fix(proxy): address review findings on admission control (#64 )

- tracing: responseWriter now forwards Flush and exposes Unwrap. Without it w.(http.Flusher) failed for every request passing the tracing middleware, so the parked-stream keep-alive never started: a parked streaming request wrote nothing for up to ADMISSION_MAX_PARK_SECONDS, which is exactly the client timeout the keep-alive was added to prevent. - proxy: once the SSE status line is spent, ModifyResponse frames an upstream response >= 400 as an event: error instead of letting a raw error body land inside the stream, where a streaming client reads it as an empty completion and reports success. net/http keeps the first status, so the upstream status cannot be forwarded after the park has written one. - engine_state: new Snapshot() returns the state and its tick channel under one lock. Get and Tick taken separately can miss the poll that lands between the two calls, parking a request for an extra interval. - engine_state: observer_degraded now tracks current parsability rather than being raised once on the first poll and never cleared, which pinned it at 1 after a single truncated response and never reported drift that appeared later. markDegraded only touches the exporter when the value changes. - deployment: ADMISSION_MAX_INFLIGHT is set explicitly to 4, documenting the coupling to the engine's --max-num-seqs, instead of silently defaulting to VLLM_MAX_NUM_SEQS.
feat(proxy): engine-wide LLM admission control via vLLM gauges (#64 )
2026-09-30 19:32:10 +08:00 · 2026-09-30 17:39:47 +08:00
11 changed files with 65 additions and 796 deletions
@@ -5,51 +5,39 @@ go 1.26.0
 require (
 	github.com/MicahParks/keyfunc/v2 v2.1.0
 	github.com/golang-jwt/jwt/v5 v5.3.1
-	github.com/google/uuid v1.6.0
-	github.com/zitadel/oidc/v3 v3.49.2
-	github.com/zitadel/zitadel-go/v3 v3.30.0
 	go.opentelemetry.io/otel v1.46.0
 	go.opentelemetry.io/otel/exporters/otlp/otlptrace/otlptracegrpc v1.46.0
 	go.opentelemetry.io/otel/sdk v1.46.0
 	go.opentelemetry.io/otel/trace v1.46.0
 	go.temporal.io/api v1.63.5
 	go.temporal.io/sdk v1.48.0
-	golang.org/x/net v0.58.0
 	google.golang.org/grpc v1.83.2
 	google.golang.org/protobuf v1.36.12
 	gopkg.in/yaml.v3 v3.0.1
 )

 require (
-	github.com/bmatcuk/doublestar/v4 v4.10.0 // indirect
 	github.com/cenkalti/backoff/v5 v5.0.3 // indirect
 	github.com/cespare/xxhash/v2 v2.3.0 // indirect
-	github.com/envoyproxy/protoc-gen-validate v1.3.3 // indirect
 	github.com/facebookgo/clock v0.0.0-20150410010913-600d898af40a // indirect
-	github.com/go-chi/chi/v5 v5.3.1 // indirect
-	github.com/go-jose/go-jose/v4 v4.1.4 // indirect
 	github.com/go-logr/logr v1.4.4 // indirect
 	github.com/go-logr/stdr v1.2.2 // indirect
 	github.com/gogo/protobuf v1.3.2 // indirect
 	github.com/golang/mock v1.6.0 // indirect
-	github.com/gorilla/securecookie v1.1.2 // indirect
-	github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.3 // indirect
+	github.com/google/uuid v1.6.0 // indirect
+	github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.2 // indirect
 	github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0 // indirect
-	github.com/muhlemmer/gu v0.3.1 // indirect
-	github.com/muhlemmer/httpforwarded v0.1.0 // indirect
 	github.com/nexus-rpc/nexus-proto-annotations v0.1.0 // indirect
 	github.com/nexus-rpc/sdk-go v0.7.0 // indirect
 	github.com/robfig/cron v1.2.0 // indirect
-	github.com/rs/cors v1.11.1 // indirect
 	github.com/stretchr/objx v0.5.3 // indirect
 	github.com/stretchr/testify v1.12.1 // indirect
-	github.com/zitadel/schema v1.3.2 // indirect
 	go.opentelemetry.io/auto/sdk v1.2.1 // indirect
 	go.opentelemetry.io/otel/exporters/otlp/otlptrace v1.46.0 // indirect
 	go.opentelemetry.io/otel/metric v1.46.0 // indirect
 	go.opentelemetry.io/proto/otlp v1.11.0 // indirect
 	go.yaml.in/yaml/v3 v3.0.5 // indirect
-	golang.org/x/oauth2 v0.36.0 // indirect
+	golang.org/x/net v0.58.0 // indirect
 	golang.org/x/sync v0.22.0 // indirect
 	golang.org/x/sys v0.47.0 // indirect
 	golang.org/x/text v0.41.0 // indirect
@@ -1,19 +1,11 @@
 github.com/MicahParks/keyfunc/v2 v2.1.0 h1:6ZXKb9Rp6qp1bDbJefnG7cTH8yMN1IC/4nf+GVjO99k=
 github.com/MicahParks/keyfunc/v2 v2.1.0/go.mod h1:rW42fi+xgLJ2FRRXAfNx9ZA8WpD4OeE/yHVMteCkw9k=
-github.com/bmatcuk/doublestar/v4 v4.10.0 h1:zU9WiOla1YA122oLM6i4EXvGW62DvKZVxIe6TYWexEs=
-github.com/bmatcuk/doublestar/v4 v4.10.0/go.mod h1:xBQ8jztBU6kakFMg+8WGxn0c6z1fTSPVIjEY1Wr7jzc=
 github.com/cenkalti/backoff/v5 v5.0.3 h1:ZN+IMa753KfX5hd8vVaMixjnqRZ3y8CuJKRKj1xcsSM=
 github.com/cenkalti/backoff/v5 v5.0.3/go.mod h1:rkhZdG3JZukswDf7f0cwqPNk4K0sa+F97BxZthm/crw=
 github.com/cespare/xxhash/v2 v2.3.0 h1:UL815xU9SqsFlibzuggzjXhog7bL6oX9BbNZnL2UFvs=
 github.com/cespare/xxhash/v2 v2.3.0/go.mod h1:VGX0DQ3Q6kWi7AoAeZDth3/j3BFtOZR5XLFGgcrjCOs=
-github.com/envoyproxy/protoc-gen-validate v1.3.3 h1:MVQghNeW+LZcmXe7SY1V36Z+WFMDjpqGAGacLe2T0ds=
-github.com/envoyproxy/protoc-gen-validate v1.3.3/go.mod h1:TsndJ/ngyIdQRhMcVVGDDHINPLWB7C82oDArY51KfB0=
 github.com/facebookgo/clock v0.0.0-20150410010913-600d898af40a h1:yDWHCSQ40h88yih2JAcL6Ls/kVkSE8GFACTGVnMPruw=
 github.com/facebookgo/clock v0.0.0-20150410010913-600d898af40a/go.mod h1:7Ga40egUymuWXxAe151lTNnCv97MddSOVsjpPPkityA=
-github.com/go-chi/chi/v5 v5.3.1 h1:3j4HZLGZQ3JpMCrPJF/Jl3mYJfWLKBfNJ6quurUGCf8=
-github.com/go-chi/chi/v5 v5.3.1/go.mod h1:R+tYY2hNuVUUjxoPtqUdgBqevM9s9njzkTLutVsOCto=
-github.com/go-jose/go-jose/v4 v4.1.4 h1:moDMcTHmvE6Groj34emNPLs/qtYXRVcd6S7NHbHz3kA=
-github.com/go-jose/go-jose/v4 v4.1.4/go.mod h1:x4oUasVrzR7071A4TnHLGSPpNOm2a21K9Kf04k1rs08=
 github.com/go-logr/logr v1.2.2/go.mod h1:jdQByPbusPIv2/zmleS9BjJVeZ6kBagPoEUsqbVz/1A=
 github.com/go-logr/logr v1.4.4 h1:tG4xh9yMsRCAiodLVTxyrkzSZ9+o0L1Kg/+cPVcbP/8=
 github.com/go-logr/logr v1.4.4/go.mod h1:9T104GzyrTigFIr8wt5mBrctHMim0Nb2HLGrmQ40KvY=
@@ -31,12 +23,8 @@ github.com/google/go-cmp v0.7.0 h1:wk8382ETsv4JYUZwIsn6YpYiWiBsYLSJiTsyBybVuN8=
 github.com/google/go-cmp v0.7.0/go.mod h1:pXiqmnSA92OHEEa9HXL2W4E7lf9JzCmGVUdgjX3N/iU=
 github.com/google/uuid v1.6.0 h1:NIvaJDMOsjHA8n1jAhLSgzrAzy1Hgr+hNrb57e+94F0=
 github.com/google/uuid v1.6.0/go.mod h1:TIyPZe4MgqvfeYDBFedMoGGpEw/LqOeaOT+nhxU+yHo=
-github.com/gorilla/securecookie v1.1.2 h1:YCIWL56dvtr73r6715mJs5ZvhtnY73hBvEF8kXD8ePA=
-github.com/gorilla/securecookie v1.1.2/go.mod h1:NfCASbcHqRSY+3a8tlWJwsQap2VX5pwzwo4h3eOamfo=
 github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.2 h1:sGm2vDRFUrQJO/Veii4h4zG2vvqG6uWNkBHSTqXOZk0=
 github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.2/go.mod h1:wd1YpapPLivG6nQgbf7ZkG1hhSOXDhhn4MLTknx2aAc=
-github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.3 h1:B+8ClL/kCQkRiU82d9xajRPKYMrB7E0MbtzWVi1K4ns=
-github.com/grpc-ecosystem/go-grpc-middleware/v2 v2.3.3/go.mod h1:NbCUVmiS4foBGBHOYlCT25+YmGpJ32dZPi75pGEUpj4=
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0 h1:/Tnpcb2E0Pz/tN9s3bfEY2Q8ePCEX9iuS+cneUwncnw=
 github.com/grpc-ecosystem/grpc-gateway/v2 v2.30.0/go.mod h1:zOBXOsUaBSjKgmH4OGzV1esUpR3oUSCPYVd2cUBjKYY=
 github.com/kisielk/errcheck v1.5.0/go.mod h1:pFxgyoBC7bSaBwPgfKdkLd5X25qrDl4LWUI2bnpBCr8=
@@ -45,10 +33,6 @@ github.com/kr/pretty v0.3.1 h1:flRD4NNwYAUpkphVc1HcthR4KEIFJ65n8Mw5qdRn3LE=
 github.com/kr/pretty v0.3.1/go.mod h1:hoEshYVHaxMs3cyo3Yncou5ZscifuDolrwPKZanG3xk=
 github.com/kr/text v0.2.0 h1:5Nx0Ya0ZqY2ygV366QzturHI13Jq95ApcVaJBhpS+AY=
 github.com/kr/text v0.2.0/go.mod h1:eLer722TekiGuMkidMxC/pM04lWEeraHUUmBw8l2grE=
-github.com/muhlemmer/gu v0.3.1 h1:7EAqmFrW7n3hETvuAdmFmn4hS8W+z3LgKtrnow+YzNM=
-github.com/muhlemmer/gu v0.3.1/go.mod h1:YHtHR+gxM+bKEIIs7Hmi9sPT3ZDUvTN/i88wQpZkrdM=
-github.com/muhlemmer/httpforwarded v0.1.0 h1:x4DLrzXdliq8mprgUMR0olDvHGkou5BJsK/vWUetyzY=
-github.com/muhlemmer/httpforwarded v0.1.0/go.mod h1:yo9czKedo2pdZhoXe+yDkGVbU0TJ0q9oQ90BVoDEtw0=
 github.com/nexus-rpc/nexus-proto-annotations v0.1.0 h1:2fELd+9sqUtNu6Fg//pw8YFsxOvp8vZ8hfP0nHhNI80=
 github.com/nexus-rpc/nexus-proto-annotations v0.1.0/go.mod h1:n3UjF1bPCW8llR8tHvbxJ+27yPWrhpo8w/Yg1IOuY0Y=
 github.com/nexus-rpc/sdk-go v0.7.0 h1:38NrfY5rLnZAiMMs2ZfCKI/CSDzdfJG+27iAgfA8bUI=
@@ -57,8 +41,6 @@ github.com/robfig/cron v1.2.0 h1:ZjScXvvxeQ63Dbyxy76Fj3AT3Ut0aKsyd2/tl3DTMuQ=
 github.com/robfig/cron v1.2.0/go.mod h1:JGuDeoQd7Z6yL4zQhZ3OPEVHB7fL6Ka6skscFHfmt2k=
 github.com/rogpeppe/go-internal v1.14.1 h1:UQB4HGPB6osV0SQTLymcB4TgvyWu6ZyliaW0tI/otEQ=
 github.com/rogpeppe/go-internal v1.14.1/go.mod h1:MaRKkUm5W0goXpeCfT7UZI6fk/L7L7so1lCWt35ZSgc=
-github.com/rs/cors v1.11.1 h1:eU3gRzXLRK57F5rKMGMZURNdIG4EoAmX8k94r9wXWHA=
-github.com/rs/cors v1.11.1/go.mod h1:XyqrcTp5zjWr1wsJ8PIRZssZ8b/WMcMf71DJnit4EMU=
 github.com/stretchr/objx v0.5.3 h1:jmXUvGomnU1o3W/V5h2VEradbpJDwGrzugQQvL0POH4=
 github.com/stretchr/objx v0.5.3/go.mod h1:rDQraq+vQZU7Fde9LOZLr8Tax6zZvy4kuNKF+QYS+U0=
 github.com/stretchr/testify v1.12.1 h1:EuwCh5fleGS7H32xRwO3wRGT7DxrDhLAT6FF8MpWDWE=
@@ -66,12 +48,6 @@ github.com/stretchr/testify v1.12.1/go.mod h1:MDEgiDPPsNp5cuIrHPPCyornHKgEVbtFUm
 github.com/yuin/goldmark v1.1.27/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.2.1/go.mod h1:3hX8gzYuyVAZsxl0MRgGTJEmQBFcNTphYh9decYSb74=
 github.com/yuin/goldmark v1.3.5/go.mod h1:mwnBkeHKe2W/ZEtQ+71ViKU8L12m81fl3OWwC1Zlc8k=
-github.com/zitadel/oidc/v3 v3.49.2 h1:yKvB2Hx6rVWp0vOTk1pEyXAPxJusgIHNmr20AxOohmQ=
-github.com/zitadel/oidc/v3 v3.49.2/go.mod h1:HwoguOGo0eem0RK5Gb+P6Q4aQLVinJ9LhomlVEA57ck=
-github.com/zitadel/schema v1.3.2 h1:gfJvt7dOMfTmxzhscZ9KkapKo3Nei3B6cAxjav+lyjI=
-github.com/zitadel/schema v1.3.2/go.mod h1:IZmdfF9Wu62Zu6tJJTH3UsArevs3Y4smfJIj3L8fzxw=
-github.com/zitadel/zitadel-go/v3 v3.30.0 h1:yk/surQ+atLS/ocGaaJdHD/R6VbveFScihxJhIkGZ3Q=
-github.com/zitadel/zitadel-go/v3 v3.30.0/go.mod h1:4R+aki0JFMwIPsr4CIVLYq5DRTVEAfLbehJpFNhFKQw=
 go.opentelemetry.io/auto/sdk v1.2.1 h1:jXsnJ4Lmnqd11kwkBV2LgLoFMZKizbCi5fNZ/ipaZ64=
 go.opentelemetry.io/auto/sdk v1.2.1/go.mod h1:KRTj+aOaElaLi+wW1kO/DZRXwkF4C5xPbEe3ZiIhN7Y=
 go.opentelemetry.io/otel v1.46.0 h1:FHt5/CDyVxi/8IM1CH7VE/rRgq3kLHa2mSTVMO8AWyc=
@@ -111,8 +87,6 @@ golang.org/x/net v0.0.0-20201021035429-f5854403a974/go.mod h1:sp8m0HH+o8qH0wwXwY
 golang.org/x/net v0.0.0-20210405180319-a5a99cb37ef4/go.mod h1:p54w0d4576C0XHj96bSt6lcn1PtDYWL6XObtHCRCNQM=
 golang.org/x/net v0.58.0 h1:ynWG7rqYi4ccpTEuPZ2QGWHktVEM9DMCj9yzDE0Q7To=
 golang.org/x/net v0.58.0/go.mod h1:YwCddHnFlT7eLQqVprV19OnhLGtc5xOKgE0RyqgfWAU=
-golang.org/x/oauth2 v0.36.0 h1:peZ/1z27fi9hUOFCAZaHyrpWG5lwe0RJEEEeH0ThlIs=
-golang.org/x/oauth2 v0.36.0/go.mod h1:YDBUJMTkDnJS+A4BP4eZBjCqtokkg1hODuPjwiGPO7Q=
 golang.org/x/sync v0.0.0-20190423024810-112230192c58/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20190911185100-cd5d95a43a6e/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
 golang.org/x/sync v0.0.0-20201020160332-67f06af15bc9/go.mod h1:RxMgew5VJxzue5/jJTE5uejpjVlOe/izrB70Jof72aM=
@@ -1,21 +0,0 @@
-package auth
-
-import (
-	"strings"
-	"testing"
-)
-
-func TestValidateBearerToken_JWEGetsClearError(t *testing.T) {
-	ti := newTestIssuer(t)
-	host, port, _ := strings.Cut(strings.TrimPrefix(ti.server.URL, "http://"), ":")
-	v, err := NewZitadelValidator(t.Context(), ti.server.URL, nil, ZitadelConfig{
-		Domain: host, InsecurePort: port, HTTPClient: ti.server.Client(),
-	})
-	if err != nil {
-		t.Fatalf("NewZitadelValidator: %v", err)
-	}
-	_, err = v.ValidateBearerToken("Bearer a.b.c.d.e")
-	if err == nil || !strings.Contains(err.Error(), "encrypted JWE") {
-		t.Fatalf("expected a JWE-specific error, got %v", err)
-	}
-}
@@ -3,141 +3,37 @@ package auth
 import (
 	"context"
 	"fmt"
-	gohttp "net/http"
 	"strings"
 	"sync"
 	"time"

 	"github.com/MicahParks/keyfunc/v2"
 	"github.com/golang-jwt/jwt/v5"
-	"github.com/zitadel/zitadel-go/v3/pkg/authorization/oauth"
 )

-// issuerMatches reports whether iss is the configured issuer or a sub-path of
-// it. Zitadel emits exactly "https://zitadel.riotpiao.com"; providers that
-// mount an application path append a suffix (Authentik used
-// "/application/o/<app>/"). A trailing slash on either side is ignored.
-//
-// An empty configured issuer disables the domain check; the JWKS signature
-// check still applies. Production configuration always sets an issuer.
-func issuerMatches(iss, configured string) bool {
-	if configured == "" {
-		return iss != ""
-	}
-	iss = strings.TrimSuffix(iss, "/")
-	configured = strings.TrimSuffix(configured, "/")
-	return iss == configured || strings.HasPrefix(iss, configured+"/")
+// isValidIssuer checks if issuer is from Authentik (any provider/app).
+// Accepts: https://authentik.riotpiao.com/application/o/{provider}/
+func isValidIssuer(iss string) bool {
+	return strings.Contains(iss, "authentik.riotpiao.com/application/o/") &&
+		strings.HasSuffix(iss, "/")
 }

-// audienceValues normalises the aud claim. Zitadel access tokens carry an
-// array, older providers carry a single string.
-func audienceValues(claims jwt.MapClaims) ([]string, bool) {
-	switch v := claims["aud"].(type) {
-	case string:
-		return []string{v}, v != ""
-	case []string:
-		return v, len(v) > 0
-	case []interface{}:
-		out := make([]string, 0, len(v))
-		for _, item := range v {
-			if s, ok := item.(string); ok {
-				out = append(out, s)
-			}
-		}
-		return out, len(out) > 0
-	case jwt.ClaimStrings:
-		return []string(v), len(v) > 0
-	}
-	return nil, false
-}
-
-// Validator validates JWTs against the JWKS of an OIDC provider (Zitadel).
-// The issuer is configuration-driven: tokens whose iss claim equals the
-// configured issuer, or lives under it as a sub-path, are accepted.
-//
-// When a ZitadelVerifier is attached (see NewZitadelValidator), validation
-// delegates to the official ZITADEL Go SDK and this type only adapts the
-// result to jwt.MapClaims for the rest of the gateway.
+// Validator validates JWTs against Authentik JWKS.
+// Supports multi-issuer: any Authentik service account provider is accepted
+// (portfolio-agent, memory-agent, api-gw, etc.) because all share the same
+// JWKS signing key.
 type Validator struct {
-	issuer   string // Expected issuer (iss claim); empty disables the domain check
-	audience string // Expected audience; currently logged only (any non-empty aud is accepted)
+	issuer   string // Not used for validation (kept for logging); issuer regex check is sufficient
+	audience string // Not used for validation; any audience from valid Authentik issuer is accepted
 	jwksURL  string
 	jwks     *keyfunc.JWKS
 	mu       sync.Mutex
-	zitadel  *ZitadelVerifier // non-nil when backed by the ZITADEL SDK
-}
-
-// NewZitadelValidator builds a Validator backed by the ZITADEL Go SDK.
-// issuer is the expected iss claim, used for logging and for the legacy
-// claim checks if the SDK verifier is ever bypassed. clientIDs is the
-// audience allow-list. Discovery runs here, so a wrong domain fails at
-// startup.
-func NewZitadelValidator(ctx context.Context, issuer string, clientIDs []string, cfg ZitadelConfig) (*Validator, error) {
-	return NewZitadelValidatorWithClient(ctx, issuer, clientIDs, cfg, nil)
-}
-
-// NewZitadelValidatorWithClient is NewZitadelValidator with an injected HTTP
-// client, for tests that serve discovery and JWKS from httptest.
-func NewZitadelValidatorWithClient(ctx context.Context, issuer string, clientIDs []string, cfg ZitadelConfig, httpClient *gohttp.Client) (*Validator, error) {
-	cfg.ClientIDs = clientIDs
-	if cfg.HTTPClient == nil {
-		cfg.HTTPClient = httpClient
-	}
-	sdk, err := NewZitadelVerifier(ctx, cfg)
-	if err != nil {
-		return nil, err
-	}
-	return &Validator{
-		issuer:  issuer,
-		jwksURL: "", // unused on the SDK path
-		zitadel: sdk,
-	}, nil
-}
-
-// claimsFromIntrospection converts the SDK's verified claim set into
-// jwt.MapClaims so the rest of the gateway keeps one claim representation.
-// The SDK has already checked the signature, exp, nbf and iss; this only
-// copies values across.
-func claimsFromIntrospection(authCtx *oauth.IntrospectionContext) jwt.MapClaims {
-	claims := jwt.MapClaims{}
-	if authCtx == nil {
-		return claims
-	}
-	for k, v := range authCtx.Claims {
-		claims[k] = v
-	}
-	if authCtx.Issuer != "" {
-		claims["iss"] = authCtx.Issuer
-	}
-	if authCtx.Subject != "" {
-		claims["sub"] = authCtx.Subject
-	}
-	if len(authCtx.Audience) > 0 {
-		aud := make([]interface{}, 0, len(authCtx.Audience))
-		for _, a := range authCtx.Audience {
-			aud = append(aud, a)
-		}
-		claims["aud"] = aud
-	}
-	if authCtx.Expiration != 0 {
-		claims["exp"] = float64(authCtx.Expiration)
-	}
-	if authCtx.NotBefore != 0 {
-		claims["nbf"] = float64(authCtx.NotBefore)
-	}
-	if authCtx.IssuedAt != 0 {
-		claims["iat"] = float64(authCtx.IssuedAt)
-	}
-	if authCtx.ClientID != "" {
-		claims["client_id"] = authCtx.ClientID
-	}
-	return claims
 }

 // NewValidator creates a new JWT validator for a service.
-// issuer is the expected iss claim (exact value or prefix). audience is kept
-// for compatibility; only non-emptiness is enforced. JWKS fetching is lazy
-// (deferred until first validation).
+// issuer and audience params are deprecated (ignored for validation) but kept
+// for backward compatibility. Multi-issuer validation via isValidIssuer() is used instead.
+// JWKS fetching is lazy (deferred until first validation).
 func NewValidator(issuer, audience, jwksURL string) *Validator {
 	return &Validator{
 		issuer:   issuer,   // deprecated param, kept for compat
@@ -191,31 +87,8 @@ func (v *Validator) ValidateBearerToken(authHeader string) (jwt.MapClaims, error
 	}

 	// Ensure JWKS is loaded (lazy)
-	if v.jwks != nil || v.zitadel == nil {
-		if err := v.ensureJWKS(); err != nil {
-			return nil, err
-		}
-	}
-
-	// SDK path: ZITADEL Go SDK performs signature, expiry and issuer checks.
-	if v.zitadel != nil {
-		// ZITADEL issues two access-token shapes. A plain JWS has three
-		// dot-separated segments and can be verified locally. A JWE has five,
-		// is encrypted with the instance key, and only ZITADEL can read it.
-		// Report that case explicitly so the operator changes the client to
-		// OIDC_TOKEN_TYPE_BEARER instead of debugging a signature failure.
-		if strings.Count(tokenString, ".") == 4 {
-			return nil, fmt.Errorf("token is an encrypted JWE (5 segments); " +
-				"set accessTokenType=OIDC_TOKEN_TYPE_BEARER on the ZITADEL client so the gateway can verify it locally")
-		}
-
-		ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-		defer cancel()
-		authCtx, err := v.zitadel.CheckAuthorization(ctx, authHeader)
-		if err != nil {
-			return nil, fmt.Errorf("token validation failed: %v", err)
-		}
-		return claimsFromIntrospection(authCtx), nil
+	if err := v.ensureJWKS(); err != nil {
+		return nil, err
 	}

 	// Parse and validate
@@ -247,35 +120,34 @@ func (v *Validator) ValidateBearerToken(authHeader string) (jwt.MapClaims, error
 		}
 	}

-	// Check iss against the configured issuer (exact match or sub-path).
+	// Check iss (issuer) - accept any Authentik provider issuer
+	// (portfolio-agent, memory-agent, api-gw, etc.)
+	// All use same signing key so JWKS validation is sufficient
 	if iss, ok := claims["iss"].(string); !ok {
 		return nil, fmt.Errorf("missing issuer claim")
-	} else if !issuerMatches(iss, v.issuer) {
-		return nil, fmt.Errorf("invalid issuer: %s (want %s)", iss, v.issuer)
+	} else if !isValidIssuer(iss) {
+		return nil, fmt.Errorf("invalid issuer: %s", iss)
 	}

-	// Check aud - Zitadel sends an array; accept a string or array.
-	if auds, ok := audienceValues(claims); !ok {
+	// Check aud (audience) - accept any Authentik-provided audience
+	// since all Authentik service accounts use the same signing key.
+	// The issuer check above is sufficient to ensure JWT came from Authentik.
+	if aud, ok := claims["aud"].(string); !ok {
 		return nil, fmt.Errorf("missing audience claim")
-	} else if len(auds) == 0 {
+	} else if aud == "" {
 		return nil, fmt.Errorf("empty audience claim")
 	}
-	// The signature check plus the issuer check establish the trust boundary;
-	// the audience value itself is not hardcoded.
+	// Note: Not hardcoding expected audience. Any audience from a valid Authentik
+	// issuer is accepted, since all service accounts are under the same trust boundary.

 	return claims, nil
 }

 // CheckPermissions checks if claims contain required permission(s).
-// Understands three token shapes:
-//   - "permissions": []string            (generic / legacy tokens)
-//   - "roles": []string or map[string]…  (legacy service-account tokens)
-//   - "urn:zitadel:iam:org:project:roles": map[string]…  (Zitadel access tokens,
-//     where the map keys are role keys)
-//
+// Checks both "permissions" claim (for users) and "roles" claim (for service accounts).
 // Returns true if any required permission is found or wildcard "*" exists.
 func (v *Validator) CheckPermissions(claims jwt.MapClaims, required ...string) bool {
-	// 1. permissions claim (array of strings)
+	// Try permissions claim first (for user tokens)
 	if permsIface, ok := claims["permissions"]; ok {
 		if perms, ok := permsIface.([]interface{}); ok {
 			if v.checkPermList(perms, required...) {
@@ -284,45 +156,15 @@ func (v *Validator) CheckPermissions(claims jwt.MapClaims, required ...string) b
 		}
 	}

-	// 2. roles claim: array (legacy) or map keyed by role name
+	// Fall back to roles claim (for service account tokens)
 	if rolesIface, ok := claims["roles"]; ok {
-		switch roles := rolesIface.(type) {
-		case []interface{}:
+		if roles, ok := rolesIface.([]interface{}); ok {
 			if v.checkPermList(roles, required...) {
 				return true
 			}
-		case map[string]interface{}:
-			if v.checkPermMap(roles, required...) {
-				return true
-			}
 		}
 	}

-	// 3. Zitadel access-token role assertion: map keyed by role key
-	if rolesIface, ok := claims["urn:zitadel:iam:org:project:roles"]; ok {
-		if roles, ok := rolesIface.(map[string]interface{}); ok {
-			if v.checkPermMap(roles, required...) {
-				return true
-			}
-		}
-	}
-
-	return false
-}
-
-// checkPermMap reports whether any key in a role map matches a required
-// permission, or the map carries a wildcard key.
-func (v *Validator) checkPermMap(roles map[string]interface{}, required ...string) bool {
-	for roleKey := range roles {
-		if roleKey == "*" {
-			return true
-		}
-		for _, req := range required {
-			if roleKey == req {
-				return true
-			}
-		}
-	}
 	return false
 }

@@ -1,157 +0,0 @@
-package auth
-
-import (
-	"context"
-	"errors"
-	"fmt"
-	gohttp "net/http"
-	"slices"
-	"strings"
-
-	"github.com/zitadel/oidc/v3/pkg/client"
-	"github.com/zitadel/oidc/v3/pkg/client/rp"
-	"github.com/zitadel/oidc/v3/pkg/oidc"
-	"github.com/zitadel/oidc/v3/pkg/op"
-	"github.com/zitadel/zitadel-go/v3/pkg/authorization/oauth"
-	"github.com/zitadel/zitadel-go/v3/pkg/zitadel"
-)
-
-// ErrAudienceNotAllowed reports a token whose aud claim names none of the
-// configured resource-server client IDs.
-var ErrAudienceNotAllowed = errors.New("audience not allowed")
-
-// ZitadelConfig configures the ZITADEL-backed verifier.
-type ZitadelConfig struct {
-	// Domain is the ZITADEL instance host, for example "zitadel.riotpiao.com".
-	Domain string
-	// InsecurePort, when non-empty, speaks plain HTTP to that port instead of
-	// HTTPS on 443. Use only for a local port-forward.
-	InsecurePort string
-	// ClientIDs is the audience allow-list: a token is accepted when its aud
-	// claim contains at least one of these client IDs. Empty accepts any
-	// non-empty audience, which trusts every client of the same instance.
-	ClientIDs []string
-	// HTTPClient overrides the client used for discovery and JWKS fetches.
-	HTTPClient *gohttp.Client
-}
-
-// ZitadelVerifier validates ZITADEL access tokens with the official
-// github.com/zitadel/zitadel-go SDK.
-//
-// It reuses the SDK's building blocks rather than reimplementing JWT parsing:
-//   - zitadel.New + Origin to address the instance
-//   - client.Discover for OIDC discovery (issuer and JWKS URI)
-//   - rp.NewRemoteKeySet for cached, auto-refreshing signing keys
-//   - op.NewAccessTokenVerifier / op.VerifyAccessToken for signature, expiry
-//     and issuer checks
-//   - oauth.IntrospectionContext for the verified claim set and its
-//     IsGrantedRole helper, which already understands the
-//     urn:zitadel:iam:org:project:roles claim
-//
-// The only behaviour added on top of the SDK is a multi-value audience
-// allow-list, because this gateway is a shared resource server: callers hold
-// their own client IDs (paperless, poimen, api-gw, ...) and the SDK's
-// oauth.WithJWT helper accepts exactly one.
-type ZitadelVerifier struct {
-	instance  *zitadel.Zitadel
-	verifier  *op.AccessTokenVerifier
-	clientIDs []string
-}
-
-// NewZitadelVerifier performs OIDC discovery and returns a ready verifier.
-// Discovery happens once at startup so a misconfigured domain fails fast
-// instead of on the first request.
-func NewZitadelVerifier(ctx context.Context, cfg ZitadelConfig) (*ZitadelVerifier, error) {
-	if cfg.Domain == "" {
-		return nil, errors.New("zitadel: domain is required")
-	}
-
-	httpClient := cfg.HTTPClient
-	if httpClient == nil {
-		httpClient = gohttp.DefaultClient
-	}
-
-	var opts []zitadel.Option
-	if cfg.InsecurePort != "" {
-		opts = append(opts, zitadel.WithInsecure(cfg.InsecurePort))
-	}
-	instance := zitadel.New(cfg.Domain, opts...)
-
-	discovery, err := client.Discover(ctx, instance.Origin(), httpClient)
-	if err != nil {
-		return nil, fmt.Errorf("zitadel: OIDC discovery at %s: %w", instance.Origin(), err)
-	}
-
-	keySet := rp.NewRemoteKeySet(httpClient, discovery.JwksURI)
-	verifier := op.NewAccessTokenVerifier(discovery.Issuer, keySet)
-
-	return &ZitadelVerifier{
-		instance:  instance,
-		verifier:  verifier,
-		clientIDs: append([]string(nil), cfg.ClientIDs...),
-	}, nil
-}
-
-// Origin returns the instance origin used for discovery, for logging.
-func (z *ZitadelVerifier) Origin() string {
-	if z == nil || z.instance == nil {
-		return ""
-	}
-	return z.instance.Origin()
-}
-
-// CheckAuthorization validates an "Authorization: Bearer <token>" header and
-// returns the SDK's verified introspection context.
-func (z *ZitadelVerifier) CheckAuthorization(ctx context.Context, authHeader string) (*oauth.IntrospectionContext, error) {
-	accessToken, ok := strings.CutPrefix(authHeader, oidc.BearerToken)
-	if !ok {
-		return nil, fmt.Errorf("%w: expected %q prefix", oauth.ErrInvalidAuthorizationHeader, oidc.BearerToken)
-	}
-	accessToken = strings.TrimSpace(accessToken)
-	if accessToken == "" {
-		return nil, fmt.Errorf("%w: empty bearer token", oauth.ErrInvalidAuthorizationHeader)
-	}
-
-	claims, err := op.VerifyAccessToken[*oidc.AccessTokenClaims](ctx, accessToken, z.verifier)
-	if err != nil {
-		return nil, fmt.Errorf("%w: %v", oauth.ErrInvalidToken, err)
-	}
-
-	if len(claims.Audience) == 0 {
-		return nil, fmt.Errorf("%w: empty aud", oauth.ErrInvalidToken)
-	}
-	if !z.audienceAllowed(claims.Audience) {
-		return nil, fmt.Errorf("%w: %v", ErrAudienceNotAllowed, claims.Audience)
-	}
-
-	authCtx := &oauth.IntrospectionContext{
-		IntrospectionResponse: oidc.IntrospectionResponse{
-			Active:     true,
-			Issuer:     claims.Issuer,
-			Subject:    claims.Subject,
-			Audience:   claims.Audience,
-			Expiration: claims.Expiration,
-			IssuedAt:   claims.IssuedAt,
-			NotBefore:  claims.NotBefore,
-			ClientID:   claims.ClientID,
-			JWTID:      claims.JWTID,
-			Claims:     claims.Claims,
-		},
-	}
-	authCtx.SetToken(accessToken)
-	return authCtx, nil
-}
-
-// audienceAllowed reports whether any aud entry is in the allow-list.
-// An empty allow-list accepts every non-empty audience.
-func (z *ZitadelVerifier) audienceAllowed(audience []string) bool {
-	if len(z.clientIDs) == 0 {
-		return true
-	}
-	for _, aud := range audience {
-		if slices.Contains(z.clientIDs, aud) {
-			return true
-		}
-	}
-	return false
-}
@@ -1,294 +0,0 @@
-package auth
-
-import (
-	"context"
-	"crypto/rand"
-	"crypto/rsa"
-	"encoding/base64"
-	"encoding/json"
-	"fmt"
-	"math/big"
-	"net/http"
-	"net/http/httptest"
-	"strings"
-	"testing"
-	"time"
-
-	"github.com/golang-jwt/jwt/v5"
-)
-
-// testIssuer mimics one ZITADEL instance: OIDC discovery on
-// /.well-known/openid-configuration and the JWKS on /oauth/v2/keys.
-type testIssuer struct {
-	server  *httptest.Server
-	key     *rsa.PrivateKey
-	kid     string
-	jwksHit int
-}
-
-func newTestIssuer(t *testing.T) *testIssuer {
-	t.Helper()
-
-	key, err := rsa.GenerateKey(rand.Reader, 2048)
-	if err != nil {
-		t.Fatalf("generate key: %v", err)
-	}
-
-	ti := &testIssuer{key: key, kid: "test-kid-1"}
-
-	mux := http.NewServeMux()
-	mux.HandleFunc("/.well-known/openid-configuration", func(w http.ResponseWriter, r *http.Request) {
-		_ = json.NewEncoder(w).Encode(map[string]any{
-			"issuer":   ti.server.URL,
-			"jwks_uri": ti.server.URL + "/oauth/v2/keys",
-		})
-	})
-	mux.HandleFunc("/oauth/v2/keys", func(w http.ResponseWriter, r *http.Request) {
-		ti.jwksHit++
-		pub := key.PublicKey
-		_ = json.NewEncoder(w).Encode(map[string]any{
-			"keys": []map[string]any{{
-				"kty": "RSA",
-				"kid": ti.kid,
-				"use": "sig",
-				"alg": "RS256",
-				"n":   base64.RawURLEncoding.EncodeToString(pub.N.Bytes()),
-				"e":   base64.RawURLEncoding.EncodeToString(big.NewInt(int64(pub.E)).Bytes()),
-			}},
-		})
-	})
-
-	ti.server = httptest.NewServer(mux)
-	t.Cleanup(ti.server.Close)
-	return ti
-}
-
-// sign mints an access token with the given audience and project roles.
-func (ti *testIssuer) sign(t *testing.T, aud []string, roles map[string]any) string {
-	t.Helper()
-
-	claims := jwt.MapClaims{
-		"iss": ti.server.URL,
-		"sub": "396757800000000001",
-		"exp": time.Now().Add(time.Hour).Unix(),
-		"iat": time.Now().Unix(),
-	}
-	if len(aud) > 0 {
-		audClaim := make([]interface{}, 0, len(aud))
-		for _, a := range aud {
-			audClaim = append(audClaim, a)
-		}
-		claims["aud"] = audClaim
-	}
-	if roles != nil {
-		claims["urn:zitadel:iam:org:project:roles"] = roles
-	}
-
-	tok := jwt.NewWithClaims(jwt.SigningMethodRS256, claims)
-	tok.Header["kid"] = ti.kid
-	signed, err := tok.SignedString(ti.key)
-	if err != nil {
-		t.Fatalf("sign token: %v", err)
-	}
-	return signed
-}
-
-// verifierFor builds a ZitadelVerifier pointed at the httptest instance.
-// The SDK is given the loopback host and its port, so discovery goes to the
-// test server rather than a real domain.
-func verifierFor(t *testing.T, ti *testIssuer, clientIDs []string) *ZitadelVerifier {
-	t.Helper()
-
-	host, port, ok := strings.Cut(strings.TrimPrefix(ti.server.URL, "http://"), ":")
-	if !ok {
-		t.Fatalf("unexpected test server URL %q", ti.server.URL)
-	}
-
-	v, err := NewZitadelVerifier(context.Background(), ZitadelConfig{
-		Domain:       host,
-		InsecurePort: port,
-		ClientIDs:    clientIDs,
-		HTTPClient:   ti.server.Client(),
-	})
-	if err != nil {
-		t.Fatalf("NewZitadelVerifier: %v", err)
-	}
-	return v
-}
-
-func TestZitadelVerifier_AcceptsTokenWithAllowedAudience(t *testing.T) {
-	ti := newTestIssuer(t)
-	v := verifierFor(t, ti, []string{"396757803559027071", "396757800874672511"})
-
-	token := ti.sign(t, []string{"396757803559027071"}, map[string]any{
-		"llm:inference": map[string]any{"396757796898471937": "homelab"},
-	})
-
-	authCtx, err := v.CheckAuthorization(context.Background(), "Bearer "+token)
-	if err != nil {
-		t.Fatalf("expected token to pass, got %v", err)
-	}
-	if !authCtx.IsAuthorized() {
-		t.Error("expected IsAuthorized() true")
-	}
-	if want := "396757800000000001"; authCtx.UserID() != want {
-		t.Errorf("UserID() = %q, want %q", authCtx.UserID(), want)
-	}
-	if !authCtx.IsGrantedRole("llm:inference") {
-		t.Error("expected SDK IsGrantedRole to see llm:inference from the project roles claim")
-	}
-	if ti.jwksHit == 0 {
-		t.Error("expected the SDK to fetch the JWKS")
-	}
-}
-
-func TestZitadelVerifier_RejectsForeignAudience(t *testing.T) {
-	ti := newTestIssuer(t)
-	v := verifierFor(t, ti, []string{"396757803559027071"})
-
-	token := ti.sign(t, []string{"999999999999999999"}, nil)
-
-	_, err := v.CheckAuthorization(context.Background(), "Bearer "+token)
-	if err == nil {
-		t.Fatal("expected rejection for an audience outside the allow-list")
-	}
-	if !strings.Contains(err.Error(), ErrAudienceNotAllowed.Error()) {
-		t.Errorf("expected ErrAudienceNotAllowed, got %v", err)
-	}
-}
-
-func TestZitadelVerifier_AcceptsAnyAudienceWhenAllowListEmpty(t *testing.T) {
-	ti := newTestIssuer(t)
-	v := verifierFor(t, ti, nil)
-
-	token := ti.sign(t, []string{"396757803877794175"}, nil)
-
-	if _, err := v.CheckAuthorization(context.Background(), "Bearer "+token); err != nil {
-		t.Fatalf("empty allow-list should accept any audience, got %v", err)
-	}
-}
-
-func TestZitadelVerifier_RejectsMissingBearerPrefix(t *testing.T) {
-	ti := newTestIssuer(t)
-	v := verifierFor(t, ti, nil)
-
-	token := ti.sign(t, []string{"396757803877794175"}, nil)
-
-	if _, err := v.CheckAuthorization(context.Background(), token); err == nil {
-		t.Fatal("expected rejection when the Bearer prefix is missing")
-	}
-}
-
-func TestZitadelVerifier_RejectsExpiredToken(t *testing.T) {
-	ti := newTestIssuer(t)
-	v := verifierFor(t, ti, nil)
-
-	claims := jwt.MapClaims{
-		"iss": ti.server.URL,
-		"sub": "u1",
-		"aud": []interface{}{"396757803877794175"},
-		"exp": time.Now().Add(-2 * time.Hour).Unix(),
-		"iat": time.Now().Add(-3 * time.Hour).Unix(),
-	}
-	tok := jwt.NewWithClaims(jwt.SigningMethodRS256, claims)
-	tok.Header["kid"] = ti.kid
-	signed, err := tok.SignedString(ti.key)
-	if err != nil {
-		t.Fatalf("sign token: %v", err)
-	}
-
-	if _, err := v.CheckAuthorization(context.Background(), "Bearer "+signed); err == nil {
-		t.Fatal("expected rejection for an expired token")
-	}
-}
-
-// TestValidator_SDKPathFeedsCheckPermissions proves the adapter from the SDK
-// introspection context to jwt.MapClaims keeps the capability gate working.
-func TestValidator_SDKPathFeedsCheckPermissions(t *testing.T) {
-	ti := newTestIssuer(t)
-	host, port, _ := strings.Cut(strings.TrimPrefix(ti.server.URL, "http://"), ":")
-
-	validator, err := NewZitadelValidator(context.Background(), ti.server.URL, []string{"396757803559027071"}, ZitadelConfig{
-		Domain:       host,
-		InsecurePort: port,
-		HTTPClient:   ti.server.Client(),
-	})
-	if err != nil {
-		t.Fatalf("NewZitadelValidator: %v", err)
-	}
-
-	good := ti.sign(t, []string{"396757803559027071"}, map[string]any{
-		"llm:inference": map[string]any{"396757796898471937": "homelab"},
-	})
-	claims, err := validator.ValidateBearerToken("Bearer " + good)
-	if err != nil {
-		t.Fatalf("ValidateBearerToken: %v", err)
-	}
-	if !validator.CheckPermissions(claims, "llm:inference") {
-		t.Error("expected CheckPermissions to accept llm:inference via the SDK path")
-	}
-	if validator.CheckPermissions(claims, "sqs:write") {
-		t.Error("expected CheckPermissions to reject a role the token does not carry")
-	}
-
-	noRole := ti.sign(t, []string{"396757803559027071"}, nil)
-	claims, err = validator.ValidateBearerToken("Bearer " + noRole)
-	if err != nil {
-		t.Fatalf("ValidateBearerToken without roles: %v", err)
-	}
-	if validator.CheckPermissions(claims, "llm:inference") {
-		t.Error("expected CheckPermissions to reject a token with no project roles")
-	}
-}
-
-func TestIssuerMatches(t *testing.T) {
-	cases := []struct {
-		iss        string
-		configured string
-		want       bool
-	}{
-		{"https://zitadel.riotpiao.com", "https://zitadel.riotpiao.com", true},
-		{"https://zitadel.riotpiao.com/", "https://zitadel.riotpiao.com", true},
-		{"https://zitadel.riotpiao.com/oauth/v2", "https://zitadel.riotpiao.com", true},
-		{"https://evil.example.com", "https://zitadel.riotpiao.com", false},
-		{"https://zitadel.riotpiao.com.evil.com", "https://zitadel.riotpiao.com", false},
-		{"anything", "", true},
-	}
-	for _, c := range cases {
-		if got := issuerMatches(c.iss, c.configured); got != c.want {
-			t.Errorf("issuerMatches(%q, %q) = %v, want %v", c.iss, c.configured, got, c.want)
-		}
-	}
-}
-
-func TestClaimsFromIntrospection_Nil(t *testing.T) {
-	if got := claimsFromIntrospection(nil); got == nil || len(got) != 0 {
-		t.Errorf("expected empty non-nil claims, got %v", got)
-	}
-}
-
-func TestAudienceValues(t *testing.T) {
-	cases := []struct {
-		name  string
-		claim any
-		want  int
-		ok    bool
-	}{
-		{"string", "api-gw", 1, true},
-		{"empty string", "", 1, false},
-		{"string slice", []string{"a", "b"}, 2, true},
-		{"iface slice", []interface{}{"a", 5}, 1, true},
-		{"missing", nil, 0, false},
-	}
-	for _, c := range cases {
-		claims := jwt.MapClaims{}
-		if c.claim != nil {
-			claims["aud"] = c.claim
-		}
-		got, ok := audienceValues(claims)
-		if ok != c.ok || len(got) != c.want {
-			t.Errorf("%s: audienceValues = (%v, %v), want len %d ok %v", c.name, got, ok, c.want, c.ok)
-		}
-	}
-	_ = fmt.Sprint()
-}
@@ -66,23 +66,12 @@ type AuthConfig struct {
 	JWKSURL string
 	// RequiredCapability is the permission required for LLM inference (e.g., "llm:inference").
 	RequiredCapability string
-	// TokenURL is the OIDC token endpoint used for token exchange
-	// (client_credentials / password grants).
+	// TokenURL is the Authentik token endpoint for password/refresh grants.
 	TokenURL string
 	// ClientID is the OAuth2 client ID for token exchange.
 	ClientID string
 	// ClientSecret is the OAuth2 client secret (loaded from env, never from config file).
 	ClientSecret string
-	// ZitadelDomain is the ZITADEL instance host (for example
-	// "zitadel.riotpiao.com"). When set, token validation uses the official
-	// ZITADEL Go SDK instead of the raw JWKS path.
-	ZitadelDomain string
-	// ZitadelInsecurePort, when non-empty, makes the SDK speak plain HTTP to
-	// that port. Only for a local port-forward.
-	ZitadelInsecurePort string
-	// ZitadelClientIDs is the audience allow-list for the SDK verifier. Empty
-	// accepts any audience issued by the same instance.
-	ZitadelClientIDs []string
 }

 // Route represents a single route and its upstream configuration.
@@ -20,16 +20,13 @@ type rawConfig struct {

 // rawAuth represents auth configuration in YAML.
 type rawAuth struct {
-	Enabled             bool     `yaml:"enabled"`
-	Issuer              string   `yaml:"issuer"`
-	Audience            string   `yaml:"audience"`
-	JWKSURL             string   `yaml:"jwksUrl"`
-	RequiredCapability  string   `yaml:"requiredCapability"`
-	TokenURL            string   `yaml:"tokenUrl"`
-	ClientID            string   `yaml:"clientId"`
-	ZitadelDomain       string   `yaml:"zitadelDomain"`
-	ZitadelInsecurePort string   `yaml:"zitadelInsecurePort"`
-	ZitadelClientIDs    []string `yaml:"zitadelClientIds"`
+	Enabled            bool   `yaml:"enabled"`
+	Issuer             string `yaml:"issuer"`
+	Audience           string `yaml:"audience"`
+	JWKSURL            string `yaml:"jwksUrl"`
+	RequiredCapability string `yaml:"requiredCapability"`
+	TokenURL           string `yaml:"tokenUrl"`
+	ClientID           string `yaml:"clientId"`
 }

 // rawRoute represents a single route in the YAML configuration.
@@ -187,17 +184,14 @@ func LoadRoutesAndModelsFromFile(path string) (map[string]*Route, map[string]*Mo

 	// Parse auth config
 	authConfig := AuthConfig{
-		Enabled:             raw.Auth.Enabled,
-		Issuer:              raw.Auth.Issuer,
-		Audience:            raw.Auth.Audience,
-		JWKSURL:             raw.Auth.JWKSURL,
-		RequiredCapability:  raw.Auth.RequiredCapability,
-		TokenURL:            raw.Auth.TokenURL,
-		ClientID:            raw.Auth.ClientID,
-		ClientSecret:        os.Getenv("AUTH_CLIENT_SECRET"),
-		ZitadelDomain:       raw.Auth.ZitadelDomain,
-		ZitadelInsecurePort: raw.Auth.ZitadelInsecurePort,
-		ZitadelClientIDs:    raw.Auth.ZitadelClientIDs,
+		Enabled:            raw.Auth.Enabled,
+		Issuer:             raw.Auth.Issuer,
+		Audience:           raw.Auth.Audience,
+		JWKSURL:            raw.Auth.JWKSURL,
+		RequiredCapability: raw.Auth.RequiredCapability,
+		TokenURL:           raw.Auth.TokenURL,
+		ClientID:           raw.Auth.ClientID,
+		ClientSecret:       os.Getenv("AUTH_CLIENT_SECRET"),
 	}

 	return routes, models, adapters, authConfig, nil
@@ -98,29 +98,8 @@ func New(cfg *config.Config) *Handler {
 		defaultMaxBodySize:    100 * 1024 * 1024,
 	}

-	// Initialize JWT validator if auth is enabled.
-	//
-	// Two paths, in order of preference:
-	//  1. ZitadelDomain set -> official ZITADEL Go SDK (discovery + remote JWKS
-	//     + op.AccessTokenVerifier). This is the production path.
-	//  2. JWKSURL set -> legacy raw-JWKS validation, kept for tests and for
-	//     any deployment that has not moved to ZITADEL yet.
-	if cfg.Auth.Enabled && cfg.Auth.ZitadelDomain != "" {
-		ctx, cancel := context.WithTimeout(context.Background(), 30*time.Second)
-		defer cancel()
-		v, err := auth.NewZitadelValidator(ctx, cfg.Auth.Issuer, cfg.Auth.ZitadelClientIDs, auth.ZitadelConfig{
-			Domain:       cfg.Auth.ZitadelDomain,
-			InsecurePort: cfg.Auth.ZitadelInsecurePort,
-			ClientIDs:    cfg.Auth.ZitadelClientIDs,
-		})
-		if err != nil {
-			logging.Errorf("zitadel: verifier init failed; auth disabled until restart", err, map[string]string{
-				"domain": cfg.Auth.ZitadelDomain,
-			})
-		} else {
-			h.jwtValidator = v
-		}
-	} else if cfg.Auth.Enabled && cfg.Auth.JWKSURL != "" {
+	// Initialize JWT validator if auth is enabled
+	if cfg.Auth.Enabled && cfg.Auth.JWKSURL != "" {
 		h.jwtValidator = auth.NewValidator(
 			cfg.Auth.Issuer,
 			cfg.Auth.Audience,
@@ -743,17 +722,16 @@ func (h *Handler) authenticateV1(w http.ResponseWriter, r *http.Request) bool {
 		return false
 	}
 	identity.Inject(r, claims)
-	if caps := requiredCapabilities(h.config.Auth.RequiredCapability); len(caps) > 0 {
-		args := append(caps, "*")
-		if !h.jwtValidator.CheckPermissions(claims, args...) {
+	if h.config.Auth.RequiredCapability != "" {
+		if !h.jwtValidator.CheckPermissions(claims, h.config.Auth.RequiredCapability, "*") {
 			writeProblemDetail(w, http.StatusForbidden,
 				"https://api.example.com/problems/insufficient-permissions",
 				"Insufficient Permissions",
-				fmt.Sprintf("Required capability: %s", strings.Join(caps, " or ")),
+				fmt.Sprintf("Required capability: %s", h.config.Auth.RequiredCapability),
 				nil)
 			logging.Errorf("auth failed", fmt.Errorf("insufficient permissions"), map[string]string{
 				"path":     r.URL.Path,
-				"required": strings.Join(caps, ","),
+				"required": h.config.Auth.RequiredCapability,
 			})
 			return false
 		}
@@ -761,22 +739,6 @@ func (h *Handler) authenticateV1(w http.ResponseWriter, r *http.Request) bool {
 	return true
 }

-// requiredCapabilities splits the comma-separated requiredCapability setting
-// into a list. Several spellings are accepted because the same capability is
-// named differently by each identity provider: Authentik used "llm:inference"
-// as a permission string, while ZITADEL uses role keys such as
-// "llm-inference". A token satisfies the gate when it carries any one of them.
-func requiredCapabilities(raw string) []string {
-	parts := strings.Split(raw, ",")
-	out := make([]string, 0, len(parts))
-	for _, p := range parts {
-		if p = strings.TrimSpace(p); p != "" {
-			out = append(out, p)
-		}
-	}
-	return out
-}
-
 // enforceBodyLimit caps request body size for non-chat routes.
 // Returns false when the request is rejected (response already written).
 func (h *Handler) enforceBodyLimit(w http.ResponseWriter, r *http.Request, route *Route) bool {
@@ -835,14 +797,8 @@ func (h *Handler) setupCAGProxy(ctx context.Context, w http.ResponseWriter, r *h
 			if resp.StatusCode < http.StatusBadRequest {
 				return nil
 			}
-			body, readErr := io.ReadAll(io.LimitReader(resp.Body, 4096))
+			body, _ := io.ReadAll(io.LimitReader(resp.Body, 4096))
 			resp.Body.Close()
-			if readErr != nil {
-				logging.Warnf("reading upstream error body failed", map[string]string{
-					"upstream": capturedRoute.Upstream.Address,
-					"err":      readErr.Error(),
-				})
-			}
 			logging.Warnf("upstream error after parked stream started", map[string]string{
 				"upstream": capturedRoute.Upstream.Address,
 				"status":   strconv.Itoa(resp.StatusCode),
@@ -99,21 +99,19 @@ spec:
    ports:
    - protocol: TCP
      port: 8080
-  # Allow to IAM (Zitadel) for JWKS fetch + API
-  # Zitadel API listens on 8080 (h2c/plaintext in-cluster); the login UI
-  # listens on 3000. Public access goes through ingress-nginx on 443 (already
-  # covered by the 10.96.0.0/12 CIDR below), but 8080 keeps direct in-cluster
-  # calls working if CoreDNS rewrite is bypassed.
-  # NetworkPolicy matches destination pod port, not service port.
+  # Allow to IAM (Authentik) for JWKS fetch + API
+  # Authentik pod listens on 9000 (http) and 9443 (https)
+  # Service translates 80→9000, 443→9443
+  # NetworkPolicy matches destination pod port, not service port
  - to:
    - namespaceSelector:
        matchLabels:
          kubernetes.io/metadata.name: iam
    ports:
    - protocol: TCP
-      port: 8080
+      port: 9000
    - protocol: TCP
-      port: 3000
+      port: 9443
  # Allow to SQS (queue management service)
  - to:
    - namespaceSelector: