Major release to v1.0.0, aligned with the v1.0.0 framework. The MCP documented the wiring conventions but not the config half of a project, and its example main.go omitted the godotenv autoload — so an assistant starting a service from zero still hand-rolled main.go and the launcher, and got config wrong. This adds a first-class scaffold, completes the config/.env.example conventions, and adds rules that catch the "mess in main" pattern. internal/tools: - New get_scaffold: returns the canonical minimum application scaffold as ready-to-write files (main.go with godotenv autoload + wire.Run(), wire.go, a composed config.go, a health hook, .env.example), with import paths filled from a `module` argument. Registered in tools.go. internal/rules: - Three new validate_snippet rules, appended in scaffold_rules.go: main.dirty (launcher/components built in main instead of internal/wire), main.no-godotenv-autoload (a wire-convention main that never loads .env), and config.raw-getenv (an EINHERJAR_* var read via os.Getenv instead of composing the component Config; EINHERJAR_LOG_* stays with logz.direct-env-read). - scaffold_rules_test.go — internal/rules had no tests; asserts each new rule fires and that a clean main is not flagged. internal/index (builtins): - The synthetic wire module gains a Config section (compose the framework's component configs, load with caarlos0/env, APP_* app fields / EINHERJAR_* framework fields) and a Config & .env.example discipline (every env var the config reads is documented in .env.example, kept in lock-step). - main.go now shows the `_ "github.com/joho/godotenv/autoload"` blank import, previously omitted. The assembly file is renamed launcher.go -> wire.go. - Migrations and seeding removed from the documented scaffold — developer choices, not framework conventions. Re-synced against iron-dough-api / pei-api. Version: - Badge and serverVersion const were stale at v0.1.0; both now v1.0.0. Docs: - README (eleven tools, eleven validation rules) and CHANGELOG updated. No new dependencies. The wire conventions are embedded at build time (//go:embed builtins/README.md) and the new tool and rules are compiled in, so a deployment must be rebuilt to serve them; a server still running the v0.2.0 binary keeps serving the old conventions until redeployed. Reviewed-on: #2 Co-authored-by: Rene Nochebuena Guerrero <rene@nochebuena.dev> Co-committed-by: Rene Nochebuena Guerrero <rene@nochebuena.dev>
14 KiB
Wiring Conventions
Forging a service is mostly wiring. Do it the same way every time.
This is not an Einherjar module — it is the canonical application shape that uses
Einherjar modules. Apps live in their own repository with an internal/wire/ package that
mirrors this template. The conventions here are distilled from production services built on
Einherjar v1 (iron-dough-api, pei-api) and describe the one opinionated minimum a
scaffolded app should have. You can hand-roll something else — but then it is yours to
maintain, and it will not match what the rest of the ecosystem reads at a glance.
The opinionated minimum
Every scaffolded Einherjar application has, at minimum:
- A clean
main.go— nothing but.envautoload and a call towire.Run(). - An
internal/wire/package — one file per feature pluswire.go, which assembles everything. - An
internal/config/config.go— one globalConfigthat composes the framework's component configs, loaded from the environment withcaarlos0/env. - A
.env.examplekept in lock-step with that config (see Config & .env.example).
Anything a developer freely chooses — how migrations run, how the first admin is seeded, an init-by-endpoint/webhook/email flow — is not part of this convention and is left to the app.
Project layout
cmd/<app>/main.go one-line entrypoint: godotenv autoload + wire.Run()
internal/wire/wire.go Run() — loads config, builds infra, registers feature hooks
internal/wire/<feature>.go one file per feature, hosts a with<Feature> hook
internal/wire/middleware.go authz, skipPublicPaths, skipMethodPath helpers
internal/config/config.go global Config composing framework component configs
.env.example every env var the config reads, documented, in sync
internal/<feature>/dto/ request/response DTOs
internal/<feature>/handler/ HTTP handlers
internal/<feature>/repository/ data access
internal/<feature>/service/ domain logic
main.go
cmd/<app>/main.go contains nothing but the .env autoload and the call to wire.Run().
The blank import _ "github.com/joho/godotenv/autoload" is the standard, documented way to load
a local .env — it never overrides variables already set in the real environment, and a missing
file is not an error, so deployments (vars injected by the platform, no .env) are unaffected.
package main
import (
"fmt"
"os"
_ "github.com/joho/godotenv/autoload"
"myapp/internal/wire"
)
func main() {
if err := wire.Run(); err != nil {
fmt.Fprintln(os.Stderr, "fatal:", err)
os.Exit(1)
}
}
No config parsing, no component construction, no logging setup — all of that lives in
internal/wire/. A main.go that builds anything itself is the single most common scaffolding
mistake.
Config
internal/config/config.go is one Config struct that composes the framework's component
configs as nested fields, alongside the app's own settings. caarlos0/env recurses into the
nested fields, so each Einherjar component's EINHERJAR_* env tags load automatically next to
the app-owned fields. App-owned fields use the APP_* prefix so they never collide with the
framework's EINHERJAR_* namespace. There is exactly one Load().
package config
import (
"time"
"github.com/caarlos0/env/v11"
"code.nochebuena.dev/einherjar/db-postgres"
"code.nochebuena.dev/einherjar/web/server"
)
// JWTConfig is app-owned: the secret is handed to the signer in code, not consumed
// by a framework component, so it carries no EINHERJAR_ prefix.
type JWTConfig struct {
Secret string `env:"APP_JWT_SECRET,required,notEmpty"`
Issuer string `env:"APP_JWT_ISSUER" envDefault:"myapp"`
AccessTTL time.Duration `env:"APP_JWT_ACCESS_TTL" envDefault:"1h"`
RefreshTTL time.Duration `env:"APP_JWT_REFRESH_TTL" envDefault:"168h"`
}
// Config is the fully-resolved startup configuration. Einherjar component configs
// are nested fields; caarlos0/env recurses into them, populating their
// EINHERJAR_SERVER_* / EINHERJAR_PG_* tags from the environment.
type Config struct {
AppEnv string `env:"APP_ENV" envDefault:"local"`
CORSOrigins []string `env:"APP_CORS_ORIGINS" envSeparator:","`
JWT JWTConfig
// Framework component configs — composed verbatim. Their own EINHERJAR_* tags
// load through this one env.Parse call.
Server server.Config // EINHERJAR_SERVER_*
PG postgres.Config // EINHERJAR_PG_*
}
func Load() (Config, error) {
var cfg Config
if err := env.Parse(&cfg); err != nil {
return Config{}, err
}
return cfg, nil
}
Never read framework env vars (EINHERJAR_*) with os.Getenv — compose the component's Config
type and let caarlos0/env load it. A raw os.Getenv("EINHERJAR_PG_HOST") in application code is
the mistake this convention removes.
Config & .env.example
Every environment variable the config package reads must also appear in .env.example at
the repo root, documented. The two are kept in lock-step: when a feature introduces a new env
var, the same change adds its env:"..." tag to config and a documented line to
.env.example. This is not optional bookkeeping — it is what stops a long feature from shipping
and then failing at boot because nobody knew which variables to set.
# .env.example — copy to .env for local dev. Every var the app reads lives here.
# ── App ───────────────────────────────────────────────────────────────────
APP_ENV=local
APP_CORS_ORIGINS=*
APP_JWT_SECRET=change-me
APP_JWT_ISSUER=myapp
# ── Einherjar: HTTP server (EINHERJAR_SERVER_*) ───────────────────────────
EINHERJAR_SERVER_ADDR=:8080
# ── Einherjar: PostgreSQL (EINHERJAR_PG_*) ────────────────────────────────
EINHERJAR_PG_HOST=localhost
EINHERJAR_PG_PORT=5432
EINHERJAR_PG_USER=postgres
EINHERJAR_PG_PASSWORD=postgres
EINHERJAR_PG_DATABASE=myapp
To discover the full set, walk every env:"..." tag reachable from config.Config (including the
nested framework configs) — every one of them belongs in .env.example.
wire.go — Run()
The application entry point. The order below is load-bearing: configuration first, observability
second, infrastructure third, cross-cutting helpers fourth, then the launcher with every component
appended, then feature hooks, then lc.Run().
package wire
import (
"strings"
"github.com/google/uuid"
authjwt "code.nochebuena.dev/einherjar/auth-jwt"
"code.nochebuena.dev/einherjar/auth/authmw"
"code.nochebuena.dev/einherjar/auth/rbac"
"code.nochebuena.dev/einherjar/core/launcher"
"code.nochebuena.dev/einherjar/core/logz"
"code.nochebuena.dev/einherjar/core/valid"
"code.nochebuena.dev/einherjar/db-postgres"
"code.nochebuena.dev/einherjar/web/mw"
"code.nochebuena.dev/einherjar/web/server"
"myapp/internal/config"
)
func Run() error {
cfg, err := config.Load()
if err != nil {
return err
}
logger := logz.New(logz.Config{
JSON: !strings.EqualFold(cfg.AppEnv, "local"),
StaticArgs: []any{"service", "myapp", "env", cfg.AppEnv},
})
signer := authjwt.NewHMACSigner([]byte(cfg.JWT.Secret))
publicPaths := []string{
"/health",
"/api/v1/auth/login",
"/api/v1/auth/refresh",
}
db := postgres.New(logger, cfg.PG)
srv := server.New(logger, cfg.Server,
server.WithMiddleware(
mw.RequestID(uuid.NewString),
mw.Recover(logger),
mw.CORS(cfg.CORSOrigins),
mw.RequestLogger(logger),
authjwt.AuthMiddleware(logger, signer, publicPaths),
authmw.EnrichmentMiddleware(logger, &claimsEnricher{}),
),
)
v := valid.New(valid.WithMessageProvider(valid.SpanishMessages))
provider := rbac.NewClaimsPermissionProvider("masks", claimsFromCtx)
lc := launcher.New(logger)
lc.Append(db, srv)
withHealth(lc, srv, logger, db)
withUsers(lc, srv, db, logger, provider, v)
// … one withFeature(...) call per feature in your domain.
return lc.Run()
}
Feature hook
One file per feature in internal/wire/. The function signature is fixed:
launcher.Launcher first, server.Server second when registering routes, deps last. The body is
one call to lc.BeforeStart. Everything else — repository, service, handler construction, route
registration — lives inside the closure.
package wire
import (
"code.nochebuena.dev/einherjar/contracts/security"
"code.nochebuena.dev/einherjar/core/launcher"
"code.nochebuena.dev/einherjar/core/logz"
"code.nochebuena.dev/einherjar/core/valid"
"code.nochebuena.dev/einherjar/db-postgres"
"code.nochebuena.dev/einherjar/web/server"
"myapp/internal/domains"
userhandler "myapp/internal/user/handler"
userrepo "myapp/internal/user/repository"
usersvc "myapp/internal/user/service"
)
func withUsers(
lc launcher.Launcher,
srv server.Server,
db postgres.Component,
logger logz.Logger,
provider security.PermissionProvider,
v valid.Validator,
) {
lc.BeforeStart(func() error {
repo := userrepo.New(db)
uow := postgres.NewUnitOfWork(logger, db)
svc := usersvc.New(repo, uow)
h := userhandler.New(svc, v)
// Literal-segment routes register BEFORE parametrised siblings.
srv.Put("/api/v1/users/me/password", h.ChangeOwnPassword)
srv.With(authz(provider, domains.ResourceUsers, domains.GrantReadUser)).
Get("/api/v1/users", h.ListUsers)
srv.With(authz(provider, domains.ResourceUsers, domains.GrantCreateUser)).
Post("/api/v1/users", h.CreateUser)
srv.With(authz(provider, domains.ResourceUsers, domains.GrantUpdateUser)).
Put("/api/v1/users/{user_id}", h.UpdateUser)
return nil
})
}
Route ordering
chi matches paths in registration order. Always register literal-segment routes before parametrised-segment routes that share the same prefix.
✅ Correct:
srv.Put("/api/v1/users/me/password", h.ChangeOwnPassword)
srv.Put("/api/v1/users/{user_id}", h.UpdateUser)
❌ Wrong — chi binds me to {user_id} and the literal route is unreachable:
srv.Put("/api/v1/users/{user_id}", h.UpdateUser)
srv.Put("/api/v1/users/me/password", h.ChangeOwnPassword)
Authorization
Every protected route registers with .With(authz(provider, resource, grant)):
srv.With(authz(provider, domains.ResourceUsers, domains.GrantReadUser)).
Get("/api/v1/users", h.ListUsers)
Resource constants and grant bits live in internal/domains/. Routes that the caller owns
(/me/...) intentionally skip authz — they are reachable to any authenticated user.
Middleware helpers
These belong in internal/wire/middleware.go and are used across every feature hook.
// authz returns a per-route authorization middleware that checks one bit.
func authz(p security.PermissionProvider, resource string, bit int) func(http.Handler) http.Handler {
return authmw.AuthzMiddleware(nil, p, resource, security.Permission(bit))
}
// skipPublicPaths wraps mw so it is bypassed for any path that matches publicPaths.
// Use this for middleware that must not run on unauthenticated endpoints
// (e.g. EnrichmentMiddleware).
func skipPublicPaths(publicPaths []string, mw func(http.Handler) http.Handler) func(http.Handler) http.Handler {
return func(next http.Handler) http.Handler {
inner := mw(next)
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
for _, p := range publicPaths {
if matched, _ := path.Match(p, r.URL.Path); matched {
next.ServeHTTP(w, r)
return
}
}
inner.ServeHTTP(w, r)
})
}
}
// skipMethodPath bypasses mw only when BOTH method and path match. Use this to
// expose ONE method on an otherwise-authenticated path (e.g. GET /api/v1/config
// public while PUT is not). Adding such a path to publicPaths would silently
// strip identity from context on the protected methods, breaking authz().
func skipMethodPath(method, pathPattern string, mw func(http.Handler) http.Handler) func(http.Handler) http.Handler {
return func(next http.Handler) http.Handler {
inner := mw(next)
return http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
if r.Method == method {
if matched, _ := path.Match(pathPattern, r.URL.Path); matched {
next.ServeHTTP(w, r)
return
}
}
inner.ServeHTTP(w, r)
})
}
}
Adapters at the wire boundary
When a framework type does not match a service-layer port, write a small typed adapter in
internal/wire/. Always compile-time assert with var _ TargetIface = (*adapter)(nil).
The framework intentionally exposes only Signer.Sign(claims) (string, error) — the framework
gives you a signing primitive; the access/refresh strategy, claim layout, and response shape are
application concerns. A "helper" that returned a fixed {access, refresh, type, expiresIn} struct
would silently decide for every app whether refresh tokens exist, what fields to expose, and what
casing to use. Those are wire-format choices the app owns.
import (
"time"
"github.com/golang-jwt/jwt/v5"
"github.com/google/uuid"
authjwt "code.nochebuena.dev/einherjar/auth-jwt"
)
type tokenSignerAdapter struct {
signer authjwt.Signer
cfg authjwt.TokenConfig
}
var _ authsvc.TokenSigner = (*tokenSignerAdapter)(nil)
func (a *tokenSignerAdapter) IssueTokenPair(subject string, custom map[string]any) (authdto.TokenPairResponse, error) {
now := time.Now()
access := jwt.MapClaims{
"sub": subject,
"iss": a.cfg.Issuer,
"iat": now.Unix(),
"exp": now.Add(a.cfg.AccessTTL).Unix(),
}
for k, v := range custom {
access[k] = v
}
accessToken, err := a.signer.Sign(access)
if err != nil {
return authdto.TokenPairResponse{}, err
}
refreshToken, err := a.signer.Sign(jwt.MapClaims{
"sub": subject,
"jti": uuid.NewString(),
"iat": now.Unix(),
"exp": now.Add(a.cfg.RefreshTTL).Unix(),
})
if err != nil {
return authdto.TokenPairResponse{}, err
}
return authdto.TokenPairResponse{
AccessToken: accessToken,
RefreshToken: refreshToken,
TokenType: "Bearer",
ExpiresIn: int(a.cfg.AccessTTL.Seconds()),
}, nil
}