ConnectOnion 1.9.0a33 (preview)

Install this preview

Stable remains 1.8.10. Install this opt-in preview by exact pin, then follow the first-run guide:

python -m pip install --upgrade 'connectonion==1.9.0a33'

Keep the start of a relationship honest

The Person investigation prompt now distinguishes the earliest retained message from the first contact. If that message refers to an earlier application, introduction or conversation whose date is missing, it asks the writer to keep First contact as Unknown and name the gap in Uncertainties. A dated acknowledgement can still appear in History with its own citation. See #2261 and the wider claim audit #2223.

Fact extraction no longer labels the earliest retained mail as First contact. The People index takes that date only from an explicit page fact, including when it reads an index built by an earlier preview. Existing pages with a previously written first-contact claim still need their sources reviewed.

The Person instruction now keeps a booking's displayed time zone separate from each participant's own time zone unless an original labels the person directly. On narrow phones, the Person lead has room for five shorter lines so the decisive outcome can remain visible beside the source link. See #2263 and #2264. The Now summary names a source-backed reason when a thread closes, because Home and the Person first fold use that sentence directly.

First-run costs and release wait

The first-run preview now uses planning rates from a partial 90-day concurrent run and separate full Person and Skill reads. It uses the notebook's own completed-run medians when available. The terminal shows the configured investigation target (35% of a weekly allowance by default) and the configured safety floor (90% by default). The target is advisory; the floor stops new pages, while pages already running finish. Input-token estimates include cached tokens and are not weekly quota points. See #2260 for the remaining measured-progress work.

The release workflow now builds the tagged package beside the test matrix. PyPI publication still waits for both jobs to pass, and the built wheel still gets its installed-package check. This removes a serial wait; the actual time saved will be measured on this release.

Trial limits

On one protected five-source Person sample, the public a32 wheel produced an accepted page with a false First contact date. An independent role-based AI review compared all five cited originals in the rendered desktop and phone reader and found that P2. The first a33 rerun corrected that date, but review found an unsupported participant time-zone attribution and a clipped phone summary. A second run cited an attachment that explicitly labels the invitee's time zone; review then found that its short Now summary omitted the reason the thread closed. A third fresh run left First contact and the person's own time zone Unknown, stated the cited closure reason in Now, and kept the People table's first contact empty. It used 188,486 input tokens (165,504 cached) and 5,430 output tokens. An independent role-based AI reviewer opened all three cited originals and checked Home, the Person page, Full memory, People, privacy, and 1440/375-pixel layouts. No P1/P2 remained in that sample; a one-off user-authored booking note was not included as an example in the communication context (P3 #2265). The full 258-page first-run cohort has not been completed on this preview. One corrected page does not establish that all Person claims or pages are accurate.