check-style

check-style

Role: review the shard against the codified standards.

Failure model: the code reads differently from its surroundings in a way that costs a future reader, or carries an AI tell or process scaffolding that should never reach committed prose.

The standards are cited, not duplicated:

• Prose (docs, ADRs, docstrings, comments, commit and changelog messages): skills/shared/references/prose-style.md, backed by skills/shared/references/style-foundations.md.
• C: skills/write-c/references/c-style.md.
• Zig: skills/write-zig/references/zig-style.md.
• Clojure: skills/write-clj/references/clj-style.md.
• Elixir: skills/write-elixir/references/elixir-style.md.

Those files are the checklist, but apply them for judgment, not for what a lane already gates. An em dash, a prose arrow, or a process ID is the lint spine task's job; a format nit is the formatter's; an unused binding is the language linter's. This is the nit tax the dimension rejects: flag only the style judgment no lane can make.

Look for

1. Inconsistency with the surrounding file. A function that names, comments, or structures differently from its unit without reason.
2. Pure, shell, or native leakage as a style signal. The highest-value style finding, because the three-way split is the load-bearing structure. IO, a native call, a persistence write, or state mutation inside a core unit; pure decision logic buried in the shell; domain logic inside a native wrapper. Factoring owns the structural fix; style flags it when the seam is readable as a style break.
3. Boundary contract drift. A native call whose signature disagrees with the edge contract, or that reaches into a handle's internals from the calling language.
4. Comment and docstring debt. The standard and the concrete budget live in prose-style.md and the per-language style file. Apply both directions:
   • Missing: an ownership or lifetime constraint that is true but unstated where it matters (an allocator ownership transfer, a borrow duration, a transaction invariant).
   • Stale: a comment or docstring that contradicts the code.
   • Narration: a comment that restates the next line, or a docstring that describes mechanism instead of returns and input shape.
   • Verbose: an inline comment block longer than three lines; a file whose inline comment density exceeds one line per fifty code lines; a rationale essay that belongs in an ADR; a change note ("used to", "previously") that belongs in :deprecated metadata.
   • Docstring overweight: a docstring longer than three lines on a fn whose arguments have no surface area. An options map, a spec, or a public primitive earns the length; a clever implementation does not.
   • Banner: a --- rule, ASCII art, or any decorated separator. A section marker is a single ;;;; label line, label only, no prose underneath.
5. Leaked process identifiers. A phase, task, slice, run, or plan identifier, a plan.edn or decisions.edn reference, or an internal ticket label in a comment, docstring, test description, or commit message. The rule and the allowed durable refs (ADR, the docs tree) live in write-prose. Flag the leak; suggest the code-state rephrasing.
6. Naming that misleads. A name that promises a behavior the code does not deliver, or a name whose meaning has drifted from the domain the surrounding code uses.
7. Macro and metaprogramming overreach. A macro or a compile-time construct doing work that belongs in a data function, so the result is unreachable without it.