Attest v0.6.0 — Drift Detection & SDK Parity

Released: April 17, 2026

Attest v0.6.0 closes all Phase 4 SDK gaps. The Go engine was feature-complete at v0.5.0 — it implemented dynamic thresholds, drift detection, user-message simulation, and a full persona library. But the Python and TypeScript SDKs didn’t expose all of those capabilities. This release brings both SDKs to full parity with the engine.

Highlights

• queryDrift() / query_drift() — first-class drift reports from both SDKs
• generateUserMessage() / generate_user_message() — simulated users via the engine’s LLM provider
• Dynamic thresholds in the expect DSL — σ-based drift detection with threshold: "dynamic"
• Vitest simulation mode — attestGlobalSetup() respects ATTEST_SIMULATION=1 and skips engine startup entirely
• Persona parity — 4 built-in personas aligned across engine, Python, and TypeScript

---

Drift Detection Reaches the SDKs

The engine has shipped σ-based drift detection since v0.4.0 via the query_drift RPC. Until now, there was no SDK-level method to call it — you had to hand-write JSON-RPC payloads. That ends in v0.6.0.

Python

from attest import AttestClient

report = await client.query_drift("my_assertion_id", window_size=50)

print(f"mean: {report.mean}, stddev: {report.stddev}")
print(f"latest: {report.latest_score}, deviation: {report.deviation}σ")
print(f"status: {report.status}")  # "stable" | "drift_detected" | "no_data"

TypeScript

import type { DriftReport } from "@attest-ai/core";

const report: DriftReport = await client.queryDrift("my_assertion_id", 50);

console.log(`mean: ${report.mean}, stddev: ${report.stddev}`);
console.log(`latest: ${report.latest_score}, deviation: ${report.deviation}σ`);

Both methods return a stub DriftReport in simulation mode (ATTEST_SIMULATION=1), so tests don’t hit the engine.

---

Dynamic Thresholds in the Expect DSL

The engine’s ClassifyDynamic compares scores against a historical window using configurable σ-based thresholds. The SDKs now expose this via a string literal on any assertion that takes a threshold.

Python

expect(result).output_similar_to("reference", threshold="dynamic")
expect(result).passes_judge("is polite", threshold="dynamic")

Python validates the string: threshold="bad" raises ValueError. Only "dynamic" and floats (0.0–1.0) are accepted.

TypeScript

attestExpect(result).outputSimilarTo("reference", { threshold: "dynamic" });
attestExpect(result).passesJudge("is polite", { threshold: "dynamic" });

The TypeScript type is now number | "dynamic", so the compiler catches typos at call sites.

When threshold is "dynamic", the engine consults the history store, computes the rolling mean and stddev, and fails the assertion if the latest score deviates more than N σ from baseline. Useful for catching quality regressions that a static threshold would miss.

---

Vitest Simulation Mode

The TypeScript examples repo had a papercut in CI: running @attest-ai/vitest tests always required an engine binary, even for deterministic assertions that didn’t need one. attestGlobalSetup() would spawn the engine unconditionally, and there was no way to skip it.

v0.6.0 teaches the vitest plugin about simulation mode:

vitest.config.ts

import { defineConfig } from "vitest/config";

export default defineConfig({
  test: {
    globalSetup: ["@attest-ai/vitest/setup"],
  },
});

ATTEST_SIMULATION=1 pnpm test

In simulation mode:

• attestGlobalSetup().setup() returns immediately without spawning the engine
• evaluate() returns deterministic pass results via simulationEvaluateBatch()
• useAttest() returns a simulation-aware fixture with no engine dependency

This unblocks CI in the attest-examples repo and any downstream project that wants to run Attest tests in a container without downloading the engine binary.

---

Simulated Users from the SDK

The engine has generated simulated user messages since v0.3 — generate_user_message returns a persona-driven response from the configured LLM provider. The SDKs now wrap this directly.

Python

from attest import AttestClient, ConversationMessage
from attest.simulation import FRIENDLY_USER, ADVERSARIAL_USER

history = [
    ConversationMessage(role="assistant", content="How can I help?"),
]

message = await client.generate_user_message(
    persona=ADVERSARIAL_USER,
    conversation_history=history,
)

TypeScript

import { ADVERSARIAL_USER } from "@attest-ai/core/personas";

const history = [{ role: "assistant", content: "How can I help?" }];

const message = await client.generateUserMessage(ADVERSARIAL_USER, history);

Both SDKs accept an optional faultConfig for injecting errors, latency jitter, content corruption, or timeouts — useful for resilience testing.

---

Persona Parity

Four personas are now aligned across all three platforms:

Persona	Temperature	Style	Use Case
FRIENDLY_USER	0.7	friendly	Happy path — clear, well-formed requests
ADVERSARIAL_USER	0.9	adversarial	Edge cases, boundary probing
CONFUSED_USER	0.8	confused	Vague requests, contradictions
COOPERATIVE_USER	0.6	cooperative	Actively helpful, provides clear context

CooperativeUser is new in the Go engine for v0.6.0. The TypeScript SDK gets a dedicated personas.ts module exporting all four. Python’s export had a small bug — COOPERATIVE_USER existed in personas.py but wasn’t in attest.simulation.__all__ — that’s fixed now.

---

TypeScript Config Parity

Python’s config(sample_rate=...) has existed since continuous eval shipped. TypeScript only had config({ simulation }). v0.6.0 brings full parity:

import { config, getSampleRate } from "@attest-ai/core";

config({ sampleRate: 0.1 });  // 10% sampling
console.log(getSampleRate());  // 0.1

// Or via env var
process.env["ATTEST_SAMPLE_RATE"] = "0.05";
console.log(getSampleRate());  // 0.05

Validation catches out-of-range values: config({ sampleRate: 1.5 }) throws RangeError.

---

Migration

No breaking changes. All new APIs are additive.

# Python
uv add attest-ai@0.6.0

# TypeScript
pnpm add @attest-ai/core@0.6.0 @attest-ai/vitest@0.6.0

The engine binary is also at v0.6.0 but has no behavior changes from v0.5.0 — this release is purely SDK-side. Both SDKs auto-download the matching engine version on first use.

---

What’s Next

With Phase 4 complete, the roadmap moves to Phase 5: Go SDK, Attest Cloud MVP, and a benchmark registry. Ahead of that, expect a GTM push — the code has been stable since v0.5.0 in February, and the bottleneck is now visibility.

If you’ve been running Attest in production or CI, we’d love to hear from you. Open an issue on GitHub or jump in on the existing discussions.