Kotlin Toolchain Plugin Authoring

Local plugins are the official escape hatch from declarative YAML: a jvm/amper-plugin module shipping task actions, settings, and generated sources/resources alongside your project. Code patterns to adapt are in references/examples.md.

When to write a plugin

Write one when you need:

• A build-time step a library's workflow expects (code generation, schema compilation, resource transformation, version stamping).
• Custom verification wired into the build (pre-release checks, schema validation, contract tests).
• A build-time value published into the JAR classpath or downstream tasks — the closest analog to Gradle's project.version.
• A named CLI command for a repeated workflow (./kotlin do release).

Don't write one when module.yaml already covers it (dependencies, JDK provisioning, source layouts, basic packaging), and never to reuse a Gradle plugin — the Kotlin Toolchain cannot consume them.

Layout

repo-root/
├── kotlin, kotlin.bat            # wrappers (from `kotlin init`)
├── project.yaml                  # registers the plugin
├── plugins/<name>/
│   ├── module.yaml               # product: jvm/amper-plugin
│   ├── plugin.yaml               # tasks: + commands: + generated:
│   └── src/
│       ├── Settings.kt           # @Configurable interface
│       ├── tasks/                # one @TaskAction per file
│       │   ├── Foo.kt
│       │   └── FooSteps.kt       # internal shared helpers (not @TaskAction)
│       └── <domain logic>/
└── <consumer-module>/
    └── module.yaml               # plugins: { <name>: enabled: true, ... }

Keep at least one consumer module in the repo — it is the only way to exercise the plugin end-to-end, and plugins cannot be published to any public registry yet.

project.yaml

modules:
  - consumer-app
  - plugins/<name>

plugins:
  - ./plugins/<name>

Without the top-level plugins: block the plugin id is unresolvable from any consumer.

module.yaml — the plugin module

product: jvm/amper-plugin          # marks the module as a plugin

dependencies:
  - <coordinate>:<version>
  - <coordinate>:<version>: runtime-only   # required only at runtime
  - <coordinate>:<version>: compile-only

pluginInfo:
  id: <plugin-id>                                  # what consumers write under `plugins:`
  settingsClass: <fully.qualified.Settings>        # the @Configurable interface

settings:
  jvm:
    jdk:
      version: 21
  kotlin:
    languageVersion: 2.1

@Configurable interface Settings

Defaults go in interface property getters; nested blocks become nested @Configurable interfaces.

@Configurable
interface Settings {
    val someValue: String get() = "default"
    val checks: ChecksSettings
}

@Configurable
interface ChecksSettings {
    val strict: Boolean get() = true
}

Consumers override what they need in module.yaml; omitted values fall back to the getter default:

plugins:
  <plugin-id>:
    enabled: true
    someValue: "override"
    checks:
      strict: false

@TaskAction

Task actions are top-level funs, called when the matching plugin.yaml entry executes.

@TaskAction
fun foo(
    @Input moduleRootDir: Path,
    @Output outputDir: Path,
    settings: Settings,
) {
    // body
}

• @Input path: Path — declared input; Kotlin Toolchain snapshots its contents for execution avoidance.
• @Output path: Path — declared output directory; Kotlin Toolchain creates it and passes the path in. Write to the exact Path you received, or downstream references won't find the result.
• settings: Settings (or any @Configurable) — typed configuration, wired in plugin.yaml.
• Plain Path / primitives — passed literally from plugin.yaml.
• println(...) is the output channel; Kotlin Toolchain captures stdout.

Execution avoidance

A @TaskAction is skipped when its declared inputs are unchanged. Tasks whose real inputs are Git history, the network, or environment variables cannot be fingerprinted, so opt out:

@TaskAction(executionAvoidance = ExecutionAvoidance.Disabled)
fun foo(@Output outputDir: Path, settings: Settings) { /* ... */ }

Tasks with no @Output are never cached and always re-run — correct for purely side-effecting tasks (releases, deployments, pushes).

plugin.yaml

tasks:
  foo:
    action: !<fully.qualified.foo>
      moduleRootDir: ${module.rootDir}
      outputDir: ${taskOutputDir}
      settings: ${pluginSettings}

  bar:
    action: !<fully.qualified.bar>
      input: ${tasks.foo.action.outputDir}/result.txt
      settings: ${pluginSettings}

generated:
  resources:
    - directory: ${tasks.foo.action.outputDir}

commands:
  - foo

generated.resources / generated.sources

Both register a directory (usually a task's @Output) as a contribution to the consumer's build, and both auto-wire the producing task to run first:

• generated.resources — added to the JAR classpath, reachable via getResourceAsStream("/path/in/jar").
• generated.sources — added as a Kotlin source root and compiled with the consumer's src/.

Tasks vs commands

Tasks are the implementation, addressed as ./kotlin task :<module>:<task>@<plugin-id> — the docs advise against relying on that mangled name. Commands are the public API: ./kotlin do <command-name>, listed via ./kotlin show commands (-m <module> to scope).

• A task whose @Output feeds generated.resources/generated.sources is a build-graph contributor. Keep it out of commands:; it runs automatically and exposing it invites users to run it by hand.
• A task users invoke directly must be in commands:.

File-based task communication

There is no shared mutable build state — no project.version, no extension property maps. Tasks talk through matched paths:

1. The producer takes @Output outputDir: Path and writes files into it.
2. plugin.yaml points a consumer task's @Input at ${tasks.<producer>.action.outputDir}/<file>.
3. The Toolchain infers the dependency from the path match — no dependsOn API needed.

The same @Output directory can serve build-time consumers (via @Input) and runtime consumers (registered under generated.resources, read via getResourceAsStream) simultaneously.

Runtime overrides via environment variables

There is no -Pkey=value. Read env vars inside the action for ephemeral overrides:

val forced = System.getenv("MYPLUGIN_FORCE_VALUE")?.takeIf { it.isNotBlank() }
val skipChecks = System.getenv("MYPLUGIN_SKIP_CHECKS")?.equals("true", ignoreCase = true) == true

Pass the env map in as a constructor parameter rather than calling System.getenv() deep in the call stack, so logic stays unit-testable. Document every recognised variable in the plugin's README. Env vars are ephemeral overrides, not a trust boundary — validate a value before using it in a file path or process argument.

Sharing logic across task actions

Tasks often share steps (verify → create → push). Don't compose an atomic user-facing task from a chain of build-graph tasks: separate invocations re-open shared resources and open a window where another process observes intermediate state.

Limitations to design around

• Plugins are local-only; no public registry publishing yet.
• Plugins are module-level; there is no project-wide plugin. Every consumer module lists it under plugins:, and cross-module effects flow through files.
• No ${...} interpolation in module.yaml (as of 0.11) — consumer settings are literal values.
• No Project.afterEvaluate, no lazy Provider/Property graph. Compute derived values in the action body.
• -h/--help does not list plugin commands; use ./kotlin show commands.

Validate against a consumer

1. Add a small consumer module (demo-app/, sample/) enabling the plugin with realistic settings.
2. Have its main.kt or a test read whatever the plugin publishes.
3. Put the exact commands and expected output in the plugin's README, so a fresh clone can paste and compare.

Plugin docs: https://kotlin-toolchain.org/dev/user-guide/plugins/