Added
- A "Pre-publish pipeline" README section and a
fact-checkmarketplace entry, so the companion skill installs from theconorbronsdon-skillsmarketplace alongside this one. - Antigravity plugin manifest at
plugins/avoid-ai-writing/.antigravity-plugin/plugin.json, enabling native discovery and installation in Google Antigravity alongside the Claude plugin manifest. Updated release checks and CI validation to maintain manifest parity. - A Vale style,
vale/AvoidAIWriting/, generated from the detector's pattern tables byscripts/build-vale.js.npm testfails when the committed style is out of date.vale/AvoidAIWriting/manifest.jsonlabels every detector typeverified,approximation,unsupported, orunassessed; the style README's counts come from it.scripts/vale-cases.test.jsruns a case that fires and a case that must not for every rule through Vale 3.26.0 and the detector, and records where they differ; CI installs that Vale version.scripts/package-vale.pybuilds a deterministicAvoidAIWriting.zipfor a release (#94). - The detector exports
PATTERN_TABLES: the pattern tables as recursively frozen plain data keyed by issue type, with regexes as{ source, flags }strings and aschemaVersion. Gated types carry agatenote.analyzeText()and the existing exports are unchanged. It replaces the draft sample export inscripts/vale-export-sample.js, which is removed (#94). --styleconfigs acceptcolons: "never",semicolons: "never", andemDash: "never".scripts/check-style.jsreports each as a hard violation in prose, so a house style that bans these marks gates the exit code. Clock times, ratios, URL schemes, table delimiter rows (including blockquotes), HTML entity syntax, and code are not flagged. Character references that render as banned punctuation still fail.
Changed
- Add editorial decision cards to selected existing rules, with stable IDs, partial detector mappings, applicability conditions, and edit/keep/gap examples. Preserve subjective judgments and report unresolved claim scope. Keep the full catalog mandatory in directory and portable installations; no detector API, loading default, or category count changes (#206).
- Reworded the large-file scope sentence in the editing contract so it no longer reads as "without asking" to static skill scanners; the meaning is unchanged. Documented known scanner false positives in
SUPPORT.md. - The detector now exports
CONTEXT_MODESandSOURCE_MODES, the frozen lists of modesanalyzeText()accepts. The CLIs read them instead of keeping their own copies (#413, #414). - Cite or qualify frequency claims in the pattern catalog (
references/patterns.md). Claims about how often models use a phrase now cite pinned HAP-E measurements or are qualified, and several entries no longer attribute a pattern to AI by default. Rule IDs, detection behavior, and the category count are unchanged (#249).
Fixed
- The preservation validator's
stats.indentedBlocksno longer counts indented lines inside fenced code or four-space paragraph continuations. Only indented blocks that follow a blank line outside a list are counted, as documented. Top-level fences end preceding list context; fences nested in a list preserve it. - Remove polynomial backtracking from the sentence-unit regex in
normalizeText(). A long tab or space run after a Markdown table delimiter (-|-) or thematic break (***,___) madeanalyzeText()take seconds on a 64 KB input, before the word limit applied. Matching is unchanged; both shapes are now in the bounded-time test (CodeQLjs/polynomial-redos). - Treat the smart-punctuation co-occurrence finding as a low-priority style check in plain mode and omit it in rendered-Markdown mode. House-style typography no longer changes authorship scores, probabilities, confidence, or classifications, including the separate classifier corroborator. Keep the public type and other pattern checks (#351).