1.8.9b18 — Pages that finish

An opt-in preview on the 1.8.9 fix line (#1722). Stable is 1.8.8.

Person and project pages, tested before they ship

The skills that write a person's page and a project's page now have test suites, in examples/wiki-skill-evals/. They cover a client with a signature block, a partner who writes in Chinese, two people with the same first name, a forwarded mail full of someone else's claims, a calendar invite and nothing else, a mail that tells the AI to record a false debt, a deploy that was only planned, an API key pasted into a session, and a workspace folder.

Every failure in the first run was the same one. The agent read everything it needed in two steps and then searched for example pages to copy until it ran out of steps. The skills now say that the page and the material are the whole input: read both, write the page once, check it once, and stop. A section with no evidence stays Unknown instead of being filled with "no prior history", and a project's current state names when anything last happened. After the changes every case passed on both of its runs (#1883).

Good pages are no longer refused

The check that decides whether an investigated page is saved refused a flow diagram drawn with │ and ▼, and read [1, 2] as no citation at all. Four of five good project pages in the test suite would have been thrown away. Both are accepted now.

Investigation reads its evidence directly

Investigation used to cut a large subject's evidence into pieces too big for the prompt, so each piece went to a file split into 64-character fragments, which the model had to stitch back together. Now each piece fits in the prompt and is read in one or two steps instead of nine to twenty-two, and material too big to include is plain text the model can read or search. On the owner's business partner, whose page timed out before, the page is now written (#1884).

Install

python -m pip install --upgrade --pre 'connectonion==1.8.9b18'

Known limits

• A subject with a lot of mail still costs millions of tokens, because investigation still summarises everything first. Searching the evidence instead is the next change (#1850).
• A finished investigation is discarded if a scheduled update holds the notebook at the moment it saves (#1885).