cfxmark

Bidirectional Markdown ↔ Confluence Storage XHTML converter — with lossless opaque preservation for everything cfxmark doesn't explicitly know how to convert.

import cfxmark

# Markdown → Confluence storage XHTML
result = cfxmark.to_cfx(markdown_text)
result.xhtml          # str    — ready for Confluence REST PUT
result.attachments    # tuple  — local file refs the caller should upload
result.warnings       # tuple  — human-readable conversion warnings

# Confluence storage XHTML → Markdown
result = cfxmark.to_md(xhtml_text)
result.markdown       # str    — canonical markdown
result.warnings       # tuple

# Markdown or Confluence XHTML → Jira wiki markup
result = cfxmark.to_jira_wiki(markdown_text)
result.jira_wiki      # str | None — Jira wiki markup

# Jira wiki markup → Markdown  (EXPERIMENTAL — v0.3+, enhanced v0.4)
from cfxmark.jira import from_jira_wiki
result = from_jira_wiki(jira_issue_description)
result.markdown       # str
result.attachments    # tuple  — filenames referenced via [^file] / !file!
result.warnings       # tuple  — unsupported user mentions, …

ConversionResult is the same dataclass for all directions — xhtml is populated for to_cfx, markdown for to_md / from_jira_wiki, and jira_wiki for to_jira_wiki.

Why another converter?

Two existing projects inspired this one — md2cf and md2conf — but both are one-directional (md → cf) and neither preserves unknown macros across a round trip. cfxmark fills both gaps:

1. Bidirectional. to_md(to_cfx(m)) is byte-identical to canonicalize(m) for every construct in the supported subset.
2. Opaque preservation. Confluence content cfxmark doesn't understand (custom plugins, drawio diagrams, exotic table cells) round-trips byte-for-byte, including the ac:macro-id UUID. Confluence treats the round-tripped macro as the same instance, so comments, attachments, and permissions stay attached.
3. Pure text-in / text-out. No Confluence API, no network, no attachment upload. The caller owns REST I/O. (See "Image assets" below for the helper function that lets the caller plug in network-bound logic without bloating cfxmark.)

Install

cfxmark ships in two modes:

# Core: Markdown ↔ Confluence XHTML converter + Jira wiki renderer
pip install cfxmark

# With uv (recommended):
uv add cfxmark

# With the optional Confluence REST client (zero additional deps —
# the extra is namespace-only and reserves a stable upgrade slot):
pip install 'cfxmark[confluence]'

The confluence extra declares zero third-party runtime dependencies — from cfxmark.confluence import ConfluenceClient works even without it. The extra exists to signal intent in requirements files and to reserve a stable upgrade slot for future convenience helpers.

cfxmark depends on lxml and mistletoe. Python 3.10+.

The contract

cfxmark grades every Confluence construct into one of three buckets:

Grade	Description	Behaviour
I — Native	Standard CommonMark / GFM (headings, lists, tables, code fences, links, images, blockquote, hr, inline emphasis)	Lossless round-trip after canonicalization.
II — Directive	Confluence macros with a known Markdown directive mapping (info, note, warning, tip, jira, expand, toc)	Lossless after canonicalization. Pluggable via MacroRegistry.
III — Opaque	Everything else	Captured byte-for-byte through cfxmark's opaque-block / inline-opaque mechanism. Never dropped, never rewritten.

See docs/SPEC.md for the full mapping table and docs/OPAQUE.md for the opaque-block format.

Usage

Round-trip a Confluence page through Markdown

import cfxmark

# Whatever fetched the page (REST API call, exported XML file, …)
xhtml = my_confluence_client.get_storage_format(page_id)

# Convert to Markdown
md_result = cfxmark.to_md(xhtml)
markdown = md_result.markdown

# … user edits the Markdown …

# Convert back to Confluence storage XHTML
cfx_result = cfxmark.to_cfx(markdown)
my_confluence_client.update_page(page_id, cfx_result.xhtml)

# Optionally upload any newly referenced local images
for filename in cfx_result.attachments:
    my_confluence_client.upload_attachment(page_id, filename)

Image assets

When you convert a Confluence page that references uploaded attachments, the resulting Markdown looks like this:

![](image-3.png#cfxmark:w=700)<!-- cfxmark:asset src="image-3.png" -->

The image link still points at the original Confluence filename (broken in any local Markdown viewer until you fetch the bytes), and the <!-- cfxmark:asset --> HTML comment carries enough metadata for a follow-up step to fetch and embed.

cfxmark.resolve_assets is that follow-up step. You provide a fetcher callback that returns bytes for one filename at a time, and choose between two output strategies:

import cfxmark
from pathlib import Path

def fetcher(filename: str) -> bytes:
    # Whatever you use to download from Confluence:
    return my_confluence_client.download_attachment(page_id, filename)

# Strategy A — sidecar directory (recommended for git-tracked docs).
# Saves bytes to ./assets/ and rewrites links to relative paths.
md = cfxmark.resolve_assets(
    md_result.markdown,
    fetcher,
    mode="sidecar",
    asset_dir="docs/page-42/assets",
    md_path="docs/page-42.md",
)
Path("docs/page-42.md").write_text(md)
# docs/page-42/assets/image-3.png exists
# md link: ![](assets/image-3.png#cfxmark:w=700)<!-- cfxmark:asset src="image-3.png" -->

# Strategy B — inline data URIs (single self-contained file).
md = cfxmark.resolve_assets(md_result.markdown, fetcher, mode="inline")
# md link: ![](data:image/png;base64,iVBORw0K...)<!-- cfxmark:asset src="image-3.png" -->

The asset markers are preserved through both strategies, so resolve_assets is idempotent and a subsequent to_cfx call always recovers the original Confluence filename — even if the visible link target has been rewritten to a sidecar path or a data URI.

Mermaid diagrams

cfxmark maps Markdown's ```mermaid fenced code block to Confluence's code macro with language=mermaid. If your Confluence instance has a Mermaid plugin installed (e.g. Mermaid Diagrams for Confluence) it will render the diagram automatically; otherwise the content is shown as a syntax-highlighted code block.

​```mermaid
graph LR
  A --> B --> C
​```

Inline opaque references

Inline elements that have no native Markdown form — Confluence user mentions, inline Jira issue macros, custom widget invocations, … — become a short Markdown link with a cfx:op-... URL:

Contact the purchaser ([@user-2c9402cc](cfx:op-4fab0f8d))

The [label] is auto-derived from the underlying element type (@user-…, jira:PROJ-1, cfx:status, …) and the op-XXXXXXXX ID is a SHA-256 prefix of the original XML payload. The full XML lives in a cfxmark:payloads sidecar at the bottom of the same Markdown file:

<!-- cfxmark:payloads -->
<!-- op-4fab0f8d
<ac:link><ri:user ri:userkey="2c9402cc83d4bcc40183d976ef730001"/></ac:link>
-->
<!-- /cfxmark:payloads -->

The SHA-256 fingerprint means a user who types that exact link syntax in their own Markdown is not silently re-interpreted as an opaque payload — the verification fails and the region falls back to ordinary text.

Block opaque blocks

Block-level Confluence content cfxmark doesn't know how to convert (e.g. drawio diagrams, plantuml, complex tables) is wrapped in a fenced code block with sentinel comments:

<!-- cfxmark:opaque id="op-1188e2b4" -->
```cfx-storage
<ac:structured-macro ac:name="drawio" ac:macro-id="...">
  <ac:parameter ac:name="diagramName">flow</ac:parameter>
  ...
</ac:structured-macro>
```
<!-- /cfxmark:opaque -->

Editors render this as a clearly visible code block — a "do not touch" signal for human readers. The Markdown parser detects the sentinels first and round-trips the contents byte-for-byte, including the original ac:macro-id UUID that Confluence uses to identify macro instances.

Header notice

When a converted Markdown document contains any opaque or directive markers, cfxmark prepends a single-line HTML comment explaining the conventions to humans and AI agents:

<!-- cfxmark:notice Converted from Confluence storage format. Inline
[label](cfx:op-XXXXXXXX) references preserve Confluence content that
has no native Markdown form; the raw XML for each lives in the
cfxmark:payloads sidecar at the bottom of this file. Do not edit
those references or the sidecar — tampering invalidates a SHA-256
fingerprint and the round trip falls back to plain text. -->

The comment is invisible in any Markdown viewer.

Custom macros

Promote a Confluence macro from "opaque" to "directive" by registering a custom handler:

import cfxmark
from cfxmark.macros import MacroRegistry
from cfxmark.macros.builtins import AdmonitionHandler

# Start from the default registry and add your own.
my_registry = cfxmark.default_registry.copy()
# Built-in AdmonitionHandler accepts one of: "info", "note", "warning", "tip".
# To promote a previously-opaque macro, write a small MacroHandler subclass —
# see cfxmark/macros/builtins/admonition.py for a complete example.
my_registry.register(AdmonitionHandler("warning"))

result = cfxmark.to_md(xhtml, macros=my_registry)

Implementing a MacroHandler from scratch requires a small amount of lxml knowledge — see cfxmark/macros/builtins/admonition.py for a complete example. A higher-level handler API that hides lxml is planned for v0.3.

Canonicalization helpers

cfxmark ships two canonicalization helpers, one for each side of the pipeline. Both are idempotent: f(f(x)) == f(x).

canonicalize_cfx(xhtml) — compare two storage fragments

Two Confluence storage fragments are "the same" only after a deep normalization pass that strips volatile attributes, editor noise, and rendering hints. Use canonicalize_cfx to compare two snapshots:

import cfxmark

c1 = cfxmark.canonicalize_cfx(original_xhtml)
c2 = cfxmark.canonicalize_cfx(round_tripped_xhtml)
assert c1 == c2  # passes for any document in the supported subset

canonicalize_cfx is the same function the test suite uses to verify byte-identical round trips against real Confluence pages. A good push pipeline calls it before the REST PUT so an unchanged body is skipped entirely:

remote = cfxmark.canonicalize_cfx(my_client.get_page(page_id))
local = cfxmark.canonicalize_cfx(cfxmark.to_cfx(local_md).xhtml)
if remote != local:
    my_client.update_page(page_id, ...)

normalize_md(markdown) — converge hand-edited Markdown

normalize_md is the Markdown-side counterpart: it runs the document through parse_md → render_md so the output is exactly the form cfxmark would have produced. Applying it before push flattens any drift introduced by hand edits, a different editor's Markdown autoformatter, or a historical cfxmark version.

import cfxmark

# Pre-push recipe: normalize hand-edited Markdown so the canonical
# XHTML body is stable across authors and editor plugins.
clean_md = cfxmark.normalize_md(local_md_from_disk)
xhtml = cfxmark.to_cfx(clean_md).xhtml

The key property: a document produced by to_md is already a fixed point of normalize_md, so round-trippers pay nothing. Hand-edited documents converge in a single pass, and that pass is enough to eliminate the "local file drifted from the round-trip form" class of bug (for example, stray ** delimiters in positions where cfxmark would have emitted raw <strong> HTML because of CommonMark's CJK word-boundary rule).

If you only push normalize_md(text) rather than raw hand-edits, the canonicalize_cfx diff above stays stable across collaborators.

Jira wiki output

to_jira_wiki converts Markdown or Confluence storage XHTML to Jira wiki markup. It accepts the same source formats as to_cfx / to_md and auto-detects which format it received.

import cfxmark

result = cfxmark.to_jira_wiki(markdown_text)
print(result.jira_wiki)   # h2. Heading\n\n*bold* text …

Two optional parameters cover common push-pipeline patterns:

import re

# Only render the body of the first H2 section titled "Summary".
result = cfxmark.to_jira_wiki(markdown_text, section="Summary")

# Drop a leading cfxmark:notice comment before rendering
# (useful when pushing a round-tripped Confluence page to Jira).
result = cfxmark.to_jira_wiki(
    markdown_text,
    drop_leading_notice=(re.compile(r"cfxmark:notice"),),
)

Code block language identifiers can be normalised for Jira Server compatibility:

result = cfxmark.to_jira_wiki(
    markdown_text,
    code_language_map={"ts": "javascript", "kotlin": "java"},
)

result.jira_wiki is None when section= is specified but not found in the document.

Confluence client (optional extra)

Install with the confluence extra to signal intent — the client itself is always importable because it is built on Python's standard library:

pip install 'cfxmark[confluence]'

The extra declares zero additional runtime dependencies. It exists to:

1. Signal the dependency in your requirements.txt / pyproject.toml so readers see that you rely on the optional subsystem.
2. Reserve a stable upgrade slot — if future convenience helpers (credential stores, rich CLI) gain third-party deps, the extra is the place they'll land.

from cfxmark.confluence import ConfluenceClient, BearerTokenFile

client = ConfluenceClient(
    host="https://confluence.example.com",
    auth=BearerTokenFile("~/.secrets/confluence_pat"),
    dialect="server",
)

# Canonical-aware push — skips the REST PUT entirely when the remote
# body is byte-equivalent to the rendered local Markdown.
result = client.push_markdown(
    page_id="12345",
    md_text=my_markdown,
    md_path="docs/my_page.md",
    on_conflict="abort",
)
if result.changed:
    print(f"Pushed. Uploaded {len(result.uploaded_attachments)} new attachments.")
    if result.has_partial_failure:
        for name, ex in result.failed_attachments:
            print(f"  ! attachment {name} failed: {ex!r}")
else:
    print("No-op; remote is already current.")

# Canonical-aware pull with resolved assets in a sidecar directory.
pull = client.pull_markdown(
    page_id="12345",
    md_path="docs/my_page.md",
    resolve_assets_mode="sidecar",
    asset_dir="docs/my_page-assets",
)

Logging. The client uses logging.getLogger("cfxmark.confluence") exclusively — no direct writes to sys.stdout or sys.stderr. Enable progress output with:

import logging
logging.getLogger("cfxmark").setLevel(logging.INFO)

Confluence dialect. The default is dialect="server" because Confluence Server / Data Center is the reference test target. Confluence Cloud users should pass dialect="cloud" — the X-Atlassian-Token: no-check XSRF bypass header (mandatory on Server, unsupported on Cloud) is gated on this setting. Cloud support is best-effort; if you hit a Cloud-only regression, please open an issue.

Jira wiki (experimental, lossy)

The Confluence round-trip (to_cfx / to_md) is lossless after canonicalization — every construct in the supported subset round-trips byte-for-byte, and everything else is preserved through the opaque-block mechanism. The Jira wiki direction is not. Jira wiki markup is a looser dialect without opaque-macro identity preservation, and several constructs have no equivalent on the Markdown side.

Keep the two contracts at different import sites so the asymmetry is visible:

from cfxmark import to_cfx, to_md           # lossless, Confluence side
from cfxmark.jira import from_jira_wiki, to_jira_wiki  # experimental

Contract

The strongest guarantee the Jira pipeline offers is:

from_jira_wiki(jira_wiki) produces a :class:ConversionResult whose Markdown representation reaches a fixed point after at most two wiki → md → wiki iterations. Pass 1 (wiki → md) is a one-way canonicalization; pass 2 (md → wiki → md) must be idempotent. The real-world fixture corpus (6 Jira issue descriptions drawn from production) is exercised in tests/unit/test_jira_wiki_parser.py to pin this contract.

Explicitly allowed canonicalization (not considered a diff):

• Heading spacing, list indent, trailing whitespace normalisation
• Soft-break inside a paragraph collapsed into a single line
• Jira {panel} macro mapped to {note} admonition
• _italic_ canonicalised to Markdown *italic*
• -strike- canonicalised to Markdown ~~strike~~
• ~sub~ → <sub>, ^sup^ → <sup>, +ins+ → <ins> (v0.4+)
• {color:#hex}text{color} → <span style="color:#hex">text</span> (v0.4+)
• ??text?? → <cite>text</cite> (v0.4+)

Explicitly forbidden (would break the contract):

• Content loss inside headings, paragraphs, list items, code blocks
• Re-ordering of top-level blocks
• URL rewriting in links
• Renaming attachments (filenames in [^file] / !file! are preserved verbatim in ConversionResult.attachments)

Lossy mapping table

Jira wiki	Markdown	Note
h1.…h6.	#…######	1:1 identity (no promotion on parse)
*bold*	**bold**	boundary-aware
_italic_	*italic*
-strike-	~~strike~~	boundary-aware
{{mono}}	`mono`
~sub~	<sub>sub</sub>	Subscript node (v0.4+)
^sup^	<sup>sup</sup>	Superscript node (v0.4+)
+ins+	<ins>ins</ins>	Underline node (v0.4+)
{color:#hex}x{color}	<span style="color:#hex">x</span>	ColorSpan node (v0.4+)
??text??	<cite>text</cite>	Citation node (v0.4+)
[url]	[url](url) (empty label form)
[label|url]	[label](url)	label may contain nested [...]
[^file.png]		when extension is image-like
[^file.msg]	[file.msg](attachment:file.msg)	otherwise
[~user]	(dropped)	warning recorded
{code:python}body{code}	```python\nbody\n```
{noformat}body{noformat}	```\nbody\n```
{quote}body{quote}, bq.	> body
{info}/{note}/{warning}/{tip}	> [!INFO] callout	GitHub / Obsidian style
{panel:title=X}body{panel}	> [!NOTE] X + warning	D4 mapping
h3. *Title*	### **Title**	literal nested bold preserved
Multi-line table cell	GFM cell with <br> soft break
||h1||h2|| header row	GFM header row
----	---

Warnings accumulate on ConversionResult.warnings; fixture tests in tests/unit/test_jira_wiki_parser.py pin the behaviour of every entry in this table.

Heading promotion on the output side

to_jira_wiki has a heading_promotion keyword that controls the heading level mapping:

• "confluence" (default) — Markdown H3 collapses to Jira h2 (and H4 → h3, …). Use when the Jira wiki output will be pushed to a Confluence page whose title already occupies the top slot.
• "jira" — identity mapping. Use when the output is pushed to a Jira issue description, because the issue title lives in a separate field and the body can start at h1.
• "none" — alias for "jira".

from cfxmark import to_jira_wiki

# Confluence push — historical default
to_jira_wiki(md)

# Jira issue description push
to_jira_wiki(md, heading_promotion="jira")

HTML comment passthrough (wrapper metadata)

If your wrapper embeds caller-owned metadata as HTML comments in the local Markdown file — for example a workflow manifest:

# My feature

<!-- workflow:meta
  key: TASK-42
  type: Story
  last_synced_version: 15
-->

body...

— opt in via ConversionOptions.passthrough_html_comment_prefixes. Matching comments are preserved verbatim across parse_md / render_md, and silently dropped by to_cfx / to_jira_wiki so they never leak to Confluence or Jira:

from cfxmark import ConversionOptions, to_cfx, to_md
from cfxmark.normalize import strip_passthrough_comments

opts = ConversionOptions(
    passthrough_html_comment_prefixes=("workflow:",)
)

# Push: comment is dropped on the way to Confluence
result = to_cfx(local_md, options=opts)

# Canonical compare: strip comments on both sides before diffing
left = strip_passthrough_comments(local_md, ("workflow:",))
right = strip_passthrough_comments(pulled_md, ("workflow:",))
assert left == right

cfxmark: prefixes are filtered out so cfxmark's own sentinel comments (cfxmark:opaque, cfxmark:notice) cannot be hijacked.

Security

cfxmark hardens its XML parser against XXE and billion-laughs attacks:

• Inputs containing <!DOCTYPE> or <!ENTITY> declarations are rejected before lxml ever sees them.
• The lxml parser is configured with no_network=True, load_dtd=False, and huge_tree=False.
• Opaque-block sentinels are SHA-256 verified — accidental sentinel syntax in user-typed Markdown does not become a real opaque block.

If you find a security issue, please open a GitHub issue.

Stability contract

The following names are covered by semantic versioning and will not be removed or incompatibly changed without a major version bump:

cfxmark package — to_cfx, to_md, to_jira_wiki, from_jira_wiki, canonicalize_cfx, normalize_md, strip_passthrough_comments, resolve_assets, ConversionResult, ConversionOptions, DEFAULT_OPTIONS, AssetFetcher, ResolveMode, CfxmarkError, ConversionError, MacroError, ParseError, AssetSecurityError, MacroRegistry, default_registry.

cfxmark.jira — to_jira_wiki, from_jira_wiki. Experimental in v0.3: the Jira parser contract is "converges after at most three wiki → md → wiki iterations", not byte-identical. The top-level to_jira_wiki, from_jira_wiki, and ConversionResult.jira_wiki field ARE stable — only the quality of the round-trip itself is experimental. See the "Jira wiki (experimental, lossy)" section above for the full contract.

cfxmark.confluence — ConfluenceClient, PushResult, PullResult, Auth, BearerToken, BearerTokenFile, BasicAuth, EnvBearerToken, HTTPError, ConfluenceVersionConflict.

Guarantees:

• Breaking changes bump the minor version for 0.x.y releases.
• canonicalize_cfx normalization rules are cumulative — each release is a strict superset of the previous release's canonicalization.
• Deprecations are announced one minor version before removal.

Not covered: underscore-prefixed symbols, parsers.* / renderers.* / ast.* internals, logging message wording, ConversionResult.document AST shape, warning message wording.

Note: 0.x.y versioning is looser than 1.x.y — minor version bumps may carry breaking changes as noted above.

Development

git clone https://github.com/eunsanMountain/cfxmark
cd cfxmark
uv sync --all-extras

# Run all tests
uv run pytest

# Type-check
uv run mypy src/

# Lint
uv run ruff check .

# Build
uv build

The corpus tests look for .cfx files in tests/corpus/ (gitignored to keep your own private samples out of version control). Drop your own Confluence storage XHTML there and they will be exercised by pytest tests/test_corpus.py.

License

MIT. See LICENSE.