Skip to content
Shelzo

Chat and agents

Chat keeps a readable conversation with one engine. Agent adds a name, a brief, memory, and routines for work you repeat.

Chat and approvals

  1. Select Chat (⌘3), create a chat (⌘N), and choose its folder, engine, and model.
  2. Choose the permission mode in the composer: Supervised, Accept edits, Auto, or Full access. Read the mode’s description; the engine controls which requests it raises.
  3. Send a message with Return. Tool calls and results appear in the transcript. When approval is needed, use Approve, Always allow this session, or Decline.
  4. Open Inbox (⇧⌘I) for requests and results across chats, routines, and delegations. Connector permissions remain separate from the engine’s permission mode.

Switch engines or hand off

The composer’s engine control switches the current chat to another supported engine and carries conversation context forward. Model and effort choices depend on the engine. If Codex asks for a restart to change permissions, follow that message.

Use the chat menu’s Review with or Continue in entries to hand work to another installed chat engine. This opens a linked chat with conversation context and the uncommitted diff. Long context is limited; include important requirements in your latest message.

Create a named agent

  1. Select Agent (⌘1) and New Agent… (⌥⌘N).
  2. Choose a name, engine, home folder, and brief. For example: “Docs Keeper: keep the setup guide consistent with the source.”
  3. Use the agent’s memory editor for information it should keep across chats.
  4. Create a routine with a saved prompt and schedule, or use Run now. Keep Shelzo running for scheduled work. Missed occurrences are caught up once after launch.

Pause agents queues new routines and agent questions; running chats continue. Local composers also offer @ agent questions when the named agent’s grants allow them.

If it fails

Check Inbox for a pending approval, then check the engine’s sign-in and exit details. For a routine, open its run history, fix the reported problem, and use Retry. Open the agent’s home workspace and resume paused agents before retrying.