Documentation Why Install Try It Kotlin Portability Atlas Architecture Shared Code Limits

Architecture

Modules

Module	Owns
`core`	Capsule model, canonical JSON, SHA-256, fixed-point math, diff, ranking, rules, planner, Lab. Common code only.
`host`	Operating system access, one implementation per target (JVM, Linux, macOS, Windows, Wasm, Android, iOS).
`scan`	Probes, toolchain scanners, the Atlas dataset builder and the probe classification.
`cli`	The `drift` command and its subcommands, including `upgrade-check`, `triage`, `contract`, `install`, `uninstall` and `serve`.
`studio`	The Compose Multiplatform app for desktop, web, Android and iOS: Home, Case, Atlas, Investigation, Upgrade, Contract and Bisect screens, and the installer packaging.
`bench`	DriftBench, a seeded benchmark of container pairs with an injected cause.
`tools`	Fixture freezing, Atlas ladder tooling, the triage corpus builder and scorer, the performance rig and the installer test scripts.
`androidApp`, `iosApp`	The Android and iOS app shells. They exist only under `-Pdrift.mobile=true`.
`packaging`	Homebrew formula and cask and Chocolatey package templates, rendered by `./gradlew renderPackaging` after a release.
`atlas`	Hand-written probe classification, column descriptions and repro drafts in YAML. The build embeds the classification and builds the static site from it.
`build-core`	Gradle convention plugins and build tasks: rules, Atlas site, packaging templates, Dokka.
`rules`	Hand-written rule files in YAML. The build generates the JSON that `core` reads.
`triage`	The manifest of the issue corpus that `drift triage` is measured on, and the claim-check labels.
`fixtures`	Recorded capsules, Atlas transcripts, solved cases, flagship cases, project trees, issue bodies and upgrade-check captures.
`install.sh`, `install.ps1`	Download, verify and install a release.

Dependencies

host depends on core, scan depends on core and host, and cli and studio depend on all three. bench and tools depend on core alone. androidApp wraps studio.

flowchart TD
    cli --> scan
    studio --> scan
    cli --> host
    studio --> host
    scan --> host
    scan --> core
    host --> core
    cli --> core
    studio --> core
    bench --> core
    tools --> core

Platform Code

core and scan have no platform source sets. Every expect and actual pair lives in host, which has one source set per target: JVM, Windows, Wasm, Android and iOS, plus a posix set that Linux and macOS share. Common code in the other modules reaches the operating system only through host.

flowchart LR
    common["common code in core, scan, cli and studio"] --> host["host: expect declarations"]
    host --> jvm
    host --> posix["posix (Linux and macOS)"]
    host --> mingw["mingw (Windows)"]
    host --> wasmJs
    host --> android
    host --> ios

Investigation Chain

An investigation runs in five steps. drift capture records an environment as a capsule. drift diagnose diffs a passing and a failing capsule and ranks the differences. The planner prices each possible experiment and proposes one. drift solve runs it in containers under a decision rule fixed before the first trial. drift pack writes the result as a .driftcase certificate, and drift verify re-derives its counts and verdict from the file.

flowchart TD
    capture["capture: record each environment"] --> diff["diagnose: diff and rank"]
    diff --> plan["plan: price the experiments"]
    plan --> run["solve: run the chosen experiment"]
    run --> certify["pack and verify: certificate"]

Shared Code lists how much of this code is common to every target.