eval-support-agent

Support Agent Eval Suite

Pytest-based evals that prove Statewave produces correct, ranked, traceable context for support-agent scenarios. The tests run against a live Statewave server (no mocks).

What's covered

File	Scenario	Asserts
test_support_context.py	Returning enterprise customer, 3 sessions	identity recall, preference recall, history recall, token-budget respected, provenance traces, idempotent compile, memory-count sanity
test_handoff.py	3 sessions with 2 resolved + 1 open issue	active-issue extraction, customer facts surfaced, resolution deprioritization, signal preserved, provenance, determinism
test_support_advanced.py	4 sessions with a recurring billing-gateway timeout	session-aware ranking, repeat-issue detection, health endpoint shape, health-aware handoff, resolution-aware ranking, compactness, determinism, provenance

Run

# Requires Statewave server at http://localhost:8100
pip install statewave httpx pytest

pytest -v                                    # all evals
pytest test_support_context.py               # one file
pytest -k handoff                            # by keyword

Why this matters

Memory quality is measurable, not anecdotal. Run the suite against any Statewave deployment to validate the context bundle, the handoff pack, and the health endpoint shape. The tests intentionally sit at the public API boundary, so they double as integration coverage.