Keep a copy of what was sent, where the sender looks for it

[rutgerhofste]

The bug

mail_send from a Soverin mailbox reported Sent with a real message id, the
recipient got the mail — and mail_search on Sent found nothing. It never
had. SMTP delivers a message and keeps nothing, and most IMAP hosts (Soverin
included) file no copy server-side either, so the sender's own record of the
message did not exist. A send you cannot find afterwards reads as a send that
never happened.

The fix

provider.send finishes the job. smtp.send hands back the exact RFC822
bytes it delivered plus the moment it did, and provider.py — the one place
that sees both halves, exactly as with threading — APPENDs them to the Sent
folder \Seen, with the send time as INTERNALDATE so the copy sorts where it
belongs. Rebuilding the message instead would mint a fresh Message-ID and Date,
and the copy would no longer be the mail the recipient got.

Four decisions hold it up:

• The folder is read, not guessed. The \Sent SPECIAL-USE attribute
  (RFC 6154) off the same LIST mail_list_folders already returns. The name is
  localized (Verzonden items), sometimes under INBOX, and Sent Items /
  Sent Messages are both common — Sent is only the fallback when a server
  advertises nothing. Resolved once per session and cached; a failed listing
  is deliberately not cached.
• A failed APPEND does not fail the send. The message is with the recipient
  by then, so raising would report a delivered mail as undelivered and invite a
  retry that sends it twice. The outcome rides back as saved_to_sent /
  sent_folder on SendResult, and the tool docstring tells the client to say
  so and not resend.
• The Bcc header stays on the copy and only comes off what leaves: your own
  Sent folder is the only record of whom you blind-copied.
• A host that files its own copy gets one, not two — the Message-ID is
  searched for in the folder before the APPEND.

A reply travels this same path, so it is filed like any other message,
threading headers and all.

Tests

• tests/test_sent_copy.py (new): SPECIAL-USE discovery incl. a localized
  name, the Sent fallback, resolve-once, a failed listing not being cached,
  the APPEND's bytes/flags/INTERNALDATE, the no-duplicate probe, the provider
  joining SMTP and IMAP, a reply being filed, an APPEND failure being reported
  rather than raised, and Bcc on the copy but not on the wire.
• tests/test_tools.py: mail_send reports the copy; a failed copy is a
  successful send with saved_to_sent=False.
• GreenMail e2e: sends for real and reads the copy back out of Sent, for a
  new message and for a reply. (Not run here — no Docker in this environment;
  make test is green: 221 passed. Ruff + ty clean.)

Companion change

pantalytics/squirrel-mcp-admin#… — the Graph backend reports
saved_to_sent=True / Sent Items instead of filing anything, because
/send already leaves the message in Sent Items server-side. That report is
what keeps a second copy from being filed on top of Graph's. Shipping the fix
to the hosted product also needs ARG SQUIRREL_MCP_REF in the admin
Dockerfile bumped to this commit once merged.

🤖 Generated with Claude Code

https://claude.ai/code/session_01PPDJU6WSgJ7cQdfp8GoHXC

---

Generated by Claude Code