the task with zero eligible nodes, so we skip it and let least_loaded
+ *     eligibility take over (a warning is logged so the misconfiguration is visible).
+ * Idempotent: normalizeMeshCapabilityTags dedupes.
+ */
+export function resolveTaskAffinityRequiredTags(
+    meshId: string,
+    taskMode: MeshTaskMode | undefined,
+    explicitRequiredTags: string[],
+    opts?: { targetNodeId?: string; callerSpecifiedRequiredTags?: boolean },
+): string[] {
+    // Caller pinned the node, or brought their own required_tags → respect it verbatim.
+    if (typeof opts?.targetNodeId === 'string' && opts.targetNodeId.trim()) return explicitRequiredTags;
+    if (opts?.callerSpecifiedRequiredTags === true) return explicitRequiredTags;
+
+    let mesh;
     try {
-        const content = readFileSync(path, 'utf-8');
-        return JSON.parse(content) as MeshWorkQueueEntry[];
+        mesh = getMesh(meshId);
     } catch {
-        return [];
+        return explicitRequiredTags;
+    }
+    const role = resolveTaskAffinityRole(taskMode, mesh?.policy?.taskAffinity);
+    if (!role) return explicitRequiredTags;
+
+    // SOFT affinity: only hard-filter when the role actually exists in the mesh.
+    if (!meshHasNodeAdvertisingRole(mesh?.nodes, role)) {
+        try {
+            console.warn(`[mesh] task_affinity: no node advertises role=${role} for taskMode=${taskMode} in mesh ${meshId}; skipping injection (least_loaded fallback)`);
+        } catch { /* logging is best-effort */ }
+        return explicitRequiredTags;
     }
+    return normalizeMeshCapabilityTags([...explicitRequiredTags, `role=${role}`]);
+}
+
+function withQueueLock<T>(_meshId: string, fn: () => T): T {
+    return MeshRuntimeStore.getInstance().transaction(fn);
+}
+
+function readQueue(meshId: string): MeshWorkQueueEntry[] {
+    return MeshRuntimeStore.getInstance().getQueueEntries(meshId);
 }
 
 function writeQueue(meshId: string, queue: MeshWorkQueueEntry[]): void {
-    const path = getQueuePath(meshId);
-    writeFileSync(path, JSON.stringify(queue, null, 2), 'utf-8');
+    MeshRuntimeStore.getInstance().replaceQueue(meshId, queue);
+}
+
+function normalizeDependsOn(value: unknown): string[] {
+    if (!Array.isArray(value)) return [];
+    const seen = new Set<string>();
+    return value
+        .map(id => typeof id === 'string' ? id.trim() : '')
+        .filter(Boolean)
+        .filter(id => {
+            if (seen.has(id)) return false;
+            seen.add(id);
+            return true;
+        });
+}
+
+/**
+ * M1: detect dependency cycles before enqueue. Walks the dependency graph of
+ * existing queue entries plus the new task's edges. Fail-closed: a cycle
+ * rejects the enqueue entirely. Synchronous and bounded by queue size.
+ */
+export function assertNoDependencyCycle(meshId: string, newTaskId: string, dependsOn: string[]): void {
+    if (dependsOn.length === 0) return;
+    if (dependsOn.includes(newTaskId)) {
+        throw new Error(`dependency_cycle_detected: task '${newTaskId}' cannot depend on itself`);
+    }
+    const adjacency = new Map<string, string[]>();
+    for (const entry of readQueue(meshId)) {
+        adjacency.set(entry.id, normalizeDependsOn(entry.dependsOn));
+    }
+    adjacency.set(newTaskId, dependsOn);
+    // DFS from the new task: if we can reach newTaskId again, the edges form a cycle.
+    const stack = [...dependsOn];
+    const visited = new Set<string>();
+    while (stack.length > 0) {
+        const current = stack.pop()!;
+        if (current === newTaskId) {
+            throw new Error(`dependency_cycle_detected: task '${newTaskId}' is part of a dependency cycle via '${dependsOn.join(', ')}'`);
+        }
+        if (visited.has(current)) continue;
+        visited.add(current);
+        stack.push(...(adjacency.get(current) ?? []));
+    }
 }
 
 /**
@@ -97,22 +449,133 @@ function writeQueue(meshId: string, queue: MeshWorkQueueEntry[]): void {
 export function enqueueTask(
     meshId: string,
     message: string,
-    opts?: { targetNodeId?: string; targetSessionId?: string }
+    opts?: {
+        targetNodeId?: string;
+        targetSessionId?: string;
+        taskMode?: MeshTaskMode | string;
+        requiredTags?: string[];
+        /** M1: tasks that must complete before this one is claimable. */
+        dependsOn?: string[];
+        /** M1/M3: mission this task belongs to. */
+        missionId?: string;
+        /** Explicit task id for batch/template flows (M5). Random UUID when omitted. */
+        id?: string;
+    } & MeshQueueMutationOptions,
 ): MeshWorkQueueEntry {
+    requireMeshHostQueueOwner(opts);
+    const modeValidation = validateMeshTaskModeRequest(opts?.taskMode, message);
+    if (!modeValidation.valid) {
+        throw new Error(`live_debug_readonly_guardrail_violation: forbidden operations (${modeValidation.violations.join(', ')})`);
+    }
+    const id = typeof opts?.id === 'string' && opts.id.trim() ? opts.id.trim() : randomUUID();
+    const dependsOn = normalizeDependsOn(opts?.dependsOn);
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
+        if (MeshRuntimeStore.getInstance().findQueueEntryById(meshId, id)) {
+            throw new Error(`duplicate_task_id: task '${id}' already exists in mesh '${meshId}'`);
+        }
+        assertNoDependencyCycle(meshId, id, dependsOn);
+        // Did the caller bring their own routing tags? Affinity auto-injection is
+        // suppressed when so (caller's required_tags win), matching the spec — measured
+        // against the ORIGINAL caller tags, before convergence augmentation below.
+        const callerTags = normalizeMeshCapabilityTags(opts?.requiredTags);
+        const callerSpecifiedRequiredTags = callerTags.length > 0;
+        // Convergence routing (opt-in): auto-inject converge=refine for code_change
+        // tasks so they hard-filter onto refine-capable worktree nodes. No-op unless
+        // the mesh opts in; explicit target_node_id / required_tags are preserved.
+        let resolvedRequiredTags = resolveConvergeRequiredTags(
+            meshId,
+            modeValidation.taskMode,
+            callerTags,
+            { targetNodeId: opts?.targetNodeId },
+        );
+        // Declarative task_mode → role affinity (precedent: convergence routing above):
+        // inject role=<role> for the task's taskMode unless the caller pinned a node or
+        // supplied their own required_tags. SOFT — skipped when no node advertises the
+        // role (least_loaded fallback). Built-in default mapping applies even without a
+        // taskAffinity config block; { enabled:false } disables it.
+        resolvedRequiredTags = resolveTaskAffinityRequiredTags(
+            meshId,
+            modeValidation.taskMode,
+            resolvedRequiredTags,
+            { targetNodeId: opts?.targetNodeId, callerSpecifiedRequiredTags },
+        );
         const entry: MeshWorkQueueEntry = {
-            id: randomUUID(),
+            id,
             meshId,
             message,
             status: 'pending',
+            taskMode: modeValidation.taskMode,
             targetNodeId: opts?.targetNodeId,
             targetSessionId: opts?.targetSessionId,
+            requiredTags: resolvedRequiredTags,
+            ...(dependsOn.length > 0 ? { dependsOn } : {}),
+            ...(typeof opts?.missionId === 'string' && opts.missionId.trim() ? { missionId: opts.missionId.trim() } : {}),
             createdAt: new Date().toISOString(),
             updatedAt: new Date().toISOString(),
         };
-        queue.push(entry);
-        writeQueue(meshId, queue);
+        MeshRuntimeStore.getInstance().insertQueueEntry(entry);
+        return entry;
+    });
+}
+
+/**
+ * Record a direct-dispatch task (mesh_send_task) as an already-assigned queue
+ * entry so it is attributable to a mission.
+ *
+ * Direct dispatch normally bypasses the queue entirely — the task lives only in
+ * the ledger + mesh_direct_dispatches table, neither of which carries a
+ * missionId, so {@link summarizeMissionTasks}/{@link computeMeshTaskStats}
+ * (which both scan the queue for `task.missionId`) count it as 0. When a
+ * mission is attached, we materialise the same queue entry shape an enqueued
+ * task would have, but pre-assigned to the dispatched node/session and stamped
+ * with the dispatch timestamp. The terminal event path (updateSessionTaskStatus
+ * → findAssignedBySession) then flips it to completed/failed exactly like a
+ * pulled task, so mission total + completed aggregates work with no extra wiring.
+ *
+ * Intentionally separate from {@link enqueueTask}: enqueue creates `pending`
+ * work for the queue to assign, whereas this records work already dispatched
+ * out-of-band. They share the missionId stamping rule and mode validation.
+ */
+export function recordDirectDispatchTask(
+    meshId: string,
+    message: string,
+    opts: {
+        id: string;
+        missionId: string;
+        assignedNodeId?: string;
+        assignedSessionId?: string;
+        taskMode?: MeshTaskMode | string;
+        dispatchedAt?: string;
+    },
+): MeshWorkQueueEntry | null {
+    const missionId = typeof opts.missionId === 'string' ? opts.missionId.trim() : '';
+    if (!missionId) return null;
+    const taskId = typeof opts.id === 'string' ? opts.id.trim() : '';
+    if (!taskId) return null;
+    const modeValidation = validateMeshTaskModeRequest(opts.taskMode, message);
+    if (!modeValidation.valid) {
+        throw new Error(`live_debug_readonly_guardrail_violation: forbidden operations (${modeValidation.violations.join(', ')})`);
+    }
+    const now = opts.dispatchedAt && opts.dispatchedAt.trim() ? opts.dispatchedAt : new Date().toISOString();
+    return withQueueLock(meshId, () => {
+        if (MeshRuntimeStore.getInstance().findQueueEntryById(meshId, taskId)) {
+            // Already materialised (e.g. retry of the same dispatch) — leave it untouched.
+            return null;
+        }
+        const entry: MeshWorkQueueEntry = {
+            id: taskId,
+            meshId,
+            message,
+            status: 'assigned',
+            ...(modeValidation.taskMode ? { taskMode: modeValidation.taskMode } : {}),
+            missionId,
+            ...(opts.assignedNodeId ? { targetNodeId: opts.assignedNodeId, assignedNodeId: opts.assignedNodeId } : {}),
+            ...(opts.assignedSessionId ? { targetSessionId: opts.assignedSessionId, assignedSessionId: opts.assignedSessionId } : {}),
+            dispatchTimestamp: now,
+            createdAt: now,
+            updatedAt: now,
+        };
+        MeshRuntimeStore.getInstance().insertQueueEntry(entry);
         return entry;
     });
 }
@@ -121,43 +584,83 @@ export function enqueueTask(
  * Get all tasks in the queue, optionally filtered by status.
  */
 export function getQueue(meshId: string, opts?: { status?: MeshTaskStatus[] }): MeshWorkQueueEntry[] {
-    let queue = readQueue(meshId);
-    if (opts?.status?.length) {
-        const statuses = new Set(opts.status);
-        queue = queue.filter(q => statuses.has(q.status));
-    }
-    return queue;
+    return MeshRuntimeStore.getInstance().getQueueEntries(meshId, opts?.status?.length ? opts.status : undefined);
+}
+
+export function getMeshQueueRevision(meshId: string): string {
+    return MeshRuntimeStore.getInstance().getQueueRevision(meshId);
 }
 
 /**
  * Find the next pending task that this node is allowed to claim, and mark it as assigned.
+ *
+ * `opts.providerType` is stamped onto the claimed entry (assignedProviderType) so
+ * per-(node, provider) caps can be counted. `opts.providerMaxParallel`, when set,
+ * is the enforced per-(node, provider) cap from RepoMeshNodePolicy.providerRoles:
+ * a task is not assigned to this (node, provider) once it already has that many
+ * active assignments. This composes with the global/taskMode caps (stricter wins).
  */
-export function claimNextTask(meshId: string, nodeId: string, sessionId: string): MeshWorkQueueEntry | null {
-    return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const hasActiveAssignment = queue.some(q => q.status === 'assigned' && (
-            q.assignedSessionId === sessionId || q.assignedNodeId === nodeId
-        ));
-        if (hasActiveAssignment) return null;
-        let targetIdx = queue.findIndex(q => q.status === 'pending' && q.targetSessionId === sessionId);
-        if (targetIdx === -1) {
-            targetIdx = queue.findIndex(q => q.status === 'pending' && q.targetNodeId === nodeId && !q.targetSessionId);
-        }
-        if (targetIdx === -1) {
-            targetIdx = queue.findIndex(q => q.status === 'pending' && !q.targetNodeId && !q.targetSessionId);
+export function claimNextTask(
+    meshId: string,
+    nodeId: string,
+    sessionId: string,
+    capabilityTags?: string[],
+    opts?: { providerType?: string; providerMaxParallel?: number },
+): MeshWorkQueueEntry | null {
+    return MeshRuntimeStore.getInstance().claimNextQueueTask(meshId, nodeId, sessionId, capabilityTags, opts);
+}
+
+// ─── M1: Dependency Failure Propagation ─────────
+
+export type DependencyFailurePolicy = 'block' | 'cancel';
+
+function resolveDependencyFailurePolicy(meshId: string): DependencyFailurePolicy {
+    try {
+        const policy = (getMesh(meshId)?.policy ?? {}) as Record<string, unknown>;
+        return policy.onDependencyFailure === 'cancel' ? 'cancel' : 'block';
+    } catch {
+        return 'block';
+    }
+}
+
+/**
+ * Apply the mesh's onDependencyFailure policy to pending dependents of a task
+ * that just reached a failed/cancelled terminal state.
+ *
+ * - 'block' (default): dependents stay pending with blockedReason
+ *   "dependency_failed:<taskId>" so an operator can requeue/cancel them.
+ * - 'cancel': dependents are cancelled (cascading to their own dependents).
+ *
+ * Must be called inside the queue lock of the triggering transition.
+ */
+function propagateDependencyFailure(meshId: string, failedTaskId: string): void {
+    const policy = resolveDependencyFailurePolicy(meshId);
+    const store = MeshRuntimeStore.getInstance();
+    const frontier = [failedTaskId];
+    const seen = new Set<string>(frontier);
+    while (frontier.length > 0) {
+        const currentId = frontier.pop()!;
+        const dependents = store.getQueueEntries(meshId, ['pending'])
+            .filter(entry => Array.isArray(entry.dependsOn) && entry.dependsOn.includes(currentId));
+        for (const dependent of dependents) {
+            if (seen.has(dependent.id)) continue;
+            seen.add(dependent.id);
+            if (policy === 'cancel') {
+                dependent.status = 'cancelled';
+                dependent.cancelledAt = new Date().toISOString();
+                dependent.cancelReason = `dependency_failed:${currentId}`;
+                store.updateQueueEntry(dependent);
+                frontier.push(dependent.id); // cascade to transitive dependents
+            } else {
+                dependent.blockedReason = `dependency_failed:${currentId}`;
+                store.updateQueueEntry(dependent);
+            }
         }
-        if (targetIdx === -1) return null;
-        const entry = queue[targetIdx];
-        entry.status = 'assigned';
-        entry.assignedNodeId = nodeId;
-        entry.assignedSessionId = sessionId;
-        entry.dispatchTimestamp = new Date().toISOString();
-        entry.updatedAt = new Date().toISOString();
-        writeQueue(meshId, queue);
-        return entry;
-    });
+    }
 }
 
+const DEPENDENCY_FAILURE_TERMINALS = new Set<MeshTaskStatus>(['failed', 'cancelled']);
+
 /**
  * Update the status of a specific task.
  * Used when a session completes, fails, or stalls.
@@ -166,15 +669,16 @@ export function updateTaskStatus(
     meshId: string,
     taskId: string,
     status: MeshTaskStatus,
+    opts?: MeshQueueMutationOptions,
 ): MeshWorkQueueEntry | null {
+    requireMeshHostQueueOwner(opts);
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const idx = queue.findIndex(q => q.id === taskId);
-        if (idx === -1) return null;
-        queue[idx].status = status;
-        queue[idx].updatedAt = new Date().toISOString();
-        writeQueue(meshId, queue);
-        return queue[idx];
+        const entry = MeshRuntimeStore.getInstance().findQueueEntryById(meshId, taskId);
+        if (!entry) return null;
+        entry.status = status;
+        MeshRuntimeStore.getInstance().updateQueueEntry(entry);
+        if (DEPENDENCY_FAILURE_TERMINALS.has(status)) propagateDependencyFailure(meshId, taskId);
+        return entry;
     });
 }
 
@@ -184,14 +688,12 @@ export function recordTaskAutoLaunch(
     autoLaunch: Omit<NonNullable<MeshWorkQueueEntry['autoLaunch']>, 'updatedAt'>,
 ): MeshWorkQueueEntry | null {
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const idx = queue.findIndex(q => q.id === taskId);
-        if (idx === -1) return null;
+        const entry = MeshRuntimeStore.getInstance().findQueueEntryById(meshId, taskId);
+        if (!entry) return null;
         const now = new Date().toISOString();
-        queue[idx].autoLaunch = { ...autoLaunch, updatedAt: now };
-        queue[idx].updatedAt = now;
-        writeQueue(meshId, queue);
-        return queue[idx];
+        entry.autoLaunch = { ...autoLaunch, updatedAt: now };
+        MeshRuntimeStore.getInstance().updateQueueEntry(entry);
+        return entry;
     });
 }
 
@@ -201,19 +703,19 @@ export function recordTaskAutoLaunch(
 export function cancelTask(
     meshId: string,
     taskId: string,
-    opts?: { reason?: string },
+    opts?: { reason?: string } & MeshQueueMutationOptions,
 ): MeshWorkQueueEntry | null {
+    requireMeshHostQueueOwner(opts);
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const idx = queue.findIndex(q => q.id === taskId);
-        if (idx === -1) return null;
+        const entry = MeshRuntimeStore.getInstance().findQueueEntryById(meshId, taskId);
+        if (!entry) return null;
         const now = new Date().toISOString();
-        queue[idx].status = 'cancelled';
-        queue[idx].updatedAt = now;
-        queue[idx].cancelledAt = now;
-        if (opts?.reason) queue[idx].cancelReason = opts.reason;
-        writeQueue(meshId, queue);
-        return queue[idx];
+        entry.status = 'cancelled';
+        entry.cancelledAt = now;
+        if (opts?.reason) entry.cancelReason = opts.reason;
+        MeshRuntimeStore.getInstance().updateQueueEntry(entry);
+        propagateDependencyFailure(meshId, taskId);
+        return entry;
     });
 }
 
@@ -221,6 +723,11 @@ export function cancelTask(
  * Return a queue task to pending for retry. By default, dead session targeting
  * and assigned ownership are cleared so stale assignments do not strand again.
  */
+export type RequeueResult =
+    | { status: 'requeued'; entry: MeshWorkQueueEntry }
+    | { status: 'failed_max_retries'; entry: MeshWorkQueueEntry; maxRetries: number; requeueCount: number }
+    | { status: 'not_found' };
+
 export function requeueTask(
     meshId: string,
     taskId: string,
@@ -230,15 +737,34 @@ export function requeueTask(
         targetSessionId?: string;
         clearTargetNode?: boolean;
         clearTargetSession?: boolean;
-    },
+        /**
+         * Override the retry cap for this call. Use only for explicit operator actions.
+         * If true, the task is requeued even when requeueCount >= maxRetries.
+         */
+        force?: boolean;
+        /** Per-task retry cap override. Falls back to mesh policy maxTaskRetries (default 1). */
+        maxRetries?: number;
+    } & MeshQueueMutationOptions,
 ): MeshWorkQueueEntry | null {
+    requireMeshHostQueueOwner(opts);
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const idx = queue.findIndex(q => q.id === taskId);
-        if (idx === -1) return null;
-        const entry = queue[idx];
-        const now = new Date().toISOString();
+        const entry = MeshRuntimeStore.getInstance().findQueueEntryById(meshId, taskId);
+        if (!entry) return null;
+        const currentCount = entry.requeueCount || 0;
+        const maxRetries = opts?.maxRetries ?? entry.maxRetries ?? 1;
+        if (!opts?.force && currentCount >= maxRetries) {
+            // Auto-fail: cap exceeded without explicit force override.
+            entry.status = 'failed';
+            entry.cancelReason = `max_retries_exceeded: requeued ${currentCount} time(s), limit is ${maxRetries}`;
+            entry.updatedAt = new Date().toISOString();
+            MeshRuntimeStore.getInstance().updateQueueEntry(entry);
+            propagateDependencyFailure(meshId, taskId);
+            return entry;
+        }
         entry.status = 'pending';
+        // Operator requeue clears a dependency-failure block — the operator is
+        // explicitly overriding the held-back state.
+        delete entry.blockedReason;
         delete entry.assignedNodeId;
         delete entry.assignedSessionId;
         delete entry.cancelledAt;
@@ -247,11 +773,10 @@ export function requeueTask(
         if (typeof opts?.targetNodeId === 'string') entry.targetNodeId = opts.targetNodeId;
         if (opts?.clearTargetSession !== false) delete entry.targetSessionId;
         if (typeof opts?.targetSessionId === 'string') entry.targetSessionId = opts.targetSessionId;
-        entry.updatedAt = now;
-        entry.requeuedAt = now;
-        entry.requeueCount = (entry.requeueCount || 0) + 1;
+        entry.requeuedAt = new Date().toISOString();
+        entry.requeueCount = currentCount + 1;
         if (opts?.reason) entry.requeueReason = opts.reason;
-        writeQueue(meshId, queue);
+        MeshRuntimeStore.getInstance().updateQueueEntry(entry);
         return entry;
     });
 }
@@ -266,25 +791,42 @@ export function updateSessionTaskStatus(
     opts?: { occurredAt?: string },
 ): MeshWorkQueueEntry | null {
     return withQueueLock(meshId, () => {
-        const queue = readQueue(meshId);
-        const occurredAtTime = opts?.occurredAt ? new Date(opts.occurredAt).getTime() : Number.NaN;
-        const hasOccurredAt = Number.isFinite(occurredAtTime);
-        let bestIdx = -1;
-        let bestTime = 0;
-        for (let i = queue.length - 1; i >= 0; i--) {
-            if (queue[i].assignedSessionId !== sessionId || queue[i].status !== 'assigned') continue;
-            const time = new Date(queue[i].dispatchTimestamp || queue[i].updatedAt).getTime();
-            if (hasOccurredAt && Number.isFinite(time) && time > occurredAtTime) continue;
-            if (time > bestTime) { bestTime = time; bestIdx = i; }
-        }
-        if (bestIdx === -1) return null;
-        queue[bestIdx].status = status;
-        queue[bestIdx].updatedAt = new Date().toISOString();
-        writeQueue(meshId, queue);
-        return queue[bestIdx];
+        const occurredAtIso = opts?.occurredAt ? new Date(opts.occurredAt).toISOString() : undefined;
+        const entry = MeshRuntimeStore.getInstance().findAssignedBySession(meshId, sessionId, occurredAtIso);
+        if (!entry) return null;
+        entry.status = status;
+        MeshRuntimeStore.getInstance().updateQueueEntry(entry);
+        if (DEPENDENCY_FAILURE_TERMINALS.has(status)) propagateDependencyFailure(meshId, entry.id);
+        return entry;
     });
 }
 
+/**
+ * M1-3: true when at least one pending task is waiting on the given task.
+ * Used by the completion event path to decide whether to wake the queue.
+ */
+export function hasPendingDependents(meshId: string, taskId: string): boolean {
+    return MeshRuntimeStore.getInstance().getQueueEntries(meshId, ['pending'])
+        .some(entry => Array.isArray(entry.dependsOn) && entry.dependsOn.includes(taskId));
+}
+
+/**
+ * M1-4: view-time dependency state for a task — unmet dependency ids and
+ * whether the task is currently claimable from a dependency standpoint.
+ * Not stored (truth stays in task statuses).
+ */
+export function describeTaskDependencyState(
+    entry: Pick<MeshWorkQueueEntry, 'dependsOn' | 'blockedReason'>,
+    statusById: Map<string, MeshTaskStatus | string>,
+): { waitingOn: string[]; dependenciesSatisfied: boolean } {
+    const deps = Array.isArray(entry.dependsOn) ? entry.dependsOn : [];
+    const waitingOn = deps.filter(depId => statusById.get(depId) !== 'completed');
+    return {
+        waitingOn,
+        dependenciesSatisfied: waitingOn.length === 0 && !entry.blockedReason,
+    };
+}
+
 export interface MeshWorkQueueStats {
     total: number;
     active: number;
@@ -310,14 +852,16 @@ export interface MeshWorkQueueStats {
  * Return aggregate queue statistics for the given mesh.
  */
 export function getMeshQueueStats(meshId: string): MeshWorkQueueStats {
-    const queue = readQueue(meshId);
-    const pending = queue.filter(q => q.status === 'pending').length;
-    const assigned = queue.filter(q => q.status === 'assigned').length;
-    const completed = queue.filter(q => q.status === 'completed').length;
-    const failed = queue.filter(q => q.status === 'failed').length;
-    const cancelled = queue.filter(q => q.status === 'cancelled').length;
+    const rows = MeshRuntimeStore.getInstance().getQueueStatsByStatus(meshId);
+    const counts: Record<string, number> = {};
+    for (const r of rows) counts[r.status] = r.count;
+    const pending = counts['pending'] ?? 0;
+    const assigned = counts['assigned'] ?? 0;
+    const completed = counts['completed'] ?? 0;
+    const failed = counts['failed'] ?? 0;
+    const cancelled = counts['cancelled'] ?? 0;
     return {
-        total: queue.length,
+        total: pending + assigned + completed + failed + cancelled,
         active: pending + assigned,
         historical: completed + failed + cancelled,
         pending,
@@ -325,22 +869,119 @@ export function getMeshQueueStats(meshId: string): MeshWorkQueueStats {
         completed,
         failed,
         cancelled,
-        activeCounts: {
-            pending,
-            assigned,
-        },
-        historicalCounts: {
-            completed,
-            failed,
-            cancelled,
-        },
-        activeAssignments: queue
-            .filter(q => q.status === 'assigned')
-            .map(q => ({
-                id: q.id,
-                nodeId: q.assignedNodeId,
-                sessionId: q.assignedSessionId,
-                message: q.message,
-            })),
+        activeCounts: { pending, assigned },
+        historicalCounts: { completed, failed, cancelled },
+        activeAssignments: MeshRuntimeStore.getInstance().getActiveAssignmentDetails(meshId),
     };
 }
+
+export function __replaceMeshQueueForTests(meshId: string, queue: MeshWorkQueueEntry[]): void {
+    MeshRuntimeStore.getInstance().transaction(() => {
+        MeshRuntimeStore.getInstance().replaceQueue(meshId, queue);
+    });
+}
+
+export function __clearMeshQueueForTests(meshId: string): void {
+    MeshRuntimeStore.getInstance().deleteQueue(meshId);
+}
+
+export function __clearDirectDispatchesForTests(meshId: string): void {
+    MeshRuntimeStore.getInstance().deleteDirectDispatches(meshId);
+}
+
+export function __resetMeshRuntimeStoreForTests(): void {
+    MeshRuntimeStore.resetForTests();
+}
+
+// ── Direct Dispatch Tracking ─────────────────────────────────────────────────
+// Persists direct (non-queue) task dispatches so buildMeshActiveWork can read
+// active work from MeshRuntimeStore instead of scanning ledger JSONL entries.
+
+export type DirectDispatchRecord = ReturnType<MeshRuntimeStore['getActiveDirectDispatches']>[number];
+
+export function insertDirectDispatch(
+    meshId: string,
+    data: {
+        taskId: string;
+        nodeId?: string;
+        sessionId?: string;
+        providerType?: string;
+        message: string;
+        taskMode?: string;
+        via: string;
+        dispatchedToIdleSession?: boolean;
+        dispatchedAt: string;
+    },
+): void {
+    try {
+        MeshRuntimeStore.getInstance().insertDirectDispatch({ ...data, meshId });
+    } catch (e: any) {
+        process.stderr.write(`[adhdev-mesh] insertDirectDispatch failed for task ${data.taskId}: ${e?.message || e}\n`);
+    }
+}
+
+export function getActiveDirectDispatches(meshId: string): DirectDispatchRecord[] {
+    try {
+        return MeshRuntimeStore.getInstance().getActiveDirectDispatches(meshId);
+    } catch {
+        return [];
+    }
+}
+
+export function updateDirectDispatchStatus(
+    meshId: string,
+    sessionId: string,
+    status: 'acked' | 'completed' | 'failed' | 'stale',
+): void {
+    try {
+        MeshRuntimeStore.getInstance().updateDirectDispatchStatus(meshId, sessionId, status);
+    } catch { /* best-effort */ }
+}
+
+export function cleanupTerminalDirectDispatches(olderThanMs = 7 * 24 * 60 * 60_000): void {
+    try {
+        MeshRuntimeStore.getInstance().cleanupTerminalDirectDispatches(olderThanMs);
+    } catch { /* best-effort */ }
+}
+
+export function markStaleDirectDispatches(meshId: string, olderThanMs = 60 * 60_000): void {
+    try {
+        MeshRuntimeStore.getInstance().markStaleDirectDispatches(meshId, olderThanMs);
+    } catch { /* best-effort */ }
+}
+
+/**
+ * Delete specific direct dispatch rows by taskId. Returns the number of rows deleted.
+ * Used by the staleDirect prune path to evict orphaned/terminal dispatch records from the
+ * active staleDirect surface while leaving the append-only mesh ledger (audit history) intact.
+ */
+export function deleteDirectDispatchesByTaskId(meshId: string, taskIds: string[]): number {
+    try {
+        return MeshRuntimeStore.getInstance().deleteDirectDispatchesByTaskId(meshId, taskIds);
+    } catch {
+        return 0;
+    }
+}
+
+export type MeshToolCallRateResult = { rateLimitExceeded: boolean; callsInWindow: number; advisory: string | null };
+
+/**
+ * Record a coordinator tool call and return a rate-limit advisory when the
+ * call rate for that tool exceeds the allowed threshold.
+ *
+ * Defaults: 10-second sliding window, max 5 calls before advisory is raised.
+ * Returns { rateLimitExceeded: false } on any store error so callers are not blocked.
+ */
+export function recordMeshToolCall(opts: {
+    meshId: string;
+    tool: string;
+    sessionId?: string | null;
+    windowMs?: number;
+    maxCalls?: number;
+}): MeshToolCallRateResult {
+    try {
+        return MeshRuntimeStore.getInstance().recordMeshToolCall(opts);
+    } catch {
+        return { rateLimitExceeded: false, callsInWindow: 0, advisory: null };
+    }
+}