| name | local-subagent-delegation |
| description | Use when local Qwen/Copilot should delegate isolated research, review, or summarization to subagents without bloating parent context. |
| argument-hint | [subtask to delegate] |
Local Subagent Delegation
Use subagents for context isolation, not for every task.
Procedure
- Delegate only when isolation helps.
- Good targets: codebase exploration, independent review, log inspection, alternative-plan comparison, focused documentation lookup.
- Poor targets: tiny edits, single-file changes, anything where the subagent would need the full parent transcript.
- Default to no delegation for short tasks.
- Give the subagent a narrow contract.
- Exact question.
- Only the files, commands, or symptoms it needs.
- Output format and length cap.
- Whether editing is allowed (default: read-only).
- Require a compact return format.
- Paths, facts, risks, recommendation.
- No raw logs, no full files, no long diffs.
- Stop as soon as evidence is sufficient.
- Synthesize in the parent context.
- Treat the subagent result as a compressed evidence packet.
- Re-open original files only when the summary is ambiguous or the change is high-risk.
- Avoid nested subagents unless the user explicitly wants divide-and-conquer.
- Hand off long sessions instead of growing them.
- If the parent context is already large, prefer a session handoff over dispatching a subagent that has to re-read everything.
Output Contract
For each subagent result, preserve:
- Subtask name.
- Files or sources inspected.
- Findings in priority order.
- Recommendation.
- Confidence or uncertainty.
References