# 1.9.0a4 — what the real notebook found

An opt-in preview of 1.9.0. Stable is 1.8.9. Everything here was found by
running 1.9.0a2 and 1.9.0a3 on the owner's own notebook (#1974) and on a copy
of it (#1981–#1990), and fixed with a test that failed first.

## Pages you can trust

- **No "investigated" without material.** A page whose model saw nothing about
  its subject was stamped investigated anyway (one cost 919k tokens and stayed
  empty). Now the run stops before any model call, says why, and the page stays
  as it was; pages stamped that way go back in the queue. A page investigated
  before whose window has nothing new calls no model either (#1974, #1984).
- **One bad citation drops one line.** A page was thrown away after 199
  seconds of work for one miscopied id. Now only the line resting on it goes,
  and the review records how many (#1974).
- **Services are not people.** Apple Developer, Airbnb, Google Cloud and the
  like get no people page; an address only the owner writes to is held for
  review (#1987).

## Pages worth reading

- **A person page opens on who they are to you**, what is open between you and
  who owes it, and the last contact date, before the contact fields. What was
  or wasn't searched stays off the page (#1974).
- **A project page from your messages** says what the product is, not the
  loudest thread, keeps "Where it stands" to a few current bullets, hedges once,
  and uses one language (#1974). A main checkout weeks behind your newest session
  is named as stale, with the version from the current line (#1982).
- **A skill is one catalog page**, however many copies are installed, linking
  its source instead of pasting it, with how often and when you last used it:
  on the owner's notebook 419 pages became 153 (#1974).
- **A repository is one project page**: worktrees fold into it, and one-session
  chat folders and caches are not projects (#1974).

## Reading what you actually wrote

- **Codex Desktop.** Desktop marks every message with a key co rem took to mean
  "injected", so it dropped 94% of what the owner typed there. Reading the
  message's own kinds instead, the owner's material grew from 1,020 to 1,682
  messages, and workspace messages placed in a repository from 258 to 830
  (#1978).
- **One people queue** for `investigate` and `investigate people`: the owner's
  page first while it is unwritten, then people you write to, then people who
  write often, then one-mail contacts; counts are named as the map's (#1974).
- **Org pages work again**: Outlook is asked for the domain's first label, and
  one mailbox failing is a gap, not the end of the page (#1981). Extracted
  project notes reach the page they name, and a written page starts the next
  window (#1983, #1985).

## Running it

- `co rem status` and `doctor` say whether each mailbox is actually read by the
  daily round, and `co rem start` offers the connected ones for your approval.
- Task folders are owner-only and scrubbed, even after Ctrl-C.
- The model's `co ai` gets an absolute PYTHONPATH, so a checkout run with
  `PYTHONPATH=.` no longer imports an older installed connectonion.

## Install

```sh
python -m pip install --upgrade 'connectonion==1.9.0a4'
co rem status
```

## Known limits

- Not yet fixed from the same run: an investigation page-size rule (#1988),
  eight CLI text items (#1989), and the budget meter (#1990).
- The benchmark lets the agent under test read the repository history, and rem
  suites test the core skill rather than the composed instructions (#1992).