0008

Status: Accepted (supersedes the spago pin in 0001)

Context

The set was built for the legacy spago (0.21.0): a Dhall overlay (src/packages.dhall) holding only the Lua forks, released as a packages.dhall asset that a consumer merged with the upstream PureScript set (upstream-ps // upstream-lua). The right operand won, so each fork overrode its JavaScript-targeting upstream. ADR 0001 pinned spago 0.21.0 across the forks to make this work.

The PureScript toolchain has moved to the new spago (1.x), which drives spago.yaml + spago.lock and the PureScript Registry, and dropped Dhall. pslua itself migrated its test toolchain to it (pslua #54, #55). The new spago has no Dhall and no set-merge operator: workspace.packageSet takes a single source — a registry version, a URL, or a path. So the old "merge two Dhall sets" model cannot be expressed, and a forks-only overlay is no longer directly consumable.

A spike (pslua #115) confirmed three ways — the spago docs, the core/src/Config.purs codecs (RemotePackage = git | registry-version | legacy), and a working PoC — that the new spago consumes a custom set through packageSet.url / packageSet.path, and that such a set may carry git-based package entries. So the set can become a single, self-contained, consumable input again.

Decision

Each release publishes a packages.json asset: a RemotePackageSet ({ compiler, version, packages }) where the pure packages come from a pinned upstream registry package set baseline (the proven-compatible 77.x set; see the generator) and the Lua forks are overlaid as git entries ({ git, ref, dependencies }). src/packages.json is the single source of truth for the forks (a small overlay); scripts/gen-package-set-json reshapes it onto the baseline. Consumers point workspace.packageSet.url at the published packages.json and drop extraPackages entirely.

The forks and the set's own coherence check migrate to the new spago (1.x): spago.yaml replaces spago.dhall/packages.dhall, and the toolchain pin moves from spago-bin.spago-0_21_0 to spago-bin.spago-1_0_4. This supersedes the spago 0.21.0 pin in ADR 0001; the purs 0.15.16 and Lua 5.1 pins there stand.

Dhall is dropped wholesale: the overlay source becomes src/packages.json, the legacy packages.dhall release asset is removed, and the dhall-to-json step leaves the generator and the README-table script. spago 0.21 / Dhall consumers are no longer supported — the whole ecosystem moves to the new spago.

Consequences

• The set is a live build input again — one URL, no // merge, no inlined extraPackages dependency lists in every consumer.
• packages.json is generated, never committed: scripts/gen-package-set-json builds it from src/packages.json + the pinned registry baseline. The coherence check (test-set.yml) regenerates it and builds against it, so the generator is exercised on every push.
• The published version/compiler fields are the registry baseline's (0.15.15 / its set version); the GitHub release tag (psc-0.15.15-<date>) identifies our revision. purs 0.15.16 against a 0.15.15-built set is the same proven-compatible pairing pslua's test/ps already relies on.
• A fork tag bump flows through unchanged: bump src/packages.json, regenerate the README table, refresh latest-compatible-sets.json, push a psc-* tag — the release workflow emits packages.json (see 0006).
• Each fork's own build migrates to spago.yaml + spago 1.x (rolled out per fork; CONTRIBUTING reflects the new toolchain). The Dhall formatter and the dhall dev-shell tool leave each fork as part of that pass.