fix: eliminate per-request global mutexes in Yaegi hot paths

The v1.0.14 fix replaced one contended sync.RWMutex (RefreshCoordinator.
refreshMutex) with sync.Map. Production showed the same death-spiral
signature recurring ~2 hours later — same shape, different mutex:
65 goroutines stuck on a sync.(*RWMutex).Lock at one address, pod
pinned at 1000m CPU, identical Yaegi runCfg/reflect.Value.Call stack
pattern. The mutex was RefreshCoordinator.attemptsMutex.

Generalising: under Yaegi (interpreted Go for traefik plugins), any
per-request global mutex acquisition is a latent serialization point.
reflect.Value.Call dispatch on a held lock turns a microsecond
critical section into a multi-millisecond one, and on a GOMAXPROCS=1
pod the queue is unbounded.

This commit removes every per-request global mutex on the hot path:

1. RefreshCoordinator.attemptsMutex (sync.RWMutex)
   sessionRefreshAttempts: map -> sync.Map.
   refreshAttemptTracker: all fields atomic (int32, int64 UnixNano,
   cooldownEndNano == 0 as the not-in-cooldown sentinel, replacing
   the inCooldown bool).
   isInCooldown / recordRefreshAttempt / recordRefreshSuccess /
   recordRefreshFailure all become lock-free. Cooldown entry uses
   CompareAndSwapInt64 so only one goroutine logs the transition.

2. RefreshCircuitBreaker.mutex (sync.RWMutex)
   lastFailureTime / lastSuccessTime -> atomic.Int64 UnixNano.
   state and failures already atomic.
   AllowRequest / RecordSuccess / RecordFailure now pure atomic ops.

3. TraefikOidc.firstRequestMutex (sync.Mutex)
   firstRequestReceived bool -> firstRequestStarted int32.
   metadataRefreshStarted bool -> metadataRefreshStartedAtomic int32.
   ServeHTTP bootstrap path uses CompareAndSwapInt32 — fires once,
   zero steady-state cost. Previously the mutex was acquired on
   every non-health request forever.

4. TraefikOidc.metadataRetryMutex (sync.Mutex)
   lastMetadataRetryTime time.Time -> lastMetadataRetryNano int64.
   The 30-second retry throttle is now a CAS on lastMetadataRetryNano.

cleanupStaleEntries iterates via sync.Map.Range; eviction is a
CompareAndDelete by pointer identity so a tracker freshly re-used by
a concurrent caller is not lost.

Empirical evidence (3 specialist-agent analysis of the v1.0.14 spike,
profiles in /tmp/traefik-spike-1779511683/):
  * mutex profile: 97% delay in sync.(*Mutex).Unlock via
    HTTPHandlerSwitcher -> accesslog -> metrics -> backoff.RetryNotify
  * 65 stuck goroutines at one RWMutex address (0x40022eb648),
    identical Yaegi CFG pointer, all on rc.attemptsMutex via
    recordRefreshAttempt + isInCooldown
  * traffic driver: long-lived in-cluster Go-http-client doing
    ~5.4 req/s POST embeddings via OIDC cookie session → same
    sessionID → contention all funnels to one tracker entry

Yaegi support for sync/atomic confirmed at
github.com/traefik/yaegi@v0.16.1/stdlib/go1_22_sync_atomic.go:
AddInt32/Int64, LoadInt32/Int64, StoreInt32/Int64,
CompareAndSwapInt32/Int64 all exposed via reflect.ValueOf. Yaegi
dispatches each call through reflect.Value.Call to the COMPILED
atomic.* function, which executes a single hardware CAS/LOCK-XADD
instruction. Each atomic op still pays Yaegi dispatch cost but
cannot block — no queueing, no death spiral.

Trade-off acknowledged: v1.0.15 issues ~6-8 atomic/sync.Map ops per
leader-path request vs the 4 mutex ops of v1.0.14. Under low
contention this is a modest CPU bump. Under high contention it's
an unbounded → bounded transformation. Net win.

All tests pass with -race; golangci-lint clean.

.github/FUNDING.yml

@@ -0,0 +1,15 @@
+# These are supported funding model platforms
+
+github: lukaszraczylo
+patreon: # Replace with a single Patreon username
+open_collective: # Replace with a single Open Collective username
+ko_fi: # Replace with a single Ko-fi username
+tidelift: # Replace with a single Tidelift platform-name/package-name e.g., npm/babel
+community_bridge: # Replace with a single Community Bridge project-name e.g., cloud-foundry
+liberapay: # Replace with a single Liberapay username
+issuehunt: # Replace with a single IssueHunt username
+lfx_crowdfunding: # Replace with a single LFX Crowdfunding project-name e.g., cloud-foundry
+polar: # Replace with a single Polar username
+buy_me_a_coffee: # Replace with a single Buy Me a Coffee username
+thanks_dev: # Replace with a single thanks.dev username
+custom: https://monzo.me/lukaszraczylo

.traefik.yml

@@ -80,3 +80,25 @@ testData:
   #   address: redis:6379
   #   password: urn:k8s:secret:redis:password
   #   cacheMode: hybrid
+
+  # Optional: bearer-token authentication for M2M (machine-to-machine) API
+  # clients. Default off. When enabled, requests presenting
+  # "Authorization: Bearer <jwt>" are validated against the configured OIDC
+  # provider (signature/issuer/audience/exp) and forwarded without creating
+  # a cookie session. The bearer path REJECTS ID tokens, requires a non-
+  # default audience, and never trusts the `email` claim as the identifier.
+  # See docs/BEARER_AUTH.md for the full threat model.
+  #
+  # enableBearerAuth: true                # opt-in
+  # audience: https://api.example.com     # REQUIRED when bearer is enabled
+  # bearerIdentifierClaim: sub            # default; used as X-Forwarded-User. `email` is rejected.
+  # stripAuthorizationHeader: true        # default; drops the raw token before forwarding
+  # bearerEmitWWWAuthenticate: true       # default; RFC 6750 hint on 401s
+  # bearerOverridesCookie: false          # default; cookie wins when both are present
+  # requireTokenIntrospection: false      # opt-in; calls RFC 7662 introspection per request
+  # maxTokenAgeSeconds: 86400             # 24h cap on iat (rejects clock-skew/forever tokens)
+  # maxIdentifierLength: 256              # cap on the sanitised principal identifier
+  # bearerFailureThreshold: 20            # consecutive 401s/IP that trip the throttle
+  # bearerFailureWindowSeconds: 60        # rolling window over which 401s are counted
+  # bearerFailurePenaltySeconds: 60       # 429 + Retry-After duration after threshold trips
+

README.md

@@ -9,6 +9,7 @@ manages sessions, and forwards user identity to downstream services.
 - [Configuration reference](docs/CONFIGURATION.md) — every parameter
 - [Provider guide](docs/PROVIDERS.md) — Google, Azure, Auth0, Okta, Keycloak, Cognito, GitLab, GitHub, generic
 - [Auth0 audience guide](docs/AUTH0_AUDIENCE_GUIDE.md) — custom APIs, opaque tokens, token confusion
+- [Bearer-token (M2M) auth](docs/BEARER_AUTH.md) — opt-in `Authorization: Bearer` path, threat model
 - [Redis cache](docs/REDIS.md) — multi-replica deployments
 - [Dynamic Client Registration](docs/DCR.md) — RFC 7591
 - [Development](docs/DEVELOPMENT.md) · [Testing](docs/TESTING.md)
@@ -171,6 +172,92 @@ Each instance must use a unique `cookiePrefix` **and** `sessionEncryptionKey`,
 otherwise a session minted by one instance can grant access through another.
 See [issue #87](https://github.com/lukaszraczylo/traefikoidc/issues/87).
 
+### Bearer-token (M2M) authentication
+
+Opt-in path for API clients that present `Authorization: Bearer <jwt>` instead
+of logging in via the browser flow. Default off. When enabled, the middleware
+validates the bearer JWT against the configured OIDC provider (signature,
+issuer, audience, expiry) and forwards the request downstream with the
+principal headers — no cookie session is created.
+
+```yaml
+enableBearerAuth: true
+audience: https://api.example.com   # REQUIRED when bearer is enabled
+# optional, defaults shown:
+bearerIdentifierClaim: sub          # claim used as X-Forwarded-User
+stripAuthorizationHeader: true      # drop the raw token before forwarding
+bearerEmitWWWAuthenticate: true     # RFC 6750 hint on 401s
+bearerOverridesCookie: false        # cookie wins when both are present (safer)
+maxTokenAgeSeconds: 86400           # 24h cap on iat
+bearerFailureThreshold: 20          # consecutive 401s/IP before 429 throttle
+```
+
+Hardening built in by default:
+
+- **Audience required.** Startup fails if `enableBearerAuth=true` and
+  `audience` is unset. Eliminates the "token issued for service B accepted
+  by A" confusion vector.
+- **ID tokens explicitly rejected.** Bearer is access-token-only. ID tokens
+  (detected via `nonce`, `typ: at+jwt`, `token_use`, `scope`, or audience
+  shape) return `401`.
+- **`alg` and `kid` pinned at the entrypoint.** Asymmetric-only allowlist
+  (`RS256/384/512`, `PS256/384/512`, `ES256/384/512`); `kid` length and
+  charset capped — both checked **before** any JWKS fetch so attacker noise
+  can't amplify into upstream calls.
+- **Identifier sanitised.** Default identifier source is `sub`; `email` is
+  rejected unless explicitly opted in (which the middleware still refuses to
+  avoid the unverified-email spoofing footgun). Control characters, bidi-
+  override codepoints, and the delimiters `, ; =` are all rejected before
+  the value reaches `X-Forwarded-User`.
+- **Multi-audience tokens require `azp`.** When `aud` is an array of more
+  than one element, the token must carry `azp == clientID`.
+- **`iat` upper-age bound.** Tokens older than `maxTokenAgeSeconds` are
+  rejected even if `exp` is far in the future.
+- **Per-IP 401 throttle.** After `bearerFailureThreshold` consecutive 401s
+  from one source IP, further bearer requests from that IP are rejected
+  with `429 Too Many Requests` + `Retry-After`.
+- **Cookie-wins by default.** When both a session cookie and an
+  `Authorization: Bearer` header arrive on the same request, the cookie path
+  runs (safer against browser/extension/proxy bearer injection). Set
+  `bearerOverridesCookie: true` for the AWS/GCP/Kubernetes convention.
+- **Replay protection preserved.** The bearer path skips the JTI **Set**
+  (so the same token can be reused) but the **Get** stays active —
+  `RevokeToken` still terminates a bearer token immediately.
+- **Excluded URLs strip Authorization.** When `enableBearerAuth=true`,
+  excluded paths (e.g. `/health`, `/metrics`) get the `Authorization` header
+  removed before forwarding so the token can't leak into public endpoint
+  logs.
+- **Optional real-time revocation.** Set `requireTokenIntrospection: true`
+  to call RFC 7662 introspection on every cache miss; revoked tokens fail
+  immediately. Introspection endpoint failures return `503` (distinguishes
+  infra outage from credential rejection).
+
+**Obtaining bearer tokens** — minting is the IdP's job, not the
+middleware's. The canonical M2M flow is OAuth 2.0 `client_credentials`
+(RFC 6749 §4.4); Google requires JWT bearer assertion (RFC 7523) instead.
+Minimal Auth0-shape request:
+
+```bash
+curl -s -X POST https://issuer.example.com/oauth/token \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "grant_type":    "client_credentials",
+    "client_id":     "your-m2m-client-id",
+    "client_secret": "your-m2m-client-secret",
+    "audience":      "https://api.example.com",
+    "scope":         "api:read api:write"
+  }'
+```
+
+The `audience` you request from the IdP **must match** the `audience` you
+configured on the middleware. Per-provider endpoints, parameter names, and
+gotchas (Entra v2 endpoint, Cognito Resource Servers, Keycloak audience
+mappers, Google's opaque-token quirk) are documented in
+[docs/BEARER_AUTH.md](docs/BEARER_AUTH.md#obtaining-bearer-tokens-from-your-oidc-provider).
+
+Full threat model, configuration matrix, and follow-up gaps in
+[docs/BEARER_AUTH.md](docs/BEARER_AUTH.md).
+
 ### SSE and WebSocket endpoints
 
 Browser clients cannot follow an OIDC `302` redirect on an SSE stream or a
@@ -324,6 +411,19 @@ namespaced claims, Cognito regions, GitLab self-hosted) live in
 
 Set `logLevel: debug` to surface detail.
 
+## Telemetry
+
+On first plugin instantiation this middleware sends a single anonymous
+adoption ping — project name, version, timestamp; no identifiers, no
+request data, no token contents. Fire-and-forget with a 2-second timeout;
+cannot block plugin load or panic.
+
+Local source: [`telemetry.go`](./telemetry.go). Disclosure mirrors
+**[oss-telemetry — Disabling telemetry](https://github.com/lukaszraczylo/oss-telemetry#disabling-telemetry)**.
+
+Quick opt-out: set any of `DO_NOT_TRACK=1`, `OSS_TELEMETRY_DISABLED=1`,
+or `TRAEFIKOIDC_DISABLE_TELEMETRY=1`.
+
 ## License
 
 See [LICENSE](LICENSE).

bearer_auth.go

@@ -0,0 +1,592 @@
+// Package traefikoidc — bearer-token (M2M) authentication path.
+//
+// Disabled by default. When enabled via Config.EnableBearerAuth, requests
+// presenting "Authorization: Bearer <jwt>" are validated against the
+// configured OIDC provider (signature, issuer, audience, exp, replay-Get)
+// and the request is forwarded downstream without creating a cookie session.
+//
+// Design rules (kept here in code as the single source of truth):
+//   - Access tokens only. ID tokens are rejected via detectTokenType.
+//   - Audience is mandatory (enforced at startup in main.go).
+//   - alg + kid pinned BEFORE JWKS fetch to deny amplification probes.
+//   - iat upper-age cap bounds clock-skew / forever-token abuse.
+//   - Multi-audience tokens require matching azp.
+//   - Per-IP 401 throttle returns 429 + Retry-After after a threshold.
+//   - JTI Set is suppressed (skipReplayMarking) but JTI Get stays — revoked
+//     tokens (RevokeToken adds to blacklist) are still rejected.
+//   - Identifier is read from BearerIdentifierClaim (default "sub"), never
+//     from UserIdentifierClaim, to avoid the unverified-email spoofing path.
+//   - Identifier is sanitized: length cap, control chars, bidi-override,
+//     delimiter chars (, ; =) rejected.
+//   - On excluded URLs the Authorization header is stripped before forwarding.
+//
+// See docs/superpowers/specs/2026-05-18-bearer-token-auth-design.md and
+// docs/BEARER_AUTH.md for the full threat model.
+package traefikoidc
+
+import (
+	"crypto/sha256"
+	"encoding/base64"
+	"encoding/hex"
+	"encoding/json"
+	"fmt"
+	"net"
+	"net/http"
+	"strings"
+	"sync"
+	"time"
+	"unicode"
+)
+
+const bearerPrefix = "Bearer "
+
+// bearerAlgAllowlist is the set of JWS algorithms accepted on the bearer
+// path. Asymmetric-only — HS* would allow public-key-as-HMAC-secret attacks
+// if any operator ever rotates a key into the symmetric branch by mistake;
+// "none" is obvious. Matches the allowlist enforced inside jwt.Verify but is
+// checked here BEFORE the JWKS fetch so attacker noise can't amplify.
+var bearerAlgAllowlist = map[string]struct{}{
+	"RS256": {}, "RS384": {}, "RS512": {},
+	"PS256": {}, "PS384": {}, "PS512": {},
+	"ES256": {}, "ES384": {}, "ES512": {},
+}
+
+// bearerKidMaxLen caps the JOSE kid header length to keep memory and cache-key
+// usage bounded against attacker-controlled values.
+const bearerKidMaxLen = 256
+
+// validKidChar is the allowlist for kid header characters. Letters, digits,
+// dot, underscore, hyphen, equals. Intentionally narrow; real-world kid
+// values are short URL-safe-base64-ish identifiers.
+func validKidChar(r rune) bool {
+	if r >= 'a' && r <= 'z' {
+		return true
+	}
+	if r >= 'A' && r <= 'Z' {
+		return true
+	}
+	if r >= '0' && r <= '9' {
+		return true
+	}
+	switch r {
+	case '.', '_', '-', '=':
+		return true
+	}
+	return false
+}
+
+// bearerError categorizes failure modes for the response builder. Categories
+// map 1:1 to the table in docs/superpowers/specs/2026-05-18-bearer-token-auth-design.md
+// §9 so behavior is auditable from spec to code.
+type bearerErrorKind int
+
+const (
+	bearerErrInvalidRequest bearerErrorKind = iota
+	bearerErrInvalidToken
+	bearerErrTokenInactive
+	bearerErrInvalidIdentifier
+	bearerErrForbidden
+	bearerErrThrottled
+	bearerErrIntrospectionUnavailable
+)
+
+type bearerError struct {
+	kind   bearerErrorKind
+	reason string
+}
+
+func (e *bearerError) Error() string { return e.reason }
+
+func newBearerError(kind bearerErrorKind, reason string) *bearerError {
+	return &bearerError{kind: kind, reason: reason}
+}
+
+// joseHeader is the minimal subset of the JWS protected header we inspect
+// BEFORE running the full verification pipeline. Lifted out so the alg+kid
+// pin can run without paying for parseJWT's full claim decode.
+type joseHeader struct {
+	Alg string `json:"alg"`
+	Kid string `json:"kid"`
+	Typ string `json:"typ"`
+}
+
+// parseBearerJOSEHeader decodes the first JWT segment for early alg/kid pinning.
+// Does not touch the payload or signature — those are the verifier's job.
+// Returns nil on success; *bearerError on rejection so the handler can map
+// directly to a status code. The decoded header itself is not surfaced because
+// callers don't need it (verifyTokenWithOpts re-parses internally).
+func parseBearerJOSEHeader(token string) *bearerError {
+	dot := strings.IndexByte(token, '.')
+	if dot <= 0 {
+		return newBearerError(bearerErrInvalidToken, "malformed JWT: no header segment")
+	}
+	raw, err := base64.RawURLEncoding.DecodeString(token[:dot])
+	if err != nil {
+		// Some IdPs pad with '='; tolerate by retrying with StdEncoding.
+		raw, err = base64.URLEncoding.DecodeString(token[:dot])
+		if err != nil {
+			return newBearerError(bearerErrInvalidToken, "malformed JWT: header not base64url")
+		}
+	}
+	var hdr joseHeader
+	if err := json.Unmarshal(raw, &hdr); err != nil {
+		return newBearerError(bearerErrInvalidToken, "malformed JWT: header not JSON")
+	}
+	if _, ok := bearerAlgAllowlist[hdr.Alg]; !ok {
+		return newBearerError(bearerErrInvalidToken, fmt.Sprintf("disallowed alg %q on bearer path", hdr.Alg))
+	}
+	if hdr.Kid == "" {
+		return newBearerError(bearerErrInvalidToken, "missing kid header")
+	}
+	if len(hdr.Kid) > bearerKidMaxLen {
+		return newBearerError(bearerErrInvalidToken, "kid header exceeds max length")
+	}
+	for _, r := range hdr.Kid {
+		if !validKidChar(r) {
+			return newBearerError(bearerErrInvalidToken, "kid header contains disallowed characters")
+		}
+	}
+	return nil
+}
+
+// sanitizeBearerIdentifier validates and trims a principal identifier before
+// it is injected into request headers. Layered defense: net/http will reject
+// CRLF on the wire too, but rejecting early gives clearer error logs and
+// prevents bidi-override / delimiter chars that pass net/http's narrower
+// checks but confuse downstream parsers and admin UIs.
+func sanitizeBearerIdentifier(raw string, maxLen int) (string, *bearerError) {
+	identifier := strings.TrimSpace(raw)
+	if identifier == "" {
+		return "", newBearerError(bearerErrInvalidIdentifier, "identifier claim empty")
+	}
+	if maxLen > 0 && len(identifier) > maxLen {
+		return "", newBearerError(bearerErrInvalidIdentifier, "identifier exceeds max length")
+	}
+	for _, r := range identifier {
+		if unicode.IsControl(r) {
+			return "", newBearerError(bearerErrInvalidIdentifier, "identifier contains control character")
+		}
+		// Unicode bidi-override range (RTL spoofing of admin UI / SIEM).
+		if (r >= 0x202A && r <= 0x202E) || (r >= 0x2066 && r <= 0x2069) {
+			return "", newBearerError(bearerErrInvalidIdentifier, "identifier contains bidi-override character")
+		}
+		if r == ',' || r == ';' || r == '=' {
+			return "", newBearerError(bearerErrInvalidIdentifier, "identifier contains delimiter character")
+		}
+	}
+	return identifier, nil
+}
+
+// resolveBearerIdentifier picks the principal identifier from claims using
+// the configured BearerIdentifierClaim (default "sub"). Decoupled from
+// userIdentifierClaim (cookie path) to avoid the unverified-email spoofing
+// vector documented in the spec §13.
+func resolveBearerIdentifier(claims map[string]interface{}, claimName string) (string, *bearerError) {
+	if claimName == "" {
+		claimName = "sub"
+	}
+	raw, ok := claims[claimName]
+	if !ok {
+		return "", newBearerError(bearerErrInvalidIdentifier, fmt.Sprintf("missing claim %q", claimName))
+	}
+	str, ok := raw.(string)
+	if !ok {
+		return "", newBearerError(bearerErrInvalidIdentifier, fmt.Sprintf("claim %q not a string", claimName))
+	}
+	return str, nil
+}
+
+// enforceMultiAudienceAzp implements the spec hardening: when aud is a
+// multi-element array, require an azp claim equal to clientID. Single-string
+// aud is unaffected (existing verifyAudience handles it).
+func enforceMultiAudienceAzp(claims map[string]interface{}, clientID string) *bearerError {
+	audRaw, ok := claims["aud"]
+	if !ok {
+		return nil // verifyToken already rejects missing aud
+	}
+	arr, ok := audRaw.([]interface{})
+	if !ok {
+		return nil // single-string aud
+	}
+	if len(arr) <= 1 {
+		return nil
+	}
+	azpRaw, ok := claims["azp"]
+	if !ok {
+		return newBearerError(bearerErrInvalidToken, "multi-audience token missing azp")
+	}
+	azp, ok := azpRaw.(string)
+	if !ok || azp == "" {
+		return newBearerError(bearerErrInvalidToken, "multi-audience token has empty/non-string azp")
+	}
+	if azp != clientID {
+		return newBearerError(bearerErrInvalidToken, "multi-audience token azp does not match clientID")
+	}
+	return nil
+}
+
+// enforceIatAge implements the spec MaxTokenAgeSeconds bound on iat. Bounds
+// clock-manipulation / forever-token abuse without rejecting tokens with a
+// normal iat just because the issuer's clock skews a few seconds.
+func enforceIatAge(claims map[string]interface{}, maxAge time.Duration) *bearerError {
+	if maxAge <= 0 {
+		return nil
+	}
+	iatRaw, ok := claims["iat"].(float64)
+	if !ok {
+		// jwt.Verify already requires iat; this branch shouldn't be reached.
+		return newBearerError(bearerErrInvalidToken, "missing iat claim")
+	}
+	iat := time.Unix(int64(iatRaw), 0)
+	if time.Since(iat) > maxAge {
+		return newBearerError(bearerErrInvalidToken, "token iat outside age bound")
+	}
+	return nil
+}
+
+// hashIdentifierForLog returns a short SHA-256 prefix safe for info-level
+// logs. Full identifier is only emitted at debug. Satisfies the audit
+// requirement (trace which principal was rejected) without leaking PII.
+func hashIdentifierForLog(identifier string) string {
+	if identifier == "" {
+		return "(none)"
+	}
+	sum := sha256.Sum256([]byte(identifier))
+	return hex.EncodeToString(sum[:4]) // 8 hex chars
+}
+
+// --- Per-IP failure throttle ---
+
+// bearerFailureTracker records consecutive bearer-auth 401s per source IP and
+// parks repeat offenders in a 429 penalty box. Limits offline-guessing-style
+// attacks and protects the shared rate-limiter / JWKS endpoint from being
+// burned by a single source.
+type bearerFailureTracker struct {
+	mu      sync.Mutex
+	entries map[string]*bearerFailureEntry
+	// Configuration snapshot. Captured at construction so a hot reconfigure
+	// doesn't race with the per-request paths.
+	threshold int
+	window    time.Duration
+	penalty   time.Duration
+}
+
+type bearerFailureEntry struct {
+	firstFailureAt time.Time
+	penaltyUntil   time.Time
+	count          int
+}
+
+func newBearerFailureTracker(threshold int, window, penalty time.Duration) *bearerFailureTracker {
+	if threshold <= 0 {
+		threshold = 20
+	}
+	if window <= 0 {
+		window = 60 * time.Second
+	}
+	if penalty <= 0 {
+		penalty = 60 * time.Second
+	}
+	return &bearerFailureTracker{
+		entries:   make(map[string]*bearerFailureEntry),
+		threshold: threshold,
+		window:    window,
+		penalty:   penalty,
+	}
+}
+
+// blocked reports whether the source IP is currently in the penalty box.
+// Returns (true, retryAfter) when blocked; (false, 0) when allowed.
+func (b *bearerFailureTracker) blocked(ip string) (bool, time.Duration) {
+	if b == nil || ip == "" {
+		return false, 0
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	e, ok := b.entries[ip]
+	if !ok {
+		return false, 0
+	}
+	now := time.Now()
+	if !e.penaltyUntil.IsZero() && now.Before(e.penaltyUntil) {
+		return true, time.Until(e.penaltyUntil)
+	}
+	return false, 0
+}
+
+// recordFailure increments the failure counter for the given IP and trips
+// the penalty box once threshold-within-window is exceeded.
+func (b *bearerFailureTracker) recordFailure(ip string) {
+	if b == nil || ip == "" {
+		return
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	now := time.Now()
+	e, ok := b.entries[ip]
+	if !ok || now.Sub(e.firstFailureAt) > b.window {
+		e = &bearerFailureEntry{firstFailureAt: now}
+		b.entries[ip] = e
+	}
+	e.count++
+	if e.count >= b.threshold {
+		e.penaltyUntil = now.Add(b.penalty)
+	}
+}
+
+// recordSuccess clears the failure counter for the given IP after a
+// successful bearer auth.
+func (b *bearerFailureTracker) recordSuccess(ip string) {
+	if b == nil || ip == "" {
+		return
+	}
+	b.mu.Lock()
+	defer b.mu.Unlock()
+	delete(b.entries, ip)
+}
+
+// clientIPForBearer returns the source IP used to key the failure tracker.
+// Trusts only the request's transport-level RemoteAddr; X-Forwarded-For is
+// intentionally ignored to avoid attacker-controlled key spoofing. Behind a
+// trusted reverse proxy where every request shares one IP, the throttle is
+// still useful (caps attacker churn through that proxy) — operators wanting
+// per-real-client throttling must terminate at this middleware.
+func clientIPForBearer(req *http.Request) string {
+	if req == nil {
+		return ""
+	}
+	host, _, err := net.SplitHostPort(req.RemoteAddr)
+	if err != nil {
+		return req.RemoteAddr
+	}
+	return host
+}
+
+// --- Bearer auth entrypoint ---
+
+// detectBearerToken returns (token, true) when the request carries a usable
+// Authorization: Bearer header. Case-insensitive on the scheme. Returns
+// ("", false) for any other shape.
+func detectBearerToken(req *http.Request) (string, bool) {
+	if req == nil {
+		return "", false
+	}
+	h := req.Header.Get("Authorization")
+	if len(h) < len(bearerPrefix) {
+		return "", false
+	}
+	if !strings.EqualFold(h[:len(bearerPrefix)], bearerPrefix) {
+		return "", false
+	}
+	token := strings.TrimSpace(h[len(bearerPrefix):])
+	if token == "" {
+		return "", false
+	}
+	return token, true
+}
+
+// hasSessionCookie reports whether the request carries any cookie matching
+// the session prefix. Used to implement the cookie-wins-by-default
+// precedence rule when both bearer and cookie are present.
+func (t *TraefikOidc) hasSessionCookie(req *http.Request) bool {
+	if t.sessionManager == nil {
+		return false
+	}
+	prefix := t.sessionManager.GetCookiePrefix()
+	if prefix == "" {
+		return false
+	}
+	for _, c := range req.Cookies() {
+		if strings.HasPrefix(c.Name, prefix) {
+			return true
+		}
+	}
+	return false
+}
+
+// writeBearerError writes the canonical 401/403/429/503 response per spec §9.
+// Body is always generic; reason is logged at debug only. The
+// WWW-Authenticate hint is gated by config (default on, RFC 6750 compliant).
+func (t *TraefikOidc) writeBearerError(rw http.ResponseWriter, req *http.Request, err *bearerError) {
+	var (
+		status     int
+		errCode    string
+		body       string
+		retryAfter time.Duration
+	)
+	switch err.kind {
+	case bearerErrInvalidRequest:
+		status = http.StatusUnauthorized
+		errCode = "invalid_request"
+		body = "Unauthorized"
+	case bearerErrInvalidToken, bearerErrTokenInactive, bearerErrInvalidIdentifier:
+		status = http.StatusUnauthorized
+		errCode = "invalid_token"
+		body = "Unauthorized"
+	case bearerErrForbidden:
+		status = http.StatusForbidden
+		body = "Access denied"
+	case bearerErrThrottled:
+		status = http.StatusTooManyRequests
+		body = "Too Many Requests"
+		retryAfter = t.bearerFailurePenalty
+	case bearerErrIntrospectionUnavailable:
+		status = http.StatusServiceUnavailable
+		body = "Service Unavailable"
+	default:
+		status = http.StatusUnauthorized
+		body = "Unauthorized"
+	}
+
+	if t.bearerEmitWWWAuthenticate && errCode != "" {
+		rw.Header().Set("WWW-Authenticate", fmt.Sprintf(`Bearer error=%q`, errCode))
+	}
+	if retryAfter > 0 {
+		rw.Header().Set("Retry-After", fmt.Sprintf("%d", int(retryAfter.Seconds())))
+	}
+	rw.Header().Set("Content-Type", "text/plain; charset=utf-8")
+	rw.WriteHeader(status)
+	_, _ = rw.Write([]byte(body)) // Safe to ignore: best-effort error body write
+
+	if t.logger != nil {
+		t.logger.Debugf("bearer auth rejected: status=%d category=%v reason=%q path=%s",
+			status, err.kind, err.reason, req.URL.Path)
+	}
+}
+
+// handleBearerRequest is the entry point invoked by ServeHTTP when the
+// EnableBearerAuth flag is set, the request carries an Authorization: Bearer
+// header, and the (configurable) cookie-precedence rule allows the bearer
+// path to run.
+func (t *TraefikOidc) handleBearerRequest(rw http.ResponseWriter, req *http.Request) {
+	ip := clientIPForBearer(req)
+
+	if blocked, retryAfter := t.bearerFailureTracker.blocked(ip); blocked {
+		throttled := newBearerError(bearerErrThrottled, "ip in penalty box")
+		// Preserve the actual retry-after even if it diverged from the
+		// configured default (clock-skew, partial-window expiry).
+		if retryAfter > 0 {
+			rw.Header().Set("Retry-After", fmt.Sprintf("%d", int(retryAfter.Seconds())))
+		}
+		t.writeBearerError(rw, req, throttled)
+		return
+	}
+
+	token, ok := detectBearerToken(req)
+	if !ok {
+		t.bearerFailureTracker.recordFailure(ip)
+		t.writeBearerError(rw, req, newBearerError(bearerErrInvalidRequest, "missing or empty bearer token"))
+		return
+	}
+	if len(token) > AccessTokenConfig.MaxLength {
+		t.bearerFailureTracker.recordFailure(ip)
+		t.writeBearerError(rw, req, newBearerError(bearerErrInvalidToken, "token exceeds max length"))
+		return
+	}
+	if strings.Count(token, ".") != 2 {
+		t.bearerFailureTracker.recordFailure(ip)
+		t.writeBearerError(rw, req, newBearerError(bearerErrInvalidToken, "token is not a 3-segment JWT"))
+		return
+	}
+
+	if bErr := parseBearerJOSEHeader(token); bErr != nil {
+		t.bearerFailureTracker.recordFailure(ip)
+		t.writeBearerError(rw, req, bErr)
+		return
+	}
+
+	p, bErr := t.buildPrincipalFromBearerToken(token)
+	if bErr != nil {
+		t.bearerFailureTracker.recordFailure(ip)
+		t.writeBearerError(rw, req, bErr)
+		return
+	}
+
+	t.bearerFailureTracker.recordSuccess(ip)
+	if t.logger != nil {
+		t.logger.Debugf("bearer auth success: identifier_hash=%s path=%s",
+			hashIdentifierForLog(p.Identifier), req.URL.Path)
+	}
+	t.forwardAuthorized(rw, req, p)
+}
+
+// buildPrincipalFromBearerToken runs the full bearer verification pipeline
+// described in spec §7.3 and returns a principal ready for forwardAuthorized.
+// Returns a typed *bearerError on failure so the caller can map to status.
+func (t *TraefikOidc) buildPrincipalFromBearerToken(token string) (*principal, *bearerError) {
+	if err := t.verifyTokenWithOpts(token, verifyOpts{skipReplayMarking: true}); err != nil {
+		return nil, newBearerError(bearerErrInvalidToken, "token verification failed: "+err.Error())
+	}
+
+	parsed, err := parseJWT(token)
+	if err != nil {
+		return nil, newBearerError(bearerErrInvalidToken, "post-verify parseJWT failed: "+err.Error())
+	}
+	claims := parsed.Claims
+
+	// Token-type guard. Reuse the well-tested classifier which already
+	// checks nonce / typ=at+jwt / token_use / scope / aud-vs-clientID.
+	if t.detectTokenType(parsed, token) {
+		return nil, newBearerError(bearerErrInvalidToken, "ID tokens are not accepted on the bearer path")
+	}
+	// Belt-and-braces explicit rejection (cheap, catches edge cases not
+	// covered by detectTokenType's heuristic).
+	if nonce, ok := claims["nonce"].(string); ok && nonce != "" {
+		return nil, newBearerError(bearerErrInvalidToken, "nonce claim present (ID-token shape)")
+	}
+	if tu, ok := claims["token_use"].(string); ok && tu == "id" {
+		return nil, newBearerError(bearerErrInvalidToken, "token_use=id rejected")
+	}
+
+	if bErr := enforceMultiAudienceAzp(claims, t.clientID); bErr != nil {
+		return nil, bErr
+	}
+	if bErr := enforceIatAge(claims, t.maxTokenAge); bErr != nil {
+		return nil, bErr
+	}
+
+	if t.requireTokenIntrospection {
+		if bErr := t.introspectOnBearerPath(token); bErr != nil {
+			return nil, bErr
+		}
+	}
+
+	rawIdentifier, bErr := resolveBearerIdentifier(claims, t.bearerIdentifierClaim)
+	if bErr != nil {
+		return nil, bErr
+	}
+	identifier, bErr := sanitizeBearerIdentifier(rawIdentifier, t.maxIdentifierLength)
+	if bErr != nil {
+		return nil, bErr
+	}
+
+	subject, _ := claims["sub"].(string)
+	clientID, _ := claims["azp"].(string)
+	if clientID == "" {
+		clientID, _ = claims["client_id"].(string)
+	}
+
+	return &principal{
+		Source:      sourceBearer,
+		Identifier:  identifier,
+		Subject:     subject,
+		ClientID:    clientID,
+		Claims:      claims,
+		AccessToken: token,
+	}, nil
+}
+
+// introspectOnBearerPath calls the existing RFC 7662 introspector when the
+// operator demands real-time revocation. Distinguishes "token revoked" (401)
+// from "endpoint unavailable" (503) so transient infra failures don't look
+// like credential failures.
+func (t *TraefikOidc) introspectOnBearerPath(token string) *bearerError {
+	resp, err := t.introspectToken(token)
+	if err != nil {
+		return newBearerError(bearerErrIntrospectionUnavailable, "introspection failed: "+err.Error())
+	}
+	if !resp.Active {
+		return newBearerError(bearerErrTokenInactive, "introspection reports token inactive")
+	}
+	return nil
+}

bearer_auth_test.go

@@ -0,0 +1,812 @@
+package traefikoidc
+
+import (
+	"context"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync/atomic"
+	"testing"
+	"time"
+
+	"golang.org/x/time/rate"
+)
+
+// =============================================================================
+// Helper builders
+// =============================================================================
+
+// makeBearerJWT constructs a JWT with explicit header + claims for tests.
+// Signature is opaque (b64("signature")) — bearer tests don't exercise the
+// real cryptographic verifier; verification is bypassed via tokenCache pre-
+// seed so the bearer pipeline under test sees a "verified" token.
+func makeBearerJWT(t *testing.T, header, claims map[string]interface{}) string {
+	t.Helper()
+	hb, err := json.Marshal(header)
+	if err != nil {
+		t.Fatalf("marshal header: %v", err)
+	}
+	cb, err := json.Marshal(claims)
+	if err != nil {
+		t.Fatalf("marshal claims: %v", err)
+	}
+	return fmt.Sprintf("%s.%s.%s",
+		base64.RawURLEncoding.EncodeToString(hb),
+		base64.RawURLEncoding.EncodeToString(cb),
+		base64.RawURLEncoding.EncodeToString([]byte("signature")),
+	)
+}
+
+// defaultBearerHeader produces the standard RS256+kid header used in tests.
+func defaultBearerHeader() map[string]interface{} {
+	return map[string]interface{}{"alg": "RS256", "kid": "test-kid"}
+}
+
+// defaultBearerClaims produces a baseline access-token claim set. Tests
+// shallow-clone and override fields as needed.
+func defaultBearerClaims() map[string]interface{} {
+	return map[string]interface{}{
+		"iss":   "https://issuer.example.com",
+		"aud":   "https://api.example.com",
+		"sub":   "service-account-1",
+		"scope": "api:read api:write",
+		"exp":   float64(time.Now().Add(time.Hour).Unix()),
+		"iat":   float64(time.Now().Unix()),
+	}
+}
+
+// makeBearerOIDC constructs a TraefikOidc wired for bearer auth tests. The
+// real verifyTokenWithOpts pipeline is short-circuited via tokenCache pre-
+// seed: any token Set into t.tokenCache returns nil from VerifyToken,
+// letting tests exercise the post-verify bearer logic (classifier, identifier,
+// throttle, header forwarding) without standing up JWKs.
+func makeBearerOIDC(t *testing.T, next http.Handler) *TraefikOidc {
+	t.Helper()
+	sm := createTestSessionManager(t)
+	oidc := &TraefikOidc{
+		next:                      next,
+		logger:                    NewLogger("error"),
+		initComplete:              make(chan struct{}),
+		sessionManager:            sm,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
+		issuerURL:                 "https://issuer.example.com",
+		audience:                  "https://api.example.com",
+		clientID:                  "https://api.example.com",
+		tokenCache:                NewTokenCache(),
+		excludedURLs:              map[string]struct{}{"/favicon.ico": {}},
+		allowedRolesAndGroups:     map[string]struct{}{},
+		limiter:                   rate.NewLimiter(rate.Every(time.Second), 1000),
+		ctx:                       context.Background(),
+		enableBearerAuth:          true,
+		stripAuthorizationHeader:  true,
+		bearerEmitWWWAuthenticate: true,
+		bearerOverridesCookie:     false,
+		bearerIdentifierClaim:     "sub",
+		maxIdentifierLength:       256,
+		maxTokenAge:               24 * time.Hour,
+		bearerFailureThreshold:    20,
+		bearerFailureWindow:       60 * time.Second,
+		bearerFailurePenalty:      60 * time.Second,
+		bearerFailureTracker:      newBearerFailureTracker(20, 60*time.Second, 60*time.Second),
+	}
+	oidc.extractClaimsFunc = extractClaims
+	close(oidc.initComplete)
+	return oidc
+}
+
+// seedVerified pre-populates the tokenCache so verifyTokenWithOpts short-
+// circuits to nil for the given token. Mirrors the production fast-return
+// path at token_manager.go for previously-verified tokens.
+func seedVerified(t *testing.T, oidc *TraefikOidc, token string, claims map[string]interface{}) {
+	t.Helper()
+	if oidc.tokenCache == nil {
+		oidc.tokenCache = NewTokenCache()
+	}
+	oidc.tokenCache.Set(token, claims, time.Hour)
+}
+
+// =============================================================================
+// Unit tests — small helpers
+// =============================================================================
+
+func TestDetectBearerToken(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		name   string
+		header string
+		want   string
+		ok     bool
+	}{
+		{"missing header", "", "", false},
+		{"basic auth", "Basic abc", "", false},
+		{"bearer with token", "Bearer abc.def.ghi", "abc.def.ghi", true},
+		{"lowercase bearer", "bearer abc.def.ghi", "abc.def.ghi", true},
+		{"mixed case", "BeArEr abc.def.ghi", "abc.def.ghi", true},
+		{"empty token after prefix", "Bearer    ", "", false},
+		{"bearer no space", "Bearerabc", "", false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			req := httptest.NewRequest("GET", "/", nil)
+			if tc.header != "" {
+				req.Header.Set("Authorization", tc.header)
+			}
+			got, ok := detectBearerToken(req)
+			if ok != tc.ok || got != tc.want {
+				t.Fatalf("got=(%q, %v), want=(%q, %v)", got, ok, tc.want, tc.ok)
+			}
+		})
+	}
+}
+
+func TestParseBearerJOSEHeader(t *testing.T) {
+	t.Parallel()
+	mk := func(t *testing.T, h map[string]interface{}) string {
+		return makeBearerJWT(t, h, map[string]interface{}{"sub": "x"})
+	}
+	cases := []struct {
+		header  map[string]interface{}
+		name    string
+		wantErr bool
+	}{
+		{name: "valid RS256", header: map[string]interface{}{"alg": "RS256", "kid": "k1"}, wantErr: false},
+		{name: "valid ES512", header: map[string]interface{}{"alg": "ES512", "kid": "abc-_.="}, wantErr: false},
+		{name: "alg=none rejected", header: map[string]interface{}{"alg": "none", "kid": "k1"}, wantErr: true},
+		{name: "alg=HS256 rejected", header: map[string]interface{}{"alg": "HS256", "kid": "k1"}, wantErr: true},
+		{name: "missing kid", header: map[string]interface{}{"alg": "RS256"}, wantErr: true},
+		{name: "kid too long", header: map[string]interface{}{"alg": "RS256", "kid": strings.Repeat("a", bearerKidMaxLen+1)}, wantErr: true},
+		{name: "kid bad chars", header: map[string]interface{}{"alg": "RS256", "kid": "evil/../etc/passwd"}, wantErr: true},
+		{name: "kid with space", header: map[string]interface{}{"alg": "RS256", "kid": "key one"}, wantErr: true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			token := mk(t, tc.header)
+			err := parseBearerJOSEHeader(token)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err=%v wantErr=%v", err, tc.wantErr)
+			}
+		})
+	}
+}
+
+func TestSanitiseBearerIdentifier(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		name    string
+		in      string
+		want    string
+		wantErr bool
+	}{
+		{"normal sub", "service-account-1", "service-account-1", false},
+		{"email-like", "alice@example.com", "alice@example.com", false},
+		{"trim whitespace", "  abc  ", "abc", false},
+		{"empty", "", "", true},
+		{"only whitespace", "   ", "", true},
+		{"control char (newline)", "alice\nbob", "", true},
+		{"control char (CR)", "alice\rbob", "", true},
+		{"control char (NUL)", "alice\x00bob", "", true},
+		{"bidi override", "alice\u202ebob", "", true},
+		{"bidi isolate", "alice\u2066bob", "", true},
+		{"comma delimiter", "alice,bob", "", true},
+		{"semicolon delimiter", "alice;bob", "", true},
+		{"equals delimiter", "alice=bob", "", true},
+		{"over length", strings.Repeat("a", 257), "", true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := sanitizeBearerIdentifier(tc.in, 256)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err=%v wantErr=%v", err, tc.wantErr)
+			}
+			if !tc.wantErr && got != tc.want {
+				t.Fatalf("got=%q want=%q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestResolveBearerIdentifier(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		claims  map[string]interface{}
+		name    string
+		claim   string
+		want    string
+		wantErr bool
+	}{
+		{name: "default sub", claims: map[string]interface{}{"sub": "abc"}, claim: "", want: "abc"},
+		{name: "explicit sub", claims: map[string]interface{}{"sub": "abc"}, claim: "sub", want: "abc"},
+		{name: "custom client_id claim", claims: map[string]interface{}{"client_id": "svc"}, claim: "client_id", want: "svc"},
+		{name: "missing claim", claims: map[string]interface{}{"other": "x"}, claim: "sub", wantErr: true},
+		{name: "non-string claim", claims: map[string]interface{}{"sub": 123}, claim: "sub", wantErr: true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got, err := resolveBearerIdentifier(tc.claims, tc.claim)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err=%v wantErr=%v", err, tc.wantErr)
+			}
+			if !tc.wantErr && got != tc.want {
+				t.Fatalf("got=%q want=%q", got, tc.want)
+			}
+		})
+	}
+}
+
+func TestEnforceMultiAudienceAzp(t *testing.T) {
+	t.Parallel()
+	const cid = "https://api.example.com"
+	cases := []struct {
+		claims  map[string]interface{}
+		name    string
+		wantErr bool
+	}{
+		{name: "single string aud", claims: map[string]interface{}{"aud": "x"}, wantErr: false},
+		{name: "single element array", claims: map[string]interface{}{"aud": []interface{}{"x"}}, wantErr: false},
+		{name: "multi-aud with matching azp", claims: map[string]interface{}{"aud": []interface{}{"a", "b"}, "azp": cid}, wantErr: false},
+		{name: "multi-aud missing azp", claims: map[string]interface{}{"aud": []interface{}{"a", "b"}}, wantErr: true},
+		{name: "multi-aud empty azp", claims: map[string]interface{}{"aud": []interface{}{"a", "b"}, "azp": ""}, wantErr: true},
+		{name: "multi-aud wrong azp", claims: map[string]interface{}{"aud": []interface{}{"a", "b"}, "azp": "other"}, wantErr: true},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			err := enforceMultiAudienceAzp(tc.claims, cid)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err=%v wantErr=%v", err, tc.wantErr)
+			}
+		})
+	}
+}
+
+func TestEnforceIatAge(t *testing.T) {
+	t.Parallel()
+	now := time.Now()
+	cases := []struct {
+		name    string
+		iat     float64
+		maxAge  time.Duration
+		wantErr bool
+	}{
+		{name: "fresh", iat: float64(now.Unix()), maxAge: time.Hour, wantErr: false},
+		{name: "23h59m old, max 24h", iat: float64(now.Add(-23*time.Hour - 59*time.Minute).Unix()), maxAge: 24 * time.Hour, wantErr: false},
+		{name: "25h old, max 24h", iat: float64(now.Add(-25 * time.Hour).Unix()), maxAge: 24 * time.Hour, wantErr: true},
+		{name: "1970 token", iat: float64(0), maxAge: 24 * time.Hour, wantErr: true},
+		{name: "maxAge disabled (0)", iat: float64(0), maxAge: 0, wantErr: false},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			err := enforceIatAge(map[string]interface{}{"iat": tc.iat}, tc.maxAge)
+			if (err != nil) != tc.wantErr {
+				t.Fatalf("err=%v wantErr=%v", err, tc.wantErr)
+			}
+		})
+	}
+}
+
+func TestBearerFailureTracker(t *testing.T) {
+	t.Parallel()
+	tr := newBearerFailureTracker(3, 60*time.Second, 60*time.Second)
+	const ip = "10.0.0.1"
+	// Below threshold: not blocked.
+	for i := 0; i < 2; i++ {
+		tr.recordFailure(ip)
+		if b, _ := tr.blocked(ip); b {
+			t.Fatalf("blocked too early after %d failures", i+1)
+		}
+	}
+	// Threshold reached: blocked.
+	tr.recordFailure(ip)
+	if b, retry := tr.blocked(ip); !b || retry <= 0 {
+		t.Fatalf("expected blocked with positive retry, got=%v retry=%v", b, retry)
+	}
+	// Success clears the counter.
+	tr.recordSuccess(ip)
+	if b, _ := tr.blocked(ip); b {
+		t.Fatalf("expected unblocked after success")
+	}
+	// Other IPs are unaffected.
+	if b, _ := tr.blocked("10.0.0.2"); b {
+		t.Fatalf("unrelated IP should not be blocked")
+	}
+}
+
+// =============================================================================
+// Integration tests — full ServeHTTP via the bearer pipeline
+// =============================================================================
+
+func TestServeHTTP_Bearer_HappyPath(t *testing.T) {
+	t.Parallel()
+	var nextCalled atomic.Bool
+	var capturedHeaders http.Header
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		nextCalled.Store(true)
+		capturedHeaders = r.Header.Clone()
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if !nextCalled.Load() {
+		t.Fatalf("expected next handler to run; got status=%d body=%q", rw.Code, rw.Body.String())
+	}
+	if rw.Code != http.StatusOK {
+		t.Fatalf("status=%d, want 200", rw.Code)
+	}
+	if got := capturedHeaders.Get("X-Forwarded-User"); got != "service-account-1" {
+		t.Fatalf("X-Forwarded-User=%q, want service-account-1", got)
+	}
+	if got := capturedHeaders.Get("Authorization"); got != "" {
+		t.Fatalf("Authorization should be stripped, got=%q", got)
+	}
+}
+
+func TestServeHTTP_Bearer_StripAuthDisabled(t *testing.T) {
+	t.Parallel()
+	var capturedAuth string
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		capturedAuth = r.Header.Get("Authorization")
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	oidc.stripAuthorizationHeader = false
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if !strings.HasPrefix(capturedAuth, "Bearer ") {
+		t.Fatalf("expected Authorization to be forwarded, got=%q", capturedAuth)
+	}
+}
+
+func TestServeHTTP_Bearer_RejectIDToken(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for ID token rejection")
+	})
+	oidc := makeBearerOIDC(t, next)
+	// ID-token shape: nonce claim present and no scope. detectTokenType
+	// returns true.
+	claims := map[string]interface{}{
+		"iss":   "https://issuer.example.com",
+		"aud":   "https://api.example.com",
+		"sub":   "user-1",
+		"nonce": "n-0S6_WzA2Mj",
+		"exp":   float64(time.Now().Add(time.Hour).Unix()),
+		"iat":   float64(time.Now().Unix()),
+	}
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+	if wa := rw.Header().Get("WWW-Authenticate"); !strings.Contains(wa, `error="invalid_token"`) {
+		t.Fatalf("expected WWW-Authenticate invalid_token, got=%q", wa)
+	}
+}
+
+func TestServeHTTP_Bearer_AlgNoneRejected(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for alg=none")
+	})
+	oidc := makeBearerOIDC(t, next)
+	header := map[string]interface{}{"alg": "none", "kid": "k1"}
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, header, claims)
+	// Even if we pre-seeded the cache, the early alg pin runs FIRST.
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_KidTooLongRejected(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for oversized kid")
+	})
+	oidc := makeBearerOIDC(t, next)
+	header := map[string]interface{}{"alg": "RS256", "kid": strings.Repeat("a", bearerKidMaxLen+1)}
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, header, claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_MultiAudRequiresAzp(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for multi-aud without azp")
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	claims["aud"] = []interface{}{"https://api.example.com", "https://other.example.com"}
+	delete(claims, "azp")
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_MultiAudWithAzpAccepted(t *testing.T) {
+	t.Parallel()
+	var nextCalled atomic.Bool
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		nextCalled.Store(true)
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	claims["aud"] = []interface{}{"https://api.example.com", "https://other.example.com"}
+	claims["azp"] = oidc.clientID
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusOK || !nextCalled.Load() {
+		t.Fatalf("expected 200 + next called; got status=%d called=%v", rw.Code, nextCalled.Load())
+	}
+}
+
+func TestServeHTTP_Bearer_IatTooOldRejected(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for old iat")
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	claims["iat"] = float64(time.Now().Add(-25 * time.Hour).Unix())
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_IdentifierWithBidiRejected(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for bidi identifier")
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	claims["sub"] = "alice\u202ebob"
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_ReplayRegression(t *testing.T) {
+	t.Parallel()
+	var successCount atomic.Int32
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		successCount.Add(1)
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	claims["jti"] = "regression-jti"
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	for i := 0; i < 100; i++ {
+		req := httptest.NewRequest("GET", "/api/work", nil)
+		req.Header.Set("Authorization", "Bearer "+token)
+		rw := httptest.NewRecorder()
+		oidc.ServeHTTP(rw, req)
+		if rw.Code != http.StatusOK {
+			t.Fatalf("iteration %d: status=%d, want 200", i, rw.Code)
+		}
+	}
+	if successCount.Load() != 100 {
+		t.Fatalf("successCount=%d, want 100", successCount.Load())
+	}
+}
+
+func TestServeHTTP_Bearer_ThrottleTrips429(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run during throttle test")
+	})
+	oidc := makeBearerOIDC(t, next)
+	oidc.bearerFailureTracker = newBearerFailureTracker(3, 60*time.Second, 60*time.Second)
+
+	// Send malformed bearers from the same RemoteAddr until threshold trips.
+	send := func() *httptest.ResponseRecorder {
+		req := httptest.NewRequest("GET", "/api/work", nil)
+		req.RemoteAddr = "10.0.0.5:1234"
+		req.Header.Set("Authorization", "Bearer not-a-jwt")
+		rw := httptest.NewRecorder()
+		oidc.ServeHTTP(rw, req)
+		return rw
+	}
+	for i := 0; i < 3; i++ {
+		rw := send()
+		if rw.Code != http.StatusUnauthorized {
+			t.Fatalf("pre-throttle iteration %d: status=%d, want 401", i, rw.Code)
+		}
+	}
+	// 4th request: throttled.
+	rw := send()
+	if rw.Code != http.StatusTooManyRequests {
+		t.Fatalf("expected 429 after threshold, got %d", rw.Code)
+	}
+	if ra := rw.Header().Get("Retry-After"); ra == "" {
+		t.Fatalf("expected Retry-After header on 429")
+	}
+}
+
+func TestServeHTTP_Bearer_ExcludedURLStripsAuth(t *testing.T) {
+	t.Parallel()
+	var capturedAuth string
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		capturedAuth = r.Header.Get("Authorization")
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	oidc.excludedURLs = map[string]struct{}{"/favicon.ico": {}}
+
+	req := httptest.NewRequest("GET", "/favicon.ico", nil)
+	req.Header.Set("Authorization", "Bearer abc.def.ghi")
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if rw.Code != http.StatusOK {
+		t.Fatalf("excluded path should pass; got %d", rw.Code)
+	}
+	if capturedAuth != "" {
+		t.Fatalf("Authorization must be stripped on excluded paths, got=%q", capturedAuth)
+	}
+}
+
+func TestServeHTTP_Bearer_RolesGate(t *testing.T) {
+	t.Parallel()
+	cases := []struct {
+		name       string
+		rolesClaim []interface{}
+		want       int
+	}{
+		{name: "matching role", rolesClaim: []interface{}{"admin"}, want: http.StatusOK},
+		{name: "no matching role", rolesClaim: []interface{}{"viewer"}, want: http.StatusForbidden},
+	}
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+				w.WriteHeader(http.StatusOK)
+			})
+			oidc := makeBearerOIDC(t, next)
+			oidc.allowedRolesAndGroups = map[string]struct{}{"admin": {}}
+			oidc.roleClaimName = "roles"
+			claims := defaultBearerClaims()
+			claims["roles"] = tc.rolesClaim
+			token := makeBearerJWT(t, defaultBearerHeader(), claims)
+			seedVerified(t, oidc, token, claims)
+
+			req := httptest.NewRequest("GET", "/api/work", nil)
+			req.Header.Set("Authorization", "Bearer "+token)
+			rw := httptest.NewRecorder()
+			oidc.ServeHTTP(rw, req)
+			if rw.Code != tc.want {
+				t.Fatalf("status=%d, want %d", rw.Code, tc.want)
+			}
+		})
+	}
+}
+
+func TestServeHTTP_Bearer_CookieWinsByDefault(t *testing.T) {
+	t.Parallel()
+	// Both cookie and bearer present: cookie path runs (which will redirect
+	// to /authorize since the cookie is empty/unauthenticated).
+	var nextCalled atomic.Bool
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		nextCalled.Store(true)
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	prefix := oidc.sessionManager.GetCookiePrefix()
+	req.AddCookie(&http.Cookie{Name: prefix + "main", Value: "irrelevant"})
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	// Cookie path consumed the request; bearer was ignored. Since the
+	// cookie is empty, the cookie path will either 302 to /authorize or
+	// return 401 — in either case, next must NOT be called.
+	if nextCalled.Load() {
+		t.Fatalf("next must not be called when bearer is ignored due to cookie precedence")
+	}
+}
+
+func TestServeHTTP_Bearer_BearerOverridesCookie(t *testing.T) {
+	t.Parallel()
+	var nextCalled atomic.Bool
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		nextCalled.Store(true)
+		w.WriteHeader(http.StatusOK)
+	})
+	oidc := makeBearerOIDC(t, next)
+	oidc.bearerOverridesCookie = true
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	prefix := oidc.sessionManager.GetCookiePrefix()
+	req.AddCookie(&http.Cookie{Name: prefix + "main", Value: "irrelevant"})
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+
+	if !nextCalled.Load() || rw.Code != http.StatusOK {
+		t.Fatalf("expected bearer to win with override; status=%d called=%v", rw.Code, nextCalled.Load())
+	}
+}
+
+func TestServeHTTP_Bearer_OversizedToken(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for oversized token")
+	})
+	oidc := makeBearerOIDC(t, next)
+	huge := strings.Repeat("a", AccessTokenConfig.MaxLength+1)
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+huge)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_MalformedJWT(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		t.Fatalf("next must not run for malformed JWT")
+	})
+	oidc := makeBearerOIDC(t, next)
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer not.jwt") // 1 dot
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+	if rw.Code != http.StatusUnauthorized {
+		t.Fatalf("status=%d, want 401", rw.Code)
+	}
+}
+
+func TestServeHTTP_Bearer_FeatureOffPassesThrough(t *testing.T) {
+	t.Parallel()
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		// Should not be reached: cookie path runs and (with no session)
+		// will redirect or 401. We assert no panic / next not called.
+		t.Fatalf("next must not run when bearer is off and no valid session exists")
+	})
+	oidc := makeBearerOIDC(t, next)
+	oidc.enableBearerAuth = false
+	claims := defaultBearerClaims()
+	token := makeBearerJWT(t, defaultBearerHeader(), claims)
+	seedVerified(t, oidc, token, claims)
+	req := httptest.NewRequest("GET", "/api/work", nil)
+	req.Header.Set("Authorization", "Bearer "+token)
+	rw := httptest.NewRecorder()
+	oidc.ServeHTTP(rw, req)
+	// Expect non-200: either 302 to /authorize or 401. The point is the
+	// bearer pipeline didn't run.
+	if rw.Code == http.StatusOK {
+		t.Fatalf("expected non-200 when bearer is off; got %d", rw.Code)
+	}
+}
+
+// =============================================================================
+// Startup validation tests
+// =============================================================================
+
+func TestStartupValidation_BearerRequiresAudience(t *testing.T) {
+	t.Parallel()
+	cfg := CreateConfig()
+	cfg.ProviderURL = "https://issuer.example.com"
+	cfg.ClientID = "id"
+	cfg.ClientSecret = "secret"
+	cfg.CallbackURL = "/oauth/callback"
+	cfg.SessionEncryptionKey = "0123456789abcdef0123456789abcdef0123456789abcdef"
+	cfg.EnableBearerAuth = true
+	cfg.Audience = ""
+	_, err := New(context.Background(), http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {}), cfg, "bearer-test")
+	if err == nil || !strings.Contains(err.Error(), "requires Audience") {
+		t.Fatalf("expected audience-required error, got %v", err)
+	}
+}
+
+func TestStartupValidation_BearerRejectsEmailIdentifier(t *testing.T) {
+	t.Parallel()
+	cfg := CreateConfig()
+	cfg.ProviderURL = "https://issuer.example.com"
+	cfg.ClientID = "id"
+	cfg.ClientSecret = "secret"
+	cfg.CallbackURL = "/oauth/callback"
+	cfg.SessionEncryptionKey = "0123456789abcdef0123456789abcdef0123456789abcdef"
+	cfg.EnableBearerAuth = true
+	cfg.Audience = "https://api.example.com"
+	cfg.BearerIdentifierClaim = "email"
+	_, err := New(context.Background(), http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {}), cfg, "bearer-test")
+	if err == nil || !strings.Contains(err.Error(), "bearerIdentifierClaim=\"email\"") {
+		t.Fatalf("expected email-identifier rejection, got %v", err)
+	}
+}
+
+// =============================================================================
+// Principal invariants
+// =============================================================================
+
+func TestBuildPrincipalFromSession_NoIdentifier(t *testing.T) {
+	t.Parallel()
+	oidc := &TraefikOidc{logger: NewLogger("error")}
+	if p := oidc.buildPrincipalFromSession(nil); p != nil {
+		t.Fatalf("nil session must produce nil principal")
+	}
+}

docs/BEARER_AUTH.md

@@ -0,0 +1,250 @@
+# Bearer Token (M2M) Authentication
+
+Opt-in path that lets API clients present `Authorization: Bearer <jwt>` to
+authenticate without going through the cookie-based OIDC redirect flow.
+Designed for machine-to-machine (M2M) traffic — services calling other
+services with tokens minted by your OIDC provider.
+
+The bearer path lives next to the cookie path: both go through the same
+post-auth pipeline (`forwardAuthorized`) that injects identity headers,
+checks `allowedRolesAndGroups`, applies security headers, and forwards to
+the backend. The only thing that differs is how the principal is established
+for that single request.
+
+## Quick start
+
+```yaml
+enableBearerAuth: true
+audience: https://api.example.com  # REQUIRED when bearer is enabled
+clientID: my-api-client-id
+providerURL: https://issuer.example.com
+sessionEncryptionKey: <32+-byte secret>
+callbackURL: /oauth2/callback
+```
+
+That is the minimum. Everything else has a secure default.
+
+## Obtaining bearer tokens from your OIDC provider
+
+The middleware only **validates** bearer tokens — minting them is the IdP's job. For M2M traffic the canonical mint flow is OAuth 2.0 **`client_credentials`** (RFC 6749 §4.4); some providers require **JWT bearer assertion** (RFC 7523) instead.
+
+```
+┌────────────┐  POST /token                    ┌──────────┐
+│  client    │ ───────────────────────────────►│  IdP     │
+│  (service) │   grant_type=client_credentials │ /token   │
+│            │   client_id=…                   │          │
+│            │   client_secret=…  (or JWT)     │          │
+│            │   audience=https://api.…   ←── critical    │
+│            │   scope=api:read …                         │
+│            │ ◄───────────────────────────────│          │
+│            │     access_token (JWT)          │          │
+└────────────┘                                 └──────────┘
+         │
+         │  GET /protected
+         │  Authorization: Bearer <access_token>
+         ▼
+   Your service (behind Traefik + this plugin)
+```
+
+The IdP returns a JWT signed by the same JWKs the middleware already trusts (it discovers them from `providerURL`/.well-known). On the first protected request, the middleware verifies signature + issuer + **audience** + `exp` + identifier claim, then forwards downstream with `X-Forwarded-User` set.
+
+### Minimal worked example (Auth0-shape)
+
+```bash
+# 1. Mint a token
+curl -s -X POST https://issuer.example.com/oauth/token \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "grant_type":    "client_credentials",
+    "client_id":     "your-m2m-client-id",
+    "client_secret": "your-m2m-client-secret",
+    "audience":      "https://api.example.com",
+    "scope":         "api:read api:write"
+  }'
+# → {"access_token":"eyJhbGciOiJSUzI1NiIs…","token_type":"Bearer","expires_in":86400,…}
+
+# 2. Use it
+curl -H 'Authorization: Bearer eyJhbGciOiJSUzI1NiIs…' https://api.example.com/protected
+```
+
+The `audience` field in the token request **must match** the `audience` you configured on the middleware. Mismatch → 401 with `Bearer error="invalid_token"`.
+
+### Per-provider quick reference
+
+| Provider | Grant | Token endpoint | Audience parameter | Notes |
+|---|---|---|---|---|
+| **Auth0** | `client_credentials` | `https://TENANT.auth0.com/oauth/token` | `audience=<your API identifier>` | Register an "API" + "Machine to Machine Application" authorised against that API. Without `audience` you get an opaque /userinfo token, which the bearer path rejects. See `docs/AUTH0_AUDIENCE_GUIDE.md`. |
+| **Okta** | `client_credentials` | `https://TENANT.okta.com/oauth2/default/v1/token` | Configured in the authorization server; default `aud` is the auth-server URL | Service app must enable the `client_credentials` flow and be granted the requested scopes. |
+| **Keycloak** | `client_credentials` | `https://kc/realms/REALM/protocol/openid-connect/token` | Configure an "Audience" mapper on a client scope, or use `client_id` as the audience | Client must have `serviceAccountsEnabled: true` plus role mappings. |
+| **Entra ID / Azure AD** | `client_credentials` (v2.0 endpoint) | `https://login.microsoftonline.com/TENANT/oauth2/v2.0/token` | Pass `scope=<App ID URI>/.default`; `aud` ends up being the API's App ID URI | Requires an App Registration + API permissions + admin consent. **Use the v2.0 endpoint** — v1 issues Microsoft-proprietary access tokens that are opaque to non-Microsoft clients. |
+| **AWS Cognito** | `client_credentials` | `https://YOUR_DOMAIN.auth.REGION.amazoncognito.com/oauth2/token` | Scopes from a "Resource Server" attached to your User Pool | App client must have `client_credentials` flow enabled. Use HTTP **Basic** auth header for `client_id:client_secret`. |
+| **GitLab** | `client_credentials` | `https://gitlab.com/oauth/token` | Audience matches the GitLab issuer | Rarely used for protecting external APIs; better suited for GitLab's own resources. |
+| **Google** | **JWT bearer (RFC 7523)** — *not* `client_credentials` | `https://oauth2.googleapis.com/token` | Signed assertion JWT carries `aud=https://oauth2.googleapis.com/token`; resulting access token is **opaque** unless you specifically request a Google-issued JWT for your API | Google service-account flow is not the best fit for this middleware (opaque tokens are rejected on the bearer path). Run Auth0 / Okta / Keycloak in front, or use ID-token-based flows on the cookie path. |
+
+### RFC 7523 (JWT bearer assertion) — secretless alternative
+
+When shared secrets are forbidden (FAPI, internal compliance), swap `client_secret` for a signed JWT assertion:
+
+```
+POST /token
+grant_type=urn:ietf:params:oauth:grant-type:jwt-bearer
+assertion=<JWT signed by the client's private key>
+```
+
+The assertion JWT carries `iss=<client_id>`, `sub=<client_id>`, `aud=<token endpoint>`, `exp`. The IdP verifies the signature against a public key you've pre-registered and returns an access token.
+
+This middleware already supports JWT assertions on the *middleware → IdP* hop via `clientAuthMethod: private_key_jwt` (see `docs/CONFIGURATION.md`). For the *client → IdP* hop, the same pattern applies — the client signs its own assertion.
+
+### Operational notes
+
+- **Token TTL is typically 1–24 hours.** Clients should refresh on `401`, not on a polling timer — saves the IdP.
+- **Cache and reuse tokens.** The middleware caches verified tokens too, so repeated presentations are cheap. Clients SHOULD reuse a token until ~80 % of `expires_in`.
+- **JWKS rotation is transparent.** The middleware auto-refreshes its JWKS cache when the IdP rotates keys. Clients don't need to do anything.
+- **Revocation is generally not per-token** with `client_credentials`. If you need real-time revocation, set `requireTokenIntrospection: true` on the middleware and the IdP is consulted on every cache miss.
+- **`scope` vs `audience`.** Scope says *what the client may do*; audience says *which service the token is for*. The middleware enforces audience; the backend service should enforce scope.
+- **Secret hygiene.** Store `client_secret` in a secrets manager (Vault, AWS Secrets Manager, Kubernetes `Secret`). For higher assurance, switch the client to `private_key_jwt` (no shared secret at all).
+
+### Quickest validation loop
+
+```bash
+# 1. Mint
+TOKEN=$(curl -s -X POST https://issuer.example.com/oauth/token \
+  -H 'Content-Type: application/json' \
+  -d '{"grant_type":"client_credentials","client_id":"…","client_secret":"…","audience":"https://api.example.com"}' \
+  | jq -r .access_token)
+
+# 2. Inspect claims to confirm aud/iss/exp match the middleware config
+echo "$TOKEN" | cut -d. -f2 | base64 -d 2>/dev/null | jq
+
+# 3. Hit the protected route
+curl -i -H "Authorization: Bearer $TOKEN" https://api.example.com/protected
+```
+
+`HTTP/1.1 200` with `X-Forwarded-User` on the backend confirms the loop works end-to-end. `401` with `WWW-Authenticate: Bearer error="invalid_token"` plus a middleware debug log explaining the rejection (audience mismatch, ID token presented, `iat` outside the 24h window, etc.) confirms the hardening is firing as designed.
+
+## Threat model and design rules
+
+Bearer authentication has materially different security properties from
+cookie sessions: no `HttpOnly`/`Secure`/`SameSite` shielding, the token is
+visible in headers and logs, and it's easier to exfiltrate. The bearer path
+treats every one of these as a first-class concern.
+
+| Property | Behaviour | Why |
+|---|---|---|
+| Default state | `enableBearerAuth=false` | Bearer is opt-in; existing deployments observe no change. |
+| Audience | **Mandatory.** Startup fails if `audience` is empty when bearer is enabled. | Eliminates the "token issued for service B accepted by service A" confusion attack. |
+| Token format | JWT only (3 segments, JOSE-encoded). Opaque tokens are not accepted on the bearer path. | Matches the validation pipeline; opaque tokens require introspection only and bypass JWT-specific defences. |
+| `alg` allowlist | Hard-pinned asymmetric: `RS256/384/512`, `PS256/384/512`, `ES256/384/512`. Checked **before** any JWKS fetch. | Denies `alg=none` and `alg=HS*` probes; prevents attacker noise from amplifying into JWKS round-trips. |
+| `kid` hardening | Max 256 bytes; charset `[A-Za-z0-9._\-=]`. Checked **before** JWKS fetch. | Prevents cache-key explosion / pathological-`kid` JWKS amplification. |
+| Token type | ID tokens are explicitly rejected (`nonce` claim, `typ: at+jwt`, `token_use=id`, scope/aud heuristics — reuses the existing `detectTokenType` helper). | ID tokens are not API credentials; treating them as such is classic token confusion. |
+| Multi-audience | When `aud` is an array of length > 1, the token must carry `azp == clientID`. | OIDC §2 hardening against tokens minted for one client being replayed by another. |
+| `iat` upper-age | Rejects tokens older than `maxTokenAgeSeconds` (default 24h). | Bounds clock-manipulation / forever-token abuse, even if `exp` is far in the future. |
+| Identifier claim | `bearerIdentifierClaim` (default `"sub"`). Resolved value drives `X-Forwarded-User`. | Decoupled from the cookie path's `UserIdentifierClaim` (default `email`) so the M2M flow can never accidentally trust an unverified email. |
+| Identifier sanitisation | Length cap (`maxIdentifierLength`, default 256). Rejects control chars, Unicode bidi-overrides (U+202A–U+202E, U+2066–U+2069), and the delimiters `, ; =`. | Defence in depth against downstream header injection / log injection / admin-UI spoofing. |
+| JTI replay marking | Bearer path skips the JTI **Set** (so the same token can be reused until `exp`) but the **Get** stays active. | Allows legitimate bearer reuse without false-positive replay detection; revoked tokens (added to the blacklist by `RevokeToken`) still fail immediately. |
+| Mixed bearer + cookie | **Cookie wins by default.** Flip to bearer-wins with `bearerOverridesCookie=true`. | Safer against browser/extension/proxy bearer injection scenarios. The cookie is the authoritative authenticator when present. |
+| `Authorization` strip | `stripAuthorizationHeader=true` by default. | Keeps the raw token out of downstream services and their logs. |
+| Excluded URLs | `Authorization` is stripped on excluded paths when `enableBearerAuth=true`. | Prevents bearer leakage into public health/metrics endpoint logs and prevents recon via excluded paths. |
+| Per-IP throttle | After `bearerFailureThreshold` consecutive 401s from one source IP within `bearerFailureWindowSeconds`, further bearer requests from that IP return `429 Too Many Requests` + `Retry-After` for `bearerFailurePenaltySeconds`. | Limits offline-guessing-style attacks and protects the shared rate-limiter / JWKS endpoint. |
+| Optional introspection | `requireTokenIntrospection=true` calls RFC 7662 introspection on every cache miss. Introspection result is cached briefly. Endpoint failure returns `503` (distinguishes infra outage from credential rejection). | Real-time revocation for high-assurance environments. Adds per-request IdP latency. |
+| Response shape | `401 Unauthorized` with generic body. `WWW-Authenticate: Bearer error="invalid_token"` per RFC 6750 §3 (toggleable via `bearerEmitWWWAuthenticate`). `403` for roles/groups denial. `429` for throttle. `503` for introspection-endpoint outage. | Auditable from spec to code; reason categories never leak into the response body. |
+| Logging | Failure reason + identifier hash (SHA-256 truncated to 8 hex chars) logged at debug. Raw tokens are never logged. | Audit trail without secrets-in-logs. |
+
+## Configuration reference
+
+| Field | Default | Description |
+|---|---|---|
+| `enableBearerAuth` | `false` | Master switch for the bearer path. |
+| `audience` | (unset) | **Required** when `enableBearerAuth=true`. Reuses the existing global `audience` field. |
+| `bearerIdentifierClaim` | `"sub"` | JWT claim used as the principal identifier. `"email"` is rejected at startup. |
+| `stripAuthorizationHeader` | `true` | Remove the `Authorization` header before forwarding to the backend. Disable only when a downstream needs to re-verify the bearer. |
+| `bearerEmitWWWAuthenticate` | `true` | Include `WWW-Authenticate: Bearer error="..."` on 401 responses (RFC 6750 §3). Disable to reduce recon signal. |
+| `bearerOverridesCookie` | `false` | Cookie wins when both are present (default). Set `true` for the AWS/GCP/Kubernetes bearer-wins convention. |
+| `maxTokenAgeSeconds` | `86400` | Upper bound on `iat` claim age (24h). Set `0` to disable the check (not recommended). |
+| `maxIdentifierLength` | `256` | Length cap for the post-sanitisation identifier. |
+| `bearerFailureThreshold` | `20` | Consecutive 401s from one IP that trip the throttle. |
+| `bearerFailureWindowSeconds` | `60` | Rolling window over which 401s are counted. |
+| `bearerFailurePenaltySeconds` | `60` | Duration of the 429 penalty box after the threshold trips. |
+| `requireTokenIntrospection` | `false` | Call RFC 7662 introspection on every cache miss. Adds per-request IdP latency. |
+
+## What the bearer path does NOT do
+
+- **Human-user / browser flows.** The bearer path is M2M-only in this
+  iteration. Browser SPAs that want to attach a bearer to fetch calls work
+  if your backend treats them as machine clients, but the spec defaults are
+  tuned for service-to-service traffic.
+- **Opaque access tokens.** Tokens must be JWTs. Introspection is a
+  revocation overlay on top of JWT verification, not a substitute for it.
+- **`email_verified` enforcement.** The bearer path rejects `email` as the
+  identifier claim at startup precisely because `email_verified` is not
+  enforced in this iteration. Adding human-user bearer support is a
+  follow-up that must include this check.
+- **mTLS / API keys.** Out of scope. The `principal` abstraction enables
+  adding these later as additional auth methods that produce a principal
+  for the shared `forwardAuthorized` pipeline.
+- **SSE / WebSocket bypass with bearer.** Bypass paths keep their existing
+  cookie-only behaviour; bearer headers are ignored on those endpoints.
+  Documented limitation; widen by removing the bypass if you need bearer on
+  streaming endpoints.
+
+## Operational guidance
+
+- **Always set `strictAudienceValidation: true` when bearer is enabled.**
+  Startup logs a recommendation if you don't.
+- **Set a tight `maxTokenAgeSeconds`** for environments where tokens are
+  expected to be minted frequently — the default 24h is conservative.
+- **Enable `requireTokenIntrospection`** if your IdP supports it and
+  revocation latency matters. Bearer-path introspection caches results for
+  a short window per token.
+- **Monitor 429s.** Sustained 429 traffic indicates either a buggy client
+  loop or an active credential-stuffing attempt. The throttle is your
+  primary signal for both.
+- **`stripAuthorizationHeader=false` extends the token's blast radius** to
+  every downstream service that sees the request. Treat those services'
+  logs as token stores.
+- **Bearer reuse is normal.** Don't enable per-token rate limiting; that's
+  what `bearerFailureThreshold` is for (per-IP, not per-token).
+- **Cookie-wins is the safer default.** Only flip `bearerOverridesCookie`
+  if you control all clients and have audited that none of them present a
+  cookie alongside a bearer they don't intend to authenticate with.
+
+## Failure response matrix
+
+| Trigger | Status | Body | `WWW-Authenticate` |
+|---|---|---|---|
+| Empty bearer after prefix | 401 | `Unauthorized` | `Bearer error="invalid_request"` |
+| Token over `MaxLength` | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Not a 3-segment JWT | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Disallowed `alg` (e.g. none, HS*) | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Missing / oversized / bad-charset `kid` | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Signature / issuer / audience / `exp` failure | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| `iat` older than `maxTokenAgeSeconds` | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Multi-audience token without matching `azp` | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Detected as ID token | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| JTI blacklisted (revoked) | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Introspection reports `active=false` | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Introspection endpoint failure | 503 | `Service Unavailable` | (none) |
+| Identifier claim missing / empty | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Identifier fails sanitisation | 401 | `Unauthorized` | `Bearer error="invalid_token"` |
+| Per-IP failure threshold tripped | 429 | `Too Many Requests` | (none); `Retry-After: <bearerFailurePenaltySeconds>` |
+| Roles / groups not allowed | 403 | `Access denied` | (none) |
+
+## Known follow-ups (deferred)
+
+These are documented as future work, not blockers:
+
+- **Human-user bearer with `email_verified` enforcement.** Requires
+  decoupling the email-claim guard from the startup rejection and adding a
+  per-request `email_verified=true` check.
+- **Introspection respects `client_assertion`.** The existing introspection
+  helper uses `client_secret_basic` only; operators on `private_key_jwt`
+  will see introspection silently use basic auth.
+- **Per-route bearer configuration.** Single middleware-wide setting in this
+  iteration.
+
+## References
+
+- [PR design spec](superpowers/specs/2026-05-18-bearer-token-auth-design.md) — full design rationale, alternatives considered, and per-section sign-off history.
+- [RFC 6750](https://www.rfc-editor.org/rfc/rfc6750) — Bearer Token Usage.
+- [RFC 7662](https://www.rfc-editor.org/rfc/rfc7662) — OAuth 2.0 Token Introspection.
+- [RFC 9068](https://www.rfc-editor.org/rfc/rfc9068) — JWT Profile for OAuth 2.0 Access Tokens.

docs/CONFIGURATION.md

@@ -261,6 +261,26 @@ strictAudienceValidation: true
 | `disableReplayDetection` | bool | `false` | Disable JTI-based replay attack detection |
 | `allowPrivateIPAddresses` | bool | `false` | Allow private IPs in provider URLs |
 
+### Bearer-token (M2M) authentication
+
+Opt-in path that accepts `Authorization: Bearer <jwt>` instead of the cookie
+session flow. M2M-only, default off, audience-mandatory. See
+[docs/BEARER_AUTH.md](BEARER_AUTH.md) for the threat model and operational
+guidance.
+
+| Parameter | Type | Default | Description |
+|-----------|------|---------|-------------|
+| `enableBearerAuth` | bool | `false` | Master switch. Startup fails if true with empty `audience` or with `bearerIdentifierClaim=email`. |
+| `bearerIdentifierClaim` | string | `"sub"` | JWT claim used as the principal identifier. `"email"` is rejected at startup. |
+| `stripAuthorizationHeader` | bool | `true` | Strip `Authorization` from forwarded requests after successful bearer auth. |
+| `bearerEmitWWWAuthenticate` | bool | `true` | Emit RFC 6750 `WWW-Authenticate: Bearer error="..."` hints on 401. |
+| `bearerOverridesCookie` | bool | `false` | Cookie wins when both bearer and cookie are present (default). Set true for bearer-wins. |
+| `maxTokenAgeSeconds` | int64 | `86400` | Upper bound on `iat` claim age (24h). 0 disables the check. |
+| `maxIdentifierLength` | int | `256` | Length cap on the sanitised principal identifier. |
+| `bearerFailureThreshold` | int | `20` | Consecutive 401s from one source IP that trip the throttle. |
+| `bearerFailureWindowSeconds` | int | `60` | Rolling window for counting 401s. |
+| `bearerFailurePenaltySeconds` | int | `60` | 429 + `Retry-After` duration after the threshold trips. |
+
 ---
 
 ## Session Management

docs/superpowers/specs/2026-05-18-bearer-token-auth-design.md

@@ -0,0 +1,459 @@
+# Bearer Token Authentication — Design Spec
+
+- **Date**: 2026-05-18
+- **Status**: Design — pending implementation plan
+- **Supersedes**: PR #93 (broken implementation; recommended to close in favour of this design)
+
+## 1. Summary
+
+Add an opt-in path that lets API clients (machine-to-machine) authenticate by presenting a signed access token in the `Authorization: Bearer <token>` header, bypassing the cookie-based OIDC redirect flow. Identity, roles, and authorization checks remain consistent with the existing cookie path; the only thing that changes is how the principal is established for that single request.
+
+The feature is implemented by extracting a shared `forwardAuthorized` pipeline from the existing `processAuthorizedRequest`, introducing a `principal` value type, and adding a small bearer-specific entrypoint that builds a principal directly from a verified JWT — without synthesising a fake `SessionData`.
+
+## 2. Motivation
+
+PR #93 attempted this feature by building an in-memory `SessionData` from JWT claims and reusing `processAuthorizedRequest`. The approach has three latent defects:
+
+1. The synthetic session omits `mainSession.Values["user_identifier"]`. `processAuthorizedRequest` reads it via `GetUserIdentifier()`; when empty it bails to `defaultInitiateAuthentication` and issues an OIDC redirect. The feature is non-functional in practice despite the unit test passing.
+2. `verifyToken` accepts both ID tokens (audience match against `clientID`) and access tokens. ID tokens are not API credentials; treating them as such is a classic token-confusion vector.
+3. `verifyToken` adds JTI to the replay blacklist on first verify. Once the verified-token cache evicts, subsequent reuse of the same bearer token triggers a false-positive replay rejection.
+
+Rather than patch a synthetic-session approach that will keep generating bugs as `SessionData` evolves, this spec replaces it with a cleaner abstraction where session lifecycle and post-auth header injection live in separate units.
+
+## 3. Goals
+
+- Accept `Authorization: Bearer <jwt>` from M2M clients, validate the token, and forward the request downstream with identity headers populated.
+- Enforce the same `allowedRolesAndGroups` policy as the cookie path.
+- Default-off; safe defaults when enabled (audience required, ID tokens rejected, identifier sanitised).
+- No behavioural change to the cookie path. Existing tests must continue to pass without modification.
+
+## 4. Non-Goals
+
+- Human-user / browser flows. Bearer is M2M-only in this iteration.
+- Pure opaque access tokens on the bearer path. Tokens must be JWTs; introspection (RFC 7662) is supported *on top of* JWT verification for revocation state, not as a substitute for it.
+- mTLS, API keys, or any other auth method. The `principal` abstraction enables them later, but they are not delivered here.
+- Per-route bearer configuration. Single middleware-wide setting.
+
+## 5. Decided Requirements
+
+| Topic | Decision |
+|---|---|
+| Consumer type | Machine-to-machine (M2M) only |
+| Token format | JWT only (signature, issuer, audience, exp) |
+| Audience | Mandatory when feature enabled; startup fails if `Audience == ""` |
+| Token type | Access tokens only; ID tokens explicitly rejected |
+| Revocation | JWT-only verification by default; introspection (RFC 7662) opt-in via existing `RequireTokenIntrospection` |
+| Identity claim | New `BearerIdentifierClaim` config (string, default `"sub"`). Bearer path reads this claim exclusively; does NOT use `UserIdentifierClaim` (which defaults to `"email"` and drives the cookie path). Resolved value must be a non-empty string. `sub` is mandatory per `jwt.go:416` regardless, so even with a different `BearerIdentifierClaim` the token must still carry a valid `sub`. Decoupling avoids the M2M-vs-human-user identity-claim conflict and the email-spoofing footgun. |
+| Identifier sanitisation | Reject value containing any `unicode.IsControl` char, any Unicode bidi-override (U+202A–U+202E, U+2066–U+2069), leading/trailing whitespace, commas, semicolons, equals signs. Max length 256 bytes. |
+| Token classifier | **Reuse existing `detectTokenType(jwt, token)` at `token_manager.go:187-303`** which already handles `nonce`, `typ: at+jwt`, `token_use`, `scope`, and aud-vs-clientID priority. Bearer path rejects any token where `detectTokenType == true` (ID token). Do not invent a parallel classifier. |
+| Algorithm pinning | Hard-pin `alg ∈ {RS256, RS384, RS512, PS256, PS384, PS512, ES256, ES384, ES512}`, enforced **before** JWKS lookup on the bearer path. Prevents wasted JWKS fetches for `alg=none`/HS attacker probes. |
+| `kid` hardening | `kid` ≤ 256 bytes, charset `[A-Za-z0-9._\-=]`. Reject before JWKS lookup. |
+| Token age | Bearer path enforces `now - iat <= MaxTokenAgeSeconds` (default 86400 / 24h, configurable). Cookie path unchanged. |
+| Multi-audience policy | If `aud` is an array (length > 1), require `azp` claim to be present and equal to `clientID`. Single-string `aud` unaffected. |
+| Mixed bearer + cookie precedence | **Cookie wins by default** when both are presented (safer for browser scenarios). Operator opt-in: `BearerOverridesCookie=true` to flip. Either way, a warning is logged on the request. |
+| Bearer + excluded URL | `Authorization` header is **stripped** before forwarding when the request hits an excluded URL. Prevents bearer leaking into public endpoints' downstream logs and prevents recon via excluded paths. |
+| Per-source bearer 401 throttle | New sharded cache `failedBearerAttempts` keyed by client IP. After N (default 20) consecutive 401s from one IP within 1 minute, reject further bearer requests from that IP with 429 for 60s. Applied BEFORE `verifyToken` to deny JWKS amplification. |
+| `Authorization` header passthrough | New `StripAuthorizationHeader` config, default `true` |
+| Roles/groups gating | Same `allowedRolesAndGroups` rules as cookie path |
+| Default state | `EnableBearerAuth` = `false` |
+| JTI replay marking | Suppressed on bearer path; cookie path unchanged |
+| Failure response shape | 401 with generic body; `WWW-Authenticate: Bearer error="invalid_token"` per RFC 6750 |
+| Introspection endpoint outage | 503 (distinguishes infra outage from token rejection) |
+| Mixed bearer + cookie | Bearer wins; cookie ignored on that request |
+| SSE/WS bypass + bearer | Bypass paths keep cookie-only check; bearer header ignored on SSE/WS |
+
+## 6. Architecture
+
+```
+                ┌──────────────────┐
+   HTTP req ──► │   ServeHTTP      │  (existing entry; adds bearer detection)
+                └─────────┬────────┘
+              ┌───────────┴────────────┐
+              ▼                        ▼
+        cookie / session         bearer (Authorization: Bearer …)
+              │                        │
+              ▼                        ▼
+    ┌────────────────┐        ┌────────────────────┐
+    │ buildPrincipal │        │ buildPrincipal     │
+    │ FromSession()  │        │ FromBearerToken()  │
+    └────────┬───────┘        └─────────┬──────────┘
+             │     produces *principal  │
+             └──────────────┬───────────┘
+                            ▼
+              ┌────────────────────────────┐
+              │ forwardAuthorized(rw,req,p)│  (shared pipeline)
+              │  • roles/groups gate       │
+              │  • header injection        │
+              │  • header templates        │
+              │  • security headers        │
+              │  • cookie stripping        │
+              │  • next.ServeHTTP          │
+              └────────────────────────────┘
+```
+
+**Invariant**: `forwardAuthorized` never touches session storage. Session-specific concerns (Save, IsDirty, backchannel-logout invalidation) stay inside `processAuthorizedRequest` around the call to `forwardAuthorized`.
+
+**Feature gate**: when `EnableBearerAuth == false`, the bearer-detection check in `ServeHTTP` is a no-op. Existing deployments observe byte-identical behaviour.
+
+## 7. Components
+
+### 7.1 `principal` type (new file `principal.go`)
+
+```go
+type principalSource int
+
+const (
+    sourceSession principalSource = iota
+    sourceBearer
+)
+
+type principal struct {
+    Identifier   string                 // drives X-Forwarded-User
+    Email        string                 // optional, "" for M2M
+    Subject      string                 // sub claim
+    ClientID     string                 // azp / client_id, M2M caller
+    Claims       map[string]interface{} // raw claims for templates / groups
+    AccessToken  string                 // for X-Auth-Request-Token (gated by minimalHeaders)
+    IDToken      string                 // "" on bearer path
+    RefreshToken string                 // "" on bearer path
+    Source       principalSource
+}
+```
+
+Pure data. No methods that mutate it. No I/O. No manager pointer.
+
+### 7.2 `buildPrincipalFromSession(*SessionData) *principal` (new in `principal.go`)
+
+Read-only adapter over existing `SessionData` getters: `GetUserIdentifier`, `GetEmail`, `GetAccessToken`, `GetIDToken`, `GetRefreshToken`, cached claims via `GetIDTokenClaims`. Does not write back to the session. This is the only function that still knows about `SessionData`.
+
+### 7.3 `buildPrincipalFromBearerToken(token string) (*principal, error)` (new in `bearer_auth.go`)
+
+1. **Length / format guards**: `len(token) <= AccessTokenConfig.MaxLength`, exactly two dots, non-empty after trim.
+2. **Parse header for early alg/kid pinning** (without trusting payload): decode JOSE header; reject if `alg` ∉ asymmetric allowlist; reject if `kid` missing, > 256 bytes, or contains chars outside `[A-Za-z0-9._\-=]`. This happens **before** JWKS lookup so attacker noise doesn't amplify into JWKS fetches.
+3. **Per-IP 401 throttle check**: if this IP is in the `failedBearerAttempts` penalty box, return 429 immediately.
+4. `t.verifyToken(token, verifyOpts{skipReplayMarking: true})` — reuses signature, issuer, audience, expiration, JTI Get (replay detection). The `skipReplayMarking` flag gates ONLY the JTI Set at `token_manager.go:108-143`; the JTI Get at `token_manager.go:44-47, 80-89` remains active so revoked tokens (via `RevokeToken` adding to blacklist) are still rejected.
+5. **Re-parse claims** (`parseJWT(token)` is cheap and already done internally; reuse via a single decode if practical).
+6. **Token-type guard**: call existing `detectTokenType(jwt, token)` (`token_manager.go:187-303`). Reject when it returns `true` (ID token). Belt-and-braces: also reject if `claims["nonce"]` is a non-empty string or `claims["token_use"] == "id"`.
+7. **Multi-audience hardening**: if `claims["aud"]` is a `[]interface{}` with length > 1, require `claims["azp"]` to be a non-empty string equal to `t.clientID`; reject otherwise.
+8. **`iat` upper-age bound**: reject when `time.Now().Unix() - int64(claims["iat"].(float64)) > MaxTokenAgeSeconds` (default 86400).
+9. **Optional introspection**: if `requireTokenIntrospection` is set, call `introspectToken`; reject if `active == false` (401); surface 503 on transport failure. Bearer-path introspection cache TTL is capped at 60s (not 5min) to keep the "real-time revocation" promise close to true.
+10. **Identifier resolution**: read `t.bearerIdentifierClaim` (defaults to `"sub"`); do NOT use `t.userIdentifierClaim` (cookie path's setting, default `email`). The bearer path does NOT fall back to other claims because `jwt.Verify` already enforces non-empty `sub` (`jwt.go:416-419`). Empty/missing identifier → 401.
+11. **Identifier sanitisation**: trim, then reject if length > 256 OR contains any of: `unicode.IsControl`, bidi-override (U+202A–U+202E, U+2066–U+2069), `,`, `;`, `=`.
+12. Return `&principal{ Source: sourceBearer, … }`.
+
+On any failure path: increment the per-IP `failedBearerAttempts` counter; return the appropriate HTTP status (401 / 403 / 429 / 503) without revealing the failure reason in the response body. Reason is logged at debug only, with the identifier (if resolved) hashed via SHA-256 truncated to 8 hex chars.
+
+### 7.4 `forwardAuthorized(rw, req, *principal)` (new in `middleware.go`, extracted)
+
+The shared post-auth pipeline. Lifted verbatim from the existing `processAuthorizedRequest`:
+
+1. Roles/groups extraction via existing `extractGroupsAndRolesFromClaims`.
+2. `allowedRolesAndGroups` gate (existing logic).
+3. Inject `X-Forwarded-User`, `X-User-Groups`, `X-User-Roles`.
+4. Inject `X-Auth-Request-*` (gated by `minimalHeaders`).
+5. Header templates.
+6. Security headers.
+7. Cookie strip when `stripAuthCookies`.
+8. **New**: `Authorization` header strip when `stripAuthorizationHeader` AND `principal.Source == sourceBearer`.
+9. `t.next.ServeHTTP(rw, req)`.
+
+Does not call `Save`, does not check `IsDirty`. Session persistence stays with the cookie-path caller.
+
+### 7.5 `handleBearerRequest(rw, req)` (new in `bearer_auth.go`)
+
+```
+1. Detect "Authorization: Bearer <token>" (case-insensitive prefix).
+2. token = TrimSpace(authHeader[7:]); reject empty.
+3. p, err := buildPrincipalFromBearerToken(token).
+   On err → 401 with WWW-Authenticate, log reason at debug.
+4. forwardAuthorized(rw, req, p).
+```
+
+Target: ~40 lines.
+
+### 7.6 Refactor of `processAuthorizedRequest` (modify `middleware.go`)
+
+Splits along the principal boundary:
+- Session-specific part (backchannel-logout invalidation, `IsDirty` / `Save`) stays in `processAuthorizedRequest`.
+- Everything else moves to `forwardAuthorized`.
+- `processAuthorizedRequest` ends with `forwardAuthorized(rw, req, buildPrincipalFromSession(session))`.
+
+### 7.7 `verifyOpts` extension to `verifyToken` (modify `token_manager.go`)
+
+Add a parameter struct:
+```go
+type verifyOpts struct {
+    skipReplayMarking bool // suppress JTI Set (token_manager.go:108-143); blacklist Get stays active
+}
+```
+
+Both the type and field are unexported (internal-only knob). Signature change: `verifyToken(token string)` becomes `verifyToken(token string, opts verifyOpts)`. Existing callers pass `verifyOpts{}` (zero value = current behaviour). Bearer path passes `verifyOpts{skipReplayMarking: true}`.
+
+**Critical semantics — must be reflected in implementation and tests:**
+- `skipReplayMarking` only gates the **Set** at `token_manager.go:108-143` (the call adding the JTI to the blacklist and replay cache).
+- The blacklist **Get** at `token_manager.go:44-47, 80-89` stays unconditionally active on the bearer path. Tokens revoked via `RevokeToken` (which adds the JTI to the blacklist) MUST still be rejected on the bearer path.
+- Must NOT be implemented by mutating `t.disableReplayDetection` (struct field) — that would create a cross-request race that disables replay protection globally.
+
+A targeted regression test exercises: bearer token verified once → admin calls `RevokeToken` adding the JTI to the blacklist → same token replayed → 401.
+
+### 7.8 Config additions (modify `settings.go`)
+
+```go
+EnableBearerAuth              bool   `json:"enableBearerAuth,omitempty"`
+BearerIdentifierClaim         string `json:"bearerIdentifierClaim,omitempty"`
+StripAuthorizationHeader      bool   `json:"stripAuthorizationHeader,omitempty"`
+BearerEmitWWWAuthenticate     bool   `json:"bearerEmitWWWAuthenticate,omitempty"`
+BearerOverridesCookie         bool   `json:"bearerOverridesCookie,omitempty"`
+MaxTokenAgeSeconds            int64  `json:"maxTokenAgeSeconds,omitempty"`
+MaxIdentifierLength           int    `json:"maxIdentifierLength,omitempty"`
+BearerFailureThreshold        int    `json:"bearerFailureThreshold,omitempty"`
+BearerFailureWindowSeconds    int    `json:"bearerFailureWindowSeconds,omitempty"`
+BearerFailurePenaltySeconds   int    `json:"bearerFailurePenaltySeconds,omitempty"`
+```
+
+Defaults (applied in `CreateConfig` for the bearer-related fields; values >0 only honoured when `EnableBearerAuth=true`):
+- `EnableBearerAuth`: `false`.
+- `BearerIdentifierClaim`: `"sub"`.
+- `StripAuthorizationHeader`: `true`.
+- `BearerEmitWWWAuthenticate`: `true` (RFC 6750 hint enabled by default; flip to false if recon-exposure is a concern).
+- `BearerOverridesCookie`: `false` (cookie wins when both present; flip to `true` for the legacy/industry-default behaviour).
+- `MaxTokenAgeSeconds`: `86400` (24h upper bound on `iat`).
+- `MaxIdentifierLength`: `256`.
+- `BearerFailureThreshold`: `20` (consecutive 401s per IP before throttle).
+- `BearerFailureWindowSeconds`: `60`.
+- `BearerFailurePenaltySeconds`: `60` (429 reply for this long after threshold tripped).
+
+### 7.9 Startup validation (modify `main.go` `New()`)
+
+- `EnableBearerAuth && Audience == ""` → fatal error.
+- `EnableBearerAuth && !StrictAudienceValidation` → warning log (recommended hardening).
+- `EnableBearerAuth && BearerIdentifierClaim == "email"` → fatal error (the bearer path is M2M and an `email` identifier without `email_verified` enforcement is a spoofing vector; default `BearerIdentifierClaim=sub` avoids this; explicit override to `email` is rejected).
+- `EnableBearerAuth && MaxTokenAgeSeconds <= 0` → reset to default 86400 with info log.
+- `EnableBearerAuth && BearerFailureThreshold <= 0` → reset to default 20 with info log.
+
+## 8. Data Flow
+
+### 8.1 Bearer path
+
+```
+ServeHTTP entry (pre-init paths unchanged: logout, backchannel, frontchannel, excluded URLs, SSE/WS bypass)
+  │
+  ├─ enableBearerAuth == false?  → fall through to cookie path
+  │
+  └─ enableBearerAuth == true AND Authorization starts with "Bearer "
+       │
+       ▼
+  handleBearerRequest
+       │
+       ├─ format guards (empty, length, segment count)
+       │
+       ▼
+  verifyToken(token, verifyOpts{SkipReplayMarking: true})
+       │  signature, issuer, audience (strict), exp
+       │
+       ▼
+  classifyToken(claims) → reject ID tokens
+       │
+       ▼
+  if requireTokenIntrospection: introspectToken → active check
+       │
+       ▼
+  resolveIdentifier(claims) → sanitiseIdentifier
+       │
+       ▼
+  principal{Source: sourceBearer, …}
+       │
+       ▼
+  forwardAuthorized(rw, req, principal)
+       │
+       ├─ roles/groups gate (403 on deny)
+       ├─ header injection
+       ├─ header templates
+       ├─ security headers
+       ├─ strip OIDC cookies (existing)
+       ├─ strip Authorization header (new, when configured)
+       └─ next.ServeHTTP(rw, req)
+```
+
+### 8.2 Cookie path (refactored, semantically unchanged)
+
+```
+processAuthorizedRequest
+  1. Session validity / backchannel-logout invalidation (unchanged).
+  2. principal := buildPrincipalFromSession(session).
+  3. forwardAuthorized(rw, req, principal).
+  4. if session.IsDirty(): session.Save().
+```
+
+## 9. Error Handling
+
+| Trigger | Status | Body | WWW-Authenticate | Debug log reason |
+|---|---|---|---|---|
+| Empty bearer after prefix | 401 | `Unauthorized` | `Bearer error="invalid_request"` | empty bearer token |
+| Token over MaxLength | 401 | `Unauthorized` | `Bearer error="invalid_token"` | token exceeds max length |
+| Not a 3-segment JWT | 401 | `Unauthorized` | `Bearer error="invalid_token"` | malformed JWT |
+| Disallowed `alg` (e.g. none, HS*) | 401 | `Unauthorized` | `Bearer error="invalid_token"` | unsupported alg |
+| Missing/oversized/bad-charset `kid` | 401 | `Unauthorized` | `Bearer error="invalid_token"` | invalid kid |
+| Signature / issuer / aud / exp fail | 401 | `Unauthorized` | `Bearer error="invalid_token"` | reason from verifyToken (category only) |
+| `iat` older than MaxTokenAgeSeconds | 401 | `Unauthorized` | `Bearer error="invalid_token"` | token too old (iat outside age bound) |
+| Multi-aud without matching `azp` | 401 | `Unauthorized` | `Bearer error="invalid_token"` | multi-aud token without azp match |
+| Detected as ID token | 401 | `Unauthorized` | `Bearer error="invalid_token"` | ID tokens not accepted on bearer path |
+| JTI blacklisted (revoked) | 401 | `Unauthorized` | `Bearer error="invalid_token"` | token JTI in blacklist |
+| Introspection `active=false` | 401 | `Unauthorized` | `Bearer error="invalid_token"` | token inactive at IdP |
+| Introspection endpoint failure | 503 | `Service Unavailable` | (none) | introspection unavailable |
+| Identifier claim missing/empty | 401 | `Unauthorized` | `Bearer error="invalid_token"` | no identifier claim |
+| Identifier fails sanitisation | 401 | `Unauthorized` | `Bearer error="invalid_token"` | invalid identifier characters |
+| Per-IP failure threshold tripped | 429 | `Too Many Requests` | (none); `Retry-After: <BearerFailurePenaltySeconds>` | source IP in penalty box |
+| Roles/groups not allowed | 403 | `Access denied` | (none) | user not in allowedRolesAndGroups |
+
+Responses never include token contents, never include the raw failure reason, and never set `Location` headers (API clients cannot follow redirects).
+
+## 10. Edge Cases
+
+1. **Both bearer header and cookie session present.** Cookie wins by default (safer against browser/extension/proxy bearer injection). `BearerOverridesCookie=true` flips to bearer-wins. Either way: WARN log includes both source markers so operators can audit.
+2. **`Authorization: Basic …`.** Not bearer; cookie path runs as today.
+3. **`Authorization: Bearer ` (trailing space, no value).** Empty after trim → 401.
+4. **Mixed-case prefix (`bearer`, `BEARER`, `BeArEr`).** Case-insensitive prefix check; token value preserved verbatim.
+5. **Multiple `Authorization` headers.** Use only the first (Go `http.Header.Get` default). Documented.
+6. **Bearer during OIDC init wait.** Bearer requests also block on init: we need `issuerURL`, `audience`, JWKs ready. If init fails, bearer requests return 503 just like cookie requests.
+7. **SSE / WebSocket bypass with bearer.** Bypass paths keep cookie-only behaviour. Operators who want bearer on streaming endpoints must remove SSE/WS bypass. Documented.
+8. **Logout endpoint with bearer.** Logout runs before bearer detection. Treated as cookie-session logout; bearer token revocation requires IdP-side action.
+9. **Excluded URLs with bearer.** Bypass excluded URLs as today; bearer not validated on excluded paths. ADDITIONALLY: `Authorization: Bearer` is stripped from the request before forwarding so the token can't leak into the excluded endpoint's downstream logs / metrics scrapers / health checks.
+10. **Concurrent identical bearer requests.** Existing `tokenCache` is concurrency-safe; no new locking.
+11. **Client rotates token between requests.** Independent verification per token; independent cache entries.
+12. **Clock skew.** Use existing `jwt.Verify` leeway. (If absent, add ±30s as a separate change; out of scope here.)
+
+## 11. Testing Strategy
+
+### 11.1 Integration tests (new `bearer_auth_test.go`)
+
+Table-driven test against a real `httptest.Server` and the full `ServeHTTP` flow. Coverage matrix:
+
+- Valid access token + allowed roles → 200, `next` ran, `X-Forwarded-User` set.
+- Valid token without configured roles → 200.
+- Wrong audience, expired, tampered signature → 401, `next` did not run.
+- ID token presented → 401 (`ID tokens not accepted`).
+- Malformed JWT (2 segments) → 401.
+- Oversized token (> MaxLength) → 401.
+- Empty bearer → 401.
+- Missing identifier claim → 401.
+- Identifier containing `\r\n` → 401.
+- `allowedRolesAndGroups` mismatch → 403.
+- `allowedRolesAndGroups` match → 200.
+- `EnableBearerAuth=false` + bearer header → cookie path runs (302 to `/authorize`).
+- Bearer + valid cookie session → bearer wins, 200.
+- `StripAuthorizationHeader=true` → downstream sees no `Authorization`.
+- `StripAuthorizationHeader=false` → downstream sees `Authorization`.
+- Case variants (`bearer`, `BEARER`) → 200.
+- SSE bypass + bearer → cookie-only check applies (bearer ignored).
+- **Replay regression**: same token 1000 times in a row → all 200.
+- **Cache-evict regression**: same token, force-evict `tokenCache` between iterations (call `tokenCache.Delete` directly), replay → still 200 (verifies `skipReplayMarking` doesn't poison the blacklist).
+- **Revocation-while-bearer regression**: bearer token verified once → admin calls `RevokeToken` adding JTI to blacklist → same token presented → 401 (verifies blacklist Get stays active on bearer path even with `skipReplayMarking` set).
+- **Alg-pin: token signed with `alg=none`** → 401, no JWKS fetch happens (verify with a counting mock).
+- **`kid` injection: 50KB random kid** → 401 immediately, no JWKS fetch.
+- **Per-IP throttle**: 21 bad bearer requests from same IP within 1 minute → 22nd returns 429 + Retry-After.
+- **`iat` upper-age**: token with `iat = now - 25h` → 401 (older than 24h default).
+- **Multi-aud without azp**: aud = `["a", "b"]`, no azp → 401.
+- **Multi-aud with matching azp**: aud = `["api-aud", "other"]`, azp = clientID → 200.
+- **Identifier with bidi-override**: sub contains U+202E → 401.
+- **Identifier with comma**: sub = `"alice,bob"` → 401.
+- **Identifier over 256 bytes** → 401.
+- **`UserIdentifierClaim=email` at startup with EnableBearerAuth=true** → startup fails.
+- **Excluded URL + bearer**: bearer header presented on excluded URL → request forwarded, downstream sees no `Authorization` header (stripped).
+
+### 11.2 Unit tests (in `bearer_auth_test.go`)
+
+- `classifyToken`: ID-token detection, access-token detection by `scope`/`scp`/`token_use`, ambiguous → reject.
+- `resolveIdentifier`: precedence (`userIdentifierClaim` → `sub` → `client_id`/`azp`); missing → error; empty string → error.
+- `sanitizeIdentifier`: rejects all `unicode.IsControl`; accepts email/sub-style values.
+
+### 11.3 Introspection tests (`bearer_auth_introspection_test.go`)
+
+- Token valid + introspection `active=true` → 200.
+- Token valid + introspection `active=false` → 401.
+- Introspection endpoint 500 → 503.
+- Second request hits introspection cache (no second HTTP call).
+
+### 11.4 Startup validation tests (extend `settings_test.go` / `main_test.go`)
+
+- `EnableBearerAuth=true, Audience=""` → `New()` errors.
+- `EnableBearerAuth=true, StrictAudienceValidation=false` → succeeds with warning.
+- `EnableBearerAuth=false` → no validation; existing tests untouched.
+
+### 11.5 Cookie-path regression suite
+
+- All existing `TestServeHTTP_*` tests in `main_servehttp_test.go` pass unmodified.
+- Add: cookie session, `EnableBearerAuth=true`, no bearer header → identical behaviour to baseline.
+- Add: dirty session still triggers `Save()` after refactor.
+
+### 11.6 Principal invariants
+
+- `buildPrincipalFromSession`: `Source == sourceSession`; `IDToken` / `RefreshToken` populated when present in session.
+- `buildPrincipalFromBearerToken`: `Source == sourceBearer`; `IDToken == ""`, `RefreshToken == ""`.
+- `forwardAuthorized` produces identical headers for equivalent principals regardless of source.
+
+### 11.7 Coverage gate
+
+- New code in `bearer_auth.go` and `principal.go`: ≥ 90% line coverage.
+- `forwardAuthorized` coverage ≥ existing `processAuthorizedRequest` coverage baseline.
+
+### 11.8 Out of scope (follow-ups)
+
+- Load test of bearer vs cookie hot path.
+- Fuzzing the JWT parser.
+- Additional auth methods (mTLS, API keys) — design enables them, but they are separate work.
+
+## 12. Migration / Rollout
+
+Default-off. Existing deployments observe no behavioural change. Operators opt in by setting:
+
+```yaml
+enableBearerAuth: true
+audience: https://api.example.com   # required when bearer enabled
+# optional:
+stripAuthorizationHeader: true       # default
+requireTokenIntrospection: false     # default; set true for real-time revocation
+userIdentifierClaim: client_id       # optional override; defaults to sub fallback chain
+```
+
+Documentation: update `docs/CONFIGURATION.md` with a bearer-auth section, and add a new `docs/BEARER_AUTH.md` covering the security model, threat assumptions (token issuer is trusted; audience must be set; bearer means trust the issuer's revocation policy unless introspection enabled), and recommended configurations for common IdPs.
+
+## 13. Security Considerations
+
+| Concern | Mitigation |
+|---|---|
+| Token confusion (ID token used as bearer) | Reuse `detectTokenType` (`token_manager.go:187-303`) which checks `nonce`, `typ: at+jwt`, `token_use`, `scope`, aud-vs-clientID. Belt-and-braces: explicit `nonce` + `token_use == "id"` rejection on top. |
+| Audience confusion (token for service B accepted by A) | `Audience` mandatory at startup; verified via existing `VerifyJWTSignatureAndClaims`; multi-aud tokens require matching `azp == clientID`. |
+| Replay-via-blacklist false positive | `verifyOpts{skipReplayMarking: true}` on bearer path. Gates ONLY the Set; the Get stays so revoked tokens still fail. |
+| Revocation lag | Optional RFC 7662 introspection. Bearer-path introspection cache TTL capped at 60s. Set `RequireTokenIntrospection=true` for real-time revocation. |
+| `alg`-confusion / `alg=none` attacks | Hard-pin asymmetric allowlist at bearer entry, **before** JWKS fetch. Prevents wasted upstream calls and locks out HS/none probes. |
+| `kid` injection / JWKS amplification | `kid` length cap (256 bytes) + charset allowlist enforced at bearer entry. |
+| Bearer 401 brute-force / oracle | Per-IP `failedBearerAttempts` cache; configurable threshold + penalty box returning 429 + `Retry-After`. |
+| `iat` clock-manipulation / forever-tokens | `MaxTokenAgeSeconds` upper bound (default 24h); cookie path unchanged. |
+| Identifier-driven header injection | `sanitizeIdentifier`: length cap, control-char + bidi-override + `,;=` rejection. `net/http` rejects CRLF on the wire too (defence in depth). |
+| Token leakage downstream | `StripAuthorizationHeader=true` by default. Also: `Authorization` stripped on excluded-URL requests so bearer can't leak into health/metrics downstream logs. |
+| Token-in-logs | All log paths log reason categories, not raw tokens. Identifier hashed via SHA-256 truncated to 8 hex chars before any info/warn-level emission (full identifier only at debug). New `safeLogAuthEvent(category, hashedIdentifier, reasonCode)` helper makes this hard to misuse. |
+| `email` claim spoofing | Startup fails if `EnableBearerAuth && UserIdentifierClaim == "email"`. Future human-user bearer iteration must add `email_verified` enforcement. |
+| Bypass on SSE / WS endpoints | SSE/WS bypass keeps cookie-only behaviour; bearer ignored. Operators choose to widen if needed. |
+| Mixed bearer + cookie precedence | Cookie wins by default (safer for browser scenarios); `BearerOverridesCookie=true` flips. WARN log on both-present requests. |
+| Configuration drift (operator forgets audience) | Startup fails when `EnableBearerAuth=true && Audience==""`. |
+| Downstream blast radius when `StripAuthorizationHeader=false` | Documented: forwarded bearer extends token's blast radius to all downstream services. Logs at those services become token stores. Operators must treat downstream log policy accordingly. |
+| Introspection auth method (pre-existing gap, called out) | `token_introspection.go:80` uses `client_secret_basic` only; does not honour `private_key_jwt`. Out of scope for this PR but documented as a follow-up; operators using `ClientAuthMethod=private_key_jwt` + `RequireTokenIntrospection=true` should be aware introspection will use basic auth. |
+
+## 14. Open Questions
+
+None — all design decisions resolved during brainstorming + security review. Implementation may surface incidental questions (e.g. exact clock-skew leeway in `jwt.Verify`); those are out of scope for this spec and handled in the implementation plan.
+
+## 14a. Security Review Reference
+
+This design was reviewed by the `security-reviewer` subagent on 2026-05-18. Findings incorporated:
+
+- **Critical**: C1 (classifier reuses `detectTokenType`), C2 (sub fallback dropped — unreachable due to `jwt.go:416`), C3 (replay-marking gates only Set, not Get; revocation regression test added).
+- **High**: H1 (alg pinned at bearer entry), H2 (kid length + charset), H3 (cookie wins by default, configurable), H4 (per-IP 401 throttle), H5 (multi-aud requires azp).
+- **Medium**: M1 (identifier max-length + bidi reject + delimiter chars), M2 (introspection cache TTL capped at 60s on bearer path), M4 (log-hashing via SHA-256[:8]), M5 (StripAuth blast-radius documented), M6 (iat upper-age bound), M7 (Authorization stripped on excluded URLs).
+- **Low/Nit**: L2 (renamed to `BearerEmitWWWAuthenticate`), N3 (startup rejects `UserIdentifierClaim=email`).
+- **Documented as pre-existing gaps (follow-up PRs)**: M3 (introspection auth method doesn't honour `private_key_jwt`).
+
+## 15. Implementation Plan Reference
+
+To be produced by the `writing-plans` skill in a follow-up document at `docs/superpowers/plans/2026-05-18-bearer-token-auth-plan.md`. The plan decomposes this design into ordered, independently-testable PRs.

internal/cache/backends/hybrid.go

@@ -164,7 +164,7 @@ func (h *HybridBackend) Set(ctx context.Context, key string, value []byte, ttl t
 
 	// Check if we're in fallback mode
 	if h.fallbackMode.Load() {
-		h.logger.Debugf("Operating in fallback mode, skipping L2 write for key: %s", key)
+		h.logger.Debugf("Operating in fallback mode, skipping L2 write for key: %s", redactKey(key))
 		return nil // Don't fail the operation if L2 is down
 	}
 
@@ -176,13 +176,13 @@ func (h *HybridBackend) Set(ctx context.Context, key string, value []byte, ttl t
 		// Synchronous write for critical cache types
 		if err := h.secondary.Set(ctx, key, value, ttl); err != nil {
 			h.errors.Add(1)
-			h.logger.Warnf("Failed to write to L2 cache (sync) for key %s: %v", key, err)
+			h.logger.Warnf("Failed to write to L2 cache (sync) for key %s: %v", redactKey(key), err)
 			h.recordL2Error()
 			// Don't fail the operation - L1 write succeeded
 			return nil
 		}
 		h.l2Writes.Add(1)
-		h.logger.Debugf("Synchronous write to L2 completed for critical key: %s", key)
+		h.logger.Debugf("Synchronous write to L2 completed for critical key: %s", redactKey(key))
 	} else {
 		// Asynchronous write for non-critical cache types
 		select {
@@ -192,10 +192,10 @@ func (h *HybridBackend) Set(ctx context.Context, key string, value []byte, ttl t
 			ttl:   ttl,
 			ctx:   ctx,
 		}:
-			h.logger.Debugf("Queued async write to L2 for key: %s", key)
+			h.logger.Debugf("Queued async write to L2 for key: %s", redactKey(key))
 		default:
 			// Buffer is full, log and continue
-			h.logger.Warnf("Async write buffer full, dropping L2 write for key: %s", key)
+			h.logger.Warnf("Async write buffer full, dropping L2 write for key: %s", redactKey(key))
 			h.errors.Add(1)
 		}
 	}
@@ -209,7 +209,7 @@ func (h *HybridBackend) Get(ctx context.Context, key string) ([]byte, time.Durat
 	value, ttl, exists, err := h.primary.Get(ctx, key)
 	if err != nil {
 		h.errors.Add(1)
-		h.logger.Debugf("L1 get error for key %s: %v", key, err)
+		h.logger.Debugf("L1 get error for key %s: %v", redactKey(key), err)
 	}
 
 	if exists {
@@ -227,7 +227,7 @@ func (h *HybridBackend) Get(ctx context.Context, key string) ([]byte, time.Durat
 	value, ttl, exists, err = h.secondary.Get(ctx, key)
 	if err != nil {
 		h.errors.Add(1)
-		h.logger.Debugf("L2 get error for key %s: %v", key, err)
+		h.logger.Debugf("L2 get error for key %s: %v", redactKey(key), err)
 		h.recordL2Error()
 		h.misses.Add(1)
 		return nil, 0, false, nil // Don't propagate L2 errors
@@ -544,7 +544,7 @@ func (h *HybridBackend) queueL1Backfill(key string, value []byte, ttl time.Durat
 	case h.l1BackfillBuffer <- &l1BackfillItem{key: key, value: value, ttl: ttl}:
 	default:
 		h.l1BackfillDrops.Add(1)
-		h.logger.Debugf("L1 backfill buffer full, dropping for key: %s", key)
+		h.logger.Debugf("L1 backfill buffer full, dropping for key: %s", redactKey(key))
 	}
 }
 
@@ -576,9 +576,9 @@ func (h *HybridBackend) l1BackfillWorker() {
 			}
 			writeCtx, cancel := context.WithTimeout(context.Background(), 100*time.Millisecond)
 			if err := h.primary.Set(writeCtx, item.key, item.value, item.ttl); err != nil {
-				h.logger.Debugf("Failed to populate L1 cache from L2 for key %s: %v", item.key, err)
+				h.logger.Debugf("Failed to populate L1 cache from L2 for key %s: %v", redactKey(item.key), err)
 			} else {
-				h.logger.Debugf("Populated L1 cache from L2 for key: %s", item.key)
+				h.logger.Debugf("Populated L1 cache from L2 for key: %s", redactKey(item.key))
 			}
 			cancel()
 		}
@@ -619,11 +619,11 @@ func (h *HybridBackend) asyncWriteWorker() {
 			writeCtx, cancel := context.WithTimeout(item.ctx, 500*time.Millisecond)
 			if err := h.secondary.Set(writeCtx, item.key, item.value, item.ttl); err != nil {
 				h.errors.Add(1)
-				h.logger.Debugf("Async write to L2 failed for key %s: %v", item.key, err)
+				h.logger.Debugf("Async write to L2 failed for key %s: %v", redactKey(item.key), err)
 				h.recordL2Error()
 			} else {
 				h.l2Writes.Add(1)
-				h.logger.Debugf("Async write to L2 completed for key: %s", item.key)
+				h.logger.Debugf("Async write to L2 completed for key: %s", redactKey(item.key))
 			}
 			cancel()
 		}

internal/cache/backends/log_redact.go

@@ -0,0 +1,26 @@
+// Package backends provides cache backend implementations for the Traefik OIDC plugin.
+package backends
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+)
+
+// redactKey returns a short, deterministic hash prefix of a cache key for use
+// in debug/info log lines. Cache keys in this plugin can include raw access /
+// refresh / id tokens (any caller may pass an arbitrary string), and CodeQL
+// flags `key=%s` formatters as a clear-text-logging sink for HTTP-header-
+// sourced taint. The hash preserves cache-key uniqueness in logs (same key →
+// same hash, useful for correlating a problematic key across log lines) while
+// keeping the raw value out of disk-resident log streams.
+//
+// 8 hex chars (32 bits) is enough to disambiguate at human-debugging scale
+// without making the hash itself a useful lookup primitive for an attacker
+// who only has the log stream.
+func redactKey(key string) string {
+	if key == "" {
+		return "(empty)"
+	}
+	sum := sha256.Sum256([]byte(key))
+	return hex.EncodeToString(sum[:4])
+}

internal/cache/cache.go

@@ -190,7 +190,7 @@ func (c *Cache) Set(key string, value interface{}, ttl time.Duration) error {
 	c.currentSize++
 	atomic.AddInt64(&c.sets, 1)
 
-	c.logger.Debugf("Cache: Set key=%s, size=%d, ttl=%v", key, size, ttl)
+	c.logger.Debugf("Cache: Set key=%s, size=%d, ttl=%v", redactKey(key), size, ttl)
 	return nil
 }
 
@@ -346,7 +346,7 @@ func (c *Cache) evictLRU() {
 		item, _ := elem.Value.(*Item) // Safe to ignore: type assertion from known type
 		c.removeItem(item.Key, item)
 		atomic.AddInt64(&c.evictions, 1)
-		c.logger.Debugf("Cache: Evicted LRU item key=%s", item.Key)
+		c.logger.Debugf("Cache: Evicted LRU item key=%s", redactKey(item.Key))
 	}
 }
 

internal/cache/log_redact.go

@@ -0,0 +1,22 @@
+// Package cache provides the in-memory cache implementation for the Traefik
+// OIDC plugin.
+package cache
+
+import (
+	"crypto/sha256"
+	"encoding/hex"
+)
+
+// redactKey returns a short, deterministic hash prefix of a cache key for use
+// in debug/info log lines. Cache keys may include raw access / refresh / id
+// tokens (callers pass arbitrary strings) and CodeQL flags `key=%s`
+// formatters as a clear-text-logging sink for HTTP-header-sourced taint.
+// The hash preserves uniqueness in logs (same key → same hash) while keeping
+// the raw value out of disk-resident log streams.
+func redactKey(key string) string {
+	if key == "" {
+		return "(empty)"
+	}
+	sum := sha256.Sum256([]byte(key))
+	return hex.EncodeToString(sum[:4])
+}

issue134_followup_graph_test.go

@@ -0,0 +1,449 @@
+package traefikoidc
+
+import (
+	"bytes"
+	"crypto"
+	"crypto/rand"
+	"crypto/rsa"
+	"crypto/sha256"
+	"encoding/base64"
+	"encoding/json"
+	"fmt"
+	"io"
+	"log"
+	"math/big"
+	"net/http"
+	"testing"
+	"time"
+
+	"github.com/gorilla/sessions"
+	"github.com/stretchr/testify/assert"
+	"github.com/stretchr/testify/require"
+	"golang.org/x/time/rate"
+)
+
+// signGraphStyleAccessToken builds a JWT in Microsoft's Graph proprietary
+// nonce-header form: bytes that get signed contain the SHA256 hash of the
+// nonce, while the wire token ships the original nonce. A standard JWS
+// verifier always rejects these with `crypto/rsa: verification error`, which
+// is why Microsoft documents Graph access tokens as opaque to client apps:
+//
+//	https://learn.microsoft.com/en-us/entra/identity-platform/access-tokens
+//	"you can't validate tokens for Microsoft Graph according to these rules
+//	 due to their proprietary format"
+func signGraphStyleAccessToken(t *testing.T, key *rsa.PrivateKey, kid, originalNonce string, claims map[string]any) string {
+	t.Helper()
+
+	wireHeader := map[string]any{
+		"alg":   "RS256",
+		"kid":   kid,
+		"typ":   "JWT",
+		"nonce": originalNonce,
+	}
+	wireHeaderJSON, err := json.Marshal(wireHeader)
+	require.NoError(t, err)
+
+	hashed := sha256.Sum256([]byte(originalNonce))
+	signedHeader := map[string]any{
+		"alg":   "RS256",
+		"kid":   kid,
+		"typ":   "JWT",
+		"nonce": fmt.Sprintf("%x", hashed),
+	}
+	signedHeaderJSON, err := json.Marshal(signedHeader)
+	require.NoError(t, err)
+
+	claimsJSON, err := json.Marshal(claims)
+	require.NoError(t, err)
+
+	wireHeaderB64 := base64.RawURLEncoding.EncodeToString(wireHeaderJSON)
+	signedHeaderB64 := base64.RawURLEncoding.EncodeToString(signedHeaderJSON)
+	claimsB64 := base64.RawURLEncoding.EncodeToString(claimsJSON)
+
+	signedInput := signedHeaderB64 + "." + claimsB64
+	hSign := sha256.Sum256([]byte(signedInput))
+	sig, err := rsa.SignPKCS1v15(rand.Reader, key, crypto.SHA256, hSign[:])
+	require.NoError(t, err)
+
+	return wireHeaderB64 + "." + claimsB64 + "." + base64.RawURLEncoding.EncodeToString(sig)
+}
+
+// newAzureFollowupOIDC produces a TraefikOidc instance wired for an Azure
+// AD tenant with a captured error log buffer. Used by the issue #134 followup
+// tests to assert log behavior during validateAzureTokens flows.
+func newAzureFollowupOIDC(t *testing.T, jwks *JWKSet) (*TraefikOidc, *bytes.Buffer) {
+	t.Helper()
+	tc := newTestCleanup(t)
+
+	errBuf := &bytes.Buffer{}
+	logger := &Logger{
+		logError: log.New(errBuf, "", 0),
+		logInfo:  log.New(io.Discard, "", 0),
+		logDebug: log.New(io.Discard, "", 0),
+	}
+
+	tokenCache := tc.addTokenCache(NewTokenCache())
+	tokenBlacklist := tc.addCache(NewCache())
+
+	oidc := &TraefikOidc{
+		issuerURL:         "https://login.microsoftonline.com/tenant-id/v2.0",
+		clientID:          "test-client-id",
+		audience:          "test-client-id",
+		jwksURL:           "https://login.microsoftonline.com/tenant-id/discovery/v2.0/keys",
+		limiter:           rate.NewLimiter(rate.Every(time.Second), 100),
+		logger:            logger,
+		httpClient:        &http.Client{Timeout: 10 * time.Second},
+		jwkCache:          &MockJWKCache{JWKS: jwks},
+		tokenCache:        tokenCache,
+		tokenBlacklist:    tokenBlacklist,
+		extractClaimsFunc: extractClaims,
+	}
+	oidc.tokenVerifier = oidc
+	oidc.jwtVerifier = oidc
+	require.True(t, oidc.isAzureProvider(), "fixture must be detected as Azure provider")
+	return oidc, errBuf
+}
+
+// authedSessionWithTokens returns a SessionData populated with the supplied
+// access and ID tokens, marked authenticated and recently created. The
+// SessionManager carries a real ChunkManager so that GetAccessToken /
+// GetIDToken / GetRefreshToken behave like the production code path.
+func authedSessionWithTokens(t *testing.T, accessToken, idToken string) *SessionData {
+	t.Helper()
+
+	chunkLogger := NewLogger("error")
+	chunkManager := NewChunkManager(chunkLogger)
+	t.Cleanup(chunkManager.Shutdown)
+
+	sd := CreateMockSessionData()
+	sd.manager = &SessionManager{
+		sessionMaxAge: 24 * time.Hour,
+		chunkManager:  chunkManager,
+		logger:        chunkLogger,
+	}
+
+	sd.mainSession = sessions.NewSession(nil, "main")
+	sd.mainSession.Values["authenticated"] = true
+	sd.mainSession.Values["created_at"] = time.Now().Unix()
+
+	sd.accessSession = sessions.NewSession(nil, "access")
+	sd.accessSession.Values["token"] = accessToken
+	sd.accessSession.Values["compressed"] = false
+
+	sd.idTokenSession = sessions.NewSession(nil, "id")
+	sd.idTokenSession.Values["token"] = idToken
+	sd.idTokenSession.Values["compressed"] = false
+
+	sd.refreshSession = sessions.NewSession(nil, "refresh")
+	sd.refreshSession.Values["token"] = ""
+	sd.refreshSession.Values["compressed"] = false
+
+	return sd
+}
+
+// TestIssue134_Followup_GraphAccessTokenReproducesUsersError sanity-checks
+// that our crafted Graph-style token reproduces the exact rsa error string
+// quoted on the issue thread (dada-engineer 2026-05-08, friek 2026-05-11).
+//
+// Sanity test: must always pass, regardless of the issue #134 followup fix.
+// It exists so a future contributor does not accidentally weaken the
+// reproducer and assume the followup fix is no longer needed.
+func TestIssue134_Followup_GraphAccessTokenReproducesUsersError(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-followup-kid"
+	graphToken := signGraphStyleAccessToken(t, rsaKey, kid, "wire-only-nonce", map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "00000003-0000-0000-c000-000000000000",
+		"exp": time.Now().Add(time.Hour).Unix(),
+		"iat": time.Now().Unix(),
+		"sub": "user-azure-id",
+		"scp": "User.Read",
+	})
+
+	parsedJWT, err := parseJWT(graphToken)
+	require.NoError(t, err)
+	pubKey := &rsaKey.PublicKey
+	alg, _ := parsedJWT.Header["alg"].(string)
+	verifyErr := verifySignatureWithKey(graphToken, pubKey, alg)
+	require.Error(t, verifyErr)
+	assert.Contains(t, verifyErr.Error(), "crypto/rsa: verification error",
+		"reproducer must emit the exact error string reported on issue #134")
+}
+
+// TestIssue134_Followup_ValidateAzureTokensSkipsGraphAccessToken is the
+// failing-then-passing test for the followup fix.
+//
+// Symptom (before fix): validateAzureTokens calls verifyToken on every
+// JWT-shaped access token. For Microsoft Graph access tokens (the default
+// when no custom resource is registered), verification always fails with
+// `crypto/rsa: verification error`, generating two error log lines per
+// request:
+//
+//	UNKNOWN token verification failed: signature verification failed:
+//	  crypto/rsa: verification error
+//	DIAGNOSTIC: Signature verification failed for kid=<kid>, alg=RS256:
+//	  crypto/rsa: verification error
+//
+// Microsoft's own documentation tells client apps not to validate Graph
+// access tokens. The fix matches that guidance: when an Azure access token
+// carries Microsoft's proprietary `nonce` JWT header, treat it as opaque
+// (skip JWT verification, fall through to ID token validation).
+func TestIssue134_Followup_ValidateAzureTokensSkipsGraphAccessToken(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-followup-kid"
+	jwk := JWK{
+		Kty: "RSA",
+		Use: "sig",
+		Alg: "RS256",
+		Kid: kid,
+		N:   base64.RawURLEncoding.EncodeToString(rsaKey.N.Bytes()),
+		E:   base64.RawURLEncoding.EncodeToString(big.NewInt(int64(rsaKey.E)).Bytes()),
+	}
+	jwks := &JWKSet{Keys: []JWK{jwk}}
+
+	now := time.Now()
+	exp := now.Add(time.Hour).Unix()
+
+	graphAccessToken := signGraphStyleAccessToken(t, rsaKey, kid, "wire-only-nonce-azure-graph", map[string]any{
+		"iss":   "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud":   "00000003-0000-0000-c000-000000000000",
+		"exp":   exp,
+		"iat":   now.Unix(),
+		"sub":   "user-azure-id",
+		"appid": "test-client-id",
+		"scp":   "User.Read",
+	})
+
+	idToken, err := createTestJWT(rsaKey, "RS256", kid, map[string]any{
+		"iss":   "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud":   "test-client-id",
+		"exp":   exp,
+		"iat":   now.Add(-2 * time.Minute).Unix(),
+		"nbf":   now.Add(-2 * time.Minute).Unix(),
+		"sub":   "user-azure-id",
+		"email": "user@example.com",
+		"nonce": "id-token-oidc-nonce",
+		"jti":   "id-token-jti-followup",
+	})
+	require.NoError(t, err)
+
+	oidc, errBuf := newAzureFollowupOIDC(t, jwks)
+	session := authedSessionWithTokens(t, graphAccessToken, idToken)
+
+	authenticated, needsRefresh, expired := oidc.validateAzureTokens(session)
+
+	output := errBuf.String()
+	assert.NotContains(t, output, "crypto/rsa: verification error",
+		"validateAzureTokens must not log rsa verification error for Graph-style access tokens; got: %q", output)
+	assert.NotContains(t, output, "DIAGNOSTIC: Signature verification failed",
+		"DIAGNOSTIC line must not fire for Graph-style access tokens; got: %q", output)
+	assert.NotContains(t, output, "UNKNOWN token verification failed",
+		"UNKNOWN classification log must not fire for Graph-style access tokens; got: %q", output)
+
+	assert.True(t, authenticated, "session must remain authenticated via the ID token fallback")
+	assert.False(t, needsRefresh, "valid ID token must not signal a refresh need")
+	assert.False(t, expired, "valid ID token must not be reported as expired")
+}
+
+// TestIssue134_Followup_IsUnverifiableAzureAccessToken_Detection covers the
+// classifier added by the followup fix. Pure-function unit test for the
+// Microsoft proprietary marker we rely on (nonce in JWT header).
+func TestIssue134_Followup_IsUnverifiableAzureAccessToken_Detection(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-detection-kid"
+	standardToken, err := createTestJWT(rsaKey, "RS256", kid, map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "test-client-id",
+		"exp": time.Now().Add(time.Hour).Unix(),
+		"iat": time.Now().Unix(),
+		"sub": "user-azure-id",
+	})
+	require.NoError(t, err)
+
+	graphToken := signGraphStyleAccessToken(t, rsaKey, kid, "wire-only-nonce", map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "00000003-0000-0000-c000-000000000000",
+		"exp": time.Now().Add(time.Hour).Unix(),
+		"iat": time.Now().Unix(),
+		"sub": "user-azure-id",
+		"scp": "User.Read",
+	})
+
+	oidc, _ := newAzureFollowupOIDC(t, &JWKSet{})
+
+	cases := []struct {
+		name           string
+		token          string
+		wantUnverified bool
+	}{
+		{name: "standard JWT without nonce header", token: standardToken, wantUnverified: false},
+		{name: "Microsoft proprietary token (nonce in header)", token: graphToken, wantUnverified: true},
+		{name: "garbage token treated as unverifiable", token: "not-a-jwt-at-all", wantUnverified: true},
+		{name: "empty token treated as unverifiable", token: "", wantUnverified: true},
+	}
+
+	for _, tc := range cases {
+		t.Run(tc.name, func(t *testing.T) {
+			got := oidc.isUnverifiableAzureAccessToken(tc.token)
+			assert.Equal(t, tc.wantUnverified, got)
+		})
+	}
+}
+
+// TestIssue134_Followup_StandardAzureAccessTokenStillVerifies guards against
+// regression in the happy path: an access token issued for our own clientID
+// (custom Azure-registered API) — no proprietary nonce header, signed normally
+// — must still flow through the standard verification path and authenticate.
+func TestIssue134_Followup_StandardAzureAccessTokenStillVerifies(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-standard-kid"
+	jwk := JWK{
+		Kty: "RSA", Use: "sig", Alg: "RS256", Kid: kid,
+		N: base64.RawURLEncoding.EncodeToString(rsaKey.N.Bytes()),
+		E: base64.RawURLEncoding.EncodeToString(big.NewInt(int64(rsaKey.E)).Bytes()),
+	}
+	jwks := &JWKSet{Keys: []JWK{jwk}}
+
+	now := time.Now()
+	exp := now.Add(time.Hour).Unix()
+
+	// Custom-resource access token: aud points to the app, no nonce header.
+	accessToken, err := createTestJWT(rsaKey, "RS256", kid, map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "test-client-id",
+		"exp": exp,
+		"iat": now.Add(-2 * time.Minute).Unix(),
+		"nbf": now.Add(-2 * time.Minute).Unix(),
+		"sub": "user-azure-id",
+		"scp": "api.read",
+		"jti": "standard-access-jti",
+	})
+	require.NoError(t, err)
+
+	idToken, err := createTestJWT(rsaKey, "RS256", kid, map[string]any{
+		"iss":   "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud":   "test-client-id",
+		"exp":   exp,
+		"iat":   now.Add(-2 * time.Minute).Unix(),
+		"nbf":   now.Add(-2 * time.Minute).Unix(),
+		"sub":   "user-azure-id",
+		"email": "user@example.com",
+		"nonce": "id-token-oidc-nonce",
+		"jti":   "standard-id-jti",
+	})
+	require.NoError(t, err)
+
+	oidc, errBuf := newAzureFollowupOIDC(t, jwks)
+	session := authedSessionWithTokens(t, accessToken, idToken)
+
+	authenticated, needsRefresh, expired := oidc.validateAzureTokens(session)
+
+	assert.True(t, authenticated, "standard Azure access token must verify and authenticate")
+	assert.False(t, needsRefresh)
+	assert.False(t, expired)
+	assert.NotContains(t, errBuf.String(), "crypto/rsa: verification error",
+		"standard Azure token must not produce signature errors")
+}
+
+// TestIssue134_Followup_GraphAccessTokenWithoutIDToken covers the edge where
+// the session has only a Graph access token (no ID token). The classifier must
+// preserve the existing "treat as opaque" semantics for backward compatibility:
+// authenticated=true even when there is no ID token to verify.
+func TestIssue134_Followup_GraphAccessTokenWithoutIDToken(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-no-idt-kid"
+	jwk := JWK{
+		Kty: "RSA", Use: "sig", Alg: "RS256", Kid: kid,
+		N: base64.RawURLEncoding.EncodeToString(rsaKey.N.Bytes()),
+		E: base64.RawURLEncoding.EncodeToString(big.NewInt(int64(rsaKey.E)).Bytes()),
+	}
+	jwks := &JWKSet{Keys: []JWK{jwk}}
+
+	graphAccessToken := signGraphStyleAccessToken(t, rsaKey, kid, "wire-only-nonce-no-idt", map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "00000003-0000-0000-c000-000000000000",
+		"exp": time.Now().Add(time.Hour).Unix(),
+		"iat": time.Now().Unix(),
+		"sub": "user-azure-id",
+		"scp": "User.Read",
+	})
+
+	oidc, errBuf := newAzureFollowupOIDC(t, jwks)
+	session := authedSessionWithTokens(t, graphAccessToken, "")
+
+	authenticated, needsRefresh, expired := oidc.validateAzureTokens(session)
+
+	assert.True(t, authenticated, "Graph token without ID token must remain authenticated (matches existing opaque-token semantics)")
+	assert.False(t, needsRefresh)
+	assert.False(t, expired)
+	assert.NotContains(t, errBuf.String(), "crypto/rsa: verification error")
+}
+
+// TestIssue134_Followup_ConfusedDeputyAttackDoesNotBypassVerification proves
+// the classifier is not a security regression. An attacker who forges a JWT
+// with a `nonce` JWT header (Microsoft's proprietary marker) but a payload
+// claiming `aud=our-clientID` should NOT gain authenticated status simply by
+// triggering the "treat as opaque" branch.
+//
+// This is the confused-deputy guardrail Microsoft warns about
+// (https://cwe.mitre.org/data/definitions/441.html): we treat the access token
+// as opaque, which means we DO NOT authorize from it — authorization comes
+// only from a separately verifiable ID token.  An attacker without a valid ID
+// token must not be authenticated.
+func TestIssue134_Followup_ConfusedDeputyAttackDoesNotBypassVerification(t *testing.T) {
+	rsaKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+	attackerKey, err := rsa.GenerateKey(rand.Reader, 2048)
+	require.NoError(t, err)
+
+	const kid = "azure-attack-kid"
+	jwk := JWK{
+		Kty: "RSA", Use: "sig", Alg: "RS256", Kid: kid,
+		N: base64.RawURLEncoding.EncodeToString(rsaKey.N.Bytes()),
+		E: base64.RawURLEncoding.EncodeToString(big.NewInt(int64(rsaKey.E)).Bytes()),
+	}
+	jwks := &JWKSet{Keys: []JWK{jwk}}
+
+	// Forged: attacker uses their OWN key, sets aud = our clientID, plants a
+	// `nonce` header to trip the opaque-detection path.
+	forgedAccessToken := signGraphStyleAccessToken(t, attackerKey, kid, "attacker-nonce", map[string]any{
+		"iss": "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud": "test-client-id",
+		"exp": time.Now().Add(time.Hour).Unix(),
+		"iat": time.Now().Unix(),
+		"sub": "attacker",
+		"scp": "admin",
+	})
+
+	// Forged ID token signed with the attacker's key — must fail verification
+	// against the tenant JWKS.
+	forgedIDToken, err := createTestJWT(attackerKey, "RS256", kid, map[string]any{
+		"iss":   "https://login.microsoftonline.com/tenant-id/v2.0",
+		"aud":   "test-client-id",
+		"exp":   time.Now().Add(time.Hour).Unix(),
+		"iat":   time.Now().Add(-2 * time.Minute).Unix(),
+		"nbf":   time.Now().Add(-2 * time.Minute).Unix(),
+		"sub":   "attacker",
+		"email": "attacker@evil.example",
+		"nonce": "id-token-oidc-nonce",
+		"jti":   "attacker-id-jti",
+	})
+	require.NoError(t, err)
+
+	oidc, _ := newAzureFollowupOIDC(t, jwks)
+	session := authedSessionWithTokens(t, forgedAccessToken, forgedIDToken)
+
+	authenticated, _, _ := oidc.validateAzureTokens(session)
+	assert.False(t, authenticated,
+		"attacker's forged tokens must not authenticate even when the access token has a nonce header — ID token verification rejects the wrong-key signature")
+}

issue67_regression_test.go

@@ -478,11 +478,10 @@ func TestRefreshCoordinatorIntegration(t *testing.T) {
 
 	// Test 3: Rate limiting
 	t.Run("RateLimiting", func(t *testing.T) {
-		// Reset circuit breaker to closed state for this test
-		coordinator.circuitBreaker.mutex.Lock()
+		// Reset circuit breaker to closed state for this test. All fields are
+		// atomic so we don't need any mutex.
 		atomic.StoreInt32(&coordinator.circuitBreaker.state, 0) // closed
 		atomic.StoreInt32(&coordinator.circuitBreaker.failures, 0)
-		coordinator.circuitBreaker.mutex.Unlock()
 
 		// Temporarily increase circuit breaker threshold to not interfere
 		oldMaxFailures := coordinator.circuitBreaker.config.MaxFailures
@@ -525,9 +524,11 @@ func TestRefreshCoordinatorIntegration(t *testing.T) {
 		time.Sleep(config.CleanupInterval * 3)
 
 		// Old sessions should be cleaned up
-		coordinator.attemptsMutex.RLock()
-		count := len(coordinator.sessionRefreshAttempts)
-		coordinator.attemptsMutex.RUnlock()
+		count := 0
+		coordinator.sessionRefreshAttempts.Range(func(_, _ interface{}) bool {
+			count++
+			return true
+		})
 
 		// Should have fewer sessions after cleanup
 		if count > 10 {

logout_test.go

@@ -415,8 +415,8 @@ func TestMiddlewareBackchannelLogoutRouting(t *testing.T) {
 		clientID:                 "test-client",
 		issuerURL:                "https://provider.example.com",
 		initComplete:             make(chan struct{}),
-		firstRequestReceived:     true,
-		metadataRefreshStarted:   true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		logoutURLPath:            "/logout",
 	}
 	close(oidc.initComplete)
@@ -457,8 +457,8 @@ func TestMiddlewareFrontchannelLogoutRouting(t *testing.T) {
 		clientID:                 "test-client",
 		issuerURL:                "https://provider.example.com",
 		initComplete:             make(chan struct{}),
-		firstRequestReceived:     true,
-		metadataRefreshStarted:   true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		logoutURLPath:            "/logout",
 	}
 	close(oidc.initComplete)

main.go

@@ -89,6 +89,7 @@ var defaultExcludedURLs = map[string]struct{}{
 //   - The configured TraefikOidc handler ready to process requests.
 //   - An error if essential configuration is missing or invalid (e.g., short encryption key).
 func New(ctx context.Context, next http.Handler, config *Config, name string) (http.Handler, error) {
+	sendTelemetry(pluginVersion)
 	return NewWithContext(ctx, config, next, name)
 }
 
@@ -239,23 +240,63 @@ func NewWithContext(ctx context.Context, config *Config, next http.Handler, name
 			}
 			return 0
 		}(),
-		tokenCleanupStopChan:     make(chan struct{}),
-		metadataRefreshStopChan:  make(chan struct{}),
-		ctx:                      pluginCtx,
-		cancelFunc:               cancelFunc,
-		suppressDiagnosticLogs:   isTestMode(),
-		securityHeadersApplier:   config.GetSecurityHeadersApplier(),
-		scopeFilter:              NewScopeFilter(logger), // NEW - for discovery-based scope filtering
-		dcrConfig:                config.DynamicClientRegistration,
-		allowPrivateIPAddresses:  config.AllowPrivateIPAddresses,
-		minimalHeaders:           config.MinimalHeaders,
-		stripAuthCookies:         config.StripAuthCookies,
-		enableBackchannelLogout:  config.EnableBackchannelLogout,
-		enableFrontchannelLogout: config.EnableFrontchannelLogout,
-		backchannelLogoutPath:    normalizeLogoutPath(config.BackchannelLogoutURL),
-		frontchannelLogoutPath:   normalizeLogoutPath(config.FrontchannelLogoutURL),
-		sessionInvalidationCache: cacheManager.GetSharedSessionInvalidationCache(),
-		refreshResultCache:       cacheManager.GetSharedRefreshResultCache(),
+		tokenCleanupStopChan:      make(chan struct{}),
+		metadataRefreshStopChan:   make(chan struct{}),
+		ctx:                       pluginCtx,
+		cancelFunc:                cancelFunc,
+		suppressDiagnosticLogs:    isTestMode(),
+		securityHeadersApplier:    config.GetSecurityHeadersApplier(),
+		scopeFilter:               NewScopeFilter(logger), // NEW - for discovery-based scope filtering
+		dcrConfig:                 config.DynamicClientRegistration,
+		allowPrivateIPAddresses:   config.AllowPrivateIPAddresses,
+		minimalHeaders:            config.MinimalHeaders,
+		stripAuthCookies:          config.StripAuthCookies,
+		enableBackchannelLogout:   config.EnableBackchannelLogout,
+		enableFrontchannelLogout:  config.EnableFrontchannelLogout,
+		backchannelLogoutPath:     normalizeLogoutPath(config.BackchannelLogoutURL),
+		frontchannelLogoutPath:    normalizeLogoutPath(config.FrontchannelLogoutURL),
+		sessionInvalidationCache:  cacheManager.GetSharedSessionInvalidationCache(),
+		refreshResultCache:        cacheManager.GetSharedRefreshResultCache(),
+		enableBearerAuth:          config.EnableBearerAuth,
+		stripAuthorizationHeader:  config.StripAuthorizationHeader,
+		bearerEmitWWWAuthenticate: config.BearerEmitWWWAuthenticate,
+		bearerOverridesCookie:     config.BearerOverridesCookie,
+		bearerIdentifierClaim: func() string {
+			if config.BearerIdentifierClaim != "" {
+				return config.BearerIdentifierClaim
+			}
+			return "sub"
+		}(),
+		maxIdentifierLength: func() int {
+			if config.MaxIdentifierLength > 0 {
+				return config.MaxIdentifierLength
+			}
+			return 256
+		}(),
+		maxTokenAge: func() time.Duration {
+			if config.MaxTokenAgeSeconds > 0 {
+				return time.Duration(config.MaxTokenAgeSeconds) * time.Second
+			}
+			return 24 * time.Hour
+		}(),
+		bearerFailureThreshold: func() int {
+			if config.BearerFailureThreshold > 0 {
+				return config.BearerFailureThreshold
+			}
+			return 20
+		}(),
+		bearerFailureWindow: func() time.Duration {
+			if config.BearerFailureWindowSeconds > 0 {
+				return time.Duration(config.BearerFailureWindowSeconds) * time.Second
+			}
+			return 60 * time.Second
+		}(),
+		bearerFailurePenalty: func() time.Duration {
+			if config.BearerFailurePenaltySeconds > 0 {
+				return time.Duration(config.BearerFailurePenaltySeconds) * time.Second
+			}
+			return 60 * time.Second
+		}(),
 	}
 
 	// Log audience configuration
@@ -265,6 +306,31 @@ func NewWithContext(ctx context.Context, config *Config, next http.Handler, name
 		t.logger.Debugf("No custom audience specified, using clientID as audience: %s", t.clientID)
 	}
 
+	// Bearer-auth startup validation. The bearer path is M2M-only and demands
+	// a non-default audience so tokens issued for a different resource cannot
+	// be replayed against this service. The BearerIdentifierClaim guard blocks
+	// the `email` claim explicitly — without email_verified enforcement (out of
+	// scope for M2M), trusting email is a spoofing vector for federated IdPs.
+	// See spec §7.9 / §13.
+	if config.EnableBearerAuth {
+		if config.Audience == "" {
+			cancelFunc()
+			return nil, fmt.Errorf("EnableBearerAuth=true requires Audience to be set explicitly (cannot default to clientID — that path accepts ID tokens)")
+		}
+		if t.bearerIdentifierClaim == "email" {
+			cancelFunc()
+			return nil, fmt.Errorf("enableBearerAuth=true with bearerIdentifierClaim=%q is rejected: email-based identity without email_verified enforcement is a spoofing vector for federated IdPs (use \"sub\" or a custom claim; cookie-path userIdentifierClaim is unaffected)", t.bearerIdentifierClaim)
+		}
+		if !config.StrictAudienceValidation {
+			t.logger.Infof("EnableBearerAuth=true with StrictAudienceValidation=false: recommend enabling strict audience validation for hardening")
+		}
+		t.bearerFailureTracker = newBearerFailureTracker(
+			t.bearerFailureThreshold, t.bearerFailureWindow, t.bearerFailurePenalty,
+		)
+		t.logger.Infof("Bearer-token auth enabled: audience=%q identifierClaim=%q stripAuthz=%t bearerOverridesCookie=%t maxTokenAge=%s",
+			config.Audience, t.bearerIdentifierClaim, t.stripAuthorizationHeader, t.bearerOverridesCookie, t.maxTokenAge)
+	}
+
 	// Convert sessionMaxAge from seconds to duration (0 will use default 24 hours)
 	sessionMaxAge := time.Duration(config.SessionMaxAge) * time.Second
 	t.sessionManager, _ = NewSessionManager(config.SessionEncryptionKey, config.ForceHTTPS, config.CookieDomain, config.CookiePrefix, sessionMaxAge, t.logger) // Safe to ignore: session manager creation with fallback to defaults

main_initialization_test.go

@@ -8,6 +8,7 @@ import (
 	"net/http/httptest"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 )
@@ -484,9 +485,8 @@ func TestFirstRequestHandling(t *testing.T) {
 		defer server.Close()
 
 		oidc := &TraefikOidc{
-			providerURL:          server.URL,
-			firstRequestReceived: false,
-			firstRequestMutex:    sync.Mutex{},
+			providerURL:         server.URL,
+			firstRequestStarted: 0,
 			httpClient: &http.Client{
 				Timeout: 5 * time.Second,
 			},
@@ -508,19 +508,13 @@ func TestFirstRequestHandling(t *testing.T) {
 			},
 		}
 
-		// Simulate first request processing
-		oidc.firstRequestMutex.Lock()
-		if !oidc.firstRequestReceived {
-			oidc.firstRequestReceived = true
-			oidc.firstRequestMutex.Unlock()
-
+		// Simulate first request processing — single-firing via CAS.
+		if atomic.CompareAndSwapInt32(&oidc.firstRequestStarted, 0, 1) {
 			// This would normally be called asynchronously
 			go func() {
 				oidc.initializeMetadata(server.URL)
 				// initComplete is closed internally by initializeMetadata
 			}()
-		} else {
-			oidc.firstRequestMutex.Unlock()
 		}
 
 		// Wait for initialization
@@ -556,9 +550,8 @@ func TestFirstRequestHandling(t *testing.T) {
 		defer server.Close()
 
 		oidc := &TraefikOidc{
-			providerURL:          server.URL,
-			firstRequestReceived: false,
-			firstRequestMutex:    sync.Mutex{},
+			providerURL:         server.URL,
+			firstRequestStarted: 0,
 			httpClient: &http.Client{
 				Timeout: 5 * time.Second,
 			},
@@ -580,31 +573,22 @@ func TestFirstRequestHandling(t *testing.T) {
 			},
 		}
 
-		// Simulate multiple concurrent "first" requests
+		// Simulate multiple concurrent "first" requests — only one CAS winner
+		// fires the bootstrap path.
 		const numRequests = 10
 		var wg sync.WaitGroup
 		wg.Add(numRequests)
 
-		initStarted := 0
-		var initMu sync.Mutex
+		var initStarted int32
 
 		for i := 0; i < numRequests; i++ {
 			go func() {
 				defer wg.Done()
 
-				oidc.firstRequestMutex.Lock()
-				if !oidc.firstRequestReceived {
-					oidc.firstRequestReceived = true
-					oidc.firstRequestMutex.Unlock()
-
-					initMu.Lock()
-					initStarted++
-					initMu.Unlock()
-
+				if atomic.CompareAndSwapInt32(&oidc.firstRequestStarted, 0, 1) {
+					atomic.AddInt32(&initStarted, 1)
 					// Only one should actually start initialization
 					oidc.initializeMetadata(server.URL)
-				} else {
-					oidc.firstRequestMutex.Unlock()
 				}
 			}()
 		}
@@ -612,8 +596,8 @@ func TestFirstRequestHandling(t *testing.T) {
 		wg.Wait()
 
 		// Verify only one initialization was started
-		if initStarted != 1 {
-			t.Errorf("expected exactly 1 initialization, got %d", initStarted)
+		if atomic.LoadInt32(&initStarted) != 1 {
+			t.Errorf("expected exactly 1 initialization, got %d", atomic.LoadInt32(&initStarted))
 		}
 
 		// The metadata endpoint might be called once or not at all depending on timing

main_servehttp_test.go

@@ -61,8 +61,8 @@ func TestServeHTTP_ExcludedURLs(t *testing.T) {
 				logger:                 NewLogger("debug"),
 				initComplete:           make(chan struct{}),
 				sessionManager:         createTestSessionManager(t),
-				firstRequestReceived:   true,
-				metadataRefreshStarted: true,
+				firstRequestStarted: 1,
+				metadataRefreshStartedAtomic: 1,
 				issuerURL:              "https://provider.example.com", // Required for initialization check
 			}
 			close(oidc.initComplete)
@@ -92,8 +92,8 @@ func TestServeHTTP_EventStream(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         sessionManager,
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 		}
 		close(oidc.initComplete)
@@ -175,8 +175,8 @@ func TestServeHTTP_WebSocketUpgrade(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         sessionManager,
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 		}
 		close(oidc.initComplete)
@@ -272,8 +272,8 @@ func TestServeHTTP_InitializationTimeout(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}), // Never close this to simulate timeout
 			sessionManager:         createTestSessionManager(t),
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 		}
 
 		req := httptest.NewRequest("GET", "/protected", nil)
@@ -307,8 +307,8 @@ func TestServeHTTP_InitializationTimeout(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         createTestSessionManager(t),
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 			redirURLPath:           "/callback",
 			logoutURLPath:          "/logout",
@@ -337,8 +337,8 @@ func TestServeHTTP_CallbackAndLogout(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         createTestSessionManager(t),
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 			redirURLPath:           "/callback",
 			logoutURLPath:          "/logout",
@@ -367,8 +367,8 @@ func TestServeHTTP_CallbackAndLogout(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         createTestSessionManager(t),
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 			redirURLPath:           "/callback",
 			logoutURLPath:          "/logout",
@@ -740,8 +740,8 @@ func TestMinimalHeaders(t *testing.T) {
 				logger:                 NewLogger("debug"),
 				initComplete:           make(chan struct{}),
 				sessionManager:         sessionManager,
-				firstRequestReceived:   true,
-				metadataRefreshStarted: true,
+				firstRequestStarted: 1,
+				metadataRefreshStartedAtomic: 1,
 				issuerURL:              "https://provider.example.com",
 				minimalHeaders:         tt.minimalHeaders,
 				extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -817,8 +817,8 @@ func TestMinimalHeaders_TokenHeaderNotSet(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sessionManager,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		minimalHeaders:         true, // Enable minimal headers
 		extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -903,8 +903,8 @@ func TestStripAuthCookies(t *testing.T) {
 				logger:                 NewLogger("debug"),
 				initComplete:           make(chan struct{}),
 				sessionManager:         sessionManager,
-				firstRequestReceived:   true,
-				metadataRefreshStarted: true,
+				firstRequestStarted: 1,
+				metadataRefreshStartedAtomic: 1,
 				issuerURL:              "https://provider.example.com",
 				stripAuthCookies:       tt.stripAuthCookies,
 				extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -987,8 +987,8 @@ func TestStripAuthCookies_NoCookies(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sessionManager,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		stripAuthCookies:       true,
 		extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -1034,8 +1034,8 @@ func TestStripAuthCookies_OnlyOIDCCookies(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sessionManager,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		stripAuthCookies:       true,
 		extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -1085,8 +1085,8 @@ func TestStripAuthCookies_OnlyAppCookies(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sessionManager,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		stripAuthCookies:       true,
 		extractClaimsFunc: func(token string) (map[string]interface{}, error) {
@@ -1148,8 +1148,8 @@ func TestStripAuthCookies_CustomPrefix(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sm,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		stripAuthCookies:       true,
 		extractClaimsFunc: func(token string) (map[string]interface{}, error) {

main_test.go

@@ -16,6 +16,7 @@ import (
 	"net/url"
 	"strings"
 	"sync"
+	"sync/atomic"
 	"testing"
 	"time"
 
@@ -2685,10 +2686,9 @@ func TestMetadataRecoveryOnProviderFailure(t *testing.T) {
 	providerAvailable = true
 	mu.Unlock()
 
-	// Reset the retry timer to allow immediate retry
-	m.metadataRetryMutex.Lock()
-	m.lastMetadataRetryTime = time.Time{} // Reset to zero time
-	m.metadataRetryMutex.Unlock()
+	// Reset the retry timer to allow immediate retry. The field is atomic
+	// now, so no lock is needed.
+	atomic.StoreInt64(&m.lastMetadataRetryNano, 0)
 
 	// Second request should trigger recovery attempt
 	req2 := httptest.NewRequest("GET", "/protected", nil)

middleware.go

@@ -8,6 +8,7 @@ import (
 	"fmt"
 	"net/http"
 	"strings"
+	"sync/atomic"
 	"time"
 
 	"github.com/lukaszraczylo/traefikoidc/internal/utils"
@@ -145,19 +146,20 @@ func (t *TraefikOidc) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
 	}
 
 	if !strings.HasPrefix(req.URL.Path, "/health") {
-		t.firstRequestMutex.Lock()
-		if !t.firstRequestReceived {
-			t.firstRequestReceived = true
+		// Lock-free one-shot bootstrap. The previous firstRequestMutex.Lock()
+		// fired on EVERY non-health request forever (even after the boolean
+		// flipped true), which under Yaegi added a per-request serialization
+		// point. CAS gives single-firing semantics with zero steady-state cost.
+		if atomic.CompareAndSwapInt32(&t.firstRequestStarted, 0, 1) {
 			t.logger.Debug("Starting background tasks on first request")
 			t.startTokenCleanup()
 
-			if !t.metadataRefreshStarted && t.providerURL != "" {
-				t.metadataRefreshStarted = true
+			if t.providerURL != "" &&
+				atomic.CompareAndSwapInt32(&t.metadataRefreshStartedAtomic, 0, 1) {
 				// Metadata refresh is handled by singleton resource manager
 				t.startMetadataRefresh(t.providerURL)
 			}
 		}
-		t.firstRequestMutex.Unlock()
 	}
 
 	// Evaluate auth-bypass once, before waiting for initialization. Excluded
@@ -168,6 +170,14 @@ func (t *TraefikOidc) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
 	// unauthenticated traffic would silently expose the backend.
 	if bypass, reason := t.shouldBypassAuth(req); bypass {
 		t.logger.Debugf("Bypassing OIDC for %s (%s)", req.URL.Path, reason)
+		// When bearer auth is enabled, strip the Authorization header on
+		// bypassed paths so a bearer token can't leak into health/metrics/
+		// public endpoint logs via downstream services that don't expect it.
+		// Excluded URLs are explicitly public; bearer is an artifact of the
+		// API auth flow that doesn't belong on them.
+		if t.enableBearerAuth {
+			req.Header.Del("Authorization")
+		}
 		switch reason {
 		case bypassReasonExcluded:
 			// Operator-declared excluded URLs forward unconditionally.
@@ -205,14 +215,14 @@ func (t *TraefikOidc) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
 		t.metadataMu.RUnlock()
 
 		if issuerURL == "" {
-			// Provider metadata initialization failed - try to recover
-			// Retry every 30 seconds to allow automatic recovery when provider comes back online
-			t.metadataRetryMutex.Lock()
-			shouldRetry := time.Since(t.lastMetadataRetryTime) >= 30*time.Second
-			if shouldRetry {
-				t.lastMetadataRetryTime = time.Now()
-			}
-			t.metadataRetryMutex.Unlock()
+			// Provider metadata initialization failed - try to recover.
+			// Retry every 30 seconds to allow automatic recovery. Lock-free
+			// throttle via CAS on lastMetadataRetryNano: one goroutine wins
+			// the window, others see shouldRetry=false.
+			nowNano := time.Now().UnixNano()
+			last := atomic.LoadInt64(&t.lastMetadataRetryNano)
+			shouldRetry := time.Duration(nowNano-last) >= 30*time.Second &&
+				atomic.CompareAndSwapInt64(&t.lastMetadataRetryNano, last, nowNano)
 
 			if shouldRetry && t.providerURL != "" {
 				t.logger.Info("Attempting to recover OIDC provider metadata...")
@@ -236,6 +246,24 @@ func (t *TraefikOidc) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
 	// Bypass checks already ran before the init wait; no need to repeat them.
 	t.sessionManager.CleanupOldCookies(rw, req)
 
+	// Bearer-token auth (opt-in). Runs after init (we need issuer+JWKs+aud
+	// available) and after bypass (excluded URLs always win). Cookie-vs-
+	// bearer precedence is configurable; the safe default is cookie-wins.
+	// See bearer_auth.go for the full pipeline.
+	if t.enableBearerAuth {
+		if _, hasBearer := detectBearerToken(req); hasBearer {
+			cookiePresent := t.hasSessionCookie(req)
+			if !cookiePresent || t.bearerOverridesCookie {
+				if cookiePresent {
+					t.logger.Infof("Both Authorization: Bearer and session cookie present on %s; bearer-wins per BearerOverridesCookie=true", req.URL.Path)
+				}
+				t.handleBearerRequest(rw, req)
+				return
+			}
+			t.logger.Infof("Both Authorization: Bearer and session cookie present on %s; cookie-wins (default); bearer ignored", req.URL.Path)
+		}
+	}
+
 	session, err := t.sessionManager.GetSession(req)
 	if err != nil {
 		t.logger.Errorf("Error getting session: %v. Initiating authentication.", err)
@@ -401,10 +429,17 @@ func (t *TraefikOidc) ServeHTTP(rw http.ResponseWriter, req *http.Request) {
 	t.defaultInitiateAuthentication(rw, req, session, redirectURL)
 }
 
-// processAuthorizedRequest processes requests for authenticated users.
-// It extracts claims, validates roles/groups if configured, sets authentication headers,
-// processes header templates, and forwards the request to the next handler.
-// Domain checks should be performed before calling this method.
+// processAuthorizedRequest processes requests for authenticated cookie/session
+// users. It performs session-specific checks (identifier presence, backchannel-
+// logout invalidation, claims extraction with potential re-auth), persists
+// dirty session state, then delegates the post-auth pipeline (roles/groups,
+// header injection, security headers, cookie strip, forward) to
+// forwardAuthorized.
+//
+// The bearer-token path uses the same forwardAuthorized helper but takes a
+// different route to it (see bearer_auth.go). Keeping forwardAuthorized
+// session-agnostic is what lets the two auth methods share one pipeline.
+//
 // Parameters:
 //   - rw: The HTTP response writer.
 //   - req: The HTTP request to process.
@@ -442,8 +477,7 @@ func (t *TraefikOidc) processAuthorizedRequest(rw http.ResponseWriter, req *http
 	// the parsed claims keyed on the raw ID token, so concurrent dashboard
 	// panel requests on the same session don't repeatedly base64-decode and
 	// JSON-unmarshal the same JWT (a real cost under the yaegi interpreter
-	// that hosts Traefik plugins). idClaims is reused below by the
-	// header-templates branch.
+	// that hosts Traefik plugins).
 	idToken := session.GetIDToken()
 	var (
 		idClaims    map[string]interface{}
@@ -472,18 +506,76 @@ func (t *TraefikOidc) processAuthorizedRequest(rw http.ResponseWriter, req *http
 		return
 	}
 
-	var groups, roles []string
+	if groupClaimsErr != nil && len(t.allowedRolesAndGroups) > 0 {
+		// Claims couldn't be extracted but roles checks are required:
+		// re-authenticate rather than 403 (session may be salvageable on
+		// re-issue). Bearer path uses 401 for the equivalent failure.
+		t.logger.Errorf("Failed to extract claims for roles/groups check: %v", groupClaimsErr)
+		session.ResetRedirectCount()
+		t.defaultInitiateAuthentication(rw, req, session, redirectURL)
+		return
+	}
 
-	if groupClaimsErr == nil && groupClaims != nil {
-		var err error
-		groups, roles, err = t.extractGroupsAndRolesFromClaims(groupClaims)
-		if err != nil && len(t.allowedRolesAndGroups) > 0 {
-			t.logger.Errorf("Failed to extract groups and roles: %v", err)
-			session.ResetRedirectCount()
-			t.defaultInitiateAuthentication(rw, req, session, redirectURL)
+	// Persist any dirty session state BEFORE forwardAuthorized writes the
+	// response. Once next.ServeHTTP fires, Set-Cookie can no longer reach
+	// the client. The forwardAuthorized pipeline does not mutate session
+	// state, so saving here is safe.
+	if session.IsDirty() {
+		if err := session.Save(req, rw); err != nil {
+			t.logger.Errorf("Failed to save session after processing headers: %v", err)
+		}
+	} else {
+		t.logger.Debug("Session not dirty, skipping save in processAuthorizedRequest")
+	}
+
+	// Build the source-agnostic principal. ID-token claims drive header
+	// templates and roles when present; otherwise fall back to access-token
+	// claims (matches prior behavior for opaque-ID-token providers).
+	p := &principal{
+		Source:       sourceSession,
+		Identifier:   userIdentifier,
+		AccessToken:  session.GetAccessToken(),
+		IDToken:      idToken,
+		RefreshToken: session.GetRefreshToken(),
+		Claims:       groupClaims,
+	}
+
+	t.forwardAuthorized(rw, req, p)
+}
+
+// forwardAuthorized completes the post-authentication pipeline shared by the
+// cookie/session path and the bearer-token path. It performs:
+//
+//  1. Roles/groups extraction from p.Claims (idempotent; existing
+//     extractGroupsAndRolesFromClaims helper).
+//  2. allowedRolesAndGroups gate — writes a 403 and returns if denied.
+//  3. Identity-header injection (X-Forwarded-User, X-User-Groups, X-User-Roles,
+//     plus X-Auth-Request-* when !minimalHeaders).
+//  4. Operator-defined header templates.
+//  5. Security headers (delegated to t.securityHeadersApplier or fallback).
+//  6. OIDC session-cookie strip (stripAuthCookies).
+//  7. Authorization header strip on bearer source when stripAuthorizationHeader.
+//  8. next.ServeHTTP.
+//
+// Session persistence is the CALLER's responsibility — it must happen before
+// this function so Set-Cookie reaches the response.
+func (t *TraefikOidc) forwardAuthorized(rw http.ResponseWriter, req *http.Request, p *principal) {
+	var (
+		groups, roles []string
+		extractErr    error
+	)
+	if p.Claims != nil {
+		groups, roles, extractErr = t.extractGroupsAndRolesFromClaims(p.Claims)
+		if extractErr != nil && len(t.allowedRolesAndGroups) > 0 {
+			// Bearer path: 403 (caller already verified the token; principal
+			// claims are present but malformed for roles purposes).
+			// Cookie path can't reach here because processAuthorizedRequest
+			// catches groupClaimsErr earlier.
+			t.logger.Errorf("Failed to extract groups and roles: %v", extractErr)
+			t.sendErrorResponse(rw, req, "Access denied", http.StatusForbidden)
 			return
 		}
-		if err == nil {
+		if extractErr == nil {
 			if len(groups) > 0 {
 				req.Header.Set("X-User-Groups", strings.Join(groups, ","))
 			}
@@ -502,62 +594,46 @@ func (t *TraefikOidc) processAuthorizedRequest(rw http.ResponseWriter, req *http
 			}
 		}
 		if !allowed {
-			t.logger.Infof("User %s does not have any allowed roles or groups", userIdentifier)
+			t.logger.Infof("User %s does not have any allowed roles or groups", p.Identifier)
 			errorMsg := fmt.Sprintf("Access denied: You do not have any of the allowed roles or groups. To log out, visit: %s", t.logoutURLPath)
 			t.sendErrorResponse(rw, req, errorMsg, http.StatusForbidden)
 			return
 		}
 	}
 
-	req.Header.Set("X-Forwarded-User", userIdentifier)
+	req.Header.Set("X-Forwarded-User", p.Identifier)
 
 	// When minimalHeaders is enabled, skip extra headers to prevent 431 errors
 	if !t.minimalHeaders {
 		req.Header.Set("X-Auth-Request-Redirect", req.URL.RequestURI())
-		req.Header.Set("X-Auth-Request-User", userIdentifier)
-		if idToken != "" {
-			req.Header.Set("X-Auth-Request-Token", idToken)
+		req.Header.Set("X-Auth-Request-User", p.Identifier)
+		if p.IDToken != "" {
+			req.Header.Set("X-Auth-Request-Token", p.IDToken)
 		}
 	}
 
 	if len(t.headerTemplates) > 0 {
-		if idClaimsErr != nil {
-			t.logger.Errorf("Failed to extract claims from ID Token for template headers: %v", idClaimsErr)
-		} else {
-			// idClaims may be nil when no ID token is present; templates
-			// referencing .Claims.* will simply produce empty values, which
-			// matches the prior behavior.
-			templateData := map[string]interface{}{
-				"AccessToken":  session.GetAccessToken(),
-				"IDToken":      idToken,
-				"RefreshToken": session.GetRefreshToken(),
-				"Claims":       idClaims,
-			}
-
-			for headerName, tmpl := range t.headerTemplates {
-				var buf bytes.Buffer
-				if err := tmpl.Execute(&buf, templateData); err != nil {
-					t.logger.Errorf("Failed to execute template for header %s: %v", headerName, err)
-					continue
-				}
-				headerValue := buf.String()
-				req.Header.Set(headerName, headerValue)
-				t.logger.Debugf("Set templated header %s = %s", headerName, headerValue)
-			}
-			// NOTE: templates only mutate request headers (not session state),
-			// so we deliberately do NOT MarkDirty / Save here. Previously every
-			// authenticated request with header templates re-encrypted and
-			// rewrote all session cookies, which was a measurable CPU and
-			// Set-Cookie tax on dashboards that poll many panels per second.
+		// p.Claims may be nil (e.g. session without an ID token). Templates
+		// referencing .Claims.* will simply produce empty values — matches
+		// the prior behavior. Bearer-source principals always carry access-
+		// token claims (post-verifyToken).
+		templateData := map[string]interface{}{
+			"AccessToken":  p.AccessToken,
+			"IDToken":      p.IDToken,
+			"RefreshToken": p.RefreshToken,
+			"Claims":       p.Claims,
 		}
-	}
 
-	if session.IsDirty() {
-		if err := session.Save(req, rw); err != nil {
-			t.logger.Errorf("Failed to save session after processing headers: %v", err)
+		for headerName, tmpl := range t.headerTemplates {
+			var buf bytes.Buffer
+			if err := tmpl.Execute(&buf, templateData); err != nil {
+				t.logger.Errorf("Failed to execute template for header %s: %v", headerName, err)
+				continue
+			}
+			headerValue := buf.String()
+			req.Header.Set(headerName, headerValue)
+			t.logger.Debugf("Set templated header %s = %s", headerName, headerValue)
 		}
-	} else {
-		t.logger.Debug("Session not dirty, skipping save in processAuthorizedRequest")
 	}
 
 	// Apply security headers if configured
@@ -573,7 +649,7 @@ func (t *TraefikOidc) processAuthorizedRequest(rw http.ResponseWriter, req *http
 
 	// Strip OIDC session cookies before forwarding to the backend to prevent
 	// HTTP 431 "Request Header Fields Too Large" errors (GitHub issue #122).
-	if t.stripAuthCookies {
+	if t.stripAuthCookies && t.sessionManager != nil {
 		prefix := t.sessionManager.GetCookiePrefix()
 		filtered := make([]*http.Cookie, 0, len(req.Cookies()))
 		for _, c := range req.Cookies() {
@@ -587,7 +663,14 @@ func (t *TraefikOidc) processAuthorizedRequest(rw http.ResponseWriter, req *http
 		}
 	}
 
-	t.logger.Debugf("Request authorized for user %s, forwarding to next handler", userIdentifier)
+	// Bearer source: strip the Authorization header to keep the raw token
+	// out of downstream service logs. Off-by-config for operators who chain
+	// services that each re-verify the bearer.
+	if p.Source == sourceBearer && t.stripAuthorizationHeader {
+		req.Header.Del("Authorization")
+	}
+
+	t.logger.Debugf("Request authorized for user %s (source=%d), forwarding to next handler", p.Identifier, p.Source)
 
 	t.next.ServeHTTP(rw, req)
 }

middleware_edge_cases_test.go

@@ -13,8 +13,8 @@ func TestMiddlewareContextCancellation(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}), // Never close to simulate waiting
 		sessionManager:         createTestSessionManager(t),
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 	}
 
 	// Create request with canceled context
@@ -39,8 +39,8 @@ func TestMiddlewareSessionErrorRecovery(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         createTestSessionManager(t),
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		redirURLPath:           "/callback",
 		logoutURLPath:          "/logout",
@@ -73,8 +73,8 @@ func TestMiddlewareAJAXRequestHandling(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         createTestSessionManager(t),
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		redirURLPath:           "/callback",
 		logoutURLPath:          "/logout",
@@ -102,8 +102,8 @@ func TestLogoutWorksWithoutOIDCInitialization(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}), // Never close to simulate provider unavailable
 		sessionManager:         createTestSessionManager(t),
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		logoutURLPath:          "/logout",
 		postLogoutRedirectURI:  "/",
 		forceHTTPS:             false,
@@ -142,8 +142,8 @@ func TestMiddlewareDomainRestrictions(t *testing.T) {
 			logger:         NewLogger("debug"),
 			initComplete:   make(chan struct{}),
 			sessionManager: sessionManager,
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:      "https://provider.example.com",
 			redirURLPath:   "/callback",
 			logoutURLPath:  "/logout",
@@ -187,8 +187,8 @@ func TestMiddlewareDomainRestrictions(t *testing.T) {
 			logger:                 NewLogger("debug"),
 			initComplete:           make(chan struct{}),
 			sessionManager:         sessionManager,
-			firstRequestReceived:   true,
-			metadataRefreshStarted: true,
+			firstRequestStarted: 1,
+			metadataRefreshStartedAtomic: 1,
 			issuerURL:              "https://provider.example.com",
 			redirURLPath:           "/callback",
 			logoutURLPath:          "/logout",
@@ -236,8 +236,8 @@ func TestMiddlewareOpaqueTokenHandling(t *testing.T) {
 		logger:                 NewLogger("debug"),
 		initComplete:           make(chan struct{}),
 		sessionManager:         sessionManager,
-		firstRequestReceived:   true,
-		metadataRefreshStarted: true,
+		firstRequestStarted: 1,
+		metadataRefreshStartedAtomic: 1,
 		issuerURL:              "https://provider.example.com",
 		redirURLPath:           "/callback",
 		logoutURLPath:          "/logout",

principal.go

@@ -0,0 +1,58 @@
+// Package traefikoidc — principal abstraction for the shared post-auth
+// pipeline. A principal carries the resolved identity + tokens + claims
+// produced by EITHER the cookie session path or the bearer-token path, so
+// downstream header injection / roles checks / forwarding can be implemented
+// once and reused.
+package traefikoidc
+
+// principalSource indicates which auth path produced a principal. Used by
+// forwardAuthorized to decide source-specific behavior (e.g. only strip the
+// Authorization header for bearer-source principals).
+type principalSource int
+
+const (
+	sourceSession principalSource = iota
+	sourceBearer
+)
+
+// principal is the immutable post-auth value passed to forwardAuthorized.
+// No methods mutate it; no manager pointer; no I/O. Pure data.
+type principal struct {
+	Claims       map[string]interface{}
+	Identifier   string
+	Subject      string
+	ClientID     string
+	AccessToken  string
+	IDToken      string
+	RefreshToken string
+	Source       principalSource
+}
+
+// buildPrincipalFromSession adapts an authenticated SessionData into a
+// principal value WITHOUT writing back to the session. This is the only
+// function that still knows about SessionData; the rest of the pipeline is
+// session-agnostic. Returns nil when the session has no usable identity.
+func (t *TraefikOidc) buildPrincipalFromSession(session *SessionData) *principal {
+	if session == nil {
+		return nil
+	}
+	identifier := session.GetUserIdentifier()
+	if identifier == "" {
+		return nil
+	}
+
+	var claims map[string]interface{}
+	if idToken := session.GetIDToken(); idToken != "" && t.extractClaimsFunc != nil {
+		// Best-effort: cached on the session, never blocking.
+		claims, _ = session.GetIDTokenClaims(t.extractClaimsFunc) // Safe to ignore: claims-error path handled by header-template branch
+	}
+
+	return &principal{
+		Source:       sourceSession,
+		Identifier:   identifier,
+		AccessToken:  session.GetAccessToken(),
+		IDToken:      session.GetIDToken(),
+		RefreshToken: session.GetRefreshToken(),
+		Claims:       claims,
+	}
+}

refresh_coordinator.go

@@ -15,17 +15,29 @@ import (
 // It implements request coalescing, rate limiting, and circuit breaking
 // specifically for token refresh operations.
 type RefreshCoordinator struct {
-	inFlightRefreshes      map[string]*refreshOperation
+	// inFlightRefreshes maps tokenHash -> *refreshOperation. sync.Map is used
+	// instead of a plain map + RWMutex so concurrent refreshes do not
+	// serialize on a single global lock. Under Yaegi the previous
+	// refreshMutex.Lock() was held for tens of milliseconds per request due
+	// to interpreter overhead on the work inside the critical section,
+	// causing dozens of goroutines to stack up on it and pin one CPU core.
+	inFlightRefreshes sync.Map
+	// sessionRefreshAttempts maps sessionID -> *refreshAttemptTracker.
+	// sync.Map + atomic tracker fields means isInCooldown/recordRefreshAttempt/
+	// recordRefreshSuccess/recordRefreshFailure are lock-free. Previously
+	// these used attemptsMutex sync.RWMutex; under Yaegi every Lock() acquisition
+	// adds 10-50ms of dispatch overhead, and they were called twice per leader
+	// request (once for recordRefreshAttempt, once for isInCooldown). That
+	// serializing pattern caused the v1.0.15 death spiral after v1.0.14
+	// removed the refreshMutex (same architectural shape, different mutex).
+	sessionRefreshAttempts sync.Map
 	cleanupTimers          map[string]*time.Timer
-	sessionRefreshAttempts map[string]*refreshAttemptTracker
 	circuitBreaker         *RefreshCircuitBreaker
 	metrics                *RefreshMetrics
 	logger                 *Logger
 	stopChan               chan struct{}
 	config                 RefreshCoordinatorConfig
 	wg                     sync.WaitGroup
-	attemptsMutex          sync.RWMutex
-	refreshMutex           sync.RWMutex
 	cleanupTimerMu         sync.Mutex
 }
 
@@ -84,14 +96,22 @@ type refreshResult struct {
 	fromCache     bool
 }
 
-// refreshAttemptTracker tracks refresh attempts for a session
+// refreshAttemptTracker tracks refresh attempts for a session. All fields are
+// accessed via sync/atomic so isInCooldown/recordRefreshAttempt/Success/Failure
+// can run without holding any per-coordinator lock. Times are UnixNano so they
+// fit in an int64 and can be read with a single atomic.LoadInt64.
+//
+// cooldownEndNano == 0 means "not in cooldown". This sentinel replaces the
+// inCooldown bool that the previous implementation kept under attemptsMutex —
+// under Yaegi any per-request global mutex turns into a serializing bottleneck
+// (the v1.0.14 refreshMutex -> sync.Map fix removed only one such bottleneck;
+// attemptsMutex was the next one in the queue).
 type refreshAttemptTracker struct {
-	lastAttemptTime     time.Time
-	windowStartTime     time.Time
-	cooldownEndTime     time.Time
-	attempts            int32
-	consecutiveFailures int32
-	inCooldown          bool
+	lastAttemptNano     int64 // atomic, UnixNano of last attempt
+	windowStartNano     int64 // atomic, UnixNano of attempt-window start
+	cooldownEndNano     int64 // atomic, UnixNano; 0 = not in cooldown
+	attempts            int32 // atomic
+	consecutiveFailures int32 // atomic
 }
 
 // RefreshMetrics tracks coordinator performance metrics
@@ -106,14 +126,18 @@ type RefreshMetrics struct {
 	currentInFlightRefreshes int32
 }
 
-// RefreshCircuitBreaker implements a circuit breaker specifically for refresh operations
+// RefreshCircuitBreaker implements a circuit breaker specifically for refresh
+// operations. All mutable fields are atomic so AllowRequest/RecordSuccess/
+// RecordFailure run without any mutex. The previous sync.RWMutex.RLock() was
+// taken on every CoordinateRefresh — under Yaegi this added 10-50ms of
+// interpreter dispatch per call, which compounded with attemptsMutex to keep
+// the pod's single CPU core saturated.
 type RefreshCircuitBreaker struct {
-	lastFailureTime time.Time
-	lastSuccessTime time.Time
+	lastFailureNano int64 // atomic, UnixNano of most recent failure
+	lastSuccessNano int64 // atomic, UnixNano of most recent success
 	config          RefreshCircuitBreakerConfig
-	mutex           sync.RWMutex
-	state           int32
-	failures        int32
+	state           int32 // atomic: 0=closed, 1=open, 2=half-open
+	failures        int32 // atomic
 }
 
 // RefreshCircuitBreakerConfig configures the refresh circuit breaker
@@ -130,13 +154,13 @@ func NewRefreshCoordinator(config RefreshCoordinatorConfig, logger *Logger) *Ref
 	}
 
 	rc := &RefreshCoordinator{
-		inFlightRefreshes:      make(map[string]*refreshOperation),
-		sessionRefreshAttempts: make(map[string]*refreshAttemptTracker),
-		config:                 config,
-		metrics:                &RefreshMetrics{},
-		logger:                 logger,
-		stopChan:               make(chan struct{}),
-		cleanupTimers:          make(map[string]*time.Timer),
+		// inFlightRefreshes and sessionRefreshAttempts are both sync.Map;
+		// their zero values are ready to use.
+		config:        config,
+		metrics:       &RefreshMetrics{},
+		logger:        logger,
+		stopChan:      make(chan struct{}),
+		cleanupTimers: make(map[string]*time.Timer),
 		circuitBreaker: &RefreshCircuitBreaker{
 			config: RefreshCircuitBreakerConfig{
 				MaxFailures:      3,
@@ -227,13 +251,28 @@ func (rc *RefreshCoordinator) getOrCreateOperation(
 	tokenHash string,
 	refreshToken string,
 ) (*refreshOperation, bool, error) {
-	rc.refreshMutex.Lock()
-	defer rc.refreshMutex.Unlock()
+	// Speculatively construct the operation we WOULD register if we win the
+	// race. Allocating here keeps the LoadOrStore call below atomic and
+	// avoids any global lock — under Yaegi the previous map+RWMutex design
+	// held the write lock long enough (tens of ms per call) that concurrent
+	// refreshes on the same coordinator serialized into a queue that grew
+	// without bound. See struct comment on inFlightRefreshes.
+	candidate := &refreshOperation{
+		refreshToken: refreshToken,
+		done:         make(chan struct{}),
+		startTime:    time.Now(),
+		waiterCount:  1,
+	}
 
-	// Check for existing operation while holding the lock
-	if existingOp, exists := rc.inFlightRefreshes[tokenHash]; exists {
+	if existing, loaded := rc.inFlightRefreshes.LoadOrStore(tokenHash, candidate); loaded {
+		existingOp, ok := existing.(*refreshOperation)
+		if !ok {
+			// Defensive: anything stored here is always *refreshOperation, but
+			// keep the typed assert so a programming error elsewhere doesn't
+			// surface as a confusing panic in an interpreter frame.
+			return nil, false, fmt.Errorf("inFlightRefreshes corrupt: unexpected type %T", existing)
+		}
 		if existingOp.refreshToken == refreshToken {
-			// Join existing operation
 			atomic.AddInt32(&existingOp.waiterCount, 1)
 			return existingOp, false, nil
 		}
@@ -241,41 +280,60 @@ func (rc *RefreshCoordinator) getOrCreateOperation(
 		return nil, false, fmt.Errorf("refresh token mismatch")
 	}
 
-	// No existing operation - check if we can create a new one
-	// All checks happen while holding the lock to prevent races
+	// We won the race and registered `candidate`. Apply gates now. If any
+	// gate fails we must remove our entry from the map and signal failure
+	// to any joiners that snuck in between LoadOrStore and now.
+	if err := rc.applyLeaderGates(sessionID); err != nil {
+		rc.failCandidate(tokenHash, candidate, err)
+		return nil, false, err
+	}
 
-	// Check and record refresh attempt for rate limiting
+	// Reserve concurrent slot via CAS — without the old global lock we can
+	// no longer rely on mutex-mediated check-then-increment. If we lose the
+	// CAS race we retry; if the limit has since been reached we back out.
+	for {
+		current := atomic.LoadInt32(&rc.metrics.currentInFlightRefreshes)
+		if int(current) >= rc.config.MaxConcurrentRefreshes {
+			err := fmt.Errorf("maximum concurrent refresh operations reached")
+			rc.failCandidate(tokenHash, candidate, err)
+			return nil, false, err
+		}
+		if atomic.CompareAndSwapInt32(&rc.metrics.currentInFlightRefreshes, current, current+1) {
+			break
+		}
+	}
+
+	return candidate, true, nil
+}
+
+// applyLeaderGates runs the rate-limit, cooldown, and memory-pressure checks
+// that previously ran under the global refreshMutex. Only the leader (the
+// goroutine that just registered the operation) runs them; joiners share the
+// leader's outcome via operation.done.
+func (rc *RefreshCoordinator) applyLeaderGates(sessionID string) error {
 	rc.recordRefreshAttempt(sessionID)
 	if rc.isInCooldown(sessionID) {
 		atomic.AddInt64(&rc.metrics.cooldownsTriggered, 1)
-		return nil, false, fmt.Errorf("refresh attempts exceeded for session, in cooldown period")
+		return fmt.Errorf("refresh attempts exceeded for session, in cooldown period")
 	}
-
-	// Check memory pressure
 	if rc.config.EnableMemoryPressureDetection && rc.isUnderMemoryPressure() {
 		atomic.AddInt64(&rc.metrics.memoryPressureEvents, 1)
-		return nil, false, fmt.Errorf("system under memory pressure, refresh denied")
+		return fmt.Errorf("system under memory pressure, refresh denied")
 	}
+	return nil
+}
 
-	// Check and reserve concurrent refresh slot atomically
-	current := atomic.LoadInt32(&rc.metrics.currentInFlightRefreshes)
-	if int(current) >= rc.config.MaxConcurrentRefreshes {
-		return nil, false, fmt.Errorf("maximum concurrent refresh operations reached")
-	}
-
-	// Reserve the slot - we're still holding the lock so this is safe
-	atomic.AddInt32(&rc.metrics.currentInFlightRefreshes, 1)
-
-	// Create and register new operation
-	operation := &refreshOperation{
-		refreshToken: refreshToken,
-		done:         make(chan struct{}),
-		startTime:    time.Now(),
-		waiterCount:  1,
-	}
-	rc.inFlightRefreshes[tokenHash] = operation
-
-	return operation, true, nil
+// failCandidate removes the leader's just-registered operation from the
+// in-flight map and signals the error to any joiners by recording the result
+// and closing the done channel. This keeps the (nil, false, err) return path
+// equivalent to the pre-sync.Map version: callers see the error directly,
+// joiners see it via operation.done.
+func (rc *RefreshCoordinator) failCandidate(tokenHash string, op *refreshOperation, err error) {
+	rc.inFlightRefreshes.Delete(tokenHash)
+	op.mutex.Lock()
+	op.result = &refreshResult{err: err}
+	op.mutex.Unlock()
+	close(op.done)
 }
 
 // executeRefreshAsync performs the actual refresh operation asynchronously
@@ -367,100 +425,108 @@ func (rc *RefreshCoordinator) scheduleDelayedCleanup(tokenHash string) {
 
 // performCleanup removes the operation from the in-flight map.
 // Idempotent: only decrements the in-flight counter if an entry was actually
-// removed. This guards against any future path accidentally calling cleanup
-// twice for the same tokenHash (which would corrupt the refresh budget).
+// removed. LoadAndDelete is atomic so any concurrent failCandidate or repeat
+// cleanup call will see exactly one removal — the budget cannot be corrupted
+// by double-decrement.
 func (rc *RefreshCoordinator) performCleanup(tokenHash string) {
-	rc.refreshMutex.Lock()
-	_, existed := rc.inFlightRefreshes[tokenHash]
-	if existed {
-		delete(rc.inFlightRefreshes, tokenHash)
-	}
-	rc.refreshMutex.Unlock()
-	if existed {
+	if _, existed := rc.inFlightRefreshes.LoadAndDelete(tokenHash); existed {
 		atomic.AddInt32(&rc.metrics.currentInFlightRefreshes, -1)
 	}
 }
 
-// isInCooldown checks if a session is in cooldown after recording an attempt
-func (rc *RefreshCoordinator) isInCooldown(sessionID string) bool {
-	rc.attemptsMutex.Lock()
-	defer rc.attemptsMutex.Unlock()
+// getOrCreateTracker fetches the tracker for sessionID or atomically creates a
+// fresh one. The sync.Map.LoadOrStore semantics make this lock-free even under
+// concurrent first-touch races: at most one tracker per sessionID survives.
+//
+// trackerFromMapValue centralizes the type assertion so the lint-mandated
+// two-value form lives in one place; the stored type is always
+// *refreshAttemptTracker by construction.
+func trackerFromMapValue(v interface{}) *refreshAttemptTracker {
+	t, _ := v.(*refreshAttemptTracker)
+	return t
+}
 
-	tracker, exists := rc.sessionRefreshAttempts[sessionID]
-	if !exists {
+func (rc *RefreshCoordinator) getOrCreateTracker(sessionID string) *refreshAttemptTracker {
+	if v, ok := rc.sessionRefreshAttempts.Load(sessionID); ok {
+		return trackerFromMapValue(v)
+	}
+	fresh := &refreshAttemptTracker{
+		windowStartNano: time.Now().UnixNano(),
+	}
+	actual, _ := rc.sessionRefreshAttempts.LoadOrStore(sessionID, fresh)
+	return trackerFromMapValue(actual)
+}
+
+// isInCooldown checks if a session is in cooldown. Lock-free read with a
+// best-effort cooldown-reset CAS on the cooldownEndNano sentinel. If the
+// reset races with another goroutine we accept the loser's view (the winner's
+// reset still happens). The attempt-window expiry and limit-exceeded paths
+// are write-mostly but use atomic.StoreInt64/AddInt32 — never a held lock.
+func (rc *RefreshCoordinator) isInCooldown(sessionID string) bool {
+	v, ok := rc.sessionRefreshAttempts.Load(sessionID)
+	if !ok {
 		return false // No tracker means first attempt, not in cooldown
 	}
-
+	tracker := trackerFromMapValue(v)
 	now := time.Now()
+	nowNano := now.UnixNano()
 
-	// Check if already in cooldown
-	if tracker.inCooldown {
-		if now.After(tracker.cooldownEndTime) {
-			// Cooldown expired, reset tracker
-			tracker.inCooldown = false
-			tracker.attempts = 1 // Already recorded one attempt
-			tracker.consecutiveFailures = 0
-			tracker.windowStartTime = now
-			return false
+	// Already in cooldown?
+	if cooldownEnd := atomic.LoadInt64(&tracker.cooldownEndNano); cooldownEnd != 0 {
+		if nowNano <= cooldownEnd {
+			return true // still in cooldown
+		}
+		// Cooldown expired. Best-effort reset (a concurrent caller may also
+		// reset; the result is equivalent — fresh window + one recorded
+		// attempt — so the CAS race is benign).
+		if atomic.CompareAndSwapInt64(&tracker.cooldownEndNano, cooldownEnd, 0) {
+			atomic.StoreInt32(&tracker.attempts, 1)
+			atomic.StoreInt32(&tracker.consecutiveFailures, 0)
+			atomic.StoreInt64(&tracker.windowStartNano, nowNano)
 		}
-		return true // Still in cooldown
-	}
-
-	// Check if window expired
-	if now.Sub(tracker.windowStartTime) > rc.config.RefreshAttemptWindow {
-		// Reset window
-		tracker.attempts = 1 // Already recorded one attempt
-		tracker.windowStartTime = now
 		return false
 	}
 
-	// Check if just exceeded attempt limit
-	if int(tracker.attempts) >= rc.config.MaxRefreshAttempts {
-		// Enter cooldown now
-		tracker.inCooldown = true
-		tracker.cooldownEndTime = now.Add(rc.config.RefreshCooldownPeriod)
-		rc.logger.Infof("Session %s entering refresh cooldown after %d attempts",
-			sessionID, tracker.attempts)
+	// Window expired?
+	if windowStart := atomic.LoadInt64(&tracker.windowStartNano); time.Duration(nowNano-windowStart) > rc.config.RefreshAttemptWindow {
+		atomic.StoreInt32(&tracker.attempts, 1)
+		atomic.StoreInt64(&tracker.windowStartNano, nowNano)
+		return false
+	}
+
+	// Just exceeded attempt limit?
+	if int(atomic.LoadInt32(&tracker.attempts)) >= rc.config.MaxRefreshAttempts {
+		end := now.Add(rc.config.RefreshCooldownPeriod).UnixNano()
+		// Only one CAS winner publishes the cooldown end + logs.
+		if atomic.CompareAndSwapInt64(&tracker.cooldownEndNano, 0, end) {
+			rc.logger.Infof("Session %s entering refresh cooldown after %d attempts",
+				sessionID, atomic.LoadInt32(&tracker.attempts))
+		}
 		return true
 	}
 
 	return false
 }
 
-// recordRefreshAttempt records a refresh attempt for rate limiting
+// recordRefreshAttempt records a refresh attempt for rate limiting. Lock-free:
+// LoadOrStore for the tracker, atomic counters/timestamps for fields.
 func (rc *RefreshCoordinator) recordRefreshAttempt(sessionID string) {
-	rc.attemptsMutex.Lock()
-	defer rc.attemptsMutex.Unlock()
-
-	tracker, exists := rc.sessionRefreshAttempts[sessionID]
-	if !exists {
-		tracker = &refreshAttemptTracker{
-			windowStartTime: time.Now(),
-		}
-		rc.sessionRefreshAttempts[sessionID] = tracker
-	}
-
+	tracker := rc.getOrCreateTracker(sessionID)
 	atomic.AddInt32(&tracker.attempts, 1)
-	tracker.lastAttemptTime = time.Now()
+	atomic.StoreInt64(&tracker.lastAttemptNano, time.Now().UnixNano())
 }
 
-// recordRefreshSuccess records a successful refresh
+// recordRefreshSuccess records a successful refresh. Lock-free.
 func (rc *RefreshCoordinator) recordRefreshSuccess(sessionID string) {
-	rc.attemptsMutex.Lock()
-	defer rc.attemptsMutex.Unlock()
-
-	if tracker, exists := rc.sessionRefreshAttempts[sessionID]; exists {
-		tracker.consecutiveFailures = 0
+	if v, ok := rc.sessionRefreshAttempts.Load(sessionID); ok {
+		atomic.StoreInt32(&trackerFromMapValue(v).consecutiveFailures, 0)
 	}
 }
 
-// recordRefreshFailure records a failed refresh
+// recordRefreshFailure records a failed refresh. Lock-free.
 func (rc *RefreshCoordinator) recordRefreshFailure(sessionID string) {
-	rc.attemptsMutex.Lock()
-	defer rc.attemptsMutex.Unlock()
-
-	if tracker, exists := rc.sessionRefreshAttempts[sessionID]; exists {
-		atomic.AddInt32(&tracker.consecutiveFailures, 1)
+	if v, ok := rc.sessionRefreshAttempts.Load(sessionID); ok {
+		atomic.AddInt32(&trackerFromMapValue(v).consecutiveFailures, 1)
 	}
 }
 
@@ -512,20 +578,22 @@ func (rc *RefreshCoordinator) cleanupRoutine() {
 	}
 }
 
-// cleanupStaleEntries removes outdated tracking entries
+// cleanupStaleEntries removes outdated tracking entries. Lock-free iteration
+// via sync.Map.Range; safe to race with concurrent reads/writes.
 func (rc *RefreshCoordinator) cleanupStaleEntries() {
-	now := time.Now()
-
-	rc.attemptsMutex.Lock()
-	defer rc.attemptsMutex.Unlock()
-
-	// Clean up old session trackers
-	for sessionID, tracker := range rc.sessionRefreshAttempts {
-		// Remove trackers that haven't been used recently
-		if now.Sub(tracker.lastAttemptTime) > 2*rc.config.RefreshAttemptWindow {
-			delete(rc.sessionRefreshAttempts, sessionID)
+	cutoff := time.Now().Add(-2 * rc.config.RefreshAttemptWindow).UnixNano()
+	rc.sessionRefreshAttempts.Range(func(key, value interface{}) bool {
+		tracker := trackerFromMapValue(value)
+		if tracker == nil {
+			return true
 		}
-	}
+		if atomic.LoadInt64(&tracker.lastAttemptNano) < cutoff {
+			// Compare-and-delete to avoid evicting a tracker that was just
+			// re-used by a concurrent caller. We compare by pointer identity.
+			rc.sessionRefreshAttempts.CompareAndDelete(key, value)
+		}
+		return true
+	})
 }
 
 // GetMetrics returns current coordinator metrics
@@ -558,63 +626,51 @@ func (rc *RefreshCoordinator) Shutdown() {
 	rc.wg.Wait()
 }
 
-// AllowRequest checks if the circuit breaker allows a request
+// AllowRequest reports whether the circuit breaker allows a request. Lock-free.
 func (cb *RefreshCircuitBreaker) AllowRequest() bool {
-	cb.mutex.RLock()
-	defer cb.mutex.RUnlock()
-
-	state := atomic.LoadInt32(&cb.state)
-
-	switch state {
-	case 0: // Closed
+	switch atomic.LoadInt32(&cb.state) {
+	case 0: // closed
 		return true
-	case 1: // Open
-		if time.Since(cb.lastFailureTime) > cb.config.OpenDuration {
-			// Try to transition to half-open
+	case 1: // open
+		lastFail := atomic.LoadInt64(&cb.lastFailureNano)
+		if time.Duration(time.Now().UnixNano()-lastFail) > cb.config.OpenDuration {
+			// Transition to half-open; first CAS winner gets the probe.
 			if atomic.CompareAndSwapInt32(&cb.state, 1, 2) {
 				return true
 			}
 		}
 		return false
-	case 2: // Half-open
+	case 2: // half-open
 		return true
 	default:
 		return false
 	}
 }
 
-// RecordSuccess records a successful operation
+// RecordSuccess records a successful operation. Lock-free.
 func (cb *RefreshCircuitBreaker) RecordSuccess() {
-	cb.mutex.Lock()
-	defer cb.mutex.Unlock()
-
-	state := atomic.LoadInt32(&cb.state)
-	if state == 2 { // Half-open
-		// Close the circuit
+	switch atomic.LoadInt32(&cb.state) {
+	case 2: // half-open -> close
 		atomic.StoreInt32(&cb.state, 0)
 		atomic.StoreInt32(&cb.failures, 0)
-	} else if state == 0 { // Closed
-		// Reset failure count on success
+	case 0: // closed
 		atomic.StoreInt32(&cb.failures, 0)
 	}
-	cb.lastSuccessTime = time.Now()
+	atomic.StoreInt64(&cb.lastSuccessNano, time.Now().UnixNano())
 }
 
-// RecordFailure records a failed operation
+// RecordFailure records a failed operation. Lock-free.
 func (cb *RefreshCircuitBreaker) RecordFailure() {
-	cb.mutex.Lock()
-	defer cb.mutex.Unlock()
-
 	failures := atomic.AddInt32(&cb.failures, 1)
-	cb.lastFailureTime = time.Now()
+	atomic.StoreInt64(&cb.lastFailureNano, time.Now().UnixNano())
 
-	state := atomic.LoadInt32(&cb.state)
-
-	if state == 0 && int(failures) >= cb.config.MaxFailures {
-		// Open the circuit
-		atomic.StoreInt32(&cb.state, 1)
-	} else if state == 2 {
-		// Half-open failed, return to open
+	switch atomic.LoadInt32(&cb.state) {
+	case 0:
+		if int(failures) >= cb.config.MaxFailures {
+			atomic.StoreInt32(&cb.state, 1)
+		}
+	case 2:
+		// Half-open probe failed -> back to open.
 		atomic.StoreInt32(&cb.state, 1)
 	}
 }

refresh_coordinator_test.go

@@ -365,10 +365,12 @@ func TestMemoryLeakPrevention(t *testing.T) {
 		}
 	}
 
-	// Verify cleanup is working
-	coordinator.attemptsMutex.RLock()
-	sessionCount := len(coordinator.sessionRefreshAttempts)
-	coordinator.attemptsMutex.RUnlock()
+	// Verify cleanup is working. sync.Map has no Len(); count via Range.
+	sessionCount := 0
+	coordinator.sessionRefreshAttempts.Range(func(_, _ interface{}) bool {
+		sessionCount++
+		return true
+	})
 
 	// Should have cleaned up old sessions (only recent ones remain)
 	if sessionCount > numWorkers*2 {
@@ -650,24 +652,23 @@ func TestCleanupRoutine(t *testing.T) {
 		coordinator.recordRefreshAttempt(fmt.Sprintf("session_%d", i))
 	}
 
-	// Verify sessions exist
-	coordinator.attemptsMutex.RLock()
-	initialCount := len(coordinator.sessionRefreshAttempts)
-	coordinator.attemptsMutex.RUnlock()
+	countSessions := func() int {
+		n := 0
+		coordinator.sessionRefreshAttempts.Range(func(_, _ interface{}) bool {
+			n++
+			return true
+		})
+		return n
+	}
 
-	if initialCount != 5 {
+	if initialCount := countSessions(); initialCount != 5 {
 		t.Errorf("Expected 5 sessions, got %d", initialCount)
 	}
 
 	// Wait for cleanup to run (2x window + cleanup interval)
 	time.Sleep(2*config.RefreshAttemptWindow + 2*config.CleanupInterval)
 
-	// Verify sessions were cleaned up
-	coordinator.attemptsMutex.RLock()
-	finalCount := len(coordinator.sessionRefreshAttempts)
-	coordinator.attemptsMutex.RUnlock()
-
-	if finalCount != 0 {
+	if finalCount := countSessions(); finalCount != 0 {
 		t.Errorf("Expected 0 sessions after cleanup, got %d", finalCount)
 	}
 }

settings.go

@@ -63,23 +63,23 @@ type Config struct {
 	// IdPs do not expose RT TTL on the wire, so this is intentionally a
 	// conservative heuristic; tune to match your provider configuration.
 	// Default 21600 (6h). Set to 0 to disable the check.
-	MaxRefreshTokenAgeSeconds int                              `json:"maxRefreshTokenAgeSeconds"`
-	SessionMaxAge             int                              `json:"sessionMaxAge"`
-	RateLimit                 int                              `json:"rateLimit"`
-	OverrideScopes            bool                             `json:"overrideScopes"`
-	DisableReplayDetection    bool                             `json:"disableReplayDetection,omitempty"`
-	RequireTokenIntrospection bool                             `json:"requireTokenIntrospection,omitempty"`
-	AllowOpaqueTokens         bool                             `json:"allowOpaqueTokens,omitempty"`
-	StrictAudienceValidation  bool                             `json:"strictAudienceValidation,omitempty"`
-	EnablePKCE                bool                             `json:"enablePKCE"`
-	ForceHTTPS                bool                             `json:"forceHTTPS"`
-	AllowPrivateIPAddresses   bool                             `json:"allowPrivateIPAddresses,omitempty"`
-	MinimalHeaders            bool                             `json:"minimalHeaders,omitempty"`
-	StripAuthCookies          bool                             `json:"stripAuthCookies,omitempty"`
-	EnableBackchannelLogout   bool                             `json:"enableBackchannelLogout,omitempty"`
-	EnableFrontchannelLogout  bool                             `json:"enableFrontchannelLogout,omitempty"`
-	BackchannelLogoutURL      string                           `json:"backchannelLogoutURL,omitempty"`
-	FrontchannelLogoutURL     string                           `json:"frontchannelLogoutURL,omitempty"`
+	MaxRefreshTokenAgeSeconds int    `json:"maxRefreshTokenAgeSeconds"`
+	SessionMaxAge             int    `json:"sessionMaxAge"`
+	RateLimit                 int    `json:"rateLimit"`
+	OverrideScopes            bool   `json:"overrideScopes"`
+	DisableReplayDetection    bool   `json:"disableReplayDetection,omitempty"`
+	RequireTokenIntrospection bool   `json:"requireTokenIntrospection,omitempty"`
+	AllowOpaqueTokens         bool   `json:"allowOpaqueTokens,omitempty"`
+	StrictAudienceValidation  bool   `json:"strictAudienceValidation,omitempty"`
+	EnablePKCE                bool   `json:"enablePKCE"`
+	ForceHTTPS                bool   `json:"forceHTTPS"`
+	AllowPrivateIPAddresses   bool   `json:"allowPrivateIPAddresses,omitempty"`
+	MinimalHeaders            bool   `json:"minimalHeaders,omitempty"`
+	StripAuthCookies          bool   `json:"stripAuthCookies,omitempty"`
+	EnableBackchannelLogout   bool   `json:"enableBackchannelLogout,omitempty"`
+	EnableFrontchannelLogout  bool   `json:"enableFrontchannelLogout,omitempty"`
+	BackchannelLogoutURL      string `json:"backchannelLogoutURL,omitempty"`
+	FrontchannelLogoutURL     string `json:"frontchannelLogoutURL,omitempty"`
 	// CACertPath is an optional filesystem path to a PEM-encoded CA bundle used
 	// to verify the OIDC provider's TLS certificate. Use this when the provider
 	// is signed by an internal/private CA that is not in the system trust store.
@@ -125,6 +125,52 @@ type Config struct {
 	// ClientAssertionAlg is the JWS signing algorithm. Defaults to RS256.
 	// Supported: RS256/384/512, PS256/384/512, ES256/384/512.
 	ClientAssertionAlg string `json:"clientAssertionAlg,omitempty"`
+
+	// --- Bearer-token auth (opt-in M2M path) ---
+
+	// EnableBearerAuth turns on the Authorization: Bearer <jwt> auth path.
+	// Default false. When true, Audience MUST be set or startup fails. The
+	// bearer path is M2M-only: it accepts validated access-token JWTs, rejects
+	// ID tokens, and forwards principal headers downstream without creating a
+	// cookie session. See docs/BEARER_AUTH.md for the threat model.
+	EnableBearerAuth bool `json:"enableBearerAuth,omitempty"`
+	// BearerIdentifierClaim names the JWT claim used as the principal identifier
+	// on the bearer-token auth path. Default "sub". Decoupled from
+	// UserIdentifierClaim (which defaults to "email" and drives the cookie path)
+	// so M2M bearer flow never accidentally relies on an unverified email.
+	BearerIdentifierClaim string `json:"bearerIdentifierClaim,omitempty"`
+	// StripAuthorizationHeader removes the Authorization header from the
+	// forwarded request after successful bearer auth, so downstream services
+	// never see the raw token. Default true. Disable only when a downstream
+	// explicitly needs to re-validate the bearer.
+	StripAuthorizationHeader bool `json:"stripAuthorizationHeader,omitempty"`
+	// BearerEmitWWWAuthenticate controls whether 401 responses on the bearer
+	// path include a WWW-Authenticate: Bearer error="invalid_token" hint per
+	// RFC 6750 §3. Default true. Disable to reduce reconnaissance signal.
+	BearerEmitWWWAuthenticate bool `json:"bearerEmitWWWAuthenticate,omitempty"`
+	// BearerOverridesCookie controls precedence when both Authorization:
+	// Bearer and a session cookie are present. Default false: cookie wins
+	// (safer against browser/extension/proxy bearer injection). Set true for
+	// the bearer-wins convention used by AWS/GCP/Kubernetes API gateways.
+	BearerOverridesCookie bool `json:"bearerOverridesCookie,omitempty"`
+	// MaxTokenAgeSeconds caps how old (iat-based) a bearer token may be.
+	// Default 86400 (24h). Bounds clock-manipulation tokens with implausibly
+	// distant iat values.
+	MaxTokenAgeSeconds int64 `json:"maxTokenAgeSeconds,omitempty"`
+	// MaxIdentifierLength bounds the post-sanitisation length of the bearer
+	// principal identifier (the value injected as X-Forwarded-User). Default
+	// 256.
+	MaxIdentifierLength int `json:"maxIdentifierLength,omitempty"`
+	// BearerFailureThreshold is the number of consecutive 401s from one
+	// source IP within BearerFailureWindowSeconds that trips the throttle.
+	// Default 20.
+	BearerFailureThreshold int `json:"bearerFailureThreshold,omitempty"`
+	// BearerFailureWindowSeconds is the rolling window (seconds) over which
+	// 401s are counted for throttling. Default 60.
+	BearerFailureWindowSeconds int `json:"bearerFailureWindowSeconds,omitempty"`
+	// BearerFailurePenaltySeconds is how long an IP is parked in the 429
+	// penalty box after BearerFailureThreshold is exceeded. Default 60.
+	BearerFailurePenaltySeconds int `json:"bearerFailurePenaltySeconds,omitempty"`
 }
 
 // loadCACertPool assembles an x509.CertPool from CACertPath and CACertPEM.
@@ -291,6 +337,19 @@ func CreateConfig() *Config {
 		MaxRefreshTokenAgeSeconds: 21600, // 6h - conservative heuristic, see field doc
 		SecurityHeaders:           createDefaultSecurityConfig(),
 		Redis:                     nil, // Redis is disabled by default, configure via Traefik or env vars
+
+		// Bearer-auth defaults. EnableBearerAuth=false leaves the feature
+		// dormant; the rest are values that apply only when bearer is enabled.
+		EnableBearerAuth:            false,
+		BearerIdentifierClaim:       "sub",
+		StripAuthorizationHeader:    true,
+		BearerEmitWWWAuthenticate:   true,
+		BearerOverridesCookie:       false,
+		MaxTokenAgeSeconds:          86400,
+		MaxIdentifierLength:         256,
+		BearerFailureThreshold:      20,
+		BearerFailureWindowSeconds:  60,
+		BearerFailurePenaltySeconds: 60,
 	}
 
 	return c

telemetry.go

@@ -0,0 +1,142 @@
+package traefikoidc
+
+import (
+	"bytes"
+	"context"
+	"net/http"
+	"os"
+	"strconv"
+	"strings"
+	"sync"
+	"time"
+)
+
+// pluginVersion is bumped manually on each release. Keep in sync with the
+// most recent git tag (see `git tag --sort=-v:refname | head -1`).
+const pluginVersion = "1.0.11"
+
+const (
+	telemetryProject = "traefikoidc"
+	telemetryTimeout = 2 * time.Second
+)
+
+// telemetryEndpoint is intentionally a var rather than a const so the test
+// suite in this package can retarget it at an httptest server. Production
+// code never mutates it.
+var telemetryEndpoint = "https://oss.raczylo.com/v1/ping"
+
+// telemetryOnce guarantees a single anonymous "plugin loaded" ping per
+// process lifetime. Traefik can instantiate a middleware many times per
+// process (one per route using the plugin); the sync.Once gate keeps the
+// fire-and-forget call from amplifying into many pings.
+//
+// Reset in tests via `telemetryOnce = sync.Once{}`.
+var telemetryOnce sync.Once
+
+// telemetryInflight tracks any background goroutine started by sendTelemetry.
+// Tests Wait on it to drain in-flight goroutines before mutating package
+// state. Production code never calls Wait — the goroutine is fire-and-forget.
+var telemetryInflight sync.WaitGroup
+
+// sendTelemetry fires one anonymous usage ping in the background. It is
+// failproof by contract:
+//
+//   - never blocks the caller
+//   - never panics (the goroutine recovers internally)
+//   - never returns errors
+//   - silently dropped on invalid input, env-driven opt-out, or network failure
+//
+// Opt-out is honored via any of:
+//
+//   - DO_NOT_TRACK=1
+//   - OSS_TELEMETRY_DISABLED=1
+//   - TRAEFIKOIDC_DISABLE_TELEMETRY=1
+//
+// Yaegi note: this file deliberately avoids generics (atomic.Pointer[T]) and
+// range-over-int (Go 1.22) so it interprets under any reasonably recent
+// Traefik yaegi runtime.
+func sendTelemetry(version string) {
+	telemetryOnce.Do(func() {
+		if telemetryDisabledByEnv() {
+			return
+		}
+		if !validTelemetryVersion(version) {
+			return
+		}
+		telemetryInflight.Add(1)
+		go func() {
+			defer telemetryInflight.Done()
+			defer func() { _ = recover() }()
+			doTelemetryPost(version)
+		}()
+	})
+}
+
+func telemetryDisabledByEnv() bool {
+	keys := []string{
+		"DO_NOT_TRACK",
+		"OSS_TELEMETRY_DISABLED",
+		"TRAEFIKOIDC_DISABLE_TELEMETRY",
+	}
+	for _, k := range keys {
+		v := strings.ToLower(strings.TrimSpace(os.Getenv(k)))
+		if v == "1" || v == "true" || v == "yes" || v == "on" {
+			return true
+		}
+	}
+	return false
+}
+
+// validTelemetryVersion mirrors the server-side regex ^[A-Za-z0-9.+_-]{1,32}$
+// using a byte loop. No allocation, no regexp dependency.
+//
+// Yaegi note: written as an `||` chain rather than `switch{case A,B,C:}` —
+// some yaegi releases mis-evaluate comma-separated case expressions in
+// switch-true blocks, returning false for all inputs.
+func validTelemetryVersion(v string) bool {
+	if len(v) == 0 || len(v) > 32 {
+		return false
+	}
+	for i := 0; i < len(v); i++ {
+		c := v[i]
+		ok := (c >= 'A' && c <= 'Z') ||
+			(c >= 'a' && c <= 'z') ||
+			(c >= '0' && c <= '9') ||
+			c == '.' || c == '+' || c == '_' || c == '-'
+		if !ok {
+			return false
+		}
+	}
+	return true
+}
+
+// doTelemetryPost builds the JSON body manually. The project name is a
+// constant and the version is pre-validated against an ASCII-only allowlist,
+// so direct concatenation needs no JSON escaping.
+func doTelemetryPost(version string) {
+	body := make([]byte, 0, 96)
+	body = append(body, `{"project":"`...)
+	body = append(body, telemetryProject...)
+	body = append(body, `","version":"`...)
+	body = append(body, version...)
+	body = append(body, `","ts":`...)
+	body = strconv.AppendInt(body, time.Now().Unix(), 10)
+	body = append(body, '}')
+
+	ctx, cancel := context.WithTimeout(context.Background(), telemetryTimeout)
+	defer cancel()
+
+	url := telemetryEndpoint
+	req, err := http.NewRequestWithContext(ctx, http.MethodPost, url, bytes.NewReader(body))
+	if err != nil {
+		return
+	}
+	req.Header.Set("Content-Type", "application/json")
+
+	client := &http.Client{Timeout: telemetryTimeout}
+	resp, err := client.Do(req)
+	if err != nil {
+		return
+	}
+	_ = resp.Body.Close()
+}

telemetry_test.go

@@ -0,0 +1,167 @@
+package traefikoidc
+
+import (
+	"encoding/json"
+	"io"
+	"net/http"
+	"net/http/httptest"
+	"strings"
+	"sync"
+	"sync/atomic"
+	"testing"
+	"time"
+)
+
+// resetTelemetryState restores package-level mutable state so tests do not
+// contaminate one another. The cleanup waits for any in-flight ping goroutine
+// to finish before restoring telemetryEndpoint — without that drain step the
+// goroutine and the cleanup would race on the var.
+func resetTelemetryState(t *testing.T) {
+	t.Helper()
+	telemetryOnce = sync.Once{}
+	prev := telemetryEndpoint
+	t.Cleanup(func() {
+		telemetryInflight.Wait()
+		telemetryEndpoint = prev
+		telemetryOnce = sync.Once{}
+	})
+}
+
+func newTelemetryServer(t *testing.T, status int) (hits *int32, lastBody func() string) {
+	t.Helper()
+	var counter int32
+	var mu sync.Mutex
+	var body string
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		atomic.AddInt32(&counter, 1)
+		b, _ := io.ReadAll(r.Body)
+		_ = r.Body.Close()
+		mu.Lock()
+		body = string(b)
+		mu.Unlock()
+		w.WriteHeader(status)
+	}))
+	telemetryEndpoint = srv.URL
+	t.Cleanup(srv.Close)
+	return &counter, func() string {
+		mu.Lock()
+		defer mu.Unlock()
+		return body
+	}
+}
+
+func TestValidTelemetryVersion(t *testing.T) {
+	good := []string{"1.2.3", "1.4.0-beta1", "2.0", "v1.0.0", "1.0.0+meta", "dev"}
+	for _, v := range good {
+		if !validTelemetryVersion(v) {
+			t.Errorf("validTelemetryVersion(%q) = false, want true", v)
+		}
+	}
+	bad := []string{"", "has space", "semi;colon", strings.Repeat("1", 33)}
+	for _, v := range bad {
+		if validTelemetryVersion(v) {
+			t.Errorf("validTelemetryVersion(%q) = true, want false", v)
+		}
+	}
+}
+
+func TestTelemetryDisabledByEnv(t *testing.T) {
+	for _, k := range []string{"DO_NOT_TRACK", "OSS_TELEMETRY_DISABLED", "TRAEFIKOIDC_DISABLE_TELEMETRY"} {
+		t.Run(k, func(t *testing.T) {
+			t.Setenv(k, "1")
+			if !telemetryDisabledByEnv() {
+				t.Fatalf("%s=1 should disable", k)
+			}
+		})
+	}
+	t.Run("falsy_values_do_not_disable", func(t *testing.T) {
+		t.Setenv("DO_NOT_TRACK", "0")
+		t.Setenv("OSS_TELEMETRY_DISABLED", "false")
+		t.Setenv("TRAEFIKOIDC_DISABLE_TELEMETRY", "no")
+		if telemetryDisabledByEnv() {
+			t.Fatal("falsy env values should not disable")
+		}
+	})
+}
+
+func TestSendTelemetry_FiresOnceAcrossManyCalls(t *testing.T) {
+	resetTelemetryState(t)
+	hits, lastBody := newTelemetryServer(t, http.StatusNoContent)
+
+	for i := 0; i < 50; i++ {
+		sendTelemetry("1.2.3")
+	}
+	telemetryInflight.Wait()
+
+	if got := atomic.LoadInt32(hits); got != 1 {
+		t.Fatalf("expected exactly 1 hit, got %d", got)
+	}
+
+	var payload struct {
+		Project string `json:"project"`
+		Version string `json:"version"`
+		Ts      int64  `json:"ts"`
+	}
+	if err := json.Unmarshal([]byte(lastBody()), &payload); err != nil {
+		t.Fatalf("server received non-JSON body: %q (err: %v)", lastBody(), err)
+	}
+	if payload.Project != "traefikoidc" || payload.Version != "1.2.3" || payload.Ts <= 0 {
+		t.Fatalf("unexpected payload: %+v", payload)
+	}
+}
+
+func TestSendTelemetry_RespectsDisableEnv(t *testing.T) {
+	resetTelemetryState(t)
+	hits, _ := newTelemetryServer(t, http.StatusNoContent)
+	t.Setenv("DO_NOT_TRACK", "1")
+
+	sendTelemetry("1.2.3")
+	telemetryInflight.Wait()
+
+	if got := atomic.LoadInt32(hits); got != 0 {
+		t.Fatalf("DO_NOT_TRACK should suppress; got %d hits", got)
+	}
+}
+
+func TestSendTelemetry_DropsInvalidVersion(t *testing.T) {
+	resetTelemetryState(t)
+	hits, _ := newTelemetryServer(t, http.StatusNoContent)
+
+	sendTelemetry("has space")
+	telemetryInflight.Wait()
+
+	if got := atomic.LoadInt32(hits); got != 0 {
+		t.Fatalf("invalid version should suppress; got %d hits", got)
+	}
+}
+
+func TestSendTelemetry_DoesNotBlock(t *testing.T) {
+	resetTelemetryState(t)
+	// Hanging server proves the caller is never blocked. The 2s context
+	// timeout in doTelemetryPost ensures the goroutine eventually exits;
+	// resetTelemetryState's cleanup waits for that drain before restoring
+	// telemetryEndpoint so there is no race with this test's mutation.
+	hung := httptest.NewServer(http.HandlerFunc(func(_ http.ResponseWriter, _ *http.Request) {
+		time.Sleep(5 * time.Second)
+	}))
+	t.Cleanup(hung.Close)
+	telemetryEndpoint = hung.URL
+
+	start := time.Now()
+	sendTelemetry("1.2.3")
+	if elapsed := time.Since(start); elapsed > 50*time.Millisecond {
+		t.Fatalf("sendTelemetry blocked for %v, expected near-instant return", elapsed)
+	}
+}
+
+func TestSendTelemetry_SurvivesServerError(t *testing.T) {
+	resetTelemetryState(t)
+	hits, _ := newTelemetryServer(t, http.StatusInternalServerError)
+
+	sendTelemetry("1.2.3")
+	telemetryInflight.Wait()
+
+	if got := atomic.LoadInt32(hits); got != 1 {
+		t.Fatalf("request should still reach server even on 500; got %d hits", got)
+	}
+}

token_manager.go

@@ -29,6 +29,29 @@ import (
 //
 //nolint:gocognit,gocyclo // Complex token verification logic requires multiple security checks
 func (t *TraefikOidc) VerifyToken(token string) error {
+	return t.verifyTokenWithOpts(token, verifyOpts{})
+}
+
+// verifyOpts are internal-only knobs for verifyTokenWithOpts. Kept unexported
+// because they expose subtle replay-protection semantics that are dangerous
+// to misuse.
+type verifyOpts struct {
+	// skipReplayMarking suppresses the JTI -> blacklist Set near the bottom
+	// of verifyTokenWithOpts. The Get at the top remains active, so revoked
+	// tokens (added to the blacklist by RevokeToken) are still rejected.
+	// Used exclusively by the bearer-auth path, where bearer tokens are
+	// designed to be reused until exp.
+	skipReplayMarking bool
+}
+
+// verifyTokenWithOpts runs the full token verification pipeline used by both
+// the cookie path and the bearer path. The cookie path uses the zero-value
+// opts; the bearer path sets skipReplayMarking=true. See the security spec
+// (docs/superpowers/specs/2026-05-18-bearer-token-auth-design.md §7.7) for
+// the exact contract: skipReplayMarking gates ONLY the JTI Set, never the Get.
+//
+//nolint:gocognit,gocyclo // Complex token verification logic requires multiple security checks
+func (t *TraefikOidc) verifyTokenWithOpts(token string, opts verifyOpts) error {
 	if token == "" {
 		return fmt.Errorf("invalid JWT format: token is empty")
 	}
@@ -76,7 +99,9 @@ func (t *TraefikOidc) VerifyToken(token string) error {
 	}
 
 	// Only check JTI blacklist for tokens that aren't already in the cache
-	// This is for FIRST-TIME validation to detect replay attacks
+	// This is for FIRST-TIME validation to detect replay attacks. The
+	// blacklist Get is ALWAYS active on the bearer path too — only the
+	// Set below is gated by opts.skipReplayMarking.
 	if jti, ok := parsedJWT.Claims["jti"].(string); ok && jti != "" {
 		// Skip JTI blacklist check if replay detection is disabled
 		if !t.disableReplayDetection {
@@ -105,8 +130,12 @@ func (t *TraefikOidc) VerifyToken(token string) error {
 
 	t.cacheVerifiedToken(token, jwt.Claims)
 
-	if jti, ok := jwt.Claims["jti"].(string); ok && jti != "" && !t.disableReplayDetection {
-		// Only add to blacklist if replay detection is enabled
+	// Replay marking: add JTI to blacklist so subsequent presentations of
+	// the SAME token can short-circuit via cache. Bearer path suppresses
+	// this Set (opts.skipReplayMarking=true) because bearer tokens are
+	// designed for reuse until exp; the cache-evict-then-replay scenario
+	// would otherwise trigger false replay detection.
+	if jti, ok := jwt.Claims["jti"].(string); ok && jti != "" && !t.disableReplayDetection && !opts.skipReplayMarking {
 		expiry := time.Now().Add(defaultBlacklistDuration)
 		if expClaim, expOk := jwt.Claims["exp"].(float64); expOk {
 			expTime := time.Unix(int64(expClaim), 0)
@@ -341,7 +370,17 @@ func (t *TraefikOidc) VerifyJWTSignatureAndClaims(jwt *JWT, token string) error
 
 	if err := verifySignatureWithKey(token, pubKey, alg); err != nil {
 		if !t.suppressDiagnosticLogs {
-			t.safeLogErrorf("DIAGNOSTIC: Signature verification failed for kid=%s, alg=%s: %v", kid, alg, err)
+			// Microsoft Graph access tokens carry a `nonce` JWT header and are
+			// signed in a proprietary form Microsoft documents as unverifiable
+			// by client applications. They reach this path only when the
+			// per-provider classifier (validateAzureTokens) didn't catch them,
+			// so log at debug to keep the error stream actionable while still
+			// surfacing the cause for diagnostics.
+			if _, isMSProprietary := jwt.Header["nonce"]; isMSProprietary {
+				t.safeLogDebugf("DIAGNOSTIC: Signature verification failed for kid=%s, alg=%s (Microsoft proprietary nonce header — token is opaque to clients): %v", kid, alg, err)
+			} else {
+				t.safeLogErrorf("DIAGNOSTIC: Signature verification failed for kid=%s, alg=%s: %v", kid, alg, err)
+			}
 		}
 		return fmt.Errorf("signature verification failed: %w", err)
 	}
@@ -785,6 +824,27 @@ func (t *TraefikOidc) isGoogleProvider() bool {
 	return strings.Contains(issuerURL, "google") || strings.Contains(issuerURL, "accounts.google.com")
 }
 
+// isUnverifiableAzureAccessToken reports whether a JWT-shaped access token
+// matches the Microsoft proprietary format that client applications must not
+// validate. Microsoft injects a `nonce` value into the JWT header, signs over
+// the SHA256 hash of that nonce, and ships the original nonce on the wire,
+// guaranteeing that any standard JWS verifier rejects the signature. This is
+// the documented mechanism that keeps access tokens opaque to non-resource
+// holders (Microsoft Graph, Azure Management API).
+//
+// https://learn.microsoft.com/en-us/entra/identity-platform/access-tokens
+//
+// Returns true on parse failure as well — a token we cannot parse should not
+// be passed through the verification path that emits ERROR logs.
+func (t *TraefikOidc) isUnverifiableAzureAccessToken(token string) bool {
+	parsed, err := parseJWT(token)
+	if err != nil {
+		return true
+	}
+	_, hasProprietaryNonce := parsed.Header["nonce"]
+	return hasProprietaryNonce
+}
+
 // isAzureProvider detects if the configured OIDC provider is Azure AD.
 // It checks the issuer URL for Microsoft Azure AD domains.
 // Returns:
@@ -827,6 +887,31 @@ func (t *TraefikOidc) validateAzureTokens(session *SessionData) (bool, bool, boo
 
 	if accessToken != "" {
 		if strings.Count(accessToken, ".") == 2 {
+			// Microsoft documents that client apps cannot validate access
+			// tokens issued for Microsoft-owned APIs (Graph, Azure Mgmt) due
+			// to their proprietary signing format (nonce in JWT header is
+			// the marker — signed bytes hash the nonce, wire bytes ship the
+			// raw value, so rsa verification always fails). Treat such
+			// tokens as opaque, matching Microsoft's guidance and avoiding
+			// per-request signature-error log spam (issue #134 followup).
+			//
+			// https://learn.microsoft.com/en-us/entra/identity-platform/access-tokens
+			//   "you can't validate tokens for Microsoft Graph according to
+			//    these rules due to their proprietary format"
+			if t.isUnverifiableAzureAccessToken(accessToken) {
+				t.logger.Debug("Azure access token is Microsoft-proprietary (Graph/Mgmt) — treating as opaque per Microsoft guidance")
+				if idToken != "" {
+					if err := t.verifyToken(idToken); err != nil {
+						t.logger.Debugf("Azure: ID token validation failed while access token was opaque: %v", err)
+						if session.GetRefreshToken() != "" {
+							return false, true, false
+						}
+						return false, false, true
+					}
+					return t.validateTokenExpiry(session, idToken)
+				}
+				return true, false, false
+			}
 			if err := t.verifyToken(accessToken); err != nil {
 				if idToken != "" {
 					if err := t.verifyToken(idToken); err != nil {

types.go

@@ -65,7 +65,19 @@ type ProviderMetadata struct {
 // the complete authentication flow. It's designed to work seamlessly with Traefik's
 // plugin system and provides flexible configuration options.
 type TraefikOidc struct {
-	lastMetadataRetryTime      time.Time
+	// lastMetadataRetryNano is the UnixNano timestamp of the last metadata
+	// recovery attempt. Stored atomically so the hot ServeHTTP path can
+	// throttle retries without acquiring metadataRetryMutex on every request.
+	lastMetadataRetryNano      int64
+	// firstRequestStarted is 0 until the very first non-health request fires
+	// the background-task bootstrap; then it flips to 1 via CAS. Replaces the
+	// firstRequestMutex + firstRequestReceived combo which previously took
+	// a write lock on every non-health request forever.
+	firstRequestStarted        int32
+	// metadataRefreshStartedAtomic is the CAS-only variant of the old
+	// metadataRefreshStarted bool. Both flags live under the same atomic so
+	// concurrent first-request goroutines race exactly once.
+	metadataRefreshStartedAtomic int32
 	jwkCache                   JWKCacheInterface
 	jwtVerifier                JWTVerifier
 	ctx                        context.Context
@@ -130,17 +142,13 @@ type TraefikOidc struct {
 	maxRefreshTokenAge         time.Duration
 	metadataMu                 sync.RWMutex
 	shutdownOnce               sync.Once
-	metadataRetryMutex         sync.Mutex
-	firstRequestMutex          sync.Mutex
 	sessionInvalidationCache   CacheInterface
 	refreshResultCache         CacheInterface
 	minimalHeaders             bool
 	stripAuthCookies           bool
 	enableBackchannelLogout    bool
 	enableFrontchannelLogout   bool
-	firstRequestReceived       bool
 	requireTokenIntrospection  bool
-	metadataRefreshStarted     bool
 	allowPrivateIPAddresses    bool
 	disableReplayDetection     bool
 	allowOpaqueTokens          bool
@@ -149,4 +157,17 @@ type TraefikOidc struct {
 	enablePKCE                 bool
 	forceHTTPS                 bool
 	suppressDiagnosticLogs     bool
+
+	// Bearer-auth runtime state (populated only when EnableBearerAuth=true).
+	bearerIdentifierClaim     string
+	bearerFailureTracker      *bearerFailureTracker
+	maxTokenAge               time.Duration
+	maxIdentifierLength       int
+	bearerFailureThreshold    int
+	bearerFailureWindow       time.Duration
+	bearerFailurePenalty      time.Duration
+	enableBearerAuth          bool
+	stripAuthorizationHeader  bool
+	bearerEmitWWWAuthenticate bool
+	bearerOverridesCookie     bool
 }