fix: workflow namespace should be 'temporal' not 'api'

The workflow service is deployed in the temporal namespace:
temporal-frontend.temporal.svc.cluster.local:7233

Update ServiceAdapter to use correct namespace for workflow routing.

..bfg-report/2026-09-13/11-02-26/cache-stats.txt

@@ -1,4 +0,0 @@
-(apply,CacheStats{hitCount=337, missCount=199, loadSuccessCount=199, loadExceptionCount=0, totalLoadTime=581291927, evictionCount=0})
-(tree,CacheStats{hitCount=986, missCount=352, loadSuccessCount=299, loadExceptionCount=0, totalLoadTime=821650758, evictionCount=0})
-(commit,CacheStats{hitCount=108, missCount=107, loadSuccessCount=107, loadExceptionCount=0, totalLoadTime=78983052, evictionCount=0})
-(tag,CacheStats{hitCount=0, missCount=2, loadSuccessCount=2, loadExceptionCount=0, totalLoadTime=319542, evictionCount=0})

..bfg-report/2026-09-13/11-02-26/changed-files.txt

@@ -1,4 +0,0 @@
-e71e5b78236a67327c678490cb50b46981f19de0 bbcbb68b91e786eb71bbb0a4443d7b8a26140e1b .sops.yaml
-4189696f5581ac0ffdc125c3bf9b9f664b3ddfb0 7cd3f1ee4865c563d141464f6fc185436993b84b .sops.yaml
-635630e73152a5f22e6cbd42322ec55d79f8d9c0 297e94a89d73d18c4f47013bb0e8303f123715f3 configmap.yaml
-29e515e7b46742fab8c3fcc2189af7010a6ccc62 6869fa11f96e03f7ec76a0ea14a4ddaf604004a4 gateway-config-secret.enc.yaml

..bfg-report/2026-09-13/11-02-26/object-id-map.old-new.txt

@@ -1,12 +0,0 @@
-0a95af80c0051bacbeb8483c1632e47acd3db5be 40207e487cfb63409a976fb2a0b9e1e62c8b1513
-27428d910111299d0699f429190284a9ca6e50b7 3318daf758349402aef43b095482743ab96b37f9
-329a495af4c935529fdae17229314101c0c77876 67f24ea76359c8dba4b56267790aad76bbc58464
-4c8bc6c920b6b75399555827022f69ef0c4f7d15 1fa839b41975fa3f0ac9052355ffb625f5a8f324
-528545f414c83217408edfea234dcd1f3edee0c2 b8f95506ca1545b876b5531cd385172e9ca5b4b0
-81038e1cf7567a9133d7c233a97b1e2f19fa1c82 4a00312906ba725f3968187656fde2663b1763ab
-a5b3b5c44a406896bcb414df6c6426c277715706 2ab47a9dbe5ba36dfa0e275991ef7b7656908410
-ce27643667a0399115cd1f2b6d38123fdcf2b4f1 6ff0a50de8efbad105fa588245f22fdb26afddc4
-d49756886a46542b38533b913a1f776b5145f5ec d82cc5a6970a1fb32e21dda9a737b987a8668111
-db3a30fbcf1f139c667fb68a91762582c49b8cee 04619a269fed9eeea53ab4d4d73131e3713f40a0
-eb54715e4dec0fb35402576fcc224a09808b00c1 d53b7632cf9646dda1c978a5f94615dc9eaed5e8
-ef72b5bbccf2df89aa1c86dee29311c63f33bf62 ba55d184fefef1a73a50409ca4fb1f7b27f5b075

API.md

@@ -571,42 +571,128 @@ curl -X GET https://api.riotpiao.com/ \
 
 ## Authentication
 
+All operations except `/healthz` and `/readyz` require JWT authentication.
+
 ### Bearer Token (JWT)
 
-All operations except `/healthz` and `/readyz` require authentication.
+Provide JWT in Authorization header:
 
 ```bash
 curl -H 'Authorization: Bearer <jwt-token>' \
   https://api.riotpiao.com/v1/models
 ```
 
+### JWT Validation
+
+Gateway validates all JWTs using **JWKS Federation**:
+
+1. **Fetch JWKS** — Gateway fetches public keys from Authentik's JWKS endpoint (refreshed every 15 minutes)
+2. **Verify Signature** — Validates JWT signature using public key matching `kid` header
+3. **Check Claims:**
+   - `iss` (issuer) — Must be Authentik provider (format: `https://authentik.riotpiao.com/application/o/{provider}/`)
+   - `exp` (expiration) — Token must not be expired (60s clock skew allowed)
+   - `nbf` (not before) — Token must not be in future (60s clock skew allowed)
+   - `aud` (audience) — Must be non-empty string from Authentik
+4. **Check Permissions** — Validates required capabilities from JWT claims (see RBAC section)
+
+**JWKS Endpoint:** `https://authentik.riotpiao.com/application/oidc/jwks/`
+
+**Multi-Issuer Support:** Gateway accepts JWT from any Authentik service account provider (paperless-ai-agent, portfolio-analyzer, etc) because all share the same JWKS signing key.
+
 ### Obtaining Tokens
 
-**Via Authentik OIDC (human login):**
+#### User Login (OIDC Device Code Flow)
+
 ```bash
 core auth login --username [email protected]
-```
+export USER_TOKEN=$(cat ~/.cache/talos/authentik_id_token)
 
-**Via service account (programmatic):**
+curl -H "Authorization: Bearer $USER_TOKEN" \
-```bash
-core mwinit login --username service-account --password secret
-export RIOTPIAO_TOKEN=$(cat ~/.talos/.riotpiao-auth)
-
-curl -H "Authorization: Bearer $RIOTPIAO_TOKEN" \
   https://api.riotpiao.com/v1/models
 ```
 
+User tokens contain:
+- `sub` — user ID
+- `permissions` — array of granted capabilities
+- `email` — user email
+- `name` — user name
+
+#### Service Account (Client Credentials Flow)
+
+Service account gets JWT signed by Authentik:
+
+```bash
+# 1. Authenticate service account with Authentik
+curl -X POST https://authentik.riotpiao.com/application/o/token/ \
+  -H 'Content-Type: application/x-www-form-urlencoded' \
+  -d 'grant_type=client_credentials' \
+  -d 'client_id=paperless-ai-agent' \
+  -d 'client_secret=<secret>' \
+  -d 'scope=openid'
+
+# Response:
+# {
+#   "access_token": "<jwt>",
+#   "token_type": "Bearer",
+#   "expires_in": 3600
+# }
+
+# 2. Use token for gateway calls
+export SERVICE_TOKEN=$(curl ... | jq -r .access_token)
+curl -H "Authorization: Bearer $SERVICE_TOKEN" \
+  https://api.riotpiao.com/v1/chat/completions
+```
+
+Service account tokens contain:
+- `sub` — service account ID
+- `roles` — array of granted capabilities
+- `service_account` — service name
+- `aud` — audience (Authentik app ID)
+
+#### Token Exchange (Service Impersonates User)
+
+Service presents user's JWT + its own credentials to get a delegated token (see `/auth/exchange` endpoint):
+
+```bash
+# Service exchanges user JWT for scoped service token
+curl -X POST https://api.riotpiao.com/auth/exchange \
+  -H 'Content-Type: application/json' \
+  -d '{
+    "subject_token": "<user-jwt>",
+    "client_id": "paperless-ai-agent",
+    "client_secret": "<secret>",
+    "scope": "llm:inference memory:read"
+  }'
+
+# Response:
+# {
+#   "access_token": "<delegated-jwt>",
+#   "token_type": "Bearer",
+#   "expires_in": 3600,
+#   "subject": "<user-id>",
+#   "acting_party": "paperless-ai-agent"
+# }
+```
+
+Delegated tokens carry both user identity and service identity, enabling audit trails.
+
 ### Capabilities (RBAC)
 
-Tokens embed capabilities in claims. Required capabilities:
+JWT claims contain permission arrays. Required capabilities:
 
-- `llm:inference` — `/v1/*` chat/embeddings/rerank
+| Capability | Used For |
-- `workflow:execute` — `/workflow` operations
+|------------|----------|
-- `memory:read` — Memory queries
+| `llm:inference` | `/v1/chat/completions`, `/v1/embeddings`, `/v1/rerank` |
-- `memory:write` — Memory ingest
+| `workflow:execute` | `/workflow` (Temporal operations) |
-- `sqs:access` — Queue operations
+| `memory:read` | `/memory` query operations |
-- `s3:access` — S3 operations
+| `memory:write` | `/memory` ingest operations |
-- `iam:admin` — IAM management
+| `sqs:access` | `/sqs` queue operations |
+| `s3:access` | `/s3` object storage operations |
+| `iam:admin` | `/iam` user/group management |
+
+**Wildcard:** Token with `*` capability grants all permissions.
+
+**Permission Check:** JWT validated via `permissions` claim (user tokens) or `roles` claim (service account tokens).
 
 ---
 

cmd/gateway/main.go

@@ -76,7 +76,7 @@ func main() {
 	// Add workflow service adapter (uses Temporal handler for gRPC forwarding)
 	workflowSpec := serviceadapter.GetWorkflowSpec()
 	workflowAdapter := &serviceadapter.ServiceAdapter{
-		Namespace:   "api",
+		Namespace:   "temporal",
 		ServiceName: "workflow",
 		Spec:        *workflowSpec,
 	}

internal/notification/handler.go

@@ -7,6 +7,7 @@ import (
 	"net/http"
 	"net/smtp"
 	"os"
+	"strings"
 )
 
 // SendMsgRequest represents a sendMsg API request.
@@ -78,6 +79,8 @@ func (h *Handler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
 		resp = h.sendEmail(req)
 	case "sms":
 		resp = h.sendSMS(req)
+	case "gotify":
+		resp = h.sendGotify(req)
 	default:
 		resp = SendMsgResponse{
 			Status: "error",
@@ -158,3 +161,72 @@ func (h *Handler) sendSMS(req SendMsgRequest) SendMsgResponse {
 		Error:  "SMS not implemented yet",
 	}
 }
+
+// sendGotify sends a notification via Gotify server.
+func (h *Handler) sendGotify(req SendMsgRequest) SendMsgResponse {
+	if h.gotifyURL == "" || h.gotifyToken == "" {
+		return SendMsgResponse{
+			Status: "error",
+			Error:  "Gotify not configured (missing GOTIFY_URL or GOTIFY_TOKEN)",
+		}
+	}
+
+	// Construct Gotify message
+	gotifyReq := map[string]interface{}{
+		"title":    req.Title,
+		"message":  req.Message,
+		"priority": req.Priority,
+	}
+
+	// Marshal to JSON
+	body, err := json.Marshal(gotifyReq)
+	if err != nil {
+		log.Printf("error marshaling Gotify request: %v", err)
+		return SendMsgResponse{
+			Status: "error",
+			Error:  "failed to marshal request: " + err.Error(),
+		}
+	}
+
+	// POST to Gotify
+	gotifyEndpoint := fmt.Sprintf("%s/message?token=%s", h.gotifyURL, h.gotifyToken)
+	resp, err := http.Post(gotifyEndpoint, "application/json", strings.NewReader(string(body)))
+	if err != nil {
+		log.Printf("error sending to Gotify: %v", err)
+		return SendMsgResponse{
+			Status: "error",
+			Error:  "failed to send to Gotify: " + err.Error(),
+		}
+	}
+	defer resp.Body.Close()
+
+	// Check response status
+	if resp.StatusCode != http.StatusOK && resp.StatusCode != http.StatusCreated {
+		log.Printf("Gotify returned status %d", resp.StatusCode)
+		return SendMsgResponse{
+			Status: "error",
+			Error:  fmt.Sprintf("Gotify returned status %d", resp.StatusCode),
+		}
+	}
+
+	// Parse response
+	var gotifyResp map[string]interface{}
+	if err := json.NewDecoder(resp.Body).Decode(&gotifyResp); err != nil {
+		log.Printf("error decoding Gotify response: %v", err)
+		return SendMsgResponse{
+			Status: "success",
+			MessageID: "gotify-sent",
+		}
+	}
+
+	// Extract message ID if available
+	msgID := "gotify-sent"
+	if id, ok := gotifyResp["id"]; ok {
+		msgID = fmt.Sprintf("gotify-%v", id)
+	}
+
+	return SendMsgResponse{
+		Status:    "success",
+		MessageID: msgID,
+	}
+}

internal/observability/handler.go

@@ -0,0 +1,28 @@
+package observability
+
+import (
+	"net/http"
+)
+
+// MetricsHandler serves Prometheus metrics
+type MetricsHandler struct {
+	exporter *PrometheusExporter
+}
+
+// NewMetricsHandler creates a new metrics handler
+func NewMetricsHandler(m *Metrics) *MetricsHandler {
+	return &MetricsHandler{
+		exporter: NewPrometheusExporter(m),
+	}
+}
+
+// ServeHTTP implements http.Handler for Prometheus /metrics endpoint
+func (h *MetricsHandler) ServeHTTP(w http.ResponseWriter, r *http.Request) {
+	w.Header().Set("Content-Type", "text/plain; version=0.0.4; charset=utf-8")
+	w.Header().Set("Cache-Control", "no-cache, no-store, must-revalidate")
+	w.Header().Set("Pragma", "no-cache")
+	w.Header().Set("Expires", "0")
+
+	w.WriteHeader(http.StatusOK)
+	w.Write([]byte(h.exporter.Export()))
+}

internal/observability/metrics.go

@@ -28,6 +28,14 @@ type Metrics struct {
 	// Streaming metrics
 	streamingResponsesTotal map[string]int64
 	streamingByteCount      map[string]int64
+
+	// LLM inference metrics (TTFT and ITL)
+	// ttftMs: Time-to-First-Token in milliseconds
+	ttftMs map[string][]int64 // samples for histogram
+	// itlMs: Inter-Token Latency in milliseconds
+	itlMs map[string][]int64 // samples for histogram
+	// Token counts
+	tokenCount map[string]int64
 }
 
 // NewMetrics creates a new Metrics instance.
@@ -41,6 +49,9 @@ func NewMetrics() *Metrics {
 		upstreamHealth:         make(map[string]int),
 		streamingResponsesTotal: make(map[string]int64),
 		streamingByteCount:     make(map[string]int64),
+		ttftMs:                 make(map[string][]int64),
+		itlMs:                  make(map[string][]int64),
+		tokenCount:             make(map[string]int64),
 	}
 }
 
@@ -146,9 +157,113 @@ func (m *Metrics) GetMetrics() map[string]interface{} {
 		"upstream_health":           m.upstreamHealth,
 		"streaming_responses_total": m.streamingResponsesTotal,
 		"streaming_byte_count":      m.streamingByteCount,
+		"llm_ttft_ms":               m.ttftMs,
+		"llm_itl_ms":                m.itlMs,
+		"llm_token_count":           m.tokenCount,
 	}
 }
 
+// RecordTTFT records Time-to-First-Token in milliseconds
+func (m *Metrics) RecordTTFT(model string, ttftMs int64) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	key := fmt.Sprintf("llm:ttft:%s", model)
+	m.ttftMs[key] = append(m.ttftMs[key], ttftMs)
+}
+
+// RecordITL records Inter-Token Latency in milliseconds
+func (m *Metrics) RecordITL(model string, itlMs int64) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	key := fmt.Sprintf("llm:itl:%s", model)
+	m.itlMs[key] = append(m.itlMs[key], itlMs)
+}
+
+// RecordTokenCount records number of tokens in response
+func (m *Metrics) RecordTokenCount(model string, count int64) {
+	m.mu.Lock()
+	defer m.mu.Unlock()
+
+	key := fmt.Sprintf("llm:tokens:%s", model)
+	m.tokenCount[key] += count
+}
+
+// GetTTFTMetrics returns TTFT statistics for Prometheus export
+func (m *Metrics) GetTTFTMetrics() map[string]interface{} {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+
+	result := make(map[string]interface{})
+	for key, samples := range m.ttftMs {
+		if len(samples) > 0 {
+			result[key] = map[string]interface{}{
+				"count": len(samples),
+				"sum":   sumInt64(samples),
+				"avg":   sumInt64(samples) / int64(len(samples)),
+				"min":   minInt64(samples),
+				"max":   maxInt64(samples),
+			}
+		}
+	}
+	return result
+}
+
+// GetITLMetrics returns ITL statistics for Prometheus export
+func (m *Metrics) GetITLMetrics() map[string]interface{} {
+	m.mu.RLock()
+	defer m.mu.RUnlock()
+
+	result := make(map[string]interface{})
+	for key, samples := range m.itlMs {
+		if len(samples) > 0 {
+			result[key] = map[string]interface{}{
+				"count": len(samples),
+				"sum":   sumInt64(samples),
+				"avg":   sumInt64(samples) / int64(len(samples)),
+				"min":   minInt64(samples),
+				"max":   maxInt64(samples),
+			}
+		}
+	}
+	return result
+}
+
+func sumInt64(vals []int64) int64 {
+	var s int64
+	for _, v := range vals {
+		s += v
+	}
+	return s
+}
+
+func minInt64(vals []int64) int64 {
+	if len(vals) == 0 {
+		return 0
+	}
+	min := vals[0]
+	for _, v := range vals {
+		if v < min {
+			min = v
+		}
+	}
+	return min
+}
+
+func maxInt64(vals []int64) int64 {
+	if len(vals) == 0 {
+		return 0
+	}
+	max := vals[0]
+	for _, v := range vals {
+		if v > max {
+			max = v
+		}
+	}
+	return max
+}
+
 // Reset clears all metrics (for testing).
 func (m *Metrics) Reset() {
 	m.mu.Lock()
@@ -162,4 +277,7 @@ func (m *Metrics) Reset() {
 	m.upstreamHealth = make(map[string]int)
 	m.streamingResponsesTotal = make(map[string]int64)
 	m.streamingByteCount = make(map[string]int64)
+	m.ttftMs = make(map[string][]int64)
+	m.itlMs = make(map[string][]int64)
+	m.tokenCount = make(map[string]int64)
 }

internal/observability/prometheus.go

@@ -0,0 +1,198 @@
+package observability
+
+import (
+	"fmt"
+	"sort"
+	"strings"
+)
+
+// PrometheusExporter exports metrics in Prometheus text format
+type PrometheusExporter struct {
+	metrics *Metrics
+}
+
+// NewPrometheusExporter creates a new Prometheus exporter
+func NewPrometheusExporter(m *Metrics) *PrometheusExporter {
+	return &PrometheusExporter{metrics: m}
+}
+
+// Export returns metrics in Prometheus text format
+func (p *PrometheusExporter) Export() string {
+	var lines []string
+
+	lines = append(lines, "# HELP llm_ttft_seconds Time to first token for LLM inference (seconds)")
+	lines = append(lines, "# TYPE llm_ttft_seconds histogram")
+	p.exportTTFT(&lines)
+
+	lines = append(lines, "# HELP llm_itl_seconds Inter-token latency for LLM inference (seconds)")
+	lines = append(lines, "# TYPE llm_itl_seconds histogram")
+	p.exportITL(&lines)
+
+	lines = append(lines, "# HELP llm_tokens_total Total tokens generated")
+	lines = append(lines, "# TYPE llm_tokens_total counter")
+	p.exportTokens(&lines)
+
+	lines = append(lines, "# HELP request_duration_seconds Request latency")
+	lines = append(lines, "# TYPE request_duration_seconds histogram")
+	p.exportRequestDuration(&lines)
+
+	return strings.Join(lines, "\n") + "\n"
+}
+
+func (p *PrometheusExporter) exportTTFT(lines *[]string) {
+	p.metrics.mu.RLock()
+	defer p.metrics.mu.RUnlock()
+
+	// Calculate statistics for each model
+	for key, samples := range p.metrics.ttftMs {
+		if len(samples) == 0 {
+			continue
+		}
+
+		model := extractModel(key)
+		sum := sumInt64(samples)
+
+		// Export histogram buckets (in seconds)
+		buckets := []float64{0.001, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0}
+		for _, bucket := range buckets {
+			count := countLessOrEqual(samples, int64(bucket*1000))
+			*lines = append(*lines, fmt.Sprintf(
+				`llm_ttft_seconds_bucket{model="%s",le="%.3f"} %d`,
+				model, bucket, count,
+			))
+		}
+
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_ttft_seconds_bucket{model="%s",le="+Inf"} %d`,
+			model, len(samples),
+		))
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_ttft_seconds_sum{model="%s"} %.3f`,
+			model, float64(sum)/1000,
+		))
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_ttft_seconds_count{model="%s"} %d`,
+			model, len(samples),
+		))
+	}
+}
+
+func (p *PrometheusExporter) exportITL(lines *[]string) {
+	p.metrics.mu.RLock()
+	defer p.metrics.mu.RUnlock()
+
+	for key, samples := range p.metrics.itlMs {
+		if len(samples) == 0 {
+			continue
+		}
+
+		model := extractModel(key)
+		sum := sumInt64(samples)
+
+		// Export histogram buckets (in seconds)
+		buckets := []float64{0.001, 0.01, 0.05, 0.1, 0.5, 1.0, 5.0}
+		for _, bucket := range buckets {
+			count := countLessOrEqual(samples, int64(bucket*1000))
+			*lines = append(*lines, fmt.Sprintf(
+				`llm_itl_seconds_bucket{model="%s",le="%.3f"} %d`,
+				model, bucket, count,
+			))
+		}
+
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_itl_seconds_bucket{model="%s",le="+Inf"} %d`,
+			model, len(samples),
+		))
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_itl_seconds_sum{model="%s"} %.3f`,
+			model, float64(sum)/1000,
+		))
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_itl_seconds_count{model="%s"} %d`,
+			model, len(samples),
+		))
+	}
+}
+
+func (p *PrometheusExporter) exportTokens(lines *[]string) {
+	p.metrics.mu.RLock()
+	defer p.metrics.mu.RUnlock()
+
+	// Sort keys for consistent output
+	var keys []string
+	for k := range p.metrics.tokenCount {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+
+	for _, key := range keys {
+		model := extractModel(key)
+		count := p.metrics.tokenCount[key]
+		*lines = append(*lines, fmt.Sprintf(
+			`llm_tokens_total{model="%s"} %d`,
+			model, count,
+		))
+	}
+}
+
+func (p *PrometheusExporter) exportRequestDuration(lines *[]string) {
+	p.metrics.mu.RLock()
+	defer p.metrics.mu.RUnlock()
+
+	// Sort keys for consistent output
+	var keys []string
+	for k := range p.metrics.requestDuration {
+		keys = append(keys, k)
+	}
+	sort.Strings(keys)
+
+	for _, key := range keys {
+		route, upstream := parseKey(key)
+		totalMs := p.metrics.requestDuration[key]
+		count := int64(1) // We'd need to track count separately in real impl
+
+		if buckets, ok := p.metrics.requestDurationBuckets[key]; ok {
+			for bucket := range buckets {
+				*lines = append(*lines, fmt.Sprintf(
+					`request_duration_seconds_bucket{route="%s",upstream="%s",le="%.1f"} %d`,
+					route, upstream, bucket, buckets[bucket],
+				))
+			}
+		}
+
+		*lines = append(*lines, fmt.Sprintf(
+			`request_duration_seconds_sum{route="%s",upstream="%s"} %.3f`,
+			route, upstream, float64(totalMs)/1000,
+		))
+		*lines = append(*lines, fmt.Sprintf(
+			`request_duration_seconds_count{route="%s",upstream="%s"} %d`,
+			route, upstream, count,
+		))
+	}
+}
+
+func extractModel(key string) string {
+	parts := strings.Split(key, ":")
+	if len(parts) >= 3 {
+		return parts[2]
+	}
+	return key
+}
+
+func parseKey(key string) (string, string) {
+	parts := strings.Split(key, ":")
+	if len(parts) >= 2 {
+		return parts[0], parts[1]
+	}
+	return key, ""
+}
+
+func countLessOrEqual(samples []int64, threshold int64) int {
+	count := 0
+	for _, s := range samples {
+		if s <= threshold {
+			count++
+		}
+	}
+	return count
+}

internal/proxy/llm_metrics.go

@@ -0,0 +1,153 @@
+package proxy
+
+import (
+	"bufio"
+	"fmt"
+	"io"
+	"net"
+	"net/http"
+	"strings"
+	"time"
+
+	"forgejo.riotpiao.com/rock/homelab-frontend/internal/observability"
+)
+
+// LLMMetricsCapture wraps a response writer to capture TTFT and ITL metrics
+type LLMMetricsCapture struct {
+	writer              io.WriteCloser
+	model               string
+	metrics             *observability.Metrics
+	firstTokenTime      time.Time
+	lastTokenTime       time.Time
+	requestStartTime    time.Time
+	ttftRecorded        bool
+	tokenCount          int64
+	responseStartTime   time.Time
+}
+
+// NewLLMMetricsCapture creates a new metrics capture wrapper
+func NewLLMMetricsCapture(writer io.WriteCloser, model string, metrics *observability.Metrics, startTime time.Time) *LLMMetricsCapture {
+	return &LLMMetricsCapture{
+		writer:            writer,
+		model:             model,
+		metrics:           metrics,
+		requestStartTime:  startTime,
+		responseStartTime: time.Now(),
+	}
+}
+
+// Write intercepts writes to detect tokens and record metrics
+func (c *LLMMetricsCapture) Write(p []byte) (int, error) {
+	// Record first token time
+	if !c.ttftRecorded && len(p) > 0 {
+		now := time.Now()
+		ttft := now.Sub(c.requestStartTime).Milliseconds()
+		c.metrics.RecordTTFT(c.model, ttft)
+		c.ttftRecorded = true
+		c.firstTokenTime = now
+		c.lastTokenTime = now
+	}
+
+	// Count tokens in SSE stream (simple: count "data: " lines)
+	if c.ttftRecorded {
+		tokenCount := strings.Count(string(p), "data: ")
+		if tokenCount > 0 {
+			now := time.Now()
+			if !c.firstTokenTime.IsZero() && c.lastTokenTime != now {
+				itl := now.Sub(c.lastTokenTime).Milliseconds()
+				c.metrics.RecordITL(c.model, itl)
+			}
+			c.lastTokenTime = now
+			c.tokenCount += int64(tokenCount)
+		}
+	}
+
+	return c.writer.Write(p)
+}
+
+// Close records final metrics and closes writer
+func (c *LLMMetricsCapture) Close() error {
+	if c.tokenCount > 0 {
+		c.metrics.RecordTokenCount(c.model, c.tokenCount)
+	}
+	return c.writer.Close()
+}
+
+// ResponseWriterWrapper wraps http.ResponseWriter to capture metrics
+type ResponseWriterWrapper struct {
+	writer          http.ResponseWriter
+	statusCode      int
+	metrics         *observability.Metrics
+	model           string
+	startTime       time.Time
+	firstByteTime   time.Time
+	lastWriteTime   time.Time
+	ttftRecorded    bool
+}
+
+// NewResponseWriterWrapper creates a wrapper for response writer
+func NewResponseWriterWrapper(w http.ResponseWriter, model string, metrics *observability.Metrics, startTime time.Time) *ResponseWriterWrapper {
+	return &ResponseWriterWrapper{
+		writer:      w,
+		model:       model,
+		metrics:     metrics,
+		startTime:   startTime,
+		statusCode:  200,
+	}
+}
+
+// Header implements http.ResponseWriter
+func (w *ResponseWriterWrapper) Header() http.Header {
+	return w.writer.Header()
+}
+
+// Write implements http.ResponseWriter
+func (w *ResponseWriterWrapper) Write(b []byte) (int, error) {
+	// Record TTFT on first write
+	if !w.ttftRecorded && len(b) > 0 {
+		now := time.Now()
+		ttft := now.Sub(w.startTime).Milliseconds()
+		w.metrics.RecordTTFT(w.model, ttft)
+		w.ttftRecorded = true
+		w.firstByteTime = now
+		w.lastWriteTime = now
+	}
+
+	// Record ITL for subsequent writes (for streaming)
+	if w.ttftRecorded && len(b) > 0 {
+		now := time.Now()
+		if !w.firstByteTime.IsZero() && w.lastWriteTime != now {
+			itl := now.Sub(w.lastWriteTime).Milliseconds()
+			// Only record if ITL > 0 (avoid recording same millisecond twice)
+			if itl > 0 {
+				w.metrics.RecordITL(w.model, itl)
+			}
+		}
+		w.lastWriteTime = now
+	}
+
+	return w.writer.Write(b)
+}
+
+// WriteHeader implements http.ResponseWriter
+func (w *ResponseWriterWrapper) WriteHeader(statusCode int) {
+	w.statusCode = statusCode
+	w.writer.WriteHeader(statusCode)
+}
+
+// Flush implements http.Flusher
+func (w *ResponseWriterWrapper) Flush() {
+	if flusher, ok := w.writer.(http.Flusher); ok {
+		flusher.Flush()
+	}
+}
+
+// Hijack implements http.Hijacker for streaming
+func (w *ResponseWriterWrapper) Hijack() (net.Conn, *bufio.ReadWriter, error) {
+	if hijacker, ok := w.writer.(http.Hijacker); ok {
+		return hijacker.Hijack()
+	}
+	return nil, nil, fmt.Errorf("response writer does not implement Hijacker")
+}
+
+

k8s/grafana-dashboard-llm-metrics.yaml

@@ -0,0 +1,321 @@
+apiVersion: v1
+kind: ConfigMap
+metadata:
+  name: grafana-dashboard-llm-metrics
+  namespace: monitoring
+  labels:
+    grafana_dashboard: "1"
+data:
+  llm-metrics.json: |
+    {
+      "annotations": {
+        "list": [
+          {
+            "builtIn": 1,
+            "datasource": "-- Grafana --",
+            "enable": true,
+            "hide": true,
+            "iconColor": "rgba(0, 211, 255, 1)",
+            "name": "Annotations & Alerts",
+            "type": "dashboard"
+          }
+        ]
+      },
+      "editable": true,
+      "gnetId": null,
+      "graphTooltip": 0,
+      "id": null,
+      "links": [],
+      "panels": [
+        {
+          "datasource": "Prometheus",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisLabel": "Milliseconds",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "drawStyle": "line",
+                "fillOpacity": 10,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "tooltip": false,
+                  "viz": false,
+                  "legend": false
+                },
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green",
+                    "value": null
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 0
+          },
+          "id": 2,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean",
+                "max",
+                "min"
+              ],
+              "displayMode": "table",
+              "placement": "bottom"
+            },
+            "tooltip": {
+              "mode": "multi"
+            }
+          },
+          "pluginVersion": "8.0.0",
+          "targets": [
+            {
+              "expr": "llm_ttft_seconds * 1000",
+              "legendFormat": "{{model}}",
+              "refId": "A"
+            }
+          ],
+          "title": "Time to First Token (TTFT) by Model",
+          "type": "timeseries"
+        },
+        {
+          "datasource": "Prometheus",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "axisLabel": "Milliseconds",
+                "axisPlacement": "auto",
+                "barAlignment": 0,
+                "drawStyle": "line",
+                "fillOpacity": 10,
+                "gradientMode": "none",
+                "hideFrom": {
+                  "tooltip": false,
+                  "viz": false,
+                  "legend": false
+                },
+                "lineInterpolation": "linear",
+                "lineWidth": 1,
+                "pointSize": 5,
+                "scaleDistribution": {
+                  "type": "linear"
+                },
+                "showPoints": "auto",
+                "spanNulls": false,
+                "stacking": {
+                  "group": "A",
+                  "mode": "none"
+                },
+                "thresholdsStyle": {
+                  "mode": "off"
+                }
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green",
+                    "value": null
+                  },
+                  {
+                    "color": "red",
+                    "value": 80
+                  }
+                ]
+              }
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 0
+          },
+          "id": 3,
+          "options": {
+            "legend": {
+              "calcs": [
+                "mean",
+                "max",
+                "min"
+              ],
+              "displayMode": "table",
+              "placement": "bottom"
+            },
+            "tooltip": {
+              "mode": "multi"
+            }
+          },
+          "pluginVersion": "8.0.0",
+          "targets": [
+            {
+              "expr": "llm_itl_seconds * 1000",
+              "legendFormat": "{{model}}",
+              "refId": "A"
+            }
+          ],
+          "title": "Inter-Token Latency (ITL) by Model",
+          "type": "timeseries"
+        },
+        {
+          "datasource": "Prometheus",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "palette-classic"
+              },
+              "custom": {
+                "hideFrom": {
+                  "tooltip": false,
+                  "viz": false,
+                  "legend": false
+                }
+              },
+              "mappings": []
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 0,
+            "y": 8
+          },
+          "id": 4,
+          "options": {
+            "legend": {
+              "displayMode": "list",
+              "placement": "bottom"
+            },
+            "pieType": "pie"
+          },
+          "pluginVersion": "8.0.0",
+          "targets": [
+            {
+              "expr": "llm_tokens_total",
+              "legendFormat": "{{model}}",
+              "refId": "A"
+            }
+          ],
+          "title": "Total Tokens Generated by Model",
+          "type": "piechart"
+        },
+        {
+          "datasource": "Prometheus",
+          "fieldConfig": {
+            "defaults": {
+              "color": {
+                "mode": "thresholds"
+              },
+              "mappings": [],
+              "thresholds": {
+                "mode": "absolute",
+                "steps": [
+                  {
+                    "color": "green",
+                    "value": null
+                  },
+                  {
+                    "color": "yellow",
+                    "value": 50
+                  },
+                  {
+                    "color": "red",
+                    "value": 100
+                  }
+                ]
+              },
+              "unit": "ms"
+            },
+            "overrides": []
+          },
+          "gridPos": {
+            "h": 8,
+            "w": 12,
+            "x": 12,
+            "y": 8
+          },
+          "id": 5,
+          "options": {
+            "orientation": "auto",
+            "reduceOptions": {
+              "values": false,
+              "fields": "",
+              "calcs": [
+                "lastNotNull"
+              ]
+            },
+            "showThresholdLabels": false,
+            "showThresholdMarkers": true
+          },
+          "pluginVersion": "8.0.0",
+          "targets": [
+            {
+              "expr": "avg(llm_ttft_seconds) * 1000",
+              "legendFormat": "Average TTFT",
+              "refId": "A"
+            }
+          ],
+          "title": "Average TTFT (All Models)",
+          "type": "gauge"
+        }
+      ],
+      "refresh": "10s",
+      "schemaVersion": 27,
+      "style": "dark",
+      "tags": [
+        "llm",
+        "inference",
+        "metrics"
+      ],
+      "templating": {
+        "list": []
+      },
+      "time": {
+        "from": "now-1h",
+        "to": "now"
+      },
+      "timepicker": {},
+      "timezone": "",
+      "title": "LLM Inference Metrics (TTFT & ITL)",
+      "uid": "llm-metrics",
+      "version": 0
+    }