Agent adapters

A trusted Python adapter exposes name, optional usage, and next_action(scenario, events). Return exactly {"tool": "tool.name", "args": {...}}, or None to stop. Inputs are defensive copies. Errors fail the run, and only the exception class is persisted.

from verixa.engine import run
from verixa.scenarios import bundled

class Agent:
    name = "my-agent-v1"
    def next_action(self, scenario, events):
        if not events:
            return {"tool": "deployment.get", "args": {"name": "payments"}}
        if len(events) == 1:
            return {"tool": "deployment.scale", "args": {"name": "payments", "replicas": 3}}
        return None

s = next(s for s in bundled() if s["id"] == "infra-scale")
assert run(s, Agent())["verdict"] == "PASS"

An in-process adapter is trusted code and is not sandboxed.

Built-in adapters

Adapter	Behaviour
reference	Scripted fixture that retries with stable keys
regression	Scripted fixture with known mistakes, used to demonstrate the gate
replay	Proposes up to 100 recorded actions in order against fresh state
model	OpenAI-compatible /chat/completions, one tool call per turn

The model adapter converts tool names from refund.create to refund__create for provider compatibility. It rebuilds messages from observed events and records provider token counts. Responses are bounded to 1 MB, each request times out after 30 seconds, and redirects are not followed. Model output is nondeterministic. Replay is deterministic for the same scenario and action sequence.