# ConnectOnion 1.9.3rc3 (release candidate)

## Install this preview

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

```bash
python -m pip install --upgrade 'connectonion==1.9.3rc3'
```

## Deepen one page: enrich and muse

`co rem investigate` writes a page from what its gather found. Two new commands
go further on one page that is already written:

```bash
co rem enrich projects/aurora.md   # follow the leads the page has not followed
co rem muse projects/aurora.md     # think the page through; write only on insight
```

- **`co rem enrich PAGE` is discovery.** It reads the page, its cited
  originals, and a reading ledger: every source the saved mail archive and
  your coding sessions hold about the subject, laid out as searchable files,
  with the ones the page does not cite yet listed first. The model follows
  links (it may open them on the web), unread attachments, names, the start
  or end of a thread, older periods and other channels, adds what they
  establish with citations, and ends with a lead list: `followed`, `open`
  (the next run's work, specific enough to act on) and `blocked`. No live
  mailbox is queried; the saved archive is read.
- **`co rem muse PAGE` is thinking.** One conversation of seven turns in the
  same Codex or Claude Code thread: turn 1 asks at least twenty specific
  questions of the page and its originals and picks the five worth pursuing;
  turns 2 to 6 study one each and edit the page only when the answer is an
  insight that needed two or more sources; the last turn looks back, brings
  the lead and the page's shape in line with what changed, and lists the
  questions left open. It is given the latest enrich lead list and muse open
  questions for the page, so the next run starts from them.
- **The same review gate as investigate.** Every citation must trace to a
  source, every quote must be word for word in the source it cites, a contact
  must be the subject's own, and the cited claims must pass the audit; what
  review names goes back to the writer. Muse adds one check: if the page lost
  any `“…”` quote or any `Sources` id it had at the start of the run, it is
  put back as it was and the result says so.
- Both work on people, projects and organisation pages, change no other page,
  print whether the page changed and its size before and after, and record a
  run in `co rem logs` with its usage, counted in the weekly budget like
  investigate.

`co ai --json --resume <session>` now resumes a delegated Codex or Claude Code
thread as well as ConnectOnion's own loop; that is how muse keeps its seven
turns in one conversation.

## Verification and limits

- The full unit suite and the REM help contract pass on this branch, apart
  from five tests that also fail on 1.9.3rc2 on the release machine (an
  optional WhatsApp library not installed, and a local docs-site checkout
  behind). New unit tests cover the command wiring, the seven-turn sequence
  in one resumed thread, the review happening once after the last turn, and
  the page being put back when a quote or source id is lost.
- One end-to-end run of each command on a copy of a test notebook, on a
  project page, with Codex `gpt-6.1-sol`:
  - enrich: 5,164 to 9,032 characters, quotes 3 to 6, cited sources 10 to 20,
    none lost; 2.06M input tokens (1.73M cached), 15k output; 31 minutes, of
    which most was gathering about 4,800 local coding sessions.
  - muse, on the enriched page: 9,032 to 8,514 characters (shorter, as the
    last turn removes what the insights superseded), quotes 6 kept, cited
    sources 20 to 21, none lost; all seven turns in one Codex thread; 16.1M
    input tokens (15.1M cached), 67k output; 25 minutes.

Known and open:

- **Muse is expensive**: seven turns re-send a growing context, so a run costs
  several times an investigation of the same page, most of it cached input.
- **Gathering sessions is slow** on a machine with thousands of coding
  sessions: enrich reads all of them to build the ledger.
- **What is printed is everything the model said in the run**, ending with
  the reply the Skill asks for (the lead list, or the insights and open
  questions); a delegate's progress notes come before it.
- No schedule runs enrich or muse yet; they run when you ask.
- Everything listed as open in the [1.9.3rc2 notes](1.9.3rc2.md) still
  applies.