Task Dispatch
Caboose can break complex tasks into parallel work streams using its dispatch system. Three tools work together: the agent tool for spawning subagents, the explore tool for autonomous codebase navigation, and the todo tools for tracking progress.
Subagents
Section titled “Subagents”The agent tool lets the AI spawn independent subagents that work in parallel. When the primary agent identifies tasks that don’t depend on each other — implementing separate features, updating unrelated files, running independent analyses — it dispatches each to a subagent.
Each subagent runs in its own isolated git worktree, keeping its changes completely separate from the main working tree and from other agents. The primary agent continues working or waits, and coordinates the results when subagents finish.
Approval flow
Section titled “Approval flow”Subagent spawning is gated by your current permission mode:
- Chug mode — subagents spawn automatically
- Create / AutoEdit mode — a dialog appears with
y(approve),n(deny), ora(always-approve for this session)
Conflict detection
Section titled “Conflict detection”Before merging a subagent’s worktree back into the main tree, Caboose performs hunk-level conflict analysis:
- Non-overlapping edits — merged automatically
- Overlapping edits — a conflict report surfaces the affected files and line ranges from each agent for your review
Sidebar status
Section titled “Sidebar status”Active subagents are shown in the sidebar with live status indicators:
- Blinking dot — running
- Yellow dot — waiting for approval
- Green dot — complete
- Elapsed time shown for each agent
After completion, a clear button dismisses the subagent entry.
Codebase exploration
Section titled “Codebase exploration”The explore tool performs multi-step codebase exploration autonomously. Rather than a single search, it chains reads, greps, and glob searches to answer structural questions: “Where is this interface implemented?”, “What calls this function?”, “How does data flow from X to Y?” The agent uses explore when it needs deep understanding of unfamiliar code before making changes.
Progress tracking
Section titled “Progress tracking”The todo_write and todo_read tools give the agent a lightweight task list during execution. When working on a multi-step plan, the agent writes checkpoints and marks them complete as it progresses. The task list is visible in the session UI so you can monitor progress in real time. The list is automatically cleared when you send a new message, so stale tasks don’t linger after a topic change.