Skip to content

TypeSafe ​

TypeSafe hosts Jev, a decision model: it answers typed questions about a state with a probability for every option, rather than writing text. Rho uses TypeSafe only for decision models, such as the permission classifier's screen. It is not a chat provider, so /model does not list it, and selecting it as the chat model is an error.

Provider details ​

SettingValue
Providertypesafe
Authtypesafe-api-key
Environment overrideTYPESAFE_API_KEY
API basehttps://api.typesafe.ai/v1 (System One API at /systemone)
Modelsjev-latest, jev-preview, discovered from /models

Create an API key in the TypeSafe dashboard, then store it with Rho. Do not put the key in config.toml.

Interactive login ​

In the TUI, run:

text
/login typesafe

Login discovers TypeSafe's models; /refresh-models updates them. Then pick Jev where a feature takes a decision model, for example the permission screen model, by naming it in config:

toml
[internal_agents.permission-classifier-screen]
provider = "typesafe"
model = "jev-latest"
auth = "typesafe-api-key"

Remove the stored key with:

text
/logout typesafe

Environment and automation ​

For CI or headless runs, set TYPESAFE_API_KEY. A nonblank value takes precedence over the stored key.

Limits ​

TypeSafe rejects a request past about 32,500 input tokens. Rho checks the state before sending, so a permission-screen transcript over about 20,500 estimated tokens is shortened, oldest tool calls first, or the screen is skipped and the review decides. Jev reports probabilities rounded to two decimals, so the screen's allow threshold of 95% admits a true probability of about 0.945 or more.