Skip to main content
Version: main 🚧

Specification Quality Checklist: Centralise LLM Routing, Quotas and Budgets

Created: 2026-10-10 · Feature: spec.md · Issue: #2874

Content quality​

  • Problem and goal stated without committing to one gateway product
  • User journeys for platform admin, agent author, end user and agent runtime
  • Bullets and Mermaid diagrams over prose

Design criteria (from #2874)​

  • Plug-and-play: any OpenAI-compatible gateway; swap needs no agent change (FR-001, FR-002, SC-001)
  • Budgets: JWT mode enforces user, agent and team limits; key mode enforces agent and team limits, with best-effort per-user limits (FR-004, FR-005, contract ②)
  • AuthN: JWT mode verifies user, agent and team identity; key mode verifies agent/team bindings but leaves user attribution unverified. Neither mode uses provider or shared admin keys in agents (FR-004–FR-007)
  • AuthZ: OpenFGA governs agent and model access (FR-010)
  • Journeys include first-use onboarding (J1–J4, research Decision 6)
  • Data: one source of truth each, no drifting copies (research Decision 8, FR-015)
  • Rollout: off by default, phased, reversible (FR-018, FR-019, plan)
  • Reliability: fail closed, streaming and tool calls, clear quota errors (FR-012, FR-013)
  • Observability: OTel GenAI per user, agent and team (FR-009)
  • Reuse: llm_wrapper, Keycloak, CAS, llm_models (no parallel mechanisms)

Readiness​

  • Success criteria measurable
  • Open decisions explicit (OD-1, spike S1)
  • Implementation split into follow-up issues (tasks.md)