lebek

Headlessly A/B-test how a harness change (system prompt, tool, tool-result text, model) affects the agent's behavior, by replaying a REAL ModMixer chat transcript and running the agent's next turn against a chosen model — no GUI, no RimWorld, no manual clicking. Use when you've changed something that should change what the model DOES (e.g. "the agent should now ask before relaunching") and want runtime evidence across variants, not just a typecheck.