The task
dreamDEX, a perp exchange on the Somnia network, opened a brief with two tracks: reports on problems found, paid per accepted report, and an MVP, a useful tool on top of the exchange. For us it was income plus an audit case for a DeFi product, a service that can be sold to other exchanges.
By the time we came in, the brief already had 447 reports from others, so the obvious had long been found. We had to look for what others had missed and prove every finding with a transaction hash, not a description.
Why, and what the idea was
The key finding from live tests in the app: a trader sets TP/SL in ‘Entire’ mode, meaning the whole position, then adds to the position, and the app keeps saying ‘Entire’. When the price reaches the stop, only the old size closes and the rest is left unprotected. The loss estimate in the dialog is also computed on the old size, so it is nearly impossible to notice until it is too late.
Then came a fork: build the MVP on our own finding, or make yet another ‘guardian’ bot, of which the brief already had five. We ran the decision through a council of five AI advisors and went with our finding. That is how StopCheck came about: paste a wallet address, sign nothing, the page reads the chain and shows what percentage of the position is really covered by a stop. Next to it, a bot that closes the uncovered remainder when the stop fires or the price reaches it.
A second council went over the whole plan just before publishing: soften the posts, never display other people’s positions, add a shareable result card and a 50-second video.
How it went
| When | What happened |
|---|---|
| Day 1, afternoon | Registration, whitelist ticket, first reports via the SDK and bot kit |
| Day 1, evening | Review of 447 reports by others: 8 of our topics had been filed earlier. New findings cross-checked before writing up |
| Day 1, evening | Live tests in the app: reports 019, 020, 022 found |
| Day 1, night | Council of 5 AI advisors: build the MVP on our own finding, not a sixth ‘guardian’ |
| Day 1, night | StopCheck: public repository, page, bot, live demo with hashes, scan of 463 accounts |
| Day 1, night | Second council on the whole plan: softer posts, no other people’s positions, result card, 50 s video |
The whole job fit into one day.
Tools and method
Reading: the @somnia-chain/markets-sdk SDK; positions and stop sizes come from the chain, the indexer only hints where to look. The page is static on GitHub Pages, asks for no keys and signs nothing. Built with esbuild, TypeScript and viem.
Bot: a linked key (the exchange’s own ‘Link a bot’ feature) and perp-core from the official bot kit, closing via placeOrderFor. The protocol does not let a bot place stops on the owner’s behalf, so the bot watches on its own. It tells a triggered stop from a cancelled one by the status in the indexer.
Reports: 21 of them across the app, SDK, bot kit, docs and contracts, each reproducible with transaction hashes. They live in a separate repository, dreamdex-perps-feedback. The 50-second video was assembled in ffmpeg from real screenshots captured in headless Chrome.
Result
The page at dmitriytopka.github.io/stopcheck: paste an address, get a card saying ‘the stop covers 50% of the position’, a list of problems and coverage bars per position, with Copy link and Share on X buttons. The code is open: github.com/DmitriyTopka/stopcheck.
A live demo on chain, both transactions linked in the README: the stop closed half the position at 19:34:35, the bot closed the other half at 19:34:38. Without the bot, half the position would have stayed open with no stop.
A scan of 463 exchange accounts: 19 with a partial stop loss, from 6% to 99.6%; 129 with stops but no position; 6.19 STT waiting to be claimed across 30 accounts. We do not publish other people’s addresses, that was a council decision.
Infrastructure cost $0: GitHub Pages and our own server. Testnet gas: the bot spent about 0.013 STT. The weak spots, stated plainly: no users yet, the bot needs its own server, and if the exchange fixes finding 019 the bot will be needed less often, while the checker stays useful.
Mistakes and what we took from them
