Testing

An honest account of what test coverage exists, what does not, and what to actually run before shipping.

What exists

Suite	Location	Count	Runs without a chain?
Autopilot contract	bz-backend/contract-deployment/test/autopilot.test.cjs	61 passing	Yes — Hardhat local
Autopilot trigger engine	bz-backend/scripts/test-autopilot-engine.js	13 passing	Yes
Policy toolkit	bz-backend/scripts/test-policy.js	—	Yes
Wallet-auth digest	bz-backend/scripts/test-wallet-auth.js	—	Yes
Blog API	bz-backend/scripts/testBlogAPI.js	—	Needs a running backend

What does not exist

There is no backend test suite

npm test in bz-backend runs echo "Error: no test specified" && exit 1. There are no unit tests for controllers, services, the quote fan-out, the provider proxies, the DCA cron, or anything in the ecosystem layer.

Also absent:

Missing	Consequence
Frontend tests	App.test.tsx and setupTests.ts are Create React App leftovers. vite build is the only gate
bz-agent tests	Classification quality has no evaluation harness. The only signal is the non-fallback rate — which does not catch a message routed to the wrong intent
swap-sdk tests	26 aggregators, none covered. A provider silently breaking is invisible
Relayer contract tests	The contract handling every interactive trade on 22 chains has no test suite in the repo
DCA contract tests	Same
Integration tests	No end-to-end coverage of quote → build → sign → settle

The relayer and DCA contracts have neither tests nor an adversarial audit

Autopilot got Slither plus an adversarial review (which found a critical drain vector). The relayer — which handles every swap on 22 chains — has had a hardening pass but no equivalent review, and no test suite. That is the largest coverage gap in the platform.

Before shipping — the practical checklist

Any change

# frontend compiles?
cd defi-dex && npm run build:app

# agent types check?
cd bz-agent && npm run check-types

npx tsc --noEmit in defi-dex is not a gate

It reports roughly 3,300 pre-existing errors originating in node_modules type definitions. vite build is the real gate. Do not chase tsc to zero as part of an unrelated change.

Contract change

cd bz-backend/contract-deployment
npx hardhat compile
npx hardhat test test/autopilot.test.cjs      # must stay at 61

# storage layout — MANDATORY for any upgrade
npx hardhat run scripts/simulate-upgrade.js --network <net>

Never upgrade a proxy without validating the storage layout

upgrades.validateUpgrade(V1 → V2) was run for the DCA upgrade and state was preserved on mainnet. A layout break on a live vault corrupts every user's position irrecoverably. simulate-upgrade.js and simulate-migration-upgrade.js exist for this.

Autopilot change

cd bz-backend
node scripts/test-policy.js
node scripts/test-wallet-auth.js
node scripts/test-autopilot-engine.js

Plus the hash-parity check, which is the highest-value single test in the codebase:

Verify the FE/BE hash parity manually

// backend
import { hashPolicy } from './utils/policy.js'; console.log(hashPolicy(policy))

// frontend browser console
import { hashPolicy } from './utils/autopilotPolicy'; console.log(hashPolicy(policy))

They must be identical strings. If not, the mirror has drifted and every Autopilot creation will fail.

swap-sdk change

There are no tests, so verification is manual:

□ Quote appears in the SSE stream on the target chain
□ Output amount and units are correct vs the provider's own UI
□ usdAmount populated (else the fee display shows $0)
□ allowanceTo is the real spender
□ sortQuotes gas estimate passes
□ A real trade completes and output arrives
□ getTxStatus resolves pending → success
□ Killing the provider's API does not break the stream

AI intent change

□ 5 different phrasings all classify to your intent
□ A neighbouring intent's phrasings do NOT hit yours
□ Missing a required field produces a question, not a guess
□ A hallucinated token name degrades gracefully
□ The data field exists in the Mongoose schema — reload and check the card
□ Old history still renders (optional chaining)

SEO change

cd defi-dex && npm run build && npm run check:routes

□ dist/{path}/index.html contains real content, not the empty shell
□ Exactly one FAQPage per page
□ Canonical trailing slash matches the sitemap loc
□ dist/og/{key}.png generated
□ The footer renders in the static HTML with JS off

One-liner for the FAQPage audit: walk dist/, parse every ld+json block, count FAQPage per page. More than one means the dedupe rules were broken.

Testing against production data

Local dev points at production Atlas

There is no staging database. Set CRON=false before starting the backend, or your laptop becomes a second CSIP and Autopilot executor using the production PRIVATE_KEY_EXECUTOR.

Useful read-only scripts:

cd bz-backend
node scripts/checkRelayerDeployed.js       # relayer presence per chain
node scripts/check-contract-owner.js       # owner/signer verification
node scripts/check-contract-state.js       # fee config, pause state
node scripts/checkDeployerBalance.js       # executor gas
node scripts/blazai-analytics.js           # chat usage figures

Many scripts in scripts/ and at the repo root are mutations

curate-token-order.js, enableChain.js, enableRelayerFees.js, and roughly 30 files at the repo root (fix_admin_and_txs.js, delete_remaining_anomalies.js, migrate_exempt.js, …). Several delete data and are not idempotent — they were written for a specific incident. Read before running.

Live end-to-end testing

The pattern used for Autopilot: a small real trade on mainnet.

□ ~$5 of value
□ Watch the transaction on the explorer
□ Confirm the measured delta credited matches
□ Confirm the fee accrued (check /defi/fees/summary?force=1)
□ Confirm the record in Mongo
□ Confirm the UI reflects it

The full Autopilot loop was proven this way on 2026-07-03: a probed route executed, credited the exact balance delta (0.000866… ETH), autoClaimFor delivered to the user's wallet, and platformEarnings accrued its first 0.00003 ETH.

Testnets exist but are thin

AUTOPILOT_ADDRESSES includes Arbitrum Sepolia (0x7cA5A09E…) and hardhat.config.js has arbitrum_sepolia. But there is no aggregator liquidity on testnets, so a route cannot be quoted — which means the interesting half of the flow cannot be tested there. A tiny mainnet trade is more informative than a testnet one.

Browser-driven verification

Several bugs were found this way and are hard to catch otherwise:

Bug	Found by
?q= left in the URL with an empty composer	CDP testing
The vendor-react split being TDZ-safe	Headless Chrome on /defi/swap and /
~96 px dead space on the AI site	Visual inspection at multiple widths

Headless Chrome on macOS clamps --window-size to ~500 px wide

So mobile screenshots look like overflow bugs when they are not. Verify at 500 px or above.

What to add first, if you get time

Ranked by risk reduction per hour:

1. Relayer contract tests. It handles every trade on 22 chains and has none.
2. An adversarial audit of the relayer and DCA contracts. Autopilot's found a critical issue; there is no reason to assume these are cleaner.
3. Per-provider quote smoke tests in swap-sdk. A broken provider currently disappears silently — see monitoring.
4. An intent-classification eval set. 50 labelled messages, run on every prompt change. The current signal cannot detect mis-classification.
5. A hash-parity test in CI. Automate the manual check above; the failure mode it prevents is total.
6. Contract tests for the DCA vault, particularly around the V1 accounting legacy.