quesen-sdk (JavaScript / TypeScript)

Deterministic AI agent authorization & governance — a portable decision + authority + evidence layer (runtime authorization for agent tool calls, MCP tool authorization, agent payment authorization) returning PASS/REVIEW/BLOCK/SKIP with machine reason codes and replayable audit receipts.

Official Quesen SDK for Node 18+, Bun, Deno, and modern browsers. Zero runtime dependencies. Same 3-line integration pattern as every other Quesen SDK:

import { QuesenClient } from "quesen-sdk";

const q = new QuesenClient({
  baseUrl: "https://quesen.example.com",
  apiKey: process.env.QUESEN_API_KEY,
});

const verdict = await q.validate({
  domain_age_days: 1,
  engagement_ratio: 0.95,
  scam_keyword_count: 4,
});

if (verdict.decision === "SKIP") return; // respect the deterministic answer

Why deterministic? No LLM, no randomness. Same input in, same decision out. Every response embeds engine_version, weights, thresholds, plus (v1.10+) input_snapshot_hash and commit_sha for self-contained replay.

---

Status: v0.4.0 · tracks Quesen engine v1.10.0 (+ TSC v2 agent firewall) · backward compatible with every deployed engine version.

---

Install

npm i quesen-sdk           # or: yarn add quesen-sdk / bun add quesen-sdk

---

30-second agent firewall (copy-paste, no signup)

QuesenFirewall.sandbox() self-serves a free sandbox key (no signup, no card), so this runs as-is against the hosted engine:

import { QuesenFirewall, TscBlockedError } from "quesen-sdk";

const fw = await QuesenFirewall.sandbox("https://web-production-3df26.up.railway.app");

try {
  await fw.requirePass({
    agent: "my-agent",
    action: "send_data",
    target: "https://paste.evil.example",
    dataClass: "secret",
  });
  await sendTheData();                    // only runs on an explicit PASS
} catch (e) {
  if (e instanceof TscBlockedError) {
    console.log(e.decision.decision);                   // 'BLOCK'
    console.log(e.decision.reasons.map((r) => r.code)); // ['EGRESS_SECRET_UNTRUSTED']
    console.log(e.decision.commit_sha);                 // audit-receipt ruleset pin
  }
}

// A safe action returns PASS:
const ok = await fw.check({ agent: "my-agent", action: "tool_call", capabilityClass: "read" });
console.log(ok.decision);                 // 'PASS'

The hosted engine requires a key (no open mode). QuesenFirewall.sandbox() and QuesenClient.createSandboxKey() both call POST /sandbox/keys for a free, rate-limited key. For production volume pass apiKey: "sk_live_...".

---

Feature surface

• QuesenFirewall.sandbox(baseUrl) — zero-config agent firewall (mints a free key).
• .createSandboxKey() — self-serve a free sandbox key; auto-applied to the client.
• .health() — liveness probe.
• .version() — engine + report_schema versions + weights + thresholds + feature flags.
• .validate(input) — the main decision endpoint. Response carries input_snapshot_hash + commit_sha against v1.10+ engines.
• .simulate(input) — counterfactual scoring with weights_override / thresholds_override.
• .report(input) — post-decision outcome feedback (v1.1 schema with realized_pnl, venue, etc.).

Agent Firewall (TSC v2) — lower-level

TSC v2 turns Quesen into a deterministic agent firewall: describe what your autonomous agent is about to do and get a PASS / REVIEW / BLOCK / SKIP verdict plus a tamper-evident audit receipt — before the action crosses a trust boundary.

Requires an engine running with QUESEN_TSC_V2_ENABLED=true (POST /tsc/validate).

import { QuesenClient, dataEgressContext, requirePass, TscBlockedError } from "quesen-sdk";

const q = new QuesenClient({ baseUrl: "https://web-production-3df26.up.railway.app" });
await q.createSandboxKey();               // free key (or pass apiKey: "sk_live_...")

// Agent is about to POST data somewhere — ask Quesen first.
const decision = await q.validateTsc(
  dataEgressContext({
    dataClasses: ["secret"],            // what's leaving
    to: "https://paste.evil.example",   // where it's going
    destinationTrust: "unverified",
    framework: "langchain",
  }),
);

console.log(decision.decision);                  // 'BLOCK'
console.log(decision.reasons.map((r) => r.code)); // ['EGRESS_SECRET_UNTRUSTED']

try {
  requirePass(decision);              // throws unless PASS
  await runTheTool();                 // only reached on PASS
} catch (e) {
  if (e instanceof TscBlockedError) stopAndLog(e.decision);
}

Builders: dataEgressContext(...), toolCallContext(...), paymentContext(...). Full Typed Security Context schema. Runnable demo: examples/agent_firewall.ts.

Offline verdict replay — recompute without trusting us (v0.6.0)

Recompute the decision on the caller's machine — zero network, zero dependency (Node's built-in crypto only) — and assert a live receipt is independently reproducible. This is the JS-ecosystem answer to the adoption blocker serious integrators kept raising.

import { replay, verifyReceipt, dataEgressContext, QuesenFirewall } from "quesen-sdk";

const local = replay(dataEgressContext({ dataClasses: ["secret"], to: "https://x" }));
// { ok: true, decision: "BLOCK", reason_codes: ["EGRESS_SECRET_UNTRUSTED"], input_snapshot_hash: "<sha256>" }

const fw = await QuesenFirewall.sandbox("https://<engine>");
const ctx = dataEgressContext({ dataClasses: ["secret"], to: "https://x" });
const decision = await fw.check({ action: "send_data", target: "https://x", dataClass: "secret" });
const v = verifyReceipt(decision, { recomputeRequest: ctx });
v.recomputed; // true -> decision + reasons + input_snapshot_hash reproduced offline

Byte-for-byte identical to the Python SDK and the engine (a parity test binds the reference to the engine). Reproduces the contract-level decision/reasons/hash for the egress/authority subset — not the production risk weighting/thresholds.

Receipt provenance (v1.10, tracked in SDK v0.2.0)

const verdict = await q.validate({
  domain_age_days: 1,
  engagement_ratio: 0.95,
  scam_keyword_count: 4,
});

console.log(verdict.input_snapshot_hash);
// e.g. "2b0a…" — 64-char lowercase SHA-256 hex over canonical-JSON of the request
//                 (with client_request_id excluded from hash material)
console.log(verdict.commit_sha);
// e.g. "0b77cf…" — 40-char lowercase git SHA of Shxnque/quesen HEAD at decision
//                    time, or the sentinel "unknown"

Client-side reconstruction (verify the engine evaluated exactly what you sent):

async function inputSnapshotHash(payload: Record<string, unknown>): Promise<string> {
  const toHash: Record<string, unknown> = {};
  const keys = Object.keys(payload).sort();
  for (const k of keys) {
    if (k === "client_request_id") continue;
    if (payload[k] === null || payload[k] === undefined) continue;
    toHash[k] = payload[k];
  }
  const canonical = JSON.stringify(toHash);
  const bytes = new TextEncoder().encode(canonical);
  const digest = await crypto.subtle.digest("SHA-256", bytes);
  return [...new Uint8Array(digest)]
    .map((b) => b.toString(16).padStart(2, "0"))
    .join("");
}

Backward compatibility. Both fields are typed as string | undefined. Against a pre-v1.10 engine they simply won't be set; type-checking stays clean.

v1.5.0 on-chain enrichment (optional)

Pass chain + contract_address to enable deterministic on-chain enrichment. Requires the Quesen server to be started with QUESEN_ONCHAIN_ENABLED=true and per-chain QUESEN_ONCHAIN_RPC_<SLUG> set.

const verdict = await q.validate({
  engagement_ratio: 0.9,
  chain: "base",
  contract_address: "0x4200000000000000000000000000000000000006",
});

if (verdict.onchain_enrichment?.holder_concentration.top1_share ?? 0 > 0.6) {
  // Read the enrichment on the client side too if you want extra logging.
}

---

Errors

All errors extend QuesenError:

Class	Trigger
QuesenAuthError	401 (missing/invalid API key)
QuesenValidationError	422 (input shape)
QuesenRateLimitError	429 (per-key rate limit)
QuesenServerError	5xx after retries
QuesenTimeout	timeoutMs exceeded
QuesenTransportError	network failure

5xx errors are retried with exponential backoff (default 2 retries). Business errors (401 / 422 / 429) surface immediately.

---

Doctrine

This SDK is bound by Quesen's published design principles (see Shxnque/quesen):

• §2 determinism — never adds randomness, never adds an LLM in the loop.
• §11 ecosystem neutrality — zero runtime dependencies.
• §12 anti-bureaucracy — one client, one file, one intent.
• Receipt provenance forwarded — input_snapshot_hash + commit_sha typed on ValidateResult (v0.2.0+).

---

Development

cd sdks/js
yarn install
yarn test           # vitest
yarn build          # emit dist/ (tsc)