xmutfyh/dsh-plugin-writing-guard

A deterministic academic-writing linter for DeepSeek Harness (DSH).

Keeps your academic writing clean: it scans manuscripts for revision-process residue, claim-calibration issues, rhetorical patterns, LLM-associated vocabulary, and formatting tells — with document-type awareness and density-based thresholds. Auto-audits every paper file you write. Zero network, zero LLM cost (pure local regex/statistics).

Not an "AI detector" — a linter that knows what document it is checking and can explain why it flags something.

Install

# from GitHub (lib/ is committed — no build step needed)
dsh plugin --profile web add github:xmutfyh/dsh-plugin-writing-guard

# or from the GitHub tarball directly
dsh plugin --profile web add https://github.com/xmutfyh/dsh-plugin-writing-guard/archive/refs/heads/master.tar.gz

# restart to load
dsh web

Repo: https://github.com/xmutfyh/dsh-plugin-writing-guard

Document profiles (v0.3)

The same phrase means different things in different documents. Rules are scoped by document type:

profile	meaning	e.g. "as requested by the reviewer"
manuscript	paper body	🔴 revision residue — flagged
rebuttal	response letter	✅ normal — not flagged
cover_letter	submission letter	🔴 residue — flagged
review / notes / unknown	other	conservative

Pass profile to writing_audit, or it is auto-detected from the file path.

What it catches

Category	Typical issues
Process residue	"revised model", "as requested", "we have updated", CN "本轮/投稿前"
Claim calibration	"we do not claim", CN "本文并非要证明", self-deprecation; legitimate limitations statements are NOT flagged (ICMJE requires them)
Rhetorical patterns	"not X but Y", "rather than" abuse, absolutist definitions, rule of three
LLM-associated words	delve / tapestry / testament / leverage / harness… (density rule — a single occurrence is fine)
Academic style	"we believe/think", hedges, abstract adverbs; "significantly" only prompts review of non-statistical uses
Formatting	em-dash density (range en-dashes excluded), colon-title abuse

Density thresholds (v0.3.3)

Frequency rules use per-1,000 language units: English rules use the English word count, Chinese rules use CJK char count (bilingual files do not dilute each other), with a double gate: count >= minCount AND count/denominator*1000 >= perK. E.g. rather than: ≥4 and ≥1.0/1k; em-dash: ≥5 and ≥0.5/1k; LLM words: ≥2 and ≥0.4/1k; Chinese connectives: ≥8 and ≥2.0/1k chars. A 500-word abstract and a 12,000-word full paper no longer share one absolute threshold.

Preprocessing (v0.3.3, on by default)

Before auditing, non-prose content is stripped: YAML frontmatter, code fences, inline code, LaTeX inline/block math, Markdown links (anchor text kept), bare URLs, LaTeX commands; the References/Bibliography section is cut (# References, References:, \section{References}, \begin{thebibliography} all supported). Rules scan prose only — references/code/math/URLs never enter hits or density denominators.

Confidence & evidence (v0.3)

Every rule carries confidence (high/medium/low) and evidence (literature/style-guide/heuristic/project-specific). Reports show 🔴 HIGH · conf high, so you know which flags are deterministic rules (e.g. revision residue) and which are probabilistic signals (e.g. LLM-word density).

Tools

Tool	Purpose
writing_audit	Scan text or file; args: text/filePath, profile, verbose; returns severity+confidence sorted issues and full-text statistics
writing_rules	Return the discipline quick-reference (profiles + density) before drafting

Auto audit (on by default)

Listens on tools/post-execute: after write/edit touches a paper-like file (.md/.tex/.txt under manuscript/paper/revision/response/论文/修订/返修… or KB dirs), it runs the audit automatically (profile auto-detected) and injects high-severity findings as additionalContexts for the model's next request.

Configure in cordis.patch.yml:

- id: dsh-plugin-writing-guard
  config:
    autoAuditOnWrite: true
    autoAuditMinSeverity: high
    maxAutoInjectPerTurn: 2
    verboseByDefault: false
    autoBrief: false
    projectResidueTerms: []   # project-internal terms appended to defaults

Tests

npm test   # builds first, then 50 TP/TN/edge cases, no framework needed

Development

pnpm install && pnpm build   # TypeScript -> lib/
# rule engine: src/rules.ts (zero dependencies, regex + statistics)

License

MIT