What's new in 0.12.2
Bazilion 0.12 makes long-lived Agent teams easier to supervise and safer to
leave running. In that release, the three public packages
move together: bazilion@0.12.2, @bazilion/client@0.12.2, and
@bazilion/api-types@0.12.2.
One Attention Center for operator work
Section titled “One Attention Center for operator work”The web navigation now shows one source-owned queue for items that need human attention:
- pending communication approvals,
- reviewed-learning proposals,
- terminal review and trigger failures,
- Agent message loop breaks.
Filter the queue, open the canonical decision screen, and acknowledge or restore informational failures without copying sensitive payloads into a second audit store. The same state is available through authenticated HTTP and CLI surfaces.
Reviewed learning for long-lived Agents
Section titled “Reviewed learning for long-lived Agents”Successful user turns can enqueue restricted background reviews. A review produces an evidence-backed proposal rather than changing prompts immediately. Operators can inspect and edit the proposal, approve or reject it, and later revoke an approved lesson.
Private lessons enter only that Agent’s prompt. Shared lessons become deterministic notes in the Team’s qmd-indexed memory. Review cadence, model, and reasoning are configurable from the Agent Learning surface.
Agent message-loop protection
Section titled “Agent message-loop protection”Agent messages retain causal-chain and hop metadata across inbox wake turns.
The daemon rejects over-budget sends before another LLM turn can start, even if
an Agent omits reply_to. Configure the ceiling with
BAZILION_AGENT_LOOP_MAX_HOPS; inspect payload-free stop events through the
Agent API, bazilion inbox loop-breaks, or the web inbox.
Pi 0.83 and the current model catalog
Section titled “Pi 0.83 and the current model catalog”Bazilion 0.12.2 updates the bundled Pi packages to 0.83.0. Direct provider chat
and full Agent sessions now share one public ModelRuntime boundary for model
resolution, provider aliases, authentication, local endpoints, and arbitrary
model IDs. Sessions use Pi’s public DefaultResourceLoader; runtime and tests no
longer depend on Pi’s former compatibility entry points.
The provider catalog adds Qwen Token Plan international and China endpoints and refreshes curated examples for current model families including GPT-5.6, Claude 5, Gemini 3.6, Kimi K3, Grok 4.5, and Qwen 3.8. ChatGPT OAuth credentials remain encrypted and daemon-owned while the runtime preserves refresh state.
See the GitHub release for the published artifacts and package-level changelog.