zoning: An Import-Topology Gate

• Overview
• Why Not a Code Review?
• Should You Be Using This?
• Support
• Install
• Editor Language
• The Language
  • The Package Block
  • Zones
  • Seals
  • Keeps
  • Grants
  • Structural Laws
  • Variances
• The Seven Laws
• Reading a Map
• Adopting It
• The Verbs
• Wiring It Into CI
• Languages
• Build and Test
• Where This Came From

Overview

Every language hands you some boundary. Go has internal/, Rust has pub(crate), Python has a package graph, TypeScript has an exports map.

Inside a single Zig package there is nothing at all. Every import is a filesystem path, any file may name any other, and because analysis is lazy a genuine import cycle compiles clean. Architecture there is a convention with nothing standing behind it.

zoning is what stands behind it: a package declares the shape it means to have in a .zone file, and the tool judges that declaration against the real @import graph.

✗ zone [acme]: 310 files, 1436 imports, 2 violation(s), 3 allowed
src/core/parser/glean/check.zig:32:1: [zone] zone `parser` imports up into `query` (`core/query/query.zig`) — imports may only point down the stack
src/runtime/cold/emit/render.zig:31:1: [seal] reaches past the seal on `core/scan/` into `core/scan/simd.zig` — enter through `core/scan/scan.zig`

zone: Move the dependency down the stack, or ratify the edge with `variance zone … because "…"`.

seal: Re-export what the caller needs from the seal's entry file, or widen that seal's `open to` list.

Every failure closes with the remedy for its law. A gate whose output does not say what to do next is a gate somebody eventually silences.

It reads the tree, not a build system; there is no project model to configure, no graph to rebuild, nothing to keep in step. The judge stays std-only. The installed executable also carries a narrow, MSRV-checked protocol stack for its in-process language server.

Why Not a Code Review?

A reviewer catches the import that points the wrong way. A reviewer does not catch the fourth one, in the file nobody opened, eleven weeks later, once the person who drew the line has moved on.

Architecture decays by increments that each look reasonable in isolation. That is the whole failure mode, and it is exactly the kind a machine is good at.

The other half is that a declaration is a document. zones { … } read top to bottom is the architecture, in the order it actually stacks, on one screen.

A README that says the same thing drifts. The .zone file cannot: it fails the build the day it stops being true.

Should You Be Using This?

• Java – ArchUnit already does this, over the boundary Java gives you.
• Go, wanting the boundary Go draws – go vet and internal/, and nothing else to install.
• A language that draws no boundary, or a boundary coarser than the one you want – here. Zig has no module system inside a package at all; Python's package graph exists but says nothing about which of your packages may import which — import-linter covers that ground too, if you would rather stay in one ecosystem's tooling.

The dividing line is whether your language already separates the parts you mean to keep apart, at the grain you actually want. Zig draws none; Python's is real but coarser than a zone stack, which is why both dialects exist here.

Support

• Bugs and feature requests go in the issue tracker, with the contract and the output. A verdict without its .zone file is a verdict nobody can reproduce.
• Security vulnerabilities never go in a public issue. Mail griffin@billylives.com instead.
• A governed package usually lives in a repository of its own. An argument about where a boundary should fall belongs there; a wrong reading of the graph belongs here.

Install

The binary ships from crates.io and PyPI, prebuilt, under two names — zone is the command to type, zoning is the same executable installed alongside it for anyone who typed the package name out of habit:

cargo install zoning          # the static binary — installs `zone` and `zoning`
uv tool install zoning        # the same thing, through PyPI
pipx install zoning

Or build it here:

cargo build --release         # target/release/{zone,zoning}

The first interactive run detects Cursor, VS Code, Zed, Neovim, and Vim and installs the matching adapter. It never mutates an editor home under CI, from a non-terminal process, when ZONING_NO_SETUP=1, or while serving LSP. The explicit lifecycle is always available:

zone setup status
zone setup run
zone setup repair
zone setup uninstall

Editor Language

.zone is a real editor language, not a filename with borrowed highlighting. The same executable serves diagnostics and language intelligence:

zone lsp --stdio

Cursor and VS Code receive the exact SVG file identity plus completion, hover, navigation, symbols, folding, semantic tokens, formatting, zone rename, and safe code actions. The embedded VSIX launches the separately installed zoning binary; the extension never downloads another executable.

Zed receives a native extension, Tree-sitter grammar, and the same LSP once its registry submission lands. Zed does not expose a local-extension install CLI, so setup adds zoning to auto_install_extensions without reformatting the rest of the user's JSONC. Its default icon theme submission is separate because Zed forbids language extensions from carrying icon themes.

Vim and Neovim receive file detection, syntax, indentation, folding, and LSP client registration. Neovim 0.11 starts the server natively; Vim connects through an installed supported client. Terminal editors cannot render SVG file icons, so the runtime uses a Nerd Font glyph with Unicode and ASCII fallbacks.

Setup state is versioned and owned by zoning. Repair is idempotent; uninstall removes only files recorded by zoning and leaves unrelated editor settings alone.

The Language

A package opts in by writing <name>.zone at its own root, beside the manifest and every other file that configures it, and here is a whole one:

package acme {
    root     src
    language zig
    facade   root.zig
}

// Low to high. An import may not point up the page.
zones {
    portal   portal.zig
    assay    assay/**
    math     core/math/**
    parser   core/parser/**
    session  runtime/session/**
    ffi      surface/ffi/**
}

seal core/parser through parser.zig     // enter a deep module by its door
keep surface/api.zig to root.zig        // and this region has a guest list

use build_options                       // and these imports may leave
use acme by ffi

limit  reach to 5 hops
forbid cycles across directories

variance zone a.zig -> b.zig
    because "…and here is exactly what would retire this"

Comments are //. Globs mean what they mean to a Python reader; the matcher is CPython's glob.translate(..., recursive=True, include_hidden=True), reimplemented and pinned by tests against that reference.

Where The File Goes

A contract sits at the root of what it organizes: acme/acme.zone governs acme/. That is where the manifest, the formatter config, and the CI config already live, and a boundary tool that demanded its own directory would be asking for a folder to hold one page.

A file inside a contract/ drawer still governs the drawer's parent, because that was the only layout zoning accepted before and contracts are checked in. A tool does not get to invalidate a repository's on-disk shape to tidy its own rules. Both spellings resolve to the same anchor and nothing downstream can tell them apart; zone draft --write only mints the first.

What To Call It

The filename is yours. Identity lives in the package block, which is what every verdict, every --package filter, and every workspace lookup reads, so nothing downstream ever asks what the file was called. zone draft --write mints <name>.zone because a first contract has to be called something and the package's own name is the one guess that is never wrong.

Once a tree has more than one, naming each after its package stops paying. The name is already in the file, on the first line, and it is already in the path; spending the filename on a third copy means every contract in the fleet has a different name for the same kind of document. Naming it for the role instead gives you one word you can grep, one word a reviewer recognises across repositories, and a filename that survives the package being renamed:

None of this is enforced, and it is not meant to be — a tool that dictated filenames would be back to demanding a directory. It is what the sibling repos this was written for settled on after every one of them had a contract named after its own binary.

.zone is also BIND's extension for DNS data, and now that a contract sits where a nameserver file might, the extension cannot be what identifies one. So a contract leads with package or workspace — what it governs, before anything it says about it — and a sweep skips any *.zone that opens with something else. That is a claim of authorship rather than a guess about content, and it only governs sweeping: a file named on the command line is parsed, and faults, like anything else.

The Package Block

root names the source directory, relative to the directory the contract governs. language names the dialect this package is read in, so a monorepo holding several is still one run. facade names the files that may reach anywhere: the module's public face, which by construction re-exports everything and therefore imports everything. exclude drops paths from judgment entirely.

The whole block is optional. A member of a workspace can inherit every setting in it, and then package hearth on one line is the entire header — a mandatory empty { } is exactly the boilerplate the workspace was written to delete.

The file that declares the package — build.zig, pyproject.toml — is never judged as part of it, in any dialect. A build script legitimately imports things no module file may, and a package whose contract made declaring it illegal would be a joke.

Workspaces

Ten packages in one tree are ten copies of the same four lines, and boilerplate is not a cosmetic problem: four lines nobody reads are four lines nobody notices are wrong, and a per-package fact that is really a whole-repository fact will eventually disagree with itself in one package and pass. So a file above them may claim them and say the shared part once:

workspace {
    member   libs/kernels/*
    root     src
    language zig
    facade   root.zig
    use      hyper by core/**
    limit    reach to 1 hop
}

The link points down, the way [workspace] members and [tool.uv.workspace] do: the greater document claims its members, and a package is a member because something above it said so, never because it declared a parent. One less thing to keep in sync, and a package cannot quietly attach itself to a policy nobody granted it. Inheritance is one hop — a member finds the nearest workspace above it that claims it, and that is the whole search, because a chain of overriding defaults is a thing you debug rather than a thing you read.

A member's own word always wins, so inheritance can delete a line somebody had to repeat but can never change the meaning of one they wrote. What a workspace cannot share is anything naming a file: zones, seals, keeps, and variances are claims about one graph, and a blanket exception written once for a whole monorepo is the accretion this language exists to prevent. Nor can a shared grant be scoped by zone name — zones belong to a package and each member's are its own — so use hyper by core/** is a path glob or nothing.

A file may hold both blocks: a root package with members below it. A file that holds only a workspace governs no graph of its own, so a sweep finds it and never offers it up to be judged.

Zones

Declared bottom-up. Each zone is a name and the globs that belong to it, and their order on the page is the stack: a file may import anything at or below its own height, and nothing above.

Declare zones at the granularity the architecture actually has, not the granularity of your top-level folders. Collapsing six kernel tiers into one kernel zone is how math grows a dependency on slate and calls it legal.

Seals

seal <dir> through <entry> says that directory is a deep module: outsiders enter through the entry file and may not reach past it. The entry may live inside the directory (regex/regex.zig) or beside it (sqrt.zig next to sqrt/), because both spellings of "front door" are in use and a contract should not have to care. open to <globs> widens the guest list for one seal.

A seal is the strongest statement in the language, and the one that pays most. It is how "there is exactly one parser for this grammar" stops being a claim and becomes a fact a fork cannot route around.

Keeps

Zones order the stack and say nothing about peers, because peers at one height are unordered by construction. keep <subject> to <importers> is where that silence ends: only the named importers may reach the subject at all.

Two importers are always implicit and need no naming, the region's own insiders and a file's directory sibling, because render.zig importing render_test.zig is an aggregation idiom rather than an architectural crossing.

keep <subject> to nobody is the limit case, for the region whose whole point is that nothing reaches it.

Grants

Everything above governs where an import may point inside the package. use governs the ones that leave:

use build_options            // anywhere in the package
use httpx by session cold    // only these zones

That scope is the reason the law exists. "The CLI face may talk to the network" and "any file in this package may talk to the network" are different architectures, and without a scope they are spelled the same way. A zone stack tells you kernel/ sits under surface/; it will never tell you that a leaf three directories down started dialing an HTTP client, and that is the dependency that ends up hardest to remove.

The standard library is exempt by construction, per dialect — std, builtin, root in Zig; the whole standard library, version-independent, in Python. Every zone has it, no zone chose it, and a contract that spent its lines declaring it would bury the handful of grants that are decisions.

A grant nobody exercises is stale, and stale is a failure. It is a permission somebody forgot to withdraw.

A grant a member inherited is judged once, against the whole workspace. Asking each member to exercise every shared line would make sharing a line strictly worse than repeating it, since one grant of nine members would fail eight times. So it is stale only when no member exercises it, and the report names the workspace that wrote it rather than a package that merely lives under it — and a run that judged only part of the membership (--under, or a contract that would not parse) says nothing at all, because absence of evidence across an unknown remainder is not evidence.

Structural Laws

limit reach to N hops caps how many ../ an import may climb. It is a ceiling you lower deliberately, never one you raise to go green.

forbid cycles across directories bans import cycles that cross a directory boundary, which is the class of cycle a lazy compiler will happily accept.

Variances

Every exception is a variance, and every variance must carry a reason:

variance cycle {
    runtime/cold/engine/serial.zig
    runtime/cold/engine/swarm/swarm.zig
} because
        \\Work distribution recursion: `serial` hands a multi-file query to
        \\`swarm`, and each swarm worker runs the identical per-file path back
        \\in `serial`, which is what keeps the parallel and sequential answers
        \\byte-identical. Retire by extracting that per-file path into a leaf
        \\both import.

The \\ block is a folded paragraph, delimited per line so a missing terminator cannot swallow the rest of the file.

A variance that stops matching is a hard failure. Pay the debt and the build tells you to delete the entry; exception lists in most tools accrete into folklore nobody dares touch, and this one can only shrink.

zone verify --suggest drafts the stanzas for today's violations and writes nothing. A machine can find the edge; it cannot supply the reason, and the reason is the entire value.

The Seven Laws

Seven, and closed on purpose. A boundary language whose vocabulary grows per project stops being a language and becomes a config file.

• zone – an import points up the stack.
• seal – an import reaches past a sealed directory's entry file.
• keep – an importer not on the guest list reaches into a kept region.
• cycle – an import cycle crosses a directory boundary.
• reach – an import climbs more ../ than the ceiling allows.
• escape – an import climbs out of the module root entirely.
• use – a zone imports an outside module no grant covers.

Each exists because a compiler structurally cannot enforce it.

Six of them are claims about how a package's files sit relative to each other, so a single-file module has almost nothing for a contract to say; use and escape are the two that still bind. zone list says so rather than letting you write the file and wonder.

Reading a Map

zone map draws the contract the way gravity works, so "imports point down the page" stops being a rule you memorise and becomes a thing you can see:

zone map · acme · 26 zones, high to low
──────────────────────────────────────────────────────────────────────────
 25 │ ffi      ██············  7 ↓5      surface/ffi/**
 24 │ api      █·············  2 ↓3    ⊘ surface/api.zig surface/api_test.…
 23 │ session  ██████········ 35 ↓13  ⊙  runtime/session/**
 22 │ cold     ███████······· 44 ↓16  ⊙  runtime/cold/**
 21 │ cli      █·············  6 ↓4      surface/cli/**
  …
  8 │ parser   ██████████████ 92 ↓4   ⊙  core/parser/**
  7 │ scan     ███··········· 16 ↓4      core/scan/**
  5 │ math     ███··········· 18 ↓3      core/math/**
  3 │ fault    █·············  1 ↓1      fault.zig
  2 │ assay    █·············  5      ⊙  assay/**
  1 │ portal   █·············  1         portal.zig
──────────────────────────────────────────────────────────────────────────
 310 files · 1436 imports · 5 seals · 3 keeps · 2 grants · reach ≤ 5 hops
 3 ratified variance(s) — each one names what would retire it

The bar is how many files the zone holds. ↓N is how many distinct zones beneath it that zone actually reaches into, which is the number worth staring at: a zone that reaches into every zone below it is not a layer, it is a pile. ⊙ marks a sealed directory, ⊘ an anchored guest list.

Adopting It

Every boundary tool is easy to love on a greenfield package and miserable to adopt on a real one. A tree with nine hundred files has an architecture already — it is simply undeclared — and the first contract somebody writes for it arrives red, which teaches the reader exactly one lesson: the gate is noise.

So don't write the first one. Take it:

zone list                     # every package here, governed or not
zone draft . --write          # the contract this graph already obeys
zone verify                   # green, on the first run

list prints the exact draft invocation for each ungoverned package, so the middle line is a paste rather than a guess — and it names the package what the package's own manifest names it, not what its directory happens to be called.

draft derives the stack from a topological sort over directories, the grants from the modules the code already imports, and the reach ceiling from the reach the tree actually needs. Everything it emits is true today. Then the cleanup begins, and every step of it — merge two zones, seal a directory, drop a grant, lower the ceiling — is a decision somebody made on purpose rather than a fight with a wall.

It refuses to guess at two things. Seals and keeps are claims — "this directory is a deep module", "these peers are independent" — and a machine inferring them from today's call sites would guess wrong the first time somebody adds a second legitimate caller. And a real import cycle comes out as a variance stanza with an empty reason, which does not parse: a draft over a genuinely tangled package cannot be adopted until a person has written why each tangle stays.

Directories that import each other cannot be ordered, so they land in one zone, and that zone is called tangle. A zone nobody enjoys reading is a zone somebody eventually splits.

The question you actually have day to day is narrower, and explain answers it without a contract edit or a full run:

zone explain src/runtime/cold/emit/render.zig         # where does this file stand?
zone explain src/core/math/sqrt.zig src/portal.zig    # may I write this import?

The second form works whether or not the import exists yet, which is the point: every tool in this class makes you write the line and run the whole gate to find out. Paths are taken as typed, resolved against your shell's own directory — the way an editor tab has it. And when a path is real but unjudged it says which of the four reasons applies, because "untracked by git" and "excluded by the contract" call for opposite actions.

The verdict reaches the exit code, so the question is also a shell question:

zone explain from.zig to.zig && $EDITOR from.zig

The Verbs

• verify – does the code obey the contract. The default, and what CI runs.
• status – verify, plus the census: files per zone, hop histogram, what is sealable for free, and the entry-file bypass count for everything still open.
• list – every package in the tree, governed or not, with the next command for each.
• show – the resolved contract, as zoning understood it rather than as you typed it.
• map – the stack, drawn.
• explain FILE – one file's zone, reach, grants, and importers.
• explain FROM TO – whether that one import is legal, and the clause that decides.
• draft DIR – the contract DIR's graph already obeys. --write files it at DIR's root, and refuses if either layout already governs DIR.

Options: --package NAME, --under DIR (monorepos), --root PATH, --language NAME, --complete, --write, --untracked, --suggest, --json, --no-color.

Exit codes: 0 clean, 1 a violation or a stale declaration, 2 the contract or the invocation is malformed. explain uses the same three, so 1 there means the file is in violation or the import is not allowed.

Wiring It Into CI

One step, and no Rust toolchain in the job:

- name: Import topology
  run: uv run --no-project --with zoning==1.1.0 zone verify --complete

No build, no compile database, no network beyond fetching itself. Pin the version: a gate whose verdict can change without a commit is not a gate.

--complete adds the one claim no law can make: every package in scope has a contract. Without it a clean run says nothing whatsoever about the package somebody added last week, and adoption that cannot notice a new ungoverned package rots back toward zero, one package at a time.

It forgives a vendored dependency, and on the right authority. A vendored package is a package by every test this tool can run — manifest, source, an import graph — and it is nonetheless not yours: its architecture is decided in the repository it came from, which is where its contract lives. The obvious fix is an allowlist, which drifts the moment somebody vendors a second thing, so the dialect reads the manifest instead. build.zig.zon spells one .vendor = .{ .path = "vendor" }, and a build that had not said so would not link. The fact is already written down and the compiler maintains it.

With a toolchain already in the job, the same binary comes from crates.io:

- run: cargo install zoning --locked && zone verify --complete

Languages

Zig first, because Zig is where the absence of any boundary is total. Nothing above is about Zig, though: zones, seals, keeps, cycles, and reach are statements about a graph of files, and a graph of files has the same shape in every language.

So the language-specific surface is deliberately tiny, and a Dialect carries only what genuinely varies:

• which extensions are source,
• how an import is spelled,
• whether a given spec names a path inside the module or a dependency outside it,
• the comment and string conventions, so imports are read from code alone rather than from a line that mentions one,
• which filenames declare a package, so list can find one nobody has governed,
• what name a manifest gives the package it declares, so a contract is called what the build already calls it,
• which modules the language always provides, so no contract spends a use line on the standard library,
• and which directories a manifest calls an in-tree dependency, so coverage knows whose package is whose.

Resolution, the graph, all seven laws, and every rendering are shared. That is on purpose: a dialect that could resolve paths its own way is a dialect that could disagree with the others about what a cycle is.

A contract names its own language, so a polyglot monorepo is one run rather than one run per dialect. --language sets the default for a package that has not said.

Python is the second dialect, and it earns its keep on the same absence: a package's imports reach any module on sys.path, and nothing about the language says which of your packages may reach which. from a.b import c and import a.b.c both resolve dotted names against real files — a module or a package's __init__.py, whichever exists — and a relative import's leading dots climb exactly as many directories as they count, independent of how deep from nests. The standard library grant is a fixed list read out of sys across supported versions, not a per-repo guess, so zone verify never asks a contract to use json or pathlib: those names are not the repository's to grant, in any Python version the interpreter has shipped for years and will keep shipping.

Build and Test

Everything is cargo, except the parity gate:

cargo build --release
cargo test
cargo clippy --all-targets

tools/differential.py is the parity gate against the Python implementation this was rewritten from. It mutates each real contract the way a person breaks one; drop every seal, drop every guest list, squeeze the reach ceiling, revoke every variance, invert the entire stack; and requires both implementations to produce the same set of findings, law by law and file by file. It covers the six laws both implementations have; use postdates the rewrite and has no Python twin to check it against. crates/zoning/tests/properties.rs closes that gap with an independent oracle instead: the_use_law_flags_exactly_the_outside_imports_no_grant_covers grows randomized packages with real outside imports, drafts a randomized grant table over them, and hand-computes which imports should be refused — the same role the_cycle_law_finds_what_a_slower_algorithm_finds plays for cycle (Tarjan checked against an O(n³) fixed-point relaxation). use is pinned by fixtures and generative coverage now, not fixtures alone.

Agreeing on a clean tree proves nothing, because every gate agrees that nothing is wrong.

Where This Came From

This began as a Python tool called ward, inside a monorepo, guarding one Zig package. It was correct and it was invisible: a gate that lives in your repo is a gate only your repo runs, and the four packages that most needed it were the four that had been split out into repositories of their own.

The .zone files were sitting there in each of them, declaring boundaries, judged by nobody. The rewrite is a static binary with no dependencies for exactly that reason: a gate has to be cheaper to install than to ignore.

Apache-2.0. Built at The Billy Company.