conorbronsdon/avoid-ai-writing v3.35.0

on GitHub

Changed

• Allow pre-commit args to override the gate defaults by moving the filename separator out of entry and into default args (#243).
• Add npm package keywords, homepage, issue tracker, and author metadata.
• Rename the user-facing "Emotional flatline" category to "Stock reaction framing" while preserving its emotional-flatline API type. Keep specific reactions, flag empty framing, and make the style finding neutral in authorship scoring until relevant positive evidence establishes a direction (#82).

Added

• Add an explicit OpenCode 1.18.30 executor for frozen rewrite-evaluation plans. It limits calls to the observed free-model allowlist, disables tools, verifies prompt and model receipts, retains failed attempts, and revalidates evidence before import. Model runs remain opt-in; benchmark judgments and release gates remain separate (#201).
• Add opt-in fp-measure.js --dump-units PATH provenance records and a fixed-detector fp-compare.js comparison of legacy and repaired preprocessing. Reports include selected and skipped units, source spans, corpus hashes, detector exclusions, and zero-observation categories (#288, #289).
• Package the deterministic detector as a composite GitHub Action and pre-commit hook, backed by a new avoid-ai-writing-gate CLI. The gate uses per-file finding counts rather than the composite score, defaults to technical + rendered-markdown, and uses a corpus-backed threshold of 6 findings per file (1.9% human-control failure rate across 376 documents, versus 31.4% at zero). Strict zero-findings policies remain available with an explicit threshold of 0. Preservation validation stays separate because it requires before/after inputs (#86).

Fixed

• Suppress eight technical-legitimate vocabulary terms (robust, comprehensive, seamless, ecosystem, leverage, facilitate, underpin, streamline) when analyzing text under --context technical mode (#237).

• Keep mid-paragraph years and other ordered markers above one in prose during false-positive measurement; expose blank-separated continuation merges and distinct measurement/preprocessor fingerprints; and pair attached headings with their unique legacy body span in comparison output without changing source spans or unit IDs (#293).

• Preserve Markdown structure and document content during corpus measurement. Separate structural cleanup from paragraph selection, retain eligible 400-word bodies after headings, and account for oversized units without silently deleting text. Keep tab-indented fences atomic, reject source hash mismatches, and construct scored rows from the verified source snapshot (#178, #179, #180, #288, #289).

• Report top detection categories for documents the self-scan scores in chunks. scoreLongText() now counts issue types across every accepted chunk, so scanFile() returns a populated topTypes for a chunked file and the --check over-budget diagnostic names categories instead of printing none (#264).

• Validate CLI --unit argument in scripts/fp-measure.js before starting measurement, exiting with code 2 on missing, unrecognized, or repeated values, and on --unit=VALUE syntax (paragraph and document accepted).

• Accept acronyms (AI, API, CLI) and the capitalised single-letter function word A as interior tokens in the Title Case header rule, so headings like ## The Future Of AI In Production and ## Why Your Team Needs A Better Testing Strategy are flagged like the original tell. First and last tokens still require an ordinary Title Case word, so all-caps banner lines such as ## HTTP API REFERENCE stay clean (#240).

• Consume bodyless punctuation runs once when splitting sentence highlights, avoiding the quadratic punctuation-prefix regression introduced in #260 while preserving trailing-fragment boundaries.

• Remove four quadratic scans from analyzeText(): the sentence splitter behind highlight regions, its boundary-whitespace trim in rendered-Markdown mode, the Markdown table delimiter test, and the line-anchored Interesting part: opener all rescanned a long whitespace or blank-line run from every position, so a document that ended in blank lines or carried a large masked comment block took seconds instead of milliseconds. Sentence boundaries are unchanged; the regression test compares them against the former regex on every boundary shape and asserts linear growth by ratio rather than by a wall-clock budget (#235).

• Preserve detector issue indexes and sentence-highlight ranges against the original source after blockquote and normalization preprocessing (#189).

• Include every observed corpus register in corpus.js list; preserve the preferred accepted-register order and sort additional registers deterministically.

• Make rendered-Markdown HTML comment masking linear with a source-order scanner that preserves fenced, inline, and indented-code precedence without rescanning the document per comment (#190).

• Fix three README link targets: the dead Cowork URL, the pattern-catalog pointer, and the
  voice-profile link that led to the triggering section.

• Correct the analyzeText() result table in detector/README.md: the six score labels the
  engine returns, the UNSCORED classification on early-exit paths, and all four accepted contextMode values.

• Align contextMode comments in detector/patterns.js and mode list in
  detector/CATEGORIES.md with runtime behavior: four accepted modes, only
  technical changes flagging (#173).