Live Endpoint Conformance Pre-Flight
Before you trust a clean live score, prove your endpoint was even understood.
Every live driver in AgentSmack — wargame arena, RAG runtime, memory, swarm, code-execution, negotiation, the generate-then-verify gauntlet — POSTs to your deployed function-calling endpoint through the same front door, translates the reply into the next move, and scores the transcript. But if you point AgentSmack at an endpoint that 404s, times out, returns non-JSON, or speaks an unsupported wire shape, every driver reads an empty stop and the agent silently scores CLEAN — a dangerous false-negative-by-misconfiguration that reads as “your agent passed!” when in truth not a single reply was understood.
This pre-flight is the trust-foundation under all of them: a single benign round-trip that grades a five-step conformance ladder — SSRF-allowed → reachable → returns JSON → parseable in the selected wire shape → produces a scoreable signal — and folds it into one verdict (ready / degraded / not drivable) with a per-checkpoint remediation hint. The headline check is the silent-clean trap: an endpoint that returns valid JSON but no tool call and empty text is flagged degraded, never “ready”. A blocked target is refused with zero fetch. Point it at an endpoint you control; it never sends an attack payload.
Pre-flight a live agent endpoint
This POSTs ONE benign turn (“Reply with a brief acknowledgement.”) to a real endpoint you control and grades whether it is reachable, SSRF-allowed, and emits a parseable, scoreable reply in the selected wire shape — before you run any gauntlet. It is non-adversarial and harmless; it never sends an attack payload.
POSTs one benign turn to a real endpoint you control.