Operations

The connector exposes 11 operations. Provider-backed operations call Jev through the configured route. Policy, capability, and validation operations run locally.

Decision operations

Operation	Alias	Provider calls	Purpose
[Decide] Evaluate	evaluate	1	Evaluate a complete named question set against one state.
[Decide] Ask Yes/No	ask-noul	1	Return the yes probability for one Noul question.
[Decide] Choose	choose	1	Classify state into one of a fixed set of options.
[Decide] Score	score	1	Grade state against ordered rubric levels.
[Select] Candidate	select-candidate	1	Select and rank the best row from runtime candidates.

Scale operations

Operation	Alias	Provider calls	Purpose
[Decide] Evaluate Batch	evaluate-batch	One per unique cache miss	Apply one question set to many states with bounded concurrency.
[Select] Filter	filter	One per chunk	Retain items whose Noul probability clears a threshold.

Policy and utility operations

Operation	Alias	Provider calls	Purpose
[Policy] Apply	apply-policy	0	Convert answers to ACCEPT, REVIEW, or REJECT.
[Util] Connection Get Capabilities	get-capabilities	0	Describe primary and fallback route capabilities.
[Util] Connection List Models	list-models	One per supporting route	Merge model cards across the connection.
[Util] Validate Question Set	validate-question-set	0	Validate questions before a billed call.

Shared request options

Provider-backed operations include:

Option	Default	Description
Model override	Connection model	Select a model for this call.
Include raw response	false	Attach the provider body to attributes.rawResponse.
Use cache	false	Use the connector cache when caching is enabled globally.
Step	Empty	Copy a flow label into attributes.traceEntry.step.

Decision attributes

Provider-backed decision operations return metadata separately from the payload:

• Provider, requested model, and actual model
• Input and output token usage
• Estimated cost and cost source
• Latency, attempts, and failed-over routes
• Cache-hit status
• Question-set ID and version
• State hash and provider request ID
• Optional raw response
• Audit-ready trace entry

Provider endpoints

• TypeSafe, OpenRouter, Vercel, and compatible decision routes use POST /{apiVersion}/systemone (connection API version, default v1).
• Cloudflare uses POST /client/v4/accounts/{accountId}/ai/run/{model}.
• Model discovery uses GET /{apiVersion}/models.
• The mock route, Apply Policy, Get Capabilities, and Validate Question Set make no HTTP request.

See TypeSafe API documentation and model-list documentation.

Errors

Operations can raise TYPESAFE:* errors including:

• UNAUTHORIZED, RATE_LIMITED, OVERLOADED, TIMEOUT, and CONNECTIVITY
• PROVIDER_VALIDATION, PROVIDER_ERROR, and INVALID_RESPONSE
• INVALID_QUESTION_SET, INVALID_STATE, and TOO_MANY_OPTIONS
• BATCH_TOO_LARGE, BUDGET_EXCEEDED, and UNSUPPORTED_BY_PROVIDER
• BELOW_THRESHOLD and REJECTED

Transient connectivity, timeout, rate-limit, and overload failures can use ordered fallback routes. Validation and authorization failures do not fail over.