Markdown Processor (marcus)

A zero-dependency Markdown → HTML CLI and library for Node.js — the converter that verifies its own output.

Pandoc tells you nothing. Marked trusts everything. marcus checks the input, escapes untrusted HTML by default, and reports everything it changes.

Install

npm install -g @ashusevim/markdown-processor   # exposes `marcus`

Requires Node.js >= 20. No runtime dependencies.

Usage

marcus file.md                # HTML on stdout
cat notes.md | marcus         # stdin (implicit with no file argument)
marcus a.md b.md              # multiple files, in order
marcus --strict docs/*.md     # CI gate: exit 1 if warnings are found
marcus --report json doc.md   # machine-readable report instead of HTML
llm | marcus --fix-llm        # repair known LLM artifacts, logging each one

Why marcus

• Safe by default. Raw HTML from users, scrapers or models is escaped to visible text, so a <script> can't execute. --unsafe passes it through for trusted documents. Most converters do the opposite and leave you to "sanitize it yourself".

• It tells you the truth. Render-affecting problems come back as line-numbered diagnostics instead of silently wrong HTML:

  marcus: draft.md:4: warning [unclosed-bold]: odd number of '**' — bold segment may be unclosed
  

• Repairs LLM slop on the record. --fix-llm fixes the well-documented artifacts and logs every repair as an info issue, so nothing is rewritten behind your back.

• Tiny and auditable. A few hundred dependency-free lines; the CLI is a thin wrapper around a pure parser you can import.

Options

Option	Description
-h, --help · -V, --version	help and version
- · --	read stdin explicitly · treat the rest as file names
--strict	exit 1 on warning-level issues (CI gate)
--report json	JSON report (html, issues, stats) instead of HTML
-q, --quiet	don't mirror issues to stderr
--unsafe	pass raw HTML through unescaped (trusted input)
--fix-llm	repair LLM artifacts before parsing, logging each repair

Exit codes: 0 success · 1 unreadable file or --strict warnings · 2 usage error.

Diagnostics

Two severities: warning means the output probably doesn't match your intent (and --strict fails on it); info is a repair or unsupported-syntax notice, which never fails a build.

Code	Problem
unclosed-fence · unclosed-inline-code · unclosed-bold · unclosed-strikethrough	a delimiter is never closed, so everything after it renders wrong
heading-skip	heading level jumps (e.g. h1 → h4), breaking the outline
list-interrupted · mixed-list-markers	implicit list boundary, or marker style changed mid-list
table-ragged	a table row has more cells than the header — the extras are dropped
frontmatter-unsupported · html-escaped	--- frontmatter (renders as <hr>) · raw HTML neutralized by safe mode
repaired-*	what --fix-llm changed (heading space, frontmatter fence, proved, table delimiter)

Reports are deterministic — identical input produces byte-identical output (no timestamps), so they can be cached and diffed:

{
  "tool": "marcus", "version": "2.3.0",
  "files": [{ "file": "doc.md", "html": "…", "words": 5, "headings": [] }],
  "issues": [{ "file": "doc.md", "line": 5, "code": "unclosed-fence", "severity": "warning", "message": "…" }],
  "stats": { "files": 1, "words": 5, "readingTimeMinutes": 1, "warnings": 1, "infos": 0, "repairs": 0 }
}

Library

import { parseMarkdown, parseMarkdownDetail } from "@ashusevim/markdown-processor";

parseMarkdown("# hi");                     // '<h1>hi</h1>'
const { html, issues, headings, words } = parseMarkdownDetail(source, { fixLlm: true });

Syntax

CommonMark-style blocks and GFM extensions: ATX headings, thematic breaks, blockquotes (parsed recursively), ordered/unordered lists nested to any depth, task lists (- [x]), GFM pipe tables with :---/:-:/---: alignment, fenced code with language classes, and inline code (including multi-backtick spans), bold/italic/strikethrough, links, angle autolinks and images. A leading BOM and CRLF/CR line endings are normalized.

Limitations

• No footnotes or setext headings yet.
• Frontmatter is not parsed; it renders as a thematic break and is reported as info.
• Safe mode is escape-all, not an allowlist sanitizer. Pipe --unsafe output through a dedicated sanitizer if you need fine-grained control.
• Task-list checkboxes are inert (disabled) by design.
• Lists render tight: a blank line between items doesn't force <p> wrappers, and a blank line inside an item only wraps the paragraphs that follow it.

Development

npm test        # builds, then runs the node:test suite (165 tests)
npm run build   # tsc → dist/

src/parse.ts is the whole parser and diagnostics engine (no I/O, no deps) — that's the library. src/cli.ts is the argument/stdin/exit-code wrapper. docs/demo.sh regenerates the demo GIF:

asciinema rec -q --overwrite --cols 124 --rows 32 -c "bash docs/demo.sh" docs/demo.cast
.tools/agg --font-size 14 --theme monokai docs/demo.cast docs/demo.gif

License

ISC