Proxy

The proxy is the public data-plane edge. It authenticates PATs, verifies agent identity, applies admission controls, optionally assembles context, and forwards to mock or live providers.

Start with the request path

• Overview — endpoints, middleware order, headers, and failure modes.
• Authentication — bearer token, agent header, probes, and error semantics.
• Request routing — model registry, provider selection, and fallbacks.
• Idempotency and budgets — replay, admission holds, and usage accounting.
• Sessions — sticky session IDs and checkpoint behavior.
• Directives — org directive resolution and system-prompt injection.
• Rate limiting — Redis budgets and fail-closed behavior.
• Provider adapters — mock, OpenAI, Anthropic, and self-hosted.
• Provider routing and resilience — model policy, capability manifests, breakers, and fallback.
• Auth caching — bloom + LRU and the revocation gate.
• Configuration — environment variables and startup validation.

Implemented vs provisional

Never infer

A successful mock response does not validate live provider behavior, context quality, billing correctness, or production throughput. A 501 PROVIDER_NOT_CONFIGURED means the model is not in the active registry — not that forwarding is broken.