The ordering is what decides whether the rest of your server ever executes. kubectl will not guess at what a server supports, so it asks GET /api and GET /apis first, requesting aggregated discovery from apidiscovery.k8s.io and falling back to plain application/json if that is not offered [4]. Against a server that 404s, the author's capture shows five attempts to each endpoint and then a clean give-up [5]. That is ten requests, none of them for a pod [6]. The PodList handler you are proudest of is dead code until two near-empty documents exist [7].
Then the content type. Once discovery succeeds, `kubectl get pods` sends Accept: application/json;as=Table;v=v1;g=meta.k8s.io before it will accept plain JSON [8]. The server answers with column definitions and rows of already-rendered cells, and the client prints them nearly verbatim, which means the NAME READY STATUS RESTARTS AGE line is a string the API server picked [9]. Sixty lines of Node that answer the three discovery paths and return a one-row Table are enough to make a real kubectl print NAME and demo-pod [10]. A control plane with no concept of a pod can still display one convincingly, which is worth knowing before you treat kubectl output as evidence.
The author calls the design genuinely good, and the reason holds up: display logic lives on the server, so `kubectl get` renders resource types the binary has never heard of, including CRDs that shipped after it was built [11]. The bill goes to whoever writes the server. Around 55 kinds in that repo each need their own table() with their own columns, and every column carries a type, a format, a priority and a description [12]. Pods alone define five columns, so twenty metadata values for one kind's list view [15]. Streaming adds a state requirement: column definitions belong on the first watch event only, because repeating them on each update makes the client re-print headers [13].
What would have to be true for these captures to transfer. They were taken against kubectl v1.34.1 [3], so the five retries and the exact Accept lists are properties of one client build, not constants of the protocol. What generalises is the shape: discovery is negotiated, list output is negotiated, and both happen before any resource logic runs. The demonstration also covers `kubectl get pods` and nothing else [14]. Treat the sixty-line stub as a proof about one verb.
The repo itself is a Node reimplementation of the core APIs on MongoDB instead of etcd, running pods as sibling Docker containers [1]. Its author's own read on why strangers finally looked is that the README made a claim you can falsify in one command, while his earlier repos led with how they were built and went unread [2].
