kert ed55a8f883 fix(llm): empty-text items are stamped as final and count toward docket completion (refs #615) ...

Treating a blank load() as unindexable defeated the feature it was
meant to protect: four of the eleven historical dockets hold 1-2
comments with no text at all, so none of them could ever be marked
complete and ~130k items would be re-listed on every run.

An item with no text is a finished state, not a failure. Stamp it into
index_state with content_hash='' / chunk_count=0 and its fingerprint —
the empty hash cannot collide with a real sha256 — so the next run
fingerprint-skips it without a load, and it no longer blocks its
docket's completion. Text appearing later changes the fingerprint and
the item is embedded normally.

Only a ref that loaded *with* text and chunked to nothing stays
unindexable and blocks index_docket_state; the embed/write path already
aborts the whole run on error. The index_state upsert both callers need
is now _record_state().

2026-09-08 15:52:34 -04:00

6.7 KiB

Raw Blame History

View Raw