Minor release, and a breaking one. Migration 0023 deletes rows. Project knowledge is now
written by agents rather than assembled from threads, and the entries the old mechanism
produced are removed.
Removed
The automatic thread extractor. It took a closed thread, used its subject as the title and
its last agent message as the body, cut to 2000 characters, and stored that as knowledge.
There was no distillation step anywhere in it - what accumulated was a copy of the
conversation, and enough of it to bury anything worth reading.
The numbers are why it went rather than got improved. Across the two hubs this was measured
on, 1128 entries carried the thread's subject verbatim and 3 had been written by an agent.
The three read as instructions - one titled "없는 경로 404 통합테스트는 반드시 permitAll
경로로 짠다" against a thread called "[01] 없는 경로 404/405 - 적대리뷰 재요청(bong)". Those
three were written with no guidance at all, from the tool schema alone.
Changed
close carrying a lesson is the only automatic way knowledge is created, and the playbook
finally says so. That argument shipped in 0.6.0 and the playbook never mentioned it, which is
why nearly nobody passed one and everybody saw the fallback. The new section says distil rather
than summarise: the title is the query a future agent would search for, the body is why it is
true and then what to do. It also says plainly that closing without a lesson is correct when
a thread taught nothing durable - guidance that only pushes toward writing produces filler, and
filler occupies the place a real lesson would.
Lessons are stored as source_kind = 'lesson'. They previously shared 'thread' with
extractor output, which made the two indistinguishable in the data - the deletion below would
have taken the replacement along with the trash.
Knowledge entries have somewhere to lead. Each opens on its own page and links back to the
thread it came from, so an entry can be short without that being a loss. Where a source cannot
be resolved the page says the thread is gone and that the entry is what remains of it, rather
than showing nothing: a missing link and a lost source are different facts.
Migrations
0023 identifies rows by reconstructing them, not by a label. For each candidate it rebuilds
what the extractor would have produced from the same thread and deletes only byte-exact
matches. Rows whose title is not the thread's subject are relabelled as lessons. Anything it
cannot attribute is left alone - including rows whose thread is gone, and rows whose body was
redacted at extraction time and so no longer matches the message.
The asymmetry is deliberate: deleting an agent's lesson cannot be undone, while leftover
clutter is visible and removable. Rows that were promoted or marked trusted are never touched,
because a human who read one and approved it made a later and more specific judgement than this
migration can. The extraction watermarks go with the deleted rows - leaving them would mean
those threads could never receive the lesson meant to replace what was removed.
A note on our own verification
The machinery around this extractor was the most heavily reviewed work in 0.6.0: a watermark
table, a sweep guard, a lock-order fix, a migration backfill tested against its own SQL, and
fourteen adversarial review rounds. All of it was correct. None of it asked whether the rows
coming out were worth reading, and that question was answered in a sentence by the first
person who opened the screen.
Precision about a mechanism is not evidence about its output. The reviews measured one and we
reported the other.