Skip to content

Overview

This page is the one-paragraph summary of what the gateway is, who owns it, and what is and is not in scope for v1.0.0.

What it is

The Smart LLM Supervisor Gateway is a small OpenAI-compatible HTTP service that decides which underlying model should answer a request, then proxies the request and response to OpenRouter. It is a single FastAPI process with a deterministic policy engine at its core.

What it is not

  • It is not an agent framework. It does not execute tools, run shell commands, or write files. Your coding client is still the agent.
  • It is not a generic LLM router across arbitrary providers. It routes inside the OpenRouter ecosystem.
  • It is not multi-tenant. The production deployment is a personal VPS gateway with a single API key.
  • It is not a training, fine-tuning, or evaluation platform. The evaluation artifacts under artifacts/ are how the maintainers measured this release, not a public service.

Release status

Field Value
Release v1.0.0 (annotated tag, commit 7e0d24f)
Architecture freeze SGW-V3-DESIGN-FREEZE-2026-08-20-01
Production endpoint https://smart-openrounter.bee1x.one
Status PROJECT_COMPLETE (per PROJECT-COMPLETION-REPORT.md)
Local tests 118/118 PASS (make check)
Independent review PASS, 0 BLOCKER, 0 HIGH

What's in scope for v1.0.0

  • OpenAI Chat Completions API on the wire.
  • A small set of routing aliases that map to the frozen model registry.
  • Streaming responses (Server-Sent Events) for the chat completions endpoint.
  • A health probe and a Prometheus /metrics endpoint.
  • A privacy gate that refuses to send sensitive content to the free Scout model.
  • A budget policy with per-goal soft and hard caps (stateful path only; degraded mode requires human approval).
  • A model health registry with cooldown, half-open probes, and quarantine semantics.
  • An emergency fallback to z-ai/glm-5.2 when the senior solver fails.

What is intentionally out of scope

  • The OpenAI Responses API (no client uses it on the production deployment; see User guide → Model aliases).
  • The Anthropic /v1/messages shape.
  • Multimodal / vision input. Requests with image parts are rejected by the policy layer.
  • Multi-user RBAC, virtual keys, and per-tenant spend dashboards.

Where the source of truth lives

  • The policy (priority order, hard Pro rules, thresholds) is in frozen-v3/config/routing-policy.yaml. The gateway refuses to start if the freeze ID is changed.
  • The model registry (aliases → real OpenRouter IDs) is in frozen-v3/config/models.yaml.
  • The operational state (production, staging, evidence) is in PROJECT-COMPLETION-REPORT.md and the artifacts/ tree.

If anything in the documentation disagrees with those files, the files win. Please open an issue (or follow the internal change process documented in docs-internal/development/) rather than silently editing prose.

Next steps