A protocol you did not author is not yours to improve

Interoperating with someone else's implementation inverts the usual instincts. The names are not descriptive enough, the numbering has gaps, half the messages are ones you draw no UI for — and every one of those is load-bearing. A renamed field or a reordered number is not a refactor; it is a client that connects successfully and silently cannot join anything.

Write that down where the schema lives, because it is the only defence against a well-meant cleanup:

/*
 * Source of truth: <the schema you are implementing>. Every field number below is that file's
 * number and every constant is its string spelled exactly. NOTHING here may be "improved": a
 * renamed field or a reordered number is a client that silently cannot join.
 */

Traps

The constants the schema does not state are the ones that will cost you a day. A schema describes message shapes; it frequently does not name the envelope type strings that carry them, the ordering requirements, or the size thresholds. Those live only in the counterpart's source. Read them there and cite the file and symbol you read them from in a comment — a guessed type string is answered with an "unknown message type" error, so the socket stays open, frames keep flowing, and the handshake simply never completes. That failure looks like a network problem and is not one.

Decode an unrecognised message type to null instead of throwing. The other implementation will ship new types before you do. A client that dies on an unknown frame cannot share a session with a newer peer, and the failure arrives as "it worked until someone updated". The dispatch's else branch returns null, the layer above forwards the type with a null payload, and nothing breaks.

Parse message types you have no feature for. A text-message frame in a client with no such UI still has to travel through the codec without disturbing the stream, because a peer running the fuller implementation will send one. "We don't support that feature" is a UI decision; it is not permission to reject the bytes.

Narrow a negotiated capability, never widen it. When the handshake says the peer cannot do something, turn it off locally. When it says the peer can, that is not an instruction to turn it on — your own thresholds and your own reasons still apply, and a capability you enable on the strength of the peer's claim is one you have no fallback for. One directional branch, no else:

// adapted — the capability is a placeholder
if (!peerCaps.supportsCompression) {
    Logger.w(TAG, "Peer reports no compression support — sending uncompressed")
    codec = MessageCodec(compressionEnabled = false)
}

Split the peer's errors into "about this message" and "about this connection" before reacting to either. They arrive on the same channel and look identical. A rate-limit or a malformed-payload error concerns one frame; treating it as a connection failure tears down a working session and, if it feeds a retry policy, burns the whole budget on a single bad send. A refusal of the client itself — an unsupported version, a rejected credential — is the opposite: retrying repeats it forever. Keep an explicit set of the non-recoverable codes and default everything else to recoverable, because the peer will add codes you have never seen.

Do not surface every peer error to the user, either. An error the user cannot act on — a rate-limit on a background message — becomes an alarming banner about nothing. Filter at the state machine, log the rest.

Field numbers and type strings are the contract, so pin them in tests rather than trusting review. A test that asserts the literal string of each handshake type, and one that asserts the raw tag bytes of the envelope, converts the entire class of "someone tidied the schema" into a build failure. Without them the first symptom is a user saying they cannot connect, and the change that caused it looks harmless in a diff. protobuf-without-codegen-kmp covers the byte-level half of this; api-ok-but-ignored is the neighbouring lesson that a successful-looking response is not proof of acceptance.

Keep the borrowed types behind a module boundary from the first commit. Because they are ordinary classes, they spread like ordinary classes — and once a screen imports one, the "nothing here may be improved" rule has to hold in a file nobody reads it in. response-to-domain-flow is the placement discipline.

Verifying it

The discipline is mostly about where strings live, which greps answer well:

grep -rnE 'const val [A-Z0-9_]+ = "[a-z][a-z0-9_]*"' --include='*.kt' . | grep -v '/build/' | cut -d: -f1 | sort | uniq -c | sort -rn | head
grep -rnE '[=!]= "[a-z][a-z0-9]*_[a-z0-9_]+"|setOf\("[a-z][a-z0-9]*_[a-z0-9_]+"' --include='*.kt' . | grep -v '/build/'
grep -rnE 'supports[A-Z][A-Za-z]*' --include='*.kt' . | grep -v '/build/' | grep -v Test

The first ranks files by how many lowercase wire literals they declare; the protocol's constants object should tower over everything else in that list, and a second file with a handful of them is either an unrelated preference key set or the beginning of a duplicate vocabulary. The second is the sharper check: it finds wire strings compared inline rather than referenced by name. Expect a small number of hits and read every one — in practice message types get constants while error codes stay inline literals in whichever file reacts to them, which is exactly the population that drifts apart between the send side and the receive side. The third shows the negotiation mixed in with every unrelated capability predicate in the tree — range support on an HTTP client, a loader's mime-type override, a query parameter — because that pattern is ordinary English rather than protocol vocabulary. Read only the hits within reach of the handshake: each of those should be a schema field or a narrowing branch, and a handshake branch that enables something because the peer supports it is the widening mistake. The rest look identical and are correct code.

Then read the schema file top to bottom once and check that every constant not present in the published schema carries a comment naming where it was read from. A constant with no provenance is a guess, and a guess in this file is the failure that presents as a connection that opens and does nothing.