ConnectOnion 1.9.0a41 (preview)

Install this preview

Stable remains 1.8.10. Install this opt-in preview by exact pin:

python -m pip install --upgrade 'connectonion==1.9.0a41'

New notebooks run on Claude Code with Sonnet

A new notebook's config.yaml now says runner: claude-code and model: claude-sonnet-5-5, so model turns use the owner's own Claude plan instead of their Codex week. A notebook made before this preview keeps the runner and model it saved; switch one with co rem config set runner claude-code model claude-sonnet-5-5.

The first run's ten workers also work on this runner now. Every model turn starts from the same task folder, and the Claude Code adapter used to lock that folder for every headless run, so nine of ten pages failed with "Claude Code already owns this workspace". Only a turn that resumes a session or runs through the Hook bridge takes the lock now (#2294).

The choice was measured on one owner's real mailboxes (Gmail and Outlook, 90-day map: 69 people, 18 organisations, 18 projects, 164 skills) with the same three hand-checked people pages as a benchmark:

	Codex gpt-6-luna	Haiku 4.5	Sonnet 5.5
Pages written	earlier, wider map	4 of 17 projects, 121 of 164 skills	16 of 17 projects, all 158 skills, 67 people
Benchmark	17/25	15/25	17/25
Privacy labels written	34	1	31
Wall clock	37 min	about 63 min	52 min

The Sonnet run took about five points of the account's seven-day Claude window. Headless claude -p turns draw from the subscription's usage limits today; Anthropic paused a separate Agent SDK credit on 2026-06-15 and has said it will announce any change in advance. Remaining page-quality gaps, including unlabelled private sentences, are tracked in #2283 and #2118.

Find the people outside the recent window

co rem init --investigate-all --estimate-only scans available connected mail metadata since 1970 and local session history, then shows the selected owner, People, Project, Organization, and Skill work with an input-token and time forecast. It performs no model turns, mail-body archive, or scheduling. A failed mailbox or time window is named and makes the estimate a lower bound. The ordinary first run still uses its shorter source window.

The People reader now separates written person memories from one-off correspondents. Historical contacts without a page remain searchable, with source coverage and a command to prepare a memory. A phone shows those contacts as cards with a copyable command. Long citation tips stay inside the viewport. The design review records the sampled before and after states.

Gmail's REM history path now uses bounded direct profile, alias, listing and metadata requests. It reads up to eight message headers at once in a window; full provider pages split before headers are fetched. The sender, recipients, date and conversation identity remain available to the map. A failed read does not become a complete result.

Nightly upkeep follows init

After its foreground pass, co rem init presents the connected sources and schedule for approval and then installs the existing nightly REM upkeep. The first batch is not repeated. --yes explicitly approves this in a script; --no-start leaves background work off. In a noninteractive run without --yes, the schedule waits for co rem start. Estimate-only never schedules.

What was checked

A private Gmail and Outlook lifetime metadata estimate completed in 844 seconds. It observed 22,411 mail headers, mapped 634 People records and ten projects, and retained 803 low-signal contacts for review. Both mailboxes finished with zero reported errors and zero incomplete windows. This was a map and estimate only; no mail body or model turn was read by the trial. It selected the owner's page, 589 person pages and eight project pages. With ten workers and no completed local page samples, those 598 pages forecast about 660 million billed input tokens and 297 minutes. Including organizations and Skills yields 886 pages, about 751 million tokens and 364 minutes. These are planning estimates, not observed investigation durations or weekly quota percentages. No private contact data was published. The reader review used invented content on 375/390-pixel phones and desktop, in light and dark states, including contact search, the command state and a long citation tip.

Trial limits

The lifetime scan counts addresses seen in connected mail; it does not prove that every address is a person or that the written memories are accurate. The cost forecast uses defaults until the notebook has completed page samples. The configured investigation target is advisory, while the configured weekly safety floor still stops new model work. A full all-person model run, claim-by-claim review and a public-wheel real-data reader trial remain open. Gmail rate limits and resumability of a stopped lifetime scan are tracked in #2287.