Files
kong-platform-llm/dev/proxy-kustomization.yaml
T
Stéphane Tailland f9b73fda9f fix(llm): validate tokens against the classic-models realm, not the stale VPS one
The AI gateway's openid-connect plugin still carried the cluster-wide
dataplane-template-defaults issuer, https://keycloak.sttlab.eu/realms/sttlab.
Every other domain gateway overrides it to the lab realm; this one was missed.

The plugin therefore fetched JWKS from a realm that never signed the token and
rejected it with "invalid signature", which reaches the caller as a bare 401
"Authentication required". Combined with the missing llm-proxy audience in the
realm, the OIDC path on this gateway had never authenticated anyone — only the
key-auth path LibreChat uses was ever exercised.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
2026-08-28 11:09:39 +02:00

62 lines
2.9 KiB
YAML

apiVersion: kustomize.toolkit.fluxcd.io/v1
kind: Kustomization
metadata:
name: platform-llm-dev
namespace: llm
spec:
interval: 10m
sourceRef:
kind: GitRepository
name: platform-llm-dev
path: ./kong
prune: true
dependsOn:
- name: llm-dev-ai-gateway
# The AI gateway Kustomization moved out of flux-system into this
# namespace. A dependsOn still naming flux-system resolves to nothing
# and holds this Kustomization not ready rather than failing loudly.
namespace: llm
postBuild:
substitute:
NAME: llm-dev
NAMESPACE: llm
CONTROL_PLANE_REF_NAME: llm-dev-cp
# The classic-models realm on the lab Keycloak, not the stale VPS realm
# inherited from dataplane-template-defaults. Every domain gateway
# overrides it the same way; this one was missed, so its openid-connect
# plugin validated against the wrong JWKS and answered 401 to every
# token — visible only as "invalid signature" in the gateway log.
OIDC_ISSUER: https://keycloak.sttlab.local/realms/classic-models
AGENT_AUDIENCE: '["llm-proxy"]'
OPENROUTER_UPSTREAM_URL: https://openrouter.ai/api/v1/chat/completions
# Cost per 1M tokens for the model behind OPENROUTER_UPSTREAM_URL, used by
# ai-proxy-advanced to compute kong_ai_llm_cost_total. Checked against
# OpenRouter's price list for google/gemini-3.7-flash on 2026-08-27
# ($0.375 in / $1.875 out). These are a snapshot, not a live rate: change
# the model or the contract and they silently become wrong money.
LLM_INPUT_COST: "0.375"
LLM_OUTPUT_COST: "1.875"
OPENROUTER_EMBEDDINGS_UPSTREAM_URL: https://openrouter.ai/api/v1/embeddings
VECTORDB_HOST: llm-dev-vectordb
VECTORDB_PORT: "6379"
RATE_LIMIT_GLOBAL_TOKENS_PER_MINUTE: "500000"
RATE_LIMIT_PER_CLIENT_TOKENS_PER_MINUTE: "50000"
# Looked up after consumer.yaml's llm-anonymous KongConsumer first
# reconciled — config.anonymous needs the real Consumer ID, not the
# username (Konnect stores it as a literal string otherwise, silently
# never resolving — verified live, request-termination never fired).
LLM_ANONYMOUS_CONSUMER_ID: "1efd79ee-fd2d-4f2d-beb8-e6e9b24185a1"
# The URI path is the collector's routing key: Fluent Bit tags the
# record `llm` and turns that into the Loki label domain=llm.
AUDIT_LOG_ENDPOINT: http://audit-log-collector.observability.svc.cluster.local/llm
substituteFrom:
# Readable again now that k8s-platform places this ConfigMap in every
# domain namespace. It used to be flux-system only, which is why this
# Kustomization — which must stay here, holding real credentials — had to
# inline OTEL_TRACES_ENDPOINT instead. Listed first so the Secret still
# wins on any key it defines.
- kind: ConfigMap
name: dataplane-template-defaults
- kind: Secret
name: platform-llm-dev-secrets