ConnectOnion 1.9.2b11 (beta)

Install this preview

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

python -m pip install --upgrade 'connectonion==1.9.2b11'

A handoff that keeps what you said

A co handoff is now one message, built the way a compaction is: every message you typed in the session, word for word, the AI's replies in summary, and where the code is. Tool output and the raw session stay on your machine; if the other side needs more, they ask, and the question comes back to you.

• The whole session, across every compaction. Codex and Claude Code keep the full history on disk, so the handoff reads all of it instead of the last forty turns and a clipped summary. Messages typed while the agent was mid-turn are kept, and so is a Codex /goal, which an earlier filter threw away as injected text.
• The summary on top keeps your words. The task's first line is what you said you were handing off. A review stays a review instead of being turned into a fix, later decisions override earlier ones, and permissions you gave your own AI are not passed on. Decided and Rejected lists are back.
• Where the code is. Repository, branch and the commit the session was on at the time, whether it is pushed, and a note when it is not. Issues and PRs the session mentions arrive with their titles, not bare numbers.
• Private details are flagged. The preview lists home paths, email addresses, phone numbers and IP addresses. --since N starts at message N when a session covered other work first, and a message over 30,000 characters says so.
• Wait in the background. co handoff status <id> --wait returns when the other side accepts, asks or answers. The paste-in block uses it, so the install floor is now 1.9.2b11.
• Faster on long sessions. The AI's replies are summarised in parallel pieces: a session that took 241 seconds now takes 110.

Also in this preview

• Live REM starts its own Host (#2308): co rem open --live reuses a running co ai, or starts one, and waits until it is ready, instead of telling you to run co ai yourself.

Tested on real sessions

Fourteen real compacted sessions were handed off with 1.9.2b10 and with this release, and 131 facts in them were checked one by one. 1.9.2b10 kept 60 percent of them; 1.9.2b11 kept 130. In a second test, sessions were cut back to before the work was done and handed to a fresh Codex on another machine. One release-CI failure was fixed correctly from the handoff alone in 144 seconds (662 with 1.9.2b10). A second task reached the right issue but only part of the real fix: the missing requirement had appeared only in the original AI's tool calls, which a handoff does not send.

A handoff from a Mac to Codex on a Linux machine ran end to end over real agent mail: accept, a question, an answer, and --wait returning on each.

Verification and limits

• The handoff suites pass (117 tests), with new tests that failed first on 1.9.2b10.
• Every message you typed is sent. If a session covered something unrelated, remove it with --edit or --since; the scan does not recognise things like a client's name.
• Issue numbers are matched by number, so #12 in two repositories can get the wrong title.