Config in code (Starlark)

YAML is the default and right for most setups. When a config has logic — many near-identical services, per-environment conditionals, computed ports/URLs — write tarjan.star in Starlark (a small, sandboxed Python-like language) instead. It produces the same config, so every feature works unchanged; the interpreter is pure Go, so tarjan stays a single binary with no extra toolchain.

# N near-identical services become a loop, not N copy-pasted blocks
services = [
    service(
        name = "db",
        docker = docker(image = "postgres:16", ports = ["5432:5432"]),
        health = health(tcp = "localhost:5432"),
    ),
]
for i, name in enumerate(["catalog", "orders", "gateway"]):
    services.append(service(
        name = name,
        workdir = name,
        command = "go run ./cmd/server",
        env = {"PORT": str(8080 + i)},
        depends_on = ["db"],
        restart = "on-failure",
    ))

tarjan = config(name = "shop", services = services, workspace = workspace(vscode = True))

tarjan auto-detects tarjan.star (preferred) or tarjan.yaml. The builtins mirror the YAML schema: config, repo, tool, service, docker, health, watch, hooks, remote, workspace. Keyword arguments are snake_case (e.g. depends_on, min_version, identity_file). Pass remotes = {...} to config(...) and remote = "name" to service(...) to run a service on a remote target. The script must assign its result to a top-level tarjan. tarjan reload works with Starlark too.