ServerKit is a lightweight, modern server control panel for managing web applications, databases, and services on your VPS or dedicated server. Built with Python/Flask backend and React frontend.
Python astracloudflarecontrol-panelcyberpaneldnsfablehosting-control-panelhosting-panelkimi-k3server-managementserverkitubuntu-web-panelwebserver
docs(roadmap): record safe deploys, fleet health and extension trust
Juan Denis67884e2483cd
diff --git a/ROADMAP.md b/ROADMAP.mdindex 0ab523b7..a78973e7 100644--- a/ROADMAP.md+++ b/ROADMAP.md@@ -18,6 +18,9 @@ This document outlines the development roadmap for ServerKit. Features are organ- **Extensions Platform Maturation** - Marketplace registry backed by standalone extension repos (Kubernetes, Automations, Analytics, Mail, GPU, CrowdSec, DNS server, and more), with built-in features progressively extracted to keep the default panel lean. Installed extensions that ship a prebuilt bundle now render their UI on a production panel with no rebuild.- **Privacy-First Web Analytics** - Cookieless, self-hosted analytics extension with WordPress and nginx integration and optional geo lookup.- **Agent Fleet & Peering** - ServerKit-to-ServerKit peering (embedded agent mode), fleet doctor tooling, and agent capability upgrades.+- **Honest Extension Trust** - Installed extensions get a Permissions view showing what their manifest declared against what the panel has actually observed them do — including calls it refused. It is deliberately candid about its own limits: for capabilities the panel cannot mediate it says so outright rather than showing a reassuring zero, because an extension runs inside the panel process and no permission string changes that. The Python packages an extension shipped that the panel declined to install (they are opt-in, since pip runs with the backend's privileges) are now listed with the file itself and the exact setting to change, instead of being mentioned only in an install log nobody reads.+- **A Failed Deploy No Longer Takes the Site Down** - Redeploys used to stop the running version first and only then discover that the new one could not start: a broken compose file, an image that no longer pulls, or a port something else had taken left the site down until someone intervened. Every deploy path now proves the new version can run *before* anything live is stopped — the compose file is validated, images are built and pulled to completion, and host port conflicts are checked — and a failure aborts with the running deployment untouched and a log line saying so. On single-container apps the registry pull in particular used to happen *after* the old container was already removed, so a registry hiccup left the app with nothing running at all.+- **Fleet Health & Working Repair Buttons** - The health doctor now sweeps connected servers as well as the panel's own host, with per-server findings that offer the same one-click repairs, an offline server reported as offline rather than missing, and a daily fleet-wide sweep. The backup checks' repair buttons now actually work: a restore drill can be started from the doctor and tracked as a job, and an unverified backup can be verified in place or against its offsite copy, whichever it has.- **Secure by Default on a Fresh Install** - A stock Debian/Ubuntu install now ends with a firewall instead of a warning: UFW is installed and enabled with default-deny inbound, allowing only the ports sshd actually answers on plus 80/443 and the panel's own port. The SSH port is discovered from the effective sshd config rather than assumed to be 22, and if it cannot be determined the installer refuses to enable anything rather than risk locking you out. fail2ban with an SSH jail moved from the "full" profile into the standard one, watching the same port. Setup Health gained firewall and brute-force-protection checks, surfaced on the dashboard, so a box that skipped either says so instead of looking fine.- **Auth & Proxy Hardening** - Trusted reverse-proxy client-IP handling and per-IP login brute-force throttling.- **Certificate & DNS Reliability** - Certificates stop failing for reasons the panel could see and didn't say. certbot is found wherever it is installed, including the snap location its own docs recommend, and requesting a certificate checks the nginx plugin first and names the missing package instead of passing certbot's internal confusion through. A stray AAAA record — the classic case where the IPv4 record is perfect and issuance still fails, because Let's Encrypt prefers IPv6 — is now detected and explained by both domain verification and the doctor. One Let's Encrypt contact address is remembered panel-wide rather than re-typed for every certificate, and the doctor catches an nginx that is running but serving a config it can no longer parse.