mirror of
https://github.com/zap-proto/zip.git
synced 2026-10-19 15:38:57 +00:00
Ten of eleven examples declared their surface with untyped handlers. An untyped route registers no op, so every one of them modelled the single pattern that leaves an endpoint out of the OpenAPI document, out of the MCP tool list, out of the CLI and out of the call plane — while the README two files away calls that registry the one source of truth. The examples are where the default is learned, and they were teaching the wrong one. hello, subsystem-mount, migrate-from-chi, migrate-from-beego, plugin/billing, plugin/host and internal/testplugin now declare typed ops. The migrate-from-* examples show the port as the two steps it is: mechanical first (same handler shape, one edit, and invisible to every projection), typed second, because stopping after the first leaves exactly the surface you were migrating away from. sse-streaming and websocket stay untyped and now say why in the file header: a typed op is one decode, one call, one serialize, and a stream has no single value to return while an upgrade has no response body at all. That is the whole list of reasons — a stream, an upgrade, a proxied byte range, a non-JSON body. express-in-zip registered two routes on app.Fiber(), BELOW zip entirely: no *zip.Ctx, no error handler, no middleware, no Authorizer, and no projection — served by the app but not governed by it, which is a security seam and not a shortcut. Both move onto the zip router. /runtime/:lang becomes a typed op, so "run this source in this language" is now in the document and is a tool an agent can call; the legacy JS wildcard stays a wildcard and says what that costs. testplugin's catch-all stays untyped for the reason it exists: it answers paths that do not exist, so a host mounting it at two prefixes can tell its own 404 from the plugin's. A wildcard is deliberately not one named endpoint with one schema. Every example was run and probed, not just built: 30 assertions across all eleven, including the openapi/mcp routes the typed ops now light up. Co-authored-by: Hanzo Dev <[email protected]>
47 lines
1.3 KiB
Go
47 lines
1.3 KiB
Go
// websocket example — chat-style echo server.
|
|
//
|
|
// go run ./examples/websocket
|
|
// wscat -c ws://localhost:8080/ws
|
|
//
|
|
// # Untyped on purpose — the request never completes
|
|
//
|
|
// An upgrade hands the connection to another protocol: there is no response
|
|
// body for a typed op's Out to be, and the exchange that follows is frames in
|
|
// both directions for as long as the peer stays. A typed op is one decode, one
|
|
// call, one serialize, so it cannot express this — which is why wsx.Upgrade
|
|
// takes an untyped handler and not a TypedHandler.
|
|
//
|
|
// The cost is the same one every untyped route pays: no OpenAPI entry, no MCP
|
|
// tool, not reachable by zip.Call. Anything with a request-and-response shape
|
|
// belongs in a typed op instead — see examples/hello.
|
|
package main
|
|
|
|
import (
|
|
"log"
|
|
|
|
"github.com/zap-proto/zip"
|
|
"github.com/zap-proto/zip/middleware"
|
|
"github.com/zap-proto/zip/wsx"
|
|
)
|
|
|
|
func main() {
|
|
app := zip.New(zip.Config{AppName: "websocket"})
|
|
app.Use(middleware.Recover(), middleware.RequestID())
|
|
|
|
app.Get("/ws", wsx.Upgrade(func(c *wsx.Conn) error {
|
|
log.Printf("ws connection from %s", c.RemoteAddr())
|
|
for {
|
|
mt, msg, err := c.ReadMessage()
|
|
if err != nil {
|
|
log.Printf("ws read: %v", err)
|
|
return err
|
|
}
|
|
if err := c.WriteMessage(mt, msg); err != nil {
|
|
return err
|
|
}
|
|
}
|
|
}))
|
|
|
|
log.Fatal(app.Listen("http://:8080"))
|
|
}
|