config.yaml:"# PRODUCTION GATEWAY CONFIGURATION\n# ==========================================\n# All upstream services MUST use Kubernetes internal service DNS names\n# Format: <service>.<namespace>.svc.cluster.local\n# \n# This ensures:\n# - Communication within cluster network only (no external IP exposure)\n# - Pod-to-pod service discovery via internal DNS\n# - Security policy enforcement at network level\n# - Service-level load balancing via kube-proxy\n#\n# Routing Pattern:\n# PREFERRED: X-Service header routing (e.g., X-Service: workflow)\n# Legacy: Path-based routing (e.g., /workflow) - being deprecated\n#\nauth:\n enabled: true\n # ZITADEL is the only identity provider. CoreDNS rewrites\n # zitadel.riotpiao.com to ingress-nginx, so discovery and JWKS stay inside\n # the cluster and TLS is the wildcard certificate.\n issuer: \"https://zitadel.riotpiao.com\"\n # zitadelDomain enables the official ZITADEL Go SDK path: OIDC discovery,\n # remote JWKS cache, and op.AccessTokenVerifier.\n zitadelDomain: \"zitadel.riotpiao.com\"\n # Audience allow-list: a caller token must carry one of these client IDs.\n # Add a consumer here before it may call the LLM models.\n zitadelClientIds:\n - \"396757800874672511\" # api-gw\n - \"396757803559027071\" # paperless (paperless-ai-agent)\n - \"396757803877794175\" # poimen\n - \"396757801210282367\" # argocd\n - \"396757804213403649\" # queue-sqs\n # Two spellings accepted. ZITADEL does store role keys with a colon (this\n # project holds iam:read, memory:write and llm:inference), so the colon form\n # is primary and the hyphen form is the legacy alias.\n requiredCapability: \"llm-inference,llm:inference\"\n # token exchange (client_credentials) against the same instance\n tokenUrl: \"https://zitadel.riotpiao.com/oauth/v2/token\"\n clientId: \"396757800874672511\"\n # jwksUrl MUST be set. cmd/gateway/main.go:144 builds the dispatcher\n # validator only when cfg.Auth.Enabled and cfg.Auth.JWKSURL are both set, so\n # an unset value leaves jwtValidator nil and every adapter request passes with\n # only a warning in the log. The SDK discovers the keys on its own path, but\n # the adapter dispatcher does not use that path.\n # In-cluster endpoint, not the public host. The public host resolves through\n # the CoreDNS rewrite to ingress-nginx and the fetch timed out there, so the\n # validator answered 403 with a context-deadline error on every caller. The\n # ZITADEL service answers JWKS on 8080 and the NetworkPolicy already allows\n # egress from this namespace to iam:8080.\n # The public host, not the service. The ZITADEL service port 8080 answers 404\n # for /oauth/v2/keys, while https://zitadel.riotpiao.com/oauth/v2/keys answers\n # 200 from a pod in this namespace.\n jwksUrl: \"https://zitadel.riotpiao.com/oauth/v2/keys\"\nroutes: []\nmodels:\n# All model services use internal Kubernetes DNS (llm-serving namespace)\n- name: \"reasoning\"\n address: \"reasoning-predictor.llm-serving.svc.cluster.local:80\"\n path: \"/v1/chat/completions\"\n# clef-flash is a DECISION model, not a chat model. It takes a `state` plus a\n# schema of typed `questions` and returns one probability per allowed option in\n# a single prefill pass. There is no decoding and no free-form text.\n#\n# Its request and response shapes are Clef's own (System One / Jev), NOT\n# OpenAI's: no `messages`, no `choices`. A standard OpenAI client cannot call\n# it, and the public path is /v1/systemone rather than /v1/chat/completions.\n#\n# The dispatcher resolves this model by the body's `model` field, then forwards\n# to the REQUEST path. The `path` value below records intent and must stay in\n# sync with the dispatch list in internal/proxy/router.go.\n- name: \"clef-flash\"\n address: \"clef-flash-predictor.llm-serving.svc.cluster.local:80\"\n path: \"/v1/systemone\"\n# `ornith:35b` retired: the Ollama Llama-35B predictor is gone and its GPU now\n# serves clef-flash. Callers must move to `reasoning`, which keeps the same\n# serving name while running Qwen3-30B-A
