RIPDPI: Skill for Claude Code

CI Workflow Authoring

Workflow count, jobs, dispatch inputs, action versions, and concurrency rules are source-controlled in .github/workflows/. Inspect the current manifests and copy the nearest existing job; do not maintain a second hand-written inventory here.

CI architecture (ci.yml)

Derive the current job graph, triggers, dispatch inputs, and concurrency: policy directly from .github/workflows/ci.yml; these change too often for a second inventory here.

Environment Setup Pattern

Most Android/native jobs reuse .github/actions/setup-android-rust. Copy that composite action or the nearest current job, including every exact action SHA; never substitute a floating tag from an example.

Caching Strategy

Cache	Implementation	Notes
Gradle	Exact pinned gradle/actions/setup-gradle SHA from the nearest job	Preferred over hand-rolled cache blocks
Rust workspace	Exact pinned Swatinem/rust-cache SHA from the nearest job	Caches native/rust -> target
Rust compiler cache	Exact pinned sccache action SHA from the nearest job	Used on native-heavy jobs
Benchmark baselines	actions/cache/restore / save	Used by criterion baselines on PRs and main
Tool installs	taiki-e/install-action@v2	Used for cargo-nextest, cargo-llvm-cov, cargo-bloat

Manual Dispatch Inputs

Derive the current inputs from the workflow_dispatch.inputs mapping in ci.yml; it changes as optional lanes evolve. Typical inputs include:

workflow_dispatch:
  inputs:
    soak_profile: smoke|full
    run_native_soak: true|false

If you add a new manual-only lane, wire its input into both the job if: condition and the step logic that consumes it.

Adding or Modifying a CI Job

Checklist:

1. Decide which workflow the job belongs in -- inspect .github/workflows/ for the current file list and pick the nearest match; the repo has grown well past a small fixed set and this skill does not maintain a second inventory of it.
2. Copy environment setup from the nearest existing job instead of inventing a new setup pattern.
3. Read the NDK version from gradle.properties; never hardcode it in YAML.
4. Add timeout-minutes unless the job is trivially short.
5. Use if: always() on artifact upload steps.
6. Keep artifact names stable when downstream debugging depends on them.
7. If the job is nightly-only, gate it explicitly on schedule or workflow_dispatch.
8. If the job runs :app or :baselineprofile instrumented tests, use the Gradle Managed Device pattern: a <device>...AndroidTest task plus the shared device registry in build-logic/convention/src/main/kotlin/RipDpiManagedDevices.kt (see the android-instrumented-tests and android-macrobenchmark jobs). Specialized emulator lanes (journeys, relay-smoke, network-e2e) still use the scripts/ci/start-android-emulator.sh harness.