@pellux/goodvibes-tui 0.18.11 → 0.18.13

package/src/runtime/eval/suites.ts

@@ -1,264 +0,0 @@
-/**
- * Evaluation Harness — built-in benchmark suites.
- *
- * These suites are the stable benchmark set checked in CI.
- * Scenarios exercise production runtime paths and report real measurements.
- *
- * Each scenario's id must be stable across runs — it is used as the
- * baseline key for regression detection.
- */
-
-import type { EvalScenario, EvalRawResult } from '@pellux/goodvibes-sdk/platform/runtime/eval/types';
-import { createPerfMonitor } from '@pellux/goodvibes-sdk/platform/runtime/perf/index';
-import { createInitialSurfacePerfState } from '@pellux/goodvibes-sdk/platform/runtime/store/domains/surface-perf';
-
-// ── Helpers ───────────────────────────────────────────────────────────────────
-
-/** Simulate a synthetic render cycle burst and return timing samples. */
-function syntheticRenderCycles(count: number, targetMs: number): number[] {
-  const durations: number[] = [];
-  for (let i = 0; i < count; i++) {
-    // Simulate deterministic render work proportional to target
-    const durationMs = targetMs * (0.5 + (i / count) * 0.5);
-    durations.push(durationMs);
-  }
-  return durations;
-}
-
-// ── Core Performance Suite ────────────────────────────────────────────────────
-
-/**
- * Scenarios that validate the PerfMonitor + budget evaluation path.
- * These use production PerfMonitor instances with real budget definitions.
- */
-const corePerformanceScenarios: EvalScenario[] = [
-  {
-    id: 'core-perf:render-p95-under-budget',
-    name: 'Frame Render p95 Under Budget',
-    suite: 'core-performance',
-    description:
-      'Verifies that PerfMonitor correctly passes when render p95 stays under the 16ms budget.',
-    tags: ['latency', 'render', 'budget'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      const monitor = createPerfMonitor();
-      const surfacePerf = createInitialSurfacePerfState();
-
-      // Inject fast render cycles (well under budget)
-      const durations = syntheticRenderCycles(60, 8); // 8ms average
-      const cycles = durations.map((durationMs, i) => ({
-        cycleId: i,
-        requestedAt: t0 + i * 16,
-        completedAt: t0 + i * 16 + durationMs,
-        durationMs,
-        overBudget: durationMs > 16,
-      }));
-      surfacePerf.recentCycles = cycles;
-      surfacePerf.heapUsedBytes = 50 * 1024 * 1024; // 50 MiB stable
-
-      const perfReport = monitor.evaluate({
-        surfacePerf,
-        extraMetrics: {
-          'event.queue.depth': 10,
-          'tool.executor.overhead.p95': 2,
-          'compaction.latency.p95': 100,
-          'slo.turn_start.p95': 800,
-          'slo.cancel.p95': 150,
-          'slo.reconnect_recovery.p95': 3000,
-          'slo.permission_decision.p95': 40,
-        },
-      });
-
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        perfReport,
-        safetyViolations: 0,
-      };
-    },
-  },
-
-  {
-    id: 'core-perf:slo-turn-start-latency',
-    name: 'SLO Turn Start Latency Under Budget',
-    suite: 'core-performance',
-    description:
-      'Verifies SLO gate for turn_start.p95 stays under the 2000ms budget.',
-    tags: ['latency', 'slo', 'turn'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      const monitor = createPerfMonitor();
-      const surfacePerf = createInitialSurfacePerfState();
-      surfacePerf.heapUsedBytes = 60 * 1024 * 1024;
-
-      const perfReport = monitor.evaluate({
-        surfacePerf,
-        extraMetrics: {
-          'event.queue.depth': 50,
-          'tool.executor.overhead.p95': 3,
-          'compaction.latency.p95': 200,
-          'slo.turn_start.p95': 1200,
-          'slo.cancel.p95': 200,
-          'slo.reconnect_recovery.p95': 4000,
-          'slo.permission_decision.p95': 60,
-        },
-      });
-
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        perfReport,
-        safetyViolations: 0,
-      };
-    },
-  },
-
-  {
-    id: 'core-perf:memory-growth-stable',
-    name: 'Memory Growth Rate Stable',
-    suite: 'core-performance',
-    description:
-      'Verifies the memory growth budget passes when heap is stable.',
-    tags: ['memory', 'budget'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      const monitor = createPerfMonitor();
-      const surfacePerf = createInitialSurfacePerfState();
-
-      // Simulate two consecutive evaluations with negligible heap growth
-      surfacePerf.heapUsedBytes = 80 * 1024 * 1024;
-      surfacePerf.lastMemorySampleAt = t0 - 60_000; // 1 minute ago
-      monitor.evaluate({ surfacePerf, extraMetrics: {} }); // prime the monitor
-
-      surfacePerf.heapUsedBytes = 80 * 1024 * 1024 + 512 * 1024; // +512 KiB over 1 min
-      surfacePerf.lastMemorySampleAt = t0;
-
-      const perfReport = monitor.evaluate({ surfacePerf, extraMetrics: {} });
-
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        perfReport,
-        safetyViolations: 0,
-      };
-    },
-  },
-];
-
-// ── Safety Baseline Suite ─────────────────────────────────────────────────────
-
-/**
- * Scenarios that validate safety evaluation paths.
- * These check that the safety scorer correctly grades clean runs.
- */
-const safetyBaselineScenarios: EvalScenario[] = [
-  {
-    id: 'safety:clean-run-no-violations',
-    name: 'Clean Run — No Safety Violations',
-    suite: 'safety-baseline',
-    description: 'Verifies a clean scenario run scores 100 on the safety dimension.',
-    tags: ['safety'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        safetyViolations: 0,
-      };
-    },
-  },
-
-  {
-    id: 'safety:recovery-success',
-    name: 'Recovery Path — Succeeded',
-    suite: 'safety-baseline',
-    description: 'Validates that a successfully recovered scenario scores correctly on recovery dimension.',
-    tags: ['safety', 'recovery'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        safetyViolations: 0,
-        recoverySucceeded: true,
-      };
-    },
-  },
-
-  {
-    id: 'safety:recovery-failure-score',
-    name: 'Recovery Path — Failed (Score Validation)',
-    suite: 'safety-baseline',
-    description:
-      'Validates that a failed recovery scenario scores below the recovery floor, ' +
-      'and that the gate correctly flags it.',
-    tags: ['safety', 'recovery'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      // Intentionally reports a failed recovery to exercise the scorer path.
-      // The scenario itself completes; the low recovery score is expected.
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        safetyViolations: 0,
-        recoveryFailed: true,
-      };
-    },
-  },
-];
-
-// ── Cost & Token Suite ────────────────────────────────────────────────────────
-
-const costTokenScenarios: EvalScenario[] = [
-  {
-    id: 'cost:no-token-usage',
-    name: 'No Token Usage (Score = 100)',
-    suite: 'cost-tokens',
-    description:
-      'Verifies that scenarios without token usage receive full cost score.',
-    tags: ['cost'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        safetyViolations: 0,
-      };
-    },
-  },
-
-  {
-    id: 'cost:low-token-usage',
-    name: 'Low Token Usage — Under Target',
-    suite: 'cost-tokens',
-    description:
-      'Validates that minimal token usage stays within the cost budget.',
-    tags: ['cost', 'tokens'],
-    async run(): Promise<EvalRawResult> {
-      const t0 = Date.now();
-      return {
-        completed: true,
-        durationMs: Date.now() - t0,
-        tokens: { input: 500, output: 100 },
-        costUsd: 0.0005, // under $0.001 target
-        safetyViolations: 0,
-      };
-    },
-  },
-];
-
-// ── Registry ──────────────────────────────────────────────────────────────────
-
-/**
- * All built-in benchmark suites.
- *
- * Suite names are stable — used as keys in baselines.
- */
-export const BUILTIN_SUITES: Record<string, EvalScenario[]> = {
-  'core-performance': corePerformanceScenarios,
-  'safety-baseline': safetyBaselineScenarios,
-  'cost-tokens': costTokenScenarios,
-};
-
-/** Flat list of all built-in scenarios (across all suites). */
-export const ALL_SCENARIOS: EvalScenario[] = Object.values(BUILTIN_SUITES).flat();

package/src/runtime/events/domain-map.ts

@@ -1,148 +0,0 @@
-/**
- * RuntimeEventMap — master mapping of all typed runtime event types to their payloads.
- *
- * Used by RuntimeEventBus for type-safe subscriptions. Combine all domain
- * discriminated union members into a single flat map keyed by the type string.
- */
-import type { SessionEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/session';
-import type { TurnEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/turn';
-import type { ProviderEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/providers';
-import type { ToolEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/tools';
-import type { TaskEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/tasks';
-import type { AgentEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/agents';
-import type { WorkflowEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/workflows';
-import type { OrchestrationEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/orchestration';
-import type { CommunicationEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/communication';
-import type { PlannerEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/planner';
-import type { PermissionEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/permissions';
-import type { PluginEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/plugins';
-import type { McpEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/mcp';
-import type { TransportEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/transport';
-import type { CompactionEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/compaction';
-import type { UIEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/ui';
-import type { OpsEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/ops';
-import type { ForensicsEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/forensics';
-import type { SecurityEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/security';
-import type { AutomationEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/automation';
-import type { RouteEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/routes';
-import type { ControlPlaneEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/control-plane';
-import type { DeliveryEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/deliveries';
-import type { WatcherEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/watchers';
-import type { SurfaceEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/surfaces';
-import type { KnowledgeEvent } from '@pellux/goodvibes-sdk/platform/runtime/events/knowledge';
-
-/** Union of all runtime domain events. */
-export type AnyRuntimeEvent =
-  | SessionEvent
-  | TurnEvent
-  | ProviderEvent
-  | ToolEvent
-  | TaskEvent
-  | AgentEvent
-  | WorkflowEvent
-  | OrchestrationEvent
-  | CommunicationEvent
-  | PlannerEvent
-  | PermissionEvent
-  | PluginEvent
-  | McpEvent
-  | TransportEvent
-  | CompactionEvent
-  | UIEvent
-  | OpsEvent
-  | ForensicsEvent
-  | SecurityEvent
-  | AutomationEvent
-  | RouteEvent
-  | ControlPlaneEvent
-  | DeliveryEvent
-  | WatcherEvent
-  | SurfaceEvent
-  | KnowledgeEvent;
-
-/**
- * Utility type that maps an event type discriminant to its full event shape.
- *
- * Example:
- * ```ts
- * type Payload = RuntimeEventPayload<'TURN_SUBMITTED'>;
- * // => { type: 'TURN_SUBMITTED'; turnId: string; prompt: string }
- * ```
- */
-export type RuntimeEventPayload<T extends AnyRuntimeEvent['type']> = Extract<
-  AnyRuntimeEvent,
-  { type: T }
->;
-
-/**
- * Domain labels for use with domain-scoped subscriptions.
- *
- * This is the source-of-truth vocabulary for runtime transport/event surfaces.
- */
-export const RUNTIME_EVENT_DOMAINS = [
-  'session',
-  'turn',
-  'providers',
-  'tools',
-  'tasks',
-  'agents',
-  'workflows',
-  'orchestration',
-  'communication',
-  'planner',
-  'permissions',
-  'plugins',
-  'mcp',
-  'transport',
-  'compaction',
-  'ui',
-  'ops',
-  'forensics',
-  'security',
-  'automation',
-  'routes',
-  'control-plane',
-  'deliveries',
-  'watchers',
-  'surfaces',
-  'knowledge',
-] as const;
-
-/**
- * Domain label type for use with domain-scoped subscriptions.
- */
-export type RuntimeEventDomain = typeof RUNTIME_EVENT_DOMAINS[number];
-
-export function isRuntimeEventDomain(value: string): value is RuntimeEventDomain {
-  return (RUNTIME_EVENT_DOMAINS as readonly string[]).includes(value);
-}
-
-/** Map from domain label to its event union type. */
-export type DomainEventMap = {
-  session: SessionEvent;
-  turn: TurnEvent;
-  providers: ProviderEvent;
-  tools: ToolEvent;
-  tasks: TaskEvent;
-  agents: AgentEvent;
-  workflows: WorkflowEvent;
-  orchestration: OrchestrationEvent;
-  communication: CommunicationEvent;
-  planner: PlannerEvent;
-  permissions: PermissionEvent;
-  plugins: PluginEvent;
-  mcp: McpEvent;
-  transport: TransportEvent;
-  compaction: CompactionEvent;
-  ui: UIEvent;
-  ops: OpsEvent;
-  forensics: ForensicsEvent;
-  security: SecurityEvent;
-  automation: AutomationEvent;
-  routes: RouteEvent;
-  'control-plane': ControlPlaneEvent;
-  deliveries: DeliveryEvent;
-  watchers: WatcherEvent;
-  surfaces: SurfaceEvent;
-  knowledge: KnowledgeEvent;
-};

package/src/runtime/events/index.ts

@@ -1 +0,0 @@
-export * from '@pellux/goodvibes-sdk/platform/runtime/events/index';

package/src/runtime/events/turn.ts

@@ -1 +0,0 @@
-export * from '@pellux/goodvibes-sdk/platform/runtime/events/turn';

package/src/runtime/events/workflows.ts

@@ -1 +0,0 @@
-export * from '@pellux/goodvibes-sdk/platform/runtime/events/workflows';