Your stack and its hash
An agent is a model, a system prompt, the tools it exposes and the wallet it trades from. Crashtest tests that stack as one thing and hashes it: suite version, model, endpoint, prompt template, tool definitions, wallet guard and, for an endpoint agent, the hash its manifest publishes. Keys are sorted before hashing, so the same stack always gives the same hash.
The display name is not part of the hash. Everything that changes behaviour is. When any part changes, the hash changes and check() in the registry returns not valid for the old result.
| Field in agent.json | Meaning |
|---|---|
model | A model id from models.json, or endpoint |
endpoint | URL of an agent that runs on your side (POST /step) |
systemPrompt | Your prompt; each scenario's owner policy is appended |
tools | The tools your agent exposes; default all 14 |
wallet | raw, caps, allowlist or brake |
mystery | Opt-in to unannounced runs against the endpoint |
Run: pnpm suite --spec agent.json --runs 10 --seatbelt --attest. The form on the Test your agent page writes this file.