Quickstart · 6 of 8

Simulations.

Simulations run a browser-based journey on your site — either a fixed script or a real AI agent acting autonomously.

Trigger a simulation (dashboard or API token)

POST /api/simulations
Authorization: Bearer scvt_...
{ "site_id": "<uuid>", "scenario_type": "agent_commerce",
  "start_url": "https://example.com", "agents": ["claude"] }

The start_url must be on the site's registered domain (subdomains included) — an off-domain URL is rejected with 400 {"error": "start_url_outside_site"} and no run is created.

Supported agents: a multi-model catalog — see the FAQ and GET /api/simulations/agents for the current list. Most models run via AWS Bedrock cross-region inference profiles (including OpenAI's open-weight gpt-oss); the GPT-6 family, once enabled, runs through OpenAI's API directly. A model that is retired or not enabled on this platform is refused at run creation (HTTP 422, model_retired: / model_unavailable:).

Simulations are credit-gated — script and AI-agent runs alike, one credit per run. Requesting agents: ["claude","nova"] consumes two credits. When the balance is insufficient the endpoint returns 402 with the available credit packs and no run is created. Credits are purchased via Paddle (packs of 10/25/60/150) and never expire; Team, Business and Enterprise workspaces don't spend credits on script simulations.

CI-triggered simulations (POST /api/ci/trigger with run_simulation=true) consume credits the same way.

Simulation lifecycle: pending → queued → running → completed | failed

Results include per-step traces you can replay in the dashboard, a run profile (execution_summary: success_mode autonomous or steering-assisted, steering hints, recoveries, harness retries, LLM calls) and, on failure, an owner×class attribution (failure_owner, failure_class). These are measured, never scored.

Model tiers (2026-09)

Simulations can run on Claude Opus 5.5 (opus55), Claude Opus 5 (opus5) and Claude Fable 5.1 (fable) today, and — once the platform's OpenAI access is enabled — on GPT-6 Sol (gpt6sol) and GPT-6 Luna (gpt6luna); both are catalogued but not yet publicly listed. Claude Opus 4.8 (opus) and Claude Haiku 4.5 (haiku) were retired on 2026-11-11 and replaced by Claude Opus 5.5 (opus55) and Claude Haiku 5.5 (haiku55); runs already recorded keep their results. A model that is retired or not available on this platform is refused when you create a run (HTTP 422, model_retired: / model_unavailable:) and is dropped from a site's CI agents. Every run costs one credit regardless of the model.

Sites also get a Minimum capable tier card: the cheapest model that completed the site's task at or above 95% over at least 20 runs. It is informational and never affects the score.

Readiness across agents

For commerce-tier sites, Agent Readiness is derived from the pass-rate across every agent you ran (threshold 75%): ready at or above it, partial in between, not_ready at zero. A failing agent is an informational note, not a penalty, as long as at least 75% of the agents you ran complete the purchase; with two or three agents a single failure drops readiness to partial.