| name | subagent-first |
| description | Orchestrate work by delegating to sub-agents. Main agent handles only planning, judgment, and user communication. Always active for multi-step work. |
| user-invocable | true |
| disable-model-invocation | false |
Subagent-First
Overview
Preserve the main agent's context window by delegating all investigation, analysis, execution, and validation
to sub-agents.
The main agent acts strictly as an orchestrator: planning, high-level judgment, user communication, and state tracking.
Main Agent Boundaries
- DO NOT read or edit codebase files directly.
All file modifications, tests, and analysis must be done via sub-agents.
- Exception: Small, self-contained inline tasks (under 50 lines of expected output) that are directly needed for the
next immediate decision may be handled by the main agent. However,
lint, test, and build commands must ALWAYS
be delegated to a sub-agent regardless of size.
Dispatch Rules
- Minimize Return Context: Sub-agents must write verbose outputs (e.g., test logs, lint results, raw search results) to
{session_dir}/ and return ONLY:
- Success or failure status
- File path to the full log in
{session_dir}/
- A concise summary of items requiring action (e.g., specific test failures)
- Explicit Context: Always provide each sub-agent with:
- Exact file paths (never let them guess)
- Specific scope: what to do AND what NOT to do
- Expected output format or success signal
session_dir resolves to ~/.copilot/session-state/{session_id}/files/.
Model Selection
When invoking a sub-agent, choose the cheapest model that is sufficient for the task:
claude-haiku-4.5: Default for trivial work: status checks, simple reads/searches, basic commands, or short summaries.gpt-5.3-codex: Default for coding work:
codebase inspection, targeted edits, debugging, test-fix loops, and medium-to-complex implementation.claude-sonnet-4.6: Use for deep refactoring, architecture analysis, complex test design,
or decisions requiring broad engineering judgment.gpt-5.4: Use for second opinions, alternative perspectives, or review of another model's plan/output.
