# ConnectOnion 1.9.2b4 (beta)

## Install this preview

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

```bash
python -m pip install --upgrade 'connectonion==1.9.2b4'
```

## Hand work to someone else

The new `handoff` skill is loaded in every `co ai` session. Say "hand this to
Bob" and your agent:

1. finds Bob with `co rem show "Bob"`, and asks you to choose if more than one
   page matches;
2. writes a brief from the session: the task, where it stands, what was
   decided, what was rejected and why, open questions, and the code by branch,
   commit or PR;
3. removes keys, tokens, private paths and other people's data;
4. shows you the exact text and sends it only after you approve it;
5. sends it by email, or straight to Bob's agent when you have his full agent
   address.

An emailed handoff carries your agent's address and asks Bob to reply with
his. It never carries an invite code. An invite code admits whoever holds it
as a contact, and contacts may run commands on your machine; an email can be
forwarded. Receiving works the same way in reverse: the brief is saved, you
are told what arrived, and nothing runs until you accept.

`co copy handoff` copies the skill into a project.

## Found by running 1.9.2b3 on a real machine

- **REM no longer reads its own work back as yours.** Codex records REM's own
  task sessions under the same client name as `co ai`. A real owner page cited
  REM's model check ("Reply READY only…") as the user's request, and the next
  sync would have read 583 more. Sessions run in a notebook's `.state/tasks`
  are now skipped, for Codex and for Claude Code.
- **Outlook searches with a quoted phrase work.** 67 of 584 investigation
  turns had a follow-up search refused with HTTP 400, all of them with a
  phrase in quotes. The phrase is now escaped. Checked against live Microsoft
  Graph: a phrase that was refused now returns its mail.
- **`co rem status` says when macOS never started the background job.** A job
  was loaded for four days with launchd reporting zero runs, while status said
  "the last scheduled investigation failed; run co rem logs". It now says the
  job has not started since it was installed, where to look, and that
  `co rem sync` works by hand.
- **`co rem config set model …` no longer crashes** after saving the setting.
  The model check investigates a sample page. When the model asked for a mail
  search, the sample mailbox could not answer.
- **Intel Macs install without compiling.** cryptography 49+ ships macOS wheels
  for Apple silicon only, and pip sat compiling 50.0.2 for over thirty minutes.
  Intel Macs now get 48.x, which has a universal wheel. A fresh install took 25
  seconds.

## Verification and limits

- Every fix has a regression test that fails before the fix. The Outlook
  search and the Intel Mac install were also checked live.
- **Known and not fixed in this beta:**
  - Investigation turns still run with full file access, and on a real run
    they read files outside their material: other notebooks, raw session
    transcripts, notes.
  - Agents search for a skill document at a path that does not exist.
  - A run record does not name the Codex thread that ran it.

  All three are next.