kimi-code

Large multi-step repository refactoring can be automated by delegating the implementation passes to a CLI for subagent-style execution, passing the scope and constraints via prompt context handoff. This reduces context cost and fatigue by orchestrating isolated non-interactive runs.

Deep codebase exploration is handled by delegating the heavy lifting to a CLI in non-interactive mode with controlled output formatting. You provide the original goal, prior decisions, and verification expectations, allowing isolated runs to perform batch code edits.

Long-horizon planning context is passed using prompt-based context handoff that includes the original goal, prior decisions, scope, constraints, and verification expectations. This cognition-style handoff ensures isolated automated runs stay aligned with the overarching repository objectives.

Batch code edits can run in non-interactive mode with auto-approval by delegating repo-wide edits to a CLI. You enforce a verification command that the agent must pass before returning, ensuring guardrailed automation and reconciling results back into the main thread.

Repo-wide code delegation supports model selection to handle complex feature slices and migrations. It requires non-interactive programmatic CLI usage where you specify the chosen constraints and the exact test command the agent must pass.

An orchestrator context pass is unnecessary for codebase migrations when a single pass is sufficient to cover the required changes. You should only use delegation for broad refactors and deep exploration where a single orchestrator context pass is insufficient.