Daily operations

Diagnose an Agent that is not answering

bazilion doctor
bazilion agent list
bazilion attention list

1. bazilion doctor checks the home, the database, skills, providers, web search and the Docker image. Make sure the CLI talks to the daemon you mean (BAZILION_SERVER, BAZILION_HOME).

2. Check that the Agent’s provider is enabled, has a saved model and a working credential: bazilion provider test <provider:model>.

3. Open the Agent’s chat and read the error or the failed tool result. For Telegram, also run bazilion telegram health.

4. If a message to another Agent failed, check the Team’s policy: a denied message is an error returned to the sender.

5. If a turn is stuck, cancel it:

   bazilion agent cancel <agent>

Cancelling does not undo files the Agent wrote or messages it sent. Look at the workspace and the conversation before asking again.

Schedules

bazilion trigger add <agent> --every 300 --message 'Summarise project status' --disabled
bazilion trigger list <agent>
bazilion trigger enable <trigger-id>
bazilion trigger history <trigger-id>
bazilion trigger disable <trigger-id>

On the web, use the Agent’s Triggers tab. Each occurrence runs at most once and is never retried: a busy Agent leaves it pending, then it ends succeeded, failed or cancelled. The history shows each run with its real error, and a failure also shows in Attention. Disabling a trigger stops future runs, not a turn that is already running.

Attention

Attention lists what needs you: lesson proposals, failed or cancelled learning reviews, failed triggers, stopped Agent message loops and Agent queues paused after a daemon restart. Open the Attention page (the badge in the top bar shows the count) or:

bazilion attention summary
bazilion attention list
bazilion attention acknowledge <key> --yes
bazilion attention acknowledge-all --yes

Acknowledging a failure or loop break records that you saw it; it repairs nothing. Proposals and paused queues leave the list when you act on them. Attention items can also go to Telegram; see Attention notifications.

Reviewed learning

An Agent can turn its recent conversations into lessons, but only the lessons you approve change its behaviour. Learning is off by default.

1. On the Agent’s Learning tab, switch reviews on and choose how often (every N completed turns you started, default 8), a review model and a reasoning level. From the CLI:

   bazilion agent review-config <agent> --enable --every 8 --model agent --reasoning low
   bazilion agent review <agent>            # run one now

2. A review runs in a restricted worker whose only tool is propose_lesson, and proposes at most five lessons. A review runs once; a failure shows in Attention.

3. Read each proposal, edit its text or scope, then approve or reject it. A private lesson is added to the Agent’s system prompt; a shared one is written to the Team’s memory as lessons/<id>.md.

   bazilion agent lessons <agent> --status pending
   bazilion agent lesson edit <proposal-id> --text '…' --scope private
   bazilion agent lesson approve <proposal-id> --yes
   bazilion agent lesson reject <proposal-id> --yes

4. Revoke an approved lesson later with Revoke or bazilion agent lesson revoke <id> --yes. Earlier conversations are not changed.

Reviews cost model calls; pick a cheap review model for busy Agents. Stable facts about you or the project belong in the Team’s USER.md, not in a lesson.

Around a running Agent

• Conversations: read earlier ones or start a new one.
• Follow-up queue: messages sent while the Agent is busy.
• Agent questions: questions the Agent asks you during a turn.
• Saved results: files the Agent delivered.

Before an upgrade, read Upgrade to 0.26.0 and Backup and recovery.