[backmeupplz]

GET /v2/farcaster/notifications?fid=N&include_quotes_and_deep_replies=true also returns:

• casts quoting the user's casts, as type: "quote" (the quoting cast as cast, its author as user), found through the cast quotes index;
• replies deeper in the threads under the user's casts (up to five levels below the direct replies), as type: "reply";
• recasts as type: "recasts" instead of likes.

They are merged into the same time ordering and cursor as the other notifications. Without the flag the response is unchanged. Clients otherwise have to look up quotes with one request per cast.

While in there, two fixes for why the quotes index looked like it missed many quotes:

• CastQuotesIndexer::get_quotes returned true (stop) from the iterator callback after the first key, so /v2/farcaster/cast/quotes returned at most one quote per cast. The same inverted check was in UserDataIndexer::get_fids_by_value_prefix.
• The block backfills (run_block_backfill and the hyper backfill) passed end as the stop block to get_shard_chunks, which is exclusive, so the last block of every 100-block batch was never replayed. Already-backfilled nodes keep that gap until their checkpoints are reset.

[martinvr]

Does this fix require existing nodes to rerun the backfill to recover previously skipped notifications, or are those blocks eventually indexed through another mechanism?

But probably not worth rebuilding indexes for. Users generally don't care about notifications from weeks or months ago.

[backmeupplz]

Merged main (0.16.0) in, so the conflict is gone. It was only in server_tests.rs, where both sides added tests at the end; I kept both. cargo test --lib api:: passes (301), and the notification tests (92).

On the backfill: notifications themselves don't depend on it. Replies, deep replies, mentions, likes, recasts and follows are read straight from the cast, reaction and link stores, so they're complete on every node, old blocks included. Only the quote notifications (behind the new flag) and /v2/farcaster/cast/quotes read the cast quotes index.

The off-by-one only affected the index backfill: it skipped the last block of every batch, so about 1 in 101 historical blocks. Blocks indexed live after a node finished its backfill are complete. A node that already backfilled keeps that ~1% gap in its indexes (quotes, and the other indexers that replay the same way) unless its indexer checkpoints are reset so they replay. Agreed, not worth a rebuild for old notifications. New nodes get it right from the start.