What problem does it solve? When building agents that work across Anthropic, OpenAI, Google, and xAI, teams must decide which API differences to normalize and which leak into core architecture. Getting this wrong causes silent degradation of multi-turn tool use, where dropped reasoning state produces worse agents with no visible error. ## Core Features & Use Cases - Two-tier design rule: Specifies a thin universal driver per provider with opinionated primitives on top, and enumerates which of 14 concerns are trivial to normalize versus which leak into core types. - Opaque continuation state contract: Provides a checklist for round-tripping provider-specific reasoning blobs (Anthropic thinking blocks, OpenAI reasoning items, Gemini thought signatures) without dropping, merging, or inspecting them. - Anti-pattern catalog: Documents measured failures of middleware frameworks like LiteLLM, hosted-tool normalization, and provider-specific budget units. - Use Case: When adding a second model provider to an agent system, use this Skill to decide what belongs in the adapter versus the core types, and to verify that opaque state survives serialization, checkpointing, and resume. ## Quick Start Ask the agent to review your provider adapter design against the provider-abstraction skill's normalize-versus-leaks table and opaque state checklist.
