No blockers raised; residual cleanup:
- processor.py: guard int(doc_task.doc_id) with is_valid_nextcloud_doc_id in
the mail_message branch (consistent with search/context.py + the verifier).
- mail metadata symmetry: store `bcc` in file_metadata and the Qdrant payload
alongside cc (build_mail_content already emits a Bcc: line).
- server/mail.py: extract _cap_attachment_content helper (byte-accurate cap)
and unit-test it (small/None/oversized/multibyte).
- client/mail.py: give the synthetic OCS-error Response an explicit empty body;
add a test that a traversal-style attachment_id is percent-encoded.
- models/mail.py: clarify ListMessagesResponse.total_count is the page count,
not the mailbox total.
Deferred (Deck #376): _potentially_deleted doc_type-in-key. It's pre-existing
and spans ~30 sites across the notes/news/deck/file/mail scanners (whose
deletion paths have no unit coverage), so it belongs in its own focused PR
rather than expanding this mail PR's blast radius into other doc types.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
No blockers raised; hardening + clarity:
- client/mail.py: URL-encode the caller-supplied attachment_id
(quote(..., safe="")) — defense-in-depth against path traversal.
- server/mail.py: measure attachment content in UTF-8 bytes (not characters)
for the size cap and the sentinel message.
- scanner.py: bound _mail_cap_logged (insertion-ordered dict + oldest-first
eviction at 50k, mirroring _consent_backstop_done) so the cap-log dedup set
can't leak in a long-running multi-tenant process; reword the cap log to not
imply MAIL_SCAN_MAX_PER_MAILBOX is operator-tunable (it's the Mail OCS max).
- models/mail.py: comment why GetAttachmentResponse doesn't nest MailAttachment
(different OCS endpoint shape).
- mail_content.py: document format_mail_addresses' empty-entry skip contract.
_potentially_deleted doc_type-in-key remains tracked as Deck #376 (pre-existing
cross-cutting; reviewer confirmed deferral).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Address PR #935 round-3 review:
- search/verification.py: rewrite _verify_mail_messages to batch by mailbox.
get_message triggers a server-side IMAP body fetch, so per-result verify
issued one IMAP FETCH per hit; now it calls the DB-cached list_messages once
per mailbox (mailbox_id comes from the Qdrant payload via result.metadata)
and intersects — O(unique mailboxes) light calls instead of O(results) IMAP.
- vector/mail_content.py: include Cc/Bcc in the indexed text so recipient
queries match; move MAIL_SCAN_MAX_PER_MAILBOX here (shared by scanner index
window + verifier presence window) with a note that it equals the Mail OCS
per-request max (100), so it's a fixed constant not a config knob.
- client/mail.py: clamp list_messages limit to 1..100 at the client layer.
- tests: add test_mail_content.py (exact-layout contract for build_mail_content);
rewrite the mail verifier tests for the batch-per-mailbox shape.
Left as-is: ValidationError isn't caught in the list-endpoint tools — consistent
with nc_notes_*/nc_deck_* and not a regression.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- server/mail.py: guard nc_mail_get_message against an empty OCS payload so
MailMessage(**{}) can't raise an uncaught ValidationError (returns a clean
'not found' instead).
- server/mail.py: cap inlined attachment content at MAX_ATTACHMENT_CONTENT_BYTES
(5 MiB), replacing oversized bodies with a sentinel so a large attachment
can't blow up the MCP response.
- client/mail.py: harden the OCS meta statuscode parse against a non-numeric
value (treat as success) instead of letting int() raise an uncaught
ValueError.
- scanner.py: log the newest-N cap hit once per (user, mailbox) at info level
(discoverable without flooding multi-tenant logs on every scan tick).
- tests: add incremental-sync scanner cases (new message queued, reappeared
message clears grace, deletion after grace expiry).
Deferred (tracked, card #376): include doc_type in the _potentially_deleted
grace-period key — a pre-existing cross-cutting collision the reviewer flagged
as a follow-up, not a blocker.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
- models/mail.py: lowercase `list` generics per CLAUDE.md convention.
- client/mail.py: _ocs_get now inspects ocs.meta.statuscode (re-raises >=400
as HTTPStatusError carrying the OCS code so callers' 404/403 handling
applies) and guards response.json() against non-JSON bodies (RequestError).
- Extract the duplicated _format_addresses + content reconstruction into
vector/mail_content.py, used by both processor.py and context.py (fixes the
SonarCloud new_duplicated_lines_density gate).
- processor.py: add the missing mail_message Qdrant payload block so the
computed mail metadata (subject/from/to/cc/date_int/has_attachments/
account_id/mailbox_id) is actually stored, not dropped.
- Rename the list_messages `filter` param to `search_filter` (avoid shadowing
builtins.filter); still maps to the OCS `filter` query param.
- Docstring notes: has_more heuristic, attachment content size.
- Tests: OCS meta-failure + non-JSON client paths; initial-sync scanner tests
(tests/unit/vector/test_scanner_mail.py).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Add read-only support for the Nextcloud Mail app, plus semantic indexing
of mail messages. The MCP server never speaks IMAP/POP3 itself: it calls
the Mail app's CSRF-free OCS API (/ocs/v2.php/apps/mail/api/...) with the
existing Basic-Auth app-password flow and an OCS-APIRequest header, and the
Mail app handles IMAP server-side.
- client/mail.py: MailClient (accounts, mailboxes, messages, message,
attachment), OCS-envelope aware.
- models/mail.py: Pydantic models with the API's camelCase aliases.
- server/mail.py: 5 read-only MCP tools (mail.read scope), registered in
AVAILABLE_APPS.
- Vector pipeline: new "mail_message" doc_type wired into scanner
(scan_mail_messages, newest-N per mailbox), processor (body -> markdown
embedding), per-id verifier, and context expansion.
- Tests: client API, model round-trips, verifier behavior; consent-backstop
test now derives its allowed set from INDEXED_DOC_TYPES.
- README + semantic-search docstrings updated.
Requires Mail 5.x / Nextcloud 32+ and a mail account configured in the
Mail app. Follow-up: astrolabe must advertise "mail_message" in its
enabled_doc_types capability for search under admin doc_type restriction.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
