What problem does it solve? Running multi-agent coding workflows on frontier models is expensive, and routing tasks to oversized models wastes money while undersized models compact mid-task and lose state. This Skill defines a cost-optimized orchestration structure where an orchestrator delegates pre-digested work slices to cheap worker models and reserves expensive models for planning and validation gates. ## Core Features & Use Cases - Three-role architecture: GLM-5.2 acts as the oracle for planning and validation gates, mimo-v2.5-pro coordinates as the orchestrator, and mimo-v2.5 or deepseek-v4-flash execute bounded edits as workers. - Context-based routing: Measures the exact working set with wc -l and routes each task to the cheapest model whose window holds it under 75% utilization, with a return contract (out_of_scope, too_big, mismatch) that triggers re-routing instead of silent compaction. - Hard delegation rule: The orchestrator never edits files itself; every code change goes through a worker with a scoped prompt containing exact files, one precise task, and a success check. - Use Case: When refactoring a large codebase on a budget, the orchestrator greps and measures the affected files, splits the work into slices under 240K tokens, spawns parallel mimo-v2.5 workers, and calls GLM-5.2 once at the end to validate coherence. ## Quick Start Ask the agent to orchestrate this refactoring task using cheap open models, delegating all edits to workers and validating the result with the oracle.
