Markdown in.
Confluence out.

The problem

Your doc lives in markdown. Confluence stores its pages as ADF — Atlassian Document Format, a JSON tree. Nothing bridges those two by hand, so "publish it to the wiki" degenerates into pasting into the editor and watching tables, links and headings come out wrong. Every republish, again.

The turn

One command converts markdown → ADF and replaces the page body. The doc stays the source of truth on disk; Confluence becomes a render target. Point it at a page URL (or a raw page ID), name any rows to strip, push.

What survives the conversion

• Headings #…######, pipe tables (first row is the header, separator discarded), links and _italic_, single-line > blockquotes, --- rules. YAML frontmatter is stripped.
• Date pills. A bare YYYY-MM-DD inside a table cell becomes a real ADF date node — a native Confluence pill. Dates in narrative paragraphs stay plain text, which is what you actually want.
• Row-dropping. --drop-row "NAME" strips field-table rows that only make sense locally — provenance, debug metadata — before anything is published.
• Confluence-native elements you added in the UI. A fresh ADF body would wipe ToC macros, info panels, Smart Links and Jira embeds — markdown can't express them, so the converter can't regenerate them. preserve-extensions.py lifts them off the current page, anchors each to its nearest preceding heading, and re-splices them into the new body.

The receipt

Fetch the current page (and keep it — the preserve step needs it), convert, splice, push:

# 1. getConfluencePage with contentFormat:"adf" → save to /tmp/current-page.json
# 2. convert + preserve
python3 <skill-dir>/md-to-adf.py <source> --drop-row "Source" > /tmp/adf-full-raw.json
python3 <skill-dir>/preserve-extensions.py /tmp/current-page.json /tmp/adf-full-raw.json > /tmp/adf-full.json

# 3. push from disk — never inline the body
python3 <skill-dir>/mcp-http-call.py updateConfluencePage /tmp/update-args.json

The push goes through a direct Atlassian MCP HTTP client that reads its bearer token from the keychain and loads args from a file. Inline tool-call args flow through the model's output tokens — fragile above a few KB (character-level corruption, escaping bugs, hard size limits). The disk path is tested at ≥600KB of args.

The limits

• Needs python3, the Atlassian MCP tools loaded, and edit access to the target page.
• No page named? It asks. It doesn't guess — and a 404 stops the run rather than picking a different page.
• It's content-agnostic on purpose. Domain conventions (a journal format with provenance rows) belong in a thin caller that supplies the right --drop-row flags and target page — not in this skill.