1/ launching https://t.co/320ApbJWvA today.
one VPS. one Go binary. one SQLite file. continuous backup to R2. caddy for HTTPS. systemd to keep it alive. a shell script to deploy.
this is the whole stack. the rest is what each piece is doing and what it replaces.
๐งต
Probe your app's public /healthz from another machine. Check TLS and the expected body. Localhost can stay green while a broken certificate or misrouted proxy keeps users out.
https://t.co/DyhsKgLemo
@DanielLockyer I'd cancel superseded PR test runs before tuning the cache. Group by workflow and PR, with cancel-in-progress enabled, so each push doesn't leave the older run chewing through the suite.
@yongfook For a solo founder, I'd want the bot to bring back the buyer's exact objection. โToo expensiveโ and โI don't understand what this doesโ need different fixes, and you lose that feedback if all you get is a call count.
With Go's built-in slog handlers, passing ctx doesn't add a request ID. Add it as an attribute, then trigger a failed save and check that its log includes the ID you need to trace the request.
https://t.co/HojJVJkIme
@PovilasKorop I'd keep the form filled in after that redirect too. Being told when to retry is less helpful if the user has to type the whole submission again.
@housecor For a suspected bug, I'd ask the reviewer to include the input and the wrong result. A reviewer can rewrite an AI warning perfectly without checking whether it happens.
@mattpocockuk I'd test this with search filters that need to survive a copied URL. The constraints need to say whether the URL or Zustand owns those values, or the agent can follow the state rule and still create two sources of truth.
Logging the same save error in repository, service and handler produces three entries for one failure. Let the caller that knows the final outcome own the completion event.
https://t.co/HojJVJkIme
@garrytan For API-backed reports, I'd keep a redacted response fixture alongside the queries when data permissions allow it. That lets someone test the calculations against the original inputs after the live data changes.
For a Go save wrapped in a transaction, emit outcome=ok after Commit succeeds. A successful insert alone doesn't confirm the save. That keeps a failed commit from leaving a misleading success event.
https://t.co/HojJVJkIme
@DmytroKrasun When retiring a provider, I'd check webhooks and scheduled exports too. A forgotten nightly export can keep sending it customer data after the replacement ships.
@MaxHirsch13 For the free/trial experiments, I'd separate users who never got a useful result from those who did and still wouldn't pay. I'd check onboarding for the first group before changing the paywall again.
For Go logs, outcome=failed doesn't tell you what broke. Add error_kind from documented error types or codes, with unknown as the fallback. Avoid matching error-message text.
https://t.co/HojJVJkIme
@yongfook For the room-to-video workflow, I'd keep the furnished still as a reusable output. If the voiceover fails, regenerating the room is an expensive way to retry audio.
@yongfook A one-person business can outgrow free long before it needs SSO. I'd like more $30 plans that charge for a useful recurring job, like scheduled exports.
If every saved change needs an audit trail, write the audit row in the same database transaction as the change. Your Go process can die after commit and before the completion log runs.
https://t.co/HojJVJkIme
@Swizec Ten lines can change who gets billed. I'd make design review depend on the decision and rollback cost. A billing-rule change and a button-label fix shouldn't need the same paperwork.
Keep Go log messages constant, such as expense_save, and put request IDs and outcomes in attributes. That gives you one event name to search when a customer reports a failed save.
https://t.co/HojJVJkIme
For Go save logs, keep duration_ms numeric on both success and failure paths. Mixing 250, "250ms" and 0.25 seconds under one key means cleaning up your logs before you can compare slow saves.
https://t.co/HojJVJkIme