ConnectOnion 1.9.2b7 (beta)

Install this preview

Stable remains 1.9.1. Install this opt-in beta by exact pin:

python -m pip install --upgrade 'connectonion==1.9.2b7'

A first run that does not trip over itself

These fixes come from a real co rem init of 1.9.2b3 installed from PyPI: 180 days of Gmail and Outlook, 312 pages, Codex with 16 workers. All 312 pages were written. Its run records and all 637 Codex turns were then read back one by one.

• Another co command no longer fails the run. 14 investigations failed within the same two minutes with "The selected credential record changed". A co command in another terminal had refreshed the Microsoft sign-in, and Microsoft issues a new refresh token every time. Each process remembered only the rotations it made itself. Rotations are now recorded next to the env file as token digests (the tokens themselves are never stored), so every co process recognises them. A sign-in for a different account written over the record is still refused.
• A new notebook runs on Codex. With both coding agents signed in, a fresh install picked the other runner. It then announced about 460M input tokens on a plan the user had not chosen. Codex with gpt-6-luna is now the default. The other runner is used only when Codex is not installed or not signed in. A notebook keeps the runner saved in its config.yaml.
• Turns stop looking for a file the package does not ship. Each REM Skill named the document explaining its rules. Codex loads a Skill exactly as written, so 127 of 637 turns went looking for that document. The documents now name their Skill, and the Skills no longer name the documents.
• A run record names its Codex threads. Every turn of an investigation, including each round, the search follow-up and the editing turn, is listed under threads. A page now leads straight to the transcripts that wrote it. Before this, about two thirds of runs could not be matched to their traces by timestamps alone.

Checking a run after it finishes

scripts/rem_trial_audit.py NOTEBOOK reads what a run left behind and exits 1 when it finds something severe:

• a page citing REM's own session as the user's words;
• a refused page missing from the refusal ledger;
• a background job macOS never started.

It also reports the following:

• where investigation turns read outside the material they were given;
• MCP, web and sub-agent calls;
• refused Outlook searches and missing files;
• page problems: citations with no source, gaps in source numbering, more than five Uncertainties, Also known as holding only addresses, and organisations titled by their domain.

Reading outside the material is measured, not treated as a failure. Investigations keep full disk access by choice. --strict-reads makes it severe.

On the 1.9.2b3 notebook, the script found what had been found by reading the logs by hand:

• the owner page citing REM's own model check;
• 3 refusals missing from the ledger;
• 44 turns with a refused Outlook search.

Verification and limits

• Every fix has a regression test that fails before the fix. The co rem unit suite passed.
• The real run that found these used 1.9.2b3. These fixes have not had their own real run yet.
• The run took 259 minutes on a machine that was busy with other work throughout, against 122 minutes on a quiet one the day before. Duration here is not a measure of this release.
• Open, in issue 2349:
  • The first page of your own takes about 15 minutes, not the 4 the help promises.
  • Some Open threads are months old with no sign they still matter.
  • Two project pages cite a source they do not list.