1.9.0a1 — the notebook looks things up instead of summarising everything
The first preview of the 1.9.0 line, where the personal notebook (being renamed co rem, #1932) becomes long-term supported. Stable is 1.8.9.
Investigation searches files (#1850)
Investigating the owner's own page on a real notebook took 40 model calls, 16.2M input tokens and about 75 minutes, 39 of those calls summarising chunks of material before the one call that wrote the page.
- When the gathered material does not fit one turn, it is written to files
instead: one per mail or attachment, one per coding session or chat, and an
index.mdlisting them by date, sender and subject. - Every entry starts with its source id, which is what the page cites and what the validator accepts.
- The one investigate turn gets the page and the index and searches with
rg/sed, reading only what eachUnknownfield needs. No digest pass. - The files are deleted when the run ends.
Every turn carries less (#1851)
Each Wiki Skill now holds only rules; the stories behind them moved to
docs/wiki-skills/. A one-page turn's instructions:
| Turn | Before | After |
|---|---|---|
| Investigate a person | 30.6k | 14.6k |
| Investigate a project | 31.1k | 14.9k |
| Investigate an org | 63.7k | 14.6k |
| Maintain one person page | 22.9k | 14.4k |
A test holds every one-page turn under 15k characters, and
co wiki logs --usage reports each stage's instruction size.
Smaller fixes

co wiki list peoplelists the people you mail most first, the order investigate works through them (#1670).- Investigating one page reads the WhatsApp chats you chose; on your own page it takes only what you wrote (#1625).
Install
python -m pip install --upgrade 'connectonion==1.9.0a1'
co wiki status
Known limits
- The #1850 and #1851 acceptance runs (the four measured subjects re-run on the same real data, compared for cost and for the facts found) are not done yet. This preview is how they get done.
- The command is still
co wiki; the rename toco rem(#1932) comes in a later 1.9.0 preview.
ConnectOnion