ADR 0001: LLM Public Compatibility Profile

• Status: Accepted
• Date: 2026-07-18
• Gate: LF-2 before LF-3

Decision

The first public surface is OpenAI Chat Completions: buffered JSON, SSE, and GET /v1/models. Models are authorized aliases, never provider deployment identifiers. Errors use the OpenAI envelope while retaining a typed, sanitized internal category.

A typed parse failure is terminal. Same-format OpenAI forwarding may retain unknown fields only in a size-bounded compatibility envelope and only for an alias/deployment allowlist. Cross-format OpenAI-to-Anthropic routing uses canonical typed content and rejects unrepresentable fields. No detect-and-opaque fallback is allowed.

The checked-in corpus under benchmarks/llm-gateway/payloads is the Phase 0 compatibility baseline. Account/CLI providers are not eligible for the shared gateway.

Consequences

LF-3 codecs can define one closed canonical model. Compatibility behavior is observable and cannot turn parse failures into arbitrary upstream forwarding.