2.0.14 Release notes: AI call contract + Agent OS governance

Shipping production LLM features takes more than a thin provider wrapper. Teams need named models, tool loops, budgets, streams, and — for HazelJS — wiring into context layers, Gatekeeper, and eval.

This release ships two stories: a production @hazeljs/​ai call contract on native providers, and Agent OS governance surfaces — Decision Plane, governed runtime, and governed memory — that enforce one policy across harnesses.

---

Why this release

Teams already using HazelJS for HTTP + agents were choosing between:

1. DIY provider glue (timeouts, budgets, FakeModel, SSE) in every app, or
2. Adopting another call layer that ignores CAG/RAG/KAG, Gatekeeper, and eval.

We closed that gap without rewriting providers onto a third-party AI SDK. Named models, retry/fallback, @AgentTool, budgets, and UI streams land on Hazel’s existing OpenAI / Anthropic / Gemini / Cohere / Ollama adapters — then plug into context, Gatekeeper, and @hazeljs/​eval.

---

What’s new in @hazeljs/​ai

Call contract (native providers)

Capability	API
Named models + pricing + fallback	AIModule.register({ models }) + modelName
Timeout / selective retry / abort	Call policy + abortSignal
Token & cost budgets	budget: { maxTokens, maxCost }
Structured output	generateObject with OUTPUT_INVALID (not blind retry)
Tools	@AgentTool + maxSteps loop
Streams	Text + UI SSE with keep-alive and client abort
Tests	FakeModel

Docs: AI package.

Three differentiators

1. Context-native calls — Pass a duck-typed @hazeljs/​context assembler so complete() / ChatBuilder / HCEL can assemble CAG + RAG + KAG + Memory before the model runs. Soft-fail matches Agent Runtime. Response includes contextMeta.

await ai.chat('Refund policy?').context(contextService, { tenantId: 'acme' }).send();

2. Eval-native goldens — runAiGolden / runAiGoldenCi wrap @hazeljs/​eval with a fresh FakeModel per case and optional metadata.expectedCalls matchers. Optional LLM judge blends into runGoldenDataset.

3. Durable Gatekeeper HITL on the tool loop — ApprovalDecision may be { pending: { approvalId } }. The loop suspends (finishReason: 'approval-required'), checkpoints, and resumes with resumeAfterApproval. Use GatekeeperApprovalPort + InMemoryCallCheckpointStore or a HumanTask-backed store.

Full Agent OS durability (DNA, run store, ledger) remains on createGovernedRuntime.

---

What’s new in Agent OS governance

Decision Plane (@hazeljs/​decision)

Opt-in defineDecision / createDecisionPlane — finite answers, discriminated PlaneResult union, rules + structured providers, Redis cache port. Does not replace DecisionRuntime.

Docs: Decision package — Decision Plane.

Governed runtime (@hazeljs/​agent-gatekeeper)

createGovernedRuntime / createGovernedToolExecutor: server-built TrustedCallerContext, optional argument digest on approvals, Redis lock + tool attempt ledger, reject hosted computer-use / shell without toolMediation.

Ships today: Redis lock/ledger; in-memory run store. OpenAI Agents adapter is opt-in (HAZEL_OPENAI_AGENTS_LIVE=1).

Docs: Agent Gatekeeper — governed runtime.

Governed memory (@hazeljs/​memory)

GovernedMemoryService lifecycle: capture → accept / quarantine / reject → retrieve → correct / forget / promote. Tenant filters before rank. Runtime store: InMemoryGovernedMemoryStore. SQL migration exists; no Prisma store adapter yet.

Docs: Memory — governed learning.

Shipped vs roadmap

See the governance slice table in the Agent OS guide.

---

What we are not claiming

• Drop-in compatibility with third-party AI SDK call surfaces
• Multi-instance Postgres for governed memory out of the box
• Governed SQL run store (Agent createSqlDurableRunStore is a separate API)
• HIPAA/GDPR certification or complete computer-use mediation

---

Upgrade path

1. Register named models on AIModule and switch hot paths to modelName + budgets.
2. Add @AgentTool for tool loops; keep InMemoryApprovalPort in tests.
3. Wire contextService when you already use @hazeljs/​context.
4. Add runAiGolden to CI for prompt/tool regressions.
5. Adopt Decision Plane for routing labels, then GovernedRuntime for destructive tools.

Existing DecisionRuntime, AgentGatekeeper.evaluate / execute, and MemoryService paths stay unchanged if you never call the new entry points.

---

Links

• AI package docs
• Eval package
• Decision package
• Agent Gatekeeper
• Memory package
• RAG vs CAG vs KAG
• What is an Agent OS?

Orchestration gets agents running. Agent OS — and a serious AI call contract — gets them shippable.