## Summary
Fixes two bugs in
`MessagingMessageService.saveMessagesWithinTransaction`.
### Bug 1 — `ON CONFLICT DO UPDATE command cannot affect row a second
time`
Reported in production logs from `MessagingMessagesImportService`.
Postgres rejects a single `INSERT … ON CONFLICT DO UPDATE` when the same
conflict target appears twice in the values list, and that's exactly
what was happening to `messageThread`.
Where it came from: #19351 added the message-thread subject refresh
feature and, in doing so, switched the existing thread `insert` to a
bulk `upsert(['id'])` over a list built by concatenating two sources:
```ts
const threadsToUpsert = [
...messageThreadsToCreate, // brand-new thread rows
...threadSubjectUpdates entries, // subject refreshes for existing threads
];
await messageThreadRepository.upsert(threadsToUpsert, ['id'], txManager);
```
Each list is internally unique, but they are **not disjoint**.
`enrichMessageAccumulatorWithMessageThreadToCreate`, when it sees two
messages in the same batch sharing a brand-new thread external id,
copies the first sibling's freshly-minted thread id into the second
sibling's `existingThreadInDB`. The subject-update gate later in the
loop then trusts that field and queues a subject refresh for that id —
which is also already in `messageThreadsToCreate`. Same id, same
statement, two rows → Postgres aborts the transaction and the import
retries forever on the same batch.
**Fix:** stop merging the two lists. Issue creates and subject updates
as two separate statements within the same transaction:
```ts
if (messageThreadsToCreate.length > 0) {
await messageThreadRepository.insert(messageThreadsToCreate, txManager);
}
if (threadSubjectUpdates.size > 0) {
await messageThreadRepository.upsert(
Array.from(threadSubjectUpdates.entries()).map(([id, { subject }]) => ({ id, subject })),
['id'],
txManager,
);
}
```
This is closer to the pre-#19351 shape (`insert` for new rows) and
side-steps the duplicate-row constraint entirely: each statement is
internally unique (creates use freshly minted UUIDs; updates are keyed
by a `Map<id, …>`), and within the same transaction Postgres happily
applies a subject update to a row inserted by a previous statement.
### Bug 2 —
`enrichMessageAccumulatorWithExistingMessageChannelMessageAssociations`
clobbers the accumulator
Independent latent bug spotted while tracing the flow. The helper did:
```ts
if (existingMessageChannelMessageAssociation) {
messageAccumulatorMap.set(message.externalId, {
existingMessageInDB: existingMessage,
existingMessageChannelMessageAssociationInDB: existingMessageChannelMessageAssociation,
});
}
```
i.e. it **replaces** the accumulator object, dropping the
`existingThreadInDB` set just before by
`enrichMessageAccumulatorWithExistingMessageThreadIds`.
The branch only fires when re-encountering a message that's already been
fully synced on this channel (matched on `headerMessageId` AND already
has an association row) — i.e. routinely on Gmail/IMAP incremental syncs
whenever the connector re-delivers an existing message (label change,
read/unread, archive, full-resync after error, …).
When it fires, the next enrichment step sees `existingThreadInDB` as
`undefined`, falls into the "create a new thread" branch, mints a fresh
`threadToCreate`, but the main loop never queues a `messageToCreate` or
association for it (because both `existingMessageInDB` and the existing
association are still set). Net effect: **one orphan `messageThread` row
inserted per re-encountered message, with nothing referencing it.**
The existing message in the DB keeps pointing at its real thread, so
this is invisible to users — no thread fragmentation, no UI symptoms, no
error logs. Just slow accumulation of orphan thread rows that no query
joins onto. Probably worth running
```sql
SELECT COUNT(*)
FROM "messageThread" mt
WHERE NOT EXISTS (
SELECT 1 FROM message m WHERE m."messageThreadId" = mt.id
);
```
on a busy production workspace once this lands to size whether a cleanup
migration is warranted.
**Fix:** mutate the existing accumulator in place instead of replacing
it.
## Test plan
- [x] `oxlint --type-aware` clean on touched file
- [x] `prettier` clean
🤖 Generated with [Claude Code](https://claude.com/claude-code)
