ConnectOnion 1.9.3b5 (beta)

Install this preview

Stable remains 1.9.2. Install this opt-in beta by exact pin:

python -m pip install --upgrade 'connectonion==1.9.3b5'

REM pages you can trust on the first run

This preview is about co rem's first run on a real notebook. The 1.9.2b3 notebook (2026-10-10 to 10-15, 180 days of mail and sessions) and two later trials were audited claim by claim. The problems were wrong facts that looked right, and good pages lost because of one small fault. Both are fixed here.

What a page says is checked against where it came from

• Quotes are checked word for word. Pages now quote the words people may act on (an ask, a term, a decision) instead of paraphrasing them. Every “…” on a person, organisation or project page must appear in a source its line cites. On the 1.9.2b3 notebook 8 of 89 quotes were not in their source: a fee quoted as gross where the owner had written net, and Chinese turned into English.
• Every page about people and work is audited. The cited-claim audit used to run only on project investigations. It now runs on every write of a person, organisation or project page, reads the material that turn read, and names each claim it could not find.
• Contact facts are the person's own, or Unknown (#2348). A phone number stays only when a mail the person sent shows them giving it. Before, one page carried a facilities manager's mobile from an estate notice the person had forwarded, and another carried someone else's mobile from an event host's message. A relay address (Luma, lu.ma, Airbnb's message relay) is never anyone's email. Company comes only from the newest domain someone writes from, never a mailbox provider, relay, agent or student address. Last contact is the last human mail, not a notice or a mail to more than five people.
• Sources are dated by the runner. Each source is listed once and dated from its original. The audit had found one mail cited as [1] and [2] with two different dates. Every page writer and the audit are now told today's date: one page kept the 13th's and 14th's events as open threads on the 15th.
• Only what you typed is cited as you. Session prompts a program sent (a scheduled job's prompt, REM's own model-access check) are no longer quoted as your words, and old citations of them are tidied once.
• Your own company gets no organisation page. Domains that are yours, or named like your Company, are not queued as a counterparty.

Good pages are repaired, not refused

• Two repair turns before a refusal. Every page type now gets two repair turns that name what review refused. What is still refuted after that is removed with its sentence (a contact field says Unknown), and the rest of the page is saved. review.json says how many claims went. On a 1.9.3 trial run, one unmatched claim had cost 10 of 28 pages and 29M input tokens; a refuted claim that no line holds now refuses nothing.
• Code settles what code can. Glued or repeated sections are merged, a citation to mail REM already holds is resolved from the archive, inline source ids are numbered before Facts are checked, and a leftover "not investigated yet" is removed. Before, one person's page was refused for two real mails the turn had not been handed, and seven true Facts fields were lost on another.
• Maintenance sees the rest of the thread. Upkeep and the audit get the later messages in a batch's threads, so a question answered the next day is not left open.
• One refusal ledger, retried with a cap. Every path that can refuse a page writes .state/refused-investigations.json. Before, one of eight refused pages reached it. Later runs retry ledger pages once every 20 hours; a page refused three times on the same material waits for new material. Naming a page (co rem investigate "Name") retries its refused material.
• No shrinking rewrites. A maintenance turn that keeps less than 60% of a page's size or cited sources keeps the old page. One project page had been cut from 14,356 to 4,132 characters.
• Fewer wasted turns. Each sync starts at the source after the one it read last (in 26 syncs Claude Code, Gmail and Outlook were never read). Project pages with folders are routed by folder only and written by the project writer at most once in 20 hours unless 10 messages wait (one page had 12 turns in 26 syncs, 9 of them NO CHANGE, 5.36M tokens).

Read a page in your own language

• co rem config set language zh writes new pages in your language. Headings, Facts labels, source ids and quotes stay as they are.
• Your languages are detected from your own sent mail and typed session messages, one vote per message. The owner's notebook came out 59% Chinese and 41% English. co rem config set languages zh,en overrides it.
• co rem open starts the live reader on this computer through co ai and opens it at 127.0.0.1. Under a page's title, Read in offers your languages first, then the person's own, then the rest. A translation is one co ai turn, cached until the page changes. co rem translate "Name" --to zh does the same from the terminal. co rem open --local is the offline snapshot, and it falls back to it, saying why, when co ai cannot start.

Activity instead of Run logs and AI notes

The reader's sidebar listed "AI notes" and "Run logs" as memory categories; Run logs were skill reports and maps, while the real run history sat at the bottom of Today. Both are now one Activity view: each run's time, kind, outcome, items read, tokens, the pages it changed and the pages it refused, with the reason.

The terminal

• co rem status leads with the open threads waiting on you, counted the way the reader counts them, and "To write next" counts only pages its command would write.
• One census: investigate and status no longer disagree on how many pages are written.
• co rem list people --table shows names instead of Markdown links, fits the terminal, and never shows a relay as an email.
• "Organisations" everywhere, as the reader spells it.

Also in this preview

• Claude Code projects found after session preambles (#2347, fixes #2336, thanks to Jasper, @jipeng6036-del): Claude Code can begin a session with queue or mode records that carry no working directory. Project discovery stopped at the first record and missed the project; it now reads ahead to the first working directory, bounded to 256 records or 1 MB.
• A cited project file that is not UTF-8 (cp1252, for example) is read instead of crashing the investigation.
• An unexpected investigation failure keeps its traceback in the run record, for co rem logs.
• The scheduled job imports the checkout it runs, never the shell's PYTHONPATH.
• The handoff prompt pins 1.9.3b5. Handoffs already sent keep their version.

Verification and limits

• Unit tests for REM, handoff, release and CLI discovery, and the REM CLI end-to-end tests, pass on this branch.
• Screenshots use the invented notebook from the reader's tests, rendered by 1.9.3b4 and by this release in headless Chrome.
• Removal is only as good as the audit: a wrong claim the audit does not flag stays. A quote whose source is not held is left to the audit.
• Some true contact facts now read Unknown, such as a number given only in a quoted reply. Add one with co rem stub.
• Translation needs co ai and a working runner (Codex or Claude Code). Only the live reader can make a new one; the snapshot and O Chat show cached ones.