Daily operations

Start with the affected Agent and the way its turn was initiated. A local chat can work while Telegram or a scheduled turn fails protected-runtime preflight.

Diagnose an Agent that is not responding

bazilion doctor
bazilion agent list
bazilion attention summary
bazilion attention list

1. Confirm the CLI points at the intended daemon/home. Check provider enablement, curated model selection, credentials, and the Agent’s active/archive state.

2. For Telegram, triggers, inbox, or private-gateway chat, inspect protected-work readiness and the tool boundary.

3. Open the Agent’s Chat and inspect the error or tool result. For Telegram, also run bazilion telegram health and bazilion telegram pairing status.

4. Check Operations → Approvals and the Team’s Activity tab. A held attempt or policy denial is different from a provider failure.

5. If a turn is stuck, cancel it and inspect any partial work before retrying:

   bazilion agent cancel <agent-id>

Cancellation does not undo tool writes or sent messages. Retrying may repeat side effects; check the workspace and conversation first.

Inspect a schedule

bazilion trigger list <agent-id>
bazilion trigger history <trigger-id>

The trigger’s last-fired timestamp means an occurrence was recorded, not that the Agent completed it. Read dispatch history for attempts, errors, and terminal results. Busy Agents defer delivery; retries are bounded. Interval occurrences coalesce while a dispatch is open instead of building an unlimited queue.

Before testing a new schedule, complete Docker/provider readiness. Create a disabled trigger, inspect it, and explicitly enable it when ready:

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

To pause future scheduled work, use bazilion trigger disable <trigger-id>. Disabling is not cancellation of an already active turn. After repairing a terminally failed occurrence, decide whether to run a new explicit turn or wait for the next occurrence; do not assume editing settings replays terminal work.

Held communication versus shell approval

What you see	Where to act	Meaning
Pending communication	Operations → Approvals	One policy-gated attempt awaits a decision
Team Activity denial	Team → Policy	Missing or invalid permission; fix policy before a new attempt
Inline dangerous-command prompt	Interactive web chat or TTY CLI	Ephemeral shell decision for the active turn
Trigger retry/error	Agent → Triggers	Inspect delivery failure and readiness

Inspect the payload and current policy before approving communication:

bazilion approval list
bazilion approval show <approval-id>
bazilion approval approve <approval-id> --yes

Use deny or cancel instead of approve when appropriate. Approval rechecks membership and policy and dispatches at most once. For a scheduled occurrence, approval grants that occurrence; the scheduler still executes it under its lease/retry rules. A communication approval never grants a shell command.

Work through Attention

Open Operations → Attention or use bazilion attention list. Follow an item’s source to the failed turn, trigger, approval, or lesson. Resolve that source first. Acknowledging an informational item records that you saw it; it does not repair a failure or approve work. CLI acknowledgement requires the item’s displayed key and --yes.

Review and apply a lesson

1. Open the Agent’s Learning section. Check the review model and reasoning settings and ensure protected/restricted review readiness is available.
2. Choose Review now, or enable periodic reviews. Review inspects a bounded set of successful user turns and may legitimately produce no proposals.
3. Read each proposed lesson and its evidence. Edit its text and choose whether it belongs to the Agent’s future prompts or searchable Team memory. Click Save edit before approving changed text or scope.
4. Approve only the scoped lesson you intend to keep, or reject the proposal. No proposed lesson is applied without approval.
5. To withdraw an applied lesson, use Revoke and confirm the lesson removal. Agent-scoped revocation removes it from future prompts; Team-scoped revocation removes the applied memory note. Evidence and decision history remain. Previous conversations and actions are not undone.

Stable information about the human or project belongs in Team Context (USER.md). A review proposal is not a replacement for that shared context.

Before an upgrade or recovery

Use Backup and recovery and the version-specific upgrade steps. If Node changes, reinstall native dependencies with the runtime used by the daemon. Do not reset a healthy home just to update the model catalog.

Follow up, answer and inspect retained work

Use Conversations to inspect prior work and start a new conversation safely. Saved results are immutable outputs, independent of later workspace edits. When the Agent is busy, queue a follow-up rather than assuming the current turn has consumed a new instruction. Stop pauses future claims before cancelling the active turn.

An Agent question is live clarification; its accepted and consumed states are separate from task completion and from communication approval. Optional Attention notifications report existing sources without resolving them. An uncertain delivery needs an explicit retry and may duplicate a Telegram message.