@image-cascade/conformance

Conformance suite and language-neutral corpus for image-context-cascade adapters.

If you wire the cascade into a new agent, framework, or provider format, this package tells you whether your integration actually preserves the invariants that make the pattern safe:

• current-turn images are never touched;
• historical images are downgraded to byte-stable placeholders;
• telemetry never leaks image data;
• malformed input fails open.

Usage

import { runConformance, loadCorpusCases, runCorpusConformance } from "@image-cascade/conformance";

// Wrap your adapter in a harness: (payload) => ({ payload, telemetry })
const report = await runConformance(myAdapterHarness);
if (!report.ok) console.error(report.failures);

The corpus

corpus/ ships JSON case files (importable via @image-cascade/conformance/corpus/<name>.json) that pin expected behavior for Anthropic Messages, OpenAI Chat Completions, OpenAI Responses, document attachments, and boundary edge cases. The cases are language-neutral: a Python or Go port can consume the same fixtures and claim conformance against the same expectations.

See the adapter guide and verification doc for the full contract.

Apache-2.0.