feat(make-run): resolve and run a repo's make targets from its own Makefile

[hunterdsp]

Why

Anything that needs a project's command — CI docs, skill docs, tooling — has been transcribing it by hand, and transcriptions drift silently. One example, found while writing this:

source	what it says make docs runs
mcp-store skills/zensical.md:131	uv run zensical build --clean
standard.mk:353 default	uv run --group dev zensical build --clean --strict
live doppler	uv run --group docs zensical build --clean --strict

Three answers, two of them wrong, and nothing that makes them disagree out loud.

A repo bases its Makefile on standard.mk, which supplies the sanctioned defaults with ?=; the repo then overrides what it must. This library reads the result of that layering rather than re-deriving it — so there is no second copy to vendor and nothing to keep in sync.

What

make-run.sh is a sourced library, so jbx dispatches its functions against whatever repo you point it at:

jbx just-bashit:make-run mk-var -C ~/doppler DOCS_BUILD_CMD
# uv run --group docs zensical build --clean --strict

function	purpose
mk-var [-C DIR] NAME	one variable, fully expanded after defaults + overrides
mk-vars [-C DIR] [REGEX]	the whole command surface in one make invocation
mk-run [-C DIR] TARGET	run a target using that repo's own recipe
mk-has / mk-origin	the predicates the above are built on

mk-vars deliberately carries no built-in list of interesting variables — such a list would be exactly the hand-maintained copy this abolishes. A test asserts that a variable the library has never heard of still appears.

Two properties the tests pin down

Both exist so that an empty result means something specific rather than nothing:

• An undefined variable is an error, never an empty string. DOCS_CHECK_CMD ?= is legitimately empty; a typo'd name is not. Collapsing the two is how a blank command comes to read as a real one. $(origin) is the only reliable discriminator.
• An unparseable makefile fails loudly. make -pRrq exits 1 in question mode whenever the default goal is out of date — normal — but exits 2 when it cannot read the makefiles at all. Treating those alike hands the caller an empty database, and the empty answer then reads as real.

Implementation notes

• Queries never run a recipe. Values resolve inside a sentinel target's recipe, because a --eval string is parsed before any makefile is read, so variables referenced there are still empty.
• The database is captured before any pipe. make -pRrq's exit 1 would otherwise poison every pipeline under set -o pipefail, which bats and most CI set — the failure mode that passes in a plain shell and breaks in CI.
• The test fixture builds its makefiles with printf, not a <<-EOF heredoc: <<- strips leading tabs, and a make recipe is defined by its leading tab.

Verification

• make test — 321/321, including 13 new
• make lint — clean (shellcheck, shfmt, standard-check, help-check, ghost-check, gates-check)
• mk-var/mk-vars/mk-run exercised against live doppler, just-makeit, and just-bashit

Once this lands on main, the org Pages mirror picks it up automatically — mirror.yml copies src/just_bashit/*.sh into jbs/ by glob, so nothing needs registering.

🤖 Generated with Claude Code

[hunterdsp]

Pushed 76b302f — a portability fix found while adding workflow_dispatch in #22.

--eval is GNU make 3.82+ (NEWS, Version 3.82, 28 Jul 2010), and macOS ships 3.81 as /usr/bin/make. The test-macos job installs bash coreutils diffutils bats-core — not make — so make test there runs 3.81 and every query in this library would have failed on that one runner while passing everywhere else.

The expression now goes into a sentinel target in a throwaway makefile, handed to make alongside the repo's own via -f. That option has always existed. -f disables make's automatic makefile search, so the repo's own makefile is named explicitly — otherwise the query resolves against nothing but the sentinel and reports every variable undefined. $(info ...) needed no change; it landed in 3.81.

A test now asserts the library stays off --eval, since the only other thing that would catch a regression is a macOS-only CI failure.

Worth noting how this was found: CI is still dark from the webhook throttle, so this would have shipped green-on-my-machine and broken on merge. It came out of reading the matrix by hand rather than from any run. Local verification: 322/322 (14 in this file), make lint clean.