ADR-200: nikaia serve Is Cut, And The MVP Is The Application Server Alone

Status: Superseded by ADR-289 — its D1-D3 are D12, D13 and D16 there Date: September 23, 2026 Target Version: Nikaia 0.0.166 Narrows: ADR-194 D2 (the first of its two products) and D3’s operator half; ADR-038 §4.5’s order loses one step Related: ADR-194 D5 (the MVP’s protocol, which this leaves whole), ADR-069 D1 (http is a package), ADR-018 (what a handler is given and gives back), ADR-198 (the socket), ADR-199 (the readiness layer under it)

1. Context

ADR-194 answered who decides what is exposed by finding the question false: there are two products, and each is right about itself. The standard server — nikaia serve [dir] — has no author present at run time, so everything about it is the operator’s; an application that serves has an author who decided already.

That reasoning is not withdrawn. What has changed is the project’s own position: the MVP is behind schedule, and one of the two products is worth much less than the other.

2. Decisions

D1 — nikaia serve is cut.

There is no serve subcommand, and the standard file server is not a thing this project is going to build. The record that decided it stands; what is cut is the product, by the owner, for two reasons stated together:

• Schedule. The MVP is late, and serve is the half of D2 that nothing else waits on. Every record listed under #90 — ADR-018 entire, ADR-058’s body rows, the roadmap’s route hashing — waits on the application server. Not one of them waits on a file server.
• Value. python -m http.server is the thing it would have imitated, and the people who reach for that already have it. A Nikaia one would be a second way to serve a directory, and what it would demonstrate about this language is the part it has least of: no handler, no routing, no ledger, no provenance — a command that reads files and writes them out.

A cut is written down, which is the whole of why this record exists. The alternative was to leave step 4 on a list where it would be read as not yet for as long as the list exists, and a step nobody intends to take is worse than an absent one: it makes the list wrong about what is coming rather than merely incomplete.

D2 — So the caps are the program’s, and there is no operator half.

ADR-194 D5 ends the limits are the operator’s in the first product and the program’s in the second. With the first product gone, the second clause is the whole sentence: body_cap, head_cap, head_wait and connections are options on http::listen, with defaults that are a server rather than a denial of service, and a program that wants others names only what it changes.

The same subtraction reaches D3. It binds to localhost, and anything wider is asked for was written for both products, and localhost was the fail-closed default because a standard server’s author is not there to choose. An application’s author is: http::listen(at) takes the address, the program says which, and there is no default to make safe. What D3 leaves behind is the --trust listing of a wide bind, which it named as belonging to ADR-108’s shape rather than to itself, and that is unaffected — a program that binds every interface is a thing a reviewer may want listed whether or not a serve command exists.

D3 — What replaces it is examples/hello-http/, and it is a real server.

The demonstration a reader needs is one address and one function, and that is what the example is: http::listen(at) fn(request) { … }, with the handler deciding. It used to render one response and print it, because there was nothing to hand a response to.

It is driven by a test over a real socket (crates/nikaia/tests/project.rs): port 0, the bound address read back off the line the server prints, and then the two methods, a body by Content-Length and each of the refusals asked of it in turn. That test is what serve would also have needed, and it is spent once here instead of twice.

3. Consequences

• ADR-038 §4.5’s order is four steps and not five. A socket layer that keeps registrations (ADR-199), a minimal HTTP/1.1 server (this version), the parsing moved into Nikaia (ADR-196 D2’s route), then rustls, then HTTP/2. serve was never in that order; it was ADR-194 D2’s other half, and it leaves with this record.
• ADR-194 is not rewritten. A record says what was decided when it was decided; the superseded and narrowed table in README.md is where a reader of D2 is sent.
• Nothing about the language changes, and nothing about std. The socket, the readiness layer and http1 were all built for the application server and are untouched.
• What it costs: the python -m http.server ergonomics of the original ask are not answered by this project. That was one of two things the ask wanted and it is the one being given up, said plainly rather than deferred.

4. What this does not decide

• Whether a route may be refused by what its handler touches. ADR-194 §4 carries it, the owner asked to be asked again with a fuller write-up when the work reached it, and the work has now reached it. It is a question and not a cut.
• Keep-alive, chunked bodies, TLS and HTTP/2, each a later step of the order above and none of them touched by this.
• Whether some other command ever serves a directory. This cuts the product from the MVP; it does not reserve a promise against the day somebody wants one for a different reason.

This site is open source. Improve this page.