# ConnectOnion 1.9.3rc9 (release candidate)

## Install this preview

Stable remains 1.9.2. Install this release candidate by exact pin:

```bash
python -m pip install --upgrade 'connectonion==1.9.3rc9'
```

## Pages that quote people

- **The words, not a retelling.** The quote rule now lives in the page skills,
  which every stage that writes a page composes, and extraction notes carry
  `Quote:` lines, so daily upkeep has words to keep. A person page that cites
  more than three mails and quotes fewer than six (ten past five mails), or
  has fewer than two Insight bullets, gets one more turn with only the page
  and its originals; if that turn loses a quote, it is undone.
- **Dates in code.** First contact is the earliest direct mail read, marked as
  such. History is sorted newest first by code, not by the model.
- **Pronouns.** A he, his, she or her whose nearest named actor is the owner
  becomes "the user"; a pronoun after another name or inside a quote is left
  alone.
- **Measured** on notebook copies, writing person pages from scratch: mean
  quotes per person page from 6.8 to 22.4.

## Deeper pages

- `co rem enrich` reads a long project history one period per turn, each
  about 15k characters of timeline, six at most, in one resumed thread. The
  repository's issues and pull requests are all listed, and history is found
  in the owner's other repositories too.
- Enrich and muse are judged on the lines they wrote; findings on carried lines
  are recorded as `carried_findings` instead of refusing the run. Repairs
  resume the thread that wrote the page, and a lost quote is sent back by name
  or put back as it was.
- Each scheduled deepen slot takes the most valuable thin pages one round deeper,
  uses the owner's own words per history period, and rests a
  page whose rounds stop adding.

## Faster first runs

- One page-concurrency setting, `limits.workers`, for init and investigate
  CATEGORY.
- Sessions are scanned once before page workers start, and only notes that name
  the subject are excerpted. At 16 workers, gathering is about twice as fast.
- Overlapping investigations count the Codex meter's movement once.

## The reader

Quotes stand out from the prose around them; citations are a small date after
the word they support; the owner's language reaches the reader, so a Chinese
notebook gets a Chinese interface; people cards; and a "last night" entrance.

## Handoff

A handoff pasted into a conversation that was already about something else now
stops and points to a new conversation, or to the sender's thread, instead of
doing the work there. Without a repository, the work goes into
`handoff-<id>/`. The recipient's install floor is 1.9.3rc9.

## Verification and limits

- Every REM pull request in this candidate merged with all CI checks green and
  was measured on notebook copies. Over repeated deepening rounds, no quote was
  lost.
- A real first `co rem init` (16 subjects and the owner, 16 workers) on main
  at 52b7e3aa wrote 17 of 17 pages with none failed, in 109 minutes. The
  later REM merges in this candidate (#2563, #2567, #2568) came after that run.
- The session cache is upgraded from v1 to v2 in place on the first run after
  installing.

Known in this candidate, fixes planned for rc10:

- A transient Gmail SSL error while mapping stops `co rem init`; running it
  again resumes.
- The first-run time and token estimate is about half of what it takes: 42
  minutes and 50M tokens predicted, 82 minutes and 102M measured.
- Organisation pages are the slow tail of a first run.
- Everything listed as open in the [1.9.3rc8 notes](1.9.3rc8.md) still applies.