ConnectOnion 1.9.0 (stable)

Install

python -m pip install --upgrade 'connectonion==1.9.0'
co rem init

co rem init maps the last 90 days by default, then attempts every eligible mapped Person, Project, Organisation and installed Skill. It does not stop at a REM page count or weekly allowance percentage. The CLI shows source totals, an estimate before model work, and a running total for the page queue and each category. Explicit --first-* options still cap a deliberate test run. Model provider limits and account access still apply.

The selected model gets a source-free access check before page dispatch. A refusal is recorded, the map and saved mail remain available, and init exits without installing nightly upkeep. An access denial during the queue stops new jobs while pages already in flight finish. Ordinary individual page failures remain visible and can be retried.

On macOS, co rem start installs a root-specific launchd tick. Status, doctor and the local reader now surface a loaded-job failure or a scheduled warning. If a coding-session transcript rewrites an already consumed prefix, that file keeps its checkpoint and is skipped while unaffected sessions continue. A reviewed reindex path for the skipped file remains 1.9.1 work.

The reader makes phone People facts readable, names source actions explicitly, and puts the complete finding expansion near the lead. A mapped Skill no longer claims that a source mention proves its version or outcome. The role-based founder/UI review records the inspected pages, changes, remaining issues and recheck criteria. The visual evidence uses invented contacts: Today desktop, Person phone, before and after, Project desktop, Skill phone, and first-run CLI progress. The screenshot manifest records their origins.

Real trial and verification

The public 1.9.0a43 wheel was exercised on a private 90-day owner notebook with Gmail, Outlook and coding-session sources. It mapped 3,362 mail messages (1,298 Gmail, 2,064 Outlook), 70 People entries (including the owner and two held for identity review), 18 Projects, 18 Organisations and 154 installed Skills. The first-run queue selected 67 eligible People, 17 Projects, 18 Organisations and 154 Skills, plus the owner page. The 258 page runs (including the owner's quick and full passes) all completed with zero failed, refused or skipped pages in 52m 57s wall time. Ten workers were active through the long Skill tail; no provider throttling or concurrency failure appeared. A sample of four written pages in each category resolved 161 of 161 numbered source IDs to retained local excerpts. That checks source availability, not every claim's meaning. Counts are aggregate; private names and source text are not published.

An installed 1.9.0 candidate wheel was then run against the existing default 90-day notebook. It re-listed the 3,362 mail messages, reused 3,278 saved bodies, and showed category progress for 17 pending pages with ten workers. Seven People, six Skills and three Projects were accepted. The fourth Project was refused by its cited-claim audit because two citations were missing; an explicit --retry-refused attempt also failed its audit, so the prior page was kept. Init exited nonzero and left nightly upkeep off. The CLI now points to that exact retry command; the bounded automatic-retry improvement is tracked in #2310. The owner explicitly enabled upkeep with co rem start --yes after seeing the partial result. Its first sync completed a 40-item batch and launchd loaded the root-specific job. A controlled scheduled pass processed another 40 items, then the same Project's cited-claim audit refused its daily investigation. The command exited nonzero. This exposed a status bug: the worker had saved the successful source-sync result instead of the failed end-to-end result. The 1.9.0 fix records the entire scheduled pass and marks background upkeep as needing attention. The page refusal and bounded retry remain #2310 for 1.9.1; the prior Project page remains intact.

The REM unit suite passed 1,563 tests including the scheduled-status and retry-next regressions. The local Chrome reader suite passed 14 scenarios. A clean Python 3.13 full repository run passed 14,610 tests with 101 skips; 13 installed-wheel tests passed. The docs site built and its blog checks passed. Package artifact and publication checks remain release gates.

Limits and 1.9.1

The 90-day default covers eligible mapped people, not every historical contact. People held for identity review require owner confirmation. Historical scan checkpoint recovery is tracked in #2298 and broader historical People coverage in #2176. Mobile reading-space and partial Project-state refinements are in #2309. A cited-claim audit may still refuse an individual page; its retry is explicit and the saved page remains intact. These limits are visible; the release does not claim all-history or all-page quality acceptance.