Agents & Sessions

Multi-agent

Tapioca runs independent agents side by side, each with its own provider, model, system prompt, goal, history and stats; they stream concurrently. Each agent gets a tab in the title bar — ctrl+n opens a new one, ctrl+→/ctrl+← switch, ctrl+w closes.

/fork branches the current conversation into a new agent, and /agent [name] creates a fresh one.

Subagents

The model can delegate a self-contained task with spawn_agent: it runs in its own tab with a fresh context window and only its final answer comes back, so wide searches never flood the main conversation. Subagents cannot spawn their own, and they run under the same permission mode, executor and grants as the agent that spawned them.

Conversation mechanics

• /regen regenerates the last response.
• /edit pulls the last prompt back into the input.
• up recalls prompt history.
• Prompts typed while the agent is busy are queued and sent automatically when it finishes.
• /goal text pins a goal into the system prompt (/goal clear removes it).
• /btw note adds context without asking for a reply.
• /compact summarizes the conversation to free context (also automatic when context nears the limit, if auto_compact is on).
• /clear clears the current agent's conversation.

Sessions

The whole workspace autosaves. /resume opens a picker whose filter searches across the text of all saved conversations.

Sessions remember the project they belong to: tapioca -c continues this directory's latest, and /resume lists it first. tapioca -r <id> resumes a specific session, and --fork-session -c continues into a new session id.

Titles are written by the model from your first prompt (title_model in the config points that at a cheaper one), falling back to truncation.

Sessions live in ~/.local/share/tapioca/sessions/ as JSON snapshots with full-text search blobs, stored with owner-only permissions.