gemini-cli

Delegating large refactoring tasks requires passing a single handoff prompt with the original goal, prior decisions, and repo scope to an isolated programmatic agent. This preserves alignment by reconciling the resulting file changes and decisions back into your main orchestrator thread.

Running a security audit when local context is insufficient involves delegating the task to an isolated agent using specialist prompting. The delegated run executes with auto-approved tool use and post-run verification to ensure isolation-safe checks are reconciled back to the main thread.

Performing deep codebase investigation and web-grounded documentation searches without context flooding is possible by delegating the multi-step work to an isolated agent run. This non-interactive execution handles the heavy exploration and returns only the reconciled output to your main thread.

Verifying file changes after delegating multi-file edits requires the delegated prompt to run explicit checks like tests, lint, or typecheck. The agent must execute these commands during the isolated run and reconcile the diffs and decisions back into the main orchestrator.

Delegating deep work to an external agent does require auto-approved tool use for non-interactive programmatic execution. This ensures the isolated run can autonomously perform strict model selection and apply isolation-safe prompts without manual intervention.

You should not use an isolated agent run for broad repository refactoring when your primary orchestrator has sufficient local context to handle the task. Isolated delegation is specifically for heavy multi-file edits and long-horizon work that would otherwise flood the main context window.