Structured decisions
A structured decision model handles questions that have explicit candidates and criteria, such as tool / Skill recommendation, memory ranking, Agent selection and approval assistance; it can also be the model of the interjection judge. It returns reasoning evidence only and never widens permissions; activation, approval, and execution stay with the business module.
Configure several models
Install the matching Driver plugin, for example @zhin.js/service-typesafe, then create a Service, an Account, and a decision Profile in Console. They share the same catalog and persistence as chat models and are no longer bound to one global decision owner.
TypeSafe import proposal example:
# ai.plan.yml
version: 1
services:
decisions:
driver: typesafe
models:
your-decision-model:
id: your-decision-model
kinds: [decision]
protocol: typesafe-system-one
accounts:
first:
service: decisions
auth:
method: api-key
source: env://TYPESAFE_API_KEY
profiles:
fast:
account: first
model: your-decision-model
parameters:
timeoutMs: 3000
careful:
account: first
model: your-decision-model
parameters:
timeoutMs: 5000
purposes:
decisions:
tools:
profile: fast
mode: shadow
approval:
profile: careful
mode: activeReplace the example model ID with a value the service actually supports. The currently installed instance of a Driver must be uniquely identifiable; when Drivers with the same name come from different plugins, the preview requires { id, owner, contractVersion } and never picks the newest one.
Each purpose works independently
The five purposes are skills, tools, memory, agents, and approval. Each one selects a Profile and keeps its existing purpose policy (the former sixth purpose ambient moved into the interjection judge purposes.ambient, which takes a decision model directly or mixed with language models in its fallback chain, see the migration notes):
offdoes not call the model.shadowrecords the suggestion while the business keeps its original way of deciding.activehands the suggestion to that business's apply flow.
Thresholds, candidate counts, timeouts, and similar settings are still validated by the matching purpose policy, and switching Profiles never relaxes permissions. A legitimate "choose nothing" is different from a service failure; a successful apply is recorded after the business confirms it and never marked as success when the model answers. The interjection judge is not one of these five purposes: purposes.ambient has one model choice with a mixed fallback chain and one threshold, purposes.ambient.minConfidence.
Calling from an ordinary plugin
const models = context.use(aiInferenceClientToken);
const profiles = await models.list({ kind: 'decision' });
const selected = profiles.find(profile => profile.label === 'Careful');
if (!selected) throw new Error('Decision Profile unavailable');
const evidence = await models.evaluate({
profile: selected.reference,
state: { topic: 'synthetic task' },
questions: {
relevant: { type: 'choice', criteria: { read: 'Read existing facts', search: 'Find missing facts' } },
},
});The question schema must follow the real contract of the Driver in use. evaluate does not pose as one of the five purposes above and returns no applied / off state. Choice, Score, and Noul use the vendor's real results; a chat model that produces JSON is not enough to claim this contract.
A failed call never switches to a model or Account that was not selected. A personal quota or other pre-submit admission refusal is not retried automatically by the TypeSafe SDK as a network error. An unknown result for a submitted request and a request that was definitely not submitted are recorded separately; a single HTTP success or a mock result is not proof that the real vendor accepted it.