Installation

composer require --dev padosoft/eval-harness-ai-bridge

--dev is the usual choice: evals run in CI and in your test suite, not in the request path. Install it in require only if you promote production interactions or run evals from a scheduled job inside the same deployment.

There is nothing to configure

No service provider, no config file, no published assets.

A package whose job is to connect two other packages should not become a third thing to configure. Everything here is a class you construct where you use it:

use Padosoft\EvalHarnessAiBridge\Datasets\ConversationDataset;
use Padosoft\EvalHarnessAiBridge\Runners\AgentSampleRunner;
use Padosoft\EvalHarnessAiBridge\Trajectories\AgentResponseTrajectory;

The one exception is the Pest expectation, registered through a Composer files
autoload entry — Pest has no service provider to hook.

That is best-effort rather than a guarantee, and it is worth being precise about
why: the order of files entries across sibling packages is not specified,
and this package only suggests Pest, so there is no dependency edge to order
them by. In the load order where this file runs before Pest’s own function file,
expect() does not exist yet, and Composer will not re-run the file.

So the registration is a named, idempotent function you can call yourself. If
toPassEval ever comes back as an unknown expectation, one line in
tests/Pest.php settles it:

\Padosoft\EvalHarnessAiBridge\Testing\registerPestExpectations();

Calling it twice is safe; calling it without Pest is a no-op returning false.
Without Pest installed nothing runs at all, and the AssertsEvals trait is the
surface that works everywhere.

Configuration lives in eval-harness

Judges, embeddings, report disks, cost rates and budgets are all eval-harness concerns. See its configuration guide.

Verifying the install

use Padosoft\EvalHarnessAiBridge\Trajectories\AgentResponseTrajectory;

$trajectory = AgentResponseTrajectory::fromResponse(
    Ai::agent(SupportAgent::class)->prompt('When does order 44192 ship?')
);

dump($trajectory->toolNames());   // ['lookup_order']
dump($trajectory->stepCount());   // 2

If toolNames() comes back empty on an agent you know uses tools, the agent returned before the tool loop ran — that is a real finding, not an install problem.