# Agent Collaboration Policy Use subagents for implementation work that introduces multiple independent features or components. Keep basic, single-purpose tasks in the main agent unless delegation clearly improves the outcome. ## Model selection - **Basic tasks:** delegate to a fast, economical model such as `gpt-5.6-luna`. - **Complex implementation tasks:** delegate to a balanced model such as `gpt-5.6-terra`. - **Code reviews and very complex tasks:** delegate to a frontier model such as `gpt-5.6-sol`. When the available model names differ, choose the closest equivalent by the same capability tier: fast/economical, balanced/complex, then frontier/review. ## Delegation practice - Split only work that is independently actionable, with clear file or module ownership, to avoid overlapping edits. - Retain final integration, verification, and user communication in the primary agent. - Do not use subagents for simple edits, one-off questions, or narrowly scoped diagnostics unless the user explicitly requests delegation. ## Feature delivery and pull requests - Deliver new features through pull requests. - Once a requested task is complete and verified, automatically commit the relevant changes, push the branch, and open a pull request unless the user explicitly asks not to. - Do not create a pull request for a question, investigation, review-only task, or an intentionally uncommitted work-in-progress. ## Pull request descriptions Write a detailed but straightforward PR body for an AI engineer who does not need to read the implementation to understand the change. Use plain language and explain: - what changed from a user's perspective; - why it matters; - how to verify it, including any important limitations or follow-up work. Avoid implementation jargon, internal file names, framework details, and code walkthroughs unless they are essential to using or reviewing the feature. Keep the description focused: include enough context to make the decision clear, without turning it into a technical design document.