Writing a spell

A spell is how magus learns something it does not know. Everything below is Buzz in a single file, and the whole contract is a set of exported functions magus looks up by name.

There are two kinds, and picking the right one first saves rewriting the file:

A spell can be both, but rarely wants to be. A provider contributes no operation a target could compose, which is why the provider spells in this repo declare no mgs_listTargets at all.

Where the file lives, and the one constraint that decides it

spells/<dir>/spell.buzz

The directory name and the registered name are independent - the go spell lives in spells/golang/. What magus registers is whatever mgs_getName() returns.

The constraint that decides how your spell is loaded:

A spell that imports any host module (os, http, fs, vcs, ...) cannot be compiled into the magus binary. Built-ins are bare-compiled to bytecode at build time (cmd/magus-utils spells), and that compile has no host modules to link against.

So there are two shapes, and you do not get to choose - the imports choose for you:

Almost every provider is workspace-local, because reaching a backend means os\exec or http. That is expected, not a downgrade: spells/github/actions backs this repo's own remote cache that way.

The mgs_ contract

Every function is optional except mgs_getName. magus looks each one up by name and uses a default when it is absent, so a minimal spell is two functions.

Readiness

A version probe answers "what is installed". It cannot answer "is it usable", and for a client/server tool those are different questions: docker --version is client-only and exits 0 with no daemon running at all. Without a readiness probe the op forked, docker failed on its own terms, and the run reported a build failure for a project with nothing wrong with it.

export fun mgs_getTools() > {str: Tool} {
    return {
        "docker": Tool{
            probe = Command{bin = "docker", args = ["--version"]},
            key   = VersionKey{upTo = VersionComponent.patch},
            ready = Command{bin = "docker", args = ["info"]},
        },
        "hadolint": Tool{probe = Command{bin = "hadolint", args = ["--version"]}},
    };
}

Keyed by tool, and resolved through an op's own bin, so no op restates which tool it runs. The docker spell gates docker and deliberately not hadolint - linting a Dockerfile talks to no daemon, and a spell-scoped probe would make a lint wait on a service it never uses.

A failing probe raises MGS3004 before the op forks, carrying the probe's own output. At a terminal magus retries for 30 seconds first, so starting the daemon in another window lets the run continue; without a TTY it fails at once, because nobody starts a daemon mid-run in CI.

The result never enters a cache key: it is a precondition, not an input. docker info reports running containers and disk usage, so keying on it would invalidate every entry on every run. magus doctor lists every declared gate without running any of them.

Most spells need none - go, rustc, and node are self-contained.

A version probe is worth more thought than it looks. If a tool changes what passes and nothing else in the cache key changes with it, every cached entry replays the old verdict. Anything pinned by a manifest the spell already reads (a go.mod the go spell claims) needs no probe; anything that is just "whatever is on PATH" does.

Declaring how a tool reports findings

diagnostics names the convention a binary prints its findings in, so magus reads them as records (file, line, severity, message) rather than scraping prose. DiagnosticFormat.gnu is the GNU Coding Standards shape, [program:]file:line[:column]: severity: message. hadolint spells it -f gnu and shellcheck --format=gcc; gcc and ruff emit the same skeleton.

magus implements the standard once and tools opt in, so it carries no per-tool patterns to rot. It also never rewrites argv: put the flag in the op's own args, beside the declaration.

fun hadolint(target: Target) > Command {
    return Command{bin = "hadolint", args = ["-f", "gnu", "Dockerfile"]};
}

export fun mgs_getTools() > {str: Tool} {
    return {"hadolint": Tool{
        probe       = Command{bin = "hadolint", args = ["--version"]},
        diagnostics = DiagnosticFormat.gnu,
    }};
}

Declare nothing and the output stays prose: a mis-parsed line claims a file and a line that do not exist.

A toolchain spell

An operation is a function from a Target to a Command. magus runs it, caches it, and lets a magusfile compose it.

import "magus/spell";

export fun mgs_getName() > str { return "shellcheck"; }

export fun mgs_listRequiredGlobs() > [Path] { return [Path{value = "**/*.sh"}]; }

export fun mgs_getTools() > {str: Tool} {
    return {"shellcheck": Tool{probe = Command{bin = "shellcheck", args = ["--version"]}}};
}

fun lint(target: Target) > Command {
    return Command{bin = "shellcheck", args = ["--severity", "warning"]};
}

export fun mgs_listTargets() > {str: fun(Target) Command} {
    return {"shellcheck": lint};
}

Add charms to reshape the argv without a second operation - each is an RFC 6902 JSON Patch built by the magus/charm constructors, so adding one later never shifts another's indices.

A provider spell

A provider exports handler ops instead. They take (target: Target, cb: fun(any)), fill a payload map by calling cb, and return data rather than a Command.

Three subsystems accept one, and each detects its ops by name:

A secret provider is the smallest of the three, and the whole contract is one op:

import "magus/spell";
import "os";

export fun mgs_getName() > str { return "onepassword"; }

export fun mgs_getTools() > {str: Tool} {
    return {"op": Tool{probe = Command{bin = "op", args = ["--version"]}}};
}

export fun resolve_secret(target: Target, cb: fun(any)) > str {
    var io = {};
    cb(io);                       // fills io with the payload; io["ref"] is the reference
    final ref = "" + io["ref"];
    return os\exec("op", args: ["read", "op://" + ref], dir: ".", opts: {}).stdout;
}

Make your failures teach. This is the difference between a provider people adopt and one they abandon: someone wiring a backend for the first time hits "not installed", "not authenticated", or "wrong path", and a message naming which one and what to type is worth more than any amount of documentation. Use opts.allow_failure to classify the exit rather than raising a bare status:

if (os\which("op") == "") {
    throw "onepassword: the `op` CLI is not on PATH.\n  mise: mise use -g op";
}
final res = os\exec("op", args: ["read", uri], dir: ".", opts: {"allow_failure": true});
if (res["code"] != 0) { /* classify res["stderr"], then throw something actionable */ }

spells/onepassword/spell.buzz in this repo is the worked version of exactly this.

Testing it

Buzz has in-file test blocks, so a spell is testable without magus running it:

magus buzz -t spells/myspell/spell.buzz

Toolchain spells are mostly assertions about the Command an op returns, which needs no tool installed. Provider spells are worth testing against a stub binary on PATH - that is how this repo's secret provider covers its not-installed, not-signed-in, and wrong-path branches without a 1Password account.

Gotchas worth knowing before you hit them

• magus buzz runs upstream-strict. Every argument after the first must be labeled (os\exec("op", args: [...], dir: ".", opts: {})), and there is no top-level control flow.
• resolve is a reserved keyword, along with yield and resume. Member access on a keyword does not parse, which is why the secret namespace is magus\secret.read.
• str.indexOf returns null when absent, not -1. Comparing the result to a number raises "cannot compare null and int".
• Namespace access is a backslash, object literals use =, and a mut list literal is mut [<str>].

See also

• Spells - what a spell is and how it binds to a project
• Operations - what an op contributes to a target
• Charms - reshaping an op's argv
• Buzz module reference - every host module a spell can import
• Wards - spell-authored diagnostics over resolved ops