How It Works

goninja has two halves that never run at the same time: a generator you invoke from the command line, and a small runtime package the generated code calls into. Understanding which half does what explains most of the design.

The pipeline

Parse

goninja generate scans the -models directory with go/parser and go/ast — it reads source text, it does not build or load your package. Any struct with at least one goninja-tagged field becomes a model. Structs with no tagged fields are skipped silently.

Build an IR

Each model becomes an intermediate representation: the model’s name, its ID type, and one entry per tagged field carrying the field’s Go type, JSON name, validate tag, whether it is a relation, and which verbs it opted into.

Because the IR is decoupled from both parsing and templating, the same description drives the handlers, the queries and the OpenAPI fragment. That is why the generated document cannot disagree with the generated code.

Render and format

The IR is rendered through text/template, passed through go/format, and written as <model>_generated.go — one file per model. Formatting happens before the write, so generated output is always gofmt-clean.

Compile

The result is ordinary Go. Your normal go build type-checks the generated handlers against your models. A tag that names a field wrongly, or a relation type goninja cannot handle, fails here — not at runtime.

What runs at request time

Only plain Go. A generated handler decodes JSON into a typed struct, validates it against its validate tags, builds a GORM query, maps the row into a typed response struct, and writes JSON. There is no route table built by reflection, no schema resolved from a type at first request, no any-typed model registry.

This is the trade goninja makes. A reflection-driven CRUD library can adapt to a type it has never seen, at the cost of doing that work per process (or per request) and of failing at runtime when a type does not fit. goninja needs a regeneration step whenever a model changes, and in exchange the failure mode moves to compile time and the request path stays boring.

Regenerating is the step people forget. Run goninja generate -watch during development and it happens on every save.

The two halves

	Generator	Runtime
Package	`internal/codegen`, driven by `cmd/goninja`	`goninja` plus `openapi`, `docsui`, `id`
Runs	when you invoke the CLI	on every request
Sees your models	as source text	as Go types, already compiled
Failure shows up as	a generate or build error	an HTTP error response

The runtime is deliberately small. It holds the pieces that would otherwise be duplicated into every generated file: BaseResource (database access, transactions, config, the Self() dispatch point), the error types and their mapping to HTTP responses, validation, pagination, and the auth contract. Generated code calls these directly, so there is no per-model runtime scaffolding to deduplicate.

Two views per model, by design

Every model produces four output types, and the split between two of them is a guarantee rather than a default:

<Model>List is the collection view. It never preloads a relation.
<Model>Retrieve is the detail view. It preloads every relation field it carries.

A list endpoint therefore cannot issue one query per row, because the code that would do it is never generated. If you want a relation in the detail view but not a join, tag it byid and the field becomes a bare ID.

<Model>Create and <Model>Update are the request bodies, and they are the only types that carry validate tags — validation applies to input only.

All four are separate structs from your GORM model. The model is never reused as a response type, so a column cannot appear in a response merely because it exists on the struct.

Extending without editing generated files

Generated files carry a DO NOT EDIT header and are meant to be regenerated freely. Every extension point works from the outside:

Hooks — implement an interface like BeforeCreateHook on a wrapper type to run logic inside the operation’s transaction.
Method overrides — each resource exposes a <Model>Ops interface; wrap the resource and replace a single method.
Config — implement Configurer to change the mount path or restrict which routes are registered.
Actions — attach non-CRUD endpoints with SetActions.

All of these resolve through SetSelf/Self(), because Go has no dynamic dispatch through embedding: a generated method has to ask “has someone wrapped me?” explicitly. See Hooks & Overrides.

Struct Tags

The full tag vocabulary.

Generated Code

What a generated file contains.

Comparison