@lssm/lib.agent-runtime-conformance

Cross-runtime lifecycle conformance suite for ContractSpec agent runtime adapters. Makes "portable across runtimes" falsifiable instead of asserted: any adapter that claims compatibility must pass this suite.

Usage

import { runAgentRuntimeConformanceSuite } from '@lssm/lib.agent-runtime-conformance';
import { createSqliteWorkflowDevkitTarget } from '@lssm/lib.agent-runtime-conformance/fixtures';

const report = await runAgentRuntimeConformanceSuite(
  createSqliteWorkflowDevkitTarget({ file: './agent.db' })
);

console.log(report.summary.single_adapter_lifecycle_conformance_rate); // 16/19

The 19 cases (spec 04)

Cases 1-15 plus the mandatory case 5b (effect-replay safety) are testable for every durable target. Cases 16-17 remain NOT_APPLICABLE_THIS_RELEASE; case 18 becomes mandatory when an integration-local two-adapter target supplies runtime-probed portable evidence.

• 1-3 create session, stream lifecycle events, invoke an operation-backed tool
• 4 checkpoint round-trip through AgentCheckpointPort
• 5 survive forced process death (real SIGKILL child; durable backing only)
• 5b effect-replay safety on resume — the recorded result is replayed via the (nonce, stepInstanceId) key so an approval-gated sendMessage is never double-sent. Its inverse sub-case (5b-inverse) proves two distinct approval-authorized identical-input sends both fire (nonce-scoped, not input-scoped).
• 6 resume from durable recovery
• 7-8 suspend for human input / approval
• 9-12 reject missing / wrong-input / replayed approval; resume with valid approval
• 13 cancel (side effect must not fire once cancelled)
• 14-15 emit artifacts and evidence refs
• 16-17 execution receipt / outcome linkage — NOT_APPLICABLE_THIS_RELEASE
• 18 cross-adapter portability — compares runtime-probed lifecycle, approval, effect replay, artifact/evidence, replay, and explicitly contracted loss diagnostics

Metric

single_adapter_lifecycle_conformance_rate = passing / 19. Target this release is 16/19 ≈ 0.84 for the Phase-1 Workflow DevKit gate. Integration-local EVE and Workflow DevKit targets additionally run case 18, producing 17/19 with cases 16-17 explicitly not applicable.

Targets

• createInMemoryWorkflowDevkitTarget() — baseline; cannot survive process death, so cases 5/5b/6 report NOT_APPLICABLE.
• createSqliteWorkflowDevkitTarget({ file }) — durable; the canonical gate.

Vendor-specific targets remain in their integration packages. Case 18 uses runPortableAdapterProbe() and fails on any unclassified field or declared-loss difference; integrations must supply an exact expected-loss contract.