Lint

Lint decides everything a rule can decide. A deck is not done while dekc lint --visual fails; passing it means nothing measurable is wrong, not that the deck is good. The balance of the slides, the argument, and the timing are judged by reading the sheet and the script, and finally by the author. The dev server lints on every save, so a deck you are working on is always passing or telling you why not. You rarely run dekc lint by hand; it exists for CI and for checking one slide.

dekc lint
dekc lint --fix
dekc lint --visual
dekc lint --format sarif
dekc check architecture --shot

--fix syncs first: it creates skeleton HTML for missing slides (DEK001). It never touches a slide you have edited; only an untouched skeleton is rewritten or removed.

Layers

dek delegates general Markdown hygiene to rumdl and writes only the rules rumdl cannot know about. rumdl is optional: dek uses DEK_RUMDL when it is set, else looks for it on PATH and then in the node_modules/.bin dek itself is installed in (never the current directory's, which a cloned repository could fill), and skips it when absent: the text output says rumdl: skipped (rumdl is not installed) with a help: line on how to install it, --json lists it in "skipped" with the same reason and hint, and SARIF reports it as a tool execution notification. dekc init writes a .rumdl.toml that disables the first-line-heading rule, since a script starts with frontmatter and ##.

Layer	Covers
Markdown style	rumdl
Schema	Frontmatter and config, validated with Zod
Consistency	script.md ↔ slides/*.html, beats ↔ data-step
Theme contract	Unknown classes, inline styles, tokens, scoping, slide stylesheets
Slide scripts	slides/<id>.ts evaluated apart from Node and Bun
Self-containment	Remote URLs, missing files, paths outside the deck
Accessibility	Pictures with nothing for a screen reader to say; contrast is under Rendering
Rendering	Overflow and contrast, measured in a browser
Narration	Only for decks with voice/
Length	Only for decks with a duration

--visual

Overflow (DEK030) and contrast (DEK031) are measured in a real browser through Playwright. Both are questions of geometry with definite answers. dek does not show a screenshot and ask whether the text fits; it measures. Every beat of every slide is rendered in one browser session, and what each page measured is kept in the deck's .cache/visual, named by a hash of all that decides it: the page as rendered, with its theme, CSS, script, and assets, and the code that measures. A page measured before is not measured again, so after the first run --visual costs only the slides you changed. Contrast is read from the pixels as drawn, so text over a gradient, an image, or a glow is judged by what the audience sees; see DEK031.

A finding names what to fix: the element, the start of its text, and the amount.

slides/objection.html: DEK030 li "https://example.com/very…" overflows the right edge by 102px at steps slow, vague
  help: shorten it, or let it wrap with overflow-wrap: anywhere in slides/objection.css
slides/objection.html: DEK031 p.note "補足" has contrast 1.5 (#333333 on #111111), below 4.5:1 at steps slow, vague
  help: theme.css alone draws it below 4.5:1: fix the pair in theme.css, where one change reaches every slide that uses it

A list that runs off the bottom is one finding for the list, not one per item: a child is reported only for an edge its parent stays inside. The same finding on several beats is reported once, with every beat named. A string that cannot wrap, such as a URL, counts even when its box fits.

Contrast thresholds follow WCAG AA. Body text needs 4.5:1. Large text, meaning 24px or larger, or 18.66px and bold, needs 3:1. That is what lets a big number in a soft color pass while the same color on body text fails.

Without --visual, dekc lint says so: visual is in "skipped", and the text output ends with visual: skipped, with the command that measures. A clean lint that did not measure is never mistaken for one that did. Without Playwright, only the commands that need it fail, each with the install command in its hint. The rest of the CLI runs.

• dekc lint --visual is a verdict. It returns diagnostics (SARIF with --format sarif), and it is the primary way an agent checks its own output.
• dekc shot is an observation, not a verdict. Whether a slide looks good stays a human call.

Voice does not change what failing means

DEK040 (an English word missing from the pronunciation dictionary) and DEK042 (a beat that shows something but says nothing) apply only to decks with voice/. DEK041 compares the talk's length with the duration budget: the narrated length when a Timeline exists, otherwise the reading-time estimate dekc ls shows, which gets a wider margin (35% instead of 20%) because it leaves out pauses and demos. DEK043 (a [beats] key in voice.toml that matches nothing) applies only to decks with voice/. All four are warnings: they are reported, and they do not fail lint. Adding voice never turns a passing live-only deck into a failing one.

Output

Diagnostics are SARIF 2.1.0 so that VS Code, CI, and agents all read the same format. rumdl's results are merged in as a second run. Humans get ESLint-style text by default.

Wherever the fix is known, a diagnostic carries a hint: the beat ids a data-step may use, the classes a slide may use, the assets/ path for a remote image. Text output prints it as a help: line, --json as a hint field, and SARIF as properties.hint, next to slug and data.

Every diagnostic has a severity, error or warning. Only errors fail: dekc lint and dekc check exit 1 and return "ok": false when at least one error remains, and warnings alone exit 0 with "ok": true. Text output labels a warning after its rule id, and SARIF sets level. Diagnostics from rumdl count as errors.

A diagnostic points at the file to change: DEK001 at the section heading in script.md, rules about slide HTML at the line and column of the offending element, attribute, or URL. Lint reports every occurrence, not the first of each kind, so one run lists everything to fix. The values the message names are also in a data field, so an agent reads { "class": "headline" } or { "edges": { "bottom": 591 }, "steps": ["1"] } instead of parsing the message.

{
  "id": "DEK031",
  "severity": "error",
  "message": "p.note \"補足\" has contrast 1.5 (#333333 on #111111), below 4.5:1 at step 1",
  "path": "slides/objection.html",
  "slug": "objection",
  "hint": "theme.css alone draws it below 4.5:1: fix the pair in theme.css, where one change reaches every slide that uses it",
  "data": { "box": "p.note", "text": "補足", "ratio": 1.5, "threshold": 4.5, "fg": "#333333", "bg": "#111111", "origin": "theme", "steps": ["1"] }
}

dekc lint --format sarif > results.sarif

Every rule is listed in Lint Rules.

Next

Step through the deck and export it: Presenting.