ConnectOnion 1.9.2b3 (beta)

Install this preview

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

python -m pip install --upgrade 'connectonion==1.9.2b3'

1.9.2b2 was built and tested but not published. Everything in the 1.9.2b2 notes is part of this beta.

A first run you can read straight away

Measured on the same real mailbox (180 days of Gmail and Outlook, about 295 pages, Codex gpt-6-luna, 16 workers):

	1.9.2b2 build	1.9.2b3 build
Wall clock	126 min	122 min
API-equivalent cost	$9.54	$8.42
Pages not written	8	3
Lost to the notebook lock	0	0

Of the 3 pages not written, 2 left out a required section. A missing section is now added as Unknown instead of the page being refused. The third cited files outside its own material and was rightly refused.

An audit of every page, against the 1.9.2b1 notebook from the same mailbox:

	1.9.2b1	1.9.2b3
Pages that pass every check	213 of 296	217 of 295
Sources numbered with gaps	40 of 123	0 of 284
Skill pages with a broken run-evidence block	38	1
People with Company left Unknown	53	31
Pages that refer to you only by name	58	1
Private file paths cited as sources	2	0

• Sources read 1, 2, 3. When a page is written its sources are renumbered in order, and every citation in the text follows.
• Company comes from the organisation page. When a person's Company is Unknown and their mail domain has an organisation page, the page links it.
• Probable duplicates are named. At the end of the first run, init lists pages that are probably one person (a name and a handle, two spellings) with the command that merges them.
• People are contacts you can name. co rem show "Ody Zhou" finds a person by name, alias or email. If several pages match, it lists them with their emails and shows none, so a command never reaches the wrong address. co rem --json list people --table gives every contact as data.
• Also known as holds names. It no longer repeats the person's email addresses, which 64 pages did.
• One voice. Pages refer to you as "the user", which the reader shows as "you", instead of by name. The opening line says who the person is before what happened lately.
• Skill pages count co ai runs. The usage header and the run list read from the same records, and a skill page keeps exactly one run-evidence block.
• Quoted replies stay out of the page. A two-line Outlook or Apple Mail From: / Sent: header is cut with the quote below it.
• An assistant's memory files are not sources. Paths under ~/.claude/projects or ~/.codex are rejected as citations.
• A second pass with nothing older to read is reported as "nothing older" and is not retried.
• Project pages: a line whose citation is outside the material is dropped, not the whole page; a mapped project that waits for a run says how many sessions it will read.
• The estimate before the first run is calibrated from the measured run.

Reading the page

• Facts labels never break mid-word. REPOSITORY and ORGANISATION used to wrap as REPOSITOR / Y.
• Long addresses, handles and paths break after @ and before . or /, not inside a word.
• A 64-character source id shows its first 12 characters with an ellipsis; the full id is in the tooltip and in the Markdown.
• A connected-context card ends on a whole word with ….
• On a phone, a project's source action stays on the first screen.

Verification and limits

• Tests:
  • The co rem unit suite passed (2,499 tests).
  • Each fix has a regression test that failed before the fix.
  • The reader browser suite passed (65 tests).
• The measured run used this beta's code before the last two fixes: a missing section added as Unknown, and a second pass with nothing older counted as nothing new. Both are covered by tests, not yet by a real run.
• Pages read by hand after the run were mostly detailed and cited. What remains is tracked in issue 2349: a thin page for a large project, four organisations still titled by their domain, the owner's own page opening on the latest task instead of who they are, and minor items listed as open threads.