elumenotion/guideants/transcript-to-pdf

transcript-to-pdf

Paths — fixed layout, do not probe or re-derive. The sandbox CWD is the notebook's output directory. This skill's scripts live under Skills/transcript-to-pdf/scripts/ relative to it. Write PDF output with a bare filename (e.g. -o report.pdf); never prefix with Output/.

Turns a markdown conversation transcript (the kind produced by saving a GuideAnts / Code-Executor session) into a well-formatted PDF.

The renderer understands the transcript system, not generic markdown: it recognises **User** / **Assistant** / **Tool** turns, **Tool Call:** input blocks, and tool OUTPUT JSON, and lays each out as a distinct card. Heredoc script inputs (e.g. cat > file <<EOF) are split so the bash header and the written-file body are highlighted with the correct language lexer.

Stdlib + weasyprint only. No network. No model download. Runs in seconds.

Dependencies

Required (already present in the sandbox image):

weasyprint   # HTML -> PDF
pygments     # code highlighting
pypdf        # (optional) page inspection

Check with python3 -c "import weasyprint, pygments". If missing: pip install weasyprint pygments.

What to run

Single command, one input, one output:

python3 Skills/transcript-to-pdf/scripts/md_transcript_to_pdf.py \
  <transcript>.md -o <out>.pdf

• Output defaults to <input>.pdf when -o is omitted.
• The script prints a short report (title, turn counts, tool-call counts,

fence fixes, output path) and exits 0 on success.

Useful flags

python3 .../md_transcript_to_pdf.py in.md --no-cover   # drop the cover page
python3 .../md_transcript_to_pdf.py in.md --html out.html   # also dump the intermediate HTML (debug)
python3 .../md_transcript_to_pdf.py in.md --quiet       # suppress the report

Layout guarantees

• Cover page — title, subtitle, stat grid (messages / users / assistants /

tool calls / heredocs / source lines), metadata table, and a list of the session's user requests. The list region is a bounded flex box (flex:1 1 0; min-height:0; overflow:hidden) so it cannot collide with the footer regardless of how long the session is. Snippets are capped (4 shown, "+ N more" otherwise) and shortened to ~140 chars to keep the common case clean.

• Turns — every message is a role-badged card (User / Assistant / Tool).
• Tool-call inputs — **Tool Call:** blocks are anchored to the following

JSON fence; the script field is unescaped and, if it is a heredoc, rendered as (bash header) + (file body in the right language) + (bash footer).

• Tool outputs — the **Tool** result JSON renders in an amber panel.
• Markdown — headings, bold/italic, inline code, lists, and pipe tables.
• Truncated-fence repair — turns that end mid-code-fence get a closing

fence inserted so the document balances.

Inputs it expects

A .md file in the saved-conversation shape:

# Title
**Created:** ...            <- optional "Key: value" metadata
**Last Activity:** ...
**Assistant:** ...
---
**User**
<message>
---
**Assistant** (Code Executor)
<message / reasoning>
**Tool Call:** `run_bash`

{ "script": "cat > file <<'EOF' ... EOF" }

---
**Tool** (run_bash)
{"StandardOutput":"...","ExitCode":0,...}
---

See references/transcript-format.md for the exact structural contract and the quirks the parser auto-repairs.

Failure modes

• No `# Title` — the base filename is used as the title. Safe.
• Unbalanced code fences (log cut mid-block) — auto-closed; the report

shows Fences fixed: N.

• Malformed tool-call JSON — falls back to rendering the raw fence body as

text; the rest of the document still renders.

• weasyprint missing — clear ModuleNotFoundError; install it and re-run.

Reporting

End by stating the output path, the turn/tool-call counts from the report, and how many truncated fences were auto-repaired.