| name | summon |
| description | Summon the right subagent for a task: survey those available and dispatch the best fit. |
Map the specialized subagents available in this environment and, if $ARGUMENTS
describes a task, recommend the best-fit agent(s) and show exactly how to dispatch them.
1. Enumerate what's available
Gather every agent on hand from all sources:
- Host-exposed types: the agent types your environment already lets you launch
(e.g.
Explore, or whatever the Task tool offers).
- Project agents: definitions on disk at
.claude/agents/*.md.
- Personal agents: definitions at
~/.claude/agents/*.md.
- Plugin agents: any agents contributed by installed plugins. Godmode ships five:
oracle (read-only code recon), demiurge (implementation), inquisitor (review),
arbiter (test/verify judgment), and seer (debugging / root-cause).
Glob those paths and read each definition's frontmatter for its name and description.
De-duplicate by name; note where each one comes from.
2. Summarize each
For every agent, give a one-line summary of what it's best at, pulled from its
description, not invented. Group them by source (host, project, personal, plugin) so the
user can see what's built in versus what this repo adds.
3. Recommend for the task
If $ARGUMENTS names a task, pick the best-fit agent(s):
- State which agent and why it fits; match the task's shape to the agent's strength.
- Show the exact dispatch: the agent type and a tight prompt describing the goal.
- If the work splits into independent pieces, recommend launching multiple agents in
a single batch so they run in parallel. Spell out the split.
- If nothing fits well, say so and suggest handling it in the main thread instead.
4. Otherwise, just present the list
If no task is given, present the categorized list from steps 1–2 so the user knows what's
on hand, and invite them to re-run with a task description for a recommendation.
The principle: delegate broad exploration and parallelizable, well-bounded work to
subagents; keep tightly-coupled reasoning that needs full context in the main thread.