Render HTML

/render-html: Markdown → single-file HTML for human reading

Markdown is for writers. HTML is for readers. ARIS workflow nodes write Markdown (canonical, audit-trail-friendly, machine-parseable). /render-html turns selected artifacts into a polished single-file HTML view for the human who actually has to read them. The Markdown stays the source of truth.

When to use this skill

Use /render-html for ARIS artifacts that have a real human reader:

Do NOT use for:

• LaTeX paper output — the final reader-facing artifact is PDF, not HTML.
• SKILL.md files — those are internal LLM-facing protocol.
• .aris/traces/* review traces — forensic debug, not human display.
• Every Markdown file in your project — only artifacts that benefit from sticky TOC, callouts, math, or score progressions.

Core invariants

• MD / JSON is canonical, HTML is generated view. Edit the source, then re-render. Do not hand-edit the HTML.
• Cross-model review at the artifact boundary (ARIS invariant). Academic-template HTML — used for the artifacts humans actually read (IDEA_REPORT, AUTO_REVIEW, KILL_ARGUMENT, PAPER_PLAN) — is reviewed by a fresh cross-family Codex thread before being claimed as a finished view. Dashboard-template HTML (cockpit / debug views) skips review by default but accepts --review to force it. See § HTML Review Gate below.
• Drift detection. Every rendered HTML embeds the source path, SHA256, and generation timestamp in <meta> tags AND in the visible page header. If the HTML and source diverge, the meta tells you which version of the source produced it.
• Single-file output. No build system, no separate CSS, no node_modules. Just one .html.
• CDN-friendly default, --offline fallback. MathJax 3 and highlight.js load from cdn.jsdelivr.net by default. Pass --offline to skip both — math will appear as raw $x$, code blocks won't get syntax highlighting, but everything stays readable.
• Pure stdlib helper. render_html.py uses only re, html, hashlib, json, datetime, pathlib, argparse, sys. No pip install required.
• Defense-in-depth XSS sanitization. The helper strips <script>/<style>/<iframe>/<object>/<embed>/<form>/<input>/<button>/<link>/<meta>/<base> tags, all on* event-handler attributes (onclick, onload, …), and rewrites javascript:/vbscript:/data: href/src/action schemes to #blocked-unsafe-url:. ARIS workflow artifacts should not contain these in the first place, but the sanitizer is the safety net in case an LLM hallucinates one. Markdown text content is HTML-escaped separately and never reaches the sanitizer.

Tool Location

In this repo (ARIS-in-AI-Offer) the renderer lives at tools/render_html.py, with its templates at tools/templates/{academic,dashboard}.html. It is pure stdlib — no install step, no scripts/ subdirectory. Resolve $RENDER_HTML from the repo root:

# Run from the ARIS-in-AI-Offer repo root (where tools/ lives).
RENDER_HTML="tools/render_html.py"
[ -f "$RENDER_HTML" ] || RENDER_HTML="$(git rev-parse --show-toplevel 2>/dev/null)/tools/render_html.py"
[ -f "$RENDER_HTML" ] || {
  echo "ERROR: tools/render_html.py not found — run from the ARIS-in-AI-Offer repo root." >&2
  exit 1
}

Vendoring note. If you copy this skill into a self-contained skills/render-html/scripts/ layout (the upstream ARIS convention), move render_html.py + templates/ there and repoint $RENDER_HTML accordingly. In this repo the single source of truth is tools/ — do not introduce a second copy.

Invocation

# Default: academic template, output to <input>.html alongside source
python3 "$RENDER_HTML" idea-stage/IDEA_REPORT.md

# Dashboard template for cockpit-style views
python3 "$RENDER_HTML" research-wiki/SUMMARY.md --template dashboard

# Custom output path + title + eyebrow
python3 "$RENDER_HTML" review-stage/AUTO_REVIEW.md \
  --out review-stage/AUTO_REVIEW.html \
  --title "Auto Review — overnight run" \
  --eyebrow "Workflow 2"

# Embed sidecar state JSON (rendered as a folded <details> JSON block at end)
python3 "$RENDER_HTML" review-stage/AUTO_REVIEW.md \
  --state review-stage/REVIEW_STATE.json

# Embed sidecar JSON (e.g., for KILL_ARGUMENT.md + KILL_ARGUMENT.json)
python3 "$RENDER_HTML" paper/KILL_ARGUMENT.md \
  --json paper/KILL_ARGUMENT.json \
  --title "Kill Argument — adversarial review"

# Offline (no CDN; math + code render as plain text)
python3 "$RENDER_HTML" idea-stage/IDEA_REPORT.md --offline

# JSON-only input (wrapped in a <pre><code class="language-json"> block)
python3 "$RENDER_HTML" review-stage/REVIEW_STATE.json --template dashboard

# Language attr (default zh-CN; set for English-primary artifacts)
python3 "$RENDER_HTML" docs/SKILLS_CATALOG.md --lang en

# Skip review (academic template otherwise reviews by default)
python3 "$RENDER_HTML" idea-stage/IDEA_REPORT.md
# … then the skill skips the mcp__codex__codex review step if --no-review
# was on the command line. Pass --review to a dashboard render to force it.

The --review / --no-review flags are parsed by the SKILL orchestrator (Claude Code), not by render_html.py. The helper itself stays pure stdlib and never calls MCP. See § HTML Review Gate below for the exact resolution and prompt.

Workflow

Step 1: Identify the artifact

From $ARGUMENTS, determine what to render. Common patterns:

• "render IDEA_REPORT" → idea-stage/IDEA_REPORT.md, academic template
• "make AUTO_REVIEW readable" → review-stage/AUTO_REVIEW.md + --state review-stage/REVIEW_STATE.json
• "show the kill argument as HTML" → paper/KILL_ARGUMENT.md + --json paper/KILL_ARGUMENT.json
• "research-wiki dashboard" → look for research-wiki/SUMMARY.md or generate from raw entity counts (Phase 2 work; for Phase 1 just render SUMMARY.md if present, else fall back to listing top-level wiki structure)

Step 2: Pick template

• academic (default): linear long-form. Sticky TOC sidebar + serif body + callouts + tables + math + Q&A <details>. Use for IDEA_REPORT, AUTO_REVIEW, KILL_ARGUMENT, PAPER_PLAN.
• dashboard: grid layout, smaller font, denser metrics cards, no TOC sidebar. Use for research-wiki cockpit, RESUBMIT_REPORT, multi-project overviews.

Step 3: Run the helper

Use the resolver above to get $RENDER_HTML, then invoke. The script writes the HTML alongside the source by default (or wherever --out says) and prints a one-line confirmation including the source SHA256 prefix.

Step 4: HTML Review Gate (cross-model)

Decide whether to run review. Per ARIS invariant "executor must not judge its own output", the academic-template HTML is reviewed by a fresh cross-family Codex thread before being claimed as a delivered view. Resolution:

should_review = explicit --review present
             or (template == "academic" and --no-review NOT present)

So:

• --template academic (default) → review by default. Skip with --no-review.
• --template dashboard → no review by default. Force with --review.
• Phase 2 workflow auto-emit (planned) follows the same rule.

If should_review is true, fire a fresh mcp__codex__codex thread (NEVER codex-reply) with the prompt below. The reviewer reads the source MD + generated HTML directly; it does not see this skill's intermediate state.

Scope of review (narrow on purpose). The HTML reviewer audits render fidelity / safety / structure only — not claim truthfulness. Claim audit belongs upstream (/paper-claim-audit, /research-review, /result-to-claim). Specifically the reviewer checks:

1. Information fidelity — every section, claim, table, code block, list, math snippet, sidecar payload is present in the HTML (no silent drop)
2. Structural integrity — heading hierarchy / nested lists / table cells / code fences / <details> blocks / math delimiters survive parsing
3. Callout routing — > 🚨 … got .callout-bad, > 💡 … got .callout-info, etc.
4. Safety / escaping — no raw <script> / onclick= / javascript: / unreplaced placeholders / template-leak survives
5. Expected-difference allowance — frontmatter strip, generated header/footer/meta, TOC insert, sanitized unsafe HTML are all expected, not flagged

Codex prompt (mandatory shape). Send this as a fresh thread (mcp__codex__codex, NOT codex-reply):

You are an independent ARIS HTML render auditor. This is a fresh review thread.

Read these files directly:
- Source artifact: <ABS path to source.md or source.json>
- Generated HTML:  <ABS path to out.html>
- Optional sidecars: <state.json>, <kill_argument.json> (if any)

Task: Audit whether the generated HTML is a faithful, safe, structurally
usable view of the source artifact. Do NOT judge whether the research
claims are true. Judge only rendering fidelity.

Checks:
1. Information fidelity — sections / claims / tables / code blocks /
   lists / math / sidecars not silently dropped or materially altered.
2. Structural integrity — heading hierarchy, tables, nested lists,
   code fences, details/summary, math delimiters preserved.
3. Callout routing — warning/critical/good/info blockquotes map to
   appropriate CSS classes when present.
4. Safety/escaping — no unexpected raw script/style/iframe/form/
   event-handler/javascript/data URL survives from source.
5. Placeholder/template leakage — no unreplaced {{PLACEHOLDER}} or
   parser private placeholder character appears.
6. Expected differences — frontmatter strip, generated header/footer/
   meta, TOC insertion, sanitized unsafe HTML are EXPECTED and not a
   defect.

Return STRICT JSON first, then a short prose note:

{
  "verdict": "PASS|WARN|FAIL|ERROR",
  "checks": {
    "source_hash_match": "pass|warn|fail|unknown",
    "information_fidelity": "pass|warn|fail",
    "structure": "pass|warn|fail",
    "math_code_tables": "pass|warn|fail",
    "callouts": "pass|warn|fail|not_applicable",
    "safety_escaping": "pass|warn|fail",
    "placeholder_leak": "pass|warn|fail"
  },
  "blocking_issues": [
    {"severity": "fail",
     "source_location": "L<n>...",
     "html_location": "<selector or near-text>",
     "issue": "...",
     "suggested_fix": "..."}
  ],
  "warnings": [
    {"severity": "warn", "issue": "...", "suggested_fix": "..."}
  ],
  "summary": "one paragraph"
}

Verdict rules:
- PASS: no material fidelity/safety issue.
- WARN: readable output with minor unsupported-Markdown or cosmetic
  degradation only.
- FAIL: missing/altered meaningful content, broken tables/math/code/
  callouts that change interpretation, unsafe executable HTML, source
  hash mismatch, or placeholder leakage.
- ERROR: files could not be read or audit could not complete.

Save outputs:

1. Write the JSON verdict to <out_path>.review.json (sibling to the HTML).
2. Save the raw codex trace to .aris/traces/render-html/<YYYY-MM-DD>_run<NN>/review.{txt,json} (a per-run forensic copy of the reviewer's JSON verdict + prose).
3. Print a one-line summary to the user: verdict, N blocking, N warnings, trace: <path>.

If verdict == FAIL: the HTML is NOT a delivered review-passed view. Tell the user the blocking issues, point them at the source (fix MD or template, not the HTML), and re-render. Do not silently overwrite or mark as complete.

If verdict == WARN: deliver the HTML but surface the warning list. User decides whether to fix or accept.

If mcp__codex__codex is not available (e.g., user runs /render-html on a Codex-CLI-only setup where Codex MCP isn't wired): emit verdict: REVIEW_UNAVAILABLE to the sidecar, do not fabricate PASS, and tell the user the HTML was generated but not independently reviewed. The user can manually invoke /research-review on the source MD or re-run with Codex MCP available.

Step 5: (Optional) Verify in browser

open <out.html> on macOS, xdg-open on Linux, start on Windows. Math + code highlighting need internet (CDN); offline mode degrades gracefully to readable text.