What we shipped.

Every recipe install now records pass/fail in the DB. Form refuses to take a customer's money for a modpack whose last 3 install attempts failed — even if the static recipe still says 'verified'. Nightly sweep tests 3 stale recipes per night against fresh VMs.

Stripe webhook + signup forms now enqueue provision jobs in a DB-backed queue instead of fire-and-forget background calls. A panel restart mid-provision is recoverable; nothing strands the customer on a half-built server.

Parallel client-side fetch against Modrinth + CurseForge — results trickle in as they arrive instead of waiting for the slowest source. FTB + ATLauncher are behind a 'Show more sources' opt-in.

Mark @aws-sdk/client-s3 and ssh2-sftp-client as runtime require() so Next.js stops trying to bundle them. Unblocks the off-site backup deploy.

Live TPS / RAM / cold-start data for ATM10, Prominence II, Better MC and friends. Numbers anyone can verify rather than 'lag-free' marketing copy.

/r/{slug} short-link redirects with first-party click attribution so modpack authors can drive traffic without third-party trackers.

Real customers, real numbers — TPS-before / TPS-after, mods running, players online. Placeholder slots ready for first wave of approvals.

Public partner page for modpack authors — revenue share, hosted demo servers, pre-configured one-click installs of their pack.

Who actually runs CoalHost, where we're based, what we believe about reliability. Hosting is a trust category and faceless support emails don't build that trust.

Console and mobile players can join Java servers with no client mods. Toggle on, share the address, done.

Side-by-side pricing vs Apex, Bisect and Shockbyte — RAM, player slots, hidden fees made visible. No 'tier locked' tricks.

Deep-dive at /ai-reliability — what the predictor / leak-detector / pattern-matcher actually catch, with screenshots from real customer crashes.

Premium VM backups now restore into a sandbox VM every month and pass an end-to-end smoke check. Theatre-grade backups don't count.

Long-window TPS retention so 'stable 20 TPS' claims can be verified across a month, not a five-minute snapshot.

Every crash that triggers a restart shows up on the public status page. We don't quietly memory-hole regressions.

Server-list MOTD now carries live TPS + uptime so players see real numbers before they connect, not just a server name.

Long-window TPS graphs make stability claims falsifiable across a real time window, not just a snapshot screenshot.

When the modpack you're running gets a new release on Modrinth / CurseForge, we email you with the changelog and a one-click upgrade button.

Every customer gets a shareable link showing their server's live TPS, uptime and incidents. Useful for answering player questions without screenshots.

One toggle installs Lithium / Krypton-style perf mods — only the safe ones, only on packs they're known compatible with.

Backups that never restore are theatre. A random backup is restored into a sandbox container every week and smoke-tested end-to-end.

Backups copy to an external object store so a full Hetzner outage doesn't take customer worlds with it.

Real uptime per service over the last 30 days, no hand-edited 'all systems operational' green dot. Per-region health included.

Parses every modpack's mods/ folder before install and blocks the known-bad combinations that crash on first boot. No more failed deploys at 2am.

A bot logs in and walks the spawn chunks before peak hours, warming JIT and chunk cache. Same bot does daily reachability checks across the fleet.

Windows console window hidden, classic Tauri polish pass — smaller binary, cleaner first-run.

Watches tick time vs simulation-distance under live load and lowers automatically when TPS approaches the safe floor.

Spark is installed on every Java server. We sample TPS dips, hand the report to an LLM, and surface concrete config / mod changes.

Detects slow heap-growth signatures, captures a JVM heap dump, then runs an AI diagnosis to identify the offending mod or plugin.

Launcher release binaries mirrored to our own CDN so anonymous users can download without a GitHub redirect dance.

Smaller fixes on the flock-locked deploy.sh — clearer error output and a few edge cases on the systemd restart path.

deploy.sh now refuses to push if HEAD is dirty, eliminating the 'someone deployed their WIP' class of incident.

Watches heap gradient + GC frequency and warns ~30 min before an out-of-memory crash. Pairs with a restart-during-idle window to defragment the heap.

Fixed the launcher landing page download URLs that pointed at a stale release tag — anonymous downloads work again.

Refactor: server-detail pages now use the AppShell layout and the new ServerSubNav tab strip. One layout, fewer duplicated sidebars.

Introduced a single AppShell layout across non-server pages so the nav, sidebar and footer come from one place instead of six.

Replaced the marketing hero on the authenticated dashboard with a compact header. Customers wanted faster access to their servers, not a re-pitch.