feat: per-group cache TTL

GroupConfig gains CacheTTL (json cacheTTL, empty inherits the global
cache.ttl). Parsed and validated at construction (newAuthStore,
groupCacheTTL in cache.go): a malformed duration, a duration <= 0, or
CacheTTL set while the global cache block is not configured is a
constructor error naming the group, mirroring Cache:true's existing
"nothing to inherit from" check.

group carries the resolved time.Duration (0 means inherit).
responseCache.store gains an explicit ttl parameter; routes_unified.go
computes effectiveTTL(g.cache, grp) immediately before each store call,
so setEx's Redis-native EX seconds reflect the group's own override.

Tests: constructor-error table, empty-inherits and valid-override
resolution, effectiveTTL inheritance/override, a scripted-RESP-server
proof that store's SET...EX carries a different value per group, and an
end-to-end Gateway.ServeHTTP test with two groups asserting the actual
wire EX value (30 vs 60) differs per group.

README: GroupConfig.cacheTTL config row and a Caching-section paragraph.

README.md

@@ -211,6 +211,7 @@ also a construction error, never a panic (`providers.go`, `mcp_a2a.go`,
 | `providers` | `[]string` | `[]` (all) | Glob-matched (`path.Match`) against a provider's configured name. |
 | `models` | `[]string` | `[]` (all) | Glob-matched, exactly as given — no automatic `provider/` prefix stripping. Both the bare and provider-prefixed forms of a model id are checked, so a pattern like `gpt-*` matches a client request for either `gpt-5-mini` or `openai/gpt-5-mini`. |
 | `cache` | `*bool` | `nil` (inherit) | Overrides the top-level `cache.enabled` setting for this group's requests — see [Caching](#caching). `nil` inherits the global setting; `false` opts the group out even when caching is globally on; `true` opts the group in, but only when the top-level `cache` block is actually configured (a construction error otherwise — there is nothing to inherit `ttl`/`maxBodyBytes` from). |
+| `cacheTTL` | `string` | `""` (inherit) | Overrides the top-level `cache.ttl` for this group's cache entries — see [Caching](#caching). Empty inherits the global TTL. A set value must be a valid, positive Go duration (`"30s"`, `"5m"`), and requires the top-level `cache` block to be configured — a construction error otherwise, the same "nothing to inherit from" reasoning as `cache: true` above. |
 | `mcpServers` | `[]string` | `[]` (all) | Glob-matched against a configured MCP server name. |
 | `agents` | `[]string` | `[]` (all) | Glob-matched against a configured agent name. |
 
@@ -584,6 +585,11 @@ cached bodies for the same request.
   enabled — the block omitted entirely, or present with `enabled: false`,
   both read the same way — is a construction error: there is nothing
   configured to inherit `ttl`/`maxBodyBytes` from.
+- **Group TTL override**: `GroupConfig.cacheTTL` overrides the top-level
+  `cache.ttl` for one group's cache entries only, independently of that
+  group's own `cache` setting — an empty value inherits the global TTL, a
+  set value must parse as a positive Go duration and requires the
+  top-level `cache` block to be configured, or construction fails.
 - **Storage**: values are stored via the same hand-rolled RESP2 client
   (`resp.go`) the limiter's distributed counters use — the identical
   Redis/Valkey/Dragonfly connection, not a second one. A Redis error

auth.go

@@ -29,9 +29,19 @@ type group struct {
 	// cache is GroupConfig.Cache carried through unchanged: nil inherits
 	// the global cache.enabled setting, non-nil overrides it for this
 	// group's requests. Resolved by groupCacheEnabled (cache.go).
-	cache                                 *bool
-	name                                  string
-	providers, models, mcpServers, agents []string
+	cache      *bool
+	name       string
+	providers  []string
+	models     []string
+	mcpServers []string
+	agents     []string
+	// cacheTTL is GroupConfig.CacheTTL parsed and validated at construction
+	// (newAuthStore below): 0 inherits the global responseCache's TTL
+	// (effectiveTTL, cache.go), a positive value overrides it for this
+	// group's cache entries only. GroupConfig.CacheTTL == "" is the only
+	// input that produces 0 here — every other value either becomes a
+	// positive duration or fails newAuthStore as a constructor error.
+	cacheTTL time.Duration
 }
 
 // allowsProvider reports whether name matches one of the group's provider
@@ -143,9 +153,22 @@ func newAuthStore(cfg *Config) (*authStore, error) {
 		if gc.Cache != nil && *gc.Cache && !cfg.Cache.Enabled {
 			return nil, fmt.Errorf("llmgateway: group %q enables cache but global cache is not configured", name)
 		}
+		// CacheTTL follows the identical "nothing to inherit from" reasoning
+		// as Cache:true above: overriding a TTL only means something when
+		// there is a global cache.ttl to override, so CacheTTL set while
+		// the global cache block itself is not configured is a constructor
+		// error, independent of this group's own Cache setting (a group
+		// may set CacheTTL for when caching later becomes enabled globally
+		// is NOT supported — cfg.Cache.Enabled is checked here, not
+		// per-group, matching Cache:true's own check above).
+		cacheTTL, err := groupCacheTTL(name, gc.CacheTTL, cfg.Cache.Enabled)
+		if err != nil {
+			return nil, err
+		}
 		a.groups[name] = &group{
 			limits:     gc.Limits,
 			cache:      gc.Cache,
+			cacheTTL:   cacheTTL,
 			name:       name,
 			providers:  gc.Providers,
 			models:     gc.Models,

auth_test.go

@@ -190,6 +190,74 @@ func TestNewAuthStore_NoUsers_OK(t *testing.T) {
 	}
 }
 
+// --- GroupConfig.CacheTTL: constructor validation (item B) ---
+
+// TestNewAuthStore_GroupCacheTTL_ConstructorErrors is the constructor-error
+// table for GroupConfig.CacheTTL: a malformed duration, a zero duration, a
+// negative duration, and CacheTTL set while the global cache block is not
+// configured (nothing to inherit ttl from, the same reasoning
+// newAuthStore already applies to Cache:true) must all fail construction.
+func TestNewAuthStore_GroupCacheTTL_ConstructorErrors(t *testing.T) {
+	tests := []struct {
+		name        string
+		cacheTTL    string
+		globalCache bool
+	}{
+		{name: "malformed duration", cacheTTL: "not-a-duration", globalCache: true},
+		{name: "zero duration", cacheTTL: "0s", globalCache: true},
+		{name: "negative duration", cacheTTL: "-5s", globalCache: true},
+		{name: "set without global cache enabled", cacheTTL: "5m", globalCache: false},
+	}
+	for _, tt := range tests {
+		t.Run(tt.name, func(t *testing.T) {
+			cfg := &Config{
+				Providers: map[string]*ProviderConfig{"openai": {Type: "openai", APIKey: "k"}},
+				Groups:    map[string]*GroupConfig{"default": {CacheTTL: tt.cacheTTL}},
+				Cache:     CacheConfig{Enabled: tt.globalCache},
+			}
+			if _, err := newAuthStore(cfg); err == nil {
+				t.Fatalf("newAuthStore: want a constructor error for cacheTTL=%q (globalCache=%v)", tt.cacheTTL, tt.globalCache)
+			}
+		})
+	}
+}
+
+// TestNewAuthStore_GroupCacheTTL_EmptyInheritsZero proves an omitted
+// CacheTTL resolves to group.cacheTTL == 0 — effectiveTTL's (cache.go)
+// signal to inherit the global cache TTL rather than override it.
+func TestNewAuthStore_GroupCacheTTL_EmptyInheritsZero(t *testing.T) {
+	cfg := &Config{
+		Providers: map[string]*ProviderConfig{"openai": {Type: "openai", APIKey: "k"}},
+		Groups:    map[string]*GroupConfig{"default": {}},
+	}
+	a, err := newAuthStore(cfg)
+	if err != nil {
+		t.Fatalf("newAuthStore: %v", err)
+	}
+	if got := a.groups["default"].cacheTTL; got != 0 {
+		t.Errorf("groups[default].cacheTTL = %v, want 0 (inherit)", got)
+	}
+}
+
+// TestNewAuthStore_GroupCacheTTL_ValidOverride_ResolvedOnGroup proves a
+// well-formed, positive CacheTTL with the global cache block configured
+// both constructs cleanly and resolves onto the group struct exactly,
+// ready for effectiveTTL to read.
+func TestNewAuthStore_GroupCacheTTL_ValidOverride_ResolvedOnGroup(t *testing.T) {
+	cfg := &Config{
+		Providers: map[string]*ProviderConfig{"openai": {Type: "openai", APIKey: "k"}},
+		Groups:    map[string]*GroupConfig{"default": {CacheTTL: "30s"}},
+		Cache:     CacheConfig{Enabled: true},
+	}
+	a, err := newAuthStore(cfg)
+	if err != nil {
+		t.Fatalf("newAuthStore: %v", err)
+	}
+	if got := a.groups["default"].cacheTTL; got != 30*time.Second {
+		t.Errorf("groups[default].cacheTTL = %v, want 30s", got)
+	}
+}
+
 func TestGroup_AllowsProvider_EmptyListAllowsAll(t *testing.T) {
 	grp := &group{name: "eng"}
 	if !grp.allowsProvider("openai") {

cache.go

@@ -107,6 +107,46 @@ func groupCacheEnabled(grp *group) bool {
 	return true
 }
 
+// groupCacheTTL parses and validates a GroupConfig.CacheTTL string for
+// group name, returning the resolved time.Duration — 0 when raw is empty,
+// meaning "inherit the global cache TTL" (effectiveTTL below). cacheEnabled
+// is cfg.Cache.Enabled: a non-empty raw while the global cache block is not
+// configured is a constructor error, since there is then no global TTL to
+// override — the identical "nothing to inherit from" reasoning newAuthStore
+// already applies to Cache:true (auth.go). Called once per group by
+// newAuthStore (auth.go), independent of that group's own Cache setting.
+func groupCacheTTL(name, raw string, cacheEnabled bool) (time.Duration, error) {
+	if raw == "" {
+		return 0, nil
+	}
+	if !cacheEnabled {
+		return 0, fmt.Errorf("llmgateway: group %q sets cacheTTL but global cache is not configured", name)
+	}
+	ttl, err := time.ParseDuration(raw)
+	if err != nil {
+		return 0, fmt.Errorf("llmgateway: group %q: cacheTTL: %w", name, err)
+	}
+	if ttl <= 0 {
+		return 0, fmt.Errorf("llmgateway: group %q: cacheTTL must be positive, got %q", name, raw)
+	}
+	return ttl, nil
+}
+
+// effectiveTTL resolves the TTL to use for grp's cache entries: grp's own
+// resolved cacheTTL when set (non-zero — GroupConfig.CacheTTL parsed and
+// validated at construction, groupCacheTTL above), otherwise c's global TTL
+// (buildResponseCache's cc.TTL, validateCacheConfig). Called by
+// routes_unified.go's runUnified immediately before store, so every stored
+// entry's Redis-native expiry (setEx's ttl argument, resp.go) already
+// reflects the group-specific override — no separate "shrink the group's
+// TTL later" step exists.
+func effectiveTTL(c *responseCache, grp *group) time.Duration {
+	if grp.cacheTTL > 0 {
+		return grp.cacheTTL
+	}
+	return c.ttl
+}
+
 // cachedResponse is the value responseCache stores per key: enough to
 // replay a cached response verbatim. json.Marshal/Unmarshal base64-encode
 // Body automatically (it is a []byte field), matching spec §2's "value:
@@ -245,14 +285,16 @@ func (c *responseCache) lookup(key string) (*cachedResponse, bool) {
 	return &cr, true
 }
 
-// store saves status/contentType/body under key with the cache's
-// configured TTL. A body exceeding maxBodyBytes is skipped (logged, not
-// treated as an error) before ever touching the network; a down-latch
+// store saves status/contentType/body under key with ttl (the caller's
+// resolved TTL — effectiveTTL above; every existing caller before the
+// per-group override passed c.ttl itself, so behavior for a group with no
+// override is unchanged). A body exceeding maxBodyBytes is skipped (logged,
+// not treated as an error) before ever touching the network; a down-latch
 // already open (latched) also skips the network entirely; and a Redis
 // error writing it is logged (rate-limited) and opens the latch — none of
 // which surface to the caller: a cache write must never fail a request
 // that already succeeded and was already written to the client.
-func (c *responseCache) store(key string, status int, contentType string, body []byte) {
+func (c *responseCache) store(key string, status int, contentType string, body []byte, ttl time.Duration) {
 	if len(body) > c.maxBodyBytes {
 		// Pre-formatted with fmt.Sprintf, then logged as a single "%s"
 		// argument, not c.logf(format, key, len(body), c.maxBodyBytes)
@@ -277,7 +319,7 @@ func (c *responseCache) store(key string, status int, contentType string, body [
 		panic(fmt.Sprintf("llmgateway: cache store: marshal cached response: %v", err))
 	}
 
-	if err := c.client.setEx(key, val, c.ttl); err != nil {
+	if err := c.client.setEx(key, val, ttl); err != nil {
 		c.recordFailure(err)
 	}
 }

cache_test.go

@@ -425,7 +425,7 @@ func TestResponseCache_Store_SendsSETWithTTL(t *testing.T) {
 	c := newResponseCache(client, 90*time.Second, defaultCacheMaxBodyBytes, noop, noop)
 	c.nowFn = time.Now
 
-	c.store("llmgw:cache:k", 200, "application/json", []byte(`{"ok":true}`))
+	c.store("llmgw:cache:k", 200, "application/json", []byte(`{"ok":true}`), 90*time.Second)
 }
 
 func TestResponseCache_Store_OversizeBody_SkipsSilentlyLogged(t *testing.T) {
@@ -443,7 +443,7 @@ func TestResponseCache_Store_OversizeBody_SkipsSilentlyLogged(t *testing.T) {
 	c := newResponseCache(client, time.Minute, 4, logf, errorf) // maxBodyBytes=4
 	c.nowFn = time.Now
 
-	c.store("k", 200, "text/plain", []byte("way too big"))
+	c.store("k", 200, "text/plain", []byte("way too big"), time.Minute)
 
 	if len(logged) != 1 {
 		t.Fatalf("logged = %d lines, want 1 (oversize skip is logged via logf, not errorf)", len(logged))
@@ -464,12 +464,12 @@ func TestResponseCache_Store_RedisDown_LoggedRateLimited(t *testing.T) {
 	now := time.Date(2026, 8, 20, 10, 0, 0, 0, time.UTC)
 	c.nowFn = func() time.Time { return now }
 
-	c.store("k", 200, "application/json", []byte(`{}`))
+	c.store("k", 200, "application/json", []byte(`{}`), time.Minute)
 	if len(logged) != 1 {
 		t.Fatalf("logged = %d lines after the first failed store, want 1", len(logged))
 	}
 
-	c.store("k", 200, "application/json", []byte(`{}`))
+	c.store("k", 200, "application/json", []byte(`{}`), time.Minute)
 	if len(logged) != 1 {
 		t.Fatalf("logged = %d lines still inside the rate-limit window, want 1", len(logged))
 	}
@@ -526,7 +526,7 @@ func TestResponseCache_DownLatch_SkipsNetworkCallsWithinWindow(t *testing.T) {
 	}
 
 	// store while latched: also zero network calls.
-	c.store("k", 200, "application/json", []byte(`{}`))
+	c.store("k", 200, "application/json", []byte(`{}`), time.Minute)
 	if stub.setCalls != 0 {
 		t.Fatalf("setCalls = %d while latched, want 0 (no network call attempted)", stub.setCalls)
 	}
@@ -630,3 +630,62 @@ func TestNewAuthStore_GroupCacheTrueWithGlobalCacheEnabled_NoError(t *testing.T)
 		t.Fatalf("newAuthStore: %v, want no error when the global cache block is configured", err)
 	}
 }
+
+// --- effectiveTTL: per-group cache TTL (item B) ---
+
+// TestEffectiveTTL_NilInheritsGlobal proves a group with no CacheTTL
+// override (cacheTTL == 0, GroupConfig.CacheTTL left empty) resolves to
+// the cache's own global TTL.
+func TestEffectiveTTL_NilInheritsGlobal(t *testing.T) {
+	c := &responseCache{ttl: 5 * time.Minute}
+	grp := &group{name: "g"}
+	if got := effectiveTTL(c, grp); got != 5*time.Minute {
+		t.Errorf("effectiveTTL = %v, want the global TTL 5m (group has no override)", got)
+	}
+}
+
+// TestEffectiveTTL_GroupOverrideHonored proves a group's own resolved
+// cacheTTL wins over the cache's global TTL.
+func TestEffectiveTTL_GroupOverrideHonored(t *testing.T) {
+	c := &responseCache{ttl: 5 * time.Minute}
+	grp := &group{name: "g", cacheTTL: 30 * time.Second}
+	if got := effectiveTTL(c, grp); got != 30*time.Second {
+		t.Errorf("effectiveTTL = %v, want the group override 30s", got)
+	}
+}
+
+// TestResponseCache_Store_EffectiveTTLDiffersPerGroup_SETCarriesGroupTTL is
+// item B's "override honored" proof against a real (scripted) RESP wire:
+// store's ttl argument — effectiveTTL's per-group resolution, computed by
+// the caller (routes_unified.go) before store is ever invoked — reaches
+// the SET...EX command as a genuinely different EX value for a group that
+// overrides the global TTL (30) versus one that inherits it (60, the
+// cache's own 1m global TTL). runFakeRESPServer fails the test outright if
+// either SET carries the wrong EX value.
+func TestResponseCache_Store_EffectiveTTLDiffersPerGroup_SETCarriesGroupTTL(t *testing.T) {
+	val1, err := json.Marshal(cachedResponse{Status: 200, ContentType: "application/json", Body: []byte(`{"a":1}`)})
+	if err != nil {
+		t.Fatalf("json.Marshal: %v", err)
+	}
+	val2, err := json.Marshal(cachedResponse{Status: 200, ContentType: "application/json", Body: []byte(`{"b":2}`)})
+	if err != nil {
+		t.Fatalf("json.Marshal: %v", err)
+	}
+
+	ln := newFakeListener(t)
+	runFakeRESPServer(t, ln, []respStep{
+		{wantArgs: []string{"SELECT", "0"}, reply: []byte("+OK\r\n")},
+		{wantArgs: []string{"SET", "k-override", string(val1), "EX", "30"}, reply: []byte("+OK\r\n")},
+		{wantArgs: []string{"SET", "k-inherit", string(val2), "EX", "60"}, reply: []byte("+OK\r\n")},
+	})
+	client := newRESPClient(ln.Addr().String(), "", 0)
+	noop := func(string, ...any) {}
+	c := newResponseCache(client, time.Minute, defaultCacheMaxBodyBytes, noop, noop) // global ttl = 60s
+	c.nowFn = time.Now
+
+	overrideGrp := &group{name: "override", cacheTTL: 30 * time.Second}
+	inheritGrp := &group{name: "inherit"} // cacheTTL 0 → inherits c.ttl (60s)
+
+	c.store("k-override", 200, "application/json", []byte(`{"a":1}`), effectiveTTL(c, overrideGrp))
+	c.store("k-inherit", 200, "application/json", []byte(`{"b":2}`), effectiveTTL(c, inheritGrp))
+}

llmgateway.go

@@ -71,7 +71,18 @@ type GroupConfig struct {
 	// group in — but only when the global cache block is actually
 	// configured (newAuthStore rejects true otherwise, a constructor
 	// error, since there is nothing to inherit ttl/maxBodyBytes from).
-	Cache      *bool    `json:"cache,omitempty"`
+	Cache *bool `json:"cache,omitempty"`
+	// CacheTTL overrides the global cache.ttl (a Go duration string, e.g.
+	// "5m") for this group's cache entries only. Empty (the default)
+	// inherits the global TTL. Set means: parsed via time.ParseDuration at
+	// construction (newAuthStore, auth.go); a malformed duration, a
+	// duration <= 0, or CacheTTL set while the global cache block is not
+	// configured (cfg.Cache.Enabled false — there is no global TTL to
+	// override) is a constructor error naming the group. CacheTTL is
+	// independent of Cache above: it names the TTL to use WHEN this
+	// group's requests are cached, whether that caching comes from the
+	// global default or from this group's own Cache:true.
+	CacheTTL   string   `json:"cacheTTL,omitempty"`
 	MCPServers []string `json:"mcpServers,omitempty"`
 	Agents     []string `json:"agents,omitempty"`
 }

routes_unified.go

@@ -210,7 +210,7 @@ func (g *Gateway) runUnified(w http.ResponseWriter, r *http.Request, u *user, gr
 	// but only after being handed a silently truncated body — skip the
 	// call outright instead of ever constructing a corrupt cache entry.
 	if cacheable && callErr == nil && capture.status == http.StatusOK && !capture.oversize {
-		g.cache.store(cacheKeyStr, capture.status, capture.contentType, capture.buf.Bytes())
+		g.cache.store(cacheKeyStr, capture.status, capture.contentType, capture.buf.Bytes(), effectiveTTL(g.cache, grp))
 	}
 
 	if callErr != nil {

routes_unified_test.go

@@ -1002,18 +1002,45 @@ func TestHandleChat_MidStreamDropAfterUsageChunk_AccountsPartialUsage(t *testing
 // resp_test.go's scripted tests already pin.
 type behavioralRedisServer struct {
 	data map[string]string
-	mu   sync.Mutex
+	// lastSetEX is the EX seconds argument from the most recent SET...EX
+	// command this server has handled — item B's per-group cache TTL
+	// override needs a way to observe the actual wire value a real
+	// Gateway.ServeHTTP request produced, not just the reply data's
+	// content. Empty until the first such SET arrives.
+	lastSetEX string
+	mu        sync.Mutex
 }
 
 // newBehavioralRedisServer starts the server on an OS-assigned port,
 // closed automatically at test cleanup (via newFakeListener), and returns
 // its listener.
 func newBehavioralRedisServer(t *testing.T) net.Listener {
+	t.Helper()
+	ln, _ := newBehavioralRedisServerAndHandle(t)
+	return ln
+}
+
+// newBehavioralRedisServerAndHandle is newBehavioralRedisServer, additionally
+// returning the *behavioralRedisServer itself so a caller can inspect state
+// the wire protocol alone does not expose — namely lastSetEX (see setEX
+// below), used by TestHandleChat_CachePerGroupTTL_EndToEnd_DifferentEXPerGroup
+// to prove effectiveTTL's per-group override (cache.go) actually reaches
+// the SET...EX command a real Gateway.ServeHTTP request issues.
+func newBehavioralRedisServerAndHandle(t *testing.T) (net.Listener, *behavioralRedisServer) {
 	t.Helper()
 	ln := newFakeListener(t)
 	srv := &behavioralRedisServer{data: make(map[string]string)}
 	go srv.acceptLoop(ln)
-	return ln
+	return ln, srv
+}
+
+// setEX returns the EX seconds argument from the most recent SET...EX
+// command this server has handled, and whether any such SET has arrived
+// yet.
+func (s *behavioralRedisServer) setEX() (string, bool) {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+	return s.lastSetEX, s.lastSetEX != ""
 }
 
 func (s *behavioralRedisServer) acceptLoop(ln net.Listener) {
@@ -1062,6 +1089,9 @@ func (s *behavioralRedisServer) handle(args []string) []byte {
 	case "SET":
 		s.mu.Lock()
 		s.data[args[1]] = args[2]
+		if len(args) >= 5 && strings.EqualFold(args[3], "EX") {
+			s.lastSetEX = args[4]
+		}
 		s.mu.Unlock()
 		return []byte("+OK\r\n")
 	case "INCRBY":
@@ -1522,6 +1552,79 @@ func TestHandleChat_CacheableRequest_UpstreamDown_502HasNoCacheHeader(t *testing
 	}
 }
 
+// --- per-group cache TTL: end-to-end through a real Gateway (item B) ---
+
+// TestHandleChat_CachePerGroupTTL_EndToEnd_DifferentEXPerGroup is item B's
+// e2e proof: two groups share one global cache.ttl ("1m" = EX 60), one
+// ("override") overrides it to "30s", the other ("inherit") leaves it
+// unset. A cache-miss request from each group's own user must produce a
+// SET...EX command carrying that group's own resolved TTL — 30 for
+// "override", 60 for "inherit" — proving effectiveTTL's per-group
+// resolution (cache.go) actually reaches the wire through the full
+// ServeHTTP path, not just the unit level
+// (TestResponseCache_Store_EffectiveTTLDiffersPerGroup_SETCarriesGroupTTL,
+// cache_test.go). The two requests carry deliberately different bodies —
+// cacheKey has no group/user component (routes_unified.go), so identical
+// bodies from different groups would collide into one cache entry and the
+// second request would be a hit, never issuing a second SET at all.
+func TestHandleChat_CachePerGroupTTL_EndToEnd_DifferentEXPerGroup(t *testing.T) {
+	const respBody = `{"id":"c1","object":"chat.completion","model":"gpt-test","choices":[{"index":0,"message":{"role":"assistant","content":"hi"},"finish_reason":"stop"}],"usage":{"prompt_tokens":1,"completion_tokens":1}}`
+	srv := httptest.NewServer(http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {
+		w.Header().Set("Content-Type", "application/json")
+		w.WriteHeader(http.StatusOK)
+		_, _ = w.Write([]byte(respBody))
+	}))
+	defer srv.Close()
+
+	redisLn, redisSrv := newBehavioralRedisServerAndHandle(t)
+	cfg := CreateConfig()
+	cfg.Providers = map[string]*ProviderConfig{
+		"openai": {Type: "openai", BaseURL: srv.URL, APIKey: "k", Models: []string{"gpt-test"}},
+	}
+	cfg.Groups = map[string]*GroupConfig{
+		"override": {CacheTTL: "30s"},
+		"inherit":  {},
+	}
+	cfg.Users = &UsersConfig{Inline: []*UserConfig{
+		{Name: "bob", Group: "override", APIKey: "sk-bob", Limits: &LimitsConfig{}},
+		{Name: "alice", Group: "inherit", APIKey: "sk-alice", Limits: &LimitsConfig{}},
+	}}
+	cfg.Redis = &RedisConfig{Address: redisLn.Addr().String()}
+	cfg.Cache = CacheConfig{Enabled: true, TTL: "1m", MaxBodyBytes: defaultCacheMaxBodyBytes}
+
+	next := http.HandlerFunc(func(w http.ResponseWriter, r *http.Request) {})
+	h, err := New(context.Background(), next, cfg, "llmgw")
+	if err != nil {
+		t.Fatalf("New: %v", err)
+	}
+
+	// "override" group's user (bob): expect EX 30.
+	overrideBody := map[string]any{"model": "gpt-test", "messages": []any{map[string]any{"role": "user", "content": "override-group"}}}
+	reqOverride := newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-bob", overrideBody)
+	recOverride := httptest.NewRecorder()
+	h.ServeHTTP(recOverride, reqOverride)
+	if recOverride.Code != http.StatusOK {
+		t.Fatalf("override-group request: status = %d, want 200", recOverride.Code)
+	}
+	if ex, ok := redisSrv.setEX(); !ok || ex != "30" {
+		t.Errorf("override-group SET...EX = %q (ok=%v), want %q", ex, ok, "30")
+	}
+
+	// "inherit" group's user (alice), a genuinely different body so it
+	// misses (and issues its own SET) rather than hitting bob's entry:
+	// expect EX 60 (the global 1m).
+	inheritBody := map[string]any{"model": "gpt-test", "messages": []any{map[string]any{"role": "user", "content": "inherit-group"}}}
+	reqInherit := newUnifiedRequest(t, http.MethodPost, "/v1/chat/completions", "sk-alice", inheritBody)
+	recInherit := httptest.NewRecorder()
+	h.ServeHTTP(recInherit, reqInherit)
+	if recInherit.Code != http.StatusOK {
+		t.Fatalf("inherit-group request: status = %d, want 200", recInherit.Code)
+	}
+	if ex, ok := redisSrv.setEX(); !ok || ex != "60" {
+		t.Errorf("inherit-group SET...EX = %q (ok=%v), want %q", ex, ok, "60")
+	}
+}
+
 // --- cacheCaptureWriter: bounded capture (review finding 2) ---
 
 // TestCacheCaptureWriter_BoundsBufferAtMaxBodyBytes_StillWritesFullBodyThrough