@artemiskit/core 0.5.3 → 0.6.1

package/dist/provenance/execution-provenance.d.ts

@@ -1,5 +1,5 @@
 /** Bounded execution provenance derived from declared and observed run evidence. */
-import type { CaseResult, ExecutionProvenance } from '../artifacts/types';
+import type { CaseResult, ExecutionProvenance } from '../artifacts/types.js';
 export declare function createExecutionProvenance(options: {
     provider: string;
     requestedModel?: string;

package/dist/provenance/git.d.ts

@@ -1,7 +1,7 @@
 /**
  * Git provenance utilities
  */
-import type { GitInfo } from '../artifacts/types';
+import type { GitInfo } from '../artifacts/types.js';
 /**
  * Get git information for the current repository
  */

package/dist/provenance/index.d.ts

@@ -1,8 +1,8 @@
 /**
  * Provenance module exports
  */
-export { getGitInfo } from './git';
-export { getEnvironmentInfo } from './environment';
-export { createWorkloadIdentity } from './workload-identity';
-export { createExecutionProvenance } from './execution-provenance';
+export { getGitInfo } from './git.js';
+export { getEnvironmentInfo } from './environment.js';
+export { createWorkloadIdentity } from './workload-identity.js';
+export { createExecutionProvenance } from './execution-provenance.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/provenance/workload-identity.d.ts

@@ -1,8 +1,8 @@
 /**
  * Canonical, redacted workload identities for reproducible assessment evidence.
  */
-import type { WorkloadIdentity } from '../artifacts/types';
-import type { Scenario } from '../scenario/schema';
+import type { WorkloadIdentity } from '../artifacts/types.js';
+import type { Scenario } from '../scenario/schema.js';
 /**
  * Create independent identities for the declared workload and its evaluation
  * criteria. Input is sorted, redacted, and hashed locally; raw material is

package/dist/redaction/index.d.ts

@@ -1,3 +1,3 @@
-export { BUILTIN_PATTERNS, BUILTIN_REGEX_PATTERNS, DEFAULT_REDACTION_PATTERNS, RedactionConfigSchema, type BuiltinPatternName, type CaseRedactionDetails, type RedactionConfig, type RedactionMetadata, type RedactionOptions, type RedactionResult, } from './types';
-export { Redactor, createDefaultRedactor, createNoOpRedactor, createRedactionOptions, hashText, redactText, redactWithHash, resolvePatterns, } from './redactor';
+export { BUILTIN_PATTERNS, BUILTIN_REGEX_PATTERNS, DEFAULT_REDACTION_PATTERNS, RedactionConfigSchema, type BuiltinPatternName, type CaseRedactionDetails, type RedactionConfig, type RedactionMetadata, type RedactionOptions, type RedactionResult, } from './types.js';
+export { Redactor, createDefaultRedactor, createNoOpRedactor, createRedactionOptions, hashText, redactText, redactWithHash, resolvePatterns, } from './redactor.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/redaction/redactor.d.ts

@@ -1,4 +1,4 @@
-import { type RedactionConfig, type RedactionOptions, type RedactionResult } from './types';
+import { type RedactionConfig, type RedactionOptions, type RedactionResult } from './types.js';
 /**
  * Resolve pattern strings to RegExp objects
  * Supports built-in pattern names and custom regex strings

package/dist/runner/executor.d.ts

@@ -1,9 +1,9 @@
 /**
  * Test case executor
  */
-import type { CaseResult } from '../artifacts/types';
-import type { TestCase } from '../scenario/schema';
-import type { ExecutorContext } from './types';
+import type { CaseResult } from '../artifacts/types.js';
+import type { TestCase } from '../scenario/schema.js';
+import type { ExecutorContext } from './types.js';
 /**
  * Execute a single test case
  */

package/dist/runner/index.d.ts

@@ -1,7 +1,7 @@
 /**
  * Runner module exports
  */
-export * from './types';
-export { executeCase } from './executor';
-export { runScenario, runScenarios } from './runner';
+export * from './types.js';
+export { executeCase } from './executor.js';
+export { runScenario, runScenarios } from './runner.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/runner/runner.d.ts

@@ -1,7 +1,7 @@
 /**
  * Scenario runner - main entry point for running test scenarios
  */
-import type { RunOptions, RunResult } from './types';
+import type { RunOptions, RunResult } from './types.js';
 /**
  * Run a test scenario
  */

package/dist/runner/types.d.ts

@@ -1,11 +1,11 @@
 /**
  * Runner types and interfaces
  */
-import type { ModelClient } from '../adapters/types';
-import type { CaseResult, CostProvenance, ResolvedConfig, RunManifest } from '../artifacts/types';
-import type { RedactionConfig } from '../redaction/types';
-import type { Scenario } from '../scenario/schema';
-import type { ToolExecutor } from '../tools';
+import type { ModelClient } from '../adapters/types.js';
+import type { CaseResult, CostProvenance, ResolvedConfig, RunManifest } from '../artifacts/types.js';
+import type { RedactionConfig } from '../redaction/types.js';
+import type { Scenario } from '../scenario/schema.js';
+import type { ToolExecutor } from '../tools/index.js';
 /**
  * Options for running a scenario
  */

package/dist/scenario/index.d.ts

@@ -1,8 +1,8 @@
 /**
  * Scenario module exports
  */
-export * from './schema';
-export * from './parser';
-export * from './variables';
-export * from './discovery';
+export * from './schema.js';
+export * from './parser.js';
+export * from './variables.js';
+export * from './discovery.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/scenario/parser.d.ts

@@ -1,7 +1,7 @@
 /**
  * Scenario parser for YAML files
  */
-import { type Scenario } from './schema';
+import { type Scenario } from './schema.js';
 /**
  * Parse a scenario from a YAML file
  */

package/dist/scenario/variables.d.ts

@@ -2,7 +2,7 @@
  * Variable substitution for scenario templates
  * Supports {{variable}} syntax in strings
  */
-import type { Variables } from './schema';
+import type { Variables } from './schema.js';
 /**
  * Substitute variables in a string using {{variable}} syntax
  */

package/dist/storage/factory.d.ts

@@ -1,7 +1,7 @@
 /**
  * Storage factory for creating storage adapters
  */
-import type { StorageAdapter, StorageConfig } from './types';
+import type { StorageAdapter, StorageConfig } from './types.js';
 /**
  * Create a storage adapter from configuration
  */

package/dist/storage/index.d.ts

@@ -1,8 +1,8 @@
 /**
  * Storage module exports
  */
-export * from './types';
-export { createStorageAdapter, createStorageFromEnv } from './factory';
-export { SupabaseStorageAdapter, type SupabaseStorageConfig } from './supabase';
-export { LocalStorageAdapter } from './local';
+export * from './types.js';
+export { createStorageAdapter, createStorageFromEnv } from './factory.js';
+export { SupabaseStorageAdapter, type SupabaseStorageConfig } from './supabase.js';
+export { LocalStorageAdapter } from './local.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/storage/local.d.ts

@@ -1,8 +1,8 @@
 /**
  * Local filesystem storage adapter
  */
-import { type AnyManifest, type RedTeamManifest, type RunManifest, type StressManifest } from '../artifacts/types';
-import type { BaselineMetadata, BaselineStorageAdapter, ComparisonResult, ListOptions, RunListItem } from './types';
+import { type AnyManifest, type RedTeamManifest, type RunManifest, type StressManifest } from '../artifacts/types.js';
+import type { BaselineMetadata, BaselineStorageAdapter, ComparisonResult, ListOptions, RunListItem } from './types.js';
 export declare class LocalStorageAdapter implements BaselineStorageAdapter {
     private basePath;
     private baselinesPath;

package/dist/storage/supabase.d.ts

@@ -1,8 +1,8 @@
 /**
  * Supabase storage adapter with analytics capabilities
  */
-import { type RunManifest } from '../artifacts/types';
-import type { AnalyticsStorageAdapter, BaselineMetadata, CaseResultQueryOptions, CaseResultRecord, ComparisonResult, ListOptions, MetricsSnapshot, MetricsTrendOptions, RunListItem, TrendDataPoint } from './types';
+import { type RunManifest } from '../artifacts/types.js';
+import type { AnalyticsStorageAdapter, BaselineMetadata, CaseResultQueryOptions, CaseResultRecord, ComparisonResult, ListOptions, MetricsSnapshot, MetricsTrendOptions, RunListItem, TrendDataPoint } from './types.js';
 export interface SupabaseStorageConfig {
     url: string;
     anonKey: string;

package/dist/storage/types.d.ts

@@ -1,8 +1,8 @@
 /**
  * Storage types and interfaces
  */
-import type { AnyManifest, CaseEvaluationEvidence, RedTeamManifest, RunManifest, StressManifest } from '../artifacts/types';
-import type { ComparisonEligibility } from '../comparison';
+import type { AnyManifest, CaseEvaluationEvidence, RedTeamManifest, RunManifest, StressManifest } from '../artifacts/types.js';
+import type { ComparisonEligibility } from '../comparison/index.js';
 /**
  * Run listing item
  */

package/dist/tools/fixture-executor.d.ts

@@ -1,5 +1,5 @@
-import type { ToolCall } from '../adapters/types';
-import type { FixtureExecutorOptions, ToolExecutionContext, ToolExecutionResult } from './types';
+import type { ToolCall } from '../adapters/types.js';
+import type { FixtureExecutorOptions, ToolExecutionContext, ToolExecutionResult } from './types.js';
 export declare class FixtureToolExecutor {
     private readonly tools;
     private readonly fixtures;

package/dist/tools/index.d.ts

@@ -1,4 +1,4 @@
-export * from './fixture-executor';
-export * from './schema-validator';
-export * from './types';
+export * from './fixture-executor.js';
+export * from './schema-validator.js';
+export * from './types.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/tools/types.d.ts

@@ -1,4 +1,4 @@
-import type { ToolCall, ToolDefinition } from '../adapters/types';
+import type { ToolCall, ToolDefinition } from '../adapters/types.js';
 export type ToolLoopTerminationReason = 'completed' | 'max_steps' | 'timeout' | 'duplicate_call' | 'invalid_arguments' | 'unknown_tool' | 'tool_error';
 export interface ToolLoopPolicy {
     enabled: boolean;

package/dist/utils/index.d.ts

@@ -1,6 +1,6 @@
 /**
  * Utilities module exports
  */
-export * from './errors';
-export * from './logger';
+export * from './errors.js';
+export * from './logger.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/validator/index.d.ts

@@ -1,6 +1,6 @@
 /**
  * Validator module exports
  */
-export * from './types';
-export { ScenarioValidator } from './validator';
+export * from './types.js';
+export { ScenarioValidator } from './validator.js';
 //# sourceMappingURL=index.d.ts.map

package/dist/validator/validator.d.ts

@@ -7,7 +7,7 @@
  * 3. Semantic errors (duplicate IDs, undefined variables)
  * 4. Warnings (deprecated patterns)
  */
-import type { ValidationResult, ValidatorOptions } from './types';
+import type { ValidationResult, ValidatorOptions } from './types.js';
 /**
  * Scenario validator class
  */

package/package.json

@@ -1,6 +1,6 @@
 {
   "name": "@artemiskit/core",
-  "version": "0.5.3",
+  "version": "0.6.1",
   "description": "Core runner, evaluators, and storage for ArtemisKit LLM evaluation toolkit",
   "type": "module",
   "license": "Apache-2.0",
@@ -30,12 +30,12 @@
   "types": "./dist/index.d.ts",
   "exports": {
     ".": {
-      "import": "./dist/index.js",
-      "types": "./dist/index.d.ts"
+      "types": "./dist/index.d.ts",
+      "import": "./dist/index.js"
     }
   },
   "scripts": {
-    "build": "tsc --emitDeclarationOnly && bun build ./src/index.ts --outdir ./dist --target bun",
+    "build": "tsc --emitDeclarationOnly && bun build ./src/index.ts --outdir ./dist --target node --external '@artemiskit/*' && node ../../scripts/finalize-declarations.mjs ./dist",
     "typecheck": "tsc --noEmit",
     "clean": "rm -rf dist",
     "test": "bun test"

package/src/adapters/types.ts

@@ -31,6 +31,10 @@ export interface ToolCall {
  * Options for generating a completion
  */
 export interface GenerateOptions {
+  /** Optional transport cancellation; consult capabilities before relying on it. */
+  signal?: AbortSignal;
+  /** Override retries per request when a host owns cumulative admission budgets. */
+  maxRetries?: number;
   prompt: string | ChatMessage[];
   model?: string;
   maxTokens?: number;
@@ -78,6 +82,8 @@ export interface TokenUsage {
  * Result from a generation request
  */
 export interface GenerateResult {
+  /** False when zero counts are placeholders rather than provider-reported usage. */
+  usageAvailable?: boolean;
   id: string;
   model: string;
   text: string;
@@ -96,6 +102,7 @@ export interface GenerateResult {
  * Model capabilities
  */
 export interface ModelCapabilities {
+  transportCancellation?: boolean;
   streaming: boolean;
   functionCalling: boolean;
   toolUse: boolean;

package/src/agent-workflow/catalog.ts

@@ -0,0 +1,220 @@
+/** Versioned run-local capabilities shared by simulated and disposable sandbox environments. */
+export const WORKFLOW_TOOL_IDS = [
+  'search',
+  'read_document',
+  'query_records',
+  'read_file',
+  'write_file',
+  'calculator',
+  'get_workflow_state',
+  'request_approval',
+  'record_decision',
+  'draft_message',
+  'delegate_task',
+  'get_task_status',
+] as const;
+
+export type WorkflowToolId = (typeof WORKFLOW_TOOL_IDS)[number];
+export type WorkflowResource =
+  | 'documents'
+  | 'records'
+  | 'files'
+  | 'computation'
+  | 'workflow_state'
+  | 'communication'
+  | 'coordination';
+
+export interface WorkflowToolDescriptor {
+  id: WorkflowToolId;
+  version: '1';
+  family: string;
+  description: string;
+  inputSchema: Record<string, unknown>;
+  outputSchema: Record<string, unknown>;
+  authority: {
+    resource: WorkflowResource;
+    access: 'read' | 'write' | 'none';
+    network: 'denied';
+    /** V1's simulated label means run-local effects; file storage follows the selected environment. */
+    sideEffects: 'none' | 'simulated';
+  };
+  evidence: { mode: 'metadata_only'; maxBytes: 1024 };
+  failureModes: string[];
+}
+
+const text = { type: 'string', minLength: 1, maxLength: 16_384 };
+const identifier = {
+  type: 'string',
+  pattern: '^(?!(?:__proto__|prototype|constructor)$)[a-zA-Z0-9_-]{1,128}$',
+};
+const relativePath = {
+  type: 'string',
+  maxLength: 512,
+  pattern:
+    '^(?!.*(?:^|/)(?:\\.\\.?|__proto__|prototype|constructor)(?:/|$))[a-zA-Z0-9_-][a-zA-Z0-9_./-]*$',
+};
+
+function object(properties: Record<string, unknown>, required = Object.keys(properties)) {
+  return { type: 'object', properties, required, additionalProperties: false };
+}
+
+function descriptor(
+  id: WorkflowToolId,
+  family: string,
+  description: string,
+  resource: WorkflowResource,
+  access: 'read' | 'write' | 'none',
+  inputSchema: Record<string, unknown>,
+  outputSchema: Record<string, unknown>
+): WorkflowToolDescriptor {
+  return {
+    id,
+    version: '1',
+    family,
+    description,
+    inputSchema,
+    outputSchema,
+    authority: {
+      resource,
+      access,
+      network: 'denied',
+      sideEffects: access === 'write' ? 'simulated' : 'none',
+    },
+    evidence: { mode: 'metadata_only', maxBytes: 1024 },
+    failureModes: [
+      'undeclared_tool',
+      'permission_denied',
+      'invalid_policy',
+      'invalid_input',
+      'invalid_state',
+      'not_found',
+      'output_limit',
+      'tool_error',
+    ],
+  };
+}
+
+const tools: WorkflowToolDescriptor[] = [
+  descriptor(
+    'search',
+    'retrieval',
+    'Search declared documents by case-insensitive literal text.',
+    'documents',
+    'read',
+    object({ query: text }),
+    object({ matches: { type: 'array', maxItems: 100, items: object({ id: identifier }) } })
+  ),
+  descriptor(
+    'read_document',
+    'documents',
+    'Read one document from simulated state.',
+    'documents',
+    'read',
+    object({ id: identifier }),
+    object({ id: identifier, content: { type: 'string', maxLength: 16_384 } })
+  ),
+  descriptor(
+    'query_records',
+    'records',
+    'Read a bounded declared collection of structured records.',
+    'records',
+    'read',
+    object({ collection: identifier }),
+    object({ records: { type: 'array', maxItems: 100, items: { type: 'object' } } })
+  ),
+  descriptor(
+    'read_file',
+    'files',
+    'Read one relative file from the isolated workflow environment.',
+    'files',
+    'read',
+    object({ path: relativePath }),
+    object({ content: { type: 'string', maxLength: 16_384 } })
+  ),
+  descriptor(
+    'write_file',
+    'files',
+    'Write one relative file in the isolated workflow environment.',
+    'files',
+    'write',
+    object({ path: relativePath, content: { type: 'string', maxLength: 16_384 } }),
+    object({ path: relativePath, written: { const: true } })
+  ),
+  descriptor(
+    'calculator',
+    'computation',
+    'Perform one finite arithmetic operation without evaluating code.',
+    'computation',
+    'none',
+    object({
+      operation: { enum: ['add', 'subtract', 'multiply', 'divide'] },
+      a: { type: 'number' },
+      b: { type: 'number' },
+    }),
+    object({ value: { type: 'number' } })
+  ),
+  descriptor(
+    'get_workflow_state',
+    'workflow',
+    'Read declared workflow state.',
+    'workflow_state',
+    'read',
+    object({}),
+    object({ state: { type: 'object' } })
+  ),
+  descriptor(
+    'request_approval',
+    'approvals',
+    'Record a pending simulated request; never grant approval.',
+    'workflow_state',
+    'write',
+    object({ reason: text }),
+    object({ requested: { const: true }, status: { const: 'pending' } })
+  ),
+  descriptor(
+    'record_decision',
+    'workflow',
+    'Record a proposed decision without granting authority.',
+    'workflow_state',
+    'write',
+    object({ decision: text }),
+    object({ recorded: { const: true } })
+  ),
+  descriptor(
+    'draft_message',
+    'communication',
+    'Create a local message draft; never send it.',
+    'communication',
+    'write',
+    object({ recipient: text, body: text }),
+    object({ id: identifier, status: { const: 'draft' } })
+  ),
+  descriptor(
+    'delegate_task',
+    'coordination',
+    'Record a pending local task; never invoke a worker or external system.',
+    'coordination',
+    'write',
+    object({ task: text }),
+    object({ id: identifier, status: { const: 'pending' } })
+  ),
+  descriptor(
+    'get_task_status',
+    'coordination',
+    'Read the status of a declared simulated task.',
+    'coordination',
+    'read',
+    object({ id: identifier }),
+    object({ id: identifier, status: { enum: ['pending', 'completed', 'failed'] } })
+  ),
+];
+
+/** Defensive copies prevent callers or adapters from modifying the catalog's authority. */
+export function listWorkflowTools(): WorkflowToolDescriptor[] {
+  return structuredClone(tools);
+}
+
+export function getWorkflowTool(id: string): WorkflowToolDescriptor | undefined {
+  const tool = tools.find((candidate) => candidate.id === id);
+  return tool ? structuredClone(tool) : undefined;
+}