m-cli

The cross-engine M toolchain — the m busybox. A single static Go binary that delivers fmt/lint/lsp/test/coverage/watch for M (MUMPS) source, working across YottaDB and IRIS (spec §1). It's the Go rewrite of the Python m-cli, built on the m-parse substrate (tree-sitter-m via wazero — no CGO).

Status. The native commands fmt/lint/lsp/test/coverage/watch are built, and m is now the busybox dispatcher: it fronts the sibling binaries irissync (IRIS source axis) and kids-vc (KIDS round-trip) under one m, with one aggregated m schema. See the implementation plan.

m fmt --rules=canonical --check .    # CI gate: exit 3 if any file differs
m lint --profile xindex routine.m    # query-driven rules over the parse tree
m test --engine ydb …                # run *TST.m suites through the engine
m pull --full                        # → irissync pull (materialize IRIS source)
m kids decompose patch.KID           # → kids-vc decompose
m schema | jq .                      # the full aggregated tree as JSON

---

m fmt

An AST-preserving formatter over the m-parse syntax tree (spec §3.1). It works as edits-over-source: rules emit byte-span edits guided by the parse tree, applied to the original bytes. Two key properties:

• Identity by default. With the default --rules=identity, nothing changes — unformatted input round-trips byte-for-byte. Formatting is opt-in (mirrors the Python m-cli's identity default + canonical layer).
• AST-preserving. Rules only change what they must (e.g. keyword letter-case), verified by an internal tree-shape check: parse(format(src)) has the same shape as parse(src).

Preset (--rules)	What it does
identity (default)	No-op; the round-trip baseline.
canonical	uppercase-command-keywords — set→SET, w→W, leaving arguments untouched. (More rules + the pythonic/pythonic-lower/compact/sac presets follow.)

Flags: --check (report files needing formatting; exit 3 if any — the CI gate), --write/-w (rewrite in place), --stdin (format stdin → stdout as a raw filter). With no flags it reports what would change (exit 0).

File discovery walks paths (default .) for .m / .mac / .int — .int is included because VistA loaded via ^%RI stores its routine source as .int (there .int is the source, not compiler output). Explicit file arguments are formatted as given.

The busybox — dispatched subcommands

m keeps its native commands and dispatches the rest to sibling binaries, so each sibling keeps its own SBOM and release cadence — a small attestable family, not one mixed-dep blob (spec §2.2, ADR §5):

m command	Forwards to	Notes
m list · m pull · m status · m verify · m push	irissync	the IRIS source axis (push is the sole DB writer)
m kids <decompose|assemble|roundtrip|canonicalize|lint|parse>	kids-vc	KIDS round-trip + PIKS data-class gate

• Discovery. Each sibling is resolved in order: the M_<NAME>_BIN override (e.g. M_IRISSYNC_BIN, M_KIDS_VC_BIN), then alongside the m binary, then $PATH. A miss is a deterministic SIBLING_NOT_FOUND error object (exit 1) — never a panic or a raw exec error.
• Faithful passthrough. Args, stdin/stdout/stderr, and the child's exit code pass through unchanged; the toolchain-wide globals (--output, --verbose, --no-color) are re-forwarded so the sibling renders identically.
• One tree. m schema aggregates each available sibling's sub-schema under its dispatched namespace, so an agent sees a single command tree. Siblings that aren't installed degrade to a discoverable stub (the schema stays valid).

Extension point. Adding a dispatched namespace (e.g. m meta … → vista-meta, deferred) is a one-line Spec in internal/dispatch; nothing else changes. Discovery, forwarding, and schema aggregation all key off that registry.

Architecture

   m fmt … ──► discover .m/.mac/.int ──► for each file:
                                           parse (m-parse: tree-sitter-m via
                                                  wazero, embedded WASM, no CGO)
                                           rules emit byte-span edits ──► apply
                                           ──► --check (exit 3) / --write / report

The whole binary stays static (CGO_ENABLED=0) because the parser is the embedded grammar WASM run in wazero (m-parse), not a CGO tree-sitter binding.

Repository layout

m-cli/
├── main.go                 # the `m` CLI grammar (Kong struct) + `m fmt`, version, schema
├── internal/mfmt/          # the formatter
│   ├── format.go           #   Format() + the edits-over-source engine (applyEdits)
│   ├── rules.go            #   Rule interface; identity/canonical presets; uppercase-command-keywords
│   └── shape.go            #   SameShape — the AST-preserving check
├── clikit/                 # shared CLI conventions (from go-cli-template)
├── Makefile · .golangci.yml · .github/workflows/ci.yml
└── LICENSE · NOTICE        # Apache-2.0 (Go); see Licensing

Build, test, CI

Target	What it does
make build	dist/m, static (CGO_ENABLED=0), -trimpath, version-stamped
make test	go test -race -cover ./... (race needs CGO; the rest is CGO-free)
make lint	golangci-lint run ./...
make schema	build + emit the JSON schema (CI conformance artifact)
make dist	cross-compile linux/{amd64,arm64}, darwin/arm64, windows/amd64

CI (the org's reusable go-ci workflow) runs golangci-lint, race tests, the schema contract, and a static CGO_ENABLED=0 cross-compile matrix.

Licensing

The Go code here is Apache-2.0 (LICENSE/NOTICE). The binary links m-parse, whose embedded grammar WASM is currently AGPL-3.0, so a built m transitively includes an AGPL-derived artifact. Per project policy, all licensing reconciliation is deferred to project completion — end-state Apache-2.0 for every artifact except the VS Code extensions (MIT). The interim AGPL status is not a blocker. See the m-parse NOTICE and the toolchain spec.