Ollama und Persistenz-Patch

.env.example

@@ -12,11 +12,44 @@ BRAIN_STAGING_DIRS=../glpi-ai-knowledgebase/staging
 # Optional read-only audit streams. Multiple files may be comma-separated.
 BRAIN_AGENT_RUNS_FILES=../glpi-ai-agent/data/runs.jsonl
 
-# Local models
+# Batched disk persistence. AI-THINK drafts, the GLPI-KB cache and graph-state
+# are written sequentially at this interval and once more during shutdown.
+BRAIN_PERSIST_INTERVAL=5m
+
+# Ollama pool. OLLAMA_URLS takes precedence over legacy OLLAMA_URL.
 OLLAMA_URL=http://localhost:11434
+OLLAMA_URLS=http://localhost:11434
+OLLAMA_NODE_NAMES=local-gpu
+OLLAMA_NODE_WEIGHTS=1
+OLLAMA_ROUTING_MODE=least_inflight
+OLLAMA_NODE_MAX_INFLIGHT=1
+OLLAMA_NODE_HEALTH_INTERVAL=15s
+OLLAMA_NODE_FAILURE_COOLDOWN=30s
+OLLAMA_NODE_REQUEST_TIMEOUT=8m
+OLLAMA_FAILOVER_ENABLED=true
+# 0 = try all configured nodes.
+OLLAMA_FAILOVER_ATTEMPTS=0
+OLLAMA_REQUIRE_SAME_MODEL_DIGEST=true
+OLLAMA_REQUIRE_EMBEDDING_MODEL=true
 OLLAMA_CHAT_MODEL=qwen3:8b
 OLLAMA_EMBEDDING_MODEL=embeddinggemma
 
+# Optional read-only GLPI Knowledge Base ingest.
+GLPI_KB_ENABLED=false
+GLPI_URL=https://glpi.example.invalid
+GLPI_API_VERSION=v2.3
+GLPI_CLIENT_ID=
+GLPI_CLIENT_SECRET=
+GLPI_USERNAME=
+GLPI_PASSWORD=
+GLPI_ALLOW_INSECURE_HTTP=false
+GLPI_TIMEOUT=20s
+GLPI_KB_PATH=auto
+GLPI_KB_FILTER=
+GLPI_KB_LIMIT=500
+GLPI_KB_SYNC_INTERVAL=10m
+GLPI_KB_SOURCE=GLPI Knowledge Base
+
 # Sequential enrichment
 BRAIN_AUTO_ENRICH=true
 BRAIN_SCAN_INTERVAL=20s