garvgupta2711/taskflow-agent-evals ? reverse-engineered prompt

Build me a Python project that evaluates a simple sales AI agent called Alex.

I want two kinds of testing in one repo. First, a fixed regression benchmark with a set of conversation test cases, where the agent is run, the full conversation is saved, and the results are scored for things like qualification, pricing accuracy, discount policy, state retention, objection handling, and tool use. Second, a more dynamic buyer simulation where different customer personas can talk to the agent and reveal behaviors we would not think to hardcode.

Please include deterministic checks for hard rules like minimum team size, discount limits, and demo eligibility, plus a lighter conversational judge for things like response quality. Also track safety issues such as prompt disclosure, unauthorized discounts, and obvious policy violations, along with simple telemetry like latency, token counts if available, and tool calls.

Keep the project clean and testable, save structured outputs to JSON, and make sure the whole thing can be run from the command line with a clear path from test cases to final evaluation report.