cli

modelspec

modelspec validates ModelSpec models and exports them to their JSON interchange form. It is one static binary that needs no project, no network and no other tool.

• modelspec lint checks models written in HCL (*.modelspec.hcl) or in the JSON form (*.modelspec.json).
• modelspec export writes the JSON form of an HCL model, and export --check fails when a committed JSON file is not what its HCL exports to.
• modelspec version and modelspec self-update.

The checks are also an importable Go library, github.com/modelspec-org/cli/pkg/modelspec.

Install

Pin a release in CI by version and SHA-256. Every release publishes its archives and modelspec_<version>_checksums.txt on the releases page; take the SHA-256 of your platform's archive from that file and keep it in your repository next to the version:

VERSION=X.Y.Z    # the release you pin, for example 0.1.0
SHA256=...       # that archive's line in modelspec_${VERSION}_checksums.txt
ARCHIVE="modelspec_${VERSION}_linux_amd64.tar.gz"   # or darwin_arm64, darwin_amd64, linux_arm64
curl -fsSL -o "$ARCHIVE" "https://github.com/modelspec-org/cli/releases/download/v${VERSION}/${ARCHIVE}"
echo "${SHA256}  ${ARCHIVE}" | sha256sum -c -      # on macOS: shasum -a 256 -c -
tar -xzf "$ARCHIVE" modelspec
./modelspec version

Archives exist for linux, darwin and windows on amd64 and arm64 (no windows/arm64); windows archives are .zip.

With a Go toolchain:

go install github.com/modelspec-org/cli/cmd/modelspec@vX.Y.Z

A binary installed from a release archive updates itself with modelspec self-update (--check only reports; see modelspec self-update --help).

Lint

modelspec lint                                   # the current directory, recursively
modelspec lint model/                            # a directory
modelspec lint model/chinook.modelspec.hcl model/chinook.modelspec.json
modelspec lint spec/                             # a SpecScore tree
modelspec lint sales.modelspec.hcl --module core=shared/core/
modelspec lint --profile publish --format json models/ | jq .

Each path is a file or a directory. A directory is searched recursively for *.modelspec.hcl and *.modelspec.json (and, inside a SpecScore models/ directory, for any *.hcl). Hidden directories and node_modules are skipped. A search follows no symbolic link, to a directory or to a file (a repository can hold a link to /dev/zero, or to a file outside it, and lint . on someone else's branch must not read it): a link that would have been a model file is an error (rule skipped-file, exit 1, under both profiles) that names it, and so is a model file that is not a regular file (a named pipe, a device). It is an error, not a warning, because the run did not check what it was asked to: a warning would let an invalid model reached through a link pass, and let export write half a module. The module the file belongs to is not checked further, so what the missing file would have answered is not reported as an unresolved reference; the other modules of the run are checked, and export and export --check refuse the module (exit 1, the finding on standard error). Replace the link with the file, or name it on the command line. A file named on the command line may be a symbolic link to a regular file and is read through; one that is not a regular file (a device, a pipe, a directory, or a link to one) is an error, exit 2, whose message says what it is. A file reached by two names (a relative and an absolute path, a symbolic link named on the command line) is read once.

Text output, sorted by file, line and rule, then a summary:

model/shop.modelspec.hcl:12: error: entity "Order" property "customer" entity reference "Customer" does not resolve to an entity in module "shop" [reference]
failed: 1 file checked, 1 error, 0 warnings

--format json prints one object: files, errors, warnings, notes and findings, each finding with file, line (omitted when unknown), rule, severity and message.

A module is the unit of checking. Given one file of a module that has more files on disk (a SpecScore models directory, or X.modelspec.hcl and its X.modelspec.json), lint loads and checks the whole module and reports the other files' findings with their paths; a note: line (the notes array in JSON) says so once per module. The same module reached through two files is checked once. export loads a module the same way. A module directory that cannot be listed is an error (exit 2), not a quiet partial check.

Exit codes

With --format json, every exit-2 path also writes one JSON object to standard output, {"error": "...", "exit": 2}, so a script never finds an empty or success-looking standard output; the message is on standard error as well. modelspec makes no network request except self-update, and sends no telemetry.

Profiles

The default profile implements the standard as written in specscore/modelspec (spec/core-model.md, spec/hcl-authoring.md, spec/json-format.md and the decisions) and nothing else. A model the standard allows is never refused because one consumer's reader is narrower. Narrowing for one consumer is opt-in.

--profile publish adds the rules a model needs to be listed in public catalogues. These are exactly what the JSON reader of openvaultdb/directory (parseModelSpec) requires on top of the standard; the corpus tests run that reader's recorded verdicts against this profile.

The publish profile includes every rule of the default profile.

Modules

A module is a set of files, not one file. Concept names are unique per module across all its files, and a reference resolves against the whole module.

1. --module <name>=<path> (repeatable; a file or a directory, any .hcl name) assigns files to module <name> and wins over the other rules. Assigned files are linted too, so modelspec lint --module core=shared/ is a complete command.
2. SpecScore layout. Every *.hcl file directly inside …/modules/<id>/models/ belongs to module <id>, whatever the files are called. The layout is detected from the path of each file as given or found (symbolic links are not resolved first), with no need for the directory to be searched: a single file named on the command line is enough.
3. Standalone. <name>.modelspec.hcl is module <name> by itself. A JSON file is module module.name, or its file name without .modelspec.json when it has none.

--module wins in every case: a file assigned to a module is in that module whatever its name or place, and a JSON file in an assigned directory joins the module as a copy of it. Assigning part of a layout module's directory, or one file to two modules, is a usage error (exit 2), because it would split a module.

X.modelspec.json beside X.modelspec.hcl is the interchange copy of the same module, not a second module (this is the layout export --out produces); so is any JSON file in the models directory of a layout module, or in the directory of an explicit module that has HCL. Both are linted, references into the module resolve to the HCL, and the pair is never "ambiguous". A copy is checked against its HCL: when it is not what the HCL exports to, stale-twin (a warning) says so and which key differs. The copy's own module.name is accepted for references inside it. Two genuinely different sources claiming one module name are an error where the module is referenced.

A module-qualified reference such as entity = "core.Space" (decision 0014) resolves against the modules among the files linted together. When the module is not among them, the error says so and how to supply it (--module core=<path>, or name its files too). A key or a bind sees the fields a component contributes through use, including a component of another module when that module is supplied; when it is not, the key and bind are not judged.

What the default profile checks

Both forms, on the same typed model:

A JSON document needs no entities, and an entity needs no properties: a module of components and enums is valid, and so is an entity whose key comes from a component it uses. JSON that export writes from HCL that lints clean lints clean (a test runs that round trip over the whole accepting HCL corpus).

What the default profile does not check

• Attributes. The accepted attributes are the lists in spec/core-model.md (types, constraints, bind, source, query, kind, key, use, values). The standard says attributes include "similar metadata", so an attribute outside those lists is refused; if you need one, that is a point for the standard (see "Open points").
• Values it only reads. pattern is not compiled, format is not checked against a list, query text and recordset source expressions are not interpreted, an enum's values are not checked against the type of the property it constrains.
• Constructs with no defined JSON form. The bodies of index, projection and migration blocks are not read (the blocks lint clean; export refuses them), and the contents of the JSON projections and migrations objects are carried through unchecked.
• Other repositories. A module-qualified reference is resolved only against the files linted together.
• SpecScore's own rules, such as a module's dependsOn, and any comparison with data.

Literal values only, and limits

What a value can be. ModelSpec v0 HCL has literal values only: decision 0009 and spec/hcl-authoring.md ("V0 Grammar Scope") allow singular named blocks, attributes, and strings, numbers, booleans and lists (object literals only "where explicitly specified"; the standard specifies none, a search of spec/ and examples/ finds only the non-canonical properties = { … }, so modelspec reads one and refuses it with the map-style message of decision 0007), and "do not use dynamic HCL expressions or functions". A reference to another module is a string (entity = "core.Space", decision 0014), so it is allowed. modelspec takes that to mean this grammar:

• block types and labels (a label is a quoted string; a bare word is read as a label too);
• attribute names, =, and these values: a quoted string or a heredoc, a number (- only as its sign), true, false, null (read, and refused as "not a ModelSpec value"), and a list of those, with commas, newlines and comments between items; a bare word (a reference) is read and refused with the message that it must be a literal.

What is refused before parsing. The HCL parser is recursive, and a stack overflow in Go is fatal, so the source is lexed first (the lexer does not recurse) and refused, without being parsed, if it holds any token a literal cannot contain. That closes the whole class instead of the recursive constructs one by one. Each is a finding with rule literal that names the construct and the line (at most one for each construct on a line, and 50 for a file), exit 1:

Brackets, quotes and operators inside strings, heredocs and comments are text, so a file whose pattern = <<EOT … [[[ … EOT holds unbalanced brackets is read normally.

What is bounded. One table, every number in the code:

Output is bounded. One run lists at most 1,001 findings (the 1,000 kept and the one that counts the rest), each at most 1,024 bytes of message, plus its path, line number, severity and rule (at most 64 bytes besides the path): at most 1,001 x (1,088 + the length of the path) bytes, about 1.1 MB, in the text format; in --format json a byte can be written as six (\u003c), so at most about 6.5 MB. A test builds a model with 1,001 findings that echo the longest names and values and asserts the bound (it gives 395 KB).

Real models nest four or five levels in HCL and six to eight in JSON, and no query or pattern comes near a thousand lines.

Why strings and heredocs are linear. The HCL parser joins the pieces of a string or heredoc one at a time, copying the text and shifting the list of pieces each time, which takes time that grows with the square of their number; and the lexer starts a new piece at each $ and % (and, in a heredoc, at each line). A pattern of 200,000 $a (400 KB) took 24 seconds, one of 1 MiB nearly three minutes, one of 4 MiB of $${ nearly seven. modelspec rewrites the $ and % inside literal text before the parser sees it, so that they stay inside their piece (in a quoted string as the escapes \u0024 and \u0025, which the parser reads as the same text; in a heredoc, which has no escapes, as two private-use characters that are turned back into $ and % in the value), and the pieces left are the lines of a heredoc, which are capped. A backslash and the character after it in a quoted string are one unit that the rewrite does not touch, so "\$" reaches the library as written and is refused (Invalid escape sequence) as it is without the rewrite; and in a heredoc a carriage return right after a $ or % stays with it, as the library's lexer takes it. A test compares the CLI with the unmodified library in both directions over 800 generated strings and heredocs full of $, %, their escapes, a backslash before every kind of character, carriage returns and the private-use characters themselves: whatever the library refuses the CLI refuses, and whatever the library reads the CLI reads to the same value (374 were read by both, with the same value, and 426 were refused by both); there is no known input on which they differ. Over the corpus and 5.8 million valid files of an earlier review the values were identical; that is a measurement, not a proof. A string of 520,000 $a (1 MiB) takes 0.3 seconds.

The checker is linear too. A reference used to be found by scanning every concept of the module, and an entity's properties rebuilt for every bind that named it: valid models of 4 MB took minutes (see the table below). Concepts are now indexed once for each module and the properties of an entity once for each entity (the fields of a component are kept once, not copied into each entity that uses it). Tests count the memory allocated at two sizes of each shape, which does not depend on the load of the machine, and fail when eight times the model costs more than fourteen times the memory. A scan allocates nothing, so other tests count steps: concepts visited to build a module's index (each exactly once, however many lookups there are), members listed to build a set (once), and sets and names consulted to answer whether a name is a property of an entity (bounded by twice the fields of the components it uses, however many lookups).

Numbers and names. The HCL library reads a number into 512 bits (about 154 digits) and takes time that grows with the exponent: values = [1e10000000], 37 bytes, took 12 seconds and lint passed it; an integer of 4.19 million digits took 12 seconds; two integers that differ after the 154th digit were read as equal; a name of 2 million characters was repeated in every finding that mentioned it. So modelspec bounds the token and not the behaviour after it, in precheck, from the lexer's tokens, before the parser runs (and in the JSON reader as it reads): a number of at most 40 characters and, written as digits (without trailing zeros) times a power of ten, an exponent of at most 100 either way, so every number that is accepted is read exactly (40 digits need 133 bits); and a name of at most 255 bytes. The standard sets neither limit, and a model that exceeds one is valid by the standard and refused here, with a limit finding that says what was counted and the limit. Where databases limit an identifier it is between 63 and 128 bytes, so 255 leaves room; no number in a real model comes near 40 characters or 1e100. The JSON form has the same two limits (every object key and the same name strings are names), so a model and its twin are refused alike. Text that is not a name (a pattern, a format, a query, a source, an enum value) is limited only by the size of the file; whatever of it a message repeats is cut.

One form for a number. Writing a number with a fraction or a negative exponent used to cost 20 to 35 microseconds (the library prints a 512-bit value as the shortest decimal), so a megabyte of 1e-99, took 5.7 seconds; and a number had no single form (-0 was 0 in HCL and stayed -0 in JSON, JSON kept 1e2 and 1.0 as written, and 1e41 to 1e100 were exported as integers of 42 to 101 digits that lint then refused). One function now turns the text of an accepted number into its canonical form by string arithmetic, with no floating point, and both readers, the comparison of equal values (duplicate enum values), the comparison of a JSON copy with its HCL and export all use it. A number is read as digits times a power of ten; the canonical form is 0 for zero (no sign), the plain decimal when it has at most 40 characters, and otherwise the digits, an e and the exponent: 1, 1.0, 1e0 and 10e-1 are 1; -0 is 0; 1e+100 is 1e100; 1e40 is 1e40 and 1e39 is a 1 and 39 zeros; 0.5 is 0.5 and 15e-50 is 15e-50. export writes that form, so a model exported and read back is the same model and is within the limits it was read under. JSON numbers are the same: 1e2 and 1.0 are the integers 100 and 1 there too. The one change this brings to what export wrote is for a number of more than 40 characters as an integer or decimal, which it used to write in full and now writes with an exponent; none in the corpus, the standard's examples or Chinook changes a byte.

The exponent limit is on that normal form, not on the exponent as written: 100e99 is refused (it is 1e101, and the message says so: "the number 100e99 is 1e101, whose exponent 101 is past the limit of 100"), 0.1e101 is accepted (it is 1e100), and so are 10e99 and 1e100. Zero keeps the exponent it was written with, so 0e100 is accepted and 0e101 and 0e2147483648 are refused alike.

Time and memory, measured. modelspec lint of the inputs below, every one within the 1 MiB limit, at the head of this change, on a laptop shared with other work (load average 2.8 to 4.2, so the figures are upper bounds; the peak is the process's resident memory). The first table is the shapes that were once slow, the second what the limits and the reader refuse:

A list of numbers, enum "E" { values = [ ... ] }, by size (seconds and peak memory; before is the previous head, which printed each number as the shortest decimal of its 512 bits, after is the canonical form):

The slowest input measured is 0.44 seconds and the largest peak 521 MiB, both for a megabyte of 1, (an enum of 524,274 repeats of it, exactly 1 MiB; half the size peaks at 228 MiB), measured at a load average of 3.0; the review measured 0.48 to 0.49 seconds and 564 to 577 MiB for it. Every other input in the tables is at most 0.30 seconds and 380 MiB. The lexer holds every token of a file at once, so memory is linear in the number of tokens, about 420 to 500 MiB for each MB of one-byte tokens (the limit of 1 MiB is what bounds that, and what bounded it at 1.7 to 1.9 GiB when it was 4 MiB). (values takes integers, so a list of fractions is a finding for each item, and each number is still read and written, which is the cost measured.) These are measurements of the inputs listed, not every input, and no time or memory is proved: the HCL library could have other paths that are slower than linear, and a new version of it needs the fuzz run below and these inputs again. A directory of such files takes the sum.

What is and is not proved. The checks are a test of the tokens, so nothing recursive in the HCL parser is reachable except the nesting of braces and brackets, which is bounded. There is a test that runs ParseHCL with the process stack limited to 4 MiB on 3,000 repeats of every construct that made the parser recurse (it fails by overflowing the stack if the pre-parse refusal is removed), and one input for each refused token. A new version of the hcl library, which could add tokens or recursion, needs scripts/fuzz.sh [seconds] (fuzz targets for the HCL and the JSON readers in scripts/fuzz/; oracles: no crash, publish refuses whatever the default profile refuses, a clean model exports to JSON that parses and lints clean, and one input costs a bounded amount of work: at most 4 MiB plus 1,000 bytes allocated for each byte of input, counted by the allocator and not by time, and a watchdog ends the process, which the fuzzer reports as a failing input, if one input is still running after ten seconds: while it runs, not after it returns). scripts/fuzz.sh runs the fuzzer with -fuzzminimizetime 5s: by default it minimises a new input for up to a minute, with no execution reported (30 to 42 seconds of nothing, with two workers), which no judge can tell from a stall. A whole run that stalls is judged by scripts/fuzz.sh through cmd/fuzzjudge (internal/fuzzjudge, with table tests over synthetic logs): it fails when nothing ran for 30 seconds anywhere in the run, including its end, or when the rate of the second half fell below 30% of the first. The fuzzer's own pauses (12 to 18 seconds with no execution) and a rate that halves from one half to the other (seen in real runs) pass. A stall of one input in one worker of several leaves most of the rate, so it is the watchdog that catches it, not the judge. The fuzz targets are skipped in go test and are not in the coverage gate.

Export

modelspec export model/chinook.modelspec.hcl \
  --module-id github.com/acme/chinook/model/chinook --module-name chinook --module-version 0.1.0 \
  --out model/chinook.modelspec.json

modelspec export --check model/chinook.modelspec.hcl model/chinook.modelspec.json

export lints the file first, as lint does under the default profile (so its module is checked whole), and refuses a file with errors (exit 1, the findings for that file on standard error); --check refuses an invalid model too, so a check cannot pass on one. It exports any .hcl file of a layout module, whatever it is called, and refuses a JSON file given as the source. A file that refers to other modules needs them supplied with --module <name>=<path> (they are used to resolve references and are not exported).

The JSON follows spec/json-format.md: modelspec, module, then components, enums, entities, collections, recordsets (concepts in source order, attributes in source order, recordset columns as an ordered array with name first). The output of export for datatug/chinookdb's model/chinook.modelspec.hcl is byte-identical to its committed model/chinook.modelspec.json.

Numbers are written in their canonical form (see "One form for a number"): 1.0 and 1e0 are 1, -0 is 0, and a number that would need more than 40 characters as a plain decimal is written as digits and an exponent (1e41), which lint reads back; so are the numbers of a JSON file when it is read, and --check compares them in that form.

--check compares documents, not bytes: whitespace does not count, but the order of keys in objects and of items in arrays does (the Directory compares a model with its registered copy the same way). The module identity is read from the committed file unless --module-* flags are given. --out cannot be combined with --check.

--out writes a regular file. The path is looked at without following a link, and one that exists and is not a regular file is refused (exit 2, nothing written, and the message says what it is): a symbolic link (even to a regular file, which would send the write elsewhere), a named pipe (which would wait for a reader), a directory, a device. /dev/null is a device, so --out /dev/null is refused; write to standard output and redirect it (export ... > /dev/null). A path that does not exist, and a regular file, are written as before.

Every file a command reads has the 1 MiB limit, the committed JSON that --check reads too (one over it is refused without being read, exit 1), and every one is read through one reader that refuses what is not a regular file (exit 2) and reads at most one byte over the limit. export refuses to write a JSON twin over the limit, since lint would refuse it: valid HCL of 0.7 to 1.0 MB can export to 1.2 to 1.7 MB (1.3 to 1.7 times the size). It exits 1, writes nothing, and says why.

Open points

The standard does not settle these, so modelspec does not invent an answer (each is listed, with what modelspec does today, in specscore/modelspec#8):

• Module identity. HCL has no place for module.id, module.name and module.version (spec/hcl-authoring.md, "Open Questions"), so export takes --module-id and --module-version (the format requires those two) and --module-name if you want one (it is written without being given when the model refers to its own module by name, so the JSON lints clean whatever file it is saved as).
• The module short name outside a SpecScore layout. lint takes it from the file name (<name>.modelspec.hcl), from module.name in JSON, or from --module.
• index, projection and migration blocks. spec/core-model.md shows index, decision 0009 lists projection, and spec/migration-metadata.md shows migration, but no document defines their HCL attributes or how they map to the JSON projections and migrations objects. A file with one lints clean and cannot be exported.
• Several files in one module. The JSON form is one document per module and nothing says how the files of a module merge, so a file that is one of several files of a module is not exported (the other files are named in the refusal). The module is linted whole whichever file is given.
• Names. Empty and blank names are errors (nothing could refer to them), names that differ only by case are warnings, and a property name with a dot is accepted although bind and migration renames use dots as separators.
• Integer enum values are accepted (decision 0013 mentions int properties); the format does not say what an enum's values may be.
• Unknown top-level JSON fields are accepted with a warning.
• The published todo example: examples/todo.modelspec.hcl names its recordset task_summary while examples/todo.modelspec.json names it taskSummary, so export --check reports that difference for the pair as published.

Parity with the other readers

go test compares modelspec lint with committed verdicts of two other readers over the corpus in testdata/corpus (121 manifest items; each is a file, a SpecScore-layout tree or a set of standalone files, or an entry under parts/ that is one file of another item given alone and must give the verdict of its whole module; testdata/corpus/manifest.json is the expected verdict of each, under both profiles; a test fails when this number is not the manifest's):

• testdata/golden/specscore.json: specscore graph lint, compared with the default profile, through the throwaway-project wrapper that datatug/chinookdb's scripts/lint-modelspec.sh uses for single files and on whole SpecScore projects for the layout trees. The file records the specscore version (0.54.2).
• testdata/golden/directory.json: parseModelSpec of the Directory's reader, compared with the publish profile. The file records the commit on the reader's main branch.

The test asserts that modelspec lint refuses everything they refuse, except where the manifest names the difference and its reason, and that every recorded difference is real. In the other direction modelspec lint is deliberately stricter in places (for example it requires entity keys and checks types); the manifest says so item by item. To refresh the golden files (needs Node, git, and specscore on PATH or SPECSCORE=...; the Directory clone must be at a commit on its main):

node scripts/regen-golden.mjs specscore
DIRECTORY_DIR=/path/to/clone-of-openvaultdb-directory node scripts/regen-golden.mjs directory

Releases

Every release is made by a push to main, and a release runs only after the coverage gate has passed for that commit, in the same workflow run. release.yml (the only workflow that runs on a push) has two jobs: gate calls ci.yml (formatting, vet, module tidiness, the tests with the race detector, the exact coverage gate, and the GoReleaser configuration and snapshot build), and release, which has needs: gate and if: github.ref == 'refs/heads/main', calls the shared strongo/cicd release workflow. That workflow tags the commit from its conventional commits (a push of only docs:, chore: or ci: commits cuts nothing) and publishes the archives and the checksums file. ci.yml has no push trigger (it runs on pull requests and when called), so a push to main runs the gate once, and there is nothing for the shared workflow to wait for: its require_workflow_success option, which waits for a workflow's run and continues without it after 180 seconds, is not used. There is no tag trigger and no manual dispatch, so a hand-pushed tag releases nothing.

The shared workflow and every action named in ci.yml and release.yml are pinned to a full commit SHA with the version in a comment (strongo/cicd v1.21.0 is 5d96b1f3fbb3; git ls-remote https://github.com/strongo/cicd refs/tags/v1.21.0 shows it). That does not make every action that runs in a release a pinned one: the shared workflow at that commit itself runs actions/checkout@v7, actions/setup-go@v7, orhun/git-cliff-action@v4.9.1 and goreleaser/goreleaser-action@v7 by tag, in the job that holds contents: write (and actions/setup-node@v7 only when its node_version input is set, which it is not here). The pin fixes the shared workflow's own steps, not the code those four tags point to at the time of a release; the fix is upstream, in strongo/cicd.

Recovery, as read from the shared workflow's source at that commit and not exercised here:

• The gate is red (gate failed in the push's run): release does not run, so there is no tag and nothing is published. Fix on main; the next push runs the gate and, if green, releases. Re-running the failed jobs of the same run is also possible when the cause was transient.
• The release failed after the tag was pushed (the tag is pushed before the archives are built): a re-run of the run finds the tag already on the commit, computes no bump and ends green with a notice, "No release published this run", without publishing. To retry that release, delete the tag on the remote by hand and re-run the run; or leave it and let the next feat: or fix: commit cut the next version.
• A push with only docs:, chore: or ci: commits since the last tag cuts no release, by design.

Development

go test -race ./...
go test -race -covermode=atomic -coverpkg=./... -coverprofile=cover.out ./... && go run ./cmd/covergate cover.out
scripts/fuzz.sh 120    # HCL and JSON reader fuzzing, outside the default run and the gate

The coverage gate is exact: it fails unless every statement of every package, including cmd/, is covered. It has no threshold, flag or environment variable to lower it. It also lists the module's packages as go list ./... does (from the files, with no subprocess) and fails if a package that has statements is absent from the profile or contributes none to it, so a package whose tests never ran (a TestMain that exits 0) cannot vanish from the count. Everything the commands touch (filesystem, output streams, build information, the update source) is injected, so the tests run in memory with no subprocess and no network.

What stops a release without the gate

The tests in internal/covergate parse .github/workflows and fail if:

• the gate job or its two steps become conditional or able to fail silently, stop running exactly the test and gate commands one after the other, run in a container or with services, or if anything touches the cover profile between them or sets GOFLAGS, GITHUB_ENV or GITHUB_PATH;
• release.yml stops being exactly two jobs, gate (calling ./.github/workflows/ci.yml, with contents: read) and release (needs: gate, if: github.ref == 'refs/heads/main', calling the shared workflow, with contents: write, the three inputs and the five optional secrets it has now, no require_workflow_success), or its concurrency stops being cancel-in-progress: false;
• the triggers are not exactly a push to main for release.yml and pull_request plus a bare workflow_call for ci.yml (no tag, dispatch, schedule or paths filter anywhere);
• either workflow gets a workflow-level env or defaults, or other permissions than release.yml: {} and ci.yml: contents: read;
• any job of ci.yml other than the two it has gets permissions, uses, needs, a container or services, runs anything that tags, pushes, calls gh, runs goreleaser release or mentions a secret, or the packaging job's steps differ in any way from the four it has (checkout, setup-go, goreleaser check, goreleaser build --snapshot --clean);
• any action, or the shared workflow, is not a full commit SHA from the table in the test, with its version in the comment after it;
• a second workflow is named CI, or any file other than ci.yml and release.yml is under .github/workflows (a new workflow has to be added to an explicit allow-list with its reason);
• any Go file has a build constraint, so nothing can hide from the gate (a test; the gate refuses one too, below).

The gate itself also reads every .go file of every package directory of the module, whatever its name or build constraint (it does not rely on the file list of the Go tool, which has only the files this platform and these tags build), and refuses a TestMain in any of them, including one behind //go:build race or in x_linux_test.go, and any build constraint (a //go:build or // +build line, or a GOOS or GOARCH file-name suffix): a TestMain can run the tests and then exit 0, which hides a failing test while every statement still counts as covered, and a file with a constraint can be left out of a test run and so of the profile. If a package needs set-up or tear-down, do it in the tests that need it with t.Cleanup, or give the code a seam (a parameter or a variable) that a test sets.

These tests are a tripwire, not the control. They run in the pull request that changes the workflows, so a change that edits them, or a workflow they do not parse, passes its own tests. The real control is a branch rule (a repository ruleset) on main that requires a pull request and the two checks of ci.yml, Test, vet, race, exact coverage and GoReleaser check and snapshot build, up to date with main, and allows no direct push, force push or deletion, with no bypass list. It is a repository setting, not a file, and nothing in this repository can set it (nor was it set by the change that wrote this text). Until it is set, main is only as safe as the people who can push to it. (On a push to main the same checks appear as gate / … inside the release run; the rule needs the plain names, which pull requests report.)

Licence

Apache-2.0, see LICENSE. NOTICE records the code derived from the SpecScore CLI.