What problem does it solve? Coordinating multi-model AI work often wastes money on oversized models or fails when tasks exceed context windows. This Skill defines a medium-tier orchestration structure where GLM-5.2 coordinates and validates while cheaper mimo and deepseek variants execute bounded tasks, routing each slice to the cheapest model whose context window can hold it. ## Core Features & Use Cases - Role-based delegation: GLM-5.2 acts as orchestrator and oracle (in a fresh instance for unbiased gate validation), while mimo-v2.5/deepseek-v4-flash serve as cheap hands and mimo-v2.5-pro/deepseek-v4-pro as skilled hands. - Measurement-driven routing: Pre-flight working-set measurement (wc -l, ~10 tokens/line, <75% window target) determines whether a task goes to a 400K or 1M model, with a return contract (out_of_scope, too_big, mismatch) triggering re-routing instead of silent compaction. - Planned and unplanned modes with a context tripwire: Big tasks require a pre-built plan; unplanned exploration stops at ~50% context fill and demands a plan. - Use Case: Refactoring a module across 30 files — the orchestrator measures each slice, spawns parallel cheap-hand workers with exact file lists and success checks, then calls a fresh GLM-5.2 oracle at the end gate to validate coherence before review. ## Quick Start Ask the agent to orchestrate a multi-file code change using the open-orchestrator-medium tier, delegating edits to cheap workers and validating the result with an oracle gate.
