hx

hx — Haskell toolchain

hx is a single CLI that replaces ghcup, cabal, stack, fourmolu, hlint, and hpc. It manages GHC itself, uses an hx.toml manifest and an hx.lock lockfile, and returns meaningful exit codes.

Detecting that this skill applies

The project is an hx project if there is an hx.toml at (or above) the working directory. If the user has a .cabal/stack.yaml but no hx.toml, you can adopt it with hx init or hx import --from stack.

Core commands

Run hx <command> --help for flags. --quiet/--verbose are global; NO_COLOR is respected.

Interpreting results — exit codes

Decide success/failure from the exit code, not by scraping output:

• 0 success · 1 general error · 2 bad arguments · 3 bad hx.toml
• 4 toolchain error (GHC/Cabal missing or mismatched — run hx doctor)
• 5 build/test failure · 6 plugin hook failure

Errors print a structured summary with a suggested fix on stderr; surface that fix to the user.

Typical workflows

# new project
hx init myapp && cd myapp && hx build && hx run

# add a dependency and rebuild
hx add aeson ">=2.0" && hx build

# pre-commit quality gate
hx fmt --check && hx lint && hx test

# reproducible build
hx lock && hx sync

Important behavior

• Toolchain auto-install: the first hx build may download a GHC (hundreds of MB). hx doctor returns exit 4 when a required tool is missing and tells you how to install it.
• hx build --native is experimental — a fast path for single-package, base-only projects; it falls back to cabal otherwise. Don't use it for multi-dependency projects.
• Plugins require trust: project-local .hx/plugins/*.scm scripts do not run on a freshly cloned repo. Never run hx plugins trust on a project you don't trust.

Programmatic / MCP use

For tool-call access instead of shelling out, run hx mcp (an MCP server over stdio) — it exposes build/check/test/run/lock/sync/fmt/lint/doctor and dependency tools. See AGENTS.md in the repo.