| name | new-mkit-spec |
| description | Scaffold a new mkit SPEC document with the repo's real structure, then register it in the spec index and wire any parity follow-up. Use when adding a docs/specs/SPEC-*.md for a new wire format, on-disk format, or subsystem; when an in-scope command or behavior needs a normative spec; or when a format change requires a spec per the CONTRIBUTING review bar. |
Author a new mkit SPEC
mkit specs are the authoritative source for on-disk formats, wire formats, and
subsystem behavior. Anything that mutates an on-disk or wire format requires a
matching docs/specs/SPEC-*.md change (CONTRIBUTING review bar item 5). Work in
the checkout, not from memory โ the specs evolve, so confirm the current
conventions before you write.
Ground truth to read first (they win over anything here if they have changed):
docs/specs/README.md โ the spec index you will register into.docs/STYLE-GUIDE.md โ the writing conventions specs follow.- Two model specs of the shape you are writing:
docs/specs/SPEC-OBJECTS.md
(a byte-layout format spec) and docs/specs/SPEC-WORKTREE.md (a subsystem
spec). Skim one more neighbor if your topic differs.
docs/PARITY.md โ only if your spec introduces or changes an in-scope
command or behavior (see step 3).
Step 1 โ Scaffold the spec file
Create docs/specs/SPEC-<NAME>.md. <NAME> is uppercase, hyphen-separated,
and names the format or subsystem (OBJECTS, PACK-SHARDS, WORKTREE). Copy
the shape from SPEC-TEMPLATE.md in this skill directory, which encodes the
structure below.
Front matter โ the four keys every spec carries, in this order:
---
spec: SPEC-<NAME>
version: 1
status: draft
audience: <who must implement or review this โ name the crate, CLI surface, or role>
---
status is one of stable, normative, or draft (README front-matter
vocabulary). A new spec normally starts draft.
Opening block, immediately after the # SPEC-<NAME> โ <short title> H1:
Status: line โ a bold state word plus a sentence (Status: **Normative** for mkit v1.). This restates the front-matter status in prose and MAY say
more than the one-word key.Scope: line โ one to three sentences fixing exactly what the document
governs, and what it explicitly does not.- Optional openers used by real specs, add the ones that apply:
Endianness:
(byte-layout specs), Reference implementation: (the authoritative module),
an Authority: sentence, and the driving issue number.
- For a format meant for external reimplementation, add the conformance
sentence the format specs use: external tools MUST be able to produce and
consume these bytes from this document alone.
Body โ numbered H2 sections (## 1. Title, ### 1.1 Subtitle), sentence-case
headings per the style guide. Numbered sections are the dominant convention;
short specs MAY use unnumbered H2s. Order sections from the model most like your
topic. The recurring ones:
- Purpose / design constraints โ why the format or subsystem exists.
- The normative core โ the byte layout (a fenced
offset size field
block), the state model, or the protocol. This is the spec's reason to exist.
- Semantics / discovery / lifecycle โ how the format is read, written, and
behaves; failure modes as typed errors.
- Out of scope / non-goals โ bound the spec so scope cannot creep.
- Version history โ a table, for on-disk and wire formats that must
migrate.
- Test vectors (formats: the inputs/outputs an implementer MUST produce,
pointing at
rust/tests/golden/) or Test anchors (subsystems: the unit
and CLI tests that pin the behavior).
- Invariants โ a two-column table mapping each invariant to the mechanism
or error that enforces it. Nearly every spec closes with this; include it.
Write to the style guide: normative RFC 2119 keywords in caps (MUST, MUST NOT,
SHOULD, MAY); present tense, second or third person; realistic 64-hex BLAKE3
digests in examples, never xxxx. Link sibling specs with relative links
([SPEC-REFS](SPEC-REFS.md)) on first reference to a term they own.
Done when: docs/specs/SPEC-<NAME>.md exists; front matter has all four keys;
the opening carries a Status: and Scope: line; the normative core is
specified precisely enough to implement from; an Invariants section closes it;
and its headings and cross-links match the two model specs (the repo has no
markdown linter, so this is an eyeball check against SPEC-OBJECTS.md).
Step 2 โ Register in the spec index
Add one bullet to the list in docs/specs/README.md, keeping the list in
alphabetical order by spec name:
- [SPEC-<NAME>](SPEC-<NAME>.md) โ <one lowercase clause naming what it specifies>.
Match the file's existing bullets exactly: linked spec name, a literal em dash
separator, a single lowercase descriptive clause, and a trailing period.
Done when: the new bullet sits in alphabetical position; every SPEC-*.md in
docs/specs/ has exactly one matching bullet (compare ls docs/specs/SPEC-*.md
against the list); and the link resolves.
Step 3 โ Wire the parity follow-up (only if in scope)
Skip this step for a pure format or internal-subsystem spec that adds no
user-visible command or flag.
Do it when the spec introduces or changes an in-scope command or behavior
(a new porcelain/plumbing command, a changed flag, a new safety guard, a
machine-output shape). Read docs/PARITY.md first โ it is the authoritative
scope gate, and the live per-command matrix renders from
apps/web/src/lib/parity-data.ts onto the web /parity page.
- If the command was previously a non-goal or unlisted, the scope changed:
add a scope-amendment paragraph to
docs/PARITY.md (follow the worktree and
git-bridge amendments as models โ state what left the non-goals list, the
semantics adopted, and any deliberate divergences), and cross-link your new
spec from it.
- Update the live matrix in
apps/web/src/lib/parity-data.ts: add or edit
the command's ParityItem (cmd, status of parity/divergent/non-goal,
and a one-line user-facing note). This file, not docs/PARITY.md, drives
the rendered table.
- Record any divergence in the matching
docs/PARITY.md section: a safety
guard under "Safety divergences", a parsed-output shape under the
"Machine-output contract", a human-output difference under "Human-facing
output parity", or a deferred flag under "Deferred flags".
Done when: the /parity command's status and note reflect the spec; any scope
move has an amendment paragraph in docs/PARITY.md linking the spec; and
divergences are logged in the right docs/PARITY.md section. Editing
apps/web/** means the path-filtered CI: Web workflow now applies to your PR
(see step 4).
Step 4 โ Validate before opening the PR
- Grep your prose for banned copy the style guide flags:
please, sorry,
unfortunately, simply, easy, and the literal โ character (the guide
wants — in prose; the index bullets in step 2 are the exception that
keep the literal em dash to match the file).
- A spec-only change is path-filtered in CI: the full Rust matrix and coverage
do not run on a docs-only push. Enumerate the checks you expect by name and
confirm each ran โ a green rollup does not prove a workflow executed. If you
touched
apps/web/** in step 3, confirm CI: Web ran; and check the
mkit-*-pr Google Cloud Build checks by name, since the docs-lint gate lives
there rather than on the GitHub gate.
- If the spec accompanies a format change, land the spec, the golden vector
under
rust/tests/golden/, and the code in the same change (CONTRIBUTING
review bar item 5). Crypto or key-handling specs need a second reviewer and a
threat-model note in the PR body (item 6).
- Commit with a
docs (or docs,specs) Conventional Commit scope, matching the
spec commits in git log.
Completion checklist
