| name | agenthub-task-manager |
| description | User-invoked AgentHub workflow for planning a multi-part request and delegating its independent subtasks to parallel AgentHub agent sessions. Use only when the user explicitly invokes this skill to plan, create, list, or clean up AgentHub task sessions. |
| disable-model-invocation | true |
| user-invocable | true |
| allowed-tools | ["ToolSearch","AskUserQuestion","mcp__agenthub__agent_hub_planning","mcp__agenthub__agenthub_create_worktree_sessions","mcp__agenthub__agenthub_list_worktrees","mcp__agenthub__agenthub_delete_worktree"] |
AgentHub Task Manager
Use this workflow only because the user explicitly invoked this skill to plan and delegate work to parallel AgentHub agent sessions.
Plan And Create Task Sessions
- Read the user's request and infer the independent subtasks yourself. Preserve explicit numbered or bulleted work streams as separate subtasks; do not split ordinary prose just because of punctuation, commas, semicolons, or conjunctions.
- Call
agent_hub_planning, passing the original prompt and the subtasks you inferred as subtasks. AgentHub does not split for you — with no subtasks it plans the whole prompt as a single task.
- Review the returned assignments and the
harnessCapabilities block — each harness's real installed skills and MCP servers. AgentHub launches a harness (Claude Code or Codex), not a model, so present only the harness, never a model name. When multiple harnesses are installed, each subtask comes with a suggested harness plus the matchedCapabilities that drove it. Confirm or override every suggestion by matching the subtask to the harness whose actual tools fit — e.g. a profiling subtask → the harness that has an xctrace/profiler tool; a SwiftUI subtask → the harness with a swiftui skill. Decide from the listed capabilities, never from a harness's general reputation, and cite the matching skill/tool in your rationale.
- Present the final assignments (task, chosen harness + why, branch, launch prompt) and wait for explicit user approval.
- Call
agenthub_create_worktree_sessions only for approved assignments. Pass one task per assignment with explicit provider, branch, and prompt. Include startPath when the task belongs to a subdirectory; AgentHub uses nearby project markers to infer a sparse worktree and then launch the agent from that same relative path. Use sparseProfile only for extra required paths and fullCheckout: true only when the task genuinely needs the whole repository materialized.
For a single-task request, still call agent_hub_planning with one subtask so the agent and branch are explicit before approval.
List Or Delete Task Sessions
Use agenthub_list_worktrees before destructive cleanup when the target or session impact is unclear. Use agenthub_delete_worktree only after the user identifies the session to delete.
Boundaries
Do not use these AgentHub task tools for generic planning, fan-out, background work, or subagent requests unless the user explicitly invoked this skill. Preserve the current harness's native subagent and background-task capabilities.
