1.8.9b17 — Gemini 3.8 is the default again

An opt-in preview on the 1.8.9 fix line (#1722). Stable is 1.8.8.

The default model

1.8.9b10 made the free co/llama the default for every Agent(), llm_do(), co ai, built-in sub-agent and project template that names no model. co/llama is Llama 3.1 8B, with a 4,096-token context and one shared GPU slot. By the owner's decision the default is co/gemini-3.8-flash again, as it is in stable 1.8.8 (#1869).

An agent that names its model is unchanged. One that relied on the default moves from Llama back to Gemini 3.8, and managed Gemini calls are billed to your credits again.

Free models are the zero-balance tip

co/gemma and co/llama stay free and selectable. When your balance reaches zero, co status now names them beside the purchase page:

⚠️  No credits left. Add credits at https://o.openonion.ai/purchase
   Or keep going for free:
   • model="co/gemma"            Google's free Gemma on ConnectOnion's GPU (no credits needed)
   • model="ollama/<model>"      a model on your own machine, e.g. ollama/llama3.2

co status at a zero balance naming co/gemma and ollama beside the purchase page

A test now fails the build if any current doc calls co/llama the default.

Wiki: a re-map fixes the pages it already made

Re-running co wiki init now retitles an uninvestigated page that an older map named with an address, using the name found now. Investigated pages keep their title. On a copy of the owner's notebook, address-titled pages went from 218 to 42 and organisation pages from 272 to 138 (134 archived, none deleted), and no investigated page changed (#1880).
The Gmail listing fetches each message once, with retries on rate limits and server errors. A 90-day init took 303 s in total, where the Gmail scan alone used to take about 10 minutes, and one refused call can no longer fail the whole scan.
A greeting sent to an agent's 0x…@ address no longer names the agent after its owner.

Fixed

A background listener loaded whatever connectonion/ sat in the directory it was started from. In b16's own end-to-end run, co whatsapp listen --restart from inside a checkout started a listener on the checkout's code, and b16's new self-restart could have re-entered it every minute. The listener now starts in its inbox directory, and restarts at most once per installed version (#1878).

Install

python -m pip install --upgrade 'connectonion==1.8.9b17'
co --version

View Markdown source