@adhdev/daemon-core 0.9.82-rc.44 → 0.9.82-rc.441

package/src/mesh/contracts.ts

@@ -0,0 +1,338 @@
+/**
+ * Mesh contract v2 — first-class identity, scopes, and protocol version.
+ *
+ * Replaces the implicit conventions that grew across mesh-events.ts,
+ * mesh-work-queue.ts, mesh-ledger.ts, and mesh-tools.ts where the same
+ * concept (a coordinator, a session, a task status) appeared in different
+ * shapes per file. The audit found:
+ *
+ *   - PendingMeshCoordinatorEvent had no targetCoordinatorDaemonId. All
+ *     events were broadcast to every coordinator that drained them. Two
+ *     coordinators sharing a mesh would each receive every event,
+ *     leading to duplicate completion handling and missed targeting.
+ *   - drainPendingMeshCoordinatorEvents accepted only meshId. The caller's
+ *     coordinator identity was never available, so per-coordinator
+ *     routing was impossible by construction.
+ *   - Session identifier keys diverged across stores: targetSessionId /
+ *     assignedSessionId / instanceId / runtimeSessionId / providerSessionId.
+ *     resolveEventSessionId() tried four fallbacks per call.
+ *   - No protocol version on the JSONL ledger or MeshRuntimeStore. Schema
+ *     evolutions had no guard rail.
+ *   - mesh_reconcile_ledger existed as a routine recovery tool, not as
+ *     an incident-response escape hatch. That itself signals the routing
+ *     layer cannot be trusted.
+ *
+ * This module introduces the types; the actual wiring lands in B2 (data
+ * model + 3-way transactional store), B3 (MCP layer enforcement), and B4
+ * (frontend consumption). B1 is intentionally non-breaking — it adds the
+ * types and leaves runtime behaviour unchanged.
+ */
+
+import { daemonIdsEquivalent, machineCoreFromDaemonId } from '@adhdev/mesh-shared';
+
+/** Provider type identifier (e.g. 'claude-cli', 'codex-cli', 'roo-code').
+ *  Free-form string; no shared enum exists in daemon-core yet. */
+type ProviderType = string;
+
+// ─── Protocol version ────────────────────────────────────────────────────
+
+export const MESH_PROTOCOL_VERSION_V1 = '1.0' as const;
+export const MESH_PROTOCOL_VERSION_V2 = '2.0' as const;
+
+export type MeshProtocolVersion =
+  | typeof MESH_PROTOCOL_VERSION_V1
+  | typeof MESH_PROTOCOL_VERSION_V2;
+
+export const SUPPORTED_MESH_PROTOCOL_VERSIONS: readonly MeshProtocolVersion[] = [
+  MESH_PROTOCOL_VERSION_V1,
+  MESH_PROTOCOL_VERSION_V2,
+] as const;
+
+export function isSupportedMeshProtocolVersion(value: unknown): value is MeshProtocolVersion {
+  return typeof value === 'string'
+    && (SUPPORTED_MESH_PROTOCOL_VERSIONS as readonly string[]).includes(value);
+}
+
+// ─── Coordinator identity ────────────────────────────────────────────────
+
+/**
+ * Identity of a mesh coordinator. Required (non-optional) on every dispatch
+ * and drain operation in v2 — opaque/missing identity is the audit's
+ * smoking gun for routing failures and is structurally banned by the v2
+ * types.
+ *
+ * - daemonId: the machineId of the daemon hosting the coordinator. Stable
+ *   across coordinator restarts on the same machine.
+ * - coordinatorRunId: a UUID generated when the coordinator process starts.
+ *   Stable for the coordinator's lifetime, fresh on restart. Required so
+ *   two coordinators on the same daemon (one CLI, one MCP) can be told
+ *   apart without conflating their drains.
+ * - sessionId: optional CLI coordinator session identifier (instanceId).
+ *   Present when the coordinator is itself a CLI session, absent for
+ *   pure MCP coordinators.
+ */
+export interface CoordinatorIdentity {
+  readonly daemonId: string;
+  readonly coordinatorRunId: string;
+  readonly sessionId?: string;
+}
+
+export function coordinatorIdentityEquals(a: CoordinatorIdentity, b: CoordinatorIdentity): boolean {
+  // daemonId compares by machine core: the same daemon answers to three
+  // interchangeable id forms (mach_X / daemon_mach_X / standalone_mach_X), so a
+  // raw `===` would treat one coordinator addressed under two forms as distinct.
+  return daemonIdsEquivalent(a.daemonId, b.daemonId)
+      && a.coordinatorRunId === b.coordinatorRunId
+      && (a.sessionId ?? '') === (b.sessionId ?? '');
+}
+
+export function coordinatorIdentityKey(identity: CoordinatorIdentity): string {
+  // Stable string form for map keys and ledger payloads. Avoids the
+  // {a:b, c:d} JSON.stringify ordering trap by enforcing field order. The
+  // daemonId is normalized to its machine core so the key stays consistent with
+  // coordinatorIdentityEquals — two equal identities addressed under different
+  // daemon-id forms must produce the SAME key (else equal-but-distinct-key bug).
+  const daemonCore = machineCoreFromDaemonId(identity.daemonId) ?? identity.daemonId;
+  return `${daemonCore}|${identity.coordinatorRunId}|${identity.sessionId ?? ''}`;
+}
+
+// ─── Session handle ──────────────────────────────────────────────────────
+
+/**
+ * Unified mesh session identifier. v1 referred to "the session" with five
+ * different field names depending on which store you were reading; v2
+ * collapses them into one explicit handle that carries every disambiguator
+ * a consumer might need.
+ *
+ * - nodeId: the mesh node this session belongs to (FK into mesh node table).
+ * - sessionId: provider-instance identifier (the runtime session id).
+ * - providerType: which provider category/type the session runs (e.g.
+ *   'claude-cli', 'codex-cli', 'roo-code').
+ * - coordinatorDaemonId: which coordinator dispatched the work that
+ *   spawned this session. Used for completion event routing.
+ * - assignedAt: ms epoch when the session was bound to its current task.
+ */
+export interface MeshSessionHandle {
+  readonly nodeId: string;
+  readonly sessionId: string;
+  readonly providerType: ProviderType;
+  readonly coordinatorDaemonId: string;
+  readonly assignedAt: number;
+}
+
+export function meshSessionHandleKey(handle: MeshSessionHandle): string {
+  return `${handle.nodeId}|${handle.sessionId}`;
+}
+
+// ─── Task status (canonical enum) ────────────────────────────────────────
+
+/**
+ * Single canonical task status enum. v1 had at least three overlapping
+ * sets (queue / ledger / direct-dispatch). v2 consumers import from here.
+ */
+export const MESH_TASK_STATUSES = [
+  'pending',
+  'assigned',
+  'in_progress',
+  'completed',
+  'failed',
+  'cancelled',
+] as const;
+export type MeshTaskStatus = typeof MESH_TASK_STATUSES[number];
+
+export function isMeshTaskStatus(value: unknown): value is MeshTaskStatus {
+  return typeof value === 'string'
+    && (MESH_TASK_STATUSES as readonly string[]).includes(value);
+}
+
+// ─── Event scope ─────────────────────────────────────────────────────────
+
+/**
+ * Explicit scope for pending coordinator events. v1 had no scope: every
+ * event was broadcast to every drainer. v2 forces producers to declare
+ * intent.
+ *
+ * - 'unicast': delivered to exactly one coordinator (intendedFor). Other
+ *   coordinators' drains skip it.
+ * - 'broadcast': delivered to every coordinator on the mesh. Used for
+ *   system-wide signals (e.g. mesh-wide policy changes). The v1 default
+ *   becomes explicit here so audit tools can flag unintended broadcasts.
+ * - 'system': delivered to the daemon-level handler, not coordinators.
+ *   Reserved for infrastructure events that no coordinator should see
+ *   (e.g. ledger reconciliation outcomes).
+ */
+export const MESH_EVENT_SCOPES = ['unicast', 'broadcast', 'system'] as const;
+export type MeshEventScope = typeof MESH_EVENT_SCOPES[number];
+
+export function isMeshEventScope(value: unknown): value is MeshEventScope {
+  return typeof value === 'string'
+    && (MESH_EVENT_SCOPES as readonly string[]).includes(value);
+}
+
+// ─── Pending coordinator event v2 ────────────────────────────────────────
+
+/**
+ * v2 shape of a pending coordinator event. Strict supersets of the v1
+ * shape — every v1 field is preserved so existing readers do not break;
+ * v2 fields (scope, dispatchedBy, intendedFor, protocolVersion) are
+ * additive and consulted by v2-aware drainers only.
+ *
+ * Mixed v1/v2 events coexist during the rollout window. B2 fully cuts
+ * over once every coordinator drain is v2-aware.
+ */
+export interface PendingMeshCoordinatorEventV2 {
+  // v1 fields (preserved exactly)
+  readonly event: string;
+  readonly meshId: string;
+  readonly nodeLabel: string;
+  readonly nodeId?: string;
+  readonly workspace?: string;
+  readonly metadataEvent: Record<string, unknown>;
+  readonly coordinatorMessage?: string;
+  readonly queuedAt: number;
+
+  // v2 additions
+  readonly protocolVersion: MeshProtocolVersion;
+  readonly scope: MeshEventScope;
+  readonly dispatchedBy: CoordinatorIdentity;
+  /**
+   * Required when scope === 'unicast', forbidden otherwise. Drainers MUST
+   * skip unicast events whose intendedFor does not equal their own identity.
+   */
+  readonly intendedFor?: CoordinatorIdentity;
+}
+
+// ─── Ledger entry v2 ─────────────────────────────────────────────────────
+
+/**
+ * v2 ledger entry additions. The originatingCoordinator field is the
+ * source of truth that lets completion events route back to the
+ * coordinator that dispatched the task. v1's "worker settings carried it
+ * if it happened to be set" approach is replaced.
+ */
+export interface MeshLedgerOriginatingCoordinatorV2 {
+  readonly originatingCoordinator: CoordinatorIdentity;
+  readonly protocolVersion: MeshProtocolVersion;
+}
+
+// ─── Validation primitives ───────────────────────────────────────────────
+
+export class MeshContractViolationError extends Error {
+  readonly violationPath: string;
+  readonly protocolVersion: MeshProtocolVersion;
+  constructor(
+    protocolVersion: MeshProtocolVersion,
+    violationPath: string,
+    detail: string,
+  ) {
+    super(`mesh contract ${protocolVersion} violation at ${violationPath}: ${detail}`);
+    this.name = 'MeshContractViolationError';
+    this.violationPath = violationPath;
+    this.protocolVersion = protocolVersion;
+  }
+}
+
+function isNonEmptyString(value: unknown): value is string {
+  return typeof value === 'string' && value.length > 0;
+}
+
+export function assertCoordinatorIdentity(raw: unknown, path: string): CoordinatorIdentity {
+  if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, path, 'must be an object');
+  }
+  const obj = raw as Record<string, unknown>;
+  if (!isNonEmptyString(obj.daemonId)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.daemonId`, 'must be a non-empty string');
+  }
+  if (!isNonEmptyString(obj.coordinatorRunId)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.coordinatorRunId`, 'must be a non-empty string');
+  }
+  const sessionId = obj.sessionId;
+  if (sessionId !== undefined && !isNonEmptyString(sessionId)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.sessionId`, 'must be a non-empty string when provided');
+  }
+  return sessionId !== undefined
+    ? { daemonId: obj.daemonId, coordinatorRunId: obj.coordinatorRunId, sessionId }
+    : { daemonId: obj.daemonId, coordinatorRunId: obj.coordinatorRunId };
+}
+
+export function assertPendingMeshCoordinatorEventV2(raw: unknown, path = '$'): PendingMeshCoordinatorEventV2 {
+  if (!raw || typeof raw !== 'object' || Array.isArray(raw)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, path, 'must be an object');
+  }
+  const obj = raw as Record<string, unknown>;
+  if (!isSupportedMeshProtocolVersion(obj.protocolVersion)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.protocolVersion`, `must be one of ${SUPPORTED_MESH_PROTOCOL_VERSIONS.join(', ')}`);
+  }
+  if (!isMeshEventScope(obj.scope)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.scope`, `must be one of ${MESH_EVENT_SCOPES.join(', ')}`);
+  }
+  if (!isNonEmptyString(obj.event)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.event`, 'must be a non-empty string');
+  }
+  if (!isNonEmptyString(obj.meshId)) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.meshId`, 'must be a non-empty string');
+  }
+  const dispatchedBy = assertCoordinatorIdentity(obj.dispatchedBy, `${path}.dispatchedBy`);
+
+  if (obj.scope === 'unicast') {
+    if (!obj.intendedFor) {
+      throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.intendedFor`, 'unicast scope requires intendedFor');
+    }
+  } else if (obj.intendedFor !== undefined) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.intendedFor`, 'only unicast scope may set intendedFor');
+  }
+  const intendedFor = obj.intendedFor
+    ? assertCoordinatorIdentity(obj.intendedFor, `${path}.intendedFor`)
+    : undefined;
+
+  const metadata = obj.metadataEvent && typeof obj.metadataEvent === 'object' && !Array.isArray(obj.metadataEvent)
+    ? (obj.metadataEvent as Record<string, unknown>)
+    : null;
+  if (!metadata) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.metadataEvent`, 'must be an object');
+  }
+  const queuedAt = typeof obj.queuedAt === 'number' && Number.isFinite(obj.queuedAt) ? obj.queuedAt : null;
+  if (queuedAt === null) {
+    throw new MeshContractViolationError(MESH_PROTOCOL_VERSION_V2, `${path}.queuedAt`, 'must be a finite number');
+  }
+
+  return {
+    event: obj.event,
+    meshId: obj.meshId,
+    nodeLabel: isNonEmptyString(obj.nodeLabel) ? obj.nodeLabel : '',
+    nodeId: typeof obj.nodeId === 'string' ? obj.nodeId : undefined,
+    workspace: typeof obj.workspace === 'string' ? obj.workspace : undefined,
+    metadataEvent: metadata,
+    coordinatorMessage: typeof obj.coordinatorMessage === 'string' ? obj.coordinatorMessage : undefined,
+    queuedAt,
+    protocolVersion: obj.protocolVersion,
+    scope: obj.scope,
+    dispatchedBy,
+    ...(intendedFor ? { intendedFor } : {}),
+  };
+}
+
+// ─── Routing helper ──────────────────────────────────────────────────────
+
+/**
+ * Decide whether a v2 pending event should be delivered to the given drainer.
+ * Centralised so every drain implementation uses the same rule.
+ *
+ *   - 'broadcast': always delivered.
+ *   - 'system': never delivered to coordinators (system handler only).
+ *   - 'unicast': delivered iff intendedFor matches drainer identity.
+ *
+ * v1 events (no scope / no protocolVersion) are treated as broadcast for
+ * backward compatibility during rollout; B3 tightens this so v1 events
+ * are quarantined to a dedicated drain endpoint instead.
+ */
+export function shouldDeliverPendingEventToCoordinator(
+  event: PendingMeshCoordinatorEventV2,
+  drainer: CoordinatorIdentity,
+): boolean {
+  if (event.scope === 'system') return false;
+  if (event.scope === 'broadcast') return true;
+  if (!event.intendedFor) return false;
+  return coordinatorIdentityEquals(event.intendedFor, drainer);
+}