feat(go): share token verification state through Redis

INSTALLATION.md

@@ -271,9 +271,11 @@ services:
 ```
 
 FCaptcha state is process-local by default. In the Go server, `REDIS_URL` now
-shares PoW challenges and their atomic single-use claim across instances. Other
-state remains local, so continue to run one instance until the remaining stores
-and the Node/Python implementations gain shared-state support.
+shares PoW challenges, token replay protection, and Siteverify idempotency
+results across instances. Challenge and token claims are atomic. Rate limits,
+suspicion, and fingerprints remain local, so continue to run one instance until
+the remaining stores and the Node/Python implementations gain shared-state
+support.
 
 Run:
 
@@ -422,10 +424,10 @@ server {
 }
 ```
 
-**Important:** Do not run multiple instances yet. The Go server can share PoW
-challenge state through Redis, but token replay protection, rate limiting,
-suspicion, fingerprints, and idempotency are still process-local. Node and
-Python remain entirely process-local.
+**Important:** Do not run multiple instances yet. The Go server shares PoW,
+token replay, and Siteverify idempotency state through Redis, but rate limiting,
+suspicion, and fingerprints are still process-local. Node and Python remain
+entirely process-local.
 
 ---
 
@@ -437,7 +439,7 @@ Python remain entirely process-local.
 |----------|----------|---------|-------------|
 | `FCAPTCHA_SECRET` | Yes | - | Secret key for signing tokens (min 16 chars) |
 | `FCAPTCHA_INSECURE_DEV_MODE` | No | off | Explicitly use the public development signing key for local-only development. Never expose a server with this enabled |
-| `REDIS_URL` | No | - | Go only: Redis URL used for shared PoW challenge and atomic replay state. Configuration is fail-closed; it does not yet make every store distributed |
+| `REDIS_URL` | No | - | Go only: Redis URL for shared PoW, token replay, and Siteverify idempotency state. Configuration is fail-closed; it does not yet make every store distributed |
 | `FCAPTCHA_VERIFY_SECRET` | No | `FCAPTCHA_SECRET` | Credential your backend sends as `secret` when verifying a token. Split it from the signing key so a leaked verify credential cannot also mint tokens |
 | `FCAPTCHA_LEGACY_UNAUTH_VERIFY` | No | off | Restore the pre-1.22.0 behaviour where token verification accepted any caller. Migration cover for one release — see [Upgrading to 1.22.0](#upgrading-to-1220) |
 | `FCAPTCHA_ALLOWED_HOSTNAMES` | No | (any) | Comma-separated hostnames permitted to mint tokens, matched against the request `Origin` (then `Referer`) |

README.md

@@ -57,15 +57,15 @@ FCAPTCHA_SECRET=my-secret docker compose -f docker/docker-compose.yml up -d
 ```
 
 FCaptcha state is currently process-local by default. The Go server can use
-`REDIS_URL` to share PoW challenges and atomically consume them across replicas;
+`REDIS_URL` to share PoW challenges, token replay protection, and Siteverify
+idempotency results across replicas; challenge and token consumption are atomic,
 it refuses to start if the configured Redis service is unavailable and returns
 503 rather than issuing an unpersisted challenge if Redis fails later.
 
-This is the first shared-state slice, not yet permission to scale the entire
-service horizontally: token replay protection, rate limits, suspicion history,
-fingerprint history, and idempotency results remain per-process, and the Node
-and Python servers do not yet use Redis. Run one replica until those stores are
-also shared.
+This is not yet permission to scale the entire service horizontally: rate
+limits, suspicion history, and fingerprint history remain per-process, and the
+Node and Python servers do not yet use Redis. Run one replica until those stores
+are also shared.
 
 Kubernetes:
 

charts/fcaptcha/README.md

@@ -38,10 +38,11 @@ to the controller and rate limiting collapses onto one address — silently. The
 default covers the usual in-cluster pod CIDRs; narrow it to your controller's
 range.
 
-**Run one replica for now.** With the Go image, `redis.url` shares PoW challenges
-and their atomic one-time claim. Token replay protection, rate limits,
-suspicion, fingerprints, and idempotency are still per-pod; Node and Python do
-not yet use Redis. `autoscaling` remains off until those remaining stores land.
+**Run one replica for now.** With the Go image, `redis.url` shares PoW challenges,
+token replay protection, and Siteverify idempotency results; challenge and token
+claims are atomic. Rate limits, suspicion, and fingerprints are still per-pod;
+Node and Python do not yet use Redis. `autoscaling` remains off until those
+remaining stores land.
 
 The full list of deployment settings with security consequences is in
 [SECURITY.md](https://github.com/WebDecoy/FCaptcha/blob/main/SECURITY.md#deployment-notes-that-are-security-relevant).
@@ -61,7 +62,7 @@ The full list of deployment settings with security consequences is in
 | `config.logVerdictsIncludeRaw` | Also log free-text detection reasons. **Can contain visitor-derived data** | `false` |
 | `image.repository` / `image.tag` | Container image; tag defaults to the chart's `appVersion` | `ghcr.io/webdecoy/fcaptcha` |
 | `replicaCount` | See the note above before raising this | `1` |
-| `redis.url` / `redis.existingSecret` | Go only: shared PoW challenge and atomic claim state; remaining stores are still local | `""` |
+| `redis.url` / `redis.existingSecret` | Go only: shared PoW, token replay, and Siteverify idempotency state; remaining stores are still local | `""` |
 | `service.type` / `service.port` | | `ClusterIP` / `80` |
 | `ingress.enabled` | | `false` |
 | `resources` | | 100m CPU / 128Mi requested |

charts/fcaptcha/values.yaml

@@ -68,9 +68,9 @@ extraEnv: []
 ## captcha chart, and the honest answer for anything you care about keeping is a
 ## managed instance or a purpose-built operator. Point this at one.
 ##
-## The Go server uses this for shared PoW challenge and atomic claim state.
-## Other stores remain in memory, and Node/Python do not use it yet, so
-## replicaCount still defaults to 1.
+## The Go server uses this for shared PoW, token replay, and Siteverify
+## idempotency state. Other stores remain in memory, and Node/Python do not use
+## it yet, so replicaCount still defaults to 1.
 redis:
   url: ""
   existingSecret: ""
@@ -147,9 +147,9 @@ tolerations: []
 affinity: {}
 topologySpreadConstraints: []
 
-## Off by default because token replay protection, rate limiting, suspicion,
-## fingerprints, and idempotency remain per-pod. Go can share PoW challenges,
-## but that alone does not make horizontal scaling safe.
+## Off by default because rate limiting, suspicion, and fingerprints remain
+## per-pod. Go shares PoW, token replay, and idempotency state, but that alone
+## does not make horizontal scaling safe.
 autoscaling:
   enabled: false
   minReplicas: 2

server-go/main.go

@@ -325,7 +325,11 @@ func main() {
 	// integration at this server should be a base-URL change and nothing else.
 	// See siteverify.go for the adapter itself.
 	log.Printf("allowed hostnames: %s", engine.allowedHostnames.Describe())
-	compat := siteverifyHandler(engine, proxyTrust, NewIdempotencyStore(), verifySecret, requireVerifySecret)
+	idempotencyStore := NewIdempotencyStore()
+	if engine.redisClient != nil {
+		idempotencyStore = NewRedisIdempotencyStore(engine.redisClient)
+	}
+	compat := siteverifyHandler(engine, proxyTrust, idempotencyStore, verifySecret, requireVerifySecret)
 	r.Post("/turnstile/v0/siteverify", compat)
 	r.Post("/recaptcha/api/siteverify", compat)
 	r.Post("/siteverify", compat)

server-go/redis_state_test.go

@@ -37,3 +37,45 @@ func TestRedisPoWFailureIsClosed(t *testing.T) {
 		t.Fatal("challenge generation must fail closed when shared state is unavailable")
 	}
 }
+
+func TestRedisTokenReplayProtectionCrossesInstances(t *testing.T) {
+	redisServer := miniredis.RunT(t)
+	redisURL := "redis://" + redisServer.Addr()
+	issuer := NewScoringEngineWithRedis("shared-test-secret", redisURL)
+	verifier := NewScoringEngineWithRedis("shared-test-secret", redisURL)
+	token := issuer.generateToken("203.0.113.5", "site-1", 0.1, TokenBinding{})
+
+	if valid, _ := verifier.VerifyToken(token)["valid"].(bool); !valid {
+		t.Fatal("another instance could not validate the token")
+	}
+	replayed := issuer.VerifyToken(token)
+	if valid, _ := replayed["valid"].(bool); valid || replayed["reason"] != "token_already_used" {
+		t.Fatalf("spent token was reusable: %#v", replayed)
+	}
+}
+
+func TestRedisTokenClaimFailsClosed(t *testing.T) {
+	redisServer := miniredis.RunT(t)
+	engine := NewScoringEngineWithRedis("shared-test-secret", "redis://"+redisServer.Addr())
+	token := engine.generateToken("203.0.113.5", "site-1", 0.1, TokenBinding{})
+	redisServer.Close()
+
+	result := engine.VerifyToken(token)
+	if valid, _ := result["valid"].(bool); valid || result["reason"] != "state_unavailable" {
+		t.Fatalf("verification did not fail closed: %#v", result)
+	}
+}
+
+func TestRedisIdempotencyResultsCrossInstances(t *testing.T) {
+	redisServer := miniredis.RunT(t)
+	client := NewScoringEngineWithRedis("shared-test-secret", "redis://"+redisServer.Addr()).redisClient
+	first := NewRedisIdempotencyStore(client)
+	second := NewRedisIdempotencyStore(client)
+	want := siteverifySuccess("2026-08-23T00:00:00.000Z", "example.com", "login", "request-1", 0.1)
+
+	first.Set("retry-key", "token-value", want)
+	got, ok := second.Get("retry-key", "token-value")
+	if !ok || got["hostname"] != want["hostname"] || got["action"] != want["action"] {
+		t.Fatalf("another instance did not read the idempotency result: %#v", got)
+	}
+}

server-go/scoring.go

@@ -294,6 +294,7 @@ type ScoringEngine struct {
 	// Which page origins may mint tokens. Unrestricted unless the operator sets
 	// FCAPTCHA_ALLOWED_HOSTNAMES; see siteverify.go.
 	allowedHostnames *HostnameAllowlist
+	redisClient      *redis.Client
 }
 
 // webBotAuthTimeout bounds the whole Web Bot Auth verification, including the
@@ -328,6 +329,7 @@ type FingerprintData struct {
 type TokenStore struct {
 	mu    sync.Mutex
 	cache *expirable.LRU[string, struct{}]
+	redis *redis.Client
 }
 
 const (
@@ -341,18 +343,36 @@ func newTokenStore() *TokenStore {
 	}
 }
 
+func newRedisTokenStore(client *redis.Client) *TokenStore {
+	return &TokenStore{redis: client}
+}
+
 func (t *TokenStore) IsUsed(sig string) bool {
+	if t.redis != nil {
+		exists, err := t.redis.Exists(context.Background(), redisStatePrefix+"token:spent:"+sig).Result()
+		return err == nil && exists > 0
+	}
 	return t.cache.Contains(sig)
 }
 
 func (t *TokenStore) MarkUsed(sig string) bool {
+	claimed, err := t.Claim(sig)
+	return err == nil && claimed
+}
+
+// Claim atomically marks a token signature as spent. Redis SET NX makes this
+// safe when two backend replicas validate the same token concurrently.
+func (t *TokenStore) Claim(sig string) (bool, error) {
+	if t.redis != nil {
+		return t.redis.SetNX(context.Background(), redisStatePrefix+"token:spent:"+sig, "1", usedTokensTTL).Result()
+	}
 	t.mu.Lock()
 	defer t.mu.Unlock()
 	if t.cache.Contains(sig) {
-		return false
+		return false, nil
 	}
 	t.cache.Add(sig, struct{}{})
-	return true
+	return true, nil
 }
 
 // NewScoringEngine creates a new engine
@@ -403,6 +423,8 @@ func NewScoringEngineWithRedis(secretKey, redisURL string) *ScoringEngine {
 	}
 	engine := NewScoringEngine(secretKey)
 	engine.powStore = newRedisPoWChallengeStore(client)
+	engine.tokenStore = newRedisTokenStore(client)
+	engine.redisClient = client
 	return engine
 }
 
@@ -922,13 +944,6 @@ func (e *ScoringEngine) VerifyTokenWithIP(token, ip string) map[string]interface
 		return result
 	}
 
-	// Check for token replay (single-use tokens)
-	if e.tokenStore.IsUsed(sig) {
-		result["valid"] = false
-		result["reason"] = "token_already_used"
-		return result
-	}
-
 	// Verify IP matches (if provided)
 	if ip != "" {
 		ipHash, _ := data["ip_hash"].(string)
@@ -941,8 +956,20 @@ func (e *ScoringEngine) VerifyTokenWithIP(token, ip string) map[string]interface
 		}
 	}
 
-	// Mark token as used (prevents replay)
-	e.tokenStore.MarkUsed(sig)
+	// Atomically claim the token only after all non-mutating validation passes.
+	// This is authoritative; a separate IsUsed/MarkUsed pair races across both
+	// goroutines and replicas.
+	claimed, err := e.tokenStore.Claim(sig)
+	if err != nil {
+		result["valid"] = false
+		result["reason"] = "state_unavailable"
+		return result
+	}
+	if !claimed {
+		result["valid"] = false
+		result["reason"] = "token_already_used"
+		return result
+	}
 
 	result["valid"] = true
 	result["site_key"] = data["site_key"]

server-go/siteverify.go

@@ -35,6 +35,7 @@ package main
 // Mirrors server-node/siteverify.js and server-python/siteverify.py.
 
 import (
+	"context"
 	"crypto/sha256"
 	"crypto/subtle"
 	"encoding/hex"
@@ -47,6 +48,7 @@ import (
 	"time"
 
 	"github.com/hashicorp/golang-lru/v2/expirable"
+	"github.com/redis/go-redis/v9"
 )
 
 const (
@@ -74,13 +76,13 @@ const (
 // The error-code vocabulary. These strings are the contract — integrators
 // branch on them — so they match Cloudflare's spelling exactly, hyphens and all.
 const (
-	errMissingSecret     = "missing-input-secret"
-	errInvalidSecret     = "invalid-input-secret"
-	errMissingResponse   = "missing-input-response"
-	errInvalidResponse   = "invalid-input-response"
-	errBadRequest        = "bad-request"
+	errMissingSecret      = "missing-input-secret"
+	errInvalidSecret      = "invalid-input-secret"
+	errMissingResponse    = "missing-input-response"
+	errInvalidResponse    = "invalid-input-response"
+	errBadRequest         = "bad-request"
 	errTimeoutOrDuplicate = "timeout-or-duplicate"
-	errInternalError     = "internal-error"
+	errInternalError      = "internal-error"
 )
 
 // reasonToErrorCode maps VerifyTokenWithIP's internal reason onto that
@@ -271,6 +273,14 @@ func (a *HostnameAllowlist) Describe() string {
 // degrades that case to an ordinary fresh verification.
 type IdempotencyStore struct {
 	entries *expirable.LRU[string, SiteverifyResponse]
+	redis   *redis.Client
+}
+
+// NewRedisIdempotencyStore shares retry responses between replicas. Keys are
+// hashed before leaving the process so caller-provided idempotency values and
+// tokens never appear in Redis key listings.
+func NewRedisIdempotencyStore(client *redis.Client) *IdempotencyStore {
+	return &IdempotencyStore{redis: client}
 }
 
 // NewIdempotencyStore builds a bounded, self-expiring result cache.
@@ -285,11 +295,27 @@ func (s *IdempotencyStore) key(idempotencyKey, token string) string {
 	return idempotencyKey + ":" + hex.EncodeToString(sum[:])[:32]
 }
 
+func (s *IdempotencyStore) redisKey(idempotencyKey, token string) string {
+	sum := sha256.Sum256([]byte(s.key(idempotencyKey, token)))
+	return redisStatePrefix + "siteverify:idempotency:" + hex.EncodeToString(sum[:])
+}
+
 // Get returns a cached response, if one is still live.
 func (s *IdempotencyStore) Get(idempotencyKey, token string) (SiteverifyResponse, bool) {
 	if idempotencyKey == "" {
 		return SiteverifyResponse{}, false
 	}
+	if s.redis != nil {
+		payload, err := s.redis.Get(context.Background(), s.redisKey(idempotencyKey, token)).Bytes()
+		if err != nil {
+			return SiteverifyResponse{}, false
+		}
+		var response SiteverifyResponse
+		if json.Unmarshal(payload, &response) != nil {
+			return SiteverifyResponse{}, false
+		}
+		return response, true
+	}
 	return s.entries.Get(s.key(idempotencyKey, token))
 }
 
@@ -298,6 +324,13 @@ func (s *IdempotencyStore) Set(idempotencyKey, token string, resp SiteverifyResp
 	if idempotencyKey == "" {
 		return
 	}
+	if s.redis != nil {
+		payload, err := json.Marshal(resp)
+		if err == nil {
+			_ = s.redis.Set(context.Background(), s.redisKey(idempotencyKey, token), payload, idempotencyTTL).Err()
+		}
+		return
+	}
 	s.entries.Add(s.key(idempotencyKey, token), resp)
 }