@adhdev/daemon-core 0.9.82-rc.41 → 0.9.82-rc.411

package/src/repo-mesh-types.ts

@@ -12,6 +12,7 @@
  */
 
 import type { GitRepoStatus, GitCompactSummary } from './git/git-types.js';
+import type { MeshMissionSummary, MeshMissionSlimSummary } from './mesh/mesh-missions.js';
 
 // ─── Core Mesh Types ────────────────────────────
 
@@ -23,11 +24,42 @@ export interface RepoMesh {
     defaultBranch?: string;
     policy: RepoMeshPolicy;
     coordinator: RepoMeshCoordinatorConfig;
+    meshHost?: RepoMeshHostMetadata;
     projectContext: ProjectContextSnapshot;
     nodes: RepoMeshNode[];
     status: 'active' | 'archived' | 'deleted';
 }
 
+export type RepoMeshDaemonRole = 'host' | 'member';
+
+export interface RepoMeshHostPairingMetadata {
+    status: 'not_configured' | 'pairing' | 'paired' | 'rejected' | 'revoked';
+    tokenId?: string;
+    joinedAt?: string;
+    lastPairedAt?: string;
+    lastRejectedAt?: string;
+    expiresAt?: string;
+}
+
+export interface RepoMeshHostMetadata {
+    /** Local daemon role for this mesh. Missing metadata defaults to host for standalone compatibility. */
+    role: RepoMeshDaemonRole;
+    /** Daemon that owns mesh truth/status/git/queue/session/ledger/coordinator ownership. */
+    hostDaemonId?: string;
+    /** Mesh node that represents the host daemon, when known. */
+    hostNodeId?: string;
+    /** Future standalone manual pairing endpoint entered by member daemons. */
+    hostAddress?: string;
+    /** Redacted pairing state only; raw join tokens must not be persisted here. */
+    pairing?: RepoMeshHostPairingMetadata;
+}
+
+export interface RepoMeshHostStatus extends RepoMeshHostMetadata {
+    canOwnCoordinator: boolean;
+    canOwnQueue: boolean;
+    defaulted: boolean;
+}
+
 export interface RepoMeshNode {
     id: string;
     daemonId: string;
@@ -42,6 +74,7 @@ export interface RepoMeshNode {
     effectiveCapabilities: RepoMeshNodeCapabilities;
     policy: RepoMeshNodePolicy;
     health: RepoMeshNodeHealth;
+    role?: RepoMeshDaemonRole;
     status: 'enabled' | 'disabled' | 'removed';
 }
 
@@ -58,26 +91,228 @@ export type RepoMeshNodeHealth =
 export type RepoMeshSessionCleanupMode = 'preserve' | 'stop' | 'delete_stopped' | 'stop_and_delete';
 export type RepoMeshSpawnedSessionVisibility = 'visible' | 'hidden';
 
+/**
+ * Mesh-wide tie-break strategy for distributing untargeted queue work across
+ * eligible nodes. This ONLY governs the final tie-break stage of the scheduler
+ * pipeline (TAG hard-filter → MAX-ALLOC capacity gate → PRIORITY soft score →
+ * TIE-BREAK); eligibility/capacity/priority are evaluated identically for every
+ * strategy.
+ *
+ * - 'first_eligible' (DEFAULT): preserve today's behavior exactly. Nodes are
+ *   visited in config/array order and the first that can launch wins. No
+ *   load-spreading. This is the strict no-change default — a mesh that never
+ *   sets schedulingStrategy behaves identically to before this feature.
+ * - 'least_loaded': prefer the eligible node with the fewest active assignments,
+ *   so untargeted work spreads instead of piling onto whichever node asks first.
+ * - 'round_robin': among nodes tied at the least load, rotate the winner using a
+ *   per-mesh cursor so distribution stays fair across passes.
+ * - 'priority_only': rank purely by schedulingPriority (then config order),
+ *   ignoring load — always send to the highest-priority eligible node.
+ *
+ * Distribution is explicit opt-in: a strategy other than 'first_eligible' must be
+ * configured for any load-spreading to occur.
+ */
+export type RepoMeshSchedulingStrategy =
+    | 'first_eligible'
+    | 'least_loaded'
+    | 'round_robin'
+    | 'priority_only';
+
+export const MESH_SCHEDULING_STRATEGIES: RepoMeshSchedulingStrategy[] = [
+    'first_eligible',
+    'least_loaded',
+    'round_robin',
+    'priority_only',
+];
+
+export const DEFAULT_MESH_SCHEDULING_STRATEGY: RepoMeshSchedulingStrategy = 'first_eligible';
+
+/**
+ * Normalize an unknown scheduling-strategy value to a valid strategy, defaulting
+ * to 'first_eligible' (strict no-change) for anything missing/blank/unrecognized.
+ */
+export function normalizeMeshSchedulingStrategy(value: unknown): RepoMeshSchedulingStrategy {
+    if (typeof value !== 'string') return DEFAULT_MESH_SCHEDULING_STRATEGY;
+    const trimmed = value.trim() as RepoMeshSchedulingStrategy;
+    return (MESH_SCHEDULING_STRATEGIES as string[]).includes(trimmed)
+        ? trimmed
+        : DEFAULT_MESH_SCHEDULING_STRATEGY;
+}
+
+// ─── User-facing distribution mode (2-mode façade over the 4-union) ──────────
+//
+// The raw RepoMeshSchedulingStrategy 4-union stays the persisted, escape-hatch
+// truth (a hand-edited meshes.json / .adhdev/mesh.json may write any of the four
+// and it is honored verbatim). The PRODUCT surface is just two modes:
+//
+//   - 'spread'   → distribute work evenly. Maps to 'least_loaded', which now also
+//                  absorbs the round_robin rotation as its internal tie-break (see
+//                  orderEligibleNodes), so a single mode covers both former
+//                  load-spreading strategies. schedulingPriority still acts as the
+//                  primary rank key automatically when set — no separate mode.
+//   - 'in_order' → assign in the order nodes were added. Maps to 'first_eligible'
+//                  (the strict no-change default).
+//
+// distributionToStrategy / strategyToDistribution are the canonical bridge the UI
+// and the .adhdev/mesh.json loader share so the 2-mode↔4-union mapping is defined
+// exactly once.
+export type RepoMeshDistribution = 'spread' | 'in_order';
+
+export const MESH_DISTRIBUTIONS: RepoMeshDistribution[] = ['spread', 'in_order'];
+
+/**
+ * Product default for the user-facing toggle. Note this is the *recommended* mode
+ * shown first, NOT the persisted default: a mesh that never sets schedulingStrategy
+ * stays 'first_eligible' (→ displays as 'in_order'), preserving strict no-change.
+ */
+export const DEFAULT_MESH_DISTRIBUTION: RepoMeshDistribution = 'spread';
+
+/** Normalize an unknown value to a valid distribution mode (defaults to 'spread'). */
+export function normalizeMeshDistribution(value: unknown): RepoMeshDistribution {
+    if (typeof value !== 'string') return DEFAULT_MESH_DISTRIBUTION;
+    const trimmed = value.trim();
+    return (MESH_DISTRIBUTIONS as string[]).includes(trimmed)
+        ? (trimmed as RepoMeshDistribution)
+        : DEFAULT_MESH_DISTRIBUTION;
+}
+
+/**
+ * Map a user-facing distribution mode to the raw scheduling strategy the scheduler
+ * acts on. 'spread' → 'least_loaded' (which absorbs round_robin rotation as its
+ * tie-break); 'in_order' → 'first_eligible'.
+ */
+export function distributionToStrategy(distribution: RepoMeshDistribution): RepoMeshSchedulingStrategy {
+    return distribution === 'spread' ? 'least_loaded' : 'first_eligible';
+}
+
+/**
+ * Map a raw scheduling strategy back to the 2-mode façade for display/migration.
+ *   - 'first_eligible'           → 'in_order'
+ *   - 'least_loaded'/'round_robin' → 'spread'
+ *   - 'priority_only'            → 'spread' when a node priority is actually
+ *     configured (priority ordering IS a spread behavior), else 'in_order'
+ *     (priority_only with no priorities is behaviorally identical to
+ *     first_eligible, so it must not flip a mesh into load-spreading on migration).
+ */
+export function strategyToDistribution(
+    strategy: unknown,
+    opts?: { priorityConfigured?: boolean },
+): RepoMeshDistribution {
+    const normalized = normalizeMeshSchedulingStrategy(strategy);
+    if (normalized === 'first_eligible') return 'in_order';
+    if (normalized === 'priority_only') return opts?.priorityConfigured ? 'spread' : 'in_order';
+    return 'spread';
+}
+
+/**
+ * Resolve a node's soft scheduling priority — a single scalar used as the PRIORITY
+ * stage rank key (higher = preferred). It is NOT an eligibility gate (the MAX-ALLOC
+ * capacity gate alone decides whether a node can take work). Missing/blank/NaN
+ * resolves to 0 so unconfigured nodes all share the same neutral priority.
+ */
+export function resolveNodeSchedulingPriority(
+    nodePolicy: Pick<RepoMeshNodePolicy, 'schedulingPriority'> | null | undefined,
+): number {
+    const raw = Number(nodePolicy?.schedulingPriority);
+    return Number.isFinite(raw) ? raw : 0;
+}
+
+/**
+ * Synthetic capability tag advertised by every mesh node describing how it can land
+ * its work onto the base branch:
+ *   - converge=refine: a local worktree node (on any machine — refine_mesh_node
+ *     forwards to the owning daemon) can run the Refinery merge → push → cleanup.
+ *   - converge=fast_forward: a non-worktree node (the machine itself) can only
+ *     fast-forward/push an already-converged branch.
+ * Emitted by buildMeshNodeCapabilityTags and matched through the ordinary
+ * required-tags filter.
+ */
+export const MESH_CONVERGE_REFINE_TAG = 'converge=refine';
+export const MESH_CONVERGE_FAST_FORWARD_TAG = 'converge=fast_forward';
+
+/**
+ * Resolve whether the load-balancing scheduler should auto-inject a
+ * `converge=refine` required tag onto code_change tasks so they hard-filter onto
+ * refine-capable (worktree) nodes only. Strict opt-in: defaults to false, so a mesh
+ * that does not set it behaves exactly as before (code_change routes to any eligible
+ * node, including a non-worktree machine node when no worktree exists).
+ */
+export function resolveAutoConvergeCodeChange(
+    policy: Pick<RepoMeshPolicy, 'autoConvergeCodeChange'> | null | undefined,
+): boolean {
+    return policy?.autoConvergeCodeChange === true;
+}
+
+export interface RepoMeshAutoFastForwardPolicy {
+    /** Defaults to true. Set false to disable daemon-initiated idle fast-forwards. */
+    enabled: boolean;
+    /** Maximum behind count eligible for automatic fast-forward. Missing means no limit. */
+    maxBehind?: number;
+    /** Defaults to true. Require submodule status to be clean before automatic fast-forward. */
+    requireCleanSubmodules?: boolean;
+}
+
 export interface RepoMeshPolicy {
     requirePreTaskCheckpoint: boolean;
     requirePostTaskCheckpoint: boolean;
     requireApprovalForPush: boolean;
+    /**
+     * Narrow Refinery opt-in: when validation and patch-equivalence have passed,
+     * allow Refinery to publish submodule gitlink commits to each submodule's
+     * configured remote main branch with a non-force push, then verify reachability.
+     * Defaults to false; root branch pushes/merges are not affected.
+     */
+    allowAutoPublishSubmoduleMainCommits?: boolean;
     requireApprovalForDestructiveGit: boolean;
     dirtyWorkspaceBehavior: 'block' | 'warn' | 'checkpoint_then_continue';
     maxParallelTasks: number;
     allowedProviders?: string[];
+    /**
+     * Mesh-wide tie-break strategy for distributing untargeted queue work across
+     * eligible nodes. Defaults to 'first_eligible' (today's exact behavior — no
+     * load-spreading). Set to 'least_loaded' / 'round_robin' / 'priority_only' to
+     * opt into distribution. Only governs the final tie-break stage; eligibility,
+     * capacity, and priority are evaluated identically regardless of strategy.
+     */
+    schedulingStrategy?: RepoMeshSchedulingStrategy;
+    /**
+     * Convergence routing opt-in: when true, the scheduler auto-injects a
+     * `converge=refine` required tag onto every code_change task at enqueue time, so
+     * code_change work hard-filters onto refine-capable worktree nodes (on any
+     * machine — refine_mesh_node forwards to the owning daemon) and never lands on a
+     * non-worktree machine node. Explicit target_node_id routing and any
+     * caller-supplied required_tags are preserved (the tag is merged, not replaced).
+     * Defaults to false: code_change routing is unchanged unless opted in.
+     */
+    autoConvergeCodeChange?: boolean;
     /**
      * Whether sessions spawned by mesh/coordinator policy should auto-open as visible
      * dashboard tabs or start hidden. Defaults to 'visible' to preserve existing
      * watch-the-agents behavior; hidden sessions remain discoverable and manually openable.
      */
     spawnedSessionVisibility?: RepoMeshSpawnedSessionVisibility;
+    /**
+     * Whether worker sessions the coordinator dispatches should auto-approve agent
+     * approval modals (tool/command prompts) without firing a user-facing approval
+     * notification. Delegated workers are coordinator-driven, so a human should not
+     * have to approve each one; defaults to true. Set to false to make delegated
+     * worker sessions stop at approval modals like an interactive session.
+     * Stamped into the worker launch settings envelope as `autoApprove`, which wins
+     * over the global per-provider-type autoApprove config via the settings merge.
+     * A node policy may override this per-node (RepoMeshNodePolicy.delegatedWorkerAutoApprove).
+     */
+    delegatedWorkerAutoApprove?: boolean;
     /**
      * What to do with delegated session-host records for a node when it is removed.
      * Defaults to 'preserve' so completed work can be reviewed later and live
      * runtimes are never stopped/deleted unless the mesh owner opts in.
      */
     sessionCleanupOnNodeRemove?: RepoMeshSessionCleanupMode;
+    /**
+     * Daemon-initiated fast-forward for idle clean nodes that are only behind
+     * their tracked upstream. Defaults to enabled.
+     */
+    autoFastForward?: RepoMeshAutoFastForwardPolicy;
     /**
      * Maximum number of automatic retry recommendations for a failed task on the
      * same node before the daemon advises the coordinator to escalate or reassign.
@@ -93,12 +328,54 @@ export interface RepoMeshRelatedRepo {
     workspace: string;
 }
 
+/**
+ * Per-(node, provider) parallelism declaration.
+ *
+ * `maxParallel` is the only enforced field: the queue will not assign a task
+ * to this (node, provider) once it already has `maxParallel` active
+ * (status='assigned') tasks. When the global parallel cap and this per-(node,
+ * provider) cap disagree, the stricter (lower effective) limit wins — a claim
+ * must satisfy both. Omitting `maxParallel` means this provider is bounded only
+ * by the global/taskMode caps (full backward compatibility).
+ *
+ * Routing is governed exclusively by required_tags (see nodeSatisfiesRequiredTags);
+ * this entry carries no routing role. To route work to a specific node, advertise an
+ * ordinary capability tag on the node and require it on the task.
+ */
+export interface RepoMeshProviderRole {
+    /** Provider type this entry governs (e.g. 'claude-cli', 'codex-cli'). */
+    providerType: string;
+    /** Max concurrent active tasks for this (node, provider). Omit = no per-provider cap. */
+    maxParallel?: number;
+}
+
 export interface RepoMeshNodePolicy {
     readOnly?: boolean;
     canPush?: boolean;
     maxConcurrentSessions?: number;
+    /**
+     * Soft scheduling priority used as the PRIORITY rank key (higher = preferred)
+     * when the mesh schedulingStrategy spreads work across nodes. Defaults to 0.
+     * This is NOT an eligibility gate — a node with a high priority that is at its
+     * capacity (MAX-ALLOC gate) is still skipped; priority only orders nodes that
+     * can actually take work. Ignored entirely under 'first_eligible'.
+     */
+    schedulingPriority?: number;
     /** Ordered provider preference used when mesh_launch_session omits an explicit type. */
     providerPriority?: string[];
+    /**
+     * Per-(node, provider) parallelism declarations. Each entry binds a
+     * providerType on THIS node to an optional maxParallel cap. maxParallel is
+     * enforced as an additional, stricter-wins constraint on top of the global
+     * maxParallelTasks/taskMode caps. Missing/empty: the node behaves exactly as
+     * before (global caps only). Routing is governed solely by required_tags.
+     */
+    providerRoles?: RepoMeshProviderRole[];
+    /**
+     * Per-node override for RepoMeshPolicy.delegatedWorkerAutoApprove. When set, takes
+     * precedence over the mesh-level policy for worker sessions launched onto this node.
+     */
+    delegatedWorkerAutoApprove?: boolean;
     /**
      * Optional associated/external repos that must be checked alongside this node.
      * These are explicit policy/config entries only; Repo Mesh does not auto-discover
@@ -126,18 +403,206 @@ export const DEFAULT_MESH_POLICY: RepoMeshPolicy = {
     requirePreTaskCheckpoint: false,
     requirePostTaskCheckpoint: true,
     requireApprovalForPush: true,
+    allowAutoPublishSubmoduleMainCommits: false,
     requireApprovalForDestructiveGit: true,
     dirtyWorkspaceBehavior: 'warn',
     maxParallelTasks: 2,
-    spawnedSessionVisibility: 'visible',
+    // Coordinator-spawned worker sessions default to hidden so the dashboard is not
+    // flooded with mesh noise tabs/notifications. Users can still surface or unmute
+    // any specific session manually; that override is preserved per-device.
+    spawnedSessionVisibility: 'hidden',
+    delegatedWorkerAutoApprove: true,
     sessionCleanupOnNodeRemove: 'preserve',
+    autoFastForward: { enabled: true },
     maxTaskRetries: 1,
 };
 
+// ─── Policy normalization (single source of truth) ──────────────────────────
+//
+// Every mesh policy passes through mergeAndNormalizePolicy exactly once on write
+// (createMesh/updateMesh) and again whenever a policy is materialized for display
+// or scheduling. Co-locating the default constant, the per-field normalizers, and
+// the merge here keeps the three former layers (DEFAULT_MESH_POLICY, the merge in
+// mesh-config, and the scattered field clamps) from drifting apart. The function
+// is idempotent: feeding it an already-normalized policy yields the same object.
+
+const SESSION_CLEANUP_MODES = new Set<RepoMeshSessionCleanupMode>([
+    'preserve', 'stop', 'delete_stopped', 'stop_and_delete',
+]);
+const SPAWNED_SESSION_VISIBILITY_MODES = new Set<RepoMeshSpawnedSessionVisibility>([
+    'visible', 'hidden',
+]);
+const DIRTY_WORKSPACE_BEHAVIORS = new Set<RepoMeshPolicy['dirtyWorkspaceBehavior']>([
+    'block', 'warn', 'checkpoint_then_continue',
+]);
+
+/** Min/max bounds for the global write-task parallel cap. */
+export const MESH_MAX_PARALLEL_TASKS_MIN = 1;
+export const MESH_MAX_PARALLEL_TASKS_MAX = 8;
+
+/**
+ * Default multiplier applied to the write cap to derive the read-only diagnosis
+ * cap. Read-only (live_debug_readonly) tasks carry no isolation/merge cost so they
+ * run under a separate, looser cap; a missing/invalid readonlyMultiplier resolves
+ * to this value, preserving the historical `max(2, write × 2)` behavior.
+ */
+export const DEFAULT_MESH_READONLY_MULTIPLIER = 2;
+
+/**
+ * SINGLE source of truth for the read-only diagnosis cap derivation. Both the live
+ * claim path (maybeAutoLaunchOneQueueSession) and the observability projection
+ * (buildMeshSchedulingRuntime) read it through here so the exposed cap and the
+ * enforced cap can never drift. Floors at 2 so a write cap of 1 still allows two
+ * concurrent read-only diagnoses. An out-of-range multiplier falls back to the
+ * default (2) — identical to the previous inline `Math.max(2, maxParallel * 2)`.
+ */
+export function resolveMaxReadonlyParallelTasks(maxParallelTasks: number, multiplier?: unknown): number {
+    const raw = Number(multiplier);
+    const mult = Number.isFinite(raw) && raw >= 1 ? Math.floor(raw) : DEFAULT_MESH_READONLY_MULTIPLIER;
+    return Math.max(2, Math.floor(maxParallelTasks) * mult);
+}
+
+/**
+ * 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 function resolveMaxParallelTasks(value: unknown): number {
+    const n = Number(value);
+    if (!Number.isFinite(n)) return DEFAULT_MESH_POLICY.maxParallelTasks;
+    return Math.max(MESH_MAX_PARALLEL_TASKS_MIN, Math.min(MESH_MAX_PARALLEL_TASKS_MAX, Math.floor(n)));
+}
+
+/**
+ * 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 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,
+    };
+}
+
+/**
+ * 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 function mergeAndNormalizePolicy(
+    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 (!DIRTY_WORKSPACE_BEHAVIORS.has(policy.dirtyWorkspaceBehavior)) {
+        policy.dirtyWorkspaceBehavior = 'warn';
+    }
+    policy.maxParallelTasks = resolveMaxParallelTasks(policy.maxParallelTasks);
+    policy.allowAutoPublishSubmoduleMainCommits = policy.allowAutoPublishSubmoduleMainCommits === true;
+    if (!SESSION_CLEANUP_MODES.has(policy.sessionCleanupOnNodeRemove as RepoMeshSessionCleanupMode)) {
+        policy.sessionCleanupOnNodeRemove = 'preserve';
+    }
+    if (!SPAWNED_SESSION_VISIBILITY_MODES.has(policy.spawnedSessionVisibility as RepoMeshSpawnedSessionVisibility)) {
+        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;
+}
+
+/**
+ * Resolve whether a delegated worker session launched onto `nodePolicy` (within a mesh
+ * governed by `meshPolicy`) should auto-approve. Precedence: node override → mesh policy
+ * → default true. The result is stamped into the worker launch settings envelope as
+ * `autoApprove`; it wins over the global per-provider-type autoApprove config because the
+ * launch path merges the envelope as a settingsOverride on top of the provider defaults.
+ */
+export function resolveDelegatedWorkerAutoApprove(
+    meshPolicy?: Pick<RepoMeshPolicy, 'delegatedWorkerAutoApprove'> | null,
+    nodePolicy?: Pick<RepoMeshNodePolicy, 'delegatedWorkerAutoApprove'> | null,
+): boolean {
+    if (typeof nodePolicy?.delegatedWorkerAutoApprove === 'boolean') {
+        return nodePolicy.delegatedWorkerAutoApprove;
+    }
+    if (typeof meshPolicy?.delegatedWorkerAutoApprove === 'boolean') {
+        return meshPolicy.delegatedWorkerAutoApprove;
+    }
+    return true;
+}
+
+/**
+ * Resolve the enforced per-(node, provider) maxParallel cap, or undefined when
+ * no finite, non-negative cap is declared for this provider. Used by the queue
+ * claim path as a stricter-wins constraint layered on top of the global caps.
+ * Case-insensitive, trimmed match on providerType. Defensive against malformed
+ * config — non-object entries and blank providerTypes are skipped rather than throwing.
+ */
+export function resolveProviderMaxParallel(
+    nodePolicy: Pick<RepoMeshNodePolicy, 'providerRoles'> | null | undefined,
+    providerType: string | null | undefined,
+): number | undefined {
+    const wanted = typeof providerType === 'string' ? providerType.trim().toLowerCase() : '';
+    if (!wanted) return undefined;
+    const roles = nodePolicy?.providerRoles;
+    if (!Array.isArray(roles)) return undefined;
+    for (const entry of roles) {
+        if (!entry || typeof entry !== 'object') continue;
+        const type = typeof entry.providerType === 'string' ? entry.providerType.trim().toLowerCase() : '';
+        if (!type || type !== wanted) continue;
+        const raw = Number(entry.maxParallel);
+        if (!Number.isFinite(raw) || raw < 0) return undefined;
+        return Math.floor(raw);
+    }
+    return undefined;
+}
+
 // ─── Capabilities ───────────────────────────────
 
 export interface RepoMeshNodeCapabilities {
+    /** Node's OS, raw NodeJS.Platform value ("darwin"/"win32"/"linux"). For
+     *  remote member nodes this is stamped by the member daemon at join time
+     *  (its own process.platform) and drives os= capability-tag routing. */
     platform?: string;
+    /** Node's CPU architecture, raw process.arch value ("arm64"/"x64"). Stamped
+     *  by the member daemon at join time; drives arch= capability-tag routing. */
+    arch?: string;
     packageManagers?: string[];
     detectedCommands?: DetectedCommand[];
     canRunLongJobs?: boolean;
@@ -207,7 +672,28 @@ export interface RepoMeshCoordinatorConfig {
     providerType?: string;
     /** Preferred node to run coordinator on (null = auto) */
     preferredNodeId?: string;
-    /** Additional system prompt context for coordinator */
+    /**
+     * Full mesh-level override for the coordinator system prompt. When set,
+     * replaces the daemon's rendered default and any user-file override
+     * (~/.adhdev/coordinator-prompts/<cli>.md). The per-launch
+     * extraSystemPrompt still composes on top — it always lands last as
+     * Additional Context. Supports the same {{placeholders}} the daemon's
+     * default template uses ({{meshName}}, {{repo}}, {{nodes}}, …).
+     */
+    systemPromptOverride?: string;
+    /**
+     * Mesh-level append. Composes after whichever base prompt won
+     * (override → user-file override → daemon default). Use this when you
+     * want extra rules for THIS mesh but otherwise the standard prompt is
+     * fine. Stacks with the user-file append (`<cli>.append.md`) — both
+     * apply if both are set.
+     */
+    systemPromptAppend?: string;
+    /**
+     * @deprecated Use systemPromptAppend. Kept as a fallback alias so
+     * existing meshes.json files keep working without a migration step;
+     * the daemon prefers systemPromptAppend when both are present.
+     */
     systemPromptSuffix?: string;
 }
 
@@ -229,6 +715,7 @@ export interface LocalMeshEntry {
     defaultBranch?: string;
     policy: RepoMeshPolicy;
     coordinator: RepoMeshCoordinatorConfig;
+    meshHost?: RepoMeshHostMetadata;
     nodes: LocalMeshNodeEntry[];
     createdAt: string;
     updatedAt: string;
@@ -241,43 +728,131 @@ export interface LocalMeshNodeEntry {
     daemonId?: string;
     /** Machine registry ID that owns this workspace, when known. */
     machineId?: string;
+    /** Operator-defined capability tags used by mesh queue matching. */
+    capabilities?: string[];
     userOverrides: Partial<RepoMeshNodeCapabilities>;
+    /**
+     * Live, self-healed platform/arch reported by the daemon that owns this
+     * node's workspace (its own process.platform/process.arch), carried on the
+     * git_status envelope and persisted by the coordinator on each direct git
+     * probe. This is auto-detected truth, kept DISTINCT from `userOverrides`
+     * (operator intent) so capability-tag derivation can prefer an explicit
+     * operator override while still self-correcting auto-detected nodes — and so
+     * a stale value is overwritten by the next report rather than sticking.
+     * Absent until the first direct probe succeeds.
+     */
+    reportedPlatform?: string;
+    reportedArch?: string;
     policy: RepoMeshNodePolicy;
+    /**
+     * Per-node instruction surfaced in the coordinator prompt so the LLM
+     * knows what each node is for (e.g. "this is the staging mirror — run
+     * only smoke tests here", or "use opus on this node, sonnet elsewhere").
+     * Empty/missing: omitted silently from the rendered prompt, no rule
+     * line about it gets added. The coordinator forwards/honors it when
+     * delegating; we don't enforce it at the daemon level.
+     */
+    systemPrompt?: string;
     /** For single-machine mesh: same daemon, different worktree */
     isLocalWorktree?: boolean;
     /** Branch this worktree tracks (set when created via clone_mesh_node) */
     worktreeBranch?: string;
     /** Node ID this worktree was cloned from */
     clonedFromNodeId?: string;
+    /** Repo-local preparation result for ADHDev-created worktree nodes. */
+    worktreeBootstrap?: {
+        status: 'ready' | 'running' | 'failed' | 'not_configured' | 'disabled' | 'stale';
+        required?: boolean;
+        configSource?: string;
+        configSourceType?: string;
+        startedAt?: string;
+        completedAt?: string;
+        lastCommand?: string;
+        exitCode?: number | null;
+        error?: string;
+        commandsRun?: Array<Record<string, unknown>>;
+        staleInputs?: string[];
+    };
     /** Optional associated/external repos configured as node metadata. */
     relatedRepos?: RepoMeshRelatedRepo[];
+    role?: RepoMeshDaemonRole;
 }
 
 // ─── Mesh Status (runtime, not persisted) ───────
 
+/**
+ * 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;
     repoIdentity: string;
     defaultBranch?: string;
     refreshedAt: string;
+    meshHost?: RepoMeshHostStatus;
     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
+     * daemons — the dashboard must treat this as optional and render an empty
+     * state when absent. Split on each entry's `status` for live vs. history.
+     *
+     * Compact (the default) status calls send the slim shape — `goalPreview` +
+     * `goalTruncated` instead of the full `goal` — while verbose sends the full
+     * `goal`. Consumers must read `goal ?? goalPreview`. Each entry may also carry
+     * an optional `stats` operational rollup (durations / retries).
+     */
+    missions?: (MeshMissionSummary | MeshMissionSlimSummary)[];
 }
 
-export interface RepoMeshSessionStatus {
-    sessionId: string;
-    providerType?: string;
-    state?: string;
-    lifecycle?: 'starting' | 'running' | 'stopping' | 'stopped' | 'failed' | 'interrupted';
-    surfaceKind?: 'live_runtime' | 'recovery_snapshot' | 'inactive_record';
-    recoveryState?: string | null;
-    workspace?: string | null;
-    title?: string | null;
-    lastActivityAt?: string | null;
-    isCached?: boolean;
-}
+// RepoMeshSessionStatus shape now lives in @adhdev/mesh-shared (shared with
+// web-core's session normalizer). Re-exported so local RepoMeshNodeStatus/
+// RepoMeshStatus and external `@adhdev/daemon-core/repo-mesh-types` consumers
+// keep resolving it unchanged.
+export type { RepoMeshSessionStatus } from '@adhdev/mesh-shared';
+import type { RepoMeshSessionStatus } from '@adhdev/mesh-shared';
 
 export type RepoMeshPeerConnectionState = 'self' | 'connected' | 'connecting' | 'disconnected' | 'failed' | 'closed' | 'unknown';
 export type RepoMeshPeerConnectionTransport = 'local' | 'direct' | 'relay' | 'unknown';
@@ -289,6 +864,12 @@ export interface RepoMeshPeerConnectionStatus {
     transport: RepoMeshPeerConnectionTransport;
     reported: boolean;
     reason?: string;
+    /**
+     * Round-trip time in ms for the selected candidate pair, as sampled by the
+     * coordinator daemon when connected. Optional — older daemons and not_reported
+     * fallbacks omit it; the dashboard must treat it as best-effort telemetry.
+     */
+    rttMs?: number;
     lastStateChangeAt?: string;
     lastConnectedAt?: string;
     lastCommandAt?: string;
@@ -301,9 +882,12 @@ export interface RepoMeshNodeStatus {
     repoRoot?: string;
     daemonId?: string;
     machineId?: string;
+    role?: RepoMeshDaemonRole;
     machineStatus?: string;
     isLocalWorktree?: boolean;
     worktreeBranch?: string;
+    /** Mirrored from LocalMeshNodeEntry.systemPrompt for coordinator-prompt rendering. */
+    systemPrompt?: string;
     health: RepoMeshNodeHealth;
     git?: GitRepoStatus;
     /**
@@ -317,9 +901,30 @@ export interface RepoMeshNodeStatus {
     activeSessionDetails?: RepoMeshSessionStatus[];
     providerPriority?: string[];
     launchReady?: boolean;
+    /** True when the node is clean, ahead=0, behind>0, and safe for fast-forward consideration. */
+    autoFastForwardEligible?: boolean;
+    /** Coordinator-facing suggestion for obvious clean catch-up work. */
+    suggestedAction?: 'auto_fast_forward';
+    worktreeBootstrap?: LocalMeshNodeEntry['worktreeBootstrap'];
+    launchBlockedReason?: string;
+    launchBlockedMessage?: string;
+    recoveryHint?: string;
     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;
 }
 
@@ -411,3 +1016,13 @@ export interface RepoMeshLedgerStatus {
     entries: RepoMeshLedgerEntryStatus[];
     summary: RepoMeshLedgerSummaryStatus;
 }
+
+// ─── Async Job Lifecycle ─────────────────────────
+// Shared base for all mesh async job types (refine jobs, bootstrap runs, etc.)
+// Each concrete type adds its own status enum and domain-specific fields.
+
+export interface MeshAsyncJobLifecycle {
+    startedAt?: string;
+    completedAt?: string;
+    error?: string;
+}