ConnectOnion 1.9.3rc4 (release candidate)

Install this preview

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

python -m pip install --upgrade 'connectonion==1.9.3rc4'

Enrich and muse, fixed

1.9.3rc3 added co rem enrich PAGE and co rem muse PAGE. Running them on more pages showed the runs were being cut short and the ledger was too thin. This candidate fixes both; the commands and their output are unchanged.

• A sub-agent's turn no longer ends the main turn. When co ai delegates to Codex, Codex reports its sub-agents' threads on the same connection. A sub-agent finishing was taken as the main turn finishing: it cut 4 of 7 muse turns off mid-sentence and the enrich turn before its verdict, and the sub-agent's reviewer notes became the run's reply. Notifications from other threads are now ignored.
• The reading ledger holds more, and the gate accepts it. For a project page, enrich's ledger now also lists the project's mapped coding sessions, its GitHub issues and pull requests with their comments, and the git log (as a reading guide, not a citable source). The ledger is kept under .state/enrich/ and muse reads it too. Before, the model found pull requests and mail on its own, cited them, and the review gate refused those citations as unidentifiable; now a source the ledger lists can be cited and audited.
• Coding sessions are read within the 730-day window that investigate already uses, instead of every session on disk. On a machine with thousands of sessions, enrich's gathering went from 29 minutes to under one.
• Token totals for resumed Codex threads no longer double-count. A resumed Codex turn reports the thread's running total, which muse was adding up turn by turn.

Verification and limits

• New unit tests cover ignoring another thread's turn completion, usage on a resumed thread, the session window, and the ledger's sessions, issues, pull requests and git log being listed and accepted by the gate.
• One end-to-end run on a copy of a test notebook, on a project page, enrich then muse with Codex:
  • quotes 3 to 15 (rc3 on the same page: 3 to 6), cited sources 10 to 40 (rc3: 10 to 21);
  • no quote lost, none that is not word for word in its source;
  • muse's turns: the opening questions, three insights, one "no insight", and a final reply in the format the skill asks for.

Known and open:

• Muse is still expensive: seven turns re-send a growing context.
• No schedule runs enrich or muse yet; they run when you ask.
• Everything listed as open in the 1.9.3rc3 notes still applies, apart from the slow session gathering fixed above.