Flow Steward Extension for TypeSafe AI (Jev)

Install the release ZIP through Extensions then Install from file, or through the Flow Steward extension catalog. Then open TypeSafe AI, go to Connections, pick a project, name the connection, and enter a key from TypeSafe Console.

The key is stored in that project's encrypted connection secrets; a project can hold several independent connections. The connector resolves the opaque reference at its own authorized boundary, so the key is never passed through Template or an agent prompt.

No. Test connection authenticates with GET /v1/models and does not send a decision request, so passing it does not by itself verify credentials beyond that call.

The graph returns manual_review when category or urgency confidence falls below min_confidence (default 0.7), when the human-required probability reaches human_threshold (default 0.8), or when the category is unrecognized. Manual review is only a routing recommendation; the example creates no ticket, approval task, or notification.

Image and Box questions are outside this version's contract, and compressed responses are unsupported. Jev is not a chat-completion model, so it cannot be added as the LLM behind an agent through Add model.

The connector has no third-party runtime dependencies and does not retry requests internally. Authentication, rate limit, timeout, TLS, transport, and malformed-response errors surface as stable error codes, without raw provider error bodies or credentials.

Yes. Test Run can call TypeSafe and consume provider credits. Importing the example itself makes no API call: it ships disabled and contains no API key, configured secret default, or machine-specific install ID.