A real run

Check TypeScript files written through a shell with an advisory Stop hook (#33)

The mechanical checks ran from a PostToolUse hook, and that hook only sees files that the Edit and Write tools wrote. A .ts file written by a shell heredoc or a script never reached the checks.

With this release, the plugin also registers a Stop hook. When a session ends, it walks the working directory for .ts and .tsx files modified since the session began and runs the same checks on them. It skips node_modules, .git, dist, build, out, coverage, .next, .turbo and .d.ts files. It is advisory: it reports through systemMessage and always exits 0, so it cannot trap a session.

New output (a .ts file written through a shell, then the session stops):

ts-patterns: mechanical violations in TypeScript files changed this session (the edit hook does not see files written through a shell):
src/list.ts:
  L1:64  no-zero-comparison  `.length` compared with zero: coerce it (`!!x.length`, `!x.length`) (see SKILL.md Core Rule 10).

Add six syntactic checks: zero-length comparison, repeated literal, colon-annotated lookup table, fail-open default, String() of an optional chain and void-returning guard (#35)

The hook now flags six more patterns from the audit of the rules. Each check is syntactic only, because the engine has no type information, so each one matches a narrow shape:

• no-zero-comparison flags .length compared with zero (> 0, === 0, >= 1, < 1 and the loose forms). It matches .length only.
• repeated-literal flags the same string or template, at least 8 characters long, written 3 or more times in one file.
• colon-annotated-lookup flags const x: Record<K, V> = { ... }. Use { ... } as const satisfies Record<K, V> so that keyof typeof x keeps the literal keys. The check ignores an empty {}.
• fail-open-default flags call() ?? "" and call() ?? 0. The check ignores a property default such as error.status ?? 0, an arithmetic operand and a template placeholder.
• string-of-optional-chain flags String(a?.b), because String(undefined) is "undefined".
• void-guard flags a function named guard or assert* that returns void. Declare asserts condition so that the narrowing reaches the caller.

Original code:

export function guard(condition: boolean, message: string): void {
  if (!condition) throw new Error(message);
}

Old output: none. The hook exited with code 0.

New output:

ts-patterns: 1 mechanical violation(s) in guard.ts:
  L1:8  void-guard  `guard` returns void: declare `asserts condition` so the narrowing reaches the caller (see SKILL.md Core Rule 1).  → SKILL.md Core Rule 1 — declare `asserts condition`

The other five checks follow the same pattern.

Affects: anyone who runs the plugin on an existing TypeScript codebase. The PostToolUse hook exits with code 2 when it finds a violation. An edit to a file that already contains one of these patterns now goes back to Claude for a fix.

To check: pipe a hook payload to the engine, as in echo '{"tool_input":{"file_path":"src/a.ts"}}' | node "${CLAUDE_PLUGIN_ROOT}/engine/dist/check.js".

Stop the magic-number check from flagging numbers inside a named constant (#35)

Previously the check flagged every number in an arithmetic expression, even when the expression is the initial value of a const that names the result. The check also flagged a number inside an as const table, although the table is the named constant.

With this release, the check skips a number in the arithmetic of a const initializer and a number inside an as const literal. This is a deliberate difference from typescript-eslint, which exempts only a bare initializer.

Original code:

export const MAX_BYTES = 16 * 1024 * 1024;

Old output:

ts-patterns: 3 mechanical violation(s) in magic.ts:
  L1:26  no-magic-numbers  No magic number: 16.  → SKILL.md Core Rule 9 — extract to a named const
  L1:31  no-magic-numbers  No magic number: 1024.  → SKILL.md Core Rule 9 — extract to a named const
  L1:38  no-magic-numbers  No magic number: 1024.  → SKILL.md Core Rule 9 — extract to a named const

New output: none. The hook exits with code 0.

Fold the audit findings into the check skill and the review agent (#32)

The check skill and the review agent gain the rules that the audit found. No mechanical check covers them. They are: hand-written library shapes, illegal states, fail-open defaults, cast laundering, repeated literals, string boundaries and small items. The review agent has new checklist items 18 to 24.

Each finding now carries the label correctness or style. A change is done when no correctness finding remains. The skill's review checklist now asks you to print each item marked checked or n/a. It also covers files written through a shell.