Support: markdown escaping leaks into the text/plain part of outbound email replies

Outbound email replies from the Support inbox are sent as multipart/alternative. The text/html part renders correctly, but the text/plain part contains markdown escape sequences as literal backslashes.

In products/conversations/backend/tasks.py, the outbound body is built as:

if comment.rich_content:
    html_body = rich_content_to_html(comment.rich_content)
    txt_body = rich_content_to_markdown(comment.rich_content, include_images=False)

text/plain isn't markdown, so escapes intended for a markdown renderer are delivered verbatim. Captured from a live reply — the two parts of the same message:

Content-Type: text/plain; charset="utf-8"

ok\. tell me more

Content-Type: text/html; charset="utf-8"

<p>ok. tell me more</p>

Note the ok\. in the plain-text alternative.

Impact. Anyone reading the text/plain part sees the backslashes: plain-text mail clients, assistive tech falling back to that part, and anything downstream parsing it. Agents typing ordinary prose in the rich editor hit this on every reply — it isn't specific to short messages, and it's invisible to the agent because the composer and the HTML part both look right.

We observed it on a period and haven't characterised which other characters the serializer escapes.

Suggested fix. Generate the text/plain alternative with a plain-text serializer (or unescape after serializing), and keep markdown output for the paths where a markdown renderer actually consumes it.

Debug info

PostHog Cloud (US) · Support email channel, Mailgun transport · observed 2026-09-03

Start in products/conversations/backend/tasks.py at the rich_content branch, then trace rich_content_to_markdown and the corresponding HTML serializer. Verify how the multipart email body is assembled and which serializer is intended for plain text. Done means the text/plain part contains no markdown escape sequences while the text/html part remains unchanged; cover the observed period case and any other escaping behavior identified.