What problem does it solve? Invoking a planner subagent in phase mode burns tokens fast because a resumed agent replays its entire prior transcript, and coordinators often duplicate scanning work or leave decision forks ungated. This protocol keeps each planner call near a 100k token budget while ensuring important architectural decisions pause for the user. ## Core Features & Use Cases - Token Budget Enforcement: Targets roughly 100k subagent tokens per phase-mode planner call and diagnoses which rule slipped when a call runs over. - Targeted Dispatch Prompts: Requires naming exact files, functions, line ranges, and CSS selectors, plus referencing exactly one prior phase as the format model. - Decision Gate Discipline: Distinguishes trivial forks (auto-decided and noted) from important forks (architecture, reuse-vs-build, contradictions of dispatch assumptions) that must halt for the user's explicit pick, with batching to minimize resume round trips. - Use Case: Before dispatching @planner to write a phase 4 spec and tasks breakdown, run a few targeted greps to pin file locations, name phase 2 as the format model, and instruct planner to read only its own tasks.md slice. ## Quick Start Invoke the planner agent for the next phase spec using this protocol, naming exact files and line ranges, citing one prior phase as the format model, and gating any architectural decisions for my approval.
