Swift client + agent middleware for the TypeSafe AI System One API (jev-latest).
Two products:
- TypeSafeKit — thin, dependency-free HTTP client for
POST /v1/systemoneandGET /v1/models. - TypeSafeMiddleware — a
DecisionProviderprotocol plus typed helpers (route,gate,rate,composite,evaluate) for wiring Jev into an agent loop.
Jev takes a state plus a map of typed questions and returns typed answers with probability distributions. There are no messages, no assistant text, no streaming, and no tool calling. This package therefore does not imitate an OpenAI/Anthropic chat surface — a shim like that would have to fabricate text Jev never produces. Use Jev next to your chat model, for the decisions: intent routing, tool-use gating, confidence checks, classification, extraction scoring.
.package(url: "https://github.com/AgentiLoop/TypeSafeKit.git", from: "1.0.0").product(name: "TypeSafeKit", package: "TypeSafeKit"),
.product(name: "TypeSafeMiddleware", package: "TypeSafeKit"),import TypeSafeKit
let client = try TypeSafeClient() // reads TYPESAFE_API_KEY, defaults to jev-latest
let reply = try await client.systemOne(
state: "Hi, I've been trying to connect my Stripe account for 3 days and it keeps failing.",
questions: [
"is_urgent": .noul(instructions: "Does this message convey urgency?"),
"department": .choice(instructions: "Which team should handle this?", criteria: [
"billing": "Payments, invoicing, refunds",
"technical": "Bugs, outages, integrations",
"sales": "Pricing, upgrades, new accounts",
]),
"frustration": .score(
instructions: "How frustrated is the customer?",
levels: ["Calm", "Frustrated", "Very angry"]),
])
reply["is_urgent"]?.noulValue // 0.92
reply["department"]?.choiceValue // "technical"
reply["department"]?.confidence // 0.82
reply["frustration"]?.scoreValue // 1.6
reply.usage.inputTokens // billed; output tokens are free
reply.model // versioned ID that answered, e.g. jev-1.13.0Every question in one request is evaluated in parallel against the same state, so pack them together (speculative fan-out) instead of making several calls.
State can be structured:
try await client.systemOne(state: State(["turn 1", "turn 2"]), questions: [...])
try await client.systemOne(state: try State.encoding(myRecord), questions: [...])List models:
for model in try await client.models.list() { print(model.name, model.releaseDate) }import TypeSafeMiddleware
enum Route: String, CaseIterable { case answerDirectly, callTool, askHuman }
let jev = try JevProvider()
let routing = try await jev.route(
State(transcript),
instructions: "What should the agent do next?",
as: Route.self,
descriptions: [
.answerDirectly: "The answer is already known",
.callTool: "External data or a side effect is needed",
.askHuman: "Destructive, ambiguous, or out of policy",
],
minimumConfidence: 0.7)
guard routing.isConfident else { return .escalate }
switch routing.route { ... }// Tool-use gate before executing a side effect
let safe = try await jev.gate(
State(["tool": .string(name), "arguments": args]),
statement: "This tool call is safe to run without user confirmation.",
threshold: 0.9)
if !safe.passed { await confirmWithUser() }// Weighted composite instead of one broad judgment
let score = try await jev.composite(
State(pitch),
components: [
"market": .score(instructions: "Market size?", levels: ["Niche", "Large", "Huge"]),
"feasibility": .score(instructions: "Technical feasibility?", levels: ["Unlikely", "Plausible", "Proven"]),
],
weights: ["market": 2, "feasibility": 1])Conform your own type to DecisionProvider (one method: decide(state:questions:)) to stub Jev in tests or swap in another decision engine — the helpers come free.
| Concern | Handling |
|---|---|
| Auth | Authorization: Bearer $TYPESAFE_API_KEY; TYPESAFE_BASE_URL overrides the host |
| Retries | RetryPolicy — 3 attempts, exponential backoff + jitter, honors retry-after, retries 408/429/500/502/503/504/529 only |
| Errors | TypeSafeError cases for 400/401/403/404/422/429/529, transport, timeout, decoding, and local validation |
| Local validation | Empty instructions, empty question map, Choice with no options, Score with fewer than two levels fail before the request |
| Networking | URLSession by default; inject any TypeSafeTransport |
- 64k tokens per request total; 32k for state plus the longest single question
- Text only — pre-process images/audio/binaries into text or fields
- Input tokens billed, output tokens free
- Rate limits are adjusting dynamically during early access; expect 429s and keep retries on
swift test # 14 tests, no network — stubbed transport
Every request/response shape asserted in the tests is taken from the published API reference. Nothing here has been exercised against the live service yet — Jev is in early access behind a waitlist, so a real TYPESAFE_API_KEY is needed to confirm end-to-end.
MIT — see LICENSE.