Kinda

Kinda is a Zig-first binding framework for Elixir.

It is designed for projects that need to bind a large C API to BEAM without hand-writing every NIF, resource wrapper, and dispatch stub.

The shortest useful description is:

• kinda treats a wrapper header as the binding surface
• extracts functions into a framework-owned manifest
• extracts named C records and field types into that same framework-owned manifest
• preserves Clang-extracted function docs when comments are available
• preserves raw C parameter/return type facts in that manifest
• generates generic Zig/Elixir wrapper outputs
• lets consumer policy project those raw C facts into public Elixir typespecs
• lets the same typed projection describe public record/field types in a machine-readable signature manifest
• lets the consumer keep product-specific policy outside the framework

This makes kinda a better fit for projects like beaver, where the problem is not “write one ergonomic NIF”, but “keep a large drifting native API coherent over time”.

Run One Example

If you want the shortest stable repo-local proof of the wrapper/reporting surface, run:

mix kinda.wrapper.example

Useful modes:

mix kinda.wrapper.example --json
mix kinda.wrapper.example --report-only

To verify the whole repo from the root, including:

• root tests
• wrapper-reporting JSON smoke
• bundled kinda_example

run:

mix kinda.verify

This is the command CI should prefer for repo-level verification.

If you only want the bundled application example from the repo root, run:

mix kinda.example.verify

If you want the same thing without compiling the project first, run:

elixir examples/wrapper_reporting.exs

It prints:

• the generated Elixir wrapper manifest
• the machine-readable typed signature manifest
• the human-readable callback-bridge backlog report
• the machine-readable callback-bridge manifest
• including dirty scheduler metadata when the example policy emits it

What Kinda Ships Today

Core runtime/building blocks:

• Kinda.ResourceKind
• Kinda.CodeGen
• Kinda.Precompiler
• Kinda.Codec
• Kinda.CallbackRuntime for the common BEAM callback/reply boundary
• kinda.callback_runtime for Zig-side native-thread callbacks into BEAM

Wrapper extraction/generation blocks:

• Kinda.Wrapper.Function
• Kinda.Wrapper.Manifest
• Kinda.Wrapper.Extract
• Kinda.Wrapper.Generate
• Kinda.Wrapper.Policy
• Kinda.Wrapper.CallbackBridge

That means kinda now has an explicit split between:

1. generic wrapper extraction/generation
2. consumer-owned policy
3. callback-bridge backlog and runtime substrate

Mental Model

Kinda is not primarily a “write NIFs directly in Zig” library.

It is a framework for binding a C library through three layers:

1. wrapper surface
   • a .h file says what native functions and types you want to expose
2. framework surface
   • kinda extracts that surface into a normalized manifest
   • kinda generates generic wrapper outputs from the manifest
3. consumer surface
   • your project decides naming, scheduler routing, diagnostics variants, unsupported APIs, and callback-heavy backlog

For a large binding, this is the important compression:

• kinda owns the generic mechanics
• your library owns the semantics

Quick Example

Define a boundary codec, generate the raw NIF surface, and bind a resource kind directly to it:

defmodule Foo.Native do
  use Kinda.Codec
end

defmodule Foo.CAPI.Raw do
  use Kinda.CodeGen,
    with: Foo.Generated,
    root: Foo.CAPI,
    codec: Foo.Native,
    surface: :raw

  @on_load :load_nif

  def load_nif do
    :erlang.load_nif(~c"path/to/foo_nif", 0)
  end
end

defmodule Foo.CAPI do
  use Kinda.CodeGen,
    with: Foo.Generated,
    root: __MODULE__,
    raw_module: __MODULE__.Raw,
    codec: Foo.Native,
    surface: :public
end

defmodule Foo.Handle do
  use Kinda.ResourceKind,
    raw_module: Foo.CAPI.Raw,
    codec: Foo.Native
end

Generated public wrappers call concrete functions in Foo.CAPI.Raw, unwrap resource arguments, and pass only the returned value through Foo.Native. The codec never selects a native function. Keeping :public wrappers and :raw NIF stubs in separate modules lets large bindings compile both surfaces independently and avoids a generated proxy layer. The raw module owns NIF loading, so the native entry module must be Elixir.Foo.CAPI.Raw.

Generated use Kinda.CodeGen modules now expose one formalized declaration surface from the same typed source:

• the resolved declaration IR through __kinda_declaration_surfaces__/0

When a generator module implements declaration_manifest/0, Kinda.CodeGen can now source generated function/type declarations directly from that unified manifest instead of reconstructing them from parallel callbacks. That callback is now the canonical declaration interface in kinda: it may return a loaded manifest value or a checked-in .ex / .json sidecar path. Kind-derived helper NIFs still come from kinds()/0.

kinda now also exposes a top-level downstream declaration facade through Kinda.Declaration. That facade sits over the same underlying Kinda.CodeGen.DeclarationSurfaces IR and formalizes the distinction between the declaration source and the generated module surface:

• Kinda.Declaration.load_source/1 reads the canonical source contract from a generator module
• Kinda.Declaration.from_generator/2 resolves that source contract into the final generated surfaces for a specific root module, including normalized nif_names, kind-derived helper entries, guaranteed/materialized generated TypeDecls, and the derived signature view
• that resolved surface lands as the formal framework-owned IR: Kinda.CodeGen.DeclarationSurfaces

Downstreams should treat Kinda.Declaration.from_generator/2 as the formalized public resolution interface. Generated modules expose that same resolved IR directly through __kinda_declaration_surfaces__/0. Inside that IR, the canonical resolved payload is the declaration manifest itself; nif_decls, type_decls, and signature_manifest are no longer stored a second time on the resolved surface. That lets downstream repos such as Beaver unify their own public rewrite/pass DSL names without re-owning the declaration contract.

When that manifest includes projected records, Kinda.CodeGen also emits deterministic public type aliases such as foo_handle_record()/0 from the same single source. These generated aliases use atom field keys derived from extracted C field names, while the machine-readable manifest keeps the original string names.

For larger generated bindings, the newer wrapper pipeline is usually more important.

Wrapper Pipeline

The wrapper pipeline is the part of kinda that scales to large native APIs.

1. Extract a manifest

manifest = Kinda.Wrapper.Extract.from_clang_ast(ast_json)

The result is a framework-owned shape:

%Kinda.Wrapper.Manifest{
  records: [
    %Kinda.Wrapper.CRecord{
      name: "MlirContext",
      kind: :struct,
      fields: [
        %Kinda.Wrapper.CField{
          name: "ptr",
          ctype: %Kinda.Wrapper.CType{spelling: "const void*", kind: :pointer}
        }
      ]
    }
  ],
  functions: [
    %Kinda.Wrapper.Function{
      name: "mlirContextCreate",
      params: [],
      arity: 0,
      doc: "Creates a new MLIR context.",
      param_ctypes: [],
      return_ctype: %Kinda.Wrapper.CType{
        spelling: "MlirContext",
        kind: :unknown
      }
    }
  ]
}

2. Apply consumer policy

defmodule MyLib.WrapperPolicy do
  @behaviour Kinda.Wrapper.Policy

  def generation_blocker_entries, do: %{}
  def generation_blocked?(_name), do: false
  def generation_blocker_reason(_name), do: nil

  def callback_bridge_entries, do: %{}
  def callback_bridge?(_name), do: false
  def callback_bridge(_name), do: nil

  def variants(name), do: [{:normal, name, name}]
  def public_name({_kind, public, _base}), do: public
  def elixir_params({_kind, _public, _base}, params), do: params
  def dirty({_kind, _public, _base}), do: false
  def typespec_field(_record, field), do: ...
  def typespec_params(_variant, function), do: ...
  def typespec_return(_variant, function), do: ...
  def zig_entry({_kind, _public, base}), do: ~s{nif("#{base}"),}
end

3. Generate outputs

Kinda.Wrapper.Generate.render_elixir_manifest(manifest, MyLib.WrapperPolicy)
Kinda.Wrapper.Generate.render_zig_nif_entries(manifest, MyLib.WrapperPolicy)
Kinda.Wrapper.Generate.declaration_surfaces_struct(manifest, MyLib.WrapperPolicy)
Kinda.Wrapper.Generate.declaration_manifest_struct(manifest, MyLib.WrapperPolicy)
Kinda.Wrapper.Generate.declaration_manifest(manifest, MyLib.WrapperPolicy)

Callback-Bridge Backlog

Not every native function should be emitted as a plain generated wrapper.

Some functions need:

• callbacks back into BEAM
• richer decoder logic
• lifetime review
• scheduler review

Kinda now represents that explicitly with Kinda.Wrapper.CallbackBridge.

Example:

Kinda.Wrapper.CallbackBridge.required(:some_function,
  scheduler: :dirty_cpu,
  facets: [:beam_callback, :scheduler_contract]
)

This metadata does not invent a consumer ABI automatically. It is the framework-owned policy layer that lets a consumer say:

• this API is not “missing”
• this API belongs to the callback-bridge backlog, or is resolved by the dispatcher runtime

Pure kind-surface gaps belong to a different path. In a consumer such as beaver, handle-like CAPIs are often unblocked by extending the Zig-side kind surface and the matching Elixir KindDecl surface; callback-bridge metadata is specifically for the remainder that kind-surface sync cannot solve.

Callback Runtime

Consumers can implement those callback bridges with the generic Zig runtime:

const kinda = @import("kinda");
const Dispatcher = kinda.callback_runtime.Dispatcher(.{ "run", "destruct" });

const dispatcher = try Dispatcher.initWithOptions(handler_pid, .{
    .timeout_ms = 30_000,
});
dispatcher.setCallback("run", run_callback_term);
const response = try dispatcher.invoke("run", message_env, .{argument_term});

On the BEAM side, the consumer supplies its own reply NIF while Kinda normalizes success, expected failure, and exceptions:

Kinda.CallbackRuntime.invoke(
  reply_token,
  fn -> {:ok, run_callback.()} end,
  &MyNative.my_raw_callback_reply/2
)

Callbacks with scalar, enum, or projected handle results use invoke_reply/4. The consumer validates its own resource before completing the shared token:

Kinda.CallbackRuntime.invoke_reply(reply_token, callback, fn token, outcome ->
  MyNative.reply_projected_result(token, outcome)
end)

The consuming NIF library exports and opens the shared reply resource:

const callback_nifs = .{
    kinda.callback_runtime.ReplyToken.nif("my_raw_callback_reply"),
    kinda.callback_runtime.ReplyToken.codeNif("my_raw_callback_reply_code"),
    kinda.callback_runtime.ReplyToken.cancelNif("my_raw_callback_cancel"),
};

export fn nif_load(env: kinda.beam.env, _: [*c]?*anyopaque, _: kinda.beam.term) c_int {
    kinda.callback_runtime.ReplyToken.open(env);
    return 0;
}

Dispatcher owns copied callback terms and its persistent environment. Each invocation sends {callback_name, reply_token, callback_fun, dispatcher_id, ...args}, then waits on a non-scheduler native thread. The process replying through the consumer-exported NIF becomes the next callback owner. Native callback signatures, argument conversion, domain state, and diagnostics remain the consumer's responsibility.

Waits are bounded to 30 seconds by default. A response reports whether it was replied, canceled, dropped, or timed out; completion after any terminal state returns stale. Dropping the reply resource or terminating its owner therefore cannot leave a native worker waiting forever. ReplyToken and consumer-owned registration resources use NIF resource-type takeover so live resources keep the correct native library generation pinned across reload and unload.

kinda.callback_adapter provides the reusable projection shapes used by consumer ABI trampolines: one handle, ranges of handles, scalar results, enum results, and validated consumer projections. It never interprets an MLIR- or library-specific handle.

Callback-bridge manifest version 2 records runtime_backed, runtime, owner, destructor, lifetime, scheduler, and timeout_ms. Entries built with Kinda.Wrapper.CallbackBridge.runtime_backed/2 no longer carry a callback_bridge_required blocker and their consumer-provided declaration variant remains in the normal resolved declaration surface.

Reporting Surface

Kinda now exposes one canonical declaration contract plus one derived signature view around that wrapper surface:

1. resolved declaration surfaces
2. unified declaration manifest
3. callback-bridge manifest

Resolved declaration surfaces:

Kinda.Wrapper.Generate.declaration_surfaces_struct(manifest, policy)

This is now the canonical wrapper-side in-memory declaration IR in kinda. It carries:

• the source declaration-manifest slot, set to nil for wrapper-generated surfaces
• the canonical resolved declaration manifest

Unified declaration manifest:

Kinda.Wrapper.Generate.declaration_manifest_struct(manifest, policy)
Kinda.Wrapper.Generate.declaration_manifest(manifest, policy)

The declaration-manifest struct is now the canonical resolved payload stored inside DeclarationSurfaces, and the map form is the JSON-friendly serialization. It keeps:

• named C records and fields from extraction
• raw C param/return type facts from extraction
• consumer-projected public params and return typespecs
• consumer-projected public record and field typespecs
• generated NIFDecl entries
• generated TypeDecl entries
• dirty scheduler metadata on emitted variants
• generation-blocker reasons when the function is not emitted as a plain generated wrapper

Internally, Kinda.CodeGen.DeclarationManifest.build/2 is now the canonical way to derive declaration metadata from typed wrapper facts, including the generated TypeDecl layer. That keeps type declarations sourced from the same declaration contract rather than from a parallel signature-only path.

When that declaration contract is consumed by use Kinda.CodeGen, the formal resolution path now lives in Kinda.CodeGen.DeclarationSurfaces.from_generator/2, so downstreams do not have to reimplement normalization or merge logic just to observe the final generated declaration surface. That resolved surface is now materialized as Kinda.CodeGen.DeclarationSurfaces, rather than an ad hoc map.

The typed signature manifest remains available as a derived compatibility view:

Kinda.Wrapper.Generate.signature_manifest(manifest, policy)

The callback-bridge backlog remains a separate reporting mode:

1. human-readable report
2. machine-readable manifest

Human-readable:

Kinda.Wrapper.Generate.render_callback_bridge_report(manifest, policy)

Machine-readable:

Kinda.Wrapper.Generate.callback_bridge_manifest(manifest, policy)

The manifest is versioned and JSON-friendly:

{
  "version": 1,
  "entries": [
    {
      "function": {
        "name": "mlirTypeConverterAddConversion",
        "arity": 1,
        "params": ["converter"]
      },
      "callback_bridge": {
        "function": "mlirTypeConverterAddConversion",
        "reason": "callback_bridge_required",
        "unblock_path": "callback_bridge_runtime",
        "scheduler": "unspecified",
        "facets": ["beam_callback", "rich_input_decoder"]
      }
    }
  ]
}

This is the main new “reporting surface” for consumers and CI.

The repo-local example above exercises this end to end.

Build / Prebuilt Surface

Kinda ships Kinda.Precompiler, which consumers can use with elixir_make precompiled builds.

Example:

def project do
  [
    make_precompiler: {:nif, Kinda.Precompiler}
  ]
end

Today this is target-selection substrate, not a full RustlerPrecompiled-style product story.

What Kinda Is Good At

• binding large C APIs with repetitive structure
• generating resource-centric Elixir/Zig surfaces
• separating framework mechanics from consumer policy
• keeping callback-heavy APIs visible as backlog instead of hiding them in a flat blacklist
• acting as a binding substrate for projects that drift with upstream native APIs

What Kinda Is Not Yet

Kinda is not yet a Rustler-complete framework.

What is still missing:

• policy-driven generation from callback metadata to runtime adapters
• richer scheduler-aware NIF declaration surface
• a complete prebuilt/download/checksum story
• a more polished one-command reporting UX

Status

Kinda is already useful as a framework substrate.

It is now moving from:

• “resource kind + codegen helper”

toward:

• “wrapper-driven Zig/Elixir binding framework with explicit policy and backlog surfaces”

That is the right frame to evaluate future work.