Testing the AI subsystem

1. FakeAiResearchProvider
2. InMemoryCache
3. Cassettes from JSON
4. Testing a real provider offline

AI tests must be deterministic and offline. The package ships a testing kit in src so you never call a live model in CI.

FakeAiResearchProvider

Returns a recorded ResearchResult per counterparty fingerprint (a “cassette”), and counts calls so you can assert caching.

use Gawrys\Counterparty\Ai\Testing\FakeAiResearchProvider;
use Gawrys\Counterparty\Ai\Research\{ResearchResult, Finding};

$provider = new FakeAiResearchProvider();
$provider->record($counterparty->fingerprint(), new ResearchResult([
    new Finding('Named in a 2023 enforcement notice.', 'https://regulator.example/114', 0.82, adverse: true),
    new Finding('Unverified forum rumour.', null, 0.3, adverse: false), // ungrounded -> ignored as evidence
], summary: 'One grounded adverse finding.'));

InMemoryCache

A minimal PSR-16 cache for asserting the strategy’s caching behaviour:

use Gawrys\Counterparty\Ai\Testing\InMemoryCache;

$cache = new InMemoryCache();
$strategy = new AiRiskStrategy($provider, new RiskPromptBuilder(), [], $cache);

$strategy->assess($cp, $report);
$strategy->assess($cp, $report);

self::assertSame(1, $provider->calls);  // second call served from cache
self::assertSame(1, $cache->writes);

Cassettes from JSON

Record a model response as JSON and load it through the same parser used in production, so a cassette that would not parse in prod cannot silently pass in tests:

use Gawrys\Counterparty\Ai\Testing\Cassette;

$result = (new Cassette())->fromFile(__DIR__ . '/cassettes/adverse-media.json');

{
  "summary": "One adverse-media hit located.",
  "findings": [
    { "claim": "Named in a regulatory notice.", "source_url": "https://regulator.example/114", "confidence": 0.82, "adverse": true }
  ]
}

Testing a real provider offline

To test AnthropicResearchProvider / OpenAiResearchProvider (or your own) without the network, back the JsonHttpClient with a mock PSR-18 client and queue the API’s JSON response - including tool_use / tool_calls turns to exercise the tool loop. See the provider tests in counterparty-ai for ready-made examples.