fix: fail close provider-scoped routes when no provider endpoints match

config/config.yaml

@@ -1,4 +1,5 @@
 # Olla Configuration (default)
+
 server:
   host: "localhost" # Use 0.0.0.0 to bind to all interfaces (so it's accessible externally)
   port: 40114 # default port for Olla - 4-OLLA
@@ -106,13 +107,14 @@ discovery:
         check_interval: 2s
         check_timeout: 1s
       - url: "http://localhost:8000"
-        name: "local-vllm"
+        name: "oblivion-gb10-vllm"
         type: "vllm"
         priority: 100
         model_url: "/v1/models"
         health_check_url: "/health"
         check_interval: 5s
         check_timeout: 2s
+        circuit_breaker_timeout: 2000s
   model_discovery:
     enabled: true
     interval: 5m
@@ -140,7 +142,7 @@ model_registry:
 
       # Discovery mode settings
       discovery_timeout: 2s  # Timeout for discovery refresh
-      discovery_refresh_on_miss: false  # Refresh discovery when model not found
+      discovery_refresh_on_miss: true  # Refresh discovery when model not found
 
 translators:
   #####
@@ -168,6 +170,8 @@ logging:
 
 engineering:
   show_nerdstats: false
+  endpoint_degraded_success_rate_threshold: 80.0
+  endpoint_degraded_minimum_requests: 10
 
 # Model aliases allow mapping a single virtual model name to multiple actual model names
 # across different backends. When a request arrives with an alias, olla finds endpoints

config/models.yaml

@@ -195,6 +195,12 @@ capabilities:
         - multimodal
         - image-understanding
 
+    - pattern: "(^|[/_-])(bge|e5|gte|minilm|jina-embeddings?|nomic-embed|mxbai-embed|text-embedding)"
+      capabilities:
+        - embeddings
+        - similarity
+        - vector-search
+
   # Context length categories (in tokens)
   context_thresholds:
     extended_context: 32000
@@ -213,4 +219,4 @@ special_rules:
     - unknown
     - test
     - temp
-    - default
+    - default

docs/content/configuration/reference.md

@@ -714,12 +714,16 @@ Debug and development features.
 | Field | Type | Default | Description |
 |-------|------|---------|-------------|
 | `show_nerdstats` | bool | `false` | Show memory stats on shutdown |
+| `endpoint_degraded_success_rate_threshold` | float | `80.0` | Mark endpoints degraded when request success rate drops below this percentage |
+| `endpoint_degraded_minimum_requests` | integer | `10` | Minimum endpoint request count before low success rate degradation is reported |
 
 Example:
 
 ```yaml
 engineering:
   show_nerdstats: true
+  endpoint_degraded_success_rate_threshold: 80.0
+  endpoint_degraded_minimum_requests: 10
 ```
 
 When enabled, displays:
@@ -731,31 +735,54 @@ When enabled, displays:
 
 ## Environment Variables
 
-All configuration can be overridden via environment variables.
-
-Pattern: `OLLA_<SECTION>_<KEY>` in uppercase with underscores.
+Supported environment variables are listed below.
 
 Examples:
 
 ```bash
 # Server settings
 OLLA_SERVER_HOST=0.0.0.0
 OLLA_SERVER_PORT=8080
-OLLA_SERVER_REQUEST_LOGGING=true
+OLLA_SERVER_READ_TIMEOUT=45s
+OLLA_SERVER_WRITE_TIMEOUT=30s
+OLLA_SERVER_MAX_BODY_SIZE=25MB
+OLLA_SERVER_MAX_HEADER_SIZE=2MB
+OLLA_SERVER_GLOBAL_RATE_LIMIT=1000
+OLLA_SERVER_PER_IP_RATE_LIMIT=100
+OLLA_SERVER_RATE_BURST_SIZE=50
+OLLA_SERVER_HEALTH_RATE_LIMIT=1000
+OLLA_SERVER_RATE_CLEANUP_INTERVAL=5m
+OLLA_SERVER_TRUST_PROXY_HEADERS=true
+OLLA_SERVER_TRUSTED_PROXY_CIDRS=127.0.0.0/8,10.0.0.0/8
 
 # Proxy settings
 OLLA_PROXY_ENGINE=olla
 OLLA_PROXY_LOAD_BALANCER=round-robin
 OLLA_PROXY_PROFILE=auto
+OLLA_PROXY_RESPONSE_TIMEOUT=600s
+OLLA_PROXY_READ_TIMEOUT=120s
+OLLA_PROXY_STICKY_SESSIONS_ENABLED=true
 
 # Logging
 OLLA_LOGGING_LEVEL=debug
 OLLA_LOGGING_FORMAT=text
 
-# Rate limits
-OLLA_SERVER_RATE_LIMITS_GLOBAL_REQUESTS_PER_MINUTE=1000
+# Model registry / translators / engine
+OLLA_MODEL_REGISTRY_TYPE=memory
+OLLA_MODEL_UNIFIER_ENABLED=true
+OLLA_MODEL_UNIFIER_CACHE_TTL=10m
+OLLA_TRANSLATORS_ANTHROPIC_ENABLED=true
+OLLA_TRANSLATORS_ANTHROPIC_MAX_MESSAGE_SIZE=10485760
+OLLA_TRANSLATORS_ANTHROPIC_PASSTHROUGH_ENABLED=true
+
+# Operational
+OLLA_SHOW_NERD_STATS=true
+OLLA_CONFIG_FILE=config/config.yaml
 ```
 
+Deprecated or unsupported variables shown in older docs (for example, `OLLA_LOG_LEVEL`, `OLLA_SERVER_REQUEST_LOGGING`,
+and `OLLA_SERVER_RATE_LIMITS_*`) are intentionally unsupported by the current env override path and must not be used.
+
 ## Duration Format
 
 Duration values use Go duration syntax:
@@ -885,4 +912,4 @@ Additionally, Olla's `Validate()` method catches dangerous zero or empty configu
 
 - [Configuration Examples](examples.md) - Common configurations
 - [Best Practices](practices/overview.md) - Production recommendations
-- [Environment Variables](#environment-variables) - Override configuration
+- [Environment Variables](#environment-variables) - Override configuration

internal/adapter/discovery/fc_repository.go

@@ -0,0 +1,138 @@
+package discovery
+
+import (
+	"context"
+	"encoding/json"
+	"fmt"
+	"net/http"
+	"time"
+
+	"github.com/thushan/olla/internal/config"
+	"github.com/thushan/olla/internal/logger"
+)
+
+// fcModelSpec mirrors the Flight Controller ModelSpec fields we care about for
+// building Olla endpoint configs. Only Name and Port are required for routing.
+type fcModelSpec struct {
+	Name      string `json:"name"`
+	Port      int    `json:"port"`
+	Framework string `json:"framework"`
+}
+
+// fcRegistryEntry mirrors the Flight Controller RegistryEntry type returned by
+// GET /registry. We only parse the fields Olla needs; extra fields are ignored.
+type fcRegistryEntry struct {
+	Host   string        `json:"host"`
+	Models []fcModelSpec `json:"models"`
+}
+
+// FCDiscoveryPoller polls the Flight Controller /registry endpoint and reconciles
+// the StaticEndpointRepository on each poll. A full replace-on-poll strategy ensures
+// that endpoints removed from the FC registry are promptly evicted from Olla's rotation,
+// meeting the <30s convergence acceptance criterion for petersimmons1972/instinct#12.
+type FCDiscoveryPoller struct {
+	repo       *StaticEndpointRepository
+	registryURL string
+	logger     logger.StyledLogger
+	client     *http.Client
+}
+
+// NewFCDiscoveryPoller creates a poller that syncs Olla endpoints from FC /registry.
+// registryBaseURL is the FC service base URL, e.g. http://ai-fleet-controller.ai-fleet.svc.cluster.local.
+func NewFCDiscoveryPoller(repo *StaticEndpointRepository, registryBaseURL string, log logger.StyledLogger) *FCDiscoveryPoller {
+	return &FCDiscoveryPoller{
+		repo:        repo,
+		registryURL: registryBaseURL + "/registry",
+		logger:      log,
+		client: &http.Client{
+			Timeout: 10 * time.Second,
+		},
+	}
+}
+
+// Poll fetches the FC registry and reconciles Olla's endpoint set.
+// On FC unreachable: logs a warning, preserves the current endpoint set (fail-open).
+// On success: fully replaces the endpoint set with the FC-derived list.
+func (p *FCDiscoveryPoller) Poll(ctx context.Context) error {
+	req, err := http.NewRequestWithContext(ctx, http.MethodGet, p.registryURL, nil)
+	if err != nil {
+		return fmt.Errorf("fc-discovery: build request: %w", err)
+	}
+
+	resp, err := p.client.Do(req)
+	if err != nil {
+		// Fail-open: preserve the existing endpoint set when FC is unreachable.
+		p.logger.Warn("fc-discovery: FC registry unreachable, preserving existing endpoints", "url", p.registryURL, "error", err)
+		return nil
+	}
+	defer resp.Body.Close()
+
+	if resp.StatusCode != http.StatusOK {
+		p.logger.Warn("fc-discovery: FC registry returned non-200, preserving existing endpoints",
+			"url", p.registryURL, "status", resp.StatusCode)
+		return nil
+	}
+
+	var entries []fcRegistryEntry
+	if err := json.NewDecoder(resp.Body).Decode(&entries); err != nil {
+		return fmt.Errorf("fc-discovery: decode registry response: %w", err)
+	}
+
+	configs := fcEntriesToEndpointConfigs(entries)
+	if err := p.repo.LoadFromConfig(ctx, configs); err != nil {
+		return fmt.Errorf("fc-discovery: load endpoint configs: %w", err)
+	}
+
+	p.logger.Info("fc-discovery: endpoint set reconciled from FC registry",
+		"endpoints", len(configs))
+	return nil
+}
+
+// RunLoop starts a background polling loop. It polls immediately on first call, then
+// at the configured interval. The loop exits when ctx is cancelled.
+func (p *FCDiscoveryPoller) RunLoop(ctx context.Context, interval time.Duration) {
+	// Immediate first poll so Olla is populated before the ticker fires.
+	if err := p.Poll(ctx); err != nil {
+		p.logger.Warn("fc-discovery: initial poll error", "error", err)
+	}
+
+	ticker := time.NewTicker(interval)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			p.logger.Info("fc-discovery: loop stopped")
+			return
+		case <-ticker.C:
+			if err := p.Poll(ctx); err != nil {
+				p.logger.Warn("fc-discovery: poll error", "error", err)
+			}
+		}
+	}
+}
+
+// fcEntriesToEndpointConfigs converts FC registry entries into Olla EndpointConfig slices.
+// Each (host, model) pair becomes one endpoint at http://<host>:<port>.
+// Type is always "openai" because all FC-managed containers serve the OpenAI-compatible API.
+// Priority defaults to 100 (standard Olla default).
+func fcEntriesToEndpointConfigs(entries []fcRegistryEntry) []config.EndpointConfig {
+	defaultPriority := 100
+	var configs []config.EndpointConfig
+
+	for _, entry := range entries {
+		for _, model := range entry.Models {
+			if model.Port == 0 {
+				continue // skip models without a port (incomplete CRD)
+			}
+			cfg := config.EndpointConfig{
+				URL:      fmt.Sprintf("http://%s:%d", entry.Host, model.Port),
+				Name:     fmt.Sprintf("%s-%s", entry.Host, model.Name),
+				Type:     "openai",
+				Priority: &defaultPriority,
+			}
+			configs = append(configs, cfg)
+		}
+	}
+	return configs
+}