Imported from outaTiME/ambito-dolar (packages/backend/AGENTS.md). Install upstream with npx skills add outaTiME/ambito-dolar --skill backend. Copyright stays with the author.

Backend

Rules for packages/backend, loaded on top of the root AGENTS.md when working here.

SST v4, Lambda handlers.

• API handlers go through Shared.wrapHandler(...) and answer with Shared.serviceResponse(...).
• packages/backend/src/routes/test.js is an API endpoint, not a test. There is no test script here.
• Public response shapes are contracts: the client, the website and both widget implementations read them. A /fetch change reaches the ios widgets first, see packages/client/docs/android-widgets.md.
• is_open in quotes.json drives how often the client asks for rates. Dropping it or changing its meaning fails silently, see docs/product-policies.md. It must not go into /fetch while storeFetchJsonObject is gated on is_updated: a gated write leaves it stale on exactly the close and the open, the two runs that move it without new rates. Anything that writes quotes.json outside Process has to carry the flag over, packages/backend/src/routes/update-rates.js reads the stored one for that.
• puppeteer-core and @sparticuz/chromium move together or not at all, both only used in packages/backend/src/libs/chrome.js. A mismatch fails when the social lambda launches Chrome, which surfaces as a missing post and not as an error. The chromium README prescribes the check: match the Chrome major that puppeteer ships, from https://pptr.dev/supported-browsers, to the chromium package major.
• One timeout in packages/core and no per call override. Every source was measured before it was set and only one is above a second, so a second number would be a guess. A success body read with .then(r => r.json()) is outside that timer, and a timeout is never retried.
• ky skips POST and PATCH, so those wrap in promiseRetry instead: every social target, the nested one in social/instagram.js that waits out a media still processing, and getBusinessDay. Retrying a publish can repeat it, and with lambda retries off in infra/defaults.ts that is the only way a post goes out twice.
• Every lambda timeout in infra/ is twice its observed duration and says so. Move the pair together.
• What survives a slow source is the catch, not the timeout. getRate catches and getObjectRates compacts the hole, so one slow source costs one rate. getBusinessDay throws uncaught and costs the whole cycle. A lambda killed after storeRatesJsonObject leaves S3 written and the push unsent, which is worse than dying before it.
• Do not reorder process.js. storeRatesJsonObject -> updateRealtimeData -> notify is load bearing, a push landing before the board is written shows stale rates to the clients still on it.
• Remove updateInstantData before 2027-08-31, the day InstantDB cloud shuts down. It carries no timeout of its own and runs before notify(), so a hung sdk eats the one minute Process budget and stops push and socials for everyone.